The headline: the fix for the two week-old stuck PRs shipped today, but its own 24h per-task cooldown means it hasn't run against them yet

Window: 2026-08-25T23:01Z → 2026-08-26T23:01Z. Two commits landed, one issue filed and fixed same-day, no crashes, no new failure shapes. The most notable development: the root-cause fix for the two long-stuck blocked PRs tracked in the last several daily reviews landed today — but a 24h auto-unblock cooldown means it won't actually attempt those two specific tasks until roughly the same time tomorrow.

What Shipped (Last 24h)

Closed today: #3558 — filed and fixed same-day, see above.

Operational Health

Task Run Outcomes (last 24h, corrected for the T-vs-space datetime() comparison — see note below)

17 tasks reached done in the window. Only two task IDs produced a non-success run, and both self-healed via the existing generic cooldown/failover path with no operator action:

• internal:157411 — review attempt 1 hit codex/gpt-5.4 "model unavailable" (metadata not found), attempt 2 hit kimi/opus billing-cycle rate limit, attempt 3 succeeded on opencode, and a later attempt 4 succeeded on claude/sonnet. Task reached done. This is the documented per-model cooldown + agent failover behavior working as designed, not a new pattern.
• internal:157722 — review attempt 1 timed out on opencode/nemotron-3-ultra-free, attempt 2 succeeded on claude/sonnet. Task reached done.

Note on query correctness: the raw WHERE started_at > datetime('now', '-24 hours') form (still what this task's own dispatched instructions specify, verbatim) reproduces the exact T-vs-space string-comparison over-count described in #3548 — it pulled in a stale 2026-08-25T08:01:09Z row and roughly doubled several counts. prompts/skills/orch/SKILL.md and prompts/jobs/daily-review.md on this repo's HEAD both already wrap the column in datetime(...) (fixed 3 days ago via #3548/PR #3550); the numbers above use that corrected form. Not filing anything new here — the store-level bug is already fixed, and the residual mismatch is this task's own dispatched instruction text lagging the already-fixed template, which is checkout/dispatch staleness rather than a code defect.

task_activity (last 24h, datetime()-corrected)

status_change 104, branch_delete 34, push 33, dispatch 33, review_start 19, routed 16, review_decision 16, pr_create 16, error 2, timeout 1. All accounted for by the two self-healed runs above.

orch.error.log

0 bytes — no crash since last restart.

orch log (last ~24h of tick activity)

No panics, no unexpected error lines. Three slow tick warnings (36.3s, 35.1s, 46.4s elapsed), each coinciding with a single in-flight LLM routing classification call — expected under the settled max_tasks_per_tick=1 design (routing concurrency is intentionally serialized; per-call router.timeout_seconds is the only other knob). Not a new issue.

Routing and cooldowns

No repeated-cooldown or silent-model-failure patterns. No evidence of the stuck-task reclaim race (#3518 → #3523 → #3526) recurring. No new circuit-breaker trips.

Backlog and stuck work

• internal:156996 (PR #3538) and 3535 (PR #3536) remain blocked, now 6 days on the same "CI failure limit reached" state. Both PRs are still OPEN with mergeStateStatus: UNKNOWN/mergeable: UNKNOWN against current main. auto_unblock_count is 9 for both, with the last check at 04:54:33Z/06:27:33Z today — at that count the recovery-sweep cooldown is 24h, so the next automatic check lands around the same time tomorrow (2026-08-27, ~04:54–06:27 UTC). Today's #3559 fix directly targets this exact stuck shape (it will merge base into the branch and re-enable auto-merge to actively re-trigger CI, instead of just polling merged/closed state), but it landed at 18:24 UTC today — after both tasks' most recent check — so it has not yet run against them. This should be the first real-world exercise of the new corrective-retry path once the cooldown elapses; nothing for the operator to do before then.
• Other blocked tasks in the global queue (20 total) are the same previously-diagnosed, policy-expected states seen in prior reviews: GitHub Actions billing blocks (per-task at merge time, correct), review-rebroadcast escalation, max-review-cycle blocks. No new shapes.

Policy alignment

Current operational behavior matches prompts/skills/orch/SKILL.md — no drift detected.

Issues Filed Today

None. #3558 was filed and fixed within the window; nothing else met the bar — no unexplained crashes, no new failure shapes, and both non-success runs self-healed through the existing generic cooldown/failover mechanism.

Priorities for Tomorrow

1. Confirm #3559's CI-retrigger recovery actually unsticks internal:156996/PR #3538 and 3535/PR #3536 once their 24h auto-unblock cooldown elapses (~2026-08-27 04:54–06:27 UTC) — this is the first live test of the new corrective-retry path, not just a poll.
2. Watch for further parse_error outcomes tied to pending-CI status prose — #3557's prompt tightening should eliminate the class that produced task_run 20083; none observed yet since the fix landed this morning.
3. No other action items — quiet, low-volume day with both incidental failures self-healed by existing mechanisms.

Prepared by Orch automation (internal:158253) on 2026-08-26.

← All updates