Hey @odrobnik litellm should already be handling this - do you have a specific model called via litellm where you're seeing this behaviour?

My use case if when you don't use litellm, like in custom_example_agent.py. In particular I am referring to tracing where it looks like this:

So my feature suggestion is not to ADD like litellm does, but rather to remove it from the final_output and possibly move it into a separate field (like DeekSeek does it) or even into a separate output item, like Responses as reasoning item:

In the least I would expect the reasoning to be logged in the tracing separately, even if nothing else is changed. But final_output from the runner shouldn't contain reasoning IMHO. The philosophy of Runner is to deal with Response objects, even when they are made up from Chat Completion results. For regular Response-based running, the reasoning is not part of final_output, so for ChatCompletion-wrapped running they shouldn't be either.

The ARE useful for tracing, so that's why they could be included, but possibly in the same way that reasoning summaries are included for Response-based traces.