HassamGani commited on
Commit ·
6dc22bd
1
Parent(s): 1b52a9e
Ops queue: the person comes back about a side job after its team is done; oracle hires back and hands over what was found
Browse files- README.md +1 -1
- reorgbench/gen/build.py +4 -3
- reorgbench/gen/queue.py +49 -4
- reorgbench/gen/templates.py +4 -3
- reorgbench/oracle.py +26 -3
- tests/test_generator.py +15 -0
README.md
CHANGED
|
@@ -179,7 +179,7 @@ From the units, the generator derives per phase:
|
|
| 179 |
| `sla-credit` | an outage took clusters offline; mark every affected customer's SLA credit and tell each the amount; who was hit comes from the incident report, not the CRM | CANCEL (partner-channel customers), MODIFY (higher SLA band), RELAX, ADD (goodwill credit, close) |
|
| 180 |
| `price-increase` | move every customer on a plan to the new price version and tell each their new monthly price; price-protection contracts, some in force, some expired | CANCEL (renewing this quarter), MODIFY (increase trimmed), RELAX, ADD (credit, close) |
|
| 181 |
| `data-retention` | archive accounts inactive past the retention period (a date comparison against today), with the archival vendor's batch reference; litigation holds | CANCEL (regulated data), MODIFY (new batch reference), RELAX, ADD (credit, close) |
|
| 182 |
-
| `ops-queue` | all three stories at once in one company: shared customers table, tickets, ledger, people and handbook; the person changes job 1 while the other two run. The other jobs are posted in a #ops-queue chat channel with their sign-offs, next to a fourth job that was never signed off by anyone who can, or was called off; its customers wait with open tickets and must be left alone | job 1's changes; every side job's facts, tags and checks namespaced, one ledger guard over all three |
|
| 183 |
|
| 184 |
The test split is generated from templates registered at run time with `register_template`, outside this repo, so whole scenarios are held out.
|
| 185 |
|
|
|
|
| 179 |
| `sla-credit` | an outage took clusters offline; mark every affected customer's SLA credit and tell each the amount; who was hit comes from the incident report, not the CRM | CANCEL (partner-channel customers), MODIFY (higher SLA band), RELAX, ADD (goodwill credit, close) |
|
| 180 |
| `price-increase` | move every customer on a plan to the new price version and tell each their new monthly price; price-protection contracts, some in force, some expired | CANCEL (renewing this quarter), MODIFY (increase trimmed), RELAX, ADD (credit, close) |
|
| 181 |
| `data-retention` | archive accounts inactive past the retention period (a date comparison against today), with the archival vendor's batch reference; litigation holds | CANCEL (regulated data), MODIFY (new batch reference), RELAX, ADD (credit, close) |
|
| 182 |
+
| `ops-queue` | all three stories at once in one company: shared customers table, tickets, ledger, people and handbook; the person changes job 1 while the other two run. The other jobs are posted in a #ops-queue chat channel with their sign-offs, next to a fourth job that was never signed off by anyone who can, or was called off; its customers wait with open tickets and must be left alone. Once job 1's changes are in, the person comes back about job 2 (a correction, an extra credit or closing its tickets), when its team may already be done and let go | job 1's changes, then one to job 2; every side job's facts, tags and checks namespaced, one ledger guard over all three |
|
| 183 |
|
| 184 |
The test split is generated from templates registered at run time with `register_template`, outside this repo, so whole scenarios are held out.
|
| 185 |
|
reorgbench/gen/build.py
CHANGED
|
@@ -1037,8 +1037,9 @@ def runbook(g: Gen, steps: str) -> str:
|
|
| 1037 |
return f"{subject}: get it done the way the current runbook for it in our docs says."
|
| 1038 |
|
| 1039 |
|
| 1040 |
-
def make_task(g: Gen, attempt: int = 0) -> tuple[Task, Gen]:
|
| 1041 |
-
"""The plan half: phases, checks, changes and mutants over a finished world.
