zsh/Etc/completion-style-guide

Conventions
-----------

There are a number of conventions related to writing completion
functions and it is useful if they are followed for functions which are
to be distributed as part of zsh to maintain a level of consistency.

Coding style:

* Use two spaces for indentation and four for continuation lines except
  where there are many continuation lines such as `_arguments' or
  `_values' specs. Lines tend to be longer than in C code so less
  indentation makes sense.

* For constructs such as `if' and `while', the `then' or `do' should be
  on the end of the line after a semi-colon and space unless there
  isn't room for it (see the next point) in which case put it on the
  next line un-indented.

* Please try not to use lines longer than 79 characters. Don't worry
  about breaking long `_arguments' or `_values' specs though.

* Never use alternative, unusual, or optional syntax in completion
  functions (or any other shell code distributed with zsh). In other
  words, do NOT use the following:

    # Short loops
    for x in $y; myfunc $x

    # Alternative forms
    if { [[ $x == $y ]] } {
      myfunc $x
    }
    foreach x in $y {
      myfunc $x
    }

    # Weird tricks
    () for 1 {
      myfunc $1
    } $x

Descriptions:

Descriptions should not have a trailing full stop and initial capital
letter. Though capitals are fine where you have an acronym which
generally appears in uppercase. Prefer the use of the imperative
mood as it is shorter and more natural in the absence of an explicit
subject for a sentence, e.g. "recurse subdirectories" instead of
"recurses subdirectories", as if orders are being given.

It is a good idea to copy descriptions from the command's man page or
--help output. If you do this, be careful that the description still
makes sense. Some commands have a description like `print help message
(this one) and exit' for --help but the `(this one)' part no longer
makes sense. A less obvious example is where the help output looks like:
  -X, --encoding=NAME        use input encoding NAME
copying this description exactly would result in completion output that
looks like this:
  --encoding   -X    -- use input encoding NAME
In the help output, it is much clearer what is meant by `NAME' because
it appears after `--encoding=' but it doesn't in the completion
listing. So it is better to use a description of this form:
  --encoding   -X    -- use specified input encoding
The word specify is commonly used with options that take arguments.

Another example of where --help output may not be suitable unedited is
where default values or units are indicated. Do not put them in
per-match descriptions; they are better placed in the group
descriptions. Put the units in parentheses after the description. So
for example, do not use:
  '--timeout[specify connection timeout in milliseconds]:timeout'
but use:
  '--timeout[specify connection timeout]:timeout (ms)'
To indicate a default value, use square brackets:
  '--timeout[specify connection timeout]:timeout (ms) [5000]'
These two conventions can be used together or individually as appropriate.
Alternatively the `_numbers' function may be used:
  '--timeout[specify connection timeout]: :_numbers -u ms -d 5000 timeout'

Group descriptions should be singular because only one thing is being
completed even though many may be listed. This applies even where you
complete a list of the things. Tags, functions for completing types of
things (such as _files), and states should have plural names.

Group descriptions can be omitted where they are handled by a helper/type
function. For example, the `file' description in the following line is
unnecessary, as `_files' provides it by default:
  '--import=[import specified file]:file:_files'
In this case, the following syntax can be used instead:
  '--import=[import specified file]: :_files'
Of course, it may make sense to add an explicit description which is
more specific than the default:
  '--config=[use specified config file]:config file:_files'

Similarly, group descriptions can and should be omitted where a `->state'
is involved, as the description in the argument spec is always ignored
in these cases. For example, instead of:
  '--config=[use specified config file]:config file:->config-files'
use:
  '--config=[use specified config file]: :->config-files'
Setting an explicit description here constitutes (a small amount of)
unnecessary noise, and may be misleading to other developers who update
the function.

In a function, allow any descriptions passed as an argument to override
the default you define. For example:
  _wanted directories expl directory _files -/ "$@" -
The "$@" adds descriptions passed as parameters and the trailing `-'
tells _wanted where to put options specifying the `directory' description.

Where two matches have identical meaning, give them the same
description so that the completion system can group them together.
Conventionally a brace expansion of this form is used:
  '(--context -C)'{--context=,-C-}'[specify lines of context]:lines'
You won't need the exclusion list if the option can be specified
multiple times. It can also be useful to use the same description for
matches which are completely opposite in their meaning if it shortens
the completion listing provided that the names of the matches makes it
clear what their effect is.

