Add files using upload-large-folder tool

ae_pipeline_review_004__long/_cua_gym_vm_bridge.sh

@@ -0,0 +1,106 @@
+#!/bin/bash
+# CUA-Gym VM-side network fix (direct-IP mock access).
+#
+# Runs INSIDE the QEMU guest. The CUA-Gym mock services are now reached by
+# their real IDC address (http://28.7.184.198:81xx) instead of the old
+# localhost:9000-9097 socat bridge, so the VM only needs two fixes:
+#
+#   1. Working DNS — the qcow2-baked /etc/resolv.conf points at the Pod's
+#      cluster DNS, which is unreachable from QEMU SLIRP NAT
+#      ("Temporary failure in name resolution").
+#   2. The mock host excluded from the woa http proxy — otherwise the VM's
+#      Chrome (gsettings proxy) and initial_setup.py's `requests` get a 403
+#      from the proxy when hitting 28.7.184.198:81xx. (The old socat path never
+#      hit this because localhost is bypassed by default.)
+#
+# Idempotent. Scoped to the CUA-Gym bridge step; does not touch the generic
+# OSWorld setup pipeline.
+#
+# Argv:
+#   $1  optional sudo password (default "password"; SetupController
+#       substitutes {CLIENT_PASSWORD} when uploaded as a cache file).
+#   $2  unused (kept for backward compat — was socat first port 9000).
+#   $3  unused (kept for backward compat — was socat last  port 9097).
+#   $4  optional mock host IP (default 28.7.184.198).
+
+set +e
+
+PASS="${1:-password}"
+MOCK_HOST="${4:-28.7.184.198}"
+# Derive the /16 IDC subnet so future port/host tweaks on the same segment
+# stay bypassed too (28.7.184.198 -> 28.7.0.0/16).
+MOCK_NET="$(echo "${MOCK_HOST}" | awk -F. 'NF==4{print $1"."$2".0.0/16"}')"
+
+run_root() {
+    echo "$PASS" | sudo -S -p "" bash -c "$1"
+}
+
+# Detect SLIRP gateway from VM's routing table; fall back to 10.0.2.2.
+GW="$(ip route 2>/dev/null | awk '/default/ {print $3; exit}')"
+GW="${GW:-10.0.2.2}"
+
+# ─── CUA-Gym network fix (Taiji-specific) ─────────────────────────────
+fix_vm_network() {
+    # 0) Discover VM network state at runtime — gateway, own IP/CIDR, SLIRP
+    #    DNS forwarder. Nothing hardcoded except the mock host above.
+    local GW_NET="${GW%.*}"
+    local SLIRP_DNS="${GW_NET}.3"        # QEMU SLIRP DNS proxy is always .3
+    local OWN_CIDR
+    OWN_CIDR="$(ip -4 -o addr show scope global 2>/dev/null \
+                | awk '{print $4; exit}')"
+    local OWN_IP="${OWN_CIDR%/*}"
+    local OWN_NET=""
+    [ -n "${OWN_IP}" ] && OWN_NET="${OWN_IP%.*}.0/24"
+
+    # 1) DNS — point at SLIRP DNS proxy + public fallbacks. chattr +i keeps
+    #    systemd-resolved/NetworkManager from clobbering it.
+    if run_root "printf 'nameserver ${SLIRP_DNS}\\nnameserver 114.114.114.114\\nnameserver 8.8.8.8\\n' > /etc/resolv.conf
+                 chattr +i /etc/resolv.conf 2>/dev/null || true"; then
+        echo "[cua_gym_vm_bridge] DNS set: ${SLIRP_DNS}, 114.114.114.114, 8.8.8.8 (gw=${GW})"
+    else
+        echo "[cua_gym_vm_bridge] WARN: failed to write /etc/resolv.conf"
+    fi
+
+    # 2) Build the proxy bypass list from detected values + the direct-IP
+    #    mock host (and its /16 subnet). THIS is what makes 28.7.184.198:81xx
+    #    reachable from inside the VM — without it the woa proxy 403s it.
+    local ignore_list="['localhost', '127.0.0.0/8', '${GW}'"
+    [ -n "${OWN_IP}" ]    && ignore_list="${ignore_list}, '${OWN_IP}'"
+    [ -n "${OWN_NET}" ]   && ignore_list="${ignore_list}, '${OWN_NET}'"
+    [ -n "${GW_NET}" ]    && ignore_list="${ignore_list}, '${GW_NET}.0/24'"
+    [ -n "${MOCK_HOST}" ] && ignore_list="${ignore_list}, '${MOCK_HOST}'"
+    [ -n "${MOCK_NET}" ]  && ignore_list="${ignore_list}, '${MOCK_NET}'"
+    ignore_list="${ignore_list}]"
+
+    # 3) GNOME proxy ignore-hosts — Chrome bypasses the woa proxy for these.
+    if gsettings set org.gnome.system.proxy ignore-hosts "${ignore_list}" 2>/tmp/gsettings_err; then
+        echo "[cua_gym_vm_bridge] gsettings ignore-hosts = ${ignore_list}"
+    else
+        echo "[cua_gym_vm_bridge] WARN: gsettings ignore-hosts failed: $(cat /tmp/gsettings_err 2>/dev/null)"
+    fi
+
+    # 4) Shell env — comma-separated form for no_proxy/NO_PROXY (requests/curl).
+    local NP="localhost,127.0.0.1,${GW}"
+    [ -n "${OWN_IP}" ]    && NP="${NP},${OWN_IP}"
+    [ -n "${OWN_NET}" ]   && NP="${NP},${OWN_NET}"
+    [ -n "${GW_NET}" ]    && NP="${NP},${GW_NET}.0/24"
+    [ -n "${MOCK_HOST}" ] && NP="${NP},${MOCK_HOST}"
+    [ -n "${MOCK_NET}" ]  && NP="${NP},${MOCK_NET}"
+
+    for var in no_proxy NO_PROXY; do
+        if ! grep -q "export ${var}=${NP}" /home/user/.bashrc 2>/dev/null; then
+            echo "export ${var}=${NP}" >> /home/user/.bashrc
+        fi
+    done
+    run_root "grep -q '^no_proxy=' /etc/environment 2>/dev/null \
+              && sed -i 's|^no_proxy=.*|no_proxy=\"${NP}\"|' /etc/environment \
+              || echo 'no_proxy=\"${NP}\"' >> /etc/environment
+              grep -q '^NO_PROXY=' /etc/environment 2>/dev/null \
+              && sed -i 's|^NO_PROXY=.*|NO_PROXY=\"${NP}\"|' /etc/environment \
+              || echo 'NO_PROXY=\"${NP}\"' >> /etc/environment" >/dev/null 2>&1
+    echo "[cua_gym_vm_bridge] no_proxy=${NP}"
+}
+
+fix_vm_network
+
+echo "[cua_gym_vm_bridge] direct-IP mode: ${MOCK_HOST} (+${MOCK_NET}) bypassed from woa proxy; no socat forwarding needed"

ae_pipeline_review_004__long/initial_setup.py

@@ -0,0 +1,799 @@
+"""
+Initial Setup: T05_salesforce_oppty_followups — Pipeline Review sync into Salesforce
+Task ID: ae_pipeline_review_004__long
+Mocks: salesforce_mock,google_sheets_mock,slack_mock
+
+The 'Pipeline Review' Google Sheet lists opportunities (rows 2+:
+OppName, Account, Amount, Stage, CloseDate). For each row the agent must find
+the matching Salesforce Opportunity by NAME and apply the decision rules:
+  - Stage 'Closed Won'  -> SF stage 'Closed Won'  + Task 'Kickoff: <Account>' (Status 'Not Started')
+                           + reassign the opportunity Owner to 'Morgan Avery' (onboarding lead)
+  - Stage 'Closed Lost' -> SF stage 'Closed Lost' (no task, owner unchanged)
+  - Stage 'Negotiation' AND Amount >= 50000 -> SF stage 'Negotiation' + Task 'Exec review: <Account>' (owner unchanged)
+  - all other rows -> leave the SF opportunity UNCHANGED
+  - rows whose OppName has NO matching SF opportunity are SKIPPED
+Then create a new Slack channel named 'backup', export the Salesforce
+Opportunities report and the 'Pipeline Review' sheet to CSV, and upload BOTH
+CSV files into the 'backup' channel.
+
+The opportunity STAGE is changed from the opportunity's own page (the stage path),
+the follow-up TASK is created from the opportunity's Activity timeline, and the
+OWNER reassignment is a bulk action that only appears once an opportunity row is
+selected in the list -- all three are controls the agent has to discover.
+
+After syncing, the agent must create a new Slack channel named 'backup', export the
+Salesforce Opportunities report to CSV and the 'Pipeline Review' sheet to CSV, and
+upload BOTH CSV files into the 'backup' channel as a pipeline backup.
+
+Ground-truth (answer key) is PRECOMPUTED below and embedded into initial_state
+so reward.py is the sole source of truth:
+  - google_sheets._task_adapter.task_sheets['Pipeline Review']  (source rows)
+  - google_sheets._task_adapter.expected   (per-row decision: matched/final_stage/task_subject)
+  - google_sheets._task_adapter.matched_count == n
+  - salesforce._task_adapter.expected / .expected_by_name / .matched_count (mirror, co-located with the SF live state reward reads)
+  - slack._task_adapter.backup_channel == 'backup'  and  .required_csv_count == 2
+"""
+import datetime
+import os
+import shlex
+import subprocess
+import time
+import uuid
+
+import requests
+
+# ---------------------------------------------------------------------------
+# sid
+# ---------------------------------------------------------------------------
+sid = str(uuid.uuid4())
+with open('/tmp/task_web_sid', 'w', encoding='utf-8') as f:
+    f.write(sid)
+print(f'Generated sid: {sid}')
+
+TASK_ID = '60c91d2c-31b7-490f-bc00-1317ca378c29'
+SHEET_NAME = 'Pipeline Review'
+HEADERS = ['OppName', 'Account', 'Amount', 'Stage', 'CloseDate']
+
+# Owner reassignment target for Closed-Won opps (the "onboarding lead").
+# Opportunities are seeded owned by 'U1'; Closed-Won rows must end owned by U4.
+SEEDED_OWNER_ID = 'U1'
+ONBOARDING_OWNER_ID = 'U4'
+ONBOARDING_OWNER_NAME = 'Morgan Avery'
+
+# ---------------------------------------------------------------------------
+# Date helpers (relative to "today" so CloseDate values are realistic)
+# ---------------------------------------------------------------------------
+TODAY = datetime.date(2026, 6, 23)
+
+
+def fdate(days_ahead):
+    return (TODAY + datetime.timedelta(days=days_ahead)).isoformat()
+
+
+# Stage -> default probability for a Salesforce opportunity.
+STAGE_PROB = {
+    'Prospecting': 10,
+    'Qualification': 25,
+    'Needs Analysis': 35,
+    'Value Proposition': 50,
+    'Proposal': 65,
+    'Negotiation': 80,
+    'Closed Won': 100,
+    'Closed Lost': 0,
+}
+
+# ---------------------------------------------------------------------------
+# Single source of truth for the scenario.  Everything else (sheet cells,
+# SF opportunities, the expected answer key, n) is DERIVED from this list, so
+# the visible data and the embedded key can never drift apart.
+#
+# Each spec:
+#   opp_name     : OppName in the sheet  (== SF Opportunity.name when matched)
+#   account      : Account in the sheet  (drives the Task subject)
+#   amount       : Amount in the sheet
+#   sheet_stage  : Stage in the sheet (the DESIRED decision input)
+#   close_off    : CloseDate = today + close_off days
+#   matched      : True  -> a SF opportunity with this name is pre-seeded
+#                  False -> NO SF opportunity (agent must skip this row)
+#   sf_initial   : the SF opportunity's stage AT INJECTION (only if matched).
+#                  Chosen != the final stage for rows that require a change so
+#                  the change is observable; == final stage for unchanged rows.
+# ---------------------------------------------------------------------------
+ROW_SPECS = [
+    # --- Branch: Closed Won  (set stage + create Kickoff task) ---------------
+    dict(opp_name='Acme Renewal',      account='Acme Corp',        amount=120000,
+         sheet_stage='Closed Won',  close_off=7,  matched=True,  sf_initial='Proposal'),
+    # --- Branch: Closed Lost (set stage, no task) ----------------------------
+    dict(opp_name='Globex Expansion',  account='Globex Inc',       amount=85000,
+         sheet_stage='Closed Lost', close_off=12, matched=True,  sf_initial='Negotiation'),
+    # --- Branch: Negotiation & Amount >= 50000 (set stage + Exec review) -----
+    dict(opp_name='Initech Platform',  account='Initech LLC',      amount=75000,
+         sheet_stage='Negotiation', close_off=18, matched=True,  sf_initial='Qualification'),
+    # --- Branch: Negotiation & Amount < 50000 (unchanged; the 30000 case) ----
+    dict(opp_name='Umbrella Rollout',  account='Umbrella Corp',    amount=30000,
+         sheet_stage='Negotiation', close_off=24, matched=True,  sf_initial='Value Proposition'),
+    # --- Branch: Closed Won (second one, observe Negotiation -> Closed Won) ---
+    dict(opp_name='Soylent Upsell',    account='Soylent Co',       amount=60000,
+         sheet_stage='Closed Won',  close_off=9,  matched=True,  sf_initial='Negotiation'),
+    # --- Branch: Closed Lost (second one) ------------------------------------
+    dict(opp_name='Stark Integration', account='Stark Industries', amount=95000,
+         sheet_stage='Closed Lost', close_off=15, matched=True,  sf_initial='Proposal'),
+    # --- Branch: other stage 'Proposal' (unchanged) --------------------------
+    dict(opp_name='Wayne Migration',   account='Wayne Enterprises', amount=40000,
+         sheet_stage='Proposal',    close_off=30, matched=True,  sf_initial='Proposal'),
+    # --- Branch: other stage 'Prospecting' (unchanged) -----------------------
+    dict(opp_name='Hooli Renewal',     account='Hooli Inc',        amount=150000,
+         sheet_stage='Prospecting', close_off=40, matched=True,  sf_initial='Prospecting'),
+    # --- No matching SF opportunity (must be SKIPPED) ------------------------
+    dict(opp_name='Pied Piper Deal',   account='Pied Piper',       amount=50000,
+         sheet_stage='Closed Won',  close_off=20, matched=False, sf_initial=None),
+    dict(opp_name='Vehement Capital',  account='Vehement Partners', amount=70000,
+         sheet_stage='Negotiation', close_off=22, matched=False, sf_initial=None),
+]
+
+
+def decide(spec):
+    """Apply the task's decision rules -> (final_stage, task_subject, task_status,
+    task_status_required, requires_change, requires_task).
+
+    final_stage is the stage the SF opp must END at; for matched rows that do
+    NOT require a change this equals sf_initial (so reward checks 'no drift').
+    task_status_required encodes that the instruction only mandates a Status
+    ('Not Started') for the Kickoff task; the Exec review task's subject is the
+    binding requirement (status defaults to the natural 'Not Started').
+    """
+    stage = spec['sheet_stage']
+    amount = spec['amount']
+    initial = spec['sf_initial']
+    account = spec['account']
+
+    if stage == 'Closed Won':
+        final_stage = 'Closed Won'
+        task_subject = f'Kickoff: {account}'
+        task_status = 'Not Started'
+        task_status_required = True
+        requires_task = True
+        final_owner = ONBOARDING_OWNER_ID          # Closed Won -> reassign owner
+    elif stage == 'Closed Lost':
+        final_stage = 'Closed Lost'
+        task_subject = None
+        task_status = None
+        task_status_required = False
+        requires_task = False
+        final_owner = SEEDED_OWNER_ID               # owner unchanged
+    elif stage == 'Negotiation' and amount >= 50000:
+        final_stage = 'Negotiation'
+        task_subject = f'Exec review: {account}'
+        task_status = 'Not Started'
+        task_status_required = False
+        requires_task = True
+        final_owner = SEEDED_OWNER_ID               # owner unchanged
+    else:
+        # all other rows -> leave the SF opportunity unchanged
+        final_stage = initial
+        task_subject = None
+        task_status = None
+        task_status_required = False
+        requires_task = False
+        final_owner = SEEDED_OWNER_ID               # owner unchanged
+
+    requires_change = bool(initial is not None and final_stage != initial)
+    requires_owner_change = bool(final_owner != SEEDED_OWNER_ID)
+    return (final_stage, task_subject, task_status,
+            task_status_required, requires_change, requires_task,
+            final_owner, requires_owner_change)
+
+
+# ---------------------------------------------------------------------------
+# PRECOMPUTE: expected answer key + matched count (n) + SF opportunities
+# ---------------------------------------------------------------------------
+expected = []                  # full per-row key (incl. unmatched rows)
+sf_opportunities = []          # matched rows -> a seeded SF opportunity
+sf_accounts = []               # one account per matched opp + decoys
+sheet_rows_dict = []           # for _task_adapter.task_sheets
+
+acc_counter = 0
+opp_counter = 0
+for spec in ROW_SPECS:
+    (final_stage, task_subject, task_status, task_status_required,
+     requires_change, requires_task, final_owner, requires_owner_change) = decide(spec)
+
+    opportunity_id = None
+    account_id = None
+    if spec['matched']:
+        opp_counter += 1
+        acc_counter += 1
+        opportunity_id = f'opp-{opp_counter}'
+        account_id = f'acc-{acc_counter}'
+        sf_accounts.append({
+            'accountId': account_id,
+            'name': spec['account'],
+            'type': 'Customer',
+            'industry': 'Technology',
+            'revenue': spec['amount'] * 10,
+            'employees': 200 + opp_counter * 25,
+            'ownerId': 'U1',
+            'billingStreet': '', 'billingCity': '', 'billingState': '',
+            'billingZip': '', 'billingCountry': 'United States',
+            'shippingStreet': '', 'shippingCity': '', 'shippingState': '',
+            'shippingZip': '', 'shippingCountry': 'United States',
+            'phone': '', 'website': '', 'description': '',
+            'createdDate': '2026-05-01T00:00:00Z',
+            'modifiedDate': '2026-05-01T00:00:00Z',
+        })
+        sf_opportunities.append({
+            'opportunityId': opportunity_id,
+            'name': spec['opp_name'],
+            'accountId': account_id,
+            'contactId': '',
+            'amount': spec['amount'],
+            'closeDate': fdate(spec['close_off']),
+            'stage': spec['sf_initial'],
+            'probability': STAGE_PROB.get(spec['sf_initial'], 10),
+            'ownerId': 'U1',
+            'type': 'Existing Customer - Upgrade',
+            'leadSource': 'Partner',
+            'description': '',
+            'createdDate': '2026-05-05T00:00:00Z',
+            'modifiedDate': '2026-05-05T00:00:00Z',
+        })
+
+    expected.append({
+        'oppName': spec['opp_name'],
+        'account': spec['account'],
+        'amount': spec['amount'],
+        'sheet_stage': spec['sheet_stage'],
+        'matched': spec['matched'],
+        'opportunityId': opportunity_id,
+        'initial_stage': spec['sf_initial'],
+        'final_stage': final_stage,
+        'task_subject': task_subject,
+        'task_status': task_status,
+        'task_status_required': task_status_required,
+        'requires_change': requires_change,
+        'requires_task': requires_task,
+        'initial_owner': SEEDED_OWNER_ID if spec['matched'] else None,
+        'final_owner': final_owner if spec['matched'] else None,
+        'requires_owner_change': requires_owner_change if spec['matched'] else False,
+    })
+
+    sheet_rows_dict.append({
+        'OppName': spec['opp_name'],
+        'Account': spec['account'],
+        'Amount': spec['amount'],
+        'Stage': spec['sheet_stage'],
+        'CloseDate': fdate(spec['close_off']),
+    })
+
+matched_count = sum(1 for e in expected if e['matched'])           # n
+expected_message = f'Synced {matched_count} opportunities'
+expected_by_name = {
+    e['oppName']: {
+        'final_stage': e['final_stage'],
+        'task_subject': e['task_subject'],
+        'task_status': e['task_status'],
+        'task_status_required': e['task_status_required'],
+        'requires_change': e['requires_change'],
+        'requires_task': e['requires_task'],
+        'initial_owner': e['initial_owner'],
+        'final_owner': e['final_owner'],
+        'requires_owner_change': e['requires_owner_change'],
+    }
+    for e in expected if e['matched']
+}
+
+# Decoy SF opportunities NOT present in the sheet (watermark; never graded).
+sf_accounts.append({
+    'accountId': 'acc-d1', 'name': 'Cyberdyne Systems', 'type': 'Prospect',
+    'industry': 'Robotics', 'revenue': 9000000, 'employees': 800, 'ownerId': 'U2',
+    'billingStreet': '', 'billingCity': '', 'billingState': '', 'billingZip': '',
+    'billingCountry': 'United States', 'shippingStreet': '', 'shippingCity': '',
+    'shippingState': '', 'shippingZip': '', 'shippingCountry': 'United States',
+    'phone': '', 'website': '', 'description': '',
+    'createdDate': '2026-05-01T00:00:00Z', 'modifiedDate': '2026-05-01T00:00:00Z',
+})
+sf_accounts.append({
+    'accountId': 'acc-d2', 'name': 'Tyrell Corporation', 'type': 'Customer',
+    'industry': 'Biotech', 'revenue': 25000000, 'employees': 1200, 'ownerId': 'U3',
+    'billingStreet': '', 'billingCity': '', 'billingState': '', 'billingZip': '',
+    'billingCountry': 'United States', 'shippingStreet': '', 'shippingCity': '',
+    'shippingState': '', 'shippingZip': '', 'shippingCountry': 'United States',
+    'phone': '', 'website': '', 'description': '',
+    'createdDate': '2026-05-01T00:00:00Z', 'modifiedDate': '2026-05-01T00:00:00Z',
+})
+sf_opportunities.append({
+    'opportunityId': 'opp-d1', 'name': 'Cyberdyne Support Renewal', 'accountId': 'acc-d1',
+    'contactId': '', 'amount': 22000, 'closeDate': fdate(55), 'stage': 'Qualification',
+    'probability': 25, 'ownerId': 'U2', 'type': 'Existing Customer - Renewal',
+    'leadSource': 'Web', 'description': '', 'createdDate': '2026-05-05T00:00:00Z',
+    'modifiedDate': '2026-05-05T00:00:00Z',
+})
+sf_opportunities.append({
+    'opportunityId': 'opp-d2', 'name': 'Tyrell New Business', 'accountId': 'acc-d2',
+    'contactId': '', 'amount': 310000, 'closeDate': fdate(70), 'stage': 'Proposal',
+    'probability': 65, 'ownerId': 'U3', 'type': 'New Business',
+    'leadSource': 'Referral', 'description': '', 'createdDate': '2026-05-05T00:00:00Z',
+    'modifiedDate': '2026-05-05T00:00:00Z',
+})
+
+# Decoy activities — present (Rule: "activities array present") but NONE has a
+# subject equal to any expected Kickoff/Exec-review subject, and none is tied to
+# a graded opportunity, so they cannot create false positives at injection.
+sf_activities = [
+    {'activityId': 'activity-1', 'type': 'event', 'subject': 'Discovery call — Cyberdyne Systems',
+     'status': 'Completed', 'priority': 'Normal',
+     'startDateTime': '2026-06-18T15:00:00Z', 'endDateTime': '2026-06-18T15:30:00Z',
+     'relatedToType': 'opportunity', 'relatedToId': 'opp-d1', 'assignedToId': 'U2',
+     'description': ''},
+    {'activityId': 'activity-2', 'type': 'task', 'subject': 'Send NDA to Tyrell Corporation',
+     'status': 'Completed', 'priority': 'High', 'dueDate': fdate(-2),
+     'relatedToType': 'opportunity', 'relatedToId': 'opp-d2', 'assignedToId': 'U3',
+     'description': ''},
+]
+
+# ---------------------------------------------------------------------------
+# Salesforce users
+# ---------------------------------------------------------------------------
+sf_users = [
+    {'userId': 'U1', 'firstName': 'John', 'lastName': 'Smith', 'email': 'john.smith@company.com',
+     'phone': '(555) 123-4567', 'title': 'RevOps Manager', 'department': 'Sales', 'role': 'Manager',
+     'avatar': 'https://i.pravatar.cc/150?u=U1', 'timezone': 'America/New_York',
+     'locale': 'en-US', 'theme': 'lightning'},
+    {'userId': 'U2', 'firstName': 'Sarah', 'lastName': 'Chen', 'email': 'sarah.chen@company.com',
+     'phone': '', 'title': 'Account Executive', 'department': 'Sales', 'role': 'Rep',
+     'avatar': 'https://i.pravatar.cc/150?u=U2', 'timezone': 'America/New_York',
+     'locale': 'en-US', 'theme': 'lightning'},
+    {'userId': 'U3', 'firstName': 'Diego', 'lastName': 'Ruiz', 'email': 'diego.ruiz@company.com',
+     'phone': '', 'title': 'Account Executive', 'department': 'Sales', 'role': 'Rep',
+     'avatar': 'https://i.pravatar.cc/150?u=U3', 'timezone': 'America/Chicago',
+     'locale': 'en-US', 'theme': 'lightning'},
+    {'userId': 'U4', 'firstName': 'Morgan', 'lastName': 'Avery', 'email': 'morgan.avery@company.com',
+     'phone': '', 'title': 'Customer Onboarding Lead', 'department': 'Customer Success', 'role': 'Manager',
+     'avatar': 'https://i.pravatar.cc/150?u=U4', 'timezone': 'America/New_York',
+     'locale': 'en-US', 'theme': 'lightning'},
+]
+
+salesforce_state = {
+    'user': sf_users[0],
+    'users': sf_users,
+    'leads': [],
+    'accounts': sf_accounts,
+    'contacts': [],
+    'opportunities': sf_opportunities,
+    'cases': [],
+    'activities': sf_activities,
+    'chatterPosts': [],
+    'files': [],
+    'dashboards': [],
+    'following': [],
+    'recentlyViewed': [],
+    'dismissedNotifications': [],
+    '_task_adapter': {
+        'task_id': TASK_ID,
+        'variant': 'eval',
+        'matched_count': matched_count,
+        'expected': expected,
+        'expected_by_name': expected_by_name,
+        'onboarding_owner_id': ONBOARDING_OWNER_ID,
+        'onboarding_owner_name': ONBOARDING_OWNER_NAME,
+        'seeded_owner_id': SEEDED_OWNER_ID,
+    },
+}
+
+# ---------------------------------------------------------------------------
+# Google Sheets — 'Pipeline Review' workbook
+# ---------------------------------------------------------------------------
+COL_LETTERS = ['A', 'B', 'C', 'D', 'E']
+
+
+def header_cell(text):
+    return {'value': text, 'formula': text, 'computed': text,
+            'style': {'bold': True, 'bg': '#F3F3F3', 'align': 'center'}}
+
+
+def text_cell(text):
+    return {'value': text, 'formula': text, 'computed': text}
+
+
+def num_cell(n):
+    return {'value': str(n), 'formula': str(n), 'computed': n}
+
+
+sheet_data = {}
+# Header row (row 1)
+for col, head in zip(COL_LETTERS, HEADERS):
+    sheet_data[f'{col}1'] = header_cell(head)
+# Data rows (rows 2..N)
+for idx, row in enumerate(sheet_rows_dict):
+    r = idx + 2
+    sheet_data[f'A{r}'] = text_cell(row['OppName'])
+    sheet_data[f'B{r}'] = text_cell(row['Account'])
+    sheet_data[f'C{r}'] = num_cell(row['Amount'])
+    sheet_data[f'D{r}'] = text_cell(row['Stage'])
+    sheet_data[f'E{r}'] = text_cell(row['CloseDate'])
+
+google_sheets_state = {
+    'id': 'workbook_t05_pipeline_review',
+    'title': 'Pipeline Review',
+    'activeSheetId': 'sheet_1',
+    'selectedCell': 'A1',
+    'selectionRange': None,
+    'clipboard': None,
+    'isDragging': False,
+    'undoStack': [],
+    'redoStack': [],
+    'namedRanges': [],
+    'conditionalFormats': [],
+    'charts': [],
+    'showGridlines': True,
+    'showFormulas': False,
+    'zoom': 100,
+    'sheets': [{
+        'id': 'sheet_1',
+        'name': SHEET_NAME,
+        'data': sheet_data,
+        'rowCount': 100,
+        'colCount': 26,
+        'frozenRows': 1,
+        'frozenCols': 0,
+        'tabColor': None,
+        'isHidden': False,
+        'columnWidths': {},
+        'rowHeights': {},
+        'filterRange': None,
+        'filterCriteria': {},
+        'sortColumn': None,
+        'sortDirection': None,
+    }],
+    '_task_adapter': {
+        'source_schema': 'rows_dict',
+        'task_id': TASK_ID,
+        'variant': 'eval',
+        'task_sheets': {SHEET_NAME: {'headers': HEADERS, 'rows': sheet_rows_dict}},
+        'headers_by_sheet': {SHEET_NAME: HEADERS},
+        'sheet_names': [SHEET_NAME],
+        'expected': expected,
+        'expected_by_name': expected_by_name,
+        'matched_count': matched_count,
+        'expected_message': expected_message,
+    },
+}
+
+# ---------------------------------------------------------------------------
+# Slack — the 'backup' channel must NOT exist yet (Rule 3); the agent creates
+# it and uploads the two CSVs. Pre-existing channels are decoy chatter only.
+# ---------------------------------------------------------------------------
+slack_state = {
+    'channels': [
+        {'channelId': 'general', 'name': 'general', 'description': 'Company-wide announcements and work-based matters', 'topic': 'Welcome to Acme Corp!', 'isPrivate': False, 'isStarred': True, 'members': ['user_1', 'user_2', 'user_3', 'user_4'], 'createdBy': 'user_2', 'createdAt': '2026-05-01T09:00:00Z', 'pinnedMessages': [], 'unreadCount': 0},
+        {'channelId': 'announcements', 'name': 'announcements', 'description': 'Important company updates', 'topic': 'Read-mostly. Big news only.', 'isPrivate': False, 'isStarred': False, 'members': ['user_1', 'user_2', 'user_3', 'user_4'], 'createdBy': 'user_2', 'createdAt': '2026-05-01T09:00:00Z', 'pinnedMessages': [], 'unreadCount': 0},
+        {'channelId': 'random', 'name': 'random', 'description': 'Non-work banter and watercooler chat', 'topic': 'Coffee, memes, weekend plans', 'isPrivate': False, 'isStarred': False, 'members': ['user_1', 'user_2', 'user_3', 'user_4'], 'createdBy': 'user_3', 'createdAt': '2026-05-01T09:00:00Z', 'pinnedMessages': [], 'unreadCount': 0},
+        {'channelId': 'revops', 'name': 'revops', 'description': 'Revenue operations — pipeline hygiene and sync confirmations', 'topic': 'Drop a note when you finish a pipeline sync', 'isPrivate': False, 'isStarred': False, 'members': ['user_1'], 'createdBy': 'user_1', 'createdAt': '2026-06-04T00:00:00Z', 'pinnedMessages': [], 'unreadCount': 0},
+    ],
+    'currentUser': {'userId': 'user_1', 'fullName': 'John Smith', 'displayName': 'John', 'email': 'john.smith@company.com', 'avatar': 'https://picsum.photos/200/200?random=1', 'title': 'RevOps Manager', 'status': 'online', 'statusMessage': '', 'statusEmoji': '', 'timeZone': 'America/New_York'},
+    'users': [
+        {'userId': 'user_1', 'fullName': 'John Smith', 'displayName': 'John', 'email': 'john.smith@company.com', 'avatar': 'https://picsum.photos/200/200?random=1', 'title': 'RevOps Manager', 'status': 'online', 'statusMessage': '', 'statusEmoji': '', 'timeZone': 'America/New_York'},
+        {'userId': 'user_2', 'fullName': 'Maya Lindqvist', 'displayName': 'Maya', 'email': 'maya.lindqvist@company.com', 'avatar': 'https://picsum.photos/200/200?random=2', 'title': 'People Ops', 'status': 'online', 'statusMessage': '', 'statusEmoji': '', 'timeZone': 'Europe/Stockholm'},
+        {'userId': 'user_3', 'fullName': 'Hiroshi Tanabe', 'displayName': 'Hiroshi', 'email': 'hiroshi.tanabe@company.com', 'avatar': 'https://picsum.photos/200/200?random=3', 'title': 'Engineering', 'status': 'away', 'statusMessage': 'In a meeting', 'statusEmoji': ':calendar:', 'timeZone': 'Asia/Tokyo'},
+        {'userId': 'user_4', 'fullName': 'Olivia Becker', 'displayName': 'Olivia', 'email': 'olivia.becker@company.com', 'avatar': 'https://picsum.photos/200/200?random=4', 'title': 'Marketing', 'status': 'online', 'statusMessage': '', 'statusEmoji': '', 'timeZone': 'America/Los_Angeles'},
+    ],
+    'workspace': {'workspaceId': 'ws_1', 'workspaceName': 'Acme Corp', 'icon': ''},
+    'messages': {
+        'general': [
+            {'messageId': 'm_g_1', 'senderId': 'user_2', 'content': 'Morning everyone — reminder that the office will be closed next Friday for the holiday.', 'timestamp': '2026-06-18T13:02:00Z', 'reactions': [], 'isEdited': False, 'threadId': None, 'attachments': []},
+            {'messageId': 'm_g_2', 'senderId': 'user_4', 'content': 'Thanks Maya! Long weekend incoming :tada:', 'timestamp': '2026-06-18T13:05:00Z', 'reactions': [{'emoji': '\U0001F389', 'users': ['user_1', 'user_3']}], 'isEdited': False, 'threadId': None, 'attachments': []},
+            {'messageId': 'm_g_3', 'senderId': 'user_3', 'content': 'Quick heads-up: the staging environment is being rebuilt today, expect slowness 10:00–11:00 JST.', 'timestamp': '2026-06-19T01:12:00Z', 'reactions': [], 'isEdited': False, 'threadId': None, 'attachments': []},
+            {'messageId': 'm_g_4', 'senderId': 'user_1', 'content': "Got it, I'll hold off on the pipeline review until after.", 'timestamp': '2026-06-19T01:15:00Z', 'reactions': [], 'isEdited': False, 'threadId': None, 'attachments': []},
+        ],
+        'announcements': [
+            {'messageId': 'm_a_1', 'senderId': 'user_2', 'content': 'Q2 all-hands has been scheduled for June 28 at 10:00 PT. Calendar invite goes out today.', 'timestamp': '2026-06-17T16:30:00Z', 'reactions': [{'emoji': '✅', 'users': ['user_1', 'user_3', 'user_4']}], 'isEdited': False, 'threadId': None, 'attachments': []},
+            {'messageId': 'm_a_2', 'senderId': 'user_2', 'content': 'New laptop refresh policy is live on the People Ops wiki. TL;DR: 3-year cycle, request via the IT portal.', 'timestamp': '2026-06-19T09:00:00Z', 'reactions': [], 'isEdited': False, 'threadId': None, 'attachments': []},
+        ],
+        'random': [
+            {'messageId': 'm_r_1', 'senderId': 'user_4', 'content': 'Anyone tried the new ramen place on 3rd? Verdict?', 'timestamp': '2026-06-18T18:50:00Z', 'reactions': [{'emoji': '\U0001F35C', 'users': ['user_3']}], 'isEdited': False, 'threadId': None, 'attachments': []},
+            {'messageId': 'm_r_2', 'senderId': 'user_3', 'content': "10/10, get the spicy miso. Bring tissues though, it's no joke.", 'timestamp': '2026-06-18T18:55:00Z', 'reactions': [{'emoji': '\U0001F605', 'users': ['user_1', 'user_4']}], 'isEdited': False, 'threadId': None, 'attachments': []},
+            {'messageId': 'm_r_3', 'senderId': 'user_1', 'content': 'My cat has decided my keyboard is her new bed. Send help.', 'timestamp': '2026-06-19T08:20:00Z', 'reactions': [{'emoji': '\U0001F63A', 'users': ['user_2', 'user_3', 'user_4']}], 'isEdited': False, 'threadId': None, 'attachments': []},
+        ],
+        # Decoy channel — not graded. The gradable artefact is the NEW 'backup'
+        # channel (created by the agent), which must be ABSENT at injection.
+        'revops': [],
+    },
+    'threads': {},
+    'dms': [],
+    'bookmarkedMessages': [],
+    'callHistory': [],
+    'settings': {'theme': 'light', 'notifications': 'all', 'displayDensity': 'comfortable', 'showAvatars': True, 'use24Hour': False},
+    'invitations': [],
+    'notifications': [],
+    '_task_adapter': {
+        'task_id': TASK_ID,
+        'variant': 'eval',
+        'backup_channel': 'backup',          # agent must CREATE this channel (absent at injection)
+        'required_csv_count': 2,             # SF opportunities CSV + Pipeline Review sheet CSV
+        'matched_count': matched_count,
+    },
+}
+
+# Rule 3 guard: the 'backup' channel must NOT exist at injection (gradable output absent).
+assert not any(str(c.get('name', '')).strip().lower() == 'backup' for c in slack_state['channels']), \
+    "'backup' channel must be absent at injection"
+
+# ---------------------------------------------------------------------------
+# (url, state) — EXACT app_urls from task_config.json
+# ---------------------------------------------------------------------------
+APP_STATES = [
+    ('http://28.7.184.198:8175', salesforce_state),      # salesforce_mock
+    ('http://28.7.184.198:8145', google_sheets_state),   # google_sheets_mock
+    ('http://28.7.184.198:8178', slack_state),           # slack_mock
+]
+
+for app_url, state in APP_STATES:
+    resp = requests.post(
+        f'{app_url}/post?sid={sid}',
+        json={'action': 'set', 'state': state},
+        timeout=30,
+    )
+    assert resp.status_code == 200, (
+        f'State injection failed for {app_url}: {resp.text}'
+    )
+    print(f'State injected: {app_url} sid={sid}')
+
+for app_url, _ in APP_STATES:
+    go_payload = requests.get(f'{app_url}/go?sid={sid}', timeout=15).json()
+    assert go_payload.get('initial_state') is not None, (
+        f'initial_state is None after injection for {app_url}'
+    )
+    print(f'Verified initial_state set: {app_url}')
+
+print(f'Precomputed matched_count (n) = {matched_count}; '
+      f'backup_channel = backup; required_csv_count = 2')
+
+
+# PATCHED_BY: patch_wait_mocks_loaded.py
+def launch_gui(command, delay_sec=1.0):
+    """Launch a GUI command, fully detached from this script.
+
+    Inject --remote-debugging-port=1337 (and Chrome quiet flags) so that
+    wait_mocks_loaded's CDP phase can verify browser render below.
+    Idempotent: if the caller already specified a debug port, keep it.
+    """
+    env = os.environ.copy()
+    env['DISPLAY'] = ':0'
+    if isinstance(command, str):
+        _parts = shlex.split(command)
+    else:
+        _parts = list(command)
+    if (_parts and 'google-chrome' in _parts[0]
+            and not any(p.startswith('--remote-debugging-port') for p in _parts)):
+        _parts[1:1] = [
+            '--no-first-run',
+            '--no-default-browser-check',
+            '--remote-debugging-port=1337',
+        ]
+    subprocess.Popen(
+        _parts,
+        stdin=subprocess.DEVNULL,
+        stdout=subprocess.DEVNULL,
+        stderr=subprocess.DEVNULL,
+        env=env,
+        start_new_session=True,
+    )
+    time.sleep(delay_sec)
+
+
+def wait_mocks_loaded(app_states, sid_value, timeout=90, render_buffer=12):
+    """2-phase readiness wait, mirrors cache/ae_pipeline_hygiene_004 style.
+
+    Phase 1 - Backend readiness:
+        Poll each mock's /state?sid=<sid> (preferred) or /go?sid=<sid>
+        (fallback) until it returns 200 with 'stored_state' or
+        'initial_state' respectively, or until timeout.
+
+    Phase 2 - Browser render verification (via CDP + Playwright):
+        Connect to http://localhost:1337 (the --remote-debugging-port
+        injected by launch_gui above), wait for every open tab to reach
+        'networkidle' AND to contain at least 20 DOM elements. This
+        guarantees the SPA fetched /go?sid= data AND actually painted,
+        not just that the mock backend is up.
+
+    Fallback:
+        If Playwright is missing or CDP cannot be reached, sleep for
+        `render_buffer` seconds as a best-effort delay so the screenshot
+        loop does not start on a still-blank tab.
+    """
+    import urllib.request as _urlreq
+    import time as _t
+    _urls = [p for p, _ in app_states]
+    _opener = _urlreq.build_opener(_urlreq.ProxyHandler({}))
+
+    # ---- Phase 1: backend readiness ----
+    _deadline = _t.time() + timeout
+    _todo = list(_urls)
+    while _todo and _t.time() < _deadline:
+        _rest = []
+        for _u in _todo:
+            _ok = False
+            for _path, _need in (('/state?sid=', b'stored_state'),
+                                 ('/go?sid=',    b'initial_state')):
+                try:
+                    _r = _opener.open(f"{_u}{_path}{sid_value}", timeout=5)
+                    if _r.status == 200:
+                        _body = _r.read()
+                        if _need in _body:
+                            _ok = True
+                            break
+                except Exception:
+                    pass
+            if _ok:
+                print(f"  {_u}: backend ready")
+                continue
+            _rest.append(_u)
+        _todo = _rest
+        if _todo:
+            _t.sleep(2)
+    _done = len(_urls) - len(_todo)
+    _pend = f" (pending: {_todo})" if _todo else ""
+    print(f"  mocks backend ready: {_done}/{len(_urls)}{_pend}")
+
+    # ---- Phase 2: browser render verification via CDP ----
+    _cdp_ok = False
+    try:
+        from playwright.sync_api import sync_playwright
+        _cdp_dl = _t.time() + max(15.0, _deadline - _t.time())
+        with sync_playwright() as _p:
+            _br = None
+            while _t.time() < _cdp_dl:
+                try:
+                    _br = _p.chromium.connect_over_cdp('http://localhost:1337')
+                    break
+                except Exception:
+                    _t.sleep(0.5)
+            if _br is None:
+                print("  wait_mocks_loaded: CDP connect to :1337 failed; is Chrome up with --remote-debugging-port?")
+            else:
+                _n_expected = len(_urls)
+                _pages = []
+                _pg_dl = _t.time() + 15
+                _last_count = -1
+                _stable_since = None
+                while _t.time() < _pg_dl:
+                    _pages = []
+                    for _c in _br.contexts:
+                        _pages.extend(_c.pages)
+                    _cnt = len(_pages)
+                    if _n_expected > 0 and _cnt >= _n_expected:
+                        break
+                    if _cnt > 0 and _cnt == _last_count:
+                        if _stable_since is None:
+                            _stable_since = _t.time()
+                        elif _t.time() - _stable_since >= 1.5:
+                            break
+                    else:
+                        _stable_since = None
+                    _last_count = _cnt
+                    _t.sleep(0.5)
+                _pages = []
+                for _c in _br.contexts:
+                    _pages.extend(_c.pages)
+                if not _pages:
+                    print("  wait_mocks_loaded: CDP connected but no open tabs found")
+                else:
+                    print(f"  wait_mocks_loaded: {len(_pages)} tab(s) open (expected {_n_expected}); polling render")
+                    _MIN_ELEMS = 20
+                    _verify_dl = _t.time() + 40.0   # bounded so total setup stays < VM 120s cap
+                    for _pg in _pages:              # bring each to front once to defeat background-tab render throttling
+                        try:
+                            _pg.bring_to_front()
+                        except Exception:
+                            pass
+                    _pending = list(_pages)
+                    while _pending and _t.time() < _verify_dl:
+                        _still = []
+                        for _pg in _pending:
+                            _ttl = ""
+                            try:
+                                _ttl = _pg.title()
+                            except Exception:
+                                pass
+                            try:
+                                _n = _pg.evaluate("document.body ? document.querySelectorAll('*').length : 0")
+                            except Exception:
+                                _n = 0
+                            if _n >= _MIN_ELEMS:
+                                print(f"  wait_mocks_loaded: tab {_ttl!r} rendered ({_n} elements)")
+                            else:
+                                _still.append(_pg)
+                        _pending = _still
+                        if _pending:
+                            _t.sleep(2)
+                    if not _pending:
+                        _cdp_ok = True
+                    else:
+                        for _pg in _pending:
+                            _ttl = ""
+                            try:
+                                _ttl = _pg.title()
+                            except Exception:
+                                pass
+                            print(f"  wait_mocks_loaded: WARNING tab {_ttl!r} still not rendered after poll, will fall back")
+    except ImportError:
+        print("  wait_mocks_loaded: playwright not installed; skipping CDP phase")
+    except Exception as _e:
+        print(f"  wait_mocks_loaded: CDP phase error ({_e!r})")
+
+    if not _cdp_ok:
+        print(f"  wait_mocks_loaded: falling back to render_buffer={render_buffer}s sleep")
+        _t.sleep(render_buffer)
+    else:
+        print("  wait_mocks_loaded: browser render verified, skipping render_buffer")
+
+
+def _open_mock_tabs(app_states, sid_value):
+    """Open the first mock as the primary Chrome window, the rest as new tabs.
+
+    Uses launch_gui once per URL so that the macOS localizer regex-extract
+    (which grabs the FIRST quoted URL per call) still works correctly.
+    """
+    _urls = [p for p, _ in app_states]
+    if not _urls:
+        return
+    launch_gui(f'google-chrome "{_urls[0]}/?sid={sid_value}"', delay_sec=2.0)
+    for _u in _urls[1:]:
+        launch_gui(f'google-chrome --new-tab "{_u}/?sid={sid_value}"', delay_sec=1.0)
+
+
+
+# PATCHED_BY: add_wait_for_mocks_loaded (colleague headless+playwright render gate)
+def wait_for_mocks_loaded(mocks, sid, timeout=60.0, render_timeout=15.0):
+    """Block until every mock serves injected state AND (best-effort) renders."""
+    import time as _t
+    deadline = _t.time() + timeout
+    pending = dict(mocks)
+    while pending and _t.time() < deadline:
+        for name in list(pending):
+            try:
+                r = requests.get(f'{pending[name]}/go?sid={sid}', timeout=5)
+                if r.status_code == 200 and r.json().get('initial_state') is not None:
+                    del pending[name]
+            except Exception:
+                pass
+        if pending:
+            _t.sleep(1.0)
+    if pending:
+        print(f'WARN: mocks not ready within {timeout}s: {list(pending)}')
+    try:
+        from playwright.sync_api import sync_playwright
+    except ImportError:
+        print('WARN: playwright not installed; skipping render check, settle 3s')
+        _t.sleep(3.0)
+        return
+    with sync_playwright() as p:
+        try:
+            browser = p.chromium.launch(channel='chrome', headless=True,
+                                        args=['--no-sandbox', '--disable-gpu', '--disable-dev-shm-usage'])
+        except Exception as e:
+            print(f'WARN: could not launch chrome for render check: {e}; settle 3s')
+            _t.sleep(3.0)
+            return
+        try:
+            for name, url in mocks.items():
+                page = browser.new_page()
+                try:
+                    page.goto(f'{url}/?sid={sid}', wait_until='domcontentloaded', timeout=int(render_timeout * 1000))
+                    page.wait_for_function(
+                        "document.body && document.body.innerText.length > 50",
+                        timeout=int(render_timeout * 1000))
+                    print(f'[{name}] rendered OK')
+                except Exception as e:
+                    print(f'WARN: [{name}] render check issue: {e}')
+                finally:
+                    page.close()
+        finally:
+            browser.close()
+
+
+# ref pattern (out_benchmark_cache): open visible tabs FIRST, then block until mocks render
+_open_mock_tabs(APP_STATES, sid)
+wait_for_mocks_loaded({u.rsplit(':', 1)[-1]: u for u, _ in APP_STATES}, sid, timeout=60, render_timeout=15)
+wait_mocks_loaded(APP_STATES, sid, timeout=90, render_buffer=12)
+print(f'GUI_READY: launched browser tabs for {len(APP_STATES)} apps')

ae_pipeline_review_004__long/reward.py

@@ -0,0 +1,586 @@
+"""
+Reward Script: T05 — Pipeline Review sync into Salesforce opportunities
+Task ID: ae_pipeline_review_004__long
+Mocks: salesforce_mock,google_sheets_mock,slack_mock
+Scoring (positives sum to 1.0; leak penalties subtracted):
+  0.40 stage set correctly over the rows that require a stage change.
+  0.25 follow-up Task with exact Subject (+ Status 'Not Started' where required)
+       created for the requires_task rows (and absent on the others).
+  0.20 owner reassigned to 'Morgan Avery' on the Closed-Won rows.
+  0.15 Slack: a NEW 'backup' channel exists (absent at injection) and holds the
+       two exported CSV files as attachments (0.4 create + 0.6 for 2 CSVs).
+  penalties: stage-drift 0.15, forbidden-task 0.10, owner-drift 0.10 (each frac).
+Answer key: read from salesforce.initial_state._task_adapter (.expected / .matched_count)
+            and slack.initial_state._task_adapter (.backup_channel / .required_csv_count).
+"""
+import copy
+import re
+import sys
+
+import requests
+
+
+try:
+    with open('/tmp/task_web_sid', encoding='utf-8') as f:
+        sid = f.read().strip()
+    if not sid:
+        raise ValueError('sid is empty')
+except Exception as e:
+    print(f'CRITICAL: Cannot read sid from /tmp/task_web_sid: {e}')
+    print('REWARD: 0.0')
+    sys.exit(0)
+
+
+APP_URLS = {'salesforce': 'http://28.7.184.198:8175', 'google_sheets': 'http://28.7.184.198:8145', 'slack': 'http://28.7.184.198:8178'}
+_GO_CACHE = {}
+
+
+def clamp01(x):
+    return max(0.0, min(1.0, float(x)))
+
+
+def frac(num, den):
+    return 0.0 if den == 0 else num / den
+
+
+def f1(true_set, pred_set):
+    if not true_set and not pred_set:
+        return 1.0
+    tp = len(true_set & pred_set)
+    if tp == 0:
+        return 0.0
+    prec = tp / len(pred_set)
+    rec = tp / len(true_set)
+    return 2 * prec * rec / (prec + rec)
+
+
+def by_id(items, key='id'):
+    return {it.get(key): it for it in (items or []) if isinstance(it, dict)}
+
+
+def has_text(haystack, *needles):
+    s = (haystack or '').lower()
+    return any(n.lower() in s for n in needles)
+
+
+def norm(s):
+    return (str(s) if s is not None else '').strip().lower()
+
+
+def _msg_text(msg):
+    if not isinstance(msg, dict):
+        return ''
+    return msg.get('content') or msg.get('text') or ''
+
+
+def _slack_channel_messages(slack_state, channel_name):
+    channels = slack_state.get('channels', []) if isinstance(slack_state, dict) else []
+    target = None
+    for ch in channels:
+        if isinstance(ch, dict) and norm(ch.get('name')) == norm(channel_name):
+            target = ch
+            break
+    if target is None:
+        return []
+    ch_msgs = target.get('messages')
+    if isinstance(ch_msgs, list):
+        return ch_msgs
+    msg_map = slack_state.get('messages', {})
+    if isinstance(msg_map, dict):
+        return msg_map.get(target.get('channelId') or target.get('id'), []) or []
+    return []
+
+
+def _to_float(v):
+    try:
+        return float(v)
+    except (TypeError, ValueError):
+        return 0.0
+
+
+def _sheet_rows(sheet):
+    if not isinstance(sheet, dict):
+        return []
+    rows = sheet.get('rows', [])
+    return rows if isinstance(rows, list) else []
+
+
+def _sheet_tabs(sheets):
+    if isinstance(sheets, dict):
+        return sheets
+    if isinstance(sheets, list):
+        out = {}
+        for sh in sheets:
+            if not isinstance(sh, dict):
+                continue
+            name = sh.get('name') or sh.get('id')
+            if name:
+                out[name] = sh
+        return out
+    return {}
+
+
+def _jaccard(a, b):
+    if not a and not b:
+        return 1.0
+    u = a | b
+    return (len(a & b) / len(u)) if u else 1.0
+
+
+def _normalize_app_name(app):
+    app = (app or '').strip().lower()
+    return app[:-5] if app.endswith('_mock') else app
+
+
+def _parse_cell_id(cell_id):
+    if not isinstance(cell_id, str) or not cell_id:
+        return None
+    idx = 0
+    while idx < len(cell_id) and cell_id[idx].isalpha():
+        idx += 1
+    if idx == 0 or idx >= len(cell_id):
+        return None
+    col = 0
+    for ch in cell_id[:idx].upper():
+        col = col * 26 + (ord(ch) - ord('A') + 1)
+    try:
+        row = int(cell_id[idx:])
+    except ValueError:
+        return None
+    return row, col - 1
+
+
+def _extract_rows_from_workbook(state):
+    if not isinstance(state, dict):
+        return None
+    sheets = state.get('sheets')
+    if not isinstance(sheets, list):
+        return None
+
+    adapter = state.get('_task_adapter') if isinstance(state.get('_task_adapter'), dict) else {}
+    headers_by_sheet = adapter.get('headers_by_sheet') if isinstance(adapter.get('headers_by_sheet'), dict) else {}
+
+    out = {}
+    for sh in sheets:
+        if not isinstance(sh, dict):
+            continue
+        name = sh.get('name') or sh.get('id') or 'Sheet1'
+        data = sh.get('data') if isinstance(sh.get('data'), dict) else {}
+
+        headers = headers_by_sheet.get(name)
+        if not headers:
+            cols = []
+            for cid, cell in data.items():
+                rc = _parse_cell_id(cid)
+                if not rc:
+                    continue
+                row_idx, col_idx = rc
+                if row_idx != 1:
+                    continue
+                value = ''
+                if isinstance(cell, dict):
+                    value = cell.get('value', '')
+                elif cell is not None:
+                    value = str(cell)
+                if value is not None and str(value) != '':
+                    cols.append((col_idx, str(value)))
+            headers = [v for _, v in sorted(cols)]
+
+        if not headers:
+            continue
+
+        by_row = {}
+        for cid, cell in data.items():
+            rc = _parse_cell_id(cid)
+            if not rc:
+                continue
+            row_idx, col_idx = rc
+            if row_idx <= 1 or col_idx >= len(headers):
+                continue
+            if isinstance(cell, dict):
+                value = cell.get('value', '')
+            else:
+                value = cell
+            by_row.setdefault(row_idx, {})[headers[col_idx]] = value
+
+        rows = []
+        for row_idx in sorted(by_row.keys()):
+            row = by_row[row_idx]
+            if not any(str(row.get(h, '')).strip() for h in headers):
+                continue
+            rows.append({h: row.get(h, '') for h in headers})
+
+        out[name] = {'headers': headers, 'rows': rows}
+
+    if not out:
+        return None
+    return {'sheets': out}
+
+
+def _materialize_google_sheets(payload):
+    for key in ('initial_state', 'current_state'):
+        state = payload.get(key)
+        if not isinstance(state, dict):
+            continue
+        maybe_rows = _extract_rows_from_workbook(state)
+        if maybe_rows is not None:
+            state['_ui_workbook'] = {
+                'sheets': copy.deepcopy(state.get('sheets', [])),
+                'title': state.get('title'),
+            }
+            state['sheets'] = maybe_rows['sheets']
+
+
+def _materialize_slack(payload):
+    for key in ('initial_state', 'current_state'):
+        state = payload.get(key)
+        if not isinstance(state, dict):
+            continue
+        channels = state.get('channels') if isinstance(state.get('channels'), list) else []
+        messages_map = state.get('messages') if isinstance(state.get('messages'), dict) else {}
+        for ch in channels:
+            if not isinstance(ch, dict):
+                continue
+            cid = ch.get('channelId') or ch.get('id')
+            if not cid:
+                continue
+            msg_list = messages_map.get(cid) if isinstance(messages_map.get(cid), list) else []
+            channel_msgs = ch.get('messages') if isinstance(ch.get('messages'), list) else []
+            merged = {
+                (m.get('messageId') or m.get('id')): dict(m)
+                for m in channel_msgs
+                if isinstance(m, dict)
+            }
+            for m in msg_list:
+                if not isinstance(m, dict):
+                    continue
+                mid = m.get('messageId') or m.get('id')
+                text = m.get('text') if m.get('text') is not None else m.get('content', '')
+                merged[mid] = {
+                    'id': mid,
+                    'messageId': mid,
+                    'text': text,
+                    'content': m.get('content', text),
+                    'senderId': m.get('senderId'),
+                    'timestamp': m.get('timestamp'),
+                    'reactions': m.get('reactions') if isinstance(m.get('reactions'), list) else [],
+                    # Preserve uploaded-file attachments (the 'backup' CSV uploads
+                    # land here when a message is sent into messages[channelId]).
+                    'attachments': m.get('attachments') if isinstance(m.get('attachments'), list) else [],
+                }
+            ch['messages'] = list(merged.values())
+
+
+def _owner_name_map(state):
+    out = {}
+    users = state.get('users') if isinstance(state.get('users'), list) else []
+    for u in users:
+        if not isinstance(u, dict):
+            continue
+        uid = u.get('userId')
+        if not uid:
+            continue
+        name = (f"{u.get('firstName', '')} {u.get('lastName', '')}".strip() or u.get('name') or uid)
+        out[uid] = name
+    return out
+
+
+def _materialize_salesforce(payload):
+    for key in ('initial_state', 'current_state'):
+        state = payload.get(key)
+        if not isinstance(state, dict):
+            continue
+
+        adapter = state.get('_task_adapter') if isinstance(state.get('_task_adapter'), dict) else {}
+        owner_names = _owner_name_map(state)
+
+        task_leads = adapter.get('task_leads') if isinstance(adapter.get('task_leads'), dict) else None
+        if task_leads and isinstance(state.get('leads'), list):
+            restored = []
+            for lead in state['leads']:
+                if not isinstance(lead, dict):
+                    continue
+                lid = str(lead.get('leadId') or lead.get('id') or '')
+                meta = task_leads.get(lid, {})
+                restored.append({
+                    'id': meta.get('id') or lead.get('leadId') or lead.get('id'),
+                    'company': lead.get('company', meta.get('company', '')),
+                    'status': lead.get('status', 'New'),
+                    'ownerName': owner_names.get(lead.get('ownerId')) or meta.get('ownerName'),
+                    'annualRevenue': meta.get('annualRevenue', lead.get('revenue')),
+                })
+            state['leads'] = restored
+
+        task_accounts = adapter.get('task_accounts') if isinstance(adapter.get('task_accounts'), dict) else None
+        if task_accounts and isinstance(state.get('accounts'), list):
+            current_by_id = {}
+            for acc in state['accounts']:
+                if isinstance(acc, dict):
+                    current_by_id[str(acc.get('accountId') or acc.get('id') or '')] = acc
+
+            restored_accounts = []
+            for aid, meta in task_accounts.items():
+                cur = current_by_id.get(str(aid), {})
+                restored_accounts.append({
+                    'id': meta.get('id') or aid,
+                    'name': meta.get('name') or cur.get('name', ''),
+                    'contact': meta.get('contact'),
+                    'wonWithin7d': bool(meta.get('wonWithin7d', False)),
+                    'onboardingStage': cur.get('onboardingStage', meta.get('onboardingStage')),
+                })
+            state['accounts'] = restored_accounts
+
+
+def _materialize_google_docs(payload):
+    for key in ('initial_state', 'current_state'):
+        state = payload.get(key)
+        if not isinstance(state, dict):
+            continue
+        adapter = state.get('_task_adapter') if isinstance(state.get('_task_adapter'), dict) else {}
+        task_docs = adapter.get('task_documents') if isinstance(adapter.get('task_documents'), dict) else None
+        if not task_docs:
+            continue
+        docs_map = state.get('documents') if isinstance(state.get('documents'), dict) else {}
+        restored = []
+        for did, meta in task_docs.items():
+            cur = docs_map.get(did, {}) if isinstance(docs_map, dict) else {}
+            restored.append({
+                'id': meta.get('id') or did,
+                'title': cur.get('title', meta.get('title')),
+                'vendor': meta.get('vendor'),
+                'folder': meta.get('folder'),
+                'content': cur.get('content', meta.get('content', '')),
+            })
+        state['documents'] = restored
+
+
+def _adapt_payload_for_reward(app, payload):
+    app_name = _normalize_app_name(app)
+    if app_name == 'google_sheets':
+        _materialize_google_sheets(payload)
+    elif app_name == 'slack':
+        _materialize_slack(payload)
+    elif app_name == 'salesforce':
+        _materialize_salesforce(payload)
+    elif app_name == 'google_docs':
+        _materialize_google_docs(payload)
+    return payload
+
+
+def go(app_name):
+    if app_name in _GO_CACHE:
+        return _GO_CACHE[app_name]
+
+    base = APP_URLS.get(app_name)
+    if not base:
+        print(f'CRITICAL: unknown app {app_name!r} (no URL mapped)')
+        print('REWARD: 0.0')
+        sys.exit(0)
+
+    try:
+        resp = requests.get(f'{base}/go?sid={sid}', timeout=15)
+        resp.raise_for_status()
+        payload = resp.json()
+    except Exception as e:
+        print(f'CRITICAL: cannot fetch {app_name} go state: {e}')
+        print('REWARD: 0.0')
+        sys.exit(0)
+
+    payload = _adapt_payload_for_reward(app_name, payload)
+    _GO_CACHE[app_name] = payload
+    return payload
+
+
+# =============================================================================
+# Task-specific reward
+#
+# Answer key (precomputed at setup) lives in salesforce.initial_state._task_adapter:
+#   .expected        -> list of 10 row dicts (8 matched + 2 unmatched). Each matched
+#                       dict carries oppName, account, opportunityId, initial_stage,
+#                       final_stage, task_subject, task_status, task_status_required,
+#                       requires_change, requires_task.
+#   .matched_count   -> n = 8.
+# slack.initial_state._task_adapter:
+#   .backup_channel  -> 'backup';  .required_csv_count -> 2.
+#
+# We score CURRENT state of each mock against this key. The salesforce materializer
+# leaves opportunities/activities/_task_adapter untouched (no task_leads/task_accounts
+# keys present), so we read them in their native shape.
+# =============================================================================
+def reward(go):
+    sf = go('salesforce')
+    sf_init = sf.get('initial_state', {}) if isinstance(sf.get('initial_state'), dict) else {}
+    sf_cur = sf.get('current_state', {}) if isinstance(sf.get('current_state'), dict) else {}
+    adapter = sf_init.get('_task_adapter', {}) if isinstance(sf_init.get('_task_adapter'), dict) else {}
+
+    expected = adapter.get('expected') if isinstance(adapter.get('expected'), list) else []
+    matched = [e for e in expected if isinstance(e, dict) and e.get('matched')]
+
+    n = adapter.get('matched_count')
+    if not isinstance(n, int) or n <= 0:
+        n = len(matched)
+    if n == 0 or not matched:
+        print('DEBUG_T05 no matched rows in answer key — cannot grade')
+        return 0.0
+
+    # --- current SF opportunities indexed by name ---
+    opp_by_name = {}
+    for o in (sf_cur.get('opportunities') or []):
+        if isinstance(o, dict):
+            opp_by_name[norm(o.get('name'))] = o
+
+    # --- current SF activities: normalized subject -> list of normalized statuses ---
+    act_subjects = set()
+    subj_statuses = {}
+    for a in (sf_cur.get('activities') or []):
+        if not isinstance(a, dict):
+            continue
+        s = norm(a.get('subject'))
+        if not s:
+            continue
+        act_subjects.add(s)
+        subj_statuses.setdefault(s, []).append(norm(a.get('status')))
+
+    # Partition matched rows by what the answer key says each REQUIRES.
+    change_rows = [e for e in matched if e.get('requires_change')]
+    keep_stage_rows = [e for e in matched if not e.get('requires_change')]
+    task_rows = [e for e in matched if e.get('requires_task')]
+    notask_rows = [e for e in matched if not e.get('requires_task')]
+    owner_rows = [e for e in matched if e.get('requires_owner_change')]
+    keep_owner_rows = [e for e in matched if not e.get('requires_owner_change')]
+
+    # ---- Component 1 (0.40): stage set correctly, scored ONLY over rows that
+    # require a stage change. Rows already at their target earn NO credit. ----
+    stage_ok = 0
+    for e in change_rows:
+        opp = opp_by_name.get(norm(e.get('oppName')))
+        if opp is not None and norm(opp.get('stage')) == norm(e.get('final_stage')):
+            stage_ok += 1
+    stage_frac = frac(stage_ok, len(change_rows)) if change_rows else 1.0
+
+    # stage-leak penalty: a must-keep-stage row whose stage drifted from injection.
+    stage_leak = 0
+    for e in keep_stage_rows:
+        opp = opp_by_name.get(norm(e.get('oppName')))
+        baseline = e.get('initial_stage')
+        if baseline is None:
+            baseline = e.get('final_stage')
+        if opp is not None and norm(opp.get('stage')) != norm(baseline):
+            stage_leak += 1
+    stage_leak_frac = frac(stage_leak, len(keep_stage_rows)) if keep_stage_rows else 0.0
+
+    # ---- Component 2 (0.25): follow-up Task created, scored ONLY over rows that
+    # require a task. Creating a forbidden task on other rows is penalized. ----
+    task_ok = 0
+    for e in task_rows:
+        want_subj = norm(e.get('task_subject'))
+        present = bool(want_subj) and want_subj in act_subjects
+        if present and e.get('task_status_required'):
+            present = any(st == norm('Not Started') for st in subj_statuses.get(want_subj, []))
+        if present:
+            task_ok += 1
+    task_frac = frac(task_ok, len(task_rows)) if task_rows else 1.0
+
+    # task-leak penalty: a no-task matched row that nonetheless gained a
+    # 'Kickoff: <Account>' / 'Exec review: <Account>' task.
+    task_leak = 0
+    for e in notask_rows:
+        account = e.get('account')
+        forbidden = {norm(f'Kickoff: {account}'), norm(f'Exec review: {account}')}
+        if forbidden & act_subjects:
+            task_leak += 1
+    task_leak_frac = frac(task_leak, len(notask_rows)) if notask_rows else 0.0
+
+    # ---- Component 3 (0.20): owner reassigned, scored ONLY over rows that
+    # require an owner change (Closed Won). Other rows must keep their owner. ----
+    owner_ok = 0
+    for e in owner_rows:
+        opp = opp_by_name.get(norm(e.get('oppName')))
+        want_owner = norm(e.get('final_owner'))
+        if opp is not None and want_owner and norm(opp.get('ownerId')) == want_owner:
+            owner_ok += 1
+    owner_frac = frac(owner_ok, len(owner_rows)) if owner_rows else 1.0
+
+    # owner-leak penalty: a must-keep-owner row whose owner drifted from seeded.
+    owner_leak = 0
+    for e in keep_owner_rows:
+        opp = opp_by_name.get(norm(e.get('oppName')))
+        baseline = norm(e.get('initial_owner'))
+        if opp is not None and baseline and norm(opp.get('ownerId')) != baseline:
+            owner_leak += 1
+    owner_leak_frac = frac(owner_leak, len(keep_owner_rows)) if keep_owner_rows else 0.0
+
+    # ---- Component 4 (0.15): Slack 'backup' channel created + 2 CSV files uploaded ----
+    slack_payload = go('slack')
+    slack_cur = slack_payload.get('current_state', {}) if isinstance(slack_payload.get('current_state'), dict) else {}
+    slack_init = slack_payload.get('initial_state', {}) if isinstance(slack_payload.get('initial_state'), dict) else {}
+    s_adapter = slack_init.get('_task_adapter', {}) if isinstance(slack_init.get('_task_adapter'), dict) else {}
+    backup_channel = s_adapter.get('backup_channel') or 'backup'
+    required_csv = s_adapter.get('required_csv_count')
+    if not isinstance(required_csv, int) or required_csv <= 0:
+        required_csv = 2
+
+    def _channel_exists(state, name):
+        chans = state.get('channels', []) if isinstance(state, dict) else []
+        return any(isinstance(c, dict) and norm(c.get('name')) == norm(name) for c in chans)
+
+    def _is_csv_attachment(att):
+        if not isinstance(att, dict):
+            return False
+        nm = norm(att.get('name') or att.get('fileName') or att.get('filename'))
+        mt = norm(att.get('mimeType') or att.get('mime') or att.get('type'))
+        url = norm(att.get('url'))
+        return nm.endswith('.csv') or 'csv' in mt or url.endswith('.csv')
+
+    # Channel must be NEWLY created (absent at injection, present now).
+    backup_exists_now = _channel_exists(slack_cur, backup_channel)
+    backup_existed_init = _channel_exists(slack_init, backup_channel)
+    channel_ok = bool(backup_exists_now and not backup_existed_init)
+
+    # Count distinct CSV files uploaded into the backup channel.
+    csv_count = 0
+    if backup_exists_now:
+        for m in _slack_channel_messages(slack_cur, backup_channel):
+            for att in (m.get('attachments') or []) if isinstance(m, dict) else []:
+                if _is_csv_attachment(att):
+                    csv_count += 1
+
+    # Sub-score: 0.4 for creating the channel, 0.6 for uploading >= required CSVs
+    # (proportional credit for 1 of 2). Only counts CSVs inside the new channel.
+    slack_ok = 0.0
+    if channel_ok:
+        slack_ok = 0.4 + 0.6 * min(csv_count, required_csv) / required_csv
+
+    # ---- Combine: positives (0 in do-nothing) minus leak penalties (0 in do-nothing) ----
+    # Weights: stage 0.40 + task 0.25 + owner 0.20 + slack 0.15 = 1.0. The Slack
+    # backup step was raised 0.05 -> 0.15 (it now requires creating a channel AND
+    # uploading two exported CSVs), so the SF-core components were trimmed 0.10 total.
+    s_stage = 0.40 * stage_frac
+    s_task = 0.25 * task_frac
+    s_owner = 0.20 * owner_frac
+    s_slack = 0.15 * slack_ok
+    positive = s_stage + s_task + s_owner + s_slack
+    penalty = 0.15 * stage_leak_frac + 0.10 * task_leak_frac + 0.10 * owner_leak_frac
+    score = clamp01(positive - penalty)
+
+    print(
+        'DEBUG_T05 '
+        f'n={n} stage_ok={stage_ok}/{len(change_rows)} stage_leak={stage_leak}/{len(keep_stage_rows)} '
+        f'task_ok={task_ok}/{len(task_rows)} task_leak={task_leak}/{len(notask_rows)} '
+        f'owner_ok={owner_ok}/{len(owner_rows)} owner_leak={owner_leak}/{len(keep_owner_rows)} '
+        f'backup_channel_ok={int(channel_ok)} csv_count={csv_count}/{required_csv} slack_ok={round(slack_ok, 3)} '
+        f'w_stage={round(s_stage, 4)} w_task={round(s_task, 4)} w_owner={round(s_owner, 4)} '
+        f'w_slack={round(s_slack, 4)} penalty={round(penalty, 4)} total={round(score, 4)}'
+    )
+
+    return clamp01(score)
+
+
+try:
+    score = float(reward(go))
+except Exception as exc:
+    print(f'ERROR: reward() raised {exc!r}')
+    score = 0.0
+
+score = clamp01(score)
+print(f'\nREWARD: {round(score, 4)}')

am_renewal_hubspot_004__long/_cua_gym_vm_bridge.sh

@@ -0,0 +1,106 @@
+#!/bin/bash
+# CUA-Gym VM-side network fix (direct-IP mock access).
+#
+# Runs INSIDE the QEMU guest. The CUA-Gym mock services are now reached by
+# their real IDC address (http://28.7.184.198:81xx) instead of the old
+# localhost:9000-9097 socat bridge, so the VM only needs two fixes:
+#
+#   1. Working DNS — the qcow2-baked /etc/resolv.conf points at the Pod's
+#      cluster DNS, which is unreachable from QEMU SLIRP NAT
+#      ("Temporary failure in name resolution").
+#   2. The mock host excluded from the woa http proxy — otherwise the VM's
+#      Chrome (gsettings proxy) and initial_setup.py's `requests` get a 403
+#      from the proxy when hitting 28.7.184.198:81xx. (The old socat path never
+#      hit this because localhost is bypassed by default.)
+#
+# Idempotent. Scoped to the CUA-Gym bridge step; does not touch the generic
+# OSWorld setup pipeline.
+#
+# Argv:
+#   $1  optional sudo password (default "password"; SetupController
+#       substitutes {CLIENT_PASSWORD} when uploaded as a cache file).
+#   $2  unused (kept for backward compat — was socat first port 9000).
+#   $3  unused (kept for backward compat — was socat last  port 9097).
+#   $4  optional mock host IP (default 28.7.184.198).
+
+set +e
+
+PASS="${1:-password}"
+MOCK_HOST="${4:-28.7.184.198}"
+# Derive the /16 IDC subnet so future port/host tweaks on the same segment
+# stay bypassed too (28.7.184.198 -> 28.7.0.0/16).
+MOCK_NET="$(echo "${MOCK_HOST}" | awk -F. 'NF==4{print $1"."$2".0.0/16"}')"
+
+run_root() {
+    echo "$PASS" | sudo -S -p "" bash -c "$1"
+}
+
+# Detect SLIRP gateway from VM's routing table; fall back to 10.0.2.2.
+GW="$(ip route 2>/dev/null | awk '/default/ {print $3; exit}')"
+GW="${GW:-10.0.2.2}"
+
+# ─── CUA-Gym network fix (Taiji-specific) ─────────────────────────────
+fix_vm_network() {
+    # 0) Discover VM network state at runtime — gateway, own IP/CIDR, SLIRP
+    #    DNS forwarder. Nothing hardcoded except the mock host above.
+    local GW_NET="${GW%.*}"
+    local SLIRP_DNS="${GW_NET}.3"        # QEMU SLIRP DNS proxy is always .3
+    local OWN_CIDR
+    OWN_CIDR="$(ip -4 -o addr show scope global 2>/dev/null \
+                | awk '{print $4; exit}')"
+    local OWN_IP="${OWN_CIDR%/*}"
+    local OWN_NET=""
+    [ -n "${OWN_IP}" ] && OWN_NET="${OWN_IP%.*}.0/24"
+
+    # 1) DNS — point at SLIRP DNS proxy + public fallbacks. chattr +i keeps
+    #    systemd-resolved/NetworkManager from clobbering it.
+    if run_root "printf 'nameserver ${SLIRP_DNS}\\nnameserver 114.114.114.114\\nnameserver 8.8.8.8\\n' > /etc/resolv.conf
+                 chattr +i /etc/resolv.conf 2>/dev/null || true"; then
+        echo "[cua_gym_vm_bridge] DNS set: ${SLIRP_DNS}, 114.114.114.114, 8.8.8.8 (gw=${GW})"
+    else
+        echo "[cua_gym_vm_bridge] WARN: failed to write /etc/resolv.conf"
+    fi
+
+    # 2) Build the proxy bypass list from detected values + the direct-IP
+    #    mock host (and its /16 subnet). THIS is what makes 28.7.184.198:81xx
+    #    reachable from inside the VM — without it the woa proxy 403s it.
+    local ignore_list="['localhost', '127.0.0.0/8', '${GW}'"
+    [ -n "${OWN_IP}" ]    && ignore_list="${ignore_list}, '${OWN_IP}'"
+    [ -n "${OWN_NET}" ]   && ignore_list="${ignore_list}, '${OWN_NET}'"
+    [ -n "${GW_NET}" ]    && ignore_list="${ignore_list}, '${GW_NET}.0/24'"
+    [ -n "${MOCK_HOST}" ] && ignore_list="${ignore_list}, '${MOCK_HOST}'"
+    [ -n "${MOCK_NET}" ]  && ignore_list="${ignore_list}, '${MOCK_NET}'"
+    ignore_list="${ignore_list}]"
+
+    # 3) GNOME proxy ignore-hosts — Chrome bypasses the woa proxy for these.
+    if gsettings set org.gnome.system.proxy ignore-hosts "${ignore_list}" 2>/tmp/gsettings_err; then
+        echo "[cua_gym_vm_bridge] gsettings ignore-hosts = ${ignore_list}"
+    else
+        echo "[cua_gym_vm_bridge] WARN: gsettings ignore-hosts failed: $(cat /tmp/gsettings_err 2>/dev/null)"
+    fi
+
+    # 4) Shell env — comma-separated form for no_proxy/NO_PROXY (requests/curl).
+    local NP="localhost,127.0.0.1,${GW}"
+    [ -n "${OWN_IP}" ]    && NP="${NP},${OWN_IP}"
+    [ -n "${OWN_NET}" ]   && NP="${NP},${OWN_NET}"
+    [ -n "${GW_NET}" ]    && NP="${NP},${GW_NET}.0/24"
+    [ -n "${MOCK_HOST}" ] && NP="${NP},${MOCK_HOST}"
+    [ -n "${MOCK_NET}" ]  && NP="${NP},${MOCK_NET}"
+
+    for var in no_proxy NO_PROXY; do
+        if ! grep -q "export ${var}=${NP}" /home/user/.bashrc 2>/dev/null; then
+            echo "export ${var}=${NP}" >> /home/user/.bashrc
+        fi
+    done
+    run_root "grep -q '^no_proxy=' /etc/environment 2>/dev/null \
+              && sed -i 's|^no_proxy=.*|no_proxy=\"${NP}\"|' /etc/environment \
+              || echo 'no_proxy=\"${NP}\"' >> /etc/environment
+              grep -q '^NO_PROXY=' /etc/environment 2>/dev/null \
+              && sed -i 's|^NO_PROXY=.*|NO_PROXY=\"${NP}\"|' /etc/environment \
+              || echo 'NO_PROXY=\"${NP}\"' >> /etc/environment" >/dev/null 2>&1
+    echo "[cua_gym_vm_bridge] no_proxy=${NP}"
+}
+
+fix_vm_network
+
+echo "[cua_gym_vm_bridge] direct-IP mode: ${MOCK_HOST} (+${MOCK_NET}) bypassed from woa proxy; no socat forwarding needed"

am_renewal_hubspot_004__long/initial_setup.py

@@ -0,0 +1,688 @@
+"""
+Initial Setup: T04 — HubSpot renewal outreach across HubSpot + Gmail + Google Calendar
+                (hidden-access-point edition)
+Task ID: am_renewal_hubspot_004__long
+Mocks: hubspot_mock, gmail_mock, google_calendar_mock
+
+Assume today is 2026-05-04 (a Monday). A deal QUALIFIES for renewal outreach iff
+ALL hold:
+  - stage == 'closed_won'  (label "Closed Won")
+  - dealType == 'existing_business'
+  - 2025-05-01 <= closeDate <= 2025-07-31  (inclusive)
+  - amount >= 10000
+
+For every QUALIFYING deal the agent must perform SIX sub-actions across the three
+apps, several of which are reachable only through HARD-TO-FIND ("hidden") controls:
+
+  HubSpot  (both edits live in the slide-in panel opened by clicking the deal NAME):
+    (H1) append the line 'Renewal outreach sent 2026-05-04.' to the deal Description
+         (keeping the existing text).
+    (H2) reassign the deal owner — the free-text "Assigned To" field — from
+         'Admin User' to 'Renewals Team'.
+  Gmail:
+    (G1) send an email to the deal's primary contact with Subject 'Renewal: <Company>'
+         and a body containing the company name, the original amount formatted as
+         $#,##0, and the original closeDate. (Obstacle: the email has to be found in HubSpot's contacts list. The contact is in the deal)
+    (G2) apply the 'Renewals' label to that sent email. The Label button only appears
+         in the toolbar AFTER an email row is selected (selection-gated control).
+  Google Calendar:
+    (C1) create a 'Renewal call: <Company>' event one week out (2026-05-11).
+    (C2) add the contact's email as a GUEST. The guest field is buried inside the
+         event editor and the contact is NOT in the autocomplete list, so the email
+         must be typed in.
+
+Distractors (5 near-miss deals) each miss exactly ONE qualifying condition and must
+be left untouched (no description edit, no owner change, no email, no label, no event).
+
+Ground truth is PRECOMPUTED below in Python so the embedded answer key (hidden `_*`
+fields on hubspot deals) is internally consistent with the visible deal/company/
+contact data.  reward.py reads `deals[*]._qualifies` (+ helper `_*` fields) from
+initial_state as the sole source of truth.
+"""
+import os
+import shlex
+import subprocess
+import time
+import uuid
+from datetime import date
+
+import requests
+
+
+sid = str(uuid.uuid4())
+with open('/tmp/task_web_sid', 'w', encoding='utf-8') as f:
+    f.write(sid)
+print(f'Generated sid: {sid}')
+
+
+# --------------------------------------------------------------------------- #
+# Task constants
+# --------------------------------------------------------------------------- #
+TODAY = date(2026, 5, 4)
+TODAY_STR = TODAY.isoformat()                     # '2026-05-04'
+RENEWAL_LINE = f'Renewal outreach sent {TODAY_STR}.'
+WINDOW_START = date(2025, 5, 1)
+WINDOW_END = date(2025, 7, 31)
+MIN_AMOUNT = 10000
+
+NEW_OWNER = 'Renewals Team'                        # H2: owner reassignment target
+ORIG_OWNER = 'Admin User'                          # owner at injection
+RENEWALS_LABEL_ID = 'l3'                            # G2: label to apply (seeded below)
+RENEWALS_LABEL_NAME = 'Renewals'
+EVENT_DATE_STR = '2026-05-11'                       # C1: one week from today (a Monday)
+
+
+# --------------------------------------------------------------------------- #
+# Builder helpers (keep many similar records terse)
+# --------------------------------------------------------------------------- #
+def company(cid, name, domain, industry, city, st, employees, revenue):
+    return {
+        'id': cid, 'name': name, 'domain': domain, 'industry': industry,
+        'phone': f'+1 (555) 1{cid[-1]}0-00{cid[-1]}0',
+        'city': city, 'state': st, 'country': 'United States',
+        'numberOfEmployees': employees, 'annualRevenue': revenue,
+        'lifecycleStage': 'customer', 'owner': ORIG_OWNER,
+        'description': f'{name} — {industry} customer account',
+        'createDate': '2023-08-10T09:00:00Z',
+    }
+
+
+def contact(cid, first, last, email, title, company_id, city, st):
+    return {
+        'id': cid, 'firstName': first, 'lastName': last, 'email': email,
+        'phone': '+1 (555) 010-0000', 'jobTitle': title, 'companyId': company_id,
+        'lifecycleStage': 'customer', 'leadStatus': 'connected',
+        'owner': ORIG_OWNER, 'city': city, 'state': st, 'country': 'United States',
+        'createDate': '2023-08-15T10:30:00Z', 'lastActivityDate': '2026-04-20T14:00:00Z',
+        'timeline': [],
+    }
+
+
+def deal(did, name, stage, amount, close_date, deal_type, company_id, contact_ids,
+         description, priority='medium'):
+    return {
+        'id': did, 'name': name, 'stage': stage, 'amount': amount,
+        'closeDate': close_date, 'dealType': deal_type, 'priority': priority,
+        'owner': ORIG_OWNER, 'companyId': company_id, 'contactIds': contact_ids,
+        'probability': 100 if stage == 'closed_won' else (0 if stage == 'closed_lost' else 60),
+        'description': description,
+        'createDate': '2024-04-01T10:00:00Z', 'lastActivityDate': '2025-07-15T14:00:00Z',
+    }
+
+
+def gmail_email(eid, frm, to, subject, body, folder='inbox', timestamp='2026-04-30T09:00:00Z'):
+    return {
+        'id': eid, 'threadId': f'thread_{eid}',
+        'from': frm, 'to': to, 'cc': [], 'bcc': [],
+        'subject': subject, 'body': body, 'snippet': body[:120],
+        'timestamp': timestamp, 'read': True, 'starred': False, 'important': False,
+        'labels': [], 'category': 'primary', 'folder': folder, 'attachments': [],
+    }
+
+
+def cal_event(eid, calendar_id, title, start_iso, end_iso, location, description, guests):
+    """A calendar event in the shape google_calendar_mock stores them."""
+    return {
+        'id': eid, 'calendarId': calendar_id, 'title': title,
+        'start': start_iso, 'end': end_iso,
+        'allDay': False, 'location': location, 'description': description,
+        'guests': guests, 'color': '#33B679', 'recurring': 'none',
+        'reminders': [{'type': 'popup', 'minutes': 10}],
+    }
+
+
+# --------------------------------------------------------------------------- #
+# Companies (one per deal; comp1..comp4 are the qualifying accounts)
+# --------------------------------------------------------------------------- #
+companies = [
+    company('comp1', 'Acme', 'acme.com', 'Technology', 'San Francisco', 'CA', 250, 15000000),
+    company('comp2', 'Globex', 'globex.com', 'Manufacturing', 'New York', 'NY', 500, 32000000),
+    company('comp3', 'Initech', 'initech.com', 'Technology', 'Austin', 'TX', 120, 8000000),
+    company('comp4', 'Umbrella', 'umbrella.com', 'Healthcare', 'Boston', 'MA', 800, 55000000),
+    company('comp5', 'Wayne', 'wayne.com', 'Finance', 'Chicago', 'IL', 1200, 90000000),
+    company('comp6', 'Hooli', 'hooli.com', 'Technology', 'Palo Alto', 'CA', 900, 60000000),
+    company('comp7', 'Stark', 'stark.com', 'Manufacturing', 'Los Angeles', 'CA', 1500, 120000000),
+    company('comp8', 'Cyberdyne', 'cyberdyne.com', 'Technology', 'Sunnyvale', 'CA', 700, 40000000),
+    company('comp9', 'Wonka', 'wonka.com', 'Other', 'Chicago', 'IL', 300, 18000000),
+]
+
+# Contacts. Each QUALIFYING company (comp1..comp4) has exactly ONE contact, so
+# "first contact whose companyId matches" is unambiguous on the graded slice.
+# Non-qualifying companies carry a couple of extra decoy contacts (realism).
+contacts = [
+    contact('c1', 'Alice', 'Chen', 'alice.chen@acme.com', 'VP Operations', 'comp1', 'San Francisco', 'CA'),
+    contact('c2', 'Bob', 'Martin', 'bob.martin@globex.com', 'Procurement Lead', 'comp2', 'New York', 'NY'),
+    contact('c3', 'Carol', 'Diaz', 'carol.diaz@initech.com', 'IT Director', 'comp3', 'Austin', 'TX'),
+    contact('c4', 'Dan', 'Wong', 'dan.wong@umbrella.com', 'Head of Research', 'comp4', 'Boston', 'MA'),
+    contact('c5', 'Eve', 'Parker', 'eve.parker@wayne.com', 'CFO', 'comp5', 'Chicago', 'IL'),
+    contact('c6', 'Frank', 'Lee', 'frank.lee@hooli.com', 'Eng Manager', 'comp6', 'Palo Alto', 'CA'),
+    contact('c7', 'Grace', 'Kim', 'grace.kim@stark.com', 'Plant Director', 'comp7', 'Los Angeles', 'CA'),
+    contact('c8', 'Henry', 'Cole', 'henry.cole@cyberdyne.com', 'CTO', 'comp8', 'Sunnyvale', 'CA'),
+    contact('c9', 'Iris', 'Shah', 'iris.shah@wonka.com', 'COO', 'comp9', 'Chicago', 'IL'),
+    # decoy second contacts on NON-qualifying accounts only
+    contact('c10', 'Mia', 'Stone', 'mia.stone@wayne.com', 'Controller', 'comp5', 'Chicago', 'IL'),
+    contact('c11', 'Leo', 'Burns', 'leo.burns@stark.com', 'Ops Analyst', 'comp7', 'Los Angeles', 'CA'),
+]
+
+
+# --------------------------------------------------------------------------- #
+# Deals — 4 qualifying + 5 near-miss distractors
+#   d1 Acme       QUALIFIES (existing, 2025-05-15, $24,000)
+#   d2 Globex     QUALIFIES (existing, 2025-06-30, $15,000)
+#   d3 Initech    QUALIFIES (existing, 2025-07-31 high boundary, $50,000)
+#   d4 Umbrella   QUALIFIES (existing, 2025-05-01 low boundary, $12,000)
+#   d5 Wayne      NO  -> new_business (right window+stage, $30,000)
+#   d6 Hooli      NO  -> amount 8000 < 10000 (existing, in window)
+#   d7 Stark      NO  -> closeDate 2025-04-15 just BEFORE window
+#   d8 Cyberdyne  NO  -> stage closed_lost (existing, in window, $35,000)
+#   d9 Wonka      NO  -> closeDate 2025-08-15 just AFTER window
+# --------------------------------------------------------------------------- #
+deals = [
+    deal('d1', 'Acme - Platform Subscription', 'closed_won', 24000, '2025-05-15',
+         'existing_business', 'comp1', ['c1'],
+         'Annual platform subscription, 250 seats. Renewed last cycle without issues.', 'high'),
+    deal('d2', 'Globex - Support Contract', 'closed_won', 15000, '2025-06-30',
+         'existing_business', 'comp2', ['c2'],
+         'Gold support contract covering all manufacturing sites.', 'medium'),
+    deal('d3', 'Initech - Enterprise Plan', 'closed_won', 50000, '2025-07-31',
+         'existing_business', 'comp3', ['c3'],
+         'Enterprise plan with premium SLA and onboarding.', 'high'),
+    deal('d4', 'Umbrella - Compliance Module', 'closed_won', 12000, '2025-05-01',
+         'existing_business', 'comp4', ['c4'],
+         'Compliance reporting module add-on for research division.', 'medium'),
+    # ---- distractors ----
+    deal('d5', 'Wayne - New Analytics Suite', 'closed_won', 30000, '2025-06-10',
+         'new_business', 'comp5', ['c5'],
+         'First-time purchase of the analytics suite for finance ops.', 'high'),
+    deal('d6', 'Hooli - Small Add-on', 'closed_won', 8000, '2025-06-20',
+         'existing_business', 'comp6', ['c6'],
+         'Minor seat expansion add-on, below renewal threshold.', 'low'),
+    deal('d7', 'Stark - Spring Renewal', 'closed_won', 40000, '2025-04-15',
+         'existing_business', 'comp7', ['c7'],
+         'Renewal that closed in April, outside the summer window.', 'high'),
+    deal('d8', 'Cyberdyne - Lost Renewal', 'closed_lost', 35000, '2025-06-05',
+         'existing_business', 'comp8', ['c8'],
+         'Renewal that lapsed; customer churned to a competitor.', 'medium'),
+    deal('d9', 'Wonka - Late Summer Renewal', 'closed_won', 22000, '2025-08-15',
+         'existing_business', 'comp9', ['c9'],
+         'Renewal closing in late August, just after the window.', 'medium'),
+]
+
+
+# --------------------------------------------------------------------------- #
+# PRECOMPUTE the ground truth and embed hidden `_*` answer-key fields.
+# --------------------------------------------------------------------------- #
+def parse_date(s):
+    y, m, d = (int(x) for x in s.split('-'))
+    return date(y, m, d)
+
+
+def primary_contact_email(company_id):
+    """companies[].name -> contacts where companyId matches, first such email."""
+    for c in contacts:
+        if c['companyId'] == company_id:
+            return c['email']
+    return None
+
+
+def fmt_amount(amount):
+    """Original amount formatted as $#,##0  (e.g. 24000 -> '$24,000')."""
+    return f'${amount:,}'
+
+
+n_qualifying = 0
+for d in deals:
+    cd = parse_date(d['closeDate'])
+    qualifies = (
+        d['stage'] == 'closed_won'
+        and d['dealType'] == 'existing_business'
+        and WINDOW_START <= cd <= WINDOW_END
+        and d['amount'] >= MIN_AMOUNT
+    )
+    company_name = next(c['name'] for c in companies if c['id'] == d['companyId'])
+    contact_email = primary_contact_email(d['companyId'])
+    d['_qualifies'] = qualifies
+    d['_company_name'] = company_name
+    d['_primary_contact_email'] = contact_email
+    d['_orig_owner'] = ORIG_OWNER
+    d['_email_subject'] = f'Renewal: {company_name}'              # ALL deals (leak shape)
+    d['_event_title'] = f'Renewal call: {company_name}'           # ALL deals (leak shape)
+    if qualifies:
+        n_qualifying += 1
+        # --- HubSpot answer key ---
+        d['_renewal_line'] = RENEWAL_LINE
+        d['_expected_description'] = d['description'] + '\n' + RENEWAL_LINE
+        d['_new_owner'] = NEW_OWNER
+        # --- Gmail answer key ---
+        d['_amount_fmt'] = fmt_amount(d['amount'])
+        d['_close_date'] = d['closeDate']
+        # Body must contain ALL of these substrings.
+        d['_email_body_must_contain'] = [company_name, fmt_amount(d['amount']), d['closeDate']]
+        d['_renewals_label_id'] = RENEWALS_LABEL_ID
+        d['_renewals_label_name'] = RENEWALS_LABEL_NAME
+        # --- Calendar answer key ---
+        d['_event_date'] = EVENT_DATE_STR
+        d['_event_guest_email'] = contact_email
+    else:
+        # Distractors must receive NO edits, NO email, NO label, NO event.
+        d['_renewal_line'] = None
+
+assert n_qualifying == 4, f'expected 4 qualifying deals, got {n_qualifying}'
+print(f'Precomputed ground truth: {n_qualifying} qualifying deals '
+      f'({[d["id"] for d in deals if d["_qualifies"]]})')
+
+
+# Standard 7-stage pipeline configuration (default labels; "Closed Won" column).
+DEAL_STAGES = {
+    'appointment_scheduled': {'id': 'appointment_scheduled', 'label': 'Appointment Scheduled', 'probability': 20, 'color': '#E5F4FF', 'order': 1},
+    'qualified_to_buy': {'id': 'qualified_to_buy', 'label': 'Qualified to Buy', 'probability': 40, 'color': '#FFF0E6', 'order': 2},
+    'presentation_scheduled': {'id': 'presentation_scheduled', 'label': 'Presentation Scheduled', 'probability': 60, 'color': '#FFF8E6', 'order': 3},
+    'decision_maker_bought_in': {'id': 'decision_maker_bought_in', 'label': 'Decision Maker Bought-In', 'probability': 80, 'color': '#E8F5E9', 'order': 4},
+    'contract_sent': {'id': 'contract_sent', 'label': 'Contract Sent', 'probability': 90, 'color': '#E6FFFA', 'order': 5},
+    'closed_won': {'id': 'closed_won', 'label': 'Closed Won', 'probability': 100, 'color': '#E6FFEC', 'order': 6},
+    'closed_lost': {'id': 'closed_lost', 'label': 'Closed Lost', 'probability': 0, 'color': '#FFE6E6', 'order': 7},
+}
+TICKET_STATUSES = {
+    'new': {'id': 'new', 'label': 'New', 'color': '#E5F4FF', 'order': 1},
+    'waiting_on_contact': {'id': 'waiting_on_contact', 'label': 'Waiting on Contact', 'color': '#FFF8E6', 'order': 2},
+    'waiting_on_us': {'id': 'waiting_on_us', 'label': 'Waiting on Us', 'color': '#FFF0E6', 'order': 3},
+    'in_progress': {'id': 'in_progress', 'label': 'In Progress', 'color': '#E6FFFA', 'order': 4},
+    'closed': {'id': 'closed', 'label': 'Closed', 'color': '#E6FFEC', 'order': 5},
+}
+
+
+# --------------------------------------------------------------------------- #
+# Gmail account (sales rep doing the outreach). NO renewal mail pre-seeded:
+# the 'sent' folder contains no 'Renewal: <Company>' subjects (Rule 3). A few
+# read inbox emails act as watermark / realism only. A 'Renewals' label is
+# pre-created (l3) so the agent can APPLY it; no email carries it at injection.
+# --------------------------------------------------------------------------- #
+gmail_user = {
+    'userId': 'u1', 'username': 'Jordan Avery',
+    'email': 'jordan.avery@ourcompany.com',
+    'avatar': 'https://picsum.photos/200/200?random=7',
+}
+gmail_emails = [
+    gmail_email('m1',
+                {'name': 'Alice Chen', 'email': 'alice.chen@acme.com', 'avatar': ''},
+                [{'name': 'Jordan Avery', 'email': 'jordan.avery@ourcompany.com'}],
+                'Re: Q1 usage report',
+                'Thanks for the usage report — numbers look healthy on our end.',
+                folder='inbox', timestamp='2026-04-28T09:15:00Z'),
+    gmail_email('m2',
+                {'name': 'Billing', 'email': 'billing@globex.com', 'avatar': ''},
+                [{'name': 'Jordan Avery', 'email': 'jordan.avery@ourcompany.com'}],
+                'Invoice received',
+                'Confirming we received the latest support invoice. No action needed.',
+                folder='inbox', timestamp='2026-04-29T11:00:00Z'),
+    gmail_email('m3',
+                {'name': 'Jordan Avery', 'email': 'jordan.avery@ourcompany.com', 'avatar': ''},
+                [{'name': 'Carol Diaz', 'email': 'carol.diaz@initech.com'}],
+                'Onboarding follow-up',
+                'Glad onboarding went smoothly. Reach out anytime with questions.',
+                folder='sent', timestamp='2026-04-22T16:30:00Z'),
+]
+gmail_labels = [
+    {'id': 'l1', 'name': 'Work', 'color': '#ef4444'},
+    {'id': 'l2', 'name': 'Personal', 'color': '#3b82f6'},
+    {'id': RENEWALS_LABEL_ID, 'name': RENEWALS_LABEL_NAME, 'color': '#16a34a'},
+]
+
+
+# --------------------------------------------------------------------------- #
+# Google Calendar account. today = 2026-05-04. A couple of decoy events for
+# realism on the Work calendar; NONE titled 'Renewal call:' (Rule 3 — the
+# gradable output must be absent at injection).
+# --------------------------------------------------------------------------- #
+# NOTE: use timezone-NAIVE local ISO (no trailing 'Z') to match the shape the
+# calendar mock stores when the user creates events through the UI. Mixing
+# UTC-suffixed injected events with naive-local user-created events makes the
+# newly created event appear on a different day/time-slot in the week grid
+# (or vanish entirely from the visible viewport). See the working reference
+# in demo_tasks/.../975bd1cb-.../initial_setup.py.
+calendar_events = [
+    cal_event('evt_standup', 'c2', 'Team Standup',
+              '2026-05-04T09:30:00', '2026-05-04T10:00:00',
+              'Conference Room A', 'Daily sync', ['teammate@ourcompany.com']),
+    cal_event('evt_pipeline', 'c2', 'Pipeline Review',
+              '2026-05-06T14:00:00', '2026-05-06T15:00:00',
+              'Zoom', 'Weekly pipeline review', []),
+    cal_event('evt_lunch', 'c1', 'Lunch with Sarah',
+              '2026-05-05T12:00:00', '2026-05-05T13:00:00',
+              'Downtown Cafe', '', []),
+]
+
+
+# --------------------------------------------------------------------------- #
+# APP_STATES — full key set per mock; answer key embedded in hubspot.deals[*].
+# --------------------------------------------------------------------------- #
+APP_STATES = [
+    ('http://28.7.184.198:8150', {  # hubspot_mock
+        'contacts': contacts,
+        'companies': companies,
+        'deals': deals,
+        'tickets': [],
+        'tasks': [],
+        'notes': [],
+        'templates': [],
+        'meetings': [],
+        'forms': [],
+        'dealStages': DEAL_STAGES,
+        'ticketStatuses': TICKET_STATUSES,
+        'appState': {
+            'sidebarOpen': True,
+            'currentUser': {'name': 'Admin User', 'email': 'admin@example.com', 'avatar': None},
+        },
+        '_task_adapter': {'task_id': 'T04_hubspot_renewal_outreach_3app', 'variant': 'eval'},
+    }),
+    ('http://28.7.184.198:8138', {  # gmail_mock — sent folder has NO renewal mail
+        'user': gmail_user,
+        'emails': gmail_emails,
+        'labels': gmail_labels,
+        'drafts': [],
+        'settings': {
+            'density': 'default', 'undoSend': 10,
+            'categoryTabs': {'primary': True, 'social': True, 'promotions': True,
+                             'updates': False, 'forums': False},
+        },
+        'today': TODAY_STR,
+    }),
+    ('http://28.7.184.198:8141', {  # google_calendar_mock — no 'Renewal call:' events
+        'user': {'id': 'u1', 'username': 'Jordan Avery',
+                 'email': 'jordan.avery@ourcompany.com',
+                 'avatar': 'https://picsum.photos/100/100?random=u1'},
+        # WeekView renders event chips with backgroundColor=<event.color OR
+        # calendar.color>. The React component only accepts hex strings ('#RRGGBB');
+        # tailwind class names like 'bg-green-500' are invalid CSS colors, so the
+        # event blocks render but are invisible on the grid (though the click
+        # popup still shows the correct data). Use hex to match the working
+        # setup in demo_tasks/.../975bd1cb-.../initial_setup.py.
+        'calendars': [
+            {'id': 'c1', 'name': 'Personal', 'color': '#039BE5',
+             'textColor': 'text-white', 'visible': True, 'userId': 'u1',
+             'isDefault': True},
+            {'id': 'c2', 'name': 'Work', 'color': '#33B679',
+             'textColor': 'text-white', 'visible': True, 'userId': 'u1',
+             'isDefault': False},
+        ],
+        'events': calendar_events,
+        'view': 'week',
+        # naive local ISO (no trailing Z), matches events[*].start above
+        'currentDate': '2026-05-04T00:00:00',
+        'sidebarOpen': True,
+        'settings': {
+            'weekStart': 0, 'defaultDuration': 60, 'defaultView': 'week',
+            'defaultReminder': {'type': 'popup', 'minutes': 10},
+        },
+        'today': TODAY_STR,
+    }),
+]
+
+
+for app_url, state in APP_STATES:
+    resp = requests.post(
+        f'{app_url}/post?sid={sid}',
+        json={'action': 'set', 'state': state},
+        timeout=30,
+    )
+    assert resp.status_code == 200, (
+        f'State injection failed for {app_url}: {resp.text}'
+    )
+    print(f'State injected: {app_url} sid={sid}')
+
+for app_url, _ in APP_STATES:
+    go_payload = requests.get(f'{app_url}/go?sid={sid}', timeout=15).json()
+    assert go_payload.get('initial_state') is not None, (
+        f'initial_state is None after injection for {app_url}'
+    )
+    print(f'Verified initial_state set: {app_url}')
+
+
+# PATCHED_BY: patch_wait_mocks_loaded.py
+def launch_gui(command, delay_sec=1.0):
+    """Launch a GUI command, fully detached from this script.
+
+    Inject --remote-debugging-port=1337 (and Chrome quiet flags) so that
+    wait_mocks_loaded's CDP phase can verify browser render below.
+    Idempotent: if the caller already specified a debug port, keep it.
+    """
+    env = os.environ.copy()
+    env['DISPLAY'] = ':0'
+    if isinstance(command, str):
+        _parts = shlex.split(command)
+    else:
+        _parts = list(command)
+    if (_parts and 'google-chrome' in _parts[0]
+            and not any(p.startswith('--remote-debugging-port') for p in _parts)):
+        _parts[1:1] = [
+            '--no-first-run',
+            '--no-default-browser-check',
+            '--remote-debugging-port=1337',
+        ]
+    subprocess.Popen(
+        _parts,
+        stdin=subprocess.DEVNULL,
+        stdout=subprocess.DEVNULL,
+        stderr=subprocess.DEVNULL,
+        env=env,
+        start_new_session=True,
+    )
+    time.sleep(delay_sec)
+
+
+def wait_mocks_loaded(app_states, sid_value, timeout=90, render_buffer=12):
+    """2-phase readiness wait, mirrors cache/ae_pipeline_hygiene_004 style.
+
+    Phase 1 - Backend readiness:
+        Poll each mock's /state?sid=<sid> (preferred) or /go?sid=<sid>
+        (fallback) until it returns 200 with 'stored_state' or
+        'initial_state' respectively, or until timeout.
+
+    Phase 2 - Browser render verification (via CDP + Playwright):
+        Connect to http://localhost:1337 (the --remote-debugging-port
+        injected by launch_gui above), wait for every open tab to reach
+        'networkidle' AND to contain at least 20 DOM elements. This
+        guarantees the SPA fetched /go?sid= data AND actually painted,
+        not just that the mock backend is up.
+
+    Fallback:
+        If Playwright is missing or CDP cannot be reached, sleep for
+        `render_buffer` seconds as a best-effort delay so the screenshot
+        loop does not start on a still-blank tab.
+    """
+    import urllib.request as _urlreq
+    import time as _t
+    _urls = [p for p, _ in app_states]
+    _opener = _urlreq.build_opener(_urlreq.ProxyHandler({}))
+
+    # ---- Phase 1: backend readiness ----
+    _deadline = _t.time() + timeout
+    _todo = list(_urls)
+    while _todo and _t.time() < _deadline:
+        _rest = []
+        for _u in _todo:
+            _ok = False
+            for _path, _need in (('/state?sid=', b'stored_state'),
+                                 ('/go?sid=',    b'initial_state')):
+                try:
+                    _r = _opener.open(f"{_u}{_path}{sid_value}", timeout=5)
+                    if _r.status == 200:
+                        _body = _r.read()
+                        if _need in _body:
+                            _ok = True
+                            break
+                except Exception:
+                    pass
+            if _ok:
+                print(f"  {_u}: backend ready")
+                continue
+            _rest.append(_u)
+        _todo = _rest
+        if _todo:
+            _t.sleep(2)
+    _done = len(_urls) - len(_todo)
+    _pend = f" (pending: {_todo})" if _todo else ""
+    print(f"  mocks backend ready: {_done}/{len(_urls)}{_pend}")
+
+    # ---- Phase 2: browser render verification via CDP ----
+    _cdp_ok = False
+    try:
+        from playwright.sync_api import sync_playwright
+        _cdp_dl = _t.time() + max(15.0, _deadline - _t.time())
+        with sync_playwright() as _p:
+            _br = None
+            while _t.time() < _cdp_dl:
+                try:
+                    _br = _p.chromium.connect_over_cdp('http://localhost:1337')
+                    break
+                except Exception:
+                    _t.sleep(0.5)
+            if _br is None:
+                print("  wait_mocks_loaded: CDP connect to :1337 failed; is Chrome up with --remote-debugging-port?")
+            else:
+                _n_expected = len(_urls)
+                _pages = []
+                _pg_dl = _t.time() + 15
+                _last_count = -1
+                _stable_since = None
+                while _t.time() < _pg_dl:
+                    _pages = []
+                    for _c in _br.contexts:
+                        _pages.extend(_c.pages)
+                    _cnt = len(_pages)
+                    if _n_expected > 0 and _cnt >= _n_expected:
+                        break
+                    if _cnt > 0 and _cnt == _last_count:
+                        if _stable_since is None:
+                            _stable_since = _t.time()
+                        elif _t.time() - _stable_since >= 1.5:
+                            break
+                    else:
+                        _stable_since = None
+                    _last_count = _cnt
+                    _t.sleep(0.5)
+                _pages = []
+                for _c in _br.contexts:
+                    _pages.extend(_c.pages)
+                if not _pages:
+                    print("  wait_mocks_loaded: CDP connected but no open tabs found")
+                else:
+                    print(f"  wait_mocks_loaded: {len(_pages)} tab(s) open (expected {_n_expected}); polling render")
+                    _MIN_ELEMS = 20
+                    _verify_dl = _t.time() + 40.0   # bounded so total setup stays < VM 120s cap
+                    for _pg in _pages:              # bring each to front once to defeat background-tab render throttling
+                        try:
+                            _pg.bring_to_front()
+                        except Exception:
+                            pass
+                    _pending = list(_pages)
+                    while _pending and _t.time() < _verify_dl:
+                        _still = []
+                        for _pg in _pending:
+                            _ttl = ""
+                            try:
+                                _ttl = _pg.title()
+                            except Exception:
+                                pass
+                            try:
+                                _n = _pg.evaluate("document.body ? document.querySelectorAll('*').length : 0")
+                            except Exception:
+                                _n = 0
+                            if _n >= _MIN_ELEMS:
+                                print(f"  wait_mocks_loaded: tab {_ttl!r} rendered ({_n} elements)")
+                            else:
+                                _still.append(_pg)
+                        _pending = _still
+                        if _pending:
+                            _t.sleep(2)
+                    if not _pending:
+                        _cdp_ok = True
+                    else:
+                        for _pg in _pending:
+                            _ttl = ""
+                            try:
+                                _ttl = _pg.title()
+                            except Exception:
+                                pass
+                            print(f"  wait_mocks_loaded: WARNING tab {_ttl!r} still not rendered after poll, will fall back")
+    except ImportError:
+        print("  wait_mocks_loaded: playwright not installed; skipping CDP phase")
+    except Exception as _e:
+        print(f"  wait_mocks_loaded: CDP phase error ({_e!r})")
+
+    if not _cdp_ok:
+        print(f"  wait_mocks_loaded: falling back to render_buffer={render_buffer}s sleep")
+        _t.sleep(render_buffer)
+    else:
+        print("  wait_mocks_loaded: browser render verified, skipping render_buffer")
+
+
+def _open_mock_tabs(app_states, sid_value):
+    """Open the first mock as the primary Chrome window, the rest as new tabs.
+
+    Uses launch_gui once per URL so that the macOS localizer regex-extract
+    (which grabs the FIRST quoted URL per call) still works correctly.
+    """
+    _urls = [p for p, _ in app_states]
+    if not _urls:
+        return
+    launch_gui(f'google-chrome "{_urls[0]}/?sid={sid_value}"', delay_sec=2.0)
+    for _u in _urls[1:]:
+        launch_gui(f'google-chrome --new-tab "{_u}/?sid={sid_value}"', delay_sec=1.0)
+
+
+
+# PATCHED_BY: add_wait_for_mocks_loaded (colleague headless+playwright render gate)
+def wait_for_mocks_loaded(mocks, sid, timeout=60.0, render_timeout=15.0):
+    """Block until every mock serves injected state AND (best-effort) renders."""
+    import time as _t
+    deadline = _t.time() + timeout
+    pending = dict(mocks)
+    while pending and _t.time() < deadline:
+        for name in list(pending):
+            try:
+                r = requests.get(f'{pending[name]}/go?sid={sid}', timeout=5)
+                if r.status_code == 200 and r.json().get('initial_state') is not None:
+                    del pending[name]
+            except Exception:
+                pass
+        if pending:
+            _t.sleep(1.0)
+    if pending:
+        print(f'WARN: mocks not ready within {timeout}s: {list(pending)}')
+    try:
+        from playwright.sync_api import sync_playwright
+    except ImportError:
+        print('WARN: playwright not installed; skipping render check, settle 3s')
+        _t.sleep(3.0)
+        return
+    with sync_playwright() as p:
+        try:
+            browser = p.chromium.launch(channel='chrome', headless=True,
+                                        args=['--no-sandbox', '--disable-gpu', '--disable-dev-shm-usage'])
+        except Exception as e:
+            print(f'WARN: could not launch chrome for render check: {e}; settle 3s')
+            _t.sleep(3.0)
+            return
+        try:
+            for name, url in mocks.items():
+                page = browser.new_page()
+                try:
+                    page.goto(f'{url}/?sid={sid}', wait_until='domcontentloaded', timeout=int(render_timeout * 1000))
+                    page.wait_for_function(
+                        "document.body && document.body.innerText.length > 50",
+                        timeout=int(render_timeout * 1000))
+                    print(f'[{name}] rendered OK')
+                except Exception as e:
+                    print(f'WARN: [{name}] render check issue: {e}')
+                finally:
+                    page.close()
+        finally:
+            browser.close()
+
+
+# ref pattern (out_benchmark_cache): open visible tabs FIRST, then block until mocks render
+_open_mock_tabs(APP_STATES, sid)
+wait_for_mocks_loaded({u.rsplit(':', 1)[-1]: u for u, _ in APP_STATES}, sid, timeout=60, render_timeout=15)
+wait_mocks_loaded(APP_STATES, sid, timeout=90, render_buffer=12)
+print(f'GUI_READY: launched browser tabs for {len(APP_STATES)} apps')

