Jujutsu

Disclaimer

This is not a Google product. It is an experimental version-control system (VCS). It was written by me, Martin von Zweigbergk (martinvonz@google.com). It is my personal hobby project. It does not indicate any commitment or direction from Google.

Introduction

I started the project mostly in order to test the viability of some UX ideas in practice. I continue to use it for that, but my short-term goal now is to make it useful as an alternative CLI for Git repos.

The command-line tool is called jj for now because it's easy to type and easy to replace (rare in English). The project is called "Jujutsu" because it matches "jj" (I initially called it "Jujube", but changed since jujutsu is more well-known).

Features:

• Compatible with Git

  Jujutsu has two backends. One of them is a Git backend (the other is a native one). This lets you use Jujutsu as an alternative interface to Git. The commits you create will look like regular Git commits. You can always switch back to Git.

• The working copy is automatically committed

  Most Jujutsu commands automatically commit the working copy. This leads to a simpler and more powerful interface, since all commands work the same way on the working copy or any other commit. It also means that you can always check out a different commit without first explicitly committing the working copy changes (you can even check out a different commit while resolving merge conflicts).

• Operations update the repo first, then possibly the working copy

  The working copy is only updated at the end of an operation, after all other changes have already been recorded. This means that you can run any command (such as jj rebase) even if the working copy is dirty.

• Entire repo is under version control

  All operations you perform in the repo are recorded, along with a snapshot of the repo state after the operation. This means that you can easily revert to an earlier repo state, or to simply undo a particular operation (which does not necessarily have to be the most recent operation).

• Conflicts can be recorded in commits

  If an operation results in conflicts, information about those conflicts will be recorded in the commit(s). The operation will succeed. You can then resolve the conflicts later. One consequence of this design is that there's no need to continue interrupted operations. Instead, you get a single workflow for resolving conflicts, regardless of which command caused them. This design also lets Jujutsu rebase merge commits correctly (unlike both Git and Mercurial).

• Supports Evolution Jujutsu copies the Evolution feature from Mercurial (see https://www.mercurial-scm.org/wiki/ChangesetEvolution). However, you will often not need to know that because the conflict design described above lets Jujutsu keep commits rebased even if there are conflicts.

Status

The tool is quite feature-complete. I have almost exclusively used jj to develop the project itself since early January 2021. However, there will be changes to workflows and backward-incompatible changes to the on-disk formats (I'll try to provide upgrade commands if requested). It's also likely that workflows and setups different from what I personally use are not well supported.

Tutorial

This text assumes that the reader is familiar with Git.

Setup

First install Jujutsu:

$ git clone https://github.com/martinvonz/jj.git jj-git
$ cd jj-git
# We need the "nightly" Rust toolchain. This command installs that without
# changing your default.
$ rustup install nightly
$ cargo +nightly install --path .
$ cd ..

You may also want to configure your name and email so commits are made in your name. Create a ~/.jjconfig file and make it look something like this:

$ cat ~/.jjconfig
[user]
name = "Martin von Zweigbergk"
email = "martinvonz@google.com"

Cloning a Git repo

Now let's clone the same repo using jj:

# Note the "git" before "clone" (there is no support for cloning native jj
# repos yet)
$ jj git clone https://github.com/martinvonz/jj.git jj-jj
Fetching into new repo in "<dir>/jj-jj"
Working copy now at: 265ecf5cab2d
$ cd jj-jj

Running jj st (short forjj status) now yields something like this:

$ jj st
Parent commit: 723ebb380971 cleanup: restructure escaped newlines to make new rustc happy
Working copy : 265ecf5cab2d
The working copy is clean

We can see from the output above that our working copy has a commit id (265ecf5cab2d in the example).

Let's check out a particular commit, so we get more predicable output:

$ jj co 080a9b37ff7e
Working copy now at: 608c179a60df
added 2 files, modified 57 files, removed 5 files
$ jj st
Parent commit: 080a9b37ff7e cli: make `jj st` show parent commit before working copy commit
Working copy : 608c179a60df
The working copy is clean

You might have noticed that even though we asked to check out some commit (080a9b37ff7e), our working copy commit ended being another commit (608c179a60df). That is because jj co (short for jj checkout) creates a new commit on top of the commit you asked it to check out. The new commit is for the working copy changes. (There's some more nuance to this. We'll go through that in a bit.)

