Lint

plugin-lint

Orchestrated quality audit for Claude Code plugins. Runs static correctness checks first (the unique value), then delegates to existing tools for depth.

Flags:

• (default): Full lint — static analysis + skill-creator description optimization
• --static-only: Skip description optimization, faster pass
• --fix: Auto-apply all HIGH severity fixes without prompting

Phase 1: Discover Target

Parse $ARGUMENTS. Extract path and flags.

Target resolution:

• No path → use current working directory
• Path ends in SKILL.md → single skill file
• Path contains hooks/ or ends in hooks.json → hooks audit only
• Path contains .claude-plugin/ or has plugin.json → whole plugin
• Otherwise → try as plugin root, skill dir, or hooks dir in that order

Enumerate components (silent):

find "$TARGET" -name "SKILL.md" 2>/dev/null
find "$TARGET" -name "hooks.json" 2>/dev/null
find "$TARGET" -name "*.sh" -path "*/hooks/*" 2>/dev/null
find "$TARGET" -name "*.md" -path "*/agents/*" 2>/dev/null
find "$TARGET" -name "plugin.json" 2>/dev/null

Store discovered files. Note counts: N skills, M hooks, P agents.

Phase 2: Static Correctness Audit

Read all three reference files before starting:

• ${CLAUDE_SKILL_DIR}/references/hook-schemas.md
• ${CLAUDE_SKILL_DIR}/references/frontmatter-fields.md
• ${CLAUDE_SKILL_DIR}/references/common-mistakes.md

Collect findings as a list. Each finding: {severity, file, line, issue, fix}.

2a. Audit hooks.json files

For each hooks.json found:

if condition syntax:

grep -n '"if"' "$HOOKS_JSON"

Flag any value containing ==, ===, ||, &&, != — these are JS expressions, not permission-rule syntax. Correct form: "Bash", "Bash(git *)", "Write".

hookSpecificOutput event names:

grep -n '"hookEventName"' "$HOOKS_JSON"

Cross-reference against the valid-events table in hook-schemas.md. Flag "PreCompact" — it has no schema entry. Flag any unknown event name.

SessionStart matcher:

grep -n '"matcher"' "$HOOKS_JSON"

If the hook event is SessionStart and the matcher doesn't include clear, flag as MEDIUM.

updatedMCPToolOutput on built-in tools: For each PostToolUse hook, check if the hook script references updatedMCPToolOutput. If so, also check the if condition — if it targets a built-in tool (Bash, Read, Write, Edit, Grep, Glob), flag HIGH.

2b. Audit hook shell scripts

For each .sh in hooks/:

Wrong input field name:

grep -n 'tool_output' "$SCRIPT"

Flag any tool_output reference — correct field is tool_response.

JSON PreCompact output:

grep -n 'hookEventName.*PreCompact' "$SCRIPT"

Flag — PreCompact must use plain stdout, not JSON hookSpecificOutput.

Placeholder scripts: If script has fewer than 5 non-comment lines and only outputs suppressOutput: true, flag as LOW (dead weight — spawns process for nothing).

2c. Audit SKILL.md files

For each SKILL.md found:

Extract frontmatter (between first and second ---):

awk 'BEGIN{f=0} /^---/{f++; next} f==1{print}' "$SKILL"

Unknown frontmatter keys: Compare all keys against the valid list in frontmatter-fields.md. Flag anything not in the list as MEDIUM.

Nonexistent fields: Flag when_not_to_use as HIGH (not in parser, dead metadata).

Missing when_to_use: If skill has a description but no when_to_use, flag as LOW (skill may not appear in model listing).

No-op paths glob: If paths is set and includes **/* or **, flag as LOW.

Body length:

wc -l < "$SKILL"

Warn at 400 lines, flag at 600 lines.

Body corruption — frontmatter keys leaking into body:

# Count when_to_use occurrences total
grep -c '^when_to_use:' "$SKILL" 2>/dev/null || echo 0

If count > 1: lines outside the frontmatter fence. Flag HIGH — indicates find-and-replace corruption.

