Tutorial: Writing a simple extension

This section is intended as a walkthrough for the creation of custom extensions. It covers the basics of writing and activating an extension, as well as commonly used features of extensions.

As an example, we will cover a “todo” extension that adds capabilities to include todo entries in the documentation, and to collect these in a central place. (A similar “todo” extension is distributed with Sphinx.)

Important objects

There are several key objects whose API you will use while writing an extension. These are:


The application object (usually called app) is an instance of Sphinx. It controls most high-level functionality, such as the setup of extensions, event dispatching and producing output (logging).

If you have the environment object, the application is available as env.app.


The build environment object (usually called env) is an instance of BuildEnvironment. It is responsible for parsing the source documents, stores all metadata about the document collection and is serialized to disk after each build.

Its API provides methods to do with access to metadata, resolving references, etc. It can also be used by extensions to cache information that should persist for incremental rebuilds.

If you have the application or builder object, the environment is available as app.env or builder.env.


The builder object (usually called builder) is an instance of a specific subclass of Builder. Each builder class knows how to convert the parsed documents into an output format, or otherwise process them (e.g. check external links).

If you have the application object, the builder is available as app.builder.


The config object (usually called config) provides the values of configuration values set in conf.py as attributes. It is an instance of Config.

The config is available as app.config or env.config.

Build Phases

One thing that is vital in order to understand extension mechanisms is the way in which a Sphinx project is built: this works in several phases.

Phase 0: Initialization

In this phase, almost nothing of interest to us happens. The source directory is searched for source files, and extensions are initialized. Should a stored build environment exist, it is loaded, otherwise a new one is created.

Phase 1: Reading

In Phase 1, all source files (and on subsequent builds, those that are new or changed) are read and parsed. This is the phase where directives and roles are encountered by docutils, and the corresponding code is executed. The output of this phase is a doctree for each source file; that is a tree of docutils nodes. For document elements that aren’t fully known until all existing files are read, temporary nodes are created.

There are nodes provided by docutils, which are documented in the docutils documentation. Additional nodes are provided by Sphinx and documented here.

During reading, the build environment is updated with all meta- and cross reference data of the read documents, such as labels, the names of headings, described Python objects and index entries. This will later be used to replace the temporary nodes.

The parsed doctrees are stored on the disk, because it is not possible to hold all of them in memory.

Phase 2: Consistency checks

Some checking is done to ensure no surprises in the built documents.

Phase 3: Resolving

Now that the metadata and cross-reference data of all existing documents is known, all temporary nodes are replaced by nodes that can be converted into output using components called tranform. For example, links are created for object references that exist, and simple literal nodes are created for those that don’t.

Phase 4: Writing

This phase converts the resolved doctrees to the desired output format, such as HTML or LaTeX. This happens via a so-called docutils writer that visits the individual nodes of each doctree and produces some output in the process.


Some builders deviate from this general build plan, for example, the builder that checks external links does not need anything more than the parsed doctrees and therefore does not have phases 2–4.

Extension Design

We want the extension to add the following to Sphinx:

  • A “todo” directive, containing some content that is marked with “TODO”, and only shown in the output if a new config value is set. (Todo entries should not be in the output by default.)
  • A “todolist” directive that creates a list of all todo entries throughout the documentation.

For that, we will need to add the following elements to Sphinx:

  • New directives, called todo and todolist.
  • New document tree nodes to represent these directives, conventionally also called todo and todolist. We wouldn’t need new nodes if the new directives only produced some content representable by existing nodes.
  • A new config value todo_include_todos (config value names should start with the extension name, in order to stay unique) that controls whether todo entries make it into the output.
  • New event handlers: one for the doctree-resolved event, to replace the todo and todolist nodes, and one for env-purge-doc (the reason for that will be covered later).

The Setup Function

The new elements are added in the extension’s setup function. Let us create a new Python module called todo.py and add the setup function:

def setup(app):
    app.add_config_value('todo_include_todos', False, 'html')

                 html=(visit_todo_node, depart_todo_node),
                 latex=(visit_todo_node, depart_todo_node),
                 text=(visit_todo_node, depart_todo_node))

    app.add_directive('todo', TodoDirective)
    app.add_directive('todolist', TodolistDirective)
    app.connect('doctree-resolved', process_todo_nodes)
    app.connect('env-purge-doc', purge_todos)

    return {'version': '0.1'}   # identifies the version of our extension

The calls in this function refer to classes and functions not yet written. What the individual calls do is the following:

  • add_config_value() lets Sphinx know that it should recognize the new config value todo_include_todos, whose default value should be False (this also tells Sphinx that it is a boolean value).

    If the third argument was 'html', HTML documents would be full rebuild if the config value changed its value. This is needed for config values that influence reading (build phase 1).

  • add_node() adds a new node class to the build system. It also can specify visitor functions for each supported output format. These visitor functions are needed when the new nodes stay until phase 4 – since the todolist node is always replaced in phase 3, it doesn’t need any.

    We need to create the two node classes todo and todolist later.

  • add_directive() adds a new directive, given by name and class.

    The handler functions are created later.

  • Finally, connect() adds an event handler to the event whose name is given by the first argument. The event handler function is called with several arguments which are documented with the event.

