Revert "gptel-rewrite: Diff buffers, not files (#731)"

[aagit]

This reverts commit 0ce917a.

The above commit while not wrong per se, prevents to use diff-mode to apply hunks, which makes it not a good tradeoff.

[karthink]

If possible, I would rather find a way to get diff-mode to apply the hunks to buffers. I'll look into it, but if you've studied this already you can let me know if this is possible.

[aagit]

That I would prefer too indeed. It didn't appear possible, but I didn't spend much time trying.

[pabl0]

Isn't this possible in ediff?

[aagit]

I'm sorry I don't find ediff a good alternative. One thing is to review code one thing is to review what changed. In this case I need to review what changed, akin to git log. When I do conflict resolution I prefer to see code not diffs as I need to review code as well not just what changed. ediff is not my favorite interface in general. diff mode is also quite nice in how it allows to split a diff and selectively apply or drop hunks.

[pabl0]

How about two separate actions: diff (buf), diff files, or something along those lines? Too confusing?

[aagit]

a diff files option sounds great with me.... it's a workaround, but it's certainly simpler than a upstream diff-mode change. It should be simple to drop if diff-mode is improved later.

The biggest upstream shortfall is not diff-mode but the scrolling over overlays, I really really would like the inference to keep showing the last character of the overlay in my window with auto scrolling, I don't even pretend to move the pointer inside the overlay if I could just show it automatically as it is extended by the inference.... Similarly to what you can achieve with gptel-auto-scroll. Alternatively if it's simpler it should be possible to move the pointer through the overlay with up/down pgup/pgdown. So I don't think it's worth spending time on diff-mode when there's a bigger usability problem in the core.