Creating our first change

Now let's say we want to edit the README.md file in the repo (i.e. what you're reading right now) to say that Jujutsu is ready for use. Let's start by describing the change (adding a commit message) so we don't forget what we're working on:

# This will bring up $EDITOR (or `pico` by default). Enter something like
# "Jujutsu is ready!" in the editor and then close it.
$ jj describe
Working copy now at: b2985d68096d Jujutsu is ready!

Now make the change in the README:

# Adjust as necessary for compatibility with your flavor of `sed`
$ sed -i 's/not ready/ready/' README.md
$ jj st
Parent commit: 080a9b37ff7e cli: make `jj st` show parent commit before working copy commit
Working copy : 5f80190c44b9
Working copy changes:
M README.md

Note that you didn't have to tell Jujutsu to add the change like you would with git add. You actually don't even need to tell it when you add new files or remove existing files. However, the flip side of that is that you need to be careful to keep your .gitignore up to date since there's currently no easy way to say that you want an already added file to not be tracked (https://github.com/martinvonz/jj/issues/14).

To see the diff, run jj diff:

$ jj diff
modified file README.md:
    ...
   4    4: ## Disclaimer
   5    5:
   6    6: This is not a Google product. It is an experimental version-control system
   7    7: (VCS). It is not ready for use. It was written by me, Martin von Zweigbergk
   8    8: (martinvonz@google.com). It is my personal hobby project. It does not indicate
   9    9: any commitment or direction from Google.
  10   10:
    ...

Jujutsu's diff format currently only has inline coloring of the diff (like git diff --color-words), which makes the diff impossible to see in the un-colorized output above (the "not" in "not ready" is red).

As you may have noticed, the working copy commit's id changed both when we edited the description and when we edited the README. However, the parent commit stayed the same. Each change to the working copy commit amends the previous version. So how do we tell Jujutsu that we are done amending the working copy commit? The answer is that we need to "close" the commit. When we close a commit, we indicate that we're done making changes to the commit. As described earlier, when we check out a commit, a new working copy commit is created on top. However, that is only true for closed commits. If the commit is open, then that commit itself will be checked out instead. (There's still more nuance to this. We'll get back to that when we talk about conflicts.)

So, let's say we're now done with this commit, so we close it:

$ jj close
Working copy now at: 192b456b024b
$ jj st
Parent commit: fb563a4c6d26 Jujutsu is ready!
Working copy : 192b456b024b
The working copy is clean

Note that a commit id printed in green indicates an open commit and blue indicates a closed commit.

If we later realize that we want to make further changes, we can make them in the working copy and then run jj squash. That command squashes the changes from a given commit into its parent commit. Like most commands, it acts on the working copy commit by default.

The log command, "revsets", and aliases

You're probably familiar with git log. Jujutsu has very similar functionality in its jj log command. It produces hundreds of lines of output, so let's pipe its output into head:

$ jj log | head
@ 192b456b024b f39aeb1a0200 martinvonz@google.com 2021-05-23 23:10:27.000 -07:00
|
o fb563a4c6d26 f63e76f175b9 martinvonz@google.com 2021-05-23 22:13:45.000 -07:00
| Jujutsu is ready!
o 080a9b37ff7e 6a91b4ba16c7 martinvonz@google.com 2021-05-23 22:08:37.000 -07:00 main
| cli: make `jj st` show parent commit before working copy commit
o ba8ff31e32fd 302257bdb7e5 martinvonz@google.com 2021-05-23 22:08:12.000 -07:00
| cli: make the working copy changes in `jj status` clearer
o dcfc888f50b3 7eddf8dfc70d martinvonz@google.com 2021-05-23 22:07:40.000 -07:00
| cli: remove "Done" message at end of git clone

The @ indicates the working copy commit. The first hash on a line is the commit id. The second hash is a "change id", which is an id that follows the commit as it's rewritten (similar to Gerrit's Change-Id). You can give either hash to commands that take revisions as arguments. We will generally prefer change ids because they stay the same when the commit is rewritten.

