Implement composes from CSS modules (#20)

This release implements the composes annotation from the CSS modules specification. It provides a way for class selectors to reference other class selectors (assuming you are using the local-css loader). And with the from syntax, this can even work with local names across CSS files. For example:

// app.js
import { submit } from './style.css'
const div = document.createElement('div')
div.className = submit
document.body.appendChild(div)

/* style.css */
.button {
  composes: pulse from "anim.css";
  display: inline-block;
}
.submit {
  composes: button;
  font-weight: bold;
}

/* anim.css */
@keyframes pulse {
  from, to { opacity: 1 }
  50% { opacity: 0.5 }
}
.pulse {
  animation: 2s ease-in-out infinite pulse;
}

Bundling this with esbuild using --bundle --outdir=dist --loader:.css=local-css now gives the following:

(() => {
  // style.css
  var submit = "anim_pulse style_button style_submit";

  // app.js
  var div = document.createElement("div");
  div.className = submit;
  document.body.appendChild(div);
})();

/* anim.css */
@keyframes anim_pulse {
  from, to {
    opacity: 1;
  }
  50% {
    opacity: 0.5;
  }
}
.anim_pulse {
  animation: 2s ease-in-out infinite anim_pulse;
}

/* style.css */
.style_button {
  display: inline-block;
}
.style_submit {
  font-weight: bold;
}

Import paths in the composes: ... from syntax are resolved using the new composes-from import kind, which can be intercepted by plugins during import path resolution when bundling is enabled.

Note that the order in which composed CSS classes from separate files appear in the bundled output file is deliberately undefined by design (see the specification for details). You are not supposed to declare the same CSS property in two separate class selectors and then compose them together. You are only supposed to compose CSS class selectors that declare non-overlapping CSS properties.

Issue #20 (the issue tracking CSS modules) is esbuild's most-upvoted issue! With this change, I now consider esbuild's implementation of CSS modules to be complete. There are still improvements to make and there may also be bugs with the current implementation, but these can be tracked in separate issues.

Fix non-determinism with tsconfig.json and symlinks (#3284)

This release fixes an issue that could cause esbuild to sometimes emit incorrect build output in cases where a file under the effect of tsconfig.json is inconsistently referenced through a symlink. It can happen when using npm link to create a symlink within node_modules to an unpublished package. The build result was non-deterministic because esbuild runs module resolution in parallel and the result of the tsconfig.json lookup depended on whether the import through the symlink or not through the symlink was resolved first. This problem was fixed by moving the realpath operation before the tsconfig.json lookup.

Add a hash property to output files (#3084, #3293)

As a convenience, every output file in esbuild's API now includes a hash property that is a hash of the contents field. This is the hash that's used internally by esbuild to detect changes between builds for esbuild's live-reload feature. You may also use it to detect changes between your own builds if its properties are sufficient for your use case.

This feature has been added directly to output file objects since it's just a hash of the contents field, so it makes conceptual sense to store it in the same location. Another benefit of putting it there instead of including it as a part of the watch mode API is that it can be used without watch mode enabled. You can use it to compare the output of two independent builds that were done at different times.

The hash algorithm (currently XXH64) is implementation-dependent and may be changed at any time in between esbuild versions. If you don't like esbuild's choice of hash algorithm then you are welcome to hash the contents yourself instead. As with any hash algorithm, note that while two different hashes mean that the contents are different, two equal hashes do not necessarily mean that the contents are equal. You may still want to compare the contents in addition to the hashes to detect with certainty when output files have been changed.

Avoid generating duplicate prefixed declarations in CSS (#3292)

There was a request for esbuild's CSS prefixer to avoid generating a prefixed declaration if a declaration by that name is already present in the same rule block. So with this release, esbuild will now avoid doing this:

/* Original code */
body {
  backdrop-filter: blur(30px);
  -webkit-backdrop-filter: blur(45px);
}

/* Old output (with --target=safari12) */
body {
  -webkit-backdrop-filter: blur(30px);
  backdrop-filter: blur(30px);
  -webkit-backdrop-filter: blur(45px);
}

/* New output (with --target=safari12) */
body {
  backdrop-filter: blur(30px);
  -webkit-backdrop-filter: blur(45px);
}

This can result in a visual difference in certain cases (for example if the browser understands blur(30px) but not blur(45px), it will be able to fall back to blur(30px)). But this change means esbuild now matches the behavior of Autoprefixer which is probably a good representation of how people expect this feature to work.

