zsh-development-guide

------------------------------
GUIDELINES FOR ZSH DEVELOPMENT
------------------------------

Zsh is currently developed and maintained by the Zsh Development Group.
This development takes place by mailing list.  Check the META-FAQ for the
various zsh mailing lists and how to subscribe to them.  The development
is very open and anyone is welcomed and encouraged to join and contribute.
Because zsh is a very large package whose development can sometimes
be very rapid, we kindly ask that people observe a few guidelines when
contributing patches and feedback to the mailing list.  These guidelines
are very simple and hopefully should make for a more orderly development
of zsh.

Tools
-----

To develop (as opposed to just build) zsh, you'll need a few specialised
tools:

* GNU autoconf, version 2.50 or later.  This version contained
  significant enhancements and earlier versions will no
  longer work.

* GNU m4.  (Required by autoconf.)

* yodl.

* texi2html.

Patches
-------

* Send all patches to the mailing list rather than directly to me.

* Send only context diffs "diff -c oldfile newfile" or unified diffs
  "diff -u oldfile newfile".  They are much easier to read and
  understand while also allowing the patch program to patch more
  intelligently.  Please make sure the filenames in the diff header
  are relative to the top-level directory of the zsh distribution; for
  example, it should say "Src/init.c" rather than "init.c" or
  "zsh/Src/init.c".  Git-style naming of diffs is also acceptable.

* Please put only one bug fix or feature enhancement in a single patch and
  only one patch per mail message.  This helps me to multiplex the many
  (possibly conflicting) patches that I receive for zsh.  You shouldn't
  needlessly split patches, but send them in the smallest LOGICAL unit.

* If a patch depends on other patches, then please say so.  Also please
  mention what version of zsh this patch is for.

* Please test your patch and make sure it applies cleanly. It takes
  considerably more time to manually merge a patch into the baseline code.

* By convention, patches should be sent with a Subject: line starting with
  one of "PATCH:", "[PATCH]" or "[PATCH n/m]" (for a patch series).

Testing
-------

* Because zsh has a huge number of different options and interacts with
  a wide range of human and artificial life, it is very difficult to
  test the shell thoroughly.  For this purpose, the Test subdirectory
  exists.  It consists of a driver script (ztst.zsh) and various test
  files (*.ztst) in a format which is described in B01cd.ztst, which acts
  as a template.  It is designed to make it easy to provide input to
  chunks of shell code and to test the corresponding standard output,
  error output and exit status.

* There is not much there yet, but please don't let that put you off adding
  tests for basic syntactic features, builtins, options etc. which you
  know to be flakey or to have had difficulties in the past.  Better
  support for testing job control and interactive features is expected
  to follow eventually.

* The directory is not part of the usual process of building and
  installation.  To run the tests, go to Test and `make check'.  Please
  report any errors with all the usual information about the zsh version
  and the system you are using.

C coding style
--------------

* The primary language is ANSI C as defined by the 1989 standard, but the
  code should always be compatible with late K&R era compilers ("The C
  Programming Language" 1st edition, plus "void" and "enum").  There are
  many hacks to avoid the need to actually restrict the code to K&R C --
  check out the configure tests -- but always bear the compatibility
  requirements in mind.  In particular, preprocessing directives must
  have the "#" unindented, and string pasting is not available.

* Conversely, there are preprocessor macros to provide safe access to some
  language features not present in pure ANSI C, such as variable-length
  arrays.  Always use the macros if you want to use these facilities.

* Avoid writing code that generates warnings under gcc with the default
  options set by the configure script.  For example, write
  "if ((foo = bar))" rather than "if (foo = bar)".

* Please try not using lines longer than 79 characters.

* The indent/brace style is Kernighan and Ritchie with 4 characters
  indentations (with leading tab characters replacing sequences of
  8 spaces).  This means that the opening brace is the last character
  in the line of the if/while/for/do statement and the closing brace
  has its own line:

      if (foo) {
	  do that
      }

* Put only one simple statement on a line.  The body of an if/while/for/do
  statement has its own line with 4 characters indentation even if there
  are no braces.

