It really seems like it's just standardizing into a first-class UI what a lot of people have already been doing.
I don't think I'm the target for this - I already use Claude Code with jj workspaces and a mostly design-doc first workflow, and I don't see why I would switch to this, but I think this could be quite useful for people who don't want to dive in so deep and combine raw tooling themselves.
Sure. Honestly I think you can get the same with git work trees, though I haven't tried.
After a couple iterations on this, I've ended up having claude code vibe-code a helper CLI in Go for me which I can invoke with `ontheside <new-workspace-name> <base-change>` and will
- create a new jj workspace based on the given change
- create a docker container configured with everything my unit tests need to run, my claude code config mounted, and jj configured
- it also sets up a claude code hook to run `jj` (no arguments) every time it changes a file, so that jj does a snapshot
- finishes by starting an interactive claude code session with `--dangerously-skip-permissions`
- it also cleans it all up once I exit the claude code session and fish shell that's running it
With this I can have Claude Code working asynchronously on something, while I can review (or peek) the changes in batch from my main editor by running `jj show <change-id>` / `jj diff -r "..."` (which in my case opens it up in the Goland multi-file diff viewer). I can also easily switch to the change it's working on in my main editor, to make some manual modifications if necessary.
This is, in general, primarily for "in the background async stuff" I want to have it work on. Most of the time I just have a dead-normal claude code session running in my main workspace.
Minor self-plug - if you want, I posted a jj intro article a while ago[0], though it doesn't include my current workspace usage.
The Agent Manager view providing a unified view of all active agents and allowing you to immediately respond to any approval requests or followup questions looks very useful regardless of which VCS you're using under the covers. Am I missing something here that jj does?
See my setup detailed in another sibling comment of yours, jj is just a small part of it, and you can probably get that with git too.
I’m already at full mental capacity planning and reviewing the work of two agents (one foreground which almost never asks for approval, and one background which never asks for approval).
I don’t really need the ability to juggle more of them, and noticing their messages is not a bottleneck for me, while I’m happy with the customizability and adaptability of my raw’er workflow.
Fair enough. I use git worktrees (with a script that creates the git branch, worktree and opens a new vs code workspace). You're right, managing more than about two active sessions at once is probably the limit though I'm somewhat hopeful that better tooling similar to the Agent Manager window here would allow me to scale a bit past that especially if some of those sessions are more design explorations.
It really seems like it's just standardizing into a first-class UI what a lot of people have already been doing.
I don't think I'm the target for this - I already use Claude Code with jj workspaces and a mostly design-doc first workflow, and I don't see why I would switch to this, but I think this could be quite useful for people who don't want to dive in so deep and combine raw tooling themselves.