PEP 258 – Docutils Design Specification
- David Goodger <goodger at python.org>
- Doc-SIG list
- Standards Track
- 256, 257
Table of Contents
- Rejection Notice
- Docutils Project Model
- Docutils Package Structure
- Front-End Tools
- Document Tree
- Error Handling
- Python Source Reader
- Project Web Site
While this may serve as an interesting design document for the now-independent docutils, it is no longer slated for inclusion in the standard library.
This PEP documents design issues and implementation details for Docutils, a Python Docstring Processing System (DPS). The rationale and high-level concepts of a DPS are documented in PEP 256, “Docstring Processing System Framework”. Also see PEP 256 for a “Road Map to the Docstring PEPs”.
Docutils is being designed modularly so that any of its components can be replaced easily. In addition, Docutils is not limited to the processing of Python docstrings; it processes standalone documents as well, in several contexts.
No changes to the core Python language are required by this PEP. Its deliverables consist of a package for the standard library and its documentation.
Docutils Project Model
Project components and data flow:
+---------------------------+ | Docutils: | | docutils.core.Publisher, | | docutils.core.publish_*() | +---------------------------+ / | \ / | \ 1,3,5 / 6 | \ 7 +--------+ +-------------+ +--------+ | READER | ----> | TRANSFORMER | ====> | WRITER | +--------+ +-------------+ +--------+ / \\ | / \\ | 2 / 4 \\ 8 | +-------+ +--------+ +--------+ | INPUT | | PARSER | | OUTPUT | +-------+ +--------+ +--------+
The numbers above each component indicate the path a document’s data takes. Double-width lines between Reader & Parser and between Transformer & Writer indicate that data sent along these paths should be standard (pure & unextended) Docutils doc trees. Single-width lines signify that internal tree extensions or completely unrelated representations are possible, but they must be supported at both ends.
docutils.core module contains a “Publisher” facade class and
several convenience functions: “publish_cmdline()” (for command-line
front ends), “publish_file()” (for programmatic use with file-like
I/O), and “publish_string()” (for programmatic use with string I/O).
The Publisher class encapsulates the high-level logic of a Docutils
system. The Publisher class has overall responsibility for
processing, controlled by the
- Set up internal settings (may include config files & command-line options) and I/O objects.
- Call the Reader object to read data from the source Input object and parse the data with the Parser object. A document object is returned.
- Set up and apply transforms via the Transformer object attached to the document.
- Call the Writer object which translates the document to the final
output format and writes the formatted data to the destination
Output object. Depending on the Output object, the output may be
returned from the Writer, and then from the
Calling the “publish” function (or instantiating a “Publisher” object)
with component names will result in default behavior. For custom
behavior (customizing component settings), create custom component
objects first, and pass them to the Publisher or
Readers understand the input context (where the data is coming from), send the whole input or discrete “chunks” to the parser, and provide the context to bind the chunks together back into a cohesive whole.
Each reader is a module or package exporting a “Reader” class with a
“read” method. The base “Reader” class can be found in the
Most Readers will have to be told what parser to use. So far (see the list of examples below), only the Python Source Reader (“PySource”; still incomplete) will be able to determine the parser on its own.
- Get input text from the source I/O.
- Pass the input text to the parser, along with a fresh document tree root.
- Standalone (Raw/Plain): Just read a text file and process it.
The reader needs to be told which parser to use.
The “Standalone Reader” has been implemented in module
- Python Source: See Python Source Reader below. This Reader is currently in development in the Docutils sandbox.
- Email: RFC 822 headers, quoted excerpts, signatures, MIME parts.
- PEP: RFC 822 headers, “PEP xxxx” and “RFC xxxx” conversion to URIs.
The “PEP Reader” has been implemented in module
docutils.readers.pep; see PEP 287 and PEP 12.
- Wiki: Global reference lookups of “wiki links” incorporated into transforms. (CamelCase only or unrestricted?) Lazy indentation?
- Web Page: As standalone, but recognize meta fields as meta tags.
Support for templates of some sort? (After
- FAQ: Structured “question & answer(s)” constructs.
- Compound document: Merge chapters into a book. Master manifest file?
