Files
openworker/tests/test_observability_engine.py
Devika Verma ff9b955494 fix(observability): a streamed model step ends as its reply arrives; a turn's end waits for shadow checks (OPE-244)
Two races found by running the schema test repeatedly:

- The streamed view ended its model step only after the engine had read
  the final piece. The engine could add the reply and end the turn first,
  and the turn's end then closed the step as interrupted, without its
  usage or finish reason. The step now ends when the final piece
  arrives, before the engine sees it.
- A shadow check runs in the background and can outlast the turn's last
  event. The turn's end is now recorded with the time of its ending
  event but sent once running shadow checks finish (at most 30 s), so
  their steps are not closed as interrupted while they run.

Also: a server test's fake provider answers the session title apart, so
the title call can no longer take the turn's only reply.
2026-11-02 10:50:08 +05:30

194 lines
7.7 KiB
Python

"""Observability E1: a real engine turn, recorded.
The engine sends `turn_started` / `turn_ended` around run, retry and resume, a tool step for
every call that runs (from `_execute_sync` to the raw result), a zero-length tool step for a
call that never ran, and ids on its UI events. The fold rules are checked by the golden test;
this checks that the engine sends the right events.
"""
from __future__ import annotations
import asyncio
import aisuite as ai
from coworker.engine import TurnEngine
from coworker.events import EventType
from coworker.observability import Recorder
from coworker.permissions import PermissionEngine
from coworker.providers import AssistantTurn, ModelCapabilities, ProviderClient, ToolCall
from coworker.tools import ToolRegistry
class ScriptedProvider(ProviderClient):
def __init__(self, turns):
self._turns = list(turns)
def complete(self, *, model, messages, tools=None, **settings):
turn = self._turns.pop(0)
if isinstance(turn, Exception):
raise turn
return turn
def capabilities(self, model):
return ModelCapabilities()
def _text(text):
return AssistantTurn(text=text, finish_reason="stop")
def _tools(*calls):
return AssistantTurn(
tool_calls=[ToolCall(id=cid, name=name, arguments=args) for cid, name, args in calls],
finish_reason="tool_calls",
)
def _engine(tmp_path, turns):
registry = ToolRegistry()
registry.register_all(ai.toolkits.files(root=str(tmp_path), allow_write=True))
engine = TurnEngine(
provider=ScriptedProvider(turns),
registry=registry,
permissions=PermissionEngine(workspace_root=tmp_path),
model="gpt-5.5",
)
engine.audit_context = {"session_id": "sess_1"}
rec = Recorder()
events: list[dict] = []
rec.add_listener(events.append)
engine.recorder = rec
return engine, rec, events
def _drain(gen):
async def go():
return [ev async for ev in gen]
return asyncio.run(go())
def test_a_turn_with_tools_is_recorded(tmp_path):
(tmp_path / "a.txt").write_text("hello", encoding="utf-8")
(tmp_path / "b.txt").write_text("world", encoding="utf-8")
engine, rec, _ = _engine(tmp_path, [
_tools(("c1", "read_file", {"path": "a.txt"}), ("c2", "read_file", {"path": "b.txt"})),
_text("done"),
])
ui = _drain(engine.run("read both", trigger="user"))
d = rec.as_dict()
[turn] = d["turns"]
user = next(m for m in engine.messages if m["role"] == "user")
assert (turn["kind"], turn["trigger"], turn["status"], turn["session_id"]) == (
"run", "user", "completed", "sess_1")
assert turn["first_message_id"] == user["mid"]
assert turn["last_message_id"] == engine.messages[-1]["mid"]
assert turn["approval_mode"] == engine.permissions.mode.value
assert turn["reviewer"] == {"active": False, "shadow": False, "kind": None, "model": None}
assert turn["duration_ms"] is not None
tools = [s for s in d["steps"] if s["kind"] == "tool"]
assert [(s["call_id"], s["call_index"], s["status"]) for s in tools] == [
("c1", 0, "ok"), ("c2", 1, "ok")]
by_call = {m.get("tool_call_id"): m["mid"] for m in engine.messages if m["role"] == "tool"}
for s in tools:
assert s["message_id"] == by_call[s["call_id"]]
assert s["attrs"]["risk_class"] == "read"
