Summary

Exceptionally productive day. 19 commits merged, 6 tasks closed, zero open issues at EOD (excluding this retro). The biggest structural change was the status-driven review workflow that eliminates the review_started sidecar flag. Agent prompts were hardened with a mandatory pre-output checklist. Service is stable and the backlog is empty.

Morning Review Recap

All priorities acknowledged.

Tasks Completed Today

Key Changes This Cycle

1. Status-Driven Review Workflow (95a2113)

The review_started sidecar flag — spread across 11 sites — is gone. Review lifecycle now uses status transitions exclusively:

• Agent done + PR exists → needs_review
• Engine spawns review agent → transitions to in_review as the guard
• Review failure → reset to needs_review (was: stuck in_review forever)
• Stale in_review with no tmux session → reset at startup and sync tick

This is a large behavioral change. Today's tasks all passed through it successfully, but edge cases (concurrent review + sync tick, rapid retry loops) should be watched.

2. Sidecar State Scoped Per-Repo (1dc75cb)

Sidecar files now land in ~/.orch/state/{repo}/{id}.json via tokio::task_local! { REPO_CONTEXT }. Brew service (cwd=/) previously caused all sidecars to write flat to ~/.orch/state/{id}.json, breaking worktree cleanup.

3. Graceful Restart (bb0e4f3)

orch restart now waits for all running agents to finish before restarting the service. Previously, an in-progress agent could be killed mid-task.

4. Agent Prompt Hardening

Two improvements to prevent the most common agent failure (work done but not pushed):

• Pre-output checklist (3cee8fd): agents must verify git status, git log, git push, and gh pr view before writing output JSON
• Route prompt cleanup (1f8245d): removed hardcoded executor descriptions that caused routing bias toward specific agents

5. Review Prompt Rewritten (404b557)

Rewrote review_task.md to be more explicit about rebase-first, CI-must-run, and decision rules. Review agents were previously approving without running CI.

6. Auto-Close on Done (c934eeb)

Issues are now auto-closed when tasks complete. Labels that were stuck at status:in_review on closed issues are also corrected.

What Didn't Go Well

Stale Integration Test (036ddbf)

Codex autonomous mode changed from --ask-for-approval never to --full-auto but the integration test was never updated. The test was #[ignore]d in the repo, masking the drift. Root cause: #[ignore]d tests aren't run in CI and can silently become stale. Filed as part of commit #272 and fixed.

Labels Not Updating on Close (historical, fixed today)

Several older closed issues showed status:in_review label — the label update code had a gap for the no-PR path. Fixed by c934eeb.

Prompt Effectiveness

No prompt changes needed beyond what's already landed.

Routing Accuracy

All 6 tasks today were routed to claude. No mis-routes observed. The route prompt cleanup (1f8245d) removes historical bias — upcoming scheduled tasks will be the first real test of neutral routing.

Performance

• GitHub API: Transient 502s at ~11:21 UTC (last reported in morning review) recovered automatically. fix: preserve active_task_id on transient GitHub API errors was merged today — this exact scenario is now handled.
• No lock contention observed in logs.
• Review latency: status-driven workflow eliminates the review_started flag polling delay. Expected improvement in review cycle time.

New Issues Filed

Tomorrow's Priorities

1. Monitor the new review workflow — needs_review → in_review status transitions are live but new. Watch for any stuck tasks in in_review without an active review session, especially under concurrency.
2. Stuck-detection threshold — the newly filed issue. Once deployed: tasks with no tmux session should be reclaimed in 10-15 min, not 30 min.
3. No other known bugs. Service is in the best shape it's been. The next scheduled jobs will reveal if the route prompt cleanup introduced any unexpected routing changes.

← All updates