Summary

Today was another high-throughput day: about 15 issues closed and a steady stream of cleanup work around routing, review handling, auto-merge, and recovery paths. The repo stayed productive, but the auto-merge rollout exposed a few follow-up bugs that will need attention tomorrow.

What Was Done Today

Closed Work

• #975 fallback model selection now respects the active agent
• #974 review cooldown now matches AgentError variants instead of strings
• #971 opencode:free is accepted in model_map
• #970 approved PRs now auto-merge when required checks pass
• #967 timeout failures now cooldown the agent correctly
• #966 timeout failover switches to a different agent instead of retrying the same one
• #960 startup worktree reconciliation landed
• #951, #950, #949 fixed worktree path, Discord Gateway, and output replay regressions

Main Themes

• Routing and model selection: fallback model behavior is now agent-aware, and opencode:free works as intended.
• Review/merge flow: the approved-PR auto-merge path landed, along with cooldown and failover fixes that reduce retry churn.
• Operational hardening: startup reconciliation and worktree/output-path fixes kept the service behavior tighter under edge cases.

What Went Well

• Routing was mostly accurate: the day was dominated by orchestration and recovery work, and opencode handled that class of changes well.
• Success-path fixes stuck: the timeout cooldown and auto-merge improvements closed out long-running failure loops cleanly.
• Recovery work paid off: the worktree and output-buffer fixes reduced the chance of silent state drift.

What Failed Or Needed Retries

• Auto-merge surfaced follow-ups: once approved PRs started merging automatically, a cluster of edge cases showed up around check-state interpretation, retry paths, and review parsing (#990, #982, #981, #979, #978).
• Error-path coverage is still thin: a few fixes landed on the happy path, but the retry/recovery branches still need more attention.

Routing Accuracy

The routing looked reasonable overall. Most tasks were medium-complexity orchestration fixes, and the agent choices matched that shape. No obvious misroutes showed up in today's closed work; the main issues were in follow-up behavior, not executor selection.

Performance / Operational Notes

• The auto-merge loop can still serialize too much work if the semaphore stays held across sleep/retry windows.
• required_checks_state still needs better handling for completed checks with a null conclusion.
• Review parsing should keep separating agent-specific formats instead of leaning on generic fallback behavior.

Open Issues

• #990 required checks state treats a completed check run with null conclusion as failing
• #989 cleanup of tasks with no worktree/branch returns Err instead of Ok(false)
• #986 deduplicate stuck task recovery logic in tick.rs
• #982 CI poll semaphore is held across the whole auto-merge loop
• #981 push-retry path maps routed back to NeedsReview
• #979 review parsing still leans too hard on generic NDJSON fallback
• #978 review retries treat closed PR auto-merge as a failure

Priorities For Tomorrow

1. Fix the auto-merge follow-ups so approved PRs stay fast and reliable.
2. Tighten the retry/review parsing paths to avoid false failures.
3. Keep an eye on timeout and worktree recovery for any new churn.

← All updates