In your screenshot you have the Codex models selected. Your text is also monspaced, which is what happens when you switch the playground to Codex models. I just tried replicating your results with the prompts and parameters from your screenshot with the text model and it does not work, it just edits the text at the top like you’d expect.
Curious about your 2nd bug but I think the above might have been a false-positive. It happens.