Human vs. AI – Diff-based line-level provenance for text under agentic editing
github.comI'm curious why people care about such tracking? I hardly use an editor anymore, and I'm not alone. All that matters is who is signing off on the commit.
> Text a human wrote or edited should be considered close to sacred: an agent should be hesitant and have a very good reason to touch it.
Not sure I agree (particularly with code, not prose). If code is risky to change for reasons that aren't obvious, it should be commented as such. It doesn't matter how the bytes were generated.
> Another use case: the README.md, originally generated, where you rewrite the opening paragraphs. The agent should feel free to redo or append parts further downwards but should really think twice changing anything in the opener.
This, I understand.
> A git repository is already a history of versions each carrying a provenance marker — every revision of the file, in order, with the author of the change that made it.
Maybe I'm missing how people use AI these days but when I have an agent working locally, all git commits have my authorship attached.
A /precommit tool can be very powerful -- not only can it run the tests, do light code review on its own, spin up sub-agents to review performance/architecture/etc impacts, but it can also write hilariously-verbose commit messages once everything passes, detailing not only what changed but why it changed.. tagged with Claude as a co-author.
I take care of that in two ways, 1) is via claude hooks, I make sure these are tagged as claude, 2) my sandboxed agents have git env vars set, with the same effect.
Does Claude run git commit on every change it makes? Doesn't that pollute your git history?
Although that would be an option (I've considered commit on every edit at one point), what I actually ended up doing is that I hook into all Bash tool calls, look if there's a git commit in there, and if so, I prepend it with the right git env variables.
Cool, we do the same thing, but we also denote when a line is "AI generated but was modified by a human" (aka human made anything upwards of a 1 character change).
I'll probably be annoyed enough soon that I want that level of granularity, too
Interesting. I use annotations for that. But even if agents ignore them 1 out of 100 cases, I have enough cases here to see it happen.