assert s["attrs"]["result_bytes"] > 0
assert s["duration_ms"] is not None and s["turn_id"] == turn["turn_id"]
# UI events carry the turn id, a time, and a tool's call and step ids.
assert all(e.data.get("turn_id") == turn["turn_id"] for e in ui)
started = [e for e in ui if e.type is EventType.TOOL_STARTED]
finished = [e for e in ui if e.type is EventType.TOOL_FINISHED]
assert {e.data["tool_call_id"] for e in started} == {"c1", "c2"}
assert {e.data["step_id"] for e in finished} == {s["step_id"] for s in tools}
assert all(isinstance(e.data["ts_ms"], int) for e in ui)
def test_every_step_ended_carries_its_identity(tmp_path):
(tmp_path / "a.txt").write_text("hello", encoding="utf-8")
engine, _, events = _engine(tmp_path, [_tools(("c1", "read_file", {"path": "a.txt"})), _text("ok")])
_drain(engine.run("go"))
ends = [e for e in events if e["type"] == "step_ended"]
assert ends
for e in ends:
assert {"v", "at_ms", "mono_ms", "kind", "name", "call_id", "started_at"} <= set(e)
assert [e["type"] for e in events][0] == "turn_started"
assert [e["type"] for e in events][-1] == "turn_ended"
def test_a_call_that_never_ran_is_a_zero_length_step(tmp_path):
engine, rec, _ = _engine(tmp_path, [
_tools(("c1", "read_file", {"_raw": "{not json"})),
_text("sorry"),
])
_drain(engine.run("go"))
[step] = [s for s in rec.as_dict()["steps"] if s["kind"] == "tool"]
assert (step["call_id"], step["status"], step["duration_ms"]) == ("c1", "error", 0)
tool_msg = next(m for m in engine.messages if m.get("tool_call_id") == "c1")
assert step["message_id"] == tool_msg["mid"]
def test_a_failed_model_call_ends_the_turn_as_error_and_retry_is_its_own_turn(tmp_path):
engine, rec, _ = _engine(tmp_path, [RuntimeError("provider down"), _text("back")])
_drain(engine.run("hello", trigger="user"))
_drain(engine.retry(trigger="user"))
first, second = rec.as_dict()["turns"]
assert (first["kind"], first["status"]) == ("run", "error")
assert (second["kind"], second["status"], second["seq"]) == ("retry", "completed", 2)
# A retry opens with the first message it adds: the model's reply.
reply = engine.messages[-1]
assert second["first_message_id"] == reply["mid"] == second["last_message_id"]
def test_an_abandoned_turn_is_closed_as_unknown(tmp_path):
(tmp_path / "a.txt").write_text("hello", encoding="utf-8")
engine, rec, _ = _engine(tmp_path, [_tools(("c1", "read_file", {"path": "a.txt"})), _text("ok")])
async def go():
gen = engine.run("go")
async for ev in gen:
if ev.type is EventType.TURN_START:
break
await gen.aclose() # the surface stopped reading
asyncio.run(go())
[turn] = rec.as_dict()["turns"]
assert turn["status"] == "unknown" and turn["duration_ms"] is not None
def test_no_recorder_means_no_ids_on_events(tmp_path):
engine, _, _ = _engine(tmp_path, [_text("hi")])
engine.recorder = None
ui = _drain(engine.run("hello", trigger="user"))
assert all("turn_id" not in e.data for e in ui)
def test_a_turn_through_the_server_lands_in_recordings_db(tmp_path):
from fastapi.testclient import TestClient
from coworker.server import SessionManager, create_app
class WithTitle(ScriptedProvider):
def complete(self, *, model, messages, tools=None, **settings):
if str(messages[0].get("content") or "").startswith("You title"):
return _text("Greeting") # the session title, which can run before the turn's call
return super().complete(model=model, messages=messages, tools=tools, **settings)
manager = SessionManager(workspace=tmp_path, provider=WithTitle([_text("hi there")]))
client = TestClient(create_app(manager))
with client.websocket_connect("/ws/session/obs1") as ws:
assert ws.receive_json()["type"] == "ready"
ws.send_json({"type": "user_message", "text": "hello"})
while ws.receive_json()["type"] != "turn_done":
pass
[turn] = manager.recordings_store.load_recording("obs1")["turns"]
assert (turn["kind"], turn["trigger"], turn["status"]) == ("run", "user", "completed")
assert manager.recordings_store.path.parent == manager.audit_store.db_path.parent
manager.delete_session("obs1")
assert manager.recordings_store.load_recording("obs1") == {"turns": [], "steps": []}