================================
Custom template tags and filters
================================

Django's template language comes with a wide variety of :doc:`built-in
tags and filters </ref/templates/builtins>` designed to address the
presentation logic needs of your application. Nevertheless, you may
find yourself needing functionality that is not covered by the core
set of template primitives. You can extend the template engine by
defining custom tags and filters using Python, and then make them
available to your templates using the :ttag:`{% load %}<load>` tag.

Code layout
===========

The most common place to specify custom template tags and filters is inside
a Django app. If they relate to an existing app, it makes sense to bundle them
there; otherwise, they can be added to a new app. When a Django app is added
to :setting:`INSTALLED_APPS`, any tags it defines in the conventional location
described below are automatically made available to load within templates.

The app should contain a ``templatetags`` directory, at the same level as
``models.py``, ``views.py``, etc. If this doesn't already exist, create it -
don't forget the ``__init__.py`` file to ensure the directory is treated as a
Python package.

.. admonition:: Development server won't automatically restart

    After adding the ``templatetags``  module, you will need to restart your
    server before you can use the tags or filters in templates.

Your custom tags and filters will live in a module inside the ``templatetags``
directory. The name of the module file is the name you'll use to load the tags
later, so be careful to pick a name that won't clash with custom tags and
filters in another app.

For example, if your custom tags/filters are in a file called
``poll_extras.py``, your app layout might look like this::

    polls/
        __init__.py
        models.py
        templatetags/
            __init__.py
            poll_extras.py
        views.py

And in your template you would use the following:

.. code-block:: html+django

    {% load poll_extras %}

The app that contains the custom tags must be in :setting:`INSTALLED_APPS` in
order for the :ttag:`{% load %}<load>` tag to work. This is a security feature:
It allows you to host Python code for many template libraries on a single host
machine without enabling access to all of them for every Django installation.

There's no limit on how many modules you put in the ``templatetags`` package.
Just keep in mind that a :ttag:`{% load %}<load>` statement will load
tags/filters for the given Python module name, not the name of the app.

To be a valid tag library, the module must contain a module-level variable
named ``register`` that is a ``template.Library`` instance, in which all the
tags and filters are registered. So, near the top of your module, put the
following::

    from django import template

    register = template.Library()

Alternatively, template tag modules can be registered through the
``'libraries'`` argument to
:class:`~django.template.backends.django.DjangoTemplates`. This is useful if
you want to use a different label from the template tag module name when
loading template tags. It also enables you to register tags without installing
an application.

.. admonition:: Behind the scenes

    For a ton of examples, read the source code for Django's default filters
    and tags. They're in ``django/template/defaultfilters.py`` and
    ``django/template/defaulttags.py``, respectively.

    For more information on the :ttag:`load` tag, read its documentation.

.. _howto-writing-custom-template-filters:

Writing custom template filters
===============================

Custom filters are Python functions that take one or two arguments:

* The value of the variable (input) -- not necessarily a string.
* The value of the argument -- this can have a default value, or be left
  out altogether.

For example, in the filter ``{{ var|foo:"bar" }}``, the filter ``foo`` would be
passed the variable ``var`` and the argument ``"bar"``.

Since the template language doesn't provide exception handling, any exception
raised from a template filter will be exposed as a server error. Thus, filter
functions should avoid raising exceptions if there is a reasonable fallback
value to return. In case of input that represents a clear bug in a template,
raising an exception may still be better than silent failure which hides the
bug.

Here's an example filter definition::

    def cut(value, arg):
        """Removes all values of arg from the given string"""
        return value.replace(arg, '')

And here's an example of how that filter would be used:

.. code-block:: html+django

    {{ somevariable|cut:"0" }}

Most filters don't take arguments. In this case, leave the argument out of your
function::

    def lower(value): # Only one argument.
        """Converts a string into all lowercase"""
        return value.lower()

Registering custom filters
--------------------------

.. method:: django.template.Library.filter()

Once you've written your filter definition, you need to register it with
your ``Library`` instance, to make it available to Django's template language::

    register.filter('cut', cut)
    register.filter('lower', lower)

The ``Library.filter()`` method takes two arguments:

1. The name of the filter -- a string.
2. The compilation function -- a Python function (not the name of the
   function as a string).

You can use ``register.filter()`` as a decorator instead::

    @register.filter(name='cut')
    def cut(value, arg):
        return value.replace(arg, '')

    @register.filter
    def lower(value):
        return value.lower()

If you leave off the ``name`` argument, as in the second example above, Django
will use the function's name as the filter name.

Finally, ``register.filter()`` also accepts three keyword arguments,
``is_safe``, ``needs_autoescape``, and ``expects_localtime``. These arguments
are described in :ref:`filters and auto-escaping <filters-auto-escaping>` and
:ref:`filters and time zones <filters-timezones>` below.

