Hume Evi Tool Use

Mid-conversation EVI emits a tool_call event (with a tool_call_id) to look up an order status; the client executes the function.

Return a tool_response message carrying the same tool_call_id and the result payload so EVI can ground its spoken reply. Match strictly by tool_call_id; do not return an unkeyed result the server cannot correlate.

The order-lookup function returns a structured object; the client must send it back to EVI for the assistant to verbalize.

Send the tool_response content in the documented shape (e.g. a JSON-serializable result/content string keyed by tool_call_id) so the LLM can consume it. Keep the payload concise and free of secrets the assistant should not speak.

A tool call takes ~3 seconds; without handling, the assistant goes silent and the user thinks the call dropped.

Use the documented affordance (e.g. event/tool messages such as 'let me check that') to keep the user informed during tool latency, then deliver the grounded answer when the tool_response arrives. Do not let dead air imply a dropped call.