am_renewal_hubspot_004__long/reward.py

@@ -0,0 +1,727 @@
+"""
+Reward Script: T04 — HubSpot renewal outreach across HubSpot + Gmail + Google Calendar
+                (hidden-access-point edition)
+Task ID: am_renewal_hubspot_004__long
+Mocks: hubspot_mock,gmail_mock,google_calendar_mock
+Scoring (per qualifying deal unless noted; frac over the qualifying set).
+Positive components sum to 1.0 and are each 0 in a do-nothing run; leakage is a
+PENALTY (subtracted), so a do-nothing agent scores exactly 0.0:
+  0.22  HubSpot: current description endswith _renewal_line AND still contains the
+        original text                                            (hidden: slide-in panel)
+  0.17  HubSpot: deal owner reassigned from 'Admin User' to 'Renewals Team'
+                                                                 (hidden: "Assigned To" field)
+  0.22  Gmail:  a sent email exists with subject == 'Renewal: <Company>' and body
+        contains company + $amount + closeDate
+  0.17  Gmail:  that renewal email carries the 'Renewals' label (hidden: selection-gated
+        Label button; matched by label NAME so an agent-created label also counts)
+  0.11  Calendar: a 'Renewal call: <Company>' event exists on 2026-05-11
+        (hidden: must create event; TZ-robust date match)
+  0.11  Calendar: that renewal-call event lists the contact's email as a guest
+        (hidden: guest field buried in the event editor; contact not in the picker)
+  -0.15 PENALTY (max): scaled by the fraction of non-qualifying deals that were
+        wrongly touched (description/owner edit, 'Renewal:' email, or 'Renewal call:' event).
+Answer key: hubspot_mock initial_state deals[*] hidden `_*` fields (sole source of truth).
+"""
+import copy
+import re
+import sys
+
+import requests
+
+from datetime import datetime, timezone, timedelta
+
+TZ_PLUS_8 = timezone(timedelta(hours=8))
+
+
+try:
+    with open('/tmp/task_web_sid', encoding='utf-8') as f:
+        sid = f.read().strip()
+    if not sid:
+        raise ValueError('sid is empty')
+except Exception as e:
+    print(f'CRITICAL: Cannot read sid from /tmp/task_web_sid: {e}')
+    print('REWARD: 0.0')
+    sys.exit(0)
+
+
+APP_URLS = {
+    'hubspot': 'http://28.7.184.198:8150',
+    'gmail': 'http://28.7.184.198:8138',
+    'google_calendar': 'http://28.7.184.198:8141',
+}
+_GO_CACHE = {}
+
+
+def clamp01(x):
+    return max(0.0, min(1.0, float(x)))
+
+
+def frac(num, den):
+    return 0.0 if den == 0 else num / den
+
+
+def f1(true_set, pred_set):
+    if not true_set and not pred_set:
+        return 1.0
+    tp = len(true_set & pred_set)
+    if tp == 0:
+        return 0.0
+    prec = tp / len(pred_set)
+    rec = tp / len(true_set)
+    return 2 * prec * rec / (prec + rec)
+
+
+def by_id(items, key='id'):
+    return {it.get(key): it for it in (items or []) if isinstance(it, dict)}
+
+
+def has_text(haystack, *needles):
+    s = (haystack or '').lower()
+    return any(n.lower() in s for n in needles)
+
+
+def norm(s):
+    return (str(s) if s is not None else '').strip().lower()
+
+
+def _msg_text(m):
+    if not isinstance(m, dict):
+        return ''
+    return m.get('content') or m.get('text') or ''
+
+
+def _slack_channel_messages(slack_state, channel_name):
+    """Return messages for a channel, handling channels[*].messages AND the
+    top-level messages[channelId] map."""
+    if not isinstance(slack_state, dict):
+        return []
+    target = norm(channel_name).lstrip('#')
+    out = []
+    messages_map = slack_state.get('messages') if isinstance(slack_state.get('messages'), dict) else {}
+    for ch in slack_state.get('channels', []) or []:
+        if not isinstance(ch, dict):
+            continue
+        nm = norm(ch.get('name')).lstrip('#')
+        if nm != target:
+            continue
+        msgs = ch.get('messages')
+        if isinstance(msgs, list):
+            out.extend(m for m in msgs if isinstance(m, dict))
+        cid = ch.get('channelId') or ch.get('id')
+        if cid and isinstance(messages_map.get(cid), list):
+            out.extend(m for m in messages_map[cid] if isinstance(m, dict))
+    return out
+
+
+def _sheet_rows(sheets_state, sheet_name=None):
+    """Rows from a materialized sheets dict {name: {headers, rows}}."""
+    if isinstance(sheets_state, dict):
+        if sheet_name and isinstance(sheets_state.get(sheet_name), dict):
+            return sheets_state[sheet_name].get('rows', []) or []
+        for v in sheets_state.values():
+            if isinstance(v, dict) and 'rows' in v:
+                return v.get('rows', []) or []
+    return []
+
+
+def _email_text(email):
+    if not isinstance(email, dict):
+        return str(email)
+
+    to = email.get('to', '')
+    if isinstance(to, list):
+        to = ' '.join(
+            (x.get('name') or x.get('email') or str(x)) if isinstance(x, dict) else str(x)
+            for x in to
+        )
+    elif isinstance(to, dict):
+        to = to.get('name') or to.get('email') or str(to)
+
+    return f"{to} {email.get('subject', '')} {email.get('body', '')}"
+
+
+def _gmail_sent_emails(gmail_state):
+    out = []
+    if not isinstance(gmail_state, dict):
+        return out
+
+    for e in gmail_state.get('emails', []):
+        if not isinstance(e, dict):
+            continue
+        folder = norm(e.get('folder'))
+        if folder in ('sent', 'sentitems', 'sent items'):
+            out.append(e)
+            continue
+        if e.get('isSent') is True or e.get('sentAt') or e.get('sentDateTime'):
+            out.append(e)
+
+    if out:
+        return out
+
+    for key in ('sent', 'sentEmails', 'outbox'):
+        items = gmail_state.get(key, [])
+        if isinstance(items, list):
+            out.extend(x for x in items if isinstance(x, dict))
+    return out
+
+
+def _normalize_app_name(app):
+    app = (app or '').strip().lower()
+    return app[:-5] if app.endswith('_mock') else app
+
+
+def _parse_cell_id(cell_id):
+    if not isinstance(cell_id, str) or not cell_id:
+        return None
+    idx = 0
+    while idx < len(cell_id) and cell_id[idx].isalpha():
+        idx += 1
+    if idx == 0 or idx >= len(cell_id):
+        return None
+    col = 0
+    for ch in cell_id[:idx].upper():
+        col = col * 26 + (ord(ch) - ord('A') + 1)
+    try:
+        row = int(cell_id[idx:])
+    except ValueError:
+        return None
+    return row, col - 1
+
+
+def _extract_rows_from_workbook(state):
+    if not isinstance(state, dict):
+        return None
+    sheets = state.get('sheets')
+    if not isinstance(sheets, list):
+        return None
+
+    adapter = state.get('_task_adapter') if isinstance(state.get('_task_adapter'), dict) else {}
+    headers_by_sheet = adapter.get('headers_by_sheet') if isinstance(adapter.get('headers_by_sheet'), dict) else {}
+
+    out = {}
+    for sh in sheets:
+        if not isinstance(sh, dict):
+            continue
+        name = sh.get('name') or sh.get('id') or 'Sheet1'
+        data = sh.get('data') if isinstance(sh.get('data'), dict) else {}
+
+        headers = headers_by_sheet.get(name)
+        if not headers:
+            cols = []
+            for cid, cell in data.items():
+                rc = _parse_cell_id(cid)
+                if not rc:
+                    continue
+                row_idx, col_idx = rc
+                if row_idx != 1:
+                    continue
+                value = ''
+                if isinstance(cell, dict):
+                    value = cell.get('value', '')
+                elif cell is not None:
+                    value = str(cell)
+                if value is not None and str(value) != '':
+                    cols.append((col_idx, str(value)))
+            headers = [v for _, v in sorted(cols)]
+
+        if not headers:
+            continue
+
+        by_row = {}
+        for cid, cell in data.items():
+            rc = _parse_cell_id(cid)
+            if not rc:
+                continue
+            row_idx, col_idx = rc
+            if row_idx <= 1 or col_idx >= len(headers):
+                continue
+            if isinstance(cell, dict):
+                value = cell.get('value', '')
+            else:
+                value = cell
+            by_row.setdefault(row_idx, {})[headers[col_idx]] = value
+
+        rows = []
+        for row_idx in sorted(by_row.keys()):
+            row = by_row[row_idx]
+            if not any(str(row.get(h, '')).strip() for h in headers):
+                continue
+            rows.append({h: row.get(h, '') for h in headers})
+
+        out[name] = {'headers': headers, 'rows': rows}
+
+    if not out:
+        return None
+    return {'sheets': out}
+
+
+def _materialize_google_sheets(payload):
+    for key in ('initial_state', 'current_state'):
+        state = payload.get(key)
+        if not isinstance(state, dict):
+            continue
+        maybe_rows = _extract_rows_from_workbook(state)
+        if maybe_rows is not None:
+            state['_ui_workbook'] = {
+                'sheets': copy.deepcopy(state.get('sheets', [])),
+                'title': state.get('title'),
+            }
+            state['sheets'] = maybe_rows['sheets']
+
+
+def _materialize_slack(payload):
+    for key in ('initial_state', 'current_state'):
+        state = payload.get(key)
+        if not isinstance(state, dict):
+            continue
+        channels = state.get('channels') if isinstance(state.get('channels'), list) else []
+        messages_map = state.get('messages') if isinstance(state.get('messages'), dict) else {}
+        for ch in channels:
+            if not isinstance(ch, dict):
+                continue
+            cid = ch.get('channelId') or ch.get('id')
+            if not cid:
+                continue
+            msg_list = messages_map.get(cid) if isinstance(messages_map.get(cid), list) else []
+            channel_msgs = ch.get('messages') if isinstance(ch.get('messages'), list) else []
+            merged = {
+                (m.get('messageId') or m.get('id')): dict(m)
+                for m in channel_msgs
+                if isinstance(m, dict)
+            }
+            for m in msg_list:
+                if not isinstance(m, dict):
+                    continue
+                mid = m.get('messageId') or m.get('id')
+                text = m.get('text') if m.get('text') is not None else m.get('content', '')
+                merged[mid] = {
+                    'id': mid,
+                    'messageId': mid,
+                    'text': text,
+                    'content': m.get('content', text),
+                    'senderId': m.get('senderId'),
+                    'timestamp': m.get('timestamp'),
+                    'reactions': m.get('reactions') if isinstance(m.get('reactions'), list) else [],
+                }
+            ch['messages'] = list(merged.values())
+
+
+def _owner_name_map(state):
+    out = {}
+    users = state.get('users') if isinstance(state.get('users'), list) else []
+    for u in users:
+        if not isinstance(u, dict):
+            continue
+        uid = u.get('userId')
+        if not uid:
+            continue
+        name = (f"{u.get('firstName', '')} {u.get('lastName', '')}".strip() or u.get('name') or uid)
+        out[uid] = name
+    return out
+
+
+def _materialize_salesforce(payload):
+    for key in ('initial_state', 'current_state'):
+        state = payload.get(key)
+        if not isinstance(state, dict):
+            continue
+
+        adapter = state.get('_task_adapter') if isinstance(state.get('_task_adapter'), dict) else {}
+        owner_names = _owner_name_map(state)
+
+        task_leads = adapter.get('task_leads') if isinstance(adapter.get('task_leads'), dict) else None
+        if task_leads and isinstance(state.get('leads'), list):
+            restored = []
+            for lead in state['leads']:
+                if not isinstance(lead, dict):
+                    continue
+                lid = str(lead.get('leadId') or lead.get('id') or '')
+                meta = task_leads.get(lid, {})
+                restored.append({
+                    'id': meta.get('id') or lead.get('leadId') or lead.get('id'),
+                    'company': lead.get('company', meta.get('company', '')),
+                    'status': lead.get('status', 'New'),
+                    'ownerName': owner_names.get(lead.get('ownerId')) or meta.get('ownerName'),
+                    'annualRevenue': meta.get('annualRevenue', lead.get('revenue')),
+                })
+            state['leads'] = restored
+
+        task_accounts = adapter.get('task_accounts') if isinstance(adapter.get('task_accounts'), dict) else None
+        if task_accounts and isinstance(state.get('accounts'), list):
+            current_by_id = {}
+            for acc in state['accounts']:
+                if isinstance(acc, dict):
+                    current_by_id[str(acc.get('accountId') or acc.get('id') or '')] = acc
+
+            restored_accounts = []
+            for aid, meta in task_accounts.items():
+                cur = current_by_id.get(str(aid), {})
+                restored_accounts.append({
+                    'id': meta.get('id') or aid,
+                    'name': meta.get('name') or cur.get('name', ''),
+                    'contact': meta.get('contact'),
+                    'wonWithin7d': bool(meta.get('wonWithin7d', False)),
+                    'onboardingStage': cur.get('onboardingStage', meta.get('onboardingStage')),
+                })
+            state['accounts'] = restored_accounts
+
+
+def _materialize_google_docs(payload):
+    for key in ('initial_state', 'current_state'):
+        state = payload.get(key)
+        if not isinstance(state, dict):
+            continue
+        adapter = state.get('_task_adapter') if isinstance(state.get('_task_adapter'), dict) else {}
+        task_docs = adapter.get('task_documents') if isinstance(adapter.get('task_documents'), dict) else None
+        if not task_docs:
+            continue
+        docs_map = state.get('documents') if isinstance(state.get('documents'), dict) else {}
+        restored = []
+        for did, meta in task_docs.items():
+            cur = docs_map.get(did, {}) if isinstance(docs_map, dict) else {}
+            restored.append({
+                'id': meta.get('id') or did,
+                'title': cur.get('title', meta.get('title')),
+                'vendor': meta.get('vendor'),
+                'folder': meta.get('folder'),
+                'content': cur.get('content', meta.get('content', '')),
+            })
+        state['documents'] = restored
+
+
+def _adapt_payload_for_reward(app, payload):
+    app_name = _normalize_app_name(app)
+    if app_name == 'google_sheets':
+        _materialize_google_sheets(payload)
+    elif app_name == 'slack':
+        _materialize_slack(payload)
+    elif app_name == 'salesforce':
+        _materialize_salesforce(payload)
+    elif app_name == 'google_docs':
+        _materialize_google_docs(payload)
+    return payload
+
+
+def go(app_name):
+    if app_name in _GO_CACHE:
+        return _GO_CACHE[app_name]
+
+    base = APP_URLS.get(app_name)
+    if not base:
+        print(f'CRITICAL: unknown app {app_name!r} (no URL mapped)')
+        print('REWARD: 0.0')
+        sys.exit(0)
+
+    try:
+        resp = requests.get(f'{base}/go?sid={sid}', timeout=15)
+        resp.raise_for_status()
+        payload = resp.json()
+    except Exception as e:
+        print(f'CRITICAL: cannot fetch {app_name} go state: {e}')
+        print('REWARD: 0.0')
+        sys.exit(0)
+
+    payload = _adapt_payload_for_reward(app_name, payload)
+    _GO_CACHE[app_name] = payload
+    return payload
+
+
+# === Task-specific reward ===
+#
+# Ground truth lives in hubspot_mock initial_state deals[*] hidden `_*` fields:
+#   _qualifies (bool)              -> THE qualifying set (d1-d4 True; d5-d9 False)
+#   _renewal_line (str)            -> exact line to append (qualifying only)
+#   _expected_description (str)    -> original + "\n" + _renewal_line (qualifying)
+#   _orig_owner (str)              -> 'Admin User' (owner at injection, ALL deals)
+#   _new_owner (str)               -> 'Renewals Team' (qualifying only)
+#   _email_subject (str)           -> "Renewal: <Company>" (ALL deals; leak shape)
+#   _email_body_must_contain       -> [company, $amount, closeDate] (qualifying only)
+#   _renewals_label_name (str)     -> 'Renewals' (qualifying only)
+#   _event_title (str)             -> "Renewal call: <Company>" (ALL deals; leak shape)
+#   _event_date (str)              -> '2026-05-11' (qualifying only)
+#   _event_guest_email (str)       -> contact email to add as guest (qualifying only)
+# Original description = initial_state deals[i].description (pre-edit visible field).
+
+# Positive components (each requires an agent action -> all 0 in do-nothing).
+# Weights sum to exactly 1.0. Leakage onto non-qualifying deals is a PENALTY
+# (subtracted, 0 in do-nothing) rather than positive credit, so a do-nothing
+# agent scores exactly 0.
+WEIGHT_DESC = 0.22
+WEIGHT_OWNER = 0.17
+WEIGHT_EMAIL = 0.22
+WEIGHT_LABEL = 0.17
+WEIGHT_EVENT = 0.11
+WEIGHT_GUEST = 0.11
+WEIGHT_LEAK_PENALTY = 0.15   # max penalty if EVERY non-qualifying deal is touched
+
+
+def _email_full_text(e):
+    """subject + HTML-stripped body + snippet, for substring search."""
+    if not isinstance(e, dict):
+        return ''
+    subj = e.get('subject') or ''
+    body = e.get('body') or ''
+    body_text = re.sub(r'<[^>]+>', ' ', body)
+    snippet = e.get('snippet') or ''
+    return f'{subj}\n{body_text}\n{snippet}'
+
+
+def _desc_match(cur_desc, orig_desc, renewal_line, expected_desc):
+    """current description endswith renewal_line AND still contains original text."""
+    cur = cur_desc or ''
+    rl = (renewal_line or '').strip()
+    if not rl:
+        return False
+    cur_stripped = cur.rstrip()
+    ends = cur_stripped.endswith(rl) or norm(cur_stripped).endswith(norm(rl))
+    orig = (orig_desc or '').strip()
+    contains_orig = (norm(orig) in norm(cur)) if orig else True
+    if ends and contains_orig:
+        return True
+    # fallback: exact match to the embedded expected description
+    if expected_desc and norm(cur) == norm(expected_desc):
+        return True
+    return False
+
+
+def _renewal_emails_by_subject(sent_emails, subject):
+    """All sent emails whose subject == subject (case-insensitive)."""
+    want_subj = norm(subject)
+    if not want_subj:
+        return []
+    return [e for e in (sent_emails or [])
+            if isinstance(e, dict) and norm(e.get('subject')) == want_subj]
+
+
+def _email_body_ok(e, needles):
+    text = norm(_email_full_text(e))
+    return all(norm(nd) in text for nd in (needles or []) if nd)
+
+
+def _label_name_set(email, label_id_to_name):
+    """Resolve an email's label ids/names to a set of normalized label NAMES."""
+    out = set()
+    labels = email.get('labels')
+    if not isinstance(labels, list):
+        return out
+    for lab in labels:
+        if isinstance(lab, dict):
+            nm = lab.get('name') or lab.get('id')
+            if nm:
+                out.add(norm(label_id_to_name.get(nm, nm)))
+        else:
+            out.add(norm(label_id_to_name.get(lab, lab)))
+    return out
+
+
+# ---- Calendar helpers (TZ-robust date matching, guest subset) ----
+def _parse_as_utc(s, default_tz=TZ_PLUS_8):
+    if not s:
+        return None
+    s = str(s).strip()
+    if s.endswith('Z'):
+        s = s[:-1] + '+00:00'
+    try:
+        dt = datetime.fromisoformat(s)
+    except Exception:
+        return None
+    if dt.tzinfo is None:
+        dt = dt.replace(tzinfo=default_tz)
+    return dt
+
+
+def _raw_date_part(s):
+    """Literal wall-clock date 'YYYY-MM-DD' (strip Z / tz offset, take date)."""
+    if not s:
+        return ''
+    t = str(s).strip()
+    if 'T' not in t:
+        return t[:10]
+    return t.split('T', 1)[0]
+
+
+def _event_on_date(ev, target_date):
+    """True if the event's start matches target_date under EITHER the literal
+    wall-clock reading OR the UTC->UTC+8 conversion (robust to datetime-local
+    -> toISOString() tz shifts the calendar mock performs)."""
+    start = ev.get('start')
+    target = norm(target_date)
+    if not target:
+        return True
+    if norm(_raw_date_part(start)) == target:
+        return True
+    dt = _parse_as_utc(start)
+    if dt is not None:
+        if dt.astimezone(TZ_PLUS_8).date().isoformat() == target:
+            return True
+        if dt.astimezone(timezone.utc).date().isoformat() == target:
+            return True
+    return False
+
+
+def _event_guests(ev):
+    g = ev.get('guests')
+    if not isinstance(g, list):
+        g = ev.get('attendees') if isinstance(ev.get('attendees'), list) else []
+    out = set()
+    for a in g:
+        if isinstance(a, dict):
+            a = a.get('email') or a.get('name')
+        if a:
+            out.add(norm(a))
+    return out
+
+
+def reward(go):
+    hub = go('hubspot')
+    hub_i = hub.get('initial_state', {}) or {}
+    hub_c = hub.get('current_state', {}) or {}
+    gmail_c = go('gmail').get('current_state', {}) or {}
+    cal_c = go('google_calendar').get('current_state', {}) or {}
+
+    deals_i = [d for d in (hub_i.get('deals', []) or []) if isinstance(d, dict)]
+    deals_c_by_id = by_id([d for d in (hub_c.get('deals', []) or []) if isinstance(d, dict)])
+    sent_emails = _gmail_sent_emails(gmail_c)
+    cur_events = [e for e in (cal_c.get('events', []) or []) if isinstance(e, dict)]
+
+    # label id -> name map (current gmail labels), for robust label-name matching
+    label_id_to_name = {}
+    for lab in (gmail_c.get('labels', []) or []):
+        if isinstance(lab, dict) and lab.get('id'):
+            label_id_to_name[lab['id']] = lab.get('name') or lab['id']
+
+    qualifying = [d for d in deals_i if d.get('_qualifies') is True]
+    nonqual = [d for d in deals_i if d.get('_qualifies') is not True]
+    n_q = len(qualifying)
+    n_nq = len(nonqual)
+
+    # --- Component 1: HubSpot description appended (0.20) ---
+    desc_ok = 0
+    desc_flags = []
+    for d in qualifying:
+        did = d.get('id')
+        cur = deals_c_by_id.get(did, {}) or {}
+        ok = _desc_match(
+            cur.get('description') or '',
+            d.get('description') or '',
+            d.get('_renewal_line') or '',
+            d.get('_expected_description') or '',
+        )
+        desc_ok += 1 if ok else 0
+        desc_flags.append(f"{did}={'Y' if ok else 'N'}")
+    s_desc = WEIGHT_DESC * frac(desc_ok, n_q)
+
+    # --- Component 2: HubSpot owner reassigned (0.15) ---
+    owner_ok = 0
+    owner_flags = []
+    for d in qualifying:
+        did = d.get('id')
+        cur = deals_c_by_id.get(did, {}) or {}
+        want = norm(d.get('_new_owner') or '')
+        ok = bool(want) and norm(cur.get('owner')) == want
+        owner_ok += 1 if ok else 0
+        owner_flags.append(f"{did}={'Y' if ok else 'N'}")
+    s_owner = WEIGHT_OWNER * frac(owner_ok, n_q)
+
+    # --- Component 3: Gmail renewal email sent (0.20) ---
+    email_ok = 0
+    email_flags = []
+    for d in qualifying:
+        did = d.get('id')
+        matches = _renewal_emails_by_subject(sent_emails, d.get('_email_subject') or '')
+        ok = any(_email_body_ok(e, d.get('_email_body_must_contain') or []) for e in matches)
+        email_ok += 1 if ok else 0
+        email_flags.append(f"{did}={'Y' if ok else 'N'}")
+    s_email = WEIGHT_EMAIL * frac(email_ok, n_q)
+
+    # --- Component 4: Gmail 'Renewals' label applied to the renewal email (0.15) ---
+    label_ok = 0
+    label_flags = []
+    for d in qualifying:
+        did = d.get('id')
+        want_label = norm(d.get('_renewals_label_name') or '')
+        matches = _renewal_emails_by_subject(sent_emails, d.get('_email_subject') or '')
+        ok = bool(want_label) and any(
+            want_label in _label_name_set(e, label_id_to_name) for e in matches
+        )
+        label_ok += 1 if ok else 0
+        label_flags.append(f"{did}={'Y' if ok else 'N'}")
+    s_label = WEIGHT_LABEL * frac(label_ok, n_q)
+
+    # --- Component 5: Calendar 'Renewal call: <Company>' event on the right date (0.10) ---
+    event_ok = 0
+    event_flags = []
+    for d in qualifying:
+        did = d.get('id')
+        title = norm(d.get('_event_title') or '')
+        date_str = d.get('_event_date') or ''
+        ok = bool(title) and any(
+            norm(ev.get('title')) == title and _event_on_date(ev, date_str)
+            for ev in cur_events
+        )
+        event_ok += 1 if ok else 0
+        event_flags.append(f"{did}={'Y' if ok else 'N'}")
+    s_event = WEIGHT_EVENT * frac(event_ok, n_q)
+
+    # --- Component 6: Calendar contact added as guest on the renewal-call event (0.10) ---
+    guest_ok = 0
+    guest_flags = []
+    for d in qualifying:
+        did = d.get('id')
+        title = norm(d.get('_event_title') or '')
+        want_guest = norm(d.get('_event_guest_email') or '')
+        ok = bool(title) and bool(want_guest) and any(
+            norm(ev.get('title')) == title and want_guest in _event_guests(ev)
+            for ev in cur_events
+        )
+        guest_ok += 1 if ok else 0
+        guest_flags.append(f"{did}={'Y' if ok else 'N'}")
+    s_guest = WEIGHT_GUEST * frac(guest_ok, n_q)
+
+    # --- Penalty: leakage onto non-qualifying deals (0 in do-nothing) ---
+    # A non-qualifying deal is "leaked" if the agent edited its description, changed
+    # its owner, sent it a 'Renewal:' email, or created its 'Renewal call:' event.
+    leaked = 0
+    leak_flags = []
+    for d in nonqual:
+        did = d.get('id')
+        cur = deals_c_by_id.get(did, {}) or {}
+        desc_leak = norm(cur.get('description') or '') != norm(d.get('description') or '')
+        owner_leak = norm(cur.get('owner')) != norm(d.get('_orig_owner') or '')
+        email_leak = len(_renewal_emails_by_subject(sent_emails, d.get('_email_subject') or '')) > 0
+        evt_title = norm(d.get('_event_title') or '')
+        event_leak = bool(evt_title) and any(norm(ev.get('title')) == evt_title for ev in cur_events)
+        if desc_leak or owner_leak or email_leak or event_leak:
+            leaked += 1
+        leak_flags.append(f"{did}=d{int(desc_leak)}o{int(owner_leak)}e{int(email_leak)}c{int(event_leak)}")
+    leak_penalty = WEIGHT_LEAK_PENALTY * (0.0 if n_nq == 0 else frac(leaked, n_nq))
+
+    positive = s_desc + s_owner + s_email + s_label + s_event + s_guest
+    score = clamp01(positive - leak_penalty)
+
+    print(
+        'DEBUG_T04_hubspot_renewal_outreach_3app '
+        f'n_qual={n_q} n_nonqual={n_nq} sent_emails={len(sent_emails)} cal_events={len(cur_events)} '
+        f'desc={desc_ok}/{n_q}[{",".join(desc_flags)}] '
+        f'owner={owner_ok}/{n_q}[{",".join(owner_flags)}] '
+        f'email={email_ok}/{n_q}[{",".join(email_flags)}] '
+        f'label={label_ok}/{n_q}[{",".join(label_flags)}] '
+        f'event={event_ok}/{n_q}[{",".join(event_flags)}] '
+        f'guest={guest_ok}/{n_q}[{",".join(guest_flags)}] '
+        f'leaked={leaked}/{n_nq}[{",".join(leak_flags)}] '
+        f'w_desc={round(s_desc, 4)} w_owner={round(s_owner, 4)} w_email={round(s_email, 4)} '
+        f'w_label={round(s_label, 4)} w_event={round(s_event, 4)} w_guest={round(s_guest, 4)} '
+        f'leak_penalty={round(leak_penalty, 4)} total={round(score, 4)}'
+    )
+
+    return clamp01(score)
+
+
+try:
+    score = float(reward(go))
+except Exception as exc:
+    print(f'ERROR: reward() raised {exc!r}')
+    score = 0.0
+
+score = clamp01(score)
+print(f'\nREWARD: {round(score, 4)}')

am_renewal_outreach_002/_cua_gym_vm_bridge.sh

@@ -0,0 +1,106 @@
+#!/bin/bash
+# CUA-Gym VM-side network fix (direct-IP mock access).
+#
+# Runs INSIDE the QEMU guest. The CUA-Gym mock services are now reached by
+# their real IDC address (http://28.7.184.198:81xx) instead of the old
+# localhost:9000-9097 socat bridge, so the VM only needs two fixes:
+#
+#   1. Working DNS — the qcow2-baked /etc/resolv.conf points at the Pod's
+#      cluster DNS, which is unreachable from QEMU SLIRP NAT
+#      ("Temporary failure in name resolution").
+#   2. The mock host excluded from the woa http proxy — otherwise the VM's
+#      Chrome (gsettings proxy) and initial_setup.py's `requests` get a 403
+#      from the proxy when hitting 28.7.184.198:81xx. (The old socat path never
+#      hit this because localhost is bypassed by default.)
+#
+# Idempotent. Scoped to the CUA-Gym bridge step; does not touch the generic
+# OSWorld setup pipeline.
+#
+# Argv:
+#   $1  optional sudo password (default "password"; SetupController
+#       substitutes {CLIENT_PASSWORD} when uploaded as a cache file).
+#   $2  unused (kept for backward compat — was socat first port 9000).
+#   $3  unused (kept for backward compat — was socat last  port 9097).
+#   $4  optional mock host IP (default 28.7.184.198).
+
+set +e
+
+PASS="${1:-password}"
+MOCK_HOST="${4:-28.7.184.198}"
+# Derive the /16 IDC subnet so future port/host tweaks on the same segment
+# stay bypassed too (28.7.184.198 -> 28.7.0.0/16).
+MOCK_NET="$(echo "${MOCK_HOST}" | awk -F. 'NF==4{print $1"."$2".0.0/16"}')"
+
+run_root() {
+    echo "$PASS" | sudo -S -p "" bash -c "$1"
+}
+
+# Detect SLIRP gateway from VM's routing table; fall back to 10.0.2.2.
+GW="$(ip route 2>/dev/null | awk '/default/ {print $3; exit}')"
+GW="${GW:-10.0.2.2}"
+
+# ─── CUA-Gym network fix (Taiji-specific) ─────────────────────────────
+fix_vm_network() {
+    # 0) Discover VM network state at runtime — gateway, own IP/CIDR, SLIRP
+    #    DNS forwarder. Nothing hardcoded except the mock host above.
+    local GW_NET="${GW%.*}"
+    local SLIRP_DNS="${GW_NET}.3"        # QEMU SLIRP DNS proxy is always .3
+    local OWN_CIDR
+    OWN_CIDR="$(ip -4 -o addr show scope global 2>/dev/null \
+                | awk '{print $4; exit}')"
+    local OWN_IP="${OWN_CIDR%/*}"
+    local OWN_NET=""
+    [ -n "${OWN_IP}" ] && OWN_NET="${OWN_IP%.*}.0/24"
+
+    # 1) DNS — point at SLIRP DNS proxy + public fallbacks. chattr +i keeps
+    #    systemd-resolved/NetworkManager from clobbering it.
+    if run_root "printf 'nameserver ${SLIRP_DNS}\\nnameserver 114.114.114.114\\nnameserver 8.8.8.8\\n' > /etc/resolv.conf
+                 chattr +i /etc/resolv.conf 2>/dev/null || true"; then
+        echo "[cua_gym_vm_bridge] DNS set: ${SLIRP_DNS}, 114.114.114.114, 8.8.8.8 (gw=${GW})"
+    else
+        echo "[cua_gym_vm_bridge] WARN: failed to write /etc/resolv.conf"
+    fi
+
+    # 2) Build the proxy bypass list from detected values + the direct-IP
+    #    mock host (and its /16 subnet). THIS is what makes 28.7.184.198:81xx
+    #    reachable from inside the VM — without it the woa proxy 403s it.
+    local ignore_list="['localhost', '127.0.0.0/8', '${GW}'"
+    [ -n "${OWN_IP}" ]    && ignore_list="${ignore_list}, '${OWN_IP}'"
+    [ -n "${OWN_NET}" ]   && ignore_list="${ignore_list}, '${OWN_NET}'"
+    [ -n "${GW_NET}" ]    && ignore_list="${ignore_list}, '${GW_NET}.0/24'"
+    [ -n "${MOCK_HOST}" ] && ignore_list="${ignore_list}, '${MOCK_HOST}'"
+    [ -n "${MOCK_NET}" ]  && ignore_list="${ignore_list}, '${MOCK_NET}'"
+    ignore_list="${ignore_list}]"
+
+    # 3) GNOME proxy ignore-hosts — Chrome bypasses the woa proxy for these.
+    if gsettings set org.gnome.system.proxy ignore-hosts "${ignore_list}" 2>/tmp/gsettings_err; then
+        echo "[cua_gym_vm_bridge] gsettings ignore-hosts = ${ignore_list}"
+    else
+        echo "[cua_gym_vm_bridge] WARN: gsettings ignore-hosts failed: $(cat /tmp/gsettings_err 2>/dev/null)"
+    fi
+
+    # 4) Shell env — comma-separated form for no_proxy/NO_PROXY (requests/curl).
+    local NP="localhost,127.0.0.1,${GW}"
+    [ -n "${OWN_IP}" ]    && NP="${NP},${OWN_IP}"
+    [ -n "${OWN_NET}" ]   && NP="${NP},${OWN_NET}"
+    [ -n "${GW_NET}" ]    && NP="${NP},${GW_NET}.0/24"
+    [ -n "${MOCK_HOST}" ] && NP="${NP},${MOCK_HOST}"
+    [ -n "${MOCK_NET}" ]  && NP="${NP},${MOCK_NET}"
+
+    for var in no_proxy NO_PROXY; do
+        if ! grep -q "export ${var}=${NP}" /home/user/.bashrc 2>/dev/null; then
+            echo "export ${var}=${NP}" >> /home/user/.bashrc
+        fi
+    done
+    run_root "grep -q '^no_proxy=' /etc/environment 2>/dev/null \
+              && sed -i 's|^no_proxy=.*|no_proxy=\"${NP}\"|' /etc/environment \
+              || echo 'no_proxy=\"${NP}\"' >> /etc/environment
+              grep -q '^NO_PROXY=' /etc/environment 2>/dev/null \
+              && sed -i 's|^NO_PROXY=.*|NO_PROXY=\"${NP}\"|' /etc/environment \
+              || echo 'NO_PROXY=\"${NP}\"' >> /etc/environment" >/dev/null 2>&1
+    echo "[cua_gym_vm_bridge] no_proxy=${NP}"
+}
+
+fix_vm_network
+
+echo "[cua_gym_vm_bridge] direct-IP mode: ${MOCK_HOST} (+${MOCK_NET}) bypassed from woa proxy; no socat forwarding needed"

am_renewal_outreach_002/initial_setup.py

@@ -0,0 +1,583 @@
+# --- MOCK_HOST_NO_PROXY_OVERRIDE: bypass woa proxy for locally-deployed mocks ---
+import os as _os
+_np = _os.environ.get("no_proxy", _os.environ.get("NO_PROXY", ""))
+for _h in ("127.0.0.1", "localhost", "28.7.184.198"):
+    if _h not in _np:
+        _np = (_np + "," + _h).lstrip(",")
+_os.environ["no_proxy"] = _np
+_os.environ["NO_PROXY"] = _np
+# --- end override ---
+
+# --- wait_mocks_loaded: wait for mock backends ready AND browser tabs rendered ---
+def wait_mocks_loaded(timeout=60, render_buffer=12):
+    """Wait for mock backends ready AND browser tabs rendered (via CDP).
+
+    Phase 1 — backend readiness: poll each mock's /state?sid= until it
+    returns 200 with 'stored_state' in the body.
+
+    Phase 2 — browser render verification: connect to Chrome's CDP
+    endpoint and wait for every open tab to reach 'networkidle' load
+    state with a non-trivial DOM element count. This confirms the SPA
+    has fetched its /go?sid= data and actually painted, not just that
+    the backend is up.
+
+    CDP port auto-detection: checks globals for _chrome_debug_port /
+    cdp_port / debug_port (set by tasks that use a dynamic port, e.g.
+    se_hard_008's launch_chrome_with_urls); defaults to 1337 for tasks
+    that use launch_gui's --remote-debugging-port injection.
+
+    Falls back to the old render_buffer sleep if Playwright or the CDP
+    endpoint is unavailable.
+    """
+    import urllib.request as _u, time as _t
+    g = globals()
+    _urls = {}
+    for _k, _v in list(g.items()):
+        if isinstance(_v, dict) and _v and all(isinstance(x, str) and x.startswith("http://") for x in _v.values()):
+            _urls.update(_v)
+        elif isinstance(_v, str) and _v.startswith("http://") and (_k.endswith("_URL") or _k.endswith("_url")):
+            _urls[_k] = _v
+    _sid = ""
+    for _k in ("sid", "SID", "task_sid", "session_id"):
+        if isinstance(g.get(_k), str) and g[_k]:
+            _sid = g[_k]; break
+    if not _sid:
+        for _k, _v in g.items():
+            if "sid" in _k.lower() and isinstance(_v, str) and _v:
+                _sid = _v; break
+    if not _urls or not _sid:
+        print("  wait_mocks_loaded: no mocks/sid found (%d urls, sid=%r), skipping" % (len(_urls), bool(_sid)))
+        return
+
+    # ---- Phase 1: backend readiness ----
+    _op = _u.build_opener(_u.ProxyHandler({}))
+    _dl = _t.time() + timeout
+    _todo = list(_urls.items())
+    while _todo and _t.time() < _dl:
+        _rest = []
+        for _n2, _uu in _todo:
+            try:
+                _r = _op.open("%s/state?sid=%s" % (_uu, _sid), timeout=5)
+                if _r.status == 200 and b"stored_state" in _r.read():
+                    print("  %s: backend ready" % _n2); continue
+            except Exception:
+                pass
+            _rest.append((_n2, _uu))
+        _todo = _rest
+        if _todo:
+            _t.sleep(2)
+    _done = len(_urls) - len(_todo)
+    _pend = (" (pending: %s)" % [n for n, _ in _todo]) if _todo else ""
+    print("  mocks backend ready: %d/%d%s" % (_done, len(_urls), _pend))
+
+    # ---- Phase 2: browser render verification via CDP ----
+    # Auto-detect CDP port from globals (set by launch_chrome_with_urls
+    # in tasks that use a dynamic port); default to 1337.
+    _cdp_port = 1337
+    for _pk in ("_chrome_debug_port", "chrome_debug_port", "cdp_port", "debug_port"):
+        _pv = g.get(_pk)
+        if isinstance(_pv, int) and 0 < _pv < 65536:
+            _cdp_port = _pv; break
+    _cdp_ok = False
+    try:
+        from playwright.sync_api import sync_playwright
+        _cdp_dl = _t.time() + max(10.0, _dl - _t.time())  # leftover budget, floor 10s
+        with sync_playwright() as _p:
+            _br = None
+            # Retry-connect: Chrome may still be starting up.
+            while _t.time() < _cdp_dl:
+                try:
+                    _br = _p.chromium.connect_over_cdp("http://localhost:%d" % _cdp_port)
+                    break
+                except Exception:
+                    _t.sleep(0.5)
+            if _br is None:
+                print("  wait_mocks_loaded: CDP connect to localhost:%d failed; "
+                      "is Chrome launched with --remote-debugging-port?" % _cdp_port)
+            else:
+                # Wait for ALL expected tabs to appear, then verify each renders.
+                _n_expected = len(_urls)
+                _pages = []
+                _pg_dl = _t.time() + 15
+                _last_count = -1
+                _stable_since = None
+                while _t.time() < _pg_dl:
+                    _pages = []
+                    for _c in _br.contexts:
+                        _pages.extend(_c.pages)
+                    _cnt = len(_pages)
+                    if _n_expected > 0 and _cnt >= _n_expected:
+                        break  # all expected tabs are open
+                    if _cnt > 0 and _cnt == _last_count:
+                        if _stable_since is None:
+                            _stable_since = _t.time()
+                        elif _t.time() - _stable_since >= 1.5:
+                            break  # tab count stable -> no more tabs opening
+                    else:
+                        _stable_since = None
+                    _last_count = _cnt
+                    _t.sleep(0.5)
+                # Re-fetch final snapshot so late-opening tabs are included.
+                _pages = []
+                for _c in _br.contexts:
+                    _pages.extend(_c.pages)
+                if not _pages:
+                    print("  wait_mocks_loaded: CDP connected but no open tabs found")
+                else:
+                    print("  wait_mocks_loaded: %d tab(s) open (expected %d); "
+                          "verifying render" % (len(_pages), _n_expected))
+                    _ok_cnt = 0
+                    for _pg in _pages:
+                        _ttl = ""
+                        try:
+                            _ttl = _pg.title()
+                        except Exception:
+                            pass
+                        try:
+                            _pg.wait_for_load_state(
+                                "networkidle",
+                                timeout=int(max(5000, (_cdp_dl - _t.time()) * 1000)))
+                            _n = _pg.evaluate("document.querySelectorAll('*').length")
+                            if _n < 20:
+                                print("  wait_mocks_loaded: WARNING tab %r has only %d "
+                                      "elements (render stall?)" % (_ttl, _n))
+                            else:
+                                print("  wait_mocks_loaded: tab %r rendered (%d elements)"
+                                      % (_ttl, _n))
+                                _ok_cnt += 1
+                        except Exception as _e:
+                            print("  wait_mocks_loaded: tab %r render wait failed: %r"
+                                  % (_ttl, _e))
+                    # Require ALL open tabs to render before skipping the
+                    # render_buffer fallback, so no tab is left half-loaded.
+                    if _ok_cnt > 0 and _ok_cnt == len(_pages):
+                        _cdp_ok = True
+                    else:
+                        print("  wait_mocks_loaded: %d/%d tabs rendered, will fall back"
+                              % (_ok_cnt, len(_pages)))
+    except ImportError:
+        print("  wait_mocks_loaded: playwright not installed; cannot verify browser render")
+    except Exception as _e:
+        print("  wait_mocks_loaded: CDP phase error (%r)" % _e)
+
+    if not _cdp_ok:
+        print("  wait_mocks_loaded: falling back to render_buffer=%ds sleep" % render_buffer)
+        _t.sleep(render_buffer)
+    else:
+        print("  wait_mocks_loaded: browser render verified, skipping render_buffer")
+# --- end wait_mocks_loaded ---
+
+# --- _open_remaining_mock_tabs: open all auto-detected mock URLs in new Chrome tabs ---
+def _open_remaining_mock_tabs(primary_url=""):
+    """Open all auto-detected mock URLs in Chrome --new-tab, except primary_url."""
+    g = globals()
+    _urls = set()
+    for _k, _v in list(g.items()):
+        if "PROBE" in _k.upper() or _k.startswith("_"):
+            continue
+        if isinstance(_v, str) and _v.startswith("http://") and (_k.endswith("_URL") or _k.endswith("_url")):
+            _urls.add(_v)
+        elif isinstance(_v, dict) and _v and all(isinstance(x, str) and x.startswith("http://") for x in _v.values()):
+            _urls.update(_v.values())
+    _sid = ""
+    for _k in ("sid", "SID", "task_sid", "session_id"):
+        if isinstance(g.get(_k), str) and g[_k]:
+            _sid = g[_k]; break
+    for _u in sorted(_urls):
+        if _u != primary_url:
+            launch_gui(f'google-chrome --new-tab "{_u}/?sid={_sid}"', delay_sec=1.0)
+# --- end _open_remaining_mock_tabs ---
+
+
+"""
+Initial Setup: Renewal check-in for Northwind Traders (Salesforce task + Gmail outreach + Slack closure)
+Task ID: am_renewal_outreach_002
+Domain: mock_websites
+Mocks: salesforce_mock, gmail_mock, slack_mock
+"""
+import os
+import shlex
+import subprocess
+import time
+import uuid
+
+import requests
+
+# --- Network egress: this VM has no direct internet; web egress to the mocks
+#     is routed through the cluster proxy. Probe once (fast) to decide whether a
+#     proxy is needed, then reuse the decision. Override the proxy with WEB_PROXY. ---
+PROXY = os.environ.get('WEB_PROXY') or None
+PROXIES = None
+
+_PROBE_URL = 'http://28.7.184.198:8138/go?sid=__probe__'
+
+
+def _decide_proxy():
+    try:
+        requests.get(_PROBE_URL, timeout=4)
+        return None  # direct works
+    except requests.exceptions.RequestException:
+        return PROXIES  # fall back to proxy
+
+
+_ACTIVE_PROXIES = _decide_proxy()
+print(f'Network mode: {"proxy " + PROXY if _ACTIVE_PROXIES else "direct"}')
+
+
+def http_post(url, payload):
+    return requests.post(url, json=payload, timeout=30, proxies=_ACTIVE_PROXIES)
+
+
+def http_get(url):
+    return requests.get(url, timeout=30, proxies=_ACTIVE_PROXIES)
+
+
+# --- Config ---
+SALESFORCE_URL = 'http://28.7.184.198:8175'
+GMAIL_URL = 'http://28.7.184.198:8138'
+SLACK_URL = 'http://28.7.184.198:8178'
+
+sid = str(uuid.uuid4())
+with open('/tmp/task_web_sid', 'w') as f:
+    f.write(sid)
+print(f'Generated sid={sid}')
+
+# Use a fresh Chrome profile per task session. This avoids stale Salesforce
+# modal/localStorage state (for example a lingering "New Task" dialog) from a
+# previous run blocking the task flow or making state detection flaky.
+CHROME_PROFILE_DIR = f'/tmp/cua_chrome_profile_{sid}'
+os.makedirs(CHROME_PROFILE_DIR, exist_ok=True)
+
+
+# ---------------------------------------------------------------------------
+# SALESFORCE STATE
+# ---------------------------------------------------------------------------
+def build_salesforce_state():
+    jordan = {
+        "userId": "user-1",
+        "firstName": "Jordan",
+        "lastName": "Avery",
+        "email": "jordan.avery@vertexcloud.example.com",
+        "phone": "(555) 204-1180",
+        "title": "Account Manager",
+        "department": "Sales",
+        "role": "Account Manager",
+        "avatar": "https://i.pravatar.cc/150?u=user-1",
+        "timezone": "America/New_York",
+        "locale": "en-US",
+        "theme": "lightning",
+    }
+    users = [
+        jordan,
+        {
+            "userId": "user-2", "firstName": "Priya", "lastName": "Nair",
+            "email": "priya.nair@vertexcloud.example.com", "phone": "(555) 204-1181",
+            "title": "Sales Director", "department": "Sales", "role": "Director",
+            "avatar": "https://i.pravatar.cc/150?u=user-2",
+            "timezone": "America/New_York", "locale": "en-US", "theme": "lightning",
+        },
+        {
+            "userId": "user-3", "firstName": "Diego", "lastName": "Morales",
+            "email": "diego.morales@vertexcloud.example.com", "phone": "(555) 204-1182",
+            "title": "Customer Success Manager", "department": "Customer Success", "role": "CSM",
+            "avatar": "https://i.pravatar.cc/150?u=user-3",
+            "timezone": "America/Chicago", "locale": "en-US", "theme": "lightning",
+        },
+    ]
+
+    accounts = [
+        {
+            "accountId": "account-1", "name": "Northwind Traders", "type": "Customer",
+            "industry": "Logistics", "revenue": 48000000, "employees": 320, "ownerId": "user-1",
+            "phone": "(206) 555-0142", "website": "https://www.northwind-traders.example.com",
+            "billingStreet": "401 Harbor Way", "billingCity": "Seattle", "billingState": "WA",
+            "billingZip": "98101", "billingCountry": "USA",
+            "shippingStreet": "401 Harbor Way", "shippingCity": "Seattle", "shippingState": "WA",
+            "shippingZip": "98101", "shippingCountry": "USA",
+            "createdDate": "2024-02-10T00:00:00.000Z", "modifiedDate": "2026-05-20T00:00:00.000Z",
+        },
+        {
+            "accountId": "account-2", "name": "Globex Corporation", "type": "Customer",
+            "industry": "Manufacturing", "revenue": 92000000, "employees": 540, "ownerId": "user-1",
+            "phone": "(312) 555-0177", "website": "https://www.globex.example.com",
+            "billingStreet": "88 Industrial Blvd", "billingCity": "Chicago", "billingState": "IL",
+            "billingZip": "60601", "billingCountry": "USA",
+            "shippingStreet": "88 Industrial Blvd", "shippingCity": "Chicago", "shippingState": "IL",
+            "shippingZip": "60601", "shippingCountry": "USA",
+            "createdDate": "2023-11-04T00:00:00.000Z", "modifiedDate": "2026-04-18T00:00:00.000Z",
+        },
+        {
+            "accountId": "account-3", "name": "Initech", "type": "Customer",
+            "industry": "Technology", "revenue": 27000000, "employees": 180, "ownerId": "user-1",
+            "phone": "(512) 555-0190", "website": "https://www.initech.example.com",
+            "billingStreet": "1200 Congress Ave", "billingCity": "Austin", "billingState": "TX",
+            "billingZip": "78701", "billingCountry": "USA",
+            "shippingStreet": "1200 Congress Ave", "shippingCity": "Austin", "shippingState": "TX",
+            "shippingZip": "78701", "shippingCountry": "USA",
+            "createdDate": "2024-06-22T00:00:00.000Z", "modifiedDate": "2026-03-30T00:00:00.000Z",
+        },
+    ]
+
+    contacts = [
+        {
+            "contactId": "contact-1", "accountId": "account-1", "firstName": "Karen",
+            "lastName": "Walsh", "title": "VP Operations", "department": "Operations",
+            "email": "karen.walsh@northwind-traders.example.com", "phone": "(206) 555-0148",
+            "ownerId": "user-1",
+        },
+        {
+            "contactId": "contact-2", "accountId": "account-2", "firstName": "Hank",
+            "lastName": "Scorpio", "title": "COO", "department": "Executive",
+            "email": "hank.scorpio@globex.example.com", "phone": "(312) 555-0178",
+            "ownerId": "user-1",
+        },
+        {
+            "contactId": "contact-3", "accountId": "account-3", "firstName": "Bill",
+            "lastName": "Lumbergh", "title": "Director of IT", "department": "IT",
+            "email": "bill.lumbergh@initech.example.com", "phone": "(512) 555-0191",
+            "ownerId": "user-1",
+        },
+    ]
+
+    opportunities = [
+        {
+            "opportunityId": "opp-1", "name": "Northwind Traders – Annual Renewal",
+            "accountId": "account-1", "contactId": "contact-1", "amount": 120000,
+            "closeDate": "2026-07-15", "stage": "Negotiation", "probability": 75,
+            "ownerId": "user-1", "type": "Renewal",
+            "createdDate": "2026-04-01T00:00:00.000Z", "modifiedDate": "2026-05-28T00:00:00.000Z",
+        },
+        {
+            "opportunityId": "opp-2", "name": "Globex Corp – Platform Renewal",
+            "accountId": "account-2", "contactId": "contact-2", "amount": 210000,
+            "closeDate": "2026-09-30", "stage": "Qualification", "probability": 40,
+            "ownerId": "user-1", "type": "Renewal",
+            "createdDate": "2026-05-02T00:00:00.000Z", "modifiedDate": "2026-05-30T00:00:00.000Z",
+        },
+        {
+            "opportunityId": "opp-3", "name": "Initech – Renewal & Upsell",
+            "accountId": "account-3", "contactId": "contact-3", "amount": 85000,
+            "closeDate": "2026-08-20", "stage": "Value Proposition", "probability": 55,
+            "ownerId": "user-1", "type": "Renewal",
+            "createdDate": "2026-05-10T00:00:00.000Z", "modifiedDate": "2026-06-01T00:00:00.000Z",
+        },
+    ]
+
+    # Distractor activities — old COMPLETED tasks unrelated to the renewal.
+    activities = [
+        {
+            "activityId": "activity-1", "type": "task", "subject": "Send onboarding deck to Globex",
+            "status": "Completed", "priority": "Normal", "dueDate": "2026-04-12",
+            "relatedToType": "opportunity", "relatedToId": "opp-2", "assignedToId": "user-1",
+        },
+        {
+            "activityId": "activity-2", "type": "task", "subject": "Update Initech billing contact",
+            "status": "Completed", "priority": "Low", "dueDate": "2026-05-05",
+            "relatedToType": "account", "relatedToId": "account-3", "assignedToId": "user-1",
+        },
+        {
+            "activityId": "activity-3", "type": "event", "subject": "Quarterly business review (internal)",
+            "status": "Completed", "priority": "Normal",
+            "startDateTime": "2026-05-18T15:00:00.000Z", "endDateTime": "2026-05-18T16:00:00.000Z",
+            "relatedToType": "account", "relatedToId": "account-1", "assignedToId": "user-1",
+        },
+    ]
+
+    return {
+        "user": jordan,
+        "users": users,
+        "leads": [],
+        "accounts": accounts,
+        "contacts": contacts,
+        "opportunities": opportunities,
+        "cases": [],
+        "activities": activities,
+        "chatterPosts": [],
+        "files": [],
+        "following": ["user-2", "user-3"],
+        "recentlyViewed": [],
+        "dismissedNotifications": [],
+    }
+
+
+# ---------------------------------------------------------------------------
+# GMAIL STATE
+# ---------------------------------------------------------------------------
+def build_gmail_state():
+    jordan_user = {
+        "userId": "u1",
+        "username": "Jordan Avery",
+        "email": "jordan.avery@vertexcloud.example.com",
+        "avatar": "https://i.pravatar.cc/150?u=jordan-avery",
+    }
+    # Distractor prior thread from Karen Walsh about general account questions.
+    emails = [
+        {
+            "id": "email_1", "threadId": "thread_1",
+            "from": {"name": "Karen Walsh", "email": "karen.walsh@northwind-traders.example.com"},
+            "to": [{"name": "Jordan Avery", "email": "jordan.avery@vertexcloud.example.com"}],
+            "cc": [], "bcc": [],
+            "subject": "Question about our account users",
+            "body": "<p>Hi Jordan,</p><p>Quick question — a couple of our new ops analysts need "
+                    "access to the reporting dashboards. Can you point me to where I add seats?</p>"
+                    "<p>Thanks,<br/>Karen</p>",
+            "timestamp": "2026-06-10T14:05:00Z",
+            "read": True, "starred": False, "important": False,
+            "labels": ["l1"], "category": "primary", "folder": "inbox", "attachments": [],
+        },
+        {
+            "id": "email_2", "threadId": "thread_1",
+            "from": {"name": "Jordan Avery", "email": "jordan.avery@vertexcloud.example.com"},
+            "to": [{"name": "Karen Walsh", "email": "karen.walsh@northwind-traders.example.com"}],
+            "cc": [], "bcc": [],
+            "subject": "Re: Question about our account users",
+            "body": "<p>Hi Karen,</p><p>Happy to help — you can add seats under Settings → "
+                    "Users → Invite. Let me know if you hit any snags.</p><p>Best,<br/>Jordan</p>",
+            "timestamp": "2026-06-10T16:20:00Z",
+            "read": True, "starred": False, "important": False,
+            "labels": ["l1"], "category": "primary", "folder": "sent", "attachments": [],
+        },
+        {
+            "id": "email_3", "threadId": "thread_2",
+            "from": {"name": "Vertex Cloud Billing", "email": "billing@vertexcloud.example.com"},
+            "to": [{"name": "Jordan Avery", "email": "jordan.avery@vertexcloud.example.com"}],
+            "cc": [], "bcc": [],
+            "subject": "June invoice summary",
+            "body": "<p>Your June account summary is now available in the billing portal.</p>",
+            "timestamp": "2026-06-15T09:00:00Z",
+            "read": False, "starred": False, "important": False,
+            "labels": ["l4"], "category": "primary", "folder": "inbox", "attachments": [],
+        },
+    ]
+    labels = [
+        {"id": "l1", "name": "Work", "color": "#ef4444"},
+        {"id": "l2", "name": "Personal", "color": "#3b82f6"},
+        {"id": "l3", "name": "Travel", "color": "#22c55e"},
+        {"id": "l4", "name": "Finance", "color": "#eab308"},
+    ]
+    return {"user": jordan_user, "emails": emails, "labels": labels, "drafts": []}
+
+
+# ---------------------------------------------------------------------------
+# SLACK STATE
+# ---------------------------------------------------------------------------
+def build_slack_state():
+    jordan = {
+        "userId": "user_1", "fullName": "Jordan Avery", "displayName": "Jordan",
+        "email": "jordan.avery@vertexcloud.example.com",
+        "avatar": "https://picsum.photos/200/200?random=1", "status": "online",
+        "statusMessage": "", "statusEmoji": "", "timeZone": "America/New_York",
+    }
+    users = [
+        jordan,
+        {"userId": "user_2", "fullName": "Priya Nair", "displayName": "Priya",
+         "email": "priya.nair@vertexcloud.example.com",
+         "avatar": "https://picsum.photos/200/200?random=2", "status": "online",
+         "statusMessage": "", "statusEmoji": "", "timeZone": "America/New_York"},
+        {"userId": "user_3", "fullName": "Diego Morales", "displayName": "Diego",
+         "email": "diego.morales@vertexcloud.example.com",
+         "avatar": "https://picsum.photos/200/200?random=3", "status": "online",
+         "statusMessage": "", "statusEmoji": "", "timeZone": "America/Chicago"},
+        {"userId": "user_4", "fullName": "Mei Lin", "displayName": "Mei",
+         "email": "mei.lin@vertexcloud.example.com",
+         "avatar": "https://picsum.photos/200/200?random=4", "status": "away",
+         "statusMessage": "", "statusEmoji": "", "timeZone": "America/Los_Angeles"},
+    ]
+    channels = [
+        {"channelId": "general", "name": "general", "description": "Company-wide announcements",
+         "topic": "", "isPrivate": False, "isStarred": False,
+         "members": ["user_1", "user_2", "user_3", "user_4"], "createdBy": "user_2",
+         "createdAt": "2025-01-05T10:00:00Z", "pinnedMessages": [], "unreadCount": 0},
+        {"channelId": "random", "name": "random", "description": "Non-work banter",
+         "topic": "", "isPrivate": False, "isStarred": False,
+         "members": ["user_1", "user_2", "user_3", "user_4"], "createdBy": "user_2",
+         "createdAt": "2025-01-05T10:00:00Z", "pinnedMessages": [], "unreadCount": 0},
+        {"channelId": "customer-success", "name": "customer-success",
+         "description": "Customer Success team — renewals, onboarding, escalations",
+         "topic": "Keeping customers happy and renewing", "isPrivate": False, "isStarred": True,
+         "members": ["user_1", "user_2", "user_3"], "createdBy": "user_3",
+         "createdAt": "2025-02-01T09:00:00Z", "pinnedMessages": [], "unreadCount": 0},
+    ]
+    messages = {
+        "general": [
+            {"messageId": "msg_1", "senderId": "user_2", "content": "Morning team! Don't forget the all-hands at 11.",
+             "timestamp": "2026-06-22T13:30:00Z", "threadId": None, "reactions": [],
+             "attachments": [], "isEdited": False},
+        ],
+        "random": [
+            {"messageId": "msg_2", "senderId": "user_4", "content": "Anyone tried the new coffee place on 3rd?",
+             "timestamp": "2026-06-21T18:10:00Z", "threadId": None, "reactions": [],
+             "attachments": [], "isEdited": False},
+        ],
+        "customer-success": [
+            {"messageId": "msg_3", "senderId": "user_3",
+             "content": "Reminder: please keep renewal notes updated in Salesforce as you work accounts.",
+             "timestamp": "2026-06-19T15:45:00Z", "threadId": None, "reactions": [],
+             "attachments": [], "isEdited": False},
+        ],
+    }
+    return {
+        "currentUser": jordan,
+        "workspace": {"workspaceId": "ws_1", "workspaceName": "Vertex Cloud", "icon": ""},
+        "users": users,
+        "channels": channels,
+        "messages": messages,
+        "threads": {},
+        "dms": [],
+        "bookmarkedMessages": [],
+        "callHistory": [],
+        "settings": {"theme": "light", "notifications": "all", "displayDensity": "comfortable",
+                     "showAvatars": True, "use24Hour": False},
+        "invitations": [],
+        "notifications": [],
+    }
+
+
+def inject(url, state, label):
+    resp = http_post(f'{url}/post?sid={sid}', {'action': 'set', 'state': state})
+    assert resp.status_code == 200, f'{label} injection failed: {resp.status_code} {resp.text}'
+    go = http_get(f'{url}/go?sid={sid}').json()
+    assert go.get('initial_state') is not None, f'{label} initial_state is None after injection'
+    print(f'{label} state injected and verified.')
+
+
+inject(SALESFORCE_URL, build_salesforce_state(), 'Salesforce')
+inject(GMAIL_URL, build_gmail_state(), 'Gmail')
+inject(SLACK_URL, build_slack_state(), 'Slack')
+
+
+# --- Launch browser (primary mock = Salesforce) ---
+def launch_gui(command, delay_sec=1.0):
+    env = os.environ.copy()
+    env['DISPLAY'] = ':0'
+    # Accept both a command string and a pre-split arg list (some tasks
+    # build arg lists themselves, e.g. se_hard_008).
+    if isinstance(command, str):
+        _parts = shlex.split(command)
+    else:
+        _parts = list(command)
+    # Inject CDP debug port so wait_mocks_loaded can verify browser render
+    # via Playwright connect_over_cdp. Idempotent: skip if any arg already
+    # starts with --remote-debugging-port (some tasks inject their own
+    # dynamic port, e.g. se_hard_008).
+    if _parts and 'google-chrome' in _parts[0] \
+       and not any(p.startswith('--remote-debugging-port') for p in _parts):
+        _parts[1:1] = ['--no-first-run', '--no-default-browser-check',
+                       '--remote-debugging-port=1337']
+    if _parts and 'google-chrome' in _parts[0] \
+       and not any(p.startswith('--user-data-dir') for p in _parts):
+        _parts[1:1] = [f'--user-data-dir={CHROME_PROFILE_DIR}']
+    _proc = subprocess.Popen(_parts, stdin=subprocess.DEVNULL,
+                             stdout=subprocess.DEVNULL,
+                             stderr=subprocess.DEVNULL, env=env,
+                             start_new_session=True)
+    time.sleep(delay_sec)
+    if _proc.poll() is not None:
+        print('  WARNING: launch_gui process exited early (code=%s)'
+              % _proc.returncode)
+    return _proc
+_primary_mock_url = SALESFORCE_URL
+launch_gui(f'google-chrome "{SALESFORCE_URL}/?sid={sid}"', delay_sec=2.0)
+_open_remaining_mock_tabs(_primary_mock_url)
+wait_mocks_loaded()
+print(f'GUI_READY: launched browser at {SALESFORCE_URL}/?sid={sid}')

am_renewal_outreach_002/reward.py

@@ -0,0 +1,427 @@
+# --- MOCK_HOST_NO_PROXY_OVERRIDE: bypass woa proxy for locally-deployed mocks ---
+import os as _os
+_np = _os.environ.get("no_proxy", _os.environ.get("NO_PROXY", ""))
+for _h in ("127.0.0.1", "localhost", "28.7.184.198"):
+    if _h not in _np:
+        _np = (_np + "," + _h).lstrip(",")
+_os.environ["no_proxy"] = _np
+_os.environ["NO_PROXY"] = _np
+# --- end override ---
+
+"""
+Reward Script: AM Renewal Outreach — Northwind Traders renewal check-in
+Task ID: am_renewal_outreach_002
+Domain: mock_websites (salesforce_mock + gmail_mock + slack_mock)
+
+Scoring (1.0 total), all programmatic / deterministic — verifies only the
+NEW artifacts the agent must create (identified by id-set diff vs initial_state,
+so pre-existing distractor records — incl. the old sent email to Karen Walsh and
+the old customer-success message — never earn points):
+
+  SALESFORCE task (0.34):
+    - 0.18  NEW 'task' activity related to the Northwind renewal (opp-1 / account-1),
+            with renewal/check-in intent and assigned to Jordan (user-1 when present)
+    - 0.16  that task has dueDate 2026-07-01 AND status 'Open' (not completed)
+  GMAIL email (0.33):
+    - 0.13  NEW email in 'sent' folder addressed to karen.walsh@northwind-traders.example.com
+    - 0.10  its subject matches the explicitly required subject, allowing case/spacing/punctuation variants
+    - 0.10  its body proposes a renewal call before the renewal timing/date
+  SLACK message (0.33):
+    - 0.16  NEW message posted by Jordan (user_1) in the 'customer-success' channel
+    - 0.17  that message says Northwind/Karen renewal outreach is underway/sent
+"""
+import html
+import os
+import re
+import sys
+from datetime import date
+
+import requests
+
+# ---------------------------------------------------------------------------
+# sid
+# ---------------------------------------------------------------------------
+try:
+    with open('/tmp/task_web_sid') as f:
+        sid = f.read().strip()
+    if not sid:
+        raise ValueError('sid is empty')
+except Exception as e:
+    print(f'CRITICAL: Cannot read sid from /tmp/task_web_sid: {e}')
+    print('REWARD: 0.0')
+    sys.exit(0)
+
+MOCKS = {
+    'salesforce': 'http://28.7.184.198:8175',
+    'gmail': 'http://28.7.184.198:8138',
+    'slack': 'http://28.7.184.198:8178',
+}
+
+# ---------------------------------------------------------------------------
+# Egress proxy helper (mandatory)
+# ---------------------------------------------------------------------------
+PROXY_CANDIDATES = [
+    os.environ.get('HTTPS_PROXY') or os.environ.get('https_proxy'),
+        None,
+]
+
+
+def resolve_proxy(probe_url):
+    try:
+        requests.get(probe_url, timeout=8)
+        return None
+    except Exception:
+        pass
+    for candidate in PROXY_CANDIDATES:
+        if not candidate:
+            continue
+        try:
+            requests.get(probe_url, timeout=12,
+                         proxies={'http': candidate, 'https': candidate})
+            return candidate
+        except Exception:
+            continue
+    return None
+
+
+PROXY = resolve_proxy(f"{MOCKS['salesforce']}/go?sid=conn-probe")
+PROXIES = {'http': PROXY, 'https': PROXY} if PROXY else None
+
+
+def fetch(name):
+    url = MOCKS[name]
+    data = requests.get(f'{url}/go?sid={sid}', timeout=20, proxies=PROXIES).json()
+    return data.get('initial_state') or {}, data.get('current_state') or {}
+
+
+# ---------------------------------------------------------------------------
+# Fetch all three mock states up front (precondition gate)
+# ---------------------------------------------------------------------------
+try:
+    sf_init, sf_cur = fetch('salesforce')
+    gm_init, gm_cur = fetch('gmail')
+    sl_init, sl_cur = fetch('slack')
+except Exception as e:
+    print(f'CRITICAL: Cannot fetch mock state: {e}')
+    print('REWARD: 0.0')
+    sys.exit(0)
+
+
+def strip_html(s):
+    return re.sub(r'<[^>]+>', ' ', str(s or ''))
+
+
+TARGET_EMAIL_SUBJECT = 'Northwind Traders renewal call before July 15'
+EXPECTED_TASK_DUE_DATE = date(2026, 7, 1)
+EXPECTED_RENEWAL_DATE = date(2026, 7, 15)
+
+
+def norm(s):
+    text = html.unescape(strip_html(s))
+    text = re.sub(r'[\u2010-\u2015]', '-', text)
+    return re.sub(r'\s+', ' ', text).strip().casefold()
+
+
+def loose(s):
+    """Normalize user-entered prose for case/whitespace/punctuation-tolerant matching."""
+    return re.sub(r'\s+', ' ', re.sub(r'[^a-z0-9]+', ' ', norm(s))).strip()
+
+
+def contains_any(text, phrases):
+    haystack = f" {loose(text)} "
+    return any(f" {loose(phrase)} " in haystack for phrase in phrases)
+
+
+def text_blob(*values):
+    return ' '.join(str(v or '') for v in values)
+
+
+def canonical_subject(s):
+    return loose(s)
+
+
+def subject_matches_target(subject):
+    return canonical_subject(subject) == canonical_subject(TARGET_EMAIL_SUBJECT)
+
+
+def norm_channel_name(s):
+    return norm(s).lstrip('#').strip()
+
+
+MONTH_NAMES = {
+    1: ('january', 'jan'), 2: ('february', 'feb'), 3: ('march', 'mar'),
+    4: ('april', 'apr'), 5: ('may', 'may'), 6: ('june', 'jun'),
+    7: ('july', 'jul'), 8: ('august', 'aug'), 9: ('september', 'sep'),
+    10: ('october', 'oct'), 11: ('november', 'nov'), 12: ('december', 'dec'),
+}
+
+
+def ordinal(n):
+    if 10 <= n % 100 <= 20:
+        suffix = 'th'
+    else:
+        suffix = {1: 'st', 2: 'nd', 3: 'rd'}.get(n % 10, 'th')
+    return f'{n}{suffix}'
+
+
+def mentions_date(text, expected):
+    t = norm(text)
+    month_full, month_short = MONTH_NAMES[expected.month]
+    day = expected.day
+    year = expected.year
+    variants = {
+        f'{month_full} {day}', f'{month_full} {ordinal(day)}',
+        f'{month_full} {day}, {year}', f'{month_full} {ordinal(day)}, {year}',
+        f'{month_full} {day} {year}', f'{month_full} {ordinal(day)} {year}',
+        f'{month_short} {day}', f'{month_short} {ordinal(day)}',
+        f'{month_short} {day}, {year}', f'{month_short} {ordinal(day)}, {year}',
+        f'{month_short} {day} {year}', f'{month_short} {ordinal(day)} {year}',
+        f'{day} {month_full}', f'{ordinal(day)} {month_full}',
+        f'{day} {month_short}', f'{ordinal(day)} {month_short}',
+        f'{day} {month_full} {year}', f'{ordinal(day)} {month_full} {year}',
+        f'{day} {month_short} {year}', f'{ordinal(day)} {month_short} {year}',
+        expected.isoformat(),
+    }
+    if any(v in t for v in variants):
+        return True
+
+    m = expected.month
+    numeric_patterns = [
+        rf'\b0?{m}[/.-]0?{day}(?:[/.-]{year})?\b',
+        rf'\b{year}[/.-]0?{m}[/.-]0?{day}\b',
+        rf'\b0?{day}[/.-]0?{m}(?:[/.-]{year})?\b',
+    ]
+    return any(re.search(p, t) for p in numeric_patterns)
+
+
+def mentions_renewal_timing(text):
+    return mentions_date(text, EXPECTED_RENEWAL_DATE) or contains_any(text, [
+        'renewal date', 'renewal deadline', 'renewal close date', 'close date',
+        'before renewal', 'before the renewal', 'before your renewal',
+        'ahead of renewal', 'ahead of the renewal', 'prior to renewal',
+        'prior to the renewal', 'before it renews', 'before the contract renews',
+    ])
+
+
+def proposes_call(text):
+    call_intent = contains_any(text, [
+        'call', 'meeting', 'meet', 'chat', 'connect', 'sync', 'discussion',
+        'discuss', 'conversation', 'touch base', 'talk', 'speak',
+        'schedule time', 'set up time',
+    ])
+    renewal_context = contains_any(text, ['renewal', 'renew', 'account']) \
+        or mentions_renewal_timing(text)
+    return call_intent and renewal_context
+
+
+def date_field_matches(value, expected):
+    if not value:
+        return False
+    text = norm(value)
+    if text[:10] == expected.isoformat():
+        return True
+    return mentions_date(text, expected)
+
+
+total_score = 0.0
+
+# ===========================================================================
+# SALESFORCE — new renewal check-in task (0.34)
+# ===========================================================================
+try:
+    init_ids = {a.get('activityId') for a in sf_init.get('activities', [])}
+    new_acts = [a for a in sf_cur.get('activities', [])
+                if a.get('activityId') not in init_ids]
+
+    # Candidate = NEW task related to the Northwind renewal (opp-1 or account-1)
+    # with renewal/check-in intent. The relation preserves the customer/opportunity
+    # signal, so the subject can use natural wording like "call Karen" or
+    # "renewal follow-up" without requiring exact tokens.
+    sf_task = None
+    for a in new_acts:
+        if norm(a.get('type')) != 'task':
+            continue
+        task_text = text_blob(a.get('subject'), a.get('description'), a.get('comments'), a.get('notes'))
+        related_type = norm(a.get('relatedToType'))
+        related_ok = (related_type == 'opportunity' and a.get('relatedToId') == 'opp-1') \
+            or (related_type == 'account' and a.get('relatedToId') == 'account-1')
+        assigned_ok = not a.get('assignedToId') or a.get('assignedToId') == 'user-1'
+        intent_ok = contains_any(task_text, [
+            'renewal', 'renew', 'check in', 'check-in', 'follow up',
+            'follow-up', 'outreach', 'call', 'meeting', 'touch base',
+            'karen', 'walsh',
+        ])
+        if related_ok and assigned_ok and intent_ok:
+            sf_task = a
+            break
+
+    # Component SF-1 (0.18): correct new task exists, right relation/owner/intent
+    if sf_task is not None:
+        print(f"PASS: SF new renewal task on Northwind (id={sf_task.get('activityId')}, "
+              f"subject={sf_task.get('subject')!r}, related={sf_task.get('relatedToId')}) (0.18)")
+        total_score += 0.18
+
+        # Component SF-2 (0.16): dueDate 2026-07-01 AND status Open
+        due_ok = date_field_matches(sf_task.get('dueDate'), EXPECTED_TASK_DUE_DATE)
+        status_ok = norm(sf_task.get('status')) in ('open', 'not started', 'in progress', 'pending')
+        if due_ok and status_ok:
+            print(f"PASS: SF task dueDate=2026-07-01 and status Open "
+                  f"(status={sf_task.get('status')!r}) (0.16)")
+            total_score += 0.16
+        else:
+            print(f"FAIL: SF task date/status — dueDate={sf_task.get('dueDate')!r} "
+                  f"(want 2026-07-01), status={sf_task.get('status')!r} (want Open)")
+    else:
+        print(f"FAIL: SF — no NEW task related to opp-1/account-1 with "
+              f"renewal/check-in intent assigned to user-1 when present "
+              f"(new activities: {[a.get('activityId') for a in new_acts]})")
+except Exception as e:
+    print(f'ERROR: Salesforce component — {e}')
+
+# ===========================================================================
+# GMAIL — new renewal outreach email to Karen Walsh (0.33)
+# ===========================================================================
+KAREN = 'karen.walsh@northwind-traders.example.com'
+try:
+    init_ids = {e.get('id') for e in gm_init.get('emails', [])}
+    new_emails = [e for e in gm_cur.get('emails', [])
+                  if e.get('id') not in init_ids]
+
+    # NEW sent email addressed to Karen Walsh; prefer the explicitly named
+    # outreach email but keep body scoring independent so one subject typo
+    # does not erase otherwise correct outreach content.
+    sent_to_karen = None
+    sent_to_karen_emails = []
+    exact_subject_email = None
+    for e in new_emails:
+        if norm(e.get('folder')) != 'sent':
+            continue
+        recips = [norm(t.get('email')) for t in e.get('to', [])]
+        if KAREN in recips:
+            sent_to_karen_emails.append(e)
+            if sent_to_karen is None:
+                sent_to_karen = e
+            if subject_matches_target(e.get('subject')):
+                exact_subject_email = e
+
+    # Component GM-1 (0.13): new sent email to Karen Walsh exists
+    if sent_to_karen is not None:
+        print(f"PASS: Gmail NEW sent email to Karen Walsh "
+              f"(id={sent_to_karen.get('id')}, subject={sent_to_karen.get('subject')!r}) (0.13)")
+        total_score += 0.13
+
+        # Component GM-2 (0.10): subject matches the explicitly required instruction.
+        if exact_subject_email is not None:
+            print(f"PASS: Gmail subject matches required subject {TARGET_EMAIL_SUBJECT!r} (0.10)")
+            total_score += 0.10
+        else:
+            print(f"FAIL: Gmail subject — expected required subject {TARGET_EMAIL_SUBJECT!r}; "
+                  f"got {[e.get('subject') for e in sent_to_karen_emails]}")
+
+        # Component GM-3 (0.10): body proposes a renewal call before the renewal timing.
+        # The fixed July-15 signal may appear in the required subject, while the body
+        # can naturally say "before your renewal date".
+        def email_body_ok(e):
+            body = strip_html(e.get('body'))
+            timing_ok = mentions_renewal_timing(body) or subject_matches_target(e.get('subject'))
+            return proposes_call(body) and timing_ok
+
+        body_email = next((e for e in ([exact_subject_email] if exact_subject_email else [])
+                           if e and email_body_ok(e)), None)
+        if body_email is None:
+            body_email = next((e for e in sent_to_karen_emails if email_body_ok(e)), None)
+
+        if body_email is not None:
+            print("PASS: Gmail body proposes a renewal call before the renewal timing/date (0.10)")
+            total_score += 0.10
+        else:
+            print("FAIL: Gmail body — need a sent-to-Karen email body proposing a renewal "
+                  "call/meeting before the renewal timing/date")
+    else:
+        print(f"FAIL: Gmail — no NEW sent email addressed to {KAREN} "
+              f"(new emails: {[e.get('id') for e in new_emails]})")
+except Exception as e:
+    print(f'ERROR: Gmail component — {e}')
+
+# ===========================================================================
+# SLACK — new message in #customer-success (0.33)
+# ===========================================================================
+try:
+    # Resolve the customer-success channel id/name; the UI may expose the
+    # channel as either "customer-success" or "#customer-success".
+    def customer_success_keys(state):
+        keys = set()
+        for c in state.get('channels', []):
+            name = c.get('name')
+            channel_id = c.get('channelId')
+            if norm_channel_name(name) == 'customer-success' \
+                    or norm_channel_name(channel_id) == 'customer-success':
+                for value in (channel_id, name, norm_channel_name(name),
+                              f"#{norm_channel_name(name)}"):
+                    if value:
+                        keys.add(value)
+        return keys
+
+    cs_keys = customer_success_keys(sl_cur) or customer_success_keys(sl_init)
+
+    if not cs_keys:
+        print("FAIL: Slack — 'customer-success' channel not found")
+    else:
+        def messages_for_channel(state, keys):
+            allowed = {norm_channel_name(k) for k in keys if k}
+            messages = []
+            seen = set()
+            for key, vals in (state.get('messages') or {}).items():
+                if key in keys or norm_channel_name(key) in allowed:
+                    for m in vals or []:
+                        msg_id = m.get('messageId') or id(m)
+                        if msg_id not in seen:
+                            seen.add(msg_id)
+                            messages.append(m)
+            return messages
+
+        init_msgs = messages_for_channel(sl_init, cs_keys)
+        cur_msgs = messages_for_channel(sl_cur, cs_keys)
+        init_msg_ids = {m.get('messageId') for m in init_msgs}
+        new_msgs = [m for m in cur_msgs if m.get('messageId') not in init_msg_ids]
+
+        # Component SL-1 (0.16): NEW message by Jordan (user_1) in customer-success
+        my_new = [m for m in new_msgs if m.get('senderId') == 'user_1']
+        if my_new:
+            print(f"PASS: Slack NEW message by user_1 in #customer-success "
+                  f"({len(my_new)} new) (0.16)")
+            total_score += 0.16
+
+            # Component SL-2 (0.17): content says renewal outreach to
+            # Northwind/Karen is underway/sent. The instruction does not require
+            # the Slack note to repeat the fixed July-15 date.
+            def content_ok(m):
+                t = m.get('content')
+                customer = contains_any(t, ['northwind', 'karen walsh', 'karen', 'walsh'])
+                renewal = contains_any(t, ['renewal', 'renew', 'renewing', 'outreach',
+                                           'check in', 'check-in', 'call', 'meeting'])
+                underway = contains_any(t, [
+                    'underway', 'kicked off', 'kicked-off', 'reached out', 'reach out',
+                    'emailed', 'email sent', 'sent email', 'contacted', 'check-in',
+                    'check in', 'sent', 'outreach', 'started', 'initiated',
+                    'in progress', 'following up', 'followed up', 'handled', 'done',
+                ])
+                return customer and renewal and underway
+
+            good = next((m for m in my_new if content_ok(m)), None)
+            if good is not None:
+                print("PASS: Slack message reports Northwind/Karen renewal outreach underway (0.17)")
+                total_score += 0.17
+            else:
+                print(f"FAIL: Slack message content — need Northwind/Karen + renewal/check-in "
+                      f"+ outreach-underway/sent. Contents: {[norm(m.get('content'))[:80] for m in my_new]}")
+        else:
+            print(f"FAIL: Slack — no NEW message by user_1 in #customer-success "
+                  f"(new msgs: {[m.get('messageId') for m in new_msgs]})")
+except Exception as e:
+    print(f'ERROR: Slack component — {e}')
+
+# ===========================================================================
+final_score = round(min(total_score, 1.0), 4)
+print(f'\nScore: {total_score}/1.0')
+print(f'REWARD: {final_score}')

am_renewal_outreach_002/reward_label.json

@@ -0,0 +1,85 @@
+{
+  "reward_file": "/apdcephfs_sh2/share_300000800/user/jackwkwang/long-horizon-tasks/mini_osworld_xiangwu/cache/am_renewal_outreach_002/reward.py",
+  "api_url": "http://28.7.184.184:8000/v1/chat/completions",
+  "model": "Kimi-K2.6",
+  "max_code_chars": 100000,
+  "created_at": "2026-07-03 20:16:52",
+  "label": {
+    "task_id": "am_renewal_outreach_002",
+    "domain": "mock_websites",
+    "summary": "验证代理是否在Salesforce中创建Northwind续签任务、在Gmail中发送给Karen Walsh的续签邮件、并在Slack customer-success频道发布续签外展消息",
+    "is_placeholder": false,
+    "data_sources": [
+      "/tmp/task_web_sid",
+      "salesforce_mock",
+      "gmail_mock",
+      "slack_mock"
+    ],
+    "scoring_components": [
+      {
+        "name": "SF-1: 新建Salesforce续签任务",
+        "weight": 0.18,
+        "description": "检查是否存在新增的Task活动，与Northwind续签相关，分配给Jordan，且意图匹配",
+        "check_logic": "通过id-set diff找出sf_cur中新增的activities；筛选type为task、relatedToId为opp-1或account-1、assignedToId为空或user-1、且subject/description/comments/notes中包含renewal/renew/check in/check-in/follow up/follow-up/outreach/call/meeting/touch base/karen/walsh等关键词的记录",
+        "pass_condition": "存在至少一条满足上述全部条件的新增task记录"
+      },
+      {
+        "name": "SF-2: 任务截止日期与状态",
+        "weight": 0.16,
+        "description": "检查上述新建任务的dueDate为2026-07-01且状态为未完成",
+        "check_logic": "对SF-1中找到的task，检查dueDate字段是否匹配2026-07-01（支持ISO格式或自然语言变体），且status归一化后为open/not started/in progress/pending之一",
+        "pass_condition": "dueDate匹配2026-07-01且status为Open（或其同义词）"
+      },
+      {
+        "name": "GM-1: 新建发送给Karen的邮件",
+        "weight": 0.13,
+        "description": "检查Gmail sent文件夹中是否存在新增的发送给Karen Walsh的邮件",
+        "check_logic": "通过id-set diff找出gm_cur中新增的emails；筛选folder为sent且to列表中包含karen.walsh@northwind-traders.example.com的邮件",
+        "pass_condition": "存在至少一封新增的发送给指定Karen邮箱的sent邮件"
+      },
+      {
+        "name": "GM-2: 邮件主题匹配",
+        "weight": 0.1,
+        "description": "检查邮件主题是否与要求的主题匹配（忽略大小写、空格、标点）",
+        "check_logic": "对发送给Karen的邮件，使用canonical_subject（loose归一化）比较是否等于'Northwind Traders renewal call before July 15'的loose归一化结果",
+        "pass_condition": "至少一封邮件的主题经loose归一化后与目标主题完全一致"
+      },
+      {
+        "name": "GM-3: 邮件正文提议续签通话",
+        "weight": 0.1,
+        "description": "检查邮件正文是否提议在续签日期前进行通话/会议",
+        "check_logic": "对候选邮件（优先主题完全匹配的邮件，否则所有发给Karen的邮件），检查body经strip_html后：proposes_call（包含call/meeting等意图词且包含renewal/account或提及续签时间）且timing_ok（正文提及renewal timing或主题已完全匹配目标主题）",
+        "pass_condition": "至少一封邮件正文表达通话/会议意图，并关联续签上下文或提及续签日期"
+      },
+      {
+        "name": "SL-1: Slack新建消息",
+        "weight": 0.16,
+        "description": "检查在customer-success频道中是否有user_1(Jordan)发布的新消息",
+        "check_logic": "通过channel名称或ID定位customer-success频道；通过id-set diff找出该频道下新增的消息；筛选senderId为user_1的消息",
+        "pass_condition": "存在至少一条user_1在customer-success频道的新增消息"
+      },
+      {
+        "name": "SL-2: 消息内容报告续签外展状态",
+        "weight": 0.17,
+        "description": "检查消息内容是否表明Northwind/Karen的续签外展正在进行或已发送",
+        "check_logic": "对SL-1中user_1的新消息，检查content是否同时满足：customer（包含northwind/karen walsh/karen/walsh）、renewal（包含renewal/renew/renewing/outreach/check in/check-in/call/meeting）、underway（包含underway/kicked off/reached out/emailed/sent email/contacted/sent/started/initiated/in progress/following up/handled/done等）",
+        "pass_condition": "至少一条消息内容同时包含客户标识、续签意图和外展已进行/已发送的表述"
+      }
+    ],
+    "total_max_score": 1.0,
+    "score_aggregation": "各组件分数累加，最终用min(total_score, 1.0)钳制到上限1.0，并四舍五入到4位小数",
+    "failure_modes": [
+      "无法从/tmp/task_web_sid读取sid时打印CRITICAL并返回0.0",
+      "无法从任一mock服务拉取状态时打印CRITICAL并返回0.0",
+      "Salesforce中未找到符合条件的新增task时SF���件不得分",
+      "Gmail中未找到新增sent邮件或收件人不是Karen时GM组件不得分",
+      "Slack中未找到customer-success频道时Slack组件不得分",
+      "Slack中未找到user_1的新消息时SL组件不得分",
+      "各子组件字段检查失败（如日期、状态、主题、正文、内容不匹配）时对应子组件不得分",
+      "各模块异常被捕获后打印ERROR，该模块不得分但不影响其他模块"
+    ],
+    "scoring_logic_explanation": "脚本首先读取/tmp/task_web_sid获取会话ID，失败则直接退出并返回0.0。接着通过HTTP GET拉取salesforce_mock、gmail_mock、slack_mock的initial_state和current_state，拉取失败同样返回0.0。评分采用id-set差分策略：将current_state中的记录ID与initial_state对比，仅对新增的artifacts进行评分，排除预存干扰项。Salesforce模块（0.34分）先检查新增task是否关联opp-1/account-1、分配给user-1且意图匹配续签/检查（0.18分），再检查该task的dueDate为2026-07-01且status为Open（0.16分）。Gmail模块（0.33分）检查新增sent邮件是否发给karen.walsh@northwind-traders.example.com（0.13分），主题是否匹配'Northwind Traders renewal call before July 15'（0.10分），正文是否提议续签通话并关联续签时间（0.10分）。Slack模块（0.33分）先定位customer-success频道，检查user_1的新消息（0.16分），再验证消息内容同时提及Northwind/Karen、续签以及外展正在进行/已发送（0.17分）。各组件独立评分，累加后通过min(total_score, 1.0)钳制上限并四舍五入输出最终REWARD。",
+    "confidence": 1.0
+  },
+  "raw_response": "用户要求我分析一个 reward.py 评分脚本的评分逻辑，并输出一个严格的 JSON 对象。\n\n让我仔细阅读代码并提取所需信息：\n\n1. **task_id**: 从 docstring 中可以看到 \"Task ID: am_renewal_outreach_002\"\n2. **domain**: 从 docstring 中可以看到 \"Domain: mock_websites (salesforce_mock + gmail_mock + slack_mock)\"，所以 domain 是 mock_websites\n3. **summary**: 一句话概括。这个脚本验证代理是否在 Salesforce 中创建了新的 Northwind 续签任务、在 Gmail 中发送了给 Karen Walsh 的续签邮件、在 Slack 的 customer-success 频道发布了续签外展消息。\n4. **is_placeholder**: false，因为脚本有详细的评分逻辑，不是占位脚本。\n5. **data_sources**: 脚本从 /tmp/task_web_sid 读取 sid，从 salesforce_mock (http://28.7.186.212:8195)、gmail_mock (http://28.7.186.212:8158)、slack_mock (http://28.7.186.212:8198) 获取状态。\n6. **scoring_components**: \n   - Salesforce 任务 (0.34 total):\n     - Component SF-1 (0.18): 新的 task 活动，与 Northwind 续签相关 (opp-1/account-1)，有续签/检查意图，分配给 Jordan (user-1)\n     - Component SF-2 (0.16): 该任务的 dueDate 是 2026-07-01 且状态为 Open (或 not started/in progress/pending)\n   - Gmail 邮件 (0.33 total):\n     - Component GM-1 (0.13): 新的 sent 邮件发送给 karen.walsh@northwind-traders.example.com\n     - Component GM-2 (0.10): 主题匹配 'Northwind Traders renewal call before July 15'（允许大小写/空格/标点变化）\n     - Component GM-3 (0.10): 正文提议在续签日期前进行续签通话\n   - Slack 消息 (0.33 total):\n     - Component SL-1 (0.16): user_1 (Jordan) 在 customer-success 频道发布的新消息\n     - Component SL-2 (0.17): 消息内容表明 Northwind/Karen 的续签外展正在进行/已发送\n\n7. **total_max_score**: 1.0\n8. **score_aggregation**: 各组件分数相加，最后用 `round(min(total_score, 1.0), 4)` 钳制到上限 1.0。\n9. **failure_modes**: \n   - 无法从 /tmp/task_web_sid 读取 sid → CRITICAL，REWARD: 0.0，sys.exit(0)\n   - 无法获取 mock 状态 → CRITICAL，REWARD: 0.0，sys.exit(0)\n   - 找不到 customer-success 频道 → 打印 FAIL，该组件不得分\n   - 各组件检查失败会打印 FAIL 并跳过加分\n   - 异常会被捕获打印 ERROR 但不退出\n\n10. **scoring_logic_explanation**: 详细解释执行顺序：\n    - 首先读取 /tmp/task_web_sid 获取会话 ID，失败则返回 0.0。\n    - 解析代理设置，尝试直连或经代理访问三个 mock 服务。\n    - 并行（顺序）拉取 salesforce、gmail、slack 的 initial_state 和 current_state，失败则返回 0.0。\n    - 通过 id-set diff 识别每个 mock 中新增的 artifacts，排除预存在的干扰记录。\n    - Salesforce: 检查新增 task 是否关联 opp-1/account-1、分配给 user-1、意图匹配续签/检查；再检查 dueDate 和状态。\n    - Gmail: 检查新增 sent 邮件是否发给 Karen Walsh；主题是否匹配目标；正文是否提议续签通话并提及续签时间。\n    - Slack: 定位 customer-success 频道，检查 user_1 的新消息；内容是否包含 Northwind/Karen、续签、外展进行中/已发送。\n    - 累加各通过组件分值，最终用 min(total_score, 1.0) 钳制并四舍五入到 4 位小数。\n\n11. **confidence**: 1.0，因为代码非常清晰完整。\n\n让我再仔细检查一下各个组件的权重和逻辑：\n\nSalesforce:\n- SF-1: 0.18 (new task, related to opp-1/account-1, assigned to user-1 when present, intent matches)\n- SF-2: 0.16 (dueDate 2026-07-01 AND status Open)\n- Total: 0.34\n\nGmail:\n- GM-1: 0.13 (new sent email to KAREN)\n- GM-2: 0.10 (subject matches target)\n- GM-3: 0.10 (body proposes renewal call before renewal timing)\n- Total: 0.33\n\nSlack:\n- SL-1: 0.16 (new message by user_1 in customer-success)\n- SL-2: 0.17 (content says Northwind/Karen renewal outreach underway/sent)\n- Total: 0.33\n\nGrand total: 0.34 + 0.33 + 0.33 = 1.0\n\nFailure modes:\n- sid read failure\n- mock fetch failure\n- slack channel not found (causes fail for slack components)\n- missing new artifacts\n- incorrect fields/dates/status/subject/body/content\n\n我需要确保 JSON 格式完全符合要求，没有 markdown 代码块，第一个字符是 {，最后一个字符是 }。\n\n让我构建 JSON：\n\n```json\n{\n  \"task_id\": \"am_renewal_outreach_002\",\n  \"domain\": \"mock_websites\",\n  \"summary\": \"验证代理是否在Salesforce中创建Northwind续签任务、在Gmail中发送给Karen Walsh的续签邮件、并在Slack customer-success频道发布续格外展消息\",\n  \"is_placeholder\": false,\n  \"data_sources\": [\n    \"/tmp/task_web_sid\",\n    \"salesforce_mock\",\n    \"gmail_mock\",\n    \"slack_mock\"\n  ],\n  \"scoring_components\": [\n    {\n      \"name\": \"SF-1: 新建Salesforce续签任务\",\n      \"weight\": 0.18,\n      \"description\": \"检查是否存在新增的Task活动，与Northwind续签相关，分配给Jordan，且意图匹配\",\n      \"check_logic\": \"通过id-set diff找出sf_cur中新增的activities；筛选type为task、relatedToId为opp-1或account-1、assignedToId为空或user-1、且subject/description/comments/notes中包含renewal/renew/check in/check-in/follow up/follow-up/outreach/call/meeting/touch base/karen/walsh等关键词的记录\",\n      \"pass_condition\": \"存在至少一条满足上述全部条件的新增task记录\"\n    },\n    {\n      \"name\": \"SF-2: 任务截止日期与状态\",\n      \"weight\": 0.16,\n      \"description\": \"检查上述新建任务的dueDate为2026-07-01且状态为未完成\",\n      \"check_logic\": \"对SF-1中找到的task，检查dueDate字段是否匹配2026-07-01（支持ISO格式或自然语言变体），且status归一化后为open/not started/in progress/pending之一\",\n      \"pass_condition\": \"dueDate匹配2026-07-01且status为Open（或其同义词）\"\n    },\n    {\n      \"name\": \"GM-1: 新建发送给Karen的邮件\",\n      \"weight\": 0.13,\n      \"description\": \"检查Gmail sent文件夹中是否存在新增的发送给Karen Walsh的邮件\",\n      \"check_logic\": \"通过id-set diff找出gm_cur中新增的emails；筛选folder为sent且to列表中包含karen.walsh@northwind-traders.example.com的邮件\",\n      \"pass_condition\": \"存在至少一封新增的发送给指定Karen邮箱的sent邮件\"\n    },\n    {\n      \"name\": \"GM-2: 邮件主题匹配\",\n      \"weight\": 0.10,\n      \"description\": \"检查邮件主题是否与要求的主题匹配（忽略大小写、空格、标点）\",\n      \"check_logic\": \"对发送给Karen的邮件，使用canonical_subject（loose归一化）比较是否等于'Northwind Traders renewal call before July 15'的loose归一化结果\",\n      \"pass_condition\": \"至少一封邮件的主题经loose归一化后与目标主题完全一致\"\n    },\n    {\n      \"name\": \"GM-3: 邮件正文提议续签通话\",\n      \"weight\": 0.10,\n      \"description\": \"检查邮件正文是否提议在续签日期前进行通话/会议\",\n      \"check_logic\": \"对候选邮件（优先主题完全匹配的邮件，否则所有发给Karen的邮件），检查body经strip_html后：proposes_call（包含call/meeting等意图词且包含renewal/account或提及续签时间）且timing_ok（正文提及renewal timing或主题已完全匹配目标主题）\",\n      \"pass_condition\": \"至少一封邮件正文表达通话/会议意图，并关联续签上下文或提及续签日期\"\n    },\n    {\n      \"name\": \"SL-1: Slack新建消息\",\n      \"weight\": 0.16,\n      \"description\": \"检查在customer-success频道中是否有user_1(Jordan)发布的新消息\",\n      \"check_logic\": \"通过channel名称或ID定位customer-success频道；通过id-set diff找出该频道下新增的消息；筛选senderId为user_1的消息\",\n      \"pass_condition\": \"存在至少一条user_1在customer-success频道的新增消息\"\n    },\n    {\n      \"name\": \"SL-2: 消息内容报告续签外展状态\",\n      \"weight\": 0.17,\n      \"description\": \"检查消息内容是否表明Northwind/Karen的续签外展正在进行或已发送\",\n      \"check_logic\": \"对SL-1中user_1的新消息，检查content是否同时满足：customer（包含northwind/karen walsh/karen/walsh）、renewal（包含renewal/renew/renewing/outreach/check in/check-in/call/meeting）、underway（包含underway/kicked off/reached out/emailed/sent email/contacted/sent/started/initiated/in progress/following up/handled/done等）\",\n      \"pass_condition\": \"至少一条消息内容同时包含客户标识、续签意图和外展已进行/已发送的表述\"\n    }\n  ],\n  \"total_max_score\": 1.0,\n  \"score_aggregation\": \"各组件分数累加，最终用min(total_score, 1.0)钳制到上限1.0，并四舍五入到4位小数\",\n  \"failure_modes\": [\n    \"无法从/tmp/task_web_sid读取sid时打印CRITICAL并返回0.0\",\n    \"无法从任一mock服务拉取状态时打印CRITICAL并返回0.0\",\n    \"Salesforce中未找到符合条件的新增task时SF组件不得分\",\n    \"Gmail中未找到新增sent邮件或收件人不是Karen时GM组件不得分\",\n    \"Slack中未找到customer-success频道时Slack组件不得分\",\n    \"Slack中未找到user_1的新消息时SL组件不得分\",\n    \"各子组件字段检查失败（如日期、状态、主题、正文、内容不匹配）时对应子组件不得分\",\n    \"各模块异常被捕获后打印ERROR，该模块不得分但不影响其他模块\"\n  ],\n  \"scoring_logic_explanation\": \"脚本首先读取/tmp/task_web_sid获取会话ID，失败则直接退出并返回0.0。接着通过HTTP GET拉取salesforce_mock、gmail_mock、slack_mock的initial_state和current_state，拉取失败同样返回0.0。评分采用id-set差分策略：将current_state中的记录ID与initial_state对比，仅对新增的artifacts进行评分，排除预存干扰项。Salesforce模块（0.34分）先检查新增task是否关联opp-1/account-1、分配给user-1且意图匹配续签/检查（0.18分），再检查该task的dueDate为2026-07-01且status为Open（0.16分）。Gmail模块（0.33分）检查新增sent邮件是否发给karen.walsh@northwind-traders.example.com（0.13分），主题是否匹配'Northwind Traders renewal call before July 15'（0.10分），正文是否提议续签通话并关联续签时间（0.10分）。Slack模块（0.33分）先定位customer-success频道，检查user_1的新消息（0.16分），再验证消息内容同时提及Northwind/Karen、续签以及外展正在进行/已发送（0.17分）。各组件独立评分，累加后通过min(total_score, 1.0)钳制上限并四舍五入输出最终REWARD。\",\n  \"confidence\": 1.0\n}\n```\n\n我需要确保：\n- 没有 markdown 代码块标记\n- 第一个字符是 {\n- 最后一个字符是 }\n- 所有字符串 value 使用中文（task_id、domain、字段名、代码标识符保留原样）\n- JSON 格式正确\n\n让我再检查一下 `data_sources` 字段。代码中 MOCKS 定义了三个服务的 URL，但通常我们称它们为 salesforce_mock、gmail_mock、slack_mock。文件路径也提到了这些。还有 `/tmp/task_web_sid`。\n\n`scoring_components` 中的 `check_logic` 和 `pass_condition` 需要准确反映代码逻辑。\n\n对于 SF-1:\n- `related_ok`: relatedToType 为 opportunity 且 relatedToId 为 opp-1，或 relatedToType 为 account 且 relatedToId 为 account-1\n- `assigned_ok`: assignedToId 为空或为 user-1\n- `intent_ok`: task_text 包含特定关键词\n- 代码中 `if related_ok and assigned_ok and intent_ok`\n\n对于 SF-2:\n- `due_ok`: date_field_matches 检查 dueDate\n- `status_ok`: norm(status) in ('open', 'not started', 'in progress', 'pending')\n- 注意：代码中 `if due_ok and status_ok` 才加分\n\n对于 GM-1:\n- `sent_to_karen`: folder 为 sent，to 列表包含 KAREN\n\n对于 GM-2:\n- `exact_subject_email`: subject_matches_target\n- `subject_matches_target`: canonical_subject(subject) == canonical_subject(TARGET_EMAIL_SUBJECT)\n\n对于 GM-3:\n- `email_body_ok`: body 的 timing_ok 和 proposes_call\n- `timing_ok`: mentions_renewal_timing(body) or subject_matches_target(e.get('subject'))\n- `proposes_call`: call_intent and renewal_context\n- 注意：代码优先检查 exact_subject_email，然后检查 sent_to_karen_emails\n\n对于 SL-1:\n- `my_new`: new_msgs 中 senderId == 'user_1'\n\n对于 SL-2:\n- `content_ok`: customer and renewal and underway\n\n一切看起来都正确。\n\n关于 `failure_modes`:\n- 代码中 Slack 频道未找到时打印 \"FAIL: Slack — 'customer-success' channel not found\"，然后进入 else 分支？不，代码是：\n```python\n    if not cs_keys:\n        print(\"FAIL: Slack — 'customer-success' channel not found\")\n    else:\n        ...\n```\n所以如果没有 cs_keys，Slack 组件不得分，但不会崩溃。\n\n关于 `score_aggregation`: 代码是 `final_score = round(min(total_score, 1.0), 4)`，所以是累加后钳制。\n\n关于 `is_placeholder`: false，因为脚本有完整的评分逻辑。\n\n关于 `confidence`: 1.0，因为代码非常清晰。\n\n让我再确认一下 domain。docstring 中写的是 \"Domain: mock_websites (salesforce_mock + gmail_mock + slack_mock)\"。按照要求，domain 应该是 mock_websites。\n\ntask_id: am_renewal_outreach_002\n\nsummary 需要一句话概括。\n\n好的，我认为 JSON 已经准备好了�� </think> {\"task_id\":\"am_renewal_outreach_002\",\"domain\":\"mock_websites\",\"summary\":\"验证代理是否在Salesforce中创建Northwind续签任务、在Gmail中发送给Karen Walsh的续签邮件、并在Slack customer-success频道发布续签外展消息\",\"is_placeholder\":false,\"data_sources\":[\"/tmp/task_web_sid\",\"salesforce_mock\",\"gmail_mock\",\"slack_mock\"],\"scoring_components\":[{\"name\":\"SF-1: 新建Salesforce续签任务\",\"weight\":0.18,\"description\":\"检查是否存在新增的Task活动，与Northwind续签相关，分配给Jordan，且意图匹配\",\"check_logic\":\"通过id-set diff找出sf_cur中新增的activities；筛选type为task、relatedToId为opp-1或account-1、assignedToId为空或user-1、且subject/description/comments/notes中包含renewal/renew/check in/check-in/follow up/follow-up/outreach/call/meeting/touch base/karen/walsh等关键词的记录\",\"pass_condition\":\"存在至少一条满足上述全部条件的新增task记录\"},{\"name\":\"SF-2: 任务截止日期与状态\",\"weight\":0.16,\"description\":\"检查上述新建任务的dueDate为2026-07-01且状态为未完成\",\"check_logic\":\"对SF-1中找到的task，检查dueDate字段是否匹配2026-07-01（支持ISO格式或自然语言变体），且status归一化后为open/not started/in progress/pending之一\",\"pass_condition\":\"dueDate匹配2026-07-01且status为Open（或其同义词）\"},{\"name\":\"GM-1: 新建发送给Karen的邮件\",\"weight\":0.13,\"description\":\"检查Gmail sent文件夹中是否存在新增的发送给Karen Walsh的邮件\",\"check_logic\":\"通过id-set diff找出gm_cur中新增的emails；筛选folder为sent且to列表中包含karen.walsh@northwind-traders.example.com的邮件\",\"pass_condition\":\"存在至少一封新增的发送给指定Karen邮箱的sent邮件\"},{\"name\":\"GM-2: 邮件主题匹配\",\"weight\":0.10,\"description\":\"检查邮件主题是否与要求的主题匹配（忽略大小写、空格、标点）\",\"check_logic\":\"对发送给Karen的邮件，使用canonical_subject（loose归一化）比较是否等于'Northwind Traders renewal call before July 15'的loose归一化结果\",\"pass_condition\":\"至少一封邮件的主题经loose归一化后与目标主题完全一致\"},{\"name\":\"GM-3: 邮件正文提议续签通话\",\"weight\":0.10,\"description\":\"检查邮件正文是否提议在续签日期前进行通话/会议\",\"check_logic\":\"对候选邮件（优先主题完全匹配的邮件，否则所有发给Karen的邮件），检查body经strip_html后：proposes_call（包含call/meeting等意图词且包含renewal/account或提及续签时间）且timing_ok（正文提及renewal timing或主题已完全匹配目标主题）\",\"pass_condition\":\"至少一封邮件正文表达通话/会议意图，并关联续签上下文或提及续签日期\"},{\"name\":\"SL-1: Slack新建消息\",\"weight\":0.16,\"description\":\"检查在customer-success频道中是否有user_1(Jordan)发布的新消息\",\"check_logic\":\"通过channel名称或ID定位customer-success频道；通过id-set diff找出该频道下新增的消息；筛选senderId为user_1的消息\",\"pass_condition\":\"存在至少一条user_1在customer-success频道的新增消息\"},{\"name\":\"SL-2: 消息内容报告续签外展状态\",\"weight\":0.17,\"description\":\"检查消息内容是否表明Northwind/Karen的续签外展正在进行或已发送\",\"check_logic\":\"对SL-1中user_1的新消息，检查content是否同时满足：customer（包含northwind/karen walsh/karen/walsh）、renewal（包含renewal/renew/renewing/outreach/check in/check-in/call/meeting）、underway（包含underway/kicked off/reached out/emailed/sent email/contacted/sent/started/initiated/in progress/following up/handled/done等）\",\"pass_condition\":\"至少一条消息内容同时包含客户标识、续签意图和外展已进行/已发送的表述\"}],\"total_max_score\":1.0,\"score_aggregation\":\"各组件分数累加，最终用min(total_score, 1.0)钳制到上限1.0，并四舍五入到4位小数\",\"failure_modes\":[\"无法从/tmp/task_web_sid读取sid时打印CRITICAL并返回0.0\",\"无法从任一mock服务拉取状态时打印CRITICAL并返回0.0\",\"Salesforce中未找到符合条件的新增task时SF组件不得分\",\"Gmail中未找到新增sent邮件或收件人不是Karen时GM组件不得分\",\"Slack中未找到customer-success频道时Slack组件不得分\",\"Slack中未找到user_1的新消息时SL组件不得分\",\"各子组件字段检查失败（如日期、状态、主题、正文、内容不匹配）时对应子组件不得分\",\"各模块异常被捕获后打印ERROR，该模块不得分但不影响其他模块\"],\"scoring_logic_explanation\":\"脚本首先读取/tmp/task_web_sid获取会话ID，失败则直接退出并返回0.0。接着通过HTTP GET拉取salesforce_mock、gmail_mock、slack_mock的initial_state和current_state，拉取失败同样返回0.0。评分采用id-set差分策略：将current_state中的记录ID与initial_state对比，仅对新���的artifacts进行评分，排除预存干扰项。Salesforce模块（0.34分）先检查新增task是否关联opp-1/account-1、分配给user-1且意图匹配续签/检查（0.18分），再检查该task的dueDate为2026-07-01且status为Open（0.16分）。Gmail模块（0.33分）检查新增sent邮件是否发给karen.walsh@northwind-traders.example.com（0.13分），主题是否匹配'Northwind Traders renewal call before July 15'（0.10分），正文是否提议续签通话并关联续签时间（0.10分）。Slack模块（0.33分）先定位customer-success频道，检查user_1的新消息（0.16分），再验证消息内容同时提及Northwind/Karen、续签以及外展正在进行/已发送（0.17分）。各组件独立评分，累加后通过min(total_score, 1.0)钳制上限并四舍五入输出最终REWARD。\",\"confidence\":1.0}"
+}

am_renewal_tracker_003/_cua_gym_vm_bridge.sh

@@ -0,0 +1,106 @@
+#!/bin/bash
+# CUA-Gym VM-side network fix (direct-IP mock access).
+#
+# Runs INSIDE the QEMU guest. The CUA-Gym mock services are now reached by
+# their real IDC address (http://28.7.184.198:81xx) instead of the old
+# localhost:9000-9097 socat bridge, so the VM only needs two fixes:
+#
+#   1. Working DNS — the qcow2-baked /etc/resolv.conf points at the Pod's
+#      cluster DNS, which is unreachable from QEMU SLIRP NAT
+#      ("Temporary failure in name resolution").
+#   2. The mock host excluded from the woa http proxy — otherwise the VM's
+#      Chrome (gsettings proxy) and initial_setup.py's `requests` get a 403
+#      from the proxy when hitting 28.7.184.198:81xx. (The old socat path never
+#      hit this because localhost is bypassed by default.)
+#
+# Idempotent. Scoped to the CUA-Gym bridge step; does not touch the generic
+# OSWorld setup pipeline.
+#
+# Argv:
+#   $1  optional sudo password (default "password"; SetupController
+#       substitutes {CLIENT_PASSWORD} when uploaded as a cache file).
+#   $2  unused (kept for backward compat — was socat first port 9000).
+#   $3  unused (kept for backward compat — was socat last  port 9097).
+#   $4  optional mock host IP (default 28.7.184.198).
+
+set +e
+
+PASS="${1:-password}"
+MOCK_HOST="${4:-28.7.184.198}"
+# Derive the /16 IDC subnet so future port/host tweaks on the same segment
+# stay bypassed too (28.7.184.198 -> 28.7.0.0/16).
+MOCK_NET="$(echo "${MOCK_HOST}" | awk -F. 'NF==4{print $1"."$2".0.0/16"}')"
+
+run_root() {
+    echo "$PASS" | sudo -S -p "" bash -c "$1"
+}
+
+# Detect SLIRP gateway from VM's routing table; fall back to 10.0.2.2.
+GW="$(ip route 2>/dev/null | awk '/default/ {print $3; exit}')"
+GW="${GW:-10.0.2.2}"
+
+# ─── CUA-Gym network fix (Taiji-specific) ─────────────────────────────
+fix_vm_network() {
+    # 0) Discover VM network state at runtime — gateway, own IP/CIDR, SLIRP
+    #    DNS forwarder. Nothing hardcoded except the mock host above.
+    local GW_NET="${GW%.*}"
+    local SLIRP_DNS="${GW_NET}.3"        # QEMU SLIRP DNS proxy is always .3
+    local OWN_CIDR
+    OWN_CIDR="$(ip -4 -o addr show scope global 2>/dev/null \
+                | awk '{print $4; exit}')"
+    local OWN_IP="${OWN_CIDR%/*}"
+    local OWN_NET=""
+    [ -n "${OWN_IP}" ] && OWN_NET="${OWN_IP%.*}.0/24"
+
+    # 1) DNS — point at SLIRP DNS proxy + public fallbacks. chattr +i keeps
+    #    systemd-resolved/NetworkManager from clobbering it.
+    if run_root "printf 'nameserver ${SLIRP_DNS}\\nnameserver 114.114.114.114\\nnameserver 8.8.8.8\\n' > /etc/resolv.conf
+                 chattr +i /etc/resolv.conf 2>/dev/null || true"; then
+        echo "[cua_gym_vm_bridge] DNS set: ${SLIRP_DNS}, 114.114.114.114, 8.8.8.8 (gw=${GW})"
+    else
+        echo "[cua_gym_vm_bridge] WARN: failed to write /etc/resolv.conf"
+    fi
+
+    # 2) Build the proxy bypass list from detected values + the direct-IP
+    #    mock host (and its /16 subnet). THIS is what makes 28.7.184.198:81xx
+    #    reachable from inside the VM — without it the woa proxy 403s it.
+    local ignore_list="['localhost', '127.0.0.0/8', '${GW}'"
+    [ -n "${OWN_IP}" ]    && ignore_list="${ignore_list}, '${OWN_IP}'"
+    [ -n "${OWN_NET}" ]   && ignore_list="${ignore_list}, '${OWN_NET}'"
+    [ -n "${GW_NET}" ]    && ignore_list="${ignore_list}, '${GW_NET}.0/24'"
+    [ -n "${MOCK_HOST}" ] && ignore_list="${ignore_list}, '${MOCK_HOST}'"
+    [ -n "${MOCK_NET}" ]  && ignore_list="${ignore_list}, '${MOCK_NET}'"
+    ignore_list="${ignore_list}]"
+
+    # 3) GNOME proxy ignore-hosts — Chrome bypasses the woa proxy for these.
+    if gsettings set org.gnome.system.proxy ignore-hosts "${ignore_list}" 2>/tmp/gsettings_err; then
+        echo "[cua_gym_vm_bridge] gsettings ignore-hosts = ${ignore_list}"
+    else
+        echo "[cua_gym_vm_bridge] WARN: gsettings ignore-hosts failed: $(cat /tmp/gsettings_err 2>/dev/null)"
+    fi
+
+    # 4) Shell env — comma-separated form for no_proxy/NO_PROXY (requests/curl).
+    local NP="localhost,127.0.0.1,${GW}"
+    [ -n "${OWN_IP}" ]    && NP="${NP},${OWN_IP}"
+    [ -n "${OWN_NET}" ]   && NP="${NP},${OWN_NET}"
+    [ -n "${GW_NET}" ]    && NP="${NP},${GW_NET}.0/24"
+    [ -n "${MOCK_HOST}" ] && NP="${NP},${MOCK_HOST}"
+    [ -n "${MOCK_NET}" ]  && NP="${NP},${MOCK_NET}"
+
+    for var in no_proxy NO_PROXY; do
+        if ! grep -q "export ${var}=${NP}" /home/user/.bashrc 2>/dev/null; then
+            echo "export ${var}=${NP}" >> /home/user/.bashrc
+        fi
+    done
+    run_root "grep -q '^no_proxy=' /etc/environment 2>/dev/null \
+              && sed -i 's|^no_proxy=.*|no_proxy=\"${NP}\"|' /etc/environment \
+              || echo 'no_proxy=\"${NP}\"' >> /etc/environment
+              grep -q '^NO_PROXY=' /etc/environment 2>/dev/null \
+              && sed -i 's|^NO_PROXY=.*|NO_PROXY=\"${NP}\"|' /etc/environment \
+              || echo 'NO_PROXY=\"${NP}\"' >> /etc/environment" >/dev/null 2>&1
+    echo "[cua_gym_vm_bridge] no_proxy=${NP}"
+}
+
+fix_vm_network
+
+echo "[cua_gym_vm_bridge] direct-IP mode: ${MOCK_HOST} (+${MOCK_NET}) bypassed from woa proxy; no socat forwarding needed"

am_renewal_tracker_003/initial_setup.py

@@ -0,0 +1,486 @@
+# --- MOCK_HOST_NO_PROXY_OVERRIDE: bypass woa proxy for locally-deployed mocks ---
+import os as _os
+_np = _os.environ.get("no_proxy", _os.environ.get("NO_PROXY", ""))
+for _h in ("127.0.0.1", "localhost", "28.7.184.198"):
+    if _h not in _np:
+        _np = (_np + "," + _h).lstrip(",")
+_os.environ["no_proxy"] = _np
+_os.environ["NO_PROXY"] = _np
+# --- end override ---
+
+# --- wait_mocks_loaded: wait for mock backends ready AND browser tabs rendered ---
+def wait_mocks_loaded(timeout=60, render_buffer=12):
+    """Wait for mock backends ready AND browser tabs rendered (via CDP).
+
+    Phase 1 — backend readiness: poll each mock's /state?sid= until it
+    returns 200 with 'stored_state' in the body.
+
+    Phase 2 — browser render verification: connect to Chrome's CDP
+    endpoint and wait for every open tab to reach 'networkidle' load
+    state with a non-trivial DOM element count. This confirms the SPA
+    has fetched its /go?sid= data and actually painted, not just that
+    the backend is up.
+
+    CDP port auto-detection: checks globals for _chrome_debug_port /
+    cdp_port / debug_port (set by tasks that use a dynamic port, e.g.
+    se_hard_008's launch_chrome_with_urls); defaults to 1337 for tasks
+    that use launch_gui's --remote-debugging-port injection.
+
+    Falls back to the old render_buffer sleep if Playwright or the CDP
+    endpoint is unavailable.
+    """
+    import urllib.request as _u, time as _t
+    g = globals()
+    _urls = {}
+    for _k, _v in list(g.items()):
+        if isinstance(_v, dict) and _v and all(isinstance(x, str) and x.startswith("http://") for x in _v.values()):
+            _urls.update(_v)
+        elif isinstance(_v, str) and _v.startswith("http://") and (_k.endswith("_URL") or _k.endswith("_url")):
+            _urls[_k] = _v
+    _sid = ""
+    for _k in ("sid", "SID", "task_sid", "session_id"):
+        if isinstance(g.get(_k), str) and g[_k]:
+            _sid = g[_k]; break
+    if not _sid:
+        for _k, _v in g.items():
+            if "sid" in _k.lower() and isinstance(_v, str) and _v:
+                _sid = _v; break
+    if not _urls or not _sid:
+        print("  wait_mocks_loaded: no mocks/sid found (%d urls, sid=%r), skipping" % (len(_urls), bool(_sid)))
+        return
+
+    # ---- Phase 1: backend readiness ----
+    _op = _u.build_opener(_u.ProxyHandler({}))
+    _dl = _t.time() + timeout
+    _todo = list(_urls.items())
+    while _todo and _t.time() < _dl:
+        _rest = []
+        for _n2, _uu in _todo:
+            try:
+                _r = _op.open("%s/state?sid=%s" % (_uu, _sid), timeout=5)
+                if _r.status == 200 and b"stored_state" in _r.read():
+                    print("  %s: backend ready" % _n2); continue
+            except Exception:
+                pass
+            _rest.append((_n2, _uu))
+        _todo = _rest
+        if _todo:
+            _t.sleep(2)
+    _done = len(_urls) - len(_todo)
+    _pend = (" (pending: %s)" % [n for n, _ in _todo]) if _todo else ""
+    print("  mocks backend ready: %d/%d%s" % (_done, len(_urls), _pend))
+
+    # ---- Phase 2: browser render verification via CDP ----
+    # Auto-detect CDP port from globals (set by launch_chrome_with_urls
+    # in tasks that use a dynamic port); default to 1337.
+    _cdp_port = 1337
+    for _pk in ("_chrome_debug_port", "chrome_debug_port", "cdp_port", "debug_port"):
+        _pv = g.get(_pk)
+        if isinstance(_pv, int) and 0 < _pv < 65536:
+            _cdp_port = _pv; break
+    _cdp_ok = False
+    try:
+        from playwright.sync_api import sync_playwright
+        _cdp_dl = _t.time() + max(10.0, _dl - _t.time())  # leftover budget, floor 10s
+        with sync_playwright() as _p:
+            _br = None
+            # Retry-connect: Chrome may still be starting up.
+            while _t.time() < _cdp_dl:
+                try:
+                    _br = _p.chromium.connect_over_cdp("http://localhost:%d" % _cdp_port)
+                    break
+                except Exception:
+                    _t.sleep(0.5)
+            if _br is None:
+                print("  wait_mocks_loaded: CDP connect to localhost:%d failed; "
+                      "is Chrome launched with --remote-debugging-port?" % _cdp_port)
+            else:
+                # Wait for ALL expected tabs to appear, then verify each renders.
+                _n_expected = len(_urls)
+                _pages = []
+                _pg_dl = _t.time() + 15
+                _last_count = -1
+                _stable_since = None
+                while _t.time() < _pg_dl:
+                    _pages = []
+                    for _c in _br.contexts:
+                        _pages.extend(_c.pages)
+                    _cnt = len(_pages)
+                    if _n_expected > 0 and _cnt >= _n_expected:
+                        break  # all expected tabs are open
+                    if _cnt > 0 and _cnt == _last_count:
+                        if _stable_since is None:
+                            _stable_since = _t.time()
+                        elif _t.time() - _stable_since >= 1.5:
+                            break  # tab count stable -> no more tabs opening
+                    else:
+                        _stable_since = None
+                    _last_count = _cnt
+                    _t.sleep(0.5)
+                # Re-fetch final snapshot so late-opening tabs are included.
+                _pages = []
+                for _c in _br.contexts:
+                    _pages.extend(_c.pages)
+                if not _pages:
+                    print("  wait_mocks_loaded: CDP connected but no open tabs found")
+                else:
+                    print("  wait_mocks_loaded: %d tab(s) open (expected %d); "
+                          "verifying render" % (len(_pages), _n_expected))
+                    _ok_cnt = 0
+                    for _pg in _pages:
+                        _ttl = ""
+                        try:
+                            _ttl = _pg.title()
+                        except Exception:
+                            pass
+                        try:
+                            _pg.wait_for_load_state(
+                                "networkidle",
+                                timeout=int(max(5000, (_cdp_dl - _t.time()) * 1000)))
+                            _n = _pg.evaluate("document.querySelectorAll('*').length")
+                            if _n < 20:
+                                print("  wait_mocks_loaded: WARNING tab %r has only %d "
+                                      "elements (render stall?)" % (_ttl, _n))
+                            else:
+                                print("  wait_mocks_loaded: tab %r rendered (%d elements)"
+                                      % (_ttl, _n))
+                                _ok_cnt += 1
+                        except Exception as _e:
+                            print("  wait_mocks_loaded: tab %r render wait failed: %r"
+                                  % (_ttl, _e))
+                    # Require ALL open tabs to render before skipping the
+                    # render_buffer fallback, so no tab is left half-loaded.
+                    if _ok_cnt > 0 and _ok_cnt == len(_pages):
+                        _cdp_ok = True
+                    else:
+                        print("  wait_mocks_loaded: %d/%d tabs rendered, will fall back"
+                              % (_ok_cnt, len(_pages)))
+    except ImportError:
+        print("  wait_mocks_loaded: playwright not installed; cannot verify browser render")
+    except Exception as _e:
+        print("  wait_mocks_loaded: CDP phase error (%r)" % _e)
+
+    if not _cdp_ok:
+        print("  wait_mocks_loaded: falling back to render_buffer=%ds sleep" % render_buffer)
+        _t.sleep(render_buffer)
+    else:
+        print("  wait_mocks_loaded: browser render verified, skipping render_buffer")
+# --- end wait_mocks_loaded ---
+
+# --- _open_remaining_mock_tabs: open all auto-detected mock URLs in new Chrome tabs ---
+def _open_remaining_mock_tabs(primary_url=""):
+    """Open all auto-detected mock URLs in Chrome --new-tab, except primary_url."""
+    g = globals()
+    _urls = set()
+    for _k, _v in list(g.items()):
+        if "PROBE" in _k.upper() or _k.startswith("_"):
+            continue
+        if isinstance(_v, str) and _v.startswith("http://") and (_k.endswith("_URL") or _k.endswith("_url")):
+            _urls.add(_v)
+        elif isinstance(_v, dict) and _v and all(isinstance(x, str) and x.startswith("http://") for x in _v.values()):
+            _urls.update(_v.values())
+    _sid = ""
+    for _k in ("sid", "SID", "task_sid", "session_id"):
+        if isinstance(g.get(_k), str) and g[_k]:
+            _sid = g[_k]; break
+    for _u in sorted(_urls):
+        if _u != primary_url:
+            launch_gui(f'google-chrome --new-tab "{_u}/?sid={_sid}"', delay_sec=1.0)
+# --- end _open_remaining_mock_tabs ---
+
+
+"""
+Initial Setup: Quarterly renewal review — Salesforce open renewals -> Notion tracker.
+Task ID: am_renewal_tracker_003
+Domain: mock_websites (multi-mock: salesforce_mock + notion_mock)
+
+Persona: Jordan Avery (user-1), Account Manager / Renewals Manager at Vertex Cloud.
+         jordan.avery@vertexcloud.example.com. Today = 2026-06-24.
+
+PRE-TASK state ONLY. The Notion 'Renewal Tracker 2026' database MUST contain
+ZERO rows (the agent's job is to add exactly three). Salesforce is read-only.
+
+DUAL-VM NOTE: /tmp is NOT shared between the initial_env and golden_env containers,
+and the orchestrator does not sync the sid. This script therefore generates its OWN
+sid, persists it on THIS VM, and injects the initial (pre-task) state under it.
+golden_patch.py does the symmetric thing on the golden VM with its own sid.
+"""
+import os
+import shlex
+import subprocess
+import time
+import uuid
+
+import requests
+
+# --- Mock registry ---
+MOCKS = {
+    'salesforce': 'http://28.7.184.198:8175',
+    'notion':     'http://28.7.184.198:8166',
+}
+
+# --- Egress proxy ---------------------------------------------------------
+# The mocks are only reachable via the cluster HTTPS proxy on this VM (no direct
+# route to the public xlang.ai hosts). Auto-detect: try direct first, then the
+# known cluster proxy. Chrome is launched through the same proxy.
+PROXY_CANDIDATES = [
+    os.environ.get('HTTPS_PROXY') or os.environ.get('https_proxy'),
+        None,
+]
+
+
+def _resolve_proxy():
+    probe = f"{MOCKS['salesforce']}/go?sid=conn-probe"
+    try:
+        requests.get(probe, timeout=8)
+        return None
+    except Exception:
+        pass
+    for cand in PROXY_CANDIDATES:
+        if not cand:
+            continue
+        try:
+            requests.get(probe, timeout=12, proxies={'http': cand, 'https': cand})
+            return cand
+        except Exception:
+            continue
+    raise RuntimeError('No working route to mock servers (direct and proxies all failed)')
+
+
+PROXY = _resolve_proxy()
+PROXIES = None
+print(f'Egress proxy: {PROXY or "direct"}')
+
+# --- Session id (this VM only) -------------------------------------------
+sid = str(uuid.uuid4())
+with open('/tmp/task_web_sid', 'w') as f:
+    f.write(sid)
+print(f'sid = {sid}')
+
+
+# ---------------------------------------------------------------------------
+# SALESFORCE initial state (read-only system of record for renewals)
+# ---------------------------------------------------------------------------
+def salesforce_state():
+    user_1 = {
+        "userId": "user-1", "firstName": "Jordan", "lastName": "Avery",
+        "email": "jordan.avery@vertexcloud.example.com", "phone": "(555) 310-7742",
+        "title": "Account Manager", "department": "Renewals", "role": "Manager",
+        "avatar": "https://i.pravatar.cc/150?u=user-1",
+        "timezone": "America/New_York", "locale": "en-US", "theme": "lightning",
+    }
+
+    accounts = [
+        {"accountId": "account-1", "name": "Northwind Traders", "type": "Customer",
+         "industry": "Logistics", "revenue": 58000000, "employees": 640, "ownerId": "user-1",
+         "billingStreet": "210 Commerce Blvd", "billingCity": "Columbus", "billingState": "OH",
+         "billingZip": "43004", "billingCountry": "USA"},
+        {"accountId": "account-2", "name": "Globex Corp", "type": "Customer",
+         "industry": "Manufacturing", "revenue": 142000000, "employees": 2100, "ownerId": "user-1",
+         "billingStreet": "1 Globex Center", "billingCity": "San Jose", "billingState": "CA",
+         "billingZip": "95110", "billingCountry": "USA"},
+        {"accountId": "account-3", "name": "Initech", "type": "Customer",
+         "industry": "Fintech", "revenue": 96000000, "employees": 880, "ownerId": "user-1",
+         "billingStreet": "4120 Office Park Dr", "billingCity": "Austin", "billingState": "TX",
+         "billingZip": "73301", "billingCountry": "USA"},
+        {"accountId": "account-4", "name": "Soylent Industries", "type": "Customer",
+         "industry": "Food Tech", "revenue": 47000000, "employees": 410, "ownerId": "user-1",
+         "billingStreet": "88 Green Way", "billingCity": "Sacramento", "billingState": "CA",
+         "billingZip": "95814", "billingCountry": "USA"},
+        {"accountId": "account-5", "name": "Umbrella Health", "type": "Customer",
+         "industry": "Healthcare", "revenue": 210000000, "employees": 3400, "ownerId": "user-1",
+         "billingStreet": "1200 Medical Pkwy", "billingCity": "Raleigh", "billingState": "NC",
+         "billingZip": "27601", "billingCountry": "USA"},
+        {"accountId": "account-6", "name": "Hooli", "type": "Customer",
+         "industry": "Technology", "revenue": 320000000, "employees": 5200, "ownerId": "user-1",
+         "billingStreet": "500 Innovation Dr", "billingCity": "Palo Alto", "billingState": "CA",
+         "billingZip": "94301", "billingCountry": "USA"},
+    ]
+
+    contacts = [
+        {"contactId": "contact-1", "accountId": "account-1", "firstName": "Karen",
+         "lastName": "Walsh", "title": "VP Operations", "department": "Operations",
+         "email": "karen.walsh@northwind-traders.example.com", "phone": "(555) 221-3380",
+         "ownerId": "user-1"},
+        {"contactId": "contact-2", "accountId": "account-2", "firstName": "David",
+         "lastName": "Okafor", "title": "COO", "department": "Executive",
+         "email": "david.okafor@globex.example.com", "phone": "(555) 552-1140",
+         "ownerId": "user-1"},
+        {"contactId": "contact-3", "accountId": "account-3", "firstName": "Priya",
+         "lastName": "Nair", "title": "VP Finance", "department": "Finance",
+         "email": "priya.nair@initech.example.com", "phone": "(555) 884-2201",
+         "ownerId": "user-1"},
+    ]
+
+    opportunities = [
+        # QUALIFIES — closes 2026-07-15 (<= 2026-09-22), open (Negotiation), prob 75 -> Low
+        {"opportunityId": "opp-1", "name": "Northwind Traders – Annual Renewal",
+         "accountId": "account-1", "contactId": "contact-1", "amount": 120000,
+         "closeDate": "2026-07-15", "stage": "Negotiation", "probability": 75,
+         "ownerId": "user-1"},
+        # QUALIFIES — closes 2026-08-30, open (Proposal), prob 55 -> Medium
+        {"opportunityId": "opp-2", "name": "Globex Corp – Platform Renewal",
+         "accountId": "account-2", "contactId": "contact-2", "amount": 85000,
+         "closeDate": "2026-08-30", "stage": "Proposal", "probability": 55,
+         "ownerId": "user-1"},
+        # QUALIFIES — closes 2026-09-10, open (Qualification), prob 35 -> High
+        {"opportunityId": "opp-3", "name": "Initech – Renewal & Upsell",
+         "accountId": "account-3", "contactId": "contact-3", "amount": 240000,
+         "closeDate": "2026-09-10", "stage": "Qualification", "probability": 35,
+         "ownerId": "user-1"},
+        # DISTRACTOR — closes 2026-11-20 (AFTER 90-day window)
+        {"opportunityId": "opp-4", "name": "Soylent Industries – Renewal",
+         "accountId": "account-4", "contactId": None, "amount": 54000,
+         "closeDate": "2026-11-20", "stage": "Negotiation", "probability": 60,
+         "ownerId": "user-1"},
+        # DISTRACTOR — closes 2026-12-31 (AFTER 90-day window)
+        {"opportunityId": "opp-5", "name": "Umbrella Health – Renewal",
+         "accountId": "account-5", "contactId": None, "amount": 300000,
+         "closeDate": "2026-12-31", "stage": "Value Proposition", "probability": 50,
+         "ownerId": "user-1"},
+        # DISTRACTOR — already Closed Won (renewed in the past, 2026-05-30)
+        {"opportunityId": "opp-6", "name": "Hooli – Renewal",
+         "accountId": "account-6", "contactId": None, "amount": 95000,
+         "closeDate": "2026-05-30", "stage": "Closed Won", "probability": 100,
+         "ownerId": "user-1"},
+    ]
+
+    return {
+        "user": user_1,
+        "users": [user_1],
+        "leads": [],
+        "accounts": accounts,
+        "contacts": contacts,
+        "opportunities": opportunities,
+        "cases": [],
+        "activities": [],
+        "chatterPosts": [],
+        "files": [],
+        "following": [],
+        "recentlyViewed": [],
+        "dismissedNotifications": [],
+    }
+
+
+# ---------------------------------------------------------------------------
+# NOTION initial state — 'Renewal Tracker 2026' database with ZERO rows
+# ---------------------------------------------------------------------------
+DB_ID = "db-renewal-tracker-2026"
+
+NOTION_DB_PROPERTIES = [
+    {"id": "prop-account", "name": "Account", "type": "text"},
+    {"id": "prop-arr", "name": "ARR", "type": "number"},
+    {"id": "prop-date", "name": "Renewal Date", "type": "date"},
+    {"id": "prop-risk", "name": "Risk", "type": "select",
+     "options": ["Low", "Medium", "High"]},
+    {"id": "prop-stage", "name": "Renewal Stage", "type": "select",
+     "options": ["Prospecting", "Qualification", "Needs Analysis",
+                 "Value Proposition", "Proposal", "Negotiation"]},
+]
+
+NOTION_VISIBLE_PROPS = ["prop-account", "prop-arr", "prop-date", "prop-risk", "prop-stage"]
+
+
+def notion_db_page(items):
+    return {
+        "id": DB_ID,
+        "title": "Renewal Tracker 2026",
+        "icon": "\U0001F4CA",
+        "cover": None,
+        "parentId": None,
+        "type": "database",
+        "viewType": "table",
+        "properties": NOTION_DB_PROPERTIES,
+        "views": [{
+            "id": "view-table", "name": "Table View", "type": "table",
+            "filters": [], "sorts": [], "groupBy": None,
+            "visibleProperties": NOTION_VISIBLE_PROPS,
+        }],
+        "items": list(items),
+        "blockIds": [],
+        "favorite": False,
+        "createdDate": "2026-06-01T09:00:00.000Z",
+    }
+
+
+def notion_state():
+    return {
+        "user": {"id": "user-1", "name": "Jordan Avery",
+                 "email": "jordan.avery@vertexcloud.example.com", "avatar": ""},
+        "workspace": {"id": "ws-1", "name": "Vertex Cloud Renewals", "icon": "\U0001F3E2",
+                      "members": ["user-1"]},
+        "pages": {
+            DB_ID: notion_db_page([]),  # ZERO rows pre-task
+        },
+        "blocks": {},
+        "trash": [],
+        "comments": {},
+        "settings": {"appearance": "light", "startWeekMonday": False, "fontSize": "default"},
+        "notifications": [],
+        "pageOrder": [DB_ID],
+    }
+
+
+# ---------------------------------------------------------------------------
+# Inject initial state into BOTH mocks (action: "set")
+# ---------------------------------------------------------------------------
+STATE_BUILDERS = {
+    'salesforce': salesforce_state,
+    'notion':     notion_state,
+}
+
+for name, build in STATE_BUILDERS.items():
+    url = MOCKS[name]
+    state = build()
+    resp = requests.post(f'{url}/post?sid={sid}',
+                         json={'action': 'set', 'state': state},
+                         timeout=30, proxies=PROXIES)
+    assert resp.status_code == 200, f'{name} injection failed: {resp.status_code} {resp.text}'
+    go = requests.get(f'{url}/go?sid={sid}', timeout=15, proxies=PROXIES).json()
+    assert go.get('initial_state') is not None, f'{name} initial_state is None after set'
+    print(f'  injected {name}: initial_state OK')
+
+# Sanity: Notion tracker has zero rows pre-task
+ng = requests.get(f'{MOCKS["notion"]}/go?sid={sid}', timeout=15, proxies=PROXIES).json()
+rows = ng['current_state']['pages'][DB_ID]['items']
+assert rows == [], f'Expected 0 rows pre-task, found {len(rows)}'
+print(f'  notion tracker rows pre-task: {len(rows)} (correct)')
+
+print(f'Both mocks injected with sid={sid}')
+
+
+# ---------------------------------------------------------------------------
+# GUI-ready: launch Chrome on Salesforce (primary app) through the proxy
+# ---------------------------------------------------------------------------
+def launch_gui(command, delay_sec=1.0):
+    env = os.environ.copy()
+    env['DISPLAY'] = ':0'
+    # Accept both a command string and a pre-split arg list (some tasks
+    # build arg lists themselves, e.g. se_hard_008).
+    if isinstance(command, str):
+        _parts = shlex.split(command)
+    else:
+        _parts = list(command)
+    # Inject CDP debug port so wait_mocks_loaded can verify browser render
+    # via Playwright connect_over_cdp. Idempotent: skip if any arg already
+    # starts with --remote-debugging-port (some tasks inject their own
+    # dynamic port, e.g. se_hard_008).
+    if _parts and 'google-chrome' in _parts[0] \
+       and not any(p.startswith('--remote-debugging-port') for p in _parts):
+        _parts[1:1] = ['--no-first-run', '--no-default-browser-check',
+                       '--remote-debugging-port=1337']
+    _proc = subprocess.Popen(_parts, stdin=subprocess.DEVNULL,
+                             stdout=subprocess.DEVNULL,
+                             stderr=subprocess.DEVNULL, env=env,
+                             start_new_session=True)
+    time.sleep(delay_sec)
+    if _proc.poll() is not None:
+        print('  WARNING: launch_gui process exited early (code=%s)'
+              % _proc.returncode)
+    return _proc
+_chrome_proxy = f'--proxy-server={PROXY}' if PROXY else ''
+_primary_mock_url = MOCKS["salesforce"]
+launch_gui(f'google-chrome "{MOCKS["salesforce"]}/?sid={sid}"', delay_sec=2.0)
+_open_remaining_mock_tabs(_primary_mock_url)
+wait_mocks_loaded()
+print(f'GUI_READY: launched browser at {MOCKS["salesforce"]}/?sid={sid}')

am_renewal_tracker_003/reward.py

@@ -0,0 +1,427 @@
+# --- MOCK_HOST_NO_PROXY_OVERRIDE: bypass woa proxy for locally-deployed mocks ---
+import os as _os
+_np = _os.environ.get("no_proxy", _os.environ.get("NO_PROXY", ""))
+for _h in ("127.0.0.1", "localhost", "28.7.184.198"):
+    if _h not in _np:
+        _np = (_np + "," + _h).lstrip(",")
+_os.environ["no_proxy"] = _np
+_os.environ["NO_PROXY"] = _np
+# --- end override ---
+
+"""
+Reward Script: Renewals tracker hygiene (Salesforce -> Notion 'Renewal Tracker 2026')
+Task ID: am_renewal_tracker_003
+Domain: mock_websites (notion_mock)
+
+This task is FULLY DETERMINISTIC. The agent must add EXACTLY 3 rows to the
+'Renewal Tracker 2026' Notion database, one per qualifying open renewal that
+closes within 90 days of 2026-06-24 (i.e. closeDate <= 2026-09-22) and is still
+open. Risk is derived from the opportunity's win probability.
+
+Expected rows (ground truth):
+  1. Northwind Traders | ARR 120000 | 2026-07-15 | Risk Low    | Stage Negotiation
+  2. Globex Corp       | ARR  85000 | 2026-08-30 | Risk Medium | Stage Proposal
+  3. Initech           | ARR 240000 | 2026-09-10 | Risk High   | Stage Qualification
+
+Distractors that MUST NOT appear: Soylent Industries (renews 2026-11-20, outside
+window), Umbrella Health (2026-12-31, outside window), Hooli (Closed Won, past).
+
+Scoring (programmatic only — NO LLM judge; the judge is unavailable here):
+  - 0.30 per correctly-populated expected row (ARR, Renewal Date, Risk,
+    Renewal Stage are parsed/normalized), keyed by Account name  -> max 0.90
+  - 0.10 cleanliness bonus: the database contains EXACTLY 3 rows AND none of the
+    distractor accounts are present.
+  Total = 1.0
+
+Initial state has 0 rows -> every component fails -> 0.0.
+"""
+
+import os
+import re
+import sys
+from datetime import date, datetime, timedelta
+
+import requests
+
+NOTION_URL = 'http://28.7.184.198:8166'
+DB_TITLE = 'Renewal Tracker 2026'
+TODAY_TEXT = 'June 24, 2026'
+WINDOW_DAYS = 90
+
+# Ground-truth expected rows (derived from task_config business rules, not from peeking).
+EXPECTED_ROWS = [
+    {'Account': 'Northwind Traders', 'ARR': 120000, 'Renewal Date': '2026-07-15', 'Risk': 'Low', 'Renewal Stage': 'Negotiation'},
+    {'Account': 'Globex Corp', 'ARR': 85000, 'Renewal Date': '2026-08-30', 'Risk': 'Medium', 'Renewal Stage': 'Proposal'},
+    {'Account': 'Initech', 'ARR': 240000, 'Renewal Date': '2026-09-10', 'Risk': 'High', 'Renewal Stage': 'Qualification'},
+]
+DISTRACTOR_ACCOUNTS = {'soylent industries', 'umbrella health', 'hooli'}
+
+# ----------------------------------------------------------------------------
+# Read sid
+# ----------------------------------------------------------------------------
+try:
+    with open('/tmp/task_web_sid') as f:
+        sid = f.read().strip()
+    if not sid:
+        raise ValueError('sid is empty')
+except Exception as e:
+    print(f'CRITICAL: Cannot read sid from /tmp/task_web_sid: {e}')
+    print('REWARD: 0.0')
+    sys.exit(0)
+
+# ----------------------------------------------------------------------------
+# Egress proxy helper (MANDATORY)
+# ----------------------------------------------------------------------------
+PROXY_CANDIDATES = [
+    os.environ.get('HTTPS_PROXY') or os.environ.get('https_proxy'),
+        None,
+]
+
+
+def resolve_proxy(probe_url):
+    try:
+        requests.get(probe_url, timeout=8)
+        return None
+    except Exception:
+        pass
+    for candidate in PROXY_CANDIDATES:
+        if not candidate:
+            continue
+        try:
+            requests.get(probe_url, timeout=12, proxies={'http': candidate, 'https': candidate})
+            return candidate
+        except Exception:
+            continue
+    return None
+
+
+PROXY = resolve_proxy(f'{NOTION_URL}/go?sid=conn-probe')
+PROXIES = None
+print(f'Egress proxy: {PROXY or "direct"}')
+
+# ----------------------------------------------------------------------------
+# Fetch notion state
+# ----------------------------------------------------------------------------
+try:
+    data = requests.get(f'{NOTION_URL}/go?sid={sid}', timeout=15, proxies=PROXIES).json()
+except Exception as e:
+    print(f'CRITICAL: Cannot fetch notion state: {e}')
+    print('REWARD: 0.0')
+    sys.exit(0)
+
+current = data.get('current_state') or {}
+if not current:
+    print('CRITICAL: current_state is None/empty')
+    print('REWARD: 0.0')
+    sys.exit(0)
+
+pages = current.get('pages') or {}
+
+
+# ----------------------------------------------------------------------------
+# Helpers
+# ----------------------------------------------------------------------------
+MONTHS = {
+    'jan': 1, 'january': 1, 'feb': 2, 'february': 2,
+    'mar': 3, 'march': 3, 'apr': 4, 'april': 4, 'may': 5,
+    'jun': 6, 'june': 6, 'jul': 7, 'july': 7,
+    'aug': 8, 'august': 8, 'sep': 9, 'sept': 9, 'september': 9,
+    'oct': 10, 'october': 10, 'nov': 11, 'november': 11,
+    'dec': 12, 'december': 12,
+}
+
+PROPERTY_ALIASES = {
+    'Account': ('Account', 'Account Name', 'Customer', 'Customer Name', 'Company'),
+    'ARR': ('ARR', 'Annual Contract Value', 'Annual Recurring Revenue', 'Amount', 'Contract Value'),
+    'Renewal Date': ('Renewal Date', 'Close Date', 'Closing Date', 'Renewal Close Date', 'Date'),
+    'Risk': ('Risk', 'Risk Level', 'Renewal Risk'),
+    'Renewal Stage': ('Renewal Stage', 'Current Sales Stage', 'Sales Stage', 'Stage'),
+}
+
+RISK_ALIASES = {
+    'Low': ('low', 'low risk', 'green', 'healthy'),
+    'Medium': ('medium', 'medium risk', 'med', 'moderate', 'yellow'),
+    'High': ('high', 'high risk', 'at risk', 'red'),
+}
+
+STAGE_ALIASES = {
+    'Negotiation': ('negotiation', 'negotiating', 'negotiate'),
+    'Proposal': ('proposal', 'proposal sent', 'proposed'),
+    'Qualification': ('qualification', 'qualified', 'qualifying'),
+}
+
+
+def text_from_value(v):
+    """Extract a display value from common Notion primitive/object shapes."""
+    if v is None:
+        return ''
+    if isinstance(v, (int, float, bool, date, datetime)):
+        return str(v)
+    if isinstance(v, list):
+        parts = [text_from_value(x) for x in v]
+        return ' '.join(p for p in parts if p)
+    if isinstance(v, dict):
+        for key in (
+            'name', 'label', 'plain_text', 'text', 'content', 'title', 'start',
+            'date', 'number', 'value', 'email', 'url',
+        ):
+            if key in v and v.get(key) not in (None, ''):
+                return text_from_value(v.get(key))
+        for key in ('rich_text', 'select', 'status'):
+            if key in v and v.get(key) not in (None, ''):
+                return text_from_value(v.get(key))
+        parts = [text_from_value(x) for x in v.values()]
+        return ' '.join(p for p in parts if p)
+    return str(v)
+
+
+def norm_str(v):
+    return re.sub(r'\s+', ' ', text_from_value(v)).strip()
+
+
+def canon_text(v):
+    return re.sub(r'[^a-z0-9]+', ' ', norm_str(v).lower()).strip()
+
+
+def norm_num(v):
+    """Parse int/float or display strings like '$120,000', '120k', or 'USD 120000'."""
+    if v is None:
+        return None
+    if isinstance(v, (int, float)):
+        return float(v)
+    s = norm_str(v).lower().replace(',', '')
+    s = s.replace('$', '').replace('usd', '').replace('arr', '')
+    match = re.search(r'-?\d+(?:\.\d+)?\s*([km])?', s)
+    if not match:
+        return None
+    try:
+        number = float(re.match(r'-?\d+(?:\.\d+)?', match.group(0)).group(0))
+    except ValueError:
+        return None
+    suffix = match.group(1)
+    if suffix == 'k':
+        number *= 1000
+    elif suffix == 'm':
+        number *= 1000000
+    return number
+
+
+def norm_date(v):
+    """Return a date object from common date/datetime strings and Notion date objects."""
+    if v is None:
+        return None
+    if isinstance(v, datetime):
+        return v.date()
+    if isinstance(v, date):
+        return v
+    if isinstance(v, dict):
+        v = v.get('start') or v.get('date') or ''
+    s = norm_str(v)
+    if not s:
+        return None
+    s = re.sub(r'(\d)(st|nd|rd|th)\b', r'\1', s, flags=re.IGNORECASE)
+    iso = s.replace('Z', '+00:00')
+    try:
+        return datetime.fromisoformat(iso).date()
+    except ValueError:
+        pass
+
+    numeric = re.search(r'\b(\d{4})[-/.](\d{1,2})[-/.](\d{1,2})\b', s)
+    if numeric:
+        return build_date(numeric.group(1), numeric.group(2), numeric.group(3))
+
+    slash = re.search(r'\b(\d{1,2})[-/.](\d{1,2})[-/.](\d{2,4})\b', s)
+    if slash:
+        first, second, year = slash.groups()
+        if int(first) > 12:
+            day, month = first, second
+        else:
+            month, day = first, second
+        return build_date(year, month, day)
+
+    month_first = re.search(r'\b([A-Za-z]+)\s+(\d{1,2})(?:,)?\s+(\d{4})\b', s)
+    if month_first and month_first.group(1).lower() in MONTHS:
+        month = MONTHS[month_first.group(1).lower()]
+        return build_date(month_first.group(3), month, month_first.group(2))
+
+    day_first = re.search(r'\b(\d{1,2})\s+([A-Za-z]+)(?:,)?\s+(\d{4})\b', s)
+    if day_first and day_first.group(2).lower() in MONTHS:
+        month = MONTHS[day_first.group(2).lower()]
+        return build_date(day_first.group(3), month, day_first.group(1))
+
+    return None
+
+
+def build_date(year, month, day):
+    try:
+        year = int(year)
+        if year < 100:
+            year += 2000
+        return date(year, int(month), int(day))
+    except (TypeError, ValueError):
+        return None
+
+
+def find_database(pages_map, title):
+    target = canon_text(title)
+    for pid, p in pages_map.items():
+        if not isinstance(p, dict):
+            continue
+        if p.get('type') == 'database' and canon_text(p.get('title')) == target:
+            return pid, p
+    return None, None
+
+
+def prop_id_by_alias(db, aliases):
+    """Resolve a property id from likely display names (case/punctuation-insensitive)."""
+    alias_keys = {canon_text(alias) for alias in aliases}
+    for prop in db.get('properties', []) or []:
+        if canon_text(prop.get('name')) in alias_keys:
+            return prop.get('id')
+    return None
+
+
+def prop_value(row, prop_id, aliases=()):
+    props = row.get('properties', {}) or {}
+    if prop_id and prop_id in props:
+        return props.get(prop_id)
+
+    alias_keys = {canon_text(alias) for alias in aliases}
+    for key, value in props.items():
+        if canon_text(key) in alias_keys:
+            return value
+    for alias in aliases:
+        if alias in row:
+            return row.get(alias)
+    return None
+
+
+def row_account_name(row, account_pid):
+    """Account may live in the Account property OR in the page title."""
+    value = prop_value(row, account_pid, PROPERTY_ALIASES['Account'])
+    if value not in (None, ''):
+        return norm_str(value)
+    return norm_str(row.get('title'))
+
+
+def account_matches(candidate, expected):
+    candidate_key = company_key(candidate)
+    expected_key = company_key(expected)
+    return (
+        candidate_key == expected_key
+        or expected_key in candidate_key
+        or candidate_key in expected_key
+    )
+
+
+def company_key(value):
+    tokens = ['corp' if token == 'corporation' else token
+              for token in canon_text(value).split()]
+    return ' '.join(tokens)
+
+
+def select_matches(value, expected, aliases):
+    value_key = canon_text(value)
+    expected_key = canon_text(expected)
+    if value_key == expected_key or expected_key in value_key:
+        return True
+    return any(canon_text(alias) == value_key or canon_text(alias) in value_key
+               for alias in aliases.get(expected, ()))
+
+
+def verify_task():
+    total_score = 0.0
+    today = norm_date(TODAY_TEXT)
+    cutoff = today + timedelta(days=WINDOW_DAYS) if today else None
+
+    # Precondition gate: the database must exist.
+    db_id, db = find_database(pages, DB_TITLE)
+    if db is None:
+        print(f"CRITICAL: database titled '{DB_TITLE}' not found")
+        print('REWARD: 0.0')
+        return 0.0
+
+    # Resolve property ids by name/alias (robust to differing prop-id schemes).
+    pid_account = prop_id_by_alias(db, PROPERTY_ALIASES['Account'])
+    pid_arr = prop_id_by_alias(db, PROPERTY_ALIASES['ARR'])
+    pid_date = prop_id_by_alias(db, PROPERTY_ALIASES['Renewal Date'])
+    pid_risk = prop_id_by_alias(db, PROPERTY_ALIASES['Risk'])
+    pid_stage = prop_id_by_alias(db, PROPERTY_ALIASES['Renewal Stage'])
+
+    # Collect rows: prefer the database's ordered items list, fall back to
+    # any page whose parentId points at this database.
+    item_ids = db.get('items') or []
+    rows = [pages[i] for i in item_ids if i in pages and isinstance(pages[i], dict)]
+    if not rows:
+        rows = [p for pid, p in pages.items()
+                if isinstance(p, dict) and p.get('parentId') == db_id and p.get('type') != 'database']
+
+    print(f'Found {len(rows)} row(s) in database "{DB_TITLE}"')
+
+    # --- Components 1-3: each expected row, all required fields correct (0.30 each) ---
+    for idx, exp in enumerate(EXPECTED_ROWS, start=1):
+        try:
+            row = next(
+                (r for r in rows
+                 if account_matches(row_account_name(r, pid_account), exp['Account'])),
+                None,
+            )
+            if row is None:
+                print(f"FAIL: Component {idx} — row for account '{exp['Account']}' not found")
+                continue
+
+            arr_value = prop_value(row, pid_arr, PROPERTY_ALIASES['ARR'])
+            date_value = prop_value(row, pid_date, PROPERTY_ALIASES['Renewal Date'])
+            risk_value = prop_value(row, pid_risk, PROPERTY_ALIASES['Risk'])
+            stage_value = prop_value(row, pid_stage, PROPERTY_ALIASES['Renewal Stage'])
+
+            expected_date = norm_date(exp['Renewal Date'])
+            actual_date = norm_date(date_value)
+            arr_ok = norm_num(arr_value) == float(exp['ARR'])
+            date_ok = (
+                actual_date == expected_date
+                and today is not None
+                and cutoff is not None
+                and today <= actual_date <= cutoff
+            )
+            risk_ok = select_matches(risk_value, exp['Risk'], RISK_ALIASES)
+            stage_ok = select_matches(stage_value, exp['Renewal Stage'], STAGE_ALIASES)
+
+            if arr_ok and date_ok and risk_ok and stage_ok:
+                print(f"PASS: Component {idx} — '{exp['Account']}' fully correct "
+                      f"(ARR={exp['ARR']}, Date={exp['Renewal Date']}, Risk={exp['Risk']}, Stage={exp['Renewal Stage']}) (0.30 pts)")
+                total_score += 0.30
+            else:
+                print(f"FAIL: Component {idx} — '{exp['Account']}' field mismatch: "
+                      f"ARR_ok={arr_ok}(got {arr_value!r}), "
+                      f"Date_ok={date_ok}(got {actual_date!r}), "
+                      f"Risk_ok={risk_ok}(got {risk_value!r}), "
+                      f"Stage_ok={stage_ok}(got {stage_value!r})")
+        except Exception as e:
+            print(f'ERROR: Component {idx} — {e}')
+
+    # --- Component 4: cleanliness — exactly 3 rows AND no distractors (0.10) ---
+    try:
+        present_accounts = [row_account_name(r, pid_account) for r in rows]
+        distractors_present = {
+            distractor
+            for distractor in DISTRACTOR_ACCOUNTS
+            if any(account_matches(account, distractor) for account in present_accounts)
+        }
+        exact_count = (len(rows) == 3)
+        if exact_count and not distractors_present:
+            print('PASS: Component 4 — exactly 3 rows, no distractor accounts (0.10 pts)')
+            total_score += 0.10
+        else:
+            print(f'FAIL: Component 4 — row_count={len(rows)} (expected 3), '
+                  f'distractors_present={sorted(distractors_present)}')
+    except Exception as e:
+        print(f'ERROR: Component 4 — {e}')
+
+    final_score = round(min(total_score, 1.0), 4)
+    print(f'\nScore: {total_score}/1.0')
+    print(f'REWARD: {final_score}')
+    return final_score
+
+
+verify_task()

am_renewal_tracker_003/reward_label.json

@@ -0,0 +1,61 @@
+{
+  "reward_file": "/apdcephfs_sh2/share_300000800/user/jackwkwang/long-horizon-tasks/mini_osworld_xiangwu/cache/am_renewal_tracker_001/reward.py",
+  "api_url": "http://28.7.184.184:8000/v1/chat/completions",
+  "model": "Kimi-K2.6",
+  "max_code_chars": 100000,
+  "created_at": "2026-07-03 19:42:48",
+  "label": {
+    "task_id": "am_renewal_tracker_003",
+    "domain": "mock_websites",
+    "summary": "验证 agent 是否向 Notion 数据库 'Renewal Tracker 2026' 正确添加了恰好 3 条符合条件的续约跟踪记录，且未包含干扰项",
+    "is_placeholder": false,
+    "data_sources": [
+      "/tmp/task_web_sid",
+      "notion_mock (http://28.7.186.212:8186)"
+    ],
+    "scoring_components": [
+      {
+        "name": "Component 1 (Northwind Traders)",
+        "weight": 0.3,
+        "description": "检查是否存在 Northwind Traders 账户对应的记录，且 ARR、Renewal Date、Risk、Renewal Stage 四个字段与预期值完全匹配",
+        "check_logic": "在数据库行中通过 account_matches 查找账户名匹配的行，提取 ARR、Renewal Date、Risk、Renewal Stage 字段值，分别使用 norm_num、norm_date、select_matches 与预期值比较；其中日期还需满足 today(2026-06-24) <= actual_date <= cutoff(2026-09-22)",
+        "pass_condition": "找到该行且 arr_ok=True、date_ok=True、risk_ok=True、stage_ok=True"
+      },
+      {
+        "name": "Component 2 (Globex Corp)",
+        "weight": 0.3,
+        "description": "检查是否存在 Globex Corp 账户对应的记录，且四个字段与预期值完全匹配",
+        "check_logic": "同 Component 1，预期 ARR=85000、Date=2026-08-30、Risk=Medium、Stage=Proposal",
+        "pass_condition": "找到该行且四个字段全部匹配"
+      },
+      {
+        "name": "Component 3 (Initech)",
+        "weight": 0.3,
+        "description": "检查是否存在 Initech 账户对应的记录，且四个字段与预期值完全匹配",
+        "check_logic": "同 Component 1，预期 ARR=240000、Date=2026-09-10、Risk=High、Stage=Qualification",
+        "pass_condition": "找到该行且四个字段全部匹配"
+      },
+      {
+        "name": "Component 4 (Cleanliness)",
+        "weight": 0.1,
+        "description": "检查数据库总行数是否恰好为 3，且不存在任何干扰账户",
+        "check_logic": "检查 len(rows)==3；遍历所有行的账户名，与 DISTRACTOR_ACCOUNTS ('soylent industries', 'umbrella health', 'hooli') 进行 account_matches 比对，确认无匹配项",
+        "pass_condition": "总行数等于 3 且 distractors_present 为空集合"
+      }
+    ],
+    "total_max_score": 1.0,
+    "score_aggregation": "各组件分数累加，最终通过 round(min(total_score, 1.0), 4) 计算，上限钳制在 1.0",
+    "failure_modes": [
+      "无法从 /tmp/task_web_sid 读取 sid 或 sid 为空 -> 打印 CRITICAL 并返回 0.0",
+      "无法通过 HTTP GET 获取 notion_mock 状态 -> 打印 CRITICAL 并返回 0.0",
+      "current_state 为空或不存在 -> 打印 CRITICAL 并返回 0.0",
+      "未找到标题为 'Renewal Tracker 2026' 的数据库 -> 打印 CRITICAL 并返回 0.0",
+      "某预期账户对应的行未找到 -> 该 Component 不得分",
+      "某预期账户字段解析或匹配失败 -> 该 Component 不得分",
+      "数据库行数不等于 3 或存在干扰账户 -> Component 4 不得分"
+    ],
+    "scoring_logic_explanation": "脚本首先读取 /tmp/task_web_sid 获取会话 ID，然后通过 HTTP 请求拉取 notion_mock 的当前状态。在返回的 JSON 中找到名为 'Renewal Tracker 2026' 的数据库，并收集其所有行记录（优先使用 db.items，否则回退到 parentId 匹配）。接着逐项检查 3 个预期账户（Northwind Traders、Globex Corp、Initech）是否存在，且 ARR、Renewal Date、Risk、Renewal Stage 四个字段与预期一致（日期还需落在 2026-06-24 至 2026-09-22 的 90 天窗口内），每通过一项加 0.30 分。最后检查清洁度：数据库必须恰好有 3 行且不含任何干扰账户（Soylent Industries、Umbrella Health、Hooli），通过则加 0.10 分。总分累加后通过 min 钳制在 1.0，再四舍五入到 4 位小数作为最终 REWARD 输出。",
+    "confidence": 1.0
+  },
+  "raw_response": "用户要求我分析一个 reward.py 评分脚本的评分逻辑，并输出一个严格的 JSON 对象。\n\n让我仔细阅读代码：\n\n1. **Task ID**: 从 docstring 中可以看到 `Task ID: am_renewal_tracker_001`\n2. **Domain**: 从 docstring 中可以看到 `Domain: mock_websites (notion_mock)`\n3. **Summary**: 这个任务要求 agent 向 Notion 数据库 \"Renewal Tracker 2026\" 添加恰好 3 行数据，对应特定的 Salesforce 续约机会。\n4. **is_placeholder**: false，因为脚本有详细的评分逻辑。\n5. **data_sources**: 脚本从 `/tmp/task_web_sid` 读取 sid，然后向 `http://28.7.186.212:8166/go?sid={sid}` 发送请求获取 Notion mock 状态。\n6. **scoring_components**:\n   - Component 1-3: 每���预期行（Northwind Traders, Globex Corp, Initech）正确填充所有字段（ARR, Renewal Date, Risk, Renewal Stage），每个 0.30 分。\n   - Component 4: 清洁度奖励，数据库恰好有 3 行且没有干扰账户，0.10 分。\n7. **total_max_score**: 1.0\n8. **score_aggregation**: 累加，然后 `min(total_score, 1.0)` 并四舍五入到 4 位小数。\n9. **failure_modes**:\n   - 无法读取 sid -> 0.0\n   - 无法获取 Notion 状态 -> 0.0\n   - current_state 为空 -> 0.0\n   - 找不到数据库 -> 0.0\n   - 行数不对或有干扰账户 -> Component 4 失败\n   - 字段不匹配 -> 对应 Component 失败\n10. **scoring_logic_explanation**: 按执行顺序说明。\n11. **confidence**: 1.0，因为代码逻辑清晰。\n\n让我详细检查每个组件：\n\n**Component 1-3**:\n- 遍历 EXPECTED_ROWS（3个账户）\n- 在数据库行中查找匹配账户名的行\n- 检查 ARR: `norm_num(arr_value) == float(exp['ARR'])`\n- 检查 Date: `actual_date == expected_date` 且 `today <= actual_date <= cutoff`（2026-06-24 到 2026-09-22）\n- 检查 Risk: `select_matches` 使用 RISK_ALIASES\n- 检查 Stage: `select_matches` 使用 STAGE_ALIASES\n- 全部通过得 0.30\n\n**Component 4**:\n- `exact_count = (len(rows) == 3)`\n- 检查是否有 DISTRACTOR_ACCOUNTS 中的账户存在\n- 如果恰好3行且无干扰账户，得 0.10\n\n**数据获取**:\n- 从 `/tmp/task_web_sid` 读取 sid\n- 通过 `requests.get(f'{NOTION_URL}/go?sid={sid}')` 获取状态\n- 解析 `current_state.pages`\n- 查找数据库 `Renewal Tracker 2026`\n- 收集行：优先使用 `db.get('items')`，否则查找 `parentId == db_id` 的页面\n\n**字段解析**:\n- 支持多种属性名别名（PROPERTY_ALIASES）\n- 支持多种风险/阶段别名\n- 支持多种日期格式和数字格式\n\n现在构建 JSON：\n\n```json\n{\n  \"task_id\": \"am_renewal_tracker_001\",\n  \"domain\": \"mock_websites\",\n  \"summary\": \"验证 agent 是否向 Notion 数据库 'Renewal Tracker 2026' 正确添加了恰好 3 条符合条件的续约跟踪记录，且未包含干扰项\",\n  \"is_placeholder\": false,\n  \"data_sources\": [\"/tmp/task_web_sid\", \"notion_mock (http://28.7.186.212:8186)\"],\n  \"scoring_components\": [\n    {\n      \"name\": \"Component 1 (Northwind Traders)\",\n      \"weight\": 0.3,\n      \"description\": \"检查是否存在 Northwind Traders 账户对应的记录，且 ARR、Renewal Date、Risk、Renewal Stage 四个字段与预期值完全匹配\",\n      \"check_logic\": \"在数据库行中查找 account_matches 的行，提取 ARR、Renewal Date、Risk、Renewal Stage 字段值，分别用 norm_num、norm_date、select_matches 与预期值比较；日期还需满足 today(2026-06-24) <= actual_date <= cutoff(2026-09-22)\",\n      \"pass_condition\": \"找到该行且 arr_ok=True, date_ok=True, risk_ok=True, stage_ok=True\"\n    },\n    {\n      \"name\": \"Component 2 (Globex Corp)\",\n      \"weight\": 0.3,\n      \"description\": \"检查是否存在 Globex Corp 账户对应的记录，且四个字段与预期值完全匹配\",\n      \"check_logic\": \"同 Component 1，预期 ARR=85000, Date=2026-08-30, Risk=Medium, Stage=Proposal\",\n      \"pass_condition\": \"找到该行且四个字段全部匹配\"\n    },\n    {\n      \"name\": \"Component 3 (Initech)\",\n      \"weight\": 0.3,\n      \"description\": \"检查是否存在 Initech 账户对应的记录，且四个字段与预期值完全匹配\",\n      \"check_logic\": \"同 Component 1，预期 ARR=240000, Date=2026-09-10, Risk=High, Stage=Qualification\",\n      \"pass_condition\": \"找到该行且四个字段全部匹配\"\n    },\n    {\n      \"name\": \"Component 4 (Cleanliness)\",\n      \"weight\": 0.1,\n      \"description\": \"检查数据库总行数是否恰好为 3，且不存在任何干扰账户\",\n      \"check_logic\": \"检查 len(rows)==3；遍历所有行的账户名，与 DISTRACTOR_ACCOUNTS ('soylent industries', 'umbrella health', 'hooli') 进行 account_matches 比对，确认无匹配\",\n      \"pass_condition\": \"总行数等于 3 且 distractors_present 为空集合\"\n    }\n  ],\n  \"total_max_score\": 1.0,\n  \"score_aggregation\": \"各组件分数累加，最终通过 round(min(total_score, 1.0), 4) 计算，上限钳制在 1.0\",\n  \"failure_modes\": [\n    \"无法从 /tmp/task_web_sid 读取 sid 或 sid 为空 -> 打印 CRITICAL 并返回 0.0\",\n    \"无法通过 HTTP GET 获取 notion mock 状态 -> 打印 CRITICAL 并返回 0.0\",\n    \"current_state 为空或不存在 -> 打印 CRITICAL 并返回 0.0\",\n    \"未找到标题为 'Renewal Tracker 2026' 的数据库 -> 打印 CRITICAL 并返回 0.0\",\n    \"某预期账户对应的行未找到 -> 该 Component 不得分\",\n    \"某预期账户字段解析或匹配失败 -> 该 Component 不得分\",\n    \"数据库行数不等于 3 或存在干扰账户 -> Component 4 不得分\"\n  ],\n  \"scoring_logic_explanation\": \"脚本首先读取 /tmp/task_web_sid 获取会话 ID，然后通过 HTTP 请求拉取 notion_mock 的当前状态。在状态中��到名为 'Renewal Tracker 2026' 的数据库，收集其所有行记录。接着逐项检查 3 个预期账户（Northwind Traders、Globex Corp、Initech）是否存在且 ARR、Renewal Date、Risk、Renewal Stage 四个字段与预期一致（日期还需落在 2026-06-24 至 2026-09-22 窗口内），每通过一项加 0.30 分。最后检查清洁度：数据库必须恰好有 3 行且不含任何干扰账户（Soylent Industries、Umbrella Health、Hooli），通过则加 0.10 分。总分累加后钳制在 1.0 并四舍五入到 4 位小数作为最终 REWARD。\",\n  \"confidence\": 1.0\n}\n```\n\n检查要求：\n- 只输出 JSON，没有 markdown 代码块\n- 第一个字符是 {，最后一个字符是 }\n- 字符串 value 使用中文，task_id、domain、字段名保留原样\n- weight 填写显式分值：0.30 和 0.10\n- is_placeholder: false\n\n确认 domain：docstring 写的是 `Domain: mock_websites (notion_mock)`，所以 domain 应该是 `mock_websites`。\n\n确认 task_id：`am_renewal_tracker_001`。\n\n确认 data_sources：`/tmp/task_web_sid` 和 `http://28.7.186.212:8166`（notion_mock）。\n\n确认评分逻辑中的日期检查：`today <= actual_date <= cutoff`，其中 today 是 2026-06-24，cutoff 是 today + 90 天 = 2026-09-22。这在代码中通过 `norm_date(TODAY_TEXT)` 和 `timedelta(days=WINDOW_DAYS)` 计算。\n\n确认数字解析：`norm_num` 会处理 `$120,000`、`120k` 等格式。\n\n确认属性别名：PROPERTY_ALIASES 中定义了多种可能的列名。\n\n一切看起来都正确。 </think> {\"task_id\":\"am_renewal_tracker_001\",\"domain\":\"mock_websites\",\"summary\":\"验证 agent 是否向 Notion 数据库 'Renewal Tracker 2026' 正确添加了恰好 3 条符合条件的续约跟踪记录，且未包含干扰项\",\"is_placeholder\":false,\"data_sources\":[\"/tmp/task_web_sid\",\"notion_mock (http://28.7.186.212:8186)\"],\"scoring_components\":[{\"name\":\"Component 1 (Northwind Traders)\",\"weight\":0.3,\"description\":\"检查是否存在 Northwind Traders 账户对应的记录，且 ARR、Renewal Date、Risk、Renewal Stage 四个字段与预期值完全匹配\",\"check_logic\":\"在数据库行中通过 account_matches 查找账户名匹配的行，提取 ARR、Renewal Date、Risk、Renewal Stage 字段值，分别使用 norm_num、norm_date、select_matches 与预期值比较；其中日期还需满足 today(2026-06-24) <= actual_date <= cutoff(2026-09-22)\",\"pass_condition\":\"找到该行且 arr_ok=True、date_ok=True、risk_ok=True、stage_ok=True\"},{\"name\":\"Component 2 (Globex Corp)\",\"weight\":0.3,\"description\":\"检查是否存在 Globex Corp 账户对应的记录，且四个字段与预期值完全匹配\",\"check_logic\":\"同 Component 1，预期 ARR=85000、Date=2026-08-30、Risk=Medium、Stage=Proposal\",\"pass_condition\":\"找到该行且四个字段全部匹配\"},{\"name\":\"Component 3 (Initech)\",\"weight\":0.3,\"description\":\"检查是否存在 Initech 账户对应的记录，且四个字段与预期值完全匹配\",\"check_logic\":\"同 Component 1，预期 ARR=240000、Date=2026-09-10、Risk=High、Stage=Qualification\",\"pass_condition\":\"找到该行且四个字段全部匹配\"},{\"name\":\"Component 4 (Cleanliness)\",\"weight\":0.1,\"description\":\"检查数据库总行数是否恰好为 3，且不存在任何干扰账户\",\"check_logic\":\"检查 len(rows)==3；遍历所有行的账户名，与 DISTRACTOR_ACCOUNTS ('soylent industries', 'umbrella health', 'hooli') 进行 account_matches 比对，确认无匹配项\",\"pass_condition\":\"总行数等于 3 且 distractors_present 为空集合\"}],\"total_max_score\":1.0,\"score_aggregation\":\"各组件分数累加，最终通过 round(min(total_score, 1.0), 4) 计算，上限钳制在 1.0\",\"failure_modes\":[\"无法从 /tmp/task_web_sid 读取 sid 或 sid 为空 -> 打印 CRITICAL 并返回 0.0\",\"无法通过 HTTP GET 获取 notion_mock 状态 -> 打印 CRITICAL 并返回 0.0\",\"current_state 为空或不存在 -> 打印 CRITICAL 并返回 0.0\",\"未找到标题为 'Renewal Tracker 2026' 的数据库 -> 打印 CRITICAL 并返回 0.0\",\"某预期账户对应的行未找到 -> 该 Component 不得分\",\"某预期账户字段解析或匹配失败 -> 该 Component 不得分\",\"数据库行数不等于 3 或存在干扰账户 -> Component 4 不得分\"],\"scoring_logic_explanation\":\"脚本首先读取 /tmp/task_web_sid 获取会话 ID，然后通过 HTTP 请求拉取 notion_mock 的当前状态。在返回的 JSON 中找到名为 'Renewal Tracker 2026' 的数据库，并收集其所有行记录（优先使用 db.items，否则回退到 parentId 匹配）。接着逐项检查 3 个预期账户（Northwind Traders、Globex Corp、Initech）是否存在，且 ARR、Renewal Date、Risk、Renewal Stage 四个字段与预期一致（日期还需落在 2026-06-24 至 2026-09-22 的 90 天窗口内），每通过一项加 0.30 分。最后检查清洁度：数据库必须恰好有 3 行且不含任何干扰账户（Soylent Industries、Umbrella Health、Hooli），通过则�� 0.10 分。总分累加后通过 min 钳制在 1.0，再四舍五入到 4 位小数作为最终 REWARD 输出。\",\"confidence\":1.0}"
+}

ar_approval_002/_cua_gym_vm_bridge.sh

@@ -0,0 +1,106 @@
+#!/bin/bash
+# CUA-Gym VM-side network fix (direct-IP mock access).
+#
+# Runs INSIDE the QEMU guest. The CUA-Gym mock services are now reached by
+# their real IDC address (http://28.7.184.198:81xx) instead of the old
+# localhost:9000-9097 socat bridge, so the VM only needs two fixes:
+#
+#   1. Working DNS — the qcow2-baked /etc/resolv.conf points at the Pod's
+#      cluster DNS, which is unreachable from QEMU SLIRP NAT
+#      ("Temporary failure in name resolution").
+#   2. The mock host excluded from the woa http proxy — otherwise the VM's
+#      Chrome (gsettings proxy) and initial_setup.py's `requests` get a 403
+#      from the proxy when hitting 28.7.184.198:81xx. (The old socat path never
+#      hit this because localhost is bypassed by default.)
+#
+# Idempotent. Scoped to the CUA-Gym bridge step; does not touch the generic
+# OSWorld setup pipeline.
+#
+# Argv:
+#   $1  optional sudo password (default "password"; SetupController
+#       substitutes {CLIENT_PASSWORD} when uploaded as a cache file).
+#   $2  unused (kept for backward compat — was socat first port 9000).
+#   $3  unused (kept for backward compat — was socat last  port 9097).
+#   $4  optional mock host IP (default 28.7.184.198).
+
+set +e
+
+PASS="${1:-password}"
+MOCK_HOST="${4:-28.7.184.198}"
+# Derive the /16 IDC subnet so future port/host tweaks on the same segment
+# stay bypassed too (28.7.184.198 -> 28.7.0.0/16).
+MOCK_NET="$(echo "${MOCK_HOST}" | awk -F. 'NF==4{print $1"."$2".0.0/16"}')"
+
+run_root() {
+    echo "$PASS" | sudo -S -p "" bash -c "$1"
+}
+
+# Detect SLIRP gateway from VM's routing table; fall back to 10.0.2.2.
+GW="$(ip route 2>/dev/null | awk '/default/ {print $3; exit}')"
+GW="${GW:-10.0.2.2}"
+
+# ─── CUA-Gym network fix (Taiji-specific) ─────────────────────────────
+fix_vm_network() {
+    # 0) Discover VM network state at runtime — gateway, own IP/CIDR, SLIRP
+    #    DNS forwarder. Nothing hardcoded except the mock host above.
+    local GW_NET="${GW%.*}"
+    local SLIRP_DNS="${GW_NET}.3"        # QEMU SLIRP DNS proxy is always .3
+    local OWN_CIDR
+    OWN_CIDR="$(ip -4 -o addr show scope global 2>/dev/null \
+                | awk '{print $4; exit}')"
+    local OWN_IP="${OWN_CIDR%/*}"
+    local OWN_NET=""
+    [ -n "${OWN_IP}" ] && OWN_NET="${OWN_IP%.*}.0/24"
+
+    # 1) DNS — point at SLIRP DNS proxy + public fallbacks. chattr +i keeps
+    #    systemd-resolved/NetworkManager from clobbering it.
+    if run_root "printf 'nameserver ${SLIRP_DNS}\\nnameserver 114.114.114.114\\nnameserver 8.8.8.8\\n' > /etc/resolv.conf
+                 chattr +i /etc/resolv.conf 2>/dev/null || true"; then
+        echo "[cua_gym_vm_bridge] DNS set: ${SLIRP_DNS}, 114.114.114.114, 8.8.8.8 (gw=${GW})"
+    else
+        echo "[cua_gym_vm_bridge] WARN: failed to write /etc/resolv.conf"
+    fi
+
+    # 2) Build the proxy bypass list from detected values + the direct-IP
+    #    mock host (and its /16 subnet). THIS is what makes 28.7.184.198:81xx
+    #    reachable from inside the VM — without it the woa proxy 403s it.
+    local ignore_list="['localhost', '127.0.0.0/8', '${GW}'"
+    [ -n "${OWN_IP}" ]    && ignore_list="${ignore_list}, '${OWN_IP}'"
+    [ -n "${OWN_NET}" ]   && ignore_list="${ignore_list}, '${OWN_NET}'"
+    [ -n "${GW_NET}" ]    && ignore_list="${ignore_list}, '${GW_NET}.0/24'"
+    [ -n "${MOCK_HOST}" ] && ignore_list="${ignore_list}, '${MOCK_HOST}'"
+    [ -n "${MOCK_NET}" ]  && ignore_list="${ignore_list}, '${MOCK_NET}'"
+    ignore_list="${ignore_list}]"
+
+    # 3) GNOME proxy ignore-hosts — Chrome bypasses the woa proxy for these.
+    if gsettings set org.gnome.system.proxy ignore-hosts "${ignore_list}" 2>/tmp/gsettings_err; then
+        echo "[cua_gym_vm_bridge] gsettings ignore-hosts = ${ignore_list}"
+    else
+        echo "[cua_gym_vm_bridge] WARN: gsettings ignore-hosts failed: $(cat /tmp/gsettings_err 2>/dev/null)"
+    fi
+
+    # 4) Shell env — comma-separated form for no_proxy/NO_PROXY (requests/curl).
+    local NP="localhost,127.0.0.1,${GW}"
+    [ -n "${OWN_IP}" ]    && NP="${NP},${OWN_IP}"
+    [ -n "${OWN_NET}" ]   && NP="${NP},${OWN_NET}"
+    [ -n "${GW_NET}" ]    && NP="${NP},${GW_NET}.0/24"
+    [ -n "${MOCK_HOST}" ] && NP="${NP},${MOCK_HOST}"
+    [ -n "${MOCK_NET}" ]  && NP="${NP},${MOCK_NET}"
+
+    for var in no_proxy NO_PROXY; do
+        if ! grep -q "export ${var}=${NP}" /home/user/.bashrc 2>/dev/null; then
+            echo "export ${var}=${NP}" >> /home/user/.bashrc
+        fi
+    done
+    run_root "grep -q '^no_proxy=' /etc/environment 2>/dev/null \
+              && sed -i 's|^no_proxy=.*|no_proxy=\"${NP}\"|' /etc/environment \
+              || echo 'no_proxy=\"${NP}\"' >> /etc/environment
+              grep -q '^NO_PROXY=' /etc/environment 2>/dev/null \
+              && sed -i 's|^NO_PROXY=.*|NO_PROXY=\"${NP}\"|' /etc/environment \
+              || echo 'NO_PROXY=\"${NP}\"' >> /etc/environment" >/dev/null 2>&1
+    echo "[cua_gym_vm_bridge] no_proxy=${NP}"
+}
+
+fix_vm_network
+
+echo "[cua_gym_vm_bridge] direct-IP mode: ${MOCK_HOST} (+${MOCK_NET}) bypassed from woa proxy; no socat forwarding needed"

ar_approval_002/initial_setup.py

@@ -0,0 +1,240 @@
+"""
+Initial Setup: Salesforce -> Slack discount-approval chain (Hooli Data Migration)
+Task ID: ar_approval_002
+Domain: mock_websites (multi-mock: salesforce_mock + slack_mock)
+
+Initial state:
+  - Salesforce: opportunity 'Hooli Data Migration' (list price $90,000,
+    proposed/discounted amount $72,000, owner Jian Wu, account Hooli).
+  - Slack: #finance-approvals channel exists but contains NO approval message.
+"""
+import os
+import shlex
+import subprocess
+import time
+import uuid
+
+import requests
+
+SF_URL = 'http://28.7.184.198:8175'
+SLACK_URL = 'http://28.7.184.198:8178'
+
+# --- Session id (shared across both mocks) ---
+sid = str(uuid.uuid4())
+with open('/tmp/task_web_sid', 'w') as f:
+    f.write(sid)
+
+
+def launch_gui(command, delay_sec=1.0):
+    env = os.environ.copy()
+    env['DISPLAY'] = ':0'
+    subprocess.Popen(
+        shlex.split(command),
+        stdout=subprocess.DEVNULL,
+        stderr=subprocess.DEVNULL,
+        env=env,
+    )
+    time.sleep(delay_sec)
+
+
+# =====================================================================
+# Salesforce state
+# =====================================================================
+sf_user = {
+    "userId": "user-1", "firstName": "John", "lastName": "Smith",
+    "email": "john.smith@company.com", "phone": "(555) 123-4567",
+    "title": "Sales Manager", "department": "Sales", "role": "Manager",
+    "avatar": "https://i.pravatar.cc/150?u=user-1",
+    "timezone": "America/New_York", "locale": "en-US", "theme": "lightning",
+}
+sf_users = [
+    sf_user,
+    {
+        "userId": "user-2", "firstName": "Jian", "lastName": "Wu",
+        "email": "jian.wu@company.com", "phone": "(555) 234-5678",
+        "title": "Account Executive", "department": "Sales", "role": "User",
+        "avatar": "https://i.pravatar.cc/150?u=user-2",
+        "timezone": "America/Los_Angeles", "locale": "en-US", "theme": "lightning",
+    },
+    {
+        "userId": "user-3", "firstName": "Priya", "lastName": "Natarajan",
+        "email": "priya.natarajan@company.com", "phone": "(555) 345-6789",
+        "title": "Account Executive", "department": "Sales", "role": "User",
+        "avatar": "https://i.pravatar.cc/150?u=user-3",
+        "timezone": "America/Chicago", "locale": "en-US", "theme": "lightning",
+    },
+]
+
+sf_accounts = [
+    {
+        "accountId": "account-1", "name": "Hooli", "phone": "(650) 555-0142",
+        "website": "https://hooli.com", "type": "Customer", "industry": "Technology",
+        "revenue": 120000000, "employees": 1200,
+        "description": "Large technology company; active enterprise account.",
+        "ownerId": "user-2",
+        "billingStreet": "1100 Bridgepointe Pkwy", "billingCity": "San Mateo",
+        "billingState": "CA", "billingZip": "94404", "billingCountry": "United States",
+        "shippingStreet": "1100 Bridgepointe Pkwy", "shippingCity": "San Mateo",
+        "shippingState": "CA", "shippingZip": "94404", "shippingCountry": "United States",
+        "createdDate": "2025-02-11T09:00:00.000Z", "modifiedDate": "2026-06-01T09:00:00.000Z",
+    },
+    {
+        "accountId": "account-2", "name": "Initech", "phone": "(512) 555-0199",
+        "website": "https://initech.com", "type": "Customer", "industry": "Finance",
+        "revenue": 40000000, "employees": 300,
+        "description": "Mid-market financial software firm.",
+        "ownerId": "user-3",
+        "billingStreet": "4120 Freidrich Ln", "billingCity": "Austin",
+        "billingState": "TX", "billingZip": "78744", "billingCountry": "United States",
+        "shippingStreet": "4120 Freidrich Ln", "shippingCity": "Austin",
+        "shippingState": "TX", "shippingZip": "78744", "shippingCountry": "United States",
+        "createdDate": "2025-08-20T09:00:00.000Z", "modifiedDate": "2026-05-15T09:00:00.000Z",
+    },
+]
+
+sf_contacts = [
+    {
+        "contactId": "contact-1", "accountId": "account-1", "firstName": "Gavin",
+        "lastName": "Belson", "title": "CEO", "department": "Executive",
+        "email": "gavin.belson@hooli.com", "phone": "(650) 555-0143", "ownerId": "user-2",
+    },
+    {
+        "contactId": "contact-2", "accountId": "account-2", "firstName": "Bill",
+        "lastName": "Lumbergh", "title": "VP Operations", "department": "Operations",
+        "email": "bill.lumbergh@initech.com", "phone": "(512) 555-0188", "ownerId": "user-3",
+    },
+]
+
+sf_opportunities = [
+    {
+        "opportunityId": "opp-1", "name": "Hooli Data Migration",
+        "accountId": "account-1", "contactId": "contact-1",
+        "amount": 72000,
+        "closeDate": "2026-09-30T00:00:00.000Z",
+        "stage": "Negotiation", "probability": 75, "type": "New Business",
+        "leadSource": "Partner Referral",
+        "nextStep": "Finalize pricing and obtain discount approval",
+        "description": (
+            "Enterprise data migration engagement for Hooli. Standard list price is "
+            "$90,000. Customer has requested a proposed (discounted) price of $72,000."
+        ),
+        "ownerId": "user-2",
+        "createdDate": "2026-05-05T09:00:00.000Z", "modifiedDate": "2026-06-28T09:00:00.000Z",
+    },
+    {
+        "opportunityId": "opp-2", "name": "Initech Cloud Upgrade",
+        "accountId": "account-2", "contactId": "contact-2",
+        "amount": 35000,
+        "closeDate": "2026-08-15T00:00:00.000Z",
+        "stage": "Qualification", "probability": 40, "type": "Existing Business",
+        "leadSource": "Website",
+        "nextStep": "Schedule technical scoping call",
+        "description": "Cloud infrastructure upgrade for Initech.",
+        "ownerId": "user-3",
+        "createdDate": "2026-06-01T09:00:00.000Z", "modifiedDate": "2026-06-20T09:00:00.000Z",
+    },
+]
+
+sf_state = {
+    "user": sf_user,
+    "users": sf_users,
+    "leads": [],
+    "accounts": sf_accounts,
+    "contacts": sf_contacts,
+    "opportunities": sf_opportunities,
+    "cases": [],
+    "activities": [],
+    "chatterPosts": [],
+    "files": [],
+    "following": [],
+    "recentlyViewed": [],
+    "dismissedNotifications": [],
+    "dashboards": [],
+    "emailDrafts": [],
+    "reportSnapshots": [],
+}
+
+# =====================================================================
+# Slack state
+# =====================================================================
+slack_current_user = {
+    "userId": "user_1", "fullName": "John Smith", "displayName": "John",
+    "email": "john.smith@company.com",
+    "avatar": "https://picsum.photos/200/200?random=1",
+    "title": "Sales Manager", "status": "online",
+    "statusMessage": "", "statusEmoji": "", "timeZone": "America/New_York",
+}
+slack_users = [
+    slack_current_user,
+    {"userId": "user_2", "fullName": "Sarah Johnson", "displayName": "Sarah",
+     "email": "sarah.johnson@company.com", "avatar": "https://picsum.photos/200/200?random=2",
+     "title": "Finance Manager", "status": "online", "statusMessage": "", "statusEmoji": "",
+     "timeZone": "America/New_York"},
+    {"userId": "user_3", "fullName": "Mike Chen", "displayName": "Mike",
+     "email": "mike.chen@company.com", "avatar": "https://picsum.photos/200/200?random=3",
+     "title": "Controller", "status": "away", "statusMessage": "", "statusEmoji": "",
+     "timeZone": "America/Los_Angeles"},
+]
+
+slack_channels = [
+    {"channelId": "general", "name": "general",
+     "description": "Company-wide announcements and general discussion",
+     "topic": "Company-wide announcements", "isPrivate": False, "isStarred": True,
+     "members": ["user_1", "user_2", "user_3"], "createdBy": "user_1",
+     "createdAt": "2024-01-01T10:00:00Z", "pinnedMessages": [], "unreadCount": 0},
+    {"channelId": "finance-approvals", "name": "finance-approvals",
+     "description": "Discount and pricing approval requests for finance sign-off",
+     "topic": "Pricing / discount approvals", "isPrivate": False, "isStarred": False,
+     "members": ["user_1", "user_2", "user_3"], "createdBy": "user_2",
+     "createdAt": "2025-11-03T08:00:00Z", "pinnedMessages": [], "unreadCount": 0},
+]
+
+slack_messages = {
+    "general": [
+        {"messageId": "msg_g1", "senderId": "user_2",
+         "content": "Reminder: Q3 pricing review is this Friday.",
+         "timestamp": "2026-06-29T15:00:00Z", "threadId": None,
+         "reactions": [], "attachments": [], "isEdited": False},
+    ],
+    "finance-approvals": [],
+}
+
+slack_state = {
+    "currentUser": slack_current_user,
+    "workspace": {"workspaceId": "ws_1", "workspaceName": "Acme Corp",
+                  "icon": "https://picsum.photos/64/64?random=workspace"},
+    "users": slack_users,
+    "channels": slack_channels,
+    "messages": slack_messages,
+    "threads": {},
+    "dms": [],
+    "bookmarkedMessages": [],
+    "callHistory": [],
+    "settings": {"theme": "light", "notifications": "all",
+                 "displayDensity": "comfortable", "showAvatars": True, "use24Hour": False},
+    "invitations": [],
+    "notifications": [],
+}
+
+# =====================================================================
+# Inject state into both mocks (action: set -> writes initial + current)
+# =====================================================================
+for name, url, state in [
+    ("salesforce", SF_URL, sf_state),
+    ("slack", SLACK_URL, slack_state),
+]:
+    resp = requests.post(f'{url}/post?sid={sid}',
+                         json={'action': 'set', 'state': state}, timeout=30)
+    assert resp.status_code == 200, f'{name} state injection failed: {resp.text}'
+    go = requests.get(f'{url}/go?sid={sid}', timeout=10).json()
+    assert go['initial_state'] is not None, f'{name} initial_state is None after injection'
+    print(f'State injected: {name} sid={sid}')
+
+# =====================================================================
+# Launch browser with both mocks (Salesforce first, then Slack tab)
+# =====================================================================
+launch_gui(
+    f'google-chrome "{SF_URL}/?sid={sid}" "{SLACK_URL}/?sid={sid}"',
+    delay_sec=3.0,
+)
+print(f'GUI_READY: launched browser with salesforce + slack (sid={sid})')

ar_approval_002/reward.py

@@ -0,0 +1,213 @@
+"""
+Reward Script: Post a structured discount-approval request in Slack #finance-approvals
+Task ID: ar_approval_002
+Domain: mock_websites (Salesforce -> Slack chaining; observable state in Slack)
+
+Task outcome (from task_config.json):
+  A single structured message in Slack #finance-approvals containing ALL of:
+    - Customer = Hooli
+    - Original (list) Amount = $90,000
+    - Adjusted (proposed) Amount = $72,000
+    - Discount = 20%
+    - Requester = Jian Wu
+    - A note that it exceeds the 15% threshold and needs approval / sign-off
+  Salesforce must remain read-only / unchanged.
+
+Scoring (all components anchored to the NEW message in #finance-approvals, which is
+absent in initial_env -> every component FAILS on initial_env -> reward(initial)=0.0):
+  C1 new message present in #finance-approvals .......... 0.15
+  C2 Customer = Hooli ................................... 0.15
+  C3 Original amount $90,000 ............................ 0.15
+  C4 Adjusted amount $72,000 ........................... 0.15
+  C5 Discount 20% ...................................... 0.15
+  C6 Requester Jian Wu ................................. 0.15
+  C7 Note: exceeds 15% threshold / needs approval ...... 0.10
+  Total = 1.0
+
+Salesforce-unchanged is a PRECONDITION (true in both initial and golden), so it is
+used only as a non-scoring integrity gate: if the Hooli opportunity's figures were
+mutated, the score is zeroed (anti-cheat). It never affects golden(1.0)/initial(0.0).
+
+All checks are deterministic. No LLM judge is used (OPENAI_API_KEY is EMPTY on the VM).
+"""
+
+import os
+import re
+import sys
+
+# The VM can only reach the mocks via the corporate proxy. Set it explicitly so
+# `requests` uses it regardless of how the script's environment was spawned.
+PROXY = 'http://star-proxy.oa.com:3128'
+os.environ['http_proxy'] = PROXY
+os.environ['https_proxy'] = PROXY
+os.environ['HTTP_PROXY'] = PROXY
+os.environ['HTTPS_PROXY'] = PROXY
+os.environ['no_proxy'] = 'localhost,127.0.0.1'
+
+import requests
+
+SLACK_URL = 'http://28.7.184.198:8178'
+SALESFORCE_URL = 'http://28.7.184.198:8175'
+APPROVAL_CHANNEL = 'finance-approvals'
+
+
+def fail(msg):
+    print(f'CRITICAL: {msg}')
+    print('REWARD: 0.0')
+    sys.exit(0)
+
+
+# --- Read sid ---
+try:
+    with open('/tmp/task_web_sid') as f:
+        sid = f.read().strip()
+    if not sid:
+        raise ValueError('sid is empty')
+except Exception as e:
+    fail(f'Cannot read sid from /tmp/task_web_sid: {e}')
+
+# --- Fetch Slack state ---
+try:
+    slack = requests.get(f'{SLACK_URL}/go?sid={sid}', timeout=20).json()
+except Exception as e:
+    fail(f'Cannot fetch Slack state: {e}')
+
+slack_initial = slack.get('initial_state') or {}
+slack_current = slack.get('current_state') or {}
+if not slack_current:
+    fail('Slack current_state is empty/None')
+
+init_msgs = (slack_initial.get('messages') or {}).get(APPROVAL_CHANNEL, []) or []
+cur_msgs = (slack_current.get('messages') or {}).get(APPROVAL_CHANNEL, []) or []
+
+# Anchor scoring to messages that are NEW relative to the initial snapshot.
+new_msgs = cur_msgs[len(init_msgs):] if len(cur_msgs) > len(init_msgs) else []
+
+# Combined raw + normalized text of the new approval message(s).
+raw_text = '\n'.join(str(m.get('content', '')) for m in new_msgs)
+text_lower = raw_text.lower()
+# Normalized: drop commas/spaces so "$90,000" and "$90000" both match "90000".
+norm = re.sub(r'[,\s]', '', text_lower)
+
+
+def verify_task():
+    total_score = 0.0
+
+    # --- Integrity gate: Salesforce must be unchanged (read-only). ---
+    # Precondition (true in both initial and golden) -> awards NO points.
+    # Only used to zero-out the score if the agent mutated the Hooli opportunity.
+    try:
+        sf = requests.get(f'{SALESFORCE_URL}/go?sid={sid}', timeout=20).json()
+        sf_init = sf.get('initial_state') or {}
+        sf_cur = sf.get('current_state') or {}
+
+        def hooli_opp(state):
+            for o in (state.get('opportunities') or []):
+                if str(o.get('name', '')).strip().lower() == 'hooli data migration':
+                    return o
+            return None
+
+        opp_i = hooli_opp(sf_init)
+        opp_c = hooli_opp(sf_cur)
+        if opp_i is not None and opp_c is not None:
+            # Compare the figures that the task says must stay read-only.
+            if (opp_i.get('amount') != opp_c.get('amount')
+                    or opp_i.get('stage') != opp_c.get('stage')):
+                print('GATE FAIL: Salesforce Hooli opportunity was modified '
+                      f'(initial amount={opp_i.get("amount")}/stage={opp_i.get("stage")} '
+                      f'-> current amount={opp_c.get("amount")}/stage={opp_c.get("stage")}). '
+                      'Salesforce must be read-only.')
+                print('\nScore: 0.0/1.0')
+                print('REWARD: 0.0')
+                return 0.0
+            print('GATE PASS: Salesforce Hooli opportunity unchanged (read-only).')
+        else:
+            print('GATE INFO: Hooli opportunity not found in SF state; skipping integrity gate.')
+    except Exception as e:
+        # Do not crash scoring if SF is unreachable; just note it.
+        print(f'GATE WARN: could not verify Salesforce integrity: {e}')
+
+    # --- Component 1: a NEW message exists in #finance-approvals (0.15) ---
+    try:
+        if new_msgs:
+            print(f'PASS: C1 — {len(new_msgs)} new message(s) in #{APPROVAL_CHANNEL} (0.15)')
+            total_score += 0.15
+        else:
+            print(f'FAIL: C1 — no new message in #{APPROVAL_CHANNEL} '
+                  f'(initial={len(init_msgs)}, current={len(cur_msgs)})')
+    except Exception as e:
+        print(f'ERROR: C1 — {e}')
+
+    # --- Component 2: Customer = Hooli (0.15) ---
+    try:
+        if 'hooli' in text_lower:
+            print('PASS: C2 — Customer "Hooli" present (0.15)')
+            total_score += 0.15
+        else:
+            print('FAIL: C2 — Customer "Hooli" not found')
+    except Exception as e:
+        print(f'ERROR: C2 — {e}')
+
+    # --- Component 3: Original/list amount $90,000 (0.15) ---
+    try:
+        if '90000' in norm:
+            print('PASS: C3 — Original amount $90,000 present (0.15)')
+            total_score += 0.15
+        else:
+            print('FAIL: C3 — Original amount 90,000 not found')
+    except Exception as e:
+        print(f'ERROR: C3 — {e}')
+
+    # --- Component 4: Adjusted/proposed amount $72,000 (0.15) ---
+    try:
+        if '72000' in norm:
+            print('PASS: C4 — Adjusted amount $72,000 present (0.15)')
+            total_score += 0.15
+        else:
+            print('FAIL: C4 — Adjusted amount 72,000 not found')
+    except Exception as e:
+        print(f'ERROR: C4 — {e}')
+
+    # --- Component 5: Discount 20% (0.15) ---
+    try:
+        # Require the 20% figure (allow "20 %" / "20percent" via normalization).
+        if '20%' in norm or '20percent' in norm:
+            print('PASS: C5 — Discount 20% present (0.15)')
+            total_score += 0.15
+        else:
+            print('FAIL: C5 — Discount 20% not found')
+    except Exception as e:
+        print(f'ERROR: C5 — {e}')
+
+    # --- Component 6: Requester Jian Wu (0.15) ---
+    try:
+        # Tolerate extra spacing; normalized form removes the space -> "jianwu".
+        if 'jianwu' in norm:
+            print('PASS: C6 — Requester "Jian Wu" present (0.15)')
+            total_score += 0.15
+        else:
+            print('FAIL: C6 — Requester "Jian Wu" not found')
+    except Exception as e:
+        print(f'ERROR: C6 — {e}')
+
+    # --- Component 7: note that it exceeds the 15% threshold / needs approval (0.10) ---
+    try:
+        mentions_15 = '15%' in norm or '15percent' in norm
+        mentions_approval = any(k in text_lower for k in
+                                ('approval', 'sign-off', 'sign off', 'signoff', 'threshold', 'exceed'))
+        if mentions_15 and mentions_approval:
+            print('PASS: C7 — Note references 15% threshold and approval/sign-off (0.10)')
+            total_score += 0.10
+        else:
+            print(f'FAIL: C7 — threshold note incomplete (mentions_15={mentions_15}, '
+                  f'mentions_approval={mentions_approval})')
+    except Exception as e:
+        print(f'ERROR: C7 — {e}')
+
+    final_score = round(min(total_score, 1.0), 4)
+    print(f'\nScore: {final_score}/1.0')
+    print(f'REWARD: {final_score}')
+    return final_score
+
+
+verify_task()

ar_approval_002/reward_label.json

@@ -0,0 +1,82 @@
+{
+  "reward_file": "/apdcephfs_sh2/share_300000800/user/jackwkwang/long-horizon-tasks/mini_osworld_xiangwu/cache/ar_approval_001/reward.py",
+  "api_url": "http://28.7.184.184:8000/v1/chat/completions",
+  "model": "Kimi-K2.6",
+  "max_code_chars": 100000,
+  "created_at": "2026-07-14 10:22:12",
+  "label": {
+    "task_id": "ar_approval_002",
+    "domain": "mock_websites",
+    "summary": "验证代理是否在 Slack #finance-approvals 频道发布了包含特定客户、金额、折扣、请求人及审批说明的结构化折扣审批请求消息，并确保 Salesforce 只读未被修改。",
+    "is_placeholder": false,
+    "data_sources": [
+      "/tmp/task_web_sid",
+      "slack_mock (http://28.7.186.212:8198)",
+      "salesforce_mock (http://28.7.186.212:8195)"
+    ],
+    "scoring_components": [
+      {
+        "name": "C1 new message present in #finance-approvals",
+        "weight": 0.15,
+        "description": "检查在 #finance-approvals 频道是否有相对于初始状态的新消息",
+        "check_logic": "比较 slack_current 与 slack_initial 中 finance-approvals 频道的消息列表，若 cur_msgs 长度大于 init_msgs，则取 new_msgs = cur_msgs[len(init_msgs):]，判断 new_msgs 是否非空",
+        "pass_condition": "new_msgs 非空（即至少有一条新消息）"
+      },
+      {
+        "name": "C2 Customer = Hooli",
+        "weight": 0.15,
+        "description": "检查新消息文本中是否包含客户名称 Hooli",
+        "check_logic": "将新消息的 raw_text 转为小写得到 text_lower，检查 'hooli' 是否在 text_lower 中",
+        "pass_condition": "text_lower 中包含子串 'hooli'"
+      },
+      {
+        "name": "C3 Original amount $90,000",
+        "weight": 0.15,
+        "description": "检查新消息中是否包含原始金额 90,000",
+        "check_logic": "对 text_lower 使用 re.sub(r'[,\\s]', '', ...) 得到 norm（移除逗号和空格），检查 '90000' 是否在 norm 中",
+        "pass_condition": "norm 中包含子串 '90000'"
+      },
+      {
+        "name": "C4 Adjusted amount $72,000",
+        "weight": 0.15,
+        "description": "检查新消息中是否包含调整后金额 72,000",
+        "check_logic": "检查 norm 中是否包含 '72000'",
+        "pass_condition": "norm 中包含子串 '72000'"
+      },
+      {
+        "name": "C5 Discount 20%",
+        "weight": 0.15,
+        "description": "检查新消息中是否包含 20% 折扣信息",
+        "check_logic": "检查 norm 中是否包含 '20%' 或 '20percent'",
+        "pass_condition": "norm 中包含 '20%' 或 '20percent'"
+      },
+      {
+        "name": "C6 Requester Jian Wu",
+        "weight": 0.15,
+        "description": "检查新消息中是否包含请求人 Jian Wu",
+        "check_logic": "检查 norm（已移除空格）中是否包含 'jianwu'",
+        "pass_condition": "norm 中包含子串 'jianwu'"
+      },
+      {
+        "name": "C7 Note: exceeds 15% threshold / needs approval",
+        "weight": 0.1,
+        "description": "检查新消息中是否包含超过 15% 阈值并需要审批的说明",
+        "check_logic": "mentions_15 = '15%' in norm or '15percent' in norm；mentions_approval = any(k in text_lower for k in ('approval', 'sign-off', 'sign off', 'signoff', 'threshold', 'exceed'))；要求两者同时为真",
+        "pass_condition": "同时满足：norm 中包含 '15%' 或 '15percent'，且 text_lower 中包含 'approval'、'sign-off'、'sign off'、'signoff'、'threshold' 或 'exceed' 之一"
+      }
+    ],
+    "total_max_score": 1.0,
+    "score_aggregation": "各组件分数相加，最后通过 round(min(total_score, 1.0), 4) 汇总并钳制到上限 1.0",
+    "failure_modes": [
+      "读取 /tmp/task_web_sid 失败或为空：调用 fail() 打印 CRITICAL 并 sys.exit(0) 返回 0.0",
+      "无法获取 Slack 状态：调用 fail() 退出并返回 0.0",
+      "Slack current_state 为空/None：调用 fail() 退出并返回 0.0",
+      "Salesforce 中 Hooli Data Migration opportunity 的 amount 或 stage 被修改：直接打印 GATE FAIL 并返回 0.0（反作弊完整性门控）",
+      "无新消息（new_msgs 为空）：C1 失败，不加分",
+      "各组件文本匹配失败：对应组件不加分，但脚本继续执行"
+    ],
+    "scoring_logic_explanation": "脚本首先读取 /tmp/task_web_sid 获取会话 ID，然后分别拉取 Slack 和 Salesforce 的 mock 状态。对于 Slack，通过比较 initial_state 和 current_state 中 #finance-approvals 频道的消息列表，提取初始状态之后的新消息（new_msgs），并将所有新消息内容拼接为 raw_text，同时生成小写和规范化（移除逗号、空格）的文本。接着执行完整性门控：检查 Salesforce 中名为 'Hooli Data Migration' 的 opportunity 在初始状态和当前状态的 amount 与 stage 是否一致，若被修改则直接返回 0.0。随后逐项检查 7 个评分组件：C1 验证新消息存在性；C2-C6 通过在 text_lower 或 norm 中搜索特定���键词验证客户、原始金额、调整后金额、折扣、请求人；C7 要求同时包含 15% 相关字样和审批/阈值/超出相关字样。每个通过组件累加对应分数，最终分数经 min(..., 1.0) 钳制并四舍五入到 4 位小数输出。",
+    "confidence": 1.0
+  },
+  "raw_response": "用户要求我分析一个 reward.py 评分脚本的评分逻辑，并输出一个严格的 JSON 对象。\n\n让我仔细阅读代码：\n\n1. **任务信息提取**：\n   - Task ID: ar_approval_001（从 docstring 中明确提取）\n   - Domain: mock_websites（从 docstring 中明确提取）\n\n2. **任务概述**：\n   - 在 Slack #finance-approvals 频道发布一个结构化的折扣审批请求消息，包含特定字段。Salesforce 必须保持只读/不变。\n\n3. **数据源**：\n   - `/tmp/task_web_sid`：读取 session ID\n   - Slack mock：`http://28.7.186.212:8198`（通过 `/go?sid={sid}` 获取状态）\n   - Salesforce mock：`http://28.7.186.212:8195`（通过 `/go?sid={sid}` 获取状态）\n\n4. **评分组件**：\n   - C1 (0.15)：#finance-approvals 中有新消息\n   - C2 (0.15)：包含 \"Hooli\"\n   - C3 (0.15)：包含原始金额 $90,000（规范化后匹配 \"90000\"）\n   - C4 (0.15)：包含调整后金额 $72,000（规范化后匹配 \"72000\"）\n   - C5 (0.15)：包含折扣 20%（规范化后匹配 \"20%\" 或 \"20percent\"）\n   - C6 (0.15)：包含请求人 Jian Wu（规范化后匹配 \"jianwu\"）\n   - C7 (0.10)：包含超过 15% 阈值/需要审批的说明（需要同时满足：包含 \"15%\" 或 \"15percent\"，以及包含 \"approval\"、\"sign-off\"、\"sign off\"、\"signoff\"、\"threshold\"、\"exceed\" 之一）\n\n5. **完整性门控（Integrity gate）**：\n   - 检查 Salesforce 中 Hooli Data Migration opportunity 的 amount 和 stage 是否被修改\n   - 如果被修改，直接返回 0.0\n   - 不贡献分数，只是反作弊\n\n6. **分数聚合**：\n   - 各组件分数相加\n   - 最后 `round(min(total_score, 1.0), 4)`\n\n7. **失败模式**：\n   - 读取 sid 失败：调用 fail()，输出 REWARD: 0.0 并 sys.exit(0)\n   - 获取 Slack 状态失败：同上\n   - Slack current_state 为空：同上\n   - Salesforce Hooli opportunity 被修改：返回 0.0\n   - 各组件检查失败：只是不加分，不会提前退出（除了 integrity gate）\n\n8. **is_placeholder**：\n   - false，因为这是一个真正的评分脚本，有详细的检查逻辑\n\n9. **confidence**：\n   - 1.0，因为代码非常清晰完整\n\n让我构建 JSON：\n\n```json\n{\n  \"task_id\": \"ar_approval_001\",\n  \"domain\": \"mock_websites\",\n  \"summary\": \"验证代理是否在 Slack #finance-approvals 频道发布了包含特定客户、金额、折扣、请求人及审批说明的结构化折扣审批请求消息，并确保 Salesforce 只读未被修改。\",\n  \"is_placeholder\": false,\n  \"data_sources\": [\"/tmp/task_web_sid\", \"slack_mock (http://28.7.186.212:8198)\", \"salesforce_mock (http://28.7.186.212:8195)\"],\n  \"scoring_components\": [\n    {\n      \"name\": \"C1 new message present in #finance-approvals\",\n      \"weight\": 0.15,\n      \"description\": \"检查在 #finance-approvals 频道是否有相对于初始状态的新消息\",\n      \"check_logic\": \"比较 slack_current 与 slack_initial 中 finance-approvals 频道的消息列表，若 cur_msgs 长度大于 init_msgs，则取 new_msgs = cur_msgs[len(init_msgs):]，判断 new_msgs 是否非空\",\n      \"pass_condition\": \"new_msgs 非空（即至少有一条新消息）\"\n    },\n    {\n      \"name\": \"C2 Customer = Hooli\",\n      \"weight\": 0.15,\n      \"description\": \"检查新消息文本中是否包含客户名称 Hooli\",\n      \"check_logic\": \"将新消息的 raw_text 转为小写得到 text_lower，检查 'hooli' 是否在 text_lower 中\",\n      \"pass_condition\": \"text_lower 中包含子串 'hooli'\"\n    },\n    {\n      \"name\": \"C3 Original amount $90,000\",\n      \"weight\": 0.15,\n      \"description\": \"检查新消息中是否包含原始金额 90,000\",\n      \"check_logic\": \"对 text_lower 使用 re.sub(r'[,\\\\s]', '', ...) 得到 norm（移除逗号和空格），检查 '90000' 是否在 norm 中\",\n      \"pass_condition\": \"norm 中包含子串 '90000'\"\n    },\n    {\n      \"name\": \"C4 Adjusted amount $72,000\",\n      \"weight\": 0.15,\n      \"description\": \"检查新消息中是否包含调整后金额 72,000\",\n      \"check_logic\": \"检查 norm 中是否包含 '72000'\",\n      \"pass_condition\": \"norm 中包含子串 '72000'\"\n    },\n    {\n      \"name\": \"C5 Discount 20%\",\n      \"weight\": 0.15,\n      \"description\": \"检查新消息中是否包含 20% 折扣信息\",\n      \"check_logic\": \"检查 norm 中是否包含 '20%' 或 '20percent'\",\n      \"pass_condition\": \"norm 中包含 '20%' 或 '20percent'\"\n    },\n    {\n      \"name\": \"C6 Requester Jian Wu\",\n      \"weight\": 0.15,\n      \"description\": \"检查新消息中是否包含请求人 Jian Wu\",\n      \"check_logic\": \"检查 norm（已移除空格）中是否包含 'jianwu'\",\n      \"pass_condition\": \"norm 中包含子串 'jianwu'\"\n    },\n    {\n      \"name\": \"C7 Note: exceeds 15% threshold / needs approval\",\n      \"weight\": 0.10,\n      \"description\": \"检查新消息中是否包含超过 15% 阈值并需要审批的说明\",\n      \"check_logic\": \"mentions_15 = '15%' in norm or '15percent' in norm；mentions_approval = any(k in text_lower for k in ('approval', 'sign-off', 'sign off', 'signoff', 'threshold', 'exceed'))；要求两者同时为真\",\n      \"pass_condition\": \"同时满足：norm 中包含 '15%' 或 '15percent'，且 text_lower 中包含 'approval'、'sign-off'、'sign off'、'signoff'、'threshold' 或 'exceed' 之一\"\n    }\n  ],\n  \"total_max_score\": 1.0,\n  \"score_aggregation\": \"各组件分数相加，最后通过 round(min(total_score, 1.0), 4) 汇总并钳制到上限 1.0\",\n  \"failure_modes\": [\n    \"读取 /tmp/task_web_sid 失败或为空：调用 fail() 打印 CRITICAL 并 sys.exit(0) 返回 0.0\",\n    \"无法获取 Slack 状态：调用 fail() 退出并返回 0.0\",\n    \"Slack current_state 为空/None：调用 fail() 退出并返回 0.0\",\n    \"Salesforce 中 Hooli Data Migration opportunity 的 amount 或 stage 被修改：直接打印 GATE FAIL 并返回 0.0（反作弊完整性门控）\",\n    \"无新消息（new_msgs 为空）：C1 失败，不加分\",\n    \"各组件文本匹配失败：对应组件不加分，但脚本继续执行\"\n  ],\n  \"scoring_logic_explanation\": \"脚本首先读取 /tmp/task_web_sid 获取会话 ID，然后分别拉取 Slack 和 Salesforce 的 mock 状态。对于 Slack，通过比较 initial_state 和 current_state 中 #finance-approvals 频道的消息列表，提取初始状态之后的新消息（new_msgs），并将所有新消息内容拼接为 raw_text，同时生成小写和规范化（移除逗号、空格）的文本。接着执行完整性门控：检查 Salesforce 中名为 'Hooli Data Migration' 的 opportunity 在初始状态和当前状态的 amount 与 stage 是否一致，若被修改则直接返回 0.0。随后逐项检查 7 个评分组件：C1 验证新消息存在性；C2-C6 通过在 text_lower 或 norm 中搜索特定关键词验证客户、原始金额、调整后金额、折扣、请求人；C7 要求同时包含 15% 相关字样和审批/阈值/超出相关字样。每个通过组件累加对应分数，最终分数经 min(..., 1.0) 钳制并四舍五入到 4 位小数输出。\",\n  \"confidence\": 1.0\n}\n```\n\n检查要求：\n- 只输出 JSON，不要 Markdown 代码块\n- 第一个字符是 {，最后一个字符是 }\n- 字符串 value 使用中文，task_id、domain、字段名、代码标识符保留原样\n- is_placeholder: false\n- weight 填写代码中显式给出的分值\n\n确认所有字段都已包含且格式正确。 </think> {\"task_id\":\"ar_approval_001\",\"domain\":\"mock_websites\",\"summary\":\"验证代理是否在 Slack #finance-approvals 频道发布了包含特定客户、金额、折扣、请求人及审批说明的结构化折扣审批请求消息，并确保 Salesforce 只读未被修改。\",\"is_placeholder\":false,\"data_sources\":[\"/tmp/task_web_sid\",\"slack_mock (http://28.7.186.212:8198)\",\"salesforce_mock (http://28.7.186.212:8195)\"],\"scoring_components\":[{\"name\":\"C1 new message present in #finance-approvals\",\"weight\":0.15,\"description\":\"检查在 #finance-approvals 频道是否有相对于初始状态的新消息\",\"check_logic\":\"比较 slack_current 与 slack_initial 中 finance-approvals 频道的消息列表，若 cur_msgs 长度大于 init_msgs，则取 new_msgs = cur_msgs[len(init_msgs):]，判断 new_msgs 是否非空\",\"pass_condition\":\"new_msgs 非空（即至少有一条新消息）\"},{\"name\":\"C2 Customer = Hooli\",\"weight\":0.15,\"description\":\"检查新消息文本中是否包含客户名称 Hooli\",\"check_logic\":\"将新消息的 raw_text 转为小写得到 text_lower，检查 'hooli' 是否在 text_lower 中\",\"pass_condition\":\"text_lower 中包含子串 'hooli'\"},{\"name\":\"C3 Original amount $90,000\",\"weight\":0.15,\"description\":\"检查新消息中是否包含原始金额 90,000\",\"check_logic\":\"对 text_lower 使用 re.sub(r'[,\\\\s]', '', ...) 得到 norm（移除逗号和空格），检查 '90000' 是否在 norm 中\",\"pass_condition\":\"norm 中包含子串 '90000'\"},{\"name\":\"C4 Adjusted amount $72,000\",\"weight\":0.15,\"description\":\"检查新消息中是否包含调整后金额 72,000\",\"check_logic\":\"检查 norm 中是否包含 '72000'\",\"pass_condition\":\"norm 中包含子串 '72000'\"},{\"name\":\"C5 Discount 20%\",\"weight\":0.15,\"description\":\"检查新消息中是否包含 20% 折扣信息\",\"check_logic\":\"检查 norm 中是否包含 '20%' 或 '20percent'\",\"pass_condition\":\"norm 中包含 '20%' 或 '20percent'\"},{\"name\":\"C6 Requester Jian Wu\",\"weight\":0.15,\"description\":\"检查新消息中是否包含请求人 Jian Wu\",\"check_logic\":\"检查 norm（已移��空格）中是否包含 'jianwu'\",\"pass_condition\":\"norm 中包含子串 'jianwu'\"},{\"name\":\"C7 Note: exceeds 15% threshold / needs approval\",\"weight\":0.10,\"description\":\"检查新消息中是否包含超过 15% 阈值并需要审批的说明\",\"check_logic\":\"mentions_15 = '15%' in norm or '15percent' in norm；mentions_approval = any(k in text_lower for k in ('approval', 'sign-off', 'sign off', 'signoff', 'threshold', 'exceed'))；要求两者同时为真\",\"pass_condition\":\"同时满足：norm 中包含 '15%' 或 '15percent'，且 text_lower 中包含 'approval'、'sign-off'、'sign off'、'signoff'、'threshold' 或 'exceed' 之一\"}],\"total_max_score\":1.0,\"score_aggregation\":\"各组件分数相加，最后通过 round(min(total_score, 1.0), 4) 汇总并钳制到上限 1.0\",\"failure_modes\":[\"读取 /tmp/task_web_sid 失败或为空：调用 fail() 打印 CRITICAL 并 sys.exit(0) 返回 0.0\",\"无法获取 Slack 状态：调用 fail() 退出并返回 0.0\",\"Slack current_state 为空/None：调用 fail() 退出并返回 0.0\",\"Salesforce 中 Hooli Data Migration opportunity 的 amount 或 stage 被修改：直接打印 GATE FAIL 并返回 0.0（反作弊完整性门控）\",\"无新消息（new_msgs 为空）：C1 失败，不加分\",\"各组件文本匹配失败：对应组件不加分，但脚本继续执行\"],\"scoring_logic_explanation\":\"脚本首先读取 /tmp/task_web_sid 获取会话 ID，然后分别拉取 Slack 和 Salesforce 的 mock 状态。对于 Slack，通过比较 initial_state 和 current_state 中 #finance-approvals 频道的消息列表，提取初始状态之后的新消息（new_msgs），并将所有新消息内容拼接为 raw_text，同时生成小写和规范化（移除逗号、空格）的文本。接着执行完整性门控：检查 Salesforce 中名为 'Hooli Data Migration' 的 opportunity 在初始状态和当前状态的 amount 与 stage 是否一致，若被修改则直接返回 0.0。随后逐项检查 7 个评分组件：C1 验证新消息存在性；C2-C6 通过在 text_lower 或 norm 中搜索特定关键词验证客户、原始金额、调整后金额、折扣、请求人；C7 要求同时包含 15% 相关字样和审批/阈值/超出相关字样。每个通过组件累加对应分数，最终分数经 min(..., 1.0) 钳制并四舍五入到 4 位小数输出。\",\"confidence\":1.0}"
+}

ar_billing_exception_012__long/_cua_gym_vm_bridge.sh

@@ -0,0 +1,106 @@
+#!/bin/bash
+# CUA-Gym VM-side network fix (direct-IP mock access).
+#
+# Runs INSIDE the QEMU guest. The CUA-Gym mock services are now reached by
+# their real IDC address (http://28.7.184.198:81xx) instead of the old
+# localhost:9000-9097 socat bridge, so the VM only needs two fixes:
+#
+#   1. Working DNS — the qcow2-baked /etc/resolv.conf points at the Pod's
+#      cluster DNS, which is unreachable from QEMU SLIRP NAT
+#      ("Temporary failure in name resolution").
+#   2. The mock host excluded from the woa http proxy — otherwise the VM's
+#      Chrome (gsettings proxy) and initial_setup.py's `requests` get a 403
+#      from the proxy when hitting 28.7.184.198:81xx. (The old socat path never
+#      hit this because localhost is bypassed by default.)
+#
+# Idempotent. Scoped to the CUA-Gym bridge step; does not touch the generic
+# OSWorld setup pipeline.
+#
+# Argv:
+#   $1  optional sudo password (default "password"; SetupController
+#       substitutes {CLIENT_PASSWORD} when uploaded as a cache file).
+#   $2  unused (kept for backward compat — was socat first port 9000).
+#   $3  unused (kept for backward compat — was socat last  port 9097).
+#   $4  optional mock host IP (default 28.7.184.198).
+
+set +e
+
+PASS="${1:-password}"
+MOCK_HOST="${4:-28.7.184.198}"
+# Derive the /16 IDC subnet so future port/host tweaks on the same segment
+# stay bypassed too (28.7.184.198 -> 28.7.0.0/16).
+MOCK_NET="$(echo "${MOCK_HOST}" | awk -F. 'NF==4{print $1"."$2".0.0/16"}')"
+
+run_root() {
+    echo "$PASS" | sudo -S -p "" bash -c "$1"
+}
+
+# Detect SLIRP gateway from VM's routing table; fall back to 10.0.2.2.
+GW="$(ip route 2>/dev/null | awk '/default/ {print $3; exit}')"
+GW="${GW:-10.0.2.2}"
+
+# ─── CUA-Gym network fix (Taiji-specific) ─────────────────────────────
+fix_vm_network() {
+    # 0) Discover VM network state at runtime — gateway, own IP/CIDR, SLIRP
+    #    DNS forwarder. Nothing hardcoded except the mock host above.
+    local GW_NET="${GW%.*}"
+    local SLIRP_DNS="${GW_NET}.3"        # QEMU SLIRP DNS proxy is always .3
+    local OWN_CIDR
+    OWN_CIDR="$(ip -4 -o addr show scope global 2>/dev/null \
+                | awk '{print $4; exit}')"
+    local OWN_IP="${OWN_CIDR%/*}"
+    local OWN_NET=""
+    [ -n "${OWN_IP}" ] && OWN_NET="${OWN_IP%.*}.0/24"
+
+    # 1) DNS — point at SLIRP DNS proxy + public fallbacks. chattr +i keeps
+    #    systemd-resolved/NetworkManager from clobbering it.
+    if run_root "printf 'nameserver ${SLIRP_DNS}\\nnameserver 114.114.114.114\\nnameserver 8.8.8.8\\n' > /etc/resolv.conf
+                 chattr +i /etc/resolv.conf 2>/dev/null || true"; then
+        echo "[cua_gym_vm_bridge] DNS set: ${SLIRP_DNS}, 114.114.114.114, 8.8.8.8 (gw=${GW})"
+    else
+        echo "[cua_gym_vm_bridge] WARN: failed to write /etc/resolv.conf"
+    fi
+
+    # 2) Build the proxy bypass list from detected values + the direct-IP
+    #    mock host (and its /16 subnet). THIS is what makes 28.7.184.198:81xx
+    #    reachable from inside the VM — without it the woa proxy 403s it.
+    local ignore_list="['localhost', '127.0.0.0/8', '${GW}'"
+    [ -n "${OWN_IP}" ]    && ignore_list="${ignore_list}, '${OWN_IP}'"
+    [ -n "${OWN_NET}" ]   && ignore_list="${ignore_list}, '${OWN_NET}'"
+    [ -n "${GW_NET}" ]    && ignore_list="${ignore_list}, '${GW_NET}.0/24'"
+    [ -n "${MOCK_HOST}" ] && ignore_list="${ignore_list}, '${MOCK_HOST}'"
+    [ -n "${MOCK_NET}" ]  && ignore_list="${ignore_list}, '${MOCK_NET}'"
+    ignore_list="${ignore_list}]"
+
+    # 3) GNOME proxy ignore-hosts — Chrome bypasses the woa proxy for these.
+    if gsettings set org.gnome.system.proxy ignore-hosts "${ignore_list}" 2>/tmp/gsettings_err; then
+        echo "[cua_gym_vm_bridge] gsettings ignore-hosts = ${ignore_list}"
+    else
+        echo "[cua_gym_vm_bridge] WARN: gsettings ignore-hosts failed: $(cat /tmp/gsettings_err 2>/dev/null)"
+    fi
+
+    # 4) Shell env — comma-separated form for no_proxy/NO_PROXY (requests/curl).
+    local NP="localhost,127.0.0.1,${GW}"
+    [ -n "${OWN_IP}" ]    && NP="${NP},${OWN_IP}"
+    [ -n "${OWN_NET}" ]   && NP="${NP},${OWN_NET}"
+    [ -n "${GW_NET}" ]    && NP="${NP},${GW_NET}.0/24"
+    [ -n "${MOCK_HOST}" ] && NP="${NP},${MOCK_HOST}"
+    [ -n "${MOCK_NET}" ]  && NP="${NP},${MOCK_NET}"
+
+    for var in no_proxy NO_PROXY; do
+        if ! grep -q "export ${var}=${NP}" /home/user/.bashrc 2>/dev/null; then
+            echo "export ${var}=${NP}" >> /home/user/.bashrc
+        fi
+    done
+    run_root "grep -q '^no_proxy=' /etc/environment 2>/dev/null \
+              && sed -i 's|^no_proxy=.*|no_proxy=\"${NP}\"|' /etc/environment \
+              || echo 'no_proxy=\"${NP}\"' >> /etc/environment
+              grep -q '^NO_PROXY=' /etc/environment 2>/dev/null \
+              && sed -i 's|^NO_PROXY=.*|NO_PROXY=\"${NP}\"|' /etc/environment \
+              || echo 'NO_PROXY=\"${NP}\"' >> /etc/environment" >/dev/null 2>&1
+    echo "[cua_gym_vm_bridge] no_proxy=${NP}"
+}
+
+fix_vm_network
+
+echo "[cua_gym_vm_bridge] direct-IP mode: ${MOCK_HOST} (+${MOCK_NET}) bypassed from woa proxy; no socat forwarding needed"