|
|
|
|
| 1042 |
template, seed, knobs = g.template, g.seed, g.knobs
|
| 1043 |
g.ledger_initial = {o for o in g.objects if o.startswith("ledger.entries/")}
|
| 1044 |
g.in_scope = list(g.cohort)
|
|
@@ -1054,7 +1055,7 @@ def make_task(g: Gen, attempt: int = 0) -> tuple[Task, Gen]:
|
|
| 1054 |
for k, kind in enumerate(knobs.changes, start=1):
|
| 1055 |
trigger = first_trigger(g) if k == 1 else NEVER # later ones are set from the oracle's runs in calibrate
|
| 1056 |
text, effects, changed = template.change(g, kind)
|
| 1057 |
-
if knobs.messy and k == len(knobs.changes) and kind != "CANCEL" and "refunded" not in g.units \
|
| 1058 |
and g.units["told"].rate is not None and template.truth_key:
|
| 1059 |
# stories whose notice is an amount, that need no invoice, and whose source of truth is billing, which
|
| 1060 |
# the new account's billing record then agrees with
|
|
|
|
| 1037 |
return f"{subject}: get it done the way the current runbook for it in our docs says."
|
| 1038 |
|
| 1039 |
|
| 1040 |
+
def make_task(g: Gen, attempt: int = 0, late: bool = True) -> tuple[Task, Gen]:
|
| 1041 |
+
"""The plan half: phases, checks, changes and mutants over a finished world. late=False keeps the last change
|
| 1042 |
+
from bringing in a late account (an ops queue's side job, whose company already has one)."""
|
| 1043 |
template, seed, knobs = g.template, g.seed, g.knobs
|
| 1044 |
g.ledger_initial = {o for o in g.objects if o.startswith("ledger.entries/")}
|
| 1045 |
g.in_scope = list(g.cohort)
|
|
|
|
| 1055 |
for k, kind in enumerate(knobs.changes, start=1):
|
| 1056 |
trigger = first_trigger(g) if k == 1 else NEVER # later ones are set from the oracle's runs in calibrate
|
| 1057 |
text, effects, changed = template.change(g, kind)
|
| 1058 |
+
if late and knobs.messy and k == len(knobs.changes) and kind != "CANCEL" and "refunded" not in g.units \
|
| 1059 |
and g.units["told"].rate is not None and template.truth_key:
|
| 1060 |
# stories whose notice is an amount, that need no invoice, and whose source of truth is billing, which
|
| 1061 |
# the new account's billing record then agrees with
|
reorgbench/gen/queue.py
CHANGED
|
@@ -10,6 +10,7 @@ from __future__ import annotations
|
|
| 10 |
|
| 11 |
import copy
|
| 12 |
import random
|
|
|
|
| 13 |
from collections import Counter
|
| 14 |
from typing import Any
|
| 15 |
|
|
@@ -17,8 +18,8 @@ from ..oracle import _args_facts
|
|
| 17 |
from ..tasks.hand.common import all_unchanged, set_replies
|
| 18 |
from ..tasks.spec import Budgets, Checkpoint, HumanFact, Intervention, Mutant, OraclePhase, Phase, PlantedFact, Step, Task
|
| 19 |
from . import company
|
| 20 |
-
from .build import (ALL_ROLES, CI, HIRE_ORDER, LOOSE, SIGNOFF, GenerationError, Knobs, base_objects,
|
| 21 |
-
make_task, make_world, tier_of)
|
| 22 |
from .templates import TEMPLATES, Template
|
| 23 |
from .world import Ids
|
| 24 |
|
|
@@ -70,6 +71,16 @@ def _own_agents(task: Task, k: int) -> None:
|
|
| 70 |
o.reset))
|
| 71 |
|
| 72 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 73 |
def _ledger_guard(cps: list[Checkpoint], initial: set[str]) -> Checkpoint:
|
| 74 |
allows = [c.allows for c in cps if getattr(c, "allows", None)]
|
