Do we have a plan to support thinking output in the future? Like the toolcalls and citations there is also thinking_delta and such for models that support thinking output