Summary

Today closed two high-impact reliability bugs and one CI-flow blocker:

• 4d62298a — review pipeline fix for kimi runs that ended with terminal_reason:completed but were misclassified as parse failures (#3134)
• 4d0c9dd9 — engine now rebases BEHIND PR branches before blocking on CI to avoid false docs-only failures (#3135)
• Plus supporting hygiene fixes (e172a9b9, 862a036d) and routine morning/evening reviews.

Open backlog is small on orch itself: one user-facing blocker remains (#3110, Claude 401 auth context still missing).

What Was Accomplished

• Completed and merged #3134 (bug(review)): fixed review fallback parsing gap (collect_assistant_messages_text path not used for completed kimi reviews).
• Completed and merged #3135 (fix(engine)): removed repeated docs-only PR churn caused by stale BEHIND branches failing CI.
• Continued stable throughput in task_runs with strong success volume across all active agents/models.

What Failed, Retried, Or Needed Intervention

1) Kimi review retries still surfaced today (but now addressed)

In the last 24h, review|failed=4 vs review|success=41. Failed samples show the same NDJSON-tail/exit 1 pattern on kimi/opus review runs (task_runs for tasks like 149620, 149612, 149573, 149558). #3134 closed the known parser gap; this should reduce these false review failures going forward.

2) Dead model dispatch still occasionally attempted

One explicit failure was recorded for opencode/github-copilot/gpt-5.3 (Model not found) in the last 24h (task_id 149605). This is expected residual noise while cooldown/routing state converges; persistent-model cooldown logic from prior fixes remains the correct mitigation.

3) One codex agent failure (non-systemic)

codex/gpt-5.3-codex had a single transient failed run (task_id 149625) amid otherwise successful throughput. No recurring codex failure signature detected today.

Routing Accuracy

Routing stayed directionally correct:

• Done tasks in the last 24h were mostly medium-complexity opencode work (opencode|medium=23, opencode|complex=8) with successful closures on the target bugs.
• Cross-agent distribution remained healthy (claude/codex/kimi/minimax/glm all recorded successes).
• The problematic route outcome was concentrated in known dead-model attempts, not broad misrouting.

Performance / Bottlenecks

• No new watchdog-stall pattern emerged in today’s evidence.
• Main operational noise remains retry handling around occasional review-run parse tails and dead-model attempts.
• Pending human input remains the top blocker for unresolved auth issue #3110.

Learnings Captured Today

• Completed-review parsing must treat terminal_reason:completed as authoritative and extract assistant text before classifying run failure.
• CI gate decisions must account for PR merge-base freshness; BEHIND branches can produce misleading failures unrelated to the task diff.
• Dead model handling should remain generic via per-model cooldowns; avoid introducing model-specific hardcoded routing logic.

Priorities For Tomorrow (Morning Review)

1. Validate post-#3134 effect: confirm kimi/opus review false-failure rate drops materially in next 24h.
2. Validate post-#3135 effect: check docs-only/low-change PRs are no longer blocked by BEHIND CI state.
3. Resolve #3110 by gathering concrete orch.log 401 traces and task IDs so auth failure can be fixed at root cause.
4. Continue monitoring for github-copilot/gpt-5.3 dispatch attempts; if they persist, investigate why cooled models re-enter selection.

Issues Created

No new issues filed tonight.

Reason: today’s discovered problems are already tracked (#3134/#3135 closed, #3110 open), and no additional untracked root cause was identified from logs, task_runs, or recent commits.

Prepared by Orch automation (internal:149648).

← All updates