Type Alias TLlmRequest<T>

Phantom field that carries the structured-output type T from outputSchema into the response. Always undefined at runtime.

Free-form for forward-compat. The known set is TLlmModel.

Optional callback fired by a provider as soon as the upstream response id is known — before the call resolves. Lets a caller persist the id mid-flight so an interrupted in-flight call can be recovered from the upstream's stored copy.

Only providers that can surface an id mid-flight invoke it (the OpenAI provider in either streaming mode, from the first response.created SSE event, when the call makes a single request: with function tools it makes one per round, and the id it returns is the last round's, so it leaves this uncalled). Recovery after an interruption is guaranteed only in background-stream mode, where the response keeps generating upstream after the caller disconnects. Synchronous and poll-only providers leave it uncalled and surface the id only at completion via TLlmResponse.rawResponseId. Optional: callers that don't set it are unaffected. Invoked at most once per provider call (per attempt).

Optional callback a provider invokes once per chunk of assistant output text, as it arrives, before the call resolves. Lets a caller show a reply as it is generated.

Chunks are passed as received and never accumulated. For a structured-output request they are pieces of the JSON envelope, not readable prose. The OpenAI provider invokes it in either streaming mode; it is never invoked on the poll-only background path, the synchronous path, or by the chat-completions provider. Each call to respond() streams from the beginning, so a pipeline retry streams its text again: a caller accumulating text must start over per attempt. With function tools, one call makes a request per round and every round's text is passed on, so text a model wrote before calling a tool precedes the final answer.

The provider response id to chain this request against. When set, the provider should continue from the specified prior response rather than starting a fresh conversation. Only meaningful for providers that support response chaining (e.g. the OpenAI Responses API). Synchronous providers (e.g. chat-completions) ignore this field — the caller folds the transcript into userMessage instead. Optional.