Permission Guard

Operation permission grading and audit system for AI agents. Four-level permissions (READ/WRITE/EXEC/DESTRUCTIVE), automatic dangerous operation interception, append-only operation log. Based on Claude Code best practices. Zero-trust design.