Repository navigation
A first review on any setup: it checks which AIs are ready, runs with those, and finishes even when a seat fails.
Added
council reviewchecks every AI CLI (Claude, Gemini throughagy, Grok) before it spends anything: a one-word prompt in an empty folder with no tools. Each line says ready, not installed, not signed in or out of quota, with the command that fixes it.--no-checkskips it.council doctorworks outside a set-up repository: it runs the same check and says which seats a review would use.- A seat that fails partway (usage limit, lost login, turn limit, error) moves to another model or AI for the rest of the review. A lost login drops that AI; any other failure drops only that model, so Claude's Engineer seat falls back from opus to sonnet before leaving Claude. A seat only moves to an AI already in the review's line-up: with a
.council/config.yml, one its seats name or one listed undercouncil.failover_providers. The review lists each swap under "Gaps in this review" and records the line-up inseats.json. When no AI is left for the Chief Architect, the review finishes without its rulings. - Reviews say who actually held each seat. A review on Claude alone is labeled single-vendor and names its models.
- When the Chief Architect overrules every objection, the review says so instead of reporting a stalled review with nothing found.
- A review ends with a verdict: PASS when nothing that blocks is left, CHANGES REQUESTED when an objection stands, needs a measurement, or was never ruled on. It replaces
STALLED, which every review showed because a review almost never reaches the PRD's zero-concern sign-off. - The Chief Architect can rule an objection MINOR: right, but it should not block (naming, style, wording, a nice-to-have). Minor notes are listed after the blocking objections and never hold up a review. STANDS now means a defect that should block.
Changed
council reviewruns one quick round by default: the Lead PM's brief, the Chief Engineer's objections, and the Chief Architect's ruling on each.--debateruns the earlier two-round debate.--roundsnow takes 1 to 4.- A seat in
.council/config.ymlwhose AI is not ready moves to a ready one for that review only; the file is not changed. - Every review, the debate included, ends with the Chief Architect ruling on whatever is still open, so the verdict rests on its rulings.
- The Chief Engineer is told to object to defects in the change and the code it touches, not to the wording, numbering or test plan of the Lead PM's brief. Those objections were why a second round never converged.
- The GitHub Action's
outcomeoutput is nowPASS,CHANGES_REQUESTED,NO_VERDICTorSKIPPED, instead ofAWAITING_APPROVALorSTALLED. A workflow that checks for the old values needs updating. - A login failure is recognised only from what a seat's CLI prints about its own login: Claude Code's own login lines on stdout only when they are the whole of it, and other login words, a bare 401 or 403 only on stderr. A model's answer that talks about 401s or logins no longer drops its AI. The Action's login error no longer assumes the seat was Claude.
- The GitHub Action still passes
--rounds 2, so it keeps the two-round debate. It skips the setup check (--no-check): a rejected token still ends the run with the Action's own "rejected the login" error, now also when the seat's failure is reported as "is not signed in". But a seat that fails mid-run now falls back to another Claude model and the comment lists the gap.