| 75 |
|
|
@@ -99,6 +110,11 @@ def sample(seed: int) -> list[tuple[Template, Knobs]]:
|
|
| 99 |
messy=True))]
|
| 100 |
jobs += [(t, Knobs(n=rng.choice([6, 8, 12]), crew=1, changes=(), traps=rng.randint(1, 3), code=False, budget=2.0,
|
| 101 |
messy=True)) for t in side]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 102 |
return jobs
|
| 103 |
|
| 104 |
|
|
@@ -136,7 +152,7 @@ def build_queue(seed: int, attempt: int = 0) -> tuple[Task, list]:
|
|
| 136 |
gens = [make_world(t, seed, k, attempt, objects, ids) for t, k in jobs]
|
| 137 |
decoy = _decoy(seed, attempt, objects, ids, gens)
|
| 138 |
gens[0].template.shuffle(gens[0])
|
| 139 |
-
tasks = [make_task(g, attempt)[0] for g in gens]
|
| 140 |
main, side = tasks[0], tasks[1:]
|
| 141 |
gens[0].template.finish_messy(gens[0], gens[1:], [company.QUEUE_RULE]) # one handbook with every job's rules
|
| 142 |
who = company.people(gens[0])
|
|
@@ -155,7 +171,9 @@ def build_queue(seed: int, attempt: int = 0) -> tuple[Task, list]:
|
|
| 155 |
side_cps: list[Checkpoint] = []
|
| 156 |
for t in side:
|
| 157 |
side_roles += Counter(r for st in t.phases[0].staffings[:1] for r in st)
|
| 158 |
-
side_cps += t.phases[
|
|
|
|
|
|
|
| 159 |
initial = {o for o in objects if o.startswith("ledger.entries/")}
|
| 160 |
about = {id(t): g.story["subject"] for t, g in zip(tasks, gens)}
|
| 161 |
queue = (" Also work through the jobs posted in the #ops-queue chat channel, doing the ones that are signed off as "
|
|
@@ -203,6 +221,33 @@ def build_queue(seed: int, attempt: int = 0) -> tuple[Task, list]:
|
|
| 203 |
phases.append(Phase(goal, cps, staffings, o))
|
| 204 |
ivs = [Intervention(iv.id, iv.change, f"About the {about[id(main)]}: {iv.text}", iv.trigger,
|
| 205 |
iv.fallback_round, iv.min_gap, iv.world_effects) for iv in main.interventions]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 206 |
human = [SIGNOFF] + ([CI] if any("code" in g.units for g in gens) else []) + [HumanFact(
|
| 207 |
"queue", "which jobs in the ops queue to do, sign-off, other requests in #ops-queue",
|
| 208 |
"Do the jobs in #ops-queue that the VP of Customer Success or the CFO signed off, unless one of them called "
|
|
|
|
| 10 |
|
| 11 |
import copy
|
| 12 |
import random
|
| 13 |
+
import re
|
| 14 |
from collections import Counter
|
| 15 |
from typing import Any
|
| 16 |
|
|
|
|
| 18 |
from ..tasks.hand.common import all_unchanged, set_replies
|
| 19 |
from ..tasks.spec import Budgets, Checkpoint, HumanFact, Intervention, Mutant, OraclePhase, Phase, PlantedFact, Step, Task
|
| 20 |
from . import company
|
| 21 |
+
from .build import (ALL_ROLES, CI, HIRE_ORDER, LOOSE, NEVER, SIGNOFF, GenerationError, Knobs, base_objects,
|
| 22 |
+
calibrate, make_task, make_world, tier_of)
|
| 23 |
from .templates import TEMPLATES, Template
|
| 24 |
from .world import Ids
|
| 25 |
|
|
|
|
| 71 |
o.reset))
|
| 72 |
|
| 73 |
|
| 74 |
+
_READS = ("query", "list", "read", "search", "lookup", "accounts", "credit_history", "list_charges", "tables")
|
| 75 |
+
|
| 76 |
+
|
| 77 |
+
def _does_work(s: Step, known: set[str]) -> bool:
|
| 78 |
+
"""A step worth hiring someone back for: not a ghost, and it changes something or finds something new."""
