Reasoning Model Deepseek Reasoner

A UI renders deepseek-reasoner output by concatenating reasoning_content and content into one block shown to the end user.

Render content as the user-facing answer and treat reasoning_content (the chain-of-thought) as separate — surface it only behind an optional 'show reasoning' affordance. Do not merge the two fields into the displayed answer.

For multi-turn deepseek-reasoner chat, the client appends the full prior assistant message — including reasoning_content — into messages[] for the next request.

When constructing the next request, include only the prior assistant content (the final answer) and strip reasoning_content from the message before resending. Per docs, reasoning_content must not be passed back into the messages history.

Code shared between deepseek-chat and deepseek-reasoner passes temperature and top_p tuned for chat directly to deepseek-reasoner and treats any difference as a bug.

deepseek-reasoner does not honor the same sampling controls as deepseek-chat — certain parameters (e.g. temperature, top_p) are unsupported / ignored for the reasoner. Branch parameter construction by model and verify supported params against the reasoning-model docs rather than assuming chat parit…