Claude Codex Handoff

codex-handoff

**[English](./README.md) | [简体中文](./README.zh-CN.md)**

  

**Claude Code plans, a Claude subagent implements, a fresh OpenAI Codex session adversarially reviews.** A three-phase workflow for code changes that deserve a written spec — the model that writes the code is never the model that grades it.

---

Why this exists

Claude Code is excellent at exploring code, asking the right clarifying questions, and — through a subagent with full host access — implementing against a spec without leaving the session. What it cannot give itself is a credible second opinion: the same model that wrote the code will rationalize its own choices. Codex CLI reading the finished diff cold is that second opinion from a **different model**.

This skill encodes the protocol so you get, every task:

• A written spec before any code is touched
• An implementer subagent bound to that spec, with no license to improvise
• An adversarial reviewer from a different model with no memory of why the code was written this way

How it works

┌────────────────────────────────────────────────────────────┐
│  PHASE 1 — PLAN    (Claude Code)                           │
│  Probe code → ask one clarifying question at a time →      │
│  write .agent/specs/.md → STOP for your approval     │
└──────────────────────────┬─────────────────────────────────┘
                           │ you reply "approved"
                           ▼
┌────────────────────────────────────────────────────────────┐
│  PHASE 2 — IMPLEMENT                                       │
│  Claude:      git switch -c feat/                    │
│  Implementer: edit files + list cmds in spec §9 (no commit)│
│  Verifier:    run §9 cmds on host → report tails           │
│  Claude:      recor