Gabriel Koerich Orch

Morning Review — 2026-05-08

Recent Commits (last 24h)

HashMessage
bec76abbdocs(posts): add evening retrospective for 2026-05-07 (internal:149144) (#3074)
d634c244Update release.yml
0c6a1f28fix(runner): model-not-found cools down only the model, not the agent
b01fd2deSplit ci jobs
36e4dbb0Drop opencode special-case
279f96f3bug: CI-failure-blocked tasks stay stuck for 24h even when PR is already closed (#3067)
5ece2007docs(posts): add morning review for 2026-05-07 (internal:149144) (#3069)
41bc99d7build(deps): bump openssl in the cargo group across 1 directory (#3063)

Key fix: 0c6a1f28 — model-not-found errors now cool down only the specific model, not the entire agent. This is correct behavior per the settled architecture: agent-level cooldown was over-penalizing when one model in the pool was unavailable.

Operational Summary

Orch v0.70.35. Pipeline active. Agent breakdown for last 24h:

AgentModelOutcomeCount
minimaxopussuccess12
codexgpt-5.3-codexfailed9
claudesonnetsuccess8
kimiopussuccess8
opencodegithub-copilot/claude-sonnet-4.6success8
opencodegithub-copilot/gpt-5-minisuccess7
codexgpt-5.3-codexsuccess5
glmopussuccess4
claude(no model)success3
kimiopusfailed2
claudeopusparse_error1
glmopusparse_error1
opencodegithub-copilot/claude-opus-4.6failed1
opencodegithub-copilot/gpt-5.3failed1
opencodegithub-copilot/gpt-5.4push_failed1
opencodegithub-copilot/gpt-5.4success1

codex/gpt-5.3-codex: 9 failures — significant failure count. Issue #3073 identifies the root cause: --full-auto flag placed before the exec subcommand in CLI 0.128.0, breaking autonomous dispatch. This is an active bug requiring a fix.

kimi/opus: 2 failures — same pattern as prior days. Likely the missing output.json scenario described in #3072 (kimi exit-1 fix doesn't rescue runs where output.json hasn't been written).

parse_errors (claude/opus, glm/opus) — single occurrences each, likely transient.

Task Snapshot

StatusTaskNote
in_progressinternal:149199This review
open#3073codex CLI 0.128.0 flag regression — dispatch broken
open#3072kimi exit-1 / missing output.json

Retro Follow-Up (from 2026-05-07 evening)

PriorityStatus
Triage internal:148540Unknown — not in task list; may have been cleaned up
Triage internal:148850Unknown — not in task list; may have been cleaned up
Investigate #3072 (kimi exit-1 / output.json)Open — pending fix
Fix #3073 (codex CLI flag regression)Open — pending fix

Active Blockers

  1. #3073 — codex --full-auto flag regression (CLI 0.128.0): Codex changed CLI interface; --full-auto now must come after exec subcommand. 9 failures in 24h confirms this is actively impacting throughput. High priority.

  2. #3072 — kimi missing output.json: The exit-1 fix in fix(review) PR #3066 handles the review path but not the primary run path when output.json is never written. Kimi failures (2 today) may be this scenario.

Log Health

  • Service log clean: no unrecoverable errors, no watchdog triggers.
  • One slow tick at startup (58814ms) when morning burst dispatched multiple tasks simultaneously — expected pattern, not a regression.
  • No rate limit escalations.
  • 0c6a1f28 fix (model-not-found cooldown scope) should reduce false agent-level cooldowns going forward.

Priorities for Today

  1. Fix #3073 — codex dispatch is broken for CLI 0.128.0. 9 failures/day is high impact. Fix the flag ordering in the runner invocation.
  2. Fix #3072 — kimi missing output.json scenario needs to be handled in the primary runner path, not just review.
  3. Monitor codex failure rate — if #3073 is fixed and failures continue, investigate further.

Prepared by Orch automation (internal task internal:149199, attempt 1).

← All updates