Soupault 4.0.0: as extensible as Jekyll, still statically linked

Estimated reading time: 25 minutes.



Soupault 4.0.0 is available for download from my own server and from GitHub releases. It introduces page processing hooks, Lua index processors (which allow creating taxonomies and paginated indices), a new option to make the site index available to content pages, an option to process certain pages before all other, a way to mark certain index fields as required, and a bunch of new Lua plugin API functions. Well, and a few bug fixes, of course.

In short, this release takes extensibility to a new level, comparable to static site generators written in interpreted languages.

If you allow me a digression, I find it odd that the static site generator “scene” tends to treat extensibility and native code implementation as mutually exclusive properties.

On one end of the spectrum, we have projects like Jekyll or Nikola that are infinitely extensible and allow plugins to redefine almost any built-in behavior but need an interpreter and have a lot of runtime dependencies. If some functionality is not there, you find or write a plugin that adds it (from scratch or using a library from gems/pypi/etc).

On the other end, there are projects like Hugo or Zola that are available as native, statically linked executables but don’t offer any extensibility other than a Turing-complete template language. If something is not there, all you can do is hope for its inclusion in the mainline code, maintain a fork, or look elsewhere.

Soupault was already the most extensible static site generator available as a native, statically linked executable. By operating at the HTML element tree level, it made all functionality work the same for any source format, and it allows automatically loading any format with user-defined page preprocessors. By embedding a Lua interpeter, it allows manipulating the page element tree (“DOM”) in the same ways as client-side JS (without interactivity, of course). It can also pipe element tree nodes through external helpers and inject their output back into the page—in place or in addition to the original node.

Still, quite a few things were only possible with inelegant workarounds.

For example, taxonomies and paginated index pages could only be made with a LaTeX-like approach: run soupault --index-only to produce a JSON dump of the index data, then run a custom script to generate new pages in site/ from it, and finally run soupault again to render a complete site.

That approach is certainly workable; I used it for my own blog for a long time. But it’s also clunky and takes more build setup than many people are willing to tolerate for a static site generator, especially when other projects provide easy (if not always very flexible) ways to do the same.

I could add a “magical” pagination and taxonomy generator that would suit most common needs. However, my goal for soupault is to give every user complete control over the website generation process. Whenever I want new functionality, I start with thinking about a general mechanism that would allow me to do that if I couldn’t modify soupault itself.

Now there are two major features that allow doing those things without external tools and multi-pass workflows:

Read on for details. But before that, let’s discuss a closely related development.

Introducing soupault “blueprints”

A common complaint about soupault is that it’s hard to get started with or even see its full capabilities. It’s a fair point. The usual way for static site generators to give people a quick start is to maintain a repository of “themes”.

I find the word “theme” deceptive, though. In reality, they are more like applications on top of the SSG framework that include both presentation and logic. Even themes with similar structures may not be interchangeable. Theme compatibility with newer SSG versions is also a very real issue for some SSGs. In any case, once you choose and download a theme, switching to another theme isn’t guaranteed to be simple, especially if you modify anything.

I suppose the real reason why soupault users aren’t actively making reusable themes is not that they also think a theme repository isn’t a solution to all beginner problems. People likely turn to soupault for use cases that no other SSG supports, and they were going to make their own setup from the start.

However, the quick start problem does exist. This is why I took the time to prepare two reusable soupault setups that you can take and build upon: one for blogs and another for books. There’s also a wiki setup in development, but it needs more design effort and work on the code. Instead of “themes”, I’m calling them “blueprints”.

So far, there are blueprints for making blogs and books. They both require soupault 4.0.0 and serve as showcases for its new features, such as Lua index generators and the index_first option.


The blog blueprint is a full-featured blog setup with support for tags and Atom feeds (global and per-tag).

Code-wise, it’s more or less a reusable version of my own blog, while the visual design resembles this website instead.


The book blueprint is meant for creating online books.

You can see a real book project based on it at It also features automatic type checking for OCaml code snippets, so you can use it for inspiration if you are writing your own programming book.

It’s similar to mdBook in functionality, but doesn’t require a file: the chapter list in the sidebar is generated automatically.

Chapter numbers are stored in file names, like One Lua hook extracts that number from the file name to add it to the page metadata, and another hook removes it from the target path, so that its URL becomes just /introduction.

Future blueprints

I’m also working on a wiki blueprint inspired by MediaWiki (e.g., it will have categories—a very underused wiki feature). It’s not ready yet and needs more work, so stay tuned for updates. I also have an idea to move the old sample site to a tongue-in-cheek homepage blueprint.

Now let’s move on to the actual release changelog.

Breaking changes

Before we get to new features, let’s discuss the breaking changes that triggered the major version bump to 4.0.0. They are minor and only affect edge cases, but semver is semver.1

Index views have no default template anymore

The index view option index_item_template does not have a default value anymore.