Command Versions:

In most cases multiple versions (releases) of commands are not
supported. The functions are merely updated to reflect the latest stable
version. Exceptions to this can be made where are particular version
continues to be commonly distributed. Where there is more than one variant
of the same command however (e.g., the command takes different options
on different platforms), the separate variants should be supported.

Contexts, tags and all that
---------------------------

The completion system keeps track of the current context in the
parameter `curcontext'. Its content is the hierarchical name for the
current context sans the `:completion:' and the last colon and the tag
currently tried. The tags represent different types of matches. So,
whenever you are about to add matches, you should use a tag for them
and test if the user wants this type of matches to be generated.
However, this only really needs to be done if no other function in the
call chain has tested that already or if you can offer different types
of matches or if you can handle tag aliases in some sophisticated way.

Most of the utility functions do the testing themselves, so you don't
have to worry about that at all. For example if you are adding matches 
with `_files', `_hosts' or functions like these, you can just call
them and they do the tests needed and the loops over the tag aliases.
The functions `_arguments' and `_values' do that too, but there is a
small difference. These functions effectively change the context
name and if you are using the `->state' form for actions, this changed
name component has to be reported back to the function calling
`_arguments' or `_values'. This is done with the parameter `context',
so you have to make that local in the calling function in the same way
as you have to make local `line', `state', and `{opt,val}_args'. This
parameter `context' should then be used when you start adding matches
by giving it to functions like `_tags' via the `-C' options, as in:

  local context ...
  ...
  _arguments ... '-foo:foo:->foo' && return 0
  ...
  if [[ "$state" = foo ]]; then
    _tags -C "$context" ...
    ...
  fi

This will put the context name given in the argument field of the
`curcontext' parameter and this context will then be used to look 
up styles for the tags.

But since this is often used, `_arguments' and `_values' have support
to make your life easier in such cases. With the `-C' option, these
functions set the parameter `curcontext', thus modifying the globally
used hierarchical context name. This means, that you have to make that 
local, but then you don't have to worry about giving the context name
reported back to functions you call. E.g.:

  local curcontext="$curcontext" ...
  ...
  _arguments -C ... 'foo:foo:->foo' && return 0
  ...
  if [[ "$state" = foo ]]; then
    _tags ...
    ...
  fi

In this case the parameter `context' is not set, so you don't have to
make that local. But make sure that `curcontext' is local so that the
value changed by `_arguments' and `_values' is only used in your
function (and make sure to initialise it to its old value as in the
example).

All this only works if the specifications given to `_arguments' define 
options and arguments that are completely separate. If there is more
than one `->state' action and more than one of them might be needed
for the same word, you'll have to use a loop:

  local state context line i expl ret=1
  ...
  _arguments \
      '::arg1:->arg1' \
      '*:args:->rest' && return 0

  while (( $#state )); do
    case "$state[1]" in
    arg1) _wanted -C "$context[1]" foo expl 'foo' compadd - foo1 foo2 && ret=0;;
    rest) _wanted -C "$context[1]" bar expl 'bar' compadd - bar1 bar2 && ret=0;;
    esac
    shift 1 state
    shift 1 context
  done

  return ret

As you can see, `state' and `context' are really arrays. In this
example, completion for the first argument has to complete both `foo's 
and `bar's.

Then, before adding the matches, see if matches of that type are
requested by the user in the current context. If you will add only one 
type of matches, this is very simple. You can use the function
`_wanted' for this. Well, you can often use it, that is. Use it as in:

  _wanted names expl 'name' compadd - alice bob

This is like testing if the tag `names' is requested by the user and
then calling `_all_labels' with the same arguments.

The `_all_labels' function implements the loop over the tag aliases and
handles the user-defined description, using (in the example) the
parameter `expl' to store options to give to the command. These options
are inserted into the command line either directly before a single
hyphen if there is such an argument or after the first word if there
is no single hyphen. Since using `_all_labels' is so much more convenient
than writing the loop with the `_next_label' function (see below), but
some functions called to generate matches don't accept a single hyphen
as an argument anywhere but want the options built as their last arguments,
`_all_labels' will *replace* the hyphen with the options if the hyphen is
the last argument. A good example for such a function is
`_combination' which can be called like:

  _all_labels foo expl 'descr...' _combination ... -

And the `-' will be replaced by the options that are to be given to
`compadd'.

