Writing Technical Documentation with Sphinx, Paver, and Cog

I’ve been working on the Python Module of the Week series since March of 2007. During the course of the project, my article style and tool chain have both evolved. I now have a fairly smooth production process in place, so the mechanics of producing a new post don’t get in the way of the actual research and writing. Most of the tools are open source, so I thought I would describe the process I go through and how the tools work together.

Editing Text: TextMate

I work on a MacBook Pro, and use TextMate for editing the articles and source for PyMOTW. TextMate is the one tool I use regularly that is not open source. When I’m doing heavy editing of hundreds of files for my day job I use Aquamacs Emacs, but TextMate is better suited for prose editing and is easier to extend with quick actions. I discovered TextMate while looking for a native editor to use for Python Magazine, and after being able to write my own “bundle” to manage magazine articles (including defining a mode for the markup language we use) I was hooked.

Some of the features that I like about TextMate for prose editing are as-you-type spell-checking (I know some people hate this feature, but I find it useful), text statistics (word count, etc.), easy block selection (I can highlight a paragraph or several sentences and move them using cursor keys), a moderately good reStructuredText mode (emacs’ is better, but TextMate’s is good enough), paren and quote matching as you type, and very simple extensibility for repetitive tasks. I also like TextMate’s project management features, since they makes it easy to open several related files at the same time.

Version Control: svn

I started out using a private svn repository for all of my projects, including PyMOTW. I’m in the middle of evaluating hosted DVCS options for PyMOTW, but still haven’t had enough time to give them all the research I think is necessary before making the move. The Python core developers are considering a similar move (PEP 374) so it will be interesting to monitor that discussion. No doubt we have different requirements (for example, they are hosting their own repository), but the experiences with the various DVCS tools will be useful input to my own decision.

Markup Language: reStructuredText

When I began posting, I wrote each article by hand using HTML. One of the first tasks that I automated was the step of passing the source code through pygments to produce a syntax colorized version. This worked well enough for me at the time, but restricted me to producing only HTML output. Eventually John Benediktsson contacted me with a version of many of the posts converted from HTML to reStructuredText.

When reStructuredText was first put forward in the ‘90’s, I was heavily into Zope development. As such, I was using StructuredText for documenting my code, and in the Zope-based wiki that we ran at ZapMedia. I even wrote my own app to extract comments and docstrings to generate library documentation for a couple of libraries I had released as open source. I really liked StructuredText and, at first, I didn’t like reStructuredText. Frankly, it looked ugly compared to what I was used to. It quickly gained acceptance in the general community though, and I knew it would give me options for producing other output formats for the PyMOTW posts, so when John sent me the markup files I took another look.

While re-acquainting myself with reST, I realized two things. First, although there is a bit more punctuation involved in the markup than with the original StructuredText, the markup language was designed with consistency in mind so it isn’t as difficult to learn as my first impressions had lead me to believe. Second, it turned out the part I thought was “ugly” was actually the part that made reST more powerful than StructuredText: It has a standard syntax for extension directives that users can define for their own documents.

Markup to Output: Sphinx

Before I made a final decision on switching from hand-coded HTML to reST, I needed a tool to convert to HTML (I still had to post the results on the blog, after all, and Blogger doesn’t support reST). I first tried David Goodger’s docutils package. The scripts it includes felt a little too much like “pieces” of a tool rather than a complete solution, though, and I didn’t really want to assemble my own wrappers if I didn’t have to – I wanted to write text for this project, not code my own tools. Around this time, Georg Brandl had made significant progress on Sphinx, which turned out to be a more complete turn-key system for converting a pile of reST files to HTML or PDF. After a few hours of experimentation, I had a sample project set up and was generating HTML from my documents using the standard templates.

I decided that reStructuredText looked like the way to go.