|
| 79 |
+
if (s.if_done or "").split(":")[-1] in ("ghost", "owed"):
|
| 80 |
+
return False
|
| 81 |
+
return not s.tool.split(".")[-1].startswith(_READS) or any(n not in known for n in s.provides)
|
| 82 |
+
|
| 83 |
+
|
| 84 |
def _ledger_guard(cps: list[Checkpoint], initial: set[str]) -> Checkpoint:
|
| 85 |
allows = [c.allows for c in cps if getattr(c, "allows", None)]
|
| 86 |
|
|
|
|
| 110 |
messy=True))]
|
| 111 |
jobs += [(t, Knobs(n=rng.choice([6, 8, 12]), crew=1, changes=(), traps=rng.randint(1, 3), code=False, budget=2.0,
|
| 112 |
messy=True)) for t in side]
|
| 113 |
+
# the person comes back about the first side job once the main job's changes are in: by then its team may be
|
| 114 |
+
# done and gone, so the change has to find its way to whoever can still act on it, or to someone hired back
|
| 115 |
+
late = [c for c in ("MODIFY", "ADD:credit", "ADD:close") if c in side[0].pool and c not in jobs[0][1].changes]
|
| 116 |
+
t, k = jobs[1]
|
| 117 |
+
jobs[1] = (t, Knobs(n=k.n, crew=1, changes=(rng.choice(late),), traps=k.traps, code=False, budget=2.0, messy=True))
|
| 118 |
return jobs
|
| 119 |
|
| 120 |
|
|
|
|
| 152 |
gens = [make_world(t, seed, k, attempt, objects, ids) for t, k in jobs]
|
| 153 |
decoy = _decoy(seed, attempt, objects, ids, gens)
|
| 154 |
gens[0].template.shuffle(gens[0])
|
| 155 |
+
tasks = [make_task(g, attempt, late=j == 0)[0] for j, g in enumerate(gens)]
|
| 156 |
main, side = tasks[0], tasks[1:]
|
| 157 |
gens[0].template.finish_messy(gens[0], gens[1:], [company.QUEUE_RULE]) # one handbook with every job's rules
|
| 158 |
who = company.people(gens[0])
|
|
|
|
| 171 |
side_cps: list[Checkpoint] = []
|
| 172 |
for t in side:
|
| 173 |
side_roles += Counter(r for st in t.phases[0].staffings[:1] for r in st)
|
| 174 |
+
side_cps += t.phases[0].checkpoints
|
| 175 |
+
changed = [t for t in side if t.interventions]
|
| 176 |
+
final_cps = [c for t in side for c in t.phases[-1].checkpoints]
|
| 177 |
initial = {o for o in objects if o.startswith("ledger.entries/")}
|
| 178 |
about = {id(t): g.story["subject"] for t, g in zip(tasks, gens)}
|
| 179 |
queue = (" Also work through the jobs posted in the #ops-queue chat channel, doing the ones that are signed off as "
|
|
|
|
| 221 |
phases.append(Phase(goal, cps, staffings, o))
|
| 222 |
ivs = [Intervention(iv.id, iv.change, f"About the {about[id(main)]}: {iv.text}", iv.trigger,
|
| 223 |
iv.fallback_round, iv.min_gap, iv.world_effects) for iv in main.interventions]
|
| 224 |
+
for t in changed:
|
| 225 |
+
p = t.phases[1]
|
| 226 |
+
cps = [c for c in main.phases[-1].checkpoints if c.id != "no_stray_ledger_entries"]
|
| 227 |
+
cps += [c for c in final_cps if c.id != "no_stray_ledger_entries"]
|
| 228 |
+
cps += [_ledger_guard([*main.phases[-1].checkpoints, *final_cps], initial), held_back]
|
| 229 |
+
o = copy.copy(p.oracle)
|
| 230 |
+
o.steps = {k: list(v) for k, v in p.oracle.steps.items()}
|
| 231 |
+
# its team was let go once its part was done: whoever the change needs is hired back, briefed with what
|
| 232 |
+
# the first team found, which the orchestrator kept from their reports
|
| 233 |
+
known = {n for q in phases for st in q.oracle.steps.values() for x in st for n in x.provides}
|
| 234 |
+
real = [l for l, st in o.steps.items() if any(_does_work(x, known) for x in st)]
|
| 235 |
+
o.hires = [l for l in dict.fromkeys([*o.hires, *real]) if l not in alive]
|
| 236 |
+
need = sorted({n for st in o.steps.values() for s in st for n in _args_facts(s.args)})
|
| 237 |
+
if need:
|
| 238 |
+
o.orchestrator_steps = [*o.orchestrator_steps, Step("docs.read", {"doc_id": ["{" + n + "}" for n in need]},
|
| 239 |
+
if_done="never:ghost")]
|
| 240 |
+
o.retire = [l for l in o.retire if l in alive or l in o.hires]
|
| 241 |
+
o.retire_now = [l for l in o.retire_now if l in alive]
|
| 242 |
+
o.reset = [l for l in o.reset if l in alive]
|
| 243 |
+
alive = (alive | set(o.hires)) - set(o.retire) - set(o.retire_now)
|
| 244 |
+
iv = t.interventions[0]
|
| 245 |
+
text = f"About the {about[id(t)]}: {iv.text}"
|
| 246 |
+
# the reference team: the main job's, plus the side agents still on once the change is done
|
| 247 |
+
side_end = Counter(l.split("#")[0] for l in alive if re.search(r"#j\d+$", l))
|
| 248 |
+
team = [sorted((Counter(a) + side_end).elements()) for a in main.phases[-1].staffings]
|
| 249 |
+
phases.append(Phase(phases[-1].goal + " " + text, cps, team, o))
|
| 250 |
+
ivs.append(Intervention(iv.id, iv.change, text, NEVER, None, iv.min_gap, iv.world_effects))
|
| 251 |
human = [SIGNOFF] + ([CI] if any("code" in g.units for g in gens) else []) + [HumanFact(
|
| 252 |
"queue", "which jobs in the ops queue to do, sign-off, other requests in #ops-queue",
|
| 253 |
"Do the jobs in #ops-queue that the VP of Customer Success or the CFO signed off, unless one of them called "
|
reorgbench/gen/templates.py
CHANGED
|
@@ -294,9 +294,10 @@ class Template:
|
|
| 294 |
def finish_messy(self, g: Gen, also: list[Gen] = (), rules_too: list[str] = ()) -> None:
|
| 295 |
"""The operations handbook, written once the changes (and so the credit cap) are known. In a shared
|
| 296 |
company, also lists the other workstreams, whose own rules go in too."""
|
| 297 |
-
|
|
|
|
| 298 |
new = cr.cap if cr is not None and cr.cap is not None else g.rng.choice([40, 60, 80])
|
| 299 |
-
cap =
|
| 300 |
rules = [
|
| 301 |
"Expenses over $500 need a purchase order.",
|
| 302 |
"Answer customer tickets within two business days, in plain language.",
|
|
@@ -328,7 +329,7 @@ class Template:
|
|
| 328 |
"(policy-updates); where the two differ, the update wins.")
|
| 329 |
g.objects["docs/operations-handbook"] = {"title": "Operations handbook", "text": " ".join(
|
| 330 |
f"{i}. {r}" for i, r in enumerate(rules, 1))}
|
| 331 |
-
fut =
|
| 332 |
updates = [
|
| 333 |
("2026-04-20", "Purchasing", "Expenses over $750 now need a purchase order (was $500)."),
|
| 334 |
("2026-05-11", "Security", "Rotate credentials every 60 days (was 90)."),
|
|
|
|
| 294 |
def finish_messy(self, g: Gen, also: list[Gen] = (), rules_too: list[str] = ()) -> None:
|
| 295 |
"""The operations handbook, written once the changes (and so the credit cap) are known. In a shared
|
| 296 |
company, also lists the other workstreams, whose own rules go in too."""
|
| 297 |
+
src = next((x for x in [g, *also] if x.units.get("credited")), g) # the job whose credits the caps govern
|
| 298 |
+
cr = src.units.get("credited")
|
| 299 |
new = cr.cap if cr is not None and cr.cap is not None else g.rng.choice([40, 60, 80])
|
| 300 |
+
cap = src.story.get("cap_old", new + 20)
|
| 301 |
rules = [
|
| 302 |
"Expenses over $500 need a purchase order.",
|
| 303 |
"Answer customer tickets within two business days, in plain language.",
|
|
|
|
| 329 |
"(policy-updates); where the two differ, the update wins.")
