add-guardrails

name	add-guardrails
description	Install a self-contained development-process guardrail kit (plan-review, code-review, and verification gates) into the current repository, so any developer's Claude Code follows the same review discipline on clone. Triggers on "add guardrails", "install guardrails", "set up quality gates", "warden gates for this repo".
user_invocable	true

/add-guardrails

Installs a portable warden quality-gate kit into the current repo. After it runs and you commit the changes, every developer who clones the repo gets three review gates enforced by Claude Code hooks — with no dependency on any external checkout:

• plan-review before any Edit / Write / MultiEdit
• code-review before git commit
• verification before git commit

It ships self-contained agents (.claude/agents/), their rule books (.claude/wardens/), a small vendored gate (.claude/hooks/warden-gate.py), a process-rules doc (.claude/rules/), the hook wiring (merged into .claude/settings.json), and a short section appended to CLAUDE.md.

How to run it

1. Confirm the target. This installs into the current working directory, which must be a git repository. If the intended target is ambiguous, confirm it with the user first.

2. Locate the installer (it ships alongside this skill):

   # The second path assumes a Deus checkout at ~/deus; adjust if yours is elsewhere.
   INSTALLER=$(find ~/.claude/skills ~/deus/.claude/skills -maxdepth 2 -name add_guardrails.py 2>/dev/null | head -1)
   

   If that finds nothing, ask the user where their Deus checkout lives and look under its .claude/skills/add-guardrails/.

3. Choose opt-in modules. Use AskUserQuestion to ask which optional modules to include beyond the always-installed baseline (plan/code/verification gates, the process-rules doc, the gate script, and the hook wiring):

   • design-logs — ADR-style decision records under docs/decisions/. Good for any repo that wants a durable trail of design decisions.
   • codegraph-first — a semantic-search-before-grep exploration rule. Include only if the repo is indexed for code-intelligence tools.

   Confirm the available modules with python3 "$INSTALLER" --list-modules.

4. Preview, then apply. Always dry-run first and show the user the planned actions:

   python3 "$INSTALLER" --target "$(pwd)" --modules <chosen,comma,separated> --dry-run
   

   On approval, re-run the same command without --dry-run.

5. Review and commit. Tell the user to inspect git diff and commit the new files so the gates travel with the repo. The gates take effect in new Claude Code sessions opened in that repo.

Updating / re-running

Re-running is safe and idempotent: existing settings.json hooks and CLAUDE.md content are merged, never replaced, and an unchanged re-run is a no-op. Pass --update to refresh kit files that have drifted from the templates.

Notes

• macOS / Linux only in v1. The hook wiring uses a bash wrapper, so on Windows the gates are inert (they fail open — no blocking, no errors). Cross-platform support is a follow-up.
• Fail-open by design. If a teammate's machine has no python3, the gates allow all actions rather than blocking work.
• Distinct from /onboard. /onboard indexes a repo for code-intelligence search; this skill installs the review-process gates. They are independent — a repo can have either or both.
• Availability in other repos. This skill must be discoverable from the session you run it in. If it only appears inside the Deus repo, make it global by symlinking it into ~/.claude/skills/add-guardrails/ so /add-guardrails works from any repo.