I still wonder how it should work otherwise if the "AI" is not running locally - it needs to see the codebase & because it runs somewhere remote in a cloud then it needs to copy the codebase over.
I'm rather surprised people would expect it work any different for non-local "AI".
It doesn't even swallow an entire code base so there is absolutely no need to see the entire codebase. Usually models search in the codebase locally and then use only the relevant parts. There is almost never the whole context just spammed by an entire code base.
I'm rather surprised people would expect it work any different for non-local "AI".