Actually, the unified diff format or context diff format for patch are somewhat independent of the line numbers. You can often apply a patch even on a modified file, because it uses the context lines to find the right place.
Oh, thanks for the insights. That changes a lot of things. I reviewed my tests manually and didn’t try to patch any diff where the lines/columns didn’t seem correct, so I never knew it could still work!
I think I’m going to give this another round next week. Thanks for the links.
@qrdl Of course that would work in this one case, but this is not widely applicable. We want the LLM to do a little more here than being able to add code, hide or override members. You wouldn’t subclass something to fix an error or to refactor code, and you wouldn’t remove code, add documentation with subclassing. And then besides code we have lots of configuration files, json, xml, yaml and etc, and nowadays every language/framework has at least one of those at the project’s root.