the thing nobody mentions about long AI coding sessions

Been talking to developers about what happens when an AI coding session gets too long or dies unexpectedly. the common assumption is "just get a bigger context window," but that's not actually what people want they want a clean way to hand off what was already figured out: what changed, what failed, what's actually verified vs. guessed.

curious if this shows up outside of coding too anyone run into the same "the session dies and the reasoning goes with it" problem in other AI use cases?