Download dev/code_HarnessRouter__harnessrouter.txt from ag2003/length-extrapolating-architecture-data: direct link, hf CLI and curl.
- Browser
- Download file 111 kB
-
https://huggingface.co/datasets/ag2003/length-extrapolating-architecture-data/resolve/main/dev/code_HarnessRouter__harnessrouter.txt
- Command line
-
hf download hf://datasets/ag2003/length-extrapolating-architecture-data/dev/code_HarnessRouter__harnessrouter.txt
-
curl -L -o code_HarnessRouter__harnessrouter.txt https://huggingface.co/datasets/ag2003/length-extrapolating-architecture-data/resolve/main/dev/code_HarnessRouter__harnessrouter.txt
111 kB
# ===== file: gateway/backing.py =====
"""Pluggable backing stores for the harness gateway.
The gateway needs three capabilities from the outside world, and nothing else:
GraphStore — session/harness/response/api-key records + ownership edges
BlobStore — traces, response records, workspace checkpoints, produced files
SecretStore — model-provider connections/policies ("bring your own keys")
Two implementations of each:
local (this file) — SQLite + filesystem + env/file, zero cloud dependencies. The whole
gateway runs on a laptop: HR_BACKING=local (or just leave
VG_GATEWAY_URL unset). Point POOL_MGMT_ENDPOINT at a single runner
process (the pool is only an identifier-routing proxy in front of
runner replicas — the routes are identical).
vg (backing_vg.py) — the production AgentStudio infra: vg-gateway (Cosmos graph + Azure
blob) and the vault service. Verbatim the code that always ran here.
Selection is one env var: HR_BACKING=vg|local; unset = vg when VG_GATEWAY_URL is set, local
otherwise. The interface is deliberately SEMANTIC (get/upsert/find/add_edge), not query-shaped:
call sites never build backend query strings, so an implementation is free to be a graph, a
relational table, or anything else that can look up records by label + property equality.
"""
from __future__ import annotations
import asyncio
import base64
import contextlib
import hashlib
import secrets as secrets_mod
import io
import json
import mimetypes
import os
import posixpath
import re
import shutil
import sqlite3
import tarfile
import time
import uuid
from pathlib import Path
from typing import Protocol
import httpx
# ── interfaces ────────────────────────────────────────────────────────────────────
class GraphStore(Protocol):
async def get(self, vid: str, label: str | None = None) -> dict | None: ...
async def upsert(self, label: str, vid: str, props: dict, *, raise_on_fail: bool = False) -> None: ...
async def add_edge(self, label: str, src: str, dst: str) -> None: ...
async def find(self, label: str, eq: dict | None = None, neq: dict | None = None) -> list[dict]: ...
async def delete(self, vid: str, label: str | None = None) -> bool:
"""Remove one vertex and its edges; False when there was none (or the label differs)."""
...
async def probe(self) -> None:
"""Raise if the store is unreachable/misconfigured. Used by /readyz, which must FAIL on a
broken plane — the normal read paths swallow errors and return empty."""
...
class BlobStore(Protocol):
async def get(self, kb: str, file_id: str) -> bytes | None: ...
async def put(self, kb: str, file_id: str, data: bytes) -> bool: ...
async def delete(self, kb: str, file_id: str) -> bool: ...
async def list(self, kb: str, prefix: str, limit: int = 20, cursor: str | None = None) -> dict: ...
async def probe(self) -> None: ...
class SecretStore(Protocol):
async def get(self, tenant: str, name: str) -> str | None: ...
async def delete(self, tenant: str, name: str) -> None:
"""Forget a stored secret; nothing happens when there is none."""
...
async def put(self, tenant: str, name: str, value: str, *,
require_encryption: bool = False) -> None:
"""`require_encryption=True` is the caller saying "this is a customer credential, not a
provider key injected by the operator" — a database connection string, for instance. A
store that cannot encrypt must raise SecretsNotConfigured rather than write it in the
clear. Part of the interface, not of one implementation: a datasource must be refused
the same way whichever backing is configured."""
...
class WorkspaceFiles(Protocol):
"""Reading and writing one file inside a session's workspace.
Two environments, genuinely different storage, one interface:
* **Self-hosted** — the runner shares this container and the workspace is a live directory on
the data volume. A write is a write.
* **Hosted** — sandboxes are ephemeral, so between turns a workspace exists ONLY as the
checkpoint tarball in blob storage (`sessions/<sid>/workspace.tgz`). A write has to rewrite
that archive, because there is no filesystem to write to until the next turn hydrates one.
This exists so an app built on a Harness can own state the agent also edits — the slides kit
keeps deck.json here — without either side needing to know which deployment it is running in.
"""
async def read(self, sid: str, path: str) -> bytes | None: ...
async def write(self, sid: str, path: str, data: bytes) -> bool: ...
class Backing:
def __init__(self, graph: GraphStore, blob: BlobStore, secrets: SecretStore, mode: str,
workspace: "WorkspaceFiles"):
self.graph, self.blob, self.secrets, self.mode = graph, blob, secrets, mode
self.workspace = workspace
# ── local: SQLite graph ───────────────────────────────────────────────────────────
class SqliteGraphStore:
"""Vertices/edges in one SQLite file. VG upsert semantics preserved: property() sets the
listed props and keeps the rest (merge, not replace); edge ids are the same uuid5(src|label|dst)
so re-adds are idempotent. find() filters by property equality in Python — dev-scale data,
no json_extract dialect games."""
def __init__(self, path: str):
self._path = path
Path(path).parent.mkdir(parents=True, exist_ok=True)
with self._conn() as c:
c.execute("CREATE TABLE IF NOT EXISTS vertices (vid TEXT PRIMARY KEY, label TEXT NOT NULL, props TEXT NOT NULL)")
c.execute("CREATE INDEX IF NOT EXISTS idx_vertices_label ON vertices(label)")
c.execute("CREATE TABLE IF NOT EXISTS edges (eid TEXT PRIMARY KEY, label TEXT NOT NULL, src TEXT NOT NULL, dst TEXT NOT NULL)")
def _conn(self) -> sqlite3.Connection:
c = sqlite3.connect(self._path, timeout=30)
c.execute("PRAGMA journal_mode=WAL")
return c
async def get(self, vid: str, label: str | None = None) -> dict | None:
def _do():
with self._conn() as c:
row = c.execute("SELECT vid, label, props FROM vertices WHERE vid=?", (vid,)).fetchone()
if not row:
return None
if label is not None and row[1] != label:
return None # label filter is load-bearing where the id is caller-controlled
return {"id": row[0], **json.loads(row[2])}
return await asyncio.to_thread(_do)
async def upsert(self, label: str, vid: str, props: dict, *, raise_on_fail: bool = False) -> None:
clean = {k: str(v) for k, v in props.items()} # VG stores string props; keep parity
def _do():
with self._conn() as c:
row = c.execute("SELECT props FROM vertices WHERE vid=?", (vid,)).fetchone()
merged = {**(json.loads(row[0]) if row else {}), **clean}
c.execute("INSERT INTO vertices(vid,label,props) VALUES(?,?,?) "
"ON CONFLICT(vid) DO UPDATE SET props=excluded.props",
(vid, label, json.dumps(merged)))
try:
await asyncio.to_thread(_do)
except Exception:
if raise_on_fail:
raise RuntimeError("graph write failed")
async def add_edge(self, label: str, src: str, dst: str) -> None:
if not (src and dst):
return
eid = str(uuid.uuid5(uuid.NAMESPACE_URL, f"{src}|{label}|{dst}"))
def _do():
with self._conn() as c:
c.execute("INSERT OR IGNORE INTO edges(eid,label,src,dst) VALUES(?,?,?,?)",
(eid, label, src, dst))
try:
await asyncio.to_thread(_do)
except Exception: # noqa: BLE001 — best-effort, matches VG edge semantics
pass
async def delete(self, vid: str, label: str | None = None) -> bool:
def _do():
with self._conn() as c:
row = c.execute("SELECT label FROM vertices WHERE vid=?", (vid,)).fetchone()
if not row or (label is not None and row[0] != label):
return False
c.execute("DELETE FROM vertices WHERE vid=?", (vid,))
c.execute("DELETE FROM edges WHERE src=? OR dst=?", (vid, vid))
return True
try:
return await asyncio.to_thread(_do)
except Exception: # noqa: BLE001
return False
async def find(self, label: str, eq: dict | None = None, neq: dict | None = None) -> list[dict]:
def _do():
with self._conn() as c:
rows = c.execute("SELECT vid, props FROM vertices WHERE label=?", (label,)).fetchall()
out = []
for vid, props_json in rows:
p = json.loads(props_json)
if eq and any(str(p.get(k, "")) != str(v) for k, v in eq.items()):
continue
if neq and any(str(p.get(k, "")) == str(v) for k, v in neq.items()):
continue
out.append({"id": vid, **p})
return out
try:
return await asyncio.to_thread(_do)
except Exception: # noqa: BLE001
return []
async def probe(self) -> None:
def _do():
with self._conn() as c:
c.execute("SELECT 1 FROM vertices LIMIT 1").fetchone()
await asyncio.to_thread(_do)
# ── local: filesystem blobs ───────────────────────────────────────────────────────
class FileBlobStore:
"""Objects as plain files under {root}/{kb}/{file_id}. Listing is ascending-lexical over the
relative keys — the SAME ordering contract the trace machinery relies on from Azure blob
listing (inverted-timestamp keys ⇒ newest-first)."""
def __init__(self, root: str):
self._root = Path(root)
def _p(self, kb: str, file_id: str) -> Path:
p = (self._root / kb / file_id).resolve()
base = (self._root / kb).resolve()
if not str(p).startswith(str(base)): # refuse path escape
raise ValueError(f"blob id escapes store: {file_id!r}")
return p
async def get(self, kb: str, file_id: str) -> bytes | None:
def _do():
try:
return self._p(kb, file_id).read_bytes()
except (OSError, ValueError):
return None
return await asyncio.to_thread(_do)
async def put(self, kb: str, file_id: str, data: bytes) -> bool:
if not data:
return False
def _do():
try:
p = self._p(kb, file_id)
p.parent.mkdir(parents=True, exist_ok=True)
tmp = p.with_suffix(p.suffix + ".tmp")
tmp.write_bytes(data)
tmp.replace(p) # atomic: a torn write never replaces a good blob
return True
except (OSError, ValueError):
return False
return await asyncio.to_thread(_do)
async def delete(self, kb: str, file_id: str) -> bool:
def _do():
try:
path = self._p(kb, file_id)
path.unlink(missing_ok=True)
except (OSError, ValueError):
return False
# The directories a key's slashes created go with the last object in them, up to the
# store root. A blob store has no directories; this one only borrows them from the file
# system, and left alone they outlive every object: hr-test held 3,696 empty
# sessions/<sid>/ shells on 2026-09-16, one per session ever deleted. rmdir refuses a
# directory that still holds anything, which is the whole check.
base = (self._root / kb).resolve()
for parent in path.parents:
if parent == base or base not in parent.parents:
break
try:
parent.rmdir()
except OSError:
break
return True
return await asyncio.to_thread(_do)
async def purge(self, kb: str, prefix: str) -> None:
"""Remove everything under a folder-shaped prefix ("sessions/<sid>/"), including the
temporary files a torn write leaves and `list` does not show. A deleted session's folder
held a lone `changed.json.tmp` after every listed object was gone (hr-test copy,
2026-10-08)."""
def _do():
base = (self._root / kb).resolve()
if not prefix.endswith("/") or prefix.strip("/") == "":
return
top = (base / prefix).resolve()
if top == base or base not in top.parents or not top.is_dir():
return
shutil.rmtree(top, ignore_errors=True)
for parent in top.parents:
if parent == base or base not in parent.parents:
break
try:
parent.rmdir()
except OSError:
break
await asyncio.to_thread(_do)
async def list(self, kb: str, prefix: str, limit: int = 20, cursor: str | None = None) -> dict:
# Walk only the subtree the prefix names: everything before its last "/" is a directory
# under the store, the rest a name prefix inside it. Walking the whole store cost every
# blob on every call, and a paged listing repeated the walk per page: with 15k blobs on a
# self-hosted volume one turns read took 55 s and the gateway sat at 150% CPU (2026-09-06).
def _do():
base = (self._root / kb).resolve()
dir_part, _, _ = prefix.rpartition("/")
top = (base / dir_part).resolve() if dir_part else base
if not str(top).startswith(str(base)) or not top.is_dir(): # escape or nothing there
return {"items": [], "cursor": None}
root_len = len(str(base)) + 1
keys: list[str] = []
for root, _dirs, files in os.walk(top):
rel = root[root_len:].replace(os.sep, "/") if len(root) > root_len else ""
for name in files:
if name.endswith(".tmp"):
continue
key = f"{rel}/{name}" if rel else name
if key.startswith(prefix) and (not cursor or key > cursor):
keys.append(key)
keys.sort()
page = keys[:limit]
nxt = page[-1] if len(keys) > limit else None
return {"items": [{"file_id": k} for k in page], "cursor": nxt}
return await asyncio.to_thread(_do)
async def probe(self) -> None:
def _do():
self._root.mkdir(parents=True, exist_ok=True)
if not os.access(self._root, os.W_OK):
raise RuntimeError(f"blob root not writable: {self._root}")
await asyncio.to_thread(_do)
# ── local: env/file secrets ("bring your own provider keys") ──────────────────────
class SecretsNotConfigured(RuntimeError):
"""Raised when a secret must be written but there is nowhere safe to write it.