* Do not use space between the function name and the opening parenthesis.
  Use space after if/for/while.  Use space after type casts.

* Do not use (unsigned char) casts since some compilers do not handle
  them properly.  Use the provided STOUC(X) macro instead.

* If you use emacs 19.30 or newer you can put the following line to your
  ~/.emacs file to make these formatting rules the default:

    (add-hook 'c-mode-common-hook (function (lambda () (c-set-style "BSD"))))

* Function declarations must look like this:

  /**/
  int
  foo(char *s, char **p)
  {
      function body
  }

  There must be an empty line, a line with "/**/", a line with the
  type of the function, and finally the name of the function with typed
  arguments.  These lines must not be indented.  The script generating
  function prototypes and the ansi2knr program depend on this format.

* Variable declarations must similarly be preceded by a
  line containing only "/**/", for the prototype generation script.
  The declaration itself should be all on one line (except for multi-line
  initialisers).

* Preprocessor directives that affect the function/variable declarations must
  also be preceded by a "/**/" line, so that they get copied into the
  prototype lists.

* There are three levels of visibility for a function or variable.  It can
  be file-local, for which it must be marked with the keyword "static" at
  the front of the declaration.  It can be visible to other object files in
  the same module, for which it requires no extra keyword.  Or it can be
  made available to the entire program (including other dynamically loaded
  modules), for which it must be marked with the pseudo-keyword "mod_export"
  at the front of the declaration.  Symbols should have the least visibility
  possible.

* Leave a blank line between the declarations and statements in a compound
  statement, if both are present.  Use blank lines elsewhere to separate
  groups of statements in the interests of clarity.  There should never
  be two consecutive blank lines.

* Each .c file *must* #include the .mdh header for the module it is a
  part of and then its own .pro file (for local prototypes).  It may
  also #include other system headers.  It *must not* #include any other
  module's headers or any other .pro files.

* The repository includes a `.editorconfig' file with whitespace/indent
  control settings. Information about text editor plugins and this file
  can be found at <http://editorconfig.org/>.


Modules
-------

Modules have hierarchical names.  Name segments are separated by `/', and
each segment consists of alphanumerics plus `_'.  Relative names are never
used; the naming hierarchy is strictly for organisational convenience.

Each module is described by a file with a name ending in `.mdd' somewhere
under the Src directory.  This file is actually a shell script that will
sourced when zsh is built. To describe the module it can/should set the
following shell variables:

  - name            name of the module
  - link            `static', `dynamic' or `no', as described in INSTALL.
		    In addition, the value `either' is allowed in the .mdd
		    file, which will be converted by configure to `dynamic'
		    if that is available, else `static'.
		    May also be a command string, which will be run within
		    configure and whose output is used to set the value
		    of `link' in config.modules.  This allows a
		    system-specific choice of modules.  For example,
		    link='case $host in *-hpux*) echo dynamic; ;;
		                        *) echo no; ;; esac'
  - load            `yes' or `no': whether the shell should include hooks
		    for loading the module automatically as necessary.
		    (This corresponds to an `L' in xmods.conf in the
		    old mechanism.)
  - moddeps         modules on which this module depends (default none)
  - nozshdep        non-empty indicates no dependence on the `zsh/main'
                    pseudo-module
  - alwayslink      if non-empty, always link the module into the executable
  - autofeatures    features defined by the module for autoloading,
                    a space-separated list.  The syntax for features is as
                    for zmodload -F, e.g. b:mybin refers to the builtin
                    mybin.  This replaces the previous mechanism with
                    separate variables for builtins, conditions, math
                    functions and parameters.  Note the features are only
                    available in zsh's native mode, not in emulation modes.
  - autofeatures_emu As autofeatures, but the features so presented are
                    available in modes that are *not* zsh's native mode.
		    The variable autofeatures must also be present.
  - objects         .o files making up this module (*must* be defined)
  - proto           .syms files for this module (default generated from $objects)
  - headers         extra headers for this module (default none)
  - hdrdeps         extra headers on which the .mdh depends (default none)
  - otherincs       extra headers that are included indirectly (default none)

Be sure to put the values in quotes. For further enlightenment have a look
at the `mkmakemod.sh' script in the Src directory of the distribution.

