caucus

name	caucus
description	Multi-LLM consensus engine. Queries multiple AI models and synthesizes a single answer using strategies like voting, judge synthesis, and multi-round debate. Use when the user wants to query multiple LLMs, compare model outputs, get a consensus answer, or needs a synthesized response from several AI models.
allowed-tools	["Bash(caucus:*)","Bash(cat:*)","Bash(mktemp:*)","Read","Write","Grep","Glob"]
user-invocable	true
argument-hint	[prompt to send to multiple models]
metadata	{"author":"bradleydwyer","version":"1.0.3","status":"experimental"}

Caucus -- Multi-LLM Consensus Engine

Queries multiple LLM models on a single prompt, then produces one consensus result. Strategies range from simple majority vote to multi-round debate with a judge synthesizing the final answer.

When to Use This Skill

• User wants to ask multiple LLMs the same question
• User wants a consensus or synthesized answer across models
• User wants to compare how different models respond
• User needs higher confidence on a tricky question
• User asks to "debate" or "vote" on something across models

Prerequisites

API keys as env vars: OPENAI_API_KEY, ANTHROPIC_API_KEY, GOOGLE_API_KEY, XAI_API_KEY. Only keys for models you want to query are required.

If caucus --help fails, tell the user to install it before proceeding.

Workflow

1. Determine Intent

User Says	Action
"ask multiple models", "get consensus"	Standard consensus query
"compare models", "how do different models answer"	Strategy comparison via compare
"debate this", "have models argue about"	Debate strategy (-s debate)
specific model names mentioned	Select with -m

2. Run the Query

Always use -f json for parseable output.

caucus "prompt" -f json                                          # all models, judge strategy
caucus "prompt" -m gpt-5.2,claude-opus-4-6 -f json              # specific models
caucus "prompt" -s debate -f json                                # debate strategy
caucus debate "prompt" --rounds 3 --live --format plain          # stream the debate live
caucus "prompt" --profile deep -f json                           # council profile (exact members)
caucus "prompt" --auto -f json                                   # zero-key council, no API keys
caucus compare "prompt" --strategies majority-vote,judge -f json # compare strategies
caucus "prompt" -v -f json                                       # verbose (individual responses)
caucus "prompt" -f supreme-court                                 # dissent/concurrence format

Strategies: majority-vote, weighted-vote, judge (default), debate, debate-then-vote

Output formats: plain (default), json, supreme-court, detailed

2a. Review Code or Diffs

For adjudicated code/text review (independent reviews → anonymized blind peer votes → judged verdicts with a provenance receipt):

caucus review --file src/main.rs --format json      # one file
caucus review --git-diff                            # working-tree diff
caucus review --staged --profile deep               # staged diff, exact council
git diff | caucus review --format json              # stdin
caucus review --staged --manifest run.json --resume # checkpoint + resume

Receipts are Markdown (default) or JSON (--format), with deterministic run ids, raw votes/dissent, adjudications, and unanimous/majority/disputed classifications. --max-requests is a hard cap; --budget-usd is advisory.

2b. Inspect Councils

caucus doctor          # rendered terminal readiness tables (--json available)
caucus profiles        # list profiles and exact members (or a name / --json)

Adapters are labeled stable (claude, codex, ollama, lmstudio) or experimental (kimi, opencode, gemini, grok, acp). The built-in deep profile: claude:opus@xhigh, claude:claude-fable-5@xhigh, codex:default@xhigh, opencode:zai-coding-plan/glm-5.2@xhigh, kimi:kimi-code/k3@high, grok:grok-4.5@high.

Profile deadline_secs is a whole-run wall-clock limit; request_timeout_secs bounds each provider request independently. Adapter timeout_secs can impose a stricter process limit.

3. Present Results

Parse the JSON and present:

• The consensus answer -- lead with this, prominently
• Strategy used and model agreement/divergence
• Individual model positions -- summarize briefly if there was disagreement
• Confidence level -- how strong was the consensus?

For compare runs, present each strategy's result and which produced the strongest consensus.

Use -v to re-run if the user wants full reasoning or individual model responses.

Tips

• Judge is the default for good reason. It synthesizes rather than counting votes -- better for nuanced questions.
• Debate works best for controversial or ambiguous topics where models should challenge each other.
• Majority vote is fastest but loses nuance. Use for factual questions with clear answers.
• Use compare when unsure which strategy fits.
• Fewer models = faster. Use -m to pick 2-3 when speed matters.