Deliberately an error rather than a fallback. A provider key injected read-only through the
environment is one risk; a customer's production database connection string written to a file
in plaintext is another entirely, and the difference is not something to decide silently on
the operator's behalf."""
class FileSecretStore:
"""Reads HR_SECRET_{TENANT}_{NAME} from the environment first (read-only injection — the
open-repo way to supply provider credentials without any vault), then falls back to files
under {root}/{tenant}/{name}.
Files are ENCRYPTED when HR_SECRET_KEY is set: AES-256-GCM under a key derived from it, with
a random nonce per write, stored as `hrenc1:<b64 nonce||ciphertext>`. Reads accept both forms,
so a store written before this existed keeps working and re-encrypts on its next write.
Without HR_SECRET_KEY, `put(..., require_encryption=True)` REFUSES. That is the whole point:
the caller that stores a database credential says so, and on an instance with no key it gets
an error telling the operator what to set, instead of a plaintext file nobody knew about."""
_MAGIC = "hrenc1:"
def __init__(self, root: str):
self._root = Path(root)
self._key = self._derive(os.environ.get("HR_SECRET_KEY", ""))
@property
def encrypts(self) -> bool:
return self._key is not None
@staticmethod
def _derive(passphrase: str) -> bytes | None:
"""A 32-byte key from the operator's passphrase. Scrypt with a fixed salt: the salt's job
is to stop cross-deployment rainbow tables, and there is nowhere to keep a random one that
would not sit beside the ciphertext anyway."""
if not passphrase:
return None
return hashlib.scrypt(passphrase.encode(), salt=b"harnessrouter.secret.v1",
n=2 ** 14, r=8, p=1, dklen=32)
def _seal(self, value: str) -> str:
from cryptography.hazmat.primitives.ciphers.aead import AESGCM
nonce = secrets_mod.token_bytes(12)
blob = nonce + AESGCM(self._key).encrypt(nonce, value.encode(), None)
return self._MAGIC + base64.b64encode(blob).decode()
def _open(self, raw: str) -> str:
if not raw.startswith(self._MAGIC):
return raw # written before encryption existed
if self._key is None:
raise SecretsNotConfigured(
"This secret is encrypted but HR_SECRET_KEY is not set. Set it to the passphrase "
"used when the secret was stored.")
from cryptography.hazmat.primitives.ciphers.aead import AESGCM
blob = base64.b64decode(raw[len(self._MAGIC):])
return AESGCM(self._key).decrypt(blob[:12], blob[12:], None).decode()
@staticmethod
def _env_key(tenant: str, name: str) -> str:
return "HR_SECRET_" + re.sub(r"[^A-Za-z0-9]", "_", f"{tenant}_{name}").upper()
def _p(self, tenant: str, name: str) -> Path:
safe = re.sub(r"[^A-Za-z0-9._-]", "_", name)
return self._root / re.sub(r"[^A-Za-z0-9._-]", "_", tenant) / safe
async def get(self, tenant: str, name: str) -> str | None:
v = os.environ.get(self._env_key(tenant, name))
if v:
return v
def _do():
try:
return self._p(tenant, name).read_text()
except OSError:
return None
raw = await asyncio.to_thread(_do)
return self._open(raw) if raw is not None else None
async def put(self, tenant: str, name: str, value: str, *, require_encryption: bool = False) -> None:
if require_encryption and self._key is None:
raise SecretsNotConfigured(
"Refusing to store this credential: HR_SECRET_KEY is not set, so it could only be "
"written to disk in plaintext. Set HR_SECRET_KEY to a passphrase and restart.")
body = self._seal(value) if self._key is not None else value
def _do():
p = self._p(tenant, name)
p.parent.mkdir(parents=True, exist_ok=True)
p.write_text(body)
with contextlib.suppress(OSError):
p.chmod(0o600) # belt to the encryption's braces
await asyncio.to_thread(_do)
async def delete(self, tenant: str, name: str) -> None:
def _do():
with contextlib.suppress(OSError):
self._p(tenant, name).unlink()
await asyncio.to_thread(_do)
async def blank(self) -> list[tuple[str, str]]:
"""Every stored secret whose value is empty, as (tenant, name). Before 0.32.0 a removed
credential was overwritten with nothing rather than deleted; the retention sweep removes
those files. A value that cannot be opened (encrypted, and no key) is not reported."""
def _do():
out: list[tuple[str, str]] = []
if not self._root.is_dir():
return out
for tdir in sorted(self._root.iterdir()):
if not tdir.is_dir():
continue
for f in sorted(tdir.iterdir()):
if not f.is_file():
continue
try:
if self._open(f.read_text()) == "":
out.append((tdir.name, f.name))
except Exception: # noqa: BLE001 — unreadable is not empty
continue
return out
return await asyncio.to_thread(_do)
# ── selection ─────────────────────────────────────────────────────────────────────
# ── workspace files ───────────────────────────────────────────────────────────────
def _safe_member(path: str) -> str | None:
"""A workspace-relative path, or None if it tries to leave the workspace.
Rejected rather than sanitised: a path that needed cleaning was not the path the caller meant,
and silently writing somewhere else is worse than refusing."""
p = (path or "").strip().lstrip("/")
if not p or p != posixpath.normpath(p) or p.startswith("../") or "\\" in p:
return None
return p
class RunnerWorkspaceFiles:
"""Self-hosted: a session's workspace is a live directory on the data volume, and the RUNNER is
the only principal that touches it. Behind the write-wall each session directory belongs to that
session's own uid and the product deliberately cannot read or write there (see _isolate_session
in the runner), so a live read or write goes through the runner, which acts as the session,
rather than through this process's own filesystem access. One door, whether the wall is up or
not: the runner is beside the gateway in every self-hosted deployment.
Live, unlike the checkpoint the listing route reads: an app sees a file the moment a tool call
creates it, rather than when the turn ends."""
def __init__(self, endpoint: str, key: str, blob=None, kb: str = ""):
self.endpoint = (endpoint or "").rstrip("/")
self.key = key
self.blob, self.kb = blob, kb
self._client: httpx.AsyncClient | None = None
def _c(self) -> httpx.AsyncClient:
if self._client is None:
self._client = httpx.AsyncClient(timeout=httpx.Timeout(60.0, connect=5.0))
return self._client
def _headers(self) -> dict:
return {"X-Harness-Internal": self.key} if self.key else {}
async def read(self, sid: str, path: str) -> bytes | None:
member = _safe_member(path)
if not member or not self.endpoint:
return None
try:
r = await self._c().get(f"{self.endpoint}/file", params={"identifier": sid, "path": member},
headers=self._headers())
except Exception: # noqa: BLE001 — a runner that is down reads as "no file", never a 500
return None
return r.content if r.status_code == 200 else None
async def write(self, sid: str, path: str, data: bytes) -> bool:
member = _safe_member(path)
if not member or not self.endpoint:
return False
try:
r = await self._c().put(f"{self.endpoint}/file", params={"identifier": sid, "path": member},
headers={**self._headers(), "content-type": "application/octet-stream"},
content=data)
except Exception: # noqa: BLE001
return False
if r.status_code != 200:
return False
await self._capture(sid, member, data)
return True
async def _capture(self, sid: str, member: str, data: bytes) -> None:
"""Persist the app's copy of this file so it outlives the live directory.
THIS is why the class takes a blob store. The live workspace is reaped after
HR_WORKSPACE_TTL_HOURS and is only ever refreshed by a TURN's checkpoint, so a file an app
wrote BETWEEN turns lived in exactly one place and was deleted with it: a sheet somebody
typed into and never asked the agent about again came back empty, and so did every agent
column they had bound by hand. The hosted sibling writes straight into the durable tarball
and never had the hole, which is the tell — one Protocol, two meanings for `write`.
The id is derived from the path, so a sheet saved a thousand times keeps ONE blob instead
of leaking one per save. `written_at` is what makes the read side deterministic: the
agent's own capture of the same filename carries one too, and the newer of the two wins
rather than whichever the blob listing happened to return last."""
if not self.blob or not self.kb:
return
cfile = "cfile_" + hashlib.sha256(f"app:{member}".encode()).hexdigest()[:32]
meta = json.dumps({"filename": member,
"media_type": mimetypes.guess_type(member)[0] or "application/octet-stream",
"written_at": time.time(),
"source": "app"}).encode()
try:
if await self.blob.put(self.kb, f"containers/{sid}/{cfile}", data):
await self.blob.put(self.kb, f"containers/{sid}/{cfile}.meta", meta)
return
except Exception as e: # noqa: BLE001
print(f"[workspace] durable capture failed sid={sid} path={member}: {e}", flush=True)
return
# A live write that was not captured is the old silent-loss shape. Say so rather than
# letting it look like a clean save.
print(f"[workspace] durable capture rejected sid={sid} path={member}", flush=True)
class CheckpointWorkspaceFiles:
"""Hosted: between turns the workspace is only the checkpoint tarball, so a write rewrites
that archive — copy every member across, substituting the one being written.
Rewriting rather than appending is deliberate: tar allows duplicate names and readers take the
LAST one, so appending would work until something read the archive differently, and the bug
would be a stale file appearing at random."""
def __init__(self, blob, kb: str, key_for):
self.blob, self.kb, self.key_for = blob, kb, key_for
async def read(self, sid: str, path: str) -> bytes | None:
member = _safe_member(path)
if not member:
return None
raw = await self.blob.get(self.kb, self.key_for(sid))
if not raw:
return None
def _extract() -> bytes | None:
with tarfile.open(fileobj=io.BytesIO(raw), mode="r:gz") as tf:
for name in (member, f"./{member}"):
try:
f = tf.extractfile(name)
except KeyError:
continue
if f:
return f.read()
return None
try:
return await asyncio.to_thread(_extract)
except Exception: # noqa: BLE001 — a corrupt archive reads as "not there"
return None
async def write(self, sid: str, path: str, data: bytes) -> bool:
member = _safe_member(path)
if not member:
return False
key = self.key_for(sid)
raw = await self.blob.get(self.kb, key)
def _rewrite() -> bytes:
out = io.BytesIO()
wrote = False
with tarfile.open(fileobj=out, mode="w:gz") as dst:
if raw:
with tarfile.open(fileobj=io.BytesIO(raw), mode="r:gz") as src:
for m in src.getmembers():
name = m.name[2:] if m.name.startswith("./") else m.name
if name == member:
continue # replaced below
f = src.extractfile(m) if m.isfile() else None
dst.addfile(m, f)
info = tarfile.TarInfo(member)
info.size = len(data)
info.mtime = int(time.time())
info.mode = 0o644
dst.addfile(info, io.BytesIO(data))
wrote = True
return out.getvalue() if wrote else b""
try:
tar = await asyncio.to_thread(_rewrite)
except Exception: # noqa: BLE001
return False
return bool(tar) and await self.blob.put(self.kb, key, tar)
def make_backing(*, client_getter, vg_url: str, vg_key: str, vg_tenant: str,
vault_url: str, vault_key: str) -> Backing:
"""Build the configured backing. HR_BACKING=vg|local; unset ⇒ vg when VG_GATEWAY_URL is
configured (production), local otherwise (a laptop clone works with zero env)."""
mode = os.environ.get("HR_BACKING", "").strip().lower() or ("vg" if vg_url else "local")
# Where a session's workspace actually lives, which is the one thing that differs most between
# the two deployments: a live directory the runner owns and serves, or a checkpoint tarball
# that is the only copy between turns.
ws_key = lambda sid: f"sessions/{sid}/workspace.tgz" # noqa: E731 — matches gateway._ws_blob
if mode == "vg":
from backing_vg import VgBlobStore, VgGraphStore, VaultSecretStore
blob = VgBlobStore(client_getter, vg_url, vg_key)
return Backing(
graph=VgGraphStore(client_getter, vg_url, vg_key, vg_tenant),
blob=blob,
secrets=VaultSecretStore(client_getter, vault_url, vault_key),
mode="vg",
workspace=CheckpointWorkspaceFiles(blob, os.environ.get("HARNESS_BLOB_KB", "harness-sessions"), ws_key))
data = os.environ.get("HR_DATA_DIR", ".hr-data")
local_blob = FileBlobStore(os.path.join(data, "blobs"))
return Backing(
graph=SqliteGraphStore(os.path.join(data, "graph.db")),
blob=local_blob,
secrets=FileSecretStore(os.path.join(data, "secrets")),
mode="local",
workspace=RunnerWorkspaceFiles(os.environ.get("POOL_MGMT_ENDPOINT", "http://127.0.0.1:8081"),
os.environ.get("HARNESS_INTERNAL_KEY", ""),
local_blob,
os.environ.get("HARNESS_RESP_BLOB_KB", "harness-responses")))
# ===== file: gateway/browser_plane.py =====
"""The browser server's plane: a real browser, in a vendor's cloud or the operator's own, driven
from this gateway over the Chrome DevTools Protocol, handed to the agent as a dozen typed tools
and nothing else.