Modules have to define six functions which will be called automatically
by the zsh core. The first one, named `setup_', should set up any data
needed in the module, at least any data other modules may be interested
in.

The next pair are `features_' and `enables_' and deal with enabling module
features.  Ensure you are familiar with the description of features under
`zmodload -F'.  The function features_ takes an argument `char
***featuresp'; *featuresp is to be set to a NULL-terminated array
containing a list of all the features.  It should then return zero.
It may return one to indicate features are not supported, but this is
not recommended.  The function featuresarray conveniently interrogates
the module's feature structures for all standard features; space
is left for abstract features at the end of the array and the names
must be added by the module.  Note that heap memory should
be used for this (zhalloc, etc.) as memory for the features array is not
freed; note also the pointers for the abstract features are not
initialised so setting them is mandatory any time there are any present.

A structure "struct features" should
be used to contain all standard features as well as the number of
abstract features (those only understood by the module itself).
See below.

enables_ takes an argument `int **enablesp'.  If *enablesp is NULL, it
should be set to an array of the same length as *featuresp without the
NULL, containing a 1 for every feature that is enabled and a zero for other
feature.  By default features are disabled.  If *enablesp is not NULL, its
values should be used to decide whether features are to be turned off.  It
should return status 0 for success, 1 on a failure to alter a feature.  The
function handlefeatures() conveniently handles all standard features
present in the module's features structure; abstract features must be
handled by the module (as with the features array, the area of the enables
array for abstract features is not even initialised by the main shell).  As
with `features_', any handling of the array by the module itself should take
into account that the array will not be freed and any allocation should
therefore be from heap memory.

The functions `features_' and `enables_' can be called at any point
after `setup_' has been called and before `cleanup_' is called.  In
particular they can be called before or after `boot_'.

The function named `boot_' should register function wrappers, hooks and
anything that will be visible to the user that is not handled by features_
and enables_ (so features should not be turned on here).  It will be called
after the `setup_'-function, and also after the initial set of features
have been set by calls to `features_' and `enables_'.

The function named `cleanup_', is called when the user tries to unload
a module and should de-register all features and hooks.  A call
to setfeatures with the final argument NULL will remove all standard
features present in the module's features structure.  Note that
`cleanup_' is called whenever `setup_' succeeded, so that `cleanup_'
must be prepared to handle any state resulting from a failed `boot_'
or initial call to `features_'.  Note also that a return code of 1
from `cleanup_' will result in the module not being unloaded, so
usually `cleanup_' will return 0 even if it has to handle an unclean
state; if it does return 1, it must be prepared to be called again
in a future attempt to unload.

The last function, `finish_' is called when the module is actually unloaded
and should finalize all the data initialized in the `setup_'-function.
However, `finish_' is called even if `setup_' failed, so it should
not rely on the module successfully being set up.
The state from `finish_' module is currently ignored; it is called
too late to prevent the module from being unloaded.
*Note* in addition to freeing memory, variables associated with allocated
memory should be set to NULL or to indicate arrays are empty, etc.  It
should not be assumed that the variables will automatically be zeroed if
the module is reloaded (though some configurations may do this).

In short, the `cleanup_'-function should undo what the `boot_'-function did
(together with handling any residual effects of `enables_'), but should
not rely on `boot_' having been successful, and the
`finish_'-function should undo what the `setup_'-function did, but
should not rely on `setup_' having been successful.

All of these functions should return zero if they succeeded and
non-zero otherwise.

Features
========

Builtins, conditions, parameters (variables) and math functions
are described as "features".  They should be made available to
the shell by declaring a `struct feature' for each module.
Below are descriptions of the individual features; first here
is generic information.