Parsers analyze their input and produce a Docutils document tree. They don’t know or care anything about the source or destination of the data.
Each input parser is a module or package exporting a “Parser” class
with a “parse” method. The base “Parser” class can be found in the
Responsibilities: Given raw input text and a doctree root node, populate the doctree by parsing the input text.
Example: The only parser implemented so far is for the
reStructuredText markup. It is implemented in the
The development and integration of other parsers is possible and encouraged.
The Transformer class, in
transforms and applies them to documents. A transformer object is
attached to every new document tree. The Publisher calls
Transformer.apply_transforms() to apply all stored transforms to
the document tree. Transforms change the document tree from one form
to another, add to the tree, or prune it. Transforms resolve
references and footnote numbers, process interpreted text, and do
other context-sensitive processing.
Some transforms are specific to components (Readers, Parser, Writers,
Input, Output). Standard component-specific transforms are specified
default_transforms attribute of component classes. After
the Reader has finished processing, the Publisher calls
Transformer.populate_from_components() with a list of components
and all default transforms are stored.
Each transform is a class in a module in the
package, a subclass of
classes each have a
default_priority attribute which is used by
the Transformer to apply transforms in order (low to high). The
default priority can be overridden when adding transforms to the
- Apply transforms to the document tree, in priority order.
- Store a mapping of component type name (‘reader’, ‘writer’, etc.) to component objects. These are used by certain transforms (such as “components.Filter”) to determine suitability.
- Modify a doctree in-place, either purely transforming one structure into another, or adding new structures based on the doctree and/or external data.
Examples of transforms (in the
- frontmatter.DocInfo: Conversion of document metadata (bibliographic information).
- references.AnonymousHyperlinks: Resolution of anonymous references to corresponding targets.
- parts.Contents: Generates a table of contents for a document.
- document.Merger: Combining multiple populated doctrees into one. (Not yet implemented or fully understood.)
- document.Splitter: Splits a document into a tree-structure of subdocuments, perhaps by section. It will have to transform references appropriately. (Neither implemented not remotely understood.)
- components.Filter: Includes or excludes elements which depend on a specific Docutils component.
Writers produce the final output (HTML, XML, TeX, etc.). Writers translate the internal document tree structure into the final data format, possibly running Writer-specific transforms first.
By the time the document gets to the Writer, it should be in final form. The Writer’s job is simply (and only) to translate from the Docutils doctree structure to the target format. Some small transforms may be required, but they should be local and format-specific.
Each writer is a module or package exporting a “Writer” class with a
“write” method. The base “Writer” class can be found in the
- Translate doctree(s) into specific output formats.
- Transform references into format-native forms.
- Write the translated output to the destination I/O.
- XML: Various forms, such as:
- Docutils XML (an expression of the internal document tree,
- DocBook (being implemented in the Docutils sandbox).
- Docutils XML (an expression of the internal document tree, implemented as
- HTML (XHTML implemented as
- PDF (a ReportLabs interface is being developed in the Docutils sandbox).
- TeX (a LaTeX Writer is being implemented in the sandbox).
- Docutils-native pseudo-XML (implemented as
docutils.writers.pseudoxml, used for testing).
- Plain text
I/O classes provide a uniform API for low-level input and output. Subclasses will exist for a variety of input/output mechanisms. However, they can be considered an implementation detail. Most applications should be satisfied using one of the convenience functions associated with the Publisher.
I/O classes are currently in the preliminary stages; there’s a lot of work yet to be done. Issues:
- How to represent multi-file input (files & directories) in the API?
- How to represent multi-file output? Perhaps “Writer” variants, one for each output distribution type? Or Output objects with associated transforms?
- Read data from the input source (Input objects) or write data to the output destination (Output objects).
Examples of input sources:
- A single file on disk or a stream (implemented as
- Multiple files on disk (
- Python source files: modules and packages.
- Python strings, as received from a client application
Examples of output destinations:
- A single file on disk or a stream (implemented as
- A tree of directories and files on disk.
- A Python string, returned to a client application (implemented as
- No output; useful for programmatic applications where only a portion
of the normal output is to be used (implemented as
- A single tree-shaped data structure in memory.
- Some other set of data structures in memory.
Docutils Package Structure
- Package “docutils”.
