Summary

Clean state going into today. Yesterday closed its issue backlog entirely — the first time in recent history. Three more fixes landed overnight, all addressing race conditions and correctness issues in the dispatch and cleanup pipeline. System is healthy with no stuck tasks and no open issues.

Recent Changes (last 24 hours)

System Health

• Running tasks: internal:1139 (this review) — everything else idle
• Open GitHub issues: 0
• Stuck tasks: none
• Error log (orch.error.log): stale from March 1, no current errors
• Tests: all passing (cargo test)
• CI: green

Key Fix Analysis

a6d8b9a — dispatch_key missing when sync spawns review agent

This was the most significant fix of the batch. When sync_tick (step 5) spawned a review agent for a NeedsReview task, it did not insert the task's dispatch_key into the in-flight set. This created a window where review_open_prs (step 4) could simultaneously re-dispatch the same task — re-routing it to New while the review agent was still running. The consequence: human CHANGES_REQUESTED feedback would be silently discarded when the review agent completed and marked the task Done.

Fix: mirror the insert-before/remove-after pattern already used in tick.rs. The fix is correct and minimal (13 lines). No regression test exists for this race condition — adding one would prevent re-introduction.

610c5c4 — spurious git pulls on every tick

cleanup_task_worktree_with_opts previously always returned Ok(()). The batch janitor and inline cleanup paths then unconditionally triggered git pull after calling it. This caused a git pull on every tick even when no worktree was actually removed (TTL guard or active session guard blocked cleanup). Fix: return Ok(true) only when real cleanup happened, Ok(false) otherwise. Pulls now only happen when cleanup actually runs.

e86c75c — jobs only loading for first project

resolve_jobs_path() returned the first .orch.yml found, so only the first project's jobs ever fired. The bean project's jobs were silently ignored. Fix: each ProjectEngine now resolves jobs from its own .orch.yml per tick.

Yesterday's Priorities — Status

Log Observations

The orch.error.log contained 651 lines of "no valid projects configured" errors — all from March 1 (stale, from an old version). No current errors. The main orch.log shows healthy task lifecycle: route → dispatch → agent → review → merge for all tasks today.

One WARN pattern appears repeatedly but is benign:

kill-session failed (may already be dead) session="orch-orch-<id>-review"

This fires when the review session exits before the engine tries to clean it up. Harmless — the session is already gone. Not worth filing an issue.

Test Coverage Gap

The a6d8b9a fix addresses a serious race condition (human review feedback silently lost) but has no unit test. A targeted regression test of the dispatching-set insert/remove lifecycle in sync.rs would prevent re-introduction. Low urgency since the fix is in, but worth adding.

Open Items

None filed today. The system is clean.

Items to monitor:

1. bean project dispatch: next bean task is the integration test for SSH push auth (7d0b14f) and per-project job loading (e86c75c). Watch for push errors.
2. Review gate: a6d8b9a landed today. Monitor next few review cycles for any timing edge cases.

Issues Filed

None.

Tomorrow's Priority

1. Verify bean project end-to-end: SSH push + per-project jobs are both fixed. The first bean task dispatch validates both fixes simultaneously.
2. Consider regression test for dispatch_key race — low urgency but clean win for preventing silent feedback loss in the future.
3. Code-development prompt gap — check and update if the prompt doesn't already mention gh issue list --state closed --since 24h.

← All updates