A Rust port and drop-in replacement of autojump

autojump, cli-utilities, rust


autojump-rs version downloads license Travis build status AppVeyor build status

A port of the wildly popular helper application autojump to Rust.


As this project is technically a fork, the license is the same as autojump, which is GPL, either version 3 or any later version. See LICENSE for details.


We have prebuilt binaries available for a while now, thanks to the trust project!

The package is a drop-in replacement of autojump. Assuming autojump is already installed, or at least the shell script part of it has been properly set up, and you have in $PATH ~/.cargo/bin before the system binary locations, all you have to do is to put a binary of your choice architecture in your PATH, overriding the original autojump script.

You may have to issue hash -r for the shell to forget previous location of autojump, if you don't want to re-exec your shell.

(Manually cloning the repository and building is okay, of course.)


Why do a port when the original version works? Primarily for two reasons:

  • The author is really tired of autojump breakage inside Python virtualenvs, and
  • Rust is simply awesome for CLI applications, with its performance and (code) slickness!

Indeed, being written in a compiled language, autojump-rs is very light on modern hardware. As the program itself is very short-running, the overhead of setting up and tearing down a whole Python VM could be overwhelming, especially on less capable hardware. With autojump-rs this latency is greatly reduced. Typical running time is like this on the author's Core i7-2670QM laptop, with a directory database of 1014 entries:

$ time ./autojump/bin/autojump au
./autojump/bin/autojump au  0.09s user 0.01s system 99% cpu 0.103 total

$ time ./autojump-rs/target/release/autojump au
./autojump-rs/target/release/autojump au  0.00s user 0.00s system 87% cpu 0.007 total

The time savings are more pronounced on less powerful hardware, where every cycle shaved off counts. The running time on a 1.4GHz Loongson 3A3000 is about 10ms, for example, which is very close to the x86 figures despite the clock frequency difference.

And, of course, the program no longer interacts with Python in any way, so the virtualenv-related crashes are no more. Say goodbye to the dreaded ImportError's showing every $PS1 in a virtualenv with the system-default Python!

# bye and you won't be missed!
Traceback (most recent call last):
  File "/usr/lib/python-exec/python2.7/autojump", line 43, in <module>
    from autojump_data import dictify
ImportError: No module named autojump_data


All of the command line flags and arguments are now implemented, and behave exactly like the original. Being a drop-in replacement, all other shell features like tab completion should work too. (Except jc and jco; see below.)

As for the text database, the on-disk format should be identical. (Actually there is a little difference in the representation of floats, but it doesn't matter.) However, as the author is developing and using this on Linux, other platforms may need a little more love, although all the libraries used are lovingly cross-platform. (Patches are welcome, of course!)

The Windows batch files shipped with the original autojump has Python hard-coded into them, and obviously that won't work with autojump-rs. Use the batch files in the windows directory instead; just replacing the original files and putting autojump.exe along with them should work. (Thanks @tomv564 for the Windows testing!)

That said, there're some IMO very minor deviations from the original Python implementation. These are:

  • Argument handling and help messages.

    Original autojump uses Python's argparse to parse its arguments. There is a Rust port of it, but it's nowhere as popular as the library, as is shown in statistics and GitHub activities. So it's necessary to re-arrange the help messages at least, as the docopt family of argument parsers mandate a specific style for them. However this shouldn't be any problem, just that it's different. Again, who looks at the usage screen all the day? XD

  • Different algorithm chosen for fuzzy matching.

    The Python version uses the difflib.SequenceMatcher algorithm for its fuzzy matches. Since it's quite a bit of complexity, I chose to leverage the strsim library instead. The Jaro-Winkler distance is computed between every filename and the last part of query needles respectively, and results are filtered based on that.

  • jc may jump outside current directory.

    Exact reason may be different filtering logic involved, but I'm not very sure about this one. The behavior is also observed on original autojump, but the frequency seems to be lower, and both implementations actually don't check if the target is below current directory. However I only use plain j mostly, so if you're heavily reliant on jc and its friends please open an issue!

Future plans

Now that platform support is mostly considered okay, next steps would be more refactoring and bug fixing. The jc behavior differences are observed on original autojump too, in that you could jump outside $(pwd), but the actual directory jumped to is different; this needs further investigation. Hell I even want to write a fasd backend too, but I don't presently have that much free time. Anyway, contributions and bug reports are welcome!