Jj

Jujutsu (jj) Version Control System

Overview

Jujutsu is a powerful Git-compatible version control system that combines ideas from Git, Mercurial, Darcs, and adds novel features. It uses Git repositories as a storage backend, making it fully interoperable with existing Git tooling.

Key differentiators from Git:

• Working copy is automatically committed (no staging area)
• Conflicts can be committed and resolved later
• Automatic rebasing of descendants when commits change
• Operation log enables easy undo of any operation
• Revsets provide powerful commit selection
• Change IDs stay stable across rewrites (unlike commit hashes)

When to Use This Skill

• User mentions "jj", "jujutsu", or "jujutsu vcs"
• Working with stacked/dependent commits, or bookmarks (jj's branches)
• Change IDs vs commit IDs; revset queries for selecting commits
• Conflict resolution, or history rewriting (squash, split, rebase, diffedit)
• Git interoperability, operation log, undo/redo

Key Concepts

Working Copy as a Commit

In jj, the working copy is always a commit and changes are snapshotted automatically. There is no git add; jj status and jj diff report the working-copy commit.

When Snapshots Are Triggered

The working copy is snapshotted into @ when running most jj commands (new, status, diff, log, describe). Force a snapshot with jj util snapshot or just run any jj command.

Change ID vs Commit ID

• Change ID: Stable identifier that persists across rewrites (e.g., kntqzsqt)
• Commit ID: Hash that changes when commit is rewritten (e.g., 5d39e19d)

Always prefer change IDs when referring to commits in commands.

Versioned access (0.37+): Use xyz/n suffix to access hidden/divergent versions:

• xyz/0 - latest version of change xyz
• xyz/1 - previous version (useful for jj restore --from xyz/1 --to xyz)
• Shown automatically in jj log for divergent changes

No Staging Area

Instead of staging, use these patterns:

• jj split - Split working copy into multiple commits
• jj squash -i - Interactively move changes to parent
• Direct editing with jj diffedit
• Hunk-level selection without a TUI: the external jj-hunk tool, if installed - see references/jj-hunk.md

First-Class Conflicts

Conflicts are recorded in commits, not blocking operations: jj rebase -s X -o Y succeeds even when it conflicts, and jj log marks the result with ×. Resolve whenever convenient - see Resolving Conflicts.

Operation Log

Every operation is recorded and can be undone or inspected - see Undoing Mistakes.

Essential Commands

Common Workflows

Starting a New Change

# Working copy changes are auto-committed
# When ready to start fresh work:
jj new                    # Create new commit on top
jj describe -m "message"  # Set description
# Or combine:
jj new -m "Start feature X"

Editing a Previous Commit

# Option 1: Edit in place
jj edit <change-id>       # Make working copy edit that commit
# Make changes, they're auto-committed
jj new                    # Return to working on new changes

# Option 2: Squash changes into parent
jj squash                 # Move all changes to parent
jj squash -i              # Interactively select changes

# Option 3: Auto-squash into correct commits in stack
jj absorb                 # Each hunk goes to commit that last changed those lines
jj absorb -i              # (0.44+) Pick which hunks to consider; rest stay in source

Non-interactively: -i opens the :builtin diff editor, which automation cannot drive. If command -v jj-hunk succeeds, jj-hunk squash 'function("handle_request")' and jj-hunk absorb --dry-run do the same work from a query. See references/jj-hunk.md.

Running a Command Across a Stack (jj run, 0.43+)

Checks out each revision in its own isolated working copy, runs the command, amends the revision, and rebases descendants so the fix propagates:

jj run -r 'trunk()..@' -- cargo fix     # Rewrite each commit in the stack
jj run -j 4 -- pre-commit run           # Parallel (start order is oldest-first)
jj run --ignore-changes -- cargo test   # Just run it; never rewrite commits
jj run --passthrough --ignore-errors -- make lint  # Stream output; don't stop on failure

Rebasing Commits

# Rebase current branch onto main (-d is a still-working alias for -o/--onto)
jj rebase -o main

# Rebase specific revision and descendants
jj rebase -s <rev> -o <destination>

# Rebase only specific revisions (not descendants)
jj rebase -r <rev> -o <destination>

# Insert commit between others
jj rebase -r X -A Y       # Insert X after Y
jj rebase -r X -B Y       # Insert X before Y

# Simplify redundant parents during rebase
jj rebase -s <rev> -o <dest> --simplify-parents

Reordering Commits with jj arrange

jj arrange <revset>       # Open TUI to reorder/abandon commits

The TUI shows selected commits with their immediate parents/children. Use swap up/down to reorder along graph edges.