ar_billing_exception_012__long/initial_setup.py

@@ -0,0 +1,934 @@
+"""
+Initial Setup: Weekly billing-exceptions reconciliation
+Task ID: ar_billing_exception_012__long
+Mocks: stripe_dashboard_mock (8179), quickbooks_mock (8172),
+       google_sheets_mock (8145), slack_mock (8178)
+
+Scenario (multi-branch finance-ops reconciliation, "today" = 2026-06-29)
+------------------------------------------------------------------------
+A finance-ops analyst clears the week's billing exceptions. The Google Sheet
+"Billing Exceptions" (tab `Exceptions`, cols: Customer | Exception Type | Amount
+| Invoice #) lists customers the billing system flagged, one row each. For every
+row the analyst derives the corrective action from the `Exception Type`, carries
+the identifier into Stripe or QuickBooks, performs the (hidden) control there,
+then adds a "Reconciliation Summary" tab to the sheet and posts + pins a one-line
+summary to Slack #finance.
+
+Branch rule (stated in the instruction; VERIFIED against current app state):
+  * Duplicate charge   -> REFUND the matching Stripe payment
+                          (skip if that payment is already refunded).
+  * Churn - downgrade  -> CANCEL the matching Stripe subscription at PERIOD END
+                          (skip if already cancel_at_period_end / canceled).
+  * Payment received   -> RECEIVE PAYMENT on the matching QuickBooks invoice
+                          (skip if the invoice is already Paid).
+  * On hold / Under review -> LEAVE (must NOT act — tempting distractor).
+  * Customer with no matching Stripe/QB record -> SKIP (unresolvable).
+
+Row plan (10 rows) => 6 require action, 4 must be skipped:
+  2 REFUND, 2 CANCEL_EOP, 2 RECEIVE_PAYMENT, 2 LEAVE (On hold / Under review),
+  1 already-refunded 'Duplicate charge' (arms the refund leak guard),
+  1 no-match 'Payment received' (unresolvable distractor).
+
+Cross-app data flow (H3):
+  Customer + Amount read in the Sheets worklist -> locate the matching Stripe
+  payment (customer_name + amount in cents) to refund, and the matching Stripe
+  subscription (customer) to cancel-at-period-end; Invoice # -> matching
+  QuickBooks invoice number to receive payment on.
+
+Ground-truth embedding (Rule 2), PRECOMPUTED in Python from the visible sheet
+rows + the injected Stripe/QB records so visible <=> key can never disagree:
+  * google_sheets.initial_state._task_adapter.expected      -> per-row worklist
+  * google_sheets.initial_state._task_adapter.summary_tab_name = 'Reconciliation Summary'
+  * slack.initial_state._task_adapter.summary_channel          = 'finance'
+
+Observable result kept ABSENT at injection (Rule 3):
+  * the 2 REFUND-target payments: amount_refunded=0, refunded=False; refunds[]
+    holds only a decoy that does NOT reference any target.
+  * the 2 CANCEL_EOP-target subscriptions: cancel_at_period_end=False, status='active'.
+  * the 2 RECEIVE_PAYMENT-target QB invoices: status Sent/Overdue, paidAmount=0,
+    paidDate=None.
+  * the Sheets workbook has NO 'Reconciliation Summary' tab (only 'Exceptions').
+  * Slack #finance starts EMPTY (no messages, no pinned).
+  * the 'already-done' distractor payment IS already refunded (must be left alone).
+
+NOTE: Stripe amounts are in CENTS, QuickBooks amounts are in DOLLARS.
+"""
+import datetime as _dt
+import os
+import shlex
+import subprocess
+import time
+import uuid
+
+import requests
+
+
+sid = str(uuid.uuid4())
+with open('/tmp/task_web_sid', 'w', encoding='utf-8') as f:
+    f.write(sid)
+print(f'Generated sid: {sid}')
+
+
+TASK_ID = 'cc91abcd-23c1-4845-afde-3bfed691bbe4'
+TODAY = _dt.date(2026, 6, 29)
+SUMMARY_TAB_NAME = 'Reconciliation Summary'
+SUMMARY_CHANNEL = 'finance'
+
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+def _ts(date_str):
+    """Unix seconds for date_str at 12:00 UTC (stable displayed date across TZs)."""
+    d = _dt.datetime.strptime(date_str, '%Y-%m-%d').replace(
+        hour=12, minute=0, second=0, microsecond=0, tzinfo=_dt.timezone.utc
+    )
+    return int(d.timestamp())
+
+
+_CREATED = _ts('2026-01-08')
+_NOW_TS = _ts('2026-06-29')
+
+
+# ===========================================================================
+# 1) STRIPE — payments + subscriptions (+ customers, payment methods)
+# ===========================================================================
+# Shared customer registry: name -> (stripe_customer_id, email, card)
+_STRIPE_CUSTS = {
+    'Acme Corporation':        ('cus_acme00000001', 'billing@acmecorp.com',       {'brand': 'visa',       'last4': '4242', 'exp_month': 12, 'exp_year': 2027}),
+    'Globex Industries':       ('cus_globex0000012', 'ap@globexind.com',           {'brand': 'mastercard', 'last4': '5555', 'exp_month':  8, 'exp_year': 2028}),
+    'TechStart Inc.':          ('cus_techstart0001', 'finance@techstart.io',       {'brand': 'visa',       'last4': '1234', 'exp_month':  3, 'exp_year': 2027}),
+    'Riverside Medical Group': ('cus_riverside0001', 'it@riversidemedical.org',    {'brand': 'visa',       'last4': '8888', 'exp_month':  1, 'exp_year': 2028}),
+    'CloudSync Pro':           ('cus_cloudsync0001', 'billing@cloudsyncpro.com',   {'brand': 'visa',       'last4': '6543', 'exp_month': 10, 'exp_year': 2028}),
+    'DataVault Systems':       ('cus_datavault0001', 'procurement@datavault.io',   {'brand': 'amex',       'last4': '0001', 'exp_month':  6, 'exp_year': 2028}),
+    'Pioneer Labs':            ('cus_pioneer000001', 'admin@pioneerlabs.io',       {'brand': 'visa',       'last4': '7777', 'exp_month': 12, 'exp_year': 2027}),
+    'Northwind Traders':       ('cus_northwind0001', 'ap@northwindtraders.com',    {'brand': 'mastercard', 'last4': '2323', 'exp_month':  5, 'exp_year': 2027}),
+    'Summit Retail':           ('cus_summit0000001', 'billing@summitretail.com',   {'brand': 'visa',       'last4': '9090', 'exp_month':  9, 'exp_year': 2028}),
+}
+
+
+def _stripe_customer(name):
+    cid, email, _card = _STRIPE_CUSTS[name]
+    return {
+        'id': cid, 'name': name, 'email': email, 'phone': None, 'description': None,
+        'address': {'line1': '200 Market St', 'line2': None, 'city': 'San Francisco',
+                    'state': 'CA', 'postal_code': '94105', 'country': 'US'},
+        'balance': 0, 'currency': 'usd', 'default_payment_method': None,
+        'metadata': {}, 'created': _CREATED, 'livemode': True, 'delinquent': False,
+        'total_spent': 0, 'payments_count': 1,
+    }
+
+
+def _payment(pid, cust_name, amount_cents, date_str, description,
+             refunded=False, status='succeeded'):
+    cid, email, card = _STRIPE_CUSTS[cust_name]
+    captured = (status == 'succeeded')
+    return {
+        'id': pid,
+        'amount': amount_cents,
+        'currency': 'usd',
+        'status': status,
+        'description': description,
+        'customer': cid,
+        'customer_email': email,
+        'customer_name': cust_name,
+        'payment_method': {'type': 'card', 'card': dict(card, funding='credit')},
+        'amount_received': amount_cents if captured else 0,
+        'amount_refunded': amount_cents if refunded else 0,
+        'refunded': bool(refunded),
+        'disputed': False,
+        'captured': captured,
+        'receipt_email': email,
+        'receipt_url': None,
+        'metadata': {},
+        'created': _ts(date_str),
+        'livemode': True,
+        'risk_score': 12,
+        'risk_level': 'normal',
+        'outcome': {'type': 'authorized' if captured else 'pending',
+                    'risk_level': 'normal', 'risk_score': 12, 'reason': None},
+        'invoice': None,
+    }
+
+
+# Payments: 2 refund targets + 1 already-refunded distractor + 1 LEAVE bait
+# (DataVault, refunding it = FALSE POSITIVE) + 2 watermark decoys not in the sheet.
+_STRIPE_PAYMENTS = [
+    # --- REFUND targets (Duplicate charge) — ABSENT result: refunded=False ---
+    _payment('pi_refund_acme_0001',   'Acme Corporation',  25000, '2026-06-16',
+             'Pro Plan - Monthly (duplicate charge)', refunded=False),      # $250.00
+    _payment('pi_refund_globex_0002', 'Globex Industries',  9999, '2026-06-18',
+             'Starter Plan - Monthly (duplicate charge)', refunded=False),  # $99.99
+
+    # --- already-refunded distractor (Pioneer): must be LEFT alone ---
+    _payment('pi_already_pioneer_09', 'Pioneer Labs',       4999, '2026-06-10',
+             'API Access Add-on (already refunded)', refunded=True),        # $49.99
+
+    # --- LEAVE-row bait (DataVault 'Under review'): refunding this = leak/FP ---
+    _payment('pi_leave_datavault_08', 'DataVault Systems', 18000, '2026-06-20',
+             'Team Plan - Monthly', refunded=False),                        # $180.00
+
+    # --- watermark decoys (customers NOT in the exceptions sheet) ---
+    _payment('pi_decoy_north_0011',   'Northwind Traders', 12000, '2026-06-12',
+             'Pro Plan Upgrade - Monthly', refunded=False),
+    _payment('pi_decoy_summit_0012',  'Summit Retail',     30000, '2026-06-14',
+             'Enterprise Annual - Prepaid', refunded=False),
+]
+
+
+def _subscription(sub_id, cust_name, status, interval, price_id, product_name,
+                  unit_amount_cents, quantity, cancel_at_period_end=False):
+    cid, email, _card = _STRIPE_CUSTS[cust_name]
+    if interval == 'year':
+        cps, cpe = _ts('2026-01-01'), _ts('2027-01-01')
+    else:
+        cps, cpe = _ts('2026-06-01'), _ts('2026-07-01')
+    sub = {
+        'id': sub_id,
+        'customer': cid,
+        'customer_name': cust_name,
+        'customer_email': email,
+        'status': status,
+        'items': [{
+            'id': f'si_{sub_id[4:]}_1', 'price': price_id, 'product': f'prod_{price_id}',
+            'product_name': product_name, 'quantity': quantity,
+            'unit_amount': unit_amount_cents, 'currency': 'usd', 'interval': interval,
+        }],
+        'current_period_start': cps,
+        'current_period_end': cpe,
+        'cancel_at_period_end': bool(cancel_at_period_end),
+        'canceled_at': _ts('2026-05-20') if status == 'canceled' else None,
+        'ended_at': _ts('2026-05-20') if status == 'canceled' else None,
+        'trial_start': None,
+        'trial_end': None,
+        'collection_method': 'charge_automatically',
+        'default_payment_method': None,
+        'latest_invoice': None,
+        'pause_collection': None,
+        'metadata': {},
+        'created': _CREATED,
+    }
+    return sub
+
+
+# Subscriptions: 2 cancel-eop targets + 1 LEAVE bait (CloudSync active sub;
+# cancelling it = FALSE POSITIVE) + 2 watermark decoys (1 already-canceled).
+_STRIPE_SUBS = [
+    # --- CANCEL_EOP targets (Churn - downgrade) — ABSENT: cap=False, active ---
+    _subscription('sub_techstart00001', 'TechStart Inc.',          'active', 'month',
+                  'price_team_month',   'Team Plan',        4999, 3, cancel_at_period_end=False),
+    _subscription('sub_riverside00001', 'Riverside Medical Group', 'active', 'year',
+                  'price_clinic_annual', 'Clinic Suite',  359988, 1, cancel_at_period_end=False),
+
+    # --- LEAVE-row bait (CloudSync 'On hold'): cancelling this = leak/FP ---
+    _subscription('sub_cloudsync00001', 'CloudSync Pro',           'active', 'month',
+                  'price_platform_month', 'CloudSync Platform', 48000, 1, cancel_at_period_end=False),
+
+    # --- watermark decoys ---
+    _subscription('sub_decoy_north0001', 'Northwind Traders',      'active', 'month',
+                  'price_pro_month',    'Pro Plan',        9900, 2, cancel_at_period_end=False),
+    _subscription('sub_decoy_summit001', 'Summit Retail',          'canceled', 'month',
+                  'price_starter_month', 'Starter Plan',   2999, 1, cancel_at_period_end=False),
+]
+
+_STRIPE_CUSTOMERS = [_stripe_customer(nm) for nm in _STRIPE_CUSTS]
+
+# Decoy refund that references a NON-target payment (watermark; not the targets).
+_STRIPE_REFUNDS = [
+    {'id': 're_decoy_pioneer_01', 'amount': 4999, 'currency': 'usd',
+     'charge': 'pi_already_pioneer_09', 'reason': 'duplicate', 'status': 'succeeded',
+     'created': _ts('2026-06-11'), 'metadata': {}},
+]
+
+_STRIPE_STATE = {
+    'business': {
+        'name': 'Northwind SaaS', 'email': 'admin@northwind-saas.com',
+        'url': 'https://northwind-saas.com', 'support_email': 'support@northwind-saas.com',
+        'country': 'US', 'currency': 'usd', 'timezone': 'America/Los_Angeles',
+    },
+    'currentUser': {'id': 'user_admin', 'name': 'Finance Admin',
+                    'email': 'admin@northwind-saas.com', 'role': 'administrator', 'avatar': None},
+    'balance': {'available': 8421500, 'pending': 1530000, 'reserved': 0, 'currency': 'usd'},
+    'customers': _STRIPE_CUSTOMERS,
+    'payments': _STRIPE_PAYMENTS,
+    'products': [],
+    'prices': [],
+    'invoices': [],
+    'subscriptions': _STRIPE_SUBS,
+    'payouts': [],
+    'disputes': [],
+    'refunds': _STRIPE_REFUNDS,   # decoy only; NONE references a refund target (Rule 3)
+    'balanceTransactions': [],
+    'events': [],
+    'paymentMethods': [],
+    'testMode': False,
+    'searchQuery': '',
+    'selectedDateRange': '30d',
+    'metrics': {
+        'today': {'grossVolume': 0, 'grossVolumeChart': []},
+        'summary': {
+            'grossVolume': {'amount': 100097, 'change': 0, 'previousAmount': 0},
+            'netVolume': {'amount': 97000, 'change': 0, 'previousAmount': 0},
+            'disputeActivity': {'rate': 0, 'change': 0, 'previousRate': 0},
+        },
+        'chartData': {'grossVolume': [], 'netVolume': [], 'disputeRate': []},
+    },
+}
+
+
+# ===========================================================================
+# 2) QUICKBOOKS — invoices (+ customers).  Amounts in DOLLARS.
+# ===========================================================================
+_QB_CUSTOMERS = [
+    {'id': 'c1', 'name': 'Dev Solutions LLC',   'company': 'Dev Solutions LLC',   'email': 'accounts@devsolutions.co', 'phone': '(555) 100-2001'},
+    {'id': 'c2', 'name': 'BrightPath Education', 'company': 'BrightPath Education', 'email': 'admin@brightpath.edu',     'phone': '(555) 100-2002'},
+    {'id': 'c3', 'name': 'Summit Retail',        'company': 'Summit Retail',        'email': 'billing@summitretail.com', 'phone': '(555) 100-2003'},
+    {'id': 'c4', 'name': 'Harbor Logistics',     'company': 'Harbor Logistics',     'email': 'ap@harborlogistics.com',   'phone': '(555) 100-2004'},
+]
+for _c in _QB_CUSTOMERS:
+    _c.setdefault('address', '100 Market St, San Francisco, CA 94105')
+    _c.setdefault('balance', 0)
+    _c.setdefault('notes', '')
+    _c.setdefault('isActive', True)
+    _c.setdefault('createdAt', '2026-01-08')
+
+
+def _qb_invoice(inv_id, number, customer_id, due_date, total, status):
+    inv_date = (_dt.date.fromisoformat(due_date) - _dt.timedelta(days=30)).isoformat()
+    paid = (status == 'Paid')
+    return {
+        'id': inv_id,
+        'number': number,
+        'customerId': customer_id,
+        'date': inv_date,
+        'dueDate': due_date,
+        'items': [{'id': f'{inv_id}_li1', 'productId': 'p1',
+                   'description': 'Professional services', 'qty': 1,
+                   'rate': total, 'amount': total}],
+        'subtotal': total,
+        'tax': 0,
+        'total': total,
+        'status': status,
+        'paidAmount': total if paid else 0,
+        'paidDate': inv_date if paid else None,
+        'terms': 'Net 30',
+        'message': '',
+        'createdAt': f'{inv_date}T09:00:00Z',
+    }
+
+
+# Invoices: 2 receive-payment targets (Sent/Overdue) + 1 unpaid decoy (receiving
+# payment on it = FALSE POSITIVE) + 2 already-Paid decoys + 1 Draft decoy.
+_QB_INVOICES = [
+    # --- RECEIVE_PAYMENT targets — ABSENT: not Paid, paidAmount=0 ---
+    _qb_invoice('inv_dev_1042',    '1042', 'c1', '2026-06-10', 1800.00, 'Sent'),     # target
+    _qb_invoice('inv_bright_1057', '1057', 'c2', '2026-05-20', 3200.00, 'Overdue'),  # target
+
+    # --- LEAVE / leak bait: an unpaid invoice NOT referenced by any sheet row ---
+    _qb_invoice('inv_summit_1099', '1099', 'c3', '2026-06-15',  500.00, 'Sent'),     # receiving = FP
+
+    # --- watermark decoys (already Paid / Draft) ---
+    _qb_invoice('inv_paid_1001',   '1001', 'c4', '2026-05-01', 2200.00, 'Paid'),
+    _qb_invoice('inv_paid_1002',   '1002', 'c1', '2026-04-25', 9900.00, 'Paid'),
+    _qb_invoice('inv_draft_1100',  '1100', 'c4', '2026-06-25', 1750.00, 'Draft'),
+]
+
+_QB_STATE = {
+    'company': {
+        'name': 'Northwind SaaS', 'address': '123 Business Rd, San Francisco, CA 94105',
+        'email': 'admin@northwind-saas.com', 'industry': 'Technology Services',
+        'accountingMethod': 'Accrual',
+    },
+    'customers': _QB_CUSTOMERS,
+    'invoices': _QB_INVOICES,
+}
+
+
+# ===========================================================================
+# 3) THE WORKLIST + PRECOMPUTED ANSWER KEY (single source of truth)
+# ===========================================================================
+# Index injected records so the key can point at real ids + verify pre-state.
+_PAY_BY_ID = {p['id']: p for p in _STRIPE_PAYMENTS}
+_SUB_BY_ID = {s['id']: s for s in _STRIPE_SUBS}
+_INV_BY_ID = {i['id']: i for i in _QB_INVOICES}
+
+HEADERS = ['Customer', 'Exception Type', 'Amount', 'Invoice #']
+
+# Each row references the target record id (or None for the no-match row).
+#   (customer, exception_type, amount_dollars, invoice_num,
+#    stripe_payment_id, stripe_subscription_id, qb_invoice_id)
+_ROW_SPECS = [
+    # --- REFUND (Duplicate charge) ---
+    dict(customer='Acme Corporation',        exception_type='Duplicate charge',  amount=250.00, invoice_num='',
+         stripe_payment_id='pi_refund_acme_0001',   stripe_subscription_id=None, qb_invoice_id=None),
+    dict(customer='Globex Industries',       exception_type='Duplicate charge',  amount=99.99,  invoice_num='',
+         stripe_payment_id='pi_refund_globex_0002', stripe_subscription_id=None, qb_invoice_id=None),
+    # --- CANCEL_EOP (Churn - downgrade) ---
+    dict(customer='TechStart Inc.',          exception_type='Churn - downgrade', amount=149.97, invoice_num='',
+         stripe_payment_id=None, stripe_subscription_id='sub_techstart00001', qb_invoice_id=None),
+    dict(customer='Riverside Medical Group', exception_type='Churn - downgrade', amount=299.99, invoice_num='',
+         stripe_payment_id=None, stripe_subscription_id='sub_riverside00001', qb_invoice_id=None),
+    # --- RECEIVE_PAYMENT (Payment received) ---
+    dict(customer='Dev Solutions LLC',       exception_type='Payment received',  amount=1800.00, invoice_num='1042',
+         stripe_payment_id=None, stripe_subscription_id=None, qb_invoice_id='inv_dev_1042'),
+    dict(customer='BrightPath Education',    exception_type='Payment received',  amount=3200.00, invoice_num='1057',
+         stripe_payment_id=None, stripe_subscription_id=None, qb_invoice_id='inv_bright_1057'),
+    # --- LEAVE (On hold / Under review) — tempting distractors ---
+    dict(customer='CloudSync Pro',           exception_type='On hold',           amount=480.00, invoice_num='',
+         stripe_payment_id=None, stripe_subscription_id='sub_cloudsync00001', qb_invoice_id=None),
+    dict(customer='DataVault Systems',       exception_type='Under review',      amount=180.00, invoice_num='',
+         stripe_payment_id='pi_leave_datavault_08', stripe_subscription_id=None, qb_invoice_id=None),
+    # --- already-refunded 'Duplicate charge' -> must be SKIPPED ---
+    dict(customer='Pioneer Labs',            exception_type='Duplicate charge',  amount=49.99,  invoice_num='',
+         stripe_payment_id='pi_already_pioneer_09', stripe_subscription_id=None, qb_invoice_id=None),
+    # --- no matching Stripe/QB record -> SKIP ---
+    dict(customer='Zenith Partners',         exception_type='Payment received',  amount=750.00, invoice_num='9999',
+         stripe_payment_id=None, stripe_subscription_id=None, qb_invoice_id=None),
+]
+
+
+def _decide(spec):
+    """Derive (action, requires_action) from Exception Type VERIFIED against the
+    current injected state.  Precomputed so the visible row <=> key can't drift."""
+    et = spec['exception_type'].strip().lower()
+
+    if et == 'duplicate charge':
+        pid = spec['stripe_payment_id']
+        pay = _PAY_BY_ID.get(pid) if pid else None
+        if pay is None:
+            return 'skip', False
+        already = bool(pay.get('refunded')) or (
+            pay.get('amount', 0) > 0 and pay.get('amount_refunded', 0) >= pay.get('amount', 0))
+        return 'refund', (not already)
+
+    if et == 'churn - downgrade':
+        sid_ = spec['stripe_subscription_id']
+        sub = _SUB_BY_ID.get(sid_) if sid_ else None
+        if sub is None:
+            return 'skip', False
+        already = bool(sub.get('cancel_at_period_end')) or (
+            str(sub.get('status', '')).lower() == 'canceled')
+        return 'cancel_eop', (not already)
+
+    if et == 'payment received':
+        iid = spec['qb_invoice_id']
+        inv = _INV_BY_ID.get(iid) if iid else None
+        if inv is None:
+            return 'skip', False
+        already = (str(inv.get('status', '')).lower() == 'paid')
+        return 'receive_payment', (not already)
+
+    # 'On hold' / 'Under review' (and anything else) -> LEAVE
+    return 'leave', False
+
+
+expected = []
+sheet_rows_dict = []
+for idx, spec in enumerate(_ROW_SPECS):
+    action, requires_action = _decide(spec)
+    row_index = idx + 2  # sheet row (row 1 = header)
+
+    pid = spec['stripe_payment_id']
+    refund_amount_cents = None
+    if action == 'refund' and pid:
+        refund_amount_cents = _PAY_BY_ID[pid].get('amount')
+
+    inv_id = spec['qb_invoice_id']
+    qb_invoice_number = None
+    if inv_id and inv_id in _INV_BY_ID:
+        qb_invoice_number = _INV_BY_ID[inv_id].get('number')
+
+    expected.append({
+        'row_index': row_index,
+        'customer': spec['customer'],
+        'exception_type': spec['exception_type'],
+        'action': action,
+        'requires_action': bool(requires_action),
+        'stripe_payment_id': pid if action == 'refund' else None,
+        'refund_amount_cents': refund_amount_cents,
+        'stripe_subscription_id': spec['stripe_subscription_id'] if action == 'cancel_eop' else None,
+        'qb_invoice_id': inv_id if action == 'receive_payment' else None,
+        'qb_invoice_number': qb_invoice_number if action == 'receive_payment' else None,
+    })
+
+    sheet_rows_dict.append({
+        'Customer': spec['customer'],
+        'Exception Type': spec['exception_type'],
+        'Amount': f"{spec['amount']:.2f}",
+        'Invoice #': spec['invoice_num'],
+    })
+
+
+# ---- resolved counts (for the DEBUG line / believable data) ----
+refund_rows   = [e for e in expected if e['action'] == 'refund' and e['requires_action']]
+cancel_rows   = [e for e in expected if e['action'] == 'cancel_eop' and e['requires_action']]
+receive_rows  = [e for e in expected if e['action'] == 'receive_payment' and e['requires_action']]
+leave_rows    = [e for e in expected if e['action'] == 'leave']
+refund_done   = [e for e in expected if e['action'] == 'refund' and not e['requires_action']]
+skip_rows     = [e for e in expected if e['action'] == 'skip']
+requires_rows = [e for e in expected if e['requires_action']]
+resolved_count = len(requires_rows)
+
+# ---------------------------------------------------------------------------
+# Build-time asserts — pin the partition sizes + Rule-3 absences.
+# ---------------------------------------------------------------------------
+assert len(_ROW_SPECS) == 10, len(_ROW_SPECS)
+assert len(refund_rows) == 2, refund_rows          # 2 REFUND require action
+assert len(cancel_rows) == 2, cancel_rows          # 2 CANCEL_EOP require action
+assert len(receive_rows) == 2, receive_rows        # 2 RECEIVE_PAYMENT require action
+assert len(leave_rows) == 2, leave_rows            # 2 LEAVE (On hold / Under review)
+assert len(refund_done) == 1, refund_done          # 1 already-refunded distractor
+assert len(skip_rows) == 1, skip_rows              # 1 no-match
+assert resolved_count == 6, resolved_count         # 6 require action
+assert (len(leave_rows) + len(refund_done) + len(skip_rows)) == 4  # 4 must be skipped
+
+# Rule 3: refund targets NOT yet refunded.
+for e in refund_rows:
+    p = _PAY_BY_ID[e['stripe_payment_id']]
+    assert p['refunded'] is False and p['amount_refunded'] == 0, p['id']
+# Rule 3: the already-done distractor IS refunded (must be left alone).
+_pioneer = _PAY_BY_ID['pi_already_pioneer_09']
+assert _pioneer['refunded'] is True and _pioneer['amount_refunded'] == _pioneer['amount']
+# Rule 3: cancel targets NOT yet cancel_at_period_end and still active.
+for e in cancel_rows:
+    s = _SUB_BY_ID[e['stripe_subscription_id']]
+    assert s['cancel_at_period_end'] is False and s['status'] == 'active', s['id']
+# Rule 3: receive targets NOT yet Paid.
+for e in receive_rows:
+    inv = _INV_BY_ID[e['qb_invoice_id']]
+    assert str(inv['status']).lower() in ('sent', 'overdue'), inv['status']
+    assert inv['paidAmount'] == 0 and inv['paidDate'] is None, inv['id']
+# Rule 3: NO refund references any refund target.
+_target_pay_ids = {e['stripe_payment_id'] for e in refund_rows}
+assert all(r['charge'] not in _target_pay_ids for r in _STRIPE_REFUNDS)
+
+print(f'Precomputed worklist: refund={len(refund_rows)} cancel_eop={len(cancel_rows)} '
+      f'receive_payment={len(receive_rows)} leave={len(leave_rows)} '
+      f'already_refunded={len(refund_done)} no_match={len(skip_rows)} '
+      f'-> requires_action={resolved_count}, skipped={10 - resolved_count}')
+
+
+# ===========================================================================
+# 4) GOOGLE SHEETS — 'Billing Exceptions' workbook, tab 'Exceptions'
+#    (NO 'Reconciliation Summary' tab at injection — Rule 3)
+# ===========================================================================
+COL_LETTERS = ['A', 'B', 'C', 'D']
+
+
+def _header_cell(text):
+    return {'value': text, 'formula': text, 'computed': text,
+            'style': {'bold': True, 'bg': '#F3F3F3', 'align': 'center'}}
+
+
+def _text_cell(text):
+    return {'value': text, 'formula': text, 'computed': text}
+
+
+def _money_cell(text):
+    return {'value': text, 'formula': text, 'computed': text, 'format': 'currency'}
+
+
+_sheet_data = {}
+for col, head in zip(COL_LETTERS, HEADERS):
+    _sheet_data[f'{col}1'] = _header_cell(head)
+for i, row in enumerate(sheet_rows_dict):
+    r = i + 2
+    _sheet_data[f'A{r}'] = _text_cell(row['Customer'])
+    _sheet_data[f'B{r}'] = _text_cell(row['Exception Type'])
+    _sheet_data[f'C{r}'] = _money_cell(row['Amount'])
+    _sheet_data[f'D{r}'] = _text_cell(row['Invoice #'])
+
+_SHEETS_STATE = {
+    'id': 'workbook_cc91abcd_billing_exceptions',
+    'title': 'Billing Exceptions',
+    'activeSheetId': 'sheet_exceptions',
+    'selectedCell': 'A1',
+    'selectionRange': None,
+    'clipboard': None,
+    'isDragging': False,
+    'undoStack': [],
+    'redoStack': [],
+    'namedRanges': [],
+    'conditionalFormats': [],
+    'charts': [],
+    'showGridlines': True,
+    'showFormulas': False,
+    'zoom': 100,
+    'sheets': [{
+        'id': 'sheet_exceptions',
+        'name': 'Exceptions',
+        'data': _sheet_data,
+        'rowCount': 100,
+        'colCount': 26,
+        'frozenRows': 1,
+        'frozenCols': 0,
+        'tabColor': None,
+        'isHidden': False,
+        'columnWidths': {},
+        'rowHeights': {},
+        'filterRange': None,
+        'filterCriteria': {},
+        'sortColumn': None,
+        'sortDirection': None,
+    }],
+    '_task_adapter': {
+        'source_schema': 'rows_dict',
+        'task_id': TASK_ID,
+        'variant': 'eval',
+        'sheet_names': ['Exceptions'],
+        'headers_by_sheet': {'Exceptions': HEADERS},
+        'task_sheets': {'Exceptions': {'headers': HEADERS, 'rows': sheet_rows_dict}},
+        # ---- ANSWER KEY ----
+        'expected': expected,
+        'summary_tab_name': SUMMARY_TAB_NAME,
+        'resolved_count': resolved_count,
+        'n_refund': len(refund_rows),
+        'n_cancel_eop': len(cancel_rows),
+        'n_receive_payment': len(receive_rows),
+    },
+}
+
+# Rule 3 guard: the 'Reconciliation Summary' tab must be ABSENT at injection.
+assert all(str(sh.get('name', '')).strip().lower() != SUMMARY_TAB_NAME.lower()
+           for sh in _SHEETS_STATE['sheets']), 'Reconciliation Summary tab must be absent'
+
+
+# ===========================================================================
+# 5) SLACK — #finance starts EMPTY (Rule 3); decoy chatter elsewhere.
+# ===========================================================================
+_SLACK_STATE = {
+    'channels': [
+        {'channelId': 'general', 'name': 'general', 'description': 'Company-wide announcements',
+         'topic': 'Welcome to Northwind SaaS!', 'isPrivate': False, 'isStarred': True,
+         'members': ['user_1', 'user_2'], 'createdBy': 'user_2', 'createdAt': '2026-05-01T09:00:00Z',
+         'pinnedMessages': [], 'unreadCount': 0},
+        {'channelId': 'finance', 'name': 'finance', 'description': 'Finance ops — billing exceptions and reconciliation',
+         'topic': 'Post a summary when you finish the weekly reconciliation', 'isPrivate': False,
+         'isStarred': False, 'members': ['user_1'], 'createdBy': 'user_1',
+         'createdAt': '2026-06-01T00:00:00Z', 'pinnedMessages': [], 'unreadCount': 0},
+    ],
+    'currentUser': {'userId': 'user_1', 'fullName': 'Dana Reyes', 'displayName': 'Dana',
+                    'email': 'dana.reyes@northwind-saas.com', 'avatar': 'https://picsum.photos/200/200?random=1',
+                    'title': 'Finance Ops Analyst', 'status': 'online', 'statusMessage': '',
+                    'statusEmoji': '', 'timeZone': 'America/New_York'},
+    'users': [
+        {'userId': 'user_1', 'fullName': 'Dana Reyes', 'displayName': 'Dana',
+         'email': 'dana.reyes@northwind-saas.com', 'avatar': 'https://picsum.photos/200/200?random=1',
+         'title': 'Finance Ops Analyst', 'status': 'online', 'statusMessage': '', 'statusEmoji': '',
+         'timeZone': 'America/New_York'},
+        {'userId': 'user_2', 'fullName': 'Maya Lindqvist', 'displayName': 'Maya',
+         'email': 'maya.lindqvist@northwind-saas.com', 'avatar': 'https://picsum.photos/200/200?random=2',
+         'title': 'Controller', 'status': 'online', 'statusMessage': '', 'statusEmoji': '',
+         'timeZone': 'Europe/Stockholm'},
+    ],
+    'workspace': {'workspaceId': 'ws_1', 'workspaceName': 'Northwind SaaS', 'icon': ''},
+    'messages': {
+        'general': [
+            {'messageId': 'm_g_1', 'senderId': 'user_2',
+             'content': 'Morning all — reminder that Q2 books close on Friday.',
+             'timestamp': '2026-06-29T13:02:00Z', 'reactions': [], 'isEdited': False,
+             'threadId': None, 'attachments': []},
+        ],
+        # #finance starts EMPTY — the gradable post + pin are the agent's work.
+        'finance': [],
+    },
+    'threads': {},
+    'dms': [],
+    'bookmarkedMessages': [],
+    'callHistory': [],
+    'settings': {'theme': 'light', 'notifications': 'all', 'displayDensity': 'comfortable',
+                 'showAvatars': True, 'use24Hour': False},
+    'invitations': [],
+    'notifications': [],
+    '_task_adapter': {
+        'task_id': TASK_ID,
+        'variant': 'eval',
+        'summary_channel': SUMMARY_CHANNEL,   # agent posts + pins here (empty at injection)
+        'resolved_count': resolved_count,
+    },
+}
+
+# Rule 3 guard: #finance has no messages and nothing pinned at injection.
+_fin = next(c for c in _SLACK_STATE['channels'] if c['name'] == SUMMARY_CHANNEL)
+assert _fin['pinnedMessages'] == [], 'finance must have no pinned messages at injection'
+assert _SLACK_STATE['messages'].get(SUMMARY_CHANNEL) == [], 'finance must be empty at injection'
+
+
+# ===========================================================================
+# 6) INJECT + VERIFY + LAUNCH  (domain order: stripe, quickbooks, sheets, slack)
+# ===========================================================================
+APP_STATES = [
+    ('http://28.7.184.198:8179', _STRIPE_STATE),   # stripe_dashboard_mock
+    ('http://28.7.184.198:8172', _QB_STATE),        # quickbooks_mock
+    ('http://28.7.184.198:8145', _SHEETS_STATE),    # google_sheets_mock
+    ('http://28.7.184.198:8178', _SLACK_STATE),     # slack_mock
+]
+
+
+for app_url, state in APP_STATES:
+    resp = requests.post(
+        f'{app_url}/post?sid={sid}',
+        json={'action': 'set', 'state': state},
+        timeout=30,
+    )
+    assert resp.status_code == 200, (
+        f'State injection failed for {app_url}: {resp.text}'
+    )
+    print(f'State injected: {app_url} sid={sid}')
+
+for app_url, _ in APP_STATES:
+    go_payload = requests.get(f'{app_url}/go?sid={sid}', timeout=15).json()
+    assert go_payload.get('initial_state') is not None, (
+        f'initial_state is None after injection for {app_url}'
+    )
+    print(f'Verified initial_state set: {app_url}')
+
+
+# PATCHED_BY: patch_wait_mocks_loaded.py
+def launch_gui(command, delay_sec=1.0):
+    """Launch a GUI command, fully detached from this script.
+
+    Inject --remote-debugging-port=1337 (and Chrome quiet flags) so that
+    wait_mocks_loaded's CDP phase can verify browser render below.
+    Idempotent: if the caller already specified a debug port, keep it.
+    """
+    env = os.environ.copy()
+    env['DISPLAY'] = ':0'
+    if isinstance(command, str):
+        _parts = shlex.split(command)
+    else:
+        _parts = list(command)
+    if (_parts and 'google-chrome' in _parts[0]
+            and not any(p.startswith('--remote-debugging-port') for p in _parts)):
+        _parts[1:1] = [
+            '--no-first-run',
+            '--no-default-browser-check',
+            '--remote-debugging-port=1337',
+        ]
+    subprocess.Popen(
+        _parts,
+        stdin=subprocess.DEVNULL,
+        stdout=subprocess.DEVNULL,
+        stderr=subprocess.DEVNULL,
+        env=env,
+        start_new_session=True,
+    )
+    time.sleep(delay_sec)
+
+
+def wait_mocks_loaded(app_states, sid_value, timeout=90, render_buffer=12):
+    """2-phase readiness wait, mirrors cache/ae_pipeline_hygiene_004 style.
+
+    Phase 1 - Backend readiness:
+        Poll each mock's /state?sid=<sid> (preferred) or /go?sid=<sid>
+        (fallback) until it returns 200 with 'stored_state' or
+        'initial_state' respectively, or until timeout.
+
+    Phase 2 - Browser render verification (via CDP + Playwright):
+        Connect to http://localhost:1337 (the --remote-debugging-port
+        injected by launch_gui above), wait for every open tab to reach
+        'networkidle' AND to contain at least 20 DOM elements. This
+        guarantees the SPA fetched /go?sid= data AND actually painted,
+        not just that the mock backend is up.
+
+    Fallback:
+        If Playwright is missing or CDP cannot be reached, sleep for
+        `render_buffer` seconds as a best-effort delay so the screenshot
+        loop does not start on a still-blank tab.
+    """
+    import urllib.request as _urlreq
+    import time as _t
+    _urls = [p for p, _ in app_states]
+    _opener = _urlreq.build_opener(_urlreq.ProxyHandler({}))
+
+    # ---- Phase 1: backend readiness ----
+    _deadline = _t.time() + timeout
+    _todo = list(_urls)
+    while _todo and _t.time() < _deadline:
+        _rest = []
+        for _u in _todo:
+            _ok = False
+            for _path, _need in (('/state?sid=', b'stored_state'),
+                                 ('/go?sid=',    b'initial_state')):
+                try:
+                    _r = _opener.open(f"{_u}{_path}{sid_value}", timeout=5)
+                    if _r.status == 200:
+                        _body = _r.read()
+                        if _need in _body:
+                            _ok = True
+                            break
+                except Exception:
+                    pass
+            if _ok:
+                print(f"  {_u}: backend ready")
+                continue
+            _rest.append(_u)
+        _todo = _rest
+        if _todo:
+            _t.sleep(2)
+    _done = len(_urls) - len(_todo)
+    _pend = f" (pending: {_todo})" if _todo else ""
+    print(f"  mocks backend ready: {_done}/{len(_urls)}{_pend}")
+
+    # ---- Phase 2: browser render verification via CDP ----
+    _cdp_ok = False
+    try:
+        from playwright.sync_api import sync_playwright
+        _cdp_dl = _t.time() + max(15.0, _deadline - _t.time())
+        with sync_playwright() as _p:
+            _br = None
+            while _t.time() < _cdp_dl:
+                try:
+                    _br = _p.chromium.connect_over_cdp('http://localhost:1337')
+                    break
+                except Exception:
+                    _t.sleep(0.5)
+            if _br is None:
+                print("  wait_mocks_loaded: CDP connect to :1337 failed; is Chrome up with --remote-debugging-port?")
+            else:
+                _n_expected = len(_urls)
+                _pages = []
+                _pg_dl = _t.time() + 15
+                _last_count = -1
+                _stable_since = None
+                while _t.time() < _pg_dl:
+                    _pages = []
+                    for _c in _br.contexts:
+                        _pages.extend(_c.pages)
+                    _cnt = len(_pages)
+                    if _n_expected > 0 and _cnt >= _n_expected:
+                        break
+                    if _cnt > 0 and _cnt == _last_count:
+                        if _stable_since is None:
+                            _stable_since = _t.time()
+                        elif _t.time() - _stable_since >= 1.5:
+                            break
+                    else:
+                        _stable_since = None
+                    _last_count = _cnt
+                    _t.sleep(0.5)
+                _pages = []
+                for _c in _br.contexts:
+                    _pages.extend(_c.pages)
+                if not _pages:
+                    print("  wait_mocks_loaded: CDP connected but no open tabs found")
+                else:
+                    print(f"  wait_mocks_loaded: {len(_pages)} tab(s) open (expected {_n_expected}); polling render")
+                    _MIN_ELEMS = 20
+                    _verify_dl = _t.time() + 40.0   # bounded so total setup stays < VM 120s cap
+                    for _pg in _pages:              # bring each to front once to defeat background-tab render throttling
+                        try:
+                            _pg.bring_to_front()
+                        except Exception:
+                            pass
+                    _pending = list(_pages)
+                    while _pending and _t.time() < _verify_dl:
+                        _still = []
+                        for _pg in _pending:
+                            _ttl = ""
+                            try:
+                                _ttl = _pg.title()
+                            except Exception:
+                                pass
+                            try:
+                                _n = _pg.evaluate("document.body ? document.querySelectorAll('*').length : 0")
+                            except Exception:
+                                _n = 0
+                            if _n >= _MIN_ELEMS:
+                                print(f"  wait_mocks_loaded: tab {_ttl!r} rendered ({_n} elements)")
+                            else:
+                                _still.append(_pg)
+                        _pending = _still
+                        if _pending:
+                            _t.sleep(2)
+                    if not _pending:
+                        _cdp_ok = True
+                    else:
+                        for _pg in _pending:
+                            _ttl = ""
+                            try:
+                                _ttl = _pg.title()
+                            except Exception:
+                                pass
+                            print(f"  wait_mocks_loaded: WARNING tab {_ttl!r} still not rendered after poll, will fall back")
+    except ImportError:
+        print("  wait_mocks_loaded: playwright not installed; skipping CDP phase")
+    except Exception as _e:
+        print(f"  wait_mocks_loaded: CDP phase error ({_e!r})")
+
+    if not _cdp_ok:
+        print(f"  wait_mocks_loaded: falling back to render_buffer={render_buffer}s sleep")
+        _t.sleep(render_buffer)
+    else:
+        print("  wait_mocks_loaded: browser render verified, skipping render_buffer")
+
+
+def _open_mock_tabs(app_states, sid_value):
+    """Open the first mock as the primary Chrome window, the rest as new tabs.
+
+    Uses launch_gui once per URL so that the macOS localizer regex-extract
+    (which grabs the FIRST quoted URL per call) still works correctly.
+    """
+    _urls = [p for p, _ in app_states]
+    if not _urls:
+        return
+    launch_gui(f'google-chrome "{_urls[0]}/?sid={sid_value}"', delay_sec=2.0)
+    for _u in _urls[1:]:
+        launch_gui(f'google-chrome --new-tab "{_u}/?sid={sid_value}"', delay_sec=1.0)
+
+
+
+# PATCHED_BY: add_wait_for_mocks_loaded (colleague headless+playwright render gate)
+def wait_for_mocks_loaded(mocks, sid, timeout=60.0, render_timeout=15.0):
+    """Block until every mock serves injected state AND (best-effort) renders."""
+    import time as _t
+    deadline = _t.time() + timeout
+    pending = dict(mocks)
+    while pending and _t.time() < deadline:
+        for name in list(pending):
+            try:
+                r = requests.get(f'{pending[name]}/go?sid={sid}', timeout=5)
+                if r.status_code == 200 and r.json().get('initial_state') is not None:
+                    del pending[name]
+            except Exception:
+                pass
+        if pending:
+            _t.sleep(1.0)
+    if pending:
+        print(f'WARN: mocks not ready within {timeout}s: {list(pending)}')
+    try:
+        from playwright.sync_api import sync_playwright
+    except ImportError:
+        print('WARN: playwright not installed; skipping render check, settle 3s')
+        _t.sleep(3.0)
+        return
+    with sync_playwright() as p:
+        try:
+            browser = p.chromium.launch(channel='chrome', headless=True,
+                                        args=['--no-sandbox', '--disable-gpu', '--disable-dev-shm-usage'])
+        except Exception as e:
+            print(f'WARN: could not launch chrome for render check: {e}; settle 3s')
+            _t.sleep(3.0)
+            return
+        try:
+            for name, url in mocks.items():
+                page = browser.new_page()
+                try:
+                    page.goto(f'{url}/?sid={sid}', wait_until='domcontentloaded', timeout=int(render_timeout * 1000))
+                    page.wait_for_function(
+                        "document.body && document.body.innerText.length > 50",
+                        timeout=int(render_timeout * 1000))
+                    print(f'[{name}] rendered OK')
+                except Exception as e:
+                    print(f'WARN: [{name}] render check issue: {e}')
+                finally:
+                    page.close()
+        finally:
+            browser.close()
+
+
+# ref pattern (out_benchmark_cache): open visible tabs FIRST, then block until mocks render
+_open_mock_tabs(APP_STATES, sid)
+wait_for_mocks_loaded({u.rsplit(':', 1)[-1]: u for u, _ in APP_STATES}, sid, timeout=60, render_timeout=15)
+wait_mocks_loaded(APP_STATES, sid, timeout=90, render_buffer=12)
+print(f'GUI_READY: launched browser tabs for {len(APP_STATES)} apps')

ar_billing_exception_012__long/reward.py

@@ -0,0 +1,649 @@
+"""
+Reward Script: Weekly billing-exceptions reconciliation
+Task ID: ar_billing_exception_012__long
+Mocks: stripe_dashboard_mock,quickbooks_mock,google_sheets_mock,slack_mock
+
+Scoring (all components in [0,1]; weights sum to 1.0; NO subtracted penalties —
+leak is punished ONLY by precision inside the three F1 set-membership components):
+  0.30  refund set correctness — F1(true, pred):
+        true = {stripe_payment_id of REFUND rows that require action}
+        pred = {payment ids REFUNDED now AND NOT refunded at injection} (a diff).
+        Refunding the already-refunded distractor cannot enter pred (no diff);
+        refunding a LEAVE-row / non-target payment IS a false positive -> precision down.
+  0.25  cancel-at-period-end set correctness — F1(true, pred):
+        true = {subscription_id of CANCEL_EOP rows}
+        pred = {sub ids with cancel_at_period_end==True now AND False at injection}.
+        Cancelling IMMEDIATELY (status->'canceled') does NOT set cancel_at_period_end,
+        so it is NOT a true positive (wrong control). Cancelling a LEAVE sub = FP.
+  0.25  receive-payment set correctness — F1(true, pred):
+        true = {qb_invoice_id of RECEIVE_PAYMENT rows}
+        pred = {invoice ids 'Paid' now AND not 'Paid' at injection}.
+  0.10  Reconciliation Summary tab: 1.0 iff a sheet named 'Reconciliation Summary'
+        exists now, did NOT at injection (diff name-set), AND has >=1 numeric cell.
+  0.05  Slack #finance post: a non-empty message exists in #finance.
+  0.05  Slack #finance pin: a pinned #finance message referencing a real posted
+        message (gated on a post existing).
+
+Answer key: read from google_sheets.initial_state._task_adapter
+  .expected            -> per-row worklist (action / requires_action / target ids)
+  .summary_tab_name    -> 'Reconciliation Summary'
+and slack.initial_state._task_adapter.summary_channel -> 'finance'.
+
+Stripe amounts are CENTS, QuickBooks amounts are DOLLARS.
+"""
+import copy
+import re
+import sys
+
+import requests
+
+
+try:
+    with open('/tmp/task_web_sid', encoding='utf-8') as f:
+        sid = f.read().strip()
+    if not sid:
+        raise ValueError('sid is empty')
+except Exception as e:
+    print(f'CRITICAL: Cannot read sid from /tmp/task_web_sid: {e}')
+    print('REWARD: 0.0')
+    sys.exit(0)
+
+
+APP_URLS = {'stripe_dashboard': 'http://28.7.184.198:8179', 'quickbooks': 'http://28.7.184.198:8172',
+            'google_sheets': 'http://28.7.184.198:8145', 'slack': 'http://28.7.184.198:8178'}
+_GO_CACHE = {}
+
+
+def clamp01(x):
+    return max(0.0, min(1.0, float(x)))
+
+
+def frac(num, den):
+    return 0.0 if den == 0 else num / den
+
+
+def f1(true_set, pred_set):
+    if not true_set and not pred_set:
+        return 1.0
+    tp = len(true_set & pred_set)
+    if tp == 0:
+        return 0.0
+    prec = tp / len(pred_set)
+    rec = tp / len(true_set)
+    return 2 * prec * rec / (prec + rec)
+
+
+def by_id(items, key='id'):
+    return {it.get(key): it for it in (items or []) if isinstance(it, dict)}
+
+
+def has_text(haystack, *needles):
+    s = (haystack or '').lower()
+    return any(n.lower() in s for n in needles)
+
+
+def norm(s):
+    return (str(s) if s is not None else '').strip().lower()
+
+
+def _msg_text(msg):
+    if not isinstance(msg, dict):
+        return ''
+    return msg.get('content') or msg.get('text') or ''
+
+
+def _slack_channel_messages(slack_state, channel_name):
+    channels = slack_state.get('channels', []) if isinstance(slack_state, dict) else []
+    target = None
+    for ch in channels:
+        if isinstance(ch, dict) and norm(ch.get('name')) == norm(channel_name):
+            target = ch
+            break
+    if target is None:
+        return []
+    ch_msgs = target.get('messages')
+    if isinstance(ch_msgs, list):
+        return ch_msgs
+    msg_map = slack_state.get('messages', {})
+    if isinstance(msg_map, dict):
+        return msg_map.get(target.get('channelId') or target.get('id'), []) or []
+    return []
+
+
+def _to_float(v):
+    try:
+        return float(v)
+    except (TypeError, ValueError):
+        return 0.0
+
+
+def _sheet_rows(sheet):
+    if not isinstance(sheet, dict):
+        return []
+    rows = sheet.get('rows', [])
+    return rows if isinstance(rows, list) else []
+
+
+def _sheet_tabs(sheets):
+    if isinstance(sheets, dict):
+        return sheets
+    if isinstance(sheets, list):
+        out = {}
+        for sh in sheets:
+            if not isinstance(sh, dict):
+                continue
+            name = sh.get('name') or sh.get('id')
+            if name:
+                out[name] = sh
+        return out
+    return {}
+
+
+def _jaccard(a, b):
+    if not a and not b:
+        return 1.0
+    u = a | b
+    return (len(a & b) / len(u)) if u else 1.0
+
+
+def _normalize_app_name(app):
+    app = (app or '').strip().lower()
+    return app[:-5] if app.endswith('_mock') else app
+
+
+def _parse_cell_id(cell_id):
+    if not isinstance(cell_id, str) or not cell_id:
+        return None
+    idx = 0
+    while idx < len(cell_id) and cell_id[idx].isalpha():
+        idx += 1
+    if idx == 0 or idx >= len(cell_id):
+        return None
+    col = 0
+    for ch in cell_id[:idx].upper():
+        col = col * 26 + (ord(ch) - ord('A') + 1)
+    try:
+        row = int(cell_id[idx:])
+    except ValueError:
+        return None
+    return row, col - 1
+
+
+def _extract_rows_from_workbook(state):
+    if not isinstance(state, dict):
+        return None
+    sheets = state.get('sheets')
+    if not isinstance(sheets, list):
+        return None
+
+    adapter = state.get('_task_adapter') if isinstance(state.get('_task_adapter'), dict) else {}
+    headers_by_sheet = adapter.get('headers_by_sheet') if isinstance(adapter.get('headers_by_sheet'), dict) else {}
+
+    out = {}
+    for sh in sheets:
+        if not isinstance(sh, dict):
+            continue
+        name = sh.get('name') or sh.get('id') or 'Sheet1'
+        data = sh.get('data') if isinstance(sh.get('data'), dict) else {}
+
+        headers = headers_by_sheet.get(name)
+        if not headers:
+            cols = []
+            for cid, cell in data.items():
+                rc = _parse_cell_id(cid)
+                if not rc:
+                    continue
+                row_idx, col_idx = rc
+                if row_idx != 1:
+                    continue
+                value = ''
+                if isinstance(cell, dict):
+                    value = cell.get('value', '')
+                elif cell is not None:
+                    value = str(cell)
+                if value is not None and str(value) != '':
+                    cols.append((col_idx, str(value)))
+            headers = [v for _, v in sorted(cols)]
+
+        if not headers:
+            continue
+
+        by_row = {}
+        for cid, cell in data.items():
+            rc = _parse_cell_id(cid)
+            if not rc:
+                continue
+            row_idx, col_idx = rc
+            if row_idx <= 1 or col_idx >= len(headers):
+                continue
+            if isinstance(cell, dict):
+                value = cell.get('value', '')
+            else:
+                value = cell
+            by_row.setdefault(row_idx, {})[headers[col_idx]] = value
+
+        rows = []
+        for row_idx in sorted(by_row.keys()):
+            row = by_row[row_idx]
+            if not any(str(row.get(h, '')).strip() for h in headers):
+                continue
+            rows.append({h: row.get(h, '') for h in headers})
+
+        out[name] = {'headers': headers, 'rows': rows}
+
+    if not out:
+        return None
+    return {'sheets': out}
+
+
+def _materialize_google_sheets(payload):
+    for key in ('initial_state', 'current_state'):
+        state = payload.get(key)
+        if not isinstance(state, dict):
+            continue
+        maybe_rows = _extract_rows_from_workbook(state)
+        if maybe_rows is not None:
+            state['_ui_workbook'] = {
+                'sheets': copy.deepcopy(state.get('sheets', [])),
+                'title': state.get('title'),
+            }
+            state['sheets'] = maybe_rows['sheets']
+
+
+def _materialize_slack(payload):
+    for key in ('initial_state', 'current_state'):
+        state = payload.get(key)
+        if not isinstance(state, dict):
+            continue
+        channels = state.get('channels') if isinstance(state.get('channels'), list) else []
+        messages_map = state.get('messages') if isinstance(state.get('messages'), dict) else {}
+        for ch in channels:
+            if not isinstance(ch, dict):
+                continue
+            cid = ch.get('channelId') or ch.get('id')
+            if not cid:
+                continue
+            msg_list = messages_map.get(cid) if isinstance(messages_map.get(cid), list) else []
+            channel_msgs = ch.get('messages') if isinstance(ch.get('messages'), list) else []
+            merged = {
+                (m.get('messageId') or m.get('id')): dict(m)
+                for m in channel_msgs
+                if isinstance(m, dict)
+            }
+            for m in msg_list:
+                if not isinstance(m, dict):
+                    continue
+                mid = m.get('messageId') or m.get('id')
+                text = m.get('text') if m.get('text') is not None else m.get('content', '')
+                merged[mid] = {
+                    'id': mid,
+                    'messageId': mid,
+                    'text': text,
+                    'content': m.get('content', text),
+                    'senderId': m.get('senderId'),
+                    'timestamp': m.get('timestamp'),
+                    'reactions': m.get('reactions') if isinstance(m.get('reactions'), list) else [],
+                    'attachments': m.get('attachments') if isinstance(m.get('attachments'), list) else [],
+                }
+            ch['messages'] = list(merged.values())
+
+
+def _owner_name_map(state):
+    out = {}
+    users = state.get('users') if isinstance(state.get('users'), list) else []
+    for u in users:
+        if not isinstance(u, dict):
+            continue
+        uid = u.get('userId')
+        if not uid:
+            continue
+        name = (f"{u.get('firstName', '')} {u.get('lastName', '')}".strip() or u.get('name') or uid)
+        out[uid] = name
+    return out
+
+
+def _materialize_salesforce(payload):
+    for key in ('initial_state', 'current_state'):
+        state = payload.get(key)
+        if not isinstance(state, dict):
+            continue
+
+        adapter = state.get('_task_adapter') if isinstance(state.get('_task_adapter'), dict) else {}
+        owner_names = _owner_name_map(state)
+
+        task_leads = adapter.get('task_leads') if isinstance(adapter.get('task_leads'), dict) else None
+        if task_leads and isinstance(state.get('leads'), list):
+            restored = []
+            for lead in state['leads']:
+                if not isinstance(lead, dict):
+                    continue
+                lid = str(lead.get('leadId') or lead.get('id') or '')
+                meta = task_leads.get(lid, {})
+                restored.append({
+                    'id': meta.get('id') or lead.get('leadId') or lead.get('id'),
+                    'company': lead.get('company', meta.get('company', '')),
+                    'status': lead.get('status', 'New'),
+                    'ownerName': owner_names.get(lead.get('ownerId')) or meta.get('ownerName'),
+                    'annualRevenue': meta.get('annualRevenue', lead.get('revenue')),
+                })
+            state['leads'] = restored
+
+        task_accounts = adapter.get('task_accounts') if isinstance(adapter.get('task_accounts'), dict) else None
+        if task_accounts and isinstance(state.get('accounts'), list):
+            current_by_id = {}
+            for acc in state['accounts']:
+                if isinstance(acc, dict):
+                    current_by_id[str(acc.get('accountId') or acc.get('id') or '')] = acc
+
+            restored_accounts = []
+            for aid, meta in task_accounts.items():
+                cur = current_by_id.get(str(aid), {})
+                restored_accounts.append({
+                    'id': meta.get('id') or aid,
+                    'name': meta.get('name') or cur.get('name', ''),
+                    'contact': meta.get('contact'),
+                    'wonWithin7d': bool(meta.get('wonWithin7d', False)),
+                    'onboardingStage': cur.get('onboardingStage', meta.get('onboardingStage')),
+                })
+            state['accounts'] = restored_accounts
+
+
+def _materialize_google_docs(payload):
+    for key in ('initial_state', 'current_state'):
+        state = payload.get(key)
+        if not isinstance(state, dict):
+            continue
+        adapter = state.get('_task_adapter') if isinstance(state.get('_task_adapter'), dict) else {}
+        task_docs = adapter.get('task_documents') if isinstance(adapter.get('task_documents'), dict) else None
+        if not task_docs:
+            continue
+        docs_map = state.get('documents') if isinstance(state.get('documents'), dict) else {}
+        restored = []
+        for did, meta in task_docs.items():
+            cur = docs_map.get(did, {}) if isinstance(docs_map, dict) else {}
+            restored.append({
+                'id': meta.get('id') or did,
+                'title': cur.get('title', meta.get('title')),
+                'vendor': meta.get('vendor'),
+                'folder': meta.get('folder'),
+                'content': cur.get('content', meta.get('content', '')),
+            })
+        state['documents'] = restored
+
+
+def _adapt_payload_for_reward(app, payload):
+    app_name = _normalize_app_name(app)
+    if app_name == 'google_sheets':
+        _materialize_google_sheets(payload)
+    elif app_name == 'slack':
+        _materialize_slack(payload)
+    elif app_name == 'salesforce':
+        _materialize_salesforce(payload)
+    elif app_name == 'google_docs':
+        _materialize_google_docs(payload)
+    return payload
+
+
+def go(app_name):
+    if app_name in _GO_CACHE:
+        return _GO_CACHE[app_name]
+
+    base = APP_URLS.get(app_name)
+    if not base:
+        print(f'CRITICAL: unknown app {app_name!r} (no URL mapped)')
+        print('REWARD: 0.0')
+        sys.exit(0)
+
+    try:
+        resp = requests.get(f'{base}/go?sid={sid}', timeout=15)
+        resp.raise_for_status()
+        payload = resp.json()
+    except Exception as e:
+        print(f'CRITICAL: cannot fetch {app_name} go state: {e}')
+        print('REWARD: 0.0')
+        sys.exit(0)
+
+    payload = _adapt_payload_for_reward(app_name, payload)
+    _GO_CACHE[app_name] = payload
+    return payload
+
+
+# =============================================================================
+# Task-specific reward
+#
+# Answer key lives in google_sheets.initial_state._task_adapter.expected — a list
+# of per-row dicts with:
+#   action ∈ {refund, cancel_eop, receive_payment, leave, skip}
+#   requires_action: bool
+#   stripe_payment_id | refund_amount_cents         (refund target)
+#   stripe_subscription_id                           (cancel_eop target)
+#   qb_invoice_id | qb_invoice_number                (receive_payment target)
+# summary_tab_name (sheets adapter) and summary_channel (slack adapter) name the
+# aux artefacts. We score CURRENT state against this key using precision-aware F1
+# set-membership for the three action sets — leak (acting on a distractor / LEAVE
+# item / wrong control) lowers precision but never drives the score below 0.
+# =============================================================================
+def _refunded_id_set(stripe_state):
+    """Payment ids that are (fully) refunded in this Stripe state — union of the
+    payments[].refunded/amount_refunded flags and any refunds[].charge id."""
+    ids = set()
+    for p in (stripe_state.get('payments') or []):
+        if not isinstance(p, dict):
+            continue
+        pid = norm(p.get('id'))
+        if not pid:
+            continue
+        amt = p.get('amount') or 0
+        ar = p.get('amount_refunded') or 0
+        try:
+            fully = (float(amt) > 0 and float(ar) >= float(amt))
+        except (TypeError, ValueError):
+            fully = False
+        if p.get('refunded') is True or fully:
+            ids.add(pid)
+    for r in (stripe_state.get('refunds') or []):
+        if isinstance(r, dict) and r.get('charge'):
+            ids.add(norm(r.get('charge')))
+    return ids
+
+
+def _cap_id_set(stripe_state):
+    """Subscription ids flagged cancel_at_period_end (an EOP cancel specifically;
+    an immediate cancel sets status='canceled' and leaves this flag False)."""
+    ids = set()
+    for s in (stripe_state.get('subscriptions') or []):
+        if not isinstance(s, dict):
+            continue
+        if s.get('cancel_at_period_end') is True and norm(s.get('status')) != 'canceled':
+            ids.add(norm(s.get('id')))
+    return ids
+
+
+def _paid_invoice_id_set(qb_state):
+    ids = set()
+    for inv in (qb_state.get('invoices') or []):
+        if isinstance(inv, dict) and norm(inv.get('status')) == 'paid':
+            ids.add(norm(inv.get('id')))
+    return ids
+
+
+def _raw_sheets(state):
+    """Raw sheet dicts (list) — prefer the pre-materialization _ui_workbook copy."""
+    if not isinstance(state, dict):
+        return []
+    ui = state.get('_ui_workbook')
+    if isinstance(ui, dict) and isinstance(ui.get('sheets'), list):
+        return ui['sheets']
+    sheets = state.get('sheets')
+    if isinstance(sheets, list):
+        return sheets
+    return []
+
+
+def _sheet_name_set(state):
+    out = set()
+    for sh in _raw_sheets(state):
+        if isinstance(sh, dict) and sh.get('name'):
+            out.add(norm(sh.get('name')))
+    # also fold in materialized-dict keys (belt & suspenders)
+    sheets = state.get('sheets') if isinstance(state, dict) else None
+    if isinstance(sheets, dict):
+        for k in sheets:
+            out.add(norm(k))
+    return out
+
+
+def _sheet_has_numeric_cell(state, tab_name):
+    for sh in _raw_sheets(state):
+        if not isinstance(sh, dict) or norm(sh.get('name')) != norm(tab_name):
+            continue
+        data = sh.get('data') if isinstance(sh.get('data'), dict) else {}
+        for cell in data.values():
+            if isinstance(cell, dict):
+                for k in ('computed', 'value', 'formula'):
+                    v = cell.get(k)
+                    if isinstance(v, bool):
+                        continue
+                    if isinstance(v, (int, float)):
+                        return True
+                    if isinstance(v, str):
+                        s = v.replace('$', '').replace(',', '').strip()
+                        if s and re.fullmatch(r'-?\d+(\.\d+)?', s):
+                            return True
+            elif isinstance(cell, (int, float)) and not isinstance(cell, bool):
+                return True
+    return False
+
+
+def reward(go):
+    W_REFUND = 0.30
+    W_CANCEL = 0.25
+    W_RECEIVE = 0.25
+    W_TAB = 0.10
+    W_POST = 0.05
+    W_PIN = 0.05
+
+    # ---- 1) Answer key from google_sheets initial_state._task_adapter ----
+    gs = go('google_sheets')
+    gs_init = gs.get('initial_state', {}) if isinstance(gs.get('initial_state'), dict) else {}
+    gs_cur = gs.get('current_state', {}) if isinstance(gs.get('current_state'), dict) else {}
+    adapter = gs_init.get('_task_adapter', {}) if isinstance(gs_init.get('_task_adapter'), dict) else {}
+
+    expected = adapter.get('expected') if isinstance(adapter.get('expected'), list) else []
+    summary_tab_name = adapter.get('summary_tab_name') or 'Reconciliation Summary'
+
+    if not expected:
+        print('DEBUG_cc91abcd fatal=no_expected_worklist total=0.0')
+        return 0.0
+
+    refund_true = {norm(e.get('stripe_payment_id')) for e in expected
+                   if isinstance(e, dict) and e.get('action') == 'refund'
+                   and e.get('requires_action') and e.get('stripe_payment_id')}
+    cancel_true = {norm(e.get('stripe_subscription_id')) for e in expected
+                   if isinstance(e, dict) and e.get('action') == 'cancel_eop'
+                   and e.get('requires_action') and e.get('stripe_subscription_id')}
+    receive_true = {norm(e.get('qb_invoice_id')) for e in expected
+                    if isinstance(e, dict) and e.get('action') == 'receive_payment'
+                    and e.get('requires_action') and e.get('qb_invoice_id')}
+
+    # ---- 2) Stripe: refund set + cancel-at-period-end set (diff vs injection) ----
+    stripe = go('stripe_dashboard')
+    st_init = stripe.get('initial_state', {}) if isinstance(stripe.get('initial_state'), dict) else {}
+    st_cur = stripe.get('current_state', {}) if isinstance(stripe.get('current_state'), dict) else {}
+
+    refunded_init = _refunded_id_set(st_init)
+    refunded_now = _refunded_id_set(st_cur)
+    refund_pred = refunded_now - refunded_init            # newly-refunded (diff)
+
+    cap_init = _cap_id_set(st_init)
+    cap_now = _cap_id_set(st_cur)
+    cancel_pred = cap_now - cap_init                      # newly flagged EOP-cancel
+
+    # ---- 3) QuickBooks: receive-payment set (diff vs injection) ----
+    qb = go('quickbooks')
+    qb_init = qb.get('initial_state', {}) if isinstance(qb.get('initial_state'), dict) else {}
+    qb_cur = qb.get('current_state', {}) if isinstance(qb.get('current_state'), dict) else {}
+    paid_init = _paid_invoice_id_set(qb_init)
+    paid_now = _paid_invoice_id_set(qb_cur)
+    receive_pred = paid_now - paid_init                   # newly-paid invoices
+
+    # F1 components (precision punishes leak; gate on a non-empty true set so a
+    # malformed/empty key can never hand out free credit at do-nothing).
+    s_refund = W_REFUND * (f1(refund_true, refund_pred) if refund_true else 0.0)
+    s_cancel = W_CANCEL * (f1(cancel_true, cancel_pred) if cancel_true else 0.0)
+    s_receive = W_RECEIVE * (f1(receive_true, receive_pred) if receive_true else 0.0)
+
+    # ---- 4) Reconciliation Summary tab: NEW named tab with >=1 numeric cell ----
+    names_init = _sheet_name_set(gs_init)
+    names_now = _sheet_name_set(gs_cur)
+    tab_is_new = norm(summary_tab_name) in names_now and norm(summary_tab_name) not in names_init
+    tab_has_numeric = _sheet_has_numeric_cell(gs_cur, summary_tab_name)
+    tab_ok = bool(tab_is_new and tab_has_numeric)
+    s_tab = W_TAB * (1.0 if tab_ok else 0.0)
+
+    # ---- 5) Slack #finance: post (non-empty) + pin (references a real message) ----
+    slack = go('slack')
+    sl_init = slack.get('initial_state', {}) if isinstance(slack.get('initial_state'), dict) else {}
+    sl_cur = slack.get('current_state', {}) if isinstance(slack.get('current_state'), dict) else {}
+    s_adapter = sl_init.get('_task_adapter', {}) if isinstance(sl_init.get('_task_adapter'), dict) else {}
+    summary_channel = s_adapter.get('summary_channel') or 'finance'
+
+    fin_msgs = _slack_channel_messages(sl_cur, summary_channel)
+    _fin_kw = ('reconcil', 'resolved', 'payment', 'invoice', 'exception', 'refund')
+    def _is_fin_summary(m):
+        t = (_msg_text(m) or '').lower()
+        return bool(t.strip()) and any(k in t for k in _fin_kw) and any(c.isdigit() for c in t)
+    post_ok = any(_is_fin_summary(m) for m in fin_msgs if isinstance(m, dict))
+    s_post = W_POST * (1.0 if post_ok else 0.0)
+
+    # Pin: at least one pinned id in #finance that matches a real posted message.
+    msg_ids = set()
+    for m in fin_msgs:
+        if isinstance(m, dict):
+            mid = m.get('messageId') or m.get('id')
+            if mid is not None:
+                msg_ids.add(norm(mid))
+    pinned = []
+    for ch in (sl_cur.get('channels') or []):
+        if isinstance(ch, dict) and norm(ch.get('name')) == norm(summary_channel):
+            pinned = ch.get('pinnedMessages') or []
+            break
+    pinned_ids = set()
+    for pm in pinned:
+        if isinstance(pm, dict):
+            pid = pm.get('messageId') or pm.get('id')
+        else:
+            pid = pm
+        if pid is not None:
+            pinned_ids.add(norm(pid))
+    pin_ok = bool(post_ok and (pinned_ids & msg_ids))
+    s_pin = W_PIN * (1.0 if pin_ok else 0.0)
+
+    score = s_refund + s_cancel + s_receive + s_tab + s_post + s_pin
+
+    print(
+        'DEBUG_cc91abcd '
+        f'refund_true={len(refund_true)} refund_pred={len(refund_pred)} '
+        f'refund_f1={round(f1(refund_true, refund_pred) if refund_true else 0.0, 4)} '
+        f'cancel_true={len(cancel_true)} cancel_pred={len(cancel_pred)} '
+        f'cancel_f1={round(f1(cancel_true, cancel_pred) if cancel_true else 0.0, 4)} '
+        f'receive_true={len(receive_true)} receive_pred={len(receive_pred)} '
+        f'receive_f1={round(f1(receive_true, receive_pred) if receive_true else 0.0, 4)} '
+        f'tab_new={int(tab_is_new)} tab_numeric={int(tab_has_numeric)} '
+        f'post_ok={int(post_ok)} pin_ok={int(pin_ok)} '
+        f'w_refund={round(s_refund, 4)} w_cancel={round(s_cancel, 4)} w_receive={round(s_receive, 4)} '
+        f'w_tab={round(s_tab, 4)} w_post={round(s_post, 4)} w_pin={round(s_pin, 4)} '
+        f'total={round(score, 4)}'
+    )
+
+    return clamp01(score)
+
+
+try:
+    score = float(reward(go))
+except Exception as exc:
+    print(f'ERROR: reward() raised {exc!r}')
+    score = 0.0
+
+score = clamp01(score)
+print(f'\nREWARD: {round(score, 4)}')

ar_deal_to_invoice_007__long/_cua_gym_vm_bridge.sh

@@ -0,0 +1,106 @@
+#!/bin/bash
+# CUA-Gym VM-side network fix (direct-IP mock access).
+#
+# Runs INSIDE the QEMU guest. The CUA-Gym mock services are now reached by
+# their real IDC address (http://28.7.184.198:81xx) instead of the old
+# localhost:9000-9097 socat bridge, so the VM only needs two fixes:
+#
+#   1. Working DNS — the qcow2-baked /etc/resolv.conf points at the Pod's
+#      cluster DNS, which is unreachable from QEMU SLIRP NAT
+#      ("Temporary failure in name resolution").
+#   2. The mock host excluded from the woa http proxy — otherwise the VM's
+#      Chrome (gsettings proxy) and initial_setup.py's `requests` get a 403
+#      from the proxy when hitting 28.7.184.198:81xx. (The old socat path never
+#      hit this because localhost is bypassed by default.)
+#
+# Idempotent. Scoped to the CUA-Gym bridge step; does not touch the generic
+# OSWorld setup pipeline.
+#
+# Argv:
+#   $1  optional sudo password (default "password"; SetupController
+#       substitutes {CLIENT_PASSWORD} when uploaded as a cache file).
+#   $2  unused (kept for backward compat — was socat first port 9000).
+#   $3  unused (kept for backward compat — was socat last  port 9097).
+#   $4  optional mock host IP (default 28.7.184.198).
+
+set +e
+
+PASS="${1:-password}"
+MOCK_HOST="${4:-28.7.184.198}"
+# Derive the /16 IDC subnet so future port/host tweaks on the same segment
+# stay bypassed too (28.7.184.198 -> 28.7.0.0/16).
+MOCK_NET="$(echo "${MOCK_HOST}" | awk -F. 'NF==4{print $1"."$2".0.0/16"}')"
+
+run_root() {
+    echo "$PASS" | sudo -S -p "" bash -c "$1"
+}
+
+# Detect SLIRP gateway from VM's routing table; fall back to 10.0.2.2.
+GW="$(ip route 2>/dev/null | awk '/default/ {print $3; exit}')"
+GW="${GW:-10.0.2.2}"
+
+# ─── CUA-Gym network fix (Taiji-specific) ─────────────────────────────
+fix_vm_network() {
+    # 0) Discover VM network state at runtime — gateway, own IP/CIDR, SLIRP
+    #    DNS forwarder. Nothing hardcoded except the mock host above.
+    local GW_NET="${GW%.*}"
+    local SLIRP_DNS="${GW_NET}.3"        # QEMU SLIRP DNS proxy is always .3
+    local OWN_CIDR
+    OWN_CIDR="$(ip -4 -o addr show scope global 2>/dev/null \
+                | awk '{print $4; exit}')"
+    local OWN_IP="${OWN_CIDR%/*}"
+    local OWN_NET=""
+    [ -n "${OWN_IP}" ] && OWN_NET="${OWN_IP%.*}.0/24"
+
+    # 1) DNS — point at SLIRP DNS proxy + public fallbacks. chattr +i keeps
+    #    systemd-resolved/NetworkManager from clobbering it.
+    if run_root "printf 'nameserver ${SLIRP_DNS}\\nnameserver 114.114.114.114\\nnameserver 8.8.8.8\\n' > /etc/resolv.conf
+                 chattr +i /etc/resolv.conf 2>/dev/null || true"; then
+        echo "[cua_gym_vm_bridge] DNS set: ${SLIRP_DNS}, 114.114.114.114, 8.8.8.8 (gw=${GW})"
+    else
+        echo "[cua_gym_vm_bridge] WARN: failed to write /etc/resolv.conf"
+    fi
+
+    # 2) Build the proxy bypass list from detected values + the direct-IP
+    #    mock host (and its /16 subnet). THIS is what makes 28.7.184.198:81xx
+    #    reachable from inside the VM — without it the woa proxy 403s it.
+    local ignore_list="['localhost', '127.0.0.0/8', '${GW}'"
+    [ -n "${OWN_IP}" ]    && ignore_list="${ignore_list}, '${OWN_IP}'"
+    [ -n "${OWN_NET}" ]   && ignore_list="${ignore_list}, '${OWN_NET}'"
+    [ -n "${GW_NET}" ]    && ignore_list="${ignore_list}, '${GW_NET}.0/24'"
+    [ -n "${MOCK_HOST}" ] && ignore_list="${ignore_list}, '${MOCK_HOST}'"
+    [ -n "${MOCK_NET}" ]  && ignore_list="${ignore_list}, '${MOCK_NET}'"
+    ignore_list="${ignore_list}]"
+
+    # 3) GNOME proxy ignore-hosts — Chrome bypasses the woa proxy for these.
+    if gsettings set org.gnome.system.proxy ignore-hosts "${ignore_list}" 2>/tmp/gsettings_err; then
+        echo "[cua_gym_vm_bridge] gsettings ignore-hosts = ${ignore_list}"
+    else
+        echo "[cua_gym_vm_bridge] WARN: gsettings ignore-hosts failed: $(cat /tmp/gsettings_err 2>/dev/null)"
+    fi
+
+    # 4) Shell env — comma-separated form for no_proxy/NO_PROXY (requests/curl).
+    local NP="localhost,127.0.0.1,${GW}"
+    [ -n "${OWN_IP}" ]    && NP="${NP},${OWN_IP}"
+    [ -n "${OWN_NET}" ]   && NP="${NP},${OWN_NET}"
+    [ -n "${GW_NET}" ]    && NP="${NP},${GW_NET}.0/24"
+    [ -n "${MOCK_HOST}" ] && NP="${NP},${MOCK_HOST}"
+    [ -n "${MOCK_NET}" ]  && NP="${NP},${MOCK_NET}"
+
+    for var in no_proxy NO_PROXY; do
+        if ! grep -q "export ${var}=${NP}" /home/user/.bashrc 2>/dev/null; then
+            echo "export ${var}=${NP}" >> /home/user/.bashrc
+        fi
+    done
+    run_root "grep -q '^no_proxy=' /etc/environment 2>/dev/null \
+              && sed -i 's|^no_proxy=.*|no_proxy=\"${NP}\"|' /etc/environment \
+              || echo 'no_proxy=\"${NP}\"' >> /etc/environment
+              grep -q '^NO_PROXY=' /etc/environment 2>/dev/null \
+              && sed -i 's|^NO_PROXY=.*|NO_PROXY=\"${NP}\"|' /etc/environment \
+              || echo 'NO_PROXY=\"${NP}\"' >> /etc/environment" >/dev/null 2>&1
+    echo "[cua_gym_vm_bridge] no_proxy=${NP}"
+}
+
+fix_vm_network
+
+echo "[cua_gym_vm_bridge] direct-IP mode: ${MOCK_HOST} (+${MOCK_NET}) bypassed from woa proxy; no socat forwarding needed"

ar_deal_to_invoice_007__long/initial_setup.py

@@ -0,0 +1,431 @@
+"""
+Initial Setup: X2 — Won deal to invoice to notification
+Source: generated from CUA-Gym-Hub/task_benchmark/tasks/lc_x2.py
+Variant: eval
+Mocks: hubspot_mock,quickbooks_mock,gmail_mock
+"""
+import os
+import shlex
+import subprocess
+import time
+import uuid
+
+import requests
+
+
+sid = str(uuid.uuid4())
+with open('/tmp/task_web_sid', 'w', encoding='utf-8') as f:
+    f.write(sid)
+print(f'Generated sid: {sid}')
+
+APP_STATES = [('http://28.7.184.198:8150',
+  {'companies': [{'id': 'comp1', 'name': 'Acme', 'domain': 'acme.com', 'industry': 'Technology', 'phone': '+1 (555) 100-0001', 'city': 'San Francisco', 'state': 'CA', 'country': 'United States', 'numberOfEmployees': 250, 'annualRevenue': 15000000, 'lifecycleStage': 'customer', 'owner': 'Admin User', 'description': 'Enterprise SaaS platform', 'createDate': '2024-01-10T09:00:00Z'},
+                 {'id': 'comp2', 'name': 'Globex', 'domain': 'globex.com', 'industry': 'Manufacturing', 'phone': '+1 (555) 100-0002', 'city': 'New York', 'state': 'NY', 'country': 'United States', 'numberOfEmployees': 500, 'annualRevenue': 32000000, 'lifecycleStage': 'customer', 'owner': 'Admin User', 'description': 'Industrial parts supplier', 'createDate': '2024-01-12T09:00:00Z'},
+                 {'id': 'comp3', 'name': 'Initech', 'domain': 'initech.com', 'industry': 'Technology', 'phone': '+1 (555) 100-0003', 'city': 'Austin', 'state': 'TX', 'country': 'United States', 'numberOfEmployees': 120, 'annualRevenue': 8000000, 'lifecycleStage': 'customer', 'owner': 'Admin User', 'description': 'Office software vendor', 'createDate': '2024-01-14T09:00:00Z'},
+                 {'id': 'comp4', 'name': 'Umbrella', 'domain': 'umbrella.com', 'industry': 'Healthcare', 'phone': '+1 (555) 100-0004', 'city': 'Boston', 'state': 'MA', 'country': 'United States', 'numberOfEmployees': 800, 'annualRevenue': 55000000, 'lifecycleStage': 'customer', 'owner': 'Admin User', 'description': 'Healthcare research', 'createDate': '2024-01-16T09:00:00Z'},
+                 {'id': 'comp5', 'name': 'Wayne', 'domain': 'wayne.com', 'industry': 'Finance', 'phone': '+1 (555) 100-0005', 'city': 'Chicago', 'state': 'IL', 'country': 'United States', 'numberOfEmployees': 1200, 'annualRevenue': 90000000, 'lifecycleStage': 'customer', 'owner': 'Admin User', 'description': 'Financial holding group', 'createDate': '2024-01-18T09:00:00Z'},
+                 {'id': 'comp6', 'name': 'Stale Co', 'domain': 'stale.com', 'industry': 'Other', 'phone': '+1 (555) 100-0006', 'city': 'Seattle', 'state': 'WA', 'country': 'United States', 'numberOfEmployees': 90, 'annualRevenue': 5000000, 'lifecycleStage': 'customer', 'owner': 'Admin User', 'description': 'Legacy customer (last quarter)', 'createDate': '2024-01-22T09:00:00Z'},
+                 {'id': 'comp7', 'name': 'Future Inc', 'domain': 'future.com', 'industry': 'Technology', 'phone': '+1 (555) 100-0007', 'city': 'Denver', 'state': 'CO', 'country': 'United States', 'numberOfEmployees': 60, 'annualRevenue': 3000000, 'lifecycleStage': 'opportunity', 'owner': 'Admin User', 'description': 'Q2 prospect', 'createDate': '2024-01-24T09:00:00Z'},
+                 {'id': 'comp8', 'name': 'LostCorp', 'domain': 'lostcorp.com', 'industry': 'Other', 'phone': '+1 (555) 100-0008', 'city': 'Miami', 'state': 'FL', 'country': 'United States', 'numberOfEmployees': 40, 'annualRevenue': 2000000, 'lifecycleStage': 'lead', 'owner': 'Admin User', 'description': 'Lost competitive bid', 'createDate': '2024-01-26T09:00:00Z'}],
+   'contacts': [{'id': 'c1', 'firstName': '', 'lastName': 'Acme', 'email': 'ap@acme.com', 'phone': '+1 (555) 100-0001', 'jobTitle': 'AP Manager', 'companyId': 'comp1', 'lifecycleStage': 'customer', 'leadStatus': 'connected', 'owner': 'Admin User', 'city': 'San Francisco', 'state': 'CA', 'country': 'United States', 'createDate': '2024-01-15T10:30:00Z', 'lastActivityDate': '2026-04-05T10:00:00Z', 'timeline': []},
+                {'id': 'c2', 'firstName': '', 'lastName': 'Globex', 'email': 'ap@globex.com', 'phone': '+1 (555) 100-0002', 'jobTitle': 'AP Manager', 'companyId': 'comp2', 'lifecycleStage': 'customer', 'leadStatus': 'connected', 'owner': 'Admin User', 'city': 'New York', 'state': 'NY', 'country': 'United States', 'createDate': '2024-01-15T10:30:00Z', 'lastActivityDate': '2026-04-08T10:00:00Z', 'timeline': []},
+                {'id': 'c3', 'firstName': '', 'lastName': 'Initech', 'email': 'ap@initech.com', 'phone': '+1 (555) 100-0003', 'jobTitle': 'AP Manager', 'companyId': 'comp3', 'lifecycleStage': 'customer', 'leadStatus': 'connected', 'owner': 'Admin User', 'city': 'Austin', 'state': 'TX', 'country': 'United States', 'createDate': '2024-01-15T10:30:00Z', 'lastActivityDate': '2026-04-12T10:00:00Z', 'timeline': []},
+                {'id': 'c4', 'firstName': '', 'lastName': 'Umbrella', 'email': 'ap@umbrella.com', 'phone': '+1 (555) 100-0004', 'jobTitle': 'AP Manager', 'companyId': 'comp4', 'lifecycleStage': 'customer', 'leadStatus': 'connected', 'owner': 'Admin User', 'city': 'Boston', 'state': 'MA', 'country': 'United States', 'createDate': '2024-01-15T10:30:00Z', 'lastActivityDate': '2026-04-18T10:00:00Z', 'timeline': []},
+                {'id': 'c5', 'firstName': '', 'lastName': 'Wayne', 'email': 'ap@wayne.com', 'phone': '+1 (555) 100-0005', 'jobTitle': 'AP Manager', 'companyId': 'comp5', 'lifecycleStage': 'customer', 'leadStatus': 'connected', 'owner': 'Admin User', 'city': 'Chicago', 'state': 'IL', 'country': 'United States', 'createDate': '2024-01-15T10:30:00Z', 'lastActivityDate': '2026-04-22T10:00:00Z', 'timeline': []},
+                {'id': 'c6', 'firstName': '', 'lastName': 'Stale Co', 'email': 'ap@stale.com', 'phone': '+1 (555) 100-0006', 'jobTitle': 'AP Manager', 'companyId': 'comp6', 'lifecycleStage': 'customer', 'leadStatus': 'connected', 'owner': 'Admin User', 'city': 'Seattle', 'state': 'WA', 'country': 'United States', 'createDate': '2024-01-15T10:30:00Z', 'lastActivityDate': '2026-03-25T10:00:00Z', 'timeline': []},
+                {'id': 'c7', 'firstName': '', 'lastName': 'Future Inc', 'email': 'ap@future.com', 'phone': '+1 (555) 100-0007', 'jobTitle': 'AP Manager', 'companyId': 'comp7', 'lifecycleStage': 'opportunity', 'leadStatus': 'connected', 'owner': 'Admin User', 'city': 'Denver', 'state': 'CO', 'country': 'United States', 'createDate': '2024-01-15T10:30:00Z', 'lastActivityDate': '2026-05-03T10:00:00Z', 'timeline': []},
+                {'id': 'c8', 'firstName': '', 'lastName': 'LostCorp', 'email': 'ap@lostcorp.com', 'phone': '+1 (555) 100-0008', 'jobTitle': 'AP Manager', 'companyId': 'comp8', 'lifecycleStage': 'lead', 'leadStatus': 'unqualified', 'owner': 'Admin User', 'city': 'Miami', 'state': 'FL', 'country': 'United States', 'createDate': '2024-01-15T10:30:00Z', 'lastActivityDate': '2026-04-15T10:00:00Z', 'timeline': []}],
+   # Each deal carries the schema fields the hubspot_mock front-end reads
+   # (name/stage/amount/closeDate/dealType/priority/owner/companyId/
+   # contactIds/probability/description). The rewritten reward.py reads the
+   # same schema (companies[id==deal.companyId].name, contacts[id==
+   # contactIds[0]].email, deal.description) -- there are NO legacy
+   # tag-along fields any more.
+   #
+   # Single-line-only design: every April Closed Won deal is invoiced with
+   # exactly one line item. The deal name follows the pattern
+   # "<Company> - <Product>" so the agent can derive the QuickBooks
+   # Product/Service name (the part after " - ") and the QB product catalog
+   # below seeds a matching product whose price equals deals[].amount.
+   #
+   # IMPORTANT: stage value is 'closedwon' (no underscore) because reward.py
+   # filters won deals via norm(d['stage']) == 'closedwon'. We add a matching
+   # 'closedwon' entry to dealStages below so the front-end pipeline shows
+   # those deals in their own labelled column.
+   'deals': [{'id': 'd1', 'name': 'Acme - Enterprise License', 'stage': 'closedwon', 'amount': 12000, 'closeDate': '2026-04-05', 'dealType': 'new_business', 'priority': 'high', 'owner': 'Admin User', 'companyId': 'comp1', 'contactIds': ['c1'], 'probability': 100, 'description': 'Annual enterprise license, 250 seats', 'createDate': '2026-03-01T10:00:00Z', 'lastActivityDate': '2026-04-05T14:00:00Z'},
+             {'id': 'd2', 'name': 'Globex - Onboarding Package', 'stage': 'closedwon', 'amount': 7000, 'closeDate': '2026-04-08', 'dealType': 'new_business', 'priority': 'high', 'owner': 'Admin User', 'companyId': 'comp2', 'contactIds': ['c2'], 'probability': 100, 'description': 'Onboarding bundle (setup + license)', 'createDate': '2026-03-01T10:00:00Z', 'lastActivityDate': '2026-04-08T14:00:00Z'},
+             {'id': 'd3', 'name': 'Initech - Annual Subscription', 'stage': 'closedwon', 'amount': 8000, 'closeDate': '2026-04-12', 'dealType': 'existing_business', 'priority': 'medium', 'owner': 'Admin User', 'companyId': 'comp3', 'contactIds': ['c3'], 'probability': 100, 'description': 'Standard annual plan', 'createDate': '2026-03-05T10:00:00Z', 'lastActivityDate': '2026-04-12T14:00:00Z'},
+             {'id': 'd4', 'name': 'Umbrella - Hybrid Plan', 'stage': 'closedwon', 'amount': 8000, 'closeDate': '2026-04-18', 'dealType': 'new_business', 'priority': 'high', 'owner': 'Admin User', 'companyId': 'comp4', 'contactIds': ['c4'], 'probability': 100, 'description': 'Hybrid plan (seats + support bundled)', 'createDate': '2026-03-08T10:00:00Z', 'lastActivityDate': '2026-04-18T14:00:00Z'},
+             {'id': 'd5', 'name': 'Wayne - Renewal', 'stage': 'closedwon', 'amount': 15000, 'closeDate': '2026-04-22', 'dealType': 'existing_business', 'priority': 'high', 'owner': 'Admin User', 'companyId': 'comp5', 'contactIds': ['c5'], 'probability': 100, 'description': 'Renewal of existing master contract', 'createDate': '2026-03-12T10:00:00Z', 'lastActivityDate': '2026-04-22T14:00:00Z'},
+             {'id': 'd6', 'name': 'Stale Co - Last Quarter', 'stage': 'closedwon', 'amount': 9000, 'closeDate': '2026-03-25', 'dealType': 'existing_business', 'priority': 'medium', 'owner': 'Admin User', 'companyId': 'comp6', 'contactIds': ['c6'], 'probability': 100, 'description': 'Q1 deal, closed in March 2026', 'createDate': '2026-02-20T10:00:00Z', 'lastActivityDate': '2026-03-25T14:00:00Z'},
+             {'id': 'd7', 'name': 'Future Inc - Q2 Pipeline', 'stage': 'contract_sent', 'amount': 11000, 'closeDate': '2026-05-03', 'dealType': 'new_business', 'priority': 'medium', 'owner': 'Admin User', 'companyId': 'comp7', 'contactIds': ['c7'], 'probability': 90, 'description': 'Q2 deal, awaiting signature', 'createDate': '2026-04-10T10:00:00Z', 'lastActivityDate': '2026-05-03T14:00:00Z'},
+             {'id': 'd8', 'name': 'LostCorp - Lost Bid', 'stage': 'closed_lost', 'amount': 7000, 'closeDate': '2026-04-15', 'dealType': 'new_business', 'priority': 'low', 'owner': 'Admin User', 'companyId': 'comp8', 'contactIds': ['c8'], 'probability': 0, 'description': 'Lost to competitor', 'createDate': '2026-02-25T10:00:00Z', 'lastActivityDate': '2026-04-15T14:00:00Z', 'closedLostReason': 'Lost to competitor pricing'}],
+   'tickets': [],
+   'tasks': [],
+   # Single-line-only design: no per-deal line-item notes are needed.
+   'notes': [],
+   'templates': [],
+   'meetings': [],
+   'forms': [],
+   # NOTE: 'closedwon' (no underscore) is intentional and matches the literal
+   # value used in reward.py (norm(d['stage']) == 'closedwon'). The default
+   # mockData.js dealStages uses 'closed_won' (with underscore); we override
+   # with this dict so the pipeline view actually has a column whose id ===
+   # 'closedwon' to render our deals into.
+   'dealStages': {'appointment_scheduled': {'id': 'appointment_scheduled', 'label': 'Appointment Scheduled', 'probability': 20, 'color': '#E5F4FF', 'order': 1},
+                  'qualified_to_buy': {'id': 'qualified_to_buy', 'label': 'Qualified to Buy', 'probability': 40, 'color': '#FFF0E6', 'order': 2},
+                  'presentation_scheduled': {'id': 'presentation_scheduled', 'label': 'Presentation Scheduled', 'probability': 60, 'color': '#FFF8E6', 'order': 3},
+                  'decision_maker_bought_in': {'id': 'decision_maker_bought_in', 'label': 'Decision Maker Bought-In', 'probability': 80, 'color': '#E8F5E9', 'order': 4},
+                  'contract_sent': {'id': 'contract_sent', 'label': 'Contract Sent', 'probability': 90, 'color': '#E6FFFA', 'order': 5},
+                  'closedwon': {'id': 'closedwon', 'label': 'Closed Won', 'probability': 100, 'color': '#E6FFEC', 'order': 6},
+                  'closed_lost': {'id': 'closed_lost', 'label': 'Closed Lost', 'probability': 0, 'color': '#FFE6E6', 'order': 7}},
+   'ticketStatuses': {'new': {'id': 'new', 'label': 'New', 'color': '#E5F4FF', 'order': 1},
+                      'waiting_on_contact': {'id': 'waiting_on_contact', 'label': 'Waiting on Contact', 'color': '#FFF8E6', 'order': 2},
+                      'waiting_on_us': {'id': 'waiting_on_us', 'label': 'Waiting on Us', 'color': '#FFF0E6', 'order': 3},
+                      'in_progress': {'id': 'in_progress', 'label': 'In Progress', 'color': '#E6FFFA', 'order': 4},
+                      'closed': {'id': 'closed', 'label': 'Closed', 'color': '#E6FFEC', 'order': 5}},
+   'appState': {'sidebarOpen': True, 'currentUser': {'name': 'Admin User', 'email': 'admin@example.com', 'avatar': None}}}),
+ # ---- QuickBooks ----
+ # Customer naming convention: B2B mapping. Each pre-seeded customer's `name`
+ # equals the corresponding HubSpot company's `name`, and `company` is set to
+ # the same value so the agent can satisfy the instruction's "use that name
+ # as both Name and Company" rule by selecting an existing record.
+ #
+ # Coverage of April 2026 Closed Won deals (d1..d5):
+ #   d1 Acme     -> qb_c1 (pre-seeded)
+ #   d2 Globex   -> qb_c2 (pre-seeded)
+ #   d3 Initech  -> qb_c3 (pre-seeded)
+ #   d4 Umbrella -> qb_c4 (pre-seeded)
+ #   d5 Wayne    -> NOT pre-seeded -> agent MUST create it first.
+ #
+ # Plus distractors (qb_c5 Hooli, qb_c6 Pied Piper) that are unrelated to any
+ # HubSpot deal -- they verify the agent picks customers by deal mapping
+ # rather than blindly choosing the first row. qb_c7 (Stale Co) is the
+ # historical customer for d6 (March 25 won deal, NOT to be invoiced this run);
+ # keeping it makes the seed look like a realistic ongoing books.
+ #
+ # `products` is pre-seeded with one Service per April Closed Won deal: each
+ # product's `name` equals the part of the corresponding deal's name after
+ # " - " (e.g. d1 "Acme - Enterprise License" -> product "Enterprise
+ # License"), and its `price` equals deals[].amount. This way, when the
+ # agent picks the product on the invoice line the form auto-fills the Rate
+ # to the correct amount with no manual override needed. Every product is
+ # `isTaxable: false` so the form's auto-tax stays at $0.
+ ('http://28.7.184.198:8172',
+  {'customers': [{'id': 'qb_c1', 'name': 'Acme', 'company': 'Acme',
+                  'email': 'ap@acme.com', 'phone': '+1 (555) 100-0001',
+                  'address': 'San Francisco, CA, United States',
+                  'balance': 0, 'isActive': True, 'createdAt': '2024-02-01',
+                  'notes': ''},
+                 # NOTE: qb_c2 (Globex) / qb_c3 (Initech) / qb_c4 (Umbrella)
+                 # are intentionally NOT pre-seeded. The agent must create
+                 # these customers via Sales -> Customers -> "New customer"
+                 # before invoicing d2/d3/d4 (same as d5 Wayne). Only d1
+                 # (Acme) starts with a pre-seeded customer in QB.
+                 # Distractors: legitimate-looking customers unrelated to any
+                 # HubSpot deal in this task. These should NOT receive any
+                 # invoice in this run.
+                 {'id': 'qb_c5', 'name': 'Hooli', 'company': 'Hooli',
+                  'email': 'billing@hooli.com', 'phone': '+1 (555) 200-0001',
+                  'address': 'Palo Alto, CA, United States',
+                  'balance': 1500, 'isActive': True, 'createdAt': '2023-09-15',
+                  'notes': 'Legacy account'},
+                 {'id': 'qb_c6', 'name': 'Pied Piper', 'company': 'Pied Piper',
+                  'email': 'ap@piedpiper.com', 'phone': '+1 (555) 200-0002',
+                  'address': 'San Francisco, CA, United States',
+                  'balance': 0, 'isActive': True, 'createdAt': '2023-11-02',
+                  'notes': ''},
+                 # Historical customer for d6 (March 25 won, OUT of April scope).
+                 {'id': 'qb_c7', 'name': 'Stale Co', 'company': 'Stale Co',
+                  'email': 'ap@stale.com', 'phone': '+1 (555) 100-0006',
+                  'address': 'Seattle, WA, United States',
+                  'balance': 0, 'isActive': True, 'createdAt': '2024-02-01',
+                  'notes': 'Q1 customer; already invoiced in March.'}],
+   # One Service per April Closed Won deal. price == deals[].amount so that
+   # selecting the product on the invoice line auto-fills the Rate correctly.
+   'products': [{'id': 'qb_p1', 'name': 'Enterprise License', 'description': 'Annual enterprise license', 'type': 'Service',
+                 'price': 12000, 'cost': 0, 'category': 'Services', 'sku': '',
+                 'isActive': True, 'isTaxable': False, 'quantityOnHand': None},
+                {'id': 'qb_p2', 'name': 'Onboarding Package', 'description': 'Onboarding bundle (setup + license)', 'type': 'Service',
+                 'price': 7000, 'cost': 0, 'category': 'Services', 'sku': '',
+                 'isActive': True, 'isTaxable': False, 'quantityOnHand': None},
+                {'id': 'qb_p3', 'name': 'Annual Subscription', 'description': 'Standard annual plan', 'type': 'Service',
+                 'price': 8000, 'cost': 0, 'category': 'Services', 'sku': '',
+                 'isActive': True, 'isTaxable': False, 'quantityOnHand': None},
+                {'id': 'qb_p4', 'name': 'Hybrid Plan', 'description': 'Hybrid plan (seats + support bundled)', 'type': 'Service',
+                 'price': 8000, 'cost': 0, 'category': 'Services', 'sku': '',
+                 'isActive': True, 'isTaxable': False, 'quantityOnHand': None},
+                {'id': 'qb_p5', 'name': 'Renewal', 'description': 'Renewal of existing master contract', 'type': 'Service',
+                 'price': 15000, 'cost': 0, 'category': 'Services', 'sku': '',
+                 'isActive': True, 'isTaxable': False, 'quantityOnHand': None}],
+   'invoices': []}),
+ ('http://28.7.184.198:8138', {'emails': []})]
+
+
+for app_url, state in APP_STATES:
+    resp = requests.post(
+        f'{app_url}/post?sid={sid}',
+        json={'action': 'set', 'state': state},
+        timeout=30,
+    )
+    assert resp.status_code == 200, (
+        f'State injection failed for {app_url}: {resp.text}'
+    )
+    print(f'State injected: {app_url} sid={sid}')
+
+for app_url, _ in APP_STATES:
+    go_payload = requests.get(f'{app_url}/go?sid={sid}', timeout=15).json()
+    assert go_payload.get('initial_state') is not None, (
+        f'initial_state is None after injection for {app_url}'
+    )
+    print(f'Verified initial_state set: {app_url}')
+
+
+# PATCHED_BY: patch_wait_mocks_loaded.py
+def launch_gui(command, delay_sec=1.0):
+    """Launch a GUI command, fully detached from this script.
+
+    Inject --remote-debugging-port=1337 (and Chrome quiet flags) so that
+    wait_mocks_loaded's CDP phase can verify browser render below.
+    Idempotent: if the caller already specified a debug port, keep it.
+    """
+    env = os.environ.copy()
+    env['DISPLAY'] = ':0'
+    if isinstance(command, str):
+        _parts = shlex.split(command)
+    else:
+        _parts = list(command)
+    if (_parts and 'google-chrome' in _parts[0]
+            and not any(p.startswith('--remote-debugging-port') for p in _parts)):
+        _parts[1:1] = [
+            '--no-first-run',
+            '--no-default-browser-check',
+            '--remote-debugging-port=1337',
+        ]
+    subprocess.Popen(
+        _parts,
+        stdin=subprocess.DEVNULL,
+        stdout=subprocess.DEVNULL,
+        stderr=subprocess.DEVNULL,
+        env=env,
+        start_new_session=True,
+    )
+    time.sleep(delay_sec)
+
+
+def wait_mocks_loaded(app_states, sid_value, timeout=90, render_buffer=12):
+    """2-phase readiness wait, mirrors cache/ae_pipeline_hygiene_004 style.
+
+    Phase 1 - Backend readiness:
+        Poll each mock's /state?sid=<sid> (preferred) or /go?sid=<sid>
+        (fallback) until it returns 200 with 'stored_state' or
+        'initial_state' respectively, or until timeout.
+
+    Phase 2 - Browser render verification (via CDP + Playwright):
+        Connect to http://localhost:1337 (the --remote-debugging-port
+        injected by launch_gui above), wait for every open tab to reach
+        'networkidle' AND to contain at least 20 DOM elements. This
+        guarantees the SPA fetched /go?sid= data AND actually painted,
+        not just that the mock backend is up.
+
+    Fallback:
+        If Playwright is missing or CDP cannot be reached, sleep for
+        `render_buffer` seconds as a best-effort delay so the screenshot
+        loop does not start on a still-blank tab.
+    """
+    import urllib.request as _urlreq
+    import time as _t
+    _urls = [p for p, _ in app_states]
+    _opener = _urlreq.build_opener(_urlreq.ProxyHandler({}))
+
+    # ---- Phase 1: backend readiness ----
+    _deadline = _t.time() + timeout
+    _todo = list(_urls)
+    while _todo and _t.time() < _deadline:
+        _rest = []
+        for _u in _todo:
+            _ok = False
+            for _path, _need in (('/state?sid=', b'stored_state'),
+                                 ('/go?sid=',    b'initial_state')):
+                try:
+                    _r = _opener.open(f"{_u}{_path}{sid_value}", timeout=5)
+                    if _r.status == 200:
+                        _body = _r.read()
+                        if _need in _body:
+                            _ok = True
+                            break
+                except Exception:
+                    pass
+            if _ok:
+                print(f"  {_u}: backend ready")
+                continue
+            _rest.append(_u)
+        _todo = _rest
+        if _todo:
+            _t.sleep(2)
+    _done = len(_urls) - len(_todo)
+    _pend = f" (pending: {_todo})" if _todo else ""
+    print(f"  mocks backend ready: {_done}/{len(_urls)}{_pend}")
+
+    # ---- Phase 2: browser render verification via CDP ----
+    _cdp_ok = False
+    try:
+        from playwright.sync_api import sync_playwright
+        _cdp_dl = _t.time() + max(15.0, _deadline - _t.time())
+        with sync_playwright() as _p:
+            _br = None
+            while _t.time() < _cdp_dl:
+                try:
+                    _br = _p.chromium.connect_over_cdp('http://localhost:1337')
+                    break
+                except Exception:
+                    _t.sleep(0.5)
+            if _br is None:
+                print("  wait_mocks_loaded: CDP connect to :1337 failed; is Chrome up with --remote-debugging-port?")
+            else:
+                _n_expected = len(_urls)
+                _pages = []
+                _pg_dl = _t.time() + 15
+                _last_count = -1
+                _stable_since = None
+                while _t.time() < _pg_dl:
+                    _pages = []
+                    for _c in _br.contexts:
+                        _pages.extend(_c.pages)
+                    _cnt = len(_pages)
+                    if _n_expected > 0 and _cnt >= _n_expected:
+                        break
+                    if _cnt > 0 and _cnt == _last_count:
+                        if _stable_since is None:
+                            _stable_since = _t.time()
+                        elif _t.time() - _stable_since >= 1.5:
+                            break
+                    else:
+                        _stable_since = None
+                    _last_count = _cnt
+                    _t.sleep(0.5)
+                _pages = []
+                for _c in _br.contexts:
+                    _pages.extend(_c.pages)
+                if not _pages:
+                    print("  wait_mocks_loaded: CDP connected but no open tabs found")
+                else:
+                    print(f"  wait_mocks_loaded: {len(_pages)} tab(s) open (expected {_n_expected}); polling render")
+                    _MIN_ELEMS = 20
+                    _verify_dl = _t.time() + 40.0   # bounded so total setup stays < VM 120s cap
+                    for _pg in _pages:              # bring each to front once to defeat background-tab render throttling
+                        try:
+                            _pg.bring_to_front()
+                        except Exception:
+                            pass
+                    _pending = list(_pages)
+                    while _pending and _t.time() < _verify_dl:
+                        _still = []
+                        for _pg in _pending:
+                            _ttl = ""
+                            try:
+                                _ttl = _pg.title()
+                            except Exception:
+                                pass
+                            try:
+                                _n = _pg.evaluate("document.body ? document.querySelectorAll('*').length : 0")
+                            except Exception:
+                                _n = 0
+                            if _n >= _MIN_ELEMS:
+                                print(f"  wait_mocks_loaded: tab {_ttl!r} rendered ({_n} elements)")
+                            else:
+                                _still.append(_pg)
+                        _pending = _still
+                        if _pending:
+                            _t.sleep(2)
+                    if not _pending:
+                        _cdp_ok = True
+                    else:
+                        for _pg in _pending:
+                            _ttl = ""
+                            try:
+                                _ttl = _pg.title()
+                            except Exception:
+                                pass
+                            print(f"  wait_mocks_loaded: WARNING tab {_ttl!r} still not rendered after poll, will fall back")
+    except ImportError:
+        print("  wait_mocks_loaded: playwright not installed; skipping CDP phase")
+    except Exception as _e:
+        print(f"  wait_mocks_loaded: CDP phase error ({_e!r})")
+
+    if not _cdp_ok:
+        print(f"  wait_mocks_loaded: falling back to render_buffer={render_buffer}s sleep")
+        _t.sleep(render_buffer)
+    else:
+        print("  wait_mocks_loaded: browser render verified, skipping render_buffer")
+
+
+def _open_mock_tabs(app_states, sid_value):
+    """Open the first mock as the primary Chrome window, the rest as new tabs.
+
+    Uses launch_gui once per URL so that the macOS localizer regex-extract
+    (which grabs the FIRST quoted URL per call) still works correctly.
+    """
+    _urls = [p for p, _ in app_states]
+    if not _urls:
+        return
+    launch_gui(f'google-chrome "{_urls[0]}/?sid={sid_value}"', delay_sec=2.0)
+    for _u in _urls[1:]:
+        launch_gui(f'google-chrome --new-tab "{_u}/?sid={sid_value}"', delay_sec=1.0)
+
+
+
+# PATCHED_BY: add_wait_for_mocks_loaded (colleague headless+playwright render gate)
+def wait_for_mocks_loaded(mocks, sid, timeout=60.0, render_timeout=15.0):
+    """Block until every mock serves injected state AND (best-effort) renders."""
+    import time as _t
+    deadline = _t.time() + timeout
+    pending = dict(mocks)
+    while pending and _t.time() < deadline:
+        for name in list(pending):
+            try:
+                r = requests.get(f'{pending[name]}/go?sid={sid}', timeout=5)
+                if r.status_code == 200 and r.json().get('initial_state') is not None:
+                    del pending[name]
+            except Exception:
+                pass
+        if pending:
+            _t.sleep(1.0)
+    if pending:
+        print(f'WARN: mocks not ready within {timeout}s: {list(pending)}')
+    try:
+        from playwright.sync_api import sync_playwright
+    except ImportError:
+        print('WARN: playwright not installed; skipping render check, settle 3s')
+        _t.sleep(3.0)
+        return
+    with sync_playwright() as p:
+        try:
+            browser = p.chromium.launch(channel='chrome', headless=True,
+                                        args=['--no-sandbox', '--disable-gpu', '--disable-dev-shm-usage'])
+        except Exception as e:
+            print(f'WARN: could not launch chrome for render check: {e}; settle 3s')
+            _t.sleep(3.0)
+            return
+        try:
+            for name, url in mocks.items():
+                page = browser.new_page()
+                try:
+                    page.goto(f'{url}/?sid={sid}', wait_until='domcontentloaded', timeout=int(render_timeout * 1000))
+                    page.wait_for_function(
+                        "document.body && document.body.innerText.length > 50",
+                        timeout=int(render_timeout * 1000))
+                    print(f'[{name}] rendered OK')
+                except Exception as e:
+                    print(f'WARN: [{name}] render check issue: {e}')
+                finally:
+                    page.close()
+        finally:
+            browser.close()
+
+
+# ref pattern (out_benchmark_cache): open visible tabs FIRST, then block until mocks render
+_open_mock_tabs(APP_STATES, sid)
+wait_for_mocks_loaded({u.rsplit(':', 1)[-1]: u for u, _ in APP_STATES}, sid, timeout=60, render_timeout=15)
+wait_mocks_loaded(APP_STATES, sid, timeout=90, render_buffer=12)
+print(f'GUI_READY: launched browser tabs for {len(APP_STATES)} apps')

ar_deal_to_invoice_007__long/reward.py

@@ -0,0 +1,826 @@
+"""
+Reward Script: X2 — Won deal to invoice to notification
+Source: generated from CUA-Gym-Hub/task_benchmark/tasks/lc_x2.py
+Variant: eval
+Mocks: hubspot_mock,quickbooks_mock,gmail_mock
+"""
+import copy
+import re
+import sys
+
+import requests
+
+
+try:
+    with open('/tmp/task_web_sid', encoding='utf-8') as f:
+        sid = f.read().strip()
+    if not sid:
+        raise ValueError('sid is empty')
+except Exception as e:
+    print(f'CRITICAL: Cannot read sid from /tmp/task_web_sid: {e}')
+    print('REWARD: 0.0')
+    sys.exit(0)
+
+
+APP_URLS = {'hubspot': 'http://28.7.184.198:8150', 'quickbooks': 'http://28.7.184.198:8172', 'gmail': 'http://28.7.184.198:8138'}
+_GO_CACHE = {}
+
+
+def clamp01(x):
+    return max(0.0, min(1.0, float(x)))
+
+
+def frac(num, den):
+    return 0.0 if den == 0 else num / den
+
+
+def f1(true_set, pred_set):
+    if not true_set and not pred_set:
+        return 1.0
+    tp = len(true_set & pred_set)
+    if tp == 0:
+        return 0.0
+    prec = tp / len(pred_set)
+    rec = tp / len(true_set)
+    return 2 * prec * rec / (prec + rec)
+
+
+def by_id(items, key='id'):
+    return {it.get(key): it for it in (items or []) if isinstance(it, dict)}
+
+
+def has_text(haystack, *needles):
+    s = (haystack or '').lower()
+    return any(n.lower() in s for n in needles)
+
+
+def norm(s):
+    return (str(s) if s is not None else '').strip().lower()
+
+
+def _deal_amount(deal):
+    if not isinstance(deal, dict):
+        return 0.0
+    line_items = deal.get('lineItems') or []
+    if isinstance(line_items, list) and line_items:
+        total = 0.0
+        for item in line_items:
+            if isinstance(item, dict):
+                try:
+                    total += float(item.get('amount', 0) or 0)
+                except (TypeError, ValueError):
+                    continue
+        return total
+    try:
+        return float(deal.get('amount', 0) or 0)
+    except (TypeError, ValueError):
+        return 0.0
+
+def _invoice_customer_name(iv):
+    if not isinstance(iv, dict):
+        return ''
+
+    customer = iv.get('customer')
+    if isinstance(customer, dict):
+        return customer.get('name') or customer.get('displayName') or customer.get('customerName') or customer.get('fullName') or ''
+
+    if isinstance(customer, list):
+        for c in customer:
+            if isinstance(c, dict):
+                name = c.get('name') or c.get('displayName') or c.get('customerName') or c.get('fullName')
+                if name:
+                    return name
+            elif c:
+                return str(c)
+
+    return customer or iv.get('customerName') or iv.get('name') or iv.get('displayName') or iv.get('customerDisplayName') or ''
+
+
+def _invoice_total(iv):
+    if not isinstance(iv, dict):
+        return 0.0
+
+    for key in ('total', 'amount', 'invoiceTotal', 'balance', 'totalAmount', 'amountDue', 'amount_total'):
+        value = iv.get(key)
+        if isinstance(value, dict):
+            value = value.get('amount') or value.get('value')
+        try:
+            return float(value or 0)
+        except (TypeError, ValueError):
+            continue
+
+    line_items = iv.get('lineItems') or iv.get('items') or iv.get('lines')
+    if isinstance(line_items, list) and line_items:
+        total = 0.0
+        for item in line_items:
+            if not isinstance(item, dict):
+                continue
+            raw = item.get('amount')
+            if raw is None:
+                qty = item.get('quantity', 1)
+                unit = item.get('unitPrice') or item.get('rate') or item.get('price')
+                try:
+                    raw = float(qty or 0) * float(unit or 0)
+                except (TypeError, ValueError):
+                    raw = 0
+            try:
+                total += float(raw or 0)
+            except (TypeError, ValueError):
+                continue
+        return total
+
+    return 0.0
+
+
+def _notes_text(deal):
+    if not isinstance(deal, dict):
+        return ''
+
+    chunks = []
+    notes = deal.get('notes', [])
+    if isinstance(notes, list):
+        for x in notes:
+            if isinstance(x, dict):
+                chunks.append(str(x.get('body') or x.get('content') or x.get('text') or x.get('note') or x))
+            else:
+                chunks.append(str(x))
+    elif notes:
+        chunks.append(str(notes))
+
+    for key in ('activity', 'activities', 'timeline', 'comments'):
+        value = deal.get(key)
+        if isinstance(value, list):
+            for v in value:
+                if isinstance(v, dict):
+                    chunks.append(str(v.get('body') or v.get('content') or v.get('text') or v.get('note') or v))
+                else:
+                    chunks.append(str(v))
+        elif value:
+            chunks.append(str(value))
+
+    return ' '.join(chunks)
+
+
+def _email_text(email):
+    if not isinstance(email, dict):
+        return str(email)
+
+    to = email.get('to', '')
+    if isinstance(to, list):
+        to = ' '.join(
+            (x.get('name') or x.get('email') or str(x)) if isinstance(x, dict) else str(x)
+            for x in to
+        )
+    elif isinstance(to, dict):
+        to = to.get('name') or to.get('email') or str(to)
+
+    to_recips = email.get('toRecipients', [])
+    if isinstance(to_recips, list):
+        to_recips = ' '.join(
+            (x.get('name') or x.get('email') or str(x)) if isinstance(x, dict) else str(x)
+            for x in to_recips
+        )
+    else:
+        to_recips = ''
+
+    return f"{to} {to_recips} {email.get('subject', '')} {email.get('body', '')}"
+
+
+def _gmail_sent_emails(gmail_state):
+    out = []
+    if not isinstance(gmail_state, dict):
+        return out
+
+    for e in gmail_state.get('emails', []):
+        if not isinstance(e, dict):
+            continue
+        folder = norm(e.get('folder'))
+        if folder in ('sent', 'sentitems', 'sent items'):
+            out.append(e)
+            continue
+        if e.get('isSent') is True or e.get('sentAt') or e.get('sentDateTime'):
+            out.append(e)
+
+    if out:
+        return out
+
+    for key in ('sent', 'sentEmails', 'outbox'):
+        items = gmail_state.get(key, [])
+        if isinstance(items, list):
+            out.extend(x for x in items if isinstance(x, dict))
+    return out
+
+
+def _normalize_app_name(app):
+    app = (app or '').strip().lower()
+    return app[:-5] if app.endswith('_mock') else app
+
+
+def _parse_cell_id(cell_id):
+    if not isinstance(cell_id, str) or not cell_id:
+        return None
+    idx = 0
+    while idx < len(cell_id) and cell_id[idx].isalpha():
+        idx += 1
+    if idx == 0 or idx >= len(cell_id):
+        return None
+    col = 0
+    for ch in cell_id[:idx].upper():
+        col = col * 26 + (ord(ch) - ord('A') + 1)
+    try:
+        row = int(cell_id[idx:])
+    except ValueError:
+        return None
+    return row, col - 1
+
+
+def _extract_rows_from_workbook(state):
+    if not isinstance(state, dict):
+        return None
+    sheets = state.get('sheets')
+    if not isinstance(sheets, list):
+        return None
+
+    adapter = state.get('_task_adapter') if isinstance(state.get('_task_adapter'), dict) else {}
+    headers_by_sheet = adapter.get('headers_by_sheet') if isinstance(adapter.get('headers_by_sheet'), dict) else {}
+
+    out = {}
+    for sh in sheets:
+        if not isinstance(sh, dict):
+            continue
+        name = sh.get('name') or sh.get('id') or 'Sheet1'
+        data = sh.get('data') if isinstance(sh.get('data'), dict) else {}
+
+        headers = headers_by_sheet.get(name)
+        if not headers:
+            cols = []
+            for cid, cell in data.items():
+                rc = _parse_cell_id(cid)
+                if not rc:
+                    continue
+                row_idx, col_idx = rc
+                if row_idx != 1:
+                    continue
+                value = ''
+                if isinstance(cell, dict):
+                    value = cell.get('value', '')
+                elif cell is not None:
+                    value = str(cell)
+                if value is not None and str(value) != '':
+                    cols.append((col_idx, str(value)))
+            headers = [v for _, v in sorted(cols)]
+
+        if not headers:
+            continue
+
+        by_row = {}
+        for cid, cell in data.items():
+            rc = _parse_cell_id(cid)
+            if not rc:
+                continue
+            row_idx, col_idx = rc
+            if row_idx <= 1 or col_idx >= len(headers):
+                continue
+            if isinstance(cell, dict):
+                value = cell.get('value', '')
+            else:
+                value = cell
+            by_row.setdefault(row_idx, {})[headers[col_idx]] = value
+
+        rows = []
+        for row_idx in sorted(by_row.keys()):
+            row = by_row[row_idx]
+            if not any(str(row.get(h, '')).strip() for h in headers):
+                continue
+            rows.append({h: row.get(h, '') for h in headers})
+
+        out[name] = {'headers': headers, 'rows': rows}
+
+    if not out:
+        return None
+    return {'sheets': out}
+
+
+def _materialize_google_sheets(payload):
+    for key in ('initial_state', 'current_state'):
+        state = payload.get(key)
+        if not isinstance(state, dict):
+            continue
+        maybe_rows = _extract_rows_from_workbook(state)
+        if maybe_rows is not None:
+            state['_ui_workbook'] = {
+                'sheets': copy.deepcopy(state.get('sheets', [])),
+                'title': state.get('title'),
+            }
+            state['sheets'] = maybe_rows['sheets']
+
+
+def _materialize_slack(payload):
+    for key in ('initial_state', 'current_state'):
+        state = payload.get(key)
+        if not isinstance(state, dict):
+            continue
+        channels = state.get('channels') if isinstance(state.get('channels'), list) else []
+        messages_map = state.get('messages') if isinstance(state.get('messages'), dict) else {}
+        for ch in channels:
+            if not isinstance(ch, dict):
+                continue
+            cid = ch.get('channelId') or ch.get('id')
+            if not cid:
+                continue
+            msg_list = messages_map.get(cid) if isinstance(messages_map.get(cid), list) else []
+            channel_msgs = ch.get('messages') if isinstance(ch.get('messages'), list) else []
+            merged = {
+                (m.get('messageId') or m.get('id')): dict(m)
+                for m in channel_msgs
+                if isinstance(m, dict)
+            }
+            for m in msg_list:
+                if not isinstance(m, dict):
+                    continue
+                mid = m.get('messageId') or m.get('id')
+                text = m.get('text') if m.get('text') is not None else m.get('content', '')
+                merged[mid] = {
+                    'id': mid,
+                    'messageId': mid,
+                    'text': text,
+                    'content': m.get('content', text),
+                    'senderId': m.get('senderId'),
+                    'timestamp': m.get('timestamp'),
+                    'reactions': m.get('reactions') if isinstance(m.get('reactions'), list) else [],
+                }
+            ch['messages'] = list(merged.values())
+
+
+def _owner_name_map(state):
+    out = {}
+    users = state.get('users') if isinstance(state.get('users'), list) else []
+    for u in users:
+        if not isinstance(u, dict):
+            continue
+        uid = u.get('userId')
+        if not uid:
+            continue
+        name = (f"{u.get('firstName', '')} {u.get('lastName', '')}".strip() or u.get('name') or uid)
+        out[uid] = name
+    return out
+
+
+def _materialize_salesforce(payload):
+    for key in ('initial_state', 'current_state'):
+        state = payload.get(key)
+        if not isinstance(state, dict):
+            continue
+
+        adapter = state.get('_task_adapter') if isinstance(state.get('_task_adapter'), dict) else {}
+        owner_names = _owner_name_map(state)
+
+        task_leads = adapter.get('task_leads') if isinstance(adapter.get('task_leads'), dict) else None
+        if task_leads and isinstance(state.get('leads'), list):
+            restored = []
+            for lead in state['leads']:
+                if not isinstance(lead, dict):
+                    continue
+                lid = str(lead.get('leadId') or lead.get('id') or '')
+                meta = task_leads.get(lid, {})
+                restored.append({
+                    'id': meta.get('id') or lead.get('leadId') or lead.get('id'),
+                    'company': lead.get('company', meta.get('company', '')),
+                    'status': lead.get('status', 'New'),
+                    'ownerName': owner_names.get(lead.get('ownerId')) or meta.get('ownerName'),
+                    'annualRevenue': meta.get('annualRevenue', lead.get('revenue')),
+                })
+            state['leads'] = restored
+
+        task_accounts = adapter.get('task_accounts') if isinstance(adapter.get('task_accounts'), dict) else None
+        if task_accounts and isinstance(state.get('accounts'), list):
+            current_by_id = {}
+            for acc in state['accounts']:
+                if isinstance(acc, dict):
+                    current_by_id[str(acc.get('accountId') or acc.get('id') or '')] = acc
+
+            restored_accounts = []
+            for aid, meta in task_accounts.items():
+                cur = current_by_id.get(str(aid), {})
+                restored_accounts.append({
+                    'id': meta.get('id') or aid,
+                    'name': meta.get('name') or cur.get('name', ''),
+                    'contact': meta.get('contact'),
+                    'wonWithin7d': bool(meta.get('wonWithin7d', False)),
+                    'onboardingStage': cur.get('onboardingStage', meta.get('onboardingStage')),
+                })
+            state['accounts'] = restored_accounts
+
+
+def _materialize_google_docs(payload):
+    for key in ('initial_state', 'current_state'):
+        state = payload.get(key)
+        if not isinstance(state, dict):
+            continue
+        adapter = state.get('_task_adapter') if isinstance(state.get('_task_adapter'), dict) else {}
+        task_docs = adapter.get('task_documents') if isinstance(adapter.get('task_documents'), dict) else None
+        if not task_docs:
+            continue
+        docs_map = state.get('documents') if isinstance(state.get('documents'), dict) else {}
+        restored = []
+        for did, meta in task_docs.items():
+            cur = docs_map.get(did, {}) if isinstance(docs_map, dict) else {}
+            restored.append({
+                'id': meta.get('id') or did,
+                'title': cur.get('title', meta.get('title')),
+                'vendor': meta.get('vendor'),
+                'folder': meta.get('folder'),
+                'content': cur.get('content', meta.get('content', '')),
+            })
+        state['documents'] = restored
+
+
+def _adapt_payload_for_reward(app, payload):
+    app_name = _normalize_app_name(app)
+    if app_name == 'google_sheets':
+        _materialize_google_sheets(payload)
+    elif app_name == 'slack':
+        _materialize_slack(payload)
+    elif app_name == 'salesforce':
+        _materialize_salesforce(payload)
+    elif app_name == 'google_docs':
+        _materialize_google_docs(payload)
+    return payload
+
+
+def go(app_name):
+    if app_name in _GO_CACHE:
+        return _GO_CACHE[app_name]
+
+    base = APP_URLS.get(app_name)
+    if not base:
+        print(f'CRITICAL: unknown app {app_name!r} (no URL mapped)')
+        print('REWARD: 0.0')
+        sys.exit(0)
+
+    try:
+        resp = requests.get(f'{base}/go?sid={sid}', timeout=15)
+        resp.raise_for_status()
+        payload = resp.json()
+    except Exception as e:
+        print(f'CRITICAL: cannot fetch {app_name} go state: {e}')
+        print('REWARD: 0.0')
+        sys.exit(0)
+
+    payload = _adapt_payload_for_reward(app_name, payload)
+    _GO_CACHE[app_name] = payload
+    return payload
+
+
+# === Task-specific reward (rewritten for new schema, per-deal weighted) ===
+#
+# Scoring model (B-mixed, decided 2026-06-15):
+#   - Each April-Closed-Won deal contributes up to 1.0 to its own score.
+#   - 6 sub-checks per deal, weights sum to 1.0:
+#       A1 customer        0.15  invoice exists for this deal's company
+#       A2 line item       0.20  exactly 1 item with the right product/qty/rate/amount
+#                                (must also match customer name; A2 is self-contained
+#                                and re-checks customer to prevent cross-deal mix-up)
+#       A3 header          0.15  same invoice has terms/date/dueDate/tax/status correct
+#                                (must also match customer name; self-contained)
+#       B1 recipient       0.10  a sent gmail email addressed to the right contact
+#       B2 email body      0.15  email's subject + body match the template anchors,
+#                                with subject/body invoice number self-consistent.
+#                                Independent of A* (does NOT require a real invoice).
+#       C  description     0.25  deal.description contains "Invoiced INV-<digits>.".
+#                                Independent of A* (does NOT require a real invoice).
+#   - total_score = sum(deal_score(d) for d in won) / n
+#   - n is the count of April-Closed-Won deals in the initial state.
+#
+# Dependency chain (B-mixed):
+#   A1 -> A2 (re-checks customer) and A3 (re-checks customer): A group is integrated.
+#   B1 -> B2 (re-checks recipient): B group is internally chained.
+#   C is fully standalone (matches the literal pattern in description).
+#   A and B and C are independent of each other.
+
+import datetime as _dt
+import re as _re
+
+_INV_NUMBER_RE = _re.compile(r'Invoiced\s+INV-(\d+)\.', _re.IGNORECASE)
+
+
+def _april_won_deals(deals_initial):
+    out = []
+    for d in deals_initial or []:
+        if not isinstance(d, dict):
+            continue
+        if norm(d.get('stage')) != 'closedwon':
+            continue
+        close_date = (d.get('closeDate') or '').strip()
+        if close_date.startswith('2026-04-'):
+            out.append(d)
+    return out
+
+
+def _company_name_for_deal(deal, hubspot_initial):
+    company_id = deal.get('companyId')
+    for c in hubspot_initial.get('companies', []) or []:
+        if isinstance(c, dict) and c.get('id') == company_id:
+            return (c.get('name') or '').strip()
+    return (deal.get('customer') or '').strip()
+
+
+def _primary_contact_for_deal(deal, hubspot_initial):
+    contact_ids = deal.get('contactIds') or []
+    if not contact_ids:
+        return None
+    cid0 = contact_ids[0]
+    for c in hubspot_initial.get('contacts', []) or []:
+        if isinstance(c, dict) and c.get('id') == cid0:
+            return c
+    return None
+
+
+def _qb_customer_name_by_id(qb_state):
+    return {
+        c.get('id'): (c.get('name') or '').strip()
+        for c in (qb_state.get('customers') or [])
+        if isinstance(c, dict) and c.get('id')
+    }
+
+
+def _qb_product_by_id(qb_state):
+    return {
+        p.get('id'): p
+        for p in (qb_state.get('products') or [])
+        if isinstance(p, dict) and p.get('id')
+    }
+
+
+def _add_days(date_str, days):
+    try:
+        d = _dt.datetime.strptime(date_str, '%Y-%m-%d').date()
+    except (TypeError, ValueError):
+        return ''
+    return (d + _dt.timedelta(days=days)).isoformat()
+
+
+def _invoice_items(iv):
+    items = iv.get('items')
+    if isinstance(items, list):
+        return items
+    items = iv.get('lineItems')
+    if isinstance(items, list):
+        return items
+    return []
+
+
+def _check_a1_customer(deal, hubspot_initial, invoices, qb_cust_by_id):
+    target_name = norm(_company_name_for_deal(deal, hubspot_initial))
+    if not target_name:
+        return None
+    for iv in invoices or []:
+        if not isinstance(iv, dict):
+            continue
+        cust_name = norm(qb_cust_by_id.get(iv.get('customerId'), ''))
+        if not cust_name:
+            cust_name = norm(_invoice_customer_name(iv))
+        if cust_name and cust_name == target_name:
+            return iv
+    return None
+
+
+def _check_a2_line_item(deal, hubspot_initial, invoices, qb_cust_by_id, qb_prod_by_id):
+    """Self-contained: find ANY invoice for the deal's company that has the right
+    single line item. Independent of A1's pick (but typically returns same one)."""
+    target_name = norm(_company_name_for_deal(deal, hubspot_initial))
+    if not target_name:
+        return False
+    deal_name = deal.get('name') or ''
+    if ' - ' not in deal_name:
+        return False
+    expected_product = deal_name.split(' - ', 1)[1].strip()
+    deal_amt = _deal_amount(deal)
+
+    for invoice in invoices or []:
+        if not isinstance(invoice, dict):
+            continue
+        cust_name = norm(qb_cust_by_id.get(invoice.get('customerId'), ''))
+        if not cust_name:
+            cust_name = norm(_invoice_customer_name(invoice))
+        if cust_name != target_name:
+            continue
+        items = _invoice_items(invoice)
+        if len(items) != 1:
+            continue
+        item = items[0]
+        if not isinstance(item, dict):
+            continue
+        product = qb_prod_by_id.get(item.get('productId')) or {}
+        if norm(product.get('name', '')) != norm(expected_product):
+            continue
+        try:
+            qty = float(item.get('qty', 0) or 0)
+        except (TypeError, ValueError):
+            qty = 0
+        if abs(qty - 1.0) > 1e-6:
+            continue
+        try:
+            rate = float(item.get('rate', 0) or 0)
+            amt = float(item.get('amount', 0) or 0)
+        except (TypeError, ValueError):
+            continue
+        if abs(rate - deal_amt) > 0.5 or abs(amt - deal_amt) > 0.5:
+            continue
+        return True
+    return False
+
+
+def _check_a3_header(deal, hubspot_initial, invoices, qb_cust_by_id):
+    """Self-contained: find ANY invoice for the deal's company with correct header."""
+    target_name = norm(_company_name_for_deal(deal, hubspot_initial))
+    if not target_name:
+        return False
+    close_date = (deal.get('closeDate') or '').strip()
+    expected_due = _add_days(close_date, 30)
+
+    for invoice in invoices or []:
+        if not isinstance(invoice, dict):
+            continue
+        cust_name = norm(qb_cust_by_id.get(invoice.get('customerId'), ''))
+        if not cust_name:
+            cust_name = norm(_invoice_customer_name(invoice))
+        if cust_name != target_name:
+            continue
+        if norm(invoice.get('terms')) != 'net 30':
+            continue
+        if (invoice.get('date') or '').strip() != close_date:
+            continue
+        if (invoice.get('dueDate') or '').strip() != expected_due:
+            continue
+        try:
+            tax = float(invoice.get('tax', 0) or 0)
+        except (TypeError, ValueError):
+            tax = 0
+        if abs(tax) > 0.01:
+            continue
+        if norm(invoice.get('status')) != 'sent':
+            continue
+        return True
+    return False
+
+
+def _check_b1_recipient(deal, hubspot_initial, sent_emails):
+    contact = _primary_contact_for_deal(deal, hubspot_initial)
+    if not contact:
+        return None
+    target_email = norm(contact.get('email'))
+    if not target_email:
+        return None
+    for e in sent_emails or []:
+        if not isinstance(e, dict):
+            continue
+        if target_email in norm(_email_text(e)):
+            return e
+    return None
+
+
+def _check_b2_email_body(deal, hubspot_initial, sent_emails):
+    """Self-contained: find ANY sent email to the deal's primary contact whose
+    subject contains 'Invoice #<n>', body contains 'INVOICE #<n>' (same <n>),
+    'Total: $<deal.amount>', and 'Status: Sent'. Independent of any QB invoice."""
+    contact = _primary_contact_for_deal(deal, hubspot_initial)
+    if not contact:
+        return False
+    target_email = norm(contact.get('email'))
+    if not target_email:
+        return False
+    deal_amt = _deal_amount(deal)
+    amt_int = 'total: ${}'.format(int(round(deal_amt)))
+    amt_dec = 'total: ${:.2f}'.format(deal_amt)
+    subject_re = _re.compile(r'invoice\s*#\s*(\d+)')
+
+    for e in sent_emails or []:
+        if not isinstance(e, dict):
+            continue
+        if target_email not in norm(_email_text(e)):
+            continue
+        subject = norm(e.get('subject') or '')
+        body = e.get('body') or ''
+        body_text = _re.sub(r'<[^>]+>', ' ', body)
+        nb = norm(body_text)
+
+        m_sub = subject_re.search(subject)
+        if not m_sub:
+            continue
+        n = m_sub.group(1)
+        # Body must reference the SAME number (self-consistent).
+        if 'invoice #{}'.format(n) not in nb:
+            continue
+        if amt_int not in nb and amt_dec not in nb:
+            continue
+        if 'status: sent' not in nb:
+            continue
+        return True
+    return False
+
+
+def _check_c_description(deal_current):
+    """Self-contained: deal.description contains 'Invoiced INV-<digits>.'.
+    Independent of any QB invoice. (Falls back to deal.notes for legacy agents.)"""
+    if not deal_current:
+        return False
+    desc = deal_current.get('description') or ''
+    if _INV_NUMBER_RE.search(desc):
+        return True
+    notes_text = _notes_text(deal_current)
+    if _INV_NUMBER_RE.search(notes_text):
+        return True
+    return False
+
+
+WEIGHTS = {
+    'A1_customer':   0.15,
+    'A2_line_item':  0.20,
+    'A3_header':     0.15,
+    'B1_recipient':  0.10,
+    'B2_email':      0.15,
+    'C_description': 0.25,
+}
+
+
+def reward(go):
+    hubspot_i = go('hubspot').get('initial_state', {}) or {}
+    hubspot_c = go('hubspot').get('current_state', {}) or {}
+    qb_c = go('quickbooks').get('current_state', {}) or {}
+    gmail_c = go('gmail').get('current_state', {}) or {}
+
+    deals_i = hubspot_i.get('deals', []) or []
+    deals_c_by_id = {
+        d.get('id'): d
+        for d in (hubspot_c.get('deals', []) or [])
+        if isinstance(d, dict)
+    }
+    invoices = qb_c.get('invoices', []) or []
+    qb_cust_by_id = _qb_customer_name_by_id(qb_c)
+    qb_prod_by_id = _qb_product_by_id(qb_c)
+    sent_emails = _gmail_sent_emails(gmail_c)
+
+    won = _april_won_deals(deals_i)
+    n = len(won)
+    if n == 0:
+        print('DEBUG_054e615f n=0 -> vacuous REWARD=1.0')
+        return 1.0
+
+    per_deal_lines = []
+    pass_counts = {k: 0 for k in WEIGHTS}
+    total = 0.0
+
+    for d in won:
+        deal_id = d.get('id')
+        deal_current = deals_c_by_id.get(deal_id, {})
+
+        # A1 picks an invoice for debug logging; A2/A3 are self-contained
+        # but conceptually still belong to the same A group.
+        invoice = _check_a1_customer(d, hubspot_i, invoices, qb_cust_by_id)
+
+        checks = {
+            'A1_customer':   invoice is not None,
+            'A2_line_item':  _check_a2_line_item(d, hubspot_i, invoices, qb_cust_by_id, qb_prod_by_id),
+            'A3_header':     _check_a3_header(d, hubspot_i, invoices, qb_cust_by_id),
+            'B1_recipient':  _check_b1_recipient(d, hubspot_i, sent_emails) is not None,
+            'B2_email':      _check_b2_email_body(d, hubspot_i, sent_emails),
+            'C_description': _check_c_description(deal_current),
+        }
+
+        deal_score = 0.0
+        for k, ok in checks.items():
+            if ok:
+                deal_score += WEIGHTS[k]
+                pass_counts[k] += 1
+        total += deal_score
+
+        flags = ' '.join(
+            '{}={}'.format(k.split('_', 1)[0], 'Y' if v else 'N')
+            for k, v in checks.items()
+        )
+        company = _company_name_for_deal(d, hubspot_i) or '?'
+        per_deal_lines.append(
+            '  {} {:<12s}: score={:.3f}  {}'.format(deal_id, company, deal_score, flags)
+        )
+
+    score = total / n
+
+    print('DEBUG_054e615f n={}'.format(n))
+    for line in per_deal_lines:
+        print(line)
+    rates = '  '.join('{}={}/{}'.format(k.split('_', 1)[0], pass_counts[k], n) for k in WEIGHTS)
+    print('sub-check pass rates: {}'.format(rates))
+    print('total = sum / n = {:.4f} / {} = {:.4f}'.format(total, n, score))
+
+    return clamp01(score)
+
+
+try:
+    score = float(reward(go))
+except Exception as exc:
+    print(f'ERROR: reward() raised {exc!r}')
+    score = 0.0
+
+score = clamp01(score)
+print(f'\nREWARD: {round(score, 4)}')

