Summary

Extremely high-output night and morning — 22+ commits landed in the last 24 hours, addressing the full kimi billing-cycle cascade, error classification improvements, routing weight configuration, and migration safety. The pipeline is healthy: minimax is the reliable workhorse (72 success runs), routing weights are now configurable, and the router now gates on cooldown availability. Remaining focus: 6 open bugs currently being worked on, plus a stale worktree metadata issue in the error log.

Recent Activity (Last 24h)

Infrastructure / Architecture

• feat: configurable per-agent routing weights — operators can now tune per-agent weights in config.yml; replaces the fragile LLM "distribute evenly" heuristic
• feat: gate router on cooldown availability (#1266) — router no longer selects cooled agents/models; prevents wasted dispatch
• feat: separate runners for MiniMax and Kimi (claude CLI wrappers) — MiniMaxClaudeRunner and KimiClaudeRunner now handle non-"type":"result" stream parsing correctly
• fix: restore migration 012 checksum and split new column to 014 + test: add migration idempotency tests — migration safety fix + idempotency tests to catch future checksum breaks

Bug Fixes

• fix: remove 24h billing-cycle cooldown fallback from parse_retry_at — cooldown fallback was interfering with vendor-specified retry-at timestamps
• fix: stash uncommitted changes before rebase on startup (#1277) — prevents startup rebase destroying agent work (#1245 companion fix)
• bug: transport last_output persists across retries, replaying stale output (#1279) — stale output from prior attempt was being replayed to review agent
• bug: auto_unblock_count should reset when block reason changes (#1276) — counter sharing between CI recovery and auto-unblock mechanisms fixed
• fix: restore ci_recovery_count usage in handle_review_changes (#1284) — counter separation fix to avoid mixing auto_unblock vs CI recovery counts
• github(http): request latest check-runs per name (filter=latest) (#1283) — avoids picking oldest stale check-run when multiple exist per name
• bug: incomplete task_runs with NULL outcome block auto-unblock (#1272) — NULL-outcome runs no longer block the auto-unblock mechanism
• fix: clear agent/model on auto-unblock (#1269) — ensures re-routing picks fresh assignment after unblock
• bug: has_commits_ahead returns false on git failures — skips push + PR creation (#1291) — git errors were silently blocking PR creation
• bug: model_for_complexity returns None when all pool models cooled (#1278) — dispatch was proceeding with empty model; opencode exits 0 silently in this case
• fix: re-route external tasks that complete without code changes — tasks that reported success without a PR now get re-routed to a different agent
• perf: single git fetch per project at startup (#1274) — startup time reduced (was N sequential fetches)
• fix: truncate stdout/stderr to 500 chars in router LLM failure log (#1273) — very long stdout was polluting router failure logs

Operational Health

Agent Performance (last 24h task_runs)

Notable: The 7 opencode/empty-model failures (dispatch with null model) are explained by #1278 which was just fixed. These should disappear. Claude has 6 NULL-outcome runs — these typically indicate mid-session cleanup races; worth watching but not alarming given the high success count.

Task Activity (last 12h)

• 1,356 status changes — high throughput day
• 401 dispatches
• 65 errors — low relative to activity
• 51 review decisions, 50 PR creates — review pipeline healthy
• 5 auto-unblocks (small, expected)

Log Errors

Recurring in brew error log:

fatal: not a git repository: /Users/gb/Projects/orch/.git/worktrees/gh-issue-1005-...
fatal: not a git repository: /Users/gb/Projects/orch/.git/worktrees/gh-issue-1130-...
fatal: not a git repository: /Users/gb/Projects/orch/.git/worktrees/gh-issue-1251-...

These are stale git worktree metadata entries in the main project directory (not ~/.orch/worktrees). The worktrees were cleaned up but .git/worktrees/ metadata entries were never pruned. This is benign (doesn't block any task) but spams the log. A git worktree prune in the main project dir would clear them. Related to #1225 (startup reconciliation prune fix) — appears the prune runs in ~/.orch/worktrees but not the user project's .git/worktrees/.

Stuck / Blocked Tasks

#1245 is blocked (15h) — the startup rebase worktree destruction bug. The companion fix (#1277, stash before rebase) landed yesterday, but #1245 itself is blocked and may need manual unblock. Worth checking if the fix would resolve the blocked state.

#1247 at 4 tries — silence detection race condition (kill vs exit-0). Claude is on attempt 4; if this fails again, the task will likely be escalated to blocked.

Retro Follow-ups

From the 2026-03-29 evening retro:

Today's Priorities

1. Monitor #1245 (blocked, 15h) — startup rebase fix (#1277) is deployed; manually unblock #1245 if the engine doesn't auto-recover it, since the fix is already in place.
2. #1244 (in-memory cooldowns) — SQLite persistence for cooldowns is still the most impactful durability fix remaining. Target closure today.
3. #1247 (silence detection spurious review, 4 tries) — high retry count; watch for resolution or escalation to blocked.
4. Watch opencode/empty-model failures — should drop to zero after #1278 is in production. If they persist, check whether the fix actually resolves the dispatch path.
5. Stale git worktree metadata — the fatal: not a git repository log spam. Investigate whether startup reconciliation should also run git worktree prune in user-managed project directories. Low urgency, but the log noise is misleading.
6. Stale KV cooldown entries — verify cooldown:kimi:k2p5 and other expired entries are cleaned up. If not, the cooldown expiry check may not be evicting them.

← All updates