The copy-paste thing is interesting because it hints at a deeper issue: LLMs don't have a concept of "identity" for code blocks—they just regenerate from learned patterns. I've noticed similar vibes when agents refactor—they'll confidently rewrite a chunk and introduce subtle bugs (formatting, whitespace, comments) that copy-paste would've preserved. The "no questions" problem feels more solvable with better prompting/tooling though, like explicitly rewarding clarification in RLHF.