skill-codex: a skill that forwards prompts to codex exec, and a README that points at SKILL.md for the real instructions

A skill with no code of its own

The purpose is stated in one line: enable Claude Code to invoke the Codex CLI, through codex exec and session resumes, for automated code analysis, refactoring and editing workflows. That is the whole design. There is no source to read and no build to run, since the repository lists .claude-plugin/, LICENSE, README.md, openpackage.yml and plugins/ at the top level, and GitHub records no detected primary language. The instructions themselves live one level down, in plugins/skill-codex/skills/codex/SKILL.md, which the file calls the place for complete operational instructions, CLI options and workflow guidance. Prerequisites are the three you would expect from a wrapper: the codex CLI installed and on PATH, Codex configured with valid credentials and settings, and a confirmation run of codex --version, with the instruction to resolve any errors before using the skill.

Two install routes, two transports

The recommended route goes through Claude Code's plugin system with a marketplace, so updates arrive on their own:

/plugin marketplace add skills-directory/skill-codex
/plugin install skill-codex@skill-codex

The second route extracts the skill by hand, and it does so over SSH rather than https:

git clone --depth 1 [email protected]:skills-directory/skill-codex.git /tmp/skills-temp && \
mkdir -p ~/.claude/skills && \
cp -r /tmp/skills-temp/plugins/skill-codex/skills/codex ~/.claude/skills/codex && \
rm -rf /tmp/skills-temp

The clone URL is [email protected], so a machine without SSH keys registered fails at the first line rather than at the copy, and the whole chain is joined with && so nothing after a failure runs. The copy path repeats the repository name twice, because the plugin lives at plugins/skill-codex/skills/codex inside the project and lands in a folder also called codex under the user's skills directory. Note also what the marketplace line produces: the marketplace name and the plugin name are the same string, which is why the install command ends in skill-codex@skill-codex.

Codex's stderr is discarded unless you ask for it

One default is worth knowing before you conclude that a delegation failed. The skill suppresses the thinking tokens, which arrive on stderr, by appending 2>/dev/null to the command it runs, and the stated reason is to keep Claude Code's context window from filling up. The escape hatch is not a flag. You have to ask Claude explicitly to show the thinking tokens, either for debugging or to read Codex's reasoning. That is a deliberate trade: the run stays cheap in tokens and quiet in the transcript, at the cost of one round trip to recover anything that was written to the error stream. Since the same redirection sits at the end of the example command, it covers the whole invocation rather than a single part of it.

The sample command pairs full-auto with a read-only sandbox

The documented invocation looks like this:

codex exec -m gpt-6-astra \
  --config model_reasoning_effort="high" \
  --sandbox read-only \
  --full-auto \
  --skip-git-repo-check \
  "Analyze this Claude Code skill repository comprehensively..." 2>/dev/null

Two of those flags point in opposite directions. --full-auto asks for unattended operation while --sandbox read-only restricts what may be touched, and the file prints them together without saying which wins. For analysis the stated default is read-only, so the sandbox choice is what keeps a delegated task from writing to your tree, and the combination above is what a reader would copy for an analysis run. --skip-git-repo-check is there because the delegated task is frequently pointed at a directory that is not a Git repository. The prompt is a plain string argument at the end, which is how the request travels: no file, no stdin, no wrapper.

Four model names and six reasoning levels

Before running anything, Claude is supposed to settle two choices, and the file is explicit that it skips both when your prompt already answers them. The first is the model: gpt-6-astra by default, with gpt-5.6-sol, gpt-5.6-terra and gpt-5.6-luna as the other current names and gpt-5.5 named as a legacy example. The second is reasoning effort, from a six-step list of low, medium, high, xhigh, max and ultra, qualified as subject to the model's ceiling. That qualification matters more than the list: a level you ask for is not a level you get, and nothing in the file says which of those levels each named model accepts. The third choice, sandbox mode, is not asked, because it has a default of read-only for analysis. The result comes back as a summary from Claude rather than as raw Codex output.

The README is a pointer, and openpackage.yml is unexplained

Read the structure and the division of labour is clear. .claude-plugin/ at the root is what makes the repository a marketplace, plugins/ holds the package, and inside it the skill directory is the part that gets copied into a user profile. The file itself then hands off, saying detailed instructions, CLI options and workflow guidance are in plugins/skill-codex/skills/codex/SKILL.md. That leaves openpackage.yml, a manifest at the top level that no line of the README mentions, so its purpose and its schema are not documented here. The repository has no GitHub releases, which is consistent with a repository whose installable unit is a plugin directory rather than a versioned archive, and the plugin route takes its updates from the marketplace instead of from a tag.

The first line of the README points somewhere else

The opening line is a recommendation to leave. If you want a more autonomous setup for agentic workflows, the file says, check out klaudworks/ralph-meets-rex. What follows is the plugin you are reading about, and the distinction the note draws is autonomy: this repository makes Claude ask which model, which reasoning level and then summarize the answer, while the other project is offered for workflows that run without that step. It is an honest boundary statement and it also tells you what this repository is not trying to be. The last recorded commit on the default branch is dated 2026-09-13, the project is not archived, and the licence is MIT, so redistribution of the skill folder in either of the two install routes is unencumbered.

Editorial conclusion

What this repository contributes is a policy layer, not a program: which model by default, which reasoning level, which sandbox, and the instruction to summarize rather than paste. That is worth reading even if you install nothing, because the defaults are stated. Two cautions before you install it. The sample command combines --full-auto with --sandbox read-only, so the effective permission level is not obvious from the flags alone, and the repository ships no code to check it against. And stderr is discarded by default, so a failing codex run looks like silence unless you ask for the output. The plugin route is the cleaner install since it updates itself; the standalone route needs Git with SSH access and copies a folder by hand. MIT licensed, no GitHub releases, last commit on record 2026-09-13.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. README
4. skills-directory/skill-codex on GitHub