The operator's own browser (HR_BROWSER_CDP_URL, a Chrome started with --remote-debugging-port or
any CDP endpoint) takes the vendor's place whole: no vendor session, no live view, no charge, one
browser context per harness session on the one browser, and the private-address rule lifted,
since that browser sits on the operator's own network and reaches what the operator's shell
already can (a customer's crawler behind a headful Chrome broker, 2026-09-30).
One browser session per harness session, made on the first call and stopped at the end of the
turn, after idle, or at the session cap. The vendor is behind one adapter (Browser Use Cloud
first; Browserbase has the same shape: a session id and a CDP address), so the tools, the guards
and the metering never learn a vendor's name.
What the agent is never given: a JavaScript evaluator, downloads, cookies, the vendor's key, the
CDP address. What every request is checked against, subresources and redirects included: the
harness's allow and deny lists and the private, loopback, link-local and metadata address ranges.
Money: the vendor's list price, no markup. A session is priced BEFORE it starts as the session cap
times the per-minute rate (that is what counts against the task's cost cap), and metered AFTER
from the vendor's own figures for the minutes it ran (rounded up, one-minute minimum). The prices
below are the estimate's; the meter reads what the vendor reports.
"""
from __future__ import annotations
import asyncio
import ipaddress
import json
import math
import os
import socket
import time
import urllib.parse
import httpx
CDP_URL = os.environ.get("HR_BROWSER_CDP_URL", "").strip() # the operator's own browser, instead of the vendor
API_KEY = os.environ.get("BROWSER_USE_API_KEY", "")
# The key as a mounted secret file, read at every use: a rotated Key Vault secret reaches a
# mounted file without a new revision, where an environment variable holds the value the process
# started with.
API_KEY_FILE = os.environ.get("BROWSER_USE_API_KEY_FILE", "")
API_BASE = os.environ.get("BROWSER_USE_API_BASE", "https://api.browser-use.com/api/v4").rstrip("/")
SESSION_CAP_MIN = int(os.environ.get("HR_BROWSER_SESSION_MINUTES", "20")) # per harness session
# Fair share of the vendor's concurrency (10 sessions until $1,000 lifetime spend, then 250):
# no org holds more than its share, and this process stops one below the vendor's ceiling so the
# refusal an agent reads is ours, in a sentence, not the vendor's 429. Per process: a second
# gateway replica has its own counts.
ORG_SESSIONS = int(os.environ.get("HR_BROWSER_ORG_SESSIONS", "3"))
MAX_SESSIONS = int(os.environ.get("HR_BROWSER_MAX_SESSIONS", "9"))
CALL_CAP_S = float(os.environ.get("HR_BROWSER_CALL_SECONDS", "60")) # per tool call
IDLE_S = float(os.environ.get("HR_BROWSER_IDLE_SECONDS", "120")) # no call for this long: stopped
TEXT_CAP = 20000 # characters of extracted text
SNAPSHOT_CAP = 150 # elements in a snapshot
# The vendor's list prices, no markup (browser-use.com/pricing and the v4 API docs, read
# 2026-09-24): browser time $0.02 per hour, rounded up to whole minutes, one-minute minimum;
# residential proxy $5 per GB (this slice runs with no proxy, so the unit never appears);
# screenshots are Playwright's and cost nothing. The meter posts dollars under one unit.
PRICES = {"browser.minute": 0.02 / 60, "browser.proxy_mb": 0.005, "browser.screenshot": 0.0}
UNIT = "browser.usd"
VENDOR = "browser-use"
transport: httpx.BaseTransport | None = None # tests: the vendor's API
connector = None # tests: async (cdp_url) -> a browser-like object
def api_key() -> str:
if API_KEY_FILE:
try:
with open(API_KEY_FILE, encoding="utf-8") as f:
key = f.read().strip()
if key:
return key
except OSError:
pass
return API_KEY
def own_browser() -> bool:
"""Whether the operator's own browser is the browser here."""
return bool(CDP_URL)
def configured() -> bool:
return own_browser() or bool(api_key())
def pricing() -> dict:
"""The browser's price line, for the page that shows it before a workspace turns it on."""
if own_browser():
return {"type": "browser", "unit": "browser hour", "usd_per_unit": 0.0, "markup": 0.0,
"source": "your own browser", "vendor": "own", "billed_as": UNIT,
"rounding": "none", "session_cap_minutes": SESSION_CAP_MIN, "session_estimate_usd": 0.0}
return {"type": "browser", "unit": "browser hour", "usd_per_unit": round(PRICES["browser.minute"] * 60, 6),
"markup": 0.0, "source": "vendor list price", "vendor": VENDOR, "billed_as": UNIT,
"rounding": "up to the minute, one-minute minimum", "session_cap_minutes": SESSION_CAP_MIN,
"session_estimate_usd": session_estimate_usd()}
def session_estimate_usd() -> float:
"""What a session may cost at most, before it starts: the cap in minutes at the list price;
nothing on the operator's own browser."""
return 0.0 if own_browser() else round(SESSION_CAP_MIN * PRICES["browser.minute"], 6)
class BrowserRefused(RuntimeError):
"""A call that must not run, with the sentence the agent reads."""
def __init__(self, code: str, message: str):
super().__init__(message)
self.code = code
class BrowserToolError(RuntimeError):
"""A call that ran and failed in a way the agent can act on."""
# ── the vendor ───────────────────────────────────────────────────────────────────────────────
def _vendor_client() -> httpx.AsyncClient:
return httpx.AsyncClient(transport=transport, timeout=60,
headers={"X-Browser-Use-API-Key": api_key(), "content-type": "application/json"})
async def vendor_create(timeout_min: int, metadata: dict) -> dict:
"""A browser in the vendor's cloud: no proxy, no CAPTCHA solving, no recording, so the only
unit that can bill is time. 402 is the vendor's wallet, 429 its concurrency. The screen is the
vendor's default, landscape: what the live view streams is that screen, whatever the shape of
the card showing it (measured on hr-test, 2026-09-28: neither an emulated viewport over CDP
nor allowResizing changed the streamed frame), so the console fits the frame inside the card.
The operator's own browser is already running: its record names the CDP address and no
vendor session (an empty id is what marks it), and there is no live view of it."""
if own_browser():
return {"id": "", "cdpUrl": CDP_URL, "liveUrl": ""}
async with _vendor_client() as c:
r = await c.post(f"{API_BASE}/browsers", json={
"timeout": int(timeout_min), "proxyCountryCode": None, "solveCaptchas": False,
"enableRecording": False, "metadata": {k: str(v)[:100] for k, v in metadata.items()}})
if r.status_code == 402:
raise BrowserRefused("vendor_credits", "The browser service has no credit left. Tell the person.")
if r.status_code == 429:
raise BrowserRefused("vendor_busy", "Every browser is in use right now. Try again in a minute.")
if r.status_code >= 400:
raise BrowserRefused("vendor_error", f"The browser service refused to start a browser ({r.status_code}).")
d = r.json()
if not d.get("id") or not d.get("cdpUrl"):
raise BrowserRefused("vendor_error", "The browser service answered without a browser address.")
return d
async def vendor_stop(vendor_id: str) -> dict:
"""Stop the browser and read what it cost: closing the CDP socket alone keeps the meter
running on the vendor's side."""
async with _vendor_client() as c:
r = await c.patch(f"{API_BASE}/browsers/{vendor_id}", json={"action": "stop"})
if r.status_code >= 400:
r = await c.get(f"{API_BASE}/browsers/{vendor_id}")
try:
return r.json() if r.status_code < 400 else {}
except ValueError:
return {}
def vendor_cost(d: dict) -> float | None:
"""Dollars the vendor reports for a session: its browser time, plus proxy traffic when any.
None when it reported no browser cost, which is the figure of record."""
if d.get("browserCost") is None:
return None
total = 0.0
for p in (d.get("browserCost"), d.get("proxyCost")):
try:
total += float(p or 0)
except (TypeError, ValueError):
pass
return round(total, 6)
# ── the address guard ────────────────────────────────────────────────────────────────────────
_PRIVATE_NAMES = ("localhost", ".localhost", ".local", ".internal", ".home.arpa", ".lan", ".corp")
def _ip_private(ip: str) -> bool:
try:
a = ipaddress.ip_address(ip)
except ValueError:
return False
return (a.is_private or a.is_loopback or a.is_link_local or a.is_multicast or a.is_reserved
or a.is_unspecified or (a.version == 6 and a.ipv4_mapped is not None and _ip_private(str(a.ipv4_mapped))))
def _suffix_match(host: str, patterns: list[str]) -> bool:
for p in patterns:
p = (p or "").strip().lower().lstrip("*").lstrip(".")
if p and (host == p or host.endswith("." + p)):
return True
return False
def host_policy(host: str, allow: list[str], deny: list[str], private_ok: bool = False) -> str | None:
"""Why this host may not be reached, or None. Name rules only; the address rule follows.
