v3.0.0

A ```mermaid fence came out of the build as syntax-highlighted CSS. highlight.js has no mermaid grammar, so the fence fell through to auto-detection, which took graph TD; A-->B; for selectors and wrapped the diagram source in hljs-selector-tag spans. Mermaid could still read it — textContent drops the spans — but only if you told it to look at code.language-mermaid, and the page flashed the wrongly-coloured source first.

Added

• A mermaid fence compiles to <pre class="mermaid">, with no <code> element and the diagram source escaped inside. That is the markup mermaid documents and the default its run() selects on, so loading mermaid on the page is all it takes. The fence info string keeps working: ```mermaid wide tab=deps still becomes class="mermaid wide" data-tab="deps".

Changed

• book-of-spells 1.5 → 2.5, poops-docs-theme 4.2 → 4.5. The one that shows: removeAccents in book-of-spells now folds letters whose mark is written through the glyph — Đ Ł Ø Ħ Þ ß and friends — which slugify inherits, and slugify is what Poops builds heading anchors, collection slugs and permalinks from. Đorđe Balašević used to slug to ore-balasevic because the unfolded letter was deleted outright; it is dorde-balasevic now. A site with those letters in a heading or a title will move the URLs they generate — worth a redirect if anything links to the old ones.

• mermaid is never syntax-highlighted, through fences, the {% highlight %} tag and the highlight filter alike. Only the language name is special — a JavaScript fence that mentions mermaid is still highlighted as JavaScript.

Poops ships no mermaid and injects no script, so a page with no diagram carries no library, and a diagram on a page that never loads one shows its source as readable text. The trade against the code-fence story is real and stated in the README: highlighting happens at build time to avoid layout shift, and a browser-rendered diagram pops in after paint.

Full Changelog: v2.5.0...v3.0.0

v2.5.0

A page whose body was built from template tags described itself with their
source. # {{ site.title }} over {{ site.description }} shipped an
og:description reading "site.description" and an anchor at #site-title,
while the page itself rendered the real title and the real description right
underneath. The build stayed green through both, and the only symptom was a link
preview nobody looks at until someone shares the page.

Added

• A description filter emits <meta name="description">, escaped.
  {{ page | description(site) }} (Liquid: {{ page | description: site }})
  reads the same chain og and jsonld do — front matter description, then
  the page's auto-excerpt, then site.description — so the three cannot
  disagree, and emits nothing when none of them is set. Poops renders with
  autoescape off, so a layout writing that tag by hand ships the front matter
  verbatim: one " in a sentence closes the attribute and truncates the
  description to the words before it, silently, on a green build. Poops' own
  v2.4.0 post described itself as A for exactly that reason. Escaping now
  happens in one named place rather than in every layout that remembers.

Fixed

• page.excerpt is taken after the template engine has run. A first
  paragraph written as {{ site.description }}, or supplied by an
  {% include %} ahead of the prose, is resolved before the text is taken, so
  the description is the one a reader gets. The excerpt now reads the rendered
  body's first <p> rather than guessing at the markdown source — headings,
  comments, code fences and tables are skipped because they are not paragraphs,
  and link text comes through without its URL. When nothing usable resolves the
  excerpt is empty and og/jsonld/<meta name="description"> fall through to
  site.description — the old behaviour was to strip the braces and ship
  site.description as a literal string, which then beat the very value it was
  standing in front of. Only pages whose first paragraph is a tag pay for the
  extra render.

  A collection item is read without a page context, so its excerpt is empty in
  that case rather than resolved; listings and feeds fall back to the item's
  description. The item's own page still resolves it.

• Heading ids are slugged from the rendered heading. # {{ site.title }},
  # {{ page.title }} and # My Site now all anchor at #my-site, so swapping
  which variable feeds a heading no longer remints every bookmarked URL. The
  toc filter slugs from the same text, so its links still land.

Full Changelog: v2.4.0...v2.5.0

v2.4.0

A "last updated" line under a post had two ways to get its date and both were
bad. Keep updated current by hand and it is one more thing to forget in every
edit. Read the file's modification time instead and git clone erases it — CI
checks out every file at once, so the whole site reads as edited today.

Added

• markup.options.lastUpdated gives every page an updated date, without
  shelling out to git. Set it to true and Poops keeps an index of content
  hashes at .poops-updates.json (a string names a different file). A page's date
  moves only when its body actually changes — front matter is outside the hash,
  so retitling or retagging a post is not editing it — and every build after
  that reads the same date back. Templates get page.updated like any other
  front-matter field; it also feeds dateModified, article:modified_time and
  the sitemap. A page carrying its own updated keeps it and stays out of the
  index. The index file has to be committed, and a build has to run before
  the commit — it is the only memory of when a page was edited, so a build says
  Updated dates changed for 3 pages — commit .poops-updates.json whenever it moves.

Fixed

• The sitemap's <lastmod> now means last modified. It read front matter
  date — a publication date — so a post revised for a year still advertised
  itself to crawlers as untouched since the day it went up. updated wins where
  a page has one, with date as the fallback.
• A date front matter cannot fail the build. A date (or updated) that
  is not a date reached toISOString() on an Invalid Date while the sitemap was
  being written, and took the whole build down with a RangeError that named no
  page. The entry now goes out without a <lastmod>.

Full Changelog: v2.3.0...v2.4.0

v2.3.0

A footer wanting to print your version had nowhere to get it. poops.json is
JSON, so a value cannot compute one, and a theme rendering site.footer renders
it verbatim — which left the number hardcoded in a second file, silently wrong
from the next release on.

