fun-writer

multi-ai-reviewer

A Claude Code plugin that orchestrates code reviews using three independent AI reviewers — Codex CLI (OpenAI), Gemini CLI (Google), and Claude Sonnet (Anthropic) — running in parallel. Every finding is validated against actual source code before presenting to you.

What It Does

/review-loop [base-branch]

Runs an iterative review cycle:

1. Gathers your branch diff against base-branch (defaults to main) 2. Launches all three reviewers in parallel 3. Validates every finding against actual source code (rejects false positives) 4. Presents a report in 繁體中文 with severity, root cause, and fix suggestions 5. Asks you what to do — auto-fix all, discuss, or skip 6. If fixes were applied, re-reviews with all three reviewers and loops back

The loop ends when all reviewers find zero issues, you say stop, or after 10 rounds.

/discuss-with-codex <topic>

A multi-turn debate between Claude and Codex for brainstorming and planning:

1. Claude frames the problem and opens discussion with Codex 2. Claude critically analyzes Codex's response — challenges weak points, proposes alternatives 3. Back-and-forth continues for 2–5 rounds until convergence 4. Presents conclusions in 繁體中文 with consensus, disagreements, and recommended approach

Prerequisites

You need these CLI tools installed and authenticated:

| Tool | Install | Auth | |------|---------|------| | Claude Code | npm install -g @anthropic-ai/claude-code | claude (follow prompts) | | Codex CLI | npm install -g @openai/codex | Set OPENAI_API_KEY | | Gemini CLI | npm install -g @anthropic-ai/gemini-cli or see repo | gemini (follow prompts) |

Verify all three work:

claude --version
codex --version
gemini --version

Installation

Option A: Install from GitHub (Recommended)

Inside Claude Code, add this repo as a marketplace and install:

/plugin marketplace add git@github.com:appleternity/multi-ai-reviewer.git
/plugin install multi-ai-reviewer@multi-ai-reviewer

That's it. The commands are now available in all your Claude Code sessions.

Option B: Local Directory Install

Clone and register as a local marketplace:

# Clone to any location you like
git clone git@github.com:appleternity/multi-ai-reviewer.git ~/multi-ai-reviewer

Then inside Claude Code:

/plugin marketplace add ~/multi-ai-reviewer
/plugin install multi-ai-reviewer@multi-ai-reviewer

Option C: One-Off Testing

For a single session without permanent installation:

claude --plugin-dir ~/multi-ai-reviewer

Scope Options

By default plugins install to user scope (available everywhere). You can also:

# Project scope — shared with team via .claude/settings.json
/plugin install multi-ai-reviewer@multi-ai-reviewer --scope project

# Local scope — just for you on this repo, gitignored
/plugin install multi-ai-reviewer@multi-ai-reviewer --scope local

Usage

Start a review loop on your current branch:

/review-loop

Or specify a base branch:

/review-loop develop

Brainstorm with Codex:

/discuss-with-codex "Should we use WebSockets or SSE for real-time updates?"

How the Review Loop Works

┌─────────────────────────────────────────┐
│           Gather branch diff            │
└──────────────────┬──────────────────────┘
                   │
    ┌──────────────┼──────────────┐
    ▼              ▼              ▼
┌────────┐  ┌──────────┐  ┌───────────┐
│ Codex  │  │  Gemini  │  │  Claude   │
│  CLI   │  │   CLI    │  │  Sonnet   │
└───┬────┘  └────┬─────┘  └─────┬─────┘
    └──────────────┼──────────────┘
                   ▼
    ┌──────────────────────────────┐
    │  Validate findings against   │
    │     actual source code       │
    └──────────────┬───────────────┘
                   ▼
    ┌──────────────────────────────┐
    │  Present report (繁體中文)    │
    └──────────────┬───────────────┘
                   ▼
    ┌──────────────────────────────┐
    │  Ask user: fix / discuss /   │
    │           stop               │
    └──────────────┬───────────────┘
                   ▼
    ┌──────────────────────────────┐
    │  Apply fixes → re-review ──────── loop
    └──────────────────────────────┘

Each reviewer brings different strengths:

| Reviewer | Access Method | Session Resume | Strengths | |----------|--------------|----------------|-----------| | Codex | codex exec review | Yes (resume --last) | Deep reasoning, code understanding | | Gemini | gemini -p --yolo | Yes (--resume latest) | Shell access for exploration | | Claude Sonnet | Task subagent | No (full context each time) | Direct file access via tools |

License

MIT