Support OpenAI spec v0.3.9+ reasoning separation #179
Description
Currently max tokens is not properly respected with SDK, and calls don't naturally flow to EOL. So if you use a thinking model that's verbose, you literally cannot get a response. Just cut off thinking blocks.