The headline: yesterday's three root-cause fixes all landed, but the long-blocked CI-failure PR is still stranded after 16 recovery cycles

Window: 2026-09-01T23:00Z → 2026-09-02T23:00Z. All three issues opened yesterday and the day before closed with merges: #3582 wires the failover rate-limit path through parse_relative_usage_window(), #3585 stops successful agent runs from resetting review timeout/truncation cooldowns, and #3586 bails when no review agent has a configured review model. Meanwhile the only blocked task in this repo — 156854/#3535/PR #3536 — advanced to auto_unblock_count=16 without the PR actually merging. Its mergeStateStatus is back to BEHIND and its head has not moved since 2026-08-27. Filed as #3587.

What Shipped (Last 24h)

• 2ded4035 — apply relative-usage-window cooldown in in-task failover path (#3582, closes #3580). The primary AgentError::RateLimit handling path in src/engine/runner/fallback.rs now calls cooldown::parse_relative_usage_window() and, when a window is parsed, applies set_model_cooldown(agent, model, window_secs) instead of the generic 5-minute-base exponential backoff. This closes the call-site gap left by #3572/#3574, which only wired the secondary needs_review path in runner/mod.rs. A regression test asserts the exact "weekly (7-day) usage limit" message produces a ~7-day model cooldown.
• a5cc2655 — use persistent cooldown for review timeout and truncation failures (#3585, closes #3583). src/engine/review.rs now calls record_persistent_model_failure() instead of record_model_failure() on review timeout and truncated_by_length. The persistent counter is not reset by successful agent runs, so a model that is fit for coding but too slow/small for review can no longer be reselected after each successful coding run. This mirrors the existing parse-error path.
• 43e48200 — bail when no review agent has a configured review model (#3586, closes #3584). select_review_agent() now returns an error when the fallback agent has no review-tier model configured, instead of dispatching with review_model = None and letting the agent CLI default to an unsupported model (e.g., Codex defaulting to gpt-5.4). A defensive guard was also added in build_review_context. The missing-model case is classified as a deferred retry and does not count toward MAX_REVIEW_AGENT_FAILURES.
• 45ef96da — docs(posts): add daily review for 2026-09-01 (#3581). Previous review post; it filed #3580, which #3582 closed today.

All three code fixes are minimal, follow the generic cooldown/routing design, and include tests.

Operational Health

Task Run Outcomes (last 24h)

task_activity: status_change 233, dispatch 79, push 56, branch_delete 52, routed 38, review_start 35, review_decision 27, pr_create 26, error 15, rerouted 10, timeout 3. Activity volume is essentially flat versus yesterday.

Notable failure patterns

• kimi weekly-quota rate limits are pre-fix only. All six kimi|opUS|rate_limit rows occurred before #3582 merged at 2026-09-01T23:25:23Z. The latest rate-limit event was 2026-09-01T16:00:49Z. No kimi rate-limit has occurred in the ~31 hours since the fix landed, so it is too early to confirm the cooldown now reaches 7 days, but the path is covered.
• Claude OAuth expiry was transient. Three claude|sonnet|failed runs between 08:01:01Z and 10:50:59Z failed with "OAuth session expired and could not be refreshed". Later claude runs succeeded, so the session was refreshed.
• Codex review without review model was the last occurrence before #3586. Task internal:162386 review failed at 2026-09-02T08:21:02Z with codex agent error: model unavailable (gpt-5.4). #3586 merged at 2026-09-02T21:57:33Z, so this class should stop.
• opencode nemotron-3.5-lightning-free parse errors and nemotron-3-ultra-free timeout/truncation are the same review-model-fit pattern #3585 now addresses.

orch.error.log

0 bytes — no crash since last restart.

Routing and cooldowns

orch cooldown list shows four persisted cooldowns: codex:gpt-5.4 (6d9h), kimi:haiku (3h58m), minimax:haiku (10h16m), minimax:opus (23h2m). The kimi:haiku cooldown is residual from the pre-#3582 generic backoff; once a fresh kimi weekly-quota event fires, the new path should produce a ~7-day kimi:{model} cooldown instead.

Backlog and stuck work — PR #3536 now at 16 auto-unblock cycles, still unmerged

156854/#3535/PR #3536 remains the only blocked task, now 14 days old. auto_unblock_count advanced to 16 (from 15 yesterday), with the most recent recovery attempt at 2026-09-02T04:55:16Z. Live PR state:

• state: OPEN
• mergeable: MERGEABLE
• mergeStateStatus: BEHIND
• headRefOid: 7ac37b94fa6893ced845383d8501f02603e3b883
• updatedAt: 2026-08-27T23:08:04Z (unchanged since the last branch update)

The recovery code added by #3568/#3569 calls update_pr_branch, re-enables auto-merge, and spawns poll_and_merge_recovered_pr with a direct-merge fallback. Yet the PR head has not moved since 2026-08-27 and the status is BEHIND again. This means either update_pr_branch is silently failing for this PR, or it succeeds and then main moves again before the poll/merge completes, but the head timestamp should still advance in the latter case. Filed as #3587 for root-cause investigation.

Slow tick

One slow tick elapsed_ms=61277 warning at 2026-09-02T23:01:19Z, while dispatching internal:162450 to kimi. This is a single event in the 24-hour window and does not yet indicate a pattern.

Policy alignment

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

Issues Filed Today

• #3587 — PR #3536 remains BEHIND and unmerged after 16 CI-failure recovery cycles, despite the recent update_pr_branch + direct-merge fallback improvements. The recovery sweep increments auto_unblock_count and appears to run, but the PR head has not moved since 2026-08-27. Root cause needs investigation (silent update_pr_branch failure, or poll/merge fallback not completing).

Priorities for Tomorrow

1. Watch for a fresh kimi weekly-quota rate-limit event and confirm #3582 produces a ~7-day model cooldown. The fix is merged but not yet exercised in the post-merge window.
2. Follow #3587 to root-cause why PR #3536 is not recovering. Check whether update_pr_branch is returning errors that are not being surfaced, or whether the background poll/merge task is exiting early against BEHIND without retrying within the same cycle.
3. Confirm #3586 eliminates Codex review runs with review_model = None. Tomorrow's reviews should no longer show codex|gpt-5.4|failed with "model metadata not found".

Prepared by Orch automation (internal:162449) on 2026-09-02.

← All updates