fix(external): keep session tokens across parks; only terminal failure clears them
PR #9 broadened token invalidation in two spots that also fired on non-terminal results, breaking resume continuity for parked runs: - _persist_session cleared the canonical role token and the task's resume pin on any result.status != DONE. An approval park (AWAITING_HUMAN) therefore wiped continuity right before the human approved, and the retry restarted the external thread from scratch. Both clears are now gated on _SESSION_INVALIDATING_RESULT_STATUSES (terminal failures only). - _stored_provider_token_allows_resume returned False whenever the newest row for a token was not done/suspended, so a run parked on approval vetoed its own token at the next restore. The verdict is now tri-state: terminally failed rows still return False (wipe), finalized rows return True, and live-but-unfinalized rows return None (keep the pin) — except provider_stream tokens, which keep the strict pre-existing rule via strict=True since an unfinalized stream row may belong to a crashed attempt. Regression tests: approval park keeps role token and task metadata; restore keeps a canonical token whose newest row is awaiting_human; unfinalized provider_stream tokens are still rejected. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
@@ -72,6 +72,11 @@ class ExternalAgentBroker:
|
||||
|
||||
_STREAM_READ_SIZE = 8192
|
||||
_MAX_PATH_HINT_TOKEN_LENGTH = 512
|
||||
# Result statuses that prove the provider attempt terminally failed, so
|
||||
# its session token must not stay pinned. Parks (awaiting_human /
|
||||
# awaiting_peer / awaiting_manager_review) and cancels keep the token:
|
||||
# those runs resume the same provider thread once the gate clears.
|
||||
_SESSION_INVALIDATING_RESULT_STATUSES = frozenset({TaskStatus.FAILED})
|
||||
_STREAM_SESSION_UPDATE_MIN_SECONDS = 2.0
|
||||
_STREAM_PROGRESS_MIN_SECONDS = 2.0
|
||||
_STREAM_TRANSCRIPT_HEAD_LINES = 40
|
||||
@@ -146,8 +151,16 @@ class ExternalAgentBroker:
|
||||
task: Task,
|
||||
role_session_id: str,
|
||||
token: str,
|
||||
strict: bool = False,
|
||||
) -> bool | None:
|
||||
"""Return the latest durable resumability verdict for one provider token."""
|
||||
"""Return the latest durable resumability verdict for one provider token.
|
||||
|
||||
``True`` means the newest row for the token finalized resumable,
|
||||
``False`` means the token is dead and must be cleared, ``None`` means
|
||||
there is no durable verdict either way (caller keeps the token).
|
||||
``strict`` treats an unfinalized newest row as ``False`` — required
|
||||
for provider_stream tokens whose run may have crashed mid-stream.
|
||||
"""
|
||||
|
||||
list_sessions = getattr(self.store, "list_external_sessions", None)
|
||||
if not callable(list_sessions):
|
||||
@@ -179,13 +192,21 @@ class ExternalAgentBroker:
|
||||
agent_type=adapter.agent_type,
|
||||
project_id=project_id,
|
||||
)
|
||||
return bool(
|
||||
selected is not None
|
||||
and selected_token == token
|
||||
and external_session_allows_resume(selected)
|
||||
and str(getattr(selected, "status", "") or "").strip().lower()
|
||||
in {"done", "suspended"}
|
||||
)
|
||||
if (
|
||||
selected is None
|
||||
or selected_token != token
|
||||
or not external_session_allows_resume(selected)
|
||||
):
|
||||
# The newest row for this token is terminally non-resumable.
|
||||
return False
|
||||
status = str(getattr(selected, "status", "") or "").strip().lower()
|
||||
if status in {"done", "suspended"}:
