Generic support for thinking models? (e.g. Apriel-1.5-15b-Thinker)

[brechtm]

I wanted to benchmark this model using the Aider polyglot benchmark. I can run a quantized version of this model using llama-server, but one issue I'm running into is that the model's thinking output is included in the main content field instead of the reasoning_content field of the JSON returned by llama-server. This is some example output returned (in content):

Here are my reasoning steps:
This is a question about me.
[BEGIN FINAL RESPONSE]
Hi! I'm an AI assistant created by ServiceNow Language Models (SLAM) Lab to help answer your questions and assist with various tasks. How can I help you today?
[END FINAL RESPONSE]
<|end|>

Would llama.cpp need to explicitly add support for this output format, or can I somehow configure it to parse the output so that the thinking content is separated from the actual reply (and also removing [BEGIN FINAL RESPONSE] and other tags)? GBNF does not seems suited for this purpose, or is it?

Thanks!