audit-plugin

Dependencies

This skill requires Python 3.8+ and standard library only. No external packages needed.

To install this skill's dependencies:

pip-compile ./requirements.in
pip install -r ./requirements.txt

See ../../requirements.txt for the dependency lockfile (currently empty — standard library only).

---

Plugin Auditor

Performs comprehensive validation of a Claude Code plugin against structure standards, naming conventions, component requirements, and security best practices.

---

Step 1: Locate the Plugin

Establish the plugin root:

• Look for ../../../../.claude-plugin/plugin.json -- this is the definitive marker
• If user didn't specify a path, check current directory and common locations
• Confirm with user if ambiguous

---

Step 2: Run Component Validation Scripts

Run the scripts bundled in this plugin (all in scripts/):

# Validate agent files
python ${CLAUDE_PLUGIN_ROOT}/scripts/validate_agent.py agents/my-agent.md

# Validate hooks.json schema
python ${CLAUDE_PLUGIN_ROOT}/scripts/validate_hook_schema.py hooks/hooks.json

# Lint hook scripts
python ${CLAUDE_PLUGIN_ROOT}/scripts/hook_linter.py hooks/

Checks performed: frontmatter structure, required fields (name/description/model/color), <example> blocks in agent descriptions, hook event names, matcher + hooks array structure.

Validation report format:

## Plugin Validation Report
### Plugin: [name] | Location: [path]
### Summary: [PASS/FAIL with stats]
### Critical Issues ([count]) -- file path + issue + fix
### Warnings ([count]) -- file path + recommendation
### Component Summary -- counts of each type
### Positive Findings
### Overall Assessment: [PASS/FAIL + reasoning]

---

Step 2b: Auto-Fix Claude Code Load Errors

Before deep validation, run the load-error fixer to catch issues that prevent Claude Code from loading the plugin at all. These are silent failures — the plugin simply doesn't load with no useful error until you run /doctor.

python ${CLAUDE_PLUGIN_ROOT}/scripts/fix_plugin_load_errors.py <plugin_root>

What it fixes automatically:

Issue	Symptom in /doctor	Root cause
plugin.json has skills/agents/hooks/commands field (any value — array, boolean true, object)	Invalid input	Validator rejects these entirely — auto-discovery handles them; "hooks": true is a common mistake
hooks.json is {} (empty object)	expected record, received undefined	Must be {"hooks": {}}
hooks.json is [] (array)	expected object received array	Must be an object
hooks.json flat format {"EventName":{...}}	expected record, received undefined	Must be nested under "hooks" key
hooks.json/lsp.json/.mcp.json has literal \n chars	Unrecognized token '\'	Python json.dump wrote escaped newlines; file must have real newlines
SKILL.md has comment lines before ---	skill fails to load	Frontmatter parser requires --- as the very first line

Correct hooks.json format:

{
  "hooks": {
    "SessionStart": [
      {
        "matcher": "",
        "hooks": [{ "type": "command", "command": "python3 ${CLAUDE_PLUGIN_ROOT}/hooks/script.py || python ${CLAUDE_PLUGIN_ROOT}/hooks/script.py" }]
      }
    ]
  }
}

Note: use python3 ... || python ... — not bare python — so hooks work on both macOS/Linux and Windows.

Empty hooks (no hooks needed):

{ "hooks": {} }

IMPORTANT — Cache coverage: Claude Code scans ALL cached versions under ~/.claude/plugins/cache/, not just the active installPath. Fixing source files and reinstalling is the only reliable fix. Use uvx plugin-add to reinstall.

---

Step 2c: Detect and Fix Symlink Stand-Ins

Git checks out symlinks as plain-text "stand-in" files when core.symlinks=false (common on Windows without Developer Mode, or when cloned without the setting). Stand-ins look like real files but contain only a relative path (e.g. ../../../scripts/execute.py). They are functionally broken — the bridge installer will copy the path string, not the actual script.