`private_ok` lifts the private-address names and literals: the operator's own browser sits on
the operator's own network (HR_BROWSER_CDP_URL), and a loopback address was still refused here
by name after the plane had opened that browser (hr-test, 2026-09-30)."""
h = (host or "").strip().lower().rstrip(".")
if not h:
return "no host"
if h.startswith("[") and h.endswith("]"):
h = h[1:-1]
if not private_ok and _ip_private(h):
return "a private or local address"
if not private_ok and (h == "localhost" or any(h.endswith(s) for s in _PRIVATE_NAMES[1:])):
return "a private or local address"
if deny and _suffix_match(h, deny):
return "a site this agent is not allowed to visit"
if allow and not _suffix_match(h, allow):
return "a site outside the ones this agent is allowed to visit"
return None
async def resolves_private(host: str) -> bool:
"""Whether any address the name resolves to is private. A name that does not resolve counts
as private: this is an address a page chose, and it is asked again on every request."""
h = (host or "").strip().lower().rstrip(".").strip("[]")
if _ip_private(h):
return True
try:
ipaddress.ip_address(h)
return False
except ValueError:
pass
try:
infos = await asyncio.wait_for(asyncio.to_thread(socket.getaddrinfo, h, None), 4)
except Exception: # noqa: BLE001
return True
return any(_ip_private(str(i[4][0])) for i in infos) or not infos
# ── the session ──────────────────────────────────────────────────────────────────────────────
# One browser per harness session, wherever the call lands. The RECORD of it (the vendor's id,
# the CDP address, when it opened and was last used, its counts, its lists) is kept once for every
# gateway process by the registry below; the ATTACHMENT (the CDP connection, the context, the
# page) is this process's own, made by the first call that reaches it and dropped when the record
# goes. A CLI's tool calls are load-balanced over every replica, and a per-process record let one
# turn open a browser on each replica it reached (four for one turn on 2026-09-25, three of them
# stopped only by the idle reaper, none of them the page the agent had opened).
class Session:
"""This process's attachment to a session's browser, built from its record."""
def __init__(self, rec: dict):
self.sid, self.hid, self.org, self.workspace = str(rec["sid"]), str(rec["hid"]), str(rec["org"]), str(rec["workspace"])
self.allow, self.deny = list(rec.get("allow") or []), list(rec.get("deny") or [])
self.vendor_id = str(rec.get("vendor_id") or "")
self.cdp_url = str(rec.get("cdp_url") or "")
self.live_url = str(rec.get("live_url") or "")
self.created = float(rec.get("created") or time.time())
self.reserved = float(rec.get("reserved") or 0.0) # what was counted against the task before the browser opened
self.browser = None
self.context = None
self.page = None
self.blocked: list[str] = [] # hosts this attachment refused, so each is recorded once
self.host_cache: dict[str, bool] = {}
self.lock = asyncio.Lock()
self.closed = False
self.opened = False # this call made the browser at the vendor: the caller writes the open row once
self.pointer: dict | None = None # where the agent last acted (the element's centre), for the console's ghost cursor; kept across navigations so the next action glides from there
def minutes(rec: dict) -> int:
return max(1, math.ceil((time.time() - float(rec.get("created") or time.time())) / 60))
def expired(rec: dict, now: float | None = None) -> str | None:
now = now or time.time()
if now - float(rec.get("created") or now) > SESSION_CAP_MIN * 60:
return "session_cap"
if now - float(rec.get("last_used") or now) > IDLE_S:
return "idle"
return None
# ── the registry ─────────────────────────────────────────────────────────────────────────────
# Where the records live. In this process by default, which is right for one gateway process (a
# self-hosted instance); a deployment that runs several installs one over its shared store, so
# the same session reads the same browser from every replica. `bump` increments the counters
# named in COUNTERS, appends to the lists named in LISTS and sets anything else.
COUNTERS = ("calls", "screenshots")
LISTS = ("blocked",)
class LocalRegistry:
def __init__(self):
self._recs: dict[str, dict] = {}
async def get(self, sid: str) -> dict | None:
return self._recs.get(sid)
async def create(self, sid: str, rec: dict) -> bool:
"""True iff this call made the record; False when one is there already."""
if sid in self._recs:
return False
self._recs[sid] = rec
return True
async def bump(self, sid: str, **fields) -> None:
rec = self._recs.get(sid)
if rec is None:
return
for k, v in fields.items():
if k in COUNTERS:
rec[k] = int(rec.get(k) or 0) + int(v)
elif k in LISTS:
rec.setdefault(k, []).append(v)
else:
rec[k] = v
async def delete(self, sid: str) -> bool:
"""True iff this call removed the record: the remover stops the browser and writes its row."""
return self._recs.pop(sid, None) is not None
async def all(self) -> list[dict]:
return list(self._recs.values())
async def lock(self, sid: str, ttl_s: int) -> bool:
return True # one process: the caller's own lock serialises its opens
async def unlock(self, sid: str) -> None:
pass
registry = LocalRegistry()
_SESSIONS: dict[str, Session] = {} # this process's attachments, by session
_pw = None
def sessions() -> dict[str, Session]:
return _SESSIONS
def _record(sid: str, hid: str, org: str, workspace: str, allow: list[str], deny: list[str], d: dict,
reserved: float) -> dict:
now = time.time()
return {"sid": sid, "hid": hid, "org": org, "workspace": workspace, "allow": list(allow), "deny": list(deny),
"vendor_id": str(d["id"]), "cdp_url": str(d["cdpUrl"]), "live_url": str(d.get("liveUrl") or ""),
"created": now, "last_used": now, "calls": 0, "screenshots": 0, "blocked": [],
"reserved": float(reserved)}
async def _connect(cdp_url: str):
global _pw
if connector is not None:
return await connector(cdp_url)
from playwright.async_api import async_playwright
if _pw is None:
_pw = await async_playwright().start()
return await _pw.chromium.connect_over_cdp(cdp_url, timeout=30000)
async def _allowed(s: Session, url: str) -> str | None:
"""Why this URL may not be requested from this session, or None; resolved once per host."""
try:
u = urllib.parse.urlsplit(url)
except ValueError:
return "not a valid address"
if u.scheme not in ("http", "https"):
return f"a {u.scheme or 'non-web'} address" if u.scheme not in ("about", "data", "blob") else None
host = u.hostname or ""
own = not s.vendor_id # the operator's own browser, on the operator's own network
why = host_policy(host, s.allow, s.deny, private_ok=own)
if why:
return why
if own:
return None
if host not in s.host_cache:
s.host_cache[host] = await resolves_private(host)
return "a private or local address" if s.host_cache[host] else None
async def open_session(sid: str, hid: str, org: str, workspace: str, allow: list[str], deny: list[str],
reserved: float = 0.0) -> Session:
"""The session's browser, attached here: the recorded one when there is one, else a new one at
the vendor, recorded for every process first. One opener at a time across processes: the
others wait for its record. `opened` on the result says this call made the browser."""
if not configured():
raise BrowserRefused("not_configured", "The browser service is not set up on this deployment. Tell the person.")
rec = await registry.get(sid)
locked = False
if rec is None:
for _ in range(60):
if await registry.lock(sid, 30):
locked = True
break
await asyncio.sleep(0.5)
rec = await registry.get(sid)
if rec is not None:
break
else:
raise BrowserRefused("busy", "The browser is still being started. Try again in a moment.")
opened = False
try:
if rec is None:
live = await registry.all()
if sum(1 for x in live if x.get("org") == org) >= ORG_SESSIONS:
raise BrowserRefused("org_busy", f"This account already has {ORG_SESSIONS} browsers open. Try again when one of them finishes.")
if len(live) >= MAX_SESSIONS:
raise BrowserRefused("busy", "Every browser is in use right now. Try again in a minute.")
d = await vendor_create(SESSION_CAP_MIN, {"harness": hid, "session": sid})
rec = _record(sid, hid, org, workspace, allow, deny, d, reserved)
if await registry.create(sid, rec):
opened = True
else:
# Made elsewhere between the lock and here (a lock the store could not keep): theirs stands.
try:
await vendor_stop(rec["vendor_id"])
except Exception: # noqa: BLE001
pass
rec = await registry.get(sid)
if rec is None:
raise BrowserRefused("busy", "The browser could not be started. Try again.")
finally:
if locked:
await registry.unlock(sid)
try:
s = await attach(rec)
except Exception:
if opened: # a browser nobody can reach: not left for the reaper
await registry.delete(sid)
try:
await vendor_stop(rec["vendor_id"])
except Exception: # noqa: BLE001
pass
raise
s.opened = opened
return s
async def attach(rec: dict) -> Session:
"""This process's attachment to the recorded browser, made on first use and kept. Every process
uses the browser's own default context and its first tab, so the page is the same page from
every replica; the request gate is installed per attachment and reads the record's lists."""
sid = str(rec["sid"])
s = _SESSIONS.get(sid)
if s is not None and not s.closed and s.vendor_id == str(rec.get("vendor_id") or ""):
return s
if s is not None:
await detach(s) # the record names another browser: this one is gone
s = Session(rec)
s.browser = await _connect(s.cdp_url)
try:
# The operator's own browser is shared by every session, so each gets a context of its own
# (its own cookies and tabs), made over this attachment's connection and closed with it; a
# vendor browser is one session's alone and its default context is the page.
contexts = [] if not s.vendor_id else list(getattr(s.browser, "contexts", None) or [])
s.context = contexts[0] if contexts else await s.browser.new_context(accept_downloads=False, viewport={"width": 1280, "height": 800})
async def gate(route, request):
why = await _allowed(s, request.url)
if why:
host = urllib.parse.urlsplit(request.url).hostname or request.url[:60]
if host not in s.blocked:
s.blocked.append(host)
await registry.bump(sid, blocked=host)
await route.abort("blockedbyclient")
else:
await route.continue_()
await s.context.route("**/*", gate)
pages = list(getattr(s.context, "pages", None) or [])
s.page = pages[0] if pages else await s.context.new_page()
s.page.set_default_timeout(30000)
on = getattr(s.context, "on", None)
if callable(on):
on("page", lambda pg: asyncio.ensure_future(_follow_new_tab(s, sid, pg)))
except Exception:
await detach(s)
raise
_SESSIONS[sid] = s
return s
async def _follow_new_tab(s: Session, sid: str, pg) -> None:
"""A link the person opened in a new tab while they have the browser is followed in the tab the
live view shows: the new page's address goes to the shown page and the new page closes, so the
view stays with them (the console's live view shows one tab, and a link marked to open in a new
one looked like a click that did nothing: the LinkedIn and X icons on a directory page,
2026-09-27). Only while the person has the browser: the agent's own tabs are its to manage
(list_tabs, switch_tab)."""
try:
rec = await registry.get(sid)
if (rec or {}).get("control") != "user" or pg is s.page:
return
try:
await pg.wait_for_load_state("domcontentloaded", timeout=8000)
except Exception: # noqa: BLE001
pass
url = str(getattr(pg, "url", "") or "")
if url and url != "about:blank":
await s.page.goto(url, wait_until="domcontentloaded", timeout=30000)
await pg.close()
except Exception: # noqa: BLE001
pass
async def detach(s: Session) -> None:
"""Drop this process's attachment; the browser itself stays as it is (a connected browser's
close disconnects it and closes only the contexts this attachment made)."""
s.closed = True
if _SESSIONS.get(s.sid) is s:
_SESSIONS.pop(s.sid, None)
closer = getattr(s.browser, "close", None)
if closer:
try:
await asyncio.wait_for(closer(), 10)
except Exception: # noqa: BLE001
pass
async def close_session_browser(rec: dict, s: Session | None) -> dict:
"""Stop the recorded browser and read the vendor's figures, dropping this process's attachment
when it has one. Every step is best effort: a browser that would not close must not keep its
minutes from being metered."""
if s is not None:
s.closed = True
if _SESSIONS.get(s.sid) is s:
_SESSIONS.pop(s.sid, None)
for closer in (getattr(s.context, "close", None), getattr(s.browser, "close", None)):
if closer:
try:
await asyncio.wait_for(closer(), 10)
except Exception: # noqa: BLE001
pass
figures: dict = {}
if rec.get("vendor_id"):
try:
figures = await vendor_stop(str(rec["vendor_id"]))
except Exception: # noqa: BLE001
figures = {}
usd = vendor_cost(figures)
source = "vendor"
if not rec.get("vendor_id"):
usd, source = 0.0, "own" # the operator's own browser: nothing to meter
elif usd is None:
usd, source = round(minutes(rec) * PRICES["browser.minute"], 6), "table"
return {"minutes": minutes(rec), "usd": usd, "usd_source": source, "calls": int(rec.get("calls") or 0),
"screenshots": int(rec.get("screenshots") or 0), "blocked": list(rec.get("blocked") or [])[:20],
"proxy_mb": figures.get("proxyUsedMb") or 0}
async def expired_sessions(now: float | None = None) -> list[tuple[dict, str]]:
"""Every recorded browser past its cap or idle, from any process."""
now = now or time.time()
return [(rec, why) for rec in await registry.all() if (why := expired(rec, now))]
async def sweep_attachments() -> None:
"""Drop the attachments whose browser was stopped from another process."""
