This hasn't been an issue for me for a while now, and I've actually had Opus 5.5 and Gemini Hydrogen build and submit a 28 commit deep stack, with all the reviewing/rebasing/shifting around that entails.
It works totally fine.
Write yourself a very good JJ skill or reference file from your AGENTS.md.
My LLMs struggled with it a lot too, but I slowly covered everything they were messing up and they haven't made a mistake in months now. Even when there are "divergent" issues, it knows how to recover now.
That was the biggest hurdle for me too. I was a 2+ year jj user, then I had to switch back to git because of all the agentic tools relying on it.
I recently went back to jj, after finding a decent skill [1] and discovering Emacs' majutsu plugin. Claude still tries to do a git command every once in a while, but it's a lot more bearable.
I can say this hasn't always been the case, and until about 2 (maybe 3?) months ago I would handle all the VCS related parts of the work. I would let the agents write code, which I've found less and less necessary to tweak as time has gone on, and write the commit messages and create the PRs by hand.
My colleagues (who use git) let Claude write commit messages, but I find those awfully long-winded and my eyes just glaze over when I open their PRs - but to each their own I guess.
At least the Claude commit messages are an improvement over the commit messages before - I gave up trying to get them to apply the recommendations of how to write good commit messages given in https://cbea.ms/git-commit/
It's a tad opinionated: my agents are not allowed to push to external repos — but that's easy to change if you need
Git and jj work trees are not interoperable, so any tool that assumes git forces you back. Looking at you, Claude app…