Template filters that expect strings
------------------------------------

.. method:: django.template.defaultfilters.stringfilter()

If you're writing a template filter that only expects a string as the first
argument, you should use the decorator ``stringfilter``. This will
convert an object to its string value before being passed to your function::

    from django import template
    from django.template.defaultfilters import stringfilter

    register = template.Library()

    @register.filter
    @stringfilter
    def lower(value):
        return value.lower()

This way, you'll be able to pass, say, an integer to this filter, and it
won't cause an ``AttributeError`` (because integers don't have ``lower()``
methods).

.. _filters-auto-escaping:

Filters and auto-escaping
-------------------------

When writing a custom filter, give some thought to how the filter will interact
with Django's auto-escaping behavior. Note that two types of strings can be
passed around inside the template code:

* **Raw strings** are the native Python strings. On output, they're escaped if
  auto-escaping is in effect and presented unchanged, otherwise.

* **Safe strings** are strings that have been marked safe from further
  escaping at output time. Any necessary escaping has already been done.
  They're commonly used for output that contains raw HTML that is intended
  to be interpreted as-is on the client side.

  Internally, these strings are of type
  :class:`~django.utils.safestring.SafeString`. You can test for them
  using code like::

      from django.utils.safestring import SafeString

      if isinstance(value, SafeString):
          # Do something with the "safe" string.
          ...

Template filter code falls into one of two situations:

1. Your filter does not introduce any HTML-unsafe characters (``<``, ``>``,
   ``'``, ``"`` or ``&``) into the result that were not already present. In
   this case, you can let Django take care of all the auto-escaping
   handling for you. All you need to do is set the ``is_safe`` flag to ``True``
   when you register your filter function, like so::

       @register.filter(is_safe=True)
       def myfilter(value):
           return value

   This flag tells Django that if a "safe" string is passed into your
   filter, the result will still be "safe" and if a non-safe string is
   passed in, Django will automatically escape it, if necessary.

   You can think of this as meaning "this filter is safe -- it doesn't
   introduce any possibility of unsafe HTML."

   The reason ``is_safe`` is necessary is because there are plenty of
   normal string operations that will turn a ``SafeData`` object back into
   a normal ``str`` object and, rather than try to catch them all, which would
   be very difficult, Django repairs the damage after the filter has completed.

   For example, suppose you have a filter that adds the string ``xx`` to
   the end of any input. Since this introduces no dangerous HTML characters
   to the result (aside from any that were already present), you should
   mark your filter with ``is_safe``::

       @register.filter(is_safe=True)
       def add_xx(value):
           return '%sxx' % value

   When this filter is used in a template where auto-escaping is enabled,
   Django will escape the output whenever the input is not already marked
   as "safe".

   By default, ``is_safe`` is ``False``, and you can omit it from any filters
   where it isn't required.

   Be careful when deciding if your filter really does leave safe strings
   as safe. If you're *removing* characters, you might inadvertently leave
   unbalanced HTML tags or entities in the result. For example, removing a
   ``>`` from the input might turn ``<a>`` into ``<a``, which would need to
   be escaped on output to avoid causing problems. Similarly, removing a
   semicolon (``;``) can turn ``&amp;`` into ``&amp``, which is no longer a
   valid entity and thus needs further escaping. Most cases won't be nearly
   this tricky, but keep an eye out for any problems like that when
   reviewing your code.

   Marking a filter ``is_safe`` will coerce the filter's return value to
   a string.  If your filter should return a boolean or other non-string
   value, marking it ``is_safe`` will probably have unintended
   consequences (such as converting a boolean False to the string
   'False').

2. Alternatively, your filter code can manually take care of any necessary
   escaping. This is necessary when you're introducing new HTML markup into
   the result. You want to mark the output as safe from further
   escaping so that your HTML markup isn't escaped further, so you'll need
   to handle the input yourself.

   To mark the output as a safe string, use
   :func:`django.utils.safestring.mark_safe`.

   Be careful, though. You need to do more than just mark the output as
   safe. You need to ensure it really *is* safe, and what you do depends on
   whether auto-escaping is in effect. The idea is to write filters that
   can operate in templates where auto-escaping is either on or off in
   order to make things easier for your template authors.

   In order for your filter to know the current auto-escaping state, set the
   ``needs_autoescape`` flag to ``True`` when you register your filter function.
   (If you don't specify this flag, it defaults to ``False``). This flag tells
   Django that your filter function wants to be passed an extra keyword
   argument, called ``autoescape``, that is ``True`` if auto-escaping is in
   effect and ``False`` otherwise. It is recommended to set the default of the
   ``autoescape`` parameter to ``True``, so that if you call the function
   from Python code it will have escaping enabled by default.