ar_invoice_003/initial_setup.py

@@ -0,0 +1,279 @@
+"""
+Initial Setup: AR invoice issuance chain (Salesforce -> Google Sheets -> Slack)
+Task ID: ar_invoice_003
+Domain: mock_websites (salesforce_mock + google_sheets_mock + slack_mock)
+
+Injects the PRE-TASK baseline into all three mocks under the INITIAL sid.
+  - Salesforce: read-only source of truth, contains the Closed-Won opportunity.
+  - Google Sheets 'AR Invoice Ledger': invoices through row 14 (INV-1042), NO INV-1043.
+  - Slack #finance: prior finance chatter, NO INV-1043 issuance message.
+"""
+import copy
+import hashlib
+import json
+import os
+import shlex
+import subprocess
+import time
+import uuid
+
+import requests
+
+TASK_ID = 'ar_invoice_001'
+SID = 'cua-' + hashlib.md5(TASK_ID.encode()).hexdigest()[:16]
+
+SF_URL = 'http://28.7.184.198:8175'
+GS_URL = 'http://28.7.184.198:8145'
+SL_URL = 'http://28.7.184.198:8178'
+PROXY = {'http': 'http://star-proxy.oa.com:3128', 'https': 'http://star-proxy.oa.com:3128'}
+
+
+# ---------------------------------------------------------------- HTTP helpers
+def http_get(url):
+    last = None
+    for attempt in range(4):
+        for proxies in (None, PROXY):
+            try:
+                r = requests.get(url, timeout=25, proxies=proxies)
+                if r.status_code == 200:
+                    return r
+                last = Exception(f'HTTP {r.status_code}: {r.text[:200]}')
+            except Exception as e:
+                last = e
+        time.sleep(1.0)
+    raise last
+
+
+def http_post(url, payload):
+    last = None
+    for attempt in range(4):
+        for proxies in (None, PROXY):
+            try:
+                r = requests.post(url, json=payload, timeout=30, proxies=proxies)
+                if r.status_code == 200:
+                    return r
+                last = Exception(f'HTTP {r.status_code}: {r.text[:200]}')
+            except Exception as e:
+                last = e
+        time.sleep(1.0)
+    raise last
+
+
+# ----------------------------------------------------------- state builders
+def build_salesforce_state():
+    """Fetch the rich default Salesforce org and patch the target opportunity."""
+    probe = 'cua-probe-' + uuid.uuid4().hex[:8]
+    d = http_get(f'{SF_URL}/go?sid={probe}').json()
+    state = copy.deepcopy(d.get('initial_state') or d.get('current_state'))
+
+    # Make user-4 the AE "Sarah Chen" (opportunity owner / Slack recipient).
+    for u in state.get('users', []):
+        if u.get('userId') == 'user-4':
+            u['firstName'] = 'Sarah'
+            u['lastName'] = 'Chen'
+            u['email'] = 'sarah.chen@company.com'
+            u['title'] = 'Account Executive'
+
+    # Patch opp-1 into the Closed-Won target opportunity (source of truth).
+    for opp in state.get('opportunities', []):
+        if opp.get('opportunityId') == 'opp-1':
+            opp['name'] = 'Acme Corp - Platform License'
+            opp['accountId'] = 'account-1'
+            opp['amount'] = 48000
+            opp['stage'] = 'Closed Won'
+            opp['probability'] = 100
+            opp['closeDate'] = '2026-06-24T00:00:00.000Z'
+            opp['paymentTerms'] = 'Net 30'
+            opp['type'] = 'New Business'
+            opp['ownerId'] = 'user-4'
+            opp['nextStep'] = 'Issue AR invoice'
+            opp['description'] = 'Platform license deal, closed won. Net 30 payment terms.'
+            break
+    return state
+
+
+def _cell(value, fmt=None, bold=False):
+    c = {'value': str(value), 'formula': str(value)}
+    style = {}
+    if bold:
+        style = {'bold': True, 'bg': '#E8EAED', 'align': 'center'}
+    if style:
+        c['style'] = style
+    if fmt:
+        c['format'] = fmt
+    return c
+
+
+# Ledger baseline rows: A=Invoice#, B=Customer, C=Deal/Contract, D=Amount,
+# E=Issue Date, F=Due Date, G=Status, H=AE/Owner.  Rows 2..14 (INV-1030..INV-1042).
+LEDGER_ROWS = [
+    ['INV-1030', 'Globex Inc', 'Globex - Annual Support', 12000, '2026-04-02', '2026-05-02', 'Paid', 'Marcus Johnson'],
+    ['INV-1031', 'Initech', 'Initech - SaaS Subscription', 8500, '2026-04-10', '2026-05-10', 'Paid', 'Sarah Chen'],
+    ['INV-1032', 'Umbrella LLC', 'Umbrella - Platform License', 30000, '2026-04-15', '2026-05-15', 'Paid', 'Priya Patel'],
+    ['INV-1033', 'Soylent Corp', 'Soylent - Data Add-on', 5400, '2026-04-22', '2026-05-22', 'Paid', 'Marcus Johnson'],
+    ['INV-1034', 'Hooli', 'Hooli - Enterprise License', 64000, '2026-05-01', '2026-05-31', 'Paid', 'Sarah Chen'],
+    ['INV-1035', 'Vandelay Industries', 'Vandelay - Consulting', 15000, '2026-05-06', '2026-06-05', 'Sent', 'Priya Patel'],
+    ['INV-1036', 'Stark Industries', 'Stark - Platform License', 52000, '2026-05-12', '2026-06-11', 'Sent', 'Marcus Johnson'],
+    ['INV-1037', 'Wonka Co', 'Wonka - Annual Support', 9800, '2026-05-18', '2026-06-17', 'Sent', 'Sarah Chen'],
+    ['INV-1038', 'Cyberdyne', 'Cyberdyne - SaaS Subscription', 22000, '2026-05-24', '2026-06-23', 'Sent', 'Priya Patel'],
+    ['INV-1039', 'Wayne Enterprises', 'Wayne - Data Add-on', 7600, '2026-05-29', '2026-06-28', 'Sent', 'Marcus Johnson'],
+    ['INV-1040', 'Tyrell Corp', 'Tyrell - Platform License', 41000, '2026-06-03', '2026-07-03', 'Sent', 'Sarah Chen'],
+    ['INV-1041', 'Gekko & Co', 'Gekko - Consulting', 18500, '2026-06-09', '2026-07-09', 'Draft', 'Priya Patel'],
+    ['INV-1042', 'Massive Dynamic', 'Massive Dynamic - Enterprise License', 73000, '2026-06-15', '2026-07-15', 'Draft', 'Marcus Johnson'],
+]
+HEADERS = ['Invoice #', 'Customer', 'Deal/Contract', 'Amount', 'Issue Date', 'Due Date', 'Status', 'AE/Owner']
+COL_FMT = [None, None, None, 'currency', 'date', 'date', None, None]
+
+
+def build_ledger_data(rows):
+    cols = ['A', 'B', 'C', 'D', 'E', 'F', 'G', 'H']
+    data = {}
+    for i, h in enumerate(HEADERS):
+        data[f'{cols[i]}1'] = _cell(h, bold=True)
+    for r, row in enumerate(rows, start=2):
+        for i, val in enumerate(row):
+            data[f'{cols[i]}{r}'] = _cell(val, fmt=COL_FMT[i])
+    return data
+
+
+def build_sheets_state(rows):
+    return {
+        'id': 'workbook_1',
+        'title': 'AR Invoice Ledger',
+        'activeSheetId': 'sheet_1',
+        'selectedCell': 'A1',
+        'selectionRange': None,
+        'clipboard': None,
+        'isDragging': False,
+        'undoStack': [],
+        'redoStack': [],
+        'namedRanges': [],
+        'conditionalFormats': [],
+        'charts': [],
+        'showGridlines': True,
+        'showFormulas': False,
+        'zoom': 100,
+        'sheets': [{
+            'id': 'sheet_1',
+            'name': 'Ledger',
+            'rowCount': 100,
+            'colCount': 26,
+            'frozenRows': 1,
+            'frozenCols': 0,
+            'tabColor': '#1A73E8',
+            'isHidden': False,
+            'columnWidths': {'0': 100, '1': 150, '2': 230, '3': 100, '4': 110, '5': 110, '6': 90, '7': 140},
+            'data': build_ledger_data(rows),
+        }],
+    }
+
+
+def build_slack_state(include_issuance=False):
+    users = [
+        {'userId': 'user_1', 'fullName': 'John Smith', 'displayName': 'John', 'email': 'john.smith@company.com',
+         'avatar': 'https://picsum.photos/200/200?random=1', 'title': 'AR Analyst', 'status': 'online',
+         'statusMessage': '', 'statusEmoji': '', 'timeZone': 'America/New_York'},
+        {'userId': 'user_2', 'fullName': 'Sarah Chen', 'displayName': 'Sarah', 'email': 'sarah.chen@company.com',
+         'avatar': 'https://picsum.photos/200/200?random=2', 'title': 'Account Executive', 'status': 'online',
+         'statusMessage': '', 'statusEmoji': '', 'timeZone': 'America/Los_Angeles'},
+        {'userId': 'user_3', 'fullName': 'Marcus Johnson', 'displayName': 'Marcus', 'email': 'marcus.johnson@company.com',
+         'avatar': 'https://picsum.photos/200/200?random=3', 'title': 'Finance Manager', 'status': 'away',
+         'statusMessage': '', 'statusEmoji': '', 'timeZone': 'America/New_York'},
+        {'userId': 'user_4', 'fullName': 'Priya Patel', 'displayName': 'Priya', 'email': 'priya.patel@company.com',
+         'avatar': 'https://picsum.photos/200/200?random=4', 'title': 'Account Executive', 'status': 'online',
+         'statusMessage': '', 'statusEmoji': '', 'timeZone': 'America/Chicago'},
+    ]
+    current_user = copy.deepcopy(users[0])
+    channels = [
+        {'channelId': 'general', 'name': 'general', 'description': 'Company-wide announcements', 'topic': '',
+         'isPrivate': False, 'isStarred': False, 'members': ['user_1', 'user_2', 'user_3', 'user_4'],
+         'createdBy': 'user_1', 'createdAt': '2026-01-01T10:00:00Z', 'pinnedMessages': [], 'unreadCount': 0},
+        {'channelId': 'finance', 'name': 'finance', 'description': 'Accounts receivable & billing',
+         'topic': 'Invoicing, AR, collections', 'isPrivate': False, 'isStarred': True,
+         'members': ['user_1', 'user_2', 'user_3', 'user_4'], 'createdBy': 'user_3',
+         'createdAt': '2026-01-05T09:00:00Z', 'pinnedMessages': [], 'unreadCount': 0},
+    ]
+    finance_msgs = [
+        {'messageId': 'msg_f1', 'senderId': 'user_3', 'content': 'Reminder: month-end AR close is this Friday.',
+         'timestamp': '2026-06-22T14:00:00Z', 'threadId': None, 'reactions': [], 'attachments': [], 'isEdited': False},
+        {'messageId': 'msg_f2', 'senderId': 'user_2', 'content': 'Can someone confirm the Tyrell invoice went out?',
+         'timestamp': '2026-06-22T15:10:00Z', 'threadId': None, 'reactions': [], 'attachments': [], 'isEdited': False},
+        {'messageId': 'msg_f3', 'senderId': 'user_1', 'content': 'Yes, INV-1040 was sent on the 3rd.',
+         'timestamp': '2026-06-22T15:18:00Z', 'threadId': None, 'reactions': [], 'attachments': [], 'isEdited': False},
+    ]
+    if include_issuance:
+        finance_msgs.append({
+            'messageId': 'msg_f_inv1043', 'senderId': 'user_1',
+            'content': ('@Sarah Chen Invoice INV-1043 for Acme Corp ($48,000, due 2026-07-24) '
+                        'has been issued for the Acme Corp - Platform License deal.'),
+            'timestamp': '2026-06-24T15:30:00Z', 'threadId': None,
+            'reactions': [], 'attachments': [], 'isEdited': False,
+        })
+    return {
+        'currentUser': current_user,
+        'workspace': {'workspaceId': 'ws_1', 'workspaceName': 'Acme Corp', 'icon': ''},
+        'users': users,
+        'channels': channels,
+        'messages': {
+            'general': [
+                {'messageId': 'msg_g1', 'senderId': 'user_2', 'content': 'Morning team!',
+                 'timestamp': '2026-06-22T09:00:00Z', 'threadId': None, 'reactions': [], 'attachments': [], 'isEdited': False},
+            ],
+            'finance': finance_msgs,
+        },
+        'threads': {},
+        'dms': [],
+        'bookmarkedMessages': [],
+        'callHistory': [],
+        'settings': {'theme': 'light', 'notifications': 'all', 'displayDensity': 'comfortable',
+                     'showAvatars': True, 'use24Hour': False},
+        'invitations': [],
+        'notifications': [],
+    }
+
+
+# ------------------------------------------------------------------- GUI launch
+def launch_gui(command, delay_sec=1.0):
+    env = os.environ.copy()
+    env['DISPLAY'] = ':0'
+    subprocess.Popen(shlex.split(command), stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL, env=env)
+    time.sleep(delay_sec)
+
+
+# --------------------------------------------------------------------- main
+def main():
+    with open('/tmp/task_web_sid', 'w') as f:
+        f.write(SID)
+    print(f'INITIAL sid = {SID}')
+
+    # Inject baseline (action:set) into all three mocks.
+    http_post(f'{SF_URL}/post?sid={SID}', {'action': 'set', 'state': build_salesforce_state()})
+    print('Salesforce baseline injected.')
+    http_post(f'{GS_URL}/post?sid={SID}', {'action': 'set', 'state': build_sheets_state(LEDGER_ROWS)})
+    print('Google Sheets baseline injected (last row INV-1042).')
+    http_post(f'{SL_URL}/post?sid={SID}', {'action': 'set', 'state': build_slack_state(include_issuance=False)})
+    print('Slack baseline injected (no INV-1043 issuance message).')
+
+    # Verify each baseline.
+    sf = http_get(f'{SF_URL}/go?sid={SID}').json()
+    opp = next(o for o in sf['current_state']['opportunities'] if o['opportunityId'] == 'opp-1')
+    assert opp['name'] == 'Acme Corp - Platform License' and opp['stage'] == 'Closed Won' and opp['amount'] == 48000
+    gs = http_get(f'{GS_URL}/go?sid={SID}').json()
+    data = gs['current_state']['sheets'][0]['data']
+    assert data['A14']['value'] == 'INV-1042', f"A14={data['A14']['value']}"
+    assert 'A15' not in data, 'baseline must NOT contain INV-1043 row'
+    sl = http_get(f'{SL_URL}/go?sid={SID}').json()
+    fin = sl['current_state']['messages']['finance']
+    assert not any('INV-1043' in m['content'] for m in fin), 'baseline Slack must NOT mention INV-1043'
+    print('Verified baseline: SF opp Closed Won, sheet ends at INV-1042, no INV-1043 anywhere.')
+
+    # GUI-ready: open all three tabs in Chrome on the VM display.
+    launch_gui(
+        f'google-chrome --no-first-run --no-default-browser-check '
+        f'"{SF_URL}/?sid={SID}" "{GS_URL}/?sid={SID}" "{SL_URL}/?sid={SID}"',
+        delay_sec=3.0,
+    )
+    print(f'GUI_READY: launched Chrome with salesforce/sheets/slack tabs (sid={SID})')
+
+
+main()

ar_invoice_003/reward.py

@@ -0,0 +1,263 @@
+"""
+Reward Script: AR invoice chained task (Salesforce -> Google Sheets -> Slack)
+Task ID: ar_invoice_003
+Domain: mock_websites (google_sheets + slack)
+
+Scoring (1.0 total) — ONLY Google Sheet + Slack are scored; Salesforce is READ-ONLY:
+  - Sheet (0.5): a genuinely NEW ledger row (INV-1043) is appended below the existing
+    data with all 8 columns A..H matching the expected values, AND existing rows are
+    left unchanged. Per-cell graded (0.5 * matched_cells/8), gated on the row being new
+    and existing rows being untouched.
+  - Slack (0.5): a NEW message in #finance addressed to Sarah Chen announcing that
+    invoice INV-1043 for Acme Corp ($48,000, due 2026-07-24) has been issued.
+    Graded over 5 keyword criteria (0.1 each).
+
+Both components are anchored to task-introduced changes: they FAIL on initial_env
+(no new row, no new message) and PASS on golden_env.
+"""
+# --- injected by remap_cua_gym_ip.py: resolve dynamic mock host -> self-hosted IP ---
+_CUA_GYM_REMAP_INJECTED = True
+_CUA_GYM_SLUG_MAP = {'12306': 'http://28.7.184.198:8100', 'canvas-lms': 'http://28.7.184.198:8101', 'expensify': 'http://28.7.184.198:8102', 'pacs-viewer': 'http://28.7.184.198:8103', 'sap': 'http://28.7.184.198:8104', 'servicenow': 'http://28.7.184.198:8105', 'tradingview': 'http://28.7.184.198:8106', 'zendesk': 'http://28.7.184.198:8107', 'adp': 'http://28.7.184.198:8108', 'airtable': 'http://28.7.184.198:8109', 'aliyun': 'http://28.7.184.198:8110', 'amazon': 'http://28.7.184.198:8111', 'amazon-seller': 'http://28.7.184.198:8112', 'amplitude': 'http://28.7.184.198:8113', 'asana': 'http://28.7.184.198:8114', 'aws-console': 'http://28.7.184.198:8115', 'azure': 'http://28.7.184.198:8116', 'bamboohr': 'http://28.7.184.198:8117', 'booking-com': 'http://28.7.184.198:8118', 'canva': 'http://28.7.184.198:8119', 'canvas': 'http://28.7.184.198:8120', 'circleci': 'http://28.7.184.198:8121', 'clio': 'http://28.7.184.198:8122', 'cloudflare': 'http://28.7.184.198:8123', 'coinbase': 'http://28.7.184.198:8124', 'confluence': 'http://28.7.184.198:8125', 'contractbook': 'http://28.7.184.198:8126', 'datadog': 'http://28.7.184.198:8127', 'dingtalk': 'http://28.7.184.198:8128', 'discord': 'http://28.7.184.198:8129', 'docusign': 'http://28.7.184.198:8130', 'ebay': 'http://28.7.184.198:8131', 'epic-health': 'http://28.7.184.198:8132', 'expedia': 'http://28.7.184.198:8133', 'facebook': 'http://28.7.184.198:8134', 'feishu': 'http://28.7.184.198:8135', 'github': 'http://28.7.184.198:8136', 'gitlab': 'http://28.7.184.198:8137', 'gmail': 'http://28.7.184.198:8138', 'google-ads': 'http://28.7.184.198:8139', 'google-analytics': 'http://28.7.184.198:8140', 'google-calendar': 'http://28.7.184.198:8141', 'google-docs': 'http://28.7.184.198:8142', 'google-drive': 'http://28.7.184.198:8143', 'google-flights': 'http://28.7.184.198:8144', 'google-sheets': 'http://28.7.184.198:8145', 'greenhouse': 'http://28.7.184.198:8146', 'gusto': 'http://28.7.184.198:8147', 'hotjar': 'http://28.7.184.198:8148', 'hubspot-marketing': 'http://28.7.184.198:8149', 'hubspot': 'http://28.7.184.198:8150', 'instacart': 'http://28.7.184.198:8151', 'instagram': 'http://28.7.184.198:8152', 'jira': 'http://28.7.184.198:8153', 'klaviyo': 'http://28.7.184.198:8154', 'lattice': 'http://28.7.184.198:8155', 'linear': 'http://28.7.184.198:8156', 'linkedin': 'http://28.7.184.198:8157', 'looker-studio': 'http://28.7.184.198:8158', 'lucidchart': 'http://28.7.184.198:8159', 'mailchimp': 'http://28.7.184.198:8160', 'meta-ads': 'http://28.7.184.198:8161', 'microsoft-teams': 'http://28.7.184.198:8162', 'miro': 'http://28.7.184.198:8163', 'mixpanel': 'http://28.7.184.198:8164', 'monday': 'http://28.7.184.198:8165', 'notion': 'http://28.7.184.198:8166', 'openreview': 'http://28.7.184.198:8167', 'outlook-web': 'http://28.7.184.198:8168', 'paypal': 'http://28.7.184.198:8169', 'pinterest': 'http://28.7.184.198:8170', 'postman': 'http://28.7.184.198:8171', 'quickbooks': 'http://28.7.184.198:8172', 'reddit': 'http://28.7.184.198:8173', 'robinhood': 'http://28.7.184.198:8174', 'salesforce': 'http://28.7.184.198:8175', 'sentry': 'http://28.7.184.198:8176', 'shopify-admin': 'http://28.7.184.198:8177', 'slack': 'http://28.7.184.198:8178', 'stripe-dashboard': 'http://28.7.184.198:8179', 'tableau': 'http://28.7.184.198:8180', 'taobao-seller': 'http://28.7.184.198:8181', 'trello': 'http://28.7.184.198:8182', 'tripadvisor': 'http://28.7.184.198:8183', 'twitter': 'http://28.7.184.198:8184', 'uber-eats': 'http://28.7.184.198:8185', 'vercel': 'http://28.7.184.198:8186', 'wandb': 'http://28.7.184.198:8187', 'wechat': 'http://28.7.184.198:8188', 'weibo': 'http://28.7.184.198:8189', 'westlaw': 'http://28.7.184.198:8190', 'woocommerce': 'http://28.7.184.198:8191', 'workday': 'http://28.7.184.198:8192', 'xiaohongshu': 'http://28.7.184.198:8193', 'youtube': 'http://28.7.184.198:8194', 'zhihu': 'http://28.7.184.198:8195', 'zillow': 'http://28.7.184.198:8196', 'zoom-web': 'http://28.7.184.198:8197'}
+def _cua_gym_base(_slug):
+    return _CUA_GYM_SLUG_MAP.get(_slug, 'https://cua-gym-%s.xlang.ai' % _slug)
+# --- end injected ---
+import json
+import re
+import sys
+import time
+
+import requests
+
+# --- Read sid (written by initial_setup.py) ---
+try:
+    with open('/tmp/task_web_sid') as f:
+        SID = f.read().strip()
+    if not SID:
+        raise ValueError('sid is empty')
+except Exception as e:
+    print(f'CRITICAL: Cannot read sid from /tmp/task_web_sid: {e}')
+    print('REWARD: 0.0')
+    sys.exit(0)
+
+PROXY = 'http://star-proxy.oa.com:3128'
+
+
+def fetch_state(app):
+    """Fetch /go for a mock app. Proxy-first, direct fallback, with retries.
+    Returns parsed JSON dict or None."""
+    url = f'{_cua_gym_base(app)}/go?sid={SID}'
+    attempts = [
+        {'proxies': {'http': PROXY, 'https': PROXY}},
+        {'proxies': {'http': PROXY, 'https': PROXY}},
+        {'proxies': None},
+        {'proxies': None},
+    ]
+    for i, kw in enumerate(attempts):
+        try:
+            r = requests.get(url, timeout=30, **kw)
+            r.raise_for_status()
+            return r.json()
+        except Exception as e:
+            print(f'WARN: fetch {app} attempt {i + 1} failed: {e}')
+            time.sleep(2)
+    return None
+
+
+def norm(s):
+    """Lowercase + collapse whitespace for robust text compare."""
+    return re.sub(r'\s+', ' ', str(s).strip().lower())
+
+
+def norm_amount(s):
+    """Strip currency symbols / commas / spaces. Drop a trailing '.0'/'.00' decimal."""
+    if s is None:
+        return ''
+    x = re.sub(r'[^0-9.]', '', str(s))
+    if '.' in x:
+        x = x.rstrip('0').rstrip('.')
+    return x
+
+
+def date_matches(value, expected):
+    """Match the expected year/month/day regardless of separators or component order."""
+    if value is None:
+        return False
+    year, month, day = int(expected[:4]), int(expected[4:6]), int(expected[6:8])
+    parts = [int(x) for x in re.findall(r'\d+', str(value))]
+    if year in parts and month in parts and day in parts:
+        return True
+    compact = re.sub(r'[^0-9]', '', str(value))
+    return compact in {
+        f'{year:04d}{month:02d}{day:02d}',
+        f'{month:02d}{day:02d}{year:04d}',
+        f'{day:02d}{month:02d}{year:04d}',
+    }
+
+
+def sheet_rows(state):
+    """Return {row_index: [A..H values or None]} for non-empty rows of sheet 0."""
+    sheets = state.get('sheets') or []
+    if not sheets:
+        return {}
+    data = sheets[0].get('data', {})
+    # determine max row referenced
+    max_r = 0
+    for cid in data:
+        m = re.match(r'^[A-Z]+(\d+)$', cid)
+        if m:
+            max_r = max(max_r, int(m.group(1)))
+    out = {}
+    for r in range(1, max_r + 1):
+        vals = []
+        for c in 'ABCDEFGH':
+            cell = data.get(f'{c}{r}')
+            vals.append(cell.get('value') if cell else None)
+        if any(v is not None and str(v) != '' for v in vals):
+            out[r] = vals
+    return out
+
+
+def verify_sheet():
+    """0.5 max. Graded per-cell, gated on row-is-new AND existing-rows-unchanged."""
+    data = fetch_state('google-sheets')
+    if not data:
+        print('FAIL: Sheet — could not fetch google-sheets state')
+        return 0.0
+    initial = data.get('initial_state') or {}
+    current = data.get('current_state') or {}
+    if not current:
+        print('FAIL: Sheet — current_state empty')
+        return 0.0
+
+    init_rows = sheet_rows(initial)
+    cur_rows = sheet_rows(current)
+
+    # invoice number -> full row, for existing (initial) rows
+    def inv_map(rows):
+        m = {}
+        for _, vals in rows.items():
+            a = (vals[0] or '').strip() if vals[0] else ''
+            if a.upper().startswith('INV-'):
+                m[a.upper()] = vals
+        return m
+
+    init_inv = inv_map(init_rows)
+    cur_inv = inv_map(cur_rows)
+
+    # Gate A: existing rows unchanged (every initial invoice row identical in current)
+    unchanged = True
+    for inv, ivals in init_inv.items():
+        cvals = cur_inv.get(inv)
+        if cvals is None or [norm(x) for x in cvals] != [norm(x) for x in ivals]:
+            print(f'FAIL: Sheet — existing row {inv} was modified or removed')
+            unchanged = False
+    if not unchanged:
+        return 0.0
+
+    # Gate B: the target row must be genuinely NEW (absent from initial)
+    if 'INV-1043' in init_inv:
+        print('FAIL: Sheet — INV-1043 already present in initial_state (not a new row)')
+        return 0.0
+    new_row = cur_inv.get('INV-1043')
+    if new_row is None:
+        print('FAIL: Sheet — no new INV-1043 row found in current_state')
+        return 0.0
+
+    # Per-cell match against expected (A..H)
+    expected = {
+        'A': ('INV-1043', 'text'),
+        'B': ('Acme Corp', 'text'),
+        'C': ('Acme Corp - Platform License', 'text'),
+        'D': ('48000', 'amount'),
+        'E': ('20260624', 'date'),
+        'F': ('20260724', 'date'),
+        'G': ('Sent', 'text'),
+        'H': ('Sarah Chen', 'text'),
+    }
+    cols = 'ABCDEFGH'
+    matched = 0
+    for i, c in enumerate(cols):
+        exp, kind = expected[c]
+        actual = new_row[i]
+        ok = False
+        if kind == 'text':
+            ok = norm(actual) == norm(exp)
+        elif kind == 'amount':
+            ok = norm_amount(actual) == exp
+        elif kind == 'date':
+            ok = date_matches(actual, exp)
+        if ok:
+            matched += 1
+            print(f'PASS: Sheet cell {c}15 = {actual!r}')
+        else:
+            print(f'FAIL: Sheet cell {c}15 — expected {exp!r} ({kind}), got {actual!r}')
+
+    score = 0.5 * (matched / len(cols))
+    print(f'Sheet score: {score:.4f} ({matched}/{len(cols)} cells)')
+    return score
+
+
+def verify_slack():
+    """0.5 max. Graded over 5 criteria on a NEW #finance message addressed to Sarah Chen."""
+    data = fetch_state('slack')
+    if not data:
+        print('FAIL: Slack — could not fetch slack state')
+        return 0.0
+    initial = data.get('initial_state') or {}
+    current = data.get('current_state') or {}
+    if not current:
+        print('FAIL: Slack — current_state empty')
+        return 0.0
+
+    init_msgs = (initial.get('messages') or {}).get('finance', [])
+    cur_msgs = (current.get('messages') or {}).get('finance', [])
+
+    if len(cur_msgs) <= len(init_msgs):
+        print(f'FAIL: Slack — no new #finance message (init={len(init_msgs)}, cur={len(cur_msgs)})')
+        return 0.0
+
+    new_msgs = cur_msgs[len(init_msgs):]
+    # criteria evaluated over the new messages (best match)
+    criteria = [
+        ('addressed to Sarah Chen', lambda t: 'sarah chen' in t or '@sarah' in t),
+        ('invoice INV-1043', lambda t: 'inv-1043' in t),
+        ('customer Acme', lambda t: 'acme' in t),
+        ('amount 48000', lambda t: '48000' in re.sub(r'[$,\s]', '', t)),
+        ('due date 2026-07-24', lambda t: date_matches(t, '20260724')),
+    ]
+    best = 0
+    best_content = ''
+    for m in new_msgs:
+        t = norm(m.get('content', ''))
+        hits = sum(1 for _, fn in criteria if fn(t))
+        if hits > best:
+            best = hits
+            best_content = m.get('content', '')
+    # report against best message
+    if best_content:
+        tb = norm(best_content)
+        for name, fn in criteria:
+            print(f"{'PASS' if fn(tb) else 'FAIL'}: Slack — {name}")
+    score = 0.5 * (best / len(criteria))
+    print(f'Slack score: {score:.4f} ({best}/{len(criteria)} criteria) — msg: {best_content!r}')
+    return score
+
+
+def main():
+    total = 0.0
+    try:
+        total += verify_sheet()
+    except Exception as e:
+        print(f'ERROR: Sheet component — {e}')
+    try:
+        total += verify_slack()
+    except Exception as e:
+        print(f'ERROR: Slack component — {e}')
+
+    final = round(min(total, 1.0), 4)
+    print(f'\nScore: {total}/1.0')
+    print(f'REWARD: {final}')
+    return final
+
+
+main()

ar_payment_004/_cua_gym_vm_bridge.sh

@@ -0,0 +1,106 @@
+#!/bin/bash
+# CUA-Gym VM-side network fix (direct-IP mock access).
+#
+# Runs INSIDE the QEMU guest. The CUA-Gym mock services are now reached by
+# their real IDC address (http://28.7.184.198:81xx) instead of the old
+# localhost:9000-9097 socat bridge, so the VM only needs two fixes:
+#
+#   1. Working DNS — the qcow2-baked /etc/resolv.conf points at the Pod's
+#      cluster DNS, which is unreachable from QEMU SLIRP NAT
+#      ("Temporary failure in name resolution").
+#   2. The mock host excluded from the woa http proxy — otherwise the VM's
+#      Chrome (gsettings proxy) and initial_setup.py's `requests` get a 403
+#      from the proxy when hitting 28.7.184.198:81xx. (The old socat path never
+#      hit this because localhost is bypassed by default.)
+#
+# Idempotent. Scoped to the CUA-Gym bridge step; does not touch the generic
+# OSWorld setup pipeline.
+#
+# Argv:
+#   $1  optional sudo password (default "password"; SetupController
+#       substitutes {CLIENT_PASSWORD} when uploaded as a cache file).
+#   $2  unused (kept for backward compat — was socat first port 9000).
+#   $3  unused (kept for backward compat — was socat last  port 9097).
+#   $4  optional mock host IP (default 28.7.184.198).
+
+set +e
+
+PASS="${1:-password}"
+MOCK_HOST="${4:-28.7.184.198}"
+# Derive the /16 IDC subnet so future port/host tweaks on the same segment
+# stay bypassed too (28.7.184.198 -> 28.7.0.0/16).
+MOCK_NET="$(echo "${MOCK_HOST}" | awk -F. 'NF==4{print $1"."$2".0.0/16"}')"
+
+run_root() {
+    echo "$PASS" | sudo -S -p "" bash -c "$1"
+}
+
+# Detect SLIRP gateway from VM's routing table; fall back to 10.0.2.2.
+GW="$(ip route 2>/dev/null | awk '/default/ {print $3; exit}')"
+GW="${GW:-10.0.2.2}"
+
+# ─── CUA-Gym network fix (Taiji-specific) ─────────────────────────────
+fix_vm_network() {
+    # 0) Discover VM network state at runtime — gateway, own IP/CIDR, SLIRP
+    #    DNS forwarder. Nothing hardcoded except the mock host above.
+    local GW_NET="${GW%.*}"
+    local SLIRP_DNS="${GW_NET}.3"        # QEMU SLIRP DNS proxy is always .3
+    local OWN_CIDR
+    OWN_CIDR="$(ip -4 -o addr show scope global 2>/dev/null \
+                | awk '{print $4; exit}')"
+    local OWN_IP="${OWN_CIDR%/*}"
+    local OWN_NET=""
+    [ -n "${OWN_IP}" ] && OWN_NET="${OWN_IP%.*}.0/24"
+
+    # 1) DNS — point at SLIRP DNS proxy + public fallbacks. chattr +i keeps
+    #    systemd-resolved/NetworkManager from clobbering it.
+    if run_root "printf 'nameserver ${SLIRP_DNS}\\nnameserver 114.114.114.114\\nnameserver 8.8.8.8\\n' > /etc/resolv.conf
+                 chattr +i /etc/resolv.conf 2>/dev/null || true"; then
+        echo "[cua_gym_vm_bridge] DNS set: ${SLIRP_DNS}, 114.114.114.114, 8.8.8.8 (gw=${GW})"
+    else
+        echo "[cua_gym_vm_bridge] WARN: failed to write /etc/resolv.conf"
+    fi
+
+    # 2) Build the proxy bypass list from detected values + the direct-IP
+    #    mock host (and its /16 subnet). THIS is what makes 28.7.184.198:81xx
+    #    reachable from inside the VM — without it the woa proxy 403s it.
+    local ignore_list="['localhost', '127.0.0.0/8', '${GW}'"
+    [ -n "${OWN_IP}" ]    && ignore_list="${ignore_list}, '${OWN_IP}'"
+    [ -n "${OWN_NET}" ]   && ignore_list="${ignore_list}, '${OWN_NET}'"
+    [ -n "${GW_NET}" ]    && ignore_list="${ignore_list}, '${GW_NET}.0/24'"
+    [ -n "${MOCK_HOST}" ] && ignore_list="${ignore_list}, '${MOCK_HOST}'"
+    [ -n "${MOCK_NET}" ]  && ignore_list="${ignore_list}, '${MOCK_NET}'"
+    ignore_list="${ignore_list}]"
+
+    # 3) GNOME proxy ignore-hosts — Chrome bypasses the woa proxy for these.
+    if gsettings set org.gnome.system.proxy ignore-hosts "${ignore_list}" 2>/tmp/gsettings_err; then
+        echo "[cua_gym_vm_bridge] gsettings ignore-hosts = ${ignore_list}"
+    else
+        echo "[cua_gym_vm_bridge] WARN: gsettings ignore-hosts failed: $(cat /tmp/gsettings_err 2>/dev/null)"
+    fi
+
+    # 4) Shell env — comma-separated form for no_proxy/NO_PROXY (requests/curl).
+    local NP="localhost,127.0.0.1,${GW}"
+    [ -n "${OWN_IP}" ]    && NP="${NP},${OWN_IP}"
+    [ -n "${OWN_NET}" ]   && NP="${NP},${OWN_NET}"
+    [ -n "${GW_NET}" ]    && NP="${NP},${GW_NET}.0/24"
+    [ -n "${MOCK_HOST}" ] && NP="${NP},${MOCK_HOST}"
+    [ -n "${MOCK_NET}" ]  && NP="${NP},${MOCK_NET}"
+
+    for var in no_proxy NO_PROXY; do
+        if ! grep -q "export ${var}=${NP}" /home/user/.bashrc 2>/dev/null; then
+            echo "export ${var}=${NP}" >> /home/user/.bashrc
+        fi
+    done
+    run_root "grep -q '^no_proxy=' /etc/environment 2>/dev/null \
+              && sed -i 's|^no_proxy=.*|no_proxy=\"${NP}\"|' /etc/environment \
+              || echo 'no_proxy=\"${NP}\"' >> /etc/environment
+              grep -q '^NO_PROXY=' /etc/environment 2>/dev/null \
+              && sed -i 's|^NO_PROXY=.*|NO_PROXY=\"${NP}\"|' /etc/environment \
+              || echo 'NO_PROXY=\"${NP}\"' >> /etc/environment" >/dev/null 2>&1
+    echo "[cua_gym_vm_bridge] no_proxy=${NP}"
+}
+
+fix_vm_network
+
+echo "[cua_gym_vm_bridge] direct-IP mode: ${MOCK_HOST} (+${MOCK_NET}) bypassed from woa proxy; no socat forwarding needed"

ar_payment_004/initial_setup.py

@@ -0,0 +1,253 @@
+"""
+Initial Setup: AR payment INV-1045 (Initech) — PRE-task state.
+Task ID: ar_payment_004
+Domain: libreoffice_calc (HYBRID: local Calc ledger + salesforce_mock + slack_mock)
+
+Pre-task state:
+  - Ledger /home/user/ar_payment_002.xlsx, sheet 'AR Invoice Ledger', INV-1045 Status='Sent'.
+  - Salesforce: Closed-Won opp 'Initech CRM Rollout' owned by Priya Nair, NO payment-received activity.
+  - Slack #finance: prior chatter only, no settlement notification to Priya Nair.
+GUI: libreoffice --calc ledger + google-chrome for salesforce & slack mocks (DISPLAY=:0).
+"""
+
+import os
+import shlex
+import subprocess
+import time
+import uuid
+
+import openpyxl
+import requests
+
+WORKDIR = '/home/user'
+TASK_ID = 'ar_payment_002'
+OUTPUT = f'{WORKDIR}/{TASK_ID}.xlsx'
+
+SF_URL = 'http://28.7.184.198:8175'
+SLACK_URL = 'http://28.7.184.198:8178'
+
+
+# ---------------------------------------------------------------------------
+# Shared baseline builders (identical in initial_setup.py and golden_patch.py)
+# ---------------------------------------------------------------------------
+def ledger_rows():
+    """Fixed columns A=Invoice #,B=Customer,C=Deal/Contract,D=Amount,
+    E=Issue Date,F=Due Date,G=Status,H=AE/Owner. 10 data rows."""
+    return [
+        ['INV-1042', 'Acme Corp', 'Acme Enterprise License', 25000, '2026-05-10', '2026-06-09', 'Paid', 'John Smith'],
+        ['INV-1043', 'Globex', 'Globex Analytics Suite', 48000, '2026-05-15', '2026-06-14', 'Paid', 'Sarah Johnson'],
+        ['INV-1044', 'Soylent', 'Soylent Data Platform', 31500, '2026-05-20', '2026-06-19', 'Sent', 'Mike Chen'],
+        ['INV-1045', 'Initech', 'Initech CRM Rollout', 72000, '2026-05-25', '2026-06-24', 'Sent', 'Priya Nair'],
+        ['INV-1046', 'Umbrella', 'Umbrella Security Audit', 15800, '2026-06-01', '2026-07-01', 'Sent', 'Emily Davis'],
+        ['INV-1047', 'Hooli', 'Hooli Cloud Migration', 96000, '2026-06-03', '2026-07-03', 'Paid', 'Priya Nair'],
+        ['INV-1048', 'Stark Industries', 'Stark IoT Integration', 54000, '2026-06-05', '2026-07-05', 'Sent', 'Alex Kim'],
+        ['INV-1049', 'Wayne Enterprises', 'Wayne Security Platform', 120000, '2026-06-08', '2026-07-08', 'Paid', 'Sarah Johnson'],
+        ['INV-1050', 'Pied Piper', 'Pied Piper Compression API', 22500, '2026-06-12', '2026-07-12', 'Sent', 'Mike Chen'],
+        ['INV-1051', 'Wonka', 'Wonka Supply Chain', 38000, '2026-06-15', '2026-07-15', 'Overdue', 'Emily Davis'],
+    ]
+
+
+def build_ledger(paid_inv1045: bool):
+    wb = openpyxl.Workbook()
+    ws = wb.active
+    ws.title = 'AR Invoice Ledger'
+    headers = ['Invoice #', 'Customer', 'Deal/Contract', 'Amount',
+               'Issue Date', 'Due Date', 'Status', 'AE/Owner']
+    for col, h in enumerate(headers, 1):
+        ws.cell(row=1, column=col, value=h)
+    for r, row in enumerate(ledger_rows(), 2):
+        vals = list(row)
+        if paid_inv1045 and vals[0] == 'INV-1045':
+            vals[6] = 'Paid'  # column G = Status
+        for c, val in enumerate(vals, 1):
+            ws.cell(row=r, column=c, value=val)
+    wb.save(OUTPUT)
+
+
+def salesforce_state():
+    """Full baseline Salesforce org with Initech Closed-Won opp owned by Priya Nair,
+    NO payment-received activity on it."""
+    user = {
+        "userId": "user-1", "firstName": "John", "lastName": "Smith",
+        "email": "john.smith@company.com", "phone": "(555) 123-4567",
+        "title": "AR Specialist", "department": "Finance", "role": "Manager",
+        "avatar": "https://i.pravatar.cc/150?u=user-1",
+        "timezone": "America/New_York", "locale": "en-US", "theme": "lightning",
+    }
+    users = [
+        user,
+        {"userId": "user-2", "firstName": "Sarah", "lastName": "Johnson",
+         "email": "sarah.johnson@company.com", "phone": "(555) 222-3333",
+         "title": "Account Executive", "department": "Sales", "role": "Rep",
+         "avatar": "https://i.pravatar.cc/150?u=user-2", "timezone": "America/New_York",
+         "locale": "en-US", "theme": "lightning"},
+        {"userId": "user-3", "firstName": "Mike", "lastName": "Chen",
+         "email": "mike.chen@company.com", "phone": "(555) 333-4444",
+         "title": "Account Executive", "department": "Sales", "role": "Rep",
+         "avatar": "https://i.pravatar.cc/150?u=user-3", "timezone": "America/Los_Angeles",
+         "locale": "en-US", "theme": "lightning"},
+        {"userId": "user-6", "firstName": "Priya", "lastName": "Nair",
+         "email": "priya.nair@company.com", "phone": "(555) 666-7788",
+         "title": "Senior Account Executive", "department": "Sales", "role": "Rep",
+         "avatar": "https://i.pravatar.cc/150?u=user-6", "timezone": "America/New_York",
+         "locale": "en-US", "theme": "lightning"},
+    ]
+    accounts = [
+        {"accountId": "account-10", "name": "Initech", "type": "Customer",
+         "industry": "Technology", "revenue": 48000000, "employees": 320,
+         "ownerId": "user-6", "phone": "(555) 010-2030", "website": "www.initech.com",
+         "billingStreet": "4120 Freidrich Lane", "billingCity": "Austin",
+         "billingState": "TX", "billingZip": "78744", "billingCountry": "USA",
+         "createdDate": "2025-11-02T10:00:00.000Z", "modifiedDate": "2026-06-01T10:00:00.000Z"},
+        {"accountId": "account-11", "name": "Globex", "type": "Customer",
+         "industry": "Manufacturing", "revenue": 92000000, "employees": 540,
+         "ownerId": "user-2", "phone": "(555) 040-5060", "website": "www.globex.com",
+         "billingStreet": "200 Market St", "billingCity": "San Francisco",
+         "billingState": "CA", "billingZip": "94105", "billingCountry": "USA",
+         "createdDate": "2025-09-15T10:00:00.000Z", "modifiedDate": "2026-05-20T10:00:00.000Z"},
+    ]
+    contacts = [
+        {"contactId": "contact-10", "accountId": "account-10", "firstName": "Bill",
+         "lastName": "Lumbergh", "title": "VP of Operations", "department": "Operations",
+         "email": "bill.lumbergh@initech.com", "phone": "(555) 010-2031", "ownerId": "user-6"},
+        {"contactId": "contact-11", "accountId": "account-11", "firstName": "Hank",
+         "lastName": "Scorpio", "title": "CEO", "department": "Executive",
+         "email": "hank.scorpio@globex.com", "phone": "(555) 040-5061", "ownerId": "user-2"},
+    ]
+    opportunities = [
+        {"opportunityId": "opp-10", "name": "Initech CRM Rollout", "accountId": "account-10",
+         "contactId": "contact-10", "amount": 72000, "closeDate": "2026-06-20T00:00:00.000Z",
+         "stage": "Closed Won", "probability": 100, "type": "New Business",
+         "leadSource": "Referral", "nextStep": "Kickoff implementation",
+         "description": "Company-wide CRM rollout for Initech, 320 seats.",
+         "ownerId": "user-6", "createdDate": "2026-02-10T10:00:00.000Z",
+         "modifiedDate": "2026-06-20T10:00:00.000Z"},
+        {"opportunityId": "opp-11", "name": "Globex Analytics Suite", "accountId": "account-11",
+         "contactId": "contact-11", "amount": 48000, "closeDate": "2026-07-30T00:00:00.000Z",
+         "stage": "Negotiation", "probability": 70, "type": "New Business",
+         "leadSource": "Website", "nextStep": "Send revised proposal",
+         "description": "Analytics suite expansion for Globex.",
+         "ownerId": "user-2", "createdDate": "2026-04-01T10:00:00.000Z",
+         "modifiedDate": "2026-06-18T10:00:00.000Z"},
+    ]
+    activities = [
+        {"activityId": "activity-10", "type": "task", "subject": "Send kickoff agenda to Initech",
+         "status": "Not Started", "priority": "Normal", "dueDate": "2026-06-26T00:00:00.000Z",
+         "relatedToType": "Opportunity", "relatedToId": "opp-10", "assignedToId": "user-6",
+         "description": "Prepare and send implementation kickoff agenda to Bill Lumbergh.",
+         "createdDate": "2026-06-21T10:00:00.000Z"},
+        {"activityId": "activity-11", "type": "event", "subject": "Globex proposal review call",
+         "status": "Scheduled", "priority": "High",
+         "startDateTime": "2026-07-02T15:00:00.000Z", "endDateTime": "2026-07-02T15:30:00.000Z",
+         "relatedToType": "Opportunity", "relatedToId": "opp-11", "assignedToId": "user-2",
+         "description": "Walk through revised analytics suite proposal.",
+         "createdDate": "2026-06-19T10:00:00.000Z"},
+    ]
+    return {
+        "user": user, "users": users, "leads": [], "accounts": accounts,
+        "contacts": contacts, "opportunities": opportunities, "cases": [],
+        "activities": activities, "chatterPosts": [], "files": [], "following": [],
+        "recentlyViewed": [], "dismissedNotifications": [],
+    }
+
+
+def slack_state():
+    """Slack workspace with a #finance channel pre-seeded with prior chatter,
+    NO settlement notification to Priya Nair. Priya Nair present as a user."""
+    current_user = {
+        "userId": "user_1", "fullName": "John Smith", "displayName": "John",
+        "email": "john.smith@company.com", "avatar": "https://picsum.photos/200/200?random=1",
+        "title": "AR Specialist", "status": "online", "statusMessage": "", "statusEmoji": "",
+        "timeZone": "America/New_York",
+    }
+    users = [
+        current_user,
+        {"userId": "user_2", "fullName": "Sarah Johnson", "displayName": "Sarah",
+         "email": "sarah.johnson@company.com", "avatar": "https://picsum.photos/200/200?random=2",
+         "title": "Account Executive", "status": "online", "statusMessage": "", "statusEmoji": "",
+         "timeZone": "America/New_York"},
+        {"userId": "user_3", "fullName": "Mike Chen", "displayName": "Mike",
+         "email": "mike.chen@company.com", "avatar": "https://picsum.photos/200/200?random=3",
+         "title": "Account Executive", "status": "away", "statusMessage": "", "statusEmoji": "",
+         "timeZone": "America/Los_Angeles"},
+        {"userId": "user_9", "fullName": "Priya Nair", "displayName": "Priya",
+         "email": "priya.nair@company.com", "avatar": "https://picsum.photos/200/200?random=9",
+         "title": "Senior Account Executive", "status": "online", "statusMessage": "", "statusEmoji": "",
+         "timeZone": "America/New_York"},
+    ]
+    channels = [
+        {"channelId": "general", "name": "general", "description": "Company-wide announcements",
+         "topic": "Welcome!", "isPrivate": False, "isStarred": False,
+         "members": ["user_1", "user_2", "user_3", "user_9"], "createdBy": "user_1",
+         "createdAt": "2026-01-01T10:00:00Z", "pinnedMessages": [], "unreadCount": 0},
+        {"channelId": "finance", "name": "finance", "description": "Accounts receivable, billing, and collections",
+         "topic": "AR / billing coordination", "isPrivate": False, "isStarred": True,
+         "members": ["user_1", "user_2", "user_3", "user_9"], "createdBy": "user_1",
+         "createdAt": "2026-01-05T10:00:00Z", "pinnedMessages": [], "unreadCount": 0},
+    ]
+    messages = {
+        "general": [
+            {"messageId": "msg_g1", "senderId": "user_3", "content": "Morning all, quarterly close is next week.",
+             "timestamp": "2026-06-29T09:00:00Z", "threadId": None, "reactions": [],
+             "attachments": [], "isEdited": False},
+        ],
+        "finance": [
+            {"messageId": "msg_f1", "senderId": "user_2", "content": "Reminder: please log payments in Salesforce as they land so AR stays clean.",
+             "timestamp": "2026-06-26T13:10:00Z", "threadId": None, "reactions": [],
+             "attachments": [], "isEdited": False},
+            {"messageId": "msg_f2", "senderId": "user_9", "content": "Still waiting on Initech's wire for the CRM Rollout, will confirm once it clears.",
+             "timestamp": "2026-06-27T16:45:00Z", "threadId": None, "reactions": [],
+             "attachments": [], "isEdited": False},
+            {"messageId": "msg_f3", "senderId": "user_1", "content": "Thanks Priya — I'll mark the invoice paid the moment treasury confirms receipt.",
+             "timestamp": "2026-06-27T17:02:00Z", "threadId": None, "reactions": [],
+             "attachments": [], "isEdited": False},
+        ],
+    }
+    return {
+        "currentUser": current_user,
+        "workspace": {"workspaceId": "ws_1", "workspaceName": "Acme Corp", "icon": ""},
+        "users": users, "channels": channels, "messages": messages, "threads": {},
+        "dms": [], "bookmarkedMessages": [], "callHistory": [],
+        "settings": {"theme": "light", "notifications": "all", "displayDensity": "comfortable",
+                     "showAvatars": True, "use24Hour": False},
+        "invitations": [], "notifications": [],
+    }
+
+
+def launch_gui(command, delay_sec=1.0):
+    env = os.environ.copy()
+    env['DISPLAY'] = ':0'
+    subprocess.Popen(shlex.split(command), stdout=subprocess.DEVNULL,
+                     stderr=subprocess.DEVNULL, env=env)
+    time.sleep(delay_sec)
+
+
+def main():
+    # --- 1. Local ledger (INV-1045 Status='Sent') ---
+    build_ledger(paid_inv1045=False)
+    print(f'Initial ledger created: {OUTPUT}')
+
+    # --- 2. Shared sid for both mocks ---
+    sid = str(uuid.uuid4())
+    with open('/tmp/task_web_sid', 'w') as f:
+        f.write(sid)
+
+    # --- 3. Inject Salesforce + Slack initial state (action 'set') ---
+    for name, url, state in [
+        ('salesforce', SF_URL, salesforce_state()),
+        ('slack', SLACK_URL, slack_state()),
+    ]:
+        r = requests.post(f'{url}/post?sid={sid}', json={'action': 'set', 'state': state}, timeout=30)
+        assert r.status_code == 200, f'{name} set failed: {r.text}'
+        go = requests.get(f'{url}/go?sid={sid}', timeout=15).json()
+        assert go['initial_state'] is not None, f'{name} initial_state is None'
+        print(f'{name} initial state injected: sid={sid}')
+
+    # --- 4. GUI-ready startup ---
+    launch_gui(f'libreoffice --calc "{OUTPUT}"', delay_sec=3.0)
+    launch_gui(f'google-chrome "{SF_URL}/?sid={sid}"', delay_sec=2.0)
+    launch_gui(f'google-chrome "{SLACK_URL}/?sid={sid}"', delay_sec=2.0)
+    print('GUI_READY: launched libreoffice calc + chrome (salesforce, slack) with DISPLAY=:0')
+
+
+main()

ar_payment_004/reward.py

@@ -0,0 +1,227 @@
+"""
+Reward Script: AR payment chain — mark INV-1045 Paid in local ledger,
+log a payment-received activity on the Salesforce opportunity 'Initech CRM Rollout',
+and notify AE Priya Nair in Slack #finance that it's settled.
+Task ID: ar_payment_004
+Domain: libreoffice_calc + mock_websites (HYBRID: local xlsx -> Salesforce -> Slack)
+
+Scoring (each component scores ONLY the task-introduced delta):
+  Component 1 — LOCAL LEDGER  (0.34): INV-1045 Status -> 'Paid' AND no other row's
+                                       Status changed from its known baseline.
+  Component 2 — SALESFORCE    (0.33): NEW activity (delta vs initial) on opp 'Initech
+                                       CRM Rollout' recording payment of $72,000 / INV-1045.
+  Component 3 — SLACK         (0.33): NEW message (delta vs initial) in #finance addressed
+                                       to Priya Nair confirming $72,000 / INV-1045 settled.
+
+Total: 1.0. Fully deterministic; no LLM judge (no OpenAI creds in this env).
+Runs on the VM: openpyxl for the local file, requests for the mock /go state.
+"""
+import sys
+import re
+
+LEDGER_PATH = '/home/user/ar_payment_002.xlsx'
+SHEET_NAME = 'AR Invoice Ledger'
+
+# Known baseline (initial) Status per invoice, from the task context.
+# INV-1045 is the ONLY row that should change (Sent -> Paid).
+BASELINE_STATUS = {
+    'INV-1042': 'Paid',
+    'INV-1043': 'Paid',
+    'INV-1044': 'Sent',
+    'INV-1045': 'Sent',   # <- the row the agent must change to 'Paid'
+    'INV-1046': 'Sent',
+    'INV-1047': 'Paid',
+    'INV-1048': 'Sent',
+    'INV-1049': 'Paid',
+    'INV-1050': 'Sent',
+    'INV-1051': 'Overdue',
+}
+
+
+def norm(s):
+    """Lowercase, strip, drop commas/$/spaces around numbers for robust matching."""
+    return str(s or '').lower()
+
+
+def amount_present(text):
+    """True if text mentions 72000 / 72,000 / $72,000."""
+    t = re.sub(r'[,\s$]', '', str(text or ''))
+    return '72000' in t
+
+
+def verify_ledger():
+    """Component 1 (0.34): INV-1045 -> Paid, identity intact, no other status changed."""
+    try:
+        import openpyxl
+    except Exception as e:
+        print(f"ERROR: Component 1 — openpyxl import failed: {e}")
+        return 0.0
+    try:
+        wb = openpyxl.load_workbook(LEDGER_PATH, data_only=True)
+    except Exception as e:
+        print(f"ERROR: Component 1 — cannot load ledger {LEDGER_PATH}: {e}")
+        return 0.0
+    try:
+        if SHEET_NAME not in wb.sheetnames:
+            print(f"FAIL: Component 1 — sheet '{SHEET_NAME}' not found; sheets={wb.sheetnames}")
+            return 0.0
+        ws = wb[SHEET_NAME]
+
+        rows_by_inv = {}
+        for r in ws.iter_rows(min_row=2, values_only=True):
+            if r and r[0]:
+                rows_by_inv[str(r[0]).strip()] = r  # A=Invoice#, B=Cust, C=Deal, D=Amt, ... G=Status, H=Owner
+
+        # 1a: INV-1045 identity + Status == 'Paid' (the core change)
+        target = rows_by_inv.get('INV-1045')
+        if not target:
+            print("FAIL: Component 1 — INV-1045 row not found")
+            return 0.0
+        cust = str(target[1]).strip() if target[1] else ''
+        deal = str(target[2]).strip() if target[2] else ''
+        amt = target[3]
+        status = str(target[6]).strip() if target[6] else ''
+        owner = str(target[7]).strip() if target[7] else ''
+
+        identity_ok = (cust == 'Initech' and deal == 'Initech CRM Rollout'
+                       and amt == 72000 and owner == 'Priya Nair')
+        if not identity_ok:
+            print(f"FAIL: Component 1 — INV-1045 identity mismatch "
+                  f"(Customer={cust!r}, Deal={deal!r}, Amount={amt!r}, Owner={owner!r})")
+            return 0.0
+        if status != 'Paid':
+            print(f"FAIL: Component 1 — INV-1045 Status is {status!r}, expected 'Paid'")
+            return 0.0
+
+        # 1b: no OTHER row's Status changed from its baseline
+        for inv, expected in BASELINE_STATUS.items():
+            if inv == 'INV-1045':
+                continue
+            row = rows_by_inv.get(inv)
+            if not row:
+                print(f"FAIL: Component 1 — expected row {inv} missing")
+                return 0.0
+            actual = str(row[6]).strip() if row[6] else ''
+            if actual != expected:
+                print(f"FAIL: Component 1 — {inv} Status changed to {actual!r}, "
+                      f"expected unchanged {expected!r}")
+                return 0.0
+
+        print("PASS: Component 1 — INV-1045 marked 'Paid', identity intact, "
+              "no other row's status changed (0.34 pts)")
+        return 0.34
+    except Exception as e:
+        print(f"ERROR: Component 1 — {e}")
+        return 0.0
+
+
+def verify_salesforce(sid):
+    """Component 2 (0.33): NEW payment activity on 'Initech CRM Rollout' opportunity."""
+    import requests
+    try:
+        data = requests.get(f'http://28.7.184.198:8175/go?sid={sid}', timeout=20).json()
+    except Exception as e:
+        print(f"ERROR: Component 2 — cannot fetch salesforce state: {e}")
+        return 0.0
+    try:
+        initial = data.get('initial_state') or {}
+        current = data.get('current_state') or {}
+
+        # Resolve the Initech opportunity id from current state.
+        opp_id = None
+        for o in (current.get('opportunities') or []):
+            if str(o.get('name', '')).strip() == 'Initech CRM Rollout':
+                opp_id = o.get('opportunityId')
+                break
+        if not opp_id:
+            print("FAIL: Component 2 — opportunity 'Initech CRM Rollout' not found")
+            return 0.0
+
+        # DELTA: activities present now but not in initial.
+        init_ids = {a.get('activityId') for a in (initial.get('activities') or [])}
+        new_acts = [a for a in (current.get('activities') or [])
+                    if a.get('activityId') not in init_ids]
+        if not new_acts:
+            print("FAIL: Component 2 — no new activity logged (delta empty)")
+            return 0.0
+
+        # A new activity must relate to the Initech opp AND record the payment.
+        for a in new_acts:
+            if a.get('relatedToId') != opp_id:
+                continue
+            blob = norm(a.get('subject', '')) + ' ' + norm(a.get('description', ''))
+            if 'inv-1045' in blob and amount_present(blob):
+                print(f"PASS: Component 2 — new payment activity {a.get('activityId')} on "
+                      f"'Initech CRM Rollout' records INV-1045 / $72,000 (0.33 pts)")
+                return 0.33
+
+        print("FAIL: Component 2 — new activity exists but none on the Initech opp "
+              "mentions both INV-1045 and $72,000")
+        return 0.0
+    except Exception as e:
+        print(f"ERROR: Component 2 — {e}")
+        return 0.0
+
+
+def verify_slack(sid):
+    """Component 3 (0.33): NEW #finance message to Priya Nair confirming settlement."""
+    import requests
+    try:
+        data = requests.get(f'http://28.7.184.198:8178/go?sid={sid}', timeout=20).json()
+    except Exception as e:
+        print(f"ERROR: Component 3 — cannot fetch slack state: {e}")
+        return 0.0
+    try:
+        initial = data.get('initial_state') or {}
+        current = data.get('current_state') or {}
+
+        init_fin = (initial.get('messages') or {}).get('finance', [])
+        cur_fin = (current.get('messages') or {}).get('finance', [])
+
+        # DELTA: messages beyond the initial #finance snapshot.
+        new_msgs = cur_fin[len(init_fin):] if len(cur_fin) > len(init_fin) else []
+        if not new_msgs:
+            print("FAIL: Component 3 — no new message in #finance (delta empty)")
+            return 0.0
+
+        for m in new_msgs:
+            c = norm(m.get('content', ''))
+            addresses_priya = ('priya' in c)  # @Priya Nair / Priya
+            confirms = ('inv-1045' in c and amount_present(c)
+                        and ('settl' in c or 'paid' in c or 'received' in c))
+            if addresses_priya and confirms:
+                print(f"PASS: Component 3 — new #finance message {m.get('messageId')} addresses "
+                      f"Priya Nair and confirms INV-1045 / $72,000 settled (0.33 pts)")
+                return 0.33
+
+        print("FAIL: Component 3 — new #finance message(s) do not address Priya Nair "
+              "AND confirm INV-1045 / $72,000 settled")
+        return 0.0
+    except Exception as e:
+        print(f"ERROR: Component 3 — {e}")
+        return 0.0
+
+
+def main():
+    # sid for the mock surfaces (per-env)
+    try:
+        with open('/tmp/task_web_sid') as f:
+            sid = f.read().strip()
+        if not sid:
+            raise ValueError('sid empty')
+    except Exception as e:
+        print(f"CRITICAL: cannot read /tmp/task_web_sid: {e}")
+        print("REWARD: 0.0")
+        return
+
+    total = 0.0
+    total += verify_ledger()
+    total += verify_salesforce(sid)
+    total += verify_slack(sid)
+
+    final = round(min(total, 1.0), 4)
+    print(f"\nScore: {total}/1.0")
+    print(f"REWARD: {final}")
+
+
+main()

ar_payment_004/reward_label.json

@@ -0,0 +1,63 @@
+{
+  "reward_file": "/apdcephfs_sh2/share_300000800/user/jackwkwang/long-horizon-tasks/mini_osworld_xiangwu/cache/ar_payment_002/reward.py",
+  "api_url": "http://28.7.184.184:8000/v1/chat/completions",
+  "model": "Kimi-K2.6",
+  "max_code_chars": 100000,
+  "created_at": "2026-07-14 10:22:47",
+  "label": {
+    "task_id": "ar_payment_004",
+    "domain": "libreoffice_calc + mock_websites (HYBRID: local xlsx -> Salesforce -> Slack)",
+    "summary": "验证代理是否将本地 Excel 账本中 INV-1045 标记为 Paid、在 Salesforce 的 'Initech CRM Rollout' 机会上新增 $72,000 付款活动，并在 Slack #finance 频道向 Priya Nair 发送确认结清的消息。",
+    "is_placeholder": false,
+    "data_sources": [
+      "/home/user/ar_payment_002.xlsx",
+      "/tmp/task_web_sid",
+      "salesforce_mock (http://28.7.186.212:8195/go)",
+      "slack_mock (http://28.7.186.212:8198/go)"
+    ],
+    "scoring_components": [
+      {
+        "name": "Component 1 — LOCAL LEDGER",
+        "weight": 0.34,
+        "description": "检查本地 Excel 账本中 INV-1045 是否被正确标记为 Paid，且其他行状态未被篡改。",
+        "check_logic": "使用 openpyxl 加载 LEDGER_PATH 的 SHEET_NAME 工作表，遍历所有行建立 Invoice# 到行的映射。对 INV-1045 检查：Customer='Initech'、Deal='Initech CRM Rollout'、Amount=72000、Owner='Priya Nair'、Status='Paid'。对 BASELINE_STATUS 中其余发票检查：行必须存在且 Status 与基准值完全一致。",
+        "pass_condition": "INV-1045 的身份字段完全匹配且 Status 为 'Paid'，其他所有发票行的 Status 与 BASELINE_STATUS 基准一致且无缺失。"
+      },
+      {
+        "name": "Component 2 — SALESFORCE",
+        "weight": 0.33,
+        "description": "检查 Salesforce 中 'Initech CRM Rollout' 机会上是否新增了记录 INV-1045 / $72,000 付款的活动。",
+        "check_logic": "通过 requests GET 请求 salesforce_mock /go 端点，获取 initial_state 与 current_state。从 current_state.opportunities 中解析 name 为 'Initech CRM Rollout' 的 opportunityId。计算 activities 的 delta（current 中存在但 initial 中 activityId 不存在的记录）。在新增活动中查找 relatedToId 等于该 opp_id，且 subject 与 description 的归一化文本同时包含 'inv-1045' 和金额 72000 的记录。",
+        "pass_condition": "存在一条新增 activity，关联到 'Initech CRM Rollout' 机会，且其内容同时提及 INV-1045 和 $72,000。"
+      },
+      {
+        "name": "Component 3 — SLACK",
+        "weight": 0.33,
+        "description": "检查 Slack #finance 频道是否新增了向 Priya Nair 确认 INV-1045 / $72,000 已结清的消息。",
+        "check_logic": "通过 requests GET 请求 slack_mock /go 端点，获取 initial_state 与 current_state。提取 #finance 频道的消息列表，取 delta（current 中超出 initial 长度的尾部消息）。在新增消息中查找归一化内容同时满足：包含 'priya'、包含 'inv-1045'、包含金额 72000、包含 'settl'/'paid'/'received' 之一。",
+        "pass_condition": "#finance 频道存在一条新增消息，内容同时提及 Priya Nair、INV-1045、$72,000 并使用确认结清的关键词。"
+      }
+    ],
+    "total_max_score": 1.0,
+    "score_aggregation": "三个组件得分直接相加，随后使用 min(total, 1.0) 钳制到上限 1.0，最后 round(..., 4) 得到最终 REWARD。",
+    "failure_modes": [
+      "读取 /tmp/task_web_sid 失败或为空：CRITICAL 提前退出，输出 REWARD: 0.0",
+      "openpyxl 导入失败：Component 1 返回 0.0",
+      "Excel 文件加载失败或工作表 'AR Invoice Ledger' 不存在：Component 1 返回 0.0",
+      "INV-1045 行缺失或身份字段（Customer/Deal/Amount/Owner）不匹配：Component 1 返回 0.0",
+      "INV-1045 Status 不是 'Paid'：Component 1 返回 0.0",
+      "其他发票行缺失或 Status 偏离 BASELINE_STATUS：Component 1 返回 0.0",
+      "Salesforce mock 请求失败：Component 2 返回 0.0",
+      "未找到 'Initech CRM Rollout' 机会：Component 2 返回 0.0",
+      "activities delta 为空：Component 2 返回 0.0",
+      "新增 activity 未同时包含 INV-1045 和 $72,000：Component 2 返回 0.0",
+      "Slack mock 请求失败：Component 3 返回 0.0",
+      "#finance 无新增消息：Component 3 返回 0.0",
+      "新增消息未同时满足 Priya + INV-1045 + 金额 + 确认词：Component 3 返回 0.0",
+      "任何未捕获异常导致对应 Component 返回 0.0"
+    ],
+    "scoring_logic_explanation": "脚本首先读取 /tmp/task_web_sid 获取会话 ID。随后按顺序执行三个验证：1) verify_ledger() 使用 openpyxl 打开本地 Excel，确认 INV-1045 被改为 Paid 且身份正确，同时确保其他行状态与基准一致，通过得 0.34；2) verify_salesforce(sid) 请求 salesforce_mock，对比 initial 与 current state，在 'Initech CRM Rollout' 机会上查找新增且同时提及 INV-1045 和 $72,000 的 activity，��过得 0.33；3) verify_slack(sid) 请求 slack_mock，对比 #finance 消息增量，查找同时提及 Priya Nair、INV-1045、$72,000 及确认结清关键词的新消息，通过得 0.33。三个组件分数累加后，经 min(total, 1.0) 钳制并四舍五入到 4 位小数输出最终分数。",
+    "confidence": 1.0
+  },
+  "raw_response": "用户要求我分析一个 reward.py 评分脚本的评分逻辑，并以 JSON 格式输出结果。\n\n让我仔细阅读代码：\n\n1. **任务信息提取**：\n   - Task ID: ar_payment_002（从 docstring 中明确提取）\n   - Domain: libreoffice_calc + mock_websites (HYBRID)（从 docstring 中提取）\n\n2. **脚本功能概述**：\n   验证三个组件：\n   - 本地账本（Excel）：将 INV-1045 标记为 Paid，且不改变其他行\n   - Salesforce：在 \"Initech CRM Rollout\" 机会上新增记录 $72,000 / INV-1045 的活动\n   - Slack：在 #finance 频道向 Priya Nair 发送新消息确认 $72,000 / INV-1045 已结清\n\n3. **is_placeholder**：false，因为脚本有详细的验证逻辑。\n\n4. **data_sources**：\n   - 本地文件：`/home/user/ar_payment_002.xlsx`（openpyxl 读取）\n   - Salesforce mock：`http://28.7.186.212:8195/go?sid={sid}`（requests 获取）\n   - Slack mock：`http://28.7.186.212:8198/go?sid={sid}`（requests 获取）\n   - `/tmp/task_web_sid`（读取 sid）\n\n5. **scoring_components**：\n   - Component 1：weight 0.34，验证本地账本\n   - Component 2：weight 0.33，验证 Salesforce\n   - Component 3：weight 0.33，验证 Slack\n\n6. **score_aggregation**：相加，然后 `min(total, 1.0)` 钳制到上限，再 round 到 4 位小数。\n\n7. **failure_modes**：\n   - 读取 /tmp/task_web_sid 失败会提前退出（CRITICAL）\n   - openpyxl 导入失败\n   - Excel 文件加载失败\n   - 工作表未找到\n   - INV-1045 行未找到\n   - INV-1045 身份不匹配（Customer, Deal, Amount, Owner）\n   - INV-1045 Status 不是 'Paid'\n   - 其他行 Status 被改变或缺失\n   - Salesforce mock 拉取失败\n   - 未找到 \"Initech CRM Rollout\" 机会\n   - 没有新活动（delta 为空）\n   - 新活动未关联到正确机会或未提及 INV-1045 和金额\n   - Slack mock 拉取失败\n   - #finance 没有新消息\n   - 新消息未同时满足：提及 Priya、包含 INV-1045、金额、确认词（settl/paid/received）\n\n8. **scoring_logic_explanation**：按执行顺序：\n   - 从 /tmp/task_web_sid 读取 sid\n   - Component 1：用 openpyxl 打开 Excel，检查 INV-1045 行身份（Customer=Initech, Deal=Initech CRM Rollout, Amount=72000, Owner=Priya Nair）且 Status='Paid'；检查其他所有发票状态与 BASELINE_STATUS 一致未改变。通过得 0.34。\n   - Component 2：用 sid 请求 Salesforce mock，对比 initial/current state，找到 \"Initech CRM Rollout\" 机会 ID，检查新增活动中是否有 relatedToId 匹配且 subject/description 同时包含 inv-1045 和 $72,000。通过得 0.33。\n   - Component 3：用 sid 请求 Slack mock，对比 initial/current 的 #finance 消息，取增量消息，检查是否有消息同时满足：内容包含 \"priya\"、包含 \"inv-1045\"、包含金额 72000、包含 \"settl\"/\"paid\"/\"received\" 之一。通过得 0.33。\n   - 汇总：三个组件分数相加，min(total, 1.0) 钳制，round(..., 4)。\n\n9. **confidence**：1.0，因为代码清晰完整。\n\n现在构建 JSON，确保：\n- 只输出 JSON，无 Markdown 代码块\n- 第一个字符是 {，最后一个字符是 }\n- 所有字符串 value 使用中文（task_id、domain、字段名、代码标识符保留原样）\n- 不要臆测\n\n让我再检查一下 weight 字段。代码中显式给出：\n- Component 1: 0.34\n- Component 2: 0.33\n- Component 3: 0.33\n\ntotal_max_score: 1.0\n\n检查 pass_condition 和 check_logic 要准确：\n\nComponent 1:\n- check_logic: 使用 openpyxl 加载 /home/user/ar_payment_002.xlsx 的 'AR Invoice Ledger' 工作表，遍历所有行。检查 INV-1045 行的 Customer='Initech'、Deal='Initech CRM Rollout'、Amount=72000、Owner='Priya Nair' 且 Status='Paid'；同时遍历 BASELINE_STATUS 中除 INV-1045 外的所有发票，确认其 Status 与基准值一致且行存在。\n- pass_condition: INV-1045 的身份字段完全匹配且 Status 为 'Paid'，且其他所有发票行的 Status 与 BASELINE_STATUS 基准一致、无缺失。\n\nComponent 2:\n- check_logic: 通过 HTTP GET 请求 salesforce_mock 的 /go 端点获取 initial_state 和 current_state。从 current_state.opportunities 中找到 name 为 'Initech CRM Rollout' 的 opportunityId。计算 activities 的 delta（current 中存在但 initial 中 activityId 不存在的记录）。在新增活动中查找 relatedToId 等于该 opp_id 且 subject 与 description 的归一化文本同时包含 'inv-1045' 和金额 72000 的活动。\n- pass_condition: 存在一条新增 activity，其关联到 'Initech CRM Rollout' 机会，且内容同时提及 INV-1045 和 $72,000。\n\nComponent 3:\n- check_logic: 通过 HTTP GET 请求 slack_mock 的 /go 端点获��� initial_state 和 current_state。提取 #finance 频道的消息列表，计算 delta（current 中超出 initial 长度的尾部消息）。在新增消息中查找归一化内容同时满足：包含 'priya'、包含 'inv-1045'、包含金额 72000、包含 'settl'/'paid'/'received' 之一。\n- pass_condition: #finance 频道存在一条新增消息，内容同时提及 Priya Nair、INV-1045、$72,000 并使用确认结清的关键词。\n\nfailure_modes:\n- /tmp/task_web_sid 读取失败或为空：CRITICAL 提前退出，REWARD 0.0\n- openpyxl 导入失败：Component 1 返回 0.0\n- Excel 文件加载失败：Component 1 返回 0.0\n- 工作表 'AR Invoice Ledger' 不存在：Component 1 返回 0.0\n- INV-1045 行缺失：Component 1 返回 0.0\n- INV-1045 身份字段不匹配：Component 1 返回 0.0\n- INV-1045 Status 不是 'Paid'：Component 1 返回 0.0\n- 其他发票行缺失或 Status 偏离基准：Component 1 返回 0.0\n- Salesforce mock 请求失败：Component 2 返回 0.0\n- 未找到 'Initech CRM Rollout' 机会：Component 2 返回 0.0\n- activities delta 为空：Component 2 返回 0.0\n- 新增 activity 未同时包含 INV-1045 和 $72,000：Component 2 返回 0.0\n- Slack mock 请求失败：Component 3 返回 0.0\n- #finance 无新增消息：Component 3 返回 0.0\n- 新增消息未同时满足 Priya + INV-1045 + 金额 + 确认词：Component 3 返回 0.0\n- 任何未捕获异常：对应 Component 返回 0.0\n\nscore_aggregation: 三个组件得分相加，使用 min(total, 1.0) 钳制上限，再 round(..., 4) 得到最终 REWARD。\n\nsummary: 验证代理是否将本地 Excel 账本中 INV-1045 标记为 Paid、在 Salesforce 机会上记录付款活动、并在 Slack #finance 通知 Priya Nair。\n\ndomain 应该是 \"libreoffice_calc + mock_websites\" 还是只取一个？docstring 中写的是 \"libreoffice_calc + mock_websites (HYBRID: local xlsx -> Salesforce -> Slack)\"。按照要求，从 docstring 中提取的领域。我保留原样 \"libreoffice_calc + mock_websites (HYBRID: local xlsx -> Salesforce -> Slack)\" 或者简化为 \"libreoffice_calc + mock_websites\"。看要求说 \"如 mock_websites / software_engineer / libreoffice_calc\"。这里 hybrid 比较特殊，我保留原文中的描述。\n\n等等，用户要求 \"domain\": \"从 docstring 中提取的领域\"。docstring 中明确写了 \"Domain: libreoffice_calc + mock_websites (HYBRID: local xlsx -> Salesforce -> Slack)\"。我就用这个。\n\n现在确保 JSON 格式正确，没有多余内容。 </think> {\"task_id\":\"ar_payment_002\",\"domain\":\"libreoffice_calc + mock_websites (HYBRID: local xlsx -> Salesforce -> Slack)\",\"summary\":\"验证代理是否将本地 Excel 账本中 INV-1045 标记为 Paid、在 Salesforce 的 'Initech CRM Rollout' 机会上新增 $72,000 付款活动，并在 Slack #finance 频道向 Priya Nair 发送确认结清的消息。\",\"is_placeholder\":false,\"data_sources\":[\"/home/user/ar_payment_002.xlsx\",\"/tmp/task_web_sid\",\"salesforce_mock (http://28.7.186.212:8195/go)\",\"slack_mock (http://28.7.186.212:8198/go)\"],\"scoring_components\":[{\"name\":\"Component 1 — LOCAL LEDGER\",\"weight\":0.34,\"description\":\"检查本地 Excel 账本中 INV-1045 是否被正确标记为 Paid，且其他行状态未被篡改。\",\"check_logic\":\"使用 openpyxl 加载 LEDGER_PATH 的 SHEET_NAME 工作表，遍历所有行建立 Invoice# 到行的映射。对 INV-1045 检查：Customer='Initech'、Deal='Initech CRM Rollout'、Amount=72000、Owner='Priya Nair'、Status='Paid'。对 BASELINE_STATUS 中其余发票检查：行必须存在且 Status 与基准值完全一致。\",\"pass_condition\":\"INV-1045 的身份字段完全匹配且 Status 为 'Paid'，其他所有发票行的 Status 与 BASELINE_STATUS 基准一致且无缺失。\"},{\"name\":\"Component 2 — SALESFORCE\",\"weight\":0.33,\"description\":\"检查 Salesforce 中 'Initech CRM Rollout' 机会上是否新增了记录 INV-1045 / $72,000 付款的活动。\",\"check_logic\":\"通过 requests GET 请求 salesforce_mock /go 端点，获取 initial_state 与 current_state。从 current_state.opportunities 中解析 name 为 'Initech CRM Rollout' 的 opportunityId。计算 activities 的 delta（current 中存在但 initial 中 activityId 不存在的记录）。在新增活动中查找 relatedToId 等于该 opp_id，且 subject 与 description 的归一化文本同时包含 'inv-1045' 和金额 72000 的记录。\",\"pass_condition\":\"存在一条新增 activity，关联到 'Initech CRM Rollout' 机会，且其内容同时提及 INV-1045 和 $72,000。\"},{\"name\":\"Component 3 — SLACK\",\"weight\":0.33,\"description\":\"检查 Slack #finance 频道是否新增了向 Priya Nair 确认 INV-1045 / $72,000 已结清的消息。\",\"check_logic\":\"通过 requests GET 请求 slack_mock /go 端点，获取 initial_state 与 current_state。提取 #finance 频道的消息列表，取 delta（current 中超出 initial 长度的尾部消息）。在新增消息中查找归一化内容同时满足：包含 'priya'、包含 'inv-1045'、包含金额 72000、包含 'settl'/'paid'/'received' 之一。\",\"pass_condition\":\"#finance 频道存在一条新增消息，内容同时提及 Priya Nair、INV-1045、$72,000 并使用确认结清的关键词。\"}],\"total_max_score\":1.0,\"score_aggregation\":\"三个组件得分直接相加，随后使用 min(total, 1.0) 钳制到上限 1.0，最后 round(..., 4) 得到最终 REWARD。\",\"failure_modes\":[\"读取 /tmp/task_web_sid 失败或为空：CRITICAL 提前退出，输出 REWARD: 0.0\",\"openpyxl 导入失败：Component 1 返回 0.0\",\"Excel 文件加载失败或工作表 'AR Invoice Ledger' 不存在：Component 1 返回 0.0\",\"INV-1045 行缺失或身份字段（Customer/Deal/Amount/Owner）不匹配：Component 1 返回 0.0\",\"INV-1045 Status 不是 'Paid'：Component 1 返回 0.0\",\"其他发票行缺失或 Status 偏离 BASELINE_STATUS：Component 1 返回 0.0\",\"Salesforce mock 请求失败：Component 2 返回 0.0\",\"未找到 'Initech CRM Rollout' 机会：Component 2 返回 0.0\",\"activities delta 为空：Component 2 返回 0.0\",\"新增 activity 未同时包含 INV-1045 和 $72,000：Component 2 返回 0.0\",\"Slack mock 请求失败：Component 3 返回 0.0\",\"#finance 无新增消息：Component 3 返回 0.0\",\"新增消息未同时满足 Priya + INV-1045 + 金额 + 确认词：Component 3 返回 0.0\",\"任何未捕获异常导致对应 Component 返回 0.0\"],\"scoring_logic_explanation\":\"脚本首先读取 /tmp/task_web_sid 获取会话 ID。随后按顺序执行三个验证：1) verify_ledger() 使用 openpyxl 打开本地 Excel，确认 INV-1045 被改为 Paid 且身份正确，同时确保其他行状态与基准一致，通过得 0.34；2) verify_salesforce(sid) 请求 salesforce_mock，对比 initial 与 current state，在 'Initech CRM Rollout' 机会上查找新增且同时提及 INV-1045 和 $72,000 的 activity，通过得 0.33；3) verify_slack(sid) 请求 slack_mock，对比 #finance 消息增量，查找同时提及 Priya Nair、INV-1045、$72,000 及确认结清关键词的新消息，通过得 0.33。三个组件分数累加后，经 min(total, 1.0) 钳制并四舍五入到 4 位小数输出最终分数。\",\"confidence\":1.0}"
+}

ar_payment_alert_010__long/_cua_gym_vm_bridge.sh

@@ -0,0 +1,106 @@
+#!/bin/bash
+# CUA-Gym VM-side network fix (direct-IP mock access).
+#
+# Runs INSIDE the QEMU guest. The CUA-Gym mock services are now reached by
+# their real IDC address (http://28.7.184.198:81xx) instead of the old
+# localhost:9000-9097 socat bridge, so the VM only needs two fixes:
+#
+#   1. Working DNS — the qcow2-baked /etc/resolv.conf points at the Pod's
+#      cluster DNS, which is unreachable from QEMU SLIRP NAT
+#      ("Temporary failure in name resolution").
+#   2. The mock host excluded from the woa http proxy — otherwise the VM's
+#      Chrome (gsettings proxy) and initial_setup.py's `requests` get a 403
+#      from the proxy when hitting 28.7.184.198:81xx. (The old socat path never
+#      hit this because localhost is bypassed by default.)
+#
+# Idempotent. Scoped to the CUA-Gym bridge step; does not touch the generic
+# OSWorld setup pipeline.
+#
+# Argv:
+#   $1  optional sudo password (default "password"; SetupController
+#       substitutes {CLIENT_PASSWORD} when uploaded as a cache file).
+#   $2  unused (kept for backward compat — was socat first port 9000).
+#   $3  unused (kept for backward compat — was socat last  port 9097).
+#   $4  optional mock host IP (default 28.7.184.198).
+
+set +e
+
+PASS="${1:-password}"
+MOCK_HOST="${4:-28.7.184.198}"
+# Derive the /16 IDC subnet so future port/host tweaks on the same segment
+# stay bypassed too (28.7.184.198 -> 28.7.0.0/16).
+MOCK_NET="$(echo "${MOCK_HOST}" | awk -F. 'NF==4{print $1"."$2".0.0/16"}')"
+
+run_root() {
+    echo "$PASS" | sudo -S -p "" bash -c "$1"
+}
+
+# Detect SLIRP gateway from VM's routing table; fall back to 10.0.2.2.
+GW="$(ip route 2>/dev/null | awk '/default/ {print $3; exit}')"
+GW="${GW:-10.0.2.2}"
+
+# ─── CUA-Gym network fix (Taiji-specific) ─────────────────────────────
+fix_vm_network() {
+    # 0) Discover VM network state at runtime — gateway, own IP/CIDR, SLIRP
+    #    DNS forwarder. Nothing hardcoded except the mock host above.
+    local GW_NET="${GW%.*}"
+    local SLIRP_DNS="${GW_NET}.3"        # QEMU SLIRP DNS proxy is always .3
+    local OWN_CIDR
+    OWN_CIDR="$(ip -4 -o addr show scope global 2>/dev/null \
+                | awk '{print $4; exit}')"
+    local OWN_IP="${OWN_CIDR%/*}"
+    local OWN_NET=""
+    [ -n "${OWN_IP}" ] && OWN_NET="${OWN_IP%.*}.0/24"
+
+    # 1) DNS — point at SLIRP DNS proxy + public fallbacks. chattr +i keeps
+    #    systemd-resolved/NetworkManager from clobbering it.
+    if run_root "printf 'nameserver ${SLIRP_DNS}\\nnameserver 114.114.114.114\\nnameserver 8.8.8.8\\n' > /etc/resolv.conf
+                 chattr +i /etc/resolv.conf 2>/dev/null || true"; then
+        echo "[cua_gym_vm_bridge] DNS set: ${SLIRP_DNS}, 114.114.114.114, 8.8.8.8 (gw=${GW})"
+    else
+        echo "[cua_gym_vm_bridge] WARN: failed to write /etc/resolv.conf"
+    fi
+
+    # 2) Build the proxy bypass list from detected values + the direct-IP
+    #    mock host (and its /16 subnet). THIS is what makes 28.7.184.198:81xx
+    #    reachable from inside the VM — without it the woa proxy 403s it.
+    local ignore_list="['localhost', '127.0.0.0/8', '${GW}'"
+    [ -n "${OWN_IP}" ]    && ignore_list="${ignore_list}, '${OWN_IP}'"
+    [ -n "${OWN_NET}" ]   && ignore_list="${ignore_list}, '${OWN_NET}'"
+    [ -n "${GW_NET}" ]    && ignore_list="${ignore_list}, '${GW_NET}.0/24'"
+    [ -n "${MOCK_HOST}" ] && ignore_list="${ignore_list}, '${MOCK_HOST}'"
+    [ -n "${MOCK_NET}" ]  && ignore_list="${ignore_list}, '${MOCK_NET}'"
+    ignore_list="${ignore_list}]"
+
+    # 3) GNOME proxy ignore-hosts — Chrome bypasses the woa proxy for these.
+    if gsettings set org.gnome.system.proxy ignore-hosts "${ignore_list}" 2>/tmp/gsettings_err; then
+        echo "[cua_gym_vm_bridge] gsettings ignore-hosts = ${ignore_list}"
+    else
+        echo "[cua_gym_vm_bridge] WARN: gsettings ignore-hosts failed: $(cat /tmp/gsettings_err 2>/dev/null)"
+    fi
+
+    # 4) Shell env — comma-separated form for no_proxy/NO_PROXY (requests/curl).
+    local NP="localhost,127.0.0.1,${GW}"
+    [ -n "${OWN_IP}" ]    && NP="${NP},${OWN_IP}"
+    [ -n "${OWN_NET}" ]   && NP="${NP},${OWN_NET}"
+    [ -n "${GW_NET}" ]    && NP="${NP},${GW_NET}.0/24"
+    [ -n "${MOCK_HOST}" ] && NP="${NP},${MOCK_HOST}"
+    [ -n "${MOCK_NET}" ]  && NP="${NP},${MOCK_NET}"
+
+    for var in no_proxy NO_PROXY; do
+        if ! grep -q "export ${var}=${NP}" /home/user/.bashrc 2>/dev/null; then
+            echo "export ${var}=${NP}" >> /home/user/.bashrc
+        fi
+    done
+    run_root "grep -q '^no_proxy=' /etc/environment 2>/dev/null \
+              && sed -i 's|^no_proxy=.*|no_proxy=\"${NP}\"|' /etc/environment \
+              || echo 'no_proxy=\"${NP}\"' >> /etc/environment
+              grep -q '^NO_PROXY=' /etc/environment 2>/dev/null \
+              && sed -i 's|^NO_PROXY=.*|NO_PROXY=\"${NP}\"|' /etc/environment \
+              || echo 'NO_PROXY=\"${NP}\"' >> /etc/environment" >/dev/null 2>&1
+    echo "[cua_gym_vm_bridge] no_proxy=${NP}"
+}
+
+fix_vm_network
+
+echo "[cua_gym_vm_bridge] direct-IP mode: ${MOCK_HOST} (+${MOCK_NET}) bypassed from woa proxy; no socat forwarding needed"

