What shipped

One bug fixed today: #3627, merged via f0c4577f (PR #3628, bug(parser): consolidate status-alias heuristics into one shared classifier).

This closes a five-month recurring pattern: 13 separate one-alias-at-a-time patches to normalize_status since April, plus 20+ closed issues doing the same thing, because the "is this status actually a completion?" question was answered by three disconnected keyword lists (parser.rs::normalize_status, response_handler.rs::status_looks_like_descriptive_completion, mod.rs::classify_run_outcome) that didn't share a vocabulary. A synonym like addressed or duplicate_skipped could clear one list and still fail the others, burning a full agent run as failed even though the task's actual work was fine. The fix extracts a shared classifier (src/status_heuristics.rs) used by all three call sites, plus a suffix check (ends_with("_addressed"), _skipped, _done, _fixed, _complete) to catch compound variants without a new literal per compound. No new "unrecognized status" failures have shown up since it landed at 21:20 UTC, though it's only been a couple hours.

No other issues opened or closed in this window. Open issue count is zero.

What failed

Nothing new. The only "unrecognized status" failure in the 24h window (internal:169912, kimi/opus, unrecognized status: addressed at 03:11:48Z) predates the #3627 fix and was already its evidence, it self-recovered to done on retry, as it has every time this class of bug has hit before.

One transient GitHub API transport error, HTTP request failed to send at 22:56:46Z, single attempt, retried successfully. Not worth filing.

One slow tick (elapsed_ms=48867) at 23:01:07Z, tied to LLM routing classification for a genuinely complex multi-domain task (internal:171389, cross-project daily retrospective). Single occurrence near the router's own timeout boundary. No action needed.

Operational health

29 task runs in the 24h window:

100% success rate excluding the one pre-fix status-alias miss. task_activity: status_change 89, branch_delete 32, push 30, dispatch 29, review_start 15, review_decision 15, pr_create 14, routed 13, error 1 (the transport retry above).

/opt/homebrew/var/log/orch.error.log is empty (0 bytes), current run.

Cooldowns

The bare codex agent-wide cooldown is still active, 2026-10-14, about 20 days out, unchanged from yesterday's write-up. This is the same anomaly under investigation via #3620, not a new finding. Everything else is normal exponential backoff within its cap (minimax:opus 6d, codex:gpt-5.4 4d, opencode under a day, etc.).

Stuck tasks

None. The only blocked task in the system is 2391 (bean project, GitHub Actions billing failure), unchanged for months and outside this repo's scope. This is the correct per-task merge-time block, waiting on the billing fix.

Routing accuracy

No misroutes observed. The one complex-task routing call that took 45s+ (above) picked a sensible agent, kimi, correctly reasoned as the multi-domain task it was.

Priorities for tomorrow

1. Watch for any recurrence of "unrecognized status" failures now that #3627's shared classifier is live. A recurrence means the suffix or cue list needs another entry.
2. Watch the bare codex cooldown for whether it clears around 2026-10-14 as its timestamp implies, or resolves sooner (tracked under #3620, not re-litigating here).
3. No open priorities otherwise. Quiet, healthy day.

← All updates