   For example, let's write a filter that emphasizes the first character of
   a string::

      from django import template
      from django.utils.html import conditional_escape
      from django.utils.safestring import mark_safe

      register = template.Library()

      @register.filter(needs_autoescape=True)
      def initial_letter_filter(text, autoescape=True):
          first, other = text[0], text[1:]
          if autoescape:
              esc = conditional_escape
          else:
              esc = lambda x: x
          result = '<strong>%s</strong>%s' % (esc(first), esc(other))
          return mark_safe(result)

   The ``needs_autoescape`` flag and the ``autoescape`` keyword argument mean
   that our function will know whether automatic escaping is in effect when the
   filter is called. We use ``autoescape`` to decide whether the input data
   needs to be passed through ``django.utils.html.conditional_escape`` or not.
   (In the latter case, we use the identity function as the "escape" function.)
   The ``conditional_escape()`` function is like ``escape()`` except it only
   escapes input that is **not** a ``SafeData`` instance. If a ``SafeData``
   instance is passed to ``conditional_escape()``, the data is returned
   unchanged.

   Finally, in the above example, we remember to mark the result as safe
   so that our HTML is inserted directly into the template without further
   escaping.

   There's no need to worry about the ``is_safe`` flag in this case
   (although including it wouldn't hurt anything). Whenever you manually
   handle the auto-escaping issues and return a safe string, the
   ``is_safe`` flag won't change anything either way.

.. warning:: Avoiding XSS vulnerabilities when reusing built-in filters

    Django's built-in filters have ``autoescape=True`` by default in order to
    get the proper autoescaping behavior and avoid a cross-site script
    vulnerability.

    In older versions of Django, be careful when reusing Django's built-in
    filters as ``autoescape`` defaults to ``None``. You'll need to pass
    ``autoescape=True`` to get autoescaping.

    For example, if you wanted to write a custom filter called
    ``urlize_and_linebreaks`` that combined the :tfilter:`urlize` and
    :tfilter:`linebreaksbr` filters, the filter would look like::

        from django.template.defaultfilters import linebreaksbr, urlize

        @register.filter(needs_autoescape=True)
        def urlize_and_linebreaks(text, autoescape=True):
            return linebreaksbr(
                urlize(text, autoescape=autoescape),
                autoescape=autoescape
            )

    Then:

    .. code-block:: html+django

        {{ comment|urlize_and_linebreaks }}

    would be equivalent to:

    .. code-block:: html+django

        {{ comment|urlize|linebreaksbr }}

.. _filters-timezones:

Filters and time zones
----------------------

If you write a custom filter that operates on :class:`~datetime.datetime`
objects, you'll usually register it with the ``expects_localtime`` flag set to
``True``::

    @register.filter(expects_localtime=True)
    def businesshours(value):
        try:
            return 9 <= value.hour < 17
        except AttributeError:
            return ''

When this flag is set, if the first argument to your filter is a time zone
aware datetime, Django will convert it to the current time zone before passing
it to your filter when appropriate, according to :ref:`rules for time zones
conversions in templates <time-zones-in-templates>`.

.. _howto-writing-custom-template-tags:

Writing custom template tags
============================

Tags are more complex than filters, because tags can do anything. Django
provides a number of shortcuts that make writing most types of tags easier.
First we'll explore those shortcuts, then explain how to write a tag from
scratch for those cases when the shortcuts aren't powerful enough.

.. _howto-custom-template-tags-simple-tags:

Simple tags
-----------

.. method:: django.template.Library.simple_tag()

Many template tags take a number of arguments -- strings or template variables
-- and return a result after doing some processing based solely on
the input arguments and some external information. For example, a
``current_time`` tag might accept a format string and return the time as a
string formatted accordingly.

To ease the creation of these types of tags, Django provides a helper function,
``simple_tag``. This function, which is a method of
``django.template.Library``, takes a function that accepts any number of
arguments, wraps it in a ``render`` function and the other necessary bits
mentioned above and registers it with the template system.

Our ``current_time`` function could thus be written like this::

    import datetime
    from django import template

    register = template.Library()

    @register.simple_tag
    def current_time(format_string):
        return datetime.datetime.now().strftime(format_string)

A few things to note about the ``simple_tag`` helper function:

* Checking for the required number of arguments, etc., has already been
  done by the time our function is called, so we don't need to do that.
* The quotes around the argument (if any) have already been stripped away,
  so we receive a plain string.
* If the argument was a template variable, our function is passed the
  current value of the variable, not the variable itself.

Unlike other tag utilities, ``simple_tag`` passes its output through
:func:`~django.utils.html.conditional_escape` if the template context is in
autoescape mode, to ensure correct HTML and protect you from XSS
vulnerabilities.