`struct feature' contains a pointer to the array that declares each
feature, followed by the number of entries in the array.  The pointer
can be NULL and the size zero for any feature that is not present in
the module.  For example, to register only builtins in zsh and thereby
make them visible to the user, the structure should contain
"bintab" where the array is declared as an array of struct builtin,
as discussed below:

  static struct feature module_features = {
    bintab, sizeof(bintab)/sizeof(*bintab),
    NULL, 0, /* declare any conditions here */
    NULL, 0, /* declare any parameters here */
    NULL, 0, /* declare any math functions here */
    0, /* number of abstract features */
  }

Within each individual table ("bintab", etc.), features should be listed
in ASCII order as no further sorting is performed by the shell when
features are listed.

Abstract features are handled by the module; the number present
in `struct features' is there to ensure the main shell allocated
space in the features and enables array in the standard
featuresarray() and handlefeatures() calls.  However, the inserting
of names in the features array and the getting and setting of
feature enables is left entirely to the module.  Note that abstract
features should not contain a colon (to avoid clashes with the
prefixes used in standard features).  It is recommended that
only alphanumerics, - and _ be used in the names of abstract
features, and - not be the first character (to avoid confusion
with disabling features) but this is not required by the main shell.

The features_ and enables_ functions for such a module will look
like:

  /**/
  int
  features_example(Module m, char ***features)
  {
     *features = featuresarray(m->nam, &module_features);
     /* fill in any abstract features in (*features) here */
     return 0;
  }
 
  /**/
  int
  enables_example(Module m, int **enables)
  {
    int ret;

    ret = handlefeatures(m->nam, &module_features, enables);
    /* handle any abstract features here */
    ...
    return ret;
  }

The functions shown take the name of the module, the set of features,


To de-register builtins, pass the features structure to
setfeatureenables with a NULL final value:

  /**/
  int
  cleanup_example(Module m)
  {
    setfeatureenables(m->nam, &module_features, NULL);
    ...
  }


Builtins
--------

Builtins are described in a table, for example:

  static struct builtin bintab[] = {
    BUILTIN("example", 0, bin_example, 0, -1, 0, "flags", NULL),
  };

Here `BUILTIN(...)' is a macro that simplifies the description. Its
arguments are:
  - the name of the builtin as a string
  - optional flags (see BINF_* in zsh.h)
  - the C-function implementing the builtin
  - the minimum number of arguments the builtin needs
  - the maximum number of arguments the builtin can handle or -1 if
    the builtin can get any number of arguments
  - an integer that is passed to the handler function and can be used
    to distinguish builtins if the same C-function is used to
    implement multiple builtins
  - the options the builtin accepts, given as a string containing the
    option characters (the above example makes the builtin accept the
    options `f', `l', `a', `g', and `s').  Passing NULL here disables
    all flag handling, i.e. even "--".
  - and finally a optional string containing option characters that
    will always be reported as set when calling the C-function (this,
    too, can be used when using one C-function to implement multiple
    builtins)

The definition of the handler function looks like:

  /**/
  static int
  bin_example(char *nam, char **args, char *ops, int func)
  {
    ...
  }

The special comment /**/ is used by the zsh Makefile to generate the
`*.pro' files. The arguments of the function are the number under
which this function was invoked (the name of the builtin, but for
functions that implement more than one builtin this information is
needed). The second argument is the array of arguments *excluding* the 
options that were defined in the struct and which are handled by the
calling code. These options are given as the third argument. It is an
array of 256 characters in which the n'th element is non-zero if the
option with ASCII-value n was set (i.e. you can easily test if an
option was used by `if (ops['f'])' etc.). The last argument is the
integer value from the table (the sixth argument to `BUILTIN(...)').
The integer return value by the function is the value returned by the
builtin in shell level.