ar_payment_alert_010__long/initial_setup.py

@@ -0,0 +1,969 @@
+"""
+Initial Setup: Clear the overnight payment-alert inbox
+Task ID: ar_payment_alert_010__long
+Mocks: stripe_dashboard_mock (8179), quickbooks_mock (8172),
+       gmail_mock (8138), slack_mock (8178)
+
+Scenario (multi-branch finance-ops, "today" = 2026-06-29)
+---------------------------------------------------------
+A finance-ops analyst clears the overnight PAYMENT-ALERT inbox in Gmail. Each
+alert email (from the billing system) flags one payment event and names the
+customer + amount (+ invoice # where relevant). For every alert the analyst
+derives the corrective action from the alert TYPE, carries the customer / amount
+/ invoice into Stripe or QuickBooks, performs the (hidden) control there, labels
+the alert with a status label, replies to the customer, then posts + pins a
+one-line summary to Slack #billing-ops.
+
+Branch rule (stated in the instruction; VERIFIED against current app state):
+  * Duplicate charge    -> REFUND the matching Stripe payment
+                           (skip if that payment is already refunded).
+  * Chargeback opened   -> SUBMIT EVIDENCE on the matching Stripe dispute
+                           (skip if its status is not 'needs_response').
+  * Payment received    -> RECEIVE PAYMENT on the matching QuickBooks invoice
+                           (skip if the invoice is already Paid).
+  * Under investigation -> LEAVE (pending — must NOT act; tempting distractor).
+  * Alert whose customer/invoice has no matching Stripe/QB record -> SKIP.
+
+Alert plan (10 emails) => 6 require action, 4 must be skipped:
+  2 REFUND (Duplicate charge), 2 SUBMIT_EVIDENCE (Chargeback opened),
+  2 RECEIVE_PAYMENT (Payment received), 2 LEAVE (Under investigation),
+  1 already-refunded 'Duplicate charge' (arms the refund skip guard),
+  1 no-match 'Payment received' (unresolvable distractor).
+
+Cross-app data flow (H3):
+  Customer + Amount read in the Gmail alert -> locate the matching Stripe payment
+  (customer_name + amount in cents) to refund, and the matching Stripe dispute
+  (by customer / charge) to submit evidence on. Invoice # in the alert -> matching
+  QuickBooks invoice number to receive payment on. Customer email (from the
+  matched Stripe payment / QB customer) -> the To of the Gmail reply.
+
+Ground-truth embedding (Rule 2), PRECOMPUTED in Python from the visible alert
+type + VERIFIED against the injected Stripe/QB records so visible <=> key can
+never disagree:
+  * gmail.initial_state.emails[*]._alert  -> per-alert worklist (inline hidden key)
+  * slack.initial_state._task_adapter.summary_channel = 'billing-ops'
+
+Observable result kept ABSENT at injection (Rule 3):
+  * the 2 REFUND-target payments: amount_refunded=0, refunded=False; refunds[]
+    holds only a decoy that does NOT reference any target.
+  * the 2 SUBMIT_EVIDENCE-target disputes: status='needs_response'.
+  * the 2 RECEIVE_PAYMENT-target QB invoices: status Sent/Overdue, paidAmount=0,
+    paidDate=None.
+  * every Gmail alert email: labels=[]; no 'sent'-folder reply to a target customer.
+  * Slack #billing-ops starts EMPTY (no messages, nothing pinned).
+  * the already-done distractor payment IS already refunded (must be left alone);
+    the LEAVE-row dispute IS 'needs_response' (must be left alone).
+
+NOTE: Stripe amounts are in CENTS, QuickBooks amounts are in DOLLARS.
+"""
+import datetime as _dt
+import os
+import shlex
+import subprocess
+import time
+import uuid
+
+import requests
+
+
+sid = str(uuid.uuid4())
+with open('/tmp/task_web_sid', 'w', encoding='utf-8') as f:
+    f.write(sid)
+print(f'Generated sid: {sid}')
+
+
+TASK_ID = 'c4b39a92-99cd-45d5-9e0c-a08817d82cb8'
+TODAY = _dt.date(2026, 6, 29)
+SUMMARY_CHANNEL = 'billing-ops'
+
+
+def norm(s):
+    return (str(s) if s is not None else '').strip().lower()
+
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+def _ts(date_str):
+    """Unix seconds for date_str at 12:00 UTC (stable displayed date across TZs)."""
+    d = _dt.datetime.strptime(date_str, '%Y-%m-%d').replace(
+        hour=12, minute=0, second=0, microsecond=0, tzinfo=_dt.timezone.utc
+    )
+    return int(d.timestamp())
+
+
+_CREATED = _ts('2026-01-08')
+_NOW_TS = _ts('2026-06-29')
+
+
+# ===========================================================================
+# 1) STRIPE — payments + disputes (+ customers, refunds)
+# ===========================================================================
+# Shared customer registry: name -> (stripe_customer_id, email, card)
+_STRIPE_CUSTS = {
+    'Acme Corporation':        ('cus_acme00000001', 'billing@acmecorp.com',       {'brand': 'visa',       'last4': '4242', 'exp_month': 12, 'exp_year': 2027}),
+    'Globex Industries':       ('cus_globex0000012', 'ap@globexind.com',           {'brand': 'mastercard', 'last4': '5555', 'exp_month':  8, 'exp_year': 2028}),
+    'TechStart Inc.':          ('cus_techstart0001', 'finance@techstart.io',       {'brand': 'visa',       'last4': '1234', 'exp_month':  3, 'exp_year': 2027}),
+    'Riverside Medical Group': ('cus_riverside0001', 'it@riversidemedical.org',    {'brand': 'visa',       'last4': '8888', 'exp_month':  1, 'exp_year': 2028}),
+    'CloudSync Pro':           ('cus_cloudsync0001', 'billing@cloudsyncpro.com',   {'brand': 'visa',       'last4': '6543', 'exp_month': 10, 'exp_year': 2028}),
+    'DataVault Systems':       ('cus_datavault0001', 'procurement@datavault.io',   {'brand': 'amex',       'last4': '0001', 'exp_month':  6, 'exp_year': 2028}),
+    'Pioneer Labs':            ('cus_pioneer000001', 'admin@pioneerlabs.io',       {'brand': 'visa',       'last4': '7777', 'exp_month': 12, 'exp_year': 2027}),
+    'Northwind Traders':       ('cus_northwind0001', 'ap@northwindtraders.com',    {'brand': 'mastercard', 'last4': '2323', 'exp_month':  5, 'exp_year': 2027}),
+    'Summit Retail':           ('cus_summit0000001', 'billing@summitretail.com',   {'brand': 'visa',       'last4': '9090', 'exp_month':  9, 'exp_year': 2028}),
+}
+
+
+def _stripe_customer(name):
+    cid, email, _card = _STRIPE_CUSTS[name]
+    return {
+        'id': cid, 'name': name, 'email': email, 'phone': None, 'description': None,
+        'address': {'line1': '200 Market St', 'line2': None, 'city': 'San Francisco',
+                    'state': 'CA', 'postal_code': '94105', 'country': 'US'},
+        'balance': 0, 'currency': 'usd', 'default_payment_method': None,
+        'metadata': {}, 'created': _CREATED, 'livemode': True, 'delinquent': False,
+        'total_spent': 0, 'payments_count': 1,
+    }
+
+
+def _payment(pid, cust_name, amount_cents, date_str, description,
+             refunded=False, disputed=False, status='succeeded'):
+    cid, email, card = _STRIPE_CUSTS[cust_name]
+    captured = (status == 'succeeded')
+    return {
+        'id': pid,
+        'amount': amount_cents,
+        'currency': 'usd',
+        'status': status,
+        'description': description,
+        'customer': cid,
+        'customer_email': email,
+        'customer_name': cust_name,
+        'payment_method': {'type': 'card', 'card': dict(card, funding='credit')},
+        'amount_received': amount_cents if captured else 0,
+        'amount_refunded': amount_cents if refunded else 0,
+        'refunded': bool(refunded),
+        'disputed': bool(disputed),
+        'captured': captured,
+        'receipt_email': email,
+        'receipt_url': None,
+        'metadata': {},
+        'created': _ts(date_str),
+        'livemode': True,
+        'risk_score': 12,
+        'risk_level': 'normal',
+        'outcome': {'type': 'authorized' if captured else 'pending',
+                    'risk_level': 'normal', 'risk_score': 12, 'reason': None},
+        'invoice': None,
+    }
+
+
+# Payments: 2 refund targets + 1 already-refunded distractor + 1 LEAVE bait
+# (DataVault, refunding it = FALSE POSITIVE) + 3 disputed charges (backing the
+# disputes) + 2 watermark decoys (backing the decoy disputes).
+_STRIPE_PAYMENTS = [
+    # --- REFUND targets (Duplicate charge) — ABSENT result: refunded=False ---
+    _payment('pi_refund_acme_0001',   'Acme Corporation',  25000, '2026-06-27',
+             'Pro Plan - Monthly (duplicate charge)', refunded=False),      # $250.00
+    _payment('pi_refund_globex_0002', 'Globex Industries',  9999, '2026-06-28',
+             'Starter Plan - Monthly (duplicate charge)', refunded=False),  # $99.99
+
+    # --- already-refunded distractor (Pioneer): must be LEFT alone ---
+    _payment('pi_already_pioneer_09', 'Pioneer Labs',       4999, '2026-06-20',
+             'API Access Add-on (already refunded)', refunded=True),        # $49.99
+
+    # --- LEAVE-row bait (DataVault 'Under investigation'): refunding = leak/FP ---
+    _payment('pi_leave_datavault_08', 'DataVault Systems', 18000, '2026-06-26',
+             'Team Plan - Monthly', refunded=False),                        # $180.00
+
+    # --- disputed charges backing the SUBMIT_EVIDENCE targets + LEAVE dispute ---
+    _payment('pi_disp_techstart_01',  'TechStart Inc.',    12000, '2026-06-10',
+             'Team Plan - Monthly', disputed=True),                         # $120.00
+    _payment('pi_disp_riverside_01',  'Riverside Medical Group', 45000, '2026-06-08',
+             'Clinic Suite - Monthly', disputed=True),                      # $450.00
+    _payment('pi_disp_cloudsync_01',  'CloudSync Pro',     48000, '2026-06-12',
+             'CloudSync Platform - Monthly', disputed=True),                # $480.00
+
+    # --- watermark decoys (backing the decoy disputes; NOT named by any alert) ---
+    _payment('pi_decoy_north_0011',   'Northwind Traders', 12000, '2026-06-12',
+             'Pro Plan Upgrade - Monthly', disputed=True),
+    _payment('pi_decoy_summit_0012',  'Summit Retail',     30000, '2026-06-14',
+             'Enterprise Annual - Prepaid', disputed=True),
+]
+
+
+def _dispute(dispute_id, cust_name, charge_id, amount_cents, reason, status,
+             due_date='2026-07-06'):
+    cid, _email, _card = _STRIPE_CUSTS[cust_name]
+    return {
+        'id': dispute_id,
+        'amount': amount_cents,
+        'currency': 'usd',
+        'charge': charge_id,
+        'customer': cid,
+        'customer_name': cust_name,
+        'reason': reason,
+        'status': status,
+        'evidence_due_by': _ts(due_date),
+        'created': _ts('2026-06-15'),
+        'metadata': {},
+    }
+
+
+# Disputes: 2 SUBMIT_EVIDENCE targets (needs_response) + 1 LEAVE bait
+# (CloudSync needs_response; responding = FALSE POSITIVE) + 2 decoys already
+# past needs_response (under_review / won — excluded from the diff).
+_STRIPE_DISPUTES = [
+    # --- SUBMIT_EVIDENCE targets — ABSENT result: status stays 'needs_response' ---
+    _dispute('dp_techstart_0001', 'TechStart Inc.',          'pi_disp_techstart_01',
+             12000, 'fraudulent',            'needs_response'),
+    _dispute('dp_riverside_0001', 'Riverside Medical Group', 'pi_disp_riverside_01',
+             45000, 'product_not_received',  'needs_response'),
+
+    # --- LEAVE-row bait (CloudSync 'Under investigation'): responding = leak/FP ---
+    _dispute('dp_cloudsync_0001', 'CloudSync Pro',           'pi_disp_cloudsync_01',
+             48000, 'unrecognized',          'needs_response'),
+
+    # --- watermark decoys (already past needs_response; never enter the diff) ---
+    _dispute('dp_decoy_north_001', 'Northwind Traders',      'pi_decoy_north_0011',
+             12000, 'duplicate',             'under_review'),
+    _dispute('dp_decoy_summit_01', 'Summit Retail',          'pi_decoy_summit_0012',
+             30000, 'fraudulent',            'won'),
+]
+
+_STRIPE_CUSTOMERS = [_stripe_customer(nm) for nm in _STRIPE_CUSTS]
+
+# Decoy refund that references a NON-target payment (watermark; not the targets).
+_STRIPE_REFUNDS = [
+    {'id': 're_decoy_pioneer_01', 'amount': 4999, 'currency': 'usd',
+     'charge': 'pi_already_pioneer_09', 'reason': 'duplicate', 'status': 'succeeded',
+     'created': _ts('2026-06-21'), 'metadata': {}},
+]
+
+_STRIPE_STATE = {
+    'business': {
+        'name': 'Northwind SaaS', 'email': 'admin@northwind-saas.com',
+        'url': 'https://northwind-saas.com', 'support_email': 'support@northwind-saas.com',
+        'country': 'US', 'currency': 'usd', 'timezone': 'America/Los_Angeles',
+    },
+    'currentUser': {'id': 'user_admin', 'name': 'Finance Admin',
+                    'email': 'admin@northwind-saas.com', 'role': 'administrator', 'avatar': None},
+    'balance': {'available': 8421500, 'pending': 1530000, 'reserved': 0, 'currency': 'usd'},
+    'customers': _STRIPE_CUSTOMERS,
+    'payments': _STRIPE_PAYMENTS,
+    'products': [],
+    'prices': [],
+    'invoices': [],
+    'subscriptions': [],
+    'payouts': [],
+    'disputes': _STRIPE_DISPUTES,
+    'refunds': _STRIPE_REFUNDS,   # decoy only; NONE references a refund target (Rule 3)
+    'balanceTransactions': [],
+    'events': [],
+    'paymentMethods': [],
+    'testMode': False,
+    'searchQuery': '',
+    'selectedDateRange': '30d',
+    'metrics': {
+        'today': {'grossVolume': 0, 'grossVolumeChart': []},
+        'summary': {
+            'grossVolume': {'amount': 100097, 'change': 0, 'previousAmount': 0},
+            'netVolume': {'amount': 97000, 'change': 0, 'previousAmount': 0},
+            'disputeActivity': {'rate': 0, 'change': 0, 'previousRate': 0},
+        },
+        'chartData': {'grossVolume': [], 'netVolume': [], 'disputeRate': []},
+    },
+}
+
+
+# ===========================================================================
+# 2) QUICKBOOKS — invoices (+ customers).  Amounts in DOLLARS.
+# ===========================================================================
+_QB_CUSTOMERS = [
+    {'id': 'c1', 'name': 'Dev Solutions LLC',    'company': 'Dev Solutions LLC',    'email': 'accounts@devsolutions.co', 'phone': '(555) 100-2001'},
+    {'id': 'c2', 'name': 'BrightPath Education',  'company': 'BrightPath Education',  'email': 'admin@brightpath.edu',     'phone': '(555) 100-2002'},
+    {'id': 'c3', 'name': 'Summit Retail',         'company': 'Summit Retail',         'email': 'billing@summitretail.com', 'phone': '(555) 100-2003'},
+    {'id': 'c4', 'name': 'Harbor Logistics',      'company': 'Harbor Logistics',      'email': 'ap@harborlogistics.com',   'phone': '(555) 100-2004'},
+]
+for _c in _QB_CUSTOMERS:
+    _c.setdefault('address', '100 Market St, San Francisco, CA 94105')
+    _c.setdefault('balance', 0)
+    _c.setdefault('notes', '')
+    _c.setdefault('isActive', True)
+    _c.setdefault('createdAt', '2026-01-08')
+
+
+def _qb_invoice(inv_id, number, customer_id, due_date, total, status):
+    inv_date = (_dt.date.fromisoformat(due_date) - _dt.timedelta(days=30)).isoformat()
+    paid = (status == 'Paid')
+    return {
+        'id': inv_id,
+        'number': number,
+        'customerId': customer_id,
+        'date': inv_date,
+        'dueDate': due_date,
+        'items': [{'id': f'{inv_id}_li1', 'productId': 'p1',
+                   'description': 'Professional services', 'qty': 1,
+                   'rate': total, 'amount': total}],
+        'subtotal': total,
+        'tax': 0,
+        'total': total,
+        'status': status,
+        'paidAmount': total if paid else 0,
+        'paidDate': inv_date if paid else None,
+        'terms': 'Net 30',
+        'message': '',
+        'createdAt': f'{inv_date}T09:00:00Z',
+    }
+
+
+# Invoices: 2 receive-payment targets (Sent/Overdue) + 1 unpaid decoy (receiving
+# payment on it = FALSE POSITIVE) + 2 already-Paid decoys + 1 Draft decoy.
+_QB_INVOICES = [
+    # --- RECEIVE_PAYMENT targets — ABSENT: not Paid, paidAmount=0 ---
+    _qb_invoice('inv_dev_1042',    '1042', 'c1', '2026-06-10', 1800.00, 'Sent'),     # target
+    _qb_invoice('inv_bright_1057', '1057', 'c2', '2026-05-20', 3200.00, 'Overdue'),  # target
+
+    # --- leak bait: an unpaid invoice NOT referenced by any alert ---
+    _qb_invoice('inv_summit_1099', '1099', 'c3', '2026-06-15',  500.00, 'Sent'),     # receiving = FP
+
+    # --- watermark decoys (already Paid / Draft) ---
+    _qb_invoice('inv_paid_1001',   '1001', 'c4', '2026-05-01', 2200.00, 'Paid'),
+    _qb_invoice('inv_paid_1002',   '1002', 'c1', '2026-04-25', 9900.00, 'Paid'),
+    _qb_invoice('inv_draft_1100',  '1100', 'c4', '2026-06-25', 1750.00, 'Draft'),
+]
+
+_QB_STATE = {
+    'company': {
+        'name': 'Northwind SaaS', 'address': '123 Business Rd, San Francisco, CA 94105',
+        'email': 'admin@northwind-saas.com', 'industry': 'Technology Services',
+        'accountingMethod': 'Accrual',
+    },
+    'customers': _QB_CUSTOMERS,
+    'invoices': _QB_INVOICES,
+}
+
+
+# ===========================================================================
+# 3) THE ALERT WORKLIST + PRECOMPUTED ANSWER KEY (single source of truth)
+# ===========================================================================
+# Index injected records so the key can point at real ids + verify pre-state.
+_PAY_BY_ID = {p['id']: p for p in _STRIPE_PAYMENTS}
+_DISP_BY_ID = {d['id']: d for d in _STRIPE_DISPUTES}
+_INV_BY_ID = {i['id']: i for i in _QB_INVOICES}
+_QB_CUST_BY_ID = {c['id']: c for c in _QB_CUSTOMERS}
+
+# Status label per action, matched by NAME.
+_LABEL_FOR_ACTION = {
+    'refund': 'Refunded',
+    'submit_evidence': 'Disputed',
+    'receive_payment': 'Reconciled',
+}
+
+# Each alert references its target record id (or None for the no-match row).
+_ALERT_SPECS = [
+    # --- REFUND (Duplicate charge) ---
+    dict(eid='m01', customer='Acme Corporation',        alert_type='Duplicate charge',
+         amount_display='$250.00', invoice_num='',
+         stripe_payment_id='pi_refund_acme_0001',   stripe_dispute_id=None, qb_invoice_id=None),
+    dict(eid='m02', customer='Globex Industries',       alert_type='Duplicate charge',
+         amount_display='$99.99', invoice_num='',
+         stripe_payment_id='pi_refund_globex_0002', stripe_dispute_id=None, qb_invoice_id=None),
+    # --- SUBMIT_EVIDENCE (Chargeback opened) ---
+    dict(eid='m03', customer='TechStart Inc.',          alert_type='Chargeback opened',
+         amount_display='$120.00', invoice_num='',
+         stripe_payment_id=None, stripe_dispute_id='dp_techstart_0001', qb_invoice_id=None),
+    dict(eid='m04', customer='Riverside Medical Group', alert_type='Chargeback opened',
+         amount_display='$450.00', invoice_num='',
+         stripe_payment_id=None, stripe_dispute_id='dp_riverside_0001', qb_invoice_id=None),
+    # --- RECEIVE_PAYMENT (Payment received) ---
+    dict(eid='m05', customer='Dev Solutions LLC',       alert_type='Payment received',
+         amount_display='$1,800.00', invoice_num='1042',
+         stripe_payment_id=None, stripe_dispute_id=None, qb_invoice_id='inv_dev_1042'),
+    dict(eid='m06', customer='BrightPath Education',     alert_type='Payment received',
+         amount_display='$3,200.00', invoice_num='1057',
+         stripe_payment_id=None, stripe_dispute_id=None, qb_invoice_id='inv_bright_1057'),
+    # --- LEAVE (Under investigation) — tempting distractors ---
+    dict(eid='m07', customer='DataVault Systems',       alert_type='Under investigation',
+         amount_display='$180.00', invoice_num='',
+         stripe_payment_id='pi_leave_datavault_08', stripe_dispute_id=None, qb_invoice_id=None),
+    dict(eid='m08', customer='CloudSync Pro',           alert_type='Under investigation',
+         amount_display='$480.00', invoice_num='',
+         stripe_payment_id=None, stripe_dispute_id='dp_cloudsync_0001', qb_invoice_id=None),
+    # --- already-refunded 'Duplicate charge' -> must be SKIPPED ---
+    dict(eid='m09', customer='Pioneer Labs',            alert_type='Duplicate charge',
+         amount_display='$49.99', invoice_num='',
+         stripe_payment_id='pi_already_pioneer_09', stripe_dispute_id=None, qb_invoice_id=None),
+    # --- no matching Stripe/QB record -> SKIP ---
+    dict(eid='m10', customer='Zenith Partners',         alert_type='Payment received',
+         amount_display='$750.00', invoice_num='9999',
+         stripe_payment_id=None, stripe_dispute_id=None, qb_invoice_id=None),
+]
+
+
+def _decide(spec):
+    """Derive (action, requires_action) from the alert TYPE, VERIFIED against the
+    current injected state.  Precomputed so the visible alert <=> key can't drift."""
+    t = norm(spec['alert_type'])
+
+    if t == 'duplicate charge':
+        pid = spec['stripe_payment_id']
+        pay = _PAY_BY_ID.get(pid) if pid else None
+        if pay is None:
+            return 'skip', False
+        already = bool(pay.get('refunded')) or (
+            pay.get('amount', 0) > 0 and pay.get('amount_refunded', 0) >= pay.get('amount', 0))
+        return 'refund', (not already)
+
+    if t == 'chargeback opened':
+        did = spec['stripe_dispute_id']
+        disp = _DISP_BY_ID.get(did) if did else None
+        if disp is None:
+            return 'skip', False
+        already = norm(disp.get('status')) != 'needs_response'
+        return 'submit_evidence', (not already)
+
+    if t == 'payment received':
+        iid = spec['qb_invoice_id']
+        inv = _INV_BY_ID.get(iid) if iid else None
+        if inv is None:
+            return 'skip', False
+        already = (norm(inv.get('status')) == 'paid')
+        return 'receive_payment', (not already)
+
+    # 'Under investigation' (and anything else) -> LEAVE
+    return 'leave', False
+
+
+def _customer_email(spec, action):
+    """Cross-app carry: the reply-To is read from the matched Stripe payment /
+    dispute customer (refund / submit_evidence) or the QB customer (receive)."""
+    if action == 'refund':
+        pay = _PAY_BY_ID.get(spec['stripe_payment_id'])
+        return pay.get('customer_email') if pay else None
+    if action == 'submit_evidence':
+        disp = _DISP_BY_ID.get(spec['stripe_dispute_id'])
+        if disp:
+            _cid, email, _card = _STRIPE_CUSTS.get(disp.get('customer_name'), (None, None, None))
+            return email
+        return None
+    if action == 'receive_payment':
+        inv = _INV_BY_ID.get(spec['qb_invoice_id'])
+        cust = _QB_CUST_BY_ID.get(inv.get('customerId')) if inv else None
+        return cust.get('email') if cust else None
+    return None
+
+
+ALERT_SUBJECTS = {
+    'Duplicate charge':    lambda s: f"Payment alert: Duplicate charge flagged for {s['customer']} ({s['amount_display']})",
+    'Chargeback opened':   lambda s: f"Payment alert: Chargeback opened by {s['customer']} ({s['amount_display']})",
+    'Payment received':    lambda s: f"Payment alert: Payment received from {s['customer']} — invoice #{s['invoice_num']} ({s['amount_display']})",
+    'Under investigation': lambda s: f"Payment alert: {s['customer']} account under investigation ({s['amount_display']})",
+}
+
+
+def _alert_body(spec, action, customer_email):
+    lines = [
+        f"Automated billing alert ({spec['alert_type']}).",
+        f"Customer: {spec['customer']}",
+        f"Amount: {spec['amount_display']}",
+    ]
+    if spec['invoice_num']:
+        lines.append(f"Invoice #: {spec['invoice_num']}")
+    if customer_email:
+        lines.append(f"Customer contact: {customer_email}")
+    lines.append("Review this event and take the corrective action its type calls for.")
+    return '\n'.join(lines)
+
+
+GMAIL_USER = {
+    'userId': 'u1',
+    'username': 'Finance Ops',
+    'email': 'billing-ops@northwind-saas.com',
+    'avatar': 'https://picsum.photos/200/200?random=1',
+}
+BILLING_BOT = {'name': 'Billing System', 'email': 'billing-bot@northwind-saas.com', 'avatar': ''}
+
+
+emails = []
+expected = []
+for idx, spec in enumerate(_ALERT_SPECS):
+    action, requires_action = _decide(spec)
+    customer_email = _customer_email(spec, action)
+    subject = ALERT_SUBJECTS[spec['alert_type']](spec)
+
+    refund_amount_cents = None
+    if action == 'refund' and spec['stripe_payment_id']:
+        refund_amount_cents = _PAY_BY_ID[spec['stripe_payment_id']].get('amount')
+    qb_invoice_number = None
+    if action == 'receive_payment' and spec['qb_invoice_id'] in _INV_BY_ID:
+        qb_invoice_number = _INV_BY_ID[spec['qb_invoice_id']].get('number')
+    label_name = _LABEL_FOR_ACTION.get(action) if requires_action else None
+
+    # ---- inline hidden answer key (reward reads emails[*]._alert) ----
+    _alert = {
+        'type': spec['alert_type'],
+        'action': action,
+        'requires_action': bool(requires_action),
+        'customer': spec['customer'],
+        'customer_email': customer_email if requires_action else None,
+        'stripe_payment_id': spec['stripe_payment_id'] if action == 'refund' else None,
+        'refund_amount_cents': refund_amount_cents,
+        'stripe_dispute_id': spec['stripe_dispute_id'] if action == 'submit_evidence' else None,
+        'qb_invoice_id': spec['qb_invoice_id'] if action == 'receive_payment' else None,
+        'qb_invoice_number': qb_invoice_number,
+        'label_name': label_name,
+        'reply_subject': f'Re: {subject}' if requires_action else None,
+    }
+
+    minutes_ago = 60 * (len(_ALERT_SPECS) - idx)  # spread overnight, all before today
+    ts_dt = _dt.datetime(2026, 6, 29, 6, 0, tzinfo=_dt.timezone.utc) - _dt.timedelta(minutes=minutes_ago)
+    body = _alert_body(spec, action, customer_email)
+    emails.append({
+        'id': spec['eid'],
+        'threadId': f"thread_{spec['eid']}",
+        'from': dict(BILLING_BOT),
+        'to': [{'name': GMAIL_USER['username'], 'email': GMAIL_USER['email']}],
+        'cc': [], 'bcc': [],
+        'folder': 'inbox',
+        'subject': subject,
+        'body': body,
+        'snippet': body[:120],
+        'timestamp': ts_dt.strftime('%Y-%m-%dT%H:%M:%SZ'),
+        'read': False, 'starred': False, 'important': False,
+        'labels': [],            # Rule 3: no status label yet
+        'category': 'primary',
+        'attachments': [],
+        '_alert': _alert,        # inline hidden answer key
+    })
+
+    expected.append({'eid': spec['eid'], **_alert})
+
+
+# ---- resolved partitions (for asserts + the DEBUG line) ----
+refund_rows   = [e for e in expected if e['action'] == 'refund' and e['requires_action']]
+dispute_rows  = [e for e in expected if e['action'] == 'submit_evidence' and e['requires_action']]
+receive_rows  = [e for e in expected if e['action'] == 'receive_payment' and e['requires_action']]
+leave_rows    = [e for e in expected if e['action'] == 'leave']
+refund_done   = [e for e in expected if e['action'] == 'refund' and not e['requires_action']]
+skip_rows     = [e for e in expected if e['action'] == 'skip']
+requires_rows = [e for e in expected if e['requires_action']]
+resolved_count = len(requires_rows)
+
+
+# ---------------------------------------------------------------------------
+# Build-time asserts — pin the partition sizes + Rule-3 absences.
+# ---------------------------------------------------------------------------
+assert len(_ALERT_SPECS) == 10, len(_ALERT_SPECS)
+assert len(refund_rows) == 2, refund_rows            # 2 REFUND require action
+assert len(dispute_rows) == 2, dispute_rows          # 2 SUBMIT_EVIDENCE require action
+assert len(receive_rows) == 2, receive_rows          # 2 RECEIVE_PAYMENT require action
+assert len(leave_rows) == 2, leave_rows              # 2 LEAVE (Under investigation)
+assert len(refund_done) == 1, refund_done            # 1 already-refunded distractor
+assert len(skip_rows) == 1, skip_rows                # 1 no-match
+assert resolved_count == 6, resolved_count           # 6 require action
+assert (len(leave_rows) + len(refund_done) + len(skip_rows)) == 4  # 4 must be skipped
+
+# every require-action alert carries a status label + reply target + customer email
+for e in requires_rows:
+    assert e['label_name'] in ('Refunded', 'Disputed', 'Reconciled'), e
+    assert e['customer_email'], e
+    assert e['reply_subject'] and e['reply_subject'].startswith('Re: '), e
+
+# Rule 3: refund targets NOT yet refunded.
+for e in refund_rows:
+    p = _PAY_BY_ID[e['stripe_payment_id']]
+    assert p['refunded'] is False and p['amount_refunded'] == 0, p['id']
+# Rule 3: the already-done distractor IS refunded (must be left alone).
+_pioneer = _PAY_BY_ID['pi_already_pioneer_09']
+assert _pioneer['refunded'] is True and _pioneer['amount_refunded'] == _pioneer['amount']
+# Rule 3: submit-evidence targets are 'needs_response' at injection.
+for e in dispute_rows:
+    d = _DISP_BY_ID[e['stripe_dispute_id']]
+    assert norm(d['status']) == 'needs_response', d['id']
+# Rule 3: the LEAVE-row dispute IS 'needs_response' (bait — must be left alone).
+assert norm(_DISP_BY_ID['dp_cloudsync_0001']['status']) == 'needs_response'
+# Rule 3: receive targets NOT yet Paid.
+for e in receive_rows:
+    inv = _INV_BY_ID[e['qb_invoice_id']]
+    assert norm(inv['status']) in ('sent', 'overdue'), inv['status']
+    assert inv['paidAmount'] == 0 and inv['paidDate'] is None, inv['id']
+# Rule 3: NO refund references any refund target.
+_target_pay_ids = {e['stripe_payment_id'] for e in refund_rows}
+assert all(r['charge'] not in _target_pay_ids for r in _STRIPE_REFUNDS)
+# Rule 3: every alert email ships with no status label.
+assert all(em['labels'] == [] for em in emails)
+
+print(f'Precomputed worklist: refund={len(refund_rows)} submit_evidence={len(dispute_rows)} '
+      f'receive_payment={len(receive_rows)} leave={len(leave_rows)} '
+      f'already_refunded={len(refund_done)} no_match={len(skip_rows)} '
+      f'-> requires_action={resolved_count}, skipped={10 - resolved_count}')
+
+
+# ===========================================================================
+# 4) GMAIL — the alert inbox.  Status labels seeded by NAME (Rule 3: unused).
+# ===========================================================================
+GMAIL_LABELS = [
+    {'id': 'l1', 'name': 'Work', 'color': '#ef4444'},
+    {'id': 'l2', 'name': 'Personal', 'color': '#3b82f6'},
+    {'id': 'l4', 'name': 'Finance', 'color': '#eab308'},
+    # status labels the SELECTION-GATED Label button applies (matched by NAME) ---
+    {'id': 'lbl-refunded',   'name': 'Refunded',   'color': '#22c55e'},
+    {'id': 'lbl-disputed',   'name': 'Disputed',   'color': '#eb5a46'},
+    {'id': 'lbl-reconciled', 'name': 'Reconciled', 'color': '#0ea5e9'},
+]
+
+_GMAIL_STATE = {
+    'user': GMAIL_USER,
+    'emails': emails,
+    'labels': GMAIL_LABELS,
+    'drafts': [],
+    'settings': {'density': 'default', 'undoSend': 10},
+    'today': '2026-06-29',
+    '_task_adapter': {
+        'task_id': TASK_ID,
+        'variant': 'eval',
+        'resolved_count': resolved_count,
+    },
+}
+
+
+# ===========================================================================
+# 5) SLACK — #billing-ops starts EMPTY (Rule 3); decoy chatter elsewhere.
+# ===========================================================================
+_SLACK_STATE = {
+    'channels': [
+        {'channelId': 'general', 'name': 'general', 'description': 'Company-wide announcements',
+         'topic': 'Welcome to Northwind SaaS!', 'isPrivate': False, 'isStarred': True,
+         'members': ['user_1', 'user_2'], 'createdBy': 'user_2', 'createdAt': '2026-05-01T09:00:00Z',
+         'pinnedMessages': [], 'unreadCount': 0},
+        {'channelId': 'billing-ops', 'name': 'billing-ops',
+         'description': 'Billing operations — payment alerts and reconciliation',
+         'topic': 'Post a summary when you finish clearing the payment-alert inbox', 'isPrivate': False,
+         'isStarred': False, 'members': ['user_1'], 'createdBy': 'user_1',
+         'createdAt': '2026-06-01T00:00:00Z', 'pinnedMessages': [], 'unreadCount': 0},
+    ],
+    'currentUser': {'userId': 'user_1', 'fullName': 'Dana Reyes', 'displayName': 'Dana',
+                    'email': 'dana.reyes@northwind-saas.com', 'avatar': 'https://picsum.photos/200/200?random=1',
+                    'title': 'Finance Ops Analyst', 'status': 'online', 'statusMessage': '',
+                    'statusEmoji': '', 'timeZone': 'America/New_York'},
+    'users': [
+        {'userId': 'user_1', 'fullName': 'Dana Reyes', 'displayName': 'Dana',
+         'email': 'dana.reyes@northwind-saas.com', 'avatar': 'https://picsum.photos/200/200?random=1',
+         'title': 'Finance Ops Analyst', 'status': 'online', 'statusMessage': '', 'statusEmoji': '',
+         'timeZone': 'America/New_York'},
+        {'userId': 'user_2', 'fullName': 'Maya Lindqvist', 'displayName': 'Maya',
+         'email': 'maya.lindqvist@northwind-saas.com', 'avatar': 'https://picsum.photos/200/200?random=2',
+         'title': 'Controller', 'status': 'online', 'statusMessage': '', 'statusEmoji': '',
+         'timeZone': 'Europe/Stockholm'},
+    ],
+    'workspace': {'workspaceId': 'ws_1', 'workspaceName': 'Northwind SaaS', 'icon': ''},
+    'messages': {
+        'general': [
+            {'messageId': 'm_g_1', 'senderId': 'user_2',
+             'content': 'Morning all — overnight payment alerts are in the finance inbox.',
+             'timestamp': '2026-06-29T13:02:00Z', 'reactions': [], 'isEdited': False,
+             'threadId': None, 'attachments': []},
+        ],
+        # #billing-ops starts EMPTY — the gradable post + pin are the agent's work.
+        'billing-ops': [],
+    },
+    'threads': {},
+    'dms': [],
+    'bookmarkedMessages': [],
+    'callHistory': [],
+    'settings': {'theme': 'light', 'notifications': 'all', 'displayDensity': 'comfortable',
+                 'showAvatars': True, 'use24Hour': False},
+    'invitations': [],
+    'notifications': [],
+    '_task_adapter': {
+        'task_id': TASK_ID,
+        'variant': 'eval',
+        'summary_channel': SUMMARY_CHANNEL,   # agent posts + pins here (empty at injection)
+        'resolved_count': resolved_count,
+    },
+}
+
+# Rule 3 guard: #billing-ops has no messages and nothing pinned at injection.
+_bo = next(c for c in _SLACK_STATE['channels'] if c['name'] == SUMMARY_CHANNEL)
+assert _bo['pinnedMessages'] == [], 'billing-ops must have no pinned messages at injection'
+assert _SLACK_STATE['messages'].get(SUMMARY_CHANNEL) == [], 'billing-ops must be empty at injection'
+
+
+# ===========================================================================
+# 6) INJECT + VERIFY + LAUNCH  (domain order: stripe, quickbooks, gmail, slack)
+# ===========================================================================
+APP_STATES = [
+    ('http://28.7.184.198:8179', _STRIPE_STATE),   # stripe_dashboard_mock
+    ('http://28.7.184.198:8172', _QB_STATE),        # quickbooks_mock
+    ('http://28.7.184.198:8138', _GMAIL_STATE),     # gmail_mock
+    ('http://28.7.184.198:8178', _SLACK_STATE),     # slack_mock
+]
+
+
+for app_url, state in APP_STATES:
+    resp = requests.post(
+        f'{app_url}/post?sid={sid}',
+        json={'action': 'set', 'state': state},
+        timeout=30,
+    )
+    assert resp.status_code == 200, (
+        f'State injection failed for {app_url}: {resp.text}'
+    )
+    print(f'State injected: {app_url} sid={sid}')
+
+for app_url, _ in APP_STATES:
+    go_payload = requests.get(f'{app_url}/go?sid={sid}', timeout=15).json()
+    assert go_payload.get('initial_state') is not None, (
+        f'initial_state is None after injection for {app_url}'
+    )
+    print(f'Verified initial_state set: {app_url}')
+
+
+# PATCHED_BY: patch_wait_mocks_loaded.py
+def launch_gui(command, delay_sec=1.0):
+    """Launch a GUI command, fully detached from this script.
+
+    Inject --remote-debugging-port=1337 (and Chrome quiet flags) so that
+    wait_mocks_loaded's CDP phase can verify browser render below.
+    Idempotent: if the caller already specified a debug port, keep it.
+    """
+    env = os.environ.copy()
+    env['DISPLAY'] = ':0'
+    if isinstance(command, str):
+        _parts = shlex.split(command)
+    else:
+        _parts = list(command)
+    if (_parts and 'google-chrome' in _parts[0]
+            and not any(p.startswith('--remote-debugging-port') for p in _parts)):
+        _parts[1:1] = [
+            '--no-first-run',
+            '--no-default-browser-check',
+            '--remote-debugging-port=1337',
+        ]
+    subprocess.Popen(
+        _parts,
+        stdin=subprocess.DEVNULL,
+        stdout=subprocess.DEVNULL,
+        stderr=subprocess.DEVNULL,
+        env=env,
+        start_new_session=True,
+    )
+    time.sleep(delay_sec)
+
+
+def wait_mocks_loaded(app_states, sid_value, timeout=90, render_buffer=12):
+    """2-phase readiness wait, mirrors cache/ae_pipeline_hygiene_004 style.
+
+    Phase 1 - Backend readiness:
+        Poll each mock's /state?sid=<sid> (preferred) or /go?sid=<sid>
+        (fallback) until it returns 200 with 'stored_state' or
+        'initial_state' respectively, or until timeout.
+
+    Phase 2 - Browser render verification (via CDP + Playwright):
+        Connect to http://localhost:1337 (the --remote-debugging-port
+        injected by launch_gui above), wait for every open tab to reach
+        'networkidle' AND to contain at least 20 DOM elements. This
+        guarantees the SPA fetched /go?sid= data AND actually painted,
+        not just that the mock backend is up.
+
+    Fallback:
+        If Playwright is missing or CDP cannot be reached, sleep for
+        `render_buffer` seconds as a best-effort delay so the screenshot
+        loop does not start on a still-blank tab.
+    """
+    import urllib.request as _urlreq
+    import time as _t
+    _urls = [p for p, _ in app_states]
+    _opener = _urlreq.build_opener(_urlreq.ProxyHandler({}))
+
+    # ---- Phase 1: backend readiness ----
+    _deadline = _t.time() + timeout
+    _todo = list(_urls)
+    while _todo and _t.time() < _deadline:
+        _rest = []
+        for _u in _todo:
+            _ok = False
+            for _path, _need in (('/state?sid=', b'stored_state'),
+                                 ('/go?sid=',    b'initial_state')):
+                try:
+                    _r = _opener.open(f"{_u}{_path}{sid_value}", timeout=5)
+                    if _r.status == 200:
+                        _body = _r.read()
+                        if _need in _body:
+                            _ok = True
+                            break
+                except Exception:
+                    pass
+            if _ok:
+                print(f"  {_u}: backend ready")
+                continue
+            _rest.append(_u)
+        _todo = _rest
+        if _todo:
+            _t.sleep(2)
+    _done = len(_urls) - len(_todo)
+    _pend = f" (pending: {_todo})" if _todo else ""
+    print(f"  mocks backend ready: {_done}/{len(_urls)}{_pend}")
+
+    # ---- Phase 2: browser render verification via CDP ----
+    _cdp_ok = False
+    try:
+        from playwright.sync_api import sync_playwright
+        _cdp_dl = _t.time() + max(15.0, _deadline - _t.time())
+        with sync_playwright() as _p:
+            _br = None
+            while _t.time() < _cdp_dl:
+                try:
+                    _br = _p.chromium.connect_over_cdp('http://localhost:1337')
+                    break
+                except Exception:
+                    _t.sleep(0.5)
+            if _br is None:
+                print("  wait_mocks_loaded: CDP connect to :1337 failed; is Chrome up with --remote-debugging-port?")
+            else:
+                _n_expected = len(_urls)
+                _pages = []
+                _pg_dl = _t.time() + 15
+                _last_count = -1
+                _stable_since = None
+                while _t.time() < _pg_dl:
+                    _pages = []
+                    for _c in _br.contexts:
+                        _pages.extend(_c.pages)
+                    _cnt = len(_pages)
+                    if _n_expected > 0 and _cnt >= _n_expected:
+                        break
+                    if _cnt > 0 and _cnt == _last_count:
+                        if _stable_since is None:
+                            _stable_since = _t.time()
+                        elif _t.time() - _stable_since >= 1.5:
+                            break
+                    else:
+                        _stable_since = None
+                    _last_count = _cnt
+                    _t.sleep(0.5)
+                _pages = []
+                for _c in _br.contexts:
+                    _pages.extend(_c.pages)
+                if not _pages:
+                    print("  wait_mocks_loaded: CDP connected but no open tabs found")
+                else:
+                    print(f"  wait_mocks_loaded: {len(_pages)} tab(s) open (expected {_n_expected}); polling render")
+                    _MIN_ELEMS = 20
+                    _verify_dl = _t.time() + 40.0   # bounded so total setup stays < VM 120s cap
+                    for _pg in _pages:              # bring each to front once to defeat background-tab render throttling
+                        try:
+                            _pg.bring_to_front()
+                        except Exception:
+                            pass
+                    _pending = list(_pages)
+                    while _pending and _t.time() < _verify_dl:
+                        _still = []
+                        for _pg in _pending:
+                            _ttl = ""
+                            try:
+                                _ttl = _pg.title()
+                            except Exception:
+                                pass
+                            try:
+                                _n = _pg.evaluate("document.body ? document.querySelectorAll('*').length : 0")
+                            except Exception:
+                                _n = 0
+                            if _n >= _MIN_ELEMS:
+                                print(f"  wait_mocks_loaded: tab {_ttl!r} rendered ({_n} elements)")
+                            else:
+                                _still.append(_pg)
+                        _pending = _still
+                        if _pending:
+                            _t.sleep(2)
+                    if not _pending:
+                        _cdp_ok = True
+                    else:
+                        for _pg in _pending:
+                            _ttl = ""
+                            try:
+                                _ttl = _pg.title()
+                            except Exception:
+                                pass
+                            print(f"  wait_mocks_loaded: WARNING tab {_ttl!r} still not rendered after poll, will fall back")
+    except ImportError:
+        print("  wait_mocks_loaded: playwright not installed; skipping CDP phase")
+    except Exception as _e:
+        print(f"  wait_mocks_loaded: CDP phase error ({_e!r})")
+
+    if not _cdp_ok:
+        print(f"  wait_mocks_loaded: falling back to render_buffer={render_buffer}s sleep")
+        _t.sleep(render_buffer)
+    else:
+        print("  wait_mocks_loaded: browser render verified, skipping render_buffer")
+
+
+def _open_mock_tabs(app_states, sid_value):
+    """Open the first mock as the primary Chrome window, the rest as new tabs.
+
+    Uses launch_gui once per URL so that the macOS localizer regex-extract
+    (which grabs the FIRST quoted URL per call) still works correctly.
+    """
+    _urls = [p for p, _ in app_states]
+    if not _urls:
+        return
+    launch_gui(f'google-chrome "{_urls[0]}/?sid={sid_value}"', delay_sec=2.0)
+    for _u in _urls[1:]:
+        launch_gui(f'google-chrome --new-tab "{_u}/?sid={sid_value}"', delay_sec=1.0)
+
+
+
+# PATCHED_BY: add_wait_for_mocks_loaded (colleague headless+playwright render gate)
+def wait_for_mocks_loaded(mocks, sid, timeout=60.0, render_timeout=15.0):
+    """Block until every mock serves injected state AND (best-effort) renders."""
+    import time as _t
+    deadline = _t.time() + timeout
+    pending = dict(mocks)
+    while pending and _t.time() < deadline:
+        for name in list(pending):
+            try:
+                r = requests.get(f'{pending[name]}/go?sid={sid}', timeout=5)
+                if r.status_code == 200 and r.json().get('initial_state') is not None:
+                    del pending[name]
+            except Exception:
+                pass
+        if pending:
+            _t.sleep(1.0)
+    if pending:
+        print(f'WARN: mocks not ready within {timeout}s: {list(pending)}')
+    try:
+        from playwright.sync_api import sync_playwright
+    except ImportError:
+        print('WARN: playwright not installed; skipping render check, settle 3s')
+        _t.sleep(3.0)
+        return
+    with sync_playwright() as p:
+        try:
+            browser = p.chromium.launch(channel='chrome', headless=True,
+                                        args=['--no-sandbox', '--disable-gpu', '--disable-dev-shm-usage'])
+        except Exception as e:
+            print(f'WARN: could not launch chrome for render check: {e}; settle 3s')
+            _t.sleep(3.0)
+            return
+        try:
+            for name, url in mocks.items():
+                page = browser.new_page()
+                try:
+                    page.goto(f'{url}/?sid={sid}', wait_until='domcontentloaded', timeout=int(render_timeout * 1000))
+                    page.wait_for_function(
+                        "document.body && document.body.innerText.length > 50",
+                        timeout=int(render_timeout * 1000))
+                    print(f'[{name}] rendered OK')
+                except Exception as e:
+                    print(f'WARN: [{name}] render check issue: {e}')
+                finally:
+                    page.close()
+        finally:
+            browser.close()
+
+
+# ref pattern (out_benchmark_cache): open visible tabs FIRST, then block until mocks render
+_open_mock_tabs(APP_STATES, sid)
+wait_for_mocks_loaded({u.rsplit(':', 1)[-1]: u for u, _ in APP_STATES}, sid, timeout=60, render_timeout=15)
+wait_mocks_loaded(APP_STATES, sid, timeout=90, render_buffer=12)
+print(f'GUI_READY: launched browser tabs for {len(APP_STATES)} apps')

ar_payment_alert_010__long/reward.py

@@ -0,0 +1,689 @@
+"""
+Reward Script: Clear the overnight payment-alert inbox
+Task ID: ar_payment_alert_010__long
+Mocks: stripe_dashboard_mock,quickbooks_mock,gmail_mock,slack_mock
+
+Scoring (all components in [0,1]; weights sum to EXACTLY 1.0; NO subtracted
+penalties — leak is punished ONLY by precision inside the three F1
+set-membership components):
+  0.24  refund set correctness — F1(true, pred):
+        true = {stripe_payment_id of REFUND alerts that require action}
+        pred = {payment ids REFUNDED now AND NOT refunded at injection} (a diff).
+        Refunding the already-refunded distractor cannot enter pred (no diff);
+        refunding the LEAVE-row / a non-target payment IS a false positive.
+  0.22  dispute-evidence set correctness — F1(true, pred):
+        true = {stripe_dispute_id of SUBMIT_EVIDENCE alerts that require action}
+        pred = {dispute ids 'under_review' (evidence submitted) now AND NOT
+                'under_review' at injection}. Submitting evidence sets a dispute's
+        status to 'under_review' (source-verified: Disputes.jsx handleSubmitEvidence
+        dispatches UPDATE_DISPUTE -> status='under_review'). The already-under_review
+        decoys cannot enter pred (no diff); responding to the LEAVE dispute = FP.
+  0.22  receive-payment set correctness — F1(true, pred):
+        true = {qb_invoice_id of RECEIVE_PAYMENT alerts that require action}
+        pred = {invoice ids 'Paid' now AND NOT 'Paid' at injection}.
+  0.12  gmail status label — frac of require-action alerts whose CURRENT gmail
+        labels include _alert.label_name (Refunded/Disputed/Reconciled), matched
+        by NAME (id-agnostic). Emails ship labels=[] so this is 0 at do-nothing.
+  0.10  customer reply — frac of require-action alerts with a 'sent' email either
+        addressed to _alert.customer_email OR whose subject == _alert.reply_subject.
+  0.05  slack #billing-ops post: a non-empty message containing a digit.
+  0.05  slack #billing-ops pin: a pinned real #billing-ops message (gated on post).
+
+Answer key: read from gmail.initial_state.emails[*]._alert
+  .action ∈ {refund, submit_evidence, receive_payment, leave, skip}
+  .requires_action: bool
+  .stripe_payment_id | .refund_amount_cents   (refund target)
+  .stripe_dispute_id                           (submit_evidence target)
+  .qb_invoice_id | .qb_invoice_number          (receive_payment target)
+  .label_name | .customer_email | .reply_subject
+and slack.initial_state._task_adapter.summary_channel -> 'billing-ops'.
+
+Stripe amounts are CENTS, QuickBooks amounts are DOLLARS.
+"""
+import copy
+import re
+import sys
+
+import requests
+
+
+try:
+    with open('/tmp/task_web_sid', encoding='utf-8') as f:
+        sid = f.read().strip()
+    if not sid:
+        raise ValueError('sid is empty')
+except Exception as e:
+    print(f'CRITICAL: Cannot read sid from /tmp/task_web_sid: {e}')
+    print('REWARD: 0.0')
+    sys.exit(0)
+
+
+APP_URLS = {'stripe_dashboard': 'http://28.7.184.198:8179', 'quickbooks': 'http://28.7.184.198:8172',
+            'gmail': 'http://28.7.184.198:8138', 'slack': 'http://28.7.184.198:8178'}
+_GO_CACHE = {}
+
+
+def clamp01(x):
+    return max(0.0, min(1.0, float(x)))
+
+
+def frac(num, den):
+    return 0.0 if den == 0 else num / den
+
+
+def f1(true_set, pred_set):
+    if not true_set and not pred_set:
+        return 1.0
+    tp = len(true_set & pred_set)
+    if tp == 0:
+        return 0.0
+    prec = tp / len(pred_set)
+    rec = tp / len(true_set)
+    return 2 * prec * rec / (prec + rec)
+
+
+def by_id(items, key='id'):
+    return {it.get(key): it for it in (items or []) if isinstance(it, dict)}
+
+
+def has_text(haystack, *needles):
+    s = (haystack or '').lower()
+    return any(n.lower() in s for n in needles)
+
+
+def norm(s):
+    return (str(s) if s is not None else '').strip().lower()
+
+
+def _msg_text(msg):
+    if not isinstance(msg, dict):
+        return ''
+    return msg.get('content') or msg.get('text') or ''
+
+
+def _slack_channel_messages(slack_state, channel_name):
+    channels = slack_state.get('channels', []) if isinstance(slack_state, dict) else []
+    target = None
+    for ch in channels:
+        if isinstance(ch, dict) and norm(ch.get('name')) == norm(channel_name):
+            target = ch
+            break
+    if target is None:
+        return []
+    ch_msgs = target.get('messages')
+    if isinstance(ch_msgs, list):
+        return ch_msgs
+    msg_map = slack_state.get('messages', {})
+    if isinstance(msg_map, dict):
+        return msg_map.get(target.get('channelId') or target.get('id'), []) or []
+    return []
+
+
+def _to_float(v):
+    try:
+        return float(v)
+    except (TypeError, ValueError):
+        return 0.0
+
+
+def _sheet_rows(sheet):
+    if not isinstance(sheet, dict):
+        return []
+    rows = sheet.get('rows', [])
+    return rows if isinstance(rows, list) else []
+
+
+def _sheet_tabs(sheets):
+    if isinstance(sheets, dict):
+        return sheets
+    if isinstance(sheets, list):
+        out = {}
+        for sh in sheets:
+            if not isinstance(sh, dict):
+                continue
+            name = sh.get('name') or sh.get('id')
+            if name:
+                out[name] = sh
+        return out
+    return {}
+
+
+def _jaccard(a, b):
+    if not a and not b:
+        return 1.0
+    u = a | b
+    return (len(a & b) / len(u)) if u else 1.0
+
+
+def _normalize_app_name(app):
+    app = (app or '').strip().lower()
+    return app[:-5] if app.endswith('_mock') else app
+
+
+def _parse_cell_id(cell_id):
+    if not isinstance(cell_id, str) or not cell_id:
+        return None
+    idx = 0
+    while idx < len(cell_id) and cell_id[idx].isalpha():
+        idx += 1
+    if idx == 0 or idx >= len(cell_id):
+        return None
+    col = 0
+    for ch in cell_id[:idx].upper():
+        col = col * 26 + (ord(ch) - ord('A') + 1)
+    try:
+        row = int(cell_id[idx:])
+    except ValueError:
+        return None
+    return row, col - 1
+
+
+def _extract_rows_from_workbook(state):
+    if not isinstance(state, dict):
+        return None
+    sheets = state.get('sheets')
+    if not isinstance(sheets, list):
+        return None
+
+    adapter = state.get('_task_adapter') if isinstance(state.get('_task_adapter'), dict) else {}
+    headers_by_sheet = adapter.get('headers_by_sheet') if isinstance(adapter.get('headers_by_sheet'), dict) else {}
+
+    out = {}
+    for sh in sheets:
+        if not isinstance(sh, dict):
+            continue
+        name = sh.get('name') or sh.get('id') or 'Sheet1'
+        data = sh.get('data') if isinstance(sh.get('data'), dict) else {}
+
+        headers = headers_by_sheet.get(name)
+        if not headers:
+            cols = []
+            for cid, cell in data.items():
+                rc = _parse_cell_id(cid)
+                if not rc:
+                    continue
+                row_idx, col_idx = rc
+                if row_idx != 1:
+                    continue
+                value = ''
+                if isinstance(cell, dict):
+                    value = cell.get('value', '')
+                elif cell is not None:
+                    value = str(cell)
+                if value is not None and str(value) != '':
+                    cols.append((col_idx, str(value)))
+            headers = [v for _, v in sorted(cols)]
+
+        if not headers:
+            continue
+
+        by_row = {}
+        for cid, cell in data.items():
+            rc = _parse_cell_id(cid)
+            if not rc:
+                continue
+            row_idx, col_idx = rc
+            if row_idx <= 1 or col_idx >= len(headers):
+                continue
+            if isinstance(cell, dict):
+                value = cell.get('value', '')
+            else:
+                value = cell
+            by_row.setdefault(row_idx, {})[headers[col_idx]] = value
+
+        rows = []
+        for row_idx in sorted(by_row.keys()):
+            row = by_row[row_idx]
+            if not any(str(row.get(h, '')).strip() for h in headers):
+                continue
+            rows.append({h: row.get(h, '') for h in headers})
+
+        out[name] = {'headers': headers, 'rows': rows}
+
+    if not out:
+        return None
+    return {'sheets': out}
+
+
+def _materialize_google_sheets(payload):
+    for key in ('initial_state', 'current_state'):
+        state = payload.get(key)
+        if not isinstance(state, dict):
+            continue
+        maybe_rows = _extract_rows_from_workbook(state)
+        if maybe_rows is not None:
+            state['_ui_workbook'] = {
+                'sheets': copy.deepcopy(state.get('sheets', [])),
+                'title': state.get('title'),
+            }
+            state['sheets'] = maybe_rows['sheets']
+
+
+def _materialize_slack(payload):
+    for key in ('initial_state', 'current_state'):
+        state = payload.get(key)
+        if not isinstance(state, dict):
+            continue
+        channels = state.get('channels') if isinstance(state.get('channels'), list) else []
+        messages_map = state.get('messages') if isinstance(state.get('messages'), dict) else {}
+        for ch in channels:
+            if not isinstance(ch, dict):
+                continue
+            cid = ch.get('channelId') or ch.get('id')
+            if not cid:
+                continue
+            msg_list = messages_map.get(cid) if isinstance(messages_map.get(cid), list) else []
+            channel_msgs = ch.get('messages') if isinstance(ch.get('messages'), list) else []
+            merged = {
+                (m.get('messageId') or m.get('id')): dict(m)
+                for m in channel_msgs
+                if isinstance(m, dict)
+            }
+            for m in msg_list:
+                if not isinstance(m, dict):
+                    continue
+                mid = m.get('messageId') or m.get('id')
+                text = m.get('text') if m.get('text') is not None else m.get('content', '')
+                merged[mid] = {
+                    'id': mid,
+                    'messageId': mid,
+                    'text': text,
+                    'content': m.get('content', text),
+                    'senderId': m.get('senderId'),
+                    'timestamp': m.get('timestamp'),
+                    'reactions': m.get('reactions') if isinstance(m.get('reactions'), list) else [],
+                    'attachments': m.get('attachments') if isinstance(m.get('attachments'), list) else [],
+                }
+            ch['messages'] = list(merged.values())
+
+
+def _owner_name_map(state):
+    out = {}
+    users = state.get('users') if isinstance(state.get('users'), list) else []
+    for u in users:
+        if not isinstance(u, dict):
+            continue
+        uid = u.get('userId')
+        if not uid:
+            continue
+        name = (f"{u.get('firstName', '')} {u.get('lastName', '')}".strip() or u.get('name') or uid)
+        out[uid] = name
+    return out
+
+
+def _materialize_salesforce(payload):
+    for key in ('initial_state', 'current_state'):
+        state = payload.get(key)
+        if not isinstance(state, dict):
+            continue
+
+        adapter = state.get('_task_adapter') if isinstance(state.get('_task_adapter'), dict) else {}
+        owner_names = _owner_name_map(state)
+
+        task_leads = adapter.get('task_leads') if isinstance(adapter.get('task_leads'), dict) else None
+        if task_leads and isinstance(state.get('leads'), list):
+            restored = []
+            for lead in state['leads']:
+                if not isinstance(lead, dict):
+                    continue
+                lid = str(lead.get('leadId') or lead.get('id') or '')
+                meta = task_leads.get(lid, {})
+                restored.append({
+                    'id': meta.get('id') or lead.get('leadId') or lead.get('id'),
+                    'company': lead.get('company', meta.get('company', '')),
+                    'status': lead.get('status', 'New'),
+                    'ownerName': owner_names.get(lead.get('ownerId')) or meta.get('ownerName'),
+                    'annualRevenue': meta.get('annualRevenue', lead.get('revenue')),
+                })
+            state['leads'] = restored
+
+        task_accounts = adapter.get('task_accounts') if isinstance(adapter.get('task_accounts'), dict) else None
+        if task_accounts and isinstance(state.get('accounts'), list):
+            current_by_id = {}
+            for acc in state['accounts']:
+                if isinstance(acc, dict):
+                    current_by_id[str(acc.get('accountId') or acc.get('id') or '')] = acc
+
+            restored_accounts = []
+            for aid, meta in task_accounts.items():
+                cur = current_by_id.get(str(aid), {})
+                restored_accounts.append({
+                    'id': meta.get('id') or aid,
+                    'name': meta.get('name') or cur.get('name', ''),
+                    'contact': meta.get('contact'),
+                    'wonWithin7d': bool(meta.get('wonWithin7d', False)),
+                    'onboardingStage': cur.get('onboardingStage', meta.get('onboardingStage')),
+                })
+            state['accounts'] = restored_accounts
+
+
+def _materialize_google_docs(payload):
+    for key in ('initial_state', 'current_state'):
+        state = payload.get(key)
+        if not isinstance(state, dict):
+            continue
+        adapter = state.get('_task_adapter') if isinstance(state.get('_task_adapter'), dict) else {}
+        task_docs = adapter.get('task_documents') if isinstance(adapter.get('task_documents'), dict) else None
+        if not task_docs:
+            continue
+        docs_map = state.get('documents') if isinstance(state.get('documents'), dict) else {}
+        restored = []
+        for did, meta in task_docs.items():
+            cur = docs_map.get(did, {}) if isinstance(docs_map, dict) else {}
+            restored.append({
+                'id': meta.get('id') or did,
+                'title': cur.get('title', meta.get('title')),
+                'vendor': meta.get('vendor'),
+                'folder': meta.get('folder'),
+                'content': cur.get('content', meta.get('content', '')),
+            })
+        state['documents'] = restored
+
+
+def _adapt_payload_for_reward(app, payload):
+    app_name = _normalize_app_name(app)
+    if app_name == 'google_sheets':
+        _materialize_google_sheets(payload)
+    elif app_name == 'slack':
+        _materialize_slack(payload)
+    elif app_name == 'salesforce':
+        _materialize_salesforce(payload)
+    elif app_name == 'google_docs':
+        _materialize_google_docs(payload)
+    return payload
+
+
+def go(app_name):
+    if app_name in _GO_CACHE:
+        return _GO_CACHE[app_name]
+
+    base = APP_URLS.get(app_name)
+    if not base:
+        print(f'CRITICAL: unknown app {app_name!r} (no URL mapped)')
+        print('REWARD: 0.0')
+        sys.exit(0)
+
+    try:
+        resp = requests.get(f'{base}/go?sid={sid}', timeout=15)
+        resp.raise_for_status()
+        payload = resp.json()
+    except Exception as e:
+        print(f'CRITICAL: cannot fetch {app_name} go state: {e}')
+        print('REWARD: 0.0')
+        sys.exit(0)
+
+    payload = _adapt_payload_for_reward(app_name, payload)
+    _GO_CACHE[app_name] = payload
+    return payload
+
+
+# =============================================================================
+# Task-specific reward
+#
+# Answer key lives in gmail.initial_state.emails[*]._alert — one dict per alert:
+#   action ∈ {refund, submit_evidence, receive_payment, leave, skip}
+#   requires_action: bool
+#   stripe_payment_id | refund_amount_cents   (refund target)
+#   stripe_dispute_id                          (submit_evidence target)
+#   qb_invoice_id | qb_invoice_number          (receive_payment target)
+#   label_name | customer_email | reply_subject
+# slack.initial_state._task_adapter.summary_channel names the aux channel.
+# We score CURRENT state against this key using precision-aware F1 set-membership
+# for the three action sets (computed as a diff vs the injected state) — acting on
+# a distractor / LEAVE item / wrong record lowers precision but never drives the
+# score below 0.
+# =============================================================================
+def _refunded_id_set(stripe_state):
+    """Payment ids that are (fully) refunded in this Stripe state — union of the
+    payments[].refunded/amount_refunded flags and any refunds[].charge id."""
+    ids = set()
+    for p in (stripe_state.get('payments') or []):
+        if not isinstance(p, dict):
+            continue
+        pid = norm(p.get('id'))
+        if not pid:
+            continue
+        amt = p.get('amount') or 0
+        ar = p.get('amount_refunded') or 0
+        try:
+            fully = (float(amt) > 0 and float(ar) >= float(amt))
+        except (TypeError, ValueError):
+            fully = False
+        if p.get('refunded') is True or fully:
+            ids.add(pid)
+    for r in (stripe_state.get('refunds') or []):
+        if isinstance(r, dict) and r.get('charge'):
+            ids.add(norm(r.get('charge')))
+    return ids
+
+
+def _evidenced_dispute_id_set(stripe_state):
+    """Dispute ids that have had evidence submitted — status flipped to
+    'under_review' (the only status the mock's Submit-Evidence flow sets) or an
+    explicit evidence_submitted_at timestamp present."""
+    ids = set()
+    for d in (stripe_state.get('disputes') or []):
+        if not isinstance(d, dict):
+            continue
+        did = norm(d.get('id'))
+        if not did:
+            continue
+        if norm(d.get('status')) == 'under_review' or d.get('evidence_submitted_at'):
+            ids.add(did)
+    return ids
+
+
+def _paid_invoice_id_set(qb_state):
+    ids = set()
+    for inv in (qb_state.get('invoices') or []):
+        if isinstance(inv, dict) and norm(inv.get('status')) == 'paid':
+            ids.add(norm(inv.get('id')))
+    return ids
+
+
+def _recipient_emails(email):
+    """All recipient emails on a gmail email (To), normalized."""
+    out = set()
+    to = email.get('to') if isinstance(email, dict) else None
+    if isinstance(to, list):
+        for x in to:
+            if isinstance(x, dict):
+                if x.get('email'):
+                    out.add(norm(x.get('email')))
+            elif x:
+                out.add(norm(x))
+    elif isinstance(to, dict):
+        if to.get('email'):
+            out.add(norm(to.get('email')))
+    elif to:
+        out.add(norm(to))
+    return out
+
+
+def _gmail_sent_emails(gmail_state):
+    """Sent-folder emails (plus isSent/sentAt fallbacks, then sent/outbox lists)."""
+    out = []
+    if not isinstance(gmail_state, dict):
+        return out
+    for e in (gmail_state.get('emails') or []):
+        if not isinstance(e, dict):
+            continue
+        if norm(e.get('folder')) in ('sent', 'sentitems', 'sent items'):
+            out.append(e)
+            continue
+        if e.get('isSent') is True or e.get('sentAt') or e.get('sentDateTime'):
+            out.append(e)
+    if out:
+        return out
+    for key in ('sent', 'sentEmails', 'outbox'):
+        items = gmail_state.get(key, [])
+        if isinstance(items, list):
+            out.extend(x for x in items if isinstance(x, dict))
+    return out
+
+
+def reward(go):
+    W_REFUND = 0.24
+    W_DISPUTE = 0.22
+    W_RECEIVE = 0.22
+    W_LABEL = 0.12
+    W_REPLY = 0.10
+    W_POST = 0.05
+    W_PIN = 0.05
+
+    # ---- 1) Answer key from gmail initial_state.emails[*]._alert ----
+    gmail = go('gmail')
+    gm_init = gmail.get('initial_state', {}) if isinstance(gmail.get('initial_state'), dict) else {}
+    gm_cur = gmail.get('current_state', {}) if isinstance(gmail.get('current_state'), dict) else {}
+
+    emails_i = gm_init.get('emails', []) or []
+    alerts = [e for e in emails_i if isinstance(e, dict) and isinstance(e.get('_alert'), dict)]
+    if not alerts:
+        print('DEBUG_c4b39a92 fatal=no_alert_worklist total=0.0')
+        return 0.0
+
+    def _key(e):
+        return e.get('_alert', {})
+
+    require_alerts = [e for e in alerts if _key(e).get('requires_action')]
+
+    refund_true = {norm(_key(e).get('stripe_payment_id')) for e in alerts
+                   if _key(e).get('action') == 'refund' and _key(e).get('requires_action')
+                   and _key(e).get('stripe_payment_id')}
+    dispute_true = {norm(_key(e).get('stripe_dispute_id')) for e in alerts
+                    if _key(e).get('action') == 'submit_evidence' and _key(e).get('requires_action')
+                    and _key(e).get('stripe_dispute_id')}
+    receive_true = {norm(_key(e).get('qb_invoice_id')) for e in alerts
+                    if _key(e).get('action') == 'receive_payment' and _key(e).get('requires_action')
+                    and _key(e).get('qb_invoice_id')}
+
+    # ---- 2) Stripe: refund set + dispute-evidence set (diff vs injection) ----
+    stripe = go('stripe_dashboard')
+    st_init = stripe.get('initial_state', {}) if isinstance(stripe.get('initial_state'), dict) else {}
+    st_cur = stripe.get('current_state', {}) if isinstance(stripe.get('current_state'), dict) else {}
+
+    refunded_init = _refunded_id_set(st_init)
+    refunded_now = _refunded_id_set(st_cur)
+    refund_pred = refunded_now - refunded_init             # newly-refunded (diff)
+
+    evidenced_init = _evidenced_dispute_id_set(st_init)
+    evidenced_now = _evidenced_dispute_id_set(st_cur)
+    dispute_pred = evidenced_now - evidenced_init          # newly-evidenced (diff)
+
+    # ---- 3) QuickBooks: receive-payment set (diff vs injection) ----
+    qb = go('quickbooks')
+    qb_init = qb.get('initial_state', {}) if isinstance(qb.get('initial_state'), dict) else {}
+    qb_cur = qb.get('current_state', {}) if isinstance(qb.get('current_state'), dict) else {}
+    paid_init = _paid_invoice_id_set(qb_init)
+    paid_now = _paid_invoice_id_set(qb_cur)
+    receive_pred = paid_now - paid_init                    # newly-paid invoices
+
+    # F1 components (precision punishes leak; gate on a non-empty true set so a
+    # malformed/empty key can never hand out free credit at do-nothing).
+    refund_f1 = f1(refund_true, refund_pred) if refund_true else 0.0
+    dispute_f1 = f1(dispute_true, dispute_pred) if dispute_true else 0.0
+    receive_f1 = f1(receive_true, receive_pred) if receive_true else 0.0
+    s_refund = W_REFUND * refund_f1
+    s_dispute = W_DISPUTE * dispute_f1
+    s_receive = W_RECEIVE * receive_f1
+
+    # ---- 4) Gmail status label among require-action alerts (matched by NAME) ----
+    emails_c = {e.get('id'): e for e in (gm_cur.get('emails', []) or []) if isinstance(e, dict)}
+    labels_i = gm_init.get('labels', []) or []
+    labels_c = gm_cur.get('labels', []) or []
+    gmail_id_to_name = {}
+    for lab in list(labels_i) + list(labels_c):
+        if isinstance(lab, dict) and lab.get('id') is not None:
+            gmail_id_to_name[lab.get('id')] = norm(lab.get('name'))
+    label_ok = 0
+    for e in require_alerts:
+        want = norm(_key(e).get('label_name'))
+        cur = emails_c.get(e.get('id'), {})
+        got_names = {gmail_id_to_name.get(lid) for lid in (cur.get('labels') or [])}
+        # tolerate a label carried as a raw name too
+        got_names |= {norm(lid) for lid in (cur.get('labels') or [])}
+        if want and want in got_names:
+            label_ok += 1
+    label_score = frac(label_ok, len(require_alerts)) if require_alerts else 0.0
+    s_label = W_LABEL * label_score
+
+    # ---- 5) Customer reply among require-action alerts ----
+    sent = _gmail_sent_emails(gm_cur)
+    reply_ok = 0
+    for e in require_alerts:
+        want_email = norm(_key(e).get('customer_email'))
+        want_subj = norm(_key(e).get('reply_subject'))
+        matched = False
+        for s in sent:
+            if want_email and want_email in _recipient_emails(s):
+                matched = True
+                break
+            subj = norm(s.get('subject'))
+            if want_subj and subj == want_subj:
+                matched = True
+                break
+        if matched:
+            reply_ok += 1
+    reply_score = frac(reply_ok, len(require_alerts)) if require_alerts else 0.0
+    s_reply = W_REPLY * reply_score
+
+    # ---- 6) Slack #billing-ops: post (non-empty w/ a digit) + pin ----
+    slack = go('slack')
+    sl_init = slack.get('initial_state', {}) if isinstance(slack.get('initial_state'), dict) else {}
+    sl_cur = slack.get('current_state', {}) if isinstance(slack.get('current_state'), dict) else {}
+    s_adapter = sl_init.get('_task_adapter', {}) if isinstance(sl_init.get('_task_adapter'), dict) else {}
+    summary_channel = s_adapter.get('summary_channel') or 'billing-ops'
+
+    bo_msgs = _slack_channel_messages(sl_cur, summary_channel)
+    post_ok = any(re.search(r'\d', _msg_text(m)) for m in bo_msgs if isinstance(m, dict))
+    s_post = W_POST * (1.0 if post_ok else 0.0)
+
+    # Pin: at least one pinned id in #billing-ops that matches a real posted message.
+    msg_ids = set()
+    for m in bo_msgs:
+        if isinstance(m, dict):
+            mid = m.get('messageId') or m.get('id')
+            if mid is not None:
+                msg_ids.add(norm(mid))
+    pinned = []
+    for ch in (sl_cur.get('channels') or []):
+        if isinstance(ch, dict) and norm(ch.get('name')) == norm(summary_channel):
+            pinned = ch.get('pinnedMessages') or []
+            break
+    pinned_ids = set()
+    for pm in pinned:
+        if isinstance(pm, dict):
+            pid = pm.get('messageId') or pm.get('id')
+        else:
+            pid = pm
+        if pid is not None:
+            pinned_ids.add(norm(pid))
+    pin_ok = bool(post_ok and (pinned_ids & msg_ids))
+    s_pin = W_PIN * (1.0 if pin_ok else 0.0)
+
+    score = s_refund + s_dispute + s_receive + s_label + s_reply + s_post + s_pin
+
+    print(
+        'DEBUG_c4b39a92 '
+        f'refund_true={len(refund_true)} refund_pred={len(refund_pred)} refund_f1={round(refund_f1, 4)} '
+        f'dispute_true={len(dispute_true)} dispute_pred={len(dispute_pred)} dispute_f1={round(dispute_f1, 4)} '
+        f'receive_true={len(receive_true)} receive_pred={len(receive_pred)} receive_f1={round(receive_f1, 4)} '
+        f'require={len(require_alerts)} label_ok={label_ok} label_score={round(label_score, 4)} '
+        f'reply_ok={reply_ok} reply_score={round(reply_score, 4)} '
+        f'post_ok={int(post_ok)} pin_ok={int(pin_ok)} '
+        f'w_refund={round(s_refund, 4)} w_dispute={round(s_dispute, 4)} w_receive={round(s_receive, 4)} '
+        f'w_label={round(s_label, 4)} w_reply={round(s_reply, 4)} w_post={round(s_post, 4)} w_pin={round(s_pin, 4)} '
+        f'total={round(score, 4)}'
+    )
+
+    return clamp01(score)
+
+
+try:
+    score = float(reward(go))
+except Exception as exc:
+    print(f'ERROR: reward() raised {exc!r}')
+    score = 0.0
+
+score = clamp01(score)
+print(f'\nREWARD: {round(score, 4)}')

csm_escalation_001/_cua_gym_vm_bridge.sh

@@ -0,0 +1,106 @@
+#!/bin/bash
+# CUA-Gym VM-side network fix (direct-IP mock access).
+#
+# Runs INSIDE the QEMU guest. The CUA-Gym mock services are now reached by
+# their real IDC address (http://28.7.184.198:81xx) instead of the old
+# localhost:9000-9097 socat bridge, so the VM only needs two fixes:
+#
+#   1. Working DNS — the qcow2-baked /etc/resolv.conf points at the Pod's
+#      cluster DNS, which is unreachable from QEMU SLIRP NAT
+#      ("Temporary failure in name resolution").
+#   2. The mock host excluded from the woa http proxy — otherwise the VM's
+#      Chrome (gsettings proxy) and initial_setup.py's `requests` get a 403
+#      from the proxy when hitting 28.7.184.198:81xx. (The old socat path never
+#      hit this because localhost is bypassed by default.)
+#
+# Idempotent. Scoped to the CUA-Gym bridge step; does not touch the generic
+# OSWorld setup pipeline.
+#
+# Argv:
+#   $1  optional sudo password (default "password"; SetupController
+#       substitutes {CLIENT_PASSWORD} when uploaded as a cache file).
+#   $2  unused (kept for backward compat — was socat first port 9000).
+#   $3  unused (kept for backward compat — was socat last  port 9097).
+#   $4  optional mock host IP (default 28.7.184.198).
+
+set +e
+
+PASS="${1:-password}"
+MOCK_HOST="${4:-28.7.184.198}"
+# Derive the /16 IDC subnet so future port/host tweaks on the same segment
+# stay bypassed too (28.7.184.198 -> 28.7.0.0/16).
+MOCK_NET="$(echo "${MOCK_HOST}" | awk -F. 'NF==4{print $1"."$2".0.0/16"}')"
+
+run_root() {
+    echo "$PASS" | sudo -S -p "" bash -c "$1"
+}
+
+# Detect SLIRP gateway from VM's routing table; fall back to 10.0.2.2.
+GW="$(ip route 2>/dev/null | awk '/default/ {print $3; exit}')"
+GW="${GW:-10.0.2.2}"
+
+# ─── CUA-Gym network fix (Taiji-specific) ─────────────────────────────
+fix_vm_network() {
+    # 0) Discover VM network state at runtime — gateway, own IP/CIDR, SLIRP
+    #    DNS forwarder. Nothing hardcoded except the mock host above.
+    local GW_NET="${GW%.*}"
+    local SLIRP_DNS="${GW_NET}.3"        # QEMU SLIRP DNS proxy is always .3
+    local OWN_CIDR
+    OWN_CIDR="$(ip -4 -o addr show scope global 2>/dev/null \
+                | awk '{print $4; exit}')"
+    local OWN_IP="${OWN_CIDR%/*}"
+    local OWN_NET=""
+    [ -n "${OWN_IP}" ] && OWN_NET="${OWN_IP%.*}.0/24"
+
+    # 1) DNS — point at SLIRP DNS proxy + public fallbacks. chattr +i keeps
+    #    systemd-resolved/NetworkManager from clobbering it.
+    if run_root "printf 'nameserver ${SLIRP_DNS}\\nnameserver 114.114.114.114\\nnameserver 8.8.8.8\\n' > /etc/resolv.conf
+                 chattr +i /etc/resolv.conf 2>/dev/null || true"; then
+        echo "[cua_gym_vm_bridge] DNS set: ${SLIRP_DNS}, 114.114.114.114, 8.8.8.8 (gw=${GW})"
+    else
+        echo "[cua_gym_vm_bridge] WARN: failed to write /etc/resolv.conf"
+    fi
+
+    # 2) Build the proxy bypass list from detected values + the direct-IP
+    #    mock host (and its /16 subnet). THIS is what makes 28.7.184.198:81xx
+    #    reachable from inside the VM — without it the woa proxy 403s it.
+    local ignore_list="['localhost', '127.0.0.0/8', '${GW}'"
+    [ -n "${OWN_IP}" ]    && ignore_list="${ignore_list}, '${OWN_IP}'"
+    [ -n "${OWN_NET}" ]   && ignore_list="${ignore_list}, '${OWN_NET}'"
+    [ -n "${GW_NET}" ]    && ignore_list="${ignore_list}, '${GW_NET}.0/24'"
+    [ -n "${MOCK_HOST}" ] && ignore_list="${ignore_list}, '${MOCK_HOST}'"
+    [ -n "${MOCK_NET}" ]  && ignore_list="${ignore_list}, '${MOCK_NET}'"
+    ignore_list="${ignore_list}]"
+
+    # 3) GNOME proxy ignore-hosts — Chrome bypasses the woa proxy for these.
+    if gsettings set org.gnome.system.proxy ignore-hosts "${ignore_list}" 2>/tmp/gsettings_err; then
+        echo "[cua_gym_vm_bridge] gsettings ignore-hosts = ${ignore_list}"
+    else
+        echo "[cua_gym_vm_bridge] WARN: gsettings ignore-hosts failed: $(cat /tmp/gsettings_err 2>/dev/null)"
+    fi
+
+    # 4) Shell env — comma-separated form for no_proxy/NO_PROXY (requests/curl).
+    local NP="localhost,127.0.0.1,${GW}"
+    [ -n "${OWN_IP}" ]    && NP="${NP},${OWN_IP}"
+    [ -n "${OWN_NET}" ]   && NP="${NP},${OWN_NET}"
+    [ -n "${GW_NET}" ]    && NP="${NP},${GW_NET}.0/24"
+    [ -n "${MOCK_HOST}" ] && NP="${NP},${MOCK_HOST}"
+    [ -n "${MOCK_NET}" ]  && NP="${NP},${MOCK_NET}"
+
+    for var in no_proxy NO_PROXY; do
+        if ! grep -q "export ${var}=${NP}" /home/user/.bashrc 2>/dev/null; then
+            echo "export ${var}=${NP}" >> /home/user/.bashrc
+        fi
+    done
+    run_root "grep -q '^no_proxy=' /etc/environment 2>/dev/null \
+              && sed -i 's|^no_proxy=.*|no_proxy=\"${NP}\"|' /etc/environment \
+              || echo 'no_proxy=\"${NP}\"' >> /etc/environment
+              grep -q '^NO_PROXY=' /etc/environment 2>/dev/null \
+              && sed -i 's|^NO_PROXY=.*|NO_PROXY=\"${NP}\"|' /etc/environment \
+              || echo 'NO_PROXY=\"${NP}\"' >> /etc/environment" >/dev/null 2>&1
+    echo "[cua_gym_vm_bridge] no_proxy=${NP}"
+}
+
+fix_vm_network
+
+echo "[cua_gym_vm_bridge] direct-IP mode: ${MOCK_HOST} (+${MOCK_NET}) bypassed from woa proxy; no socat forwarding needed"

csm_escalation_001/initial_setup.py

@@ -0,0 +1,528 @@
+# --- MOCK_HOST_NO_PROXY_OVERRIDE: bypass woa proxy for locally-deployed mocks ---
+import os as _os
+_np = _os.environ.get("no_proxy", _os.environ.get("NO_PROXY", ""))
+for _h in ("127.0.0.1", "localhost", "28.7.184.198"):
+    if _h not in _np:
+        _np = (_np + "," + _h).lstrip(",")
+_os.environ["no_proxy"] = _np
+_os.environ["NO_PROXY"] = _np
+# --- end override ---
+
+# --- wait_mocks_loaded: wait for mock backends ready AND browser tabs rendered ---
+def wait_mocks_loaded(timeout=60, render_buffer=12):
+    """Wait for mock backends ready AND browser tabs rendered (via CDP).
+
+    Phase 1 — backend readiness: poll each mock's /state?sid= until it
+    returns 200 with 'stored_state' in the body.
+
+    Phase 2 — browser render verification: connect to Chrome's CDP
+    endpoint and wait for every open tab to reach 'networkidle' load
+    state with a non-trivial DOM element count. This confirms the SPA
+    has fetched its /go?sid= data and actually painted, not just that
+    the backend is up.
+
+    CDP port auto-detection: checks globals for _chrome_debug_port /
+    cdp_port / debug_port (set by tasks that use a dynamic port, e.g.
+    se_hard_008's launch_chrome_with_urls); defaults to 1337 for tasks
+    that use launch_gui's --remote-debugging-port injection.
+
+    Falls back to the old render_buffer sleep if Playwright or the CDP
+    endpoint is unavailable.
+    """
+    import urllib.request as _u, time as _t
+    g = globals()
+    _urls = {}
+    for _k, _v in list(g.items()):
+        if isinstance(_v, dict) and _v and all(isinstance(x, str) and x.startswith("http://") for x in _v.values()):
+            _urls.update(_v)
+        elif isinstance(_v, str) and _v.startswith("http://") and (_k.endswith("_URL") or _k.endswith("_url")):
+            _urls[_k] = _v
+    _sid = ""
+    for _k in ("sid", "SID", "task_sid", "session_id"):
+        if isinstance(g.get(_k), str) and g[_k]:
+            _sid = g[_k]; break
+    if not _sid:
+        for _k, _v in g.items():
+            if "sid" in _k.lower() and isinstance(_v, str) and _v:
+                _sid = _v; break
+    if not _urls or not _sid:
+        print("  wait_mocks_loaded: no mocks/sid found (%d urls, sid=%r), skipping" % (len(_urls), bool(_sid)))
+        return
+
+    # ---- Phase 1: backend readiness ----
+    _op = _u.build_opener(_u.ProxyHandler({}))
+    _dl = _t.time() + timeout
+    _todo = list(_urls.items())
+    while _todo and _t.time() < _dl:
+        _rest = []
+        for _n2, _uu in _todo:
+            try:
+                _r = _op.open("%s/state?sid=%s" % (_uu, _sid), timeout=5)
+                if _r.status == 200 and b"stored_state" in _r.read():
+                    print("  %s: backend ready" % _n2); continue
+            except Exception:
+                pass
+            _rest.append((_n2, _uu))
+        _todo = _rest
+        if _todo:
+            _t.sleep(2)
+    _done = len(_urls) - len(_todo)
+    _pend = (" (pending: %s)" % [n for n, _ in _todo]) if _todo else ""
+    print("  mocks backend ready: %d/%d%s" % (_done, len(_urls), _pend))
+
+    # ---- Phase 2: browser render verification via CDP ----
+    # Auto-detect CDP port from globals (set by launch_chrome_with_urls
+    # in tasks that use a dynamic port); default to 1337.
+    _cdp_port = 1337
+    for _pk in ("_chrome_debug_port", "chrome_debug_port", "cdp_port", "debug_port"):
+        _pv = g.get(_pk)
+        if isinstance(_pv, int) and 0 < _pv < 65536:
+            _cdp_port = _pv; break
+    _cdp_ok = False
+    try:
+        from playwright.sync_api import sync_playwright
+        _cdp_dl = _t.time() + max(10.0, _dl - _t.time())  # leftover budget, floor 10s
+        with sync_playwright() as _p:
+            _br = None
+            # Retry-connect: Chrome may still be starting up.
+            while _t.time() < _cdp_dl:
+                try:
+                    _br = _p.chromium.connect_over_cdp("http://localhost:%d" % _cdp_port)
+                    break
+                except Exception:
+                    _t.sleep(0.5)
+            if _br is None:
+                print("  wait_mocks_loaded: CDP connect to localhost:%d failed; "
+                      "is Chrome launched with --remote-debugging-port?" % _cdp_port)
+            else:
+                # Wait for ALL expected tabs to appear, then verify each renders.
+                _n_expected = len(_urls)
+                _pages = []
+                _pg_dl = _t.time() + 15
+                _last_count = -1
+                _stable_since = None
+                while _t.time() < _pg_dl:
+                    _pages = []
+                    for _c in _br.contexts:
+                        _pages.extend(_c.pages)
+                    _cnt = len(_pages)
+                    if _n_expected > 0 and _cnt >= _n_expected:
+                        break  # all expected tabs are open
+                    if _cnt > 0 and _cnt == _last_count:
+                        if _stable_since is None:
+                            _stable_since = _t.time()
+                        elif _t.time() - _stable_since >= 1.5:
+                            break  # tab count stable -> no more tabs opening
+                    else:
+                        _stable_since = None
+                    _last_count = _cnt
+                    _t.sleep(0.5)
+                # Re-fetch final snapshot so late-opening tabs are included.
+                _pages = []
+                for _c in _br.contexts:
+                    _pages.extend(_c.pages)
+                if not _pages:
+                    print("  wait_mocks_loaded: CDP connected but no open tabs found")
+                else:
+                    print("  wait_mocks_loaded: %d tab(s) open (expected %d); "
+                          "verifying render" % (len(_pages), _n_expected))
+                    _ok_cnt = 0
+                    for _pg in _pages:
+                        _ttl = ""
+                        try:
+                            _ttl = _pg.title()
+                        except Exception:
+                            pass
+                        try:
+                            _pg.wait_for_load_state(
+                                "networkidle",
+                                timeout=int(max(5000, (_cdp_dl - _t.time()) * 1000)))
+                            _n = _pg.evaluate("document.querySelectorAll('*').length")
+                            if _n < 20:
+                                print("  wait_mocks_loaded: WARNING tab %r has only %d "
+                                      "elements (render stall?)" % (_ttl, _n))
+                            else:
+                                print("  wait_mocks_loaded: tab %r rendered (%d elements)"
+                                      % (_ttl, _n))
+                                _ok_cnt += 1
+                        except Exception as _e:
+                            print("  wait_mocks_loaded: tab %r render wait failed: %r"
+                                  % (_ttl, _e))
+                    # Require ALL open tabs to render before skipping the
+                    # render_buffer fallback, so no tab is left half-loaded.
+                    if _ok_cnt > 0 and _ok_cnt == len(_pages):
+                        _cdp_ok = True
+                    else:
+                        print("  wait_mocks_loaded: %d/%d tabs rendered, will fall back"
+                              % (_ok_cnt, len(_pages)))
+    except ImportError:
+        print("  wait_mocks_loaded: playwright not installed; cannot verify browser render")
+    except Exception as _e:
+        print("  wait_mocks_loaded: CDP phase error (%r)" % _e)
+
+    if not _cdp_ok:
+        print("  wait_mocks_loaded: falling back to render_buffer=%ds sleep" % render_buffer)
+        _t.sleep(render_buffer)
+    else:
+        print("  wait_mocks_loaded: browser render verified, skipping render_buffer")
+# --- end wait_mocks_loaded ---
+
+# --- _open_remaining_mock_tabs: open all auto-detected mock URLs in new Chrome tabs ---
+def _open_remaining_mock_tabs(primary_url=""):
+    """Open all auto-detected mock URLs in Chrome --new-tab, except primary_url."""
+    g = globals()
+    _urls = set()
+    for _k, _v in list(g.items()):
+        if "PROBE" in _k.upper() or _k.startswith("_"):
+            continue
+        if isinstance(_v, str) and _v.startswith("http://") and (_k.endswith("_URL") or _k.endswith("_url")):
+            _urls.add(_v)
+        elif isinstance(_v, dict) and _v and all(isinstance(x, str) and x.startswith("http://") for x in _v.values()):
+            _urls.update(_v.values())
+    _sid = ""
+    for _k in ("sid", "SID", "task_sid", "session_id"):
+        if isinstance(g.get(_k), str) and g[_k]:
+            _sid = g[_k]; break
+    for _u in sorted(_urls):
+        if _u != primary_url:
+            launch_gui(f'google-chrome --new-tab "{_u}/?sid={_sid}"', delay_sec=1.0)
+# --- end _open_remaining_mock_tabs ---
+
+
+"""
+Initial Setup: CSM escalation -> tracked bug across Slack/Jira/Salesforce/Notion
+Task ID: csm_escalation_001
+Domain: mock_websites
+Mocks: slack_mock, jira_mock, salesforce_mock, notion_mock
+
+NOTE on sid: initial_env and golden_env are SEPARATE VMs that do NOT share /tmp.
+We therefore use a DETERMINISTIC sid derived from the task id. This script owns the
+*_initial sid; golden_patch.py owns the *_golden sid. Both also write their sid to
+/tmp/task_web_sid so reward.py can read it on whichever VM it runs on.
+"""
+import os
+import shlex
+import subprocess
+import time
+
+import requests
+
+TASK_ID = 'csm_escalation_004'
+SID = f'cuagym-{TASK_ID}-initial'
+
+MOCKS = {
+    'slack_mock': 'http://28.7.184.198:8178',
+    'jira_mock': 'http://28.7.184.198:8153',
+    'salesforce_mock': 'http://28.7.184.198:8175',
+    'notion_mock': 'http://28.7.184.198:8166',
+}
+PRIMARY_URL = MOCKS['slack_mock']
+
+with open('/tmp/task_web_sid', 'w') as f:
+    f.write(SID)
+print(f'sid={SID}')
+
+# --- Egress proxy helper (MANDATORY) ---
+PROXY_CANDIDATES = [
+    os.environ.get('HTTPS_PROXY') or os.environ.get('https_proxy'),
+        None,
+]
+
+
+def resolve_proxy(probe_url):
+    try:
+        requests.get(probe_url, timeout=8)
+        return None
+    except Exception:
+        pass
+    for candidate in PROXY_CANDIDATES:
+        if not candidate:
+            continue
+        try:
+            requests.get(probe_url, timeout=12, proxies={'http': candidate, 'https': candidate})
+            return candidate
+        except Exception:
+            continue
+    raise RuntimeError('No working route to mock servers; direct and proxy attempts failed')
+
+
+PROXY = resolve_proxy(f'{PRIMARY_URL}/go?sid=conn-probe')
+PROXIES = None
+print(f'Egress proxy: {PROXY or "direct"}')
+
+
+# ===========================================================================
+# BASELINE BUILDERS (shared, conceptually, with golden_patch.py)
+# ===========================================================================
+def build_slack_state():
+    users = [
+        {"userId": "user_1", "fullName": "Jordan Rivera", "displayName": "Jordan", "email": "jordan.rivera@northwindcloud.com", "avatar": "https://picsum.photos/200/200?random=1", "title": "Customer Success Manager", "status": "online", "statusMessage": "", "statusEmoji": "", "timeZone": "America/New_York"},
+        {"userId": "user_2", "fullName": "Alex Park", "displayName": "Alex", "email": "alex.park@northwindcloud.com", "avatar": "https://picsum.photos/200/200?random=2", "title": "Engineering Lead", "status": "online", "statusMessage": "", "statusEmoji": "", "timeZone": "America/New_York"},
+        {"userId": "user_3", "fullName": "Priya Anand", "displayName": "Priya", "email": "priya.anand@northwindcloud.com", "avatar": "https://picsum.photos/200/200?random=3", "title": "Support Lead", "status": "online", "statusMessage": "", "statusEmoji": "", "timeZone": "America/Chicago"},
+        {"userId": "user_4", "fullName": "Diego Morales", "displayName": "Diego", "email": "diego.morales@northwindcloud.com", "avatar": "https://picsum.photos/200/200?random=4", "title": "Solutions Engineer", "status": "away", "statusMessage": "", "statusEmoji": "", "timeZone": "America/Los_Angeles"},
+        {"userId": "user_5", "fullName": "Hannah Brooks", "displayName": "Hannah", "email": "hannah.brooks@northwindcloud.com", "avatar": "https://picsum.photos/200/200?random=5", "title": "Product Manager", "status": "online", "statusMessage": "", "statusEmoji": "", "timeZone": "America/New_York"},
+        {"userId": "user_6", "fullName": "Tom Becker", "displayName": "Tom", "email": "tom.becker@cedarworks.example.com", "avatar": "https://picsum.photos/200/200?random=6", "title": "Ops Manager (Cedar Works)", "status": "online", "statusMessage": "Guest", "statusEmoji": "", "timeZone": "America/Chicago"},
+        {"userId": "user_7", "fullName": "Maria Vance", "displayName": "Maria", "email": "maria.vance@vertexmfg.example.com", "avatar": "https://picsum.photos/200/200?random=7", "title": "Admin (Vertex Manufacturing)", "status": "online", "statusMessage": "Guest", "statusEmoji": "", "timeZone": "America/New_York"},
+    ]
+    current = next(u for u in users if u['userId'] == 'user_1')
+
+    def channel(cid, name, desc, members, topic=""):
+        return {"channelId": cid, "name": name, "description": desc, "topic": topic, "isPrivate": False,
+                "isStarred": False, "members": members, "createdBy": "user_1",
+                "createdAt": "2026-01-04T09:00:00Z", "pinnedMessages": [], "unreadCount": 0}
+
+    am = ["user_1", "user_2", "user_3", "user_4", "user_5"]
+    channels = [
+        channel("general", "general", "Company-wide announcements", am),
+        channel("random", "random", "Non-work banter", am),
+        channel("engineering", "engineering", "Engineering team", ["user_1", "user_2", "user_4"]),
+        channel("design", "design", "Design team", ["user_1", "user_5"]),
+        channel("marketing", "marketing", "Marketing team", ["user_1", "user_5"]),
+        channel("project-alpha", "project-alpha", "Project Alpha workstream", ["user_1", "user_2"]),
+        channel("acct-vertex", "acct-vertex", "Shared channel with customer Vertex Manufacturing",
+                ["user_1", "user_3", "user_7"], topic="Vertex Manufacturing <> Northwind Cloud"),
+        channel("acct-cedar", "acct-cedar", "Shared channel with customer Cedar Works",
+                ["user_1", "user_3", "user_6"], topic="Cedar Works <> Northwind Cloud"),
+    ]
+    messages = {
+        "general": [
+            {"messageId": "msg_g1", "senderId": "user_5", "content": "Reminder: quarterly customer reviews are due Friday.", "timestamp": "2026-06-22T13:00:00Z", "threadId": None, "reactions": [], "attachments": [], "isEdited": False},
+            {"messageId": "msg_g2", "senderId": "user_2", "content": "Deploy window for v4.2 is tomorrow 6pm ET.", "timestamp": "2026-06-23T15:30:00Z", "threadId": None, "reactions": [], "attachments": [], "isEdited": False},
+        ],
+        "random": [
+            {"messageId": "msg_r1", "senderId": "user_4", "content": "Coffee machine on 3rd floor is fixed.", "timestamp": "2026-06-23T10:00:00Z", "threadId": None, "reactions": [], "attachments": [], "isEdited": False},
+        ],
+        "engineering": [
+            {"messageId": "msg_e1", "senderId": "user_2", "content": "Staging integration latency is being investigated, low impact.", "timestamp": "2026-06-23T11:00:00Z", "threadId": None, "reactions": [], "attachments": [], "isEdited": False},
+        ],
+        "design": [
+            {"messageId": "msg_d1", "senderId": "user_5", "content": "New dashboard mockups in Figma.", "timestamp": "2026-06-22T16:00:00Z", "threadId": None, "reactions": [], "attachments": [], "isEdited": False},
+        ],
+        "marketing": [
+            {"messageId": "msg_m1", "senderId": "user_5", "content": "Case study draft ready for review.", "timestamp": "2026-06-21T14:00:00Z", "threadId": None, "reactions": [], "attachments": [], "isEdited": False},
+        ],
+        "project-alpha": [
+            {"messageId": "msg_pa1", "senderId": "user_2", "content": "Alpha milestone 2 on track.", "timestamp": "2026-06-20T09:00:00Z", "threadId": None, "reactions": [], "attachments": [], "isEdited": False},
+        ],
+        "acct-vertex": [
+            {"messageId": "msg_v1", "senderId": "user_1", "content": "Hi Maria, welcome to our shared channel! Reach out here any time.", "timestamp": "2026-06-18T09:15:00Z", "threadId": None, "reactions": [], "attachments": [], "isEdited": False},
+            {"messageId": "msg_v2", "senderId": "user_7", "content": "Thanks Jordan, this is great. Things have been running smoothly so far.", "timestamp": "2026-06-18T09:20:00Z", "threadId": None, "reactions": [], "attachments": [], "isEdited": False},
+            {"messageId": "msg_v3", "senderId": "user_7", "content": "Our Salesforce↔Northwind integration has been DOWN since this morning — every sync fails with a 500 error and it's blocking our ops team from pushing orders. Steps to reproduce: (1) open Integrations, (2) click Sync Now, (3) it errors out with a 500. This is urgent, can someone help?", "timestamp": "2026-06-24T13:05:00Z", "threadId": None, "reactions": [], "attachments": [], "isEdited": False},
+        ],
+        "acct-cedar": [
+            {"messageId": "msg_c1", "senderId": "user_6", "content": "Hi team — not urgent at all, but it would be nice to have a dark mode option in the reporting view someday. Just a feature idea for the backlog.", "timestamp": "2026-06-24T10:30:00Z", "threadId": None, "reactions": [], "attachments": [], "isEdited": False},
+        ],
+    }
+    return {
+        "currentUser": current,
+        "workspace": {"workspaceId": "ws_northwind", "workspaceName": "Northwind Cloud", "icon": ""},
+        "users": users,
+        "channels": channels,
+        "messages": messages,
+        "threads": {},
+        "dms": [],
+        "bookmarkedMessages": [],
+        "callHistory": [],
+        "settings": {"theme": "light", "notifications": "all", "displayDensity": "comfortable", "showAvatars": True, "use24Hour": False},
+        "invitations": [],
+        "notifications": [],
+    }
+
+
+def build_jira_state():
+    users = [
+        {"id": "u1", "name": "Jordan Rivera", "email": "jordan.rivera@northwindcloud.com", "avatar": "https://picsum.photos/100/100?random=u1"},
+        {"id": "u2", "name": "Alex Park", "email": "alex.park@northwindcloud.com", "avatar": "https://picsum.photos/100/100?random=u2"},
+        {"id": "u3", "name": "Priya Anand", "email": "priya.anand@northwindcloud.com", "avatar": "https://picsum.photos/100/100?random=u3"},
+        {"id": "u4", "name": "Diego Morales", "email": "diego.morales@northwindcloud.com", "avatar": "https://picsum.photos/100/100?random=u4"},
+    ]
+    projects = [
+        {"id": "p1", "key": "KAN", "name": "Kanban Project", "leadId": "u1", "category": "Software", "icon": "https://picsum.photos/64/64?random=p1"},
+        {"id": "p2", "key": "SCRUM", "name": "Scrum Alpha", "leadId": "u2", "category": "Software", "icon": "https://picsum.photos/64/64?random=p2"},
+    ]
+    sprints = [
+        {"id": "s1", "projectId": "p2", "name": "Sprint 7", "goal": "Stabilize integrations", "startDate": "2026-06-16T12:00:00.000Z", "endDate": "2026-06-30T12:00:00.000Z", "state": "active"},
+        {"id": "s2", "projectId": "p2", "name": "Sprint 8", "goal": "Reporting improvements", "startDate": "2026-07-01T12:00:00.000Z", "endDate": "2026-07-14T12:00:00.000Z", "state": "future"},
+    ]
+
+    def issue(iid, key, pid, summary, desc, itype, status, priority, sp, reporter, assignee, sprint=None, labels=None):
+        return {"id": iid, "key": key, "projectId": pid, "summary": summary, "description": desc,
+                "type": itype, "status": status, "priority": priority, "storyPoints": sp,
+                "reporterId": reporter, "assigneeId": assignee, "sprintId": sprint, "epicId": None,
+                "labels": labels or [], "subtasks": [], "linkedIssueIds": [],
+                "createdAt": "2026-06-15T12:00:00.000Z", "updatedAt": "2026-06-20T12:00:00.000Z"}
+
+    issues = [
+        issue("i1", "KAN-1", "p1", "Set up CI/CD pipeline", "Configure build and deploy automation.", "Story", "In Progress", "High", 5, "u1", "u2"),
+        issue("i2", "KAN-2", "p1", "Add SSO login support", "Support SAML-based single sign-on.", "Story", "To Do", "Medium", 8, "u1", "u4"),
+        issue("i3", "KAN-3", "p1", "Improve onboarding docs", "Rewrite the getting-started guide.", "Task", "To Do", "Low", 3, "u3", None),
+        issue("i4", "SCRUM-1", "p2", "Integration latency on staging", "Sync jobs on the STAGING environment are slower than expected (around 2-3s added latency). Low customer impact; investigate query plan. This is NOT a production outage.", "Bug", "In Progress", "Low", 2, "u2", "u4", sprint="s1", labels=["staging", "performance"]),
+        issue("i5", "SCRUM-2", "p2", "Export to CSV button missing", "The CSV export button disappeared on the reports page.", "Bug", "To Do", "Medium", 3, "u3", "u2", sprint="s1"),
+        issue("i6", "SCRUM-3", "p2", "Dark mode for reports", "Add a dark theme to the reporting view.", "Story", "To Do", "Low", 5, "u3", None, sprint="s2"),
+    ]
+    return {
+        "currentUser": users[0],
+        "users": users,
+        "projects": projects,
+        "sprints": sprints,
+        "issues": issues,
+        "comments": [],
+        "workflows": [{"id": "w1", "name": "Software Workflow", "transitions": [
+            {"from": "To Do", "to": ["In Progress"]},
+            {"from": "In Progress", "to": ["In Review", "To Do", "Done"]},
+            {"from": "In Review", "to": ["Done", "In Progress"]},
+            {"from": "Done", "to": ["In Progress", "To Do"]}]}],
+        "notifications": [],
+    }
+
+
+def build_salesforce_state():
+    users = [
+        {"userId": "user-1", "firstName": "Jordan", "lastName": "Rivera", "email": "jordan.rivera@northwindcloud.com", "phone": "(555) 123-4567", "title": "Customer Success Manager", "department": "Customer Success", "role": "CSM", "avatar": "https://i.pravatar.cc/150?u=user-1", "timezone": "America/New_York", "locale": "en-US", "theme": "lightning"},
+        {"userId": "user-2", "firstName": "Emma", "lastName": "Wilson", "email": "emma.wilson@northwindcloud.com", "phone": "(555) 234-5678", "title": "Account Executive", "department": "Sales", "role": "Rep", "avatar": "https://i.pravatar.cc/150?u=user-2", "timezone": "America/New_York", "locale": "en-US", "theme": "lightning"},
+        {"userId": "user-3", "firstName": "Priya", "lastName": "Anand", "email": "priya.anand@northwindcloud.com", "phone": "(555) 345-6789", "title": "Support Lead", "department": "Support", "role": "Support", "avatar": "https://i.pravatar.cc/150?u=user-3", "timezone": "America/Chicago", "locale": "en-US", "theme": "lightning"},
+        {"userId": "user-4", "firstName": "Sarah", "lastName": "Davis", "email": "sarah.davis@northwindcloud.com", "phone": "(555) 456-7890", "title": "Support Engineer", "department": "Support", "role": "Support", "avatar": "https://i.pravatar.cc/150?u=user-4", "timezone": "America/Los_Angeles", "locale": "en-US", "theme": "lightning"},
+        {"userId": "user-5", "firstName": "David", "lastName": "Brown", "email": "david.brown@northwindcloud.com", "phone": "(555) 567-8901", "title": "Sales Rep", "department": "Sales", "role": "Rep", "avatar": "https://i.pravatar.cc/150?u=user-5", "timezone": "America/Denver", "locale": "en-US", "theme": "lightning"},
+    ]
+
+    def account(aid, name, industry, owner, desc):
+        return {"accountId": aid, "name": name, "phone": "(555) 100-2000", "website": "https://example.com",
+                "type": "Customer", "industry": industry, "revenue": 30000000, "employees": 400,
+                "description": desc, "ownerId": owner,
+                "billingStreet": "100 Main St", "billingCity": "Detroit", "billingState": "MI", "billingZip": "48201", "billingCountry": "United States",
+                "shippingStreet": "100 Main St", "shippingCity": "Detroit", "shippingState": "MI", "shippingZip": "48201", "shippingCountry": "United States",
+                "createdDate": "2025-02-01T00:00:00.000Z", "modifiedDate": "2026-05-01T00:00:00.000Z"}
+
+    accounts = [
+        account("account-1", "Vertex Manufacturing", "Manufacturing", "user-1", "Key mid-market customer; uses the Salesforce<->Northwind integration heavily."),
+        account("account-2", "Vertex Logistics", "Logistics", "user-2", "Unrelated prospect with a similar name; no active integration issues."),
+        account("account-3", "Cedar Works", "Consumer Goods", "user-1", "Mid-market customer; stable account."),
+        account("account-4", "Summit Retail Group", "Retail", "user-5", "Growing retail customer."),
+    ]
+    contacts = [
+        {"contactId": "contact-1", "accountId": "account-1", "firstName": "Maria", "lastName": "Vance", "title": "IT Admin", "department": "Operations", "email": "maria.vance@vertexmfg.example.com", "phone": "(555) 700-1000", "ownerId": "user-1"},
+        {"contactId": "contact-2", "accountId": "account-2", "firstName": "Greg", "lastName": "Hall", "title": "Operations Director", "department": "Operations", "email": "greg.hall@vertexlog.example.com", "phone": "(555) 700-2000", "ownerId": "user-2"},
+        {"contactId": "contact-3", "accountId": "account-3", "firstName": "Tom", "lastName": "Becker", "title": "Ops Manager", "department": "Operations", "email": "tom.becker@cedarworks.example.com", "phone": "(555) 700-3000", "ownerId": "user-1"},
+    ]
+
+    def case(cid, num, subject, status, priority, acct, contact, owner, ctype, desc):
+        return {"caseId": cid, "caseNumber": num, "subject": subject, "status": status, "priority": priority,
+                "origin": "Email", "type": ctype, "accountId": acct, "contactId": contact,
+                "description": desc, "ownerId": owner,
+                "createdDate": "2026-06-20T11:00:00.000Z", "modifiedDate": "2026-06-22T11:00:00.000Z", "closedDate": None}
+
+    cases = [
+        case("case-1", "00001001", "Dashboard access error", "New", "Medium", "account-3", "contact-3", "user-4", "Problem", "Customer reports occasional 403 on the analytics dashboard."),
+        case("case-2", "00001002", "Billing question on latest invoice", "Working", "Low", "account-4", None, "user-3", "Question", "Customer asks for a breakdown of usage charges."),
+        case("case-3", "00001003", "Feature request: bulk export", "New", "Low", "account-2", "contact-2", "user-2", "Feature Request", "Prospect would like a bulk export option."),
+    ]
+    return {
+        "user": users[0],
+        "users": users,
+        "leads": [],
+        "accounts": accounts,
+        "contacts": contacts,
+        "opportunities": [],
+        "cases": cases,
+        "activities": [],
+        "chatterPosts": [],
+        "files": [],
+        "following": ["user-2", "user-3"],
+        "recentlyViewed": [],
+        "dismissedNotifications": [],
+    }
+
+
+def build_notion_state():
+    HEALTH_DB_ID = "db-health-tracker"
+    HOME_PAGE = "page-home"
+    P_HEALTH, P_MAU, P_NOTE, P_OWNER = "prop-health", "prop-mau", "prop-risknote", "prop-owner"
+    ROW_VERTEX, ROW_CEDAR, ROW_SUMMIT = "row-vertex", "row-cedar", "row-summit"
+
+    def health_row(rid, name, health, mau, note):
+        return {"id": rid, "title": name, "icon": "\U0001f3e2", "cover": None, "parentId": HEALTH_DB_ID,
+                "blockIds": [], "favorite": False, "createdDate": "2026-06-01T09:00:00.000Z",
+                "lastEditedDate": "2026-06-20T09:00:00.000Z",
+                "properties": {P_HEALTH: health, P_MAU: mau, P_NOTE: note, P_OWNER: ["user-1"]}}
+
+    pages = {
+        HOME_PAGE: {"id": HOME_PAGE, "title": "Customer Success Home", "icon": "\U0001f4cb", "cover": None,
+                    "parentId": None, "blockIds": [], "favorite": True,
+                    "createdDate": "2026-05-01T09:00:00.000Z", "lastEditedDate": "2026-06-20T09:00:00.000Z", "properties": {}},
+        HEALTH_DB_ID: {"id": HEALTH_DB_ID, "title": "Customer Health Tracker", "icon": "\U0001f4ca", "cover": None,
+                       "parentId": None, "type": "database", "viewType": "table",
+                       "properties": [
+                           {"id": P_HEALTH, "name": "Health", "type": "select", "options": ["Green", "Yellow", "Red"]},
+                           {"id": P_MAU, "name": "MAU Trend", "type": "select", "options": ["Up", "Flat", "Down"]},
+                           {"id": P_NOTE, "name": "Risk Note", "type": "text"},
+                           {"id": P_OWNER, "name": "CSM", "type": "person"}],
+                       "views": [{"id": "view-1", "name": "All Accounts", "type": "table", "filters": [], "sorts": [], "groupBy": None, "visibleProperties": [P_HEALTH, P_MAU, P_NOTE, P_OWNER]}],
+                       "items": [ROW_VERTEX, ROW_CEDAR, ROW_SUMMIT],
+                       "blockIds": [], "favorite": True, "createdDate": "2026-05-01T09:00:00.000Z"},
+        ROW_VERTEX: health_row(ROW_VERTEX, "Vertex Manufacturing", "Yellow", "Flat", ""),
+        ROW_CEDAR: health_row(ROW_CEDAR, "Cedar Works", "Green", "Up", "Stable; expansion conversation in Q3."),
+        ROW_SUMMIT: health_row(ROW_SUMMIT, "Summit Retail Group", "Green", "Up", "Onboarding completed."),
+    }
+    return {
+        "user": {"id": "user-1", "name": "Jordan Rivera", "email": "jordan.rivera@northwindcloud.com", "avatar": ""},
+        "workspace": {"id": "ws-1", "name": "Northwind Cloud", "icon": "", "members": ["user-1", "user-2", "user-3"]},
+        "pages": pages,
+        "blocks": {},
+        "trash": [],
+        "comments": {},
+        "settings": {"appearance": "light", "startWeekMonday": False, "fontSize": "default"},
+        "notifications": [],
+        "pageOrder": [HOME_PAGE, HEALTH_DB_ID],
+    }
+
+
+STATES = {
+    'slack_mock': build_slack_state(),
+    'jira_mock': build_jira_state(),
+    'salesforce_mock': build_salesforce_state(),
+    'notion_mock': build_notion_state(),
+}
+
+# --- Inject all (reset first so re-runs are clean) ---
+for name, url in MOCKS.items():
+    requests.post(f'{url}/post?sid={SID}', json={'action': 'reset'}, timeout=60, proxies=PROXIES)
+    resp = requests.post(f'{url}/post?sid={SID}', json={'action': 'set', 'state': STATES[name]}, timeout=30, proxies=PROXIES)
+    assert resp.status_code == 200, f'{name} injection failed: {resp.status_code} {resp.text}'
+    go = requests.get(f'{url}/go?sid={SID}', timeout=15, proxies=PROXIES).json()
+    assert go.get('initial_state') is not None, f'{name} initial_state is None after injection'
+    print(f'State injected: {name} sid={SID}')
+
+
+# --- Launch browser on primary mock (Slack) ---
+def launch_gui(command, delay_sec=1.0):
+    env = os.environ.copy()
+    env['DISPLAY'] = ':0'
+    # Accept both a command string and a pre-split arg list (some tasks
+    # build arg lists themselves, e.g. se_hard_008).
+    if isinstance(command, str):
+        _parts = shlex.split(command)
+    else:
+        _parts = list(command)
+    # Inject CDP debug port so wait_mocks_loaded can verify browser render
+    # via Playwright connect_over_cdp. Idempotent: skip if any arg already
+    # starts with --remote-debugging-port (some tasks inject their own
+    # dynamic port, e.g. se_hard_008).
+    if _parts and 'google-chrome' in _parts[0] \
+       and not any(p.startswith('--remote-debugging-port') for p in _parts):
+        _parts[1:1] = ['--no-first-run', '--no-default-browser-check',
+                       '--remote-debugging-port=1337']
+    _proc = subprocess.Popen(_parts, stdin=subprocess.DEVNULL,
+                             stdout=subprocess.DEVNULL,
+                             stderr=subprocess.DEVNULL, env=env,
+                             start_new_session=True)
+    time.sleep(delay_sec)
+    if _proc.poll() is not None:
+        print('  WARNING: launch_gui process exited early (code=%s)'
+              % _proc.returncode)
+    return _proc
+chrome_proxy = f'--proxy-server={PROXY}' if PROXY else ''
+_primary_mock_url = PRIMARY_URL
+launch_gui(f'google-chrome "{PRIMARY_URL}/?sid={SID}"', delay_sec=2.0)
+_open_remaining_mock_tabs(_primary_mock_url)
+wait_mocks_loaded()
+print(f'GUI_READY: launched browser at {PRIMARY_URL}/?sid={SID}')

csm_escalation_001/reward.py

@@ -0,0 +1,466 @@
+# --- MOCK_HOST_NO_PROXY_OVERRIDE: bypass woa proxy for locally-deployed mocks ---
+import os as _os
+_np = _os.environ.get("no_proxy", _os.environ.get("NO_PROXY", ""))
+for _h in ("127.0.0.1", "localhost", "28.7.184.198"):
+    if _h not in _np:
+        _np = (_np + "," + _h).lstrip(",")
+_os.environ["no_proxy"] = _np
+_os.environ["NO_PROXY"] = _np
+# --- end override ---
+
+"""
+Reward Script: CSM escalation -> tracked engineering bug + SF case + Notion health + Slack reply
+Task ID: csm_escalation_001
+Domain: mock_websites (multi-mock: slack, jira, salesforce, notion)
+Scoring (sums to 1.0, all components measure task-introduced changes only):
+  - Component 1 (0.30): NEW Jira issue naming the Vertex integration incident, with repro
+                        evidence, high priority, a real status, and a real project.
+  - Component 2 (0.30): NEW Salesforce case on account-1 (Vertex Manufacturing),
+                        priority High/Critical, and an available active/completed status.
+  - Component 3 (0.20): Notion 'Vertex Manufacturing' health row -> Health 'Red' AND non-empty
+                        Risk Note.
+  - Component 4a (0.10): NEW Slack reply by the CSM (user_1) in 'acct-vertex' (programmatic).
+  - Component 4b (0.10): That reply acknowledges the outage AND states a concrete next-update
+                        time/ETA (programmatic pattern detection).
+All 1.0 is programmatic — fully deterministic and reproducible.
+"""
+import json
+import os
+import re
+import sys
+from datetime import datetime, timedelta
+
+import requests
+
+# --- Read sid ---
+try:
+    with open('/tmp/task_web_sid') as f:
+        sid = f.read().strip()
+    if not sid:
+        raise ValueError('sid is empty')
+except Exception as e:
+    print(f'CRITICAL: Cannot read sid from /tmp/task_web_sid: {e}')
+    print('REWARD: 0.0')
+    sys.exit(0)
+
+MOCKS = {
+    'slack': 'http://28.7.184.198:8178',
+    'jira': 'http://28.7.184.198:8153',
+    'salesforce': 'http://28.7.184.198:8175',
+    'notion': 'http://28.7.184.198:8166',
+}
+
+# --- Egress proxy helper (MANDATORY) ---
+PROXY_CANDIDATES = [
+    os.environ.get('HTTPS_PROXY') or os.environ.get('https_proxy'),
+        None,
+]
+
+
+def resolve_proxy(probe_url):
+    try:
+        requests.get(probe_url, timeout=8)
+        return None
+    except Exception:
+        pass
+    for candidate in PROXY_CANDIDATES:
+        if not candidate:
+            continue
+        try:
+            requests.get(probe_url, timeout=12,
+                         proxies={'http': candidate, 'https': candidate})
+            return candidate
+        except Exception:
+            continue
+    return None
+
+
+PROXY = resolve_proxy(f"{MOCKS['slack']}/go?sid=conn-probe")
+PROXIES = None
+
+
+def text_value(value):
+    """Flatten mock field values that may be raw strings or typed objects."""
+    if value is None:
+        return ''
+    if isinstance(value, str):
+        return value
+    if isinstance(value, (int, float, bool)):
+        return str(value)
+    if isinstance(value, dict):
+        return ' '.join(text_value(v) for v in value.values())
+    if isinstance(value, (list, tuple, set)):
+        return ' '.join(text_value(v) for v in value)
+    return str(value)
+
+
+def norm_text(value):
+    return re.sub(r'\s+', ' ', text_value(value)).strip().lower()
+
+
+def canon_option(value):
+    return re.sub(r'[^a-z0-9]+', ' ', text_value(value)).strip().lower()
+
+
+def option_in(value, allowed):
+    current = canon_option(value)
+    return current in {canon_option(v) for v in allowed}
+
+
+def contains_any(text, needles):
+    low = norm_text(text)
+    return any(needle in low for needle in needles)
+
+
+def field_blob(record, keys):
+    return ' '.join(text_value(record.get(k)) for k in keys)
+
+
+def parse_iso_datetime(value):
+    raw = text_value(value).strip()
+    if not raw:
+        return None
+    try:
+        return datetime.fromisoformat(raw.replace('Z', '+00:00'))
+    except ValueError:
+        return None
+
+
+def relative_date_tokens(reference_dt):
+    """Build date tokens dynamically instead of baking in one calendar date."""
+    if reference_dt is None:
+        return set()
+    tokens = set()
+    months = [
+        'january', 'february', 'march', 'april', 'may', 'june',
+        'july', 'august', 'september', 'october', 'november', 'december'
+    ]
+    for delta in range(0, 8):
+        cur = reference_dt + timedelta(days=delta)
+        month = months[cur.month - 1]
+        tokens.update({
+            cur.strftime('%Y-%m-%d').lower(),
+            cur.strftime('%m/%d').lstrip('0').replace('/0', '/'),
+            cur.strftime('%m/%d/%Y').lstrip('0').replace('/0', '/'),
+            f'{month} {cur.day}',
+            f'{month[:3]} {cur.day}',
+            cur.strftime('%A').lower(),
+        })
+    return tokens
+
+
+def has_next_update_eta(text, reference_dt=None):
+    low = norm_text(text)
+    number_word = (
+        r'\d+|a|an|one|two|three|four|five|six|seven|eight|nine|ten|'
+        r'eleven|twelve|couple|few'
+    )
+    eta_patterns = [
+        r'\b\d{1,2}:\d{2}\s*(?:am|pm)?\s*(?:et|pt|ct|mt|utc|gmt)?\b',
+        r'\b\d{1,2}\s*(?:am|pm)\s*(?:et|pt|ct|mt|utc|gmt)?\b',
+        r'\b(?:within|in|after|every)\s+(?:' + number_word + r')\s*'
+        r'(?:mins?|minutes?|hrs?|hours?|days?|business days?)\b',
+        r'\b(?:by|before)\s+(?:eod|eob|cob|end of (?:the )?day|'
+        r'close of business|noon|midday|midnight|today|tomorrow|tonight|'
+        r'this (?:morning|afternoon|evening)|next business day|'
+        r'(?:mon|tues|wednes|thurs|fri|satur|sun)day|'
+        r'\d{1,2}(?::\d{2})?\s*(?:am|pm)?)\b',
+        r'\b(?:later today|this (?:morning|afternoon|evening)|tomorrow|'
+        r'tonight|next business day)\b',
+        r'\b(?:next update|update you|follow up|get back to you|circle back)'
+        r'\b.{0,80}\b(?:today|tomorrow|tonight|morning|afternoon|evening|'
+        r'eod|eob|cob|hour|minute|day|business day|am|pm|\d)\b',
+        r'\b\d{4}-\d{1,2}-\d{1,2}\b',
+        r'\b\d{1,2}/\d{1,2}(?:/\d{2,4})?\b',
+        r'\b(?:jan|january|feb|february|mar|march|apr|april|may|jun|june|'
+        r'jul|july|aug|august|sep|sept|september|oct|october|nov|november|'
+        r'dec|december)\s+\d{1,2}\b',
+    ]
+    if any(re.search(pattern, low) for pattern in eta_patterns):
+        return True
+    return any(token and token in low for token in relative_date_tokens(reference_dt))
+
+
+# --- Fetch state from all mocks ---
+states = {}
+for name, url in MOCKS.items():
+    try:
+        data = requests.get(f'{url}/go?sid={sid}', timeout=15, proxies=PROXIES).json()
+        states[name] = data
+    except Exception as e:
+        print(f'CRITICAL: Cannot fetch state from {name} ({url}): {e}')
+        print('REWARD: 0.0')
+        sys.exit(0)
+
+
+def verify_task():
+    total_score = 0.0
+
+    # =====================================================================
+    # Component 1 — Jira: NEW issue for the Vertex integration outage (0.30)
+    # Measures a task change: there is NO Vertex outage bug in initial_env.
+    # =====================================================================
+    try:
+        j_init = states['jira']['initial_state']
+        j_cur = states['jira']['current_state']
+        init_ids = {i.get('id') for i in j_init.get('issues', [])}
+        new_issues = [i for i in j_cur.get('issues', []) if i.get('id') not in init_ids]
+
+        matched = None
+        for issue in new_issues:
+            blob = norm_text(field_blob(issue, ('summary', 'description')))
+            # Names the customer and the integration/sync incident without
+            # requiring an exact outage wording.
+            names_vertex = 'vertex' in blob
+            names_incident = (
+                contains_any(blob, ('integration', 'sync', 'salesforce'))
+                and contains_any(blob, (
+                    'down', 'outage', 'fail', 'failure', 'error', '500',
+                    'blocked', 'blocking', 'urgent', 'incident', 'escalat'
+                ))
+            )
+            if names_vertex and names_incident:
+                matched = issue
+                break
+
+        if matched is None:
+            print('FAIL: Component 1 — no NEW Jira issue naming the Vertex integration outage')
+        else:
+            blob = norm_text(field_blob(matched, ('summary', 'description')))
+            # Jira's mock issue types are Story/Task/Bug/Epic; Bug is preferred
+            # for this incident, while Story/Task are still reasonable tracked work.
+            good_type = option_in(matched.get('type'), ('Bug', 'Task', 'Story'))
+            # Repro evidence is accepted by core signals rather than exact
+            # phrasing: open Integrations -> Sync Now -> 500/error.
+            repro_signals = [
+                contains_any(blob, ('integration', 'integrations')),
+                contains_any(blob, ('sync', 'sync now')),
+                contains_any(blob, ('500', 'error', 'fails', 'failed', 'failure')),
+            ]
+            has_repro = (
+                sum(1 for ok in repro_signals if ok) >= 2
+                or contains_any(blob, ('repro', 'reproduce', 'steps'))
+            )
+            good_priority = option_in(matched.get('priority'), (
+                'High', 'Highest', 'Critical', 'Urgent'
+            ))
+            allowed_statuses = {
+                s for issue in j_cur.get('issues', []) for s in [issue.get('status')] if s
+            } | {'To Do', 'In Progress', 'In Review', 'Done'}
+            good_status = option_in(matched.get('status'), allowed_statuses)
+            real_project_ids = {p.get('id') for p in j_cur.get('projects', []) if p.get('id')}
+            real_project = matched.get('projectId') in real_project_ids
+            checks = {
+                'type Bug/Task/Story': good_type,
+                'has_repro_steps': has_repro,
+                'priority High/Highest': good_priority,
+                'status is a real Jira option': good_status,
+                'real Jira project': real_project,
+            }
+            if all(checks.values()):
+                print(f"PASS: Component 1 — new Jira issue {matched.get('key')} "
+                      f"'{matched.get('summary')}' (0.30 pts)")
+                total_score += 0.30
+            else:
+                failed = [k for k, v in checks.items() if not v]
+                print(f"FAIL: Component 1 — new issue {matched.get('key')} missing: {failed}")
+    except Exception as e:
+        print(f'ERROR: Component 1 (Jira) — {e}')
+
+    # =====================================================================
+    # Component 2 — Salesforce: NEW escalation case on Vertex Mfg (0.30)
+    # Measures a task change: NO Vertex outage case exists in initial_env.
+    # =====================================================================
+    try:
+        s_init = states['salesforce']['initial_state']
+        s_cur = states['salesforce']['current_state']
+        init_case_ids = {c.get('caseId') for c in s_init.get('cases', [])}
+        new_cases = [c for c in s_cur.get('cases', [])
+                     if c.get('caseId') not in init_case_ids]
+
+        account_by_id = {a.get('accountId'): a
+                         for a in s_cur.get('accounts', [])}
+
+        def is_vertex_case(case):
+            account_id = case.get('accountId')
+            account_name = (
+                account_by_id.get(account_id, {}).get('name') or '')
+            case_blob = norm_text(field_blob(case, (
+                'accountName', 'company', 'subject', 'description'
+            )))
+            return (
+                account_id == 'account-1'
+                or 'vertex manufacturing' in norm_text(account_name)
+                or 'vertex manufacturing' in case_blob
+            )
+
+        # Restrict to newly created cases tied to the Vertex Manufacturing account.
+        candidates = [c for c in new_cases if is_vertex_case(c)]
+
+        # Select the case to score WITHOUT keyword matching on
+        # subject/description/type (keyword matching was unreliable). Among the
+        # new Vertex Manufacturing cases, prefer one that already satisfies the
+        # meaningful checks (priority + real status); otherwise fall back to the
+        # first candidate so the failure is still reported.
+        def _case_scores(case):
+            return (
+                option_in(case.get('priority'), ('High', 'Critical'))
+                and option_in(case.get('status'), (
+                    'new', 'working', 'escalated', 'closed',
+                    'not started', 'not start', 'in progress', 'completed'
+                ))
+            )
+
+        matched = next((c for c in candidates if _case_scores(c)),
+                       candidates[0] if candidates else None)
+
+        if matched is None:
+            print('FAIL: Component 2 — no NEW Salesforce case on account-1 (Vertex Manufacturing)')
+        else:
+            good_account = is_vertex_case(matched)
+            good_priority = option_in(matched.get('priority'), ('High', 'Critical'))
+            good_status = option_in(matched.get('status'), (
+                'new', 'working', 'escalated', 'closed',
+                'not started', 'not start', 'in progress', 'completed'
+            ))
+            checks = {
+                'Vertex Manufacturing account': good_account,
+                'priority High/Critical': good_priority,
+                'status is a real Salesforce option': good_status,
+            }
+            if all(checks.values()):
+                print(f"PASS: Component 2 — new case {matched.get('caseNumber')} "
+                      f"'{matched.get('subject')}' (0.30 pts)")
+                total_score += 0.30
+            else:
+                failed = [k for k, v in checks.items() if not v]
+                print(f"FAIL: Component 2 — new case missing: {failed}")
+    except Exception as e:
+        print(f'ERROR: Component 2 (Salesforce) — {e}')
+
+    # =====================================================================
+    # Component 3 — Notion: Vertex health row -> Red + Risk Note (0.20)
+    # Measures a task change: initial health is Yellow/Green, Risk Note blank.
+    # =====================================================================
+    try:
+        n_init = states['notion']['initial_state']
+        n_cur = states['notion']['current_state']
+
+        def find_vertex_row(pages):
+            # Locate the Customer Health Tracker DB, then its Vertex Manufacturing row
+            for pid, p in pages.items():
+                if p.get('type') == 'database' and 'health' in norm_text(p.get('title')):
+                    # map property name -> id
+                    name_to_id = {norm_text(pr.get('name')): pr.get('id')
+                                  for pr in p.get('properties', [])}
+                    for itid in p.get('items', []):
+                        itp = pages.get(itid, {})
+                        if 'vertex manufacturing' in norm_text(itp.get('title')):
+                            return itp, name_to_id
+            return None, {}
+
+        cur_row, name_to_id = find_vertex_row(n_cur.get('pages', {}))
+        if cur_row is None:
+            print('FAIL: Component 3 — Vertex Manufacturing health row not found')
+        else:
+            props = cur_row.get('properties', {})
+            health_id = name_to_id.get('health')
+            risk_id = (
+                name_to_id.get('risk note')
+                or next((pid for name, pid in name_to_id.items()
+                         if 'risk' in name and 'note' in name), None)
+            )
+            health_val = props.get(health_id)
+            risk_val = props.get(risk_id)
+            health_red = option_in(health_val, ('Red',))
+            risk_text = norm_text(risk_val)
+            risk_ok = bool(risk_text)
+            checks = {"Health == 'Red'": health_red, 'Risk Note is non-empty': risk_ok}
+            if all(checks.values()):
+                print(f"PASS: Component 3 — Notion Vertex row Health=Red + Risk Note set (0.20 pts)")
+                total_score += 0.20
+            else:
+                failed = [k for k, v in checks.items() if not v]
+                print(f"FAIL: Component 3 — Notion row missing: {failed} "
+                      f"(health={health_val!r}, risk={risk_val!r})")
+    except Exception as e:
+        print(f'ERROR: Component 3 (Notion) — {e}')
+
+    # =====================================================================
+    # Component 4 — Slack: CSM acknowledgement reply in 'acct-vertex'
+    #   4a (0.10, programmatic): new message by user_1 in acct-vertex
+    #   4b (0.10, programmatic): acknowledges outage AND states next-update ETA
+    # Measures a task change: no CSM ack reply exists in initial_env.
+    # =====================================================================
+    new_vertex_msgs = []
+    try:
+        sl_init = states['slack']['initial_state']
+        sl_cur = states['slack']['current_state']
+        init_v = sl_init.get('messages', {}).get('acct-vertex', [])
+        cur_v = sl_cur.get('messages', {}).get('acct-vertex', [])
+        # New messages authored by the CSM (user_1). Compare message IDs so
+        # thread replies or timestamp reordering do not hide valid replies.
+        init_ids = {m.get('messageId') for m in init_v}
+        added = [m for m in cur_v if m.get('messageId') not in init_ids]
+        if not added and len(cur_v) > len(init_v):
+            added = cur_v[len(init_v):]
+        new_vertex_msgs = [m for m in added if m.get('senderId') == 'user_1']
+
+        if new_vertex_msgs:
+            print(f"PASS: Component 4a — new CSM reply in 'acct-vertex' (0.10 pts)")
+            total_score += 0.10
+        else:
+            print("FAIL: Component 4a — no new CSM (user_1) reply in 'acct-vertex'")
+    except Exception as e:
+        print(f'ERROR: Component 4a (Slack) — {e}')
+
+    # Component 4b — PROGRAMMATIC: acknowledgement + concrete next-update ETA (0.10)
+    # The task requires the reply to "acknowledge the issue and tell them when we'll
+    # update them next." We verify both intents via pattern detection rather than an
+    # LLM judge, so the score is fully deterministic and reproducible:
+    #   (a) acknowledgement: references the outage/issue/escalation it is responding to
+    #   (b) concrete next-update time/ETA: a clock time, a relative window, or a named day.
+    try:
+        if new_vertex_msgs:
+            reply_text = (new_vertex_msgs[-1].get('content') or '')
+
+            # (a) acknowledges the reported issue / outage
+            ack = contains_any(reply_text, (
+                'outage', 'integration', 'sync', 'issue', '500',
+                'escalat', 'sorry', 'flag', 'disruption', 'down',
+                'received', 'acknowledge', 'acknowledged', 'reported',
+                'looking into', 'investigating', 'engineering', 'ticket',
+                'incident', 'blocked', 'blocking'
+            ))
+
+            # (b) states a concrete next-update time / ETA. The reference date
+            # comes from the customer report so relative terms are evaluated
+            # dynamically instead of against a hard-coded calendar day.
+            customer_times = [
+                parse_iso_datetime(m.get('timestamp'))
+                for m in init_v
+                if m.get('senderId') != 'user_1'
+            ]
+            customer_times = [t for t in customer_times if t is not None]
+            reference_dt = max(customer_times) if customer_times else None
+            has_eta = has_next_update_eta(reply_text, reference_dt)
+
+            checks = {'acknowledges issue': ack, 'states next-update ETA': has_eta}
+            if all(checks.values()):
+                print(f"PASS: Component 4b — reply acknowledges outage + gives ETA (0.10 pts)")
+                total_score += 0.10
+            else:
+                failed = [k for k, v in checks.items() if not v]
+                print(f"FAIL: Component 4b — reply missing: {failed} (text={reply_text[:120]!r})")
+        else:
+            print("FAIL: Component 4b — no CSM reply to evaluate")
+    except Exception as e:
+        print(f'ERROR: Component 4b (Slack) — {e}')
+
+    final_score = round(min(total_score, 1.0), 4)
+    print(f'\nScore: {total_score}/1.0')
+    print(f'REWARD: {final_score}')
+    return final_score
+
+
+verify_task()