Note that you can also give the `-J' and `-V' options with the
optional `1' or `2' preceding them supported by `_description':

  _wanted -2V names expl 'name' compadd ...

In some cases one needs to call multiple functions or call `compadd'
more than once to generate the matches. In such a case one needs to
implement the loop over the tag aliases directly. This is done with the 
`_next_label' function. Like this:

  while _next_label names expl 'name'; do
    compadd "$expl[@]" - alice bob && ret=0
    _other_names "$expl[@]" && ret=0
  done
  return ret

Simple enough, I hope. But `_next_label' can do some more: utility
functions normally accept options which are then given to `compadd'.
Since these may contain options for the description and `_next_label' may
generate such options, too, it isn't entirely trivial to decide which
of these options should take precedence. But `_next_label' can do the work
for you here. All you have to do is to give the options your utility
function gets to `_next_label', as in:

  while _next_label names expl 'name' "$@"; do
    compadd "$expl[@]" - alice bob
    ...
  done

That's all. Note that the positional argument "$@" are *not* given to
`compadd'. They will be stuffed into the `expl' array by `_next_label'.

The most complicated case is where you can offer multiple types of
matches. In this case the user should be able to say which types he
wants to see at all and of those which he wants to see he should be
able to say which types should be tried first. The generic solution
for this uses `_tags' and `_requested':

  local expl ret=1

  _tags friends users hosts

  while _tags; do
    _requested friends expl friend compadd alice bob && ret=0
    _requested users && _users && ret=0
    _requested hosts && _hosts && ret=0

    (( ret )) || break   # leave the loop if matches were added
  done

`_tags' with tags as arguments registers those tags and checks which
of them the user wants to see and in which order the tags are to be
tried. This means that internally these tags are stored in multiple
sets. The types of matches represented by the tags from the first set
should be tried first. If that generates no matches, the second set is
tried and so on. `_tags' without arguments just makes the next set be
tried (on the first call it makes the first set be used). The function
`_requested' then tests if the tag given as its first argument is in
the set currently used and returns zero if it is,  i.e. if matches of
that type should be added now. The arguments accepted by `_requested'
are the same as for `_wanted'. I.e. you can call it with only the tag
to test, with the `tag array description' or with that plus the
command to execute.

In some cases (like the `users' and `hosts' tags in the example) you
don't need do the loop over the tag aliases yourself, because the
utility functions like `_users' and `_hosts' do it automatically.

This looks good already. But in many cases such as this one you can
also use the function `_alternative' which simply implements a loop
like the one above. It gets arguments of the form `tag:descr:action'.
E.g.:

  _alternative \
      'friends:friend:(alice bob)' \
      'users:: _users' \
      'hosts:: _hosts'

Which does the same as the previous example. (Note the empty
descriptions in the last two arguments -- the actions start with a
space so that they are executed without giving the description
build by `_alternative', i.e. we just use the description added by
`_users' and `_hosts').

In cases where you have to keep track of the context yourself, you can 
give the sub-context you want to use to `_tags', `_wanted' and
`_alternative' with the `-C' option as described above. You don't need 
to give it to `_requested' -- that function will work on the context
used by the corresponding call to `_tags' automatically.

For the names of the tags: choose simple (short, if at all possible)
names in plural. Also, first have a look at the tag names already used 
by other functions and if any of these names seem sensible for the
type of matches you are about to add, then use those names. This will
allow users to define styles for certain types of matches independent
of the place where they are added.

One final comment about when to use your own argument-contexts: do
this when the command you are writing a completion function for has
different `modes'. E.g. if it accepts host names after a `-h' option
and users or hosts after `-u' and for some reason you can't use
`_arguments' to do the work for you, then use context names as in:

  case "$1" in
  -h)
    _tags -C -h hosts && _hosts && ret=0
    ;;
  -u)
    _alternative -C -u 'users:: _users' 'hosts:: _hosts' && ret=0
    ;;
  esac


Styles
------

Users can associate patterns for hierarchical context names with
certain styles using the `zstyle' builtin. The completion code
should then use these styles to decide how matches should be added and 
to get user-configured values. This, too, is done using the builtin
`zstyle'.

