Brrrty

Changelog

0.0.10 - 2026-11-17

  • Brrrty is on npm: npm install --save-dev brrrty, then brrrty build in your scripts. The brrrty package is a small launcher that runs a prebuilt binary from @brrrty/darwin-arm64, @brrrty/darwin-x64, @brrrty/linux-arm64-musl or @brrrty/linux-x64-musl, whichever npm installed for the platform. No install scripts and no Rust toolchain needed. The Linux binaries are static, so they run on Alpine and glibc distributions alike. Arguments, output and exit codes pass through unchanged, and Ctrl-C under npm run or npx stops brrrty serve cleanly with exit 0. Ctrl-\ and every other terminating signal Node can catch are passed on too, and brrrty serve stops cleanly by itself if the launcher is killed outright, so it never stays behind holding its port.
  • Each release attaches binaries for those four platforms, with SHA-256 sums, to its GitHub release.
  • Every npm package and release tarball carries NOTICE, with the copyright and license notices of the projects Brrrty ports code and data from, next to LICENSE.
  • On Linux, a page without a date is dated by its file's creation time whatever the C library, as 11ty does. The static musl build used to fall back to the modification time, so a page edited after it was created got a different date (and a different place in date-sorted collections) than in 11ty. Where the creation time is missing, Brrrty now does what 11ty does there too: a file system that doesn't record it dates the page (and date: Last Modified) by the build time, and a system where statx is unavailable uses the file's change time.

0.0.9 - 2026-11-17

  • A build into an empty or new output folder is about 20% faster on macOS. On APFS, many threads creating files under the same folders fight over folder locks, so pages and passthrough files are now written by 4 threads of their own, and passthrough files are copied in parallel on every system. Passthrough copies whose destinations overlap (a file and a folder mapped to one path, or names that differ only by case or Unicode normalization on macOS) still run one at a time in walk order, so the output and the problems reported are the same as before.
  • Passthrough copies keep their source's access and modification times on every system, as 11ty's copy does, so rsync and other deploy tools that compare times no longer re-upload every passthrough file after a build on Linux. On macOS, copies already kept them. Rendered pages get the time they are written, as in 11ty.

0.0.8 - 2026-11-17

brrrty serve:

  • brrrty serve builds the site, serves it, and rebuilds on every change, as eleventy --serve does. It takes --port (8080, moving up to 10 ports on when it is taken), --host (localhost), --input, --output, --passthrough, --quiet and --strict. It prints Server at <url>, File changed: <path> and Watching… as 11ty does.
  • Live reload: pages get a script that patches in an edit to the page's own template, swaps an edited stylesheet, and reloads for anything else, and logs build errors in the browser console, as 11ty's dev server does.
  • The output folder is kept up to date after every build, from the build's own list of what it wrote, so an unchanged file is never read again. Passthrough sources reached through a symlink are copied again on every edit. Builds are made in <output>/.brrrty-serve/, which is never served and is removed on exit; brrrty build removes one a killed session left behind.
  • A running session holds a lock on <output>/.brrrty-serve/ (an OS file lock, with its process id in the file). brrrty build into the same output folder leaves it alone and prints a serve-staging-in-use note, and a second brrrty serve there refuses to start. The OS releases the lock when the process ends, so a killed session's folder is still cleaned up.
  • When no build ever fills the output folder (the first build failed, as under --strict), a folder the session created is removed on exit, as 11ty leaves none.
  • Ctrl-C stops cleanly and exits 0, with the last good build in the output folder and the port released. A second Ctrl-C quits at once. SIGTERM stops the same way and then ends by the signal.
  • eleventy.env.runMode is "serve" while serving.

Where brrrty serve differs from eleventy --serve:

  • It listens on localhost only, unless --host says otherwise, and on that address answers only requests whose Host names this machine.
  • It never serves a file outside the output folder: .. in any encoding and symlinks that point outside get a 404, and malformed URLs a 400.
  • A build with errors in some pages is still served, and its pages that built are written to the output folder, as brrrty build writes them. A page whose template has an error keeps the version last served (the rebuild says how many were kept), and the error shows in the browser console; a page that never built is a 404. A first build with errors keeps the server running. --strict serves no build with an error and exits with status 1 when the first build has one, as 11ty does.
  • Any change under the input folder, a deleted file included, starts a rebuild.
  • The live-reload socket refuses pages from other sites.
  • Paths under /.brrrty/ (reload-client.js, morphdom.js, live) are reserved while serving, as /.11ty/ is in 11ty.
  • --quiet prints only the Server at line, errors and warnings; 11ty's also prints the Wrote line, Watching… and changed files.
  • An added file is reported as File changed: (11ty: File added:), and a changed file outside the project folder by its full path. When the port is taken, the Server at line adds (port N was in use).
  • setServerOptions and serverPassthroughCopyBehavior have no equivalent.