for s in list(_SESSIONS.values()):
if await registry.get(s.sid) is None:
await detach(s)
# ── the tools ────────────────────────────────────────────────────────────────────────────────
_TOOLS: list[dict] = []
def _tool(name: str, risk: str, description: str, schema: dict):
_TOOLS.append({"name": name, "risk": risk, "description": description,
"inputSchema": {"type": "object", "properties": schema.get("properties", {}),
"required": schema.get("required", []), "additionalProperties": False}})
_REF = {"ref": {"type": "string", "description": "An element ref from the last snapshot, like e12."},
"selector": {"type": "string", "description": "A CSS selector, when there is no ref for the element."}}
_tool("navigate", "act", "Open a web address in the current tab and read what is on the page.",
{"properties": {"url": {"type": "string", "description": "An http or https address."}}, "required": ["url"]})
_tool("get_url", "read", "The current tab's address and title.", {})
_tool("snapshot", "read", "What is on the page: headings, links, buttons and fields, each with a ref to act on.",
{"properties": {"max_elements": {"type": "integer", "description": f"At most this many elements (default {SNAPSHOT_CAP})."}}})
_tool("extract_text", "read", "The readable text of the page, or of one element.",
{"properties": {"selector": {"type": "string", "description": "A CSS selector; the whole page when absent."},
"max_chars": {"type": "integer", "description": f"At most this many characters (default {TEXT_CAP})."}}})
_tool("click", "act", "Click an element, by ref or selector.", {"properties": dict(_REF)})
_tool("type", "act", "Type text into a field, by ref or selector, replacing what is there.",
{"properties": {**_REF, "text": {"type": "string"}, "submit": {"type": "boolean", "description": "Press Enter afterwards."}},
"required": ["text"]})
_tool("press_key", "act", "Press a key, like Enter, Tab, Escape or ArrowDown.",
{"properties": {"key": {"type": "string"}}, "required": ["key"]})
_tool("scroll", "act", "Scroll the page.",
{"properties": {"direction": {"type": "string", "enum": ["down", "up"]},
"pixels": {"type": "integer", "description": "How far (default 600)."}}})
_tool("wait_for", "read", "Wait until text or an element appears, or for a number of milliseconds.",
{"properties": {"text": {"type": "string"}, "selector": {"type": "string"},
"ms": {"type": "integer", "description": f"Milliseconds, at most {int(CALL_CAP_S * 1000)}."}}})
_tool("screenshot", "read", "A picture of the current tab, kept with the task's files and shown to you.",
{"properties": {"full_page": {"type": "boolean", "description": "The whole page, not only the visible part."}}})
_tool("back", "act", "Go back one page in the current tab.", {})
_tool("list_tabs", "read", "The open tabs, with their index, address and title.", {})
_tool("switch_tab", "act", "Make one of the open tabs the current tab.",
{"properties": {"index": {"type": "integer"}}, "required": ["index"]})
def tool_list() -> list[dict]:
return [{"name": f"browser_{t['name']}", "description": f"[browser, {t['risk']}] {t['description']}",
"inputSchema": t["inputSchema"],
"annotations": {"readOnlyHint": t["risk"] == "read", "destructiveHint": False}} for t in _TOOLS]
def find(name: str) -> dict | None:
return next((t for t in _TOOLS if t["name"] == name), None)
_SNAPSHOT_JS = """
(max) => {
const out = []; let n = 0;
const seen = new Set();
// Visible to a person: laid out, not hidden, and at least 8 px each way. Sites keep tiny proxy
// controls (DuckDuckGo's 4 x 4 search input beside the field people see); an agent that types
// into one of those types into nothing anyone can see (hr-test, 2026-09-28).
const vis = (el) => { const r = el.getBoundingClientRect(); const cs = getComputedStyle(el);
return r.width >= 8 && r.height >= 8 && cs.visibility !== 'hidden' && cs.display !== 'none'; };
const nameOf = (el) => (el.getAttribute('aria-label') || el.getAttribute('title') || el.getAttribute('placeholder')
|| (el.labels && el.labels[0] && el.labels[0].innerText) || el.getAttribute('alt') || el.innerText || el.value || '').trim().replace(/\\s+/g, ' ').slice(0, 80);
const sel = 'h1,h2,h3,a[href],button,input,select,textarea,[role=button],[role=link],[role=tab],[role=menuitem],[role=checkbox],[role=radio],[role=textbox],[contenteditable=true],summary';
for (const el of document.querySelectorAll(sel)) {
if (n >= max) { out.push({truncated: true}); break; }
if (!vis(el) || seen.has(el)) continue;
seen.add(el);
const tag = el.tagName.toLowerCase();
const role = el.getAttribute('role') || (tag === 'a' ? 'link' : tag === 'button' || tag === 'summary' ? 'button'
: tag === 'select' ? 'combobox' : tag === 'textarea' ? 'textbox' : tag === 'input' ? ((el.type === 'checkbox' || el.type === 'radio' || el.type === 'submit' || el.type === 'button') ? el.type : 'textbox')
: /^h[1-3]$/.test(tag) ? 'heading' : tag);
const row = {role, name: nameOf(el)};
if (role !== 'heading') { const ref = 'e' + (++n); el.setAttribute('data-hr-ref', ref); row.ref = ref; }
else row.level = Number(tag[1]);
if (tag === 'a') row.href = (el.getAttribute('href') || '').slice(0, 120);
if (tag === 'input' || tag === 'textarea') { if (el.value && el.type !== 'password') row.value = String(el.value).slice(0, 60); if (el.disabled) row.disabled = true; }
if (el.type === 'checkbox' || el.type === 'radio') row.checked = !!el.checked;
out.push(row);
}
return out;
}
"""
def _fmt_snapshot(rows: list[dict]) -> str:
lines = []
for r in rows:
if r.get("truncated"):
lines.append("… more elements below; scroll or ask for a larger snapshot")
continue
if r.get("role") == "heading":
lines.append(f"h{r.get('level', 2)} \"{r.get('name', '')}\"")
continue
extra = ""
if r.get("href"):
extra += f" href={r['href']}"
if "value" in r:
extra += f" value=\"{r['value']}\""
if "checked" in r:
extra += " checked" if r["checked"] else " unchecked"
if r.get("disabled"):
extra += " disabled"
lines.append(f"{r.get('ref')} {r.get('role')} \"{r.get('name', '')}\"{extra}")
return "\n".join(lines) if lines else "(nothing interactive is visible on this page)"
async def _snapshot(s: Session, max_elements: int) -> str:
rows = await s.page.evaluate(_SNAPSHOT_JS, max(1, min(int(max_elements or SNAPSHOT_CAP), 500)))
return _fmt_snapshot(rows or [])
def _target(s: Session, args: dict):
ref, selector = str(args.get("ref") or "").strip(), str(args.get("selector") or "").strip()
if ref:
return s.page.locator(f'[data-hr-ref="{ref}"]').first
if selector:
return s.page.locator(selector).first
raise BrowserToolError("Say which element: a ref from the snapshot or a selector.")
async def _where(s: Session) -> str:
try:
title = await s.page.title()
except Exception: # noqa: BLE001
title = ""
return f"{s.page.url}" + (f" ({title})" if title else "")
async def _settled(s: Session) -> None:
with_timeout = getattr(s.page, "wait_for_load_state", None)
if with_timeout:
try:
await asyncio.wait_for(s.page.wait_for_load_state("domcontentloaded"), 8)
except Exception: # noqa: BLE001
pass
async def _mark(s: Session, loc) -> None:
"""Remember the centre of the element about to be acted on, in page coordinates, so the console
can show a ghost cursor going there. Best effort: a locator that cannot be measured leaves the
last point standing."""
try:
box = await loc.bounding_box()
except Exception: # noqa: BLE001
return
if isinstance(box, dict) and box.get("width") is not None:
s.pointer = {"x": round(float(box["x"]) + float(box["width"]) / 2, 1), "y": round(float(box["y"]) + float(box["height"]) / 2, 1)}
async def viewport_of(s: Session) -> dict:
"""The page's own layout viewport (window.innerWidth/innerHeight), the coordinate space of
`pointer`. Asked of the page, not of the context: the vendor's default context carries its
window's size, not the 1280 x 800 this plane asks for when it makes a context of its own, and
a cursor placed on the assumed size landed low and right of the element the agent clicked
(hr-test, 2026-09-27)."""
try:
v = await s.page.evaluate("() => [window.innerWidth, window.innerHeight]")
if isinstance(v, (list, tuple)) and len(v) == 2 and int(v[0]) > 0:
return {"w": int(v[0]), "h": int(v[1])}
except Exception: # noqa: BLE001
pass
try:
vs = s.page.viewport_size if s.page is not None else None
except Exception: # noqa: BLE001
vs = None
return {"w": int((vs or {}).get("width") or 1280), "h": int((vs or {}).get("height") or 800)}
async def call(s: Session, name: str, args: dict):
"""Run one tool on this session. Returns text, or ("image", png bytes, caption) for a
screenshot the caller stores. Raises BrowserToolError with a sentence the agent can act on."""
await registry.bump(s.sid, last_used=time.time(), calls=1)
page = s.page
if name == "navigate":
url = str(args.get("url") or "").strip()
if not url:
raise BrowserToolError("Give a web address to open.")
if "://" not in url:
url = "https://" + url
why = await _allowed(s, url)
if why:
raise BrowserToolError(f"That address is {why}, so it cannot be opened.")
try:
await page.goto(url, wait_until="domcontentloaded")
except Exception as e: # noqa: BLE001
raise BrowserToolError(f"The page did not open: {str(e).splitlines()[0][:200]}") from None
return f"Opened {await _where(s)}\n\n{await _snapshot(s, 80)}"
if name == "get_url":
return await _where(s)
if name == "snapshot":
return f"{await _where(s)}\n\n{await _snapshot(s, args.get('max_elements') or SNAPSHOT_CAP)}"
if name == "extract_text":
selector = str(args.get("selector") or "").strip() or "body"
cap = max(200, min(int(args.get("max_chars") or TEXT_CAP), TEXT_CAP))
try:
text = await page.locator(selector).first.inner_text()
except Exception as e: # noqa: BLE001
raise BrowserToolError(f"No readable text at {selector!r}: {str(e).splitlines()[0][:160]}") from None
text = "\n".join(line.strip() for line in text.splitlines() if line.strip())
return text[:cap] + (f"\n… {len(text) - cap} more characters" if len(text) > cap else "")
if name == "click":
loc = _target(s, args)
await _mark(s, loc)
try:
await loc.click(timeout=15000)
except Exception as e: # noqa: BLE001
raise BrowserToolError(f"The click did not land: {str(e).splitlines()[0][:160]}") from None
await _settled(s)
return f"Clicked. Now at {await _where(s)}"
if name == "type":
loc = _target(s, args)
await _mark(s, loc)
text = str(args.get("text") or "")
try:
await loc.fill(text, timeout=15000)
if args.get("submit"):
await loc.press("Enter")
await _settled(s)
except Exception as e: # noqa: BLE001
raise BrowserToolError(f"Typing failed: {str(e).splitlines()[0][:160]}") from None
return f"Typed {len(text)} characters." + (f" Now at {await _where(s)}" if args.get("submit") else "")
if name == "press_key":
key = str(args.get("key") or "").strip()
if not key:
raise BrowserToolError("Say which key.")
await page.keyboard.press(key)
await _settled(s)
return f"Pressed {key}. Now at {await _where(s)}"
if name == "scroll":
px = max(50, min(int(args.get("pixels") or 600), 5000))
dy = -px if str(args.get("direction") or "down") == "up" else px
await page.mouse.wheel(0, dy)
return f"Scrolled {'up' if dy < 0 else 'down'} {abs(dy)} pixels."
if name == "wait_for":
text, selector = str(args.get("text") or "").strip(), str(args.get("selector") or "").strip()
ms = int(args.get("ms") or 0)
budget = int(CALL_CAP_S * 1000) - 2000
try:
if text:
await page.get_by_text(text).first.wait_for(timeout=budget)
return f"\"{text}\" is on the page."
if selector:
await page.locator(selector).first.wait_for(timeout=budget)
return f"{selector} is on the page."
except Exception: # noqa: BLE001
raise BrowserToolError("It did not appear within the wait.") from None
if ms > 0:
await asyncio.sleep(min(ms, budget) / 1000)
return f"Waited {min(ms, budget)} ms."
raise BrowserToolError("Say what to wait for: text, a selector or a number of milliseconds.")
if name == "screenshot":
try:
png = await page.screenshot(type="png", full_page=bool(args.get("full_page")))
except Exception as e: # noqa: BLE001
raise BrowserToolError(f"The screenshot failed: {str(e).splitlines()[0][:160]}") from None
await registry.bump(s.sid, screenshots=1)
return ("image", png, f"Screenshot of {await _where(s)}")
if name == "back":
try:
await page.go_back(wait_until="domcontentloaded")
except Exception as e: # noqa: BLE001
raise BrowserToolError(f"Could not go back: {str(e).splitlines()[0][:160]}") from None
return f"Back at {await _where(s)}"
if name == "list_tabs":
rows = []
for i, p in enumerate(s.context.pages):
try:
t = await p.title()
except Exception: # noqa: BLE001
t = ""
rows.append(f"{i}{' *' if p is s.page else ''} {p.url}" + (f" ({t})" if t else ""))
return "\n".join(rows) or "(no tabs)"
if name == "switch_tab":
pages = s.context.pages
i = int(args.get("index") or 0)
if i < 0 or i >= len(pages):
raise BrowserToolError(f"There is no tab {i}; there are {len(pages)}.")
s.page = pages[i]
s.page.set_default_timeout(30000)
return f"Current tab is now {i}: {await _where(s)}"
raise BrowserToolError(f"No tool named browser_{name}.")
def audit_line(d: dict) -> str:
"""The record's one-line shape, for the trace and the log."""
return json.dumps(d, separators=(",", ":"))
# ===== file: gateway/control_sqlite.py =====
"""SQLite control-store container — the self-hosted sibling of the hosted document store.
