The human-in-the-loop gate for destructive tools is the part most teams skip, and your PR example shows why review alone doesn't catch it. One pattern that pairs well with approval gates: log the resolved model + endpoint alongside every tool call. When an agent misbehaves you want to know not just what it did but exactly which model produced the decision — reproducibility matters as much as permissioning here.
Seven
Glad it landed — the provenance framing is exactly the runtime half. One nuance if you add it to the README: log the resolved model id from the response (what the provider says served the call), not just the requested one. On routers and gateways the two can differ, and the response field is the only one that's true after a fallback.