Conditions
----------

The definition of condition codes in modules is equally simple. First
we need a table with the descriptions:

  static struct conddef cotab[] = {
    CONDDEF("len", 0, cond_p_len, 1, 2, 0),
    CONDDEF("ex", CONDF_INFIX, cond_i_ex, 0, 0, 0),
  };

Again a macro is used, with the following arguments:

  - the name of the condition code without the leading hyphen
    (i.e. the example makes the condition codes `-len' and `-ex'
    usable in `[[...]]' constructs)
  - an optional flag which for now can only be CONDF_INFIX; if this is 
    given, an infix operator is created (i.e. the above makes
    `[[ -len str ]]' and `[[ s1 -ex s2 ]]' available)
  - the C-function implementing the conditional
  - for non-infix condition codes the next two arguments give the
    minimum and maximum number of string the conditional can handle
    (i.e. `-len' can get one or two strings); as with builtins giving
    -1 as the maximum number means that the conditional accepts any
    number of strings
  - finally as the last argument an integer that is passed to the
    handler function that can be used to distinguish different
    condition codes if the same C-function implements more than one of 
    them

The definition for the function looks like:

  /**/
  static int
  cond_p_len(char **a, int id)
  {
    ...
  }

The first argument is an array containing the strings (NULL-terminated
like the array of arguments for builtins), the second argument is the
integer value stored in the table (the last argument to `CONDDEF(...)').
The value returned by the function should be non-zero if the condition 
is true and zero otherwise.

Note that no preprocessing is done on the strings. This means that
no substitutions are performed on them and that they will be
tokenized. There are three helper functions available:

  - char *cond_str(args, num, raw)
    The first argument is the array of strings the handler function
    got as an argument and the second one is an index into this array.
    The return value is the num'th string from the array with
    substitutions performed. If the last argument is zero, the string
    will also be untokenized.
  - long cond_val(args, num)
    The arguments are the same as for cond_str(). The return value is
    the result of the mathematical evaluation of the num'th string
    form the array.
  - int cond_match(args, num, str)
    Again, the first two arguments are the same as for the other
    functions. The third argument is any string. The result of the
    function is non-zero if the num'th string from the array taken 
    as a glob pattern matches the given string.

Parameters
----------

For defining parameters, a module can call `createparam()' directly or 
use a table to describe them, e.g.:

  static struct paramdef patab[] = {
    PARAMDEF("foo", PM_INTEGER, NULL, get_foo, set_foo, unset_foo),
    INTPARAMDEF("exint", &intparam),
    STRPARAMDEF("exstr", &strparam),
    ARRPARAMDEF("exarr", &arrparam),
  };

There are four macros used:

  - PARAMDEF() gets as arguments:
    - the name of the parameter
    - the parameter flags to set for it (from the PM_* flags defined
      in zsh.h)
    - optionally a pointer to a variable holding the value of the
      parameter
    - three functions that will be used to get the value of the
      parameter, store a value in the parameter, and unset the
      parameter
  - the other macros provide simple ways to define the most common
    types of parameters; they get the name of the parameter and a
    pointer to a variable holding the value as arguments; they are
    used to define integer-, scalar-, and array-parameters, so the
    variables whose addresses are given should be of type `long',
    `char *', and `char **', respectively

For a description of how to write functions for getting or setting the 
value of parameters, or how to write a function to unset a parameter,
see the description of the following functions in the `params.c' file:

  - `intvargetfn()' and `intvarsetfn()' for integer parameters
  - `strvargetfn()' and `strvarsetfn()' for scalar parameters
  - `arrvargetfn()' and `arrvarsetfn()' for array parameters
  - `stdunsetfn()' for unsetting parameters

Note that if one defines parameters using the last two macros (for
scalars and arrays), the variable holding the value should be
initialized to either `NULL' or to a piece of memory created with
`zalloc()'. But this memory should *not* be freed in the
finish-function of the module because that will be taken care of by
the `deleteparamdefs()' function described below.

It is also possible to declare special parameters using
the macro SPECIALPMDEF().  More care is required in this case.
See, for example, many of the definitions in Src/Modules/parameter.c.