The Node Classes

Let’s start with the node classes:

from docutils import nodes

class todo(nodes.Admonition, nodes.Element):

class todolist(nodes.General, nodes.Element):

def visit_todo_node(self, node):

def depart_todo_node(self, node):

Node classes usually don’t have to do anything except inherit from the standard docutils classes defined in docutils.nodes. todo inherits from Admonition because it should be handled like a note or warning, todolist is just a “general” node.


Many extensions will not have to create their own node classes and work fine with the nodes already provided by docutils and Sphinx.

The Directive Classes

A directive class is a class deriving usually from docutils.parsers.rst.Directive. The directive interface is also covered in detail in the docutils documentation; the important thing is that the class should have attributes that configure the allowed markup, and a run method that returns a list of nodes.

The todolist directive is quite simple:

from docutils.parsers.rst import Directive

class TodolistDirective(Directive):

    def run(self):
        return [todolist('')]

An instance of our todolist node class is created and returned. The todolist directive has neither content nor arguments that need to be handled.

The todo directive function looks like this:

from sphinx.locale import _

class TodoDirective(Directive):

    # this enables content in the directive
    has_content = True

    def run(self):
        env = self.state.document.settings.env

        targetid = "todo-%d" % env.new_serialno('todo')
        targetnode = nodes.target('', '', ids=[targetid])

        todo_node = todo('\n'.join(self.content))
        todo_node += nodes.title(_('Todo'), _('Todo'))
        self.state.nested_parse(self.content, self.content_offset, todo_node)

        if not hasattr(env, 'todo_all_todos'):
            env.todo_all_todos = []
            'docname': env.docname,
            'lineno': self.lineno,
            'todo': todo_node.deepcopy(),
            'target': targetnode,

        return [targetnode, todo_node]

Several important things are covered here. First, as you can see, you can refer to the build environment instance using self.state.document.settings.env.

Then, to act as a link target (from the todolist), the todo directive needs to return a target node in addition to the todo node. The target ID (in HTML, this will be the anchor name) is generated by using env.new_serialno which returns a new unique integer on each call and therefore leads to unique target names. The target node is instantiated without any text (the first two arguments).

On creating admonition node, the content body of the directive are parsed using self.state.nested_parse. The first argument gives the content body, and the second one gives content offset. The third argument gives the parent node of parsed result, in our case the todo node.

Then, the todo node is added to the environment. This is needed to be able to create a list of all todo entries throughout the documentation, in the place where the author puts a todolist directive. For this case, the environment attribute todo_all_todos is used (again, the name should be unique, so it is prefixed by the extension name). It does not exist when a new environment is created, so the directive must check and create it if necessary. Various information about the todo entry’s location are stored along with a copy of the node.

In the last line, the nodes that should be put into the doctree are returned: the target node and the admonition node.

The node structure that the directive returns looks like this:

| target node        |
| todo node          |
     | admonition title   |
     | paragraph          |
     | ...                |

The Event Handlers

Finally, let’s look at the event handlers. First, the one for the env-purge-doc event:

def purge_todos(app, env, docname):
    if not hasattr(env, 'todo_all_todos'):
    env.todo_all_todos = [todo for todo in env.todo_all_todos
                          if todo['docname'] != docname]

Since we store information from source files in the environment, which is persistent, it may become out of date when the source file changes. Therefore, before each source file is read, the environment’s records of it are cleared, and the env-purge-doc event gives extensions a chance to do the same. Here we clear out all todos whose docname matches the given one from the todo_all_todos list. If there are todos left in the document, they will be added again during parsing.

The other handler belongs to the doctree-resolved event. This event is emitted at the end of phase 3 and allows custom resolving to be done:

def process_todo_nodes(app, doctree, fromdocname):
    if not app.config.todo_include_todos:
        for node in doctree.traverse(todo):

    # Replace all todolist nodes with a list of the collected todos.
    # Augment each todo with a backlink to the original location.
    env = app.builder.env

    for node in doctree.traverse(todolist):
        if not app.config.todo_include_todos:

        content = []

        for todo_info in env.todo_all_todos:
            para = nodes.paragraph()
            filename = env.doc2path(todo_info['docname'], base=None)
            description = (
                _('(The original entry is located in %s, line %d and can be found ') %
                (filename, todo_info['lineno']))
            para += nodes.Text(description, description)

            # Create a reference
            newnode = nodes.reference('', '')
            innernode = nodes.emphasis(_('here'), _('here'))
            newnode['refdocname'] = todo_info['docname']
            newnode['refuri'] = app.builder.get_relative_uri(
                fromdocname, todo_info['docname'])
            newnode['refuri'] += '#' + todo_info['target']['refid']
            para += newnode
            para += nodes.Text('.)', '.)')

            # Insert into the todolist


It is a bit more involved. If our new “todo_include_todos” config value is false, all todo and todolist nodes are removed from the documents.

If not, todo nodes just stay where and how they are. Todolist nodes are replaced by a list of todo entries, complete with backlinks to the location where they come from. The list items are composed of the nodes from the todo entry and docutils nodes created on the fly: a paragraph for each entry, containing text that gives the location, and a link (reference node containing an italic node) with the backreference. The reference URI is built by app.builder.get_relative_uri which creates a suitable URI depending on the used builder, and appending the todo node’s (the target’s) ID as the anchor name.