|
| 330 |
g.objects["docs/operations-handbook"] = {"title": "Operations handbook", "text": " ".join(
|
| 331 |
f"{i}. {r}" for i, r in enumerate(rules, 1))}
|
| 332 |
+
fut = src.story.get("cap_future", new + 30)
|
| 333 |
updates = [
|
| 334 |
("2026-04-20", "Purchasing", "Expenses over $750 now need a purchase order (was $500)."),
|
| 335 |
("2026-05-11", "Security", "Rotate credentials every 60 days (was 90)."),
|
reorgbench/oracle.py
CHANGED
|
@@ -25,6 +25,7 @@ _DONE = re.compile(r"^DONE PHASE (\d+)$", re.M)
|
|
| 25 |
_RESET = re.compile(r"^RESET$", re.M)
|
| 26 |
_ROLE = re.compile(r"^ROLE ([\w#]+)$", re.M)
|
| 27 |
_TEAM = re.compile(r"^TEAM ([\w#]+)=(\w+)$", re.M)
|
|
|
|
| 28 |
|
| 29 |
|
| 30 |
def role_of_label(label: str) -> str:
|
|
@@ -197,6 +198,8 @@ class OracleWorker:
|
|
| 197 |
m = _ROLE.search(text)
|
| 198 |
if m and not self.started:
|
| 199 |
self.label = m.group(1)
|
|
|
|
|
|
|
| 200 |
for name, raw in _FACT.findall(text):
|
| 201 |
self.run.facts.setdefault(name, json.loads(raw))
|
| 202 |
if sender is not None:
|
|
@@ -226,7 +229,7 @@ class OracleWorker:
|
|
| 226 |
for role in sorted(self.ctl.consumers(name, self.phase)):
|
| 227 |
if role == self.label:
|
| 228 |
continue
|
| 229 |
-
for aid in self.team.get(role, []):
|
| 230 |
if (name, aid) not in self.sent:
|
| 231 |
out.setdefault(aid, []).append(name)
|
| 232 |
return out
|
|
@@ -254,7 +257,9 @@ class OracleWorker:
|
|
| 254 |
return act
|
| 255 |
if self.run.idle and self.reported < self.phase:
|
| 256 |
self.reported = self.phase
|
| 257 |
-
|
|
|
|
|
|
|
| 258 |
return {"type": "noop"}
|
| 259 |
|
| 260 |
|
|
@@ -271,8 +276,11 @@ class OracleOrchestrator:
|
|
| 271 |
self.spawns: list[str] = []
|
| 272 |
self.pending_spawns = 0
|
| 273 |
self.pending_labels: list[str] = []
|
|
|
|
|
|
|
| 274 |
self.instruct: list[str] = []
|
| 275 |
self.resets: list[str] = []
|
|
|
|
| 276 |
self._enter_phase(0)
|
| 277 |
|
| 278 |
def _enter_phase(self, k: int) -> None:
|
|
@@ -305,12 +313,16 @@ class OracleOrchestrator:
|
|
| 305 |
k, c = it["kind"], it["content"]
|
| 306 |
if k == "spawned":
|
| 307 |
self.role_of[c["agent_id"]] = self.pending_labels.pop(0)
|
|
|
|
| 308 |
self.alive.append(c["agent_id"])
|
| 309 |
self.assigned[c["agent_id"]] = self.phase
|
| 310 |
self.pending_spawns -= 1
|
| 311 |
elif k == "message":
|
| 312 |
for d in _DONE.findall(c["text"]):
|
| 313 |
self.done.add((c["from"], int(d)))
|
|
|
|
|
|
|
|
|
|
| 314 |
for name, raw in _FACT.findall(c["text"]):
|
| 315 |
self.run.facts.setdefault(name, json.loads(raw))
|
| 316 |
elif k == "result":
|
|
@@ -336,14 +348,25 @@ class OracleOrchestrator:
|
|
| 336 |
self.pending_spawns += 1
|
| 337 |
self.pending_labels.append(role)
|
| 338 |
needed = self.ctl.needs(role, self.phase)
|
|
|
|
| 339 |
leaving = {a for a, _ in self.kills}
|
| 340 |
team = [a for a in self.alive if a not in leaving]
|
|
|
|
|
|
|
|
|
|
| 341 |
brief = "\n".join(filter(None, [f"PHASE {self.phase}", f"ROLE {role}", self._team_lines(team),
|
| 342 |
-
fact_lines(self.run.facts, needed)]))
|
| 343 |
return {"type": "spawn", "role": role_of_label(role), "brief": brief, "intro": team}
|
| 344 |
if self.pending_spawns:
|
| 345 |
return {"type": "noop"}