If you had an index view that didn’t have either index_template, index_item_template, or index_processor and it was working fine for you, you need to add its original default value to your view explicitly to make it work like before.

  index_item_template = '''<div> <a href="{{url}}">{{title}}</a> </div>'''

The reason for removing its implicit default value is that it continuted to refer to the {{title}} field long after the content model dehardcoding in soupault 2.0.0. The {{url}} field is a built-in, one of the technical fields that is guaranteed to be present. The {{title}} field is a remnant from the built-in content model of soupault 1.x.x that made title, date, and author fields special.

Now that the content model is completely user-defined, it’s very strange to have a hardcoded reference to a field that may not exist in the user’s configuration, so I decided to remove it. My survey of existing websites built with soupault shows that everyone specifies that option explicitly, so this change shouldn’t affect many people.

On a side note, I briefly thought of removing the index_item_template option itself because index_item_template = '<div> <a href="{{url}}">{{title}}</a> </div>' is merely a syntactic sugar for:

index_template = '''
{% for e in entries %}
  <div> <a href="{{e.url}}">{{e.title}}</a> </div>
{% endfor %}

However, my survey shows that surprisingly many people use index_item_template in their configs, and if people see value in it, I definitely shouldn’t remove it.

The old implementation of the absolute_links widget would just do nothing if you forgot to specify the site URL in prefix, so this configuration was valid:

  widget = "absolute_links"
  # prefix = ""

Now it causes an error because in most cases, it’s likely a configuration mistake that shouldn’t be silently ignored. Explicitly setting the prefix to "" will not cause any errors.

Empty build_dir is no longer allowed

This is another behavior that was accidentally allowed rather than intentionally designed. While an absolute_links widget with an unspecified prefix simply does nothing, build_dir = "" is more annoying: it makes soupault output pages into the current working directory.

I doubt it’s a behavior anyone would want, and I believe programs that are supposed to have a dedicated output directory should never pollute any other location, so this is an error now. If anyone ever wants it back, they can use build_dir = ".".

New features

Required index fields

It’s now possible to mark soecific index fields as required. If a required field is not present in a page, soupault will display an error and stop.


  selector = ["h1#post-title", "h1"]
  required = true

Processing certain pages before all other

By default, soupault processes content pages in essentially random order. What if a page serves as a source of persistent data for a Lua plugin? Or what if you are experimenting with some new code on that page and want the build to fail very early?

Now there’s an option to tell soupault which pages to process first:

  process_pages_first = [""]

Note that it must be a content page, not an index page (typically, not a page named index.*, unless you set your own settings.index_page option). Index pages are always processed after content pages so that extracted metadata can be inserted into them.

Page processing hooks

The first big feature in this release is the long-promised system of page processing hooks.

This release has the following hooks: pre-parse, pre-process, post-index, render, and save.

For example, this is how you can do global variable substitution with a pre-parse hook:

  lua_source = '''
soupault_release = soupault_config["custom_options"]["latest_soupault_version"]
Log.debug("running pre-parse hook")
page_source = Regex.replace_all(page_source, "\\$SOUPAULT_RELEASE\\$", soupault_release)

Lua index processors

The second big feature is an option to write site index rendering code in Lua, in addition to the old options to specify a Jingoo template or use an external script.

Here’s a reimplementation of the built-in index_template option in Lua:

  index_selector = "div#blog-index"

  index_template = """
    {% for e in entries %}
      <div class="entry">
        <a href="{{e.url}}">{{e.title}}</a> (<time>{{}}</time>)
    {% endfor %}

  lua_source = """
    env = {}
    rendered_entries = HTML.parse(String.render_template(config["index_template"], env))
    container = HTML.select_one(page, config["index_selector"])
    HTML.append_child(container, rendered_entries)

However, it’s not just a new way to do the same thing: it’s also a way to create paginated index pages and generate taxonomies.

Here’s the plot twist: Lua index processors can ask soupault to inject virtual pages into the processing queue by storing them in a special variable named pages. Each page is a record with two fields: page_file (the path it would be at if it was a normal page) and page_content (the HTML source code of the page).

Soupault will then add those pages to the queue and process them as if they were real pages loaded from the disk. Of course, it makes sure not to extract metadata from them and not to allow them to generate more pages since that would lead to a feedback loop and infinite multiplication of autogenerated pages.

This way, you can generate the complete website in a single soupault run, but you also have total control over the process and can make arbitrary taxonomy generators and paginators.

Accessing the index entry of the page from plugins

By default, soupault separates the list of all pages into “content” pages and “index” pages, extracts metadata from “content” pages and inserts rendered metadata in “index” pages.

That approach works well for most websites, but sometimes content pages need access to the site index as well. For example, if you want your soupault setup to work like Sphinx or mdBook and insert a site-wide ToC in every page, you need the complete site index metadata available to every page.

Obviously, this is a chicken and egg problem. Since soupault doesn’t use “front matter”, but rather extracts metadata directly from HTML and allows Lua plugins to generate new source material for index field extraction, every page needs to be at least partially processed to retrieve its index entry.

Until this release, the only way to do that was to dump the index data to JSON and read it from plugins on a second pass. Now a two-pass approach is a built-in feature. You can enable it with index.index_first = true.

  index_first = true

When set to true, that option will make soupault perform a reduced first pass where it does the bare minimum of work required to produce the site index. It will read pages and run widgets specified in index.extract_after_widgets, but will not finish processing any pages and will not write them to disk.

