gemini-peer-review

Gemini Peer Review

A Claude Code plugin that validates Claude's designs, code reviews, and recommendations using Google Gemini CLI as a second AI perspective.

Why Peer Review?

Two AI perspectives catch more issues than one. Before presenting implementation plans, architecture recommendations, or code review findings, this plugin validates Claude's work through Gemini CLI to:

• Catch blind spots in analysis
• Identify alternative approaches
• Surface potential issues early
• Increase confidence in recommendations

Requirements

• Claude Code CLI
• Google Gemini CLI (v0.24.5 or later)

Installing Gemini CLI

# macOS (Homebrew)
brew install gemini-cli

# Or via npm
npm install -g @anthropic-ai/gemini-cli

Verify installation:

gemini --version
# Expected: 0.24.5 or later

Installation

Via Claude Code (Recommended)

1. In Claude Code, run /plugin add
2. Navigate to Add Marketplace
3. Enter the marketplace source: slb350/gemini-peer-review
4. Select the gemini-peer-review plugin to install

Manual Installation

1. Clone the repository:

   git clone https://github.com/slb350/gemini-peer-review.git ~/.claude/plugins/marketplaces/gemini-peer-review
   

2. In Claude Code, run /plugin add

3. Select gemini-peer-review from available plugins

Usage

Automatic Validation

Once installed, the plugin automatically reminds Claude to validate:

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

Slash Command

Trigger peer review manually:

# Review current changes (interactive)
/gemini-peer-review

# Review against a specific branch
/gemini-peer-review --base main

# Review uncommitted changes
/gemini-peer-review --uncommitted

# Validate a technical question
/gemini-peer-review "Should I use composition or inheritance here?"

Subagent Dispatch

For programmatic use within Claude's workflow:

Task tool with subagent_type="gemini-peer-review:gemini-peer-reviewer"
prompt: "Validate this design: [your design summary]"

How It Works

1. Claude forms a recommendation (design, code review, architecture)
2. Plugin dispatches to Gemini via subagent (keeps main context clean)
3. Gemini validates independently using focused prompts
4. Results are compared:
   • Agreement: Synthesize and present with high confidence
   • Disagreement: Run 2-round discussion protocol
   • Critical issue: Escalate immediately (security, architecture, breaking changes)
5. Synthesized result returned to user

Discussion Protocol

When Claude and Gemini disagree:

• Round 1: Both state positions with evidence
• Round 2: Respond to evidence, attempt synthesis
• Escalation: If unresolved, use Perplexity MCP or WebSearch for external arbitration

Repository Structure

gemini-peer-review/
├── .claude-plugin/
│   └── marketplace.json         # Marketplace manifest
├── README.md
└── plugins/
    └── gemini-peer-review/
        ├── .claude-plugin/
        │   └── plugin.json      # Plugin manifest
        ├── agents/
        │   └── gemini-peer-reviewer.md
        ├── commands/
        │   └── gemini-peer-review.md
        ├── skills/
        │   └── gemini-peer-review/
        │       ├── SKILL.md
        │       ├── discussion-protocol.md
        │       ├── escalation-criteria.md
        │       └── common-mistakes.md
        └── hooks/
            ├── hooks.json
            └── *.sh

Gemini CLI Specifics

This plugin accounts for key differences between Gemini CLI and other tools:

Aspect	Approach
Model	gemini-3-pro-preview (default)
Output	Always -o json for structured parsing
Sandbox	Always -s flag for safety
Tool prevention	Every prompt ends with CRITICAL: Do not use any tools. Output text only.
Code review	No built-in review command - uses git diff + prompts
Session continuity	Context re-injection (not --resume)

Output Formats

Validated (Agreement)

## Peer Review Result
**Status:** Validated
**Confidence:** High (both AIs aligned)

[Synthesized recommendations]

Resolved Through Discussion

## Peer Review Result
**Status:** Resolved through discussion

**Initial Positions:**
- Claude: [position]
- Gemini: [position]

**Resolution:** [how resolved]
**Final Recommendation:** [synthesized view]

Escalated

## Peer Review Result
**Status:** Escalated for external research
**Source:** [Perplexity | WebSearch]

**Research Findings:** [authoritative answer]
**Final Recommendation:** [based on findings]

Configuration