Fix asset references with the --line-limit flag (#3286)

The recently-released --line-limit flag tells esbuild to terminate long lines after they pass this length limit. This includes automatically wrapping long strings across multiple lines using escaped newline syntax. However, using this could cause esbuild to generate incorrect code for references from generated output files to assets in the bundle (i.e. files loaded with the file or copy loaders). This is because esbuild implements asset references internally using find-and-replace with a randomly-generated string, but the find operation fails if the string is split by an escaped newline due to line wrapping. This release fixes the problem by not wrapping these strings. This issue affected asset references in both JS and CSS files.

Support local names in CSS for @keyframe, @counter-style, and @container (#20)

This release extends support for local names in CSS files loaded with the local-css loader to cover the @keyframe, @counter-style, and @container rules (and also animation, list-style, and container declarations). Here's an example:

@keyframes pulse {
  from, to { opacity: 1 }
  50% { opacity: 0.5 }
}
@counter-style moon {
  system: cyclic;
  symbols: 🌕 🌖 🌗 🌘 🌑 🌒 🌓 🌔;
}
@container squish {
  li { float: left }
}
ul {
  animation: 2s ease-in-out infinite pulse;
  list-style: inside moon;
  container: squish / size;
}

With the local-css loader enabled, that CSS will be turned into something like this (with the local name mapping exposed to JS):

@keyframes stdin_pulse {
  from, to {
    opacity: 1;
  }
  50% {
    opacity: 0.5;
  }
}
@counter-style stdin_moon {
  system: cyclic;
  symbols: 🌕 🌖 🌗 🌘 🌑 🌒 🌓 🌔;
}
@container stdin_squish {
  li {
    float: left;
  }
}
ul {
  animation: 2s ease-in-out infinite stdin_pulse;
  list-style: inside stdin_moon;
  container: stdin_squish / size;
}

If you want to use a global name within a file loaded with the local-css loader, you can use a :global selector to do that:

div {
  /* All symbols are global inside this scope (i.e.
   * "pulse", "moon", and "squish" are global below) */
  :global {
    animation: 2s ease-in-out infinite pulse;
    list-style: inside moon;
    container: squish / size;
  }
}

If you want to use @keyframes, @counter-style, or @container with a global name, make sure it's in a file that uses the css or global-css loader instead of the local-css loader. For example, you can configure --loader:.module.css=local-css so that the local-css loader only applies to *.module.css files.

Support strings as keyframe animation names in CSS (#2555)

With this release, esbuild will now parse animation names that are specified as strings and will convert them to identifiers. The CSS specification allows animation names to be specified using either identifiers or strings but Chrome only understands identifiers, so esbuild will now always convert string names to identifier names for Chrome compatibility:

/* Original code */
@keyframes "hide menu" {
  from { opacity: 1 }
  to { opacity: 0 }
}
menu.hide {
  animation: 0.5s ease-in-out "hide menu";
}

/* Old output */
@keyframes "hide menu" { from { opacity: 1 } to { opacity: 0 } }
menu.hide {
  animation: 0.5s ease-in-out "hide menu";
}

/* New output */
@keyframes hide\ menu {
  from {
    opacity: 1;
  }
  to {
    opacity: 0;
  }
}
menu.hide {
  animation: 0.5s ease-in-out hide\ menu;
}

Support An+B syntax and :nth-*() pseudo-classes in CSS

This adds support for the :nth-child(), :nth-last-child(), :nth-of-type(), and :nth-last-of-type() pseudo-classes to esbuild, which has the following consequences:

• The An+B syntax is now parsed, so parse errors are now reported
• An+B values inside these pseudo-classes are now pretty-printed (e.g. a leading + will be stripped because it's not in the AST)
• When minification is enabled, An+B values are reduced to equivalent but shorter forms (e.g. 2n+0 => 2n, 2n+1 => odd)
• Local CSS names in an of clause are now detected (e.g. in :nth-child(2n of :local(.foo)) the name foo is now renamed)