Added

• exec hooks can run before a stage, not only after it. A generator that has to write
  before the stage reads — fetching posts the templates render, writing a data file — had
  nowhere to go. Hijacking an earlier stage's hook worked in a build, where the pipeline order
  put it first, and then silently stopped working in watch, because a markup-only edit never
  fires the earlier stage. Every stage now also takes pre:<stage>, fired in build and watch
  alike: "pre:markup": "node script/fetch-posts.mjs". pre:build runs once before the
  pipeline starts, which is where clearing dist or fetching content belongs, and pre:styles
  fires ahead of Sass — the pair brackets the whole style pipeline, since styles fires past
  PostCSS. post:<stage> is accepted as the explicit spelling of a bare <stage>, so a config
  can read pre:markup above post:markup instead of above something that looks like a typo.

• markup.options.site values are filled from package.json, with the same tokens a
  banner uses. "footer": "MyLib v{{ version }} — MIT licensed" now prints the version
  the manifest holds, filled at build time from the one place it lives. {{ name }},
  {{ homepage }}, {{ license }}, {{ author }}, {{ description }} and {{ year }}
  fill the same way, at any depth — a title inside site.links gets them too — and the
  pkg option still says which package.json to read.

• A page can join a sidebar section without moving. The nav tree is built from urls, so
  grouping pages that already sit side by side meant a subdirectory — and a subdirectory
  changes their url, every link pointing at them and every search result that has been
  indexed, to buy one heading in a sidebar. navGroup: "No APG pattern" in front matter
  synthesizes that section beside the page's ungrouped siblings and nests the page inside it,
  url untouched. The label is used as written rather than humanized, the section sorts where
  its first child would, and grouping is per directory — the same name in two directories
  opens a section in each, instead of one reaching across the tree. A page that is itself a
  section (a directory's index page, or the homepage) is refused with a warning: its
  subpages would stay behind on the ungrouped path and split the section in two.

Fixed

• A banner token with no matching package.json field no longer prints undefined.
  {{ homepage }} in a package that declares none put that literal word at the top of
  every built file, which reads as Poops being broken rather than as a field you never
  filled in. The token is now left as written, naming what is missing.

• A PostCSS pass no longer leaves a stale source map beside its output. Poops wrote the
  processed CSS and threw away the map PostCSS had composed for it, so whatever map the Sass
  stage left on disk stayed there describing lines the pass had since moved — a debugger
  pointing confidently at the wrong rule. The composed map is now written next to the output,
  shifted past the banner line like every other stage. An input with no map still produces none.

Full Changelog: v2.2.0...v2.3.0

v2.2.0

A companion's block passed in silence only when you depended on the companion by
name, which a package that arrives through another one never is. Separately, the
permalink anchor beside every heading was hidden from screen readers and still
reachable by keyboard — a tab stop with nothing to announce, once per heading,
on every page Poops has built.

Added

• A direct dependency can vouch for a companion's config key. A septic block
  passed in silence only when septic itself sat in your package.json — but laxative
  brings septic, so a laxative app declares laxative and the warning fired on every
  build. A dependency now vouches through its own manifest,
  "poops": { "companionKeys": ["septic"] }, read one directory deep and never loaded.
  Nothing declared and nothing vouched still warns, which is what catches the typo.

Fixed

• The heading permalink is no longer a tab stop on nothing. Every heading gets an
  anchor, and it carried aria-hidden="true" while staying focusable — so a keyboard
  reached it, and a screen reader had nothing to announce when it landed. That is a link
  with no accessible name in the tab order, once per heading, on every page Poops has ever
  built. It also carried aria-label="Permalink", which aria-hidden had already made
  unreadable to anyone.

  The anchor now says the same thing three ways instead of three different things: it is
  decoration, so it takes tabindex="-1" beside the aria-hidden it always had, and the
  dead aria-label is gone. The heading beside it is already the name of the place.
  DOM change: <a class="heading-anchor" href="#id" tabindex="-1" aria-hidden="true">,
  where it was aria-label="Permalink" aria-hidden="true". Nothing changes for a mouse —
  the "#" a theme reveals through .heading-anchor::before still clicks — and a theme
  styling [aria-label="Permalink"] rather than the class is the one selector this breaks.

• Config Poops has always accepted is finally written down. exec had no section in
  the README at all, only two passing mentions of a key nothing introduced. watch: true,
  the per-entry nodePaths, markup.options.autoescape and dateFormat, the feed's
  content, the toc filter and the banner's {{ year }} were each real, tested and
  undocumented — some in the README, some on the docs site, nodePaths and {{ year }}
  in neither. Two engine-contract listings also still advertised fileExtension and
  renderString, which the pipeline stopped calling, while omitting the invalidate /
  pagesDependingOn pair that incremental rebuilds actually run on. Nothing changed in the
  code; the pages now say what it does. A new test walks every link in the README and the
  documentation site, so the eleven dead ones found writing this are the last of them.

Full Changelog: v2.1.0...v2.2.0

v2.1.0

The config file now tells your editor what belongs in it — and tells Poops,
which no longer lets a key misspelt inside a block pass in silence. It also
stops calling a companion package's block a mistake. Separately, llms was the one index
feature that could not turn itself on, and the corpus it writes carried the
machinery of the pages it was made from.

Added

• A JSON Schema for poops.json, so the editor catches a typo you would
  otherwise find in the output. A top-level "stlyes" was warned about and
  ignored, and a mistyped script option reaches esbuild, which rejects it — but
  "minfiy" in a style's options is read by nothing and warned about by nobody.
  The build stays green and the .min.css is never written.

  Poops now ships schema/poops.schema.json, covering every key, with the
  documentation for each one inline. Point $schema at it and the editor
  completes and validates as you type:

  {
    "$schema": "./node_modules/poops/schema/poops.schema.json"
  }

  The same file is published at
  https://stