Council

Karpathy's llm-council (karpathy/llm-council, backend/council.py) sends the raw question to four different models, has each rank the others' anonymised answers, then a chairman synthesises. Its value is model diversity, and it needs OpenRouter credit. Here every seat is Claude, so diversity comes from stances, not models: four seats on the raw question would return four copies of one opinion.

Cost is about 2N+1 calls (N seats, N critiques, the chairman). Say the seat count before dispatching.

1. Brief (chairman = this session)

Pull context once, here, so every seat argues from the same facts: search the workspace, repo and any connected mail or docs before asking the user anything. Write council-<slug>.md to the scratchpad:

• Question in one sentence, and the options (always include "neither / wait").
• Constraints and what the user has said they want, quoted.
• Evidence, each fact with its source path or link.
• CANNOT ANSWER: what the brief does not know and seats must not guess.
• Files seats may read, by absolute path. Nothing else is in scope.

The brief carries no lean. The chairman does not state or hint at its own view before the seats return; a chairman that has already decided turns the council into confirmation.

2. Seats (Stage 1)

Pick 3 to 5 stances from references/stances.md: at least one General, the rest from the domain, none that would agree on this question by construction. Dispatch one council-seat agent (enkr-skills:council-seat when installed as the plugin) per stance in a single message (parallel, uncorrelated). Prompt: the stance's full entry pasted verbatim, the brief path, the word cap. Give each a name so it can be continued.

Outside-model seat (opt-in only, "with gemini" or similar): the one real model diversity available without OpenRouter, via any non-Claude CLI the user has (e.g. gemini -p). Write a second brief with every personal detail stripped (names, money, health, employer), show it to the user before sending, then pass it with the seat rules from agents/council-seat.md. Never send it the full brief.

3. Critique (Stage 2)

Continue each seat with SendMessage: the other seats' answers relabelled Response A, B, C in a shuffled order, stance headers removed, its own answer left out (Karpathy's version lets a model rank its own; same-model seats recognise themselves, so that only adds self-preference). Seats critique, may revise their call, and end with FINAL RANKING:.

Tally: average rank per answer across the seats that ranked it. Parse failures count as missing, not last.

4. Verdict

Karpathy's chairman writes "the council's collective wisdom". Do not: blending is exactly where a same-model council smooths over a real split. Output, in this order:

1. Call: one option, stated plainly, with the two or three reasons that carry it.
2. Ledger: every point where seats split, who said what (by stance), and how the chairman settled it, or "unsettled" with what would settle it.
3. Ranking: the tally, one line. Where the chairman overrules the top-ranked answer, say why.
4. Change my mind: the one fact that would flip the call, and how to get it.

Unanimity is weak evidence here. If every seat agrees, say so and name the shared assumption they all made; same-model agreement is one opinion counted N times.

Show the verdict in chat. It lands in a file only when the user says where; the brief stays in the scratchpad.

Rules

• The record beats the council. If a repo file, ledger or prior decision already settled it, say so and stop: the council is arguing a closed question.
• Health and money floors win. An open health issue, or a worst case the user's buffer cannot absorb, outranks every upside argument.
• The user is the last seat. Show the verdict before acting on it; they can overrule any of it.