Ryuu's Skills

Codex

Wire the hooks into Codex and trust them.

Codex runs the same hooks as Claude Code. Every entry runs the skills CLI through the AGENTS_DIR link. Whenever ~/.codex (or CODEX_HOME) exists, the installer writes the entries below to its hooks.json. Trust new or replaced entries once. The installer replaces both owned comment and dash shims with one skills hook post-tool-use command. If you wired the shims yourself, replace both with that command.

It also replaces the owned commit guard and reply guard shell hooks with skills hook pre-tool-use and skills hook stop. If you wired either yourself, replace it with its CLI command. Do not add the CLI command beside it.

Trust the Hooks

Codex records trust per hook definition. Open codex, run /hooks, and trust the new entries once. A later install run that adds nothing leaves the file untouched, so the trust holds.

What the Installer Writes

Every entry runs the skills CLI by absolute path under AGENTS_DIR (Configuration). With a home of /Users/you, the entries are:

{
  "hooks": {
    "SessionStart": [{ "matcher": "startup|resume|clear|compact", "hooks": [{ "type": "command", "command": "/Users/you/.agents/skills/playbook/bin/skills hook session-start" }] }],
    "PreToolUse": [{ "matcher": "^Bash$", "hooks": [{ "type": "command", "command": "/Users/you/.agents/skills/playbook/bin/skills hook pre-tool-use" }] }],
    "PostToolUse": [{ "matcher": "^(Bash|apply_patch)$", "hooks": [
      { "type": "command", "command": "/Users/you/.agents/skills/playbook/bin/skills hook post-tool-use" } ] }],
    "Stop": [{ "hooks": [{ "type": "command", "command": "/Users/you/.agents/skills/playbook/bin/skills hook stop" }] }]
  }
}
  • The PostToolUse entry matches Bash as well as apply_patch, because Codex usually applies a patch by running apply_patch through the shell, and hooks see that call as Bash. The CLI checks only shell commands that apply a patch and stays silent on every other command.
  • Missing entries are added. An entry this repository shipped for a retired hook is switched to its replacement in place and needs trusting again. A personal entry is left alone. The old comment and dash entries are the exception: they sat under ^(Edit|MultiEdit|Write)$, which Codex never fires for a shell patch, so they are dropped and the CLI entry is added in a group of its own.
  • A run that adds or replaces an entry rewrites the file with two space indent, keeping its mode and group, and copying its extended attributes where the system allows. Whether ACLs survive depends on the platform. A run with nothing to change leaves the file untouched.
  • A file that is not valid JSON is left alone, and the run prints the entries to add by hand.
  • Deleting a skills entry does not stick, since the next install adds it back. To turn the hooks off, set AGENT_HOOKS=0 (Configuration). It turns off every hook but the commit guard, which skips only its comment sweep.

What each hook does, and when it stays silent, is on the Hooks pages.

Invoking Skills

Every skill whose SKILL.md sets disable-model-invocation: true also carries agents/openai.yaml, which sets allow_implicit_invocation: false. Codex keeps those skills out of the model's skill listing, so describing what one does in a prompt no longer opens it. Invoke one as $name, such as $architect, the way Claude Code takes /architect. The installer links each skill directory whole, so the file reaches ~/.codex/skills with the rest of the skill.

Rules

With AGENT_RULES set (Deny rules), the installer also writes rules/skills.rules under ~/.codex (or CODEX_HOME). Codex reads every .rules file there at startup, so restart Codex after a run that changes it. The installer owns the whole file: its first line is # Written by skills install. Edits are replaced on the next run., and a skills.rules without that line is left alone, with the reason in the output. No other rules file is touched, default.rules included.

Each guardrail group becomes prefix_rule entries at the level you chose: deny writes forbidden and ask writes prompt. Spellings that differ in one word share a rule, so the publishing group in hands-off mode reads:

# publishing: Commit, push or open a PR
prefix_rule(pattern = ["git", ["commit", "push"]], decision = "forbidden", justification = "Commit, push or open a PR")
  • A Codex rule matches the start of a command only, so an entry with a wildcard before its end (git push * --force), an Edit path, or a skills.* key has no rule. The output names every entry it skipped. The commit guard hook still covers force pushes.
  • Allow groups are never written, since a Codex allow runs the command outside the sandbox.
  • The comment group follows the commit guard in hooks.json: once that hook is wired, prs mode writes no comment rule.
  • A change that weakens a rule follows the same terminal rule as Claude settings: an agent run leaves it and names it, and you apply it by rerunning ./install.sh in your own terminal.
  • When codex is on PATH, the new file is checked with codex execpolicy check before it replaces the old one. A file that fails the check is not written. To check it yourself: codex execpolicy check --rules ~/.codex/rules/skills.rules -- gh pr merge 7.

On this page