I think it’s kind of unfortunate: I’m not terribly sure how legible Jujutsu’s default log output is for LLMs specifically. Obviously, it’s great for humans. We all love it.

Where it breaks down for me

But, you know, you sit with Codex or something, and you’re trying to get it to understand that you’re in a mega-merge and that this has implications for how it needs to do work, and… here’s what it looks like. And it’s almost impossible—at least for me, it’s impossible—to get Codex to accept the version of reality that is clearly there.

I ended up having to, like, literally inside of Codex manually run log commands with ref sets so that it could see explicitly where the parent-child relationships were. And I shouldn’t need to do that.

A tentative idea to explore

I kind of wonder if we need to develop some kind of log output that is more easily parsable by LLMs than just saying, like, here are all of the edge pairs in this.

I mean, maybe it does just need: “hey, here’s the list of all of the commits, and here’s the list of all the edges between those commits.” I don’t know what looks better. I think it’s maybe something that is worth investigating more. What would you want that “LLM-friendly” log output to optimize for—compactness, unambiguous structure, or something else?