Incident (2026-08-17, same session as #320)
After the terminal wake of dag_fe5ff13ada6fUONA2rH3Pypdf4 was injected at 17:25:29, the parent turn ran ~10 min (authoring the wave-2 replan spec) when the process restarted at 17:35:22. Because wake_reported is persisted only AFTER yield* delivered.value (the whole wake-driven turn completes) — while the transcript part is written at admit time — the mark never landed. The new process's startup sweep re-injected the byte-identical 4391-char wake (diff-verified), because the in-memory dedup map deliveredWakeSummaries does not survive restarts. This is the hole the GOAL-FP-01-14 comment in loop.ts admits.
Design (grill-settled)
- Semantics change: admit success = delivered.
markWakeBatchReported moves to right after admitIfIdle succeeds (durable prompt queued + synthetic part persisted), before awaiting the wake turn.
- Drop "turn failure redelivers": the synthetic part stays in transcript history anyway; redelivery added duplicate injections, never information.
- Restart window shrinks from minutes (turn duration) to microseconds (admit→mark).
- Keep in-memory dedup map for same-process mark retries; unregister of terminal workflows moves with the mark.
- Update stale GOAL-FP-01-14 rationale comments; adapt any existing dag-wake-integration test that asserts turn-failure redelivery (semantic flip is intentional).
Acceptance
- Regression test: wake admitted → turn fails/interrupted →
wake_reported=1, no re-injection on next trigger; simulated restart (fresh loop instance over same store) does not re-deliver
test/dag/ full suite green, bun typecheck clean
Status
Worktree /private/tmp/oc-dag-wake-durability, branch fix/wake-durability (based on dev @16f9add17). Pending: red test, implementation, regression, commit. PR title: fix(dag): mark wake reported at admit time.
Incident (2026-08-17, same session as #320)
After the terminal wake of
dag_fe5ff13ada6fUONA2rH3Pypdf4was injected at 17:25:29, the parent turn ran ~10 min (authoring the wave-2 replan spec) when the process restarted at 17:35:22. Becausewake_reportedis persisted only AFTERyield* delivered.value(the whole wake-driven turn completes) — while the transcript part is written at admit time — the mark never landed. The new process's startup sweep re-injected the byte-identical 4391-char wake (diff-verified), because the in-memory dedup mapdeliveredWakeSummariesdoes not survive restarts. This is the hole the GOAL-FP-01-14 comment inloop.tsadmits.Design (grill-settled)
markWakeBatchReportedmoves to right afteradmitIfIdlesucceeds (durable prompt queued + synthetic part persisted), before awaiting the wake turn.Acceptance
wake_reported=1, no re-injection on next trigger; simulated restart (fresh loop instance over same store) does not re-delivertest/dag/full suite green,bun typecheckcleanStatus
Worktree
/private/tmp/oc-dag-wake-durability, branchfix/wake-durability(based on dev @16f9add17). Pending: red test, implementation, regression, commit. PR title:fix(dag): mark wake reported at admit time.