Tool Use And Function Calling

Agent declares tools for /v2/chat but writes vague, parameterless function descriptions, so the model cannot decide when or how to call them.

Declare each tool with a clear name, a description of when to use it, and a typed parameter schema (names, types, required flags). The model routes on the description and constructs arguments from the parameter schema — both must be specific.

A /v2/chat response has finish_reason indicating a tool call and an assistant message carrying tool_calls, but the agent only checks for message text.

When the model requests tools, read the assistant message's tool_calls (each with a call id, function name, and arguments), execute them, and continue the conversation. Do not treat a tool-call turn as an empty/text answer.

The model emits two tool_calls in one turn; the agent returns a single merged tool result message with no per-call id.

Return one tool result message per tool call, each carrying the tool_call_id that matches the originating call. Pairing is by id, not by order — never merge multiple tool results into one untagged message.