HTML Templates: Jinja:

My next step was to work out exactly how to produce all of the outputs I needed from reST inputs. Each post for the PyMOTW series ends up going to several different places:

  • the PyMOTW source distribution (HTML)
  • my Blogger blog (HTML)
  • the PyMOTW project site (HTML)
  • O’Reilly.com (HTML)
  • the PyMOTW “book” (PDF)

Each of the four HTML outputs uses slightly different formatting, requiring separate templates (PDF is a whole different problem, covered below). The source distribution and project site are both full HTML versions of all of the documents, but use different templates. I decided to use the default Sphinx templates for the packaged version; I may change that later, but it works for the time being, and it’s one less custom template to deal with. I wanted the online version to match the appearance of the rest of my site, so I needed to create a template for it. The two blogs use a third template (O’Reilly’s site ignores a lot of the markup due to their Moveable Type configuration, but the articles come out looking good enough so I can use the same template I use for my own blog without worrying about a separate custom template).

Sphinx uses Jinja templates to produce HTML output. The syntax for Jinja is very similar to Django’s template language. As it happens, I use Django for the dynamic portion of my web site that I host myself. I lucked out, and my site’s base template was simple enough to use with Sphinx without making any changes. Yay for compatibility!

Cleaning up HTML with BeautifulSoup

The blog posts need to be relatively clean HTML that I can upload to Blogger and O’Reilly, so they could not include any html or body tags or require any markup or styles not supported by either blogging engine. The template I came up with is a stripped down version that doesn’t include the CSS and markup for sidebars, header, or footer. The result was almost exactly what I wanted, but had two problems.

The easiest problem to handle was the permalinks generated by Sphinx. After each heading on the page, Sphinx inserts an anchor tag with a ¶ character and applies CSS styles that hide/show the tag when the user hovers over it. That’s a nice feature for the main site and packaged content, but they didn’t work for the blogs. I have no control over the CSS used at O’Reilly, so the tags were always visible. I didn’t really care if they were included on the Blogger pages, so the simplest thing to do was stick with one “blogging” template and remove the permalinks.

The second, more annoying, problem, was that Blogger wanted to insert extra whitespace into the post. There is a configuration option on Blogger to treat line breaks in the post as “paragraph breaks” (I think they actually insert br tags). This is very convenient for normal posts with mostly straight text, since I can simply write each paragraph on one long line, wrapped visually by my editor, and break the paragraphs where I want them. The result is I can almost post directly from plain text input. Unfortunately, the option is applied to every post in the blog (even old posts), so changing it was not a realistic option – I wasn’t about to go back and re-edit every single post I had previously written.

The second, more annoying, problem, was that Blogger wanted to insert extra whitespace into the post.

Sphinx didn’t have an option to skip generating the permalinks, and there was no way to express that intent in the template, so I fell back to writing a little script to strip them out after the fact. I used BeautifulSoup to find the tags I wanted removed, delete them from the parse tree, then assemble the HTML text as a string again. I added code to the same script to handle the whitespace issue by removing all newlines from the input unless they were inside pre tags, which Blogger handled correctly. The result was a single blob of partial HTML without newlines or permalinks that I could post directly to either blog without editing it by hand. Score a point for automation.

def clean_blog_html(body):
    # Clean up the HTML
    import re
    import sys
    from BeautifulSoup import BeautifulSoup
    from cStringIO import StringIO

    # The post body is passed to stdin.
    soup = BeautifulSoup(body)

    # Remove the permalinks to each header since the blog does not have
    # the styles to hide them.
    links = soup.findAll('a', attrs={'class':"headerlink"})
    [l.extract() for l in links]

    # Get BeautifulSoup's version of the string
    s = soup.__str__(prettyPrint=False)

    # Remove extra newlines.  This depends on the fact that
    # code blocks are passed through pygments, which wraps each part of the line
    # in a span tag.
    pattern = re.compile(r'([^s][^p][^a][^n]>)n$', re.DOTALL|re.IGNORECASE)
    s = ''.join(pattern.sub(r'1', l) for l in StringIO(s))

    return s

Code Syntax Highlighting: pygments

I wanted my posts to look as good as possible, and an important factor in the appearance would be the presentation of the source code. I adopted pygments in the early hand-coded HTML days, because it was easy to integrate into TextMate with a simple script.

pygmentize -f html -O cssclass=syntax $@

Binding the command to a key combination meant with a few quick keypresses I had HTML ready to insert into the body of a post.

