fix(company): terminalize report-chain failure into FAILED settlement; prompt-spill hygiene

Report chain: when consecutive report cards exceed the failure limit the
parent now fails through transition_work_item (settlement, dependents,
manager visibility) instead of parking forever in AWAITING_MANAGER_REVIEW
behind a metadata hold no reconcile path could clear. The hold stamp
survives only as the quarantine fallback when the FAILED write does not
land, mirroring the attempt ledger's terminalize pattern.

Cursor prompt spill: one stable file per task (retries overwrite instead
of accumulating) and a self-ignoring .gitignore so workspace git never
picks up .opc/external_prompts.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
LZH-YS1998
2026-07-31 19:53:13 +08:00
parent 18ddc18948
commit 5f5fbf432e
4 changed files with 121 additions and 17 deletions
+68 -13
View File
@@ -6890,7 +6890,18 @@ class CompanyWorkItemExecutor:
*,
work_items: list[DelegationWorkItem] | None = None,
) -> bool:
"""True when reconcile must stop minting new report cards for parent."""
"""True when reconcile must stop minting new report cards for parent.
Over the failure limit the parent is terminalized to FAILED so the
normal settlement machinery (dependents refresh, manager visibility,
rework/resume) takes over a parent parked in
AWAITING_MANAGER_REVIEW with no spawnable report card can never
advance on its own. The ``report_chain_hold`` stamp survives only as
the quarantine fallback when even the FAILED write does not land,
mirroring the attempt ledger's terminalize pattern.
"""
if getattr(parent_item, "phase", None) in DONE_PHASES:
return True
parent_metadata = dict(getattr(parent_item, "metadata", {}) or {})
if str(parent_metadata.get("report_chain_hold", "") or "").strip():
return True
@@ -6908,31 +6919,75 @@ class CompanyWorkItemExecutor:
limit = self._report_failure_limit(parent_item)
if consecutive_failures < limit:
return False
hold_reason = (
f"{consecutive_failures} consecutive report cards failed "
f"(limit {limit}); refusing to spawn another Report #N until "
"the parent leaves AWAITING_MANAGER_REVIEW or the hold is cleared"
block_reason = (
f"report_chain_failure: {consecutive_failures} consecutive report "
f"cards failed (limit {limit})"
)
summary_text = (
"Work item failed because its report handoff pipeline is broken: "
f"{consecutive_failures} consecutive report cards died before a "
f"durable report landed (limit {limit}). The execution result is "
"preserved on the card; rework retries the handoff."
)
try:
await self.store.update_delegation_work_item(
await transition_work_item(
self.store,
parent_item.work_item_id,
target_phase=Phase.FAILED,
reason="report_chain_failure",
summary=summary_text,
metadata_updates={
"report_chain_hold": "consecutive_report_failures",
"report_chain_hold_reason": hold_reason,
"report_chain_hold_at": datetime.now().isoformat(),
"report_chain_failed_attempts": consecutive_failures,
},
release_claim=True,
)
try:
await self.store.update_delegation_work_item(
parent_item.work_item_id,
blocked_reason=block_reason,
)
except Exception:
logger.opt(exception=True).debug(
"report_chain_failure blocked_reason write failed for {}",
parent_item.work_item_id,
)
await self._emit_progress(
f"[Company:{projection_id_for_work_item(parent_item)}] {summary_text}"
)
except InvalidPhaseTransition:
# A concurrent writer moved the parent (e.g. a racing manager
# approval). Skip minting this tick; the next reconcile pass
# re-evaluates from the fresh phase.
logger.opt(exception=True).debug(
"report_chain_failure terminalize lost a phase race for {}",
parent_item.work_item_id,
)
except Exception:
logger.opt(exception=True).warning(
"Failed to stamp report_chain_hold on "
f"work_item_id={parent_item.work_item_id}"
"report_chain_failure FAILED terminalize did not land for {}; "
"quarantining via report_chain_hold",
parent_item.work_item_id,
)
try:
await self.store.update_delegation_work_item(
parent_item.work_item_id,
metadata_updates={
"report_chain_hold": "consecutive_report_failures",
"report_chain_hold_reason": block_reason,
"report_chain_hold_at": datetime.now().isoformat(),
"report_chain_failed_attempts": consecutive_failures,
},
)
except Exception:
logger.opt(exception=True).error(
"report_chain_hold quarantine write also failed for {}",
parent_item.work_item_id,
)
await self._record_work_item_runtime_diagnostic(
code="report_chain_held_after_failures",
code="report_chain_failure_terminalized",
severity="error",
work_item=parent_item,
message=hold_reason,
message=summary_text,
details={
"consecutive_failures": consecutive_failures,
"limit": limit,
+8 -3
View File
@@ -3,7 +3,6 @@
from __future__ import annotations
import asyncio
import hashlib
import re
import shutil
from pathlib import Path
@@ -128,9 +127,15 @@ class CursorAdapter(ExternalAgentAdapter):
root = Path(".").resolve()
prompt_dir = root / ".opc" / "external_prompts"
prompt_dir.mkdir(parents=True, exist_ok=True)
