ModelClient

The single LLM boundary the loop depends on.

Remarks

The loop never talks to a provider directly — it only depends on this interface, so any backend (mock, OpenAI-compatible, Anthropic, raw fetch) can be plugged in by implementing stream.

Example

const echo: ModelClient = {
  async *stream(request) {
    const message: AssistantMessage = { role: Role.Assistant, content: "hi" };
    yield { type: StreamEventType.TextDelta, text: "hi" };
    yield { type: StreamEventType.Done, message };
  },
};

Methods

stream()

stream(request): ModelStream;

Defined in: model.types.ts:155

Begin a streaming completion. Implementations MUST stream by default.

Parameters

Parameter	Type	Description
`request`	`ModelRequest`	The system prompt, history, and tools for this turn.

Returns

ModelStream

An async iterable of StreamEvents for the assistant turn.

Properties

Property	Modifier	Type	Description	Defined in
`retainsReasoning?`	`readonly`	`boolean`	Whether this model retains its FULL reasoning history across turns (GLM `clear_thinking:false`, Kimi/Qwen `preserve_thinking:true`) rather than only between tool calls. When `true`, the loop keeps reasoning on plain (non-tool-call) assistant turns too — so the resent history matches what the preserve kwarg promises — instead of stripping it. Omitted/`false` ⇒ interleaved-only: reasoning is kept on tool-call turns and dropped on plain turns (the default for every non-preserving model). See prepareRequestMessages	model.types.ts:168