GitHub - responsiblparty/cc-doubleteam

2 min read Original article ↗

Three-phase project mode for Claude Code: Claude plans, Codex executes, Claude reviews — execution burns your ChatGPT limits, not your Claude limits.

Why

You get Fable-quality planning plus Codex execution without burning your Claude usage on the heavy lifting. The models work in sequence, each at their best effort level.

Prerequisites

  • Claude Code installed
  • Codex CLI: npm install -g @openai/codex then codex login (requires paid ChatGPT Plus subscription)
  • Codex plugin in Claude Code: 
    /plugin marketplace add openai/codex-plugin-cc
/plugin install codex@openai-codex

Install

git clone https://github.com/responsiblparty/cc-doubleteam
cp -r cc-doubleteam/skills/doubleteam ~/.claude/skills/doubleteam

Optionally copy CLAUDE.md contents into your project's CLAUDE.md to enable proactive suggestions:

cat cc-doubleteam/CLAUDE.md >> /path/to/your/project/CLAUDE.md

Restart Claude Code after install. Type /doubleteam to confirm it's working.

Usage

  • /doubleteam <task> — starts immediately with your task
  • /doubleteam — Claude asks what the task is
  • Claude will also proactively suggest it when a session is heading into a substantial build

How Phase 2 works

Phase 2 runs Codex by invoking the companion script directly via Bash — not through the codex:rescue subagent. Subagents inherit the session's original permission mode and will hit a Bash permission wall before they can invoke Codex. The skill uses dynamic path discovery so it works across any Codex plugin version:

COMPANION=$(ls ~/.claude/plugins/cache/openai-codex/codex/*/scripts/codex-companion.mjs 2>/dev/null | tail -1)
node "$COMPANION" task --effort xhigh --write "<plan>"

Phase 2 runs in the background by default — Claude notifies you when Codex finishes and you can keep talking in the meantime.

The three phases

Phase Model Effort Role
1 — Plan Claude Fable (fallback: Opus) high Produces step-by-step implementation plan; pauses for your approval
2 — Execute Codex xhigh Implements the plan; burns ChatGPT limits, not Claude limits
3 — Review Claude Fable (fallback: Opus) max Reviews output against plan; flags issues without rewriting

Skipping phases

  • just plan → Phase 1 only
  • skip review → Phases 1 and 2 only
  • If Codex is unavailable → falls back to Claude execution (noted in output)

License

MIT