Basically styles map names to a bunch of strings (the `value'). In
many cases you want to treat the value as a boolean, so let's start
with that. To test if, for example, the style `verbose' is set for 
the tag `options' in the context you are currently in, you can just do:

  if zstyle -t ":completion:${curcontext}:options" verbose; then
    # yes, it is set...
  fi

I.e. with the -t option and two arguments `zstyle' takes the first one
as a context and the second one as a style name and returns zero if that
style has the boolean value `true'. Internally it checks if the style
is set to one of `yes', `true', `on', or `1' and interprets that as
`true' and every other value as `false'.

For more complicated styles for which you want to test if the value
matches a certain pattern, you can use `zstyle' with the -m option and
three arguments:

  if zstyle -m ":completion:${curcontext}:foo" bar '*baz*'; then
    ...
  fi

This tests if the value of the style `bar' for the tag `foo' matches
the pattern `*baz*' and returns zero if it does.

If you just want to see if one of the strings in the value is exactly
equal to any of a number of a strings, you can use the -t option and
give the strings after the style name:

  if zstyle -t ":completion:${curcontext}:foo" bar str1 str2; then
    ...
  fi

But sometimes you want to actually get the value stored for a certain
style instead of just testing it. For this `zstyle' supports four
options: `-b', `-s', `-a', and `-h'. After these options, three
arguments are expected, the context, the style, and a parameter name.
The parameter will then be set to the value of the style and the option
says how the strings stored as a value will be stored in the
parameter:

  - `-b': the parameter will be set to a either `yes' or `no'
  - `-s': the parameter will be set to all strings in the value
          concatenated (separated by spaces) to one string
  - `-a': the parameter will be set to an array containing the strings 
          from the value as elements
  - `-h': the parameter will be set to an association with the strings 
          from the value being interpreted alternatingly as keys and
	  values

Some random comments about style names. Use the ones already in use if 
possible. Especially, use the `verbose' style if you can add
matches in a simple and a verbose way. Use the verbose form only if
the `verbose' style is `true' for the current context. Also, if
the matches you want to add have a common prefix which is somehow
special, use the `prefix-needed' and `prefix-hidden' styles. The first 
one says if the user has to give the prefix on the line to make these
matches be added and the second one says if the prefix should be
visible in the list.

And finally, if you need a style whose value can sensibly be
interpreted as a list of words, use array or association styles with
the `-a' or `-h' options to `zstyle'. Otherwise you should only make
sure that an empty value for a style is treated in the same way as if
the style wasn't set at all (this is used elsewhere and we want to
keep things consistent).


Descriptions
------------

Always use descriptions. This is important. Really. *Always* use
descriptions. If you have just written down a `compadd' without a
"$expl[@]" (or equivalent), you have just made an error. Even in
helper functions where you use a "$@": if you can't be sure that there 
is a description in the arguments, add one. You can (and, in most
cases, should) then give the description you generated after the
"$@". This makes sure that the, probably more specific, description
given by the calling function takes precedence over the generic one
you have just generated.

And it really isn't that complicated, is it? Think about a string
people might want to see above the matches (in singular -- that's used 
throughout the completion system) and do:

  local expl

  _description tag expl <descr>
  compadd "$expl@]" - <matches ...>

Note that this function also accepts `-V' and `-J', optionally (in the 
same word) preceded by `1' or `2' to describe the type of group you
want to use. For example:

  _description tag expl '...'
  compadd "$expl[@]" -1V foo - ...    # THIS IS WRONG!!!

is *not* the right way to use a unsorted group. Instead do:

  _description -1V tag expl '...'
  compadd "$expl[@]" - ...

and everything will work fine.

Also, if you are about to add multiple different types of matches, use 
multiple calls to `_description' and add them with multiple calls to
`compadd'. But in almost all cases you should then add them using
different tags anyway, so, see above.

And since a tag directly corresponds to a group of matches, you'll
often be using the tags function that allows you to give the
explanation to the same function that is used to test if the tags are
requested (again: see above). Just as a reminder:

  _wanted [ -[1,2]V | -[1,2]J ] <tag> expl <descr> <cmd> ...

and

  _requested [ -[1,2]V | -[1,2]J ] <tag> expl <descr> [ <cmd> ... ]

is all you need to make your function work correctly with both tags
and description at the same time.