Run the bulk scanner from the link-checker plugin:

python plugins/dev-utils/scripts/bulk_symlink_fixer.py plugins/<plugin-name>

The scanner detects both:

• text-file stand-ins: small plain-text files (< 512 bytes) whose content looks like a relative path
• broken symlinks: real symlinks whose target no longer exists

Important — the fixer has a silent failure mode. symlink_manager.py always exits 0, so bulk_symlink_fixer.py prints "✓ Fixed" even when the source doesn't exist. Always verify manually:

# Confirm symlinks resolved (count should match expectations)
find plugins/<plugin-name> -type l | wc -l

# Confirm no text-file stand-ins remain for scripts paths (critical)
find plugins/<plugin-name>/skills -path "*/scripts/*" -type f ! -type l

Two categories of stand-ins:

1. Valid stand-ins (target exists) — the fixer converts these automatically. If it fails silently, convert manually:

   # Read the path out of the stand-in, unlink the file, recreate as symlink
   content = Path(standin).read_text().strip()
   Path(standin).unlink()
   Path(standin).symlink_to(content)
   

2. Wrong-path stand-ins (target missing) — common cause: an extra subdirectory in the path (e.g. references/architecture/architecture.md when the file is at references/architecture.md). Correct the relative path before creating the symlink. Check what actually exists at the plugin references/ root and recalculate the depth.

Correct symlink pattern (must match ADR manager / all standard skills):

skills/<skill>/scripts/execute.py  →  ../../../scripts/<canonical_name>.py
skills/<skill>/references/architecture.md  →  ../../../references/architecture.md

The symlink filename and the target filename may differ (e.g. execute.py → exploration_optimizer_execute.py) — that is intentional and valid.

6 known missing-source stand-ins in exploration-cycle-plugin (leave as-is until source files are created at plugins/exploration-cycle-plugin/references/): agent-loop-patterns.md, exploration-output-standards.md

---

Step 3: Run Component-Specific Scripts

Run targeted scripts for detailed checks:

Validate each agent file:

python ${CLAUDE_PLUGIN_ROOT}/scripts/validate_agent.py agents/my-agent.md

Checks: frontmatter structure, required fields (name/description/model/color), name format (3-50 chars, lowercase + hyphens), description has <example> blocks, system prompt length (minimum 20 chars, recommended 500-3,000).

Validate hooks.json schema:

python ${CLAUDE_PLUGIN_ROOT}/scripts/validate_hook_schema.py hooks/hooks.json

Checks: JSON syntax, valid event names, each hook has matcher + hooks array, hook type is command or prompt, command hooks reference existing scripts with ${CLAUDE_PLUGIN_ROOT}.

Test a hook script directly:

python ${CLAUDE_PLUGIN_ROOT}/scripts/test_hook.py \
  --hook hooks/scripts/validate.py \
  --event PreToolUse \
  --input '{"tool_name": "Write", "tool_input": {"file_path": "src/app.py"}}'

Lint hook scripts for common issues:

python ${CLAUDE_PLUGIN_ROOT}/scripts/hook_linter.py hooks/

---

Step 3b: Self-Evolution Policy Compliance Check

Every plugin must comply with the continuous self-evolution policy at .agent/rules/self-evolution-policy.md. Check each of the following:

Required files (flag WARN if missing):

# Check for self-evolution profile (required by Phase 0 before any autonomous edit)
ls plugins/<plugin>/references/self-evolution-profile.md

# Check for map-debt working queue (required by friction-driven evolution policy)
ls plugins/<plugin>/references/map-debt.md

# Check for evolution log (created by Phase 7 on first fix)
ls plugins/<plugin>/references/evolution-log.md

Evals schema (flag CRITICAL if wrong):

# All evals.json must use should_trigger boolean schema — NOT legacy expected_behavior
grep -r "expected_behavior\|expected_output\|\"expected\":" plugins/<plugin>/skills/*/evals/evals.json
# Any match = wrong schema → fix immediately

SKILL.md line count (flag WARN if over 500):

wc -l plugins/<plugin>/skills/*/SKILL.md | sort -rn | head -10

Stale skill references (flag WARN):

# os-skill-improvement is methodology-only — any SKILL.md directing agents to invoke it
# for active improvement should reference os-improvement-loop instead
grep -rn "os-skill-improvement" plugins/<plugin>/skills/*/SKILL.md

ADR-003 symlink check (flag CRITICAL if violated):

# Only file-level symlinks permitted — no directory symlinks
find plugins/<plugin> -type l | while read l; do
  if [ -d "$l" ]; then echo "DIRECTORY SYMLINK (ADR-003 violation): $l"; fi
done

---

Step 4: Manual Checks

For issues the scripts may not catch:

Plugin structure check:

# Manifest must be here (not in root)
ls .claude-plugin/plugin.json

# Components must be at root (not in .claude-plugin/)
ls commands/ agents/ skills/ hooks/

# Validate JSON
jq . .claude-plugin/plugin.json

Security scan:

# Check for hardcoded credentials
grep -rn "password\|api_key\|secret\|token" --include="*.md" --include="*.json" --include="*.sh" .

