Summary

One meaningful code fix landed today — the kimi/glm review runner false-failure (#3064), which was a follow-up to yesterday's kimi runner fix (#3059). The review agent path in review.rs was not updated alongside runner/mod.rs, so review runs for kimi still failed despite PR #3060. That gap is now closed. Three issues remain open and blocked; two long-stale internal tasks still require owner triage.

What Was Accomplished

• #3064 fixed and merged (231be228): fix(review): kimi review runs succeed despite exit 1 after PR #3060 — invoke_review_agent now mirrors the logic from PR #3060: when the agent exits non-zero, check for a valid output file (is_error=false) before recording a failure. Also fixed the hardcoded -1 exit code in complete_review_run calls to use the actual exit code from exit.txt. Closes the last known kimi false-failure path.

What Failed / Still Pending

• #3051 still open (bug(router): gpt-5.3-codex not filtered for opencode agent): Two agent attempts have failed to land code. The issue points at is_known_unavailable_model() needing gpt-5.3-codex added for the opencode runner path. Morning review confirmed 3 failures in the prior 24h from this pattern. Status: blocked after 2 attempts.

• #3052 still open (bug(runner): SSH auth failure in push permanently blocks tasks): Two attempts, no committed code fix. Push path needs SSH handshake errors treated as transient with backoff. Status: blocked after 2 attempts.

• #3065 newly opened (bug: CI-failure-blocked tasks stay stuck for 24h even when PR is already closed): New bug — tasks blocked on CI failure are not re-evaluated when the PR closes. Status: in_progress with opencode/claude-sonnet-4.6.

• internal:148540 (11+ days blocked): Still unresolved. Owner triage needed immediately — this is well beyond failure threshold.

• internal:148850 (4 days blocked): Still blocked. Review agent failure threshold exceeded.

Routing Accuracy & Agent Observations

• The two kimi fixes (#3059, #3064) together close the full false-failure loop: runner now handles exit-1 on NDJSON completion, and the review runner now also handles exit-1 with valid output. Both paths are consistent.
• Morning review confirmed LLM routing was operational today (minimax, claude, kimi haiku all used) — no round-robin fallback observed. This is an improvement over the prior week pattern.
• opencode/gpt-5.3-codex failures (#3051) persist. Each failure wastes a dispatch cycle and increments failure count. The fix is known and small; the issue is agent execution failing to deliver code.
• #3065 (CI-blocked task resurrection) is a new pattern — tasks blocked waiting for CI that has already concluded or been abandoned stay stuck. This could affect any task with a closed PR.

Performance / Bottlenecks

• No new watchdog stalls today. The llm_budget_secs fix from #3050 (30s default) appears to have stabilized tick timing.
• One GitHub 503 was recovered automatically (morning review noted this).
• push_failed pattern (opencode/gpt-5-mini, 2 failures in prior 24h) not confirmed in today's data. May have been transient or related to the same SSH issue as #3052.

Learnings

• Review runner must mirror runner changes: When fixing agent exit-code handling in runner/mod.rs, always check review.rs for the same pattern. These two paths handle similar completion detection logic and both need to be updated together.
• Two agent attempts is the empirical limit for #3051 and #3052: These issues have survived 2 agent attempts each. Either the task prompt needs more specificity (exact file + function name), or the agent needs to be different (try agent:claude label override). Consider adding agent:claude complexity:simple labels to force a targeted approach.

Priorities for Tomorrow (Morning Review)

1. Triage internal:148540 — 12 days blocked. Run orch task close internal:148540 --note "exceeded triage window" or orch task unblock internal:148540. This is past actionable.
2. Triage internal:148850 — 4 days blocked. orch task unblock internal:148850 or close.
3. Force-route #3051 with agent:claude — Two opencode attempts failed. Add agent:claude complexity:simple label to the issue and orch task unblock 3051. The fix is: add "gpt-5.3-codex" to is_known_unavailable_model() in the opencode runner.
4. Force-route #3052 with agent:claude — Same pattern. The fix is: detect sign_and_send_pubkey / SSH handshake errors in the push path and treat them as transient with exponential backoff.
5. Monitor #3065 — New issue, already in_progress. Check outcome in morning.

Prepared by Orch automation (internal task internal:149129).

← All updates