Weekly review 2026-10-17
What shipped
17 commits landed on main in the last 7 days.
- Stream: readable agent transcript (#3704) and per-entry task and attempt tags (#3705).
- CI and release: push-only CI, cow tests folded into
testandbuild-macos, arm64 macOS binary only, docs deploy through zola-docs-action with Cloudflare purge, IndexNow announcements. - Jobs: the review job now runs weekly instead of daily.
- Deps: rustls bump (#3702).
- Docs: old docs layout dropped, channels sidebar page shortened.
Closed issues include #3699 and #3700 (local squash-and-sign workflow), #3692 (review parser accepted template echo), #3688 (agent blocks when commit signing fails), #3686 (opencode error extractor), #3685 and #3652 (credit exhaustion handling), the copy-on-write worktree series (#3658 to #3682), #3676 and #3654 (language-agnostic classifier and prompts), #3655 to #3657 (startup env, module re-declaration, test cwd races).
What failed
Task runs in the last 7 days, by outcome:
| Agent / model | Outcome | Runs |
|---|---|---|
| claude / sonnet | success | 65 |
| claude / sonnet | blocked | 5 |
| opencode (free models, combined) | success | about 22 |
| minimax / opus | credit_exhausted | 4 |
| codex (gpt-5.4, gpt-5.5) | failed | 2 |
| opencode (various) | failed, rate_limit, parse_error | 4 |
| kimi / opus | rate_limit 1, success 1 | 2 |
Main failure causes: minimax credit exhaustion, codex failures, and free opencode models that rate limit or return unparsable output. The five blocked sonnet runs match the commit signing issue (#3688, now closed).
Operational health
- Open issues: none.
- Task list: only this review task is active. Nothing is stuck or blocked.
- Cooldowns handled the failures generically. No model-specific handling is needed.
- Claude sonnet carries most of the load. Free opencode models succeed often but one-off failures show they are unreliable.
Routing accuracy
Routing matches the settled policy. Credit-exhausted minimax moved to other agents with no stranded tasks. Codex has no successes in the window, so check whether it stays cooled or keeps being retried.
Priorities for next week
- Watch codex and minimax cooldown behavior until credits or models recover.
- Confirm no
blockedsonnet runs recur after the signing fix. - Keep the stream transcript changes under observation for regressions.