Building Brrrty now needs Rust 1.89 or later.

0.0.7 - 2026-11-17

Nunjucks:

  • .njk templates are built, with a Nunjucks 3.2.4 engine configured as 11ty 3.1.8 configures it: autoescaping on, _includes then the project folder as include roots, and ./ names relative to the including template. This replaces the refusal for .njk files.
  • Pages, layouts and includes can be Nunjucks, in any mix with Liquid and Markdown: a .md page in an .njk layout, and the reverse. An include of any file, such as .js or .css, is rendered as Nunjucks, as in 11ty; one without tags comes out byte for byte. Permalinks and computed data in .njk pages are Nunjucks, and markdownTemplateEngine and htmlTemplateEngine can be "njk".
  • Expressions follow JavaScript: truthiness, == and ===, + on strings and numbers, how undefined and null print, and the string, array, number and date methods a template calls. if, for (with else, loop and objects), set, macro, call and caller, import and from, include (with ignore missing), extends, block and super(), raw, verbatim, filter and switch work, with whitespace control ({%-, -%}).
  • Every Nunjucks built-in filter, 11ty's filters, the RSS plugin's filters and the config's Rhai filters run in Nunjucks. A SafeString stays safe through filters and macros as it does in Nunjucks.
  • Regular expression literals (r/.../gi) work with replace, replaceAll, split, test, the replace filter and printing. They match as JavaScript's do, with the same rules as regex_replace in scripts.
  • What Brrrty can't reproduce exactly is refused with the reason, never approximated: a method with no exact emulation, random, a sort or dictsort whose order would depend on V8's sort, async filters and tags, 11ty's setAsync, getBundle and getBundleFileUrl, printing a function (Nunjucks prints its JavaScript source), the syntax highlighting plugin's Nunjucks {% highlight %} shortcode, a SafeString passed to a config (Rhai) filter or to an RSS filter that every engine has (JavaScript would pass an object, not a string), options to slug and slugify, and a time zone argument to dateToRfc822.
  • The engine has the Liquid engine's safety limits: nesting depth, loop and output size, and a parser that never backtracks. A crash in it refuses the page instead of stopping the build.
  • A render may run 3,000,000 loop iterations in all, across nested loops, includes and macros, so nested loops are stopped in seconds.
  • A filter that lists a collection item's properties (urlencode, dictsort, list, dump) is refused, as for k, v in item is: 11ty prints its template object and content there. Counting or testing the names in collections (length, in, dictsort) is refused when the site has an 11ty config that may add collections.

Config:

  • git_head() gives the commit git rev-parse HEAD prints, for a config or data script that ports execSync("git rev-parse HEAD"). It reads the repository's files (HEAD, loose refs, packed-refs, and a linked worktree's .git file and commondir) from the project folder up, and starts no process. Outside a repository it is an error, which try/catch can turn into a default.
  • A script filter that lists the names in collections (keys(), len(), contains(), for k in) under an 11ty config, or changes a collection item, is refused.
  • map, filter and reduce with an anonymous function run about twice as fast, and regex_split with a pattern without groups is faster.
  • The regular expression matcher behind regex_replace, regex_split and regex_test moved to brrrty-jsregex, where Nunjucks uses it too.

CLI:

  • brrrty build 2>&1 | head -1 no longer aborts with exit code 134 when the pipe closes: printing stops and the build's own exit code stands.

Syntax highlighting:

  • Markdown front matter inside a fence is written highlighted as YAML when a page 11ty renders before it loaded YAML, and refused, as before, when none does. Brrrty replays the order 11ty renders pages' content in.
  • The scan for languages that change grammars counts a Nunjucks {% highlight %} whose argument isn't one string, and any {% highlight whose arguments start on the next line, as any language.

0.0.6 - 2026-11-17

Syntax highlighting:

  • Brrrty now highlights code the way @11ty/eleventy-plugin-syntaxhighlight does. Fenced code blocks in Markdown, the {% highlight %} Liquid tag and the plugin's options (lineSeparator, errorOnInvalidLanguage, alwaysWrapLineHighlights, preAttributes, codeAttributes, trim, languages, templateFormats) give the same HTML, byte for byte. config.add_plugin("syntaxhighlight") takes them as a map. This replaces the refusal for pages with a fenced code block or a {% highlight %} tag.

