effect-uai tool-call-approval

Some tool calls need a human verdict before they run. The model proposes them; your application gates execution. The wire-protocol invariant — every function_call must have a matching function_call_output — is enforced via synthesized "denied" / "cancelled" outputs.

Reach for this when the user says any of:

• "I want approval for sensitive tool calls"
• "Don't let the model send_email / delete_user without a human"
• "Show me human-in-the-loop tool gating"

Two transport flavors