paiaudit

name	PAIAudit
description	PAI security audit — scans Claude Code config (settings, MCP servers, skills, agents, hooks) for risky configurations, supply-chain drift, and weekly deltas. USE WHEN user mentions audit PAI, security audit, PAI security check, supply chain audit, skill audit, MCP audit, OR wants weekly security report. Inspired by HarmonicSecurity/claudit-sec but built for Claude Code (not Claude Desktop).

PAIAudit

Read-only audit of the PAI Claude Code surface. Catches risky configs, drift in symlinked external skills, and changes week-over-week.

What It Checks

Area	Check
Settings	~/.claude/settings.json + settings.local.json — hooks, permissions, env
MCP servers	~/.claude.json — server list, command paths, token presence (count, never values)
Skill supply chain	All symlinks in ~/.claude/skills/ — target repo + commit SHA, drift since last run
Agents	~/.claude/agents/*.md — flag broad tool access (*, Bash, Write on sensitive agents)
Hooks	Scripts wired in settings — existence, permissions, owner
File permissions	~/.claude.json should be 0600 (holds API tokens)
Delta	Diff vs last audit — new skills, new MCPs, settings changes, SHA drift

Modes

Mode	Trigger	Output
Quick	audit-summary.ts	One-line summary for Telegram daily report
Full	audit.ts	Full markdown report → ~/.claude/audit-history/YYYY-MM-DD.md + Obsidian
Diff	audit.ts --diff	Compare current state to previous audit, surface only changes

Tools

Tool	Purpose
tools/audit.ts	Full audit — markdown report with all findings
tools/audit-summary.ts	Quick check — single status line, exit code reflects severity

Usage

# Full audit (saves to audit-history + prints markdown)
bun ~/.claude/skills/PAIAudit/tools/audit.ts

# Quick summary line (used by TelegramStatus)
bun ~/.claude/skills/PAIAudit/tools/audit-summary.ts

# Diff against last week's audit
bun ~/.claude/skills/PAIAudit/tools/audit.ts --diff

Severity

• CRITICAL — exposed secrets, world-writable config, broken file perms on tokens
• HIGH — unpinned external skill repos, agents with full * access, unknown MCP servers
• MEDIUM — symlink target drifted since last audit, new agents/skills added
• LOW — settings file warnings, cosmetic config issues
• INFO — counts and inventory snapshot

Supply Chain Posture

External skill repos are pinned to specific commit SHAs, not tracking floating branches. Update procedure:

1. cd <external-repo> && git fetch
2. Review the diff: git log <pinned-sha>..origin/main
3. If safe: git checkout <new-sha>
4. Re-run audit to record new SHA baseline

Pinned repos:

• ~/Dev/grc-skills (15 GRC compliance skills)
• ~/.claude/skills/claude-obsidian (10 wiki skills)
• ~/.claude/skills/gstack (29 engineering skills)
• ~/Dev/cve-mcp-server (CVE/threat-intel MCP, pinned to a78d720)

Schedule

• Daily 07:00 CT: Quick summary appended to TelegramStatus daily report
• Weekly Monday 06:00 CT: Full audit run + Telegram digest + Obsidian save

Architecture

audit.ts (orchestrator)
  ├── checks/settings.ts      → hooks, perms, env
  ├── checks/mcp.ts           → ~/.claude.json scan
  ├── checks/supply-chain.ts  → symlinks, SHAs, drift
  ├── checks/agents.ts        → tool surface flags
  ├── checks/hooks.ts         → hook script perms
  └── report.ts               → markdown + delta vs last run

For now, all in audit.ts as one file (~300 LOC). Refactor when it grows.