Posted by doshay 22 hours ago
vcs.
i find myself struggling with questions that agents can't answer like "why did you do it?", "when did u delete this folder? why?", etc. or trying to /rewind (after a /compact...) or basically `bisect` to find when and why something was done by the agent in the current / previous session.
just like git did for code, i think we are the same core capabilities with ai agents
so...
i developed an open source solution for that (currently supporting claude code)
would love to get feedback, contribution or maybe other ideas or solutions you find for those problems.
I think it would work with “commit before you want to delete stuff” the same way.
Many people here made comments such as “why do I need another SVC since agents are pretty good with git”, which means they barely read your blurb and did not understand your project.
When you're interacting with agents, multiple prompts may reasonable culminate in a single commit. It may be useful to track or undo things between commits - at the prompt level. I personally have a workflow when I use Jujutsu (jj) for git already, and this slotted in very nicely to solve this problem. The auto-committing in jj makes it very easy and natural to compare diffs between prompts, and undo specific chunks or restore previous states without making a new commit every prompt. I only finish a commit, giving it a message and advancing the branch, once I've iteratively dialed in the changes I want.
I probably won't use this tool since I already have a flow that works for me, but maybe this will help people see why such a tool can be helpful.
Edit: fixed typo
Now impl is just a derivation of the plan, and the plan gets checked in with the same commit so that you can see the why, the intent, the objective, the research that informed the decisions.
Much simpler, and a much, much more effective process than prompt -> impl.
Also commit dag based, but we also wrote this whole knowledge graph / triple-store CRDT data format on top.[1]
We also have p2p syncing of the history so you can use it to track your local work but also to have your agents coordinate within your team.
We had our agents build their own tools on top of that substrate, that way we're vendor independent, this stuff works everywhere from claude web, to self hosted openclaw, you only need to tell your agent to use the faculties.
Because the substrate takes care of everything, every new faculty you write on top of that inherits all of the same properties.
This week I told DeepSeek v4 Flash (max variant) to scavenge for all changes and additions of a specific feature of the project and build a report of the feature timeline with example code and rationale behind the changes.
It fired a ton of read-only git commands (isolated inside Docker) and came up with a neat markdown report of the feature from inception to current state.
If DS4 Flash can do it, for SOTA LLMs like GPT 5.5 and Opus it should be a walk in the park.
That said I don't let LLMs commit. I like to take a close look at every change before committing. Early changes are cheaper.