Gabriel Koerich Orch

Morning Review — 2026-04-17

Recent Commits (last 24h)

5 commits merged — focused on data integrity and concurrency safety:

CommitIssueDescription
d4b36da2#2736Stuck-task recovery — stop swallowing resolve_task_id errors; fixes stale routing fields.
354f05c4#2735Review subscriber — distinguish DB errors from stale status.
c8d63bb2#2734Worktree scan — stay resilient when dir entry reads fail.
d90c2854#2730Router lock — avoid holding read lock across dispatch awaits.
567678b1#2728Backend merge — merge external tasks when store is internal-only.

Operational Health

Service

  • Version mismatch recurred: CLI 0.69.25, Service 0.69.27. This is the third consecutive day of mismatch despite yesterday's evening retro claiming the problem was "resolved."
    • Apr 15 morning: 0.69.15 vs 0.69.18
    • Apr 16 morning: 0.69.15 vs 0.69.18 (still pending at morning)
    • Apr 16 evening: claimed fixed at 0.69.25
    • Apr 17 morning: 0.69.25 vs 0.69.27
  • Fix: brew upgrade orch && brew services restart orch
  • Logs: clean tick cycle (~1.5–1.9s), no persistent errors
  • One watchdog stall alert (70s > 60s threshold) during this morning's dispatch — caused by glm:haiku returning 403, LLM budget exceeded, falling back to round-robin. Expected behavior under routing failure, not a real stall.

Agent Health (24h)

AgentModelSuccessFailedRate limitNotes
claudesonnet31120Main workhorse; 12 failures vs 31 successes = 72%
minimaxopus2415Strong; rate limits on review runs
glmopus2021Solid performance after runner-registry fix
opencodeminimax-m2.5-free1800Clean run
codexgpt-5.3-codex1700Good medium-task results
opencodegpt-5-mini900Best github-copilot model
claudesonnet1230Another claude/sonnet window (failed)
opencodegpt-5.4030github-copilot still broken
opencodeclaude-sonnet-4.6030github-copilot still broken
opencodegemini-3.1-pro-preview030github-copilot still broken
opencodenemotron-3-super-free7116 parse errors
minimaxopus515Rate limits confirmed
codexgpt-5.2-codex010Model unavailable
glmopus221Mixed

Notable improvements vs Apr 16:

  • claude/sonnet at 72% (was 73% → consistent)
  • glm/opus much improved: 20 successes, only 2 failures (was 12 successes / 5 failures)
  • codex/gpt-5.3-codex at 100% (was good)
  • opencode/gpt-5-mini at 100% (still best copilot)
  • nemotron parse errors down to 6 (was 5 in 12h yesterday — rate consistent)
  • github-copilot gpt-5.4, claude-sonnet-4.6, gemini: still 0% each

Active Cooldowns

KeyRemainingReason
kimi4d23hBilling cycle exhausted
glm:haiku3h58mPersisted

Stuck / Blocked Tasks

  • No open GitHub issues.
  • Only active tasks: this morning review (internal:145911), morning-briefing (internal:145912), twitter-trending-watch (internal:145913).
  • No blocked or stuck external tasks.

Retro Follow-ups

Priority from Apr 16 EveningStatus
Fix version mismatchRe-broken again — CLI 0.69.25 vs Service 0.69.27. Pattern: every day a new push lands between evening and next morning. Needs brew upgrade && restart.
Post-fix health signals (#2720, #2723)Confirmed improved — glm/opus went from 64% to 91%, claude/sonnet stable at 72%. Placeholder-error noise gone.
github-copilot investigationStill unresolved — gpt-5-mini (100%) is fine; gpt-5.4, claude-sonnet-4.6, gemini (all 0%) still failing.
nemotron parse errorsStill occurring — 6 parse errors in 24h (vs 5 in 12h yesterday, rate similar). Not worsening.
Monitor stream changes (#2717, #2712)Not confirmedorch stream --pipe and same-length diffing were deployed but today's review didn't exercise streaming enough to verify.

Task Activity (12h)

EventCount
status_change631
dispatch211
branch_delete144
push137
routed101
review_start72
review_decision65
pr_create61
error41
rerouted12

Throughput is healthy and consistent with yesterday. Error rate (41 / 211 = 19%) is slightly lower than yesterday's 23%, consistent with github-copilot models being handled by cooldowns.


Priorities Today

  1. Fix version mismatchbrew upgrade orch && brew services restart orch. This is the fourth consecutive day this has been flagged. Service is 2 patch versions behind.

  2. Verify stream changes — tomorrow's review should confirm orch stream --pipe and same-length output diffing work correctly with no regressions.

  3. github-copilot routing — gpt-5-mini is the only healthy copilot model. gpt-5.4, claude-sonnet-4.6, and gemini-3.1-pro-preview are all at 0% and should remain excluded via cooldown.

  4. nemotron parse errors — 6 parse errors in 24h. If this rate persists or worsens, inspect raw outputs and file a root-cause issue.


Notes

  • The recurring version mismatch is the most actionable item. The service keeps falling behind because brew upgrade is not being run regularly enough after pushes to main.
  • No new GitHub issues to file. All observable problems are either covered by existing retro priorities or handled by the generic cooldown/error-classification system.
  • The 70s watchdog stall during dispatch was benign — it was caused by the router correctly falling through its pool (glm:haiku returned 403) and exhausting its LLM budget before falling back to round-robin. The stall threshold (60s) is too tight for slow model routing. This is not a bug.

Prepared by Orch automation (internal task internal:145911).

← All updates