Caching
-------

Where generating matches takes a long time it can be useful to save
the list for reuse in subsequent completion attempts.  If you use
a global variable, prefix the name of it with `_cache_' and explicitly
declare it with `typeset -g'.  In many cases this is sufficient but
where generating matches takes especially long or the list is
especially large, use the `_store_cache` mechanism to allow for a
persistent cache.  When caching matches, also consider whether
generated matches might be affected by style settings for limited
contexts and adapt to allow such configuration to work.


Misc. remarks
-------------

1)  Supply match specifications to `compadd' if there are sensible ones.
2)  Use helper functions that do option completion for you (like
    `_arguments' and `_values') -- it will make your life much
    easier.
3)  Use helper functions like `_users' and `_groups' instead of some ad hoc
    mechanisms to generate such information. This ensures that users can
    change the way these things will be completed everywhere by just using
    their own implementations for these functions.
4)  Make sure that the return value of your functions is correct: zero
    if matches were added and non-zero if no matches were found.
    In some cases you'll need to test the value of `$compstate[nmatches]'
    for this. This should always be done by first saving the old value
    (`local nm="$compstate[nmatches]"') and later comparing this with
    the current value after all matches have been added (e.g. by
    writing `[[ nm -ne compstate[nmatches] ]]' at the end of your
    function).
    This guarantees that your functions will be re-usable because calling
    functions may rely on the correct return value.
5)  When writing helper functions that generate matches, the arguments
    of these should be given unchanged to `compadd' (if they are not
    used by the helper function itself).
6)  When matches with a common prefix such as option names are generated,
    add them *with* the prefix (like `-', `+', or `--' for options).
    Then check the `prefix-needed' style to see if the matches are to be
    added when the prefix is not on the line and use the `prefix-hidden'
    style to see if the prefix should be listed or not.
7)  If at all possible, completion code for a command or a suite of
    commands should go into only one file. If a command has sub-commands,
    implementing a state-machine might be a good idea. See the `_rpm' 
    and `_pbm' files for examples of different styles. Also see the
    documentation for `_arguments' and `_values' for two functions
    that may help you with this.
8)  If a completion function generates completely different types of
    completions (for example, because the command has several
    completely different modes), it should allow users to define
    functions that separately override the behavior for these
    different types. This can easily be achieved by using the
    `_call_function' utility function, as in:

      _call_function ret _command_$subcommand && return ret

    This will try to call the function `_command_$subcommand' and if
    it exists, it will be called and the completion function exits
    with its exit status. After this call to `call_function' the
    completion function would contain the code for the default way to
    generate the matches.
    See the `_email_addresses', `_rpm' and `_nslookup' files for examples.
9)  Be mindful of quoting/escaping edge cases. Notably:
    * Elements of the `$words' array are derived verbatim from the user's
      command-line input -- if they type an argument in quotes or escaped
      by backslashes, that is how it appears in the array.
    * Option-arguments are stored in `$opt_args` the same way. Further,
      since multiple arguments to the same option are represented in a
      colon-delimited fashion, backslashes and colons passed by the user
      are escaped. For instance, the option-arguments parsed from
      `-afoo -a "bar" -a 1:2:3' appear in `$opt_args[-a]` as
      `foo:"bar":1\:2\:3'.
    * The `_call_program` helper used by many completion functions is
      implemented using `eval', so arguments to it must be quoted
      accordingly. As mentioned above, most of the user's own input comes
      pre-escaped, but you may run into problems passing file names or
      data derived from another command's output to the helper. Consider
      using some variation of the `q` expansion flag to deal with this:
      `_call_program vals $words[1] ${(q-)myfile}'
10) If you are going to create a new completion function '_cmd' for a
    command 'cmd', and if the 'cmd' supports the --help option, then you
    may try
      compdef _gnu_generic cmd
      cmd -<TAB>
    _gnu_generic may not work sufficiently well for 'cmd', but the specs
    for _arguments generated from the help text are cached in a variable
    '_args_cache_cmd', and you can save them in a file '_cmd' by
      print -r -- ${(F)${(@qqq)_args_cache_cmd}} > _cmd
    and use the file as a draft of the new completion function.