Recent Commits & Progress

The reliability push continued overnight with another strong batch. Since the Apr 9 evening retrospective, commits include:

• fb57b4b9 fix(cli): handle multi-byte UTF-8 in truncate_err (#2355) — closes #2346
• fdb16b74 fix: graceful_shutdown_timeout config is parsed but never enforced — closes #2350
• 7c57cc76 fix: chunk batch issue state queries (#2353) — closes #2349
• d8e046a7 fix: accumulate token counts instead of overwriting on retries (#2352) — closes #2347
• 5cd40434 fix(security): scan markdown bullet lines instead of skipping them (#2351) — closes #2348
• 5f11441b perf: batch is_pull_request checks in scan_mentions with join_all (#2342)
• f506c686 fix(cli): fall back to rotated log files when primary log is empty (#2343)
• 6781b31f fix: add allowlist validation inside SQL-building loop in batch_set_fields
• 629bfd52 fix: propagate git rev-list parse errors in rebase_on_branch (#2338)
• c6cdfc86 fix: add allowlist validation inside SQL-building loops in set_fields (#2336)
• 56eff734 fix: recover push failures by rebasing on remote branch (#2335)
• 63f2057f feat(chat): add control session cost tracking and stats (#2334)
• 6748af50 fix: discord_ws.rs tests use unwrap() on async recv() calls (#2327)
• 1dbe5bf8 fix: fail review flow when PR URL parsing fails (#2333)
• 60915486 fix: KV helper functions in sync.rs silently discard store failures (#2325)
• f07be5eb fix: capture.rs tests use unwrap() on HashMap lookups (#2324)
• 0c577f0d fix: silence detection fires at exactly 600s for all opencode sessions (#2317)

Notable security fix: 5cd40434 — the markdown security scanner was skipping bullet lines entirely, meaning secrets embedded under bullet points (the most common documentation pattern) were invisible to the scanner. This was a meaningful leak surface.

Operational Health

Overall: healthy. Pipeline throughput is up ~30% across all metrics versus yesterday. No open issues, no blocked tasks, no active cooldowns.

Critical concern: CLI/service version boundary crossed

CLI:     0.60.159
Service: 0.61.3  ✗ mismatch — service crossed minor version boundary

The service is now on 0.61.x while the CLI is on 0.60.x. This is no longer a patch-level drift — a minor version bump signals new APIs or changed behavior that the CLI may not be compatible with. The orch log 200 command returning "No log files found" this morning is consistent with this: the CLI-side fix (#2343) landed in the repo but the installed CLI is too old to use it.

Action required:

brew upgrade orch && brew services restart orch && orch version

Apr 9 retro follow-ups

24h run outcomes

Notes:

• opencode: minimax-m2.5-free is the top opencode model (20 successes, up from previous sessions). The silence detection fix may already be showing results — opencode completions that were previously killed at 600s can now land.
• olm/gemma4 absent — was 2/3 success/failure in first 24h, now zero dispatches. Either routing changed or there's been no appropriate task type.
• claude timeouts (5) — new metric type appearing. These are distinct from failed — worth watching to see if this is a pattern or noise.
• kimi/opus rate limits (3) — same low level as yesterday. Generic backoff handling correctly.

12h task activity

Throughput is up ~30% versus yesterday's 12h window on every metric. The pipeline is processing significantly more work. Error count is down from 84 → 62 despite higher throughput, suggesting the quality of routing and execution continues to improve.

Error log

/opt/homebrew/var/log/orch.error.log is empty (0 bytes). No current service-level errors.

Open Issues

None. All recent bug reports are closed.

Priorities for Today

1. Upgrade CLI/service — minor version boundary crossed (0.60.x → 0.61.x). The orch log command is broken due to this mismatch. Priority: high.

   brew upgrade orch && brew services restart orch && orch version

2. Verify #2317 fix in production — silence detection was firing at exactly 600s for opencode sessions. The fix landed overnight. After the CLI upgrade, run orch log and check whether opencode sessions are now completing normally rather than resetting. Watch for any sessions that hit exactly 600s.

3. Investigate claude timeouts (5) — a new timeout outcome appeared for claude/sonnet (5 cases). This wasn't present in yesterday's stats. Determine whether these are silence-detection timeouts, hard timeouts, or a new failure mode, and whether the count is growing.

4. Investigate olm/gemma4 absence — disappeared from routing after first 24h appearance. Confirm whether this is expected (routing removed, agent not configured) or a silent routing failure.

5. Monitor throughput stability — 30% jump in dispatch volume is a positive sign but also a stress test. Watch error rates to ensure they don't climb proportionally with load.

← All updates