Properly handling spaces and quotes in bash completion

Question

What is the correct/best way of handling spaces and quotes in bash completion?

Here’s a simple example. I have a command called words (e.g., a dictionary lookup program) that takes various words as arguments. The supported ‘words’ may actually contain spaces, and are defined in a file called words.dat:

foo bar one bar two 

Here’s my first suggested solution:

_find_words() { search="$cur" grep -- "^$search" words.dat }  _words_complete() { local IFS=$'\n'  COMPREPLY=() cur="${COMP_WORDS[COMP_CWORD]}"  COMPREPLY=( $( compgen -W "$(_find_words)" -- "$cur" ) )  } complete -F _words_complete words 

Typing ‘words f<tab>’ correctly completes the command to ‘words foo ’ (with a trailing space), which is nice, but for ‘words b<tab>’ it suggests ‘words bar ’. The correct completion would be ‘words bar\ ’. And for ‘words "b<tab>’ and ‘words 'b<tab>’ it offers no suggestions.

This last part I have been able to solve. It’s possible to use eval to properly parse the (escaped) characters. However, eval is not fond of missing quotes, so to get everything to work, I had to change the search="$cur" to

search=$(eval echo "$cur" 2>/dev/null || eval echo "$cur'" 2>/dev/null || eval echo "$cur\"" 2>/dev/null || "") 

This actually works. Both ‘words "b<tab>’ and ‘words 'b<tab>’ correctly autocompletes, and if I add a ‘o’ and press <tab> again, it actually completes the word and adds the correct closing quote. However, if I try to complete ‘words b<tab>’ or even ‘words bar\ <tab>’, it is autocompleted to ‘words bar ’ instead of ‘words bar\ ’, and adding for instance ‘one’ would fail when the words program is run.

Now, obviously it is possible to handle this correctly. For instance, the ls command can do it for files namned ‘foo’ ‘bar one’ and ‘bar two’ (though it does have problems with some ways of expressing the filenames when one uses a (valid) combination of both ", ' and various escapes). However, I couldn’t figure out how ls does it by reading the bash completion code.

So, does anybody know of how properly handle this? The actual input quotes need not be preserved; I would be happy with a solution that changes ‘words "b<tab>’, ‘words 'b<tab>’ and ‘words b<tab>’ to ‘words bar\ ’, for instance, (though I would prefer stripping of quotes, like in this example, instead of adding them).

Answer 1

The question is rather loaded but this answer attempts to explain each aspect:

1. How to handle spaces with COMPREPLY.
2. How does ls do it.

There're also people reaching this question wanting to know how to implement the completion function in general. So:

3. How how do I implement the completion function and correctly set COMPREPLY?

How does ls do it

Moreover, why does it behave differently to when I set COMPREPLY?

Back in '12 (before I updated this answer), I was in a similar situation and searched high and low for the answer to this discrepancy myself. Here's the answer I came up with.

ls, or rather, the default completion routine does it using the -o filenames functionality. This option performs: filename-specific processing (like adding a slash to directory names or suppressing trailing spaces.

To demonstrate:

$ foo () { COMPREPLY=("bar one" "bar two"); } $ complete -o filenames -F foo words $ words ░ 

Tab

$ words bar\ ░          # Ex.1: notice the space is completed escaped 

TabTab

bar one  bar two        # Ex.2: notice the spaces are displayed unescaped $ words bar\ ░ 

Immediately there are two points I want to make clear to avoid any confusion:

• First of all, your completion function cannot be implemented simply by setting COMPREPLY to an array of your word list! The example above is hard-coded to return candidates starting with b-a-r just to show what happens when TabTab is pressed. (Don't worry, we'll get to a more general implementation shortly.)

• Second, the above format for COMPREPLY only works because -o filenames is specified. For an explanation of how to set COMPREPLY when not using -o filenames, look no further than the next heading.

Also note, there's a downside of using -o filenames: If there's a directory lying about with the same name as the matching word, the completed word automatically gets an arbitrary slash attached to the end. (e.g. bar\ one/)

How to handle spaces with COMPREPLY without using -o filenames

Long story short, it needs to be escaped.

In contrast to the above -o filenames demo:

$ foo () { COMPREPLY=("bar\ one" "bar\ two"); }     # Notice the blackslashes I've added $ complete -F foo words                             # Notice the lack of -o filenames $ words ░ 

Tab

$ words bar\ ░          # Same as -o filenames, space is completed escaped 

TabTab

bar\ one  bar\ two      # Unlike -o filenames, notice the spaces are displayed escaped $ words bar\ ░ 

How do I actually implement a completion function?

Implementing a completion functions involves:

1. Representing your word list.
2. Filtering your word list to just candidates for the current word.
3. Setting COMPREPLY correctly.

I'm not going to assume to know all the complex requirements there can be for 1 and 2 and the following is only a very basic implementation. I'm providing an explanation for each part so one can mix-and-match to fit their own requirements.

foo() {     # Get the currently completing word     local CWORD=${COMP_WORDS[COMP_CWORD]}      # This is our word list (in a bash array for convenience)     local WORD_LIST=(foo 'bar one' 'bar two')      # Commands below depend on this IFS     local IFS=$'\n'      # Filter our candidates     CANDIDATES=($(compgen -W "${WORD_LIST[*]}" -- "$CWORD"))      # Correctly set our candidates to COMPREPLY     if [ ${#CANDIDATES[*]} -eq 0 ]; then         COMPREPLY=()     else         COMPREPLY=($(printf '%q\n' "${CANDIDATES[@]}"))     fi }  complete -F foo words 

In this example, we use compgen to filter our words. (It's provided by bash for this exact purpose.) One could use any solution they like but I'd advise against using grep-like programs simply because of the complexities of escaping regex.

compgen takes the word list with the -W argument and returns the filtered result with one word per line. Since our words can contain spaces, we set IFS=$'\n' beforehand in order to only count newlines as element delimiters when putting the result into our array with the CANDIDATES=(...) syntax.

Another point of note is what we're passing for the -W argument. This argument takes an IFS delimited word list. Again, our words contain spaces so this too requires IFS=$'\n' to prevent our words being broken up. Incidentally, "${WORD_LIST[*]}" expands with elements also delimited with what we've set for IFS and is exactly what we need.

In the example above I chose to define WORD_LIST literally in code.

One could also initialize the array from an external source such as a file. Just make sure to move IFS=$'\n' beforehand if words are going to be line-delimited such as in the original question:

local IFS=$'\n' local WORD_LIST=($(cat /path/to/words.dat))` 

Finally, we set COMPREPLY making sure to escape the likes of spaces. Escaping is quite complicated but thankfully printf's %q format performs all the necessary escaping we need and that's what we use to expand CANDIDATES. (Note we're telling printf to put \n after each element because that's what we've set IFS to.)

Those observant may spot this form for COMPREPLY only applies if -o filenames is not used. No escaping is necessary if it is and COMPREPLY may be set to the same contents as CANDIDATES with COMPREPLY=("$CANDIDATES[@]").

Extra care should be taken when expansions may be performed on empty arrays as this can lead to unexpected results. The example above handles this by branching when the length of CANDIDATES is zero.