Convene a blind multi-model council on a hard problem. Fans brief out to one seat session per model — each an independent file-reading agent bound to working_dir — runs them CONCURRENTLY in the background, and returns IMMEDIATELY with a council_id and blind hat labels (hat1, hat2, ...). The hat->model mapping is hidden (revealed only via council_reveal), so you can synthesize across seats without knowing which model produced which answer. Poll for answers with council_poll.

models: OPTIONAL — omit it (or pass []) to get the default one-seat-per-family roster. working_dir defaults to the server's current working directory.

kind: what this council is FOR, e.g. "bug", "review", "architecture" (default "adhoc"). Purely descriptive record-keeping — never affects routing — but persisted so the council history shows why a council was convened.

Poll a council for seat STATUS ONLY (blind — no model names, no answer bodies, so it never bloats your context). Each hat returns its status; a DONE hat returns answer_chars (size hint); a RUNNING hat returns live progress {turns, output_tokens} read natively from its transcript so you can tell it's advancing vs wedged. With wait=True this long-polls up to timeout seconds, returning early as soon as a new seat finishes or all are done.

FIRST poll: use timeout >= 360 — seats take minutes, and a short first poll just burns turns on no-op status checks (the long-poll returns EARLY the moment any seat finishes, so a big timeout costs nothing). Subsequent polls can be shorter. Call repeatedly until done=true, THEN fetch each answer with council_answer(council_id, hat).

Fetch ONE seat's FULL answer text by blind hat label (e.g. 'hat2') — the content council_poll deliberately omits to keep your context lean. Call once a hat shows status=done in council_poll. Never reveals the model.

Cross-examine ONE seat by its blind hat label (e.g. 'hat2') within a council. Runs a follow-up turn on that seat's session (it may re-read files) and returns its reply synchronously. Never reveals the model. Use to probe disagreements before you synthesize.

De-anonymize a council: return the hat->model mapping and per-seat status. For human/debug insight AFTER synthesis — do NOT use this to weight the diagnosis.

Blind check: has model finished answering (status=done) in this council? Returns True/False only — never the hat label or answer — so you can confirm a model participated without breaking hat blindness for scoring/synthesis.

Score each seat's reply — the MANDATORY end step of every council, AFTER your synthesis and BEFORE council_reveal (so scores can't be biased by model identity). scores: [{"hat": "hat1", "score": 7, "notes": "verified root cause, thin on fix"}, ...] score is 1-10; judge correctness against the verified code, depth, and actionability. The server resolves hat->model itself, feeding the per-model quality leaderboard (see model_scores).

Per-model quality leaderboard from all council scores: avg_score, eval count, and the 5 most recent scores per model.

Dump the seat/model availability cache: per-model status (ok/cooldown), the classified reason (balance/quota), the last error, and cooldown expiry. Seats record capacity failures here; council_start skips models in cooldown. (Renamed from model_health, Decision #20; backing table renamed in P2.)

Mark a council closed. Records are KEPT (council, hats with answers/errors, and seat sessions) for history/debugging — nothing is deleted. Score the seats with council_score FIRST if you haven't.

Open a direct 1:1 chat with one seat (seats.yaml key). Returns {chat_session_id, seat, model}. model defaults to the seat's first healthy model.

Send a message to the chat's seat — ASYNC; returns immediately with {task_id}. Poll the turn with chat_poll.

Long-poll a queued chat turn; returns early when the turn completes: {status, reply?, usage?, progress?}.

Message-level history of a chat: [{role, content_preview, ts, usage}, ...].

List chat sessions, optionally filtered to a single working directory.

Mark a chat closed. History is KEPT.

List the seats defined in seats.yaml: [{name, models, runner_kind}, ...]. A seat is an LLM family served by a local agent CLI (claude/pi/codex).