The server injects tool schemas into the prompt (often via the chat template / structured output), then a model-specific tool-call parser extracts the function name and JSON args from the generated text — incrementally during streaming — and emits OpenAI-style tool_calls. Structured output can hard-constrain the args to the schema.

You must emit incremental deltas while maintaining correct semantics (role, finish_reason), and parse partial content (tool-call JSON, reasoning tags) that spans multiple chunks without committing to an interpretation too early.