- Module “__init__.py” contains: class “Component”, a base class for Docutils components; class “SettingsSpec”, a base class for specifying runtime settings (used by docutils.frontend); and class “TransformSpec”, a base class for specifying transforms.
- Module “docutils.core” contains facade class “Publisher” and convenience functions. See Publisher above.
- Module “docutils.frontend” provides runtime settings support, for programmatic use and front-end tools (including configuration file support, and command-line argument and option processing).
- Module “docutils.io” provides a uniform API for low-level input and output. See Input/Output above.
- Module “docutils.nodes” contains the Docutils document tree element class library plus tree-traversal Visitor pattern base classes. See Document Tree below.
- Module “docutils.statemachine” contains a finite state machine specialized for regular-expression-based text filters and parsers. The reStructuredText parser implementation is based on this module.
- Module “docutils.urischemes” contains a mapping of known URI schemes (“http”, “ftp”, “mail”, etc.).
- Module “docutils.utils” contains utility functions and classes, including a logger class (“Reporter”; see Error Handling below).
- Package “docutils.parsers”: markup parsers.
- Function “get_parser_class(parser_name)” returns a parser module
by name. Class “Parser” is the base class of specific parsers.
- Package “docutils.parsers.rst”: the reStructuredText parser.
- Alternate markup parsers may be added.
See Parsers above.
- Function “get_parser_class(parser_name)” returns a parser module by name. Class “Parser” is the base class of specific parsers. (
- Package “docutils.readers”: context-aware input readers.
- Function “get_reader_class(reader_name)” returns a reader module
by name or alias. Class “Reader” is the base class of specific
- Module “docutils.readers.standalone” reads independent document files.
- Module “docutils.readers.pep” reads PEPs (Python Enhancement Proposals).
- Readers to be added for: Python source code (structure & docstrings), email, FAQ, and perhaps Wiki and others.
See Readers above.
- Function “get_reader_class(reader_name)” returns a reader module by name or alias. Class “Reader” is the base class of specific readers. (
- Package “docutils.writers”: output format writers.
- Function “get_writer_class(writer_name)” returns a writer module
by name. Class “Writer” is the base class of specific writers.
- Module “docutils.writers.html4css1” is a simple HyperText Markup Language document tree writer for HTML 4.01 and CSS1.
- Module “docutils.writers.docutils_xml” writes the internal document tree in XML form.
- Module “docutils.writers.pseudoxml” is a simple internal document tree writer; it writes indented pseudo-XML.
- Writers to be added: HTML 3.2 or 4.01-loose, XML (various forms, such as DocBook), PDF, TeX, plaintext, reStructuredText, and perhaps others.
See Writers above.
- Function “get_writer_class(writer_name)” returns a writer module by name. Class “Writer” is the base class of specific writers. (
- Package “docutils.transforms”: tree transform classes.
- Class “Transformer” stores transforms and applies them to
document trees. (
- Class “Transform” is the base class of specific transforms.
- Each module contains related transform classes.
See Transforms above.
- Class “Transformer” stores transforms and applies them to document trees. (
- Package “docutils.languages”: Language modules contain
language-dependent strings and mappings. They are named for their
language identifier (as defined in Choice of Docstring Format
below), converting dashes to underscores.
- Function “get_language(language_code)”, returns matching
language module. (
- Modules: en.py (English), de.py (German), fr.py (French), it.py (Italian), sk.py (Slovak), sv.py (Swedish).
- Other languages to be added.
- Function “get_language(language_code)”, returns matching language module. (
- Third-party modules: “extras” directory. These modules are
installed only if they’re not already present in the Python
extras/textwrap.pyprovide option parsing and command-line help; from Greg Ward’s http://optik.sf.net/ project, included for convenience.
extras/roman.pycontains Roman numeral conversion routines.
tools/ directory contains several front ends for common
Docutils processing. See Docutils Front-End Tools for details.
A single intermediate data structure is used internally by Docutils,
in the interfaces between components; it is defined in the
docutils.nodes module. It is not required that this data
structure be used internally by any of the components, just
between components as outlined in the diagram in the Docutils
Project Model above.
