src/api.rs

This file is the border between loopctl and every model provider you point it at. It is object-safe, needs two methods, and its options methods fail loudly by default.

Key items

Your ApiClient requires model(), stream_messages(&StreamRequest), and create_message(&StreamRequest).

Provided methods carry defaults: set_model starts false, base_url starts empty, and stream_messages_with_options plus create_message_with_options reject non-empty options.

extract_structured defaults to tool-call input, else lenient JSON from text. supports_tool_constraints defaults to false, which gates the default strict-tool wiring.

model_context_window defaults to None, and you disclose the real window here. The OpenAI-compatible client answers from llama.cpp /props and Ollama /api/show (0.4.0).

StreamRequest { messages, system, tools } comes with new, with_system, and with_tools.

NonStreamingResponse { message, stop_reason, usage } is what a non-streaming call returns, and the aliases BoxedApiClient and SharedApiClient save you typing.

Behavior notes

Your streams must be 'static, so clone out of &self and never borrow it. create_message futures may borrow, carrying + '_.

The fail-loud options defaults are deliberate. Silently dropping your model override or a constraint would corrupt behavior invisibly.

base_url() keys your rate limiter, so custom clients that do not override it all share one bucket.

Deep dive: API client.