ccg

CCG - Claude-Codex-Gemini Tri-Model Orchestration

CCG spawns a tmux team with Codex and Gemini CLI workers running in parallel panes, then Claude synthesizes the results. Use this for tasks that benefit from multiple AI perspectives simultaneously.

When to Use

• Backend/analysis + frontend/UI work that can run truly in parallel
• Code review from multiple perspectives (architecture + style simultaneously)
• Research tasks where different models have complementary strengths
• Any task you want to split across Codex (analytical) and Gemini (design/creative) workers

Requirements

• Codex CLI: npm install -g @openai/codex (or @openai/codex)
• Gemini CLI: npm install -g @google/gemini-cli
• tmux: Must be running inside a tmux session
• If either CLI is unavailable, CCG falls back to Claude-only execution

How It Works

1. Claude decomposes the request into:
   - Backend/analytical tasks → Codex worker
   - Frontend/UI/design tasks → Gemini worker

2. mcp__team__omc_run_team_start creates a tmux session with 2 workers:
   omc-team-{name}
   ├── Leader pane (Claude orchestrates)
   ├── Worker pane 1: codex CLI (analytical tasks)
   └── Worker pane 2: gemini CLI (design tasks)

3. Workers read tasks from inbox files and write done.json on completion

4. mcp__team__omc_run_team_wait blocks until all workers finish

5. Claude reads taskResults and synthesizes into final output

Execution Protocol

When invoked, Claude MUST follow this workflow:

1. Decompose Request

Split the user's request into:

• Codex tasks: code analysis, architecture review, backend logic, security review, test strategy
• Gemini tasks: UI/UX design, documentation, visual analysis, large-context file review
• Synthesis task: Claude combines results (always done by Claude, not delegated)

Choose a short teamName slug (e.g., ccg-auth-review).

2. Start the team (non-blocking)

Call mcp__team__omc_run_team_start — spawns workers in the background and returns a jobId immediately:

mcp__team__omc_run_team_start({
  "teamName": "ccg-{slug}",
  "agentTypes": ["codex", "gemini"],
  "tasks": [
    {"subject": "Codex task: ...", "description": "Full description of analytical work..."},
    {"subject": "Gemini task: ...", "description": "Full description of design/UI work..."}
  ],
  "cwd": "{cwd}"
})

Returns: { "jobId": "omc-...", "pid": 12345, "message": "Team started in background..." }

3. Wait for completion

Call mcp__team__omc_run_team_wait — blocks internally until done:

mcp__team__omc_run_team_wait({
  "job_id": "{jobId}",
  "timeout_ms": 60000
})

Timeout guidance: timeout_ms is optional; the default wait timeout is fine. If wait times out, workers/panes keep running. Call omc_run_team_wait again to keep waiting. Use omc_run_team_cleanup only for explicit cancel intent.

Returns when done:

{
  "status": "completed|failed",
  "result": {
    "taskResults": [
      {"taskId": "1", "status": "completed", "summary": "..."},
      {"taskId": "2", "status": "completed", "summary": "..."}
    ]
  }
}

4. Synthesize Results

Parse result.taskResults and synthesize Codex + Gemini outputs into a unified response for the user.

Fallback (CLIs Not Available)

CLI availability is checked by the MCP runtime automatically. If a CLI is not installed, the worker exits with command not found. In that case, fall back to Claude Task agents:

[CCG] Codex/Gemini CLI not found. Falling back to Claude-only execution.

Use standard Claude Task agents instead:

• Task(subagent_type="oh-my-claudecode:executor", model="sonnet", ...) for analytical tasks
• Task(subagent_type="oh-my-claudecode:designer", model="sonnet", ...) for design tasks

Invocation

/oh-my-claudecode:ccg [task description]

Example:

/oh-my-claudecode:ccg Review this PR - check architecture and code quality (Codex) and UI components (Gemini)