At at least there it's more honest than GPT, although at work especially it loves to decide not to use the built in tools and instead YOLO on the terminal but doesn't realize it's in powershell not a true nix terminal, and when it gets that right there's a 50/50 shot it can actually read the output (i.e. spirals repeatedly trying to run and read the output).

I have had some success with prompting along the lines of 'document unfinished items in the plan' at least...