When I moved to Sphinx, using pygments became even easier because Sphinx automatically passes included source code through pygments as it generates its output. Syntax highlighting works for HTML and PDF, so I didn’t need any custom processing.

Automation: Paver

Automation is important for my sense of well being. I hate dealing with mundane repetitive tasks, so once an article was written I didn’t want to have to touch it to prepare it for publication of any of the final destinations. As I have written before, I started out using make to run various shell commands. I have since converted the entire process to Paver.

Automation is important for my sense of well being.

The stock Sphinx integration provided with that comes with Paver didn’t quite meet my needs, but by examining the source I was able to create my own replacement tasks in an afternoon. The main problem was the tight coupling between the code to run Sphinx and the code to find the options to pass to it. For normal projects with a single documentation output format (Paver assumes HTML with a single config file), this isn’t a problem. PyMOTW’s requirements are different, with the four output formats discussed above.

In order to produce different output with Sphinx, you need different configuration files. Since the base name for the file must always be conf.py, that means the files have to be stored in separate directories. One of the options passed to Sphinx on the command line tells it the directory to look in for its configuration file. Even though Paver doesn’t fork() before calling Sphinx, it still uses the command line options to pass instructions.

Creating separate Sphinx configuration files was easy. The problem was defining options in Paver to tell Sphinx about each configuration directory for the different output. Paver options are grouped into bundles, which are essentially a namespace. When a Paver task looks for an option, it scans through the bundles, possibly cascading to the global namespace, until it finds the option by name. The search can be limited to specific bundles, so that the same option name can be used to configure different tasks.

The html task from paver.doctools sets the options search order to look for values first in the sphinx section, then globally. Once it has retrieved the path values, via _get_paths(), it invokes Sphinx.

def _get_paths():
    """look up the options that determine where all of the files are."""
    opts = options
    docroot = path(opts.get('docroot', 'docs'))
    if not docroot.exists():
        raise BuildFailure("Sphinx documentation root (%s) does not exist."
                           % docroot)
    builddir = docroot / opts.get("builddir", ".build")
    srcdir = docroot / opts.get("sourcedir", "")
    if not srcdir.exists():
        raise BuildFailure("Sphinx source file dir (%s) does not exist"
                            % srcdir)
    htmldir = builddir / "html"
    doctrees = builddir / "doctrees"
    return Bunch(locals())

def html():
    """Build HTML documentation using Sphinx. This uses the following
    options in a "sphinx" section of the options.

      the root under which Sphinx will be working. Default: docs
      directory under the docroot where the resulting files are put.
      default: build
      directory under the docroot for the source files
      default: (empty string)
    options.order('sphinx', add_rest=True)
    paths = _get_paths()
    sphinxopts = ['', '-b', 'html', '-d', paths.doctrees,
        paths.srcdir, paths.htmldir]
    dry("sphinx-build %s" % (" ".join(sphinxopts),), sphinx.main, sphinxopts)

This didn’t work for me because I needed to pass a separate configuration directory (not handled by the default _get_paths()) and different build and output directories. The simplest solution turned out to be re-implementing the Paver-Sphinx integration to make it more flexible. I created my own _get_paths() and made it look for the extra option values and use the directory structure I needed.

def _get_paths():
    """look up the options that determine where all of the files are."""
    opts = options

    docroot = path(opts.get('docroot', 'docs'))
    if not docroot.exists():
        raise BuildFailure("Sphinx documentation root (%s) does not exist."
                           % docroot)

    builddir = docroot / opts.get("builddir", ".build")

    srcdir = docroot / opts.get("sourcedir", "")
    if not srcdir.exists():
        raise BuildFailure("Sphinx source file dir (%s) does not exist"
                            % srcdir)

    # Where is the sphinx conf.py file?
    confdir = path(opts.get('confdir', srcdir))

    # Where should output files be generated?
    outdir = opts.get('outdir', '')
    if outdir:
        outdir = path(outdir)
        outdir = builddir / opts.get('builder', 'html')

    # Where are doctrees cached?
    doctrees = opts.get('doctrees', '')
    if not doctrees:
        doctrees = builddir / "doctrees"
        doctrees = path(doctrees)

    return Bunch(locals())

Then I defined a new function, run_sphinx(), to set up the options search path, look for the option values, and invoke Sphinx. I set add_rest to False to disable searching globally for an option to avoid namespace pollution from option collisions, since I knew I was going to have options with the same names but different values for each output format. I also look for a “builder”, to support PDF generation.

def run_sphinx(*option_sets):
    """Helper function to run sphinx with common options.

