z is a tool that will remember all the directories you are visiting when using your terminal, and then make it possible to jump around those directories quickly.

Let's try and rewrite this functionality from scratch, maybe we'll learn a few things this way.

Motivation

I started using z about one year ago.

It worked fine except for one thing.

Let's say I have two directories matching spam, /path/to/spam-egss, and /other path/to/eggs-with-spam.

z will compute a score for each directory that depends on how often and how recently it was accessed.

So, in theory:

a directory that has low ranking but has been accessed recently will quickly have higher rank than a directory accessed frequently a long time ago. (excerpt of z's README, emphasis is mine)

But in practice, if you start working from spam-eggs to eggs-with-spam you will get the wrong answer for z spam a few times, ("quickly" does not means "instantly")

Also, because the algorithm uses the date of access, you cannot (easily) predict which directory it will choose.

Step one: choose a database format

z uses a database that looks like this

/other/path/to/eggs-with-spam|24|1495372880
/path/to/spam-eggs|4|1495372491
...

There's the path, the number of times it was accessed and the timestamp of last visit.

We'll use json, and since we want an algorithm that does not depend of the time, we'll only store the number of accesses:

{
  "/other/path/to/eggs-with-spam": 24,
  "/path/to/spam-eggs": 4,
  ...
}

Why json? Because it's quite easy to read and write (including pretty print) in any language, while still being editable by humans.

It still is possible to add new data should we need to.

Step two: decide how to handle non-existing directories

Let's say you work in /tmp/foo, and no other directory is called foo.
You also create the directories /tmp/foo/src and /tmp/foo/include.

With z, once the three paths, /tmp/foo , /tmp/foo/src/ and /tmp/foo/include are stored in the database, they will stay there for ever.

This means that if you remove /tmp/foo, z foo will still try to go into the non-existing directory. But, if you re-create /tmp/foo later on, z foo will work again.

In our rewrite, we'll deal with this situation an other way:

• One, if we can't cd to a directory, we'll remove it from the database immediately
• Two, we'll make it possible to explicitly "purge" the database: in our example, that means removing all three paths in one step. To do this, we'll look at every path in the database and remove those which no longer exist.

Step three: write a command line tool

I decided to name the tool cwd-history, and to use an command line syntax looking like git, with several possible "verbs" for the various actions:

• cwd-history list: to display the paths in the correct order
• cwd-history add PATH: add a path to the database
• cwd-history remove PATH: to remove a path from the database
• cwd-history edit: to edit the json file directly (could become handy)
• cwd-history clean: to remove non-existing paths from the database

The code is on github if you want to take a look.

Some notes:

def get_db_path():
    zsh_share_path = os.path.expanduser("~/.local/share/zsh")
    os.makedirs(zsh_share_path, exist_ok=True)

• We honor XDG standard (in reality we should also check for the XDG_DATA_HOME environment variable, but this is better than polluting $HOME).

• We use the exist_ok¹ argument for os.makedirs, so that the command does not fail if the directory already exists.

def add_to_db(path):
  path = os.path.realpath(path)

• We use os.path.realpath to make sure all the symlinks are resolved. This means that we won't have duplicated paths in our database.

def clean_db():
    cleaned = 0
    entries = read_db()
    for path in list(entries.keys()):
        if not os.path.exists(path):
            cleaned += 1
            del entries[path]
    if cleaned:
        print("Cleaned", cleaned, "entries")
        write_db(entries)
    else:
        print("Nothing to do")

• We try to provide as many information as possible in just one line, by displaying either the number of directories that were cleaned, or that nothing was done. Listing all the paths that were removed would be too verbose, and if we did not display anything we would never be sure the command really worked.

import operator

def list_db():
    entries = read_db()
    sorted_entries = sorted(entries.items(),
                            key=operator.itemgetter(1))
print("\n".join(x[0] for x in sorted_entries))

• We use operator.itemgetter as a shortcut to lambda x: x[1]
• Thus, we sort the paths by the number of times they were accessed.
• We only display the paths, each of them separated by one line. This will make the output of cwd-history list really easy to parse. (We could add a --verbose option to display the full database data for instance)

Step four: hook into zsh

We want to call cwd-history add $(cwd) every time zsh changes the current working directory.

This is done by writing a function and add it to a special array:

function register_cwd() {
  cwd-history add "$(pwd)"
}
typeset -gaU chpwd_functions
chpwd_functions+=register_cwd

• Note how chpwd_functions is a zsh array, so we have to use typeset. We call it with -g because chpwd_functions is a global value, and -U to make sure the list contains no duplicates. I'm not sure what -a is for, sorry.

Step five: filtering results

Instead of trying to guess the best result, we'll let the user choose by hooking into fzf.

I already talked about fzf on a previous article. The gist of it is that you can pass any command as input to fzf, and let the user interactively select one result from the list.

The implementation looks like this:

  cwd_list=$(cwd-history list)
  ret="$(echo $cwd_list| fzf --no-sort --tac --query=${1})"
  cd "${ret}"
  if [[ $? -ne 0 ]]; then
    cwd-history remove "${ret}"
  fi

Notes:

• We use fzf with --query, so that you can either type z foo, or just z, and only after type the foo pattern in fzf's window

• Since the most likely answers are at the bottom of the cwd-history list command, we use fzf with the --tac option². We also need tell fzf to not sort the input beforehand.

• As explained earlier, we remove the path from the database if we can't cd into it. (This also covers the case where we are lacking permissions to visit the folder, by the way)

Step six: from the shell to neovim and vice-versa

One last thing. Since I do most of my editing in neovim, I'm always looking for ways to achieve similar behaviors in my shell and in my editor.

So let's see how we can transfer information about visited directories from one tool to an other.

From neovim to the shell

This is kind of a ugly hack.

First, I add a auto command to write neovim's current directly into a hard-coded file in /tmp/:

" Write cwd when leaving
function! WriteCWD()
  call writefile([getcwd()], "/tmp/nvim-cwd")
endfunction

autocmd VimLeave * silent call WriteCWD()

And then, I wrap the call to neovim in a function that reads the content of the file and then calls cd, but only if neovim exited successfully.

# Change working dir when Vim exits
function neovim_wrapper() {
  nvim $*
  if [[ $? -eq 0 ]]; then
    cd "$(cat /tmp/nvim-cwd 2>/dev/null || echo .)"
  fi
}

From the shell to neovim

To go the other way, I just call fzf#run() from the fzf.vim plugin with cwd-history list as source and :tcd as sink:³

function! ListWorkingDirs()
  call fzf#run({
        \ 'source': "cwd-history list",
        \ 'sink': "tcd"
        \})
endfunction

command! -nargs=0 ListWorkingDirs :call ListWorkingDirs()
nnoremap <leader>l :ListWorkingDirs<CR>

Conclusion

And there you have it.

z is 458 lines of zsh code.

My re-implementation is 75 lignes of Python, 6 lines of zsh, and 8 lines of vimscript.

It shares the database between the shell and the editor, it is never wrong, and the database stays clean and editable by hand.

Not bad I think :)

PS: You can also use z directly with fzf with a few lines of code, as shown in fzf wiki