AskUserQuestion block integrity:

grep -n 'AskUserQuestion\|options:\|- label:' "$SKILL"

If AskUserQuestion block exists, verify each - label: is immediately followed by valid indented content (not a stray when_to_use: line).

2d. Audit agent .md files

For each agent .md in agents/:

Extract frontmatter, check valid fields.

effort: low with code tools: If effort: low and tools: includes Write, Edit, or Bash → flag MEDIUM.

Model quality mismatch: If model: haiku and tools include Write/Edit/Bash → flag MEDIUM.

Phase 3: CC Source Verification (optional deep check)

Check if source is available:

ls ~/projects/claude-code/src/types/hooks.ts 2>/dev/null && echo "available"

If available, cross-reference:

grep -n 'z.literal' ~/projects/claude-code/src/types/hooks.ts | grep -v '//'

Compare against hookSpecificOutput event names found in target's hooks. Cite actual source line numbers in findings. Upgrade confidence of all schema findings from "reference says" to "source-verified at line N".

If not available, note "source not present — findings based on reference docs only."

Phase 4: Description Quality Delegation

Skip if --static-only flag is set.

For each SKILL.md with a non-empty description:

Invoke skill-creator's description optimization. Spawn as background Agent tasks (parallel if multiple skills):

Agent task prompt:
"Use the skill-creator skill to run description optimization on this skill.
Skill path: <path>
Current description: <description>
Run the description optimization loop (not full evals — just the trigger optimization).
Find the skill-creator scripts at: ~/.claude/plugins/cache/claude-plugins-official/skill-creator/unknown/skills/skill-creator/
Report: original score, best score, and the improved description if score improved by >0.05."

Collect results as each task completes. Note: skill-creator needs the claude CLI available for run_loop.py.

If claude CLI is not found or skill-creator scripts are absent, skip this phase and note why.

Phase 5: Synthesize and Report

Present unified report:

## plugin-lint Report: <target path>
Checked: <N> skills, <M> hooks.json, <P> hook scripts, <Q> agents
Source verification: [available / not available]

### 🔴 HIGH — Must Fix (<count>)
<file>:<line>  <issue>
  → Fix: <specific correction>

### 🟡 MEDIUM — Should Fix (<count>)
...

### 🟢 LOW — Consider (<count>)
...

### ✅ Clean
<files with no issues>

### Description Optimization Results
<skill-name>: <old-score> → <new-score>
  New description: "<improved text>"
<skill-name>: <score> — no improvement found
[skipped — --static-only]

If zero findings and no description improvements: "All checks passed. No issues found."

Then — unless --fix was passed:

AskUserQuestion:
  question: "Found <N> issues. How would you like to proceed?"
  header: "plugin-lint"
  multiSelect: false
  options:
    - label: "Apply all HIGH fixes"
      description: "Auto-fix the <count> HIGH severity issues"
    - label: "Walk me through each issue"
      description: "Review and approve fixes one by one"
    - label: "Apply description improvements"
      description: "Update descriptions with optimized versions"
    - label: "Just the report"
      description: "No changes — report only"

If --fix: apply all HIGH fixes automatically, then show report.

Applying Fixes

For each HIGH finding, apply the specific correction:

• Bad if syntax → rewrite the "if" value in hooks.json
• when_not_to_use frontmatter → remove the line
• Body corruption (extra when_to_use: lines) → remove all occurrences after the frontmatter close
• Wrong tool_output field → replace with tool_response in the script
• PreCompact JSON → rewrite script to output plain text

After applying fixes, re-run Phase 2 on modified files to confirm clean.

Reference Files

Load on demand — do not load all at once:

• references/hook-schemas.md — hook event schemas, if syntax, input fields
• references/frontmatter-fields.md — valid frontmatter keys for skills and agents
• references/common-mistakes.md — top 10 failure patterns with source citations