name: audit description: Security audit of the AI agent setup - Claude Code and agy settings, permissions, hooks, MCP servers, agents, skills and instruction files - for secrets, risky permissions, download-and-run hooks, unpinned servers and prompt injection (no tokens in Claude Code; in agy run !team audit) argument-hint: "[--strict] [--json]" disable-model-invocation: true metadata: version: "1.0.0"
In Claude Code a hook answers this command without the model; you only see it when that hook did not run. Then run python3 <team skill folder>/scripts/tw.py audit $ARGUMENTS from the repository root and show its output. If the user asks for fixes, propose them one by one and change configuration only after they agree; never print a secret, and tell the user to rotate any secret the audit found.