Safety monitor for AI coding agents: watch risky commands, stuck runs, MCP changes, and cross-agent audit workflows before they burn tokens or power.