Working with Bookmarks (Branches)

jj bookmark list          # List bookmarks
jj bookmark create <name> # Create at current commit
jj bookmark set <name>    # Move bookmark to current commit
jj bookmark delete <name> # Delete bookmark
jj bookmark track <name> --remote <remote>  # Track remote bookmark
jj bookmark advance       # Move bookmark forward to @ (like "jj tug")

Gotchas: Use --allow-backwards to move a bookmark to an ancestor. The * suffix in log means diverged from remote (push to sync). Use bookmark set (not create) if the bookmark may already exist on a remote.

Searching File Contents

Breaking in 0.44: -p/--pattern is a required flag; positional args are filesets, not the pattern. Output now prints each matched line prefixed by its path (was: paths only).

jj file search -p foo                # Regex (default kind) -> "b.txt:alpha foo"
jj file search -p foo --name-only    # Paths only (pre-0.44 behavior)
jj file search -p foo -n             # Prefix each match with its line number
jj file search -p 'substring:foo'    # Substring match
jj file search -p 'glob:*foo*'       # Glob must match the WHOLE line
jj file search -p foo 'glob:*.rs'    # Limit to a fileset
jj file search -r <rev> -p foo       # Search at a specific revision

Pushing Changes

# Push specific bookmark
jj git push --bookmark <name>

# Push change by creating auto-named bookmark
jj git push --change <change-id>

# Push all bookmarks AND tags (0.44+); skips ineligible: private/conflicted
jj git push --all

# Pass push options to remote server
jj git push --bookmark <name> --option key=value

Reorder a pushed/stacked branch: jj rebase --ignore-immutable -s <change> -o <dest> (descendants auto-rebase), then jj git push --bookmark <name> — jj git push is force-with-lease-safe by default (no flag). See references/github-workflow.md.

Resolving Conflicts

# After a rebase creates conflicts:
jj log                    # Find conflicted commit (marked with ×)
jj new <conflicted>       # Create commit on top
# Edit files to resolve conflicts
jj squash                 # Move resolution into conflicted commit

# Or use external merge tool:
jj resolve                # Opens merge tool for each conflict
jj resolve --list         # List all conflicted files

Binary & Merge Conflicts

Binary files can't hold conflict markers - pick a version with jj restore --from main path/to/binary.wasm. Multi-parent merge conflicts use the same new/edit/squash loop above. Create a merge with:

jj new branch-a branch-b branch-c -m "integration: merge features"

Undoing Mistakes

jj undo                   # Undo last operation
jj redo                   # Redo undone operation
jj op log                 # View operation history
jj op revert <op-id>      # Revert specific operation
jj op restore <op-id>     # Restore to specific state

# View repo at past operation
jj --at-op=<op-id> log

# Run command without affecting repo state (read-only inspection)
jj --no-integrate-operation log -r 'trunk()..@'

Revsets Quick Reference

Examples:

jj log -r '@::'           # Working copy and descendants
jj log -r 'trunk()..@'    # Commits between trunk and working copy
jj log -r 'mine() & ::@'  # My commits in working copy ancestry
jj rebase -s 'roots(trunk()..@)' -o trunk()  # Rebase branch onto trunk

Git Interoperability

Colocated Repositories

By default, jj git clone and jj git init create colocated repos where both jj and git commands work:

jj git clone <url>        # Creates colocated repo (default)
jj git clone --no-colocate <url>  # Non-colocated (jj only)

Converting Existing Git Repo

cd existing-git-repo
jj git init --colocate    # Add jj to existing Git repo

In colocated repos, Git changes are imported/exported automatically on every jj command. Since 0.44, jj git import and jj git export are inert here — not errors: they print No import needed in colocated workspaces., exit 0, and change nothing (they were usually no-ops and had a race condition). So re-sync by running any jj command (e.g. jj st); if you genuinely suspect divergence, force one with jj git import --ignore-working-copy. Best practice: primarily use jj commands.

Tags Are First-Class (0.44+)

jj git fetch now fetches tags like bookmarks — as <name>@<remote>, auto-tracked by a local tag of the same name. Git's tagOpt is no longer respected; control fetching with remotes.<remote>.fetch-tags in jj config (a pattern, e.g. '~*' to disable). jj git push --all pushes tags too.

jj tag list                     # List tags and targets
jj tag set <name> -r <rev>      # Create/update a tag
jj tag track <name>@<remote>    # Start/stop tracking a remote tag

Gotcha: tags() is part of builtin_immutable_heads(), so newly-tracked tags can make previously-mutable commits immutable.

Colocated repos: Git sees @ as uncommitted

Git's HEAD tracks @'s parent (@-), so @'s own changes look uncommitted to git. git checkout/git switch (and git-based tooling — gh, git-chain, hooks) can abort — but only when the target branch touches the same files as the @-vs-parent diff (error: Your local changes to the following files would be overwritten by checkout … Please commit your changes or stash them before you switch branches. Aborting); with no file overlap the checkout succeeds and carries the change over. Fix — park @ so the tree is fully committed first:

jj new <bookmark>         # empty commit on the tip → clean, fully-committed tree
git checkout <bookmark>   # now succeeds

Configuration

Edit config with jj config edit --user (or --repo for per-repo config, stored outside the repo since 0.38):

[user]
name = "Your Name"
email = "your@email.com"

[ui]
default-command = "log"   # Run 'jj log' when no command given
diff-editor = ":builtin"  # For interactive diff editing (split, squash -i)

[revset-aliases]
'wip' = 'description(exact:"") & mine()'  # Custom revset alias

For automation/LLMs: Use -m flags instead of relying on editors. See Non-Interactive Workflows for patterns that work without user interaction.

