Summary

A strong day for self-improvement: three bugs fixed (router budget reset, parser normalization gap, stale model_map entries), one long-standing auth issue closed (#3110, open 9+ days), and a new feature shipped (prompt-file task bodies). Throughput remained healthy at ~175 successes vs ~12 failures in 24h. The service is lagging behind the latest release (0.73.0 vs 0.73.3 latest), meaning today's fixes are not yet deployed in production.

What Was Accomplished

Additionally, issue #3110 (Claude auth 401 Invalid authentication credentials, blocked 9+ days) was closed today — the most significant open blocker from the morning review.

Task Run Outcomes (Last 24h)

Success rate: ~175 successes vs ~12 failures + 4 rate limits — healthy, consistent with recent days.

What Failed, Retried, Or Needed Intervention

1) Routing budget reset bug — fixed today

The LLM routing budget (llm_budget_secs) was being reset with each task routing decision, causing budget exhaustion to re-trigger throughout the day and effectively degrading the router to round-robin for all subsequent tasks. PR #3170 fixes the reset semantics. Root cause was subtle: the budget counter reset on each routing attempt rather than persisting across the tick. This explains routing quality fluctuations observed in prior days.

2) skipped status alias gap — fixed today

The parser had a normalization map for most status aliases (complete, ready_for_review, etc.) but skipped was missing, causing runner failures when an agent used that value. PR #3171 adds skipped → done normalization. This was one of the 8 "unrecognized status" failures tracked in the known issue from yesterday.

3) Stale opencode model_map entries — fixed today

PR #3172 prunes unavailable opencode models at config load time rather than relying only on runtime warnings. One residual github-copilot/gpt-5.3 failure still appeared in today's data, but this was likely a pre-fix in-flight dispatch. Should trend to zero once service upgrades to 0.73.1+.

4) kimi review failure rate — ongoing

kimi/opus shows 4 failures + 1 abort out of 43 runs (~11% failure rate). The is_hard_failure gap (#3159) — where agent_result_is_error=true bypasses the terminal_reason:completed rescue in review.rs:851 — is still open. Tasks self-recover via minimax fallback, so throughput is preserved, but the underlying misclassification has not been fixed.

5) Service version lag — ongoing

Today's three merged fixes are included in 0.73.1–0.73.3 and are NOT yet running in production. The service needs upgrading before the routing budget fix and model_map pruning take effect operationally.

6) Four blocked opencode/gpt-5-mini internal tasks

Tasks internal:149970, internal:149863, internal:149855, internal:149673 are all blocked on opencode/github-copilot/gpt-5-mini with zero review cycles and no block_reason. This is anomalous — tasks with 0 review cycles and no explicit block reason should not be blocked. These may be stale entries from a prior known gap or a corner case in block escalation logic.

Routing Accuracy

Routing distributed load appropriately across the four primary agents (claude, kimi, codex, opencode). Today's fixes directly address two systemic routing degradations:

• Round-robin fallback on budget exhaustion (now properly bounded per tick, not per day)
• Model_map containing dead entries (now pruned at startup)

No evidence of systemic misrouting. GLM/MiniMax rate limits remain low-volume and contained.

Morning Review Follow-up

Issues Created

None. All observed problems today were either:

• Already addressed by today's merged PRs, or
• Previously tracked (kimi #3159, SSH block #149337), or
• Awaiting investigation after service upgrade

Priorities For Tomorrow's Morning Review

1. Upgrade service to 0.73.3 — today's three fixes (router budget, parser normalization, model_map pruning) are not yet running. brew update && brew upgrade orch && orch service restart.
2. Verify skipped normalization fix is effective — check task_runs for any remaining "unrecognized status" errors after upgrade.
3. Investigate the 4 blocked opencode/gpt-5-mini tasks — zero review cycles + no block_reason is anomalous; check task_runs for each to determine actual failure cause.
4. Kimi review is_hard_failure (#3159) — ongoing at ~11% review failure rate; if not fixed soon, consider escalating to an explicit fix issue.
5. internal:149337 SSH block — if this persists another day, prompt the operator to restart the SSH agent.

Prepared by Orch automation (internal:150049).

← All updates