setup-code-skills

Set Up Code Skills

Code ships as a full plugin (hook, subagents) and also as a vendored skill catalog. Vendoring carries the skills but not the hook or the agents, so a skills-only consumer loses the default-on entry point and the Claude-native targets that provider dispatch names. This skill installs that machinery into the target repo and, on Codex, the repo-scoped Codex config.

The machinery install is non-interactive and idempotent.

Install repo-level machinery

scripts/install-machinery.sh (in this skill's base directory) owns the idempotency and no-clobber contract. Run it rather than editing the target files by hand. It copies managed hook and agent files on every run, upserts one Code SessionStart entry while preserving unrelated hooks, and removes the legacy mandate blocks. Re-running refreshes managed files without duplicating configuration.

The installer first checks that the sibling setup-pstack skill and the patina-mode Codex tool reference are present. If any required file is missing, the install fails and tells the reader to install the full Code skill catalog.

What it materializes from assets/:

• The Code hook and its context (from assets/hooks/), plus the setup-pstack sheet decoder the hook runs, into the repo's .claude/hooks/. The context names skills by their bare names.
• The Code agents (from assets/agents/) into the repo's .claude/agents/, so subagent_type: "patina-agent" resolves, the effort agents patina-agent-<level> and effort-<level> resolve, the review agents patina-reader and patina-reader-<level> resolve, and no-comments can reach comment-sicko. patina-reader and comment-sicko set omitClaudeMd, so they start without the user, project, and local CLAUDE.md files and keep only managed policy. An installation puts the machine-specific safety rules these agents need in a managed policy file. Examples are which GitHub identity or config directory to use and a ban on pushing with the plain git command. The agents keep their bare names, so the same names resolve when Code is also installed as a plugin. Model choice per role comes from the setup-pstack sheet at runtime, not from an agent definition.
• One SessionStart entry matching startup|resume|clear|compact in .claude/settings.json; existing hooks remain in place.

Run it from the target repo:

bash "<this-skill>/scripts/install-machinery.sh"

The script defaults --repo to the current Git toplevel and the instructions file to <repo>/CLAUDE.md. Pass --repo <dir> or --instructions <file> to choose the legacy instructions file whose managed block should be removed.

Codex

The Code plugin's Codex SessionStart hook, registered in .codex-plugin/plugin.json, delivers the mandate on Codex. Codex asks the user to trust plugin hooks through /hooks before it runs them. This installer writes no mandate block to AGENTS.md, so Codex never receives the mandate twice. A skills-only Codex install carries no plugin hook, so there the user requests patina-mode explicitly. Codex has no subagent registry, so delegates route through the spawn-prompt convention in the patina-mode references/codex-tools.md (dispatch a spawn_agent told to read the patina-mode skill, or agents/comment-sicko.md, first). Subagent dispatch also needs the multi_agent feature. Add --codex to enable it:

bash "<this-skill>/scripts/install-machinery.sh" --codex

That enables multi_agent under [features] in <repo>/.codex/config.toml, idempotently. Writing the repo-scoped config rather than the user's global ~/.codex/config.toml keeps the machinery committed and shared across contributors. Codex honors a repo-scoped .codex/config.toml only for a project the user has marked trusted, so the flag takes effect after Codex's one-time trust prompt. Override the path with --codex-config <file>.

Every run, with or without --codex, removes the managed mandate block that an earlier install wrote to CLAUDE.md or AGENTS.md, and leaves all other content in those files untouched. --codex-agents <file> names a different AGENTS.md to clean. A block without its closing marker stays in place.

Skills load natively by name on Codex. The /command stubs under .codex-plugin/prompts/* come from the pstack-claude base; the installer does not write or edit them.

Configure model routing

After the installer succeeds, read its final next skill: line. Load the skill at that path and run it. Let setup-pstack own its model choices, probes, confirmation, and writes.

Confirm

Tell the user what the machinery install wrote: the session hook, the Code agents, any legacy mandate block it removed, and on Codex the multi_agent flag. Relay setup-pstack's final report. Report setup-code-skills complete only when setup-pstack succeeds. Otherwise, report it incomplete and include the reason setup-pstack reported when it stopped.