#Worktrees in Cascade (Feat. Questions)
1 messages · Page 1 of 1 (latest)
Okay, I can wrap my head around that. That is something that, through some of the research I had done, I came to understand. If you know or will humor me a little more—if you have multiple agents working on multiple things, but then you want to merge parts of that back into your code base, how do you avoid merge conflicts?
I've also been asking different AI agents to explain this to me a bit more, and they mostly say: keep your scope small and keep each work tree to a separate feature to avoid overlapping and potential conflicts.
If you have multiple Cascade agents working on the same files/scope of work, then normal git merge conflicts would occur between the different implementations. You'd need to resolve them yourself, or of course you could chat a bit longer with whatever agent to incorporate the best ideas of another agent's work before merging one chat's version of the code.
Got'cha. I think I'm getting a mental model of it now. TY.
This is ideal if you want the feeling of having multiple async agents working for you
Silly question: when would you use worktree over just another feature branch as a solo-dev? That is outside of experimentation? ;
When I don't trust the model I'm using, lol
Like, hey low-end cheapo model, can you achieve this outcome? Here's a git worktree for you to chop and change and flail about in
Could you give me a quick example of when you would personally use it to help me better understand?
If the agent figures it out, great, if not, whatever, I can discard the convo
Yes, I asked Penguin Alpha in order to test that git worktrees in Cascade worked the way I expected this prompt:
Update the @Makefile to only use dynamic help information
Because normally, I wouldn't trust a lesser model to do that without leaving me with a broken mess of a makefile
nodding
But much to my surprise, with a bunch of agentic loops and of trial and error Penguin Alpha did achieve the goal!
So now, I saved a lot of credits, where I would've spent 1 or 2 credits on a model that was much smarter and slower because I couldn't tolerate it breaking my git repo and makefile
Okay, so it's more or less for easy, quick experimentation of ideas you're unsure of so you don't mess up your working code base or have to work through reverting things back to how they once were?
Yeah, that's one use case to work with the feature as it is in Windsurf
But I'm sure you could use it to go off and try to one-shot a feature or something as well?
Thank you so much for taking the time to explain this to me. You've been a tremendous help.