See references/configuration.md for comprehensive configuration options including editor setup for interactive use.

Template Language

Customize output with -T/--template:

jj log -T 'change_id.short(8) ++ " " ++ description.first_line() ++ "\n"'
jj log -T 'if(conflict, "CONFLICT ", "") ++ description.first_line()'

Key: ++ concatenates, if(cond, then, else) conditionals, separate(sep, ...) joins non-empty. See references/templates.md.

Filesets

Select files in commands using fileset expressions:

jj diff 'glob:*.rs'              # Rust files in cwd
jj diff '~Cargo.lock'            # Everything except Cargo.lock
jj split 'glob:**/test_*'        # Test files only
jj log -r 'files("src")'         # Commits touching src/

See references/filesets.md.

Advanced Topics

For comprehensive documentation, see:

• references/github-workflow.md - GitHub/GitLab PR workflows
• references/conflicts.md - Conflict resolution deep dive
• references/faq-patterns.md - Common patterns & FAQ
• references/templates.md - Template language reference
• references/filesets.md - Fileset language reference
• references/revsets.md - Complete revset reference
• references/commands.md - Full command reference
• references/configuration.md - Configuration reference
• references/git-comparison.md - Git to jj command mapping
• references/jj-hunk.md - Non-interactive hunk selection with the optional jj-hunk tool

Troubleshooting

"Working copy is dirty" - Never happens in jj! Working copy is always a commit.

Conflicts after rebase - Normal in jj. Conflicts are recorded, resolve when convenient.

Lost commits - Use jj op log to find when commits existed, then jj op restore.

Divergent changes - Same change ID, different commits. Usually from concurrent edits:

jj log                    # Shows divergent commits with xyz/0, xyz/1 suffixes
jj diff --from xyz/0 --to xyz/1  # Compare versions
jj abandon <unwanted>     # Remove one version
jj restore --from xyz/1 --to xyz  # Restore to previous version

Immutable commit error - Can't modify trunk(), tags, or untracked remote bookmarks (e.g. branches pushed via git/gh/git-chain, not jj git push) by default:

jj --ignore-immutable <command>   # Override globally (before the subcommand)
jj <command> --ignore-immutable   # Or after the subcommand — both work
# then push

Non-Interactive Workflows

Many jj commands open an editor by default. Use these flags for automation and CLI workflows:

Commit Messages Without Editor

Squash Without Editor

jj squash -u (--use-destination-message) keeps the destination's message; jj squash -m "..." sets one explicitly. If either commit has an empty description, jj silently uses the non-empty one.

Conflict Resolution Without Merge Tool

# Use built-in tools instead of external merge tool
jj resolve --tool :ours <file>      # Take "our" version (side #1)
jj resolve --tool :theirs <file>    # Take "their" version (side #2)

# Or use restore for complete replacement
jj restore --from <rev> <file>      # Take file from specific revision

Inherently Interactive Commands

Cannot be made non-interactive: jj diffedit, jj arrange (TUI), jj resolve without --tool, and jj split without file arguments (workaround: pass paths, e.g. jj split -m "First part" src/file1.rs).

Path arguments cannot go finer than a whole file. For hunk-level selection with no editor, the optional jj-hunk tool covers split, commit, squash, diffedit, restore and absorb - see references/jj-hunk.md.

Common Pitfalls

Push Flag Combinations

Removed in 0.42: --allow-new no longer exists (error: unexpected argument). It is unnecessary — --bookmark/--tag/--change/--named now track new remote refs automatically, and --all includes new ones.

Note (0.41+): --all, --tracked, and -r REVSETS don't fail when revisions are private or conflicted - ineligible refs are silently skipped.

Working Copy Changes on Merge Commits

When you jj edit a merge commit, changes appear as "working copy changes" - this is expected. Run jj new to start a fresh commit on top. An empty, description-less, bookmark-less @ is auto-abandoned when you move off it (e.g. via jj new), so a manual jj abandon is usually unnecessary.

Bookmark Movement Refused

If jj bookmark set fails because it would move "backwards":

jj bookmark set name -r <rev> --allow-backwards

This flag is required when moving a bookmark to an ancestor of its current position.

Minimum Version Requirements

Removed: jj git push --allow-new (0.42); git_head()/git_refs() revsets & templates (0.43); jj git import/export in colocated workspaces (0.44).