|
| 346 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 347 |
while self.instruct: # an agent already let go needs no phase notice
|
| 348 |
msg = self._instruct(self.instruct.pop(0), reset=False)
|
| 349 |
if msg["to"]:
|
|
|
|
| 25 |
_RESET = re.compile(r"^RESET$", re.M)
|
| 26 |
_ROLE = re.compile(r"^ROLE ([\w#]+)$", re.M)
|
| 27 |
_TEAM = re.compile(r"^TEAM ([\w#]+)=(\w+)$", re.M)
|
| 28 |
+
_TAGS = re.compile(r"^TAGS (\S+)$", re.M) # what an earlier agent with this label already did
|
| 29 |
|
| 30 |
|
| 31 |
def role_of_label(label: str) -> str:
|
|
|
|
| 198 |
m = _ROLE.search(text)
|
| 199 |
if m and not self.started:
|
| 200 |
self.label = m.group(1)
|
| 201 |
+
for tags in _TAGS.findall(text):
|
| 202 |
+
self.run.done_tags |= set(tags.split(","))
|
| 203 |
for name, raw in _FACT.findall(text):
|
| 204 |
self.run.facts.setdefault(name, json.loads(raw))
|
| 205 |
if sender is not None:
|
|
|
|
| 229 |
for role in sorted(self.ctl.consumers(name, self.phase)):
|
| 230 |
if role == self.label:
|
| 231 |
continue
|
| 232 |
+
for aid in ([self.info.parent] if role == "orchestrator" else self.team.get(role, [])):
|
| 233 |
if (name, aid) not in self.sent:
|
| 234 |
out.setdefault(aid, []).append(name)
|
| 235 |
return out
|
|
|
|
| 257 |
return act
|
| 258 |
if self.run.idle and self.reported < self.phase:
|
| 259 |
self.reported = self.phase
|
| 260 |
+
tags = ",".join(sorted(self.run.done_tags))
|
| 261 |
+
return {"type": "msg", "to": [self.info.parent],
|
| 262 |
+
"content": f"DONE PHASE {self.phase}" + (f"\nTAGS {tags}" if tags else "")}
|
| 263 |
return {"type": "noop"}
|
| 264 |
|
| 265 |
|
|
|
|
| 276 |
self.spawns: list[str] = []
|
| 277 |
self.pending_spawns = 0
|
| 278 |
self.pending_labels: list[str] = []
|
| 279 |
+
self.pending_facts: list[list[str]] = []
|
| 280 |
+
self.tags: dict[str, set[str]] = {} # label -> what its agents reported done, for anyone hired back to it
|
| 281 |
self.instruct: list[str] = []
|
| 282 |
self.resets: list[str] = []
|
| 283 |
+
self.passed: set[tuple[str, str]] = set() # (fact, agent) the orchestrator has handed on
|
| 284 |
self._enter_phase(0)
|
| 285 |
|
| 286 |
def _enter_phase(self, k: int) -> None:
|
|
|
|
| 313 |
k, c = it["kind"], it["content"]
|
| 314 |
if k == "spawned":
|
| 315 |
self.role_of[c["agent_id"]] = self.pending_labels.pop(0)
|
| 316 |
+
self.passed |= {(n, c["agent_id"]) for n in self.pending_facts.pop(0)}
|
| 317 |
self.alive.append(c["agent_id"])
|
| 318 |
self.assigned[c["agent_id"]] = self.phase
|
| 319 |
self.pending_spawns -= 1
|
| 320 |
elif k == "message":
|
| 321 |
for d in _DONE.findall(c["text"]):
|
| 322 |
self.done.add((c["from"], int(d)))
|
| 323 |
+
for tags in _TAGS.findall(c["text"]):
|
| 324 |
+
if c["from"] in self.role_of:
|
| 325 |
+
self.tags.setdefault(self.role_of[c["from"]], set()).update(tags.split(","))
|
| 326 |
for name, raw in _FACT.findall(c["text"]):
|
| 327 |
self.run.facts.setdefault(name, json.loads(raw))
|
| 328 |
elif k == "result":
|
|
|
|
| 348 |
self.pending_spawns += 1
|
| 349 |
self.pending_labels.append(role)
|
| 350 |
needed = self.ctl.needs(role, self.phase)
|
| 351 |
+
self.pending_facts.append([n for n in needed if n in self.run.facts])
