Codex Toolkit

English | 繁體中文

Hand work to the local Codex CLI from inside the coding agent you already use (Claude Code, Codex CLI, or anything that reads agent skills): ask it for a second opinion, have it review or change code, generate images, review screenshots.

This is a community toolkit. It is not affiliated with OpenAI or Anthropic and is not an official package.

The four skills

skill	When it is used	What you get
codex	You want a second opinion, a review of uncommitted changes, or a clearly specified change done by Codex	Codex's analysis, review findings, or diff. Your agent checks them before reporting back
codex-assets	You need illustrations, backgrounds, mockups, icons, a set of assets in one style, or edits to an existing image	Files in a staging directory, manifest.json (dimensions, size, checksum computed from the files), preview.html (light/dark/checker toggle for transparency checks)
codex-visual-review	You have UI screenshots to check, or want a reference design compared with the implemented screen	report.md: per-image findings with region, impact, and suggestion, plus what a screenshot cannot verify
codex-setup	First install, or Codex is not working	A check-by-check readiness report (CLI, login, built-in image skill, jq/python3, ...). Report only; it never edits your config

Division of labor: code goes to codex, making images to codex-assets, looking at images to codex-visual-review, environment problems to codex-setup. When a screenshot review calls for a code change, it goes back to codex in implement mode. When an asset needs changes, codex-assets runs in revise mode and produces a new version; the original stays untouched.

Requirements

• Codex CLI (npm install -g @openai/codex), logged in with codex login. Developed against codex-cli 0.153.4.
• jq (needed by codex-assets); sha256sum or shasum.
• python3: reads image width, height, and alpha. Optional; without it those fields are null.
• Image generation uses Codex's built-in imagegen skill (ships with the CLI, under ~/.codex/skills/.system/imagegen). No separate API key.

After installing, let the agent run the codex-setup doctor once.

Install

All four skills live under skills/. The Claude Code plugin and the Codex plugin read the same directory.

Claude Code

claude plugin marketplace add davidleitw/codex-toolkit
claude plugin install codex-toolkit@codex-toolkit

From a local clone, replace the first line with claude plugin marketplace add /path/to/codex-toolkit.

Remove:

claude plugin uninstall codex-toolkit
claude plugin marketplace remove codex-toolkit

Codex

codex plugin marketplace add davidleitw/codex-toolkit
codex plugin add codex-toolkit@codex-toolkit

From a local clone, replace the first line with codex plugin marketplace add /path/to/codex-toolkit.

Remove:

codex plugin remove codex-toolkit@codex-toolkit
codex plugin marketplace remove codex-toolkit

Without a plugin: install skills directly

Each skill directory carries its own script, so you can install only the ones you want:

git clone https://github.com/davidleitw/codex-toolkit.git
# Claude Code
ln -s "$PWD/codex-toolkit/skills/codex" ~/.claude/skills/codex
# Codex
ln -s "$PWD/codex-toolkit/skills/codex" ~/.codex/skills/codex

Other tools that read agent skills usually look in ~/.agents/skills/; symlink there the same way.

Usage

Plain language is enough; the agent picks the skill:

• "Discuss with Codex which layer this retry logic belongs in. Don't change code yet."
• "Have Codex review my uncommitted changes, focus on error handling."
• "This rename is settled. Hand it to Codex and run the tests afterwards."
• "Use Codex to make a README cover image, Japanese woodblock style, with a terminal in it."
• "Here is a screenshot of the settings page. Ask Codex to check the layout."
• "Codex doesn't seem to work. Check my setup."

Model and speed policy

Every script passes the model, reasoning effort, and service tier explicitly on every run. Nothing is inherited from ~/.codex/config.toml. This is deliberate: even if your global config names GPT-6 or fast, neither is used unless you ask.

Work	Model / effort
Discussion, review of any size, debugging, code tracing, anything needing design judgment	gpt-5.6-sol / high
Lightweight changes with settled decisions and clear scope (including mechanical edits)	gpt-5.6-luna / max
Image generation, screenshot review	gpt-5.6-sol / high

• gpt-5.6-terra is never used; the scripts reject it.
• gpt-6-astra only when you explicitly ask for GPT-6 / Astra. A hard task, a failed run, or a config file naming it is not authorization.
• Fast mode (Codex service_tier="fast") is a separate explicit opt-in. Asking for GPT-6 does not imply fast, and the reverse.
• Settings you authorized earlier in the same conversation carry forward; you are not asked again.
• After the first delegated run, the agent reminds you once: "If you want later Codex work to go faster, say 'use fast mode from now on'; it may use more quota." Once only, non-blocking, and no job is re-sent to switch speed.

Where results go

skill	Location
codex	Printed in the agent's command output. The last line is CODEX_DELEGATION_OK ... or CODEX_DELEGATION_FAILED kind=...; the agent judges by that line, not the exit code
codex-assets	$TMPDIR/codex-toolkit/assets/job-XXXXXX/ (/tmp when TMPDIR is unset): the files, manifest.json, preview.html, request.txt, codex-stderr.log
codex-visual-review	$TMPDIR/codex-toolkit/visual-review/job-XXXXXX/report.md
codex-setup	Printed in the command output

Assets never land in your repository on their own. The agent reads the manifest and looks at the images first, then copies accepted files into the project's source asset directory.

Status

Done: SKILL.md and script for each of the four skills, plugin manifests for Claude Code and Codex, this README, the header image.

Checked: script syntax and argument error paths; claude plugin validate and the Codex plugin validator both pass; a local codex plugin marketplace add + codex plugin add installs and removes cleanly; codex-setup doctor ran; the header image was produced by the codex-assets script itself.

Not done: no automated tests; the four skills have not been exercised end to end on real tasks; the GitHub install commands (marketplace add davidleitw/codex-toolkit) could not be tried before the push.

Planned, intentionally left out of this version: job records and session resume, automated browser screenshots, an MCP interface, auto-fix in codex-setup.