codex-handoff-skill（Codex 交接技能）

Codex Handoff — Supervisor Loop

Claude Code is the supervisor/judge. Codex CLI is the executor. You orchestrate a loop: send plan to Codex, review results, re-run if incomplete, until the task is comprehensively done.

Announce at start: "Using codex-handoff to orchestrate Codex CLI execution of the plan."

Prerequisites

• Codex CLI must be installed (codex --version)
• A plan must exist (in docs/plans/, .claude/plans/, or provided inline)
• The working directory should be a git repo (for diff-based review)

Quick Start

1. Locate the plan (from docs/plans/, .claude/plans/, or user arguments)
2. Detect phases — scan for phase headings; if found, execute phase-by-phase
3. Build a structured Codex prompt with plan (or phase) + project context + coding standards
4. Execute Codex CLI in --full-auto mode
5. Review git diff, run tests, audit completion with a scorecard
6. Decide — loop with correction prompt if items remain, or advance to next phase / finalize
7. Report final status with completed/remaining items and test results

Reference

Process

Step 1: LOCATE THE PLAN

Find the plan to execute. Search in order:

1. If user provided a task description with the command, use that as context to find the relevant plan
2. Check docs/plans/ for the most recent .md file (sorted by date prefix)
3. Check .claude/plans/ for any recent plan files
4. If no plan found, tell the user: "No plan found. Please create one first using /brainstorming or /writing-plans."

Once found:

• Read the plan file completely
• Present a brief summary to the user
• Ask: "Ready to hand off to Codex CLI?"
• Wait for confirmation before proceeding

Step 1b: DETECT PHASES

After reading the plan, scan for H2 headings matching: ## Phase N:, ## Stage N:, ## Part N:, or numbered sections like ## 1. Backend.

• Phases found — Report: "Detected {N} phases. Will execute sequentially." Use phased flow.
• No phases — Single-pass execution (Steps 2-6 as normal). No changes for simple plans.
• --phase N — Optional. Re-run only phase N.

Step 2: BUILD THE CODEX PROMPT

See prompt-templates.md. Use the "Initial Execution Prompt" for single-pass, or "Phase-Scoped Execution Prompt" for phased mode (current phase only + completed phase summaries).

Step 3: EXECUTE CODEX

Parse optional arguments:

• --max-iterations N — max loop iterations (default: 5, per-phase in phased mode)
• --model MODEL — pass to codex as -m MODEL
• --phase N — execute only this phase (phased mode only)

codex exec --full-auto -s workspace-write [-m MODEL] < /tmp/codex-handoff-{timestamp}.md

Let the command run to completion. Capture stdout and exit code. Report: "Codex iteration {N} complete. Reviewing changes..."

Phased execution flow: For each phase (or single phase if --phase N): build phase-scoped prompt → run Codex → review → correction loop (up to max-iterations) → phase passes: record summary, advance → phase fails at max: ask user to continue or stop.

Step 4: REVIEW THE RESULTS

See review-process.md for the review checklist, scorecard, and decision matrix. In phased mode, the scorecard is scoped to current phase items only.

Step 5: DECIDE — LOOP OR COMPLETE

• All items DONE + tests pass: Move to Step 6 (or next phase).
• Items remain AND iterations < max: Build correction prompt and re-run Codex.
• Max iterations reached: Move to Step 6. In phased mode, ask user to continue or stop.

Step 6: FINAL REPORT

See report format in review-process.md. In phased mode, report per-phase results then aggregate. If items remain, suggest --phase N to retry.

Key Principles

1. Never modify code yourself — Your job is to supervise, not code. Codex does the coding.
2. Be a strict judge — Don't pass items as "done" unless they genuinely are.
3. Correction prompts are specific — Tell Codex exactly what's wrong and what to fix.
4. Respect the plan — Don't add or remove plan items. Execute what was planned.
5. Keep the user informed — Report status after each iteration and phase transition.