WHY THIS EXISTS
The control store holds the coordination state that keeps concurrent turns honest:
idempotency reservations, per-session execution leases (with fencing), response
cursors/terminal state, and API-key cache entries. It was originally written against a
hosted document database, which meant a self-hosted deployment had no control plane at
all — and any request carrying an `Idempotency-Key` failed closed with a 503.
Rather than fork the ~20 control-store functions into a second implementation (two copies
of subtle concurrency logic is how they drift apart), this module supplies a *container*
with the same five operations and the same error semantics the hosted client exposes:
read_item / create_item / replace_item / upsert_item / delete_item
`control_store` picks the container at runtime, so every function above it — the fencing
maths, the CAS retries, the TTL handling — is byte-for-byte the same code in both
deployments. That is the whole point: one codebase, two backings.
CONCURRENCY
Compare-and-swap is the load-bearing primitive: `replace_item(..., etag=...)` must fail if
another writer changed the document first, or two turns could both believe they hold a
session lease. Here `_etag` is a monotonically bumped counter and the CAS is performed
inside a single IMMEDIATE transaction, so a concurrent writer either loses the race
cleanly (raising the access-condition error the callers already catch) or serialises
behind it. SQLite is opened per-operation in WAL mode, which safely supports the multiple
readers + single writer this workload produces.
TTL
The hosted store expires documents by a `ttl` field. SQLite has no background expiry, so
reads treat an expired document as absent (the same observable behaviour) and writes
opportunistically sweep. Expiry is thus lazy but never *late* from a caller's point of
view, which is what the lease and idempotency logic actually depends on.
"""
from __future__ import annotations
import json
import os
import sqlite3
import time
# ── error types ───────────────────────────────────────────────────────────────
# Named to match the hosted client's exceptions so `except cex.CosmosResourceNotFoundError`
# in control_store works against either backing without a single conditional at the call
# site. control_store resolves `cex` to this module when running self-hosted.
class CosmosResourceNotFoundError(Exception):
status_code = 404
class CosmosResourceExistsError(Exception):
status_code = 409
class CosmosAccessConditionFailedError(Exception):
"""Raised when a CAS `replace_item` loses the race (etag no longer current)."""
status_code = 412
class CosmosHttpResponseError(Exception):
status_code = 500
def _connect(path: str) -> sqlite3.Connection:
con = sqlite3.connect(path, timeout=30, isolation_level=None)
con.execute("PRAGMA journal_mode=WAL")
con.execute("PRAGMA busy_timeout=30000")
con.execute("PRAGMA synchronous=NORMAL")
con.execute("CREATE TABLE IF NOT EXISTS control ("
"id TEXT PRIMARY KEY, pk TEXT NOT NULL, body TEXT NOT NULL, "
"etag INTEGER NOT NULL, expires REAL)")
con.execute("CREATE INDEX IF NOT EXISTS control_pk ON control(pk)")
return con
def _expires_at(body: dict) -> float | None:
"""Absolute expiry from the document's relative `ttl`, or None for no expiry.
`ts` is the write time the callers already stamp; falling back to now() keeps a
document with a ttl but no ts from living forever."""
try:
ttl = int(body.get("ttl") or 0)
except (TypeError, ValueError):
return None
if ttl <= 0:
return None
try:
base = float(body.get("ts") or time.time())
except (TypeError, ValueError):
base = time.time()
return base + ttl
class SqliteContainer:
"""Async-shaped container over one SQLite file.
The methods are `async def` purely to match the hosted client's surface — the work is
synchronous and fast (single-row, indexed). Making it genuinely async would buy
nothing here and would add a thread hop to every lease heartbeat.
"""
def __init__(self, path: str):
self._path = path
os.makedirs(os.path.dirname(os.path.abspath(path)) or ".", exist_ok=True)
_connect(path).close() # create schema up front, surface I/O errors early
# -- helpers ---------------------------------------------------------------
@staticmethod
def _row_to_doc(row) -> dict:
doc = json.loads(row[0])
doc["_etag"] = str(row[1])
return doc
def _live(self, con, item_id: str):
"""Fetch a row, treating an expired document as absent (and sweeping it)."""
row = con.execute("SELECT body, etag, expires FROM control WHERE id=?",
(item_id,)).fetchone()
if row is None:
return None
if row[2] is not None and row[2] <= time.time():
con.execute("DELETE FROM control WHERE id=?", (item_id,))
return None
return row
# -- container operations --------------------------------------------------
async def read_item(self, item: str, partition_key: str = "", **_kw) -> dict:
con = _connect(self._path)
try:
row = self._live(con, item)
if row is None:
raise CosmosResourceNotFoundError(item)
return self._row_to_doc(row)
finally:
con.close()
async def create_item(self, body: dict, **_kw) -> dict:
"""Insert, or raise Exists — this is the atom idempotency reservation rides on."""
item_id = str(body.get("id") or "")
con = _connect(self._path)
try:
con.execute("BEGIN IMMEDIATE")
if self._live(con, item_id) is not None:
con.execute("ROLLBACK")
raise CosmosResourceExistsError(item_id)
# An expired row may still occupy the primary key — _live swept it above.
con.execute("INSERT OR REPLACE INTO control(id, pk, body, etag, expires) "
"VALUES(?,?,?,?,?)",
(item_id, str(body.get("pk") or ""), json.dumps(body), 1,
_expires_at(body)))
con.execute("COMMIT")
return {**body, "_etag": "1"}
except CosmosResourceExistsError:
raise
except Exception as e: # noqa: BLE001
try:
con.execute("ROLLBACK")
except Exception: # noqa: BLE001
pass
raise CosmosHttpResponseError(str(e)) from e
finally:
con.close()
async def replace_item(self, item: str, body: dict, etag: str | None = None,
match_condition=None, **_kw) -> dict:
"""Replace; when `etag` is supplied this is a compare-and-swap.
Losing the CAS raises the access-condition error rather than overwriting — the
fencing logic in control_store depends on that being a hard failure, because a
silent overwrite would let two owners believe they hold the same lease.
"""
con = _connect(self._path)
try:
con.execute("BEGIN IMMEDIATE")
row = self._live(con, item)
if row is None:
con.execute("ROLLBACK")
raise CosmosResourceNotFoundError(item)
if etag is not None and str(row[1]) != str(etag):
con.execute("ROLLBACK")
raise CosmosAccessConditionFailedError(item)
new_etag = int(row[1]) + 1
stored = {k: v for k, v in body.items() if k != "_etag"}
con.execute("UPDATE control SET pk=?, body=?, etag=?, expires=? WHERE id=?",
(str(body.get("pk") or ""), json.dumps(stored), new_etag,
_expires_at(stored), item))
con.execute("COMMIT")
return {**stored, "_etag": str(new_etag)}
except (CosmosResourceNotFoundError, CosmosAccessConditionFailedError):
raise
except Exception as e: # noqa: BLE001
try:
con.execute("ROLLBACK")
except Exception: # noqa: BLE001
pass
raise CosmosHttpResponseError(str(e)) from e
finally:
con.close()
async def upsert_item(self, body: dict, **_kw) -> dict:
item_id = str(body.get("id") or "")
con = _connect(self._path)
try:
con.execute("BEGIN IMMEDIATE")
row = con.execute("SELECT etag FROM control WHERE id=?", (item_id,)).fetchone()
new_etag = (int(row[0]) + 1) if row else 1
stored = {k: v for k, v in body.items() if k != "_etag"}
con.execute("INSERT OR REPLACE INTO control(id, pk, body, etag, expires) "
"VALUES(?,?,?,?,?)",
(item_id, str(body.get("pk") or ""), json.dumps(stored), new_etag,
_expires_at(stored)))
con.execute("COMMIT")
return {**stored, "_etag": str(new_etag)}
except Exception as e: # noqa: BLE001
try:
con.execute("ROLLBACK")
except Exception: # noqa: BLE001
pass
raise CosmosHttpResponseError(str(e)) from e
finally:
con.close()
async def delete_item(self, item: str, partition_key: str = "", **_kw) -> None:
con = _connect(self._path)
try:
if self._live(con, item) is None:
raise CosmosResourceNotFoundError(item)
con.execute("DELETE FROM control WHERE id=?", (item,))
finally:
con.close()
# ===== file: gateway/control_store.py =====
"""Durable transactional control state for the harness gateway (HR-INF-010/011/012/013).
Hot control state (idempotency reservations, per-session execution leases with
fencing tokens, and monotonic response/cancellation state) lives here in a
dedicated Cosmos NoSQL container — NOT on the shared Gremlin partition, and NOT
in per-process dicts that don't survive a replica or coordinate across replicas.
Why Cosmos NoSQL: point reads, atomic create (409 on duplicate id → race-safe
reservation), ETag optimistic concurrency (if_match CAS), per-item TTL, and a
partition key so one org can't be a global hot spot. Auth is AAD via the
gateway's managed identity (no key on the hot path).
Design contract:
- Every item has `pk` (partition key) = f"org::{org}" so all of an org's control
state co-locates and no cross-org scan is ever needed.
- Idempotency: id=f"idem::{org}::{sha}". reserve() does create_item; a 409 means
another request already reserved it — the caller reads the winner and replays.
- Lease: id=f"lease::{org}::{sid}". A monotonic integer `fence` is bumped once per
turn; the turn carries its fence and only commits a checkpoint if the lease
fence still equals it (a newer turn would have advanced it → stale writer skips).
- Response control: id=f"resp::{org}::{rid}". `terminal` is a one-way latch — once
a terminal status (cancelled/completed/failed/incomplete) is set, no later write
may move it back to a live status. This makes cancellation monotonic.
FAIL-CLOSED where correctness demands it (idempotency reserve), best-effort where
it only optimizes (lease/heartbeat) — each call site decides. When COSMOS_ENDPOINT
is unset (local dev), enabled() is False and callers fall back to prior behavior.
"""
from __future__ import annotations
import os
import time
_ENDPOINT = os.environ.get("COSMOS_ENDPOINT", "").strip()
_DB = os.environ.get("HR_CONTROL_DB", "agentstudio")
_CONTAINER = os.environ.get("HR_CONTROL_CONTAINER", "hr_control")
# Optional master key (fallback); AAD via managed identity is preferred.
_KEY = os.environ.get("COSMOS_KEY", "").strip()
_client = None # CosmosClient (aio)
_cont = None # ContainerProxy
def _local_mode() -> bool:
"""Self-hosted: no hosted document store configured, so use the SQLite container.
Mirrors how backing.py chooses its storage, so one env var story covers both."""
if _ENDPOINT:
return False
return (os.environ.get("HR_BACKING", "").strip().lower() or "local") == "local"
def _exceptions():
"""The error namespace matching the ACTIVE container.
Both namespaces expose the same class names, so every `except cex.CosmosX` below is
correct under either backing and none of the control-store logic needs a conditional."""
if _local_mode():
import control_sqlite
return control_sqlite
from azure.cosmos import exceptions as cex
return cex
def enabled() -> bool:
"""True when a control plane is available. Self-hosted always has one (SQLite), so
idempotency and leases work there instead of failing closed with a 503."""
return bool(_ENDPOINT) or _local_mode()
def _pk(org: str) -> str:
return f"org::{org}"
async def _retryable(coro_factory, attempts: int = 3):
"""Run a Cosmos coroutine with a tiny backoff on TRANSIENT errors (429/503/408/
network) so a momentary hiccup doesn't turn a keyed request into a hard 503.
