name: claude-code-audit-plugin description: Audit a Claude Code plugin against best practices and suggest improvements allowed-tools: [Read, Write, Edit, Bash, Glob, Grep, AskUserQuestion]

Audit Plugin

Apply Claude Code plugin best practices from our documentation to audit a plugin repo.

See docs/claude-code-best-practices.md for the full best practices guide.

Context

!ls .claude-plugin/plugin.json 2>/dev/null && echo "Plugin detected in current directory" || echo "No plugin.json found - specify path"

Best Practices Checklist

1. Plugin Structure

plugin/
├─ .claude-plugin/
│  └─ plugin.json          # Manifest ONLY here
├─ commands/               # At root, NOT in .claude-plugin
├─ skills/
├─ hooks/
├─ agents/
└─ templates/

Check:

• plugin.json is in .claude-plugin/ (not root)
• Commands/skills/hooks are at plugin root, not nested in .claude-plugin/
• No files outside plugin root referenced via ../

2. Plugin Manifest (plugin.json)

Required fields:

{
  "name": "plugin-name",
  "version": "1.0.0",
  "description": "...",
  "author": { "name": "...", "email": "..." },
  "repository": "...",
  "license": "MIT"
}

Check:

• All required fields present
• Version follows semver
• Description is meaningful (not placeholder)

3. Commands

Frontmatter requirements:

---
description: Clear, actionable description
allowed-tools: Explicit tool list
argument-hint: [optional] Arguments if any
---

Check:

• All commands have description
• allowed-tools are explicit (not *)
• No dangerous tool patterns (Bash(*))

4. Skills

Structure:

skills/
└─ skill-name/
   ├─ SKILL.md              # Main skill file
   └─ references/           # Optional deep-dive docs

Check:

• SKILL.md exists for each skill
• Progressive disclosure (SKILL.md lean, references for detail)
• Description in frontmatter matches purpose
• No ! backtick patterns in docs (accidental execution)

5. Hooks

Registration (hooks/hooks.json):

{
  "hooks": {
    "PreToolUse": [{ "matcher": "...", "hooks": [...] }]
  }
}

Check:

• hooks.json has valid structure
• Scripts use ${CLAUDE_PLUGIN_ROOT} for paths
• Scripts are executable (chmod +x)
• Exit codes are correct (0=allow, 2=block)
• No infinite loops in Stop hooks (check stop_hook_active)

6. Token Efficiency

Check:

• SKILL.md files are lean (< 200 lines ideally)
• Large content in references/, not main files
• No large inline code dumps
• Uses file pointers instead of inline content where possible

7. Security

Check:

• No secrets in committed files
• No --no-verify patterns
• allowed-tools are scoped narrowly
• PreToolUse hooks for dangerous command blocking

Workflow

Phase 1: Locate Plugin

If path argument provided, use it. Otherwise check current directory.

PLUGIN_PATH="${1:-.}"
if [ ! -f "$PLUGIN_PATH/.claude-plugin/plugin.json" ]; then
  echo "No plugin.json found at $PLUGIN_PATH/.claude-plugin/"
  exit 1
fi

Phase 2: Run Audits

For each checklist item:

1. Run check
2. Record pass/fail/warning
3. Note specific issues

Phase 3: Generate Report

## Plugin Audit Report

**Plugin**: example-plugin v1.0.0
**Audited**: 2026-01-16

### Summary
- Checks: 20
- Passed: 15
- Warnings: 3
- Failed: 2

### Issues

#### Failed
1. **Commands missing allowed-tools**
   - `commands/deploy.md` - no allowed-tools frontmatter
   - Fix: Add `allowed-tools: Bash(deploy:*), Read`

2. **Hook script not executable**
   - `hooks/validate.sh` missing execute permission
   - Fix: `chmod +x hooks/validate.sh`

#### Warnings
1. **Large SKILL.md** (312 lines)
   - `skills/complex-skill/SKILL.md`
   - Suggestion: Move detail to references/

### Passed
- Plugin structure correct
- Manifest complete
- No secrets detected
- ...

Phase 4: Apply Fixes (if --fix)

If --fix flag provided:

1. Fix executable permissions
2. Add missing frontmatter fields
3. Restructure if needed

Report changes made.

Historical Optimizations

Search bluera-base git history for optimizations:

git log --oneline --all --grep="optimize" --grep="perf" --grep="token"

Key optimizations from our history:

• Pointer-based data passing (80-90% token savings)
• Progressive skill disclosure
• State file scoping per session
• Rate-limited type checking in hooks

Constraints

• Never modify without --fix flag
• Preserve existing functionality
• Suggest, don't force structural changes
• Report security issues prominently

Related Skills

When to consult claude-code-guide:

• Uncertain whether a pattern is best practice
• Need clarification on hook exit codes or behavior
• Plugin uses features not covered in checklist