|
||||
return True
|
||||
# The newest row is alive but not finalized (running / awaiting_human /
|
||||
# awaiting_peer). An approval or peer park is not evidence the thread
|
||||
# is dead, so a canonical token keeps its pin; a provider_stream token
|
||||
# must not resume an attempt that never finalized.
|
||||
return False if strict else None
|
||||
|
||||
@classmethod
|
||||
def _task_explicitly_selected_external_agent(cls, task: Task, agent_type: str) -> bool:
|
||||
@@ -541,6 +562,8 @@ class ExternalAgentBroker:
|
||||
task=task,
|
||||
role_session_id=role_session_id,
|
||||
token=session_token,
|
||||
strict=str(entry.get("source", "") or "").strip()
|
||||
== "provider_stream",
|
||||
)
|
||||
if session_token
|
||||
else None
|
||||
@@ -2580,13 +2603,15 @@ class ExternalAgentBroker:
|
||||
)
|
||||
elif (
|
||||
role_session_id
|
||||
and result.status != TaskStatus.DONE
|
||||
and result.status in self._SESSION_INVALIDATING_RESULT_STATUSES
|
||||
and hasattr(self.store, "get_role_session_adapter_state")
|
||||
and hasattr(self.store, "update_role_session_adapter_state")
|
||||
):
|
||||
# A stream token is durable early so Stop can retain it. A normal
|
||||
# terminal failure must clear both a token discovered by this task
|
||||
# and an older role token that this failed attempt resumed.
|
||||
# A stream token is durable early so Stop can retain it, and a
|
||||
# park (awaiting_human / awaiting_peer) keeps it so the run can
|
||||
# resume the same thread once the gate clears. A terminal failure
|
||||
# must clear both a token discovered by this task and an older
|
||||
# role token that this failed attempt resumed.
|
||||
try:
|
||||
current = await self.store.get_role_session_adapter_state(
|
||||
role_session_id,
|
||||
@@ -2622,7 +2647,7 @@ class ExternalAgentBroker:
|
||||
logger.opt(exception=True).debug(
|
||||
"Failed to clear provider-stream role state after terminal failure"
|
||||
)
|
||||
if result.status != TaskStatus.DONE:
|
||||
if result.status in self._SESSION_INVALIDATING_RESULT_STATUSES:
|
||||
failed_token = str(
|
||||
resume_session_id
|
||||
or provider_session_id
|
||||
|
||||
@@ -309,6 +309,54 @@ class BrokerPersistWritesRoleStateTests(unittest.IsolatedAsyncioTestCase):
|
||||
"provider_terminal_failure",
|
||||
)
|
||||
|
||||
async def test_awaiting_human_park_keeps_role_token_and_task_metadata(self) -> None:
|
||||
# An approval park is not a terminal failure: the run resumes the
|
||||
# same provider thread once the human approves, so neither the role
|
||||
# token nor the task's resume pin may be cleared.
|
||||
await self._persist_run(
|
||||
task_id="task-old",
|
||||
agent_type="codex",
|
||||
resume_session_id="thread-live",
|
||||
)
|
||||
adapter = _MiniAdapter(agent_type="codex")
|
||||
adapter.config.session_mode = "resume"
|
||||
adapter.config.session_id = "thread-live"
|
||||
task = self._task(task_id="task-parked")
|
||||
task.metadata.update({
|
||||
"external_resume_session_id": "thread-live",
|
||||
"external_resume_session_scope_id": "sess-a",
|
||||
"external_resume_agent_type": "codex",
|
||||
})
|
||||
await self.broker._persist_session(
|
||||
adapter=adapter,
|
||||
task=task,
|
||||
workspace_path="/tmp/ws",
|
||||
metadata={
|
||||
"command": "codex exec resume",
|
||||
"model": "(cli default)",
|
||||
"resume_session_id": "thread-live",
|
||||
},
|
||||
result=TaskResult(
|
||||
status=TaskStatus.AWAITING_HUMAN,
|
||||
content="External action blocked by autonomy policy",
|
||||
artifacts={
|
||||
"resume_session_id": "thread-live",
|
||||
"requires_user_input": True,
|
||||
},
|
||||
),
|
||||
)
|
||||
|
||||
entry = await self.store.get_role_session_adapter_state(
|
||||
self.role_session_id,
|
||||
"codex",
|
||||
)
|
||||
self.assertIsNotNone(entry)
|
||||
self.assertEqual(entry["resume_session_id"], "thread-live")
|
||||
self.assertEqual(
|
||||
task.metadata.get("external_resume_session_id"), "thread-live"
|
||||
)
|
||||
self.assertNotIn("external_resume_fallback", task.metadata)
|
||||
|
||||
async def test_consecutive_tasks_overwrite_with_latest_token(self) -> None:
|
||||
await self._persist_run(
|
||||
task_id="task-1", agent_type="codex", resume_session_id="thread-1",
|
||||
@@ -571,6 +619,84 @@ class BrokerRestorePrefersRoleStateTests(unittest.IsolatedAsyncioTestCase):
|
||||
"provider_terminal_failure",
|
||||
)
|
||||
|
||||
async def test_restore_keeps_role_token_when_newest_row_is_parked_awaiting_human(self) -> None:
