> One incredible thing was the ability to easily merge what was worth merging from forks, for instance
I agree, this is amazing, and really reduces the wasted effort. But it only works if you know what exists and where.
But IMO the primitives we need are also fundamentally different with AI coding.
Commits kind of don't matter anymore. Maybe PR's don't matter either, except as labels. But CI/hard proof that the code works as advertised is gold, and this is something git doesn't store by default.
Additionally, as most software moves to being built by agents, the "real" git history you want is the chat history with your agent, and its CoT. If you can keep that and your CI runs, you could even throw away your `git` and probably still have a functionally better AI coding system.
If we get a new Github for AI coding I hope it's a bit of a departure from current git workflows. But git is definitely extensible enough that you could build this on git (which is what I think will ultimately happen).