Non-transient errors (409/404/412) propagate immediately — the caller handles
them as control-flow. coro_factory must return a FRESH awaitable each call."""
import asyncio as _a
cex = _exceptions()
last = None
for i in range(attempts):
try:
return await coro_factory()
except cex.CosmosHttpResponseError as e:
if getattr(e, "status_code", None) not in (408, 429, 449, 503):
raise
last = e
except Exception as e: # azure.core ServiceRequestError/ServiceResponseError (network)
name = type(e).__name__
if "ServiceRequest" not in name and "ServiceResponse" not in name:
raise
last = e
await _a.sleep(0.2 * (i + 1))
raise last # exhausted transient retries — caller fails closed
async def _container():
"""Lazily build the async Cosmos container proxy. Raises on misconfig so
fail-closed call sites surface a retryable error instead of running open."""
global _client, _cont
if _cont is not None:
return _cont
if _local_mode():
# Self-hosted: one SQLite file beside the rest of the local data.
import control_sqlite
root = os.environ.get("HR_DATA_DIR", ".hr-data")
_cont = control_sqlite.SqliteContainer(os.path.join(root, "control.db"))
return _cont
if not _ENDPOINT:
raise RuntimeError("control store not configured (COSMOS_ENDPOINT unset)")
from azure.cosmos.aio import CosmosClient
if _KEY:
_client = CosmosClient(_ENDPOINT, credential=_KEY)
else:
# Managed identity — same UAMI the gateway uses for the session pool.
from azure.identity.aio import DefaultAzureCredential
cred = DefaultAzureCredential(
managed_identity_client_id=os.environ.get("AZURE_CLIENT_ID") or None)
_client = CosmosClient(_ENDPOINT, credential=cred)
db = _client.get_database_client(_DB)
_cont = db.get_container_client(_CONTAINER)
return _cont
# ── idempotency ───────────────────────────────────────────────────────────────
async def idem_reserve(org: str, sha: str, resp_id: str, req_hash: str,
ttl_s: int) -> dict | None:
"""Atomically reserve (org, sha) for resp_id. Returns None if WE won the race
(caller proceeds to execute). Returns the EXISTING record if someone else
already reserved it (caller replays that resp_id). Raises on store failure so
the caller can fail closed with a retryable 503."""
cont = await _container()
item = {"id": f"idem::{org}::{sha}", "pk": _pk(org), "kind": "idem",
"resp_id": resp_id, "req_hash": req_hash, "state": "reserved",
"ts": time.time(), "ttl": int(ttl_s)}
cex = _exceptions()
try:
await _retryable(lambda: cont.create_item(item))
return None # we won
except cex.CosmosResourceExistsError:
try:
return await _retryable(lambda: cont.read_item(item["id"], partition_key=_pk(org)))
except cex.CosmosResourceNotFoundError:
# The item existed at create time but is gone now — a TTL expiry between the
# 409 and this read. Genuinely nobody holds it → treat as won. Any OTHER error
# (network/transient) must PROPAGATE so the caller fails closed; swallowing it
# here would let BOTH racers execute and map the key to the wrong resp_id.
return None
async def idem_get(org: str, sha: str) -> dict | None:
cont = await _container()
cex = _exceptions()
try:
return await _retryable(lambda: cont.read_item(f"idem::{org}::{sha}", partition_key=_pk(org)))
except cex.CosmosResourceNotFoundError:
return None
async def idem_release(org: str, sha: str, resp_id: str) -> None:
"""Best-effort release of a reservation whose request failed BEFORE it
persisted any response record — otherwise a retry would replay a resp_id that
never produced a result (a 'poisoned' key). Only deletes the item if it still
points at OUR resp_id (never clobber a reservation that a retry already reused)."""
cex = _exceptions()
try:
cont = await _container()
cur = await cont.read_item(f"idem::{org}::{sha}", partition_key=_pk(org))
if str(cur.get("resp_id") or "") == resp_id:
await cont.delete_item(f"idem::{org}::{sha}", partition_key=_pk(org))
except cex.CosmosResourceNotFoundError:
return
except Exception: # noqa: BLE001
return
# ── per-session execution lease (fencing) ──────────────────────────────────────
async def lease_acquire(org: str, sid: str, owner: str, ttl_s: int) -> int:
"""Bump and take the session lease; returns THIS turn's fence token (monotonic
int). A newer turn will get a higher fence, so a stale turn detects it lost the
workspace at checkpoint time. Best-effort: returns 0 on contention exhaustion OR
any store error (the caller then behaves as today — no fencing, but the turn
still runs). NOTE: fencing is not yet wired into the turn path (deferred)."""
cex = _exceptions()
try:
cont = await _container()
doc_id = f"lease::{org}::{sid}"
for _ in range(5): # bounded CAS retry
try:
cur = await cont.read_item(doc_id, partition_key=_pk(org))
fence = int(cur.get("fence", 0)) + 1
cur.update(fence=fence, owner=owner, expires=time.time() + ttl_s,
ts=time.time(), ttl=max(int(ttl_s) * 4, 3600))
await cont.replace_item(doc_id, cur,
etag=cur["_etag"], match_condition=_IF_MATCH)
return fence
except cex.CosmosResourceNotFoundError:
try:
await cont.create_item({"id": doc_id, "pk": _pk(org), "kind": "lease",
"fence": 1, "owner": owner,
"expires": time.time() + ttl_s,
"ts": time.time(), "ttl": max(int(ttl_s) * 4, 3600)})
return 1
except cex.CosmosResourceExistsError:
continue # someone created it first — retry the read/bump
except cex.CosmosAccessConditionFailedError:
continue # ETag moved under us — retry
return 0 # gave up (contention) → unfenced fallback
except Exception: # noqa: BLE001 — store down: fail open per docstring
return 0
async def lease_admit(org: str, sid: str, owner: str, ttl_s: int, *,
reject_on_conflict: bool = False, fresh_s: float = 75.0) -> dict:
"""Admission gate at turn start. Reads the session lease and classifies the incumbent:
a DIFFERENT owner counts as an ACTIVE conflict only if its lease is BOTH live (expires in
the future) AND fresh (heartbeated within `fresh_s`) — so a crashed/completed-but-unreleased
lease (which stops renewing) does NOT falsely reject a legitimate follow-up.
Returns {fence, conflict, rejected, holder}.
- reject_on_conflict=False (observe): ALWAYS take ownership (bump fence) and just report the
conflict — the turn proceeds; fence identifies it for checkpoint/heartbeat.
- reject_on_conflict=True (enforce): on an ACTIVE conflict, DO NOT touch the lease (never steal
the running turn's fence — that would make ITS checkpoint stale) and return rejected=True so
the caller refuses the duplicate. Otherwise take ownership.
Best-effort: on any store error returns {fence:0, conflict:False, rejected:False} → fail OPEN."""
cex = _exceptions()
doc_id = f"lease::{org}::{sid}"
now = time.time()
try:
cont = await _container()
for _ in range(5):
try:
cur = await cont.read_item(doc_id, partition_key=_pk(org))
held_by = str(cur.get("owner") or "")
live = float(cur.get("expires") or 0) > now
fresh = (now - float(cur.get("ts") or 0)) < fresh_s
other = bool(held_by and held_by != owner)
conflict = bool(live and other)
active_conflict = bool(conflict and fresh)
if reject_on_conflict and active_conflict:
# Refuse WITHOUT modifying the lease — the incumbent keeps its fence.
return {"fence": 0, "conflict": True, "rejected": True, "holder": held_by}
fence = int(cur.get("fence", 0)) + 1
cur.update(fence=fence, owner=owner, expires=now + ttl_s, ts=now,
ttl=max(int(ttl_s) * 4, 3600))
try:
await cont.replace_item(doc_id, cur, etag=cur["_etag"], match_condition=_IF_MATCH)
return {"fence": fence, "conflict": conflict, "rejected": False, "holder": held_by}
except cex.CosmosAccessConditionFailedError:
continue # another writer moved it — re-read
except cex.CosmosResourceNotFoundError:
try:
await cont.create_item({"id": doc_id, "pk": _pk(org), "kind": "lease",
"fence": 1, "owner": owner, "expires": now + ttl_s,
"ts": now, "ttl": max(int(ttl_s) * 4, 3600)})
return {"fence": 1, "conflict": False, "rejected": False, "holder": ""}
except cex.CosmosResourceExistsError:
continue
return {"fence": 0, "conflict": False, "rejected": False, "holder": ""}
except Exception: # noqa: BLE001 — store down: fail open (turn runs unfenced)
return {"fence": 0, "conflict": False, "rejected": False, "holder": ""}
async def lease_renew(org: str, sid: str, owner: str, fence: int, ttl_s: int) -> bool:
"""Heartbeat: extend the lease expiry while WE still hold `fence`. Without this a long turn's
lease would expire and a follow-up could steal it and wipe the live workspace. Best-effort."""
if fence <= 0:
return False
cex = _exceptions()
doc_id = f"lease::{org}::{sid}"
try:
cont = await _container()
cur = await cont.read_item(doc_id, partition_key=_pk(org))
if int(cur.get("fence", 0)) != fence or str(cur.get("owner") or "") != owner:
return False # superseded — we no longer hold it
cur.update(expires=time.time() + ttl_s, ts=time.time())
try:
await cont.replace_item(doc_id, cur, etag=cur["_etag"], match_condition=_IF_MATCH)
return True
except cex.CosmosAccessConditionFailedError:
return False
except Exception: # noqa: BLE001
return False
async def lease_release(org: str, sid: str, owner: str, fence: int) -> None:
"""Turn end: relinquish the lease (expire it now) if we still hold `fence`, so a follow-up
turn admits without waiting out the TTL. Best-effort; never blocks turn completion."""
if fence <= 0:
return
cex = _exceptions()
doc_id = f"lease::{org}::{sid}"
try:
cont = await _container()
cur = await cont.read_item(doc_id, partition_key=_pk(org))
if int(cur.get("fence", 0)) != fence or str(cur.get("owner") or "") != owner:
return
cur.update(expires=0, ts=time.time())
try:
await cont.replace_item(doc_id, cur, etag=cur["_etag"], match_condition=_IF_MATCH)
except cex.CosmosAccessConditionFailedError:
return
except Exception: # noqa: BLE001
return
async def lease_still_held(org: str, sid: str, fence: int) -> bool:
"""True iff the lease fence is UNCHANGED since this turn acquired it. Used to
gate the checkpoint write so a superseded turn never clobbers newer work.
Fail-safe: if the store is unreadable we return True (write proceeds) rather
than silently drop a checkpoint — availability over the rare stale-write race,
since a single-turn session never advances its own fence."""
if fence <= 0:
return True # unfenced (store was down at acquire) — don't block
try:
cont = await _container()
cur = await cont.read_item(f"lease::{org}::{sid}", partition_key=_pk(org))
return int(cur.get("fence", 0)) == fence
except Exception:
return True
# ── response control (monotonic cancellation) ──────────────────────────────────
_TERMINAL = {"cancelled", "completed", "failed", "incomplete"}
async def resp_put_running(org: str, rid: str, sid: str, runner_turn_id: str,
ttl_s: int) -> None:
"""Upsert a response-control record as live. NEVER downgrades a terminal record
back to running (monotonic latch) — a cancel that already landed wins."""
cont = await _container()
cex = _exceptions()
doc_id = f"resp::{org}::{rid}"
try:
cur = await cont.read_item(doc_id, partition_key=_pk(org))
if str(cur.get("status")) in _TERMINAL:
return # already terminal — do not revive
cur.update(status="running", sid=sid, runner_turn_id=runner_turn_id, ts=time.time())
try:
await cont.replace_item(doc_id, cur, etag=cur["_etag"], match_condition=_IF_MATCH)
except cex.CosmosAccessConditionFailedError:
pass # a terminal write raced in — leave it terminal
except cex.CosmosResourceNotFoundError:
try:
await cont.create_item({"id": doc_id, "pk": _pk(org), "kind": "resp",
"status": "running", "sid": sid,
"runner_turn_id": runner_turn_id,
"ts": time.time(), "ttl": int(ttl_s)})
except cex.CosmosResourceExistsError:
pass
async def resp_set_cursor(org: str, rid: str, cursor: int) -> None:
"""Persist the turn's harvest cursor — how many of THIS turn's events (sandbox per-turn index)
have been durably flushed to the trace. This is the resume point if the owning replica dies and
another adopts the turn: it polls the sandbox with since=cursor, so no already-flushed event is
re-written or re-emitted. Monotonic (never moves backward); best-effort; never revives a terminal
record. Cheap single-field patch on the existing resp:: control doc."""
cont = await _container()
cex = _exceptions()
doc_id = f"resp::{org}::{rid}"
try:
curr = await cont.read_item(doc_id, partition_key=_pk(org))
except cex.CosmosResourceNotFoundError:
return # turn not registered yet — nothing to update
if str(curr.get("status")) in _TERMINAL:
return # already settled — cursor is moot
if int(cursor) <= int(curr.get("harvest_cursor", 0) or 0):
return # monotonic: never move the resume point backward
curr["harvest_cursor"] = int(cursor)
curr["ts"] = time.time()
try:
await cont.replace_item(doc_id, curr, etag=curr["_etag"], match_condition=_IF_MATCH)
except cex.CosmosAccessConditionFailedError:
pass # a concurrent write won — fine, cursor is advisory
async def resp_get_cursor(org: str, rid: str) -> int:
"""The durable harvest cursor for a turn (0 if none / unreadable). See resp_set_cursor."""
cont = await _container()
cex = _exceptions()
try:
curr = await cont.read_item(f"resp::{org}::{rid}", partition_key=_pk(org))
return int(curr.get("harvest_cursor", 0) or 0)
except cex.CosmosResourceNotFoundError:
return 0
except Exception: # noqa: BLE001
return 0
async def resp_mark_terminal(org: str, rid: str, status: str, ttl_s: int) -> dict:
"""One-way transition to a terminal status. Returns the record (with sid +
runner_turn_id) so the caller can propagate cancellation to the runner.
