Notes from JJ Con 2026

Don't expect a lot from these, I ran out of energy halfway through and tbh they're mostly for me anyway lol

Josh Steadmon: The Git and JJ Object Models

• Git's reflog exists because refs are mutable and not content addressed, so there needs to be something that keeps track of their changes
• In JJ commits explicitly point at their predecessor(s) to sidestep having to do reflog archaeology
• JJ commits can point to multiple root trees instead of just one (used for representing conflicts as a first-class concept)
  • Conflict markers are not stored anywhere in JJ's data, it instead materializes them when populating the worktree on disk from the commit's root trees
    • Conflict algebra fucks
• Views represent refs, head commits, Git metadata, and worktree commits at a given point in time in an immutable, content-addressed way
• Operations link views over time

Nick Gerace: How to Accidentally Create a Graph-Based VCS

• DAGs are DAGs and DAGs are Hard
• Oh no not stored procedures
• This reminds me of perpetual patches like in my Helix megafork, would be cool if that was a first-class feature but maybe doesn't need to be more than a technique
  • How do we represent evergreen patches in Jujutsu?
• DAGs are DAGs and DAGs are Hard
• Note to self: look into monotonic clocks and vector clocks
• DAGs are DAGs and DAGs are Hard
• Causal garbage collection is hard
• DAGs are DAGs and DAGs are Hard
• Ends make the means w/ operational transforms, basically compute the necessary diff to get to the end result
• DAGs are DAGs and DAGs are Hard
• Copy on write can be more expensive than is worth it in some situations
• Immutability is safety, treat mutability like unsafe
• DAGs are DAGs and DAGs are Hard
• Welcome back locking
• DAGs are DAGs and DAGs are Hard
• Split graphs to make it easier?
• Welcome back materialized views
• "JJ stores history, not data" strikes again
• DAGs are DAGs and DAGs are Hard
• CRDTs strike again
• Log replication?
• We keep trying to abuse Jujutsu to represent non-VCS domains, but it never quite works because it's a mismatch between data and delta. How do we resolve this?

Pandelis Zembashis: Exploring a Jujutsu without Git

• What does a forge built only for JJ look like?
  • Related to my thoughts on what code review looks like in a Jujutsu world
• Git makes rebasing and multi-author stacks difficult
• Jujutsu is not prescriptive about how a backend is implemented as long as it follows the spec, which is just an interface layer
• This feels more similar to my work on Mnemosyne than I anticipated, but without the Git indirection
• Jujutsu backends as WASM? Why compile in a backend when you could decouple it?

Laurynas Keturakis: Change IDs All the Way Down

• The abstractions that seem to survive today are those that both we humans and agents can utilize without friction
• See: https://isaaccorbrey.com/ramblings/how-accessible-to-llms-is-jujutsu
• Making code review for self better makes review for others better
• Check out jj-hunk for better non-interactive splitting
• hunkset => revset for diffs
  • My idea of property testing for revsets feels relevant here