patchgit-completion.bash: add support for path completion

The git-completion.bash script did not implemented full support for completion, for git commands that operate on files from the current working directory or the index.

For these commands, only options completion was available.

Full support for completion is now implemented, for git commands where the non-option arguments always refer to paths on the current working directory or the index, as the follow:

* the path completion for the "git mv" and "git rm" commands is provided
  using "git ls-files --exclude-standard"

* the path completion for the "git add" command is provided using
  "git ls-files --exclude-standard -o -m"

* the path completion for the "git clean" command is provided using
  "git ls-files --exclude-standard -o"

* the path completion for the "git commit" command is provides using
  "git diff-index --name-only HEAD"

Signed-off-by: Manlio Perillo <manlio.perillo@gmail.com>
---
 contrib/completion/git-completion.bash | 39 ++++++++++++++++++++++------------
 1 file changed, 26 insertions(+), 13 deletions(-)

Manlio Perillo <manlio.perillo@gmail.com> writes:

Hrm, "git mv CO<TAB>" completes it to "COPYING" for me.  Same for:

    git rm CO<TAB>
    git clean Doc<TAB>
    git commit --edit Doc<TAB>

There must be something missing from the above description, and the claim the above two paragraphs make does no make sense without that something that is missing.

Does the above mean "git mv COPYING Doc<TAB>" would complete the command line to move it to Documentation/ directory?

I think "using X" is of secondary importance. Reviewers and future developers (who are reading "git log" output) can read it from the patch. What is expected in the log message is why the implemenation was chosen, and in order to achieve what effect.

For example, something like:

    Each parameter between the first one and the one before the last
    one to 'git mv' is completed to a tracked path or a leading
    directory of a tracked path

should come first to explain what your patch wanted to do. It of course is helpful to explain how you implemented that behaviour, by appending ", using 'ls-files --exclude-standard'" at the end.

Likewise. You are adding "either a modified, or a new and unignored" and use "ls-files --exclude-standard -o -m" to achieve it.

As long as all of the above stops completion at directory boundary, I think the above sounds like a sensible thing to do. e.g. when "ls-files" gives Documentation/Makefile and Documentation/git.txt, "git cmd Doc<TAB>" first would give "git cmd Documentation/" and then the second <TAB> would offer these two paths as choices. That way, the user can choose to just execute "git cmd Documentation/" without even looking at these individual paths.

I am not sure how you would handle the last parameter to "git mv", though. That is by definition a path that does not exist, i.e. cannot be completed.

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Il 17/12/2012 05:50, Junio C Hamano ha scritto:

Thanks.

What is missing is that I forgot to explain that by "completion not supported" I meant "*git aware* completion is not supported".

If you try, as an example, "git add <TAB>", bash will suggest *all* files in your working directory, including ignored files and files with no modifications.

It will suggest all tracked files from the Documentation directory. One important detail is that it will recurse into sub directories.

This is not the behaviour of the current completion code. The problem is that current code use ls-tree, and works only for paths like <rev>:<path>.

My implementation use ls-files, that has a completly different interface from ls-tree (and since I'm very new to git, this caused be some confusion).

It seems that with ls-files it is impossible to:

* get a list of path names without recursing into sub directories.
  (this is only possible for untracked directories)

* get the file mode for all "modified" or "deleted" files.

  If I do "git ls-files --stage --modified", git ignores the modified
  option and will list all files

Ok, thanks.

Right, this is what bash usually do. However I don't know how to implement this with git.

Right, the code should be changed. No completion should be done for the second parameter.

Regards   Manlio
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.10 (GNU/Linux)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org/

iEYEARECAAYFAlDO/zoACgkQscQJ24LbaUSD1ACfXeH2GizdNG/uzlSneEpdi1Na 6IoAni3ZrhhCMXbVoTqIVjH4mEX/XmaH =w56v -----END PGP SIGNATURE-----

Manlio Perillo <manlio.perillo@gmail.com> writes:

That sounds like a regression to me.

I deliberately wrote "the last" not "the second", as you can do

	$ mkdir X
        $ git mv COPYING README X/.

You do need to expand the second parameter to README when the user types

	git mv COPYING REAMDE X

then goes back with \C-b to "M", types \C-d three times to remove "MDE", and then finally says <TAB>, to result in

	git mv COPYING README X

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Il 17/12/2012 20:42, Junio C Hamano ha scritto:

The patch is ready, however I decided to leave git mv completion simple. Pressing <TAB> will always try to autocomplete using all cached files. I have added a note to remember it needs more work.