  • It is a port of Prism 1.30.0 itself: the grammars' 464 distinct regular expressions are translated to Rust regular expressions that match as JavaScript's do, and Prism's tokenizer, hooks and line highlighting (```js/1-2,4, {% highlight js 1,3 %}) and diff- languages run as they do in Node.

  • Brrrty has the grammars of 28 of Prism's 297 languages, the ones the test sites use: bash (sh, shell), c, clike, css, diff, go, ini, javascript (js), json (webmanifest), jsx, less, liquid, lua, markdown (md), markup (html, xml, svg, mathml, ssml, atom, rss), markup-templating, nginx, perl, php, python (py), r, ruby (rb), rust, sql, toml, typescript (ts), vim and yaml (yml), plus Prism's plain (plaintext, txt).

  • A name that isn't a Prism language or alias, such as conf, zsh, console, terraform or JS, is written as the plugin writes it: the code as it is, with a language-x class. So is text. With errorOnInvalidLanguage, such a name other than text fails the page with syntax-highlight-failed, where 11ty stops the build.

  • Highlighting is never approximated. A page whose highlighting Brrrty can't reproduce exactly is refused with the reason in a syntax-highlight-unsupported problem:

    • a fence or {% highlight %} in one of the other 269 Prism languages, such as java, kotlin, swift, cpp, csharp, scss, graphql, docker or shell-session, or njk, which the plugin highlights as django;
    • a page highlighting with a grammar that a language on another page changes. In 11ty, loading regex, css-extras, js-extras, js-templates, jsdoc, javadoclike, phpdoc or php-extras (among 21 such languages) changes the grammars of javascript, css, php, markup and what is built on them for every page rendered after. Brrrty scans every file under the input folder before it builds, Nunjucks pages and pages it refuses for other reasons included, and refuses the pages those grammars would highlight, naming the file and line that loads the language. A language that reaches the highlighter only as pages render, such as a fence built from data, is caught too: every page renders before any is written, and if one meets such a language the build is redone refusing what it changes. The list comes from Prism's components.json and from loading each language in Node;
    • a pattern that would backtrack more than a budget allows (on hostile input, reaching the budget can take a few seconds), and Markdown with a fenced block or front matter inside a fence, which 11ty highlights differently depending on the order it rendered pages in.

    syntax-highlight-failed is for where the plugin itself throws, in a fence or a {% highlight %} tag.

  • Fixed: with the code rule off, Markdown still stops treating lines indented 4 or more spaces as continuing a list item, as markdown-it does.

  • Building is faster for pages with code in languages with long keyword lists, such as SQL: translating a case-insensitive pattern no longer reads Prism's whole case table for every letter.

0.0.5 - 2026-11-16

Config written in Rhai:

  • brrrty.config.rhai at the project root is Brrrty's version of eleventy.config.js. When it exists, Brrrty reads it and ignores the JavaScript config. It sets dir, passthrough copy, ignores, global data, filters, collections and plugins. PORTING.md in brrrty-config shows how to port a JavaScript config, call by call.
  • Filters (config.add_filter) get their arguments the way 11ty passes them to a JavaScript filter, keyword arguments included. A filter that fails refuses only the page that used it, where 11ty stops the build.
  • Collections (config.add_collection) get 11ty's collection API: get_all, get_all_sorted, get_filtered_by_tag, get_filtered_by_tags and get_filtered_by_glob, with 11ty's glob matching. Items are read-only.
  • Helpers for what JavaScript does differently from Rhai: object() keeps keys in the order they were added, truthy, js_string, js_concat, js_div, js_round, js_sort, js_sort_by, json_stringify, and arrays get JavaScript's sort, reverse and slice. Sorting runs V8's own TimSort, so even a comparator that isn't a consistent ordering, such as |a, b| a.date > b.date, gives the order Node gives, at any length. regex_replace, regex_split and regex_test take JavaScript patterns.
  • luxon_format(date, format, zone) formats dates as Luxon does, and markdown(s) and markdown_inline(s) render as markdownIt() does.
  • Data files can be Rhai scripts: _data/name.rhai, page.11tydata.rhai and folder.11tydata.rhai. A script's last expression, or its fn data(global), is the data. A .rhai file replaces a JavaScript data file of the same name, with a note saying so.
  • Two global data files for one key (a.json and a.rhai) merge, as in 11ty.
  • config.add_plugin("rss", #{ ... }) is 11ty's RSS feed plugin: Atom, RSS and JSON feeds, with its filters for Liquid.
  • config.add_plugin("syntaxhighlight") marks a site that uses 11ty's syntax highlighting plugin. Brrrty can't highlight yet, so pages with a fenced code block that names a language, or a {% highlight %} tag, are refused instead of written without highlighting.
  • SOURCE_DATE_EPOCH sets the time for Rhai's now() and for Liquid's 'now' and 'today'.
  • Scripts are limited in operations, time, call depth and size (what object()s hold included), and can't capture variables in closures or read the real clock through Rhai's timestamp(), so a script can't hang the build or give different output from one run to the next. Errors point at the script's file, line and column.
  • A filter written in Rhai that would be given a Liquid nil, null, empty, blank or forloop refuses its page: 11ty passes LiquidJS's own objects for those, and no script value acts the same.

