Intent IR — safe by design
Every LLM output is parsed into a typed schema before anything executes. No raw text ever reaches your shell. The schema encodes risk levels, required confirmations, and step dependencies — the architectural foundation for auditable, OS-grade execution.