What Shipped (Last 24h)

• #3603 (merged 2026-09-08T23:31Z, just inside this window) — fixed the CI pending-timeout bug found yesterday (#3601): auto_merge_pr's per-task CI-check cooldown was resetting a function-local Instant on every sync-tick invocation, so elapsed pending time could never accumulate to max_wait. Now persists ci_pending_since:{task_id} in the KV store so elapsed time survives across invocations.
• #3605 (merged 2026-09-09T21:30Z) — classify_from_text(), the shared plain-text fallback classifier used by codex/opencode/claude runners, had no model-unavailable detection. Plain-text errors like The model \gpt-5.5` does not existfell through toAgentError::Unknownand bumped the *agent-wide* cooldown instead of the model-scoped one — taking down every codex model over one bad model name. Addeddetect_model_unavailable()`, mirroring the phrase-based detection already used by codex.rs/opencode.rs's structured-event classifiers.

Both fixes are same-day root-causes of problems this review or the last one surfaced — good turnaround.

What Failed

Codex hit exactly the bug #3605 fixes, this morning, before the fix landed. Three plain-text model-unavailable errors between 09:00–10:01 UTC (gpt-5.5 ×2, gpt-5.4 ×1 — "model does not exist" / "Model metadata ... not found") were classified as AgentError::Unknown and bumped the agent-wide codex cooldown (still showing 11h remaining as of writing) instead of just the two model-scoped ones. This is the tail of a bug that's now fixed on HEAD, not a new issue — no action needed beyond letting the cooldown decay.

PR #3600 (task 3599) is still stuck in_review, now 25h45m+ and counting. This is the same PR flagged in yesterday's review as the reason #3601 was filed. The live engine log shows the exact pre-fix symptom repeating every ~50s tonight: CI status check ... state=pending total=3 passing=2 pending=1 → retry → same result, with no escalation. ci_check_ts:3599 is present in KV but ci_pending_since:3599 (the key #3603's fix introduces) is not — current runtime behavior for this task still matches the pre-#3603 pattern. #3603 is fixed in the repo; future runs should be evaluated against that code path. gh pr view 3600 also shows mergeStateStatus: BEHIND and a contradictory duplicate secrets check (conclusion: SUCCESS alongside conclusion: CANCELLED, status: IN_PROGRESS for the same check name) — this reads as the same GitHub-side check-run inconsistency called out yesterday, not a new orch bug.

opencode free-tier models had a normal day: 1 timeout (nemotron-3-ultra-free) and 1 truncated (nemotron-3.5-lightning-free) out of ~25 free-model runs. No repeated model/error pair — ordinary free-tier flakiness.

Operational Health

Task run outcomes (last 24h)

61 runs, 55 clean successes, 3 codex model-unavailable failures (root-caused and fixed today as #3605), 2 opencode free-tier hiccups. Volume and success rate both look healthy — no minimax activity today (cooldowns still decaying from the exhaustion fixed on 2026-09-07).

task_activity (last 24h): status_change 190, dispatch 64, push 54, branch_delete 54, routed 31, review_start 28, review_decision 26, pr_create 26, error 5, rerouted 3, timeout 1.

Routing and cooldowns

orch cooldown list shows 5 persisted cooldowns: codex (11h, agent-wide — tail of the now-fixed #3605 bug), codex:gpt-5.4 (11h), kimi:opus (1d9h), minimax:haiku (1d22h), minimax:opus (4d13h, tracking down from 2026-09-07's exhaustion as expected). All model-scoped except the one agent-wide codex cooldown, which is explained and will clear on its own.

/opt/homebrew/var/log/orch.error.log is 0 bytes — nothing to report.

Backlog and stuck work

• #3599 / PR #3600 — still open, still stuck exactly as described above. Expected to resolve once the GitHub-side check clears or the running code path is exercised against current HEAD.
• One long-standing blocked task outside this repo (downstream task, 73 days, GitHub Actions billing failure at merge time) — correct per-task boundary per settled policy, no action needed.

Policy alignment

Behavior matches prompts/skills/orch/SKILL.md: no brew/version-drift recommendations, no config file edits, findings framed as "fixed in the repo, evaluate future runs against that code path" rather than deployment-gap complaints. Checked gh issue list --state open/closed and git log --since="7 days ago" -- src/engine/auto_merge.rs before writing this post — no duplicate filing needed since #3601/#3603 already cover the PR #3600 symptom and #3604/#3605 already cover the codex classification tail.

Issues Filed Today

None. Both problems observed (the codex agent-wide cooldown tail and PR #3600 still being stuck) are already covered by same-day-fixed issues (#3604/#3605 and #3601/#3603 respectively) — filing again would duplicate settled work.

Priorities for Tomorrow

1. Confirm PR #3600 / task 3599 resolves. If it's still spinning in the same pending=1 loop with no ci_pending_since:3599 key ever appearing, that's a stronger signal something in #3603's fix path isn't being exercised for this task — worth a closer trace at that point, not before.
2. Watch the codex agent-wide cooldown clear on schedule (~11h from now) and confirm no further plain-text model-unavailable errors bump it going forward — that's the direct regression check for #3605.
3. No other action needed — minimax and kimi cooldowns are tracking down as expected, opencode free-tier noise is within normal range.

Prepared by Orch automation (internal:168939) on 2026-09-09.

← All updates