Regular expressions in scripts accept only what means the same in JavaScript and Rust. Text with emoji or other characters outside the BMP works with patterns where no match can stop between the two halves of such a character, such as (<([^>]+)>) or \s+. A pattern that could, or a repeated group that can match nothing, such as (a*)?, is refused with the reason.

A Markdown render that crashes refuses its page with markdown-render-failed instead of stopping the build. A crash while running a script is a problem too: config-error at the config's top level, data-rhai-error in a data file, and a refused page in a filter or collection.

A feed whose HTML nests elements more than 2000 deep is refused with feed-failed, where it used to crash Brrrty (11ty's posthtml runs out of stack a little deeper and stops the build).

Fenced code blocks whose language is only characters JavaScript's trim keeps, such as U+0085, count as naming a language, as in markdown-it.

New errors:

  • config-syntax, config-error, config-invalid, config-limit, config-unreadable and config-unsupported: a problem in brrrty.config.rhai, with its line and column.
  • data-rhai-syntax, data-rhai-error and data-rhai-limit: a problem in a .rhai data file. Only the pages that use its data are refused.
  • data-rhai-shadows-js (a note): a .rhai data file replaces a JavaScript one.
  • syntax-highlight-unsupported: a page needs syntax highlighting.
  • feed-failed and feed-unsupported: a feed that can't be built, or one Brrrty doesn't support yet, such as an unknown option or a feed path that is also a file.
  • passthrough-unsupported: a passthrough copy Brrrty can't do yet, such as a glob.
  • source-date-epoch-invalid: SOURCE_DATE_EPOCH isn't a whole number of seconds.
  • markdown-render-failed: Markdown rendering crashed on a page.

0.0.4 - 2026-11-16

Data, permalinks, pagination and collections:

  • JSON data files in _data load as 11ty loads them, including folders of data files and symbolic links. JavaScript data files (.js, .cjs, .mjs) don't run yet: only the pages that use their data are refused.
  • Template and directory data files (page.json, page.11tydata.json, folder/folder.json and the like) apply to their pages, and the whole data cascade merges deeply in 11ty's order: global data, layouts, data files, then front matter.
  • eleventyComputed values are rendered in 11ty's dependency order.
  • Permalinks can be Liquid templates, and permalink can be computed.
  • pagination works over lists and objects, with size, alias, reverse, filter, resolve: values and generatePageOnEmptyData. Any size 11ty accepts works, such as '3' or [2].
  • collections are built from tags, with collections.all, eleventyExcludeFromCollections and addAllPagesToCollections, in 11ty's order, including the order eleventyImport.collections sets. Pages can paginate a collection.
  • templateContent and content of collection items work, including pages that read each other's content, rendered in the order 11ty renders them.
  • Nunjucks templates and pages under a Nunjucks layout are collection items. Their own pages aren't written yet, but pages that list them build.
  • Front matter YAML reads values the way 11ty's YAML loader does: dates, 0123 as 123, .inf, and numbers too large for JavaScript stay strings.
  • A file named like its folder (posts/posts.md) is written as the folder's index.html, as in 11ty.
  • A date left empty, "", 0 or false counts as no date, as in 11ty: the page takes the date in its path, or its file's date. Last Modified and Created match in any case.
  • json and inspect of collection items stop at once with a problem, as 11ty does, instead of working through every item first.

Where 11ty stops the whole build, Brrrty refuses only the page or data file at fault and builds the rest: a bad data file, pagination over missing data, two pages with one permalink, a templateContent circle, or a computed data loop. A page whose date Brrrty can't read is refused with the pages that read a collection it is in; other collections still work.

