I agree with this and matches my thinking (however - the fact that this is even an option is an incredible breakthrough that I did not expect to happen preLLMs

I agree with this and matches my thinking (however - the fact that this is even an option is an incredible breakthrough that I did not expect to happen preLLMs - potentially in decades). I have basically given up on "in IDE" AI for now. I simply have a web interface on my 2nd monitor of whatever the "best" LLM (currently Gemini, was Claude) is and copy paste snippets of code back and forth or ask questions. This way I have to physically copy and paste everything back - or just redo it "my way", which seems to be enough of an overhead that I have to mentally understand it. When it's in an IDE and I just have to tick accept I just end up getting over eager with it, over accepting things and then wishing I hadn't, and spend more time reverting stuff when the penny drops later this is actually a bad way to do things. It's really more a UX/psychology problem at the moment. I don't think the 'git diff' view of suggested changes is the right one that many IDEs use for it - I'm psychologically conditioned to review these like pull requests and it seems my critique of these is just not aligned to critiquing LLM code. 99% of PR reviews I do are finding edge cases and clarifying things. They're not looking at very plausible yet subtly completely wrong code (in the most part). To give a more concrete example; if someone is doing something incredibly wrong in 'human PRs' they will tend to name the variables wrong because they clearly don't understand the concept, at which point the red flag goes off in my head. In LLM PRs the variables are named often perfectly - but just don't do what they say they will. This means my 'red flag' doesn't fire as quickly.