safety-permissions

---
name: safety-permissions
description: Run Agent Kit safety and permission checks before risky Claude Code operations. Use before destructive commands, broad writes, provider-layer changes, git operations, package installs, or release steps.
when_to_use: Use whenever blast radius is broad, operation touches protected resources, or permissions are uncertain.
argument-hint: "<operation scope>"
---

# Safety and Permissions

Evaluate `$ARGUMENTS` using Agent Kit safety rules.

## Required Checks

1. Identify affected files and directories.
2. Classify side effects: read-only, scoped write, broad write, shell command, git action, network action, destructive action.
3. Check protected resources:
   - `.git/`
   - `.claude/`, `.agents/`, `.cursor/`, `.codex/`, `.github/`
   - credentials, secrets, shell histories, sensitive dotfiles
   - package lock files
4. Decide one of:
   - `SAFE_TO_PROCEED`
   - `NEEDS_APPROVAL`
   - `BLOCKED`
5. Provide rollback or containment notes.

## Deny by Default

Block or ask before:

- Force push, hard reset, destructive deletion, broad chmod
- Package installation or uninstallation
- Publishing, release, deploy, or PR creation
- Commands touching protected namespaces without explicit scope

## Output

```text
[SAFETY CHECK]
Decision: SAFE_TO_PROCEED | NEEDS_APPROVAL | BLOCKED
Scope:
Risks:
Required approval:
Rollback:
Next safe action:
```