Custom node types are allowed, provided that either (a) a transform
converts them to standard Docutils nodes before they reach the Writer
proper, or (b) the custom node is explicitly supported by certain
Writers, and is wrapped in a filtered “pending” node. An example of
condition (a) is the Python Source Reader (see below), where a
“stylist” transform converts custom nodes. The HTML
<meta> tag is
an example of condition (b); it is supported by the HTML Writer but
not by others. The reStructuredText “meta” directive creates a
“pending” node, which contains knowledge that the embedded “meta” node
can only be handled by HTML-compatible writers. The “pending” node is
resolved by the
which checks that the calling writer supports HTML; if it doesn’t, the
“pending” node (and enclosed “meta” node) is removed from the
The document tree data structure is similar to a DOM tree, but with specific node names (classes) instead of DOM’s generic nodes. The schema is documented in an XML DTD (eXtensible Markup Language Document Type Definition), which comes in two parts:
The DTD defines a rich set of elements, suitable for many input and output formats. The DTD retains all information necessary to reconstruct the original input text, or a reasonable facsimile thereof.
See The Docutils Document Tree for details (incomplete).
When the parser encounters an error in markup, it inserts a system message (DTD element “system_message”). There are five levels of system messages:
- Level-0, “DEBUG”: an internal reporting issue. There is no effect on the processing. Level-0 system messages are handled separately from the others.
- Level-1, “INFO”: a minor issue that can be ignored. There is little or no effect on the processing. Typically level-1 system messages are not reported.
- Level-2, “WARNING”: an issue that should be addressed. If ignored, there may be minor problems with the output. Typically level-2 system messages are reported but do not halt processing
- Level-3, “ERROR”: a major issue that should be addressed. If ignored, the output will contain unpredictable errors. Typically level-3 system messages are reported but do not halt processing
- Level-4, “SEVERE”: a critical error that must be addressed. Typically level-4 system messages are turned into exceptions which halt processing. If ignored, the output will contain severe errors.
Although the initial message levels were devised independently, they have a strong correspondence to VMS error condition severity levels; the names in quotes for levels 1 through 4 were borrowed from VMS. Error handling has since been influenced by the log4j project.
Python Source Reader
The Python Source Reader (“PySource”) is the Docutils component that reads Python source files, extracts docstrings in context, then parses, links, and assembles the docstrings into a cohesive whole. It is a major and non-trivial component, currently under experimental development in the Docutils sandbox. High-level design issues are presented here.
This model will evolve over time, incorporating experience and discoveries.
- The PySource Reader uses an Input class to read in Python packages and modules, into a tree of strings.
- The Python modules are parsed, converting the tree of strings into a tree of abstract syntax trees with docstring nodes.
- The abstract syntax trees are converted into an internal representation of the packages/modules. Docstrings are extracted, as well as code structure details. See AST Mining below. Namespaces are constructed for lookup in step 6.
- One at a time, the docstrings are parsed, producing standard Docutils doctrees.
- PySource assembles all the individual docstrings’ doctrees into a Python-specific custom Docutils tree paralleling the package/module/class structure; this is a custom Reader-specific internal representation (see the Docutils Python Source DTD). Namespaces must be merged: Python identifiers, hyperlink targets.
- Cross-references from docstrings (interpreted text) to Python identifiers are resolved according to the Python namespace lookup rules. See Identifier Cross-References below.
- A “Stylist” transform is applied to the custom doctree (by the Transformer), custom nodes are rendered using standard nodes as primitives, and a standard document tree is emitted. See Stylist Transforms below.
- Other transforms are applied to the standard doctree by the Transformer.
- The standard doctree is sent to a Writer, which translates the document into a concrete format (HTML, PDF, etc.).
- The Writer uses an Output class to write the resulting data to its destination (disk file, directories and files, etc.).
Abstract Syntax Tree mining code will be written (or adapted) that scans a parsed Python module, and returns an ordered tree containing the names, docstrings (including attribute and additional docstrings; see below), and additional info (in parentheses below) of all of the following objects:
- module attributes (+ initial values)
- classes (+ inheritance)
- class attributes (+ initial values)
- instance attributes (+ initial values)
- methods (+ parameters & defaults)
- functions (+ parameters & defaults)
(Extract comments too? For example, comments at the start of a module would be a good place for bibliographic field lists.)