    Pass the names of namespaces to be used in the search path
    for options.
    if 'sphinx' not in option_sets:
        option_sets += ('sphinx',)
    kwds = dict(add_rest=False)
    options.order(*option_sets, **kwds)
    paths = _get_paths()
    sphinxopts = ['',
                  '-b', options.get('builder', 'html'),
                  '-d', paths.doctrees,
                  '-c', paths.confdir,
                  paths.srcdir, paths.outdir]
    dry("sphinx-build %s" % (" ".join(sphinxopts),), sphinx.main, sphinxopts)

With a working run_sphinx() function I could define several Sphinx-based tasks, each taking options with the same names but from different parts of the namespace. The tasks simply call run_sphinx() with the desired namespace search path. For example, to generate the HTML to include in the sdist package, the html task looks in the html bunch:

def html():
    """Build HTML documentation using Sphinx. This uses the following
    options in a "sphinx" section of the options.

      the root under which Sphinx will be working.
      default: docs
      directory under the docroot where the resulting files are put.
      default: build
      directory under the docroot for the source files
      default: (empty string)
      the location of the cached doctrees
      default: $builddir/doctrees
      the location of the sphinx conf.py
      default: $sourcedir
      the location of the generated output files
      default: $builddir/$builder
      the name of the sphinx builder to use
      default: html

while generating the HTML output for the website uses a different set of options from the website bunch:

@needs(['webtemplatebase', 'cog'])
def webhtml():
    """Generate HTML files for website.

All of the option search paths also include the sphinx bunch, so values that do not change (such as the source directory) do not need to be repeated. The relevant portion of the options from the PyMOTW pavement.py file looks like this:

    # ...

    sphinx = Bunch(
        docroot = '.',
        builder = 'html',
        confdir = 'sphinx',

    html = Bunch(

        templates = 'web',
        #outdir = 'web',
        builddir = 'web',



    # ...

To find the sourcedir for the html task, _get_paths() first looks in the html bunch, then the sphinx bunch.

Capturing Program Output: cog

As an editor at Python Magazine, and reviewer for several books, I’ve discovered that one of the most frequent sources of errors with technical writing occurs in the production process where the output of running sample code is captured to be included in the final text. This is usually done manually by running the program and copying and pasting its output from the console. It’s not uncommon for a bug to be found, or a library to change, requiring a change in the source code provided with the article. That change, in turn, means the output of commands may be different. Sometimes the change is minor, but at other times the output is different in some significant way. Since I’ve seen the problem come up so many times, I spent time thinking about and looking for a solution to avoid it in my own work.

During my research, a few people suggested that I switch to using doctests for my examples, but I felt there were several problems with that approach. First, the doctest format isn’t very friendly for users who want to copy and paste examples into their own scripts. The reader has to select each line individually, and can’t simply grab the entire block of code. Distributing the examples as separate scripts makes this easier, since they can simply copy the entire file and modify it as they want. Using individual .py files also makes it possible for some of the more complicated examples to run clients and servers at the same time from different scripts (as with SimpleXMLRPCServer, for example). But most importantly, using doctests does not solve the fundamental problem. Doctests tell me when the output has changed, but I still have to manually run the scripts to generate that output and paste it into my document in the first place. What I really wanted to be able to do was run the script and insert the output, whatever it was, without manually copying and pasting text from the console.

I finally found what I was looking for in cog, from Ned Batchelder. Ned describes cog as a “code generation tool”, and most of the examples he provides on his site are in that vein. But cog is a more general purpose tool than that. It gives you a way to include arbitrary Python instructions in your source document, have them executed, and then have the source document change to reflect the output.

