Hopefully it will be able to also reduce boilerplate and do reasonable DRY abstractions if repetition becomes too much.
E.g. I feel like it should be possible to first blast out a lot of repetitive code and then for LLM to go over all of it and abstract it reasonably, while tests are still passing.
Code generator in the editor has been around for ages and serves primarily to maximise boilerplate and minimise DRY. Expecting the opposite from a new code generator will yield disappointment.
I mean LLM can go through all the files in a source code and find repetitions that can be abstracted. Reorganize files into more appropriate structures etc. It just needs an optimal algorithm to provide optimal context for it.
E.g. I feel like it should be possible to first blast out a lot of repetitive code and then for LLM to go over all of it and abstract it reasonably, while tests are still passing.