/* Original code */
.foo:nth-child(+2n+1 of :local(.bar)) {
  color: red;
}

/* Old output (with --loader=local-css) */
.stdin_foo:nth-child(+2n + 1 of :local(.bar)) {
  color: red;
}

/* New output (with --loader=local-css) */
.stdin_foo:nth-child(2n+1 of .stdin_bar) {
  color: red;
}

Adjust CSS nesting parser for IE7 hacks (#3272)

This fixes a regression with esbuild's treatment of IE7 hacks in CSS. CSS nesting allows selectors to be used where declarations are expected. There's an IE7 hack where prefixing a declaration with a * causes that declaration to only be applied in IE7 due to a bug in IE7's CSS parser. However, it's valid for nested CSS selectors to start with *. So esbuild was incorrectly parsing these declarations and anything following it up until the next { as a selector for a nested CSS rule. This release changes esbuild's parser to terminate the parsing of selectors for nested CSS rules when a ; is encountered to fix this edge case:

/* Original code */
.item {
  *width: 100%;
  height: 1px;
}

/* Old output */
.item {
  *width: 100%; height: 1px; {
  }
}

/* New output */
.item {
  *width: 100%;
  height: 1px;
}

Note that the syntax for CSS nesting is about to change again, so esbuild's CSS parser may still not be completely accurate with how browsers do and/or will interpret CSS nesting syntax. Expect additional updates to esbuild's CSS parser in the future to deal with upcoming CSS specification changes.

Adjust esbuild's warning about undefined imports for TypeScript import equals declarations (#3271)

In JavaScript, accessing a missing property on an import namespace object is supposed to result in a value of undefined at run-time instead of an error at compile-time. This is something that esbuild warns you about by default because doing this can indicate a bug with your code. For example:

// app.js
import * as styles from './styles'
console.log(styles.buton)

// styles.js
export let button = {}

If you bundle app.js with esbuild you will get this:

▲ [WARNING] Import "buton" will always be undefined because there is no matching export in "styles.js" [import-is-undefined]

    app.js:2:19:
      2 │ console.log(styles.buton)
        │                    ~~~~~
        ╵                    button

  Did you mean to import "button" instead?

    styles.js:1:11:
      1 │ export let button = {}
        ╵            ~~~~~~

However, there is TypeScript-only syntax for import equals declarations that can represent either a type import (which esbuild should ignore) or a value import (which esbuild should respect). Since esbuild doesn't have a type system, it tries to only respect import equals declarations that are actually used as values. Previously esbuild always generated this warning for unused imports referenced within import equals declarations even when the reference could be a type instead of a value. Starting with this release, esbuild will now only warn in this case if the import is actually used. Here is an example of some code that no longer causes an incorrect warning:

// app.ts
import * as styles from './styles'
import ButtonType = styles.Button

// styles.ts
export interface Button {}

Fix a regression with whitespace inside :is() (#3265)

The change to parse the contents of :is() in version 0.18.14 introduced a regression that incorrectly flagged the contents as a syntax error if the contents started with a whitespace token (for example div:is( .foo ) {}). This regression has been fixed.

Add the --serve-fallback= option (#2904)

The web server built into esbuild serves the latest in-memory results of the configured build. If the requested path doesn't match any in-memory build result, esbuild also provides the --servedir= option to tell esbuild to serve the requested path from that directory instead. And if the requested path doesn't match either of those things, esbuild will either automatically generate a directory listing (for directories) or return a 404 error.

Starting with this release, that last step can now be replaced with telling esbuild to serve a specific HTML file using the --serve-fallback= option. This can be used to provide a "not found" page for missing URLs. It can also be used to implement a single-page app that mutates the current URL and therefore requires the single app entry point to be served when the page is loaded regardless of whatever the current URL is.

Use the tsconfig field in package.json during extends resolution (#3247)

This release adds a feature from TypeScript 3.2 where if a tsconfig.json file specifies a package name in the extends field and that package's package.json file has a tsconfig field, the contents of that field are used in the search for the base tsconfig.json file.

Implement CSS nesting without :is() when possible (#1945)

Previously esbuild would always produce a warning when transforming nested CSS for a browser that doesn't support the :is() pseudo-class. This was because the nesting transform needs to generate an :is() in some complex cases which means the transformed CSS would then not work in that browser. However, the CSS nesting transform can often be done without generating an :is(). So with this release, esbuild will no longer warn when targeting browsers that don't support :is() in the cases where an :is() isn't needed to represent the nested CSS.