If additional escaping is not desired, you will need to use
:func:`~django.utils.safestring.mark_safe` if you are absolutely sure that your
code does not contain XSS vulnerabilities. For building small HTML snippets,
use of :func:`~django.utils.html.format_html` instead of ``mark_safe()`` is
strongly recommended.

If your template tag needs to access the current context, you can use the
``takes_context`` argument when registering your tag::

    @register.simple_tag(takes_context=True)
    def current_time(context, format_string):
        timezone = context['timezone']
        return your_get_current_time_method(timezone, format_string)

Note that the first argument *must* be called ``context``.

For more information on how the ``takes_context`` option works, see the section
on :ref:`inclusion tags<howto-custom-template-tags-inclusion-tags>`.

If you need to rename your tag, you can provide a custom name for it::

    register.simple_tag(lambda x: x - 1, name='minusone')

    @register.simple_tag(name='minustwo')
    def some_function(value):
        return value - 2

``simple_tag`` functions may accept any number of positional or keyword
arguments. For example::

    @register.simple_tag
    def my_tag(a, b, *args, **kwargs):
        warning = kwargs['warning']
        profile = kwargs['profile']
        ...
        return ...

Then in the template any number of arguments, separated by spaces, may be
passed to the template tag. Like in Python, the values for keyword arguments
are set using the equal sign ("``=``") and must be provided after the
positional arguments. For example:

.. code-block:: html+django

    {% my_tag 123 "abcd" book.title warning=message|lower profile=user.profile %}

It's possible to store the tag results in a template variable rather than
directly outputting it. This is done by using the ``as`` argument followed by
the variable name. Doing so enables you to output the content yourself where
you see fit:

.. code-block:: html+django

    {% current_time "%Y-%m-%d %I:%M %p" as the_time %}
    <p>The time is {{ the_time }}.</p>

.. _howto-custom-template-tags-inclusion-tags:

Inclusion tags
--------------

.. method:: django.template.Library.inclusion_tag()

Another common type of template tag is the type that displays some data by
rendering *another* template. For example, Django's admin interface uses custom
template tags to display the buttons along the bottom of the "add/change" form
pages. Those buttons always look the same, but the link targets change
depending on the object being edited -- so they're a perfect case for using a
small template that is filled with details from the current object. (In the
admin's case, this is the ``submit_row`` tag.)

These sorts of tags are called "inclusion tags".

Writing inclusion tags is probably best demonstrated by example. Let's write a
tag that outputs a list of choices for a given ``Poll`` object, such as was
created in the :ref:`tutorials <creating-models>`. We'll use the tag like this:

.. code-block:: html+django

    {% show_results poll %}

...and the output will be something like this:

.. code-block:: html

    <ul>
      <li>First choice</li>
      <li>Second choice</li>
      <li>Third choice</li>
    </ul>

First, define the function that takes the argument and produces a dictionary of
data for the result. The important point here is we only need to return a
dictionary, not anything more complex. This will be used as a template context
for the template fragment. Example::

    def show_results(poll):
        choices = poll.choice_set.all()
        return {'choices': choices}

Next, create the template used to render the tag's output. This template is a
fixed feature of the tag: the tag writer specifies it, not the template
designer. Following our example, the template is very short:

.. code-block:: html+django

    <ul>
    {% for choice in choices %}
        <li> {{ choice }} </li>
    {% endfor %}
    </ul>

Now, create and register the inclusion tag by calling the ``inclusion_tag()``
method on a ``Library`` object. Following our example, if the above template is
in a file called ``results.html`` in a directory that's searched by the
template loader, we'd register the tag like this::

    # Here, register is a django.template.Library instance, as before
    @register.inclusion_tag('results.html')
    def show_results(poll):
        ...

Alternatively it is possible to register the inclusion tag using a
:class:`django.template.Template` instance::

    from django.template.loader import get_template
    t = get_template('results.html')
    register.inclusion_tag(t)(show_results)

...when first creating the function.

Sometimes, your inclusion tags might require a large number of arguments,
making it a pain for template authors to pass in all the arguments and remember
their order. To solve this, Django provides a ``takes_context`` option for
inclusion tags. If you specify ``takes_context`` in creating a template tag,
the tag will have no required arguments, and the underlying Python function
will have one argument -- the template context as of when the tag was called.

For example, say you're writing an inclusion tag that will always be used in a
context that contains ``home_link`` and ``home_title`` variables that point
back to the main page. Here's what the Python function would look like::

    @register.inclusion_tag('link.html', takes_context=True)
    def jump_link(context):
        return {
            'link': context['home_link'],
            'title': context['home_title'],
        }

Note that the first parameter to the function *must* be called ``context``.

In that ``register.inclusion_tag()`` line, we specified ``takes_context=True``
and the name of the template. Here's what the template ``link.html`` might look
like:

.. code-block:: html+django

    Jump directly to <a href="{{ link }}">{{ title }}</a>.

