Council v2 — Multi-Model Review System for OpenClaw

A hardened, multi-model council review skill for OpenClaw.

Spawns 3-5 independent AI reviewers across different providers, collects structured JSON verdicts, and applies mechanical synthesis — the vote count is the verdict, not the orchestrator's opinion.

Features

• Two tiers: Standard (3 reviewers) and Full (5 reviewers)
• 5 providers, 5 training biases: Anthropic, OpenAI, xAI, DeepSeek, Google
• Mechanical synthesis: Majority vote decides. Critical findings auto-block. Synthesizer cannot override.
• Anti-consensus check: Unanimous votes are flagged — agreement is hypothesis, not proof.
• Minority reports preserved: Strongest dissent is always surfaced.
• Monthly retro: Track past decisions against outcomes for continuous calibration.

Install

clawhub install council-v2

Or manually:

git clone https://github.com/zurbrick/council-v2.git ~/.openclaw/workspace/skills/council-v2

Quick Start

bash scripts/council.sh review code src/auth.py
bash scripts/council.sh review plan proposal.md --tier full
bash scripts/council.sh review architecture design.md --tier full
bash scripts/council.sh review decision options.md --options "SQLite,Postgres,Cloud SQL"

For the full workflow, review types, synthesis rules, and operational details, see SKILL.md.

Model Access

The council's strength comes from model diversity — different providers catch different blind spots. You have two options:

Option A: OpenRouter (recommended for most users)

One API key, all 5 models. Sign up at openrouter.ai.

// openclaw.json — models.providers
"openrouter": {
 "baseUrl": "https://openrouter.ai/api/v1",
 "apiKey": "${OPENROUTER_API_KEY}",
 "api": "openai-completions"
}

Then reference models as openrouter/anthropic/claude-opus-4, openrouter/openai/gpt-5.4, openrouter/x-ai/grok-4, openrouter/deepseek/deepseek-r1, openrouter/google/gemini-3.1-pro.

Option B: Direct provider subscriptions

If you have separate subscriptions (Anthropic, OpenAI, xAI, etc.), configure each provider individually. See OpenClaw model docs.

Tip: The specific models matter less than the diversity. What you want is different providers with different training data and different biases reviewing the same decision. Swap in whatever top-tier models you have access to.

Council Tiers

Standard (3 reviewers)

Role	Default Model	Provider
Architecture Synthesizer	Opus	Anthropic
Adversarial Critic	GPT-5.4	OpenAI
Security & Risk	Grok 4	xAI

Full (5 reviewers)

Role	Default Model	Provider
Architecture Synthesizer	Opus	Anthropic
Adversarial Critic	GPT-5.4	OpenAI
Security & Risk	Grok 4	xAI
First Principles	DeepSeek R1	DeepSeek
Structural Verifier	Gemini 3.1 Pro	Google

Repository Contents

File	Purpose
SKILL.md	Operational skill — workflow, when to use, interpreting results
references/review-types.md	Review type definitions and tier recommendations
references/role-prompts.md	Reviewer role prompts and shared output instructions
references/schema.md	JSON schemas for reviewer and synthesis output
references/synthesis-rules.md	Mechanical synthesis protocol and edge cases
scripts/council.sh	Orchestration script
scripts/synthesize.py	Mechanical synthesis engine
scripts/retro.sh	Monthly retrospective template generator

Origin

Built by Don Zurbrick — battle-tested across plugin evaluations, security audits, architecture reviews, and one memorable session where the council reviewed itself and called out its own biases.

License

MIT