In addition, esbuild's nested CSS transform has been updated to avoid generating an :is() in cases where an :is() is preferable but there's a longer alternative that is also equivalent. This update means esbuild can now generate a combinatorial explosion of CSS for complex CSS nesting syntax when targeting browsers that don't support :is(). This combinatorial explosion is necessary to accurately represent the original semantics. For example:

/* Original code */
.first,
.second,
.third {
  & > & {
    color: red;
  }
}

/* Old output (with --target=chrome80) */
:is(.first, .second, .third) > :is(.first, .second, .third) {
  color: red;
}

/* New output (with --target=chrome80) */
.first > .first,
.first > .second,
.first > .third,
.second > .first,
.second > .second,
.second > .third,
.third > .first,
.third > .second,
.third > .third {
  color: red;
}

This change means you can now use CSS nesting with esbuild when targeting an older browser that doesn't support :is(). You'll now only get a warning from esbuild if you use complex CSS nesting syntax that esbuild can't represent in that older browser without using :is(). There are two such cases:

/* Case 1 */
a b {
  .foo & {
    color: red;
  }
}

/* Case 2 */
a {
  > b& {
    color: red;
  }
}

These two cases still need to use :is(), both for different reasons, and cannot be used when targeting an older browser that doesn't support :is():

/* Case 1 */
.foo :is(a b) {
  color: red;
}

/* Case 2 */
a > a:is(b) {
  color: red;
}

Automatically lower inset in CSS for older browsers

With this release, esbuild will now automatically expand the inset property to the top, right, bottom, and left properties when esbuild's target is set to a browser that doesn't support inset:

/* Original code */
.app {
  position: absolute;
  inset: 10px 20px;
}

/* Old output (with --target=chrome80) */
.app {
  position: absolute;
  inset: 10px 20px;
}

/* New output (with --target=chrome80) */
.app {
  position: absolute;
  top: 10px;
  right: 20px;
  bottom: 10px;
  left: 20px;
}

Add support for the new @starting-style CSS rule (#3249)

This at rule allow authors to start CSS transitions on first style update. That is, you can now make the transition take effect when the display property changes from none to block.

/* Original code */
@starting-style {
  h1 {
    background-color: transparent;
  }
}

/* Output */
@starting-style{h1{background-color:transparent}}

This was contributed by @yisibl.

Implement local CSS names (#20)

This release introduces two new loaders called global-css and local-css and two new pseudo-class selectors :local() and :global(). This is a partial implementation of the popular CSS modules approach for avoiding unintentional name collisions in CSS. I'm not calling this feature "CSS modules" because although some people in the community call it that, other people in the community have started using "CSS modules" to refer to something completely different and now CSS modules is an overloaded term.

Here's how this new local CSS name feature works with esbuild:

• Identifiers that look like .className and #idName are global with the global-css loader and local with the local-css loader. Global identifiers are the same across all files (the way CSS normally works) but local identifiers are different between different files. If two separate CSS files use the same local identifier .button, esbuild will automatically rename one of them so that they don't collide. This is analogous to how esbuild automatically renames JS local variables with the same name in separate JS files to avoid name collisions.

• It only makes sense to use local CSS names with esbuild when you are also using esbuild's bundler to bundle JS files that import CSS files. When you do that, esbuild will generate one export for each local name in the CSS file. The JS code can import these names and use them when constructing HTML DOM. For example:

  // app.js
  import { outerShell } from './app.css'
  const div = document.createElement('div')
  div.className = outerShell
  document.body.appendChild(div)
  

  /* app.css */
  .outerShell {
    position: absolute;
    inset: 0;
  }
  

  When you bundle this with esbuild app.js --bundle --loader:.css=local-css --outdir=out you'll now get this (notice how the local CSS name outerShell has been renamed):

  // out/app.js
  (() => {
    // app.css
    var outerShell = "app_outerShell";
  
    // app.js
    var div = document.createElement("div");
    div.className = outerShell;
    document.body.appendChild(div);
  })();
  

  /* out/app.css */
  .app_outerShell {
    position: absolute;
    inset: 0;
  }
  

  This feature only makes sense to use when bundling is enabled both because your code needs to import the renamed local names so that it can use them, and because esbuild needs to be able to process all CSS files containing local names in a single bundling operation so that it can successfully rename conflicting local names to avoid collisions.