Then it will perform a second pass to actually render the website. Every plugin running on every page can access that page’s index entry via a new index_entry variable. This way, you can avoid storing index data externally and run soupault twice, even though a certain amount of work is still done twice behind the scenes.

Per-view index entry sorting settings

Before this release, if you wanted to use a different sorting key or comparison method for different index views, you’d have to do it in the index processor script or template.

Now you can just specify sort_by and sort_type options in your views.

  index_selector = "#blog-index"
  sort_by = "date"
  sort_type = "calendar"

Index view actions

Just like inclusion widgets, index views now allow you to choose how generated HTML will be inserted in the page using the same action option.

Inline Lua plugins

Lua plugin code can now be inlined in soupault.toml:

  lua_source = '''
    Log.debug("Test plugin!")
    Plugin.exit("Test plugin executed successfully")

Link rewriting widgets (relative_links and absolute_links now support all imaginable HTML elements and attributes that may contain URLs, from <bgsound> (that no browser supports anymore) to <portal> (that may never see mainstream browser support). Well, I might have missed some, but for all practical purposes, it supports everything now.

The most practically important improvement is support for srcset: <img srcset="foo1.png 1x, foo2.png 2x"> can be converted to <<img srcset="../foo1.png 1x, ../foo2.png 2x"> by relative_links or to <img srcset=" 1x, 2x"> by absolute_links.

New Lua plugin API functions and variables

New variables

New functions

Unicode string functions

New command line options


Bug fixes

Behaviour changes

Website improvements

Future plans

The introduction of the hook system is a big milestone, but there are other long-standing plans that require significant design and development effort.

Caching and incremental builds

Right now, soupault does not have any caching and always rebuilds everything. It isn’t a problem for most websites. For example, this website takes me just four seconds to build, despite its heavy use of external helpers (pandoc for Markdown pages import, highlight for syntax highlighting).

For very large sites with thousands pages, caching can make builds faster and more environmentally-friendly though. I’m by no means opposed to incremental builds, and I’m going to work on it in future versions, but I don’t have a complete design in mind yet.

The problem is that there are lots of events that must invalidate at least some parts of the cache, other than page source changes. If the soupault config changed or any of the plugin files changed, there’s a good chance that the cache is no longer entirely fully valid. And that’s just easily detectable changes! What if a third-party tool used as a page preprocessor was upgraded for a bug fix? What if a data file referenced by a Lua plugin has changed?

I haven’t studied caching architectures of any other SSGs yet, so I don’t know how they try to solve these problems, if they do. If you want incremental builds and caching, please let me know how exactly you want them to work.

Embedded scripting language improvements

Lua-ML, the Lua interpreter that soupault uses for its plugin API, was way ahead of its time in terms of integration with the host program. Since it’s written in OCaml, it automatically solves the memory safety issue since the “Lua values” are just original OCaml values in thin wrappers. It goes much further than that, though, and also allows users to add their own modules and even replace built-in modules via functors.

However, the language it implements wasn’t ahead of its time, even when Lua-ML was new. It’s roughly Lua 2.5. Fair for its day—Lua-ML is a research project from the early ’00s, revived after almost two decades of dormancy but still ancient. No anonymous functions, in particular—a maddening misfeature for any functional programming enthusiast. Its syntax and runtime error reporting is also very rudimentary.

I’m planning to spend some time modernizing its parser and the AST to provide better errors and to bring at least some features of the modern Lua to it. Then other OCaml projects can also benefit from it if they want an embedded scripting language.

Multicore support

Right now, soupault is strictly sequential and doesn’t support processing anything in parallel at all. It’s still fast despite that fact, thanks to being a native code executable written in a fast language.

Still, there may be great opportunities to make soupault even faster by parallelizing it, since in many setups it spends quite a lot of time waiting for external helpers to produce their output, and different pages can definitely be processed in parallel.

Multicore OCaml is already usable, and it’s very likely to become official this year, which makes that task easier. I also laid the groundwork for parallelization back in 2020 to structure the main processing code as a List.fold_left over a list of pages, which is fairly easy to replace with a parallel fold.

Ironically, the biggest blocker for it is Lua-ML. It has a lot of mutable state inside, and even its parser isn’t thread-safe: when I experimented with the multicore branch and tried a parallel fold, the OCaml parts worked well, but Lua code from plugins was failing to parse due to race conditions. That’s understandable because Lua-ML is a resurrected research project that was written in a time when multi-processor computers were rare and very expensive, and writing purely functional code wasn’t a common trend yet. Still, now it needs to be modernized before it can be used in multithreaded code.

We’ll see how it works out.


Special thanks to Hristos N. Triantofillou, Lulu Cathrinus Grimalkin, and Soni for testing new features and finding bugs.

1Soupault considers its config syntax and behavior a part of its public API. If any website stops building or produces functionally different output after an upgrade, it’s a breaking change—no matter the cause.