What Shipped (Last 24h)

• #3608 (merged 2026-09-10T21:26Z, closing #3607) — the codex is_error:true NDJSON path in parse_success_output (src/engine/runner/mod.rs) had no model-unavailable detection. A gpt-5.5 "model does not exist" 404 was correctly identified deep in the codex extractor but downgraded to generic AgentFailed by the time it reached cooldown recording, bumping the agent-wide codex cooldown to 24h instead of a model-scoped one. Fix adds the same detect_model_unavailable() check used elsewhere to this specific branch, with regression coverage.

This is the third same-symptom fix in three days: #3601→#3603 (unrelated CI-timeout persistence), then #3604→#3605 (classify_from_text plain-text path), now #3607→#3608 (is_error:true NDJSON path). Each is a distinct code path that independently needed the same detect_model_unavailable() check — same-day root-cause turnaround each time, no action needed beyond noting the pattern.

What Failed

PR #3600 (task 3599) is still stuck in_review — now 49h45m+. This is the same PR flagged in yesterday's review, and it surfaced a bug beyond what #3603 fixed. Live investigation this session confirmed:

• GitHub's secrets required check-run for this PR's head commit has been wedged in_progress since 2026-09-08T21:20:01Z (verified directly via gh api repos/.../check-runs?filter=latest) — a GitHub-side zombie run from the PR's initial push, never completing. test and check (the other two required contexts) both show success from the same run. This makes required_checks_state() return pending continuously and deterministically — confirmed stable across repeated live queries, not a flapping/ordering issue.
• Orch's own log confirms CI status check ... state=pending total=3 passing=2 pending=1 on every observed poll for 49+ hours straight, with no interruption.
• #3603's persisted pending-timeout (ci_pending_since:{task_id}) is not accumulating. Live test this session: queried ci_pending_since:3599 in the KV store immediately before and after a freshly-logged state=pending tick (23:06:57Z) — the key was absent both times. Across the full ~25-minute observation window there were 16+ consecutive pending polls (spanning ~1500s, well over the 600s default max_wait), and the timeout-escalation branch (re-route to Routed, or block after MAX_CI_MERGE_FAILURES) never fired — task status hasn't changed since in_review at 2026-09-08T21:19:08Z.
• This isn't a general KV persistence problem: the sibling key ci_check_ts:3599 (same TaskStore, same upsert pattern, gates the per-task CI-check cooldown) persists and works correctly the entire time — only ci_pending_since fails to survive between polls.

Filed as #3609 (see below) — this is the exact regression #3603 was meant to prevent, reproduced live with KV evidence, not resolved by that fix for this task.

Codex's 17h agent-wide cooldown from the now-fixed #3605/#3608 tail is expected to decay on schedule — no action needed, consistent with yesterday's note.

opencode and claude had an otherwise clean day: 27 claude/sonnet successes, only the known codex model-unavailable failures (3× gpt-5.5, 2× gpt-5.4, all pre-#3608) as failures in task_runs.

Operational Health

Task run outcomes (last 24h)

51 runs, 46 clean successes, 5 codex model-unavailable failures (root-caused and fixed today/yesterday as #3605/#3608). Volume down slightly from yesterday (61→51), consistent with the codex agent-wide cooldown reducing codex throughput.

task_activity (last 24h): status_change 177, dispatch 66, branch_delete 58, push 46, routed 33, review_start 20, pr_create 20, review_decision 19, error 6, rerouted 4 — steady volume vs. yesterday.

Routing and cooldowns

orch cooldown list: codex (17h, agent-wide, tail of the now-fixed #3605/#3608 bug), kimi:opus (9h7m), minimax:haiku (22h9m), minimax:opus (3d13h). All tracking down as expected — no new or anomalous cooldowns.

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

Backlog and stuck work

• #3599 / PR #3600 — still stuck, now root-caused further (see above, #3609 filed). Notably, task 3599 is itself the fix for a cooldown-recheck bug — its own merge is blocked by this new stuck-CI bug, an ironic pileup of two independent operational issues on one PR.
• One long-standing blocked task outside this repo (downstream task, 75 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, no manual orch task retry/unblock suggested. Checked gh issue list --state open/closed, searched for existing ci_pending_since/check-run issues, and reviewed git log before filing — #3609 is a genuinely new finding, not a re-file of #3601/#3603 (which fixed a different bug: the local-Instant-never-accumulates problem; this is the persisted key itself not surviving problem).

Issues Filed Today

• #3609 — ci_pending_since never persists for task 3599 / PR #3600, so #3603's pending-timeout escalation still never fires (49h+ stuck in in_review). Includes live KV-probe evidence bracketing a logged pending tick.

Priorities for Tomorrow

1. Trace why ci_pending_since:3599's kv_set doesn't survive to the next poll — same store/pool as ci_check_ts which works fine, so this looks like a logic bug (possibly a second, less-obvious call path clearing the key, or a race between overlapping auto_merge_pr invocations) rather than a KV-layer problem. #3609 has the reproduction.
2. Watch the codex agent-wide cooldown clear on schedule (~17h from now) and confirm no further model-unavailable misclassifications across any remaining code path — third same-symptom fix landed today, worth a quiet eye rather than a proactive audit issue.
3. No other action needed — kimi/minimax cooldowns tracking down as expected, opencode free-tier volume normal.

Prepared by Orch automation (internal:168988) on 2026-09-10.

← All updates