CCG - Claude + Codex + Gemini Multi-Model Collaboration

简体中文 | English

A multi-model collaboration development system where Claude Code orchestrates Codex + Gemini. Frontend tasks route to Gemini, backend tasks route to Codex, and Claude handles orchestration and code review.

Installation

npx ccg-workflow

Requirements: Claude Code CLI, Node.js 20+

Important

: This project depends on ora@9.x and string-width@8.x, which require Node.js >= 20. Using Node.js 18 will cause SyntaxError: Invalid regular expression flags. Please upgrade to Node.js 20 or higher.

Optional: Codex CLI (backend), Gemini CLI (frontend)

First-Time Setup

On first run, CCG will prompt you to select your preferred language (English or Chinese). This preference is saved and used for all subsequent sessions.

Commands

Command	Description
/ccg:workflow	Full 6-phase development workflow
/ccg:plan	Multi-model collaborative planning (Phase 1-2)
/ccg:execute	Multi-model collaborative execution (Phase 3-5)
/ccg:feat	Smart feature development
/ccg:frontend	Frontend tasks (Gemini)
/ccg:backend	Backend tasks (Codex)
/ccg:analyze	Technical analysis
/ccg:debug	Problem diagnosis
/ccg:optimize	Performance optimization
/ccg:test	Test generation
/ccg:review	Code review
/ccg:commit	Git commit
/ccg:rollback	Git rollback
/ccg:clean-branches	Clean branches
/ccg:worktree	Worktree management
/ccg:init	Initialize CLAUDE.md
/ccg:enhance	Prompt enhancement
/ccg:spec-init	Initialize OPSX environment
/ccg:spec-research	Requirements → Constraints
/ccg:spec-plan	Constraints → Zero-decision plan
/ccg:spec-impl	Execute plan + archive
/ccg:spec-review	Dual-model cross-review
/ccg:team-research	Agent Teams requirements → constraints
/ccg:team-plan	Agent Teams constraints → parallel plan
/ccg:team-exec	Agent Teams parallel execution
/ccg:team-review	Agent Teams dual-model review
/ccg:codex-exec	Codex full execution (plan → code → review)

OPSX Spec-Driven (v1.7.52+)

Integrates OPSX architecture to turn requirements into constraints, eliminating AI improvisation:

# Initialize OPSX environment
/ccg:spec-init

# Research requirements → output constraints
/ccg:spec-research implement user authentication

# Parallel analysis → zero-decision plan
/ccg:spec-plan

# Execute the plan
/ccg:spec-impl

# Independent review (available anytime)
/ccg:spec-review

Note: /ccg:spec-* commands are CCG's wrapper around OPSX, internally calling /opsx:* commands. You can /clear between phases — state is persisted in the openspec/ directory, so context overflow is not a concern.

Agent Teams Parallel Execution (v1.7.60+)

Leverage Claude Code Agent Teams experimental feature to spawn multiple Builder teammates for parallel coding:

# 1. Requirements research → constraints (recommended for complex projects)
/ccg:team-research implement real-time collaboration kanban API

# 2. /clear then plan → zero-decision parallel plan
/ccg:team-plan kanban-api

# 3. /clear then parallel execution → Builder teammates write code in parallel
/ccg:team-exec

# 4. /clear then review → dual-model cross-review
/ccg:team-review

Prerequisite: Manually enable Agent Teams (settings.json: set CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1)

vs Traditional Workflow: Team series uses /clear between steps to isolate context, passing state through files. Builders work in parallel, ideal for tasks decomposable into 3+ independent modules.