By default, jj log lists all revisions (commits) in the repo that have not been rewritten (roughly speaking). We can use the -r flag to restrict which revisions we want to list. The flag accepts a "revset", which is an expression in a simple language for specifying revision. For example, @ refers to the working copy commit, root refers to the root commit, branches() refers to all commits pointed to by branches. We can combine expression with | for union, & for intersection and - for difference. For example:

$ jj log -r '@ | root | branches()'
@ 192b456b024b f39aeb1a0200 martinvonz@google.com 2021-05-23 23:10:27.000 -07:00
:
o 080a9b37ff7e 6a91b4ba16c7 martinvonz@google.com 2021-05-23 22:08:37.000 -07:00 main
: cli: make `jj st` show parent commit before working copy commit
o 000000000000 000000000000  1970-01-01 00:00:00.000 +00:00

(The 000000000000 commit is a virtual commit that's called the "root commit". It's the root commit of every repo. The root symbol in the revset matches it.)

There are also operators for getting the parents (:foo), children foo:, ancestors (,,foo), descendants (foo,,), DAG range (foo,,bar, like git log --ancestry-path), range (foo,,,bar, like Git's foo..bar). There are also a few more functions, such as public_heads(), which is the set of revisions that have Git remote-tracking branches pointing to them, except those that are ancestors of other revisions in the set. Let's define an alias based on that by adding the following to ~/.jjconfig:

[alias]
l = ["log", "-r", "(public_heads(),,,@),,"]

The alias lets us run jj l to see the commits we have created between public heads (exclusive) and the working copy (inclusive), as well as their descendants:

$ jj l
@ 192b456b024b f39aeb1a0200 martinvonz@google.com 2021-05-23 23:10:27.000 -07:00
|
o fb563a4c6d26 f63e76f175b9 martinvonz@google.com 2021-05-23 22:13:45.000 -07:00
~ Jujutsu is ready!

Conflicts

Now let's see how Jujutsu deals with merge conflicts. We'll start by making some commits:

# Check out the grandparent of the working copy
$ jj co ::@
Working copy now at: 9164f1d6a011
added 0 files, modified 1 files, removed 0 files
$ echo a > file1; jj close -m A
Working copy now at: 5be91b2b5b69
$ echo b1 > file1; jj close -m B1
Working copy now at: a0331f1eeece
$ echo b2 > file1; jj close -m B2
Working copy now at: fd571967346e
$ echo c > file2; jj close -m C
Working copy now at: 4ae1e0587eef
$ jj co ::::@
Working copy now at: 9195b6d2e8dc
added 0 files, modified 1 files, removed 1 files
$ jj l
@ 9195b6d2e8dc 47684978bf4b martinvonz@google.com 2021-05-26 12:39:56.000 -07:00
|
| o 1769bdaa8d6d 8e6178b84ffb martinvonz@google.com 2021-05-26 12:39:35.000 -07:00
| | C
| o de5690380f40 5548374c0794 martinvonz@google.com 2021-05-26 12:39:30.000 -07:00
| | B2
| o 47e336632333 ce619d39bd96 martinvonz@google.com 2021-05-26 12:39:20.000 -07:00
|/  B1
o 661432c51c08 cf49e6bec410 martinvonz@google.com 2021-05-26 12:39:12.000 -07:00
~ A

We now have a few commits, where A, B1, and B2 modify the same file, while C modifies a different file. We checked out A in order to simplify the next steps. Let's now rebase B2 directly onto A:

$ jj rebase -s 5548374c0794 -d cf49e6bec410
Rebased 2 commits
$ jj l
o 66274d5a7d2d 8e6178b84ffb martinvonz@google.com 2021-05-26 12:39:35.000 -07:00  conflict
| C
o 0c305a9e6b27 5548374c0794 martinvonz@google.com 2021-05-26 12:39:30.000 -07:00  conflict
| B2
| @ 9195b6d2e8dc 47684978bf4b martinvonz@google.com 2021-05-26 12:39:56.000 -07:00
|/
| o 47e336632333 ce619d39bd96 martinvonz@google.com 2021-05-26 12:39:20.000 -07:00
|/  B1
o 661432c51c08 cf49e6bec410 martinvonz@google.com 2021-05-26 12:39:12.000 -07:00
~ A

There are several things worth noting here. First, the jj rebase command said "Rebased 2 commits". That's because we asked it to rebase commit B2 with the -s option, which also rebases descendants (commit C in this case). Second, because B2 modified the same file (and word) as B1, rebasing it resulted in conflicts, as the jj l output indicates. Third, the conflicts did not prevent the rebase from completing successfully, nor did it prevent C from getting rebased on top.

Now let's resolve the conflict in B2. We'll do that by checking out B2, which will create a new commit on top (even if B2 had been an open commit). That way the conflict resolution doesn't get mixed in with the changes in B2. Once we've resolved the conflict, we'll squash the conflict resolution into the conflicted B2. That might look like this:

$ jj co 5548374c0794  # Replace the hash by what you have for B2
Working copy now at: 619f58d8a988
added 0 files, modified 1 files, removed 0 files
$ cat file1
<<<<<<<
-------
+++++++
-b1
+a
+++++++
b2
>>>>>>>
$ echo resolved > file1
$ jj squash
Rebased 1 descendant commits
Working copy now at: e659edc4a9fc
$ jj l
@ e659edc4a9fc 461f38324592 martinvonz@google.com 2021-05-26 12:53:08.000 -07:00
|
| o 69dbcf76642a 8e6178b84ffb martinvonz@google.com 2021-05-26 12:39:35.000 -07:00
|/  C
o 576d647acf36 5548374c0794 martinvonz@google.com 2021-05-26 12:39:30.000 -07:00
| B2
| o 47e336632333 ce619d39bd96 martinvonz@google.com 2021-05-26 12:39:20.000 -07:00
|/  B1
o 661432c51c08 cf49e6bec410 martinvonz@google.com 2021-05-26 12:39:12.000 -07:00
~ A

Note that commit C automatically got rebased on top of the resolved B2, and that C is also resolved (since it modified only a different file).

By the way, if we want to get rid of B1 now, we can run jj prune 47e336632333. That will hide the commit from the log output and will rebase any descendants to its parent.

The operation log

Jujutsu keeps a record of all changes you've made to the repo in what's called the "operation log". Use the jj op (short for jj operation) family of commands to interact with it. To list the operations, use jj op log:

$ jj op log
@ 5bd384507342 martinvonz@<hostname> 2021-05-26 12:53:08.339 -07:00 - 2021-05-26 12:53:08.350 -07:00
| squash commit 41f0d2289b568bfcdcf35f73d4f70f3ab6696398
| args: jj squash
o 2fd266a8a2e0 martinvonz@<hostname> 2021-05-26 12:53:08.335 -07:00 - 2021-05-26 12:53:08.338 -07:00
| commit working copy
o 1e6dd15305a3 martinvonz@<hostname> 2021-05-26 12:52:39.374 -07:00 - 2021-05-26 12:52:39.382 -07:00
| check out commit 0c305a9e6b274bc09b2bca85635299dcfdc6811c
| args: jj co 0c305a9e6b27
o 401652a2f61e martinvonz@<hostname> 2021-05-26 12:44:51.872 -07:00 - 2021-05-26 12:44:51.882 -07:00
| rebase commit de5690380f40f3f7fc6b7d66d43a4f68ee606228 and descendants
| args: jj rebase -s de5690380f40 -d 661432c51c08
[many more lines]

The most useful command is jj op undo, which will undo an operation. By default, it will undo the most recent operation. Let's try it:

$ jj op undo
Working copy now at: 41f0d2289b56
$ jj l
@ 41f0d2289b56 b1e3a4afde5e martinvonz@google.com 2021-05-26 12:52:39.000 -07:00
|
| o 66274d5a7d2d 8e6178b84ffb martinvonz@google.com 2021-05-26 12:39:35.000 -07:00  conflict
|/  C
o 0c305a9e6b27 5548374c0794 martinvonz@google.com 2021-05-26 12:39:30.000 -07:00  conflict
| B2
| o 47e336632333 ce619d39bd96 martinvonz@google.com 2021-05-26 12:39:20.000 -07:00
|/  B1
o 661432c51c08 cf49e6bec410 martinvonz@google.com 2021-05-26 12:39:12.000 -07:00
~ A

As you can perhaps see, that undid the jj squash invocation we used for squashing the conflict resolution into commit B2 earlier. Notice that it also updated the working copy.

You can also view the repo the way it looked after some earlier operation. For example, if you want to see jj l output right after the jj rebase operation, try jj l --at-op=401652a2f61e but use the hash from your own jj op log.

Moving content changes between commits

You have already seen how jj squash can combine the changes from two commits into one. There are several other commands for changing the contents of existing commits. These commands assume that you have meld installed. If you prefer vimdiff, add this to your ~/.jjconfig file:

[ui]
diff-editor = "vimdiff"

We'll need some more complex content to test these commands, so let's create a few more commits:

$ jj co origin/main
Working copy now at: 61b0efa09dbe 
added 0 files, modified 0 files, removed 1 files
$ printf 'a\nb\nc\n' > file; jj close -m abc
Working copy now at: f9147a088c0d 
$ printf 'A\nB\nc\n' > file; jj close -m ABC
Working copy now at: 9d97c5018b23 
$ printf 'A\nB\nC\nD\n' > file; jj close -m ABCD
Working copy now at: c5a985bc3f41 
$ jj l
@ c5a985bc3f41 3568f6e332d5 martinvonz@google.com 2021-05-26 14:36:46.000 -07:00 
| 
o 687009839bae 874f2d307594 martinvonz@google.com 2021-05-26 14:36:38.000 -07:00 
| ABCD
o ad9b1ce3b5d0 2bbc0c1eb382 martinvonz@google.com 2021-05-26 14:36:26.000 -07:00 
| ABC
o a355fb177b21 3680117711f5 martinvonz@google.com 2021-05-26 14:36:05.000 -07:00 
~ abc

We "forgot" to capitalize "c" in the second commit when we capitalized the other letters. We then fixed that in the third commit when we also added "D". It would be cleaner to move the capitalization of "c" into the second commit. We can do that by running jj squash -i (short for jj squash --interactive) on the third commit. Remember that jj squash moves all the changes from one commit into its parent. jj squash -i moves only part of the changes into its parent. Now try that:

$ jj squash -i -r :@
Rebased 1 descendant commits
Working copy now at: 4b4c714b36aa 

That will bring up Meld with a diff of the changes in the "ABCD" commit. Modify the right side of the diff to have the desired end state in "ABC" by removing the "D" line. Then close Meld. If we look the diff of the second commit, we now see that all three lines got capitalized:

$ jj diff -r ::@
modified file file:
   1    1: aA
   2    2: bB
   3    3: cC

The child change ("ABCD" in our case) will have the same content state after the jj squash command. That means that you can move any changes you want into the parent change, even if they touch the same word, and it won't cause any conflicts.

Let's try one final command for changing the contents of an exiting commit. That command is jj edit, which lets you edit the contents of a commit without checking it out.

$ jj edit -r ::@
Created 2423c134ea70 ABC
Rebased 2 descendant commits
Working copy now at: d31c52e8ca41 

When Meld starts, edit the right side by e.g. adding something to the first line. Then close Meld. You can now inspect the rewritten commit with jj diff -r ::@ again and you should see your addition to the first line. Unlike jj squash -i, which left the content state of the commit unchanged, jj edit (typically) results in a different state, which means that descendant commits may have conflicts.

Other commands for rewriting contents of existing commits are jj restore -i, jj split, jj unsquash -i. Now that you've seen how jj squash -i and jj edit work, you can hopefully figure out how those work (with the help of the instructions in the diff).

Future plans

TODO