LLM Council

Karpathy's LLM Council pattern, provider-agnostic. dair-academy's version hardcoded Fireworks; ours reads any OpenAI-compatible endpoint via env.

When to use

• High-stakes plan review (/plan crosses N-file threshold)
• Conflicting learning-rules → re-resolve via vote
• User invokes /council "<query>" or /wiki council
• Architecture decisions where you want multiple viewpoints captured
• Persisting deliberation as a wiki page for future reference

Three phases

1. Independent: each model answers in parallel
2. Ranking: each model ranks anonymized peer responses
3. Synthesis: chairman model reads all responses + rankings → final answer

Provider config

Provider chosen via env. First-match wins:

Env var	Provider	Default base URL
ANTHROPIC_API_KEY	Anthropic	https://api.anthropic.com
OPENAI_API_KEY	OpenAI	https://api.openai.com/v1
OPENROUTER_API_KEY	OpenRouter	https://openrouter.ai/api/v1
FIREWORKS_API_KEY	Fireworks	https://api.fireworks.ai/inference/v1
LLM_COUNCIL_BASE_URL + LLM_COUNCIL_API_KEY	Custom OpenAI-compat	(user-supplied)

Override per-run with --provider openai|anthropic|openrouter|fireworks|custom.

Default model rosters per provider live in scripts/council.js and can be overridden via --models CSV and --chairman <id>.

Commands

node $SKILL_ROOT/scripts/council.js run "<query>" [--models id1,id2,id3] [--chairman id] [--provider <name>] [--wiki <slug>]
node $SKILL_ROOT/scripts/council.js providers
node $SKILL_ROOT/scripts/council.js show <session-id>

--wiki <slug> writes the full transcript to <wiki>/derived/council/<session-id>.md and registers it via wiki-cli.js page so it shows in FTS5 search.

Output

Each session writes:

~/.pro-workflow/council/<session-id>/
├── config.json           # query, models, chairman, provider
├── phase1_responses.json # raw API responses per model
├── phase2_rankings.json  # anonymized ranking outputs
├── phase3_synthesis.txt  # chairman's final answer
└── final_output.md       # human-readable bundle

Console prints the markdown bundle. Pipe to pbcopy / tee as needed.

Hard rules

1. Never skip the ranking phase. It's the core of the council pattern.
2. Save raw responses to disk verbatim. No summarization in storage.
3. Anonymize responses for ranking — models see Response A/B/C/..., not peer names.
4. The chairman sees both real names AND rankings.
5. Display all three phases to the user. No phase elision.

Cost awareness

The script logs per-call latency + tokens on supported providers. Multiply by your provider rate to estimate. Council cost grows linearly with len(models)^2 (each model ranks all others) plus the chairman.

Default council size: 3-5 models. More models = exponentially more ranking calls.

Use with wiki

/wiki council agent-memory "should we adopt episodic memory in our agents?"

Loads agent-memory wiki context as system prompt prefix, runs council, persists transcript as wiki/derived/council/<id>.md. The transcript becomes searchable via /wiki ask.