• If you are in a global CSS file (with the global-css loader) you can create a local name using :local(), and if you are in a local CSS file (with the local-css loader) you can create a global name with :global(). So the choice of the global-css loader vs. the local-css loader just sets the default behavior for identifiers, but you can override it on a case-by-case basis as necessary. For example:

  :local(.button) {
    color: red;
  }
  :global(.button) {
    color: blue;
  }
  

  Processing this CSS file with esbuild with either the global-css or local-css loader will result in something like this:

  .stdin_button {
    color: red;
  }
  .button {
    color: blue;
  }
  

• The names that esbuild generates for local CSS names are an implementation detail and are not intended to be hard-coded anywhere. The only way you should be referencing the local CSS names in your JS or HTML is with an import statement in JS that is bundled with esbuild, as demonstrated above. For example, when --minify is enabled esbuild will use a different name generation algorithm which generates names that are as short as possible (analogous to how esbuild minifies local identifiers in JS).

• You can easily use both global CSS files and local CSS files simultaneously if you give them different file extensions. For example, you could pass --loader:.css=global-css and --loader:.module.css=local-css to esbuild so that .css files still use global names by default but .module.css files use local names by default.

• Keep in mind that the css loader is different than the global-css loader. The :local and :global annotations are not enabled with the css loader and will be passed through unchanged. This allows you to have the option of using esbuild to process CSS containing while preserving these annotations. It also means that local CSS names are disabled by default for now (since the css loader is currently the default for CSS files). The :local and :global syntax may be enabled by default in a future release.

Note that esbuild's implementation does not currently have feature parity with other implementations of modular CSS in similar tools. This is only a preliminary release with a partial implementation that includes some basic behavior to get the process started. Additional behavior may be added in future releases. In particular, this release does not implement:

• The composes pragma
• Tree shaking for unused local CSS
• Local names for keyframe animations, grid lines, @container, @counter-style, etc.

Issue #20 (the issue for this feature) is esbuild's most-upvoted issue! While this release still leaves that issue open, it's an important first step in that direction.

Parse :is, :has, :not, and :where in CSS

With this release, esbuild will now parse the contents of these pseudo-class selectors as a selector list. This means you will now get syntax warnings within these selectors for invalid selector syntax. It also means that esbuild's CSS nesting transform behaves slightly differently than before because esbuild is now operating on an AST instead of a token stream. For example:

/* Original code */
div {
  :where(.foo&) {
    color: red;
  }
}

/* Old output (with --target=chrome90) */
:where(.foo:is(div)) {
  color: red;
}

/* New output (with --target=chrome90) */
:where(div.foo) {
  color: red;
}

Add the --drop-labels= option (#2398)

If you want to conditionally disable some development-only code and have it not be present in the final production bundle, right now the most straightforward way of doing this is to use the --define: flag along with a specially-named global variable. For example, consider the following code:

function main() {
  DEV && doAnExpensiveCheck()
}

You can build this for development and production like this:

• Development: esbuild --define:DEV=true
• Production: esbuild --define:DEV=false

One drawback of this approach is that the resulting code crashes if you don't provide a value for DEV with --define:. In practice this isn't that big of a problem, and there are also various ways to work around this.

However, another approach that avoids this drawback is to use JavaScript label statements instead. That's what the --drop-labels= flag implements. For example, consider the following code:

function main() {
  DEV: doAnExpensiveCheck()
}

With this release, you can now build this for development and production like this:

• Development: esbuild
• Production: esbuild --drop-labels=DEV

This means that code containing optional development-only checks can now be written such that it's safe to run without any additional configuration. The --drop-labels= flag takes comma-separated list of multiple label names to drop.

Avoid causing unhandledRejection during shutdown (#3219)

All pending esbuild JavaScript API calls are supposed to fail if esbuild's underlying child process is unexpectedly terminated. This can happen if SIGINT is sent to the parent node process with Ctrl+C, for example. Previously doing this could also cause an unhandled promise rejection when esbuild attempted to communicate this failure to its own child process that no longer exists. This release now swallows this communication failure, which should prevent this internal unhandled promise rejection. This change means that you can now use esbuild's JavaScript API with a custom SIGINT handler that extends the lifetime of the node process without esbuild's internals causing an early exit due to an unhandled promise rejection.