In order to evaluate interpreted text cross-references, namespaces for each of the above will also be required.
See the python-dev/docstring-develop thread “AST mining”, started on 2001-08-14.
Docstring Extraction Rules
- What to examine:
- If the “
__all__” variable is present in the module being documented, only identifiers listed in “
__all__” are examined for docstrings.
- In the absence of “
__all__”, all identifiers are examined, except those whose names are private (names begin with “_” but don’t begin and end with “__”).
- 1a and 1b can be overridden by runtime settings.
- If the “
Docstrings are string literal expressions, and are recognized in the following places within Python modules:
- At the beginning of a module, function definition, class
definition, or method definition, after any comments. This is
the standard for Python
- Immediately following a simple assignment at the top level of a
module, class definition, or
__init__method definition, after any comments. See Attribute Docstrings below.
- Additional string literals found immediately after the docstrings in (a) and (b) will be recognized, extracted, and concatenated. See Additional Docstrings below.
- @@@ 2.2-style “properties” with attribute docstrings? Wait for syntax?
- At the beginning of a module, function definition, class definition, or method definition, after any comments. This is the standard for Python
Whenever possible, Python modules should be parsed by Docutils, not imported. There are several reasons:
- Importing untrusted code is inherently insecure.
- Information from the source is lost when using introspection to examine an imported module, such as comments and the order of definitions.
- Docstrings are to be recognized in places where the byte-code compiler ignores string literal expressions (2b and 2c above), meaning importing the module will lose these docstrings.
Of course, standard Python parsing tools such as the “parser” library module should be used.
When the Python source code for a module is not available (i.e. only the
.pycfile exists) or for C extension modules, to access docstrings the module can only be imported, and any limitations must be lived with.
Since attribute docstrings and additional docstrings are ignored by
the Python byte-code compiler, no namespace pollution or runtime bloat
will result from their use. They are not assigned to
to any other attribute. The initial parsing of a module may take a
slight performance hit.
(This is a simplified version of PEP 224.)
A string literal immediately following an assignment statement is interpreted by the docstring extraction machinery as the docstring of the target of the assignment statement, under the following conditions:
- The assignment must be in one of the following contexts:
- At the top level of a module (i.e., not nested inside a compound statement such as a loop or conditional): a module attribute.
- At the top level of a class definition: a class attribute.
- At the top level of the “
__init__” method definition of a class: an instance attribute. Instance attributes assigned in other methods are assumed to be implementation details. (@@@
- A function attribute assignment at the top level of a module or class definition.
Since each of the above contexts are at the top level (i.e., in the outermost suite of a definition), it may be necessary to place dummy assignments for attributes assigned conditionally or in a loop.
- The assignment must be to a single target, not to a list or a tuple of targets.
- The form of the target:
- For contexts 1a and 1b above, the target must be a simple identifier (not a dotted identifier, a subscripted expression, or a sliced expression).
- For context 1c above, the target must be of the form
self.attrib”, where “
self” matches the “
__init__” method’s first parameter (the instance parameter) and “attrib” is a simple identifier as in 3a.
- For context 1d above, the target must be of the form
name.attrib”, where “
name” matches an already-defined function or method name and “attrib” is a simple identifier as in 3a.
Blank lines may be used after attribute docstrings to emphasize the connection between the assignment and the docstring.
g = 'module attribute (module-global variable)' """This is g's docstring.""" class AClass: c = 'class attribute' """This is AClass.c's docstring.""" def __init__(self): """Method __init__'s docstring.""" self.i = 'instance attribute' """This is self.i's docstring.""" def f(x): """Function f's docstring.""" return x**2 f.a = 1 """Function attribute f.a's docstring."""
(This idea was adapted from PEP 216.)
