I think this is a great idea, and I've wondered about something like this before.
I do find it sad though that the opening description has to be:
> Two agents edit different functions in the same file? Clean merge.
Why does EVERYTHING has to be geared towards agents? Humans can use this too. Why not just "two commits contain edits for different functions in the same file?"
Automated process run into race conditions more often due to their frequency. Humans can do that too, but are less likely to in practice both due to lower frequency and because they carry more awareness of global context that isn't captured in systems that aren't checking for it. The ability of your brain to read and take as context all the pull requests open in a repo that might affect your work.
> Why does EVERYTHING has to be geared towards agents?
This was also my first thought when I checked the website. I was interested in the general merge approach, and that it works with LLMs and agents fine, but that's secondary. Nowadays every product must be in-your-face AI-first somehow, often to the extent that it de-emphasizes why the product exists in the first place, its core competency and distinguishing features pushed below the fold by screaming "It supports AI" headlines. It saddens me. That something supports AI is nothing special anymore, an expected feature. Just mention it like that, in the product highlight box next to where it mentions that it supports Github or similar nothing-special features.
> Why does EVERYTHING has to be geared towards agents?
Moving forward one can expect the most amount of code to be generated by agents, so it makes sense to optimise for that use case.
(Note that i’m not saying it’s good or bad)
> the most amount of code generated by agents
Yes, I wholehartly agree. Coming from a week of agent-coding - out of pure curiosity - most code was generated by agents, which I had then to delete and rewrite to use like a quarter of statements to achieve the same, in an understandable, maintainable way.
But that's just my experience.
People will very often say they "had to" do something, when in reality they merely "chose to" do that thing.
> so it makes sense to optimise for that use case.
How do the agent and human use cases meaningfully differ here, though?
I'm pretty sure GP's complaint is about the prose description, rather than the actual functionality.
I suspect it's much worse with agents. They tend to generate a lot more code and also more likely to make large refactors.
From the parent, making edits to different functions in the same file usually doesn't cause a merge conflict unless they're close together. Usually it's when you have the same lines being replaced with 2 different options.
I think humans naturally tend to try to avoid working on the same area of the codebase at the same time and have other mitigations like daily meetings to coordinate and organize.
That said, yeah, I think they're complaining about "slap AI on everything" as well
Totally fair. The merge logic doesn't care whether a human or an agent wrote the commit, "two commits touch different functions in the same file, clean merge" is exactly how a human should read it too. We lead with agents because that's where merge volume is about to explode and the pain is worst, not because it's agent-only. Point taken on the copy, I'll make it less exclusionary.
I've genuinely had this problem in teams as small as 3 and it's a genuine headache. It's even a problem when you're working solo!
What I want is a comparison with other improved diff/merge drivers like difftastic/mergiraf.