|
| 352 |
leaving = {a for a, _ in self.kills}
|
| 353 |
team = [a for a in self.alive if a not in leaving]
|
| 354 |
+
# only what its own steps branch on: what it already did, not a list of everyone it touched
|
| 355 |
+
used = {t for st in self.ctl.steps(self.phase, role) for t in (st.if_done, st.unless_done) if t}
|
| 356 |
+
done = ",".join(sorted(self.tags.get(role, set()) & used))
|
| 357 |
brief = "\n".join(filter(None, [f"PHASE {self.phase}", f"ROLE {role}", self._team_lines(team),
|
| 358 |
+
fact_lines(self.run.facts, needed), done and f"TAGS {done}"]))
|
| 359 |
return {"type": "spawn", "role": role_of_label(role), "brief": brief, "intro": team}
|
| 360 |
if self.pending_spawns:
|
| 361 |
return {"type": "noop"}
|
| 362 |
|
| 363 |
+
for a in self.alive: # a fact a teammate reported that an agent hired alongside it still needs
|
| 364 |
+
owed = [n for n in self.ctl.needs(self.role_of[a], self.phase) if n in self.run.facts
|
| 365 |
+
and (n, a) not in self.passed]
|
| 366 |
+
if owed:
|
| 367 |
+
self.passed |= {(n, a) for n in owed}
|
| 368 |
+
return {"type": "msg", "to": [a], "content": fact_lines(self.run.facts, owed)}
|
| 369 |
+
|
| 370 |
while self.instruct: # an agent already let go needs no phase notice
|
| 371 |
msg = self._instruct(self.instruct.pop(0), reset=False)
|
| 372 |
if msg["to"]:
|
tests/test_generator.py
CHANGED
|
@@ -183,6 +183,21 @@ def test_ops_queue_runs_every_story_at_once():
|
|
| 183 |
assert "decoy_untouched" in ids # a posted job that isn't signed off
|
| 184 |
posts = [d for o, d in t.initial_objects.items() if o.startswith("chat/") and d["channel"] == "ops-queue"]
|
| 185 |
assert len(posts) == 3 and "#ops-queue" in t.phases[0].goal
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 186 |
|
| 187 |
|
| 188 |
def test_messy_overcharge_refunds_are_summed_from_invoice_lines():
|
|
|
|
| 183 |
assert "decoy_untouched" in ids # a posted job that isn't signed off
|
| 184 |
posts = [d for o, d in t.initial_objects.items() if o.startswith("chat/") and d["channel"] == "ops-queue"]
|
| 185 |
assert len(posts) == 3 and "#ops-queue" in t.phases[0].goal
|
| 186 |
+
# the person comes back about a side job last, after its team may have finished
|
| 187 |
+
side = t.title.split(": ", 1)[1].split(", ")[1]
|
| 188 |
+
assert t.interventions[-1].text.startswith(f"About the {side}: ") and len(t.phases) == len(t.interventions) + 1
|
| 189 |
+
before = {c.id for c in t.phases[-2].checkpoints if c.id.startswith("j1_")}
|
| 190 |
+
assert {c.id for c in t.phases[-1].checkpoints if c.id.startswith("j1_")} - before
|
| 191 |
+
|
| 192 |
+
|
| 193 |
+
@pytest.mark.parametrize("seed", range(6))
|
| 194 |
+
def test_queue_handbook_caps_are_the_credited_jobs(seed):
|
| 195 |
+
from reorgbench.gen.queue import build_queue
|
| 196 |
+
t, gens = build_queue(seed)
|
| 197 |
+
cr = [g.units["credited"] for g in gens if "credited" in g.units]
|
| 198 |
+
assert len(cr) <= 1
|
| 199 |
+
if cr and cr[0].cap is not None:
|
| 200 |
+
assert f"cap is now ${cr[0].cap} per account" in t.initial_objects["docs/policy-updates"]["text"]
|
| 201 |
|
| 202 |
|
| 203 |
def test_messy_overcharge_refunds_are_summed_from_invoice_lines():
|