Expose the model's reasoning from ChatCompletions API

[arunavos]

This is some low-hanging fruit that will improve the library's usability for interacting with vLLM models.

If we were to add a Reasoning field to the ChatCompletionChunkChoiceDelta struct definition, as shown below, then we would be able to access the returned reasoning, if any.

// A chat completion delta generated by streamed model responses.
type ChatCompletionChunkChoiceDelta struct {
	// The contents of the chunk message.
	Content   string `json:"content"       api:"nullable"`
	// The reasoning produced by the model.
	Reasoning string `json:"reasoning"     api:"nullable"`
	// Deprecated and replaced by `tool_calls`. The name and arguments of a function
	// that should be called, as generated by the model.
	//
	// Deprecated: deprecated
	FunctionCall ChatCompletionChunkChoiceDeltaFunctionCall `json:"function_call"`
	// The refusal message generated by the model.
	Refusal string `json:"refusal"       api:"nullable"`
	// The role of the author of this message.
	//
	// Any of "developer", "system", "user", "assistant", "tool".
	Role      string                                   `json:"role"`
	ToolCalls []ChatCompletionChunkChoiceDeltaToolCall `json:"tool_calls"`
	// JSON contains metadata for fields, check presence with [respjson.Field.Valid].
	JSON struct {
		Content      respjson.Field
		FunctionCall respjson.Field
		Refusal      respjson.Field
		Role         respjson.Field
		ToolCalls    respjson.Field
		ExtraFields  map[string]respjson.Field
		raw          string
	} `json:"-"`
}

I would find this useful since it would allow me to return thinking content to the caller so they can see the reasoning process and know that the model has not hung unexpectedly, and also monitor that it is answering the prompt as intended.

[arunavos]

Thanks for looking at this. I am still finding this problem a point of friction when using OpenAI's models or VLLM models through this SDK. It would be handy to have this 😄