Codex Delegate
codex-delegate is a Claude-oriented skill for using Codex CLI as an execution specialist for implementation-heavy coding work while keeping planning, review, and acceptance in Claude.
π Part of the agentic AI learning roadmap β a 7-stage curated path for building agentic AI, multilingual (zh-TW Β· zh-CN Β· English). This skill is referenced in Β§13 (Multi-LLM Delegation).
Positioning
This skill is for tasks that are expensive in tokens but cheap in judgment:
- multi-file implementation
- mechanical refactors
- boilerplate generation
- test scaffolding
- large batch edits
It is not meant for architecture, root-cause debugging, security review, or ambiguous product decisions.
What Changed In This Version
- clearer routing boundary between Claude, Codex, and Gemini
- explicit supervisor acceptance gate
- machine-readable wrapper output via
<log>.result.json - regression tests for bash and PowerShell wrappers
Core Pattern
- Claude writes a task file describing scope and constraints.
- Claude launches Codex synchronously through the wrapper.
- The wrapper emits sentinel files plus
result.json. - Claude reviews the diff and runs verification before accepting the result.
Wrapper success is not final acceptance. Claude still owns the judgment.
Relation to openai/codex-plugin-cc
OpenAI ships an official Codex integration for Claude Code,
openai/codex-plugin-cc. It is a
capable, broker-based plugin β and a different design point from this skill.
The two are complementary; the table below is meant to help you pick, not to
rank them.
| Aspect | codex-delegate (this repo) |
openai/codex-plugin-cc |
|---|---|---|
| Form | A single Claude Code skill | A multi-command plugin suite |
| Execution model | Thin synchronous wrapper: run β write result.json β exit |
Persistent broker process with background jobs |
| Job tracking | None by design β one run, one result | /codex:status, /codex:result, /codex:cancel |
| Invocation | Claude invokes the skill; the wrapper script runs Codex | Slash commands (/codex:review, /codex:rescue, β¦) plus a proactive subagent |
| Review gate | Claude's own acceptance gate (skills/codex-delegate/references/review-checklist.md) |
Optional Stop-hook review gate |
| Platform | bash + PowerShell wrappers, Windows-tested, no Node runtime |
Node.js 18.18+ runtime |
| Delegate routing | Three-way Claude / Codex / Gemini routing table | Codex-focused |
| Maintainer Β· License | Wenyu Chiou Β· MIT | OpenAI Β· Apache-2.0 |
In short: reach for codex-plugin-cc when you want background async jobs, a
slash-command UX, and an OpenAI-maintained integration. Reach for
codex-delegate when you want a thin, synchronous, supervisor-gated skill that
keeps acceptance in Claude, behaves the same on Windows and Linux, and routes
across Claude / Codex / Gemini.
codex-delegate also borrows from the official plugin: the prompt-engineering
reference (skills/codex-delegate/references/codex-prompt-blocks.md) is adapted
from its gpt-5-4-prompting skill (Apache-2.0).
Repository Layout
codex-delegate/
βββ SKILL.md
βββ README.md
βββ README_zh-TW.md
βββ scripts/
β βββ run_codex.sh
β βββ run_codex.ps1
βββ tests/
β βββ test_wrappers.py
βββ references/Testing
python -m pytest -qCurrent wrapper tests cover:
- success-path
result.jsongeneration - PowerShell wrapper contract behavior
Installation
1. Install the skill via the ai-research-skills Claude Code marketplace:
claude plugin marketplace add WenyuChiou/ai-research-skills
claude plugin install codex-delegate@ai-research-skillsDefault scope is user (this OS account, all projects). Add
--scope project to install only for the current project.
2. Make sure Codex CLI is on $PATH:
npm install -g @openai/codex
codex --versionLicense
MIT