setup

Set up sheepdog for this project

The sheepdog plugin is a generic engine: three hooks that read project-owned rules from .claude/sheepdog/config.json. The plugin ships no rules of its own. This skill scaffolds that config for the current project.

Do the following, in order. Adapt paths/skills to what the project actually has — do not blindly paste the examples.

1. Discover the project's skills and governed paths

• List the project's skills: look under .claude/skills/*/SKILL.md (and any plugin skills the user mentions). These are the skills the gate can require.
• Identify which directories should be gated (where hand-edits without a skill are a mistake). Ask the user if it isn't obvious. Common picks: source dirs with strong conventions, and test dirs.

2. Create .claude/sheepdog/config.json

One JSON file holds the whole gate: the authorization window, override skills, and the ordered glob→skill rules. The routing table injected on every prompt is GENERATED from rules, so the reminder and the gate can never drift apart.

Starter template (replace the example rule with this project's real globs and skill names; drop the fields the project doesn't need):

{
  "window": 300,
  "overrides": [],
  "rules": [
    { "glob": "*/path/to/governed/dir/*", "skill": "the-skill-for-that-dir", "desc": "optional label for the routing table" }
  ]
}

Field notes (JSON has no comments — explain choices to the user in chat instead):

• window — seconds a stamped skill stays valid (default 300; omit if default is fine).
• overrides — skills allowed to edit ANY governed path (orchestrator-style skills); omit if none.
• rules — first matching glob wins, so order specific → general. Globs use shell-case semantics: * matches any run of characters (including /), ? matches exactly one. They match against the absolute file path.
• reminder (optional) — replaces the default preamble injected on every prompt.
• Unknown keys are rejected and the gate fails closed — typos can't silently disable enforcement.