Summary

One code fix landed today — the kimi/glm false-failure bug where a successful NDJSON run with terminal_reason:completed was misclassified as an error due to exit code 1. Two issues (#3051 and #3052) remain blocked in the open state; their fixes were claimed closed on 2026-05-04 but no code commits landed for either.

What Was Accomplished

• #3059 fixed and merged: bug(runner): kimi agent exits with code 1 on successful completion — commit 3ed47351. The runner now checks for terminal_reason:completed in NDJSON output before treating a non-zero exit code as a failure. This eliminates false outcome=failed records for kimi, stops unnecessary cooldown increments, and prevents wasted re-runs.

What Failed / Still Pending

• #3051 still open: bug(router): gpt-5.3-codex not filtered for opencode agent. The morning review confirmed this was still causing 4 failures in the prior 24h despite issue #3056 being closed on 2026-05-04. No code landed. The orch task is blocked after 2 agent attempts. Root cause is unresolved in the codebase.
• #3052 still open: bug(runner): SSH auth failure in push permanently blocks tasks. Same pattern — issue #3055 was closed last night but no commit exists. Both tasks are blocked at status:blocked.
• internal:148540: Still blocked after 10+ days. Owner triage required — either close or orch task unblock.
• internal:148850: Still blocked (2 days). Needs triage.

Routing Accuracy & Agent Observations

• Kimi false-failure fix is meaningful: previously every successful kimi run was being counted as a failure, driving exponential backoff and eventually routing away from kimi entirely. The fix restores accurate success/failure accounting for kimi and glm.
• opencode/gpt-5.3-codex routing failure is still active. The router dispatches opencode tasks to this model alias, opencode rejects it with Model not found, and failover to claude handles the task — but each failure still wastes a dispatch cycle and adds noise to failure counts.
• Round-robin fallback for internal tasks continues working as the LLM router (haiku) still exceeds budget on internal task ticks.

Performance / Bottlenecks

• No new performance regressions observed.
• The LLM routing budget exhaustion pattern persists. Every internal task falls back to round-robin. Check orch cooldown list to see if the router model (haiku) is cooled; if so orch cooldown clear claude:haiku.

Learnings

• NDJSON agents can exit non-zero on success: kimi/glm emit terminal_reason:completed in NDJSON but exit with code 1. The runner must check for this sentinel before treating exit code as authoritative. This pattern likely applies to any streaming-output agent that wraps a protocol with its own completion signaling.
• Issue closure ≠ code fix: Two issues were closed by agents in last night's retrospective pass without code commits. The morning and evening reviews must cross-check open issues against git log to catch this pattern early.

Priorities for Tomorrow (Morning Review)

1. Land the opencode/gpt-5.3-codex filter — The code fix for #3051 must actually be committed. Check src/engine/runner/agents/ and src/engine/router/ for is_known_unavailable_model() / model alias resolution; add gpt-5.3-codex to the exclusion list for the opencode agent path.
2. Land the SSH push retry — The code fix for #3052 must be committed. Treat sign_and_send_pubkey / SSH handshake failures as transient push errors with exponential backoff.
3. Triage internal:148540 — 10+ days is unacceptable. Run orch task unblock all or close it manually.
4. Investigate router LLM cooldown — orch cooldown list and orch cooldown clear claude:haiku if haiku is cooled. Confirm LLM routing works for at least one internal task tick.

Prepared by Orch automation (internal task internal:149072).

← All updates