It would be great if IkiWiki could apply some sort of preprocessing to CSS. I'm just "thinking out loud" at the moment, but I might write a plugin.

The simplest starting point would be concatenating a list of files:

  • The actiontabs and monochrome themes are not ready for use as-is; the Makefile constructs the real stylesheet by concatenating the anti-theme's CSS and the relevant theme's CSS. It would be better if they could just drop a file in an underlay, and IkiWiki would concatenate the anti-theme stylesheet and the theme's file.

  • The blueview theme is the same, but the theme's CSS also starts with a couple of stylesheets from YUI. IkiWiki could maybe concatenate the anti-theme stylesheet, the two YUI stylesheets and the blueview-specific part? Or maybe that's too complicated.

  • The goldtype theme is the blueview theme with some overrides added to the end. Again, IkiWiki could concatenate all the pieces.

  • The album plugin needs to append some stuff to the stylesheet, making its installation more involved than it ought to be. If it could append the pieces by putting them in an underlay, installation would just be a matter of "add the files".

I'm not sure whether local.css should be concatenated too, or whether it should be separate.

It would also be great if the same mechanism could be extended to compile CSS-like languages like SASS or LESS down to CSS, but for dependency reasons, I don't think the themes shipped with IkiWiki should rely on that.

If the compiled CSS ended up with a content-based filename (perhaps ikiwiki/compiled-HASH.css where HASH is the (possibly truncated) MD5 or SHA1 hash of the content), that would mitigate stale CSS being served from cache (as long as the HTML has a short expiry, which is desirable anyway), and ikiwiki-hosting could maybe even do something like this to allow long-term caching of the files with content-based names:

<LocationMatch /ikiwiki/compiled-[a-f0-9]+\.css>
    ExpiresByType text/css "now plus 1 year"

A similar mechanism could maybe be used to minify JavaScript.

vague syntax proposal: controlled by directive

[[!css  files="""
  • css directive, placed in any page, concatenates the given files and uses them as the stylesheet for that page and its subpages, replacing <TMPL_VAR BASEURL>style.css

  • the files can be globs, which are sorted lexicographically within a glob, and do not have to match anything (so it's OK if you don't have anything that matches plugin*.css); "-" happens to sort before ".", so theme-base.css would sort before theme.css, which is nice to have

  • the default would be style.css, then ikiwiki/plugin*.css, then ikiwiki/theme*.css and maybe local.css, as in the example above

  • style.css would continue to be the anti-theme

  • themes would ship ikiwiki/theme.css instead of style.css, resulting in concatenation

  • goldtype would ship ikiwiki/theme-base.css (which is a copy of blueview, or a symlink to blueview if we can make symlinks safe) and ikiwiki/theme.css (which is the goldtype additions)

  • album would put its templates and ikiwiki/plugin-album.css in an underlay

  • any non-contrib plugin with complicated CSS requirements could also move its CSS to an underlay

I think this could be done entirely in a plugin, except for this change to page.tmpl to allow the <elyts>s to be overridden:

    <elyts type="text/css" href="<TMPL_VAR BASEURL><TMPL_VAR COMPILED_CSS>" />
    <!-- ... what it does now ... -->

The plugin could also optionally minify its output, and either pass input files through an external SASS/SCSS/LESS implementation according to their extension, or have a hook system to do so.

People who want SASS/LESS would probably just do this in their top-level page:

[[!css  files="mywiki.less"]]

and do the inclusion/concatenation logic via @import.

Security consideration: the SASS/LESS compilers will read arbitrary local files, so if you use those languages, ability to upload the appropriate extension would have to be locked-down. Perhaps it's better to implement this without that feature initially.


Although I understand the need to improve CSS inclusion, I wonder why you are proposing concatenating CSS rather than including them as several <link type="text/css" href="FILE.css" rel="stylesheet"> lines in the header: unless I am missing something, I see this as far more simpler than concatenating them.

This would imply that a template variable CSS is added to the page template, to be filled with those lines.

Whatever solution is used, I agree that such a thing would be useful: adding CSS (rather than replacing the existing one) should be easier.

-- Louis

One big request is more efficient than lots of small requests, if we model the CSS as all changing equally infrequently. In terms of bytes, each file needs some code in the HTML <head>, plus the HTTP request and response headers, plus the actual file. On the first page-view, a visitor will have to download all the CSS anyway (one request/response pair per CSS file). On subsequent page-views, there will be one request/"304 Not Modified" response per CSS file, unless the CSS files can be marked "to be cached forever" (which can be done if they have content-based filenames).

In terms of time, according to Wikipedia browsers don't generally pipeline requests, so the page won't finish loading until one round-trip time per uncached CSS file has elapsed.

Having lots of small files with content-based filenames would be the next best thing - not particularly efficient on a generic web server, but they could at least be marked as "cache forever" in server configuration. I'd be OK with doing that if it makes ikiwiki more maintainable, but I don't think concatenating all the CSS at compile time is actually going to be a problem in practice. The individual small files are still going to be available for the wiki operator to edit.

If some CSS files change with a significantly different frequency, then it might become worthwhile to separate them, but I don't think that's the case (apart from possibly local.css, which is why I'm not sure whether to include it in this). --smcv

I must admit that I am not aware of how those several CSS inclusion lines tend to make browsing less smooth. Please withdraw my comment.

As you pointed out, CSS inclusion is more painful than it should be, and your proposal seems to answer that. Go ahead! --Louis

Concatenating the theme css as is done now results in files that are unecessarily large with a doubling of a lot of selectors etc. It only makes sense for changes that should be local.css anyway. Catted css is inefficient both while downloading and while rendering. I've disabled the catting in the makefile to avoid this on my personal site. In my view it would be better for theme developers to work from the basewiki style, if lazy just add their changes to the end of it, or if speed is of secondary importance @import it.

The advanced melding of stylesheets discussed sounds quite complicated with likely useability problems when the site don't quite look as expected. Hunting down the problematic css will be difficult.

Are there parsers that remove double defined selectors according to cascading rules as is done in browser? This would at least produce cleaner css but the useability problems would remain.

When using complete themes and hunting that last bit of speed a config option to turn off local.css would probably be a good idea? Plugin css is difficult. A choice between a plugin complete theme or a local.css (or @import from it) would be a simple solution that lets you choose how you prioritize speed vs convenience. --kjs