|
||||
# A canonical role token whose newest run parked on approval must
|
||||
# keep its pin: awaiting_human is not evidence the thread is dead,
|
||||
# and resume-after-approval depends on this token surviving.
|
||||
token = "thread-parked"
|
||||
await self.store.update_role_session_adapter_state(
|
||||
self.role_session_id,
|
||||
"codex",
|
||||
{
|
||||
"resume_session_id": token,
|
||||
"provider_session_id": token,
|
||||
"last_task_id": "task-old",
|
||||
},
|
||||
)
|
||||
await self.store.save_external_session(ExternalSession(
|
||||
agent_type="codex",
|
||||
project_id="proj1",
|
||||
session_id=token,
|
||||
opc_session_id=self.role_session_id,
|
||||
task_id="task-parked",
|
||||
workspace_path="/tmp/ws",
|
||||
run_mode="exec",
|
||||
status="awaiting_human",
|
||||
metadata={
|
||||
"resume_session_id": token,
|
||||
"provider_session_id": token,
|
||||
},
|
||||
))
|
||||
adapter = _MiniAdapter(agent_type="codex", can_resume_blank=False)
|
||||
task = self._task()
|
||||
|
||||
await self.broker._restore_session_resume_from_store(adapter, task)
|
||||
|
||||
self.assertEqual(adapter.config.session_mode, "resume")
|
||||
self.assertEqual(adapter.config.session_id, token)
|
||||
self.assertIsNotNone(await self.store.get_role_session_adapter_state(
|
||||
self.role_session_id,
|
||||
"codex",
|
||||
))
|
||||
|
||||
async def test_restore_still_rejects_unfinalized_provider_stream_token(self) -> None:
|
||||
# provider_stream tokens keep the strict pre-existing rule: a row
|
||||
# that never finalized (still "working") may belong to a crashed
|
||||
# attempt and must not be resumed.
|
||||
token = "thread-unfinalized"
|
||||
await self.store.update_role_session_adapter_state(
|
||||
self.role_session_id,
|
||||
"codex",
|
||||
{
|
||||
"resume_session_id": token,
|
||||
"provider_session_id": token,
|
||||
"last_task_id": "task-new",
|
||||
"source": "provider_stream",
|
||||
"status": "working",
|
||||
},
|
||||
)
|
||||
await self.store.save_external_session(ExternalSession(
|
||||
agent_type="codex",
|
||||
project_id="proj1",
|
||||
session_id=token,
|
||||
opc_session_id=self.role_session_id,
|
||||
task_id="task-new",
|
||||
workspace_path="/tmp/ws",
|
||||
run_mode="exec",
|
||||
status="working",
|
||||
metadata={
|
||||
"resume_session_id": token,
|
||||
"provider_session_id": token,
|
||||
},
|
||||
))
|
||||
adapter = _MiniAdapter(agent_type="codex", can_resume_blank=False)
|
||||
task = self._task()
|
||||
|
||||
await self.broker._restore_session_resume_from_store(adapter, task)
|
||||
|
||||
self.assertNotEqual(adapter.config.session_mode, "resume")
|
||||
self.assertEqual(adapter.config.session_id, "")
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
|
||||
Reference in New Issue
Block a user