Idempotent: re-marking an already-terminal record just returns it."""
cont = await _container()
cex = _exceptions()
doc_id = f"resp::{org}::{rid}"
for _ in range(5):
try:
cur = await cont.read_item(doc_id, partition_key=_pk(org))
if str(cur.get("status")) in _TERMINAL:
return cur
cur.update(status=status, ts=time.time())
try:
await cont.replace_item(doc_id, cur, etag=cur["_etag"], match_condition=_IF_MATCH)
return cur
except cex.CosmosAccessConditionFailedError:
continue
except cex.CosmosResourceNotFoundError:
rec = {"id": doc_id, "pk": _pk(org), "kind": "resp", "status": status,
"sid": "", "runner_turn_id": "", "ts": time.time(), "ttl": int(ttl_s)}
try:
await cont.create_item(rec)
return rec
except cex.CosmosResourceExistsError:
continue
return {}
async def resp_is_cancelled(org: str, rid: str) -> bool:
try:
cont = await _container()
cex = _exceptions()
try:
cur = await cont.read_item(f"resp::{org}::{rid}", partition_key=_pk(org))
except cex.CosmosResourceNotFoundError:
return False
return str(cur.get("status")) == "cancelled"
except Exception:
return False
async def try_lock(name: str, ttl_s: int) -> bool:
"""One-shot distributed mutex via atomic create_item: returns True iff WE created the lock doc
(nobody else holds it), False if it already exists (another holder — skip this cycle). Per-item
TTL auto-expires the lock, so a holder that dies never wedges it. RAISES on a store error so the
caller can distinguish 'someone else has it' (skip) from 'store unavailable' (the caller's choice
to fall back). Used for single-flight periodic work (the winner runs it, losers skip)."""
cex = _exceptions()
try:
cont = await _container()
await cont.create_item({"id": f"lock::{name}", "pk": "org::_system", "kind": "lock",
"ts": time.time(), "ttl": int(ttl_s)})
return True
except cex.CosmosResourceExistsError:
return False
# ── API-key hot-read cache (HR-INF-010) ────────────────────────────────────────────────
# The public Bearer auth read (_apikey_resolve) hit the shared Gremlin partition on EVERY public
# request. These make the control store the hot-path read: a SELF-KEYED doc (pk == id == the token
# hash — the org is unknown pre-auth, so an org partition can't serve it) that is a write-through
# cache of the graph key record. The graph stays the source of truth (list/durable); dual-writes on
# mint/revoke keep this in sync (instant revocation, no TTL lag), and a lazy backfill on first
# resolve migrates pre-existing keys so Gremlin is never hit on the hot path afterward.
def _apikey_id(sha: str) -> str:
return f"apikey::{sha}"
# Per-item TTL on key docs: NOT the revocation mechanism (dual-write on revoke is, giving instant
# revocation) — it's a STALENESS BOUND. If a revoke's dual-write ever fails, the stale-valid cache
# doc expires within this window and the next resolve falls back to the authoritative graph
# (revoked). An active key is re-backfilled at most once per TTL (one Gremlin read/hour), negligible.
_APIKEY_TTL_S = int(os.environ.get("HR_APIKEY_TTL_S", "3600"))
async def apikey_get(sha: str) -> dict | None:
"""Point-read a key doc by hash. Returns None on miss OR on ANY error (incl. a missing SDK /
misconfig) — the caller falls back to the graph (the source of truth), so nothing here can
ever fail-close public auth."""
try:
cex = _exceptions()
cont = await _container()
try:
return await _retryable(lambda: cont.read_item(_apikey_id(sha), partition_key=_apikey_id(sha)))
except cex.CosmosResourceNotFoundError:
return None
except Exception: # noqa: BLE001 — store down / SDK missing → graph fallback
return None
async def apikey_put(sha: str, org: str, member: str, workspace: str,
workspace_default: bool, revoked: bool = False,
create_only: bool = False) -> None:
"""Write the key doc, TTL-bounded. mint uses upsert (authoritative fresh key). The lazy
BACKFILL uses create_only=True: a create_item that 409s (a doc already exists — possibly a
revoke tombstone written in the race window) leaves it untouched. A plain upsert here would
resurrect a just-revoked key for up to a TTL (HR-INF-010 review #1)."""
cont = await _container()
item = {"id": _apikey_id(sha), "pk": _apikey_id(sha), "kind": "apikey",
"org": org, "member": member, "workspace": workspace,
"workspace_default": bool(workspace_default), "revoked": bool(revoked),
"ts": time.time(), "ttl": _APIKEY_TTL_S}
if create_only:
cex = _exceptions()
try:
await _retryable(lambda: cont.create_item(item))
except cex.CosmosResourceExistsError:
return # a doc (maybe a tombstone) already exists — never clobber
else:
await _retryable(lambda: cont.upsert_item(item))
async def apikey_revoke(sha: str) -> None:
"""Mark a key revoked (dual-write on revoke). Upserts a revoked marker even for a key never
backfilled, so the hot read returns None immediately without consulting the graph. The revoked
marker gets a LONGER TTL than a live doc so it can't expire back to a graph re-read before the
graph revocation is certain to be visible."""
cont = await _container()
cex = _exceptions()
ttl = max(_APIKEY_TTL_S * 24, 86400) # keep the tombstone well past a live doc's lifetime
try:
cur = await _retryable(lambda: cont.read_item(_apikey_id(sha), partition_key=_apikey_id(sha)))
cur["revoked"] = True
cur["ttl"] = ttl
await _retryable(lambda: cont.upsert_item(cur))
except cex.CosmosResourceNotFoundError:
await _retryable(lambda: cont.upsert_item({
"id": _apikey_id(sha), "pk": _apikey_id(sha), "kind": "apikey", "revoked": True,
"ts": time.time(), "ttl": ttl}))
async def healthcheck() -> bool:
"""Readiness probe: can we actually reach the control store? Cheap point read
of a sentinel id (404 is a healthy answer — the store responded)."""
cont = await _container()
cex = _exceptions()
try:
await cont.read_item("health::probe", partition_key="org::_health")
except cex.CosmosResourceNotFoundError:
return True
return True
# azure-cosmos uses azure.core.MatchConditions for optimistic concurrency.
try: # imported lazily-safe so module import never fails when the SDK is absent
from azure.core import MatchConditions as _MC
_IF_MATCH = _MC.IfNotModified
except Exception: # pragma: no cover
_IF_MATCH = None
# ===== file: gateway/media_plane.py =====
"""Making a picture, a clip, a line of speech — and assembling them into one film.
Two callers, one engine, one place the provider key is resolved:
* the `media` MCP server this gateway hosts, so an agent can ask for a CAPABILITY
(text_to_video, text_to_image, …) and be told which model actually ran;
* the routes a media app calls to read its canvas, poll its jobs and export a film.
WHAT NEVER LEAVES THIS PROCESS. The provider key is resolved from the integrations document
inside the gateway, at the moment of use. The runner never receives it (an agent gets a per-
# ===== file: gateway/backing.py =====
"""Pluggable backing stores for the harness gateway.
The gateway needs three capabilities from the outside world, and nothing else:
GraphStore — session/harness/response/api-key records + ownership edges
BlobStore — traces, response records, workspace checkpoints, produced files
SecretStore — model-provider connections/policies ("bring your own keys")
Two implementations of each:
local (this file) — SQLite + filesystem + env/file, zero cloud dependencies. The whole
gateway runs on a laptop: HR_BACKING=local (or just leave
VG_GATEWAY_URL unset). Point POOL_MGMT_ENDPOINT at a single runner
process (the pool is only an identifier-routing proxy in front of
runner replicas — the routes are identical).
vg (backing_vg.py) — the production AgentStudio infra: vg-gateway (Cosmos graph + Azure
blob) and the vault service. Verbatim the code that always ran here.
Selection is one env var: HR_BACKING=vg|local; unset = vg when VG_GATEWAY_URL is set, local
otherwise. The interface is deliberately SEMANTIC (get/upsert/find/add_edge), not query-shaped:
call sites never build backend query strings, so an implementation is free to be a graph, a
relational table, or anything else that can look up records by label + property equality.
"""
from __future__ import annotations
import asyncio
import base64
import contextlib
import hashlib
import secrets as secrets_mod
import io
import json
import mimetypes
import os
import posixpath
import re
import shutil
import sqlite3
import tarfile
import time
import uuid
from pathlib import Path
from typing import Protocol
import httpx
# ── interfaces ────────────────────────────────────────────────────────────────────
class GraphStore(Protocol):
async def get(self, vid: str, label: str | None = None) -> dict | None: ...
async def upsert(self, label: str, vid: str, props: dict, *, raise_on_fail: bool = False) -> None: ...
async def add_edge(self, label: str, src: str, dst: str) -> None: ...
async def find(self, label: str, eq: dict | None = None, neq: dict | None = None) -> list[dict]: ...
async def delete(self, vid: str, label: str | None = None) -> bool:
"""Remove one vertex and its edges; False when there was none (or the label differs)."""
...
async def probe(self) -> None:
"""Raise if the store is unreachable/misconfigured. Used by /readyz, which must FAIL on a
broken plane — the normal read paths swallow errors and return empty."""
...
class BlobStore(Protocol):
async def get(self, kb: str, file_id: str) -> bytes | None: ...
async def put(self, kb: str, file_id: str, data: bytes) -> bool: ...
async def delete(self, kb: str, file_id: str) -> bool: ...
async def list(self, kb: str, prefix: str, limit: int = 20, cursor: str | None = None) -> dict: ...
async def probe(self) -> None: ...
class SecretStore(Protocol):
async def get(self, tenant: str, name: str) -> str | None: ...
async def delete(self, tenant: str, name: str) -> None:
"""Forget a stored secret; nothing happens when there is none."""
...
async def put(self, tenant: str, name: str, value: str, *,
require_encryption: bool = False) -> None:
"""`require_encryption=True` is the caller saying "this is a customer credential, not a
provider key injected by the operator" — a database connection string, for instance. A
store that cannot encrypt must raise SecretsNotConfigured rather than write it in the
clear. Part of the interface, not of one implementation: a datasource must be refused
the same way whichever backing is configured."""
...
class WorkspaceFiles(Protocol):
"""Reading and writing one file inside a session's workspace.
Two environments, genuinely different storage, one interface:
* **Self-hosted** — the runner shares this container and the workspace is a live directory on
the data volume. A write is a write.
* **Hosted** — sandboxes are ephemeral, so between turns a workspace exists ONLY as the
checkpoint tarball in blob storage (`sessions/<sid>/workspace.tgz`). A write has to rewrite
that archive, because there is no filesystem to write to until the next turn hydrates one.
This exists so an app built on a Harness can own state the agent also edits — the slides kit
keeps deck.json here — without either side needing to know which deployment it is running in.