Put a typed guardrail in front of an agent tool call

Agents fail in a specific way: they emit a tool call that looks locally reasonable and is globally wrong. A second chat prompt asking "is this safe?" is slow and still unstructured.

Send the user request and the proposed call as `state`. Ask a `noul` for confirmation, a `noul` for intent match, and a `score` for blast radius.

Your policy stays in code: if risk is the top level, or intent does not match, do not execute. Jev only judges. It never runs the tool.

The playground ships a Tool-call risk gate example you can copy.