Math functions
--------------

Modules can also define math functions. Again, they are described
using a table:

  static struct mathfunc mftab[] = {
    NUMMATHFUNC("sum", math_sum, 1, -1, 0),
    STRMATHFUNC("length", math_length, 0),
  };

The `NUMMATHFUNC()' macro defines a math function that gets an array
of mnumbers (the zsh type for representing values in arithmetic
expressions) taken from the string in parentheses at the function
call. Its arguments are the name of the function, the C-function
implementing it, the minimum and maximum number of arguments (as
usual, the later may be `-1' to specify that the function accepts any
number of arguments), and finally an integer that is given unchanged
to the C-function (to be able to implement multiple math functions in
one C-function).

The `STRMATHFUNC()' macro defines a math function that gets the string 
in parentheses at the call as one string argument (without the
parentheses). The arguments are the name of the function, the
C-function, and an integer used like the last argument of
`NUMMATHFUNC()'.

The C-functions implementing the math functions look like this:

  /**/
  static mnumber
  math_sum(char *name, int argc, mnumber *argv, int id)
  {
    ...
  }

  /**/
  static mnumber
  math_length(char *name, char *arg, int id)
  {
    ...
  }

Functions defined with `NUMMATHFUNC' get the name of the function, the 
number of numeric arguments, an array with these arguments, and the
last argument from the macro-call. Functions defined with
`STRMATHFUNC' get the name of the function, the string between the
parentheses at the call, and the last argument from the macro-call.

Both types of functions return an mnumber which is a discriminated
union looking like:

  typedef struct {
    union {
      zlong l;
      double d;
    } u;
    int type;
  } mnumber;

The `type' field should be set to `MN_INTEGER' or `MN_FLOAT' and
depending on its value either `u.l' or `u.d' contains the value.

Hooks
-----

Modules can also define function hooks. Other modules can then add
functions to these hooks to make the first module call these functions
instead of the default.  These are not handled by the features
mechanism as they are not directly visible to the user.

Again, an array is used to define hooks:

  static struct hookdef foohooks[] = {
    HOOKDEF("foo", foofunc, 0),
  };

The first argument of the macro is the name of the hook. This name
is used whenever the hook is used. The second argument is the default
function for the hook or NULL if no default function exists. The
last argument is used to define flags for the hook. Currently only one
such flag is defined: `HOOKF_ALL'. If this flag is given and more than
one function was added to the hook, all functions will be called
(including the default function). Otherwise only the last function
added will be called.

The functions that can be used as default functions or that can be
added to a hook have to be defined like:

  /**/
  static int
  foofunc(Hookdef h, void *data)
  {
    ...
  }

The first argument is a pointer to the struct defining the hook. The
second argument is an arbitrary pointer that is given to the function
used to invoke hooks (see below).

The functions to register and de-register hooks look like those for
the other things that can be defined by modules:

  /**/
  int
  boot_foo(Module m)
  {
    int ret;

    ret = addhookdefs(m->nam, foohooks, sizeof(foohooks)/sizeof(*foohooks))
    ...
  }
  ...
  /**/
  int
  cleanup_foo(Module m)
  {
    deletehookdefs(m->nam, foohooks, sizeof(foohooks)/sizeof(*foohooks));
    ...
  }

Modules that define hooks can invoke the function(s) registered for
them by calling the function `runhook(name, data)'. The first argument 
is the name of the hook and the second one is the pointer given to the 
hook functions as their second argument. Hooks that have the `HOOKF_ALL' 
flag call all function defined for them until one returns non-zero.
The return value of `runhook()' is the return value of the last hook
function called or zero if none was called.

To add a function to a hook, the function `addhookfunc(name, func)' is 
called with the name of the hook and a hook function as arguments.
Deleting them is done by calling `deletehookfunc(name, func)' with the 
same arguments as for the corresponding call to `addhookfunc()'.

Alternative forms of the last three function are provided for hooks
that are changed or called very often. These functions,
`runhookdef(def, data)', `addhookdeffunc(def, func)', and
`deletehookdeffunc(def, func)' get a pointer to the `hookdef'
structure defining the hook instead of the name and otherwise behave
like their counterparts.

