fix(store): enforce runnable-phase claim release across all write paths

Work items re-entering a runnable phase (ready/ready_for_rework) now have
ownership released unconditionally at the store layer: claim CAS no longer
consults metadata mirror keys (columns are the only ownership truth),
update_delegation_work_item blanks columns+mirror on any runnable-phase
write, review REJECT resolution releases ownership when the target phase is
runnable, and the startup sweep also covers runnable-phase residue. Closes
the 0011 rework livelock (stale four-field claim CAS vs. un-released claim).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
LZH-YS1998
2026-07-28 10:55:31 +08:00
parent 6fc5ad6be9
commit 7ae469876c
3 changed files with 255 additions and 23 deletions
+47 -4
View File
@@ -74,8 +74,10 @@ from opc.layer2_organization.phase import (
IN_PROGRESS_PHASES,
IN_REVIEW_PHASES,
InvalidPhaseTransition,
RUNNABLE_PHASES,
TODO_PHASES,
coerce_phase,
is_runnable,
is_stale_claim_releasable,
is_terminal,
kanban_column,
@@ -4579,7 +4581,11 @@ class OPCStore:
phase = coerce_phase(phase_str)
except (TypeError, ValueError):
continue
if not is_stale_claim_releasable(phase):
# Runnable phases are covered too: the phase-write invariant
# keeps READY / READY_FOR_REWORK rows unowned, so any claim
# found on one is corrupt legacy state that would starve the
# claim CAS forever.
if not (is_stale_claim_releasable(phase) or is_runnable(phase)):
continue
metadata = _json_loads(metadata_json, {})
metadata["claimed_by_role_session_id"] = ""
@@ -5465,6 +5471,18 @@ class OPCStore:
if metadata_updates:
metadata.update(dict(metadata_updates))
item.metadata = metadata
if phase is not None and item.phase in RUNNABLE_PHASES:
# Invariant: a fresh-runnable card is unowned. The claim CAS
# refuses cards with any leftover claim, so entering READY /
# READY_FOR_REWORK must release ownership in the same write —
# otherwise the card is permanently unclaimable and the
# dispatcher livelocks (project-0011 wedge).
item.claimed_by_role_runtime_session_id = ""
item.claimed_by_seat_id = ""
metadata = dict(item.metadata or {})
metadata["claimed_by_role_session_id"] = ""
metadata["claimed_task_id"] = ""
item.metadata = metadata
item.updated_at = datetime.now()
await self.save_delegation_work_item(item)
return item
@@ -5485,6 +5503,12 @@ class OPCStore:
shutdown transition. Keeping the phase, claim, queue, and durable
hold predicates in the same UPDATE prevents that stale snapshot from
resurrecting a suspended WorkItem.
Ownership truth is the two claim columns. The metadata mirror keys
(``claimed_by_role_session_id`` / ``claimed_task_id``) are written for
observability but must never gate the claim: a mirror key that a
release path forgot to blank would make a runnable card permanently
unclaimable (the project-0011 dispatcher livelock).
"""
phase = coerce_phase(expected_phase)
@@ -5528,8 +5552,6 @@ class OPCStore:
AND phase = ?
AND COALESCE(claimed_by_role_runtime_session_id, '') = ''
AND COALESCE(claimed_by_seat_id, '') = ''
AND COALESCE(json_extract(metadata, '$.claimed_by_role_session_id'), '') = ''
AND COALESCE(json_extract(metadata, '$.claimed_task_id'), '') = ''
AND COALESCE(json_extract(metadata, '$.dispatch_hold'), '') = ''
AND COALESCE(json_extract(metadata, '$.queued_behind_session'), '') = ''
AND COALESCE(
@@ -5602,12 +5624,22 @@ class OPCStore:
expected_source = str(source_report_work_item_id or "").strip()
db = self._require_db()
