Summary

Excellent day — 26 commits landed since last night's retrospective, with a dominant focus on hardening the review pipeline. Almost every recurring failure mode in the review cycle has been addressed: infinite loops, stuck states, missing git ops, empty block metadata, and missing token tracking. Agent health is strong: 90 successes out of ~108 runs in the last 12 hours (83%). No open GitHub issues — the queue is fully cleared.

Accomplished Today

Review Pipeline Overhaul (8 bugs closed)

These were the most impactful fixes of the day — the review subsystem had accumulated several silent failure modes:

• #1401 — synthesized needs_review response was skipping git commit/push/PR ops, silently losing agent work. Fixed: synthesis path now runs full git ops before transitioning to needs_review.
• #1397 — Approved PR reviews left tasks stuck in in_review indefinitely (infinite review loop). Fixed: in_review tick now transitions to done/blocked when all reviews are approved.
• #1399 — Startup in_review reset was silently abandoned after 3 retries, leaving stale tasks. Fixed: 3-retry exhaustion now falls through to a hard reset.
• #1398 — block_reason and last_error were empty when task was blocked via max review cycles. Fixed: fields now populated before the block transition.
• #1391 — review_poll no-PR reroute had no circuit breaker — tasks could loop indefinitely on non-code tasks with no PR. Fixed (d9ed5d5): circuit breaker added after 3 reroutes.
• #1389 — auto_unblock_count >= 3 early exit fired before the reason-change reset, blocking tasks prematurely. Fixed: reason-change reset now runs before the exit guard.
• #1402 — Token usage was not preserved when NDJSON parse fell back to synthesized result. Fixed: token fields threaded through synthesis path.
• #1388 — Review agent token usage was never tracked, causing orch cost reports to undercount. Fixed: review runner now records tokens like the main agent runner.

System Reliability Fixes

• Deadlock in tick_dispatch_tasks — read lock taken on a write-locked RwLock. Fixed (3f195da): lock acquisition order corrected.
• Atomic SQL increment for failure counts (#1363/#1365) — failure counter increments were non-atomic under concurrent dispatch. Fixed.
• Backoff jitter centering (#1362/#1364) — jitter was applied before capping, skewing distribution toward the cap. Fixed: jitter now centered around the capped delay.
• SqliteRow OOB panic (#1386) — try_get now used for recently-added task fields to avoid index-out-of-bounds on older schema versions.
• Immediate cooldown recording on rate limit — concurrent dispatches could hit the same rate-limited agent/model before the first cooldown was written. Fixed (213a67d): cooldowns recorded at first rate-limit signal, not on task completion.
• PR orphaned on GitHub 502 (#1393) — create_pr_if_needed was not retrying on transient 5xx, leaving tasks without a PR link. Fixed.
• token/cost data silently dropped on parse_envelope fallback (b627d0d) — synthesis fallback now propagates token metadata.
• opencode timeouts resetting to new (#1320) — silence detection was transitioning timed-out tasks to new (losing branch/worktree) instead of needs_review. Fixed.

New Features

• orch doctor (bbe6682) — new CLI subcommand that detects: done tasks without merged PRs, orphaned worktrees, stale KV cooldown entries, and tasks stuck in terminal states. Excellent foundation for automated health monitoring.
• Exponential backoff cooldown CLI (04d56ec) — orch cooldown list and orch cooldown clear now available for operator inspection and emergency resets.
• skip_limited_threshold router guard (fbeb1f3) — agents with routing weight below threshold are pre-emptively skipped before LLM routing, preventing wasted dispatch cycles.

Synthesizer Improvements

• False-positive parse failures (5f613a4) — LLM outputs containing JSON-like fragments (e.g., {"key": "value"} in prose) were triggering structured-parse failures. Fixed.
• classify_failure tuning (5a0c3bb) — "the fix is complete" and "all tests pass" phrases now correctly classified as DONE rather than needs_review.

Morning Priorities — Follow-up

Agent Performance (Last 12h)

83% success rate (90/108 runs). NULL outcomes (10) likely reflect mid-session kills from the evening retrospective task restarts — not alarming.

Current State

• Active tasks: 2 in_progress (evening retrospective jobs), 1 blocked (#30429 — trading update, stale)
• Open GitHub issues: 0 — queue fully cleared
• Pipeline health: No recurring failures observed. All major review-cycle failure modes addressed.

Patterns & Observations

What's working:

• Review pipeline is now substantially more robust — 8 bugs fixed in a single day covering the complete failure surface
• orch doctor gives operators a fast health snapshot without querying SQLite directly
• Cooldown CLI enables emergency resets without service restarts
• 83% agent success rate with claude leading (credit restoration apparent)

Still worth watching:

• NULL outcome runs (10 in 12h) — mostly benign (cleanup races) but worth confirming they trend down
• Stale worktree metadata (fatal: not a git repository in brew error log) — benign but obscures real errors; a git worktree prune in user-managed project dirs at startup would eliminate this
• Task #30429 (trading update) is blocked with no block_reason — may need manual investigation

Tomorrow's Priorities

1. Investigate #30429 (blocked, no block_reason) — check whether the block is stale or needs human intervention.
2. Monitor NULL outcome trend — should decline given today's fixes. If still elevated (>5 in 12h), investigate cause.
3. Stale worktree metadata log spam — low urgency, but a git worktree prune call in reconcile_startup_worktrees for user-managed project dirs would clear it permanently. Consider filing if not already tracked.
4. Run orch doctor — first production use of the new feature; validate outputs are accurate and actionable.
5. Review cost accounting — token tracking was broken for review agents until today's fix. Verify post-fix cost data looks correct in next morning's cost report.

← All updates