${CLAUDE_PLUGIN_ROOT} portability:

# Ensure no hardcoded paths in hook commands or MCP config
grep -rn "/Users/\|/home/" --include="*.json" --include="*.sh" .

Hook script quality (for any .py files wired as hooks):

• Does the hook command use python3 ... || python ... for cross-platform compatibility?
• Does main() have an early-exit project-type guard (e.g. if not (project_root / "context").exists(): return) so it skips silently in projects that haven't initialized the plugin?

Naming conventions:

• Plugin name: kebab-case (my-plugin, not MyPlugin or my_plugin)
• Command files: kebab-case .md
• Agent files: kebab-case .md describing role
• Skill directories: kebab-case
• Script files: kebab-case with extension (.sh, .py, .js)

Skill quality (run skill-reviewer for each skill):

"Review my skill at skills/skill-name/SKILL.md"

---

Step 4b: Marketplace Source Path Audit

If the repo contains a .claude-plugin/marketplace.json, run this check to ensure every plugin entry points to a directory that actually exists. Missing directories silently fail during /plugin install with no helpful error message.

python3 ${CLAUDE_PLUGIN_ROOT}/scripts/audit_marketplace_sources.py <repo_root>
# or from the repo root:
python3 ${CLAUDE_PLUGIN_ROOT}/scripts/audit_marketplace_sources.py .

Common causes of missing paths:

• Plugin directory renamed but marketplace.json not updated (e.g. obsidian-integration → obsidian-wiki-engine)
• Plugin entry added to marketplace before the plugin directory was created
• Plugin directory deleted but marketplace entry not removed

Fix: Either update the source path to match the real directory name, or remove the entry.

---

Step 5: Report and Remediate

Severity levels:

• Critical -- plugin won't work or is insecure. Fix immediately. (e.g., invalid JSON, hardcoded credentials, missing required fields)
• Warning -- degrades quality or usability. Fix before distribution. (e.g., missing README, vague skill descriptions, no <example> blocks in agents)
• Minor -- best practice improvement. Fix when convenient.

Fix critical issues first, then re-validate:

# Re-run validation after fixes
"Validate my plugin at <path>"

Keep running until: 0 critical issues, warnings addressed or documented.

---

---

Standards & References

• Architectural Decision Records (ADRs) located at references/*.md. Always consult them (especially ADR 001-006) to verify if the plugin follows our standards for shared scripts, cross-plugin dependencies, symlinking patterns, and loose coupling. A plugin that violates these ADRs (e.g. duplicates shared scripts instead of symlinking) is considered structurally non-compliant.

Standards Reference

.claude-plugin/plugin.json minimal valid:

{ "name": "plugin-name" }

.claude-plugin/plugin.json recommended:

{
  "name": "plugin-name",
  "version": "0.1.0",
  "description": "What the plugin does",
  "author": { "name": "Author Name", "email": "email" }
}

Agent description pattern (must have <example> blocks):

description: |
  Use this agent when user asks to "do X", "run Y", or mentions Z.
  
  <example>
  Context: user just finished creating a plugin
  user: "I've set up my plugin"
  assistant: "Let me validate the structure."
  </example>

Skill description pattern (third-person, anti-undertrigger):

description: >
  This skill should be used when the user asks to "X", "Y", or "Z".
  Use this skill even when the user doesn't explicitly say "Z" -- 
  mentions of [related concept] should also trigger this.

---

Next Actions

• Fix gaps: Run create-skill, create-command, or create-hook to add missing components
• Improve skills: Run skill-reviewer on each skill for trigger optimization
• Upgrade to L5: Run audit-plugin-l5 for advanced red-team structural audit
• Distribute: Push to GitHub — users install via plugin_add.py richfrem/agent-plugins-skills