# A rework verdict sends the card back to the dispatch queue; the
# claim CAS refuses owned cards, so ownership must be released in
# the same UPDATE that writes the runnable phase (project-0011
# wedge: REJECT kept the worker's claim and the card became
# permanently unclaimable).
release_ownership = target in RUNNABLE_PHASES
for _attempt in range(3):
item = await self.get_delegation_work_item(work_item_id)
if item is None or item.phase != Phase.AWAITING_MANAGER_REVIEW:
return None
metadata = dict(item.metadata or {})
metadata.update(dict(metadata_updates or {}))
if release_ownership:
metadata["claimed_by_role_session_id"] = ""
metadata["claimed_task_id"] = ""
if (
self._metadata_has_work_item_projection_identity(metadata)
or str(item.projection_id or "").strip()
@@ -5622,9 +5654,18 @@ class OPCStore:
)
previous_updated_at = item.updated_at.isoformat()
updated_at = datetime.now()
claimed_session = (
"" if release_ownership
else str(item.claimed_by_role_runtime_session_id or "")
)
claimed_seat = (
"" if release_ownership else str(item.claimed_by_seat_id or "")
)
cursor = await db.execute(
"""UPDATE delegation_work_items
SET phase = ?, blocked_reason = ?, metadata = ?, updated_at = ?
SET phase = ?, blocked_reason = ?, metadata = ?, updated_at = ?,
claimed_by_role_runtime_session_id = ?,
claimed_by_seat_id = ?
WHERE work_item_id = ?
AND phase = ?
AND updated_at = ?
@@ -5651,6 +5692,8 @@ class OPCStore:
str(blocked_reason or ""),
_json_dumps(metadata),
updated_at.isoformat(),
claimed_session,
claimed_seat,
work_item_id,
Phase.AWAITING_MANAGER_REVIEW.value,
previous_updated_at,
+16 -19
View File
@@ -179,14 +179,13 @@ async def transition_work_item(
if release_claim:
# Fold claim release into the same write as the phase change: the
# legacy two-call sequence could commit the phase and then fail the
# release, stranding a dead claim. Terminal phases additionally drop
# the metadata claim mirror keys so the row can never satisfy a
# future claim-CAS predicate by accident.
# release, stranding a dead claim. The metadata mirror keys are
# blanked alongside the columns so the mirror never outlives the
# ownership it mirrors.
kwargs["claimed_by_role_runtime_session_id"] = ""
kwargs["claimed_by_seat_id"] = ""
if phase in DONE_PHASES:
merged.setdefault("claimed_by_role_session_id", "")
merged.setdefault("claimed_task_id", "")
merged.setdefault("claimed_by_role_session_id", "")
merged.setdefault("claimed_task_id", "")
try:
result = await store.update_delegation_work_item(work_item_id, **kwargs)
except InvalidPhaseTransition:
@@ -1147,22 +1146,20 @@ async def refresh_dependents_for_run(
elif work_item.phase == Phase.RUNNING:
target_phase = Phase.WAITING_FOR_CHILDREN
metadata_updates["waiting_on_work_item_ids"] = dependency_ids
# Clear the parent claim whenever the parent truly leaves
# WAITING_FOR_CHILDREN toward a non-terminal phase. The old
# condition ("only when all children approved AND target is
# RUNNING") left a gap: when a child went READY_FOR_REWORK,
# the refresh now fires (per _DEPENDENT_REFRESH_TARGETS) but
# the parent stayed in WAITING_FOR_CHILDREN with a stale claim,
# so the dispatcher couldn't re-pick it even though the child
# was back on the worker's queue.
# We exclude DONE_PHASES because for terminal parents the
# claim is a historical audit record of "last executor".
# Waking a parent out of WAITING_FOR_CHILDREN must orphan it so
# the dispatcher can re-pick it. For READY targets the store's
# phase-write invariant releases ownership; only the direct
# RUNNING wake still needs the explicit clear (columns and
# mirror in the same write). Terminal targets keep the claim as
# a historical audit record of "last executor".
clear_claim_on_wake = (
work_item.phase == Phase.WAITING_FOR_CHILDREN
and target_phase != work_item.phase
and target_phase not in DONE_PHASES
and target_phase == Phase.RUNNING
)
if target_phase != work_item.phase or metadata_updates or clear_claim_on_wake:
if clear_claim_on_wake:
metadata_updates["claimed_by_role_session_id"] = ""
metadata_updates["claimed_task_id"] = ""
if target_phase != work_item.phase or metadata_updates:
try:
await store.update_delegation_work_item(
work_item.work_item_id,