Now that there are several reasoning models available it seems worth developing the abstractions for models which return thoughts alongside an answer. Here are the current implementations:

• OpenAI O1: Standard chat completions API with some limitations. Reasoning is unavailable.
• Gemini Flash thinking: returns multiple text parts with part.thought=True for one.
• QwQ puts the answer in a LaTeX box (usually) and generates all the text in one undelimited string
• DeepSeek-R1 has messages with content and reasoning_content fields which give an answer as well as intermediate reasoning.

My proposal: add a ContentBlock::Thought type that contains text and return it alongside all the other content blocks and store it alongside other content blocks. For models like O1, just don't include it. For models like Gemini and R1 where we can programmatically grab the reasoning, include the blocks in the store and returned responses. For models like QwQ, don't attempt to parse the output and just return everything as text.

In the future we might adopt formats for our own trained models to "escape" reasoning from the generation. We should easily be able to parse out a <thought> ... </thought> block (for example) and do the same behavior as Gemini / R1. This sets the stage nicely also for a Chain of Thought variant type.