agent-peer-review-marketplace

Agent Peer Review

A Claude Code plugin marketplace for AI-to-AI peer validation. Multiple perspectives catch more issues than one.

Available Plugins

codex-peer-review

Peer validation using OpenAI Codex CLI. Validates Claude's designs, code reviews, and recommendations by getting a second opinion before presenting them to users.

Installation

# Add this marketplace
/plugin marketplace add jcputney/agent-peer-review

# Install the Codex peer review plugin
/plugin install codex-peer-review

Prerequisites for codex-peer-review

Account Requirements:

• Claude Code: Requires a Claude Pro ($20/mo), Max, Team Premium, or Enterprise subscription
• OpenAI Codex CLI: Included with ChatGPT Plus/Pro/Business/Edu/Enterprise plans, or use OpenAI API credits

Install Codex CLI:

# Install Codex CLI
npm i -g @openai/codex

# Authenticate (opens browser for OAuth or prompts for API key)
codex login

Optional - Perplexity MCP (for escalation):

When Claude and Codex disagree and can't resolve through discussion, the plugin escalates to external research. If you have Perplexity MCP configured, it will be used for expert arbitration. Otherwise, the plugin falls back to WebSearch.

How Peer Review Works

1. Claude Forms Opinion: Claude analyzes the task and forms a recommendation
2. Dispatch to Reviewer: A subagent runs the peer review agent to get a second opinion
3. Compare Findings: Results are classified as agreement, disagreement, or complement
4. Resolution:
   • Agreement: Synthesize and present combined view
   • Disagreement: Enter 2-round discussion protocol
   • Persistent Conflict: Escalate to Perplexity (or WebSearch) for arbitration

codex-peer-review Features

• Automatic Reminders: Reminds Claude to run peer review before presenting implementation plans, code reviews, or architecture recommendations
• Structured Disagreement Resolution: 2-round discussion protocol with evidence-based arguments
• Expert Arbitration: Escalates to Perplexity MCP (or WebSearch fallback) for unresolved disputes or security concerns
• Slash Command: Explicit /codex-peer-review command for on-demand validation

Usage

Automatic Mode

The plugin reminds Claude to trigger peer review when about to present:

• Implementation plans or designs
• Code review results
• Architecture recommendations
• Major refactoring proposals
• Answers to broad technical questions

Slash Command

# Review current changes against default branch
/codex-peer-review

# Review against specific branch
/codex-peer-review --base develop

# Validate answer to a broad question
/codex-peer-review "Should we use microservices or monolith for this project?"

Command Selection

Validation Type	Codex Command	Use When
Code Review	codex review --base X	Reviewing actual code changes (diffs)
Design Validation	codex exec "..."	Validating proposals, plans, architecture
Question Answering	codex exec "..."	Answering broad technical questions

Permissions

The plugin uses heredoc stdin to minimize permission prompts. On first use, you'll be asked to approve:

Pattern	Purpose
codex exec*	Run focused peer review prompts
codex review*	Review git diffs against branches

Tip: When prompted, select "Always allow" to avoid repeated prompts.

Marketplace Structure

agent-peer-review/
├── .claude-plugin/
│   └── marketplace.json           # Marketplace registry
└── plugins/
    └── codex-peer-review/         # Codex CLI peer review plugin
        ├── .claude-plugin/
        │   └── plugin.json        # Plugin metadata
        ├── agents/
        │   └── codex-peer-reviewer.md
        ├── skills/
        │   └── codex-peer-review/
        │       ├── SKILL.md
        │       ├── discussion-protocol.md
        │       ├── escalation-criteria.md
        │       └── common-mistakes.md
        ├── commands/
        │   └── codex-peer-review.md
        └── hooks/
            ├── hooks.json
            ├── peer-review-reminder.sh
            ├── stop-peer-review-check.sh
            └── plan-peer-review-check.sh

Architecture

Key Design: All peer review work runs in a subagent context to keep the main conversation clean.