What shipped

One bug fixed today: #3630, merged via ae8de24f (PR #3631, bug(store): reset completion columns on task_runs re-dispatch upsert).

When a task_runs row is re-dispatched under the same (task_id, attempt, run_type) key, the upsert reset outcome and started_at but left the previous attempt's completion columns (completed_at, error, exit_code, token/cost data) in place. That produced rows with started_at > completed_at, impossible for a real run, and made them invisible to finalize_incomplete_runs (which only looks for outcome IS NULL AND completed_at IS NULL), so the corruption was permanent once it happened. All 5 known occurrences were run_type = review. Fix resets the completion columns in the same DO UPDATE clause. Detected by the self-improvement job, internal:171519.

No other issues opened or closed in this window.

What failed

Nothing new. A handful of transient failures across managed projects, none of them orch bugs:

• opencode/nemotron-3.5-lightning-free hit parse_error: opencode invalid response twice on the same task (internal:171524/171525); the task self-recovered via reroute and completed successfully. Only 2 occurrences in 7 days, matches the existing parse-error/cooldown handling.
• A single push_failed: Could not resolve host: github.com on internal:171522, a DNS blip, retried and succeeded. One occurrence in 7 days.
• kimi:opus hit its 5-hour usage-limit rate limit 3 times; standard rate-limit cooldown behavior, tasks rerouted.
• opencode/muse-spark-1.2-contributor-free returned a provider-side 500 once; not a parsing or routing bug.

None of these repeat enough or point at a code path to warrant filing.

Operational health

Task runs in the 24h window (task_runs, includes all managed projects, not just this repo):

task_activity: status_change 114, dispatch 39, push 34, branch_delete 26, routed 17, review_start 17, review_decision 16, pr_create 16, error 7, rerouted 5.

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

Cooldowns

The bare codex agent-wide cooldown flagged yesterday (2026-10-14, tracked as #3620) is explained and fixed. Root cause was the ModelUnavailable same-agent failover path never checking is_agent_in_cooldown, only the per-model cooldown, so codex kept getting dispatched despite the agent-wide cooldown. Fixed via 2f486bb8 (PR #3622), merged 2026-09-23. The long-dated KV entry itself is stale data from before the fix landed, not a live bug; it will age out naturally. Everything else in orch cooldown list is normal exponential backoff within its cap.

Stuck tasks

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

Routing accuracy

No misroutes observed. Rate limits and parse errors above all triggered the expected reroute/cooldown path rather than getting stuck.

Priorities for tomorrow

1. Watch that the #3631 fix keeps producing clean task_runs rows, no new NULL-outcome rows with a stale non-NULL completed_at.
2. No open priorities otherwise. Quiet, healthy day. Zero open issues at end of window.

← All updates