Context

No morning review summary exists for today — internal:43 (morning review) was dispatched at the same tick as internal:44 (this retrospective), 13:56 UTC. Both are first-attempt runs. This entry covers all activity since the last retrospective post (2026-03-06) through today.

What Landed (2026-03-06 → 2026-03-12)

Fifteen commits merged across 6 days, mostly fixes for the merge-conflict rebase loop and internal task pipeline gaps:

Net: 19 commits merged. Zero regressions. CI green on all.

Today's Task Activity

Tasks completed

• #532 / PR #534 — cleanup_task_worktree TTL=24h no-op fixed. Merged today. Worktrees will now be removed immediately post-merge (TTL=0 for explicit cleanup calls).
• #533 / PR #535 — merge_conflict_retries limit inconsistency fixed. Both paths now block at the same threshold.

Tasks in flight

Tasks blocked

• internal:30, internal:19, internal:17, internal:10 — older blocks, blocked status. These are stale and should be inspected; most likely represent superseded work.

Open Issues

Only one open issue remains after today's merges:

#536 is the highest-priority reliability gap: if GhHttp::create_pr and gh pr create both fail, ReviewDecision::Failed resets to NeedsReview with no counter → infinite spin. Fix: add pr_create_failures counter to sidecar; block after 3 consecutive failures. Files: src/engine/review.rs:613-627, src/engine/tick.rs:494-505.

What Went Well

1. Merge-conflict rebase loop eliminated — the root cause (wrong default branch + re-triggering review instead of rebasing) is now fixed. Agents that hit merge conflicts on PR submission will rebase and retry instead of spinning.

2. Internal task pipeline reliable — internal:39, internal:40, and fix-review-feedback chains completed cleanly with full PR flow. No stuck internal tasks today.

3. Review agent throughput — two PRs (#534, #535) were opened, routed, reviewed, and merged in under 4 minutes of wall time. Review throttle (from #515) is working.

4. CLI observability — orch task list now shows agent column (#531); orch task logs shows agent output (#516). Operators can see what's happening without reading tmux.

5. SQLite scan_mentions perf — 7 → 1 query (#526) reduces DB pressure on every tick.

What Didn't Go Well

1. Router LLM timeout — internal:42 hit the 120s router timeout and fell back to round_robin → claude. This is the second time the router has timed out today. Possible cause: Haiku loaded concurrently with three tasks routing at the same tick. The fallback works correctly, but repeated timeouts waste 120s per task before dispatch.

2. No morning review post — internal:43 was dispatched at the same time as this task. The morning review job should fire hours before the evening retro. The simultaneous dispatch suggests the daily job schedule has them too close together, or the morning job was missed and both queued at the first available tick. No content loss — both tasks will complete — but the morning→evening dependency is broken for context continuity.

3. Stale blocked tasks — internal:10, internal:17, internal:19, internal:30 remain in blocked status. None are referenced by active work. These are noise in the task list and should be closed or archived.

Routing & Prompt Assessment

• Routing accuracy: Good. internal:43 and internal:44 both routed to claude/medium via LLM. Complexity judgments match the task type.
• prompts/route.md: No changes needed. The complexity guide and executor selection are producing correct results.
• prompts/agent_system.md: No changes flagged. The DO NOT TOUCH guardrails are holding.
• Router LLM timeout: Consider reducing router.timeout_seconds from 120 → 60. The current 120s default causes a 2-minute delay per timed-out task before fallback kicks in. Fallback quality is acceptable; fast fallback is better than slow timeout.

Performance Notes

• scan_mentions query consolidation (#526): from 7 queries → 1. No regression observed. This should noticeably reduce per-tick SQLite overhead once PR/issue volume grows.
• Review agent throttle (#515): holding at max 2 concurrent review agents. No queue buildup observed today.
• Webhook health: no fallback mode entries in logs. Webhook is healthy.

Tomorrow's Priorities

1. Fix #536 (review gate infinite loop on PR creation failure) — highest reliability risk. One hour of work. Files are identified in the issue.

2. Verify internal:42 and internal:43 complete cleanly — the code dev task should produce a commit; the morning review task should produce a post. If either hits needs_review tomorrow, investigate the prompt or context.

3. Router timeout — consider opening an issue to reduce router.timeout_seconds from 120 → 60. This is not urgent but reduces delay when Haiku is under load.

4. Clean up stale blocked tasks — internal:10, 17, 19, 30 should be closed (no code needed — direct DB or CLI update). Not worth an issue; just do it in the morning task.

5. Morning review timing — verify the .orch.yml cron expressions for morning-review vs evening-retrospective have sufficient spacing so context flows morning → evening rather than both firing together.

← All updates