csm_escalation_001/reward_label.json

@@ -0,0 +1,71 @@
+{
+  "reward_file": "/apdcephfs_sh2/share_300000800/user/jackwkwang/long-horizon-tasks/mini_osworld_xiangwu/cache/csm_escalation_004/reward.py",
+  "api_url": "http://28.7.184.184:8000/v1/chat/completions",
+  "model": "Kimi-K2.6",
+  "max_code_chars": 100000,
+  "created_at": "2026-07-03 20:20:42",
+  "label": {
+    "task_id": "csm_escalation_001",
+    "domain": "mock_websites",
+    "summary": "验证 CSM 升级任务：在 Jira 创建 Vertex 集成事故工单、在 Salesforce 创建 Vertex Manufacturing 升级案例、在 Notion 更新健康状态为 Red 并填写 Risk Note、在 Slack 的 acct-vertex 频道以 CSM 身份回复确认事故并给出下次更新时间",
+    "is_placeholder": false,
+    "data_sources": [
+      "/tmp/task_web_sid",
+      "slack_mock (http://28.7.186.212:8198)",
+      "jira_mock (http://28.7.186.212:8173)",
+      "salesforce_mock (http://28.7.186.212:8195)",
+      "notion_mock (http://28.7.186.212:8186)"
+    ],
+    "scoring_components": [
+      {
+        "name": "Component 1",
+        "weight": 0.3,
+        "description": "Jira 中新建了针对 Vertex 集成事故的工单，且包含复现证据、高优先级、真实状态与真实项目",
+        "check_logic": "对比 jira 的 initial_state 与 current_state 的 issues，筛选出 id 不在初始集合中的新工单。要求新工单的 summary/description 同时包含 'vertex' 与集成/同步/事故相关词；type 为 Bug/Task/Story；priority 为 High/Highest/Critical/Urgent；status 为当前 issues 中已存在的真实选项或预定义集合之一；projectId 属于 current_state 中的真实项目；复现证据通过关键词组合（integration/integrations、sync/sync now、500/error/fails/failed/failure 中至少两项）或包含 repro/reproduce/steps 判定",
+        "pass_condition": "新增 Jira 工单同时满足：命名 Vertex 事故、类型合法、有复现证据、高优先级、status 真实、项目真实"
+      },
+      {
+        "name": "Component 2",
+        "weight": 0.3,
+        "description": "Salesforce 中在 Vertex Manufacturing 账户下新建了升级案例，优先级高，状态有效",
+        "check_logic": "对比 salesforce 的 initial_state 与 current_state 的 cases，筛选 caseId 为新的案例，并限制为关联 Vertex Manufacturing（accountId == 'account-1' 或名称/内容包含 'vertex manufacturing'）。在候选案例中优先选取 subject/description/type 包含集成/事故相关词的案例；若无则取首个候选。要求该案例 is_vertex_case 为真、priority 为 High/Critical、status 属于允许的真实选项（new/working/escalated/closed/not started/not start/in progress/completed）",
+        "pass_condition": "新增案例属于 Vertex Manufacturing、优先级 High/Critical、status 为真实 Salesforce 选项"
+      },
+      {
+        "name": "Component 3",
+        "weight": 0.2,
+        "description": "Notion 的 Customer Health Tracker 中 Vertex Manufacturing 行 Health 设为 Red 且 Risk Note 非空",
+        "check_logic": "在 notion current_state 的 pages 中查找 type 为 database 且标题含 'health' 的数据库，再定位标题含 'vertex manufacturing' 的行。读取 Health 属性值要求为 Red；读取 Risk Note 属性值（支持 'risk note' 或名称同时含 risk 与 note 的属性）要求归一化后非空",
+        "pass_condition": "Vertex Manufacturing 行的 Health 为 Red 且 Risk Note 非空"
+      },
+      {
+        "name": "Component 4a",
+        "weight": 0.1,
+        "description": "Slack 的 acct-vertex 频道中有 CSM（user_1）发送的新消息",
+        "check_logic": "对比 slack 的 initial_state 与 current_state 中 'acct-vertex' 频道的消息列表，按 messageId 差集找出新增消息；若差集为空但消息总数增加，则取末尾新增部分。筛选 senderId == 'user_1' 的消息，存在即得分",
+        "pass_condition": "存在至少一条 user_1 在 acct-vertex 频道发送的新消息"
+      },
+      {
+        "name": "Component 4b",
+        "weight": 0.1,
+        "description": "CSM 的 Slack 回复内容确认事故并给出具体的下次更新 ETA",
+        "check_logic": "取 Component 4a 中最后一条新消息的 content。acknowledgement 通过检测 outage/integration/sync/issue/500/escalat/sorry/flag/disruption/down/received/acknowledge/acknowledged/reported/looking into/investigating/engineering/ticket/incident/blocked/blocking 等关键词判定。concrete next-update ETA 通过多组正则（如具体时间、相对时间窗口、星期、日期格式等）以及基于客户最后消息时间动态生成的未来 7 天日期 token 进行判定",
+        "pass_condition": "回复内容同时满足：确认/回应事故，且包含具体的下次更新时间或 ETA"
+      }
+    ],
+    "total_max_score": 1.0,
+    "score_aggregation": "各组件得分相加，最终通过 min(total_score, 1.0) 钳制上限并四舍五入到 4 位小数",
+    "failure_modes": [
+      "/tmp/task_web_sid 读取失败或为空��打印 CRITICAL 并返回 0.0",
+      "任一 mock 服务（slack/jira/salesforce/notion）状态拉取失败：打印 CRITICAL 并返回 0.0",
+      "Jira 无新增 Vertex 集成事故工单或字段校验未通过：Component 1 失败，扣 0.30",
+      "Salesforce 无新增 Vertex Manufacturing 案例或字段校验未通过：Component 2 失败，扣 0.30",
+      "Notion 未找到 Vertex Manufacturing 健康行，或 Health 非 Red，或 Risk Note 为空：Component 3 失败，扣 0.20",
+      "Slack acct-vertex 频道无 CSM（user_1）新回复：Component 4a 失败，扣 0.10",
+      "CSM 回复未确认事故或未给出下次更新 ETA：Component 4b 失败，扣 0.10"
+    ],
+    "scoring_logic_explanation": "脚本首先读取 /tmp/task_web_sid 获取会话 ID，失败则直接返回 0.0。随后通过 HTTP GET 从 slack、jira、salesforce、notion 四个 mock 服务拉取 initial_state 与 current_state，任一服务拉取失败也直接返回 0.0。接着按顺序执行 5 项检查：1）Jira：找新增 Vertex 事故工单，验证类型、复现证据、高优先级、真实状态与项目，通过得 0.30；2）Salesforce：找 Vertex Manufacturing 账户下的新增案例，验证事故相关主题、高优先级与有效状态，通过得 0.30；3）Notion：定位 Customer Health Tracker 的 Vertex Manufacturing 行，验证 Health 为 Red 且 Risk Note 非空，通过得 0.20；4）Slack：对比 acct-vertex 频道消息，验证 CSM（user_1）发了新回复，得 0.10；再对该回复做模式检测，验证其确认事故并给出具体下次更新 ETA，得 0.10。各组件得分累加，最终用 min(total_score, 1.0) 钳制上限并四舍五入到 4 位小数输出。任一组件异常仅记录 ERROR 或 FAIL，不影响其他组件计分。",
+    "confidence": 1.0
+  },
+  "raw_response": "用户要求我分析一个 reward.py 评分脚本的评分逻辑，并输出一个严格的 JSON 对象。\n\n首先，我需要从源码中提取关键信息：\n\n1. **Task ID**: 从 docstring 中可以看到 `Task ID: csm_escalation_004`\n2. **Domain**: 从 docstring 中可以看到 `Domain: mock_websites (multi-mock: slack, jira, salesforce, notion)`\n3. **Summary**: 一句话概括。这个脚本验证 CSM 升级任务：在 Jira 创建 Vertex 集成事故工单、在 Salesforce 创建 Vertex Manufacturing 升级案例、在 Notion 更新健康状态为红色并添加风险备注、在 Slack 的 acct-vertex 频道回复客户确认事故并给出下次更新时间。\n4. **is_placeholder**: false。这个脚本有完整的评分逻辑，不是占位脚本。\n5. **data_sources**: 脚本从 `/tmp/task_web_sid` 读取 sid，然后从多个 mock 服务获取状态：slack (`http://28.7.186.212:8178`)、jira (`http://28.7.186.212:8153`)、salesforce (`http://28.7.186.212:8175`)、notion (`http://28.7.186.212:8166`)。\n6. **scoring_components**:\n   - Component 1 (0.30): Jira 新建 Vertex 集成事故工单，包含复现证据、高优先级、真实状态、真实项目。\n   - Component 2 (0.30): Salesforce 在 account-1 (Vertex Manufacturing) 新建案例，高优先级，状态有效。\n   - Component 3 (0.20): Notion 中 Vertex Manufacturing 健康行 -> Health 为 Red 且 Risk Note 非空。\n   - Component 4a (0.10): Slack 中 CSM (user_1) 在 'acct-vertex' 频道的新回复。\n   - Component 4b (0.10): 该回复确认故障并给出具体的下次更新时间/ETA。\n\n让我详细分析每个组件的检查逻辑：\n\n**Component 1**:\n- 比较 initial_state 和 current_state 的 issues，找出新创建的 issue。\n- 检查 summary/description 中是否包含 \"vertex\" 以及集成/同步/事故相关词汇。\n- 检查 type 是否为 Bug/Task/Story。\n- 检查是否有复现证据（包含 integration/integrations、sync/sync now、500/error/fails/failed/failure 中至少两个，或包含 repro/reproduce/steps）。\n- 检查 priority 是否为 High/Highest/Critical/Urgent。\n- 检查 status 是否为真实选项（从 current issues 中提取或预定义集合）。\n- 检查 projectId 是否存在于真实项目中。\n- 全部通过得 0.30。\n\n**Component 2**:\n- 比较 initial_state 和 current_state 的 cases，找出新案例。\n- 检查是否关联 Vertex Manufacturing 账户（accountId == 'account-1' 或名称/内容包含 'vertex manufacturing'）。\n- 检查 subject/description/type 是否包含集成/同步/事故相关词汇。\n- 检查 priority 是否为 High/Critical。\n- 检查 status 是否为真实 Salesforce 选项（new, working, escalated, closed, not started, not start, in progress, completed）。\n- 全部通过得 0.30。\n\n**Component 3**:\n- 在 Notion 的 current_state.pages 中查找 Customer Health Tracker 数据库。\n- 找到 Vertex Manufacturing 行。\n- 检查 Health 属性是否为 Red。\n- 检查 Risk Note 属性是否非空。\n- 全部通过得 0.20。\n\n**Component 4a**:\n- 比较 Slack initial_state 和 current_state 中 'acct-vertex' 频道的消息。\n- 找出 user_1 (CSM) 发送的新消息。\n- 有新消息得 0.10。\n\n**Component 4b**:\n- 检查 4a 中找到的回复内容。\n- 是否确认问题（包含 outage/integration/sync/issue/500/escalat/sorry 等关键词）。\n- 是否给出下次更新时间/ETA（通过正则表达式匹配时间、相对时间窗口、日期等，或动态生成的日期 token）。\n- 全部通过得 0.10。\n\n**Score aggregation**:\n- 各组件分数相加，最后 `round(min(total_score, 1.0), 4)`。所以是相加并钳制到 1.0。\n\n**Failure modes**:\n- 读取 /tmp/task_web_sid 失败或为空 -> CRITICAL, REWARD: 0.0, sys.exit(0)\n- 从任一 mock 服务获取状态失败 -> CRITICAL, REWARD: 0.0, sys.exit(0)\n- 各组件检查失败会打印 FAIL 但不退出，只是不得分。\n- 异常会被捕获并打印 ERROR，不得分。\n\n**Confidence**: 1.0，因为源码完整且逻辑清晰。\n\n现在构建 JSON。注意：\n- 只输出 JSON，不要 Markdown 代码块。\n- 第一个字符是 {，最后一个字符是 }。\n- 所有字符串 value 使用中文，但 task_id、domain、字段名、代码标识符保留原样。\n- weight 填写代码中显式给出的分值。\n\n让我确认每个组件的 weight：\n- Component 1: 0.30\n- Component 2: 0.30\n- Component 3: 0.20\n- Component 4a: 0.10\n- Component 4b: 0.10\n\ntotal_max_score: 1.0\n\nscoring_logic_explanation 需要按执行顺序详细说明。\n\n让我组织中文描述：\n\n脚本首先读取 /tmp/task_web_sid 获取会话 ID；若失败则直接返回 0.0。接着通过 HTTP GET 从 slack、jira、salesforce、notion 四个 mock 服务拉取 initial_state 与 current_state；任一服务拉取失败也直接返回 0.0。\n\n随后逐项检查：\n1. Jira：对比 initial_state 与 current_state 的 issues，找出新增工单。要求新工单 summary/description 同时包含 \"vertex\" 与集成/同步/事故相关词；type 为 Bug/Task/Story；priority 为 High/Highest/Critical/Urgent；status 为真实存在的选项；projectId 对应真实项目；且包含复现证据（特定关键词组合或 repro/reproduce/steps）。通过得 0.30。\n2. Salesforce：对比 cases 找出新增案例，且必须关联 Vertex Manufacturing（account-1）。要求 subject/description/type 包含集成/事故相关词；priority 为 High/Critical；status 为允许的真实选项。通过得 0.30。\n3. Notion：在 current_state 中定位 Customer Health Tracker 数据库的 Vertex Manufacturing 行。要求 Health 字段为 Red，Risk Note 字段非空。通过得 0.20。\n4. Slack：对比 acct-vertex 频道消息，找出 CSM（user_1）发送的新消息。\n   - 4a：存在新消息即得 0.10。\n   - 4b：取该消息内容，要求同时满足：(a) 通过关键词确认/回应事故；(b) 通过正则与动态日期 token 检测给出具体的下次更新 ETA。通过得 0.10。\n\n各组件得分累加，最终用 min(total_score, 1.0) 钳制上限并四舍五入到 4 位小数输出。\n\nfailure_modes:\n- /tmp/task_web_sid 读取失败或为空：打印 CRITICAL 并返回 0.0\n- mock 服务状态拉取失败：打印 CRITICAL 并返回 0.0\n- Jira 无新增 Vertex 事故工单：Component 1 失败，不得 0.30\n- Salesforce 无新增 Vertex Manufacturing 案例或字段不符：Component 2 失败，不得 0.30\n- Notion 未找到 Vertex 行或 Health 非 Red / Risk Note 为空：Component 3 失败，不得 0.20\n- Slack 无 CSM 新回复：Component 4a 失败，不得 0.10\n- CSM 回复未确认事故或未给出 ETA：Component 4b 失败，不得 0.10\n\n检查 JSON 格式要求：\n- 不要输出 Markdown 代码块标记。\n- 不要输出额外解释。\n- 不要输出思考过程。\n- 不要输出 <thinking> 标签。\n- 输出的第一个字符必须是 {，最后一个字符必须是 }。\n\n我需要确保 JSON 是有效的，所有字符串正确转义。\n\n让我再仔细检查一下 Component 1 的 check_logic：\n- names_vertex: 'vertex' in blob\n- names_incident: contains_any(integration, sync, salesforce) AND contains_any(down, outage, fail, failure, error, 500, blocked, blocking, urgent, incident, escalat)\n- good_type: option_in(type, (Bug, Task, Story))\n- has_repro: sum(repro_signals) >= 2 OR contains_any(repro, reproduce, steps)\n  - repro_signals: contains_any(integration, integrations), contains_any(sync, sync now), contains_any(500, error, fails, failed, failure)\n- good_priority: option_in(priority, (High, Highest, Critical, Urgent))\n- good_status: option_in(status, allowed_statuses) where allowed_statuses 从 current issues 提取 + {'To Do', 'In Progress', 'In Review', 'Done'}\n- real_project: matched.get('projectId') in real_project_ids\n\nComponent 2:\n- is_vertex_case: account_id == 'account-1' OR 'vertex manufacturing' in account_name OR 'vertex manufacturing' in case_blob\n- candidates: new_cases filtered by is_vertex_case\n- matched: first candidate with subject/description/type containing integration/sync/outage/escalat/incident/down/500/blocked/blocking; if none but candidates exist, matched = candidates[0]\n- good_account: is_vertex_case(matched)\n- good_priority: option_in(priority, (High, Critical))\n- good_status: option_in(status, (new, working, escalated, closed, not started, not start, in progress, completed))\n\nComponent 3:\n- find_vertex_row: find database with 'health' in title, then item with 'vertex manufacturing' in title\n- health_id: name_to_id.get('health')\n- risk_id: name_to_id.get('risk note') or find name containing 'risk' and 'note'\n- health_red: option_in(health_val, ('Red',))\n- risk_ok: bool(norm_text(risk_val))\n\nComponent 4a:\n- init_ids = {m.get('messageId') for m in init_v}\n- added = [m for m in cur_v if m.get('messageId') not in init_ids]\n- if not added and len(cur_v) > len(init_v): added = cur_v[len(init_v):]\n- new_vertex_msgs = [m for m in added if m.get('senderId') == 'user_1']\n\nComponent 4b:\n- reply_text = new_vertex_msgs[-1].get('content') or ''\n- ack: contains_any(outage, integration, sync, issue, 500, escalat, sorry, flag, disruption, down, received, acknowledge, acknowledged, reported, looking into, investigating, engineering, ticket, incident, blocked, blocking)\n- has_eta: has_next_update_eta(reply_text, reference_dt)\n  - reference_dt: max customer message timestamps (non-user_1)\n  - has_next_update_eta checks various regex patterns and relative_date_tokens\n\n一切看起来都正确。现在生成 JSON。 </think> {\"task_id\": \"csm_escalation_004\", \"domain\": \"mock_websites\", \"summary\": \"验证 CSM 升级任务：在 Jira 创建 Vertex 集成事故工单、在 Salesforce 创建 Vertex Manufacturing 升级案例、在 Notion 更新健康状态为 Red 并填写 Risk Note、在 Slack 的 acct-vertex 频道以 CSM 身份回复确认事故并给出下次更新时间\", \"is_placeholder\": false, \"data_sources\": [\"/tmp/task_web_sid\", \"slack_mock (http://28.7.186.212:8198)\", \"jira_mock (http://28.7.186.212:8173)\", \"salesforce_mock (http://28.7.186.212:8195)\", \"notion_mock (http://28.7.186.212:8186)\"], \"scoring_components\": [{\"name\": \"Component 1\", \"weight\": 0.3, \"description\": \"Jira 中新建了针对 Vertex 集成事故的工单，且包含复现证据、高优先级、真实状态与真实项目\", \"check_logic\": \"对比 jira 的 initial_state 与 current_state 的 issues，筛选出 id 不在初始集合中的新工单。要求新工单的 summary/description 同时包含 'vertex' 与集成/同步/事故相关词；type 为 Bug/Task/Story；priority 为 High/Highest/Critical/Urgent；status 为当前 issues 中已存在的真实选项或预定义集合之一；projectId 属于 current_state 中的真实项目；复现证据通过关键词组合（integration/integrations、sync/sync now、500/error/fails/failed/failure 中至少两项）或包含 repro/reproduce/steps 判定\", \"pass_condition\": \"新增 Jira 工单同时满足：命名 Vertex 事故、类型合法、有复现证据、高优先级、status 真实、项目真实\"}, {\"name\": \"Component 2\", \"weight\": 0.3, \"description\": \"Salesforce 中在 Vertex Manufacturing 账户下新建了升级案例，优先级高，状态有效\", \"check_logic\": \"对比 salesforce 的 initial_state 与 current_state 的 cases，筛选 caseId 为新的案例，并限制为关联 Vertex Manufacturing（accountId == 'account-1' 或名称/内容包含 'vertex manufacturing'）。在候选案例中优先选取 subject/description/type 包含集成/事故相关词的案例；若无则取首个候选。要求该案例 is_vertex_case 为真、priority 为 High/Critical、status 属于允许的真实选项（new/working/escalated/closed/not started/not start/in progress/completed）\", \"pass_condition\": \"新增案例属于 Vertex Manufacturing、优先级 High/Critical、status 为真实 Salesforce 选项\"}, {\"name\": \"Component 3\", \"weight\": 0.2, \"description\": \"Notion 的 Customer Health Tracker 中 Vertex Manufacturing 行 Health 设为 Red 且 Risk Note 非空\", \"check_logic\": \"在 notion current_state 的 pages 中查找 type 为 database 且标题含 'health' 的数据库，再定位标题含 'vertex manufacturing' 的行。读取 Health 属性值要求为 Red；读取 Risk Note 属性值（支持 'risk note' 或名称同时含 risk 与 note 的属性）要求归一化后非空\", \"pass_condition\": \"Vertex Manufacturing 行的 Health 为 Red 且 Risk Note 非空\"}, {\"name\": \"Component 4a\", \"weight\": 0.1, \"description\": \"Slack 的 acct-vertex 频道中有 CSM（user_1）发送的新消息\", \"check_logic\": \"对比 slack 的 initial_state 与 current_state 中 'acct-vertex' 频道的消息列表，按 messageId 差集找出新增消息；若差集为空但消息总数增加，则取末尾新增部分。筛选 senderId == 'user_1' 的消息，存在即得分\", \"pass_condition\": \"存在至少一条 user_1 在 acct-vertex 频道发送的新消息\"}, {\"name\": \"Component 4b\", \"weight\": 0.1, \"description\": \"CSM 的 Slack 回复内容确认事故并给出具体的下次更新 ETA\", \"check_logic\": \"取 Component 4a 中最后一条新消息的 content。acknowledgement 通过检测 outage/integration/sync/issue/500/escalat/sorry/flag/disruption/down/received/acknowledge/acknowledged/reported/looking into/investigating/engineering/ticket/incident/blocked/blocking 等关键词判定。concrete next-update ETA 通过多组正则（如具体时间、相对时间窗口、星期、日期格式等）以及基于客户最后消息时间动态生成的未来 7 天日期 token 进行判定\", \"pass_condition\": \"回复内容同时满足：确认/回应事故，且包含具体的下次更新时间或 ETA\"}], \"total_max_score\": 1.0, \"score_aggregation\": \"各组件得分相加，最终通过 min(total_score, 1.0) 钳制上限并四舍五入到 4 位小数\", \"failure_modes\": [\"/tmp/task_web_sid 读取失败或为空：打印 CRITICAL 并返回 0.0\", \"任一 mock 服务（slack/jira/salesforce/notion）状态拉取失败：打印 CRITICAL 并返回 0.0\", \"Jira 无新增 Vertex 集成事故工单或字段校验未通过：Component 1 失败，扣 0.30\", \"Salesforce 无新增 Vertex Manufacturing 案例或字段校验未通过：Component 2 失败，扣 0.30\", \"Notion 未找到 Vertex Manufacturing 健康行，或 Health 非 Red，或 Risk Note 为空：Component 3 失败，扣 0.20\", \"Slack acct-vertex 频道无 CSM（user_1）新回复：Component 4a 失败，扣 0.10\", \"CSM 回复未确认事故或未给出下次更新 ETA：Component 4b 失败，扣 0.10\"], \"scoring_logic_explanation\": \"脚本首先读取 /tmp/task_web_sid 获取会话 ID，失败则直接返回 0.0。随后通过 HTTP GET 从 slack、jira、salesforce、notion 四个 mock 服务拉取 initial_state 与 current_state，任一服务拉取失败也直接返回 0.0。接着按顺序执行 5 项检查：1）Jira：找新增 Vertex 事故工单，验证类型、复现证据、高优先级、真实状态与项目，通过得 0.30；2）Salesforce：找 Vertex Manufacturing 账户下的新增案例，验证事故相关主题、高优先级与有效状态，通过得 0.30；3）Notion：定位 Customer Health Tracker 的 Vertex Manufacturing 行，验证 Health 为 Red 且 Risk Note 非空，通过得 0.20；4）Slack：对比 acct-vertex 频道消息，验证 CSM（user_1）发了新回复，得 0.10；再对该回复做模式检测，验证其确认事故并给出具体下次更新 ETA，得 0.10。各组件得分累加，最终用 min(total_score, 1.0) 钳制上限并四舍五入到 4 位小数输出。任一组件异常仅记录 ERROR 或 FAIL，不影响其他组件计分。\", \"confidence\": 1.0}"
+}

csm_health_risk_002/_cua_gym_vm_bridge.sh

@@ -0,0 +1,106 @@
+#!/bin/bash
+# CUA-Gym VM-side network fix (direct-IP mock access).
+#
+# Runs INSIDE the QEMU guest. The CUA-Gym mock services are now reached by
+# their real IDC address (http://28.7.184.198:81xx) instead of the old
+# localhost:9000-9097 socat bridge, so the VM only needs two fixes:
+#
+#   1. Working DNS — the qcow2-baked /etc/resolv.conf points at the Pod's
+#      cluster DNS, which is unreachable from QEMU SLIRP NAT
+#      ("Temporary failure in name resolution").
+#   2. The mock host excluded from the woa http proxy — otherwise the VM's
+#      Chrome (gsettings proxy) and initial_setup.py's `requests` get a 403
+#      from the proxy when hitting 28.7.184.198:81xx. (The old socat path never
+#      hit this because localhost is bypassed by default.)
+#
+# Idempotent. Scoped to the CUA-Gym bridge step; does not touch the generic
+# OSWorld setup pipeline.
+#
+# Argv:
+#   $1  optional sudo password (default "password"; SetupController
+#       substitutes {CLIENT_PASSWORD} when uploaded as a cache file).
+#   $2  unused (kept for backward compat — was socat first port 9000).
+#   $3  unused (kept for backward compat — was socat last  port 9097).
+#   $4  optional mock host IP (default 28.7.184.198).
+
+set +e
+
+PASS="${1:-password}"
+MOCK_HOST="${4:-28.7.184.198}"
+# Derive the /16 IDC subnet so future port/host tweaks on the same segment
+# stay bypassed too (28.7.184.198 -> 28.7.0.0/16).
+MOCK_NET="$(echo "${MOCK_HOST}" | awk -F. 'NF==4{print $1"."$2".0.0/16"}')"
+
+run_root() {
+    echo "$PASS" | sudo -S -p "" bash -c "$1"
+}
+
+# Detect SLIRP gateway from VM's routing table; fall back to 10.0.2.2.
+GW="$(ip route 2>/dev/null | awk '/default/ {print $3; exit}')"
+GW="${GW:-10.0.2.2}"
+
+# ─── CUA-Gym network fix (Taiji-specific) ─────────────────────────────
+fix_vm_network() {
+    # 0) Discover VM network state at runtime — gateway, own IP/CIDR, SLIRP
+    #    DNS forwarder. Nothing hardcoded except the mock host above.
+    local GW_NET="${GW%.*}"
+    local SLIRP_DNS="${GW_NET}.3"        # QEMU SLIRP DNS proxy is always .3
+    local OWN_CIDR
+    OWN_CIDR="$(ip -4 -o addr show scope global 2>/dev/null \
+                | awk '{print $4; exit}')"
+    local OWN_IP="${OWN_CIDR%/*}"
+    local OWN_NET=""
+    [ -n "${OWN_IP}" ] && OWN_NET="${OWN_IP%.*}.0/24"
+
+    # 1) DNS — point at SLIRP DNS proxy + public fallbacks. chattr +i keeps
+    #    systemd-resolved/NetworkManager from clobbering it.
+    if run_root "printf 'nameserver ${SLIRP_DNS}\\nnameserver 114.114.114.114\\nnameserver 8.8.8.8\\n' > /etc/resolv.conf
+                 chattr +i /etc/resolv.conf 2>/dev/null || true"; then
+        echo "[cua_gym_vm_bridge] DNS set: ${SLIRP_DNS}, 114.114.114.114, 8.8.8.8 (gw=${GW})"
+    else
+        echo "[cua_gym_vm_bridge] WARN: failed to write /etc/resolv.conf"
+    fi
+
+    # 2) Build the proxy bypass list from detected values + the direct-IP
+    #    mock host (and its /16 subnet). THIS is what makes 28.7.184.198:81xx
+    #    reachable from inside the VM — without it the woa proxy 403s it.
+    local ignore_list="['localhost', '127.0.0.0/8', '${GW}'"
+    [ -n "${OWN_IP}" ]    && ignore_list="${ignore_list}, '${OWN_IP}'"
+    [ -n "${OWN_NET}" ]   && ignore_list="${ignore_list}, '${OWN_NET}'"
+    [ -n "${GW_NET}" ]    && ignore_list="${ignore_list}, '${GW_NET}.0/24'"
+    [ -n "${MOCK_HOST}" ] && ignore_list="${ignore_list}, '${MOCK_HOST}'"
+    [ -n "${MOCK_NET}" ]  && ignore_list="${ignore_list}, '${MOCK_NET}'"
+    ignore_list="${ignore_list}]"
+
+    # 3) GNOME proxy ignore-hosts — Chrome bypasses the woa proxy for these.
+    if gsettings set org.gnome.system.proxy ignore-hosts "${ignore_list}" 2>/tmp/gsettings_err; then
+        echo "[cua_gym_vm_bridge] gsettings ignore-hosts = ${ignore_list}"
+    else
+        echo "[cua_gym_vm_bridge] WARN: gsettings ignore-hosts failed: $(cat /tmp/gsettings_err 2>/dev/null)"
+    fi
+
+    # 4) Shell env — comma-separated form for no_proxy/NO_PROXY (requests/curl).
+    local NP="localhost,127.0.0.1,${GW}"
+    [ -n "${OWN_IP}" ]    && NP="${NP},${OWN_IP}"
+    [ -n "${OWN_NET}" ]   && NP="${NP},${OWN_NET}"
+    [ -n "${GW_NET}" ]    && NP="${NP},${GW_NET}.0/24"
+    [ -n "${MOCK_HOST}" ] && NP="${NP},${MOCK_HOST}"
+    [ -n "${MOCK_NET}" ]  && NP="${NP},${MOCK_NET}"
+
+    for var in no_proxy NO_PROXY; do
+        if ! grep -q "export ${var}=${NP}" /home/user/.bashrc 2>/dev/null; then
+            echo "export ${var}=${NP}" >> /home/user/.bashrc
+        fi
+    done
+    run_root "grep -q '^no_proxy=' /etc/environment 2>/dev/null \
+              && sed -i 's|^no_proxy=.*|no_proxy=\"${NP}\"|' /etc/environment \
+              || echo 'no_proxy=\"${NP}\"' >> /etc/environment
+              grep -q '^NO_PROXY=' /etc/environment 2>/dev/null \
+              && sed -i 's|^NO_PROXY=.*|NO_PROXY=\"${NP}\"|' /etc/environment \
+              || echo 'NO_PROXY=\"${NP}\"' >> /etc/environment" >/dev/null 2>&1
+    echo "[cua_gym_vm_bridge] no_proxy=${NP}"
+}
+
+fix_vm_network
+
+echo "[cua_gym_vm_bridge] direct-IP mode: ${MOCK_HOST} (+${MOCK_NET}) bypassed from woa proxy; no socat forwarding needed"

csm_health_risk_002/initial_setup.py

@@ -0,0 +1,593 @@
+# --- MOCK_HOST_NO_PROXY_OVERRIDE: bypass woa proxy for locally-deployed mocks ---
+import os as _os
+_np = _os.environ.get("no_proxy", _os.environ.get("NO_PROXY", ""))
+for _h in ("127.0.0.1", "localhost", "28.7.184.198"):
+    if _h not in _np:
+        _np = (_np + "," + _h).lstrip(",")
+_os.environ["no_proxy"] = _np
+_os.environ["NO_PROXY"] = _np
+# --- end override ---
+
+# --- wait_mocks_loaded: wait for mock backends ready AND browser tabs rendered ---
+def wait_mocks_loaded(timeout=60, render_buffer=12):
+    """Wait for mock backends ready AND browser tabs rendered (via CDP).
+
+    Phase 1 — backend readiness: poll each mock's /state?sid= until it
+    returns 200 with 'stored_state' in the body.
+
+    Phase 2 — browser render verification: connect to Chrome's CDP
+    endpoint and wait for every open tab to reach 'networkidle' load
+    state with a non-trivial DOM element count. This confirms the SPA
+    has fetched its /go?sid= data and actually painted, not just that
+    the backend is up.
+
+    CDP port auto-detection: checks globals for _chrome_debug_port /
+    cdp_port / debug_port (set by tasks that use a dynamic port, e.g.
+    se_hard_008's launch_chrome_with_urls); defaults to 1337 for tasks
+    that use launch_gui's --remote-debugging-port injection.
+
+    Falls back to the old render_buffer sleep if Playwright or the CDP
+    endpoint is unavailable.
+    """
+    import urllib.request as _u, time as _t
+    g = globals()
+    _urls = {}
+    for _k, _v in list(g.items()):
+        if isinstance(_v, dict) and _v and all(isinstance(x, str) and x.startswith("http://") for x in _v.values()):
+            _urls.update(_v)
+        elif isinstance(_v, str) and _v.startswith("http://") and (_k.endswith("_URL") or _k.endswith("_url")):
+            _urls[_k] = _v
+    _sid = ""
+    for _k in ("sid", "SID", "task_sid", "session_id"):
+        if isinstance(g.get(_k), str) and g[_k]:
+            _sid = g[_k]; break
+    if not _sid:
+        for _k, _v in g.items():
+            if "sid" in _k.lower() and isinstance(_v, str) and _v:
+                _sid = _v; break
+    if not _urls or not _sid:
+        print("  wait_mocks_loaded: no mocks/sid found (%d urls, sid=%r), skipping" % (len(_urls), bool(_sid)))
+        return
+
+    # ---- Phase 1: backend readiness ----
+    _op = _u.build_opener(_u.ProxyHandler({}))
+    _dl = _t.time() + timeout
+    _todo = list(_urls.items())
+    while _todo and _t.time() < _dl:
+        _rest = []
+        for _n2, _uu in _todo:
+            try:
+                _r = _op.open("%s/state?sid=%s" % (_uu, _sid), timeout=5)
+                if _r.status == 200 and b"stored_state" in _r.read():
+                    print("  %s: backend ready" % _n2); continue
+            except Exception:
+                pass
+            _rest.append((_n2, _uu))
+        _todo = _rest
+        if _todo:
+            _t.sleep(2)
+    _done = len(_urls) - len(_todo)
+    _pend = (" (pending: %s)" % [n for n, _ in _todo]) if _todo else ""
+    print("  mocks backend ready: %d/%d%s" % (_done, len(_urls), _pend))
+
+    # ---- Phase 2: browser render verification via CDP ----
+    # Auto-detect CDP port from globals (set by launch_chrome_with_urls
+    # in tasks that use a dynamic port); default to 1337.
+    _cdp_port = 1337
+    for _pk in ("_chrome_debug_port", "chrome_debug_port", "cdp_port", "debug_port"):
+        _pv = g.get(_pk)
+        if isinstance(_pv, int) and 0 < _pv < 65536:
+            _cdp_port = _pv; break
+    _cdp_ok = False
+    try:
+        from playwright.sync_api import sync_playwright
+        _cdp_dl = _t.time() + max(10.0, _dl - _t.time())  # leftover budget, floor 10s
+        with sync_playwright() as _p:
+            _br = None
+            # Retry-connect: Chrome may still be starting up.
+            while _t.time() < _cdp_dl:
+                try:
+                    _br = _p.chromium.connect_over_cdp("http://localhost:%d" % _cdp_port)
+                    break
+                except Exception:
+                    _t.sleep(0.5)
+            if _br is None:
+                print("  wait_mocks_loaded: CDP connect to localhost:%d failed; "
+                      "is Chrome launched with --remote-debugging-port?" % _cdp_port)
+            else:
+                # Wait for ALL expected tabs to appear, then verify each renders.
+                _n_expected = len(_urls)
+                _pages = []
+                _pg_dl = _t.time() + 15
+                _last_count = -1
+                _stable_since = None
+                while _t.time() < _pg_dl:
+                    _pages = []
+                    for _c in _br.contexts:
+                        _pages.extend(_c.pages)
+                    _cnt = len(_pages)
+                    if _n_expected > 0 and _cnt >= _n_expected:
+                        break  # all expected tabs are open
+                    if _cnt > 0 and _cnt == _last_count:
+                        if _stable_since is None:
+                            _stable_since = _t.time()
+                        elif _t.time() - _stable_since >= 1.5:
+                            break  # tab count stable -> no more tabs opening
+                    else:
+                        _stable_since = None
+                    _last_count = _cnt
+                    _t.sleep(0.5)
+                # Re-fetch final snapshot so late-opening tabs are included.
+                _pages = []
+                for _c in _br.contexts:
+                    _pages.extend(_c.pages)
+                if not _pages:
+                    print("  wait_mocks_loaded: CDP connected but no open tabs found")
+                else:
+                    print("  wait_mocks_loaded: %d tab(s) open (expected %d); "
+                          "verifying render" % (len(_pages), _n_expected))
+                    _ok_cnt = 0
+                    for _pg in _pages:
+                        _ttl = ""
+                        try:
+                            _ttl = _pg.title()
+                        except Exception:
+                            pass
+                        try:
+                            _pg.wait_for_load_state(
+                                "networkidle",
+                                timeout=int(max(5000, (_cdp_dl - _t.time()) * 1000)))
+                            _n = _pg.evaluate("document.querySelectorAll('*').length")
+                            if _n < 20:
+                                print("  wait_mocks_loaded: WARNING tab %r has only %d "
+                                      "elements (render stall?)" % (_ttl, _n))
+                            else:
+                                print("  wait_mocks_loaded: tab %r rendered (%d elements)"
+                                      % (_ttl, _n))
+                                _ok_cnt += 1
+                        except Exception as _e:
+                            print("  wait_mocks_loaded: tab %r render wait failed: %r"
+                                  % (_ttl, _e))
+                    # Require ALL open tabs to render before skipping the
+                    # render_buffer fallback, so no tab is left half-loaded.
+                    if _ok_cnt > 0 and _ok_cnt == len(_pages):
+                        _cdp_ok = True
+                    else:
+                        print("  wait_mocks_loaded: %d/%d tabs rendered, will fall back"
+                              % (_ok_cnt, len(_pages)))
+    except ImportError:
+        print("  wait_mocks_loaded: playwright not installed; cannot verify browser render")
+    except Exception as _e:
+        print("  wait_mocks_loaded: CDP phase error (%r)" % _e)
+
+    if not _cdp_ok:
+        print("  wait_mocks_loaded: falling back to render_buffer=%ds sleep" % render_buffer)
+        _t.sleep(render_buffer)
+    else:
+        print("  wait_mocks_loaded: browser render verified, skipping render_buffer")
+# --- end wait_mocks_loaded ---
+
+# --- _open_remaining_mock_tabs: open all auto-detected mock URLs in new Chrome tabs ---
+def _open_remaining_mock_tabs(primary_url=""):
+    """Open all auto-detected mock URLs in Chrome --new-tab, except primary_url."""
+    g = globals()
+    _urls = set()
+    for _k, _v in list(g.items()):
+        if "PROBE" in _k.upper() or _k.startswith("_"):
+            continue
+        if isinstance(_v, str) and _v.startswith("http://") and (_k.endswith("_URL") or _k.endswith("_url")):
+            _urls.add(_v)
+        elif isinstance(_v, dict) and _v and all(isinstance(x, str) and x.startswith("http://") for x in _v.values()):
+            _urls.update(_v.values())
+    _sid = ""
+    for _k in ("sid", "SID", "task_sid", "session_id"):
+        if isinstance(g.get(_k), str) and g[_k]:
+            _sid = g[_k]; break
+    for _u in sorted(_urls):
+        if _u != primary_url:
+            launch_gui(f'google-chrome --new-tab "{_u}/?sid={_sid}"', delay_sec=1.0)
+# --- end _open_remaining_mock_tabs ---
+
+
+"""
+Initial Setup: CSM identifies the one Red/declining account in their book and logs risk
+Task ID: csm_health_risk_002
+Domain: mock_websites
+Mocks: notion_mock, salesforce_mock, slack_mock
+"""
+import json
+import os
+import shlex
+import subprocess
+import time
+import uuid
+
+import requests
+
+# --- Config ---
+MOCKS = {
+    'notion_mock': 'http://28.7.184.198:8166',
+    'salesforce_mock': 'http://28.7.184.198:8175',
+    'slack_mock': 'http://28.7.184.198:8178',
+}
+
+sid = str(uuid.uuid4())
+with open('/tmp/task_web_sid', 'w') as f:
+    f.write(sid)
+print(f'Generated sid={sid}')
+
+NOW = '2026-06-20T09:00:00.000Z'
+
+# --- Network helper: VM egress to the mocks may require a proxy. Probe once,
+#     then use direct or the cluster proxy for all requests. ---
+PROXY = None
+PROXIES = None  # decided by _probe()
+
+
+def _probe():
+    global PROXIES
+    test = 'http://28.7.184.198:8166/go?sid=conntest'
+    try:
+        requests.get(test, timeout=5)
+        PROXIES = None
+        print('Network: direct egress works')
+        return
+    except requests.exceptions.RequestException:
+        pass
+    PROXIES = None
+    print(f'Network: using proxy {PROXY}')
+
+
+def http_post(url, payload):
+    return requests.post(url, json=payload, timeout=30, proxies=PROXIES)
+
+
+def http_get(url):
+    return requests.get(url, timeout=30, proxies=PROXIES)
+
+
+_probe()
+
+
+# =========================================================================
+# NOTION STATE
+# =========================================================================
+def build_notion_state():
+    user = {
+        'id': 'user-1',
+        'name': 'Jordan Rivera',
+        'email': 'jordan.rivera@northwindcloud.com',
+        'avatar': '',
+    }
+    workspace = {
+        'id': 'ws-1',
+        'name': 'Northwind Cloud',
+        'icon': '',
+        'members': ['user-1'],
+    }
+
+    db_id = 'db-health'
+    # Database property definitions
+    properties = [
+        {'id': 'prop-account', 'name': 'Account', 'type': 'title'},
+        {'id': 'prop-health', 'name': 'Health', 'type': 'status',
+         'options': ['Green', 'Yellow', 'Red']},
+        {'id': 'prop-mau', 'name': 'MAU Trend', 'type': 'select',
+         'options': ['Up', 'Flat', 'Down']},
+        {'id': 'prop-arr', 'name': 'ARR', 'type': 'number'},
+        {'id': 'prop-csm', 'name': 'CSM', 'type': 'text'},
+        {'id': 'prop-risk', 'name': 'Risk Note', 'type': 'text'},
+    ]
+
+    # Each database row is itself a page (item). Properties keyed by prop id.
+    rows = [
+        # (page_id, account, health, mau, arr, csm, risk_note)
+        ('item-summit', 'Summit Retail Group', 'Red', 'Down', 120000, 'Jordan Rivera', ''),
+        ('item-cedar', 'Cedar Grove Health', 'Yellow', 'Flat', 95000, 'Jordan Rivera', ''),
+        ('item-river', 'Riverstone Media', 'Green', 'Up', 64000, 'Jordan Rivera', ''),
+        ('item-blue', 'BlueHarbor Logistics', 'Yellow', 'Down', 78000, 'Jordan Rivera', ''),
+        ('item-north', 'Northstar Apparel', 'Red', 'Down', 110000, 'Dana Lee', ''),
+    ]
+
+    pages = {}
+    item_ids = []
+    for pid, account, health, mau, arr, csm, risk in rows:
+        item_ids.append(pid)
+        pages[pid] = {
+            'id': pid,
+            'title': account,
+            'icon': '',
+            'cover': None,
+            'parentId': db_id,
+            'blockIds': [],
+            'favorite': False,
+            'createdDate': '2026-01-05T00:00:00.000Z',
+            'lastEditedDate': '2026-06-01T00:00:00.000Z',
+            'properties': {
+                'prop-account': account,
+                'prop-health': health,
+                'prop-mau': mau,
+                'prop-arr': arr,
+                'prop-csm': csm,
+                'prop-risk': risk,
+            },
+        }
+
+    # The database page itself
+    pages[db_id] = {
+        'id': db_id,
+        'title': 'Customer Health Tracker',
+        'icon': '',
+        'cover': None,
+        'parentId': None,
+        'type': 'database',
+        'viewType': 'table',
+        'properties': properties,
+        'views': [{
+            'id': 'view-1',
+            'name': 'All Accounts',
+            'type': 'table',
+            'filters': [],
+            'sorts': [],
+            'groupBy': None,
+            'visibleProperties': ['prop-account', 'prop-health', 'prop-mau',
+                                  'prop-arr', 'prop-csm', 'prop-risk'],
+        }],
+        'items': item_ids,
+        'blockIds': [],
+        'favorite': True,
+        'createdDate': '2026-01-05T00:00:00.000Z',
+    }
+
+    return {
+        'user': user,
+        'workspace': workspace,
+        'pages': pages,
+        'blocks': {},
+        'trash': [],
+        'comments': {},
+        'settings': {'appearance': 'light', 'startWeekMonday': False, 'fontSize': 'default'},
+        'notifications': [],
+        'pageOrder': [db_id],
+        'focusBlockId': None,
+    }
+
+
+# =========================================================================
+# SALESFORCE STATE
+# =========================================================================
+def build_salesforce_state():
+    user = {
+        'userId': 'user-1',
+        'firstName': 'Jordan',
+        'lastName': 'Rivera',
+        'email': 'jordan.rivera@northwindcloud.com',
+        'phone': '(555) 204-7781',
+        'title': 'Customer Success Manager',
+        'department': 'Customer Success',
+        'role': 'Manager',
+        'avatar': 'https://i.pravatar.cc/150?u=user-1',
+        'timezone': 'America/New_York',
+        'locale': 'en-US',
+        'theme': 'lightning',
+    }
+    users = [
+        user,
+        {'userId': 'user-2', 'firstName': 'Dana', 'lastName': 'Lee',
+         'email': 'dana.lee@northwindcloud.com', 'phone': '(555) 204-7782',
+         'title': 'Customer Success Manager', 'department': 'Customer Success',
+         'role': 'Manager', 'avatar': 'https://i.pravatar.cc/150?u=user-2',
+         'timezone': 'America/Los_Angeles', 'locale': 'en-US', 'theme': 'lightning'},
+        {'userId': 'user-3', 'firstName': 'Priya', 'lastName': 'Nair',
+         'email': 'priya.nair@northwindcloud.com', 'phone': '(555) 204-7783',
+         'title': 'Account Executive', 'department': 'Sales',
+         'role': 'Rep', 'avatar': 'https://i.pravatar.cc/150?u=user-3',
+         'timezone': 'America/Chicago', 'locale': 'en-US', 'theme': 'lightning'},
+    ]
+
+    def acct(aid, name, atype, industry, revenue, employees, owner):
+        return {
+            'accountId': aid, 'name': name, 'type': atype, 'industry': industry,
+            'revenue': revenue, 'employees': employees, 'ownerId': owner,
+            'billingStreet': '', 'billingCity': '', 'billingState': '',
+            'billingZip': '', 'billingCountry': 'USA',
+            'shippingStreet': '', 'shippingCity': '', 'shippingState': '',
+            'shippingZip': '', 'shippingCountry': 'USA',
+            'phone': '', 'website': '',
+            'createdDate': '2025-09-01T00:00:00.000Z',
+            'modifiedDate': '2026-05-01T00:00:00.000Z',
+        }
+
+    accounts = [
+        acct('account-1', 'Summit Retail Group', 'Customer', 'Retail', 4200000, 540, 'user-1'),
+        acct('account-2', 'Cedar Grove Health', 'Customer', 'Healthcare', 3100000, 410, 'user-1'),
+        acct('account-3', 'Riverstone Media', 'Customer', 'Media', 1800000, 220, 'user-1'),
+        acct('account-4', 'Northstar Apparel', 'Customer', 'Apparel', 2600000, 300, 'user-2'),
+        acct('account-5', 'BlueHarbor Logistics', 'Customer', 'Logistics', 2900000, 360, 'user-1'),
+    ]
+
+    # Default-style activities, NONE referencing Summit churn/usage risk.
+    activities = [
+        {'activityId': 'activity-1', 'type': 'task',
+         'subject': 'Quarterly business review prep — Riverstone Media',
+         'status': 'Open', 'priority': 'Normal', 'dueDate': '2026-07-01',
+         'relatedToType': 'account', 'relatedToId': 'account-3', 'assignedToId': 'user-1'},
+        {'activityId': 'activity-2', 'type': 'event',
+         'subject': 'Onboarding sync — Cedar Grove Health',
+         'status': 'Open', 'priority': 'Normal',
+         'startDateTime': '2026-06-25T15:00:00.000Z',
+         'endDateTime': '2026-06-25T15:30:00.000Z',
+         'relatedToType': 'account', 'relatedToId': 'account-2', 'assignedToId': 'user-1'},
+        {'activityId': 'activity-3', 'type': 'task',
+         'subject': 'Send renewal paperwork — BlueHarbor Logistics',
+         'status': 'Open', 'priority': 'Normal', 'dueDate': '2026-07-10',
+         'relatedToType': 'account', 'relatedToId': 'account-5', 'assignedToId': 'user-1'},
+        {'activityId': 'activity-4', 'type': 'task',
+         'subject': 'Update contact roster — Northstar Apparel',
+         'status': 'Completed', 'priority': 'Low', 'dueDate': '2026-06-10',
+         'relatedToType': 'account', 'relatedToId': 'account-4', 'assignedToId': 'user-2'},
+    ]
+
+    return {
+        'user': user,
+        'users': users,
+        'leads': [],
+        'accounts': accounts,
+        'contacts': [],
+        'opportunities': [],
+        'cases': [],
+        'activities': activities,
+        'chatterPosts': [],
+        'files': [],
+        'following': [],
+        'recentlyViewed': [],
+        'dismissedNotifications': [],
+    }
+
+
+# =========================================================================
+# SLACK STATE
+# =========================================================================
+def build_slack_state():
+    current_user = {
+        'userId': 'user_1', 'fullName': 'Jordan Rivera', 'displayName': 'Jordan',
+        'email': 'jordan.rivera@northwindcloud.com',
+        'avatar': 'https://picsum.photos/200/200?random=1', 'title': 'Customer Success Manager',
+        'status': 'online', 'statusMessage': '', 'statusEmoji': '',
+        'timeZone': 'America/New_York',
+    }
+    users = [
+        current_user,
+        {'userId': 'user_2', 'fullName': 'Dana Lee', 'displayName': 'Dana',
+         'email': 'dana.lee@northwindcloud.com',
+         'avatar': 'https://picsum.photos/200/200?random=2', 'title': 'Customer Success Manager',
+         'status': 'online', 'statusMessage': '', 'statusEmoji': '', 'timeZone': 'America/Los_Angeles'},
+        {'userId': 'user_3', 'fullName': 'Priya Nair', 'displayName': 'Priya',
+         'email': 'priya.nair@northwindcloud.com',
+         'avatar': 'https://picsum.photos/200/200?random=3', 'title': 'Account Executive',
+         'status': 'away', 'statusMessage': '', 'statusEmoji': '', 'timeZone': 'America/Chicago'},
+    ]
+
+    def channel(cid, name, desc, members):
+        return {
+            'channelId': cid, 'name': name, 'description': desc, 'topic': '',
+            'isPrivate': False, 'isStarred': False, 'members': members,
+            'createdBy': 'user_1', 'createdAt': '2026-01-02T10:00:00Z',
+            'pinnedMessages': [], 'unreadCount': 0,
+        }
+
+    channels = [
+        channel('general', 'general', 'Company-wide announcements', ['user_1', 'user_2', 'user_3']),
+        channel('random', 'random', 'Non-work banter', ['user_1', 'user_2', 'user_3']),
+        channel('customer-success', 'customer-success', 'CS team coordination', ['user_1', 'user_2']),
+        channel('acct-summit-retail', 'acct-summit-retail',
+                'Account channel for Summit Retail Group', ['user_1', 'user_3']),
+        channel('acct-northstar', 'acct-northstar',
+                'Account channel for Northstar Apparel', ['user_2', 'user_3']),
+    ]
+
+    def msg(mid, sender, content, ts):
+        return {
+            'messageId': mid, 'senderId': sender, 'content': content,
+            'timestamp': ts, 'threadId': None, 'reactions': [],
+            'attachments': [], 'isEdited': False,
+        }
+
+    messages = {
+        'general': [
+            msg('msg_g1', 'user_2', 'Morning all — reminder that QBR season kicks off next week.',
+                '2026-06-18T13:02:00Z'),
+            msg('msg_g2', 'user_3', 'Thanks Dana, calendars are going out today.',
+                '2026-06-18T13:05:00Z'),
+        ],
+        'random': [
+            msg('msg_r1', 'user_3', 'Coffee machine on 4th floor is fixed.',
+                '2026-06-17T16:20:00Z'),
+        ],
+        'customer-success': [
+            msg('msg_cs1', 'user_2', 'Let me know if anyone needs help with renewals this quarter.',
+                '2026-06-19T14:00:00Z'),
+        ],
+        'acct-summit-retail': [
+            msg('msg_sr1', 'user_3', 'Welcome to the Summit Retail Group account channel.',
+                '2026-01-02T10:30:00Z'),
+            msg('msg_sr2', 'user_1', 'Adding our latest usage dashboard link to the bookmarks.',
+                '2026-05-12T11:15:00Z'),
+        ],
+        'acct-northstar': [
+            msg('msg_ns1', 'user_2', 'Northstar Apparel account channel — main contact is Alex Moreno.',
+                '2026-01-03T09:00:00Z'),
+        ],
+    }
+
+    return {
+        'currentUser': current_user,
+        'workspace': {'workspaceId': 'ws_1', 'workspaceName': 'Northwind Cloud', 'icon': ''},
+        'users': users,
+        'channels': channels,
+        'messages': messages,
+        'threads': {},
+        'dms': [],
+        'bookmarkedMessages': [],
+        'callHistory': [],
+        'settings': {'theme': 'light', 'notifications': 'all', 'displayDensity': 'comfortable',
+                     'showAvatars': True, 'use24Hour': False},
+        'invitations': [],
+        'notifications': [],
+    }
+
+
+# =========================================================================
+# INJECT
+# =========================================================================
+builders = {
+    'notion_mock': build_notion_state,
+    'salesforce_mock': build_salesforce_state,
+    'slack_mock': build_slack_state,
+}
+
+for name, url in MOCKS.items():
+    state = builders[name]()
+    resp = http_post(f'{url}/post?sid={sid}', {'action': 'set', 'state': state})
+    assert resp.status_code == 200, f'{name} injection failed: {resp.status_code} {resp.text}'
+    go = http_get(f'{url}/go?sid={sid}').json()
+    assert go.get('initial_state') is not None, f'{name}: initial_state is None after injection'
+    print(f'Injected {name}: sid={sid}')
+
+print('All mock states injected successfully.')
+
+
+# =========================================================================
+# GUI-READY: launch browser tabs for all three mocks
+# =========================================================================
+def launch_gui(command, delay_sec=1.0):
+    env = os.environ.copy()
+    env['DISPLAY'] = ':0'
+    # Accept both a command string and a pre-split arg list (some tasks
+    # build arg lists themselves, e.g. se_hard_008).
+    if isinstance(command, str):
+        _parts = shlex.split(command)
+    else:
+        _parts = list(command)
+    # Inject CDP debug port so wait_mocks_loaded can verify browser render
+    # via Playwright connect_over_cdp. Idempotent: skip if any arg already
+    # starts with --remote-debugging-port (some tasks inject their own
+    # dynamic port, e.g. se_hard_008).
+    if _parts and 'google-chrome' in _parts[0] \
+       and not any(p.startswith('--remote-debugging-port') for p in _parts):
+        _parts[1:1] = ['--no-first-run', '--no-default-browser-check',
+                       '--remote-debugging-port=1337']
+    _proc = subprocess.Popen(_parts, stdin=subprocess.DEVNULL,
+                             stdout=subprocess.DEVNULL,
+                             stderr=subprocess.DEVNULL, env=env,
+                             start_new_session=True)
+    time.sleep(delay_sec)
+    if _proc.poll() is not None:
+        print('  WARNING: launch_gui process exited early (code=%s)'
+              % _proc.returncode)
+    return _proc
+launch_gui(f'google-chrome "{MOCKS["notion_mock"]}/?sid={sid}"', delay_sec=2.5)
+launch_gui(f'google-chrome "{MOCKS["salesforce_mock"]}/?sid={sid}"', delay_sec=2.0)
+launch_gui(f'google-chrome "{MOCKS["slack_mock"]}/?sid={sid}"', delay_sec=2.0)
+wait_mocks_loaded()
+print('GUI_READY: launched Chrome tabs for notion, salesforce, slack with DISPLAY=:0')

csm_health_risk_002/reward.py

@@ -0,0 +1,399 @@
+# --- MOCK_HOST_NO_PROXY_OVERRIDE: bypass woa proxy for locally-deployed mocks ---
+import os as _os
+_np = _os.environ.get("no_proxy", _os.environ.get("NO_PROXY", ""))
+for _h in ("127.0.0.1", "localhost", "28.7.184.198"):
+    if _h not in _np:
+        _np = (_np + "," + _h).lstrip(",")
+_os.environ["no_proxy"] = _np
+_os.environ["NO_PROXY"] = _np
+# --- end override ---
+
+"""
+Reward Script: CSM logs churn risk for the one Red+declining account in their book
+                (Summit Retail Group) across Notion, Salesforce, and Slack.
+Task ID: csm_health_risk_002
+Domain: mock_websites (multi-mock: notion_mock, salesforce_mock, slack_mock)
+
+Scoring (1.0 total, 100% PROGRAMMATIC — LLM judge intentionally NOT used):
+  Component 1 — Notion Risk Note on Jordan Rivera's Summit Retail Group row (0.34)
+      0.17  item-summit 'Risk Note' went blank -> non-empty, AND every OTHER
+            account row's Risk Note remains blank, with the row owned by Jordan Rivera
+      0.17  the note content explains the risk (declining/Down usage + Red/churn)
+  Component 2 — Salesforce risk task on Jordan Rivera's account-1 (0.33)
+      0.13  at least one new task (type 'task') added to activities
+      0.20  the new task: relatedToType 'account', relatedToId 'account-1',
+            account owner/assignedToId 'user-1' (Jordan Rivera),
+            status != 'Completed', and text indicates health/risk or usage decline
+  Component 3 — Slack heads-up from Jordan Rivera in acct-summit-retail (0.33)
+      0.13  a new message in 'acct-summit-retail', AND 'acct-northstar'
+            and 'general' untouched, sent by 'user_1' Jordan Rivera
+      0.20  the message names Summit, states Red/at-risk + declining usage,
+            and gives a next step or owner
+
+All checks are anchored to the initial->current delta so reward(initial_env)==0.0.
+"""
+import os
+import re
+import sys
+
+import requests
+
+CURRENT_USER_NAME = 'Jordan Rivera'
+NOTION_USER_ID = 'user-1'
+SALESFORCE_USER_ID = 'user-1'
+SLACK_USER_ID = 'user_1'
+TARGET_NOTION_ITEM_ID = 'item-summit'
+TARGET_ACCOUNT_ID = 'account-1'
+TARGET_SLACK_CHANNEL_ID = 'acct-summit-retail'
+
+# ---------------------------------------------------------------------------
+# Read sid
+# ---------------------------------------------------------------------------
+try:
+    with open('/tmp/task_web_sid') as f:
+        sid = f.read().strip()
+    if not sid:
+        raise ValueError('sid is empty')
+except Exception as e:
+    print(f'CRITICAL: Cannot read sid from /tmp/task_web_sid: {e}')
+    print('REWARD: 0.0')
+    sys.exit(0)
+
+MOCKS = {
+    'notion': 'http://28.7.184.198:8166',
+    'salesforce': 'http://28.7.184.198:8175',
+    'slack': 'http://28.7.184.198:8178',
+}
+
+# ---------------------------------------------------------------------------
+# Egress proxy helper (MANDATORY in this cluster — direct egress to
+# *.xlang.ai fails from inside the Docker VMs)
+# ---------------------------------------------------------------------------
+PROXY_CANDIDATES = [
+    os.environ.get('HTTPS_PROXY') or os.environ.get('https_proxy'),
+        None,
+]
+
+
+def resolve_proxy(probe_url):
+    try:
+        requests.get(probe_url, timeout=8)
+        return None
+    except Exception:
+        pass
+    for candidate in PROXY_CANDIDATES:
+        if not candidate:
+            continue
+        try:
+            requests.get(
+                probe_url,
+                timeout=12,
+                proxies={'http': candidate, 'https': candidate},
+            )
+            return candidate
+        except Exception:
+            continue
+    return None
+
+
+def fetch_state(name, url):
+    """Return (initial_state, current_state) dicts for a mock, or (None, None)."""
+    proxy = resolve_proxy(f'{url}/go?sid=conn-probe')
+    proxies = {'http': proxy, 'https': proxy} if proxy else None
+    resp = requests.get(f'{url}/go?sid={sid}', timeout=15, proxies=proxies)
+    resp.raise_for_status()
+    data = resp.json()
+    return data.get('initial_state') or {}, data.get('current_state') or {}
+
+
+# ---------------------------------------------------------------------------
+# Fetch all three mocks up front; any failure -> 0.0
+# ---------------------------------------------------------------------------
+states = {}
+try:
+    for name, url in MOCKS.items():
+        init, curr = fetch_state(name, url)
+        if not curr:
+            print(f'CRITICAL: current_state empty for {name}')
+            print('REWARD: 0.0')
+            sys.exit(0)
+        states[name] = (init, curr)
+except Exception as e:
+    print(f'CRITICAL: Cannot fetch combined mock state: {e}')
+    print('REWARD: 0.0')
+    sys.exit(0)
+
+
+def value_text(value):
+    """Return a comparable text value for raw or lightly typed mock fields."""
+    if value is None:
+        return ''
+    if isinstance(value, str):
+        return value
+    if isinstance(value, (int, float, bool)):
+        return str(value)
+    if isinstance(value, dict):
+        parts = []
+        for key in ('name', 'label', 'title', 'text', 'content', 'value', 'id'):
+            if key in value:
+                parts.append(value_text(value.get(key)))
+        if not parts:
+            parts = [value_text(v) for v in value.values()]
+        return ' '.join(p for p in parts if p)
+    if isinstance(value, list):
+        return ' '.join(value_text(v) for v in value)
+    return str(value)
+
+
+def normalize(text):
+    text = value_text(text).lower()
+    text = text.replace('&', ' and ')
+    text = re.sub(r'[_\-/]+', ' ', text)
+    text = re.sub(r'\s+', ' ', text)
+    return text.strip()
+
+
+def has_any(text, keywords):
+    low = normalize(text)
+    return any(k in low for k in keywords)
+
+
+# Keyword groups. These intentionally allow common UI/text variations while
+# still anchoring on the requested customer, health/risk, and next action.
+JORDAN_KW = ['jordan rivera', 'jordan']
+ACCOUNT_KW = ['summit retail group', 'summit', 'srg']
+USAGE_DECLINE_KW = [
+    'declin', 'down', 'drop', 'decreas', 'fall', 'low', 'lower', 'slip',
+    'slipped', 'usage', 'mau', 'adoption', 'engagement', 'activity',
+    'utilization',
+]
+RED_RISK_KW = [
+    'red', 'at risk', 'atrisk', 'risk', 'churn', 'retention', 'health',
+    'unhealthy', 'critical', 'escalat', 'renewal concern',
+]
+NEXTSTEP_KW = [
+    'next step', 'check in', 'checkin', 'recovery', 'recover', 'plan',
+    'schedul', 'owner', 'will', 'follow up', 'followup', 'escalat',
+    'meeting', 'reach out', 'contact', 'action', 'mitigat', 'sync',
+]
+
+
+def is_jordan_name(text):
+    return has_any(text, JORDAN_KW)
+
+
+def mentions_account(text):
+    return has_any(text, ACCOUNT_KW)
+
+
+def mentions_usage_decline(text):
+    return has_any(text, USAGE_DECLINE_KW)
+
+
+def mentions_red_or_risk(text):
+    return has_any(text, RED_RISK_KW)
+
+
+def new_records(initial, current, id_key):
+    init_ids = {item.get(id_key) for item in (initial or []) if item.get(id_key)}
+    added = [item for item in (current or []) if item.get(id_key) not in init_ids]
+    if added or init_ids:
+        return added
+    return (current or [])[len(initial or []):]
+
+
+def verify_task():
+    total_score = 0.0
+
+    # =====================================================================
+    # Component 1 — Notion Risk Note on Summit Retail Group (0.34)
+    # =====================================================================
+    try:
+        n_init, n_curr = states['notion']
+        init_pages = n_init.get('pages', {})
+        curr_pages = n_curr.get('pages', {})
+
+        notion_user = n_curr.get('user', {})
+        current_user_is_jordan = (
+            notion_user.get('id') == NOTION_USER_ID
+            and is_jordan_name(notion_user.get('name'))
+        )
+
+        summit = curr_pages.get(TARGET_NOTION_ITEM_ID, {})
+        summit_props = summit.get('properties', {})
+        summit_risk = value_text(summit_props.get('prop-risk')).strip()
+        summit_owner = value_text(summit_props.get('prop-csm')).strip()
+        summit_init_risk = (
+            value_text(
+                init_pages.get(TARGET_NOTION_ITEM_ID, {})
+                .get('properties', {})
+                .get('prop-risk')
+            )
+        ).strip()
+        summit_owned_by_jordan = is_jordan_name(summit_owner)
+
+        # Other rows that MUST stay blank (negative constraint)
+        other_ids = ['item-cedar', 'item-river', 'item-blue', 'item-north']
+        others_blank = all(
+            value_text(
+                curr_pages.get(pid, {}).get('properties', {}).get('prop-risk')
+            ).strip() == ''
+            for pid in other_ids
+        )
+        other_csm_unchanged = all(
+            normalize(curr_pages.get(pid, {}).get('properties', {}).get('prop-csm')) ==
+            normalize(init_pages.get(pid, {}).get('properties', {}).get('prop-csm'))
+            for pid in other_ids
+        )
+
+        # 1a (0.17): Jordan's summit note newly non-empty AND others untouched
+        if (current_user_is_jordan and summit_owned_by_jordan and summit_init_risk == ''
+                and summit_risk != '' and others_blank and other_csm_unchanged):
+            print('PASS: Notion Jordan-owned Summit Risk Note filled, other rows untouched (0.17)')
+            total_score += 0.17
+        else:
+            print(f'FAIL: Notion 1a — summit_init={summit_init_risk!r} '
+                  f'summit_now={summit_risk!r} current_user_is_jordan={current_user_is_jordan} '
+                  f'summit_owner={summit_owner!r} others_blank={others_blank} '
+                  f'other_csm_unchanged={other_csm_unchanged}')
+
+        # 1b (0.17): note content explains the risk
+        if summit_risk != '' and mentions_usage_decline(summit_risk) and mentions_red_or_risk(summit_risk):
+            print('PASS: Notion Risk Note explains declining usage + red/churn risk (0.17)')
+            total_score += 0.17
+        else:
+            print(f'FAIL: Notion 1b — content does not name decline+red/churn: {summit_risk!r}')
+    except Exception as e:
+        print(f'ERROR: Component 1 (Notion) — {e}')
+
+    # =====================================================================
+    # Component 2 — Salesforce risk task on account-1 (0.33)
+    # =====================================================================
+    try:
+        s_init, s_curr = states['salesforce']
+        sf_user = s_curr.get('user', {})
+        current_user_is_jordan = (
+            sf_user.get('userId') == SALESFORCE_USER_ID
+            and is_jordan_name(f'{sf_user.get("firstName", "")} {sf_user.get("lastName", "")}')
+        )
+        accounts_by_id = {a.get('accountId'): a for a in (s_curr.get('accounts', []) or [])}
+        target_account = accounts_by_id.get(TARGET_ACCOUNT_ID, {})
+        target_account_owned_by_jordan = target_account.get('ownerId') == SALESFORCE_USER_ID
+
+        init_acts = s_init.get('activities', []) or []
+        curr_acts = s_curr.get('activities', []) or []
+        new_acts = new_records(init_acts, curr_acts, 'activityId')
+        new_tasks = [a for a in new_acts if normalize(a.get('type')) == 'task']
+
+        # 2a (0.13): at least one new task added
+        if len(new_tasks) >= 1:
+            print('PASS: Salesforce — at least one new task added (0.13)')
+            total_score += 0.13
+        else:
+            print(f'FAIL: Salesforce 2a — expected a new task, found {len(new_tasks)} '
+                  f'(new activities: {[a.get("activityId") for a in new_acts]})')
+
+        # 2b (0.20): the new task is the Summit churn-risk task with correct fields
+        risk_task = None
+        for t in new_tasks:
+            task_text = ' '.join(
+                value_text(t.get(k))
+                for k in ('subject', 'description', 'comments', 'body')
+            )
+            related_to_target = t.get('relatedToId') == TARGET_ACCOUNT_ID
+            if related_to_target and (
+                    mentions_red_or_risk(task_text)
+                    or mentions_usage_decline(task_text)):
+                risk_task = t
+                break
+        if risk_task is not None:
+            status = normalize(risk_task.get('status'))
+            checks = {
+                'current user=Jordan Rivera': current_user_is_jordan,
+                'account.ownerId=user-1 (Jordan Rivera)': target_account_owned_by_jordan,
+                'relatedToType=account': normalize(risk_task.get('relatedToType')) == 'account',
+                'relatedToId=account-1': risk_task.get('relatedToId') == TARGET_ACCOUNT_ID,
+                'assignedToId=user-1 (Jordan Rivera)': risk_task.get('assignedToId') == SALESFORCE_USER_ID,
+                'status!=Completed': status != 'completed',
+            }
+            if all(checks.values()):
+                print('PASS: Salesforce risk task fields correct '
+                      f'(subject={risk_task.get("subject")!r}) (0.20)')
+                total_score += 0.20
+            else:
+                failed = [k for k, v in checks.items() if not v]
+                print(f'FAIL: Salesforce 2b — task field mismatch: {failed} '
+                      f'(task={risk_task})')
+        else:
+            print('FAIL: Salesforce 2b — no new account-1 task with risk/usage context found')
+    except Exception as e:
+        print(f'ERROR: Component 2 (Salesforce) — {e}')
+
+    # =====================================================================
+    # Component 3 — Slack heads-up in acct-summit-retail (0.33)
+    # =====================================================================
+    try:
+        sl_init, sl_curr = states['slack']
+        slack_user = sl_curr.get('currentUser', {})
+        current_user_is_jordan = (
+            slack_user.get('userId') == SLACK_USER_ID
+            and is_jordan_name(slack_user.get('fullName'))
+        )
+        init_msgs = sl_init.get('messages', {}) or {}
+        curr_msgs = sl_curr.get('messages', {}) or {}
+
+        summit_init = init_msgs.get(TARGET_SLACK_CHANNEL_ID, []) or []
+        summit_curr = curr_msgs.get(TARGET_SLACK_CHANNEL_ID, []) or []
+        new_summit = new_records(summit_init, summit_curr, 'messageId')
+        jordan_messages = [m for m in new_summit if m.get('senderId') == SLACK_USER_ID]
+        sent_by_jordan = len(jordan_messages) >= 1
+
+        # negative constraints: acct-northstar and general unchanged
+        north_unchanged = not new_records(
+            init_msgs.get('acct-northstar', []) or [],
+            curr_msgs.get('acct-northstar', []) or [],
+            'messageId',
+        )
+        general_unchanged = not new_records(
+            init_msgs.get('general', []) or [],
+            curr_msgs.get('general', []) or [],
+            'messageId',
+        )
+
+        # 3a (0.13): new message in summit channel from Jordan + others untouched
+        if (current_user_is_jordan and len(new_summit) >= 1
+                and sent_by_jordan and north_unchanged and general_unchanged):
+            print('PASS: Slack — Jordan message in acct-summit-retail, '
+                  'northstar+general untouched (0.13)')
+            total_score += 0.13
+        else:
+            print(f'FAIL: Slack 3a — summit init={len(summit_init)} curr={len(summit_curr)} '
+                  f'current_user_is_jordan={current_user_is_jordan} sent_by_jordan={sent_by_jordan} '
+                  f'north_unchanged={north_unchanged} general_unchanged={general_unchanged}')
+
+        # 3b (0.20): the new message names Summit + Red/at-risk + declining usage + next step
+        matched = None
+        for m in jordan_messages:
+            txt = m.get('content')
+            if (mentions_account(txt)
+                    and mentions_red_or_risk(txt)
+                    and mentions_usage_decline(txt)
+                    and has_any(txt, NEXTSTEP_KW)):
+                matched = m
+                break
+        if matched is not None:
+            print('PASS: Slack message names Summit, Red/at-risk, declining usage, next step (0.20)')
+            total_score += 0.20
+        else:
+            print(f'FAIL: Slack 3b — new message(s) do not cover Summit+red+declining+nextstep: '
+                  f'{[m.get("content") for m in new_summit]}')
+    except Exception as e:
+        print(f'ERROR: Component 3 (Slack) — {e}')
+
+    final_score = round(min(total_score, 1.0), 4)
+    print(f'\nScore: {total_score}/1.0')
+    print(f'REWARD: {final_score}')
+    return final_score
+
+
+verify_task()

csm_health_risk_002/reward_label.json

@@ -0,0 +1,76 @@
+{
+  "reward_file": "/apdcephfs_sh2/share_300000800/user/jackwkwang/long-horizon-tasks/mini_osworld_xiangwu/cache/csm_health_risk_001/reward.py",
+  "api_url": "http://28.7.184.184:8000/v1/chat/completions",
+  "model": "Kimi-K2.6",
+  "max_code_chars": 100000,
+  "created_at": "2026-07-03 20:20:27",
+  "label": {
+    "task_id": "csm_health_risk_002",
+    "domain": "mock_websites",
+    "summary": "验证 CSM Jordan Rivera 是否在 Notion、Salesforce 和 Slack 三个系统中正确记录了 Summit Retail Group 账户的流失风险（Red 状态且使用衰退）。",
+    "is_placeholder": false,
+    "data_sources": [
+      "/tmp/task_web_sid",
+      "notion_mock (http://28.7.186.212:8186)",
+      "salesforce_mock (http://28.7.186.212:8195)",
+      "slack_mock (http://28.7.186.212:8198)"
+    ],
+    "scoring_components": [
+      {
+        "name": "Component 1a — Notion Summit 风险笔记创建",
+        "weight": 0.17,
+        "description": "检查 Notion 中 Jordan Rivera 拥有的 Summit Retail Group 行(item-summit)的 Risk Note 从空白变为非空，且其他账户行的 Risk Note 保持空白、CSM 未改动",
+        "check_logic": "current_user_is_jordan 为真；summit_owned_by_jordan 为真；summit_init_risk == '' 且 summit_risk != ''；others_blank 为真（item-cedar/river/blue/north 的 prop-risk 均为空）；other_csm_unchanged 为真",
+        "pass_condition": "当前用户是 Jordan Rivera；item-summit 的 prop-risk 初始为空且当前非空；该行 CSM 属于 Jordan；其余四行账户的 prop-risk 仍为空且 CSM 与初始状态一致"
+      },
+      {
+        "name": "Component 1b — Notion 风险笔记内容",
+        "weight": 0.17,
+        "description": "检查 Summit Retail Group 的 Risk Note 内容是否解释了使用衰退和 Red/流失风险",
+        "check_logic": "summit_risk != '' 且 mentions_usage_decline(summit_risk) 返回真且 mentions_red_or_risk(summit_risk) 返回真",
+        "pass_condition": "笔记非空，且文本经 normalize 后同时包含 decline/down/drop 等使用衰退关键词和 red/at risk/churn 等风险关键词"
+      },
+      {
+        "name": "Component 2a — Salesforce 新增任务",
+        "weight": 0.13,
+        "description": "检查 Salesforce 的 activities 中是否至少新增了一条 type 为 'task' 的记录",
+        "check_logic": "对比 initial 和 current 的 activities 列表（按 activityId 去重），new_tasks = [a for a in new_acts if normalize(a.get('type')) == 'task']，要求 len(new_tasks) >= 1",
+        "pass_condition": "至少存在一条新增 activity，其 type 字段归一化后为 'task'"
+      },
+      {
+        "name": "Component 2b — Salesforce 风险任务字段",
+        "weight": 0.2,
+        "description": "检查新增任务是否针对 account-1、字段正确且文本体现健康风险或使用衰退",
+        "check_logic": "遍历 new_tasks，拼接 subject/description/comments/body 为 task_text；找到 relatedToId == 'account-1' 且 (mentions_red_or_risk(task_text) 或 mentions_usage_decline(task_text)) 的任务；再校验 current_user_is_jordan、target_account_owned_by_jordan、relatedToType 归一化后为 'account'、assignedToId == 'user-1'、status 归一化后 != 'completed'",
+        "pass_condition": "新增任务关联 account-1、类型为 account、分配给 user-1、状态非 Completed、当前用户为 Jordan、账户所有者为 Jordan，且任务文本包含风险或使用衰退关键词"
+      },
+      {
+        "name": "Component 3a — Slack 频道消息",
+        "weight": 0.13,
+        "description": "检查 Jordan Rivera 是否在 acct-summit-retail 频道发送了新消息，且 acct-northstar 和 general 频道未被改动",
+        "check_logic": "current_user_is_jordan 为真；new_records(summit_init, summit_curr, 'messageId') 非空且其中至少一条 senderId == 'user_1'；new_records 对 acct-northstar 和 general 返回空列表（north_unchanged 和 general_unchanged 为真）",
+        "pass_condition": "当前 Slack 用户为 Jordan；acct-summit-retail 频道有新增消息且由 user_1 发送；acct-northstar 和 general 频道无新增消息"
+      },
+      {
+        "name": "Component 3b — Slack 消息内容",
+        "weight": 0.2,
+        "description": "检查新消息是否提到 Summit、Red/风险、使用衰退，并给出下一步或负责人",
+        "check_logic": "遍历 jordan_messages，检查 m.get('content') 同时满足 mentions_account(txt)、mentions_red_or_risk(txt)、mentions_usage_decline(txt)、has_any(txt, NEXTSTEP_KW)",
+        "pass_condition": "消息内容经 normalize 后同时包含 Summit 相关关键词、Red/风险关键词、使用衰退关键词以及 next step/owner/action 等下一步关键词"
+      }
+    ],
+    "total_max_score": 1.0,
+    "score_aggregation": "各子组件分数直接相加，最终通过 min(total_score, 1.0) 钳制上限为 1.0，再四舍五入保留 4 位小数",
+    "failure_modes": [
+      "读取 /tmp/task_web_sid ���败或为空：打印 CRITICAL 并返回 0.0",
+      "从任一 mock 服务获取状态失败或 current_state 为空：打印 CRITICAL 并返回 0.0",
+      "Notion 检查中当前用户非 Jordan、Summit 行未新建 Risk Note、其他行被改动：Component 1 不得分",
+      "Salesforce 中未新增 task 类型活动，或新增任务字段不匹配 account-1/非 Completed/未分配 Jordan：Component 2 不得分",
+      "Slack 中 acct-summit-retail 无 Jordan 的新消息，或其他频道被发送消息：Component 3 不得分",
+      "各组件内部异常被捕获为 ERROR，仅影响该组件得分，不会导致整体提前退出"
+    ],
+    "scoring_logic_explanation": "脚本首先读取 /tmp/task_web_sid 获取会话标识，失败则直接返回 0.0。随后通过 HTTP 请求向 notion_mock、salesforce_mock、slack_mock 三个服务的 /go?sid={sid} 端点拉取 initial_state 与 current_state，任一拉取失败或 current_state 为空即返回 0.0。接着进入 verify_task()：Component 1 针对 Notion，检查 item-summit 的 prop-risk 是否由 Jordan 从空白填为有效内容（0.17），并验证内容包含使用衰退与 Red/流失风险语义（0.17），同时要求其余四行账户的 Risk Note 仍为空白且 CSM 未变；Component 2 针对 Salesforce，对比 activities 列表找出新增 task，至少一条新增 task 得 0.13，再验证该任务关联 account-1、类型为 account、分配给 user-1、状态非 Completed 且文本含风险/衰退语义得 0.20；Component 3 针对 Slack，检查 acct-summit-retail 频道有 Jordan 发送的新消息且 acct-northstar 与 general 频道无变动得 0.13，再验证消息内容同时提及 Summit、Red/风险、使用衰退及下一步行动得 0.20。所有子项得分累加后，以 min(total_score, 1.0) 限制上限并四舍五入到 4 位小数作为最终 REWARD。",
+    "confidence": 1.0
+  },
+  "raw_response": "用户要求我分析一个 reward.py 评分脚本的评分逻辑，并输出一个严格的 JSON 对象。\n\n让我仔细阅读代码并提取所需信息：\n\n1. **task_id**: 从 docstring 中可以看到 \"Task ID: csm_health_risk_001\"\n2. **domain**: 从 docstring 中可以看到 \"Domain: mock_websites (multi-mock: notion_mock, salesforce_mock, slack_mock)\"，所以 domain 是 mock_websites\n3. **summary**: 一句话概括。这个脚本是验证 CSM (Customer Success Manager) Jordan Rivera 是否在 Notion、Salesforce 和 Slack 中正确记录了 Summit Retail Group 这个红标/衰退客户的风险信息。\n4. **is_placeholder**: false，因为脚本有详细的检查逻辑，不是占位脚本。\n5. **data_sources**: 脚本从以下数据源获取状态：\n   - `/tmp/task_web_sid` (读取 sid)\n   - `http://28.7.186.212:8166` (notion mock)\n   - `http://28.7.186.212:8175` (salesforce mock)\n   - `http://28.7.186.212:8178` (slack mock)\n   通过 `/go?sid={sid}` 端点获取 initial_state 和 current_state。\n\n6. **scoring_components**:\n   - Component 1: Notion Risk Note (0.34 total)\n     - 1a (0.17): 检查当前用户是 Jordan Rivera，Summit Retail Group 行 (item-summit) 的 Risk Note (prop-risk) 从空变为非空，且其他行 (item-cedar, item-river, item-blue, item-north) 的 Risk Note 保持空白，且其他行的 CSM 未改变。\n     - 1b (0.17): Risk Note 内容提到 usage decline 和 red/churn risk。\n   - Component 2: Salesforce risk task (0.33 total)\n     - 2a (0.13): 至少添加了一个新的 task 类型活动。\n     - 2b (0.20): 新任务关联到 account-1，relatedToType 是 account，assignedToId 是 user-1 (Jordan)，状态不是 Completed，且文本提到 health/risk 或 usage decline。\n   - Component 3: Slack heads-up (0.33 total)\n     - 3a (0.13): 在 acct-summit-retail 频道有 Jordan 发送的新消息，且 acct-northstar 和 general 频道未改变。\n     - 3b (0.20): 消息提到 Summit、Red/at-risk、declining usage，并给出 next step/owner。\n\n7. **total_max_score**: 1.0\n8. **score_aggregation**: 各组件分数相加，最后 `round(min(total_score, 1.0), 4)`，所以是相加后钳制到上限 1.0。\n9. **failure_modes**:\n   - 读取 /tmp/task_web_sid 失败或为空 -> CRITICAL, REWARD: 0.0, sys.exit(0)\n   - 获取 mock 状态失败或 current_state 为空 -> CRITICAL, REWARD: 0.0, sys.exit(0)\n   - 各组件内部异常会被捕获并打印 ERROR，但不会导致提前退出，只是该组件不得分。\n   - 注意：如果检查条件不满足，只是不加分，不会提前退出。\n\n10. **scoring_logic_explanation**: 需要详细解释执行顺序：\n    - 首先读取 /tmp/task_web_sid 获取会话 ID，失败则返回 0.0。\n    - 然后依次从 notion、salesforce、slack 三个 mock 服务获取 initial_state 和 current_state，通过 /go?sid={sid} 端点，任何失败或 current_state 为空则返回 0.0。\n    - 接着计算总分：\n      - Component 1 (Notion): 检查 item-summit 的 prop-risk 从空变非空且由 Jordan 拥有，其他行未动 (0.17)；再检查内容包含 decline 和 red/risk 关键词 (0.17)。\n      - Component 2 (Salesforce): 检查 activities 中新增至少一个 task 类型记录 (0.13)；再检查该任务关联 account-1、类型为 account、分配给 user-1、状态非 Completed、文本包含 risk/usage decline (0.20)。\n      - Component 3 (Slack): 检查 acct-summit-retail 频道有 Jordan 发送的新消息且其他频道未变 (0.13)；再检查消息内容提到 Summit、Red/at-risk、declining usage 和 next step (0.20)。\n    - 最后将各组件得分累加，用 min(total_score, 1.0) 钳制上限，四舍五入到 4 位小数输出。\n\n11. **confidence**: 1.0，因为代码逻辑非常清晰完整。\n\n现在构建 JSON。注意：\n- 所有字符串 value 使用中文，但 task_id、domain、字段名、代码标识符保留原样。\n- 不要输出 Markdown 代码块标记。\n- 不要输出额外解释。\n- 不要输出思考过程。\n- 第一个字符必须是 {，最后一个字符必须是 }。\n\n让我仔细构造每个字段：\n\ntask_id: \"csm_health_risk_001\"\ndomain: \"mock_websites\"\nsummary: \"验证 CSM Jordan Rivera 是否在 Notion、Salesforce 和 Slack 三个系统中正确记录了 Summit Retail Group 账户的流失风险（Red 状态且使用衰退）。\"\nis_placeholder: false\ndata_sources: [\"/tmp/task_web_sid\", \"notion_mock (http://28.7.186.212:8186)\", \"salesforce_mock (http://28.7.186.212:8195)\", \"slack_mock (http://28.7.186.212:8198)\"]\n\nscoring_components:\n- name: \"Component 1a — Notion Summit 风险笔记创建\"\n  weight: 0.17\n  description: \"检查 Notion 中 Jordan Rivera 拥有的 Summit Retail Group 行(item-summit)的 Risk Note 从空白变为非空，且其他账户行的 Risk Note 保持空白、CSM 未改动\"\n  check_logic: \"current_user_is_jordan 为真；summit_owned_by_jordan 为真；summit_init_risk == '' 且 summit_risk != ''；others_blank 为真（item-cedar/river/blue/north 的 prop-risk 均为空）；other_csm_unchanged 为真\"\n  pass_condition: \"当前用户是 Jordan Rivera；item-summit 的 prop-risk 初始为空且当前非空；该行 CSM 属于 Jordan；其余四行 prop-risk 仍为空且 CSM 与初始状态一致\"\n\n- name: \"Component 1b — Notion 风险笔记内容\"\n  weight: 0.17\n  description: \"检查 Summit Retail Group 的 Risk Note 内容是否解释了使用衰退和 Red/流失风险\"\n  check_logic: \"summit_risk != '' 且 mentions_usage_decline(summit_risk) 返回真且 mentions_red_or_risk(summit_risk) 返回真\"\n  pass_condition: \"笔记非空，且文本经 normalize 后同时包含 decline/down/drop 等使用衰退关键词和 red/at risk/churn 等风险关键词\"\n\n- name: \"Component 2a — Salesforce 新增任务\"\n  weight: 0.13\n  description: \"检查 Salesforce 的 activities 中是否至少新增了一条 type 为 'task' 的记录\"\n  check_logic: \"对比 initial 和 current 的 activities 列表（按 activityId 去重），new_tasks = [a for a in new_acts if normalize(a.get('type')) == 'task']，要求 len(new_tasks) >= 1\"\n  pass_condition: \"至少存在一条新增 activity，其 type 字段归一化后为 'task'\"\n\n- name: \"Component 2b — Salesforce 风险任务字段\"\n  weight: 0.20\n  description: \"检查新增任务是否针对 account-1、字段正确且文本体现健康风险或使用衰退\"\n  check_logic: \"遍历 new_tasks，拼接 subject/description/comments/body 为 task_text；找到 relatedToId == 'account-1' 且 (mentions_red_or_risk(task_text) 或 mentions_usage_decline(task_text)) 的任务；再校验 current_user_is_jordan、target_account_owned_by_jordan、relatedToType 归一化后为 'account'、assignedToId == 'user-1'、status 归一化后 != 'completed'\"\n  pass_condition: \"新增任务关联 account-1、类型为 account、分配给 user-1、状态非 Completed、当前用户为 Jordan、账户所有者为 Jordan，且任务文本包含风险或使用衰退关键词\"\n\n- name: \"Component 3a — Slack 频道消息\"\n  weight: 0.13\n  description: \"检查 Jordan Rivera 是否在 acct-summit-retail 频道发送了新消息，且 acct-northstar 和 general 频道未被改动\"\n  check_logic: \"current_user_is_jordan 为真；new_records(summit_init, summit_curr, 'messageId') 非空且其中至少一条 senderId == 'user_1'；new_records 对 acct-northstar 和 general 返回空列表（north_unchanged 和 general_unchanged 为真）\"\n  pass_condition: \"当前 Slack 用户为 Jordan；acct-summit-retail 频道有新增消息且由 user_1 发送；acct-northstar 和 general 频道无新增消息\"\n\n- name: \"Component 3b — Slack 消息内容\"\n  weight: 0.20\n  description: \"检查新消息是否提到 Summit、Red/风险、使用衰退，并给出下一步或负责人\"\n  check_logic: \"遍历 jordan_messages，检查 m.get('content') 同时满足 mentions_account(txt)、mentions_red_or_risk(txt)、mentions_usage_decline(txt)、has_any(txt, NEXTSTEP_KW)\"\n  pass_condition: \"消息内容经 normalize 后同时包含 Summit 相关关键词、Red/风险关键词、使用衰退关键词以及 next step/owner/action 等下一步关键词\"\n\ntotal_max_score: 1.0\nscore_aggregation: \"各子组件分数直接相加，最终通过 min(total_score, 1.0) 钳制上限为 1.0，再四舍五入保留 4 位小数\"\nfailure_modes: [\n  \"读取 /tmp/task_web_sid 失败或为空：打印 CRITICAL 并返回 0.0\",\n  \"从任一 mock 服务获取状态失败或 current_state 为空：打印 CRITICAL 并返回 0.0\",\n  \"Notion 检查中当前用户非 Jordan、Summit 行未新建 Risk Note、其他行被改动：Component 1 不得分\",\n  \"Salesforce 中未新增 task 类型活动，或新增任务字段不匹配 account-1/非 Completed/未分配 Jordan：Component 2 不得分\",\n  \"Slack 中 acct-summit-retail 无 Jordan 的新消息，或其他频道被发送消息：Component 3 不得分\",\n  \"各组件内部异常被捕获为 ERROR，仅影响该组件得分，不会导致整体提前退出\"\n]\n\nscoring_logic_explanation: \"脚本首先读取 /tmp/task_web_sid 获取会话标识，失败则直接返回 0.0。随后通过 HTTP 请求向 notion_mock、salesforce_mock、slack_mock 三个服务的 /go?sid={sid} 端点拉取 initial_state 与 current_state，任一拉取失败或 current_state 为空即返回 0.0。接着进入 verify_task()：Component 1 针对 Notion，检查 item-summit 的 prop-risk 是否由 Jordan 从空白填为有效内容（0.17），并验证内容包含使用衰退与 Red/流失风险语义（0.17），同时要求其余四行账户的 Risk Note 仍为空白且 CSM 未变；Component 2 针对 Salesforce，对比 activities 列表找出新增 task，至少一条新增 task 得 0.13，再验证该任务关联 account-1、类型为 account、分配给 user-1、状态非 Completed 且文本含风险/衰退语义得 0.20；Component 3 针对 Slack，检查 acct-summit-retail 频道有 Jordan 发送的新消息且 acct-northstar 与 general 频道无变动得 0.13，再验证消息内容同时提及 Summit、Red/风险、使用衰退及下一步行动得 0.20。所有子项得分累加后，以 min(total_score, 1.0) 限制上限并四舍五入到 4 位小数作为最终 REWARD。\"\n\nconfidence: 1.0\n\n检查 JSON 格式要求：\n- 只输出一个 JSON 对象\n- 不要输出 Markdown 代码块标记\n- 不要输出额外解释\n- 不要输出思考过程\n- 不要输出 <thinking> 标签\n- 输出的第一个字符必须是 {，最后一个字符必须是 }\n\n确保所有字符串值都是中文（除了代码标识符和特定名称）。\n\n再检查一下 weight 字段。代码中显式给出了分值：0.17, 0.17, 0.13, 0.20, 0.13, 0.20。总和正好是 1.0。\n\n确认 data_sources 的写法。代码中是通过 MOCKS 字典定义的，端点是 http://28.7.186.212:8166 等。可以写为 notion_mock、salesforce_mock、slack_mock，并注明端点。\n\n确认 failure_modes 是否包含提前退出路径：\n- sid 读取失败：是，sys.exit(0)\n- mock 拉取失败：是，sys.exit(0)\n- 其他是不得分，不是提前退出，但也属于失败路径。\n\n确认 score_aggregation：\"相加后钳制到上限\" 是准确的。\n\n好，我现在可以输出最终的 JSON 了。 </think> {\"task_id\":\"csm_health_risk_001\",\"domain\":\"mock_websites\",\"summary\":\"验证 CSM Jordan Rivera 是否在 Notion、Salesforce 和 Slack 三个系统中正确记录了 Summit Retail Group 账户的流失风险（Red 状态且使用衰退）。\",\"is_placeholder\":false,\"data_sources\":[\"/tmp/task_web_sid\",\"notion_mock (http://28.7.186.212:8186)\",\"salesforce_mock (http://28.7.186.212:8195)\",\"slack_mock (http://28.7.186.212:8198)\"],\"scoring_components\":[{\"name\":\"Component 1a — Notion Summit 风险笔记创建\",\"weight\":0.17,\"description\":\"检查 Notion 中 Jordan Rivera 拥有的 Summit Retail Group 行(item-summit)的 Risk Note 从空白变为非空，且其他账户行的 Risk Note 保持空白、CSM 未改动\",\"check_logic\":\"current_user_is_jordan 为真；summit_owned_by_jordan 为真；summit_init_risk == '' 且 summit_risk != ''；others_blank 为真（item-cedar/river/blue/north 的 prop-risk 均为空）；other_csm_unchanged 为真\",\"pass_condition\":\"当前用户是 Jordan Rivera；item-summit 的 prop-risk 初始为空且当前非空；该行 CSM 属于 Jordan；其余四行账户的 prop-risk 仍为空且 CSM 与初始状态一致\"},{\"name\":\"Component 1b — Notion 风险笔记内容\",\"weight\":0.17,\"description\":\"检查 Summit Retail Group 的 Risk Note 内容是否解释了使用衰退和 Red/流失风险\",\"check_logic\":\"summit_risk != '' 且 mentions_usage_decline(summit_risk) 返回真且 mentions_red_or_risk(summit_risk) 返回真\",\"pass_condition\":\"笔记非空，且文本经 normalize 后同时包含 decline/down/drop 等使用衰退关键词和 red/at risk/churn 等风险关键词\"},{\"name\":\"Component 2a — Salesforce 新增任务\",\"weight\":0.13,\"description\":\"检查 Salesforce 的 activities 中是否至少新增了一条 type 为 'task' 的记录\",\"check_logic\":\"对比 initial 和 current 的 activities 列表（按 activityId 去重），new_tasks = [a for a in new_acts if normalize(a.get('type')) == 'task']，要求 len(new_tasks) >= 1\",\"pass_condition\":\"至少存在一条新增 activity，其 type 字段归一化后为 'task'\"},{\"name\":\"Component 2b — Salesforce 风险任务字段\",\"weight\":0.2,\"description\":\"检查新增任务是否针对 account-1、字段正确且文本体现健康风险或使用衰退\",\"check_logic\":\"遍历 new_tasks，拼接 subject/description/comments/body 为 task_text；找到 relatedToId == 'account-1' 且 (mentions_red_or_risk(task_text) 或 mentions_usage_decline(task_text)) 的任务；再校验 current_user_is_jordan、target_account_owned_by_jordan、relatedToType 归一化后为 'account'、assignedToId == 'user-1'、status 归一化后 != 'completed'\",\"pass_condition\":\"新增任务关联 account-1、类型为 account、分配给 user-1、状态非 Completed、当前用户为 Jordan、账户所有者为 Jordan，且任务文本包含风险或使用衰退关键词\"},{\"name\":\"Component 3a — Slack 频道消息\",\"weight\":0.13,\"description\":\"检查 Jordan Rivera 是否在 acct-summit-retail 频道发送了新消息，且 acct-northstar 和 general 频道未被改动\",\"check_logic\":\"current_user_is_jordan 为真；new_records(summit_init, summit_curr, 'messageId') 非空且其中至少一条 senderId == 'user_1'；new_records 对 acct-northstar 和 general 返回空列表（north_unchanged 和 general_unchanged 为真）\",\"pass_condition\":\"当前 Slack 用户为 Jordan；acct-summit-retail 频道有新增消息且由 user_1 发送；acct-northstar 和 general 频道无新增消息\"},{\"name\":\"Component 3b — Slack 消息内容\",\"weight\":0.2,\"description\":\"检查新消息是否提到 Summit、Red/风险、使用衰退，并给出下一步或负责人\",\"check_logic\":\"遍历 jordan_messages，检查 m.get('content') 同时满足 mentions_account(txt)、mentions_red_or_risk(txt)、mentions_usage_decline(txt)、has_any(txt, NEXTSTEP_KW)\",\"pass_condition\":\"消息内容经 normalize 后同时包含 Summit 相关关键词、Red/风险关键词、使用衰退关键词以及 next step/owner/action 等下一步关键词\"}],\"total_max_score\":1.0,\"score_aggregation\":\"各子组件分数直接相加，最终通过 min(total_score, 1.0) 钳制上限为 1.0，再四舍五入保留 4 位小数\",\"failure_modes\":[\"读取 /tmp/task_web_sid 失败或为空：打印 CRITICAL 并返回 0.0\",\"从任一 mock 服务获取状态失败或 current_state 为空：打印 CRITICAL 并返回 0.0\",\"Notion 检查中当前用户非 Jordan、Summit 行未新建 Risk Note、其他行被改动：Component 1 不得分\",\"Salesforce 中未新增 task 类型活动，或新增任务字段不匹配 account-1/非 Completed/未分配 Jordan：Component 2 不得分\",\"Slack 中 acct-summit-retail 无 Jordan 的新消息，或其他频道被发送消息：Component 3 不得分\",\"各组件内部异常被捕获为 ERROR，仅影响该组件得分，不会导致整体提前退出\"],\"scoring_logic_explanation\":\"脚本首先读取 /tmp/task_web_sid 获取会话标识，失败则直接返回 0.0。随后通过 HTTP 请求向 notion_mock、salesforce_mock、slack_mock 三个服务的 /go?sid={sid} 端点拉取 initial_state 与 current_state，任一拉取失败或 current_state 为空即返回 0.0。接着进入 verify_task()：Component 1 针对 Notion，检查 item-summit 的 prop-risk 是否由 Jordan 从空白填为有效内容（0.17），并验证内容包含使用衰退与 Red/流失风险语义（0.17），同时要求其余四行账户的 Risk Note 仍为空白且 CSM 未变；Component 2 针对 Salesforce，对比 activities 列表找出新增 task，至少一条新增 task 得 0.13，再验证该任务关联 account-1、类型为 account、分配给 user-1、状态非 Completed 且文本含风险/衰退语义得 0.20；Component 3 针对 Slack，检查 acct-summit-retail 频道有 Jordan 发送的新消息且 acct-northstar 与 general 频道无变动得 0.13，再验证消息内容同时提及 Summit、Red/风险、使用衰退及下一步行动得 0.20。所有子项得分累加后，以 min(total_score, 1.0) 限制上限并四舍五入到 4 位小数作为最终 REWARD。\",\"confidence\":1.0}"
+}

csm_onboarding_checklist_004/_cua_gym_vm_bridge.sh

@@ -0,0 +1,106 @@
+#!/bin/bash
+# CUA-Gym VM-side network fix (direct-IP mock access).
+#
+# Runs INSIDE the QEMU guest. The CUA-Gym mock services are now reached by
+# their real IDC address (http://28.7.184.198:81xx) instead of the old
+# localhost:9000-9097 socat bridge, so the VM only needs two fixes:
+#
+#   1. Working DNS — the qcow2-baked /etc/resolv.conf points at the Pod's
+#      cluster DNS, which is unreachable from QEMU SLIRP NAT
+#      ("Temporary failure in name resolution").
+#   2. The mock host excluded from the woa http proxy — otherwise the VM's
+#      Chrome (gsettings proxy) and initial_setup.py's `requests` get a 403
+#      from the proxy when hitting 28.7.184.198:81xx. (The old socat path never
+#      hit this because localhost is bypassed by default.)
+#
+# Idempotent. Scoped to the CUA-Gym bridge step; does not touch the generic
+# OSWorld setup pipeline.
+#
+# Argv:
+#   $1  optional sudo password (default "password"; SetupController
+#       substitutes {CLIENT_PASSWORD} when uploaded as a cache file).
+#   $2  unused (kept for backward compat — was socat first port 9000).
+#   $3  unused (kept for backward compat — was socat last  port 9097).
+#   $4  optional mock host IP (default 28.7.184.198).
+
+set +e
+
+PASS="${1:-password}"
+MOCK_HOST="${4:-28.7.184.198}"
+# Derive the /16 IDC subnet so future port/host tweaks on the same segment
+# stay bypassed too (28.7.184.198 -> 28.7.0.0/16).
+MOCK_NET="$(echo "${MOCK_HOST}" | awk -F. 'NF==4{print $1"."$2".0.0/16"}')"
+
+run_root() {
+    echo "$PASS" | sudo -S -p "" bash -c "$1"
+}
+
+# Detect SLIRP gateway from VM's routing table; fall back to 10.0.2.2.
+GW="$(ip route 2>/dev/null | awk '/default/ {print $3; exit}')"
+GW="${GW:-10.0.2.2}"
+
+# ─── CUA-Gym network fix (Taiji-specific) ─────────────────────────────
+fix_vm_network() {
+    # 0) Discover VM network state at runtime — gateway, own IP/CIDR, SLIRP
+    #    DNS forwarder. Nothing hardcoded except the mock host above.
+    local GW_NET="${GW%.*}"
+    local SLIRP_DNS="${GW_NET}.3"        # QEMU SLIRP DNS proxy is always .3
+    local OWN_CIDR
+    OWN_CIDR="$(ip -4 -o addr show scope global 2>/dev/null \
+                | awk '{print $4; exit}')"
+    local OWN_IP="${OWN_CIDR%/*}"
+    local OWN_NET=""
+    [ -n "${OWN_IP}" ] && OWN_NET="${OWN_IP%.*}.0/24"
+
+    # 1) DNS — point at SLIRP DNS proxy + public fallbacks. chattr +i keeps
+    #    systemd-resolved/NetworkManager from clobbering it.
+    if run_root "printf 'nameserver ${SLIRP_DNS}\\nnameserver 114.114.114.114\\nnameserver 8.8.8.8\\n' > /etc/resolv.conf
+                 chattr +i /etc/resolv.conf 2>/dev/null || true"; then
+        echo "[cua_gym_vm_bridge] DNS set: ${SLIRP_DNS}, 114.114.114.114, 8.8.8.8 (gw=${GW})"
+    else
+        echo "[cua_gym_vm_bridge] WARN: failed to write /etc/resolv.conf"
+    fi
+
+    # 2) Build the proxy bypass list from detected values + the direct-IP
+    #    mock host (and its /16 subnet). THIS is what makes 28.7.184.198:81xx
+    #    reachable from inside the VM — without it the woa proxy 403s it.
+    local ignore_list="['localhost', '127.0.0.0/8', '${GW}'"
+    [ -n "${OWN_IP}" ]    && ignore_list="${ignore_list}, '${OWN_IP}'"
+    [ -n "${OWN_NET}" ]   && ignore_list="${ignore_list}, '${OWN_NET}'"
+    [ -n "${GW_NET}" ]    && ignore_list="${ignore_list}, '${GW_NET}.0/24'"
+    [ -n "${MOCK_HOST}" ] && ignore_list="${ignore_list}, '${MOCK_HOST}'"
+    [ -n "${MOCK_NET}" ]  && ignore_list="${ignore_list}, '${MOCK_NET}'"
+    ignore_list="${ignore_list}]"
+
+    # 3) GNOME proxy ignore-hosts — Chrome bypasses the woa proxy for these.
+    if gsettings set org.gnome.system.proxy ignore-hosts "${ignore_list}" 2>/tmp/gsettings_err; then
+        echo "[cua_gym_vm_bridge] gsettings ignore-hosts = ${ignore_list}"
+    else
+        echo "[cua_gym_vm_bridge] WARN: gsettings ignore-hosts failed: $(cat /tmp/gsettings_err 2>/dev/null)"
+    fi
+
+    # 4) Shell env — comma-separated form for no_proxy/NO_PROXY (requests/curl).
+    local NP="localhost,127.0.0.1,${GW}"
+    [ -n "${OWN_IP}" ]    && NP="${NP},${OWN_IP}"
+    [ -n "${OWN_NET}" ]   && NP="${NP},${OWN_NET}"
+    [ -n "${GW_NET}" ]    && NP="${NP},${GW_NET}.0/24"
+    [ -n "${MOCK_HOST}" ] && NP="${NP},${MOCK_HOST}"
+    [ -n "${MOCK_NET}" ]  && NP="${NP},${MOCK_NET}"
+
+    for var in no_proxy NO_PROXY; do
+        if ! grep -q "export ${var}=${NP}" /home/user/.bashrc 2>/dev/null; then
+            echo "export ${var}=${NP}" >> /home/user/.bashrc
+        fi
+    done
+    run_root "grep -q '^no_proxy=' /etc/environment 2>/dev/null \
+              && sed -i 's|^no_proxy=.*|no_proxy=\"${NP}\"|' /etc/environment \
+              || echo 'no_proxy=\"${NP}\"' >> /etc/environment
+              grep -q '^NO_PROXY=' /etc/environment 2>/dev/null \
+              && sed -i 's|^NO_PROXY=.*|NO_PROXY=\"${NP}\"|' /etc/environment \
+              || echo 'NO_PROXY=\"${NP}\"' >> /etc/environment" >/dev/null 2>&1
+    echo "[cua_gym_vm_bridge] no_proxy=${NP}"
+}
+
+fix_vm_network
+
+echo "[cua_gym_vm_bridge] direct-IP mode: ${MOCK_HOST} (+${MOCK_NET}) bypassed from woa proxy; no socat forwarding needed"

csm_onboarding_checklist_004/initial_setup.py

@@ -0,0 +1,254 @@
+"""
+Initial Setup: ce_onboarding_b106
+Task ID: csm_onboarding_checklist_004
+Domain: mock_websites
+Mocks: google_docs_mock, google_calendar_mock, gmail_mock, slack_mock (single shared sid)
+
+Persona: You are an office worker acting as a project lead. Your tools for this task are: shared documents, the team calendar, the shared email inbox, team communications.
+
+Environment / context (initial state + ground truth this setup establishes):
+Initial environment (seeded via a single shared session id across all mocks): Google Docs with documents 'Onboarding Template'; an empty Google Calendar; Gmail inbox with 1 email(s): "Excited to get started" from sarah.j@techventures.com; Slack workspace with channels #general, #random. GROUND TRUTH — the agent earns partial credit for each checkpoint: onboarding checklist doc created (+0.15); checklist has all five steps (+0.3); cal_new_event (+0.1); training on July 1 (+0.1); gmail_new_sent (+0.05); invite emailed to champion (+0.15); invite names the date (+0.05); slack_new_msg (+0.1). Scoring is absolute over each mock's current_state; new objects exclude the seeded IDs, so an untouched environment scores 0.0 and fully completing every checkpoint scores 1.0.
+"""
+import json
+import os
+import shlex
+import subprocess
+import time
+import uuid
+
+import requests
+
+MOCK_URLS = {
+    "google_docs_mock": "http://28.7.184.198:8142",
+    "google_calendar_mock": "http://28.7.184.198:8141",
+    "gmail_mock": "http://28.7.184.198:8138",
+    "slack_mock": "http://28.7.184.198:8178"
+}
+PRIMARY_URL = 'http://28.7.184.198:8142'
+
+sid = str(uuid.uuid4())
+with open('/tmp/task_web_sid', 'w') as f:
+    f.write(sid)
+print(f'Generated sid: {sid}')
+
+# Full initial state for every mock involved (all required top-level keys present).
+STATES = {   'google_docs_mock': {   'currentUser': {   'id': 'user-1',
+                                               'name': 'Demo User',
+                                               'email': 'demo@example.com',
+                                               'avatar': ''},
+                            'users': [   {   'id': 'user-1',
+                                             'name': 'Demo User',
+                                             'email': 'demo@example.com',
+                                             'avatar': ''},
+                                         {   'id': 'user-2',
+                                             'name': 'Alice Chen',
+                                             'email': 'alice@example.com',
+                                             'avatar': ''},
+                                         {   'id': 'user-3',
+                                             'name': 'Bob Smith',
+                                             'email': 'bob@example.com',
+                                             'avatar': ''}],
+                            'documents': {   'd-tmpl': {   'id': 'd-tmpl',
+                                                           'title': 'Onboarding Template',
+                                                           'content': '<p>tmpl</p>',
+                                                           'ownerId': 'user-1',
+                                                           'starred': False,
+                                                           'created': '2026-05-01T10:00:00Z',
+                                                           'updated': '2026-06-01T10:00:00Z',
+                                                           'sharedWith': [],
+                                                           'linkSharing': {   'enabled': False,
+                                                                              'permission': 'viewer'}}},
+                            'comments': [],
+                            'ui': {   'currentDocId': None,
+                                      'sidebarOpen': False,
+                                      'sidebarTab': 'comments',
+                                      'shareDialogOpen': False,
+                                      'findReplaceOpen': False,
+                                      'viewMode': 'editing',
+                                      'zoom': 100,
+                                      'documentListView': 'grid',
+                                      'searchQuery': ''}},
+    'google_calendar_mock': {   'user': {   'id': 'u1',
+                                            'username': 'Demo User',
+                                            'email': 'demo@example.com',
+                                            'avatar': ''},
+                                'calendars': [   {   'id': 'c1',
+                                                     'name': 'Personal',
+                                                     'color': '#039BE5',
+                                                     'visible': True,
+                                                     'userId': 'u1',
+                                                     'isDefault': True},
+                                                 {   'id': 'c2',
+                                                     'name': 'Work',
+                                                     'color': '#33B679',
+                                                     'visible': True,
+                                                     'userId': 'u1',
+                                                     'isDefault': False}],
+                                'events': [],
+                                'view': 'week',
+                                'currentDate': '2026-06-08T00:00:00.000Z',
+                                'sidebarOpen': True,
+                                'settings': {   'weekStart': 0,
+                                                'defaultDuration': 60,
+                                                'defaultView': 'week',
+                                                'defaultReminder': {'type': 'popup', 'minutes': 10},
+                                                'timeFormat': '12h',
+                                                'showWeekNumbers': False,
+                                                'showDeclinedEvents': False}},
+    'gmail_mock': {   'user': {   'userId': 'u1',
+                                  'username': 'John Smith',
+                                  'email': 'john.smith@company.com',
+                                  'avatar': ''},
+                      'labels': [   {'id': 'l1', 'name': 'Work', 'color': '#ef4444'},
+                                    {'id': 'l2', 'name': 'Personal', 'color': '#3b82f6'},
+                                    {'id': 'l3', 'name': 'Travel', 'color': '#22c55e'},
+                                    {'id': 'l4', 'name': 'Finance', 'color': '#eab308'}],
+                      'emails': [   {   'id': 'email_1',
+                                        'threadId': 'thread_email_1',
+                                        'from': {   'name': 'Sarah Johnson',
+                                                    'email': 'sarah.j@techventures.com',
+                                                    'avatar': ''},
+                                        'to': [   {   'name': 'John Smith',
+                                                      'email': 'john.smith@company.com'}],
+                                        'cc': [],
+                                        'bcc': [],
+                                        'subject': 'Excited to get started',
+                                        'body': "<p>Looking forward to onboarding — when's our "
+                                                'training?</p>',
+                                        'snippet': "When's training?",
+                                        'timestamp': '2026-06-08T08:15:00Z',
+                                        'read': False,
+                                        'starred': False,
+                                        'important': False,
+                                        'labels': [],
+                                        'category': 'primary',
+                                        'folder': 'inbox',
+                                        'attachments': []}],
+                      'drafts': []},
+    'slack_mock': {   'currentUser': {   'userId': 'user_1',
+                                         'fullName': 'John Smith',
+                                         'displayName': 'John',
+                                         'email': 'john.smith@company.com',
+                                         'avatar': 'https://picsum.photos/200/200?random=1',
+                                         'title': '',
+                                         'status': 'online',
+                                         'statusMessage': '',
+                                         'statusEmoji': '',
+                                         'timeZone': 'America/New_York'},
+                      'workspace': {   'workspaceId': 'ws_1',
+                                       'workspaceName': 'Acme Corp',
+                                       'icon': ''},
+                      'users': [   {   'userId': 'user_1',
+                                       'fullName': 'John Smith',
+                                       'displayName': 'John',
+                                       'email': 'john.smith@company.com',
+                                       'avatar': 'https://picsum.photos/200/200?random=1',
+                                       'title': '',
+                                       'status': 'online',
+                                       'statusMessage': '',
+                                       'statusEmoji': '',
+                                       'timeZone': 'America/New_York'},
+                                   {   'userId': 'user_2',
+                                       'fullName': 'Sarah Johnson',
+                                       'displayName': 'Sarah',
+                                       'email': 'sarah.johnson@company.com',
+                                       'avatar': 'https://picsum.photos/200/200?random=2',
+                                       'title': '',
+                                       'status': 'online',
+                                       'statusMessage': '',
+                                       'statusEmoji': '',
+                                       'timeZone': 'America/New_York'},
+                                   {   'userId': 'user_3',
+                                       'fullName': 'Mike Chen',
+                                       'displayName': 'Mike',
+                                       'email': 'mike.chen@company.com',
+                                       'avatar': 'https://picsum.photos/200/200?random=3',
+                                       'title': '',
+                                       'status': 'online',
+                                       'statusMessage': '',
+                                       'statusEmoji': '',
+                                       'timeZone': 'America/New_York'},
+                                   {   'userId': 'user_4',
+                                       'fullName': 'Lisa Park',
+                                       'displayName': 'Lisa',
+                                       'email': 'lisa.park@company.com',
+                                       'avatar': 'https://picsum.photos/200/200?random=4',
+                                       'title': '',
+                                       'status': 'online',
+                                       'statusMessage': '',
+                                       'statusEmoji': '',
+                                       'timeZone': 'America/New_York'}],
+                      'channels': [   {   'channelId': 'general',
+                                          'name': 'general',
+                                          'description': 'general channel',
+                                          'topic': '',
+                                          'isPrivate': False,
+                                          'isStarred': False,
+                                          'members': ['user_1', 'user_2', 'user_3', 'user_4'],
+                                          'createdBy': 'user_1',
+                                          'createdAt': '2026-01-01T10:00:00Z',
+                                          'pinnedMessages': [],
+                                          'unreadCount': 0},
+                                      {   'channelId': 'random',
+                                          'name': 'random',
+                                          'description': 'random channel',
+                                          'topic': '',
+                                          'isPrivate': False,
+                                          'isStarred': False,
+                                          'members': ['user_1', 'user_2', 'user_3', 'user_4'],
+                                          'createdBy': 'user_1',
+                                          'createdAt': '2026-01-01T10:00:00Z',
+                                          'pinnedMessages': [],
+                                          'unreadCount': 0}],
+                      'messages': {   'general': [   {   'messageId': 'g1',
+                                                         'senderId': 'user_2',
+                                                         'content': 'did TechVentures onboarding '
+                                                                    'start?',
+                                                         'timestamp': '2026-06-08T08:00:00Z',
+                                                         'threadId': None,
+                                                         'reactions': [],
+                                                         'attachments': [],
+                                                         'isEdited': False}],
+                                      'random': []},
+                      'threads': {},
+                      'dms': [],
+                      'bookmarkedMessages': [],
+                      'callHistory': [],
+                      'settings': {   'theme': 'light',
+                                      'notifications': 'all',
+                                      'displayDensity': 'comfortable',
+                                      'showAvatars': True,
+                                      'use24Hour': False},
+                      'invitations': [],
+                      'notifications': []}}
+
+
+def launch_gui(command, delay_sec=1.0):
+    env = os.environ.copy()
+    env['DISPLAY'] = ':0'
+    subprocess.Popen(shlex.split(command), stdin=subprocess.DEVNULL,
+                     stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL,
+                     env=env, start_new_session=True)
+    time.sleep(delay_sec)
+
+
+for name, url in MOCK_URLS.items():
+    resp = requests.post(f'{url}/post?sid={sid}',
+                         json={'action': 'set', 'state': STATES[name]}, timeout=30)
+    assert resp.status_code == 200, f'{name} injection failed: {resp.text}'
+    go = requests.get(f'{url}/go?sid={sid}', timeout=10).json()
+    assert go.get('initial_state') is not None, f'{name} initial_state is None'
+    print(f'State injected: {name} sid={sid}')
+
+# Open the PRIMARY app as the main window first, then open every other
+# involved mock as an additional browser tab so the agent can see/reach all
+# relevant apps (e.g. read the email in the Gmail tab AND post in Slack).
+# Without this, secondary-app data is injected but has no on-screen entry
+# point, and the agent wastes its budget hunting for a non-existent client.
+launch_gui(f'google-chrome "{PRIMARY_URL}/?sid={sid}"', delay_sec=2.0)
+print(f'GUI_READY: launched browser at {PRIMARY_URL}/?sid={sid}')
+for _name, _url in MOCK_URLS.items():
+    if _url == PRIMARY_URL:
+        continue
+    launch_gui(f'google-chrome --new-tab "{_url}/?sid={sid}"', delay_sec=1.0)
+    print(f'Opened tab: {_name} at {_url}/?sid={sid}')

csm_onboarding_checklist_004/reward.py

@@ -0,0 +1,154 @@
+"""
+Reward Script: ce_onboarding_b106
+Task ID: csm_onboarding_checklist_004
+Domain: mock_websites
+Mocks: google_docs_mock, google_calendar_mock, gmail_mock, slack_mock
+Scoring (all programmatic; evaluated against each mock's current_state; total = 1.0):
+  - [0.15] onboarding checklist doc created
+  - [0.3] checklist has all five steps
+  - [0.1] cal_new_event
+  - [0.1] training on July 1
+  - [0.05] gmail_new_sent
+  - [0.15] invite emailed to champion
+  - [0.05] invite names the date
+  - [0.1] slack_new_msg
+"""
+import re
+import sys
+
+import requests
+
+MOCK_URLS = {
+    "google_docs_mock": "http://28.7.184.198:8142",
+    "google_calendar_mock": "http://28.7.184.198:8141",
+    "gmail_mock": "http://28.7.184.198:8138",
+    "slack_mock": "http://28.7.184.198:8178"
+}
+SEED = {
+    "slack_mock:general": [
+        "g1"
+    ],
+    "slack_mock:random": []
+}
+
+try:
+    with open('/tmp/task_web_sid') as f:
+        sid = f.read().strip()
+    if not sid:
+        raise ValueError('sid is empty')
+except Exception as e:
+    print(f'CRITICAL: cannot read sid: {e}')
+    print('REWARD: 0.0')
+    sys.exit(0)
+
+STATES = {}
+for _name, _url in MOCK_URLS.items():
+    try:
+        _r = requests.get(f'{_url}/go?sid={sid}', timeout=15)
+        _r.raise_for_status()
+        STATES[_name] = _r.json()
+    except Exception as e:
+        print(f'CRITICAL: cannot fetch {_name}: {e}')
+        print('REWARD: 0.0')
+        sys.exit(0)
+
+
+def cur(mock):
+    return (STATES.get(mock) or {}).get('current_state') or {}
+
+
+def _lc(s):
+    return (s or '').lower()
+
+
+total = 0.0
+
+# c0: a new google doc titled about (onboarding checklist doc created)
+c0_init = set((STATES['google_docs_mock'].get('initial_state') or {}).get('documents', {}).keys())
+c0_docs = cur('google_docs_mock').get('documents') or {}
+c0_new = [d for did, d in c0_docs.items() if did not in c0_init and any(t in _lc(d.get('title','')) for t in ['techventures onboarding', 'onboarding checklist'])]
+if c0_new:
+    print('PASS: onboarding checklist doc created (0.15)')
+    total += 0.15
+else:
+    print('FAIL: onboarding checklist doc created')
+
+# c1: new google doc body (checklist has all five steps)
+c1_init = set((STATES['google_docs_mock'].get('initial_state') or {}).get('documents', {}).keys())
+c1_docs = cur('google_docs_mock').get('documents') or {}
+c1_new = [d for did, d in c1_docs.items() if did not in c1_init and any(t in _lc(d.get('title','')) for t in ['techventures onboarding', 'onboarding checklist'])]
+c1_txt = _lc(' '.join(d.get('content','') for d in c1_new))
+if c1_new and all(t in c1_txt for t in ['kickoff', 'account setup', 'admin training', 'data import', 'go-live']):
+    print('PASS: checklist has all five steps (0.3)')
+    total += 0.3
+else:
+    print('FAIL: checklist has all five steps')
+
+# c2: a new calendar event was created
+c2_init_ids = {e.get('id') for e in ((STATES['google_calendar_mock'].get('initial_state') or {}).get('events') or [])}
+c2_new = [e for e in (cur('google_calendar_mock').get('events') or []) if e.get('id') not in c2_init_ids]
+globals()['cal_new_events'] = c2_new
+if c2_new:
+    print('PASS: new calendar event (0.1)')
+    total += 0.1
+else:
+    print('FAIL: no new calendar event')
+
+# c3: new event date contains '2026-07-01'
+c3_new = globals().get('cal_new_events', [])
+if any('2026-07-01' in (e.get('start','') or '') for e in c3_new):
+    print('PASS: training on July 1 (0.1)')
+    total += 0.1
+else:
+    print('FAIL: training on July 1')
+
+# c4: a new email was sent (folder sent) or saved as draft
+c4_init = (STATES['gmail_mock'].get('initial_state') or {})
+c4_cur = cur('gmail_mock')
+c4_init_ids = {e.get('id') for e in (c4_init.get('emails') or [])}
+c4_sent = [e for e in (c4_cur.get('emails') or []) if e.get('id') not in c4_init_ids and e.get('folder') in ('sent', 'drafts')]
+c4_sent += [d for d in (c4_cur.get('drafts') or []) if d.get('id') not in {x.get('id') for x in (c4_init.get('drafts') or [])}]
+if c4_sent:
+    print('PASS: new email sent/drafted (0.05)')
+    total += 0.05
+else:
+    print('FAIL: no new sent/draft email')
+
+# c5: sent/draft email addressed to (invite emailed to champion)
+c5_init = (STATES['gmail_mock'].get('initial_state') or {})
+c5_cur = cur('gmail_mock')
+c5_init_ids = {e.get('id') for e in (c5_init.get('emails') or [])}
+c5_pool = [e for e in (c5_cur.get('emails') or []) if e.get('id') not in c5_init_ids and e.get('folder') in ('sent','drafts')] + (c5_cur.get('drafts') or [])
+c5_tos = _lc(' '.join(t.get('email','') for e in c5_pool for t in (e.get('to') or [])))
+if any(t in c5_tos for t in ['sarah.j@techventures.com']):
+    print('PASS: invite emailed to champion (0.15)')
+    total += 0.15
+else:
+    print('FAIL: invite emailed to champion')
+
+# c6: sent/draft email text contains (invite names the date)
+c6_init = (STATES['gmail_mock'].get('initial_state') or {})
+c6_cur = cur('gmail_mock')
+c6_init_ids = {e.get('id') for e in (c6_init.get('emails') or [])}
+c6_pool = [e for e in (c6_cur.get('emails') or []) if e.get('id') not in c6_init_ids and e.get('folder') in ('sent','drafts')] + (c6_cur.get('drafts') or [])
+c6_txt = _lc(' '.join((e.get('subject','') + ' ' + e.get('body','') + ' ' + e.get('snippet','')) for e in c6_pool))
+if any(t in c6_txt for t in ['july 1', 'training']):
+    print('PASS: invite names the date (0.05)')
+    total += 0.05
+else:
+    print('FAIL: invite names the date')
+
+# c7: new message by user_1 in #general
+c7_msgs = (cur('slack_mock').get('messages') or {}).get('general', []) or []
+c7_seed = set(SEED.get('slack_mock:general', []))
+c7_new = [m for m in c7_msgs if m.get('messageId') not in c7_seed and m.get('senderId') == 'user_1']
+c7_blob = _lc(' '.join((m.get('content') or '') for m in c7_new))
+if c7_new:
+    print('PASS: new message by user_1 in #general (0.1)')
+    total += 0.1
+else:
+    print('FAIL: no new user_1 message in #general')
+
+final = round(min(total, 1.0), 4)
+print(f'\nScore: {total}/1.0')
+print(f'REWARD: {final}')