# Self-ignoring directory: agents run `git add` in the workspace and
# must never commit spilled prompts.
gitignore_path = prompt_dir / ".gitignore"
if not gitignore_path.exists():
gitignore_path.write_text("*\n", encoding="utf-8")
task_id = str(getattr(task, "id", "") or "").strip() or "task"
digest = hashlib.sha256(prompt_text.encode("utf-8")).hexdigest()[:16]
prompt_path = prompt_dir / f"cursor_{task_id}_{digest}.md"
# One stable file per task: retries and resumes overwrite instead of
# accumulating a new file per prompt revision.
prompt_path = prompt_dir / f"cursor_{task_id}.md"
prompt_path.write_text(prompt_text, encoding="utf-8")
pointer = (
"Open and follow the complete task instructions in this file exactly:\n"
+11
View File
@@ -3612,6 +3612,17 @@ class ExternalAgentMonitoringTests(unittest.IsolatedAsyncioTestCase):
self.assertIn(str(prompt_file), str(cmd[-1]))
self.assertNotIn(prompt[:64], metadata["command"])
self.assertLess(len(cmd[-1].encode("utf-8")), CursorAdapter._ARGV_PROMPT_MAX_BYTES)
# Hygiene: one stable file per task (retries overwrite) and a
# self-ignoring directory so workspace git never picks it up.
self.assertEqual(prompt_file.name, "cursor_task-large.md")
_cmd2, metadata2 = adapter.build_interactive_invocation(
task, workspace_path=tmpdir
)
self.assertEqual(metadata2["prompt_file"], str(prompt_file))
self.assertEqual(
sorted(p.name for p in prompt_file.parent.iterdir()),
[".gitignore", "cursor_task-large.md"],
)
finally:
_cleanup_test_dir(tmpdir)
+34 -1
View File
@@ -13,7 +13,7 @@ from __future__ import annotations
import tempfile
import unittest
from pathlib import Path
from unittest.mock import AsyncMock, MagicMock
from unittest.mock import AsyncMock, MagicMock, patch
from opc.core.config import OPCConfig, RoleConfig
from opc.core.events import EventBus
@@ -1503,6 +1503,39 @@ class ReportFailureStormBrakeTests(unittest.IsolatedAsyncioTestCase):
)
)
refreshed = await self.store.get_delegation_work_item("wi-child")
self.assertEqual(refreshed.phase, Phase.FAILED)
self.assertIn("report_chain_failure", str(refreshed.blocked_reason or ""))
self.assertFalse(
str((refreshed.metadata or {}).get("report_chain_hold", "") or "").strip()
)
async def test_terminalize_write_failure_quarantines_via_hold(self) -> None:
parent = _build_child_work_item()
parent.phase = Phase.AWAITING_MANAGER_REVIEW
parent.metadata = {
**dict(parent.metadata or {}),
"review_owner_role_id": "cto",
"review_owner_seat_id": "seat::team::cto::cto",
"max_consecutive_report_failures": 3,
}
await self.store.save_delegation_work_item(parent)
for attempt in (1, 2, 3):
await self.store.save_delegation_work_item(self._failed_report(attempt))
with patch(
"opc.layer2_organization.company_mode.transition_work_item",
side_effect=RuntimeError("db write lost"),
):
await self.executor._reconcile_missing_review_chain(
await self.store.list_delegation_work_items("run-1")
)
self.assertIsNone(
await self.store.get_delegation_work_item(
report_work_item_id_for_attempt("wi-child", 4)
)
)
refreshed = await self.store.get_delegation_work_item("wi-child")
self.assertEqual(refreshed.phase, Phase.AWAITING_MANAGER_REVIEW)
self.assertEqual(
refreshed.metadata.get("report_chain_hold"),
"consecutive_report_failures",