Wrappers
--------

Finally, modules can define wrapper functions.  These functions are
called whenever a shell function is to be executed.  Again, they
are not handled by the features mechanism as they are not visible
to the user.

The definition is simple:

  static struct funcwrap wrapper[] = {
    WRAPDEF(ex_wrapper),
  };

The macro `WRAPDEF(...)' gets the C-function as its only argument.
This function should be defined like:

  /**/
  static int
  ex_wrapper(List list, FuncWrap w, char *name)
  {
    ...
    runshfunc(list, w, name);
    ...
    return 0;
  }

The first two arguments should only be used to pass them to
`runshfunc()' which will execute the shell function. The last argument 
is the name of the function to be executed. The arguments passed to
the function can be accessed vie the global variable `pparams' (a
NULL-terminated array of strings).

The return value of the wrapper function should be zero if it calls
`runshfunc()' itself and non-zero otherwise. This can be used for
wrapper functions that only need to run under certain conditions or
that don't need to clean anything up after the shell function has
finished:

  /**/
  static int
  ex_wrapper(List list, FuncWrap w, char *name)
  {
    if (wrapper_need_to_run) {
      ...
      runshfunc(list, w, name);
      ...
      return 0;
    }
    return 1;
  }

Inside these wrapper functions the global variable `sfcontext' will be 
set to a clue indicating the circumstances under which the shell
function was called. It can have any of the following values:

  - SFC_DIRECT:   the function was invoked directly by the user
  - SFC_SIGNAL:   the function was invoked as a signal handler
  - SFC_HOOK:     the function was automatically invoked as one of the
                  special functions known by the shell (like `chpwd')
  - SFC_WIDGET:   the function was called from the zsh line editor as a
                  user-defined widget
  - SFC_COMPLETE: the function was called from the completion code
                  (e.g. with `compctl -K func')

If a module invokes a shell function (e.g. as a hook function), the
value of this variable should only be changed temporarily and restored
to its previous value after the shell function has finished.

There is a problem when the user tries to unload a module that has
defined wrappers from a shell function. In this case the module can't
be unloaded immediately since the wrapper function is still on the
call stack. The zsh code delays unloading modules until all wrappers
from them have finished. To hide this from the user, the module's
cleanup function is run immediately so that all builtins, condition
codes, and wrapper function defined by the module are
de-registered. But if there is some module-global state that has to be 
finalized (e.g. some memory that has to be freed) and that is used by
the wrapper functions finalizing this data in the cleanup function
won't work.

This is why there are two functions each for the initialization and
finalization of modules. The `boot'- and `cleanup'-functions are run
whenever the user calls `zmodload' or `zmodload -u' and should only
register or de-register the module's interface that is visible to the
user. Anything else should be done in the `setup'- and
`finish'-functions. Otherwise modules that other modules depend upon
may destroy their state too early and wrapper functions in the latter
modules may stop working since the state they use is already destroyed.

Documentation
-------------

* Edit only the .yo files.  All other formats (man pages, TeXinfo, HTML,
  etc.) are automatically generated from the yodl source.

* Always use the correct markup.  em() is used for emphasis, and bf()
  for citations.  tt() marks text that is literal input to or output
  from the shell.  var() marks metasyntactic variables.

* In addition to appropriate markup, always use quotes (`') where
  appropriate.  Specifically, use quotes to mark text that is not a part
  of the actual text of the documentation (i.e., that it is being quoted).
  In principle, all combinations of quotes and markup are possible,
  because the purposes of the two devices are completely orthogonal.
  For example,

      Type `tt(xyzzy)' to let zsh know you have played tt(advent).
      Saying `plugh' aloud doesn't have much effect, however.

  In this case, "zsh" is normal text (a name), "advent" is a command name
  occurring in the main text, "plugh" is a normal word that is being quoted
  (it's the user that says `plugh', not the documentation), and "xyzzy"
  is some text to be typed literally that is being quoted.

* For multiple-line pieces of text that should not be filled, there are
  various models.
  - If the text is pure example, i.e. with no metasyntactic variables etc.,
    it should be included within `example(...)'.  The text will be
    indented, will not be filled and will be put into a fixed width font.
  - If the text includes mixed fonts, it should be included within
    `indent(...)'.  The text is now filled unless `nofill(...)' is also
    used, and explicit font-changing commands are required inside.
  - If the text appears inside some other format, such as for example the
    `item()' list structure, then the instruction `nofill(...)', which
    simply turns off filling should be used; as with `indent(...)',
    explicit font changing commands are required.  This can be used
    without `indent()' when no indentation is required, e.g. if the
    accumulated indentation would otherwise be too long.
  All the above should appear on their own, separated by newlines from the
  surrounding text.  No extra newlines after the opening or before the
  closing parenthesis are required.

Module names
------------

Modules have hierarchical names.  Name segments are separated by `/', and
each segment consists of alphanumerics plus `_'.  Relative names are never
used; the naming hierarchy is strictly for organisational convenience.

Top-level name segments should be organisational identifiers, assigned
by the Zsh Development Group and recorded here:

top-level identifier  organisation
--------------------  ------------
x_*                   reserved for private experimental use
zsh                   The Zsh Development Group (contact: <coordinator@zsh.org>)

Below the top level, naming authority is delegated.


Distribution of files
---------------------

zsh is distributed in two parts: a "src" distribution containing all
the source files (roughly, but not exactly, corresponding to the git
tree), and a "doc" distribution containing some pre-built files from
the documentation directory.  All the files in the "doc" distribution
may be generated from files in the "src" distribution with appropriate
freely available tools.

To indicate which files should be distributed, each directory in the git
tree includes a file .distfiles that sets any number of a set of Bourne
shell (scalar) parameters.  The value of the parameter is expanded as a
set of standard command line arguments.  Basic globbing is allowed in the
values.

The following parameters are currently used:

- DISTFILES_SRC is a list of files from the directory for the "src"
  distribution.

- DISTFILES_DOC is a list of files from the directory for the "doc"
  distribution.

- DISTFILES_NOT is a list of files that will not be included in a
  distribution, but that need to be present in the git tree.  This
  variable is not used by the zsh build process and is present for
  the convenience of external checks.


Use of Git
----------

zsh has migrated from CVS to git for version control.  We have so far
kept our workflow unchanged; to wit:

 1. change is developed and posted to the zsh-workers mailing list
 2. the zsh-workers list management software adds an X-Seq: header
 3. an entry is added to ChangeLog with details, including the X-Seq:
    header.
    [Open Question: should the first 6 or so characters of the commit
     fingerprint be included, so: "* 12345/deadbeef: frobbed the baz" ?]
 4. this is committed to git as a second commit
 5. this is pushed to the master server

Micro Git Tutorial:

  % $VISUAL file1.c file2.c new-file3.c
  % git add new-file3.c
  % git commit -a
  % git push

 "git commit -a" automatically finds files which are tracked and have
 been modified, but doesn't pick up new files; "git add" adds a file to
 the index to be part of the next commit, and can be used for new files
 or for existing files (commit -a is a shortcut for the latter)

 "git push" assumes that you're on the master branch and the repository
 was created by cloning it from some place, with default options.

Feature branch work:

  % git checkout -b feature_foo
  % $VISUAL path/to/files ...
  % git commit -a
  [lather, rinse, repeat]
  % git push origin feature_foo
  [ do mailing-list stuff here ]
  [ Switch back to master: ]
  % git checkout master
  [ and get the most recent changes: ]
  % git pull
  [ make the branch content now be relative to *new* master tip ]
  % git checkout feature_foo
  % git rebase master
  [ then bring in your changes: ]
  % git checkout master
  % git merge --ff-only feature_foo
  % $VISUAL ChangeLog
  % git commit -i ChangeLog
  % git push
  [ Cleanup: ]
  % git branch -d feature_foo
  % git push origin :feature_foo

Git further reading:
 * git help tutorial
 * git help tutorial-2
 * git help gitcore-tutorial
 * http://www-cs-students.stanford.edu/~blynn/gitmagic/