meta-audit

When to Use This Skill

Context

• Agent definitions: !find . -path '*/.claude/agents/*' -name "*.md" -not -name "settings*"
• Settings file: !find . -path '*/.claude/agents/*' -maxdepth 3 -name "settings.local.json" -type f

Your task

1. Discovery Phase

• Use Glob to find all agent definition files in .claude/agents/
• Read each agent file to extract frontmatter and configuration
• Identify the settings.local.json for permission overrides

2. Frontmatter Validation

For each agent, verify required fields are present:

• ✅ name: Agent identifier (must match filename)
• ✅ model: an alias (opus, fable, sonnet, haiku, inherit) or a current full ID (e.g. claude-fable-5-1); a dated ID from a superseded generation (claude-opus-4-7, claude-sonnet-4-6) is a finding — recommend the alias unless the pin is deliberate
• ℹ️ effort: optional low…max (overrides the session effort while the agent runs; default inherits); recommend low where the body describes mechanical work
• ✅ color: Hex color code for UI (e.g., "#E53E3E")
• ✅ description: Clear usage guidance with "Use proactively when..."
• ✅ tools: Tool list or "All" for full access

Flag issues:

• Missing required fields
• Mismatched name vs filename
• Invalid model names
• Malformed color codes
• Keys Claude Code ignores on an agent — skill fields such as context or allowed-tools, and misspelled fields. An unrecognized key is dropped without an error, so it reads as configuration and does nothing (sub-agents.md § Supported frontmatter fields)

3. Tool Assignment Analysis

Evaluate tool assignments for security and appropriateness:

Read-only agents (should NOT have write access):

• research-documentation
• code-analysis (read only)
• code-review (read + LSP only)
• security-audit (read + LSP + Bash for scanning)

Write-enabled agents (appropriate write access):

• code-refactoring (Edit, MultiEdit)
• documentation (Write, MultiEdit)
• cicd-pipelines (Write, Edit, GitHub)

Special privileges (validate necessity):

• Bash access (security-audit, system-debugging, cicd-pipelines, test-architecture)
• GitHub access (cicd-pipelines, commit-review)
• Zen MCP access (system-debugging only)
• All tools access (should be rare, validate justification)

LSP tool assignments (language-specific):

• Verify LSP tools match agent's language focus
• security-audit should have broad LSP coverage (6+ servers)
• code-review should have multi-language LSP support
• code-analysis should have appropriate LSP access

4. Security Assessment

Check for potential security issues:

Overprivileged agents:

• Agents with "All" tools without clear justification
• Read-only agents with write/edit capabilities
• Unnecessary Bash or GitHub access
• Research agents with modification permissions

Missing restrictions:

• settings.local.json should have deny rules for destructive operations
• Allow list should be minimal and specific
• No wildcard tool access unless justified

Privilege escalation risks:

• Agents that can modify other agent configs
• Agents with both read and execute permissions
• Cross-agent permission leakage

5. Consistency Checks

Validate configuration consistency:

Naming conventions:

• Agent names use kebab-case
• Filenames match agent names
• Descriptions follow consistent format

Model assignments:

• All agents use appropriate Claude models
• No deprecated model references
• Consistent model selection strategy

Tool groupings:

• Similar agents have similar tool sets
• No duplicate or redundant tool assignments
• Clear separation of concerns

6. Report Generation

Create comprehensive audit report with:

Executive Summary:

• Total agents audited
• Critical issues found
• Security concerns
• Overall health rating

Detailed Findings:

• ❌ Critical: Missing required fields, security violations
• ⚠️ Warnings: Overprivileged agents, inconsistencies
• ℹ️ Info: Best practice suggestions, optimization opportunities

Agent-by-Agent Analysis: For each agent, report:

• Configuration completeness (✅/❌ for each required field)
• Tool assignment appropriateness (✅/⚠️/❌)
• Security assessment (safe/review/risk)
• Recommendations for improvement

Action Items:

• Immediate fixes required (with file paths and line numbers)
• Optional improvements
• Configuration validation passed/failed

7. Output Format

Structure the report as:

# Claude Agent Configuration Audit

## Executive Summary

- **Total Agents**: X
- **Configuration Issues**: Y
- **Security Concerns**: Z
- **Health Rating**: [EXCELLENT|GOOD|NEEDS_WORK|CRITICAL]

## Critical Issues

[List any critical problems requiring immediate fix]

## Agent Analysis Table

| Agent      | Model | Tools    | Color | Status    | Notes  |
| ---------- | ----- | -------- | ----- | --------- | ------ |
| agent-name | ✅/❌ | ✅/⚠️/❌ | ✅/❌ | PASS/FAIL | Issues |

## Security Assessment

[Tool privilege analysis and security concerns]

## Recommendations

[Prioritized action items with file paths]

## Detailed Findings

[Per-agent breakdown with specific issues]

8. Best Practices Reference

Tool access principles:

• Least privilege principle for tool access
• Clear separation between read-only and write-enabled agents
• Appropriate LSP tool coverage for language-specific work
• Minimal Bash access (only when required)
• Restricted GitHub access (only for git-ops and CI/CD)

Configuration standards:

• All required frontmatter fields present
• Consistent model selections across agents
• Appropriate tool permissions for agent role
• Clear descriptions with proactive usage guidance

9. Optional: Verbose Mode

If --verbose flag is provided:

• Show full frontmatter for each agent
• Display complete tool lists
• Include settings.local.json content
• Provide detailed fix commands for each issue

Notes

• This is a read-only audit - modifications require explicit user request
• Focus on configuration correctness and security implications
• Provide actionable recommendations with specific file locations
• Keep the per-agent status table (§7) as your running record while checking multiple agents; the task/todo tools (TodoWrite, TaskCreate…) are unavailable on Opus 4.8+/Sonnet 5/Fable unless opted in — see .claude/rules/agentic-permissions.md § "Task-tool availability"