For each code sample, I wanted to include the Python source followed by the output it produces when run on the console. There is a reST directive to include the source file, so that part is easy:

.. include:: anydbm_whichdb.py
    :start-after: #end_pymotw_header

The include directive tells Sphinx that the file “anydbm_whichdb.py” should be treated as a literal text block (instead of more reST) and to only include the parts following the last line of the standard header I use in all my source code. Syntax highlighting comes for free when the literal block is converted to the output format.

Grabbing the command output was a little trickier. Normally with cog, one would embed the actual source to be run in the document. In my case, I had the text in an external file. Most of the source is Python, and I could just import it, but I would have to go to special lengths to capture any output and pass it to cog.out(), the cog function for including text in the processed document. I didn’t want my example code littered with calls to cog.out() instead of print, so I needed to capture sys.stdout and sys.stdin. A bigger question was whether I wanted to have all of the sample files imported into the namespace of the build process. Considering both issues, it made sense to run the script in a separate process and capture the output.

There is a bit of setup work needed to run the scripts this way, so I decided to put it all into a function instead of including the boilerplate code in every cog block. The reST source for running anydbm_whichdb.py looks like:

.. {{{cog
.. cog.out(run_script(cog.inFile, 'anydbm_whichdb.py'))
.. }}}
.. {{{end}}}

The .. at the start of each line causes the reStructuredText parser to treat the line as a comment, so it is not included in the output. After passing the reST file through cog, it is rewritten to contain:

.. {{{cog
.. cog.out(run_script(cog.inFile, 'anydbm_whichdb.py'))
.. }}}


    $ python anydbm_whichdb.py

.. {{{end}}}

The run_script() function runs the python script it is given, adds a prefix to make reST treat the following lines as literal text, then indents the script output. The script is run via Paver’s sh() function, which wraps the subprocess module and supports the dry-run feature of Paver. Because the cog instructions are comments, the only part that shows up in the output is the literal text block with the command output.

def run_script(input_file, script_name,
    """Run a script in the context of the input_file's directory,
    return the text output formatted to be included as an rst
    literal text block.


       The name of the file being processed by cog.  Usually passed as cog.inFile.

       The name of the Python script living in the same directory as input_file to be run.
       If not using an interpreter, this can be a complete command line.  If using an
       alternate interpreter, it can be some other type of file.

       Boolean controlling whether the :: prefix is included.

       Boolean controlling whether errors are ignored.  If not ignored, the error
       is printed to stdout and then the command is run *again* with errors ignored
       so that the output ends up in the cogged file.

       Boolean controlling whether the trailing newlines are added to the output.
       If False, the output is passed to rstrip() then one newline is added.  If
       True, newlines are added to the output until it ends in 2.
    rundir = path(input_file).dirname()
    if interpreter:
        cmd = '%(interpreter)s %(script_name)s' % vars()
        cmd = script_name
    real_cmd = 'cd %(rundir)s; %(cmd)s 2>&1' % vars()
        output_text = sh(real_cmd, capture=True, ignore_error=ignore_error)
    except Exception, err:
        print '*' * 50
        print 'ERROR run_script(%s) => %s' % (real_cmd, err)
        print '*' * 50
        output_text = sh(real_cmd, capture=True, ignore_error=True)
    if include_prefix:
        response = 'n::nn'
        response = ''
    response += 't$ %(cmd)snt' % vars()
    response += 'nt'.join(output_text.splitlines())
    if trailing_newlines:
        while not response.endswith('nn'):
            response += 'n'
        response = response.rstrip()
        response += 'n'
    return response

I defined run_script() in my pavement.py file, and added it to the __builtins__ namespace to avoid having to import it each time I wanted to use it from a source document.

A somewhat more complicated example shows another powerful feature of cog. Because it can run any arbitrary Python code, it is possible to establish the preconditions for a script before running it. For example, anydbm_new.py assumes that its output database does not already exist. I can ensure that condition by removing it before running the script.

.. {{{cog
.. workdir = path(cog.inFile).dirname()
.. sh("cd %s; rm -f /tmp/example.db" % workdir)
.. cog.out(run_script(cog.inFile, 'anydbm_new.py'))
.. }}}
.. {{{end}}}

Since cog is integrated into Paver, all I had to do to enable it was define the options and import the module. I chose to change the begin and end tags used by cog because the default patterns ([[[cog and ]]]) appeared in the output of some of the scripts (printing nested lists, for example).


