v0.7.7

• Fix overscroll util

v0.7.6

• Align with Tailwind 3.4.6

v0.7.5

• Align base.css with Tailwind 3.4.2
• Support named color (ie text-[red] generate a color utility)
• Skip merging of utilities that contains vendor prefixes

v0.7.4

• Actually fix plugin usage in build watch mode

v0.7.3

• Align with Tailwind 3.4: https://tailwindcss.com/blog/tailwindcss-v3-4
• Support main dynamic variants (min-*, max-*, has-*, (group/peer-)data-*, (group/peer-)aria-*)

v0.7.2

• Fix plugin usage in build watch mode

v0.7.1

• Sort defaults to get a stable output
• Internally consider media & supports variant as atRules so that the nesting output is closer to Tailwind
• Group selectors with same content. Pre 0.7 this optimization was done by Lightning CSS and esbuild does not support this optimization when minifying

v0.7.0

Remove dependency on LightningCSS

This was here mostly because I wanted to get CSS modules and Tailwind working with esbuild. Now that esbuild support CSS modules, I can remove this coupling and makes this repo easier to re-use in other bundlers. This also mean I'm dropping from this repo features related to build tools, like downwind.transform, cssModuleToJS & convertTargets. The rest of the core downwind object API has also been updated to give more flexibility outside of built-in plugins.

For usage Vite, you can get back the same behaviour by using the builtin support for lightningCSS:

export default defineConfig({
  plugins: [downwind()],
  css: {
    transformer: "lightningcss",
  },
  build: {
    cssMinify: "lightningcss",
  },
});

Interval check during builds for plugins

Using the bundler to discover file to scan is the main reason of this fork. But it poses a problem for builds: we need to be sure to have scan every UI files before generating the output.

Previously this was done line in unocss by injecting a placeholder, and inside the "onEnd" callback, replacing it with the generated output. But this mess up with sourcemap and content hashing, and the plugins where doing some hack around this.

To simplify the code, I now delayed the generation the virtual utils module until other files are scanned. But there is no API to be sure every other files is processed, so for now the build is checking at a fixed interval if some work was done in the previous Xms and if not, generate the utils. This interval is configurable and default to 50ms for the esbuild plugin and 200ms for the Vite plugin.

This is not perfect, but I think that the cost of waiting few hundred milliseconds in a build is better than having utils not processed and minified like the rest of the CSS files.

Fix important modifier with variant

There was a involuntary mismatch with Tailwind when applying the important modifier (!) with a variant, the implementation required to use !hover:font-medium instead of hover:!font-medium. This has been changed to match Tailwind syntax.

Scanning update

The read of https://marvinh.dev/blog/speeding-up-javascript-ecosystem-part-8/ make me realize the current regex approach was really inefficient: instead of search for "strings" and then for classes inside strings, we should directly grep all patterns that looks like a Tailwind classes, and by having a strict subset on the left border ('"`\s}), the first character (a-z0-9![-), the last char (a-z0-9%]) and the right border ('"`\s:$), we can get a good ratio of matches. This means that custom utils & shortcuts should be at least 2 chars.

This change uses a lookbehind assertion for the left border, which means that if a playground was made, it would not work with Safari before 16.4.

This new approach also allow for quotes inside custom values, which makes before:content-['hello_world'] & [&[data-selected="true"]]:bg-blue-100 now possible in downwind.

The parser has been modified to parse arbitrary values before modifiers (opacity, line height) so that / can be used inside arbitrary values. text-[calc(3rem/5)]/[calc(4rem/5)] is now supported.

And the nice part is that it's also quite fast. When running (mac M1) on 281 tsx files of my production codebase, time running regex went from 90ms to 13ms (and 125543 to 33992 candidates). For the total time (init, scan & CSS generation), the time went from 117ms to 36ms. The perf update is completely crushed by the 'Interval check' change, but this is nice to see that a new approach with less limitations is also faster!

v0.6.2

Align with Tailwind 3.3.3:

• Updated preflight with reset of dialog padding
• Fix gradients implementation

v0.6.1

Use exports map