P.S.: git-completion.bash has a lot of other things that may be improved:

* adding missing commands
 (as an example, there is strangely no custom support fot "git status")

* completion support for commands like "git checkout" is not complete.
  "git checkout <TAB>" will correctly try to complete the tree-ish,
  however "git checkout HEAD -- <TAB>" will try to complete the path
  using *all* files in the working directory.

  This is easy to fix, using the new functions I have added

* not all long options are supported.
  The script documentation says that only common long options are
  supported, so I'm not sure it is ok to add support for all available
  long options.

Regards   Manlio
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.10 (GNU/Linux)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org/

iEYEARECAAYFAlDQmQgACgkQscQJ24LbaUSw9QCfT1lCH/yjA4Lgmb2nMspNWM3l hMMAn26UxWesuoOxMbuwhqaypPjkmN84 =Wh4c -----END PGP SIGNATURE-----

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Il 17/12/2012 20:42, Junio C Hamano ha scritto:

Assuming X is a new untracked directory, do you think it is an usability problem if an user try to do:

	git mv COPYING README <TAB>

and X does not appear in the completion list?

As far as I know, the solution is to only support custom expansion the first parameter, unless the user will do something like:

	git mv COPYING README -- <TAB>

Regards  Manlio

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.10 (GNU/Linux)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org/

iEYEARECAAYFAlDSOWAACgkQscQJ24LbaUSOnACfds93RtX1CDOeGbwCGM5/N8HI yVwAn0AZEO6rE083gKgFimGIbiRTyN5Z =z7K5 -----END PGP SIGNATURE-----

Manlio Perillo <manlio.perillo@gmail.com> writes:

It is hard to say. Will it show "Documentation/" in the list? Will it show "git.c" in the list?

Your "git mv git.<TAB>" completing to "git mv git.c" would be an improvement compared to the stock "less git.<TAB>" that offers "git.c" and "git.o" as choices. For things like "mv" and "rm", this may not make too much of a difference, "git add <TAB>" would be a vast improvement from the stock one. The users will notice that the completion knows what it is doing, and will come to depend on it.

But at that point, if "git mv COPYING README <TAB>" offers only directories that contain tracked files, the users may get irritated, because it is likely that the destination directory was created immediately before the user started typing "git mv". You will hear "Surely I created X, it is there, why aren't you showing it to me?"

The updated completion knows what it is doing everywhere else, and it sets the user-expectation at that level. Uneven cleverness will stand out like a sore thumb and hurts the user perception, which is arguably unfair, but nothing in life is fair X-<.

I think over-showing the choices is much better than hiding some choices, if we cannot do a perfect completion in some cases. "You should know that I won't be moving these files in Y, as I already marked it to be ignored in the .gitignore file!" is less grave a gripe than "You know I created X just a minute ago, and it is there, why aren't you showing it to me?" because you can simply say "but Y is there as a directory." admitting that you are less clever than the user expects you to be, but at least you are giving the choice to the user of not picking it.

In the ideal world (read: I am *not* saying "you should implement it this way, or we won't look at your patch"), I think you would want to show tracked files (because it may be the third path that the user wants to move with the command, not the destination directory) and any directory on the filesystem (it could be the third path that is being moved if it has tracked files, or it could be the final destination directory argument).

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Il 19/12/2012 23:49, Junio C Hamano ha scritto:

Currently all cached files will be showed.

Yes, this is precisely the reason why I'm implementing it ;-).

I also use Mercurial (I discovered Git just a few weeks ago, after reading Pro Git), and Mercurial do have path completion (completion list does not stop at directory boundary, however).

When I started to use Git, one of the first thing I noticed was the lack of path completion for git add.

What about this?

* show all and only cached files for the first argument
* show all cached + untracked directories and files for all other
  arguments

This should solve most of the problem, and will still not show ignored files. And, most important, it is very easy to implement.

The only issue is that "git ls-files -o" does not show empty directories, and "git ls-files --directory -o" will add a trailing slash.

Thanks   Manlio
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.10 (GNU/Linux)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org/

iEYEARECAAYFAlDSUr4ACgkQscQJ24LbaURf0gCeJVZviwfKHgHNZ8vYBjnjwP8+ WF4AnAn3/KPJciJg9r387qIzCajx4j0s =/10k -----END PGP SIGNATURE-----