What shipped

19 commits landed on main.

Grok joined the agent pool (#3708). Task worktrees clone through the git-cow crate (#3706).

Runner and classifier:

• Read errors[] on error envelopes so credit exhaustion classifies (#3710, #3714).
• Clear the cached GitHub token and retry once on 401 (#3719). Block the task when PR creation fails with a GitHub auth error (#3718).
• Block tasks that run out of attempts, and keep those tests off the GitHub API (#3715).
• Success envelopes without terminal_reason become parse errors (#3726).
• Salvage grok runs killed at the timeout after they finished wrap-up (#3736).
• Treat a kimi review that exits mid-thought as incomplete (#3737).

Release and CI: publish as orch-cli on crates.io and run with systemd on Linux (#3711), Linux x86_64 via cargo binstall (#3712), sign and notarize the macOS binary (#3731, #3733). Agents use shipcheck when a repository has it (#3724). Board labels reconcile into GitHub Project Status (#3723). Telegram messages over the length limit split at line ends.

Closed this week: #3709, #3713, #3716, #3717, #3722, #3725, #3734, #3735.

What failed

139 runs in the last 7 days. 106 succeeded.

The remaining eight non-success runs are one each of grok-4.6 timeout, grok-4.7 parse_error, kimi failed, kimi parse_error, minimax credit_exhausted, minimax failed, codex gpt-5.4 unavailable, and codex gpt-5.5 unavailable.

Grok's first week: four empty grok failed: errors on 20 Oct, three invalid response parse errors, three 1800s timeouts, two Grok Build 402s. The timeouts and the kimi review parse error were filed and closed today (#3735, #3734). The stored invalid response string has no payload, so those three cannot be told apart from the DB.

Claude's seven failures were silence detection, max attempts, a passbox Touch ID timeout on an unattended scheduled job, and a review agent that hit GitHub 401 three times (fixed in #3717). The three blocked sonnet runs are that same scheduled import: passbox could not unlock credentials without Touch ID, then a later attempt found no new work.

Kimi hit the 5-hour usage cap four times. Codex gpt-5.4 and gpt-5.5 are unavailable. Minimax is out of credits.

Operational health

50 tasks reached done. task_activity in the window: 515 status changes, 204 dispatches, 62 PRs, 36 review starts, 33 errors, 22 reroutes, 4 review-agent fallbacks, 3 timeouts.

Open issues are design work: #3720 named profiles, #3727 cluster (blocked on draft PR #3729 for operator review), #3728 third-vote decision (no-agent), #3730 named-profile runtime acceptance (no-agent). External PRs still open: #3721 (named-profile offline foundation), #3732 (shipcheck CI).

Cooldowns now: codex:gpt-5.4 4d9h persistent_model_error, minimax 12h credit_exhaustion, minimax:haiku 1d model_error.

orch.error.log is empty (mtime today). The last 1000 log lines are INFO sync ticks, typically 1.6s to 3.7s, a few near 10s. Host suspend/resume gaps are discounted from stuck-task age.

Current behavior matches prompts/skills/orch/SKILL.md: generic cooldowns, per-task billing block at merge time, retry next tick when every agent is cooled.

Stuck tasks

Routing accuracy

Claude sonnet still does most of the work (58 successes from 68 runs). Kimi opus 29 of 35. Grok, new this week, 19 of 32; failed runs rerouted and those tasks finished. Codex and minimax recorded no successes.

No silent failures. Unavailable models and credit exhaustion took the generic cooldown path. Prompts do not need a change this week. The parser bugs that did fire were envelope shape, timeout salvage, and mid-thought reviews, all closed.

Priorities for next week

1. Watch grok after #3736 and kimi reviews after #3737. If grok invalid response keeps showing up, the next step is to keep a truncated payload on InvalidResponse so the parser can be diagnosed.
2. Confirm the 401 retry in #3719 holds. The leftover blocked review task is an operator choice after that.
3. Leave codex:gpt-5.4 on persistent cooldown until the model exists again. Leave minimax on the credit cooldown.

← All updates