consensus

 ███████╗██████╗ ███████╗██╗  ██╗
 ██╔════╝██╔══██╗██╔════╝██║ ██╔╝
 ███████╗██████╔╝███████╗█████╔╝
 ╚════██║██╔═══╝ ╚════██║██╔═██╗
 ███████║██║     ███████║██║  ██╗
 ╚══════╝╚═╝     ╚══════╝╚═╝  ╚═╝
                       consensus

Consensus

3-model consensus validation for any conclusion. Dispatches your analysis to Claude Opus, OpenAI Codex, and Google Gemini in parallel. Each model evaluates from a different analytical lens. Results synthesized with confidence scores on a 1-10 scale.

Not limited to code. Use it for architecture decisions, design direction, root cause analysis, feature prioritization, strategy calls — anything where you've formed a conclusion and want independent validation.

Install

1. Install the plugin

# Add SpSk marketplace (one time)
claude plugins marketplace add spsk-dev/marketplace

# Install consensus
claude plugins install consensus@spsk

2. Set up external models (recommended for full 3-model validation)

# Codex CLI (Devil's Advocate validator)
npm install -g @openai/codex

# Gemini CLI (Scope Analyst validator)
npm install -g @anthropic/gemini-cli

Without external CLIs, consensus falls back to 3 Claude agents with distinct personas. Works, but you lose the cross-model diversity that makes consensus valuable.

Verify installation

claude /help
# Look for: /consensus

Manual install (alternative)

git clone https://github.com/spsk-dev/consensus.git
cd consensus && bash install.sh

Usage

# Validate any conclusion
/consensus "GraphQL is better than REST for our mobile API"

# Type-aware validation — changes rubric and validator focus
/consensus --type architecture "Microservices is the right call"
/consensus --type plan "Migration plan in migration.md"
/consensus --type timeline "Ship auth rewrite by March 15"
/consensus --type security "JWT in httpOnly cookies is secure"
/consensus --type idea "Build a CLI instead of a web app"
/consensus --type design "Dark mode with warm palette"

# Include supporting evidence
/consensus --evidence @analysis.md "The auth middleware is the root cause"

# Quick mode (2 validators)
/consensus --quick "We should prioritize billing over notifications"

Validation Types

Type	Key Question	Rubric Focus
general	Is this conclusion correct?	Evidence quality, logic, alternatives, completeness
architecture	Will this architecture work?	Feasibility, dependencies, migration, scalability
plan	Will this plan succeed?	Prerequisites, sequencing, scope, rollback, evaluation
timeline	Is this timeline realistic?	Dependencies, parallelism, buffer, resources
idea	Is this worth pursuing?	Problem fit, effort/impact, alternatives, reversibility
design	Is this the right direction?	Intent match, consistency, accessibility, originality
security	Is this approach secure?	Threat surface, auth, data flow, trust boundaries

How It Works

Your conclusion
     │
     ├──► Deep Verifier (Opus)     ── Traces claims back to evidence
     ├──► Devil's Advocate (Codex) ── Finds the strongest counter-argument
     └──► Scope Analyst (Gemini)   ── Checks second-order effects & missing perspectives
     │
     ▼
Consensus Report (1-10 confidence)

Three validators, three lenses:

Validator	Model	Focus
Deep Verifier	Claude Opus	Verify every claim against evidence. Read actual files. Don't trust summaries.
Devil's Advocate	OpenAI Codex	Find the strongest counter-argument. Check for confirmation bias.
Scope Analyst	Google Gemini	Second-order effects. Missing perspectives. Reversibility.

Output

 SpSk  consensus  v1.0.0  ───  3 models  ·  opus+codex+gemini

┌─────────────────────────────────────────────────────────────┐
│  VALIDATOR VERDICTS                                         │
└─────────────────────────────────────────────────────────────┘

 Deep Verifier    (Opus)    8/10  Claims verified against codebase
 Devil's Advocate (Codex)   6/10  Found one unaddressed alternative
 Scope Analyst    (Gemini)  7/10  Migration cost not fully considered

 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 Consensus Confidence:  7.0/10  MODERATE CONFIDENCE
 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Confidence Scale

Score	Verdict	Meaning
8-10	HIGH CONFIDENCE	Strong agreement — proceed
6-7.9	MODERATE CONFIDENCE	Likely correct with caveats
4-5.9	LOW CONFIDENCE	Significant disagreement — investigate
1-3.9	NO CONSENSUS	Critical flaws found — do not proceed

Degradation

The plugin adapts to available tools:

• Full (Opus + Codex + Gemini): Maximum reasoning diversity
• Partial (Opus + 1 external): One external perspective
• Fallback (Opus + 2x Sonnet): Different personas, same model family

Requirements