Front matter whose YAML aliases (*name) expand to more than a million values is refused as front-matter-invalid, at the alias that goes over, instead of taking the build down.

New errors:

  • data-file-invalid: a data file that can't be read or parsed.
  • data-js-unsupported: a page uses data from a JavaScript data file, which Brrrty can't run yet. Pages that don't use it still build.
  • computed-cycle and computed-unsupported: computed values that depend on each other in a loop, or a computed key Brrrty can't set yet, such as layout.
  • pagination-invalid and pagination-unsupported: pagination 11ty can't build, or that Brrrty can't follow yet.
  • content-cycle and content-unavailable: a page reads templateContent in a circle, or content that 11ty wouldn't have rendered yet.
  • tags-invalid: tags that are neither a string nor a list.
  • permalink-unsupported: permalinkBypassOutputDir, which would write outside the output folder.

Removed:

  • The permalink-templated error. Templated permalinks now render.

0.0.3 - 2026-11-15

Liquid templates and layouts:

  • .liquid, .md and .html pages render through Liquid, as 11ty does by default. .md pages then go through Markdown. The Liquid engine matches LiquidJS as 11ty configures it, filters included, plus 11ty's slug, slugify, url and log.
  • {% include %} and {% render %} find files in _includes, then in the input folder. A name can't lead outside those folders, through .., an absolute path or a symbolic link.
  • The front matter layout key works: layouts are looked up in _includes with 11ty's extension order, can have their own layout, and their front matter merges under the page's. .md layouts run Markdown too.
  • Pages get 11ty's page, eleventy and pkg data, and templateEngineOverride is honored.
  • Markdown is rendered by a port of markdown-it, configured like 11ty's.
  • Dates print in the machine's local time zone, as in 11ty, and TZ is honored. Set TZ to get the same output on every machine.

New errors:

  • liquid-syntax, liquid-unknown-filter, liquid-render and liquid-limit: a Liquid problem, reported in the file it comes from (the page, a layout or an include) at that file's line and column. The rest of the site still builds.
  • layout-missing, layout-cycle, layout-invalid and layout-unsupported: a layout that can't be found, a chain of layouts that loops (reported once, naming every layout in the loop), a layout that isn't a template file, and a Nunjucks or JavaScript layout.
  • data-unsupported: a page uses something Brrrty doesn't read yet, such as _data files, template and directory data files, or collections. The page is not written, instead of being written without that data.
  • data-reserved: data sets a name 11ty sets itself, such as content or page.url.
  • template-engine-override-invalid and template-engine-unsupported: a templateEngineOverride that isn't valid, or that names Nunjucks.

Removed:

  • The template-syntax-unsupported error. Template syntax in .md and .html pages now renders.

0.0.2 - 2026-11-15

Behavior change:

  • Ignore files now follow 11ty's rules. Each line in .gitignore, .eleventyignore and <input>/.eleventyignore is a path from that file's own folder, not a gitignore pattern that matches at any depth. So a root .gitignore line drafts ignores ./drafts but not src/drafts. A new ignore-line-anchored warning points at lines that would have matched under the old rules, with the full path to write instead.

Changes:

  • <input>/.eleventyignore is now read.
  • Symlinked files and folders in the input folder are followed and built at the link's path.
  • Page dates and slugs come from the whole path, like 11ty: a date in a folder name counts, and index takes its slug from its folder. An impossible path date like 2026-13-45 is a date-invalid error for that page, and the rest of the site still builds.
  • Permalinks with no file extension are allowed when the page sets eleventyAllowMissingExtension to a truthy value, and for _redirects, .htaccess and _headers.
  • Two passthrough sources that copy to the same output file are now a passthrough-collision error.
  • An input folder outside the project root, through --input ../x or a symlink, now builds and honors its .eleventyignore.
  • Problem paths are shown relative to the project root.

New warnings:

  • ignore-line-anchored: an ignore line matches nothing where it is anchored, but would have matched pages as a gitignore pattern.
  • ignore-line-invalid: an ignore line is not a valid pattern, so it is skipped and the build goes on.
  • template-plugin-only: files such as .webc, .pug or .mdx are skipped because they are templates only when a plugin adds them.
  • symlink-loop and symlink-broken: a link that points back into its own folder, or at a missing file, is skipped.

0.0.1 - 2026-11-15

First release: builds plain Markdown and HTML sites the way 11ty does. Templates coming next.

This page is built from CHANGELOG.md in the repository each time the site is built.