Planning & Execution Separation

# 1. Generate implementation plan
/ccg:plan implement user authentication

# 2. Review the plan (editable)
# Plan saved to .claude/plan/user-auth.md

# 3a. Execute (Claude refactors) — fine-grained control
/ccg:execute .claude/plan/user-auth.md

# 3b. Execute (Codex does everything) — efficient, low Claude token usage
/ccg:codex-exec .claude/plan/user-auth.md

Configuration

Directory Structure

~/.claude/
├── commands/ccg/       # Slash commands
├── agents/ccg/         # Sub-agents
├── skills/             # Quality gates + multi-agent orchestration
├── bin/codeagent-wrapper
└── .ccg/
    ├── config.toml
    └── prompts/{codex,gemini}/

Environment Variables

Variable	Description	Default
CODEAGENT_POST_MESSAGE_DELAY	Wait time after Codex completion (seconds)	5
CODEX_TIMEOUT	codeagent-wrapper execution timeout (seconds)	7200
BASH_DEFAULT_TIMEOUT_MS	Claude Code Bash default timeout (ms)	120000
BASH_MAX_TIMEOUT_MS	Claude Code Bash max timeout (ms)	600000

Configuration in ~/.claude/settings.json:

{
  "env": {
    "CODEAGENT_POST_MESSAGE_DELAY": "1",
    "CODEX_TIMEOUT": "7200",
    "BASH_DEFAULT_TIMEOUT_MS": "600000",
    "BASH_MAX_TIMEOUT_MS": "3600000"
  }
}

codeagent-wrapper Auto-Authorization

CCG automatically installs a Hook in settings.json to auto-authorize codeagent-wrapper commands, so you don't have to manually approve each collaboration call.

Requirement: jq must be installed on your system.

# macOS
brew install jq

# Linux (Debian/Ubuntu)
sudo apt install jq

# Linux (RHEL/CentOS)
sudo yum install jq

# Windows
choco install jq
# or
scoop install jq

MCP Configuration

Code retrieval MCP (choose one):

• ace-tool (recommended) - Code search via search_context (note: enhance_prompt is no longer available). Official | Third-party proxy (recommended)
• ContextWeaver (alternative) - Local hybrid search, requires SiliconFlow API Key (free)

Optional MCP tools:

• Context7 - Latest library documentation
• Playwright - Browser automation/testing
• DeepWiki - Knowledge base queries
• Exa - Search engine (requires API Key)

# Configure MCP
npx ccg-workflow menu  # Select "Configure MCP"

Tools

npx ccg-workflow menu  # Select "Tools"

• ccusage - Claude Code usage analytics
• CCometixLine - Status bar tool (Git + usage tracking)

Install Claude Code

npx ccg-workflow menu  # Select "Install Claude Code"

Supports multiple installation methods: npm, homebrew, curl, powershell, cmd

Update / Uninstall

# Update
npx ccg-workflow@latest          # npx users
npm install -g ccg-workflow@latest  # npm global users

# Uninstall
npx ccg-workflow  # Select "Uninstall"
npm uninstall -g ccg-workflow  # npm global users need this extra step

FAQ

1. How to auto-authorize codeagent-wrapper without manual approval?

CCG automatically installs the Hook during setup (v1.7.71+). If you installed an older version, add this to ~/.claude/settings.json:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Bash",
        "hooks": [
          {
            "type": "command",
            "command": "jq -r '.tool_input.command' | grep -q 'codeagent-wrapper' && echo '{\"hookSpecificOutput\": {\"hookEventName\": \"PreToolUse\", \"permissionDecision\": \"allow\", \"permissionDecisionReason\": \"codeagent-wrapper auto-approved\"}}' || exit 1",
            "timeout": 1
          }
        ]
      }
    ]
  }
}

Note

: Requires jq installed on your system. See codeagent-wrapper Auto-Authorization for install instructions.

2. Codex CLI 0.80.0 process does not exit

In --json mode, Codex does not automatically exit after output completion.

Fix: Set CODEAGENT_POST_MESSAGE_DELAY=1

Architecture

Claude Code (Orchestrator)
       │
   ┌───┴───┐
   ↓       ↓
Codex   Gemini
(Backend) (Frontend)
   │       │
   └───┬───┘
       ↓
  Unified Patch

External models have no write access — they only return patches, which Claude reviews before applying.

Credits

• cexll/myclaude - codeagent-wrapper
• UfoMiao/zcf - Git tools
• GudaStudio/skills - Routing design
• ace-tool - MCP tool

Star History

License

MIT

---

v1.7.75 | Issues