"messages": [
{
"role": "assistant",
"content": "",
"reasoning_content": "The reasoning in the response will start with this"
}
],
When messages end with an assistant message, it should be used to prefill the generated response, which is standard behavior for self hosted llama.cpp.