I believe that's mostly for fun. Coding agents wouldn't need to interact via the same interfaces humans use, they'd be given a tool to read and write files and they'd be fine with that.
They're literally using diff/patch under the covers, at least the setup i'm currently using.
Did you get them working with diff syntax? I couldn't figure it out, so I just tried a bunch of agentic programs, found a few that actually worked, and it turned out they all use search/replace strings. There's probably other ways to do it but it seems basically everyone settled on that.
I've been trying that with smaller models and had to make some adjustments (e.g. they all really wanted to include the filename twice). So I just make a small tweak and bam suddenly I can edit code with small fast cheap models.
Unfortunately, this characterizes the entire project: "cool" examples with no practical utility. Meanwhile, the language itself is incredibly strange (functions via patterns are an example of strange language choice), extremely slow, and very unstable.
In short, it's developing in the wrong direction.
I switched from Mathematica to Matlab in my work; it was the best investment of time in the entire project
This function is user contributed. It's not official.
The thought of forcing the AI to use vim gave me a nice chuckle. Thank you sir.
I found chatGPT to be bad at VimGolfing.
``` Here is a 35 keystrokes solution that beat your 36 keystrokes solution ! <89 keystrokes> ```
And then it keeps looping in the same way if you ask it about the seahorse emoji (or sometimes just lie about the keystrokes number).
In fact that's not surprising, what is rather surprising is that some of the solutions actually work (>= 100 keystrokes)
They should probably train LLMs to be bad at vim golf. The whole point of vim’s funky language is that human keypresses are very valuable and should not be wasted. Saving keystrokes for an LLM is a non-goal at best.
Forgive my ignorance!