When using a model in thinking mode and expanding thinking tokens, it's impossible to visually tell where thinking tokens end and the answer ends.
Having a divider or a "well" for the thinking tokens would be optimal. I have styles for this in my browser:
