The agentic loop
You send a message; the model reasons, calls a tool, reads the result, and continues — up to a configurable iteration cap (15 by default). Answers stream token by token over SSE, or return in one shot if you prefer. The system prompt is project-aware: working directory, a shallow file listing, loaded skills, the active permission scope, and stack hints for Node.js, Next.js, Actix Web, and PostgreSQL.