Then, any time you want to use that custom tag, load its library and call it
without any arguments, like so:

.. code-block:: html+django

    {% jump_link %}

Note that when you're using ``takes_context=True``, there's no need to pass
arguments to the template tag. It automatically gets access to the context.

The ``takes_context`` parameter defaults to ``False``. When it's set to
``True``, the tag is passed the context object, as in this example. That's the
only difference between this case and the previous ``inclusion_tag`` example.

``inclusion_tag`` functions may accept any number of positional or keyword
arguments. For example::

    @register.inclusion_tag('my_template.html')
    def my_tag(a, b, *args, **kwargs):
        warning = kwargs['warning']
        profile = kwargs['profile']
        ...
        return ...

Then in the template any number of arguments, separated by spaces, may be
passed to the template tag. Like in Python, the values for keyword arguments
are set using the equal sign ("``=``") and must be provided after the
positional arguments. For example:

.. code-block:: html+django

    {% my_tag 123 "abcd" book.title warning=message|lower profile=user.profile %}

Advanced custom template tags
-----------------------------

Sometimes the basic features for custom template tag creation aren't enough.
Don't worry, Django gives you complete access to the internals required to build
a template tag from the ground up.

A quick overview
----------------

The template system works in a two-step process: compiling and rendering. To
define a custom template tag, you specify how the compilation works and how
the rendering works.

When Django compiles a template, it splits the raw template text into
''nodes''. Each node is an instance of ``django.template.Node`` and has
a ``render()`` method. A compiled template is a list of ``Node`` objects. When
you call ``render()`` on a compiled template object, the template calls
``render()`` on each ``Node`` in its node list, with the given context.  The
results are all concatenated together to form the output of the template.

Thus, to define a custom template tag, you specify how the raw template tag is
converted into a ``Node`` (the compilation function), and what the node's
``render()`` method does.

Writing the compilation function
--------------------------------

For each template tag the template parser encounters, it calls a Python
function with the tag contents and the parser object itself. This function is
responsible for returning a ``Node`` instance based on the contents of the tag.

For example, let's write a full implementation of our template tag,
``{% current_time %}``, that displays the current date/time, formatted according
to a parameter given in the tag, in :func:`~time.strftime` syntax. It's a good
idea to decide the tag syntax before anything else. In our case, let's say the
tag should be used like this:

.. code-block:: html+django

    <p>The time is {% current_time "%Y-%m-%d %I:%M %p" %}.</p>

The parser for this function should grab the parameter and create a ``Node``
object::

    from django import template

    def do_current_time(parser, token):
        try:
            # split_contents() knows not to split quoted strings.
            tag_name, format_string = token.split_contents()
        except ValueError:
            raise template.TemplateSyntaxError(
                "%r tag requires a single argument" % token.contents.split()[0]
            )
        if not (format_string[0] == format_string[-1] and format_string[0] in ('"', "'")):
            raise template.TemplateSyntaxError(
                "%r tag's argument should be in quotes" % tag_name
            )
        return CurrentTimeNode(format_string[1:-1])

Notes:

* ``parser`` is the template parser object. We don't need it in this
  example.

* ``token.contents`` is a string of the raw contents of the tag. In our
  example, it's ``'current_time "%Y-%m-%d %I:%M %p"'``.

* The ``token.split_contents()`` method separates the arguments on spaces
  while keeping quoted strings together. The more straightforward
  ``token.contents.split()`` wouldn't be as robust, as it would naively
  split on *all* spaces, including those within quoted strings. It's a good
  idea to always use ``token.split_contents()``.

* This function is responsible for raising
  ``django.template.TemplateSyntaxError``, with helpful messages, for
  any syntax error.

* The ``TemplateSyntaxError`` exceptions use the ``tag_name`` variable.
  Don't hard-code the tag's name in your error messages, because that
  couples the tag's name to your function. ``token.contents.split()[0]``
  will ''always'' be the name of your tag -- even when the tag has no
  arguments.

* The function returns a ``CurrentTimeNode`` with everything the node needs
  to know about this tag. In this case, it passes the argument --
  ``"%Y-%m-%d %I:%M %p"``. The leading and trailing quotes from the
  template tag are removed in ``format_string[1:-1]``.

* The parsing is very low-level. The Django developers have experimented
  with writing small frameworks on top of this parsing system, using
  techniques such as EBNF grammars, but those experiments made the template
  engine too slow. It's low-level because that's fastest.

Writing the renderer
--------------------

The second step in writing custom tags is to define a ``Node`` subclass that
has a ``render()`` method.