Update browser compatibility table scripts

The scripts that esbuild uses to compile its internal browser compatibility table have been overhauled. Briefly:

• Converted from JavaScript to TypeScript
• Fixed some bugs that resulted in small changes to the table
• Added caniuse-lite and @mdn/browser-compat-data as new data sources (replacing manually-copied information)

This change means it's now much easier to keep esbuild's internal compatibility tables up to date. You can review the table changes here if you need to debug something about this change:

• JS table changes
• CSS table changes

Fix a panic with const enum inside parentheses (#3205)

This release fixes an edge case where esbuild could potentially panic if a TypeScript const enum statement was used inside of a parenthesized expression and was followed by certain other scope-related statements. Here's a minimal example that triggers this edge case:

(() => {
  const enum E { a };
  () => E.a
})

Allow a newline in the middle of TypeScript export type statement (#3225)

Previously esbuild incorrectly rejected the following valid TypeScript code:

export type
{ T };

export type
* as foo from 'bar';

Code that uses a newline after export type is now allowed starting with this release.

Fix cross-module inlining of string enums (#3210)

A refactoring typo in version 0.18.9 accidentally introduced a regression with cross-module inlining of string enums when combined with computed property accesses. This regression has been fixed.

Rewrite .js to .ts inside packages with exports (#3201)

Packages with the exports field are supposed to disable node's path resolution behavior that allows you to import a file with a different extension than the one in the source code (for example, importing foo/bar to get foo/bar.js). And TypeScript has behavior where you can import a non-existent .js file and you will get the .ts file instead. Previously the presence of the exports field caused esbuild to disable all extension manipulation stuff which included both node's implicit file extension searching and TypeScript's file extension swapping. However, TypeScript appears to always apply file extension swapping even in this case. So with this release, esbuild will now rewrite .js to .ts even inside packages with exports.

Fix a redirect edge case in esbuild's development server (#3208)

The development server canonicalizes directory URLs by adding a trailing slash. For example, visiting /about redirects to /about/ if /about/index.html would be served. However, if the requested path begins with two slashes, then the redirect incorrectly turned into a protocol-relative URL. For example, visiting //about redirected to //about/ which the browser turns into http://about/. This release fixes the bug by canonicalizing the URL path when doing this redirect.

Fix a TypeScript code generation edge case (#3199)

This release fixes a regression in version 0.18.4 where using a TypeScript namespace that exports a class declaration combined with --keep-names and a --target of es2021 or earlier could cause esbuild to export the class from the namespace using an incorrect name (notice the assignment to X2._Y vs. X2.Y):

// Original code

// Old output (with --keep-names --target=es2021)
var X;
((X2) => {
  const _Y = class _Y {
  };
  __name(_Y, "Y");
  let Y = _Y;
  X2._Y = _Y;
})(X || (X = {}));

// New output (with --keep-names --target=es2021)
var X;
((X2) => {
  const _Y = class _Y {
  };
  __name(_Y, "Y");
  let Y = _Y;
  X2.Y = _Y;
})(X || (X = {}));

Fix a tree-shaking bug that removed side effects (#3195)

This fixes a regression in version 0.18.4 where combining --minify-syntax with --keep-names could cause expressions with side effects after a function declaration to be considered side-effect free for tree shaking purposes. The reason was because --keep-names generates an expression statement containing a call to a helper function after the function declaration with a special flag that makes the function call able to be tree shaken, and then --minify-syntax could potentially merge that expression statement with following expressions without clearing the flag. This release fixes the bug by clearing the flag when merging expression statements together.

Fix an incorrect warning about CSS nesting (#3197)

A warning is currently generated when transforming nested CSS to a browser that doesn't support :is() because transformed nested CSS may need to use that feature to represent nesting. This was previously always triggered when an at-rule was encountered in a declaration context. Typically the only case you would encounter this is when using CSS nesting within a selector rule. However, there is a case where that's not true: when using a margin at-rule such as @top-left within @page. This release avoids incorrectly generating a warning in this case by checking that the at-rule is within a selector rule before generating a warning.