To process all of the input files through cog before generating the output, I added ‘cog' to the @needs list for any task running sphinx. Then it was simply a matter of running paver html or paver webhtml to generate the output.

Paver includes an uncog task to remove the cog output from your source files before committing to a source code repository, but I decided to include the cogged values in committed versions so I would be alerted if the output ever changed.

Generating PDF: TexLive

Generating HTML using Sphinx and Jinja templates is fairly straightforward; PDF output wasn’t quite so easy to set up. Sphinx actually produces LaTeX, another text-based format, as output, along with a Makefile to run third-party LaTeX tools to create the PDF. I started out experimenting on a Linux system (normally I use a Mac, but this box claimed to have the required tools installed). Due to the age of the system, however, the tools weren’t compatible with the LaTeX produced by Sphinx. After some searching, and asking on the sphinx-dev mailing list, I installed a copy of TeX Live, a newer TeX distro. A few tweaks to my $PATH later and I was in business building PDFs right on my Mac.

Generating HTML using Sphinx and Jinja templates is fairly straightforward; PDF output wasn’t quite so easy to set up.

My pdf task runs Sphinx with the “latex” builder, then runs make using the generated Makefile.

def pdf():
    """Generate the PDF book.
    latex_dir = path(options.pdf.builddir) / 'latex'
    sh('cd %s; make' % latex_dir)

I still need to experiment with some of the LaTeX options, including templates for pages in different sizes, logos, and styles. For now I’m happy with the default look.


Once I had the “build” fully automated, it was time to address the distribution process. For each version, I need to:

  • upload HTML, PDF, and tar.gz files to my server
  • update PyPI
  • post to my blog
  • post to the O’Reilly blog

The HTML and PDF files are copied to my server using rsync, invoked from Paver. I use a web browser and the admin interface for django-codehosting to upload the tar.gz file containing the source distribution manually. That will be automated, eventually. Once the tar.gz is available, PyPI can be updated via the builtin task paver register. That just leaves the two blog posts.

For my own blog, I use MarsEdit to post and edit entries. I find the UI easy to use, and I like the ability to work on drafts of posts offline. It is much nicer than the web interface for Blogger, and has the benefit of being AppleScript-able. I have plans to automate all of the steps right up to actually posting the new blog entry, but for now I copy the generated blog entry into a new post window by hand.

O’Reilly’s blogging policy does not allow desktop clients (too much of a support issue for the tech staff), so I need to use their Moveable Type web UI to post. As with MarsEdit, I simply copy the output and paste it into the field in the browser window, then add tags.

Tying it All Together

A quick overview of my current process is:

  • Pick a module, research it, and write examples in reST and Python. Include the Python source and use cog directives to bring in the script output.
  • Use the command “paver html” to produce HTML output to verify the results look good and I haven’t messed up any markup.
  • Commit the changes to svn. When I’m done with the module, copy the “trunk” to a release branch for packaging.
  • Use “paver sdist” to create the tar.gz file containing the Python source and HTML documentation.
  • Upload the tar.gz file to my site.
  • Run “paver installwebsite” to regenerate the hosted version of the HTML and the PDF, then copy both to my web server.
  • Run “paver register” to update PyPI with the latest release information.
  • Run “paver blog” to generate the HTML to be posted to the blogs. The task opens a new TextMate window containing the HTML so it is ready to be copied.
  • Paste the blog post contents into MarsEdit, add tags, and send it to Blogger.
  • Paste the blog post contents into the MT UI for O’Reilly, add tags, verify that it renders properly, then publish.

Try It Yourself

All of the source for PyMOTW (including the pavement.py file with configuration options, task definitions, and Sphinx integration) is available from the PyMOTW web site. Sphinx, Paver, cog, and BeautifulSoup are all open source projects. I’ve only tested the PyMOTW “build” on Mac OS X, but it should work on Linux without any major alterations. If you’re on Windows, let me know if you get it working.