csm_onboarding_checklist_004/reward_label.json

@@ -0,0 +1,88 @@
+{
+  "reward_file": "/apdcephfs_sh2/share_300000800/user/jackwkwang/long-horizon-tasks/mini_osworld_xiangwu/cache/bda2b533-4396-584b-ae98-d4198dd74783/reward.py",
+  "api_url": "http://28.7.184.184:8000/v1/chat/completions",
+  "model": "Kimi-K2.6",
+  "max_code_chars": 100000,
+  "created_at": "2026-07-14 10:23:29",
+  "label": {
+    "task_id": "csm_onboarding_checklist_004",
+    "domain": "mock_websites",
+    "summary": "验证新员工入职流程是否完成，包括创建入职清单文档、安排7月1日培训日历事件、发送邀请邮件给负责人以及在Slack发布通知",
+    "is_placeholder": false,
+    "data_sources": [
+      "/tmp/task_web_sid",
+      "google_docs_mock",
+      "google_calendar_mock",
+      "gmail_mock",
+      "slack_mock"
+    ],
+    "scoring_components": [
+      {
+        "name": "Component 0",
+        "weight": 0.15,
+        "description": "创建入职清单文档",
+        "check_logic": "对比 google_docs_mock 的 initial_state 与 current_state 中的 documents，找出新增文档，检查其 title（转小写后）是否包含 'techventures onboarding' 或 'onboarding checklist'",
+        "pass_condition": "存在新增的 Google Doc，且标题包含指定关键词"
+      },
+      {
+        "name": "Component 1",
+        "weight": 0.3,
+        "description": "清单包含全部五个步骤",
+        "check_logic": "在 Component 0 筛选出的新文档中，提取 content 转小写后的文本，检查是否同时包含 'kickoff'、'account setup'、'admin training'、'data import'、'go-live'",
+        "pass_condition": "新文档内容同时包含五个指定关键词"
+      },
+      {
+        "name": "Component 2",
+        "weight": 0.1,
+        "description": "创建新的日历事件",
+        "check_logic": "对比 google_calendar_mock 的 initial_state 与 current_state 中的 events，通过 event id 判断是否有新增事件",
+        "pass_condition": "存在新增的日历事件"
+      },
+      {
+        "name": "Component 3",
+        "weight": 0.1,
+        "description": "培训安排在7月1日",
+        "check_logic": "在 Component 2 识别出的新事件中，检查任意事件的 start 字段是否包含 '2026-07-01'",
+        "pass_condition": "新增事件的 start 字段包含 '2026-07-01'"
+      },
+      {
+        "name": "Component 4",
+        "weight": 0.05,
+        "description": "发送或起草新邮件",
+        "check_logic": "对比 gmail_mock 的 initial_state 与 current_state，找出 id 不在初始集合中且 folder 为 'sent' 或 'drafts' 的邮件，或 current_state drafts 中新增的草稿",
+        "pass_condition": "存在新增的已发送或草稿邮件"
+      },
+      {
+        "name": "Component 5",
+        "weight": 0.15,
+        "description": "向负责人发送邀请邮件",
+        "check_logic": "在 Component 4 的邮件池（sent/drafts 新增邮件 + 所有 current drafts）中，提取所有 to 字段中的 email 地址转小写后，检查是否包含 'sarah.j@techventures.com'",
+        "pass_condition": "邮件收件人包含 'sarah.j@techventures.com'"
+      },
+      {
+        "name": "Component 6",
+        "weight": 0.05,
+        "description": "邮件中注明日期",
+        "check_logic": "在 Component 4 的邮件池中，将 subject、body、snippet 拼接转小写，检查是否包含 'july 1' 或 'training'",
+        "pass_condition": "邮件文本包含 'july 1' 或 'training'"
+      },
+      {
+        "name": "Component 7",
+        "weight": 0.1,
+        "description": "在 Slack #general 频道发送新消息",
+        "check_logic": "检查 slack_mock current_state 中 messages.general 列表，排除 SEED 中预置的 messageId，筛选 senderId 为 'user_1' 的新消息",
+        "pass_condition": "存在 senderId 为 'user_1' 且不在初始种子中的新消息"
+      }
+    ],
+    "total_max_score": 1.0,
+    "score_aggregation": "各组件分数相加，最后通过 min(total, 1.0) 钳制到上限 1.0，并四舍五入到 4 位小数",
+    "failure_modes": [
+      "无法读取 /tmp/task_web_sid 或 sid 为空时，打印 CRITICAL 并返回 0.0 退出",
+      "任一模 mock 服务（google_docs_mock、google_calendar_mock、gmail_mock、slack_mock）状态拉取失败时，打印 CRITICAL 并返回 0.0 退出",
+      "各评分组件检查失败仅导致该组件不得分，不会提前退出"
+    ],
+    "scoring_logic_explanation": "脚本首先读取 /tmp/task_web_sid 获取会话 ID，然后依次向四个 mock 服务（google_docs_mock、google_calendar_mock、gmail_mock、slack_mock）拉取 current_state。接着按顺序执行 8 个评分组件：1）检查是否新建了标题含 onboarding 关键词的 Google Doc；2）检查该文档内容是否包含五个指定步骤关键词；3）检查 Google Calendar 是否有新增事件；4）检查新增事件是否安排在 2026-07-01；5）检查 Gmail 是否有新增的 sent/drafts 邮件；6）检查邮件是否发送给 sarah.j@techventures.com；7）���查邮件主题或正文是否提到 july 1 或 training；8）检查 Slack #general 频道是否有 user_1 发送的新消息。每个组件独立判断，通过则累加对应分值，最终总分经 min(total, 1.0) 钳制并四舍五入后输出。",
+    "confidence": 1.0
+  },
+  "raw_response": "用户要求我分析一个 reward.py 评分脚本，并输出一个严格的 JSON 对象，不要有任何额外文字。\n\n让我仔细阅读代码：\n\n1. 任务 ID: 从 docstring 中提取 `ce_onboarding_b106`，但文件路径中的 UUID 是 `bda2b533-4396-584b-ae98-d4198dd74783`。docstring 中明确写了 `Task ID: bda2b533-4396-584b-ae98-d4198dd74783`，所以 task_id 应该是这个 UUID。等等，docstring 第一行是 `Reward Script: ce_onboarding_b106`，但后面有 `Task ID: bda2b533-4396-584b-ae98-d4198dd74783`。根据要求\"从文件 docstring 中提取的任务 ID\"，应该提取 `bda2b533-4396-584b-ae98-d4198dd74783`。\n\n2. Domain: `mock_websites`\n\n3. 数据源: `google_docs_mock`, `google_calendar_mock`, `gmail_mock`, `slack_mock`，以及 `/tmp/task_web_sid`\n\n4. 评分组件：\n   - c0: 0.15 - onboarding checklist doc created - 检查新创建的 google doc 标题包含 'techventures onboarding' 或 'onboarding checklist'\n   - c1: 0.3 - checklist has all five steps - 检查新 doc 内容包含 'kickoff', 'account setup', 'admin training', 'data import', 'go-live'\n   - c2: 0.1 - cal_new_event - 检查 google calendar 有新事件\n   - c3: 0.1 - training on July 1 - 检查新事件 start 包含 '2026-07-01'\n   - c4: 0.05 - gmail_new_sent - 检查 gmail 有新 sent 或 draft 邮件\n   - c5: 0.15 - invite emailed to champion - 检查邮件 to 包含 'sarah.j@techventures.com'\n   - c6: 0.05 - invite names the date - 检查邮件 subject/body/snippet 包含 'july 1' 或 'training'\n   - c7: 0.1 - slack_new_msg - 检查 slack #general 有新消息，senderId 为 'user_1'\n\n5. 总分计算: 各项相加，最后 `round(min(total, 1.0), 4)`\n\n6. 失败模式：\n   - 读取 sid 失败 -> 0.0 退出\n   - 获取 mock 状态失败 -> 0.0 退出\n   - 各组件检查失败只是不加分，不会提前退出\n\n7. is_placeholder: false，因为脚本有实际检查逻辑。\n\n8. confidence: 1.0，因为逻辑清晰。\n\n让我构建 JSON：\n\n```json\n{\n  \"task_id\": \"bda2b533-4396-584b-ae98-d4198dd74783\",\n  \"domain\": \"mock_websites\",\n  \"summary\": \"验证新员工入职流程是否完成，包括创建入职清单文档、安排培训日历事件、发送邀请邮件给负责人以及在Slack发布通知\",\n  \"is_placeholder\": false,\n  \"data_sources\": [\"/tmp/task_web_sid\", \"google_docs_mock\", \"google_calendar_mock\", \"gmail_mock\", \"slack_mock\"],\n  \"scoring_components\": [\n    {\n      \"name\": \"Component 0\",\n      \"weight\": 0.15,\n      \"description\": \"创建入职清单文档\",\n      \"check_logic\": \"对比 initial_state 与 current_state 中的 documents，找出新增文档，检查其 title（转小写后）是否包含 'techventures onboarding' 或 'onboarding checklist'\",\n      \"pass_condition\": \"存在新增的 Google Doc，且标题包含指定关键词\"\n    },\n    {\n      \"name\": \"Component 1\",\n      \"weight\": 0.3,\n      \"description\": \"清单包含全部五个步骤\",\n      \"check_logic\": \"在 Component 0 筛选出的新文档中，提取 content 转小写后的文本，检查是否同时包含 'kickoff'、'account setup'、'admin training'、'data import'、'go-live'\",\n      \"pass_condition\": \"新文档内容同时包含五个指定关键词\"\n    },\n    {\n      \"name\": \"Component 2\",\n      \"weight\": 0.1,\n      \"description\": \"创建新的日历事件\",\n      \"check_logic\": \"对比 google_calendar_mock 的 initial_state 与 current_state 中的 events，通过 event id 判断是否有新增事件\",\n      \"pass_condition\": \"存在新增的日历事件\"\n    },\n    {\n      \"name\": \"Component 3\",\n      \"weight\": 0.1,\n      \"description\": \"培训安排在7月1日\",\n      \"check_logic\": \"在 Component 2 识别出的新事件中，检查任意事件的 start 字段是否包含 '2026-07-01'\",\n      \"pass_condition\": \"新增事件的 start 字段包含 '2026-07-01'\"\n    },\n    {\n      \"name\": \"Component 4\",\n      \"weight\": 0.05,\n      \"description\": \"发送或起草新邮件\",\n      \"check_logic\": \"对比 gmail_mock 的 initial_state 与 current_state，找出 id 不在初始集合中且 folder 为 'sent' 或 'drafts' 的邮件，或 current_state drafts 中新增的草稿\",\n      \"pass_condition\": \"存在新增的已发送或草稿邮件\"\n    },\n    {\n      \"name\": \"Component 5\",\n      \"weight\": 0.15,\n      \"description\": \"向负责人发送邀请邮件\",\n      \"check_logic\": \"在 Component 4 的邮件池（sent/drafts 新增邮件 + 所有 current drafts）中，提取所有 to 字段中的 email 地址转小写后，检查是否包含 'sarah.j@techventures.com'\",\n      \"pass_condition\": \"邮件收件人包含 'sarah.j@techventures.com'\"\n    },\n    {\n      \"name\": \"Component 6\",\n      \"weight\": 0.05,\n      \"description\": \"邮件中注明日期\",\n      \"check_logic\": \"在 Component 4 的邮件池中，将 subject、body、snippet 拼接转小写，检查是否包含 'july 1' 或 'training'\",\n      \"pass_condition\": \"邮件文本包含 'july 1' 或 'training'\"\n    },\n    {\n      \"name\": \"Component 7\",\n      \"weight\": 0.1,\n      \"description\": \"在 Slack #general 频道发送新消息\",\n      \"check_logic\": \"检查 slack_mock current_state 中 messages.general 列表，排除 SEED 中预置的 messageId，筛选 senderId 为 'user_1' 的新消息\",\n      \"pass_condition\": \"存在 senderId 为 'user_1' 且不在初始种子中的新消息\"\n    }\n  ],\n  \"total_max_score\": 1.0,\n  \"score_aggregation\": \"各组件分数相加，最后通过 min(total, 1.0) 钳制到上限 1.0，并四舍五入到 4 位小数\",\n  \"failure_modes\": [\n    \"无法读取 /tmp/task_web_sid 或 sid 为空时，打印 CRITICAL 并返回 0.0 退出\",\n    \"任一模 mock 服务（google_docs_mock、google_calendar_mock、gmail_mock、slack_mock）状态拉取失败时，打印 CRITICAL 并返回 0.0 退出\",\n    \"各评分组件检查失败仅导致该组件不得分，不会提前退出\"\n  ],\n  \"scoring_logic_explanation\": \"脚本首先读取 /tmp/task_web_sid 获取会话 ID，然后依次向四个 mock 服务（google_docs_mock、google_calendar_mock、gmail_mock、slack_mock）拉取 current_state。接着按顺序执行 8 个评分组件：1）检查是否新建了标题含 onboarding 关键词的 Google Doc；2）检查该文档内容是否包含五个指定步骤关键词；3）检查 Google Calendar 是否有新增事件；4）检查新增事件是否安排在 2026-07-01；5）检查 Gmail 是否有新增的 sent/drafts 邮件；6）检查邮件是否发送给 sarah.j@techventures.com；7）检查邮件主题或正文是否提到 july 1 或 training；8）检查 Slack #general 频道是否有 user_1 发送的新消息。每个组件独立判断，通过则累加对应分值，最终总分经 min(total, 1.0) 钳制并四舍五入后输出。\",\n  \"confidence\": 1.0\n}\n```\n\n检查要求：\n- 只输出 JSON，没有 markdown 代码块\n- 第一个字符是 {，最后一个字符是 }\n- 字符串 value 使用中文，task_id、domain、字段名、代码标识符保留原样\n- weight 填写代码中显式给出的分值\n- is_placeholder: false\n\n确认 task_id：docstring 中明确写了 `Task ID: bda2b533-4396-584b-ae98-d4198dd74783`，所以用这个。\n\n确认 domain：`mock_websites`。\n\n确认 data_sources：包含 `/tmp/task_web_sid` 和四个 mock。\n\n确认 total_max_score：1.0。\n\n确认 score_aggregation：代码中是 `round(min(total, 1.0), 4)`，所以是相加后钳制到上限。\n\n确认 failure_modes：sid 读取失败、mock 拉取失败都会 sys.exit(0) 并打印 REWARD: 0.0。\n\n一切看起来正确。 </think> {\n  \"task_id\": \"bda2b533-4396-584b-ae98-d4198dd74783\",\n  \"domain\": \"mock_websites\",\n  \"summary\": \"验证新员工入职流程是否完成，包括创建入职清单文档、安排7月1日培训日历事件、发送邀请邮件给负责人以及在Slack发布通知\",\n  \"is_placeholder\": false,\n  \"data_sources\": [\"/tmp/task_web_sid\", \"google_docs_mock\", \"google_calendar_mock\", \"gmail_mock\", \"slack_mock\"],\n  \"scoring_components\": [\n    {\n      \"name\": \"Component 0\",\n      \"weight\": 0.15,\n      \"description\": \"创建入职清单文档\",\n      \"check_logic\": \"对比 google_docs_mock 的 initial_state 与 current_state 中的 documents，找出新增文档，检查其 title（转小写后）是否包含 'techventures onboarding' 或 'onboarding checklist'\",\n      \"pass_condition\": \"存在新增的 Google Doc，且标题包含指定关键词\"\n    },\n    {\n      \"name\": \"Component 1\",\n      \"weight\": 0.3,\n      \"description\": \"清单包含全部五个步骤\",\n      \"check_logic\": \"在 Component 0 筛选出的新文档中，提取 content 转小写后的文本，检查是否同时包含 'kickoff'、'account setup'、'admin training'、'data import'、'go-live'\",\n      \"pass_condition\": \"新文档内容同时包含五个指定关键词\"\n    },\n    {\n      \"name\": \"Component 2\",\n      \"weight\": 0.1,\n      \"description\": \"创建新的日历事件\",\n      \"check_logic\": \"对比 google_calendar_mock 的 initial_state 与 current_state 中的 events，通过 event id 判断是否有新增事件\",\n      \"pass_condition\": \"存在新增的日历事件\"\n    },\n    {\n      \"name\": \"Component 3\",\n      \"weight\": 0.1,\n      \"description\": \"培训安排在7月1日\",\n      \"check_logic\": \"在 Component 2 识别出的新事件中，检查任意事件的 start 字段是否包含 '2026-07-01'\",\n      \"pass_condition\": \"新增事件的 start 字段包含 '2026-07-01'\"\n    },\n    {\n      \"name\": \"Component 4\",\n      \"weight\": 0.05,\n      \"description\": \"发送或起草新邮件\",\n      \"check_logic\": \"对比 gmail_mock 的 initial_state 与 current_state，找出 id 不在初始集合中且 folder 为 'sent' 或 'drafts' 的邮件，或 current_state drafts 中新增的草稿\",\n      \"pass_condition\": \"存在新增的已发送或草稿邮件\"\n    },\n    {\n      \"name\": \"Component 5\",\n      \"weight\": 0.15,\n      \"description\": \"向负责人发送邀请邮件\",\n      \"check_logic\": \"在 Component 4 的邮件池（sent/drafts 新增邮件 + 所有 current drafts）中，提取所有 to 字段中的 email 地址转小写后，检查是否包含 'sarah.j@techventures.com'\",\n      \"pass_condition\": \"邮件收件人包含 'sarah.j@techventures.com'\"\n    },\n    {\n      \"name\": \"Component 6\",\n      \"weight\": 0.05,\n      \"description\": \"邮件中注明日期\",\n      \"check_logic\": \"在 Component 4 的邮件池中，将 subject、body、snippet 拼接转小写，检查是否包含 'july 1' 或 'training'\",\n      \"pass_condition\": \"邮件文本包含 'july 1' 或 'training'\"\n    },\n    {\n      \"name\": \"Component 7\",\n      \"weight\": 0.1,\n      \"description\": \"在 Slack #general 频道发送新消息\",\n      \"check_logic\": \"检查 slack_mock current_state 中 messages.general 列表，排除 SEED 中预置的 messageId，筛选 senderId 为 'user_1' 的新消息\",\n      \"pass_condition\": \"存在 senderId 为 'user_1' 且不在初始种子中的新消息\"\n    }\n  ],\n  \"total_max_score\": 1.0,\n  \"score_aggregation\": \"各组件分数相加，最后通过 min(total, 1.0) 钳制到上限 1.0，并四舍五入到 4 位小数\",\n  \"failure_modes\": [\n    \"无法读取 /tmp/task_web_sid 或 sid 为空时，打印 CRITICAL 并返回 0.0 退出\",\n    \"任一模 mock 服务（google_docs_mock、google_calendar_mock、gmail_mock、slack_mock）状态拉取失败时，打印 CRITICAL 并返回 0.0 退出\",\n    \"各评分组件检查失败仅导致该组件不得分，不会提前退出\"\n  ],\n  \"scoring_logic_explanation\": \"脚本首先读取 /tmp/task_web_sid 获取会话 ID，然后依次向四个 mock 服务（google_docs_mock、google_calendar_mock、gmail_mock、slack_mock）拉取 current_state。接着按顺序执行 8 个评分组件：1）检查是否新建了标题含 onboarding 关键词的 Google Doc；2）检查该文档内容是否包含五个指定步骤关键词；3）检查 Google Calendar 是否有新增事件；4）检查新增事件是否安排在 2026-07-01；5）检查 Gmail 是否有新增的 sent/drafts 邮件；6）检查邮件是否发送给 sarah.j@techventures.com；7）检查邮件主题或正文是否提到 july 1 或 training；8）检查 Slack #general 频道是否有 user_1 发送的新消息。每个组件独立判断，通过则累加对应分值，最终总分经 min(total, 1.0) 钳制并四舍五入后输出。\",\n  \"confidence\": 1.0\n}"
+}

csm_qbr_renewal_003/_cua_gym_vm_bridge.sh

@@ -0,0 +1,106 @@
+#!/bin/bash
+# CUA-Gym VM-side network fix (direct-IP mock access).
+#
+# Runs INSIDE the QEMU guest. The CUA-Gym mock services are now reached by
+# their real IDC address (http://28.7.184.198:81xx) instead of the old
+# localhost:9000-9097 socat bridge, so the VM only needs two fixes:
+#
+#   1. Working DNS — the qcow2-baked /etc/resolv.conf points at the Pod's
+#      cluster DNS, which is unreachable from QEMU SLIRP NAT
+#      ("Temporary failure in name resolution").
+#   2. The mock host excluded from the woa http proxy — otherwise the VM's
+#      Chrome (gsettings proxy) and initial_setup.py's `requests` get a 403
+#      from the proxy when hitting 28.7.184.198:81xx. (The old socat path never
+#      hit this because localhost is bypassed by default.)
+#
+# Idempotent. Scoped to the CUA-Gym bridge step; does not touch the generic
+# OSWorld setup pipeline.
+#
+# Argv:
+#   $1  optional sudo password (default "password"; SetupController
+#       substitutes {CLIENT_PASSWORD} when uploaded as a cache file).
+#   $2  unused (kept for backward compat — was socat first port 9000).
+#   $3  unused (kept for backward compat — was socat last  port 9097).
+#   $4  optional mock host IP (default 28.7.184.198).
+
+set +e
+
+PASS="${1:-password}"
+MOCK_HOST="${4:-28.7.184.198}"
+# Derive the /16 IDC subnet so future port/host tweaks on the same segment
+# stay bypassed too (28.7.184.198 -> 28.7.0.0/16).
+MOCK_NET="$(echo "${MOCK_HOST}" | awk -F. 'NF==4{print $1"."$2".0.0/16"}')"
+
+run_root() {
+    echo "$PASS" | sudo -S -p "" bash -c "$1"
+}
+
+# Detect SLIRP gateway from VM's routing table; fall back to 10.0.2.2.
+GW="$(ip route 2>/dev/null | awk '/default/ {print $3; exit}')"
+GW="${GW:-10.0.2.2}"
+
+# ─── CUA-Gym network fix (Taiji-specific) ─────────────────────────────
+fix_vm_network() {
+    # 0) Discover VM network state at runtime — gateway, own IP/CIDR, SLIRP
+    #    DNS forwarder. Nothing hardcoded except the mock host above.
+    local GW_NET="${GW%.*}"
+    local SLIRP_DNS="${GW_NET}.3"        # QEMU SLIRP DNS proxy is always .3
+    local OWN_CIDR
+    OWN_CIDR="$(ip -4 -o addr show scope global 2>/dev/null \
+                | awk '{print $4; exit}')"
+    local OWN_IP="${OWN_CIDR%/*}"
+    local OWN_NET=""
+    [ -n "${OWN_IP}" ] && OWN_NET="${OWN_IP%.*}.0/24"
+
+    # 1) DNS — point at SLIRP DNS proxy + public fallbacks. chattr +i keeps
+    #    systemd-resolved/NetworkManager from clobbering it.
+    if run_root "printf 'nameserver ${SLIRP_DNS}\\nnameserver 114.114.114.114\\nnameserver 8.8.8.8\\n' > /etc/resolv.conf
+                 chattr +i /etc/resolv.conf 2>/dev/null || true"; then
+        echo "[cua_gym_vm_bridge] DNS set: ${SLIRP_DNS}, 114.114.114.114, 8.8.8.8 (gw=${GW})"
+    else
+        echo "[cua_gym_vm_bridge] WARN: failed to write /etc/resolv.conf"
+    fi
+
+    # 2) Build the proxy bypass list from detected values + the direct-IP
+    #    mock host (and its /16 subnet). THIS is what makes 28.7.184.198:81xx
+    #    reachable from inside the VM — without it the woa proxy 403s it.
+    local ignore_list="['localhost', '127.0.0.0/8', '${GW}'"
+    [ -n "${OWN_IP}" ]    && ignore_list="${ignore_list}, '${OWN_IP}'"
+    [ -n "${OWN_NET}" ]   && ignore_list="${ignore_list}, '${OWN_NET}'"
+    [ -n "${GW_NET}" ]    && ignore_list="${ignore_list}, '${GW_NET}.0/24'"
+    [ -n "${MOCK_HOST}" ] && ignore_list="${ignore_list}, '${MOCK_HOST}'"
+    [ -n "${MOCK_NET}" ]  && ignore_list="${ignore_list}, '${MOCK_NET}'"
+    ignore_list="${ignore_list}]"
+
+    # 3) GNOME proxy ignore-hosts — Chrome bypasses the woa proxy for these.
+    if gsettings set org.gnome.system.proxy ignore-hosts "${ignore_list}" 2>/tmp/gsettings_err; then
+        echo "[cua_gym_vm_bridge] gsettings ignore-hosts = ${ignore_list}"
+    else
+        echo "[cua_gym_vm_bridge] WARN: gsettings ignore-hosts failed: $(cat /tmp/gsettings_err 2>/dev/null)"
+    fi
+
+    # 4) Shell env — comma-separated form for no_proxy/NO_PROXY (requests/curl).
+    local NP="localhost,127.0.0.1,${GW}"
+    [ -n "${OWN_IP}" ]    && NP="${NP},${OWN_IP}"
+    [ -n "${OWN_NET}" ]   && NP="${NP},${OWN_NET}"
+    [ -n "${GW_NET}" ]    && NP="${NP},${GW_NET}.0/24"
+    [ -n "${MOCK_HOST}" ] && NP="${NP},${MOCK_HOST}"
+    [ -n "${MOCK_NET}" ]  && NP="${NP},${MOCK_NET}"
+
+    for var in no_proxy NO_PROXY; do
+        if ! grep -q "export ${var}=${NP}" /home/user/.bashrc 2>/dev/null; then
+            echo "export ${var}=${NP}" >> /home/user/.bashrc
+        fi
+    done
+    run_root "grep -q '^no_proxy=' /etc/environment 2>/dev/null \
+              && sed -i 's|^no_proxy=.*|no_proxy=\"${NP}\"|' /etc/environment \
+              || echo 'no_proxy=\"${NP}\"' >> /etc/environment
+              grep -q '^NO_PROXY=' /etc/environment 2>/dev/null \
+              && sed -i 's|^NO_PROXY=.*|NO_PROXY=\"${NP}\"|' /etc/environment \
+              || echo 'NO_PROXY=\"${NP}\"' >> /etc/environment" >/dev/null 2>&1
+    echo "[cua_gym_vm_bridge] no_proxy=${NP}"
+}
+
+fix_vm_network
+
+echo "[cua_gym_vm_bridge] direct-IP mode: ${MOCK_HOST} (+${MOCK_NET}) bypassed from woa proxy; no socat forwarding needed"

csm_qbr_renewal_003/initial_setup.py

@@ -0,0 +1,760 @@
+# --- MOCK_HOST_NO_PROXY_OVERRIDE: bypass woa proxy for locally-deployed mocks ---
+import os as _os
+_np = _os.environ.get("no_proxy", _os.environ.get("NO_PROXY", ""))
+for _h in ("127.0.0.1", "localhost", "28.7.184.198"):
+    if _h not in _np:
+        _np = (_np + "," + _h).lstrip(",")
+_os.environ["no_proxy"] = _np
+_os.environ["NO_PROXY"] = _np
+# --- end override ---
+
+# --- wait_mocks_loaded: wait for mock backends ready AND browser tabs rendered ---
+def wait_mocks_loaded(timeout=60, render_buffer=12):
+    """Wait for mock backends ready AND browser tabs rendered (via CDP).
+
+    Phase 1 — backend readiness: poll each mock's /state?sid= until it
+    returns 200 with 'stored_state' in the body.
+
+    Phase 2 — browser render verification: connect to Chrome's CDP
+    endpoint and wait for every open tab to reach 'networkidle' load
+    state with a non-trivial DOM element count. This confirms the SPA
+    has fetched its /go?sid= data and actually painted, not just that
+    the backend is up.
+
+    CDP port auto-detection: checks globals for _chrome_debug_port /
+    cdp_port / debug_port (set by tasks that use a dynamic port, e.g.
+    se_hard_008's launch_chrome_with_urls); defaults to 1337 for tasks
+    that use launch_gui's --remote-debugging-port injection.
+
+    Falls back to the old render_buffer sleep if Playwright or the CDP
+    endpoint is unavailable.
+    """
+    import urllib.request as _u, time as _t
+    g = globals()
+    _urls = {}
+    for _k, _v in list(g.items()):
+        if isinstance(_v, dict) and _v and all(isinstance(x, str) and x.startswith("http://") for x in _v.values()):
+            _urls.update(_v)
+        elif isinstance(_v, str) and _v.startswith("http://") and (_k.endswith("_URL") or _k.endswith("_url")):
+            _urls[_k] = _v
+    _sid = ""
+    for _k in ("sid", "SID", "task_sid", "session_id"):
+        if isinstance(g.get(_k), str) and g[_k]:
+            _sid = g[_k]; break
+    if not _sid:
+        for _k, _v in g.items():
+            if "sid" in _k.lower() and isinstance(_v, str) and _v:
+                _sid = _v; break
+    if not _urls or not _sid:
+        print("  wait_mocks_loaded: no mocks/sid found (%d urls, sid=%r), skipping" % (len(_urls), bool(_sid)))
+        return
+
+    # ---- Phase 1: backend readiness ----
+    _op = _u.build_opener(_u.ProxyHandler({}))
+    _dl = _t.time() + timeout
+    _todo = list(_urls.items())
+    while _todo and _t.time() < _dl:
+        _rest = []
+        for _n2, _uu in _todo:
+            try:
+                _r = _op.open("%s/state?sid=%s" % (_uu, _sid), timeout=5)
+                if _r.status == 200 and b"stored_state" in _r.read():
+                    print("  %s: backend ready" % _n2); continue
+            except Exception:
+                pass
+            _rest.append((_n2, _uu))
+        _todo = _rest
+        if _todo:
+            _t.sleep(2)
+    _done = len(_urls) - len(_todo)
+    _pend = (" (pending: %s)" % [n for n, _ in _todo]) if _todo else ""
+    print("  mocks backend ready: %d/%d%s" % (_done, len(_urls), _pend))
+
+    # ---- Phase 2: browser render verification via CDP ----
+    # Auto-detect CDP port from globals (set by launch_chrome_with_urls
+    # in tasks that use a dynamic port); default to 1337.
+    _cdp_port = 1337
+    for _pk in ("_chrome_debug_port", "chrome_debug_port", "cdp_port", "debug_port"):
+        _pv = g.get(_pk)
+        if isinstance(_pv, int) and 0 < _pv < 65536:
+            _cdp_port = _pv; break
+    _cdp_ok = False
+    try:
+        from playwright.sync_api import sync_playwright
+        _cdp_dl = _t.time() + max(10.0, _dl - _t.time())  # leftover budget, floor 10s
+        with sync_playwright() as _p:
+            _br = None
+            # Retry-connect: Chrome may still be starting up.
+            while _t.time() < _cdp_dl:
+                try:
+                    _br = _p.chromium.connect_over_cdp("http://localhost:%d" % _cdp_port)
+                    break
+                except Exception:
+                    _t.sleep(0.5)
+            if _br is None:
+                print("  wait_mocks_loaded: CDP connect to localhost:%d failed; "
+                      "is Chrome launched with --remote-debugging-port?" % _cdp_port)
+            else:
+                # Wait for ALL expected tabs to appear, then verify each renders.
+                _n_expected = len(_urls)
+                _pages = []
+                _pg_dl = _t.time() + 15
+                _last_count = -1
+                _stable_since = None
+                while _t.time() < _pg_dl:
+                    _pages = []
+                    for _c in _br.contexts:
+                        _pages.extend(_c.pages)
+                    _cnt = len(_pages)
+                    if _n_expected > 0 and _cnt >= _n_expected:
+                        break  # all expected tabs are open
+                    if _cnt > 0 and _cnt == _last_count:
+                        if _stable_since is None:
+                            _stable_since = _t.time()
+                        elif _t.time() - _stable_since >= 1.5:
+                            break  # tab count stable -> no more tabs opening
+                    else:
+                        _stable_since = None
+                    _last_count = _cnt
+                    _t.sleep(0.5)
+                # Re-fetch final snapshot so late-opening tabs are included.
+                _pages = []
+                for _c in _br.contexts:
+                    _pages.extend(_c.pages)
+                if not _pages:
+                    print("  wait_mocks_loaded: CDP connected but no open tabs found")
+                else:
+                    print("  wait_mocks_loaded: %d tab(s) open (expected %d); "
+                          "verifying render" % (len(_pages), _n_expected))
+                    _ok_cnt = 0
+                    for _pg in _pages:
+                        _ttl = ""
+                        try:
+                            _ttl = _pg.title()
+                        except Exception:
+                            pass
+                        try:
+                            _pg.wait_for_load_state(
+                                "networkidle",
+                                timeout=int(max(5000, (_cdp_dl - _t.time()) * 1000)))
+                            _n = _pg.evaluate("document.querySelectorAll('*').length")
+                            if _n < 20:
+                                print("  wait_mocks_loaded: WARNING tab %r has only %d "
+                                      "elements (render stall?)" % (_ttl, _n))
+                            else:
+                                print("  wait_mocks_loaded: tab %r rendered (%d elements)"
+                                      % (_ttl, _n))
+                                _ok_cnt += 1
+                        except Exception as _e:
+                            print("  wait_mocks_loaded: tab %r render wait failed: %r"
+                                  % (_ttl, _e))
+                    # Require ALL open tabs to render before skipping the
+                    # render_buffer fallback, so no tab is left half-loaded.
+                    if _ok_cnt > 0 and _ok_cnt == len(_pages):
+                        _cdp_ok = True
+                    else:
+                        print("  wait_mocks_loaded: %d/%d tabs rendered, will fall back"
+                              % (_ok_cnt, len(_pages)))
+    except ImportError:
+        print("  wait_mocks_loaded: playwright not installed; cannot verify browser render")
+    except Exception as _e:
+        print("  wait_mocks_loaded: CDP phase error (%r)" % _e)
+
+    if not _cdp_ok:
+        print("  wait_mocks_loaded: falling back to render_buffer=%ds sleep" % render_buffer)
+        _t.sleep(render_buffer)
+    else:
+        print("  wait_mocks_loaded: browser render verified, skipping render_buffer")
+# --- end wait_mocks_loaded ---
+
+# --- _open_remaining_mock_tabs: open all auto-detected mock URLs in new Chrome tabs ---
+def _open_remaining_mock_tabs(primary_url=""):
+    """Open all auto-detected mock URLs in Chrome --new-tab, except primary_url."""
+    g = globals()
+    _urls = set()
+    for _k, _v in list(g.items()):
+        if "PROBE" in _k.upper() or _k.startswith("_"):
+            continue
+        if isinstance(_v, str) and _v.startswith("http://") and (_k.endswith("_URL") or _k.endswith("_url")):
+            _urls.add(_v)
+        elif isinstance(_v, dict) and _v and all(isinstance(x, str) and x.startswith("http://") for x in _v.values()):
+            _urls.update(_v.values())
+    _sid = ""
+    for _k in ("sid", "SID", "task_sid", "session_id"):
+        if isinstance(g.get(_k), str) and g[_k]:
+            _sid = g[_k]; break
+    for _u in sorted(_urls):
+        if _u != primary_url:
+            launch_gui(f'google-chrome --new-tab "{_u}/?sid={_sid}"', delay_sec=1.0)
+# --- end _open_remaining_mock_tabs ---
+
+
+"""
+Initial Setup: CSM renewal QBR prep for Aspen Dynamics
+Task ID: csm_qbr_renewal_003
+Domain: mock_websites
+Mocks: notion_mock, salesforce_mock, docusign_mock, gmail_mock
+
+INITIAL STATE (pre-task):
+  - Notion: Jordan Rivera workspace with a 'Customer Health Tracker' database that
+    holds an 'Aspen Dynamics' row (Health Yellow, MAU Trend Flat, ARR 90000, usage
+    notes). NO QBR agenda page exists yet.
+  - Salesforce: CSM Jordan Rivera (user-1, Customer Success). account-1 'Aspen
+    Dynamics' (Customer), contact-1 'Daniel Brooks' primary contact, opp-1
+    'Aspen Dynamics - FY27 Renewal' at stage 'Qualification'. Default activities
+    exist but NO QBR event for Aspen; the runtime "this Friday" 14:00-14:45
+    slot is free.
+  - DocuSign: CSM Jordan Rivera (Northwind Cloud). env_1 DRAFT 'Aspen Dynamics -
+    FY27 Renewal Agreement' with signer Daniel Brooks (status created). Distractor
+    env_2 DRAFT for a different customer.
+  - Gmail: Jordan Rivera (u1). Inbox has unrelated mail; drafts empty; NO sent
+    email to daniel.brooks@aspendynamics.com.
+"""
+import os
+import shlex
+import subprocess
+import time
+import uuid
+from datetime import date, timedelta
+
+import requests
+
+# Current mock deployments are reachable directly from the VM.
+print('Network mode: direct')
+
+
+def http_post(url, payload):
+    return requests.post(url, json=payload, timeout=30)
+
+
+def http_get(url):
+    return requests.get(url, timeout=30)
+
+
+# --- Config ---
+NOTION_URL = 'http://28.7.184.198:8166'
+SALESFORCE_URL = 'http://28.7.184.198:8175'
+DOCUSIGN_URL = 'http://28.7.184.198:8130'
+GMAIL_URL = 'http://28.7.184.198:8138'
+
+
+def this_friday(today):
+    days_until_friday = (4 - today.weekday()) % 7
+    return today + timedelta(days=days_until_friday)
+
+
+def fmt_month_day_year(d):
+    return f'{d.strftime("%b")} {d.day}, {d.year}'
+
+
+TODAY = date.today()
+TARGET_QBR_DATE = this_friday(TODAY)
+TARGET_QBR_START = f'{TARGET_QBR_DATE.isoformat()}T14:00'
+TARGET_QBR_END = f'{TARGET_QBR_DATE.isoformat()}T14:45'
+RENEWAL_DATE = TODAY + timedelta(days=45)
+
+sid = str(uuid.uuid4())
+with open('/tmp/task_web_sid', 'w') as f:
+    f.write(sid)
+with open('/tmp/task_qbr_date', 'w') as f:
+    f.write(TARGET_QBR_DATE.isoformat())
+print(f'Generated sid={sid}')
+print(f'Target QBR date={TARGET_QBR_DATE.isoformat()}')
+
+
+# ---------------------------------------------------------------------------
+# NOTION STATE — Customer Health Tracker database with Aspen Dynamics row.
+#                NO QBR agenda page yet.
+# ---------------------------------------------------------------------------
+def build_notion_state():
+    user = {
+        "id": "user-1",
+        "name": "Jordan Rivera",
+        "email": "jordan.rivera@northwindcloud.com",
+        "avatar": "",
+    }
+    workspace = {
+        "id": "ws-1",
+        "name": "Northwind Cloud — Customer Success",
+        "icon": "",
+        "members": ["user-1"],
+    }
+
+    # Database properties (column definitions)
+    db_props = [
+        {"id": "prop-name", "name": "Account", "type": "text"},
+        {"id": "prop-health", "name": "Health", "type": "select",
+         "options": ["Green", "Yellow", "Red"]},
+        {"id": "prop-mau", "name": "MAU Trend", "type": "select",
+         "options": ["Up", "Flat", "Down"]},
+        {"id": "prop-arr", "name": "ARR", "type": "text"},
+        {"id": "prop-notes", "name": "Usage Notes", "type": "text"},
+    ]
+
+    # Database rows (each is its own page referenced from items[])
+    row_aspen = {
+        "id": "row-aspen",
+        "title": "Aspen Dynamics",
+        "icon": "",
+        "cover": None,
+        "parentId": "db-health",
+        "blockIds": [],
+        "favorite": False,
+        "createdDate": "2026-04-02T00:00:00.000Z",
+        "lastEditedDate": "2026-06-18T00:00:00.000Z",
+        "properties": {
+            "prop-name": "Aspen Dynamics",
+            "prop-health": "Yellow",
+            "prop-mau": "Flat",
+            "prop-arr": "90000",
+            "prop-notes": "Adoption of reporting module up; admin seats underused; "
+                          "one open escalation last quarter (since resolved). Renewal "
+                          f"FY27 due {fmt_month_day_year(RENEWAL_DATE)}; champion is Daniel Brooks "
+                          "(VP Customer Operations).",
+        },
+    }
+    row_orion = {
+        "id": "row-orion",
+        "title": "Orion Retail Group",
+        "icon": "",
+        "cover": None,
+        "parentId": "db-health",
+        "blockIds": [],
+        "favorite": False,
+        "createdDate": "2026-03-11T00:00:00.000Z",
+        "lastEditedDate": "2026-06-10T00:00:00.000Z",
+        "properties": {
+            "prop-name": "Orion Retail Group",
+            "prop-health": "Green",
+            "prop-mau": "Up",
+            "prop-arr": "145000",
+            "prop-notes": "Strong expansion last quarter; two new teams onboarded.",
+        },
+    }
+    row_meridian = {
+        "id": "row-meridian",
+        "title": "Meridian Logistics",
+        "icon": "",
+        "cover": None,
+        "parentId": "db-health",
+        "blockIds": [],
+        "favorite": False,
+        "createdDate": "2026-02-20T00:00:00.000Z",
+        "lastEditedDate": "2026-05-29T00:00:00.000Z",
+        "properties": {
+            "prop-name": "Meridian Logistics",
+            "prop-health": "Red",
+            "prop-mau": "Down",
+            "prop-arr": "60000",
+            "prop-notes": "Usage declining; sponsor left in April. At-risk renewal.",
+        },
+    }
+
+    db_health = {
+        "id": "db-health",
+        "title": "Customer Health Tracker",
+        "icon": "",
+        "cover": None,
+        "parentId": None,
+        "type": "database",
+        "viewType": "table",
+        "properties": db_props,
+        "views": [
+            {"id": "view-1", "name": "All Accounts", "type": "table",
+             "filters": [], "sorts": [], "groupBy": None,
+             "visibleProperties": ["prop-name", "prop-health", "prop-mau",
+                                   "prop-arr", "prop-notes"]},
+        ],
+        "items": ["row-aspen", "row-orion", "row-meridian"],
+        "blockIds": [],
+        "favorite": True,
+        "createdDate": "2026-02-01T00:00:00.000Z",
+    }
+
+    # A plain notes page (distractor / realism) — not a QBR agenda.
+    block_team_1 = {
+        "id": "blk-team-1", "type": "heading-2",
+        "content": "Standing Items",
+        "properties": {}, "createdDate": "2026-06-01T00:00:00.000Z",
+        "lastEditedDate": "2026-06-01T00:00:00.000Z",
+    }
+    block_team_2 = {
+        "id": "blk-team-2", "type": "bullet-list",
+        "content": "Review at-risk accounts weekly; keep ARR figures in sync with Salesforce.",
+        "properties": {}, "createdDate": "2026-06-01T00:00:00.000Z",
+        "lastEditedDate": "2026-06-01T00:00:00.000Z",
+    }
+    page_team = {
+        "id": "page-team-notes",
+        "title": "CS Team Notes",
+        "icon": "",
+        "cover": None,
+        "parentId": None,
+        "blockIds": ["blk-team-1", "blk-team-2"],
+        "favorite": False,
+        "createdDate": "2026-06-01T00:00:00.000Z",
+        "lastEditedDate": "2026-06-01T00:00:00.000Z",
+        "properties": {},
+    }
+
+    return {
+        "user": user,
+        "workspace": workspace,
+        "pages": {
+            "db-health": db_health,
+            "row-aspen": row_aspen,
+            "row-orion": row_orion,
+            "row-meridian": row_meridian,
+            "page-team-notes": page_team,
+        },
+        "blocks": {
+            "blk-team-1": block_team_1,
+            "blk-team-2": block_team_2,
+        },
+        "trash": [],
+        "comments": {},
+        "settings": {"appearance": "light", "startWeekMonday": False, "fontSize": "default"},
+        "notifications": [],
+        "pageOrder": ["db-health", "page-team-notes"],
+        "focusBlockId": None,
+    }
+
+
+# ---------------------------------------------------------------------------
+# SALESFORCE STATE — Aspen Dynamics renewal opportunity at Qualification.
+# ---------------------------------------------------------------------------
+def build_salesforce_state():
+    jordan = {
+        "userId": "user-1",
+        "firstName": "Jordan",
+        "lastName": "Rivera",
+        "email": "jordan.rivera@northwindcloud.com",
+        "phone": "(555) 410-2200",
+        "title": "Customer Success Manager",
+        "department": "Customer Success",
+        "role": "CSM",
+        "avatar": "https://i.pravatar.cc/150?u=user-1",
+        "timezone": "America/New_York",
+        "locale": "en-US",
+        "theme": "lightning",
+    }
+    users = [
+        jordan,
+        {"userId": "user-2", "firstName": "Priya", "lastName": "Nair",
+         "email": "priya.nair@northwindcloud.com", "phone": "(555) 410-2201",
+         "title": "Director, Customer Success", "department": "Customer Success",
+         "role": "Director", "avatar": "https://i.pravatar.cc/150?u=user-2",
+         "timezone": "America/New_York", "locale": "en-US", "theme": "lightning"},
+        {"userId": "user-3", "firstName": "Marcus", "lastName": "Webb",
+         "email": "marcus.webb@northwindcloud.com", "phone": "(555) 410-2202",
+         "title": "Account Executive", "department": "Sales",
+         "role": "AE", "avatar": "https://i.pravatar.cc/150?u=user-3",
+         "timezone": "America/Chicago", "locale": "en-US", "theme": "lightning"},
+    ]
+
+    accounts = [
+        {"accountId": "account-1", "name": "Aspen Dynamics", "type": "Customer",
+         "industry": "Technology", "revenue": 38000000, "employees": 260, "ownerId": "user-1",
+         "phone": "(720) 555-0133", "website": "https://www.aspendynamics.example.com",
+         "billingStreet": "1450 Market St", "billingCity": "Denver", "billingState": "CO",
+         "billingZip": "80202", "billingCountry": "USA",
+         "shippingStreet": "1450 Market St", "shippingCity": "Denver", "shippingState": "CO",
+         "shippingZip": "80202", "shippingCountry": "USA",
+         "createdDate": "2024-03-15T00:00:00.000Z", "modifiedDate": "2026-06-10T00:00:00.000Z"},
+        {"accountId": "account-2", "name": "Orion Retail Group", "type": "Customer",
+         "industry": "Retail", "revenue": 82000000, "employees": 610, "ownerId": "user-1",
+         "phone": "(415) 555-0188", "website": "https://www.orionretail.example.com",
+         "billingStreet": "55 Mission St", "billingCity": "San Francisco", "billingState": "CA",
+         "billingZip": "94105", "billingCountry": "USA",
+         "shippingStreet": "55 Mission St", "shippingCity": "San Francisco", "shippingState": "CA",
+         "shippingZip": "94105", "shippingCountry": "USA",
+         "createdDate": "2023-09-01T00:00:00.000Z", "modifiedDate": "2026-05-22T00:00:00.000Z"},
+        {"accountId": "account-3", "name": "Meridian Logistics", "type": "Customer",
+         "industry": "Logistics", "revenue": 41000000, "employees": 300, "ownerId": "user-1",
+         "phone": "(312) 555-0144", "website": "https://www.meridianlogistics.example.com",
+         "billingStreet": "200 Wacker Dr", "billingCity": "Chicago", "billingState": "IL",
+         "billingZip": "60606", "billingCountry": "USA",
+         "shippingStreet": "200 Wacker Dr", "shippingCity": "Chicago", "shippingState": "IL",
+         "shippingZip": "60606", "shippingCountry": "USA",
+         "createdDate": "2024-01-12T00:00:00.000Z", "modifiedDate": "2026-04-30T00:00:00.000Z"},
+    ]
+
+    contacts = [
+        {"contactId": "contact-1", "accountId": "account-1", "firstName": "Daniel",
+         "lastName": "Brooks", "title": "VP Customer Operations", "department": "Operations",
+         "email": "daniel.brooks@aspendynamics.com", "phone": "(720) 555-0134",
+         "ownerId": "user-1"},
+        {"contactId": "contact-2", "accountId": "account-2", "firstName": "Lena",
+         "lastName": "Ortiz", "title": "Director of IT", "department": "IT",
+         "email": "lena.ortiz@orionretail.example.com", "phone": "(415) 555-0189",
+         "ownerId": "user-1"},
+        {"contactId": "contact-3", "accountId": "account-3", "firstName": "Greg",
+         "lastName": "Halloran", "title": "Operations Manager", "department": "Operations",
+         "email": "greg.halloran@meridianlogistics.example.com", "phone": "(312) 555-0145",
+         "ownerId": "user-1"},
+    ]
+
+    opportunities = [
+        {"opportunityId": "opp-1", "name": "Aspen Dynamics — FY27 Renewal",
+         "accountId": "account-1", "contactId": "contact-1", "amount": 90000,
+         "closeDate": RENEWAL_DATE.isoformat(), "stage": "Qualification", "probability": 30,
+         "ownerId": "user-1", "type": "Renewal",
+         "createdDate": "2026-05-15T00:00:00.000Z", "modifiedDate": "2026-06-12T00:00:00.000Z"},
+        {"opportunityId": "opp-2", "name": "Orion Retail Group — Expansion",
+         "accountId": "account-2", "contactId": "contact-2", "amount": 60000,
+         "closeDate": "2026-09-30", "stage": "Proposal", "probability": 60,
+         "ownerId": "user-1", "type": "Upsell",
+         "createdDate": "2026-04-20T00:00:00.000Z", "modifiedDate": "2026-06-05T00:00:00.000Z"},
+        {"opportunityId": "opp-3", "name": "Meridian Logistics — FY27 Renewal",
+         "accountId": "account-3", "contactId": "contact-3", "amount": 60000,
+         "closeDate": "2026-07-31", "stage": "Negotiation", "probability": 50,
+         "ownerId": "user-1", "type": "Renewal",
+         "createdDate": "2026-03-10T00:00:00.000Z", "modifiedDate": "2026-06-08T00:00:00.000Z"},
+    ]
+
+    # Default activities for the CSM: none is an Aspen QBR event, and none
+    # occupies the runtime "this Friday" 14:00-14:45 slot.
+    activities = [
+        {"activityId": "activity-1", "type": "task",
+         "subject": "Follow up on Orion expansion proposal",
+         "status": "Open", "priority": "Normal", "dueDate": "2026-06-25",
+         "relatedToType": "opportunity", "relatedToId": "opp-2", "assignedToId": "user-1"},
+        {"activityId": "activity-2", "type": "task",
+         "subject": "Prepare Meridian recovery plan",
+         "status": "Open", "priority": "High", "dueDate": "2026-06-29",
+         "relatedToType": "account", "relatedToId": "account-3", "assignedToId": "user-1"},
+        {"activityId": "activity-3", "type": "event",
+         "subject": "Weekly CS team sync",
+         "status": "Open", "priority": "Normal",
+         "startDateTime": f"{TARGET_QBR_DATE.isoformat()}T10:00",
+         "endDateTime": f"{TARGET_QBR_DATE.isoformat()}T10:30",
+         "relatedToType": "account", "relatedToId": "account-2", "assignedToId": "user-1"},
+        {"activityId": "activity-4", "type": "event",
+         "subject": "Orion quarterly check-in",
+         "status": "Open", "priority": "Normal",
+         "startDateTime": f"{(TARGET_QBR_DATE - timedelta(days=1)).isoformat()}T16:00",
+         "endDateTime": f"{(TARGET_QBR_DATE - timedelta(days=1)).isoformat()}T16:30",
+         "relatedToType": "account", "relatedToId": "account-2", "assignedToId": "user-1"},
+    ]
+
+    return {
+        "user": jordan,
+        "users": users,
+        "leads": [],
+        "accounts": accounts,
+        "contacts": contacts,
+        "opportunities": opportunities,
+        "cases": [],
+        "activities": activities,
+        "chatterPosts": [],
+        "files": [],
+        "following": ["user-2", "user-3"],
+        "recentlyViewed": [],
+        "dismissedNotifications": [],
+    }
+
+
+# ---------------------------------------------------------------------------
+# DOCUSIGN STATE — draft renewal agreement (env_1) + distractor draft (env_2).
+# ---------------------------------------------------------------------------
+def build_docusign_state():
+    user = {
+        "id": "user_1",
+        "name": "Jordan Rivera",
+        "email": "jordan.rivera@northwindcloud.com",
+        "title": "Customer Success Manager",
+        "company": "Northwind Cloud",
+        "avatar": "",
+        "memberSince": "2024-01-10",
+        "settings": {"defaultReminderDays": 3, "defaultExpirationDays": 120,
+                     "timezone": "America/New_York"},
+    }
+
+    env_1 = {
+        "id": "env_1",
+        "subject": "Aspen Dynamics — FY27 Renewal Agreement",
+        "message": "Please review and sign the FY27 renewal agreement for Aspen Dynamics.",
+        "status": "draft",
+        "createdAt": "2026-06-20T15:30:00Z",
+        "sentAt": None, "completedAt": None, "voidedAt": None, "declinedAt": None,
+        "lastActivityAt": "2026-06-20T15:30:00Z", "expiresAt": None,
+        "senderId": "user_1", "folderId": None, "templateId": None,
+        "reminderEnabled": False, "reminderDays": 3, "reminderFrequency": 2,
+        "documents": [
+            {"id": "doc_1_1", "name": "Aspen Dynamics FY27 Renewal Agreement.pdf",
+             "pageCount": 6, "order": 1,
+             "fileUrl": "https://picsum.photos/seed/aspenfy27/800/1100", "fileType": "pdf"},
+        ],
+        "recipients": [
+            {"id": "rec_1_1", "name": "Daniel Brooks", "email": "daniel.brooks@aspendynamics.com",
+             "role": "signer", "routingOrder": 1, "status": "created",
+             "signedAt": None, "viewedAt": None, "deliveredAt": None,
+             "declinedAt": None, "declineReason": None},
+        ],
+        "fields": [
+            {"id": "fld_1_1", "type": "signature", "recipientId": "rec_1_1",
+             "documentId": "doc_1_1", "pageNumber": 6, "x": 120, "y": 640,
+             "width": 200, "height": 44, "value": None, "required": True,
+             "label": "Signature", "readOnly": False, "fontSize": 12, "fontColor": "#000000"},
+            {"id": "fld_1_2", "type": "dateSigned", "recipientId": "rec_1_1",
+             "documentId": "doc_1_1", "pageNumber": 6, "x": 360, "y": 640,
+             "width": 120, "height": 24, "value": None, "required": True,
+             "label": "Date", "readOnly": True, "fontSize": 12, "fontColor": "#000000"},
+        ],
+        "history": [
+            {"id": "evt_1_1", "timestamp": "2026-06-20T15:30:00Z", "action": "created",
+             "actorName": "Jordan Rivera", "actorEmail": "jordan.rivera@northwindcloud.com",
+             "details": "Envelope created"},
+        ],
+    }
+
+    # Distractor draft for a DIFFERENT customer — must NOT be sent.
+    env_2 = {
+        "id": "env_2",
+        "subject": "Meridian Logistics — FY27 Renewal Agreement",
+        "message": "Draft renewal agreement for Meridian Logistics (not ready to send).",
+        "status": "draft",
+        "createdAt": "2026-06-18T11:05:00Z",
+        "sentAt": None, "completedAt": None, "voidedAt": None, "declinedAt": None,
+        "lastActivityAt": "2026-06-18T11:05:00Z", "expiresAt": None,
+        "senderId": "user_1", "folderId": None, "templateId": None,
+        "reminderEnabled": False, "reminderDays": 3, "reminderFrequency": 2,
+        "documents": [
+            {"id": "doc_2_1", "name": "Meridian Logistics FY27 Renewal Agreement.pdf",
+             "pageCount": 5, "order": 1,
+             "fileUrl": "https://picsum.photos/seed/meridianfy27/800/1100", "fileType": "pdf"},
+        ],
+        "recipients": [
+            {"id": "rec_2_1", "name": "Greg Halloran",
+             "email": "greg.halloran@meridianlogistics.example.com",
+             "role": "signer", "routingOrder": 1, "status": "created",
+             "signedAt": None, "viewedAt": None, "deliveredAt": None,
+             "declinedAt": None, "declineReason": None},
+        ],
+        "fields": [],
+        "history": [
+            {"id": "evt_2_1", "timestamp": "2026-06-18T11:05:00Z", "action": "created",
+             "actorName": "Jordan Rivera", "actorEmail": "jordan.rivera@northwindcloud.com",
+             "details": "Envelope created"},
+        ],
+    }
+
+    return {
+        "user": user,
+        "envelopes": [env_1, env_2],
+        "templates": [],
+        "folders": [],
+        "contacts": [
+            {"id": "c_1", "name": "Daniel Brooks", "email": "daniel.brooks@aspendynamics.com",
+             "company": "Aspen Dynamics", "title": "VP Customer Operations"},
+            {"id": "c_2", "name": "Greg Halloran",
+             "email": "greg.halloran@meridianlogistics.example.com",
+             "company": "Meridian Logistics", "title": "Operations Manager"},
+        ],
+        "auditLog": [
+            {"id": "audit_1", "timestamp": "2026-06-20T15:30:00Z", "action": "CREATE_ENVELOPE",
+             "details": "Created envelope 'Aspen Dynamics — FY27 Renewal Agreement'",
+             "envelopeId": "env_1"},
+            {"id": "audit_2", "timestamp": "2026-06-18T11:05:00Z", "action": "CREATE_ENVELOPE",
+             "details": "Created envelope 'Meridian Logistics — FY27 Renewal Agreement'",
+             "envelopeId": "env_2"},
+        ],
+    }
+
+
+# ---------------------------------------------------------------------------
+# GMAIL STATE — Jordan Rivera; inbox has unrelated mail; no sent email to Daniel.
+# ---------------------------------------------------------------------------
+def build_gmail_state():
+    user = {
+        "userId": "u1",
+        "username": "Jordan Rivera",
+        "email": "jordan.rivera@northwindcloud.com",
+        "avatar": "https://i.pravatar.cc/150?u=jordan-rivera",
+    }
+    emails = [
+        {"id": "email_1", "threadId": "thread_1",
+         "from": {"name": "Priya Nair", "email": "priya.nair@northwindcloud.com"},
+         "to": [{"name": "Jordan Rivera", "email": "jordan.rivera@northwindcloud.com"}],
+         "cc": [], "bcc": [],
+         "subject": "Renewals pipeline review — Friday",
+         "body": "<p>Hi Jordan,</p><p>Let's go through the Q3 renewals on our Friday sync. "
+                 "Please have the at-risk accounts ready.</p><p>Thanks,<br/>Priya</p>",
+         "timestamp": "2026-06-22T13:15:00Z",
+         "read": True, "starred": False, "important": True,
+         "labels": ["l1"], "category": "primary", "folder": "inbox", "attachments": []},
+        {"id": "email_2", "threadId": "thread_2",
+         "from": {"name": "Northwind Cloud Billing", "email": "billing@northwindcloud.com"},
+         "to": [{"name": "Jordan Rivera", "email": "jordan.rivera@northwindcloud.com"}],
+         "cc": [], "bcc": [],
+         "subject": "June statement available",
+         "body": "<p>Your June account statement is now available in the billing portal.</p>",
+         "timestamp": "2026-06-20T09:00:00Z",
+         "read": False, "starred": False, "important": False,
+         "labels": ["l4"], "category": "primary", "folder": "inbox", "attachments": []},
+        {"id": "email_3", "threadId": "thread_3",
+         "from": {"name": "Lena Ortiz", "email": "lena.ortiz@orionretail.example.com"},
+         "to": [{"name": "Jordan Rivera", "email": "jordan.rivera@northwindcloud.com"}],
+         "cc": [], "bcc": [],
+         "subject": "Re: Orion expansion timeline",
+         "body": "<p>Hi Jordan, thanks for the proposal — reviewing internally and will "
+                 "circle back next week.</p><p>Best,<br/>Lena</p>",
+         "timestamp": "2026-06-19T17:40:00Z",
+         "read": True, "starred": False, "important": False,
+         "labels": ["l1"], "category": "primary", "folder": "inbox", "attachments": []},
+    ]
+    labels = [
+        {"id": "l1", "name": "Work", "color": "#ef4444"},
+        {"id": "l2", "name": "Personal", "color": "#3b82f6"},
+        {"id": "l3", "name": "Travel", "color": "#22c55e"},
+        {"id": "l4", "name": "Finance", "color": "#eab308"},
+    ]
+    return {"user": user, "emails": emails, "labels": labels, "drafts": []}
+
+
+def inject(url, state, label):
+    resp = http_post(f'{url}/post?sid={sid}', {'action': 'set', 'state': state})
+    assert resp.status_code == 200, f'{label} injection failed: {resp.status_code} {resp.text}'
+    go = http_get(f'{url}/go?sid={sid}').json()
+    assert go.get('initial_state') is not None, f'{label} initial_state is None after injection'
+    print(f'{label} state injected and verified.')
+
+
+inject(NOTION_URL, build_notion_state(), 'Notion')
+inject(SALESFORCE_URL, build_salesforce_state(), 'Salesforce')
+inject(DOCUSIGN_URL, build_docusign_state(), 'DocuSign')
+inject(GMAIL_URL, build_gmail_state(), 'Gmail')
+
+
+# --- Launch browser (primary mock = Salesforce) ---
+def launch_gui(command, delay_sec=1.0):
+    env = os.environ.copy()
+    env['DISPLAY'] = ':0'
+    # Accept both a command string and a pre-split arg list (some tasks
+    # build arg lists themselves, e.g. se_hard_008).
+    if isinstance(command, str):
+        _parts = shlex.split(command)
+    else:
+        _parts = list(command)
+    # Inject CDP debug port so wait_mocks_loaded can verify browser render
+    # via Playwright connect_over_cdp. Idempotent: skip if any arg already
+    # starts with --remote-debugging-port (some tasks inject their own
+    # dynamic port, e.g. se_hard_008).
+    if _parts and 'google-chrome' in _parts[0] \
+       and not any(p.startswith('--remote-debugging-port') for p in _parts):
+        _parts[1:1] = ['--no-first-run', '--no-default-browser-check',
+                       '--remote-debugging-port=1337']
+    _proc = subprocess.Popen(_parts, stdin=subprocess.DEVNULL,
+                             stdout=subprocess.DEVNULL,
+                             stderr=subprocess.DEVNULL, env=env,
+                             start_new_session=True)
+    time.sleep(delay_sec)
+    if _proc.poll() is not None:
+        print('  WARNING: launch_gui process exited early (code=%s)'
+              % _proc.returncode)
+    return _proc
+_primary_mock_url = SALESFORCE_URL
+launch_gui(f'google-chrome "{SALESFORCE_URL}/?sid={sid}"', delay_sec=2.0)
+_open_remaining_mock_tabs(_primary_mock_url)
+wait_mocks_loaded()
+print(f'GUI_READY: launched browser at {SALESFORCE_URL}/?sid={sid}')

csm_qbr_renewal_003/reward.py

@@ -0,0 +1,339 @@
+# --- MOCK_HOST_NO_PROXY_OVERRIDE: bypass woa proxy for locally-deployed mocks ---
+import os as _os
+_np = _os.environ.get("no_proxy", _os.environ.get("NO_PROXY", ""))
+for _h in ("127.0.0.1", "localhost", "28.7.184.198"):
+    if _h not in _np:
+        _np = (_np + "," + _h).lstrip(",")
+_os.environ["no_proxy"] = _np
+_os.environ["NO_PROXY"] = _np
+# --- end override ---
+
+"""
+Reward Script: CSM Renewal QBR prep across Notion, Salesforce, DocuSign, Gmail
+Task ID: csm_qbr_renewal_003
+Domain: mock_websites (multi-mock)
+Scoring (sums to 1.0):
+  - Notion  (0.25): new Aspen QBR agenda page (0.07) + 4 sections each w/ non-empty content (0.18)
+  - Salesforce (0.30): new QBR event w/ runtime this-Friday time/relation/assignee (0.20) + opp-1 advanced forward, not closed (0.10)
+  - DocuSign (0.20): env_1 draft->sent w/ recipient sent + distractor env_2 stays draft (compound)
+  - Gmail   (0.25): new sent email to Daniel (0.12) + body walks agenda & references the QBR time (0.13)
+All scoring components are anchored to task-introduced changes: each FAILS on initial_env, PASSES on golden_env.
+"""
+import re
+import sys
+from datetime import date, timedelta
+
+import requests
+
+# --- Read sid ---
+try:
+    with open('/tmp/task_web_sid') as f:
+        sid = f.read().strip()
+    if not sid:
+        raise ValueError('sid empty')
+except Exception as e:
+    print(f'CRITICAL: cannot read /tmp/task_web_sid: {e}')
+    print('REWARD: 0.0')
+    sys.exit(0)
+
+MOCKS = {
+    'notion': 'http://28.7.184.198:8166',
+    'salesforce': 'http://28.7.184.198:8175',
+    'docusign': 'http://28.7.184.198:8130',
+    'gmail': 'http://28.7.184.198:8138',
+}
+
+# --- Fetch state from all mocks ---
+S = {}
+for name, url in MOCKS.items():
+    try:
+        data = requests.get(f'{url}/go?sid={sid}', timeout=15).json()
+    except Exception as e:
+        print(f'CRITICAL: cannot fetch {name} state: {e}')
+        print('REWARD: 0.0')
+        sys.exit(0)
+    if data.get('current_state') is None:
+        print(f'CRITICAL: {name} current_state is None')
+        print('REWARD: 0.0')
+        sys.exit(0)
+    S[name] = data
+
+
+def norm(s):
+    return re.sub(r'[^a-z0-9 ]', '', (s or '').lower()).strip()
+
+
+SECTIONS = ['business goals', 'usage highlights', 'risks', 'next actions']
+MONTH_NAMES = {
+    1: ('january', 'jan'),
+    2: ('february', 'feb'),
+    3: ('march', 'mar'),
+    4: ('april', 'apr'),
+    5: ('may', 'may'),
+    6: ('june', 'jun'),
+    7: ('july', 'jul'),
+    8: ('august', 'aug'),
+    9: ('september', 'sep'),
+    10: ('october', 'oct'),
+    11: ('november', 'nov'),
+    12: ('december', 'dec'),
+}
+
+
+def this_friday(today=None):
+    today = today or date.today()
+    days_until_friday = (4 - today.weekday()) % 7
+    return today + timedelta(days=days_until_friday)
+
+
+def load_target_qbr_date():
+    try:
+        with open('/tmp/task_qbr_date') as f:
+            return date.fromisoformat(f.read().strip())
+    except Exception:
+        return this_friday()
+
+
+TARGET_QBR_DATE = load_target_qbr_date()
+TARGET_QBR_START = (TARGET_QBR_DATE.year, TARGET_QBR_DATE.month, TARGET_QBR_DATE.day, 14, 0)
+TARGET_QBR_END = (TARGET_QBR_DATE.year, TARGET_QBR_DATE.month, TARGET_QBR_DATE.day, 14, 45)
+
+
+def parse_datetime_minute(value):
+    m = re.search(r'(\d{4})-(\d{1,2})-(\d{1,2})[T ](\d{1,2}):(\d{2})', str(value or ''))
+    if not m:
+        return None
+    return tuple(int(x) for x in m.groups())
+
+
+def datetime_matches(value, target):
+    return parse_datetime_minute(value) == target
+
+
+def text_mentions_target_date(text):
+    ntext = norm(text)
+    year, month, day = TARGET_QBR_DATE.year, TARGET_QBR_DATE.month, TARGET_QBR_DATE.day
+    full_month, short_month = MONTH_NAMES[month]
+    date_phrases = {
+        'friday',
+        f'{year} {month:02d} {day:02d}',
+        f'{year} {month} {day}',
+        f'{full_month} {day}',
+        f'{short_month} {day}',
+    }
+    return any(phrase in ntext for phrase in date_phrases)
+
+
+def text_mentions_2pm(text):
+    ntext = norm(text)
+    return bool(re.search(r'\b2\s*pm\b', ntext) or '200 pm' in ntext or '1400' in ntext)
+
+
+def verify_task():
+    total = 0.0
+
+    # ===================== NOTION (0.25) =====================
+    notion_init = S['notion']['initial_state']
+    notion_cur = S['notion']['current_state']
+    init_page_ids = set(notion_init.get('pages', {}).keys())
+    cur_pages = notion_cur.get('pages', {})
+    cur_blocks = notion_cur.get('blocks', {})
+
+    # Find NEW pages whose title references the Aspen QBR agenda
+    qbr_page = None
+    for pid, page in cur_pages.items():
+        if pid in init_page_ids:
+            continue
+        t = norm(page.get('title'))
+        if 'aspen' in t and ('qbr' in t or 'quarterly business review' in t):
+            qbr_page = page
+            break
+
+    # Component N1: new Aspen QBR agenda page exists (0.07)
+    try:
+        if qbr_page is not None:
+            print(f"PASS: Notion new QBR page '{qbr_page.get('title')}' (0.07)")
+            total += 0.07
+        else:
+            print('FAIL: No NEW Notion page referencing Aspen + QBR found')
+    except Exception as e:
+        print(f'ERROR: Notion N1 — {e}')
+
+    # Component N2: four sections, each a heading with >=1 non-empty content block (0.18, 0.045 each)
+    try:
+        section_has_content = {s: False for s in SECTIONS}
+        section_heading_seen = {s: False for s in SECTIONS}
+        if qbr_page is not None:
+            block_ids = qbr_page.get('blockIds', []) or []
+            current_section = None
+            for bid in block_ids:
+                blk = cur_blocks.get(bid)
+                if not blk:
+                    continue
+                btype = (blk.get('type') or '')
+                content = blk.get('content') or ''
+                ncontent = norm(content)
+                if btype.startswith('heading'):
+                    # does this heading start a tracked section?
+                    matched = None
+                    for s in SECTIONS:
+                        if ncontent == s or s in ncontent:
+                            matched = s
+                            break
+                    if matched:
+                        current_section = matched
+                        section_heading_seen[matched] = True
+                    else:
+                        current_section = None
+                else:
+                    # non-heading content block under the current section
+                    if current_section and content.strip():
+                        section_has_content[current_section] = True
+        for s in SECTIONS:
+            if section_heading_seen[s] and section_has_content[s]:
+                print(f"PASS: Notion section '{s}' has heading + non-empty content (0.045)")
+                total += 0.045
+            else:
+                print(f"FAIL: Notion section '{s}' heading_seen={section_heading_seen[s]} has_content={section_has_content[s]}")
+    except Exception as e:
+        print(f'ERROR: Notion N2 — {e}')
+
+    # ===================== SALESFORCE (0.30) =====================
+    sf_init = S['salesforce']['initial_state']
+    sf_cur = S['salesforce']['current_state']
+    init_act_ids = {a.get('activityId') for a in sf_init.get('activities', [])}
+    cur_acts = sf_cur.get('activities', [])
+
+    # Component S1: NEW QBR event w/ correct runtime date/time, type, relation, assignee (0.20)
+    try:
+        match = None
+        for a in cur_acts:
+            if a.get('activityId') in init_act_ids:
+                continue
+            if a.get('type') != 'event':
+                continue
+            subj = norm(a.get('subject'))
+            is_qbr = 'qbr' in subj or 'quarterly business review' in subj
+            is_aspen = 'aspen' in subj
+            if (is_qbr and is_aspen
+                    and datetime_matches(a.get('startDateTime'), TARGET_QBR_START)
+                    and datetime_matches(a.get('endDateTime'), TARGET_QBR_END)
+                    and a.get('relatedToType') == 'account'
+                    and a.get('relatedToId') == 'account-1'
+                    and a.get('assignedToId') == 'user-1'):
+                match = a
+                break
+        if match:
+            print(f"PASS: Salesforce new QBR event correct ({TARGET_QBR_DATE.isoformat()} 14:00-14:45, account-1, user-1) (0.20)")
+            total += 0.20
+        else:
+            print(f'FAIL: No NEW Salesforce QBR event matching {TARGET_QBR_DATE.isoformat()} 14:00-14:45/relation/assignee')
+    except Exception as e:
+        print(f'ERROR: Salesforce S1 — {e}')
+
+    # Component S2: opp-1 advanced FORWARD from Qualification, still open (0.10)
+    try:
+        FORWARD_OPEN = {'needs analysis', 'value proposition', 'proposal', 'negotiation'}
+        opp1 = next((o for o in sf_cur.get('opportunities', []) if o.get('opportunityId') == 'opp-1'), None)
+        stage = norm(opp1.get('stage')) if opp1 else None
+        if opp1 and stage in FORWARD_OPEN:
+            print(f"PASS: opp-1 advanced to '{opp1.get('stage')}' (forward, open) (0.10)")
+            total += 0.10
+        else:
+            print(f"FAIL: opp-1 stage='{opp1.get('stage') if opp1 else None}' not a forward open stage")
+    except Exception as e:
+        print(f'ERROR: Salesforce S2 — {e}')
+
+    # ===================== DOCUSIGN (0.20) =====================
+    # Compound: env_1 draft->sent w/ recipient sent/delivered AND distractor env_2 stays draft.
+    # Anchored to the env_1 change so it FAILS on initial (env_1 still draft).
+    try:
+        ds_cur = S['docusign']['current_state']
+        envs = {e.get('id'): e for e in ds_cur.get('envelopes', [])}
+        env1 = envs.get('env_1')
+        env2 = envs.get('env_2')
+        env1_is_aspen = bool(env1) and 'aspen' in norm(env1.get('subject'))
+        env1_recips = env1.get('recipients', []) if env1 else []
+        daniel = next((r for r in env1_recips if 'daniel.brooks@aspendynamics.com' in (r.get('email') or '').lower()), None)
+        env1_ok = (
+            env1_is_aspen
+            and env1.get('status') == 'sent'
+            and bool(env1.get('sentAt'))
+            and daniel is not None
+            and daniel.get('status') in ('sent', 'delivered')
+        )
+        env2_still_draft = bool(env2) and env2.get('status') == 'draft'
+        if env1_ok and env2_still_draft:
+            print('PASS: DocuSign env_1 sent w/ Daniel sent/delivered AND env_2 stays draft (0.20)')
+            total += 0.20
+        else:
+            print(f'FAIL: DocuSign env1_ok={env1_ok} env2_still_draft={env2_still_draft}')
+    except Exception as e:
+        print(f'ERROR: DocuSign — {e}')
+
+    # ===================== GMAIL (0.25) =====================
+    gm_init = S['gmail']['initial_state']
+    gm_cur = S['gmail']['current_state']
+    init_email_ids = {m.get('id') for m in gm_init.get('emails', [])}
+    cur_emails = gm_cur.get('emails', [])
+
+    # Find new sent email addressed to Daniel
+    new_sent = None
+    for m in cur_emails:
+        if m.get('id') in init_email_ids:
+            continue
+        if m.get('folder') != 'sent':
+            continue
+        tos = m.get('to') or []
+        if any('daniel.brooks@aspendynamics.com' in (t.get('email') or '').lower() for t in tos):
+            new_sent = m
+            break
+
+    # Component G1: new sent email to Daniel w/ non-empty subject (0.12)
+    try:
+        if new_sent and (new_sent.get('subject') or '').strip():
+            print(f"PASS: Gmail new sent email to Daniel, subject='{new_sent.get('subject')}' (0.12)")
+            total += 0.12
+        else:
+            print('FAIL: No new sent email to daniel.brooks@aspendynamics.com with non-empty subject')
+    except Exception as e:
+        print(f'ERROR: Gmail G1 — {e}')
+
+    # Component G2: body walks agenda (>=3 of 4 sections) + references runtime QBR date/time (0.13)
+    try:
+        if new_sent:
+            body = (new_sent.get('body') or '')
+            nbody = norm(body)
+            sections_hit = sum(1 for s in SECTIONS if s in nbody)
+            mentions_qbr_date = text_mentions_target_date(body)
+            mentions_qbr_time = text_mentions_2pm(body)
+            # award proportionally: agenda walk-through 0.07, date 0.03, 2PM 0.03
+            sub = 0.0
+            if sections_hit >= 3:
+                sub += 0.07
+                print(f'PASS: Gmail body walks agenda ({sections_hit}/4 sections) (0.07)')
+            else:
+                print(f'FAIL: Gmail body mentions only {sections_hit}/4 agenda sections')
+            if mentions_qbr_date:
+                sub += 0.03
+                print(f'PASS: Gmail body references QBR date {TARGET_QBR_DATE.isoformat()} (0.03)')
+            else:
+                print(f'FAIL: Gmail body does not reference QBR date {TARGET_QBR_DATE.isoformat()}')
+            if mentions_qbr_time:
+                sub += 0.03
+                print('PASS: Gmail body names 2 PM time (0.03)')
+            else:
+                print('FAIL: Gmail body does not name 2 PM time')
+            total += sub
+        else:
+            print('FAIL: Gmail G2 — no sent email to evaluate body')
+    except Exception as e:
+        print(f'ERROR: Gmail G2 — {e}')
+
+    final = round(min(total, 1.0), 4)
+    print(f'\nScore: {total}/1.0')
+    print(f'REWARD: {final}')
+    return final
+
+
+verify_task()

csm_qbr_renewal_003/reward_label.json

@@ -0,0 +1,87 @@
+{
+  "reward_file": "/apdcephfs_sh2/share_300000800/user/jackwkwang/long-horizon-tasks/mini_osworld_xiangwu/cache/csm_qbr_renewal_003/reward.py",
+  "api_url": "http://28.7.184.184:8000/v1/chat/completions",
+  "model": "Kimi-K2.6",
+  "max_code_chars": 100000,
+  "created_at": "2026-07-03 20:22:31",
+  "label": {
+    "task_id": "csm_qbr_renewal_003",
+    "domain": "mock_websites",
+    "summary": "验证在 Notion、Salesforce、DocuSign、Gmail 上完成 CSM Renewal QBR 准备任务的行为，包括创建 QBR 议程页面、安排会议事件、推进销售机会、发送 DocuSign 信封以及发送包含议程和会议时间的邮件。",
+    "is_placeholder": false,
+    "data_sources": [
+      "/tmp/task_web_sid",
+      "/tmp/task_qbr_date",
+      "http://28.7.186.212:8186 (notion)",
+      "http://28.7.186.212:8195 (salesforce)",
+      "http://28.7.186.212:8150 (docusign)",
+      "http://28.7.186.212:8158 (gmail)"
+    ],
+    "scoring_components": [
+      {
+        "name": "Notion N1",
+        "weight": 0.07,
+        "description": "新建 Aspen QBR 议程页面",
+        "check_logic": "遍历 current_state.pages，排除 initial_state 中已存在的 pageId，检查是否有页面标题经 norm 处理后同时包含 'aspen' 和 'qbr'（或 'quarterly business review'）",
+        "pass_condition": "存在一个新的 Notion 页面，其标题同时提及 Aspen 和 QBR"
+      },
+      {
+        "name": "Notion N2",
+        "weight": 0.18,
+        "description": "四个指定 section 均包含 heading 和非空内容",
+        "check_logic": "在 qbr_page 的 blockIds 中顺序遍历 block，识别 heading 类型块是否匹配 SECTIONS（business goals、usage highlights、risks、next actions）之一；匹配后，后续非 heading 块若 content.strip() 非空，则标记该 section 有内容。每个 section 独立计分 0.045。",
+        "pass_condition": "四个 section 每个都必须出现对应的 heading 且在该 heading 下至少有一个非空的非 heading 内容块"
+      },
+      {
+        "name": "Salesforce S1",
+        "weight": 0.2,
+        "description": "新建 QBR 事件，时间、关联对象和负责人正确",
+        "check_logic": "遍历 current_state.activities，排除 initial_state 中已有的 activityId，筛选 type='event' 且 subject 经 norm 后包含 'qbr' 和 'aspen' 的新活动；检查 startDateTime 匹配 TARGET_QBR_START（目标日期 14:00）、endDateTime 匹配 TARGET_QBR_END（目标日期 14:45）、relatedToType='account'、relatedToId='account-1'、assignedToId='user-1'。",
+        "pass_condition": "存在一个新的 Event，主题为 Aspen QBR，开始时间为目标日期 14:00，结束时间为 14:45，关联 account-1，负责人为 user-1"
+      },
+      {
+        "name": "Salesforce S2",
+        "weight": 0.1,
+        "description": "机会 opp-1 向前推进且仍处于开放阶段",
+        "check_logic": "在 current_state.opportunities 中查找 opportunityId='opp-1'，将其 stage 经 norm 后与集合 {'needs analysis', 'value proposition', 'proposal', 'negotiation'} 比对。",
+        "pass_condition": "opp-1 的阶段为 Needs Analysis、Value Proposition、Proposal 或 Negotiation（即向前推进且未关闭）"
+      },
+      {
+        "name": "DocuSign",
+        "weight": 0.2,
+        "description": "env_1 发送给 Daniel 且 distractor env_2 保持草稿",
+        "check_logic": "检查 current_state.envelopes 中 id='env_1' 的 envelope：subject 经 norm 后包含 'aspen'，status='sent'，sentAt 非空，收件人列表中存在 email 包含 'daniel.brooks@aspendynamics.com' 且其 status 为 'sent' 或 'delivered'；同时 id='env_2' 的 envelope status 必须为 'draft'。两者同时满足才得分。",
+        "pass_condition": "env_1 状态为 sent、主题含 Aspen、Daniel 的收件人状态为 sent/delivered，且 env_2 状态仍为 draft"
+      },
+      {
+        "name": "Gmail G1",
+        "weight": 0.12,
+        "description": "新建发送给 Daniel 的邮件且主题非空",
+        "check_logic": "遍历 current_state.emails，排除 initial_state 中已有的 id，筛选 folder='sent' 且 to 列表中任一收件人 email 包含 'daniel.brooks@aspendynamics.com' 的新邮件，并检查 subject 经 strip 后非空。",
+        "pass_condition": "存在一封新的 sent 邮件发送给 daniel.brooks@aspendynamics.com，且主题非空"
+      },
+      {
+        "name": "Gmail G2",
+        "weight": 0.13,
+        "description": "邮件正文包含议程、QBR 日期和 2PM 时间",
+        "check_logic": "对 new_sent.body 进行 norm 处理，统计包含 SECTIONS 中至少 3 个 section 名称（得 0.07）；检查正文是否通过 text_mentions_target_date 提及目标 QBR 日期（得 0.03）；检查是否通过 text_mentions_2pm 提及 2 PM（得 0.03）。三个子条件分数累加为 sub 后加入 total。",
+        "pass_condition": "正文至少提及 4 个议程 section 中的 3 个、提及目标 QBR 日期、并提及 2 PM 时间"
+      }
+    ],
+    "total_max_score": 1.0,
+    "score_aggregation": "各组件分数累加求和，最终用 min(total, 1.0) 钳制到上限后四舍五入到 4 位小数",
+    "failure_modes": [
+      "读取 /tmp/task_web_sid 失败或内容为空：打印 CRITICAL 并返回 REWARD: 0.0 后 sys.exit(0)",
+      "从任一 mock 服务（notion/salesforce/docusign/gmail）拉取状态失败：打印 CRITICAL 并返回 REWARD: 0.0 后 sys.exit(0)",
+      "任一 mock 返回的 current_state 为 None：打印 CRITICAL 并返回 REWARD: 0.0 后 sys.exit(0)",
+      "Notion 未找到新的 Aspen QBR 页面：N1 和 N2 均不得分",
+      "Salesforce 未找到符合条件的新 QBR Event：S1 不得分",
+      "opp-1 阶段未前进或已关闭：S2 不得分",
+      "DocuSign env_1 未发送或 Daniel 状态不对，或 env_2 不再是 draft：DocuSign 组件 0.20 分全部丢失",
+      "Gmail 未找到发给 Daniel 的新 sent 邮件：G1 和 G2 均不得分"
+    ],
+    "scoring_logic_explanation": "脚本首先读取 /tmp/task_web_sid 获取会话 ID，并依次从 notion、salesforce、docusign、gmail 四个 mock 服务拉取 initial_state 与 current_state；同时从 /tmp/task_qbr_date 加载目标日期（或默认取本周五）。随后按顺序执行验证：1) Notion 部分查找标题同时包含 Aspen 和 QBR 的新页面（0.07），并检查该页面是否包含 business goals、usage highlights、risks、next actions 四个 section 的 heading 及非空内容（每个 0.045）；2) Salesforce 部分查找新的 Event，验证其主题、目标日期 14:00-14:45 的时间、关联 account-1 及负责人 user-1（0.20），并检查 opp-1 的阶段是否推进到开放的前进阶段（0.10）；3) DocuSign 部分验证 env_1 变为 sent 且收件人 Daniel 状态为 sent/delivered，同时 distractor env_2 必须保持 draft（0.20）；4) Gmail 部分查找发给 Daniel 的新 sent 邮件，验证主题非空（0.12），并检查正文是否至少提及 3 个议程 section（0.07）、目标 QBR 日期（0.03）和 2 PM 时间（0.03）。所有子项分数累加，最终通过 min(total, 1.0) 限制上限并四舍五入输出。",
+    "confidence": 1.0
+  },
+  "raw_response": "用户要求我分析一个 reward.py 评分脚本的评分逻辑，并输出一个严格的 JSON 对象。\n\n让我仔细阅读代码并提取所需信息：\n\n1. **task_id**: 从 docstring 中可以看到 \"Task ID: csm_qbr_renewal_003\"\n2. **domain**: 从 docstring 中可以看到 \"Domain: mock_websites (multi-mock)\"\n3. **summary**: 一句话概括。这个脚本验证在 Notion、Salesforce、DocuSign、Gmail 上完成 CSM Renewal QBR 准备任务的行为。\n4. **is_placeholder**: false，因为脚本有详细的检查逻辑，不是占位脚本。\n5. **data_sources**: 脚本从以下 mock 服务获取状态：\n   - notion: http://28.7.186.212:8186\n   - salesforce: http://28.7.186.212:8195\n   - docusign: http://28.7.186.212:8150\n   - gmail: http://28.7.186.212:8158\n   - /tmp/task_web_sid (读取 sid)\n   - /tmp/task_qbr_date (读取目标 QBR 日期)\n\n6. **scoring_components**: 需要详细列出各个组件：\n   - Notion N1: 0.07 - 新建 Aspen QBR agenda 页面\n   - Notion N2: 0.18 (4 * 0.045) - 四个 section 每个都有 heading 和非空内容\n   - Salesforce S1: 0.20 - 新建 QBR event，正确时间/关系/负责人\n   - Salesforce S2: 0.10 - opp-1 推进到前进阶段且未关闭\n   - DocuSign: 0.20 - env_1 从 draft 变为 sent 且 Daniel 状态为 sent/delivered，env_2 保持 draft\n   - Gmail G1: 0.12 - 新建发送给 Daniel 的邮件，主题非空\n   - Gmail G2: 0.13 (0.07 + 0.03 + 0.03) - 邮件正文包含至少3个 agenda section、提及 QBR 日期、提及 2PM 时间\n\n7. **total_max_score**: 1.0\n8. **score_aggregation**: 各组件分数相加，最后用 min(total, 1.0) 钳制到上限，并四舍五入到4位小数。\n9. **failure_modes**: \n   - 读取 /tmp/task_web_sid 失败或为空 -> CRITICAL, REWARD: 0.0, sys.exit(0)\n   - 无法从任一 mock 服务获取状态 -> CRITICAL, REWARD: 0.0, sys.exit(0)\n   - 任一 mock 返回 current_state 为 None -> CRITICAL, REWARD: 0.0, sys.exit(0)\n   - 各组件检查失败只是不加分，不会提前退出（除了 DocuSign 是复合检查，必须同时满足两个条件）\n\n10. **scoring_logic_explanation**: 详细解释执行顺序：\n    - 首先读取 /tmp/task_web_sid 获取会话 ID\n    - 从四个 mock 服务（Notion、Salesforce、DocuSign、Gmail）拉取 initial_state 和 current_state\n    - 加载目标 QBR 日期（从 /tmp/task_qbr_date 或计算本周五）\n    - Notion 检查：查找标题包含 Aspen 和 QBR 的新页面；检查四个指定 section 是否有 heading 和非空内容块\n    - Salesforce 检查：查找新的 QBR event，验证时间、关联账户、负责人；检查 opp-1 阶段是否向前推进且未关闭\n    - DocuSign 检查：验证 env_1 变为 sent 且收件人 Daniel 状态为 sent/delivered，同时 env_2 保持 draft\n    - Gmail 检查：查找发给 Daniel 的新 sent 邮件，验证主题非空；验证正文包含至少3个 agenda section、提及 QBR 日期和 2PM 时间\n    - 汇总所有得分，上限 1.0，输出 REWARD\n\n11. **confidence**: 1.0，因为代码逻辑清晰完整。\n\n让我再仔细检查一下各个组件的权重和逻辑：\n\nNotion:\n- N1: 0.07 (代码中 total += 0.07)\n- N2: 每个 section 0.045，共 4 个，总计 0.18 (代码中 total += 0.045 在循环里)\n\nSalesforce:\n- S1: 0.20 (代码中 total += 0.20)\n- S2: 0.10 (代码中 total += 0.10)\n\nDocuSign:\n- 0.20 (代码中 total += 0.20) - 这是一个复合条件，必须同时满足 env1_ok 和 env2_still_draft\n\nGmail:\n- G1: 0.12 (代码中 total += 0.12)\n- G2: 0.13，但内部是细分的：\n  - sections_hit >= 3: 0.07\n  - mentions_qbr_date: 0.03\n  - mentions_qbr_time: 0.03\n  总计 0.13，代码中是 sub += ... 然后 total += sub\n\nTotal: 0.07 + 0.18 + 0.20 + 0.10 + 0.20 + 0.12 + 0.13 = 1.00\n\nScore aggregation: 各组件分数累加，最终 `round(min(total, 1.0), 4)`\n\nFailure modes:\n- /tmp/task_web_sid 读取失败\n- mock 状态拉取失败\n- current_state 为 None\n- 各检查项不满足（不加分）\n\n我需要确保 JSON 格式完全符合要求：\n- 不要 Markdown 代码块\n- 第一个字符是 {，最后一个字符是 }\n- 所有字符串 value 使用中文（task_id、domain、字段名、代码标识符保留原样）\n- 不要额外解释\n\n让我构建 JSON：\n\n```json\n{\n  \"task_id\": \"csm_qbr_renewal_003\",\n  \"domain\": \"mock_websites\",\n  \"summary\": \"验证在 Notion、Salesforce、DocuSign、Gmail 上完成 CSM Renewal QBR 准备任务的行为，包括创建 QBR 议程页面、安排会议事件、推进销售机会、发送 DocuSign 信封以及发送包含议程和会议时间的邮件。\",\n  \"is_placeholder\": false,\n  \"data_sources\": [\n    \"/tmp/task_web_sid\",\n    \"/tmp/task_qbr_date\",\n    \"http://28.7.186.212:8186 (notion)\",\n    \"http://28.7.186.212:8195 (salesforce)\",\n    \"http://28.7.186.212:8150 (docusign)\",\n    \"http://28.7.186.212:8158 (gmail)\"\n  ],\n  \"scoring_components\": [\n    {\n      \"name\": \"Notion N1\",\n      \"weight\": 0.07,\n      \"description\": \"新建 Aspen QBR 议程页面\",\n      \"check_logic\": \"遍历 current_state.pages，排除 initial_state 中已存在的 pageId，检查是否有页面标题经 norm 处理后同时包含 'aspen' 和 'qbr'（或 'quarterly business review'）\",\n      \"pass_condition\": \"存在一个新的 Notion 页面，其标题同时提及 Aspen 和 QBR\"\n    },\n    {\n      \"name\": \"Notion N2\",\n      \"weight\": 0.18,\n      \"description\": \"四个指定 section 均包含 heading 和非空内容\",\n      \"check_logic\": \"在 qbr_page 的 blockIds 中顺序遍历 block，识别 heading 类型块是否匹配 SECTIONS（business goals、usage highlights、risks、next actions）之一；匹配后，后续非 heading 块若 content.strip() 非空，则标记该 section 有内容。每个 section 独立计分 0.045。\",\n      \"pass_condition\": \"四个 section 每个都必须出现对应的 heading 且在该 heading 下至少有一个非空的非 heading 内容块\"\n    },\n    {\n      \"name\": \"Salesforce S1\",\n      \"weight\": 0.20,\n      \"description\": \"新建 QBR 事件，时间、关联对象和负责人正确\",\n      \"check_logic\": \"遍历 current_state.activities，排除 initial_state 中已有的 activityId，筛选 type='event' 且 subject 经 norm 后包含 'qbr' 和 'aspen' 的新活动；检查 startDateTime 匹配 TARGET_QBR_START（目标日期 14:00）、endDateTime 匹配 TARGET_QBR_END（目标日期 14:45）、relatedToType='account'、relatedToId='account-1'、assignedToId='user-1'。\",\n      \"pass_condition\": \"存在一个新的 Event，主题为 Aspen QBR，开始时间为目标日期 14:00，结束时间为 14:45，关联 account-1，负责人为 user-1\"\n    },\n    {\n      \"name\": \"Salesforce S2\",\n      \"weight\": 0.10,\n      \"description\": \"机会 opp-1 向前推进且仍处于开放阶段\",\n      \"check_logic\": \"在 current_state.opportunities 中查找 opportunityId='opp-1'，将其 stage 经 norm 后与集合 {'needs analysis', 'value proposition', 'proposal', 'negotiation'} 比对。\",\n      \"pass_condition\": \"opp-1 的阶段为 Needs Analysis、Value Proposition、Proposal 或 Negotiation（即向前推进且未关闭）\"\n    },\n    {\n      \"name\": \"DocuSign\",\n      \"weight\": 0.20,\n      \"description\": \"env_1 发送给 Daniel 且 distractor env_2 保持草稿\",\n      \"check_logic\": \"检查 current_state.envelopes 中 id='env_1' 的 envelope：subject 经 norm 后包含 'aspen'，status='sent'，sentAt 非空，收件人列表中存在 email 包含 'daniel.brooks@aspendynamics.com' 且其 status 为 'sent' 或 'delivered'；同时 id='env_2' 的 envelope status 必须为 'draft'。两者同时满足才得分。\",\n      \"pass_condition\": \"env_1 状态为 sent、主题含 Aspen、Daniel 的收件人状态为 sent/delivered，且 env_2 状态仍��� draft\"\n    },\n    {\n      \"name\": \"Gmail G1\",\n      \"weight\": 0.12,\n      \"description\": \"新建发送给 Daniel 的邮件且主题非空\",\n      \"check_logic\": \"遍历 current_state.emails，排除 initial_state 中已有的 id，筛选 folder='sent' 且 to 列表中任一收件人 email 包含 'daniel.brooks@aspendynamics.com' 的新邮件，并检查 subject 经 strip 后非空。\",\n      \"pass_condition\": \"存在一封新的 sent 邮件发送给 daniel.brooks@aspendynamics.com，且主题非空\"\n    },\n    {\n      \"name\": \"Gmail G2\",\n      \"weight\": 0.13,\n      \"description\": \"邮件正文包含议程、QBR 日期和 2PM 时间\",\n      \"check_logic\": \"对 new_sent.body 进行 norm 处理，统计包含 SECTIONS 中至少 3 个 section 名称（得 0.07）；检查正文是否通过 text_mentions_target_date 提及目标 QBR 日期（得 0.03）；检查是否通过 text_mentions_2pm 提及 2 PM（得 0.03）。三个子条件分数累加为 sub 后加入 total。\",\n      \"pass_condition\": \"正文至少提及 4 个议程 section 中的 3 个、提及目标 QBR 日期、并提及 2 PM 时间\"\n    }\n  ],\n  \"total_max_score\": 1.0,\n  \"score_aggregation\": \"各组件分数累加求和，最终用 min(total, 1.0) 钳制到上限后四舍五入到 4 位小数\",\n  \"failure_modes\": [\n    \"读取 /tmp/task_web_sid 失败或内容为空：打印 CRITICAL 并返回 REWARD: 0.0 后 sys.exit(0)\",\n    \"从任一 mock 服务（notion/salesforce/docusign/gmail）拉取状态失败：打印 CRITICAL 并返回 REWARD: 0.0 后 sys.exit(0)\",\n    \"任一 mock 返回的 current_state 为 None：打印 CRITICAL 并返回 REWARD: 0.0 后 sys.exit(0)\",\n    \"Notion 未找到新的 Aspen QBR 页面：N1 和 N2 均不得分\",\n    \"Salesforce 未找到符合条件的新 QBR Event：S1 不得分\",\n    \"opp-1 阶段未前进或已关闭：S2 不得分\",\n    \"DocuSign env_1 未发送或 Daniel 状态不对，或 env_2 不再是 draft：DocuSign 组件 0.20 分全部丢失\",\n    \"Gmail 未找到发给 Daniel 的新 sent 邮件：G1 和 G2 均不得分\"\n  ],\n  \"scoring_logic_explanation\": \"脚本首先读取 /tmp/task_web_sid 获取会话 ID，并依次从 notion、salesforce、docusign、gmail 四个 mock 服务拉取 initial_state 与 current_state；同时从 /tmp/task_qbr_date 加载目标日期（或默认取本周五）。随后按顺序执行验证：1) Notion 部分查找标题同时包含 Aspen 和 QBR 的新页面（0.07），并检查该页面是否包含 business goals、usage highlights、risks、next actions 四个 section 的 heading 及非空内容（每个 0.045）；2) Salesforce 部分查找新的 Event，验证其主题、目标日期 14:00-14:45 的时间、关联 account-1 及负责人 user-1（0.20），并检查 opp-1 的阶段是否推进到开放的前进阶段（0.10）；3) DocuSign 部分验证 env_1 变为 sent 且收件人 Daniel 状态为 sent/delivered，同时 distractor env_2 必须保持 draft（0.20）；4) Gmail 部分查找发给 Daniel 的新 sent 邮件，验证主题非空（0.12），并检查正文是否至少提及 3 个议程 section（0.07）、目标 QBR 日期（0.03）和 2 PM 时间（0.03）。所有子项分数累加，最终通过 min(total, 1.0) 限制上限并四舍五入输出。\",\n  \"confidence\": 1.0\n}\n```\n\n我需要确保没有使用 Markdown 代码块，且输出严格是 JSON。让我再检查一下 domain，docstring 写的是 \"mock_websites (multi-mock)\"，我应该保留为 \"mock_websites\" 还是原样？用户说 \"如 mock_websites / software_engineer / libreoffice_calc\"，所以 \"mock_websites\" 应该可以。或者我可以写 \"mock_websites (multi-mock)\"？用户说 \"从 docstring 中提取的领域\"，docstring 中明确写了 \"Domain: mock_websites (multi-mock)\"。为了准确，我应该保留原样或至少保留核心。看示例是保留主要标识，我会用 \"mock_websites\" 因为示例里就是这样用的。\n\n再检查 data_sources，用户示例里有 \"slack_mock、github_mock、salesforce_mock、/tmp/task_web_sid\"。我应该列出具体的 URL 或服务名。我会用 \"notion_mock\"、\"salesforce_mock\"、\"docusign_mock\"、\"gmail_mock\" 还是具体 URL？代码里用的是 MOCKS 字典，URL 是具体的 IP:端口。我会写 URL 并标注名称，这样更准确。\n\n再检查 weight 字段。Gmail G2 的 weight 应该是 0.13，虽然内部是 0.07+0.03+0.03，但作为一个组件整体是 0.13。或者我应该拆成三个子组件？用户要求列出组件，我可以把 G2 作为一个组件，weight 0.13，在 check_logic 中说明内部分数。这样比较合理，因为代码里它们是放在一个 try 块里并累加到 sub 再 total += sub。\n\nDocuSign 也是一个整体 0.20 的复合组件。\n\nNotion N2 是 0.18 整体，内部 4 个 0.045。我可以作为一个组件 weight 0.18。\n\n这样组件列表就是：\n- Notion N1: 0.07\n- Notion N2: 0.18\n- Salesforce S1: 0.20\n- Salesforce S2: 0.10\n- DocuSign: 0.20\n- Gmail G1: 0.12\n- Gmail G2: 0.13\n\n总和正好是 1.0。\n\n最后确认 JSON 格式正确，没有 trailing comma，所有字符串正确转义。 </think> {\"task_id\":\"csm_qbr_renewal_003\",\"domain\":\"mock_websites\",\"summary\":\"验证在 Notion、Salesforce、DocuSign、Gmail 上完成 CSM Renewal QBR 准备任务的行为，包括创建 QBR 议程页面、安排会议事件、推进销售机会、发送 DocuSign 信封以及发送包含议程和会议时间的邮件。\",\"is_placeholder\":false,\"data_sources\":[\"/tmp/task_web_sid\",\"/tmp/task_qbr_date\",\"http://28.7.186.212:8186 (notion)\",\"http://28.7.186.212:8195 (salesforce)\",\"http://28.7.186.212:8150 (docusign)\",\"http://28.7.186.212:8158 (gmail)\"],\"scoring_components\":[{\"name\":\"Notion N1\",\"weight\":0.07,\"description\":\"新建 Aspen QBR 议程页面\",\"check_logic\":\"遍历 current_state.pages，排除 initial_state 中已存在的 pageId，检查是否有页面标题经 norm 处理后同时包含 'aspen' 和 'qbr'（或 'quarterly business review'）\",\"pass_condition\":\"存在一个新的 Notion 页面，其标题同时提及 Aspen 和 QBR\"},{\"name\":\"Notion N2\",\"weight\":0.18,\"description\":\"四个指定 section 均包含 heading 和非空内容\",\"check_logic\":\"在 qbr_page 的 blockIds 中顺序遍历 block，识别 heading 类型块是否匹配 SECTIONS（business goals、usage highlights、risks、next actions）之一；匹配后，后续非 heading 块若 content.strip() 非空，则标记该 section 有内容。每个 section 独立计分 0.045。\",\"pass_condition\":\"四个 section 每个都必须出现对应的 heading 且在该 heading 下至少有一个非空的非 heading 内容块\"},{\"name\":\"Salesforce S1\",\"weight\":0.20,\"description\":\"新建 QBR 事件，时间、关联对象和负责人正确\",\"check_logic\":\"遍历 current_state.activities，排除 initial_state 中已有的 activityId，筛选 type='event' 且 subject 经 norm 后包含 'qbr' 和 'aspen' 的新活动；检查 startDateTime 匹配 TARGET_QBR_START（目标日期 14:00）、endDateTime 匹配 TARGET_QBR_END（目标日期 14:45）、relatedToType='account'、relatedToId='account-1'、assignedToId='user-1'。\",\"pass_condition\":\"存在一个新的 Event，主题为 Aspen QBR，开始时间为目标日期 14:00，结束时间为 14:45，关联 account-1，负责人为 user-1\"},{\"name\":\"Salesforce S2\",\"weight\":0.10,\"description\":\"机会 opp-1 向前推进且仍处于开放阶段\",\"check_logic\":\"在 current_state.opportunities 中查找 opportunityId='opp-1'，将其 stage 经 norm 后与集合 {'needs analysis', 'value proposition', 'proposal', 'negotiation'} 比对。\",\"pass_condition\":\"opp-1 的阶段为 Needs Analysis、Value Proposition、Proposal 或 Negotiation（即向前推进且未关闭）\"},{\"name\":\"DocuSign\",\"weight\":0.20,\"description\":\"env_1 发送给 Daniel 且 distractor env_2 保持草稿\",\"check_logic\":\"检查 current_state.envelopes 中 id='env_1' 的 envelope：subject 经 norm 后包含 'aspen'，status='sent'，sentAt 非空，收件人列表中存在 email 包含 'daniel.brooks@aspendynamics.com' 且其 status 为 'sent' 或 'delivered'；同时 id='env_2' 的 envelope status 必须为 'draft'。两者同时满足才得分。\",\"pass_condition\":\"env_1 状态为 sent、主题含 Aspen、Daniel 的收件人状态为 sent/delivered，且 env_2 状态仍为 draft\"},{\"name\":\"Gmail G1\",\"weight\":0.12,\"description\":\"新建发送给 Daniel 的邮件且主题非空\",\"check_logic\":\"遍历 current_state.emails，排除 initial_state 中已有的 id，筛选 folder='sent' 且 to 列表中任一收件人 email 包含 'daniel.brooks@aspendynamics.com' 的新邮件，并检查 subject 经 strip 后非空。\",\"pass_condition\":\"存在一封新的 sent 邮件发送给 daniel.brooks@aspendynamics.com，且主题非空\"},{\"name\":\"Gmail G2\",\"weight\":0.13,\"description\":\"邮件正文包含议程、QBR 日期和 2PM 时间\",\"check_logic\":\"对 new_sent.body 进行 norm 处理，统计包含 SECTIONS 中至少 3 个 section 名称（得 0.07）；检查正文是否通过 text_mentions_target_date 提及目标 QBR 日期（得 0.03）；检查是否通过 text_mentions_2pm 提及 2 PM（得 0.03）。三个子条件分数累加为 sub 后加入 total。\",\"pass_condition\":\"正文至少提及 4 个议程 section 中的 3 个、提及目标 QBR 日期、并提及 2 PM 时间\"}],\"total_max_score\":1.0,\"score_aggregation\":\"各组件分数累加求和，最终用 min(total, 1.0) 钳制到上限后四舍五入到 4 位小数\",\"failure_modes\":[\"读取 /tmp/task_web_sid 失败或内容为空：打印 CRITICAL 并返回 REWARD: 0.0 后 sys.exit(0)\",\"从任一 mock 服务（notion/salesforce/docusign/gmail）拉取状态失败：打印 CRITICAL 并返回 REWARD: 0.0 后 sys.exit(0)\",\"任一 mock 返回的 current_state 为 None：打印 CRITICAL 并返回 REWARD: 0.0 后 sys.exit(0)\",\"Notion 未找到新的 Aspen QBR 页面：N1 和 N2 均不得分\",\"Salesforce 未找��符合条件的新 QBR Event：S1 不得分\",\"opp-1 阶段未前进或已关闭：S2 不得分\",\"DocuSign env_1 未发送或 Daniel 状态不对，或 env_2 不再是 draft：DocuSign 组件 0.20 分全部丢失\",\"Gmail 未找到发给 Daniel 的新 sent 邮件：G1 和 G2 均不得分\"],\"scoring_logic_explanation\":\"脚本首先读取 /tmp/task_web_sid 获取会话 ID，并依次从 notion、salesforce、docusign、gmail 四个 mock 服务拉取 initial_state 与 current_state；同时从 /tmp/task_qbr_date 加载目标日期（或默认取本周五）。随后按顺序执行验证：1) Notion 部分查找标题同时包含 Aspen 和 QBR 的新页面（0.07），并检查该页面是否包含 business goals、usage highlights、risks、next actions 四个 section 的 heading 及非空内容（每个 0.045）；2) Salesforce 部分查找新的 Event，验证其主题、目标日期 14:00-14:45 的时间、关联 account-1 及负责人 user-1（0.20），并检查 opp-1 的阶段是否推进到开放的前进阶段（0.10）；3) DocuSign 部分验证 env_1 变为 sent 且收件人 Daniel 状态为 sent/delivered，同时 distractor env_2 必须保持 draft（0.20）；4) Gmail 部分查找发给 Daniel 的新 sent 邮件，验证主题非空（0.12），并检查正文是否至少提及 3 个议程 section（0.07）、目标 QBR 日期（0.03）和 2 PM 时间（0.03）。所有子项分数累加，最终通过 min(total, 1.0) 限制上限并四舍五入输出。\",\"confidence\":1.0}"
+}

csops_p1_incident_002/_cua_gym_vm_bridge.sh

@@ -0,0 +1,106 @@
+#!/bin/bash
+# CUA-Gym VM-side network fix (direct-IP mock access).
+#
+# Runs INSIDE the QEMU guest. The CUA-Gym mock services are now reached by
+# their real IDC address (http://28.7.184.198:81xx) instead of the old
+# localhost:9000-9097 socat bridge, so the VM only needs two fixes:
+#
+#   1. Working DNS — the qcow2-baked /etc/resolv.conf points at the Pod's
+#      cluster DNS, which is unreachable from QEMU SLIRP NAT
+#      ("Temporary failure in name resolution").
+#   2. The mock host excluded from the woa http proxy — otherwise the VM's
+#      Chrome (gsettings proxy) and initial_setup.py's `requests` get a 403
+#      from the proxy when hitting 28.7.184.198:81xx. (The old socat path never
+#      hit this because localhost is bypassed by default.)
+#
+# Idempotent. Scoped to the CUA-Gym bridge step; does not touch the generic
+# OSWorld setup pipeline.
+#
+# Argv:
+#   $1  optional sudo password (default "password"; SetupController
+#       substitutes {CLIENT_PASSWORD} when uploaded as a cache file).
+#   $2  unused (kept for backward compat — was socat first port 9000).
+#   $3  unused (kept for backward compat — was socat last  port 9097).
+#   $4  optional mock host IP (default 28.7.184.198).
+
+set +e
+
+PASS="${1:-password}"
+MOCK_HOST="${4:-28.7.184.198}"
+# Derive the /16 IDC subnet so future port/host tweaks on the same segment
+# stay bypassed too (28.7.184.198 -> 28.7.0.0/16).
+MOCK_NET="$(echo "${MOCK_HOST}" | awk -F. 'NF==4{print $1"."$2".0.0/16"}')"
+
+run_root() {
+    echo "$PASS" | sudo -S -p "" bash -c "$1"
+}
+
+# Detect SLIRP gateway from VM's routing table; fall back to 10.0.2.2.
+GW="$(ip route 2>/dev/null | awk '/default/ {print $3; exit}')"
+GW="${GW:-10.0.2.2}"
+
+# ─── CUA-Gym network fix (Taiji-specific) ─────────────────────────────
+fix_vm_network() {
+    # 0) Discover VM network state at runtime — gateway, own IP/CIDR, SLIRP
+    #    DNS forwarder. Nothing hardcoded except the mock host above.
+    local GW_NET="${GW%.*}"
+    local SLIRP_DNS="${GW_NET}.3"        # QEMU SLIRP DNS proxy is always .3
+    local OWN_CIDR
+    OWN_CIDR="$(ip -4 -o addr show scope global 2>/dev/null \
+                | awk '{print $4; exit}')"
+    local OWN_IP="${OWN_CIDR%/*}"
+    local OWN_NET=""
+    [ -n "${OWN_IP}" ] && OWN_NET="${OWN_IP%.*}.0/24"
+
+    # 1) DNS — point at SLIRP DNS proxy + public fallbacks. chattr +i keeps
+    #    systemd-resolved/NetworkManager from clobbering it.
+    if run_root "printf 'nameserver ${SLIRP_DNS}\\nnameserver 114.114.114.114\\nnameserver 8.8.8.8\\n' > /etc/resolv.conf
+                 chattr +i /etc/resolv.conf 2>/dev/null || true"; then
+        echo "[cua_gym_vm_bridge] DNS set: ${SLIRP_DNS}, 114.114.114.114, 8.8.8.8 (gw=${GW})"
+    else
+        echo "[cua_gym_vm_bridge] WARN: failed to write /etc/resolv.conf"
+    fi
+
+    # 2) Build the proxy bypass list from detected values + the direct-IP
+    #    mock host (and its /16 subnet). THIS is what makes 28.7.184.198:81xx
+    #    reachable from inside the VM — without it the woa proxy 403s it.
+    local ignore_list="['localhost', '127.0.0.0/8', '${GW}'"
+    [ -n "${OWN_IP}" ]    && ignore_list="${ignore_list}, '${OWN_IP}'"
+    [ -n "${OWN_NET}" ]   && ignore_list="${ignore_list}, '${OWN_NET}'"
+    [ -n "${GW_NET}" ]    && ignore_list="${ignore_list}, '${GW_NET}.0/24'"
+    [ -n "${MOCK_HOST}" ] && ignore_list="${ignore_list}, '${MOCK_HOST}'"
+    [ -n "${MOCK_NET}" ]  && ignore_list="${ignore_list}, '${MOCK_NET}'"
+    ignore_list="${ignore_list}]"
+
+    # 3) GNOME proxy ignore-hosts — Chrome bypasses the woa proxy for these.
+    if gsettings set org.gnome.system.proxy ignore-hosts "${ignore_list}" 2>/tmp/gsettings_err; then
+        echo "[cua_gym_vm_bridge] gsettings ignore-hosts = ${ignore_list}"
+    else
+        echo "[cua_gym_vm_bridge] WARN: gsettings ignore-hosts failed: $(cat /tmp/gsettings_err 2>/dev/null)"
+    fi
+
+    # 4) Shell env — comma-separated form for no_proxy/NO_PROXY (requests/curl).
+    local NP="localhost,127.0.0.1,${GW}"
+    [ -n "${OWN_IP}" ]    && NP="${NP},${OWN_IP}"
+    [ -n "${OWN_NET}" ]   && NP="${NP},${OWN_NET}"
+    [ -n "${GW_NET}" ]    && NP="${NP},${GW_NET}.0/24"
+    [ -n "${MOCK_HOST}" ] && NP="${NP},${MOCK_HOST}"
+    [ -n "${MOCK_NET}" ]  && NP="${NP},${MOCK_NET}"
+
+    for var in no_proxy NO_PROXY; do
+        if ! grep -q "export ${var}=${NP}" /home/user/.bashrc 2>/dev/null; then
+            echo "export ${var}=${NP}" >> /home/user/.bashrc
+        fi
+    done
+    run_root "grep -q '^no_proxy=' /etc/environment 2>/dev/null \
+              && sed -i 's|^no_proxy=.*|no_proxy=\"${NP}\"|' /etc/environment \
+              || echo 'no_proxy=\"${NP}\"' >> /etc/environment
+              grep -q '^NO_PROXY=' /etc/environment 2>/dev/null \
+              && sed -i 's|^NO_PROXY=.*|NO_PROXY=\"${NP}\"|' /etc/environment \
+              || echo 'NO_PROXY=\"${NP}\"' >> /etc/environment" >/dev/null 2>&1
+    echo "[cua_gym_vm_bridge] no_proxy=${NP}"
+}
+
+fix_vm_network
+
+echo "[cua_gym_vm_bridge] direct-IP mode: ${MOCK_HOST} (+${MOCK_NET}) bypassed from woa proxy; no socat forwarding needed"