Many programmers would like to make extensive use of docstrings for
API documentation. However, docstrings do take up space in the
running program, so some programmers are reluctant to “bloat up” their
code. Also, not all API documentation is applicable to interactive
__doc__ would be displayed.
Docutils’ docstring extraction tools will concatenate all string
literal expressions which appear at the beginning of a definition or
after a simple assignment. Only the first strings in definitions will
be available as
__doc__, and can be used for brief usage text
suitable for interactive sessions; subsequent string literals and all
attribute docstrings are ignored by the Python byte-code compiler and
may contain more extensive API information.
def function(arg): """This is __doc__, function's docstring.""" """ This is an additional docstring, ignored by the byte-code compiler, but extracted by Docutils. """ pass
from __future__ import
This would break “
from __future__ import” statements introduced
in Python 2.1 for multiple module docstrings (main docstring plus
additional docstring(s)). The Python Reference Manual specifies:
A future statement must appear near the top of the module. The only lines that can appear before a future statement are:
- the module docstring (if any),
- blank lines, and
- other future statements.
- Should we search for docstrings after a
__future__statement? Very ugly.
__future__statements to allow multiple preceding string literals?
- Or should we not even worry about this? There probably
__future__statements in production code, after all. Perhaps modules with
__future__statements will simply have to put up with the single-docstring limitation.
Choice of Docstring Format
Rather than force everyone to use a single docstring format, multiple
input formats are allowed by the processing system. A special
__docformat__, may appear at the top level of a module
before any function or class definitions. Over time or through
decree, a standard format or set of formats should emerge.
__docformat__ variable only applies to the objects
defined in the module’s file. In particular, the
variable in a package’s
__init__.py file does not apply to objects
defined in subpackages and submodules.
__docformat__ variable is a string containing the name of the
format being used, a case-insensitive string matching the input
parser’s module or package name (i.e., the same name as required to
“import” the module or package), or a registered alias. If no
__docformat__ is specified, the default format is “plaintext” for
now; this may be changed to the standard format if one is ever
__docformat__ string may contain an optional second field,
separated from the format name (first field) by a single space: a
case-insensitive language identifier as defined in RFC 1766. A
typical language identifier consists of a 2-letter language code from
ISO 639 (3-letter codes used only if no 2-letter code exists;
RFC 1766 is currently being revised to allow 3-letter codes). If no
language identifier is specified, the default is “en” for English.
The language identifier is passed to the parser and can be used for
language-dependent markup features.
In Python docstrings, interpreted text is used to classify and mark up program identifiers, such as the names of variables, functions, classes, and modules. If the identifier alone is given, its role is inferred implicitly according to the Python namespace lookup rules. For functions and methods (even when dynamically assigned), parentheses (‘()’) may be included:
This function uses `another()` to do its work.
For class, instance and module attributes, dotted identifiers are used when necessary. For example (using reStructuredText markup):
class Keeper(Storer): """ Extend `Storer`. Class attribute `instances` keeps track of the number of `Keeper` objects instantiated. """ instances = 0 """How many `Keeper` objects are there?""" def __init__(self): """ Extend `Storer.__init__()` to keep track of instances. Keep count in `Keeper.instances`, data in `self.data`. """ Storer.__init__(self) Keeper.instances += 1 self.data =  """Store data in a list, most recent last.""" def store_data(self, data): """ Extend `Storer.store_data()`; append new `data` to a list (in `self.data`). """ self.data = data
Each of the identifiers quoted with backquotes (“`”) will become references to the definitions of the identifiers themselves.
Stylist transforms are specialized transforms specific to the PySource Reader. The PySource Reader doesn’t have to make any decisions as to style; it just produces a logically constructed document tree, parsed and linked, including custom node types. Stylist transforms understand the custom nodes created by the Reader and convert them into standard Docutils nodes.
Multiple Stylist transforms may be implemented and one can be chosen at runtime (through a “–style” or “–stylist” command-line option). Each Stylist transform implements a different layout or style; thus the name. They decouple the context-understanding part of the Reader from the layout-generating part of processing, resulting in a more flexible and robust system. This also serves to “separate style from content”, the SGML/XML ideal.
By keeping the piece of code that does the styling small and modular, it becomes much easier for people to roll their own styles. The “barrier to entry” is too high with existing tools; extracting the stylist code will lower the barrier considerably.
Project Web Site
A SourceForge project has been set up for this work at http://docutils.sourceforge.net/.
This document has been placed in the public domain.
This document borrows ideas from the archives of the Python Doc-SIG. Thanks to all members past & present.
Last modified: 2022-02-27 22:46:36 GMT