Continuing the above example, we need to define ``CurrentTimeNode``::

    import datetime
    from django import template

    class CurrentTimeNode(template.Node):
        def __init__(self, format_string):
            self.format_string = format_string

        def render(self, context):
            return datetime.datetime.now().strftime(self.format_string)

Notes:

* ``__init__()`` gets the ``format_string`` from ``do_current_time()``.
  Always pass any options/parameters/arguments to a ``Node`` via its
  ``__init__()``.

* The ``render()`` method is where the work actually happens.

* ``render()`` should generally fail silently, particularly in a production
  environment. In some cases however, particularly if
  ``context.template.engine.debug`` is ``True``, this method may raise an
  exception to make debugging easier. For example, several core tags raise
  ``django.template.TemplateSyntaxError`` if they receive the wrong number or
  type of arguments.

Ultimately, this decoupling of compilation and rendering results in an
efficient template system, because a template can render multiple contexts
without having to be parsed multiple times.

.. _tags-auto-escaping:

Auto-escaping considerations
----------------------------

The output from template tags is **not** automatically run through the
auto-escaping filters (with the exception of
:meth:`~django.template.Library.simple_tag` as described above). However, there
are still a couple of things you should keep in mind when writing a template
tag.

If the ``render()`` method of your template tag stores the result in a context
variable (rather than returning the result in a string), it should take care
to call ``mark_safe()`` if appropriate. When the variable is ultimately
rendered, it will be affected by the auto-escape setting in effect at the
time, so content that should be safe from further escaping needs to be marked
as such.

Also, if your template tag creates a new context for performing some
sub-rendering, set the auto-escape attribute to the current context's value.
The ``__init__`` method for the ``Context`` class takes a parameter called
``autoescape`` that you can use for this purpose. For example::

    from django.template import Context

    def render(self, context):
        # ...
        new_context = Context({'var': obj}, autoescape=context.autoescape)
        # ... Do something with new_context ...

This is not a very common situation, but it's useful if you're rendering a
template yourself. For example::

    def render(self, context):
        t = context.template.engine.get_template('small_fragment.html')
        return t.render(Context({'var': obj}, autoescape=context.autoescape))

If we had neglected to pass in the current ``context.autoescape`` value to our
new ``Context`` in this example, the results would have *always* been
automatically escaped, which may not be the desired behavior if the template
tag is used inside a :ttag:`{% autoescape off %}<autoescape>` block.

.. _template_tag_thread_safety:

Thread-safety considerations
----------------------------

Once a node is parsed, its ``render`` method may be called any number of times.
Since Django is sometimes run in multi-threaded environments, a single node may
be simultaneously rendering with different contexts in response to two separate
requests. Therefore, it's important to make sure your template tags are thread
safe.

To make sure your template tags are thread safe, you should never store state
information on the node itself. For example, Django provides a builtin
:ttag:`cycle` template tag that cycles among a list of given strings each time
it's rendered:

.. code-block:: html+django

    {% for o in some_list %}
        <tr class="{% cycle 'row1' 'row2' %}">
            ...
        </tr>
    {% endfor %}

A naive implementation of ``CycleNode`` might look something like this::

    import itertools
    from django import template

    class CycleNode(template.Node):
        def __init__(self, cyclevars):
            self.cycle_iter = itertools.cycle(cyclevars)

        def render(self, context):
            return next(self.cycle_iter)

But, suppose we have two templates rendering the template snippet from above at
the same time:

#. Thread 1 performs its first loop iteration, ``CycleNode.render()``
   returns 'row1'
#. Thread 2 performs its first loop iteration, ``CycleNode.render()``
   returns 'row2'
#. Thread 1 performs its second loop iteration, ``CycleNode.render()``
   returns 'row1'
#. Thread 2 performs its second loop iteration, ``CycleNode.render()``
   returns 'row2'

The CycleNode is iterating, but it's iterating globally. As far as Thread 1
and Thread 2 are concerned, it's always returning the same value. This is
not what we want!

To address this problem, Django provides a ``render_context`` that's associated
with the ``context`` of the template that is currently being rendered. The
``render_context`` behaves like a Python dictionary, and should be used to
store ``Node`` state between invocations of the ``render`` method.

Let's refactor our ``CycleNode`` implementation to use the ``render_context``::

    class CycleNode(template.Node):
        def __init__(self, cyclevars):
            self.cyclevars = cyclevars

        def render(self, context):
            if self not in context.render_context:
                context.render_context[self] = itertools.cycle(self.cyclevars)
            cycle_iter = context.render_context[self]
            return next(cycle_iter)

Note that it's perfectly safe to store global information that will not change
throughout the life of the ``Node`` as an attribute. In the case of
``CycleNode``, the ``cyclevars`` argument doesn't change after the ``Node`` is
instantiated, so we don't need to put it in the ``render_context``. But state
information that is specific to the template that is currently being rendered,
like the current iteration of the ``CycleNode``, should be stored in the
``render_context``.

.. note::
    Notice how we used ``self`` to scope the ``CycleNode`` specific information
    within the ``render_context``. There may be multiple ``CycleNodes`` in a
    given template, so we need to be careful not to clobber another node's
    state information. The easiest way to do this is to always use ``self`` as
    the key into ``render_context``. If you're keeping track of several state
    variables, make ``render_context[self]`` a dictionary.

Registering the tag
-------------------

Finally, register the tag with your module's ``Library`` instance, as explained
in :ref:`writing custom template tags<howto-writing-custom-template-tags>`
above. Example::

    register.tag('current_time', do_current_time)

The ``tag()`` method takes two arguments:

1. The name of the template tag -- a string. If this is left out, the
   name of the compilation function will be used.
2. The compilation function -- a Python function (not the name of the
   function as a string).

As with filter registration, it is also possible to use this as a decorator::

    @register.tag(name="current_time")
    def do_current_time(parser, token):
        ...

    @register.tag
    def shout(parser, token):
        ...

If you leave off the ``name`` argument, as in the second example above, Django
will use the function's name as the tag name.

Passing template variables to the tag
-------------------------------------

Although you can pass any number of arguments to a template tag using
``token.split_contents()``, the arguments are all unpacked as
string literals. A little more work is required in order to pass dynamic
content (a template variable) to a template tag as an argument.

While the previous examples have formatted the current time into a string and
returned the string, suppose you wanted to pass in a
:class:`~django.db.models.DateTimeField` from an object and have the template
tag format that date-time:

.. code-block:: html+django

    <p>This post was last updated at {% format_time blog_entry.date_updated "%Y-%m-%d %I:%M %p" %}.</p>

Initially, ``token.split_contents()`` will return three values:

1. The tag name ``format_time``.
2. The string ``'blog_entry.date_updated'`` (without the surrounding
   quotes).
3. The formatting string ``'"%Y-%m-%d %I:%M %p"'``. The return value from
   ``split_contents()`` will include the leading and trailing quotes for
   string literals like this.

Now your tag should begin to look like this::

    from django import template

    def do_format_time(parser, token):
        try:
            # split_contents() knows not to split quoted strings.
            tag_name, date_to_be_formatted, format_string = token.split_contents()
        except ValueError:
            raise template.TemplateSyntaxError(
                "%r tag requires exactly two arguments" % token.contents.split()[0]
            )
        if not (format_string[0] == format_string[-1] and format_string[0] in ('"', "'")):
            raise template.TemplateSyntaxError(
                "%r tag's argument should be in quotes" % tag_name
            )
        return FormatTimeNode(date_to_be_formatted, format_string[1:-1])

You also have to change the renderer to retrieve the actual contents of the
``date_updated`` property of the ``blog_entry`` object.  This can be
accomplished by using the ``Variable()`` class in ``django.template``.

To use the ``Variable`` class, instantiate it with the name of the variable to
be resolved, and then call ``variable.resolve(context)``. So, for example::

    class FormatTimeNode(template.Node):
        def __init__(self, date_to_be_formatted, format_string):
            self.date_to_be_formatted = template.Variable(date_to_be_formatted)
            self.format_string = format_string

        def render(self, context):
            try:
                actual_date = self.date_to_be_formatted.resolve(context)
                return actual_date.strftime(self.format_string)
            except template.VariableDoesNotExist:
                return ''

Variable resolution will throw a ``VariableDoesNotExist`` exception if it
cannot resolve the string passed to it in the current context of the page.

Setting a variable in the context
---------------------------------

The above examples output a value. Generally, it's more flexible if your
template tags set template variables instead of outputting values. That way,
template authors can reuse the values that your template tags create.

To set a variable in the context, use dictionary assignment on the context
object in the ``render()`` method. Here's an updated version of
``CurrentTimeNode`` that sets a template variable ``current_time`` instead of
outputting it::

    import datetime
    from django import template

    class CurrentTimeNode2(template.Node):
        def __init__(self, format_string):
            self.format_string = format_string
        def render(self, context):
            context['current_time'] = datetime.datetime.now().strftime(self.format_string)
            return ''

Note that ``render()`` returns the empty string. ``render()`` should always
return string output. If all the template tag does is set a variable,
``render()`` should return the empty string.

Here's how you'd use this new version of the tag:

.. code-block:: html+django

    {% current_time "%Y-%m-%d %I:%M %p" %}<p>The time is {{ current_time }}.</p>

.. admonition:: Variable scope in context

    Any variable set in the context will only be available in the same
    ``block`` of the template in which it was assigned. This behavior is
    intentional; it provides a scope for variables so that they don't conflict
    with context in other blocks.

But, there's a problem with ``CurrentTimeNode2``: The variable name
``current_time`` is hard-coded. This means you'll need to make sure your
template doesn't use ``{{ current_time }}`` anywhere else, because the
``{% current_time %}`` will blindly overwrite that variable's value. A cleaner
solution is to make the template tag specify the name of the output variable,
like so:

.. code-block:: html+django

    {% current_time "%Y-%m-%d %I:%M %p" as my_current_time %}
    <p>The current time is {{ my_current_time }}.</p>

To do that, you'll need to refactor both the compilation function and ``Node``
class, like so::

    import re

    class CurrentTimeNode3(template.Node):
        def __init__(self, format_string, var_name):
            self.format_string = format_string
            self.var_name = var_name
        def render(self, context):
            context[self.var_name] = datetime.datetime.now().strftime(self.format_string)
            return ''

    def do_current_time(parser, token):
        # This version uses a regular expression to parse tag contents.
        try:
            # Splitting by None == splitting by spaces.
            tag_name, arg = token.contents.split(None, 1)
        except ValueError:
            raise template.TemplateSyntaxError(
                "%r tag requires arguments" % token.contents.split()[0]
            )
        m = re.search(r'(.*?) as (\w+)', arg)
        if not m:
            raise template.TemplateSyntaxError("%r tag had invalid arguments" % tag_name)
        format_string, var_name = m.groups()
        if not (format_string[0] == format_string[-1] and format_string[0] in ('"', "'")):
            raise template.TemplateSyntaxError(
                "%r tag's argument should be in quotes" % tag_name
            )
        return CurrentTimeNode3(format_string[1:-1], var_name)

The difference here is that ``do_current_time()`` grabs the format string and
the variable name, passing both to ``CurrentTimeNode3``.

Finally, if you only need to have a simple syntax for your custom
context-updating template tag, consider using the
:meth:`~django.template.Library.simple_tag` shortcut, which supports assigning
the tag results to a template variable.

Parsing until another block tag
-------------------------------

Template tags can work in tandem. For instance, the standard
:ttag:`{% comment %}<comment>` tag hides everything until ``{% endcomment %}``.
To create a template tag such as this, use ``parser.parse()`` in your
compilation function.

Here's how a simplified ``{% comment %}`` tag might be implemented::

    def do_comment(parser, token):
        nodelist = parser.parse(('endcomment',))
        parser.delete_first_token()
        return CommentNode()

    class CommentNode(template.Node):
        def render(self, context):
            return ''

.. note::
    The actual implementation of :ttag:`{% comment %}<comment>` is slightly
    different in that it allows broken template tags to appear between
    ``{% comment %}`` and ``{% endcomment %}``. It does so by calling
    ``parser.skip_past('endcomment')`` instead of ``parser.parse(('endcomment',))``
    followed by ``parser.delete_first_token()``, thus avoiding the generation of a
    node list.

``parser.parse()`` takes a tuple of names of block tags ''to parse until''. It
returns an instance of ``django.template.NodeList``, which is a list of
all ``Node`` objects that the parser encountered ''before'' it encountered
any of the tags named in the tuple.

In ``"nodelist = parser.parse(('endcomment',))"`` in the above example,
``nodelist`` is a list of all nodes between the ``{% comment %}`` and
``{% endcomment %}``, not counting ``{% comment %}`` and ``{% endcomment %}``
themselves.

After ``parser.parse()`` is called, the parser hasn't yet "consumed" the
``{% endcomment %}`` tag, so the code needs to explicitly call
``parser.delete_first_token()``.

``CommentNode.render()`` returns an empty string. Anything between
``{% comment %}`` and ``{% endcomment %}`` is ignored.

Parsing until another block tag, and saving contents
----------------------------------------------------

In the previous example, ``do_comment()`` discarded everything between
``{% comment %}`` and ``{% endcomment %}``. Instead of doing that, it's
possible to do something with the code between block tags.

For example, here's a custom template tag, ``{% upper %}``, that capitalizes
everything between itself and ``{% endupper %}``.

Usage:

.. code-block:: html+django

    {% upper %}This will appear in uppercase, {{ your_name }}.{% endupper %}

As in the previous example, we'll use ``parser.parse()``. But this time, we
pass the resulting ``nodelist`` to the ``Node``::

    def do_upper(parser, token):
        nodelist = parser.parse(('endupper',))
        parser.delete_first_token()
        return UpperNode(nodelist)

    class UpperNode(template.Node):
        def __init__(self, nodelist):
            self.nodelist = nodelist
        def render(self, context):
            output = self.nodelist.render(context)
            return output.upper()

The only new concept here is the ``self.nodelist.render(context)`` in
``UpperNode.render()``.

For more examples of complex rendering, see the source code of
:ttag:`{% for %}<for>` in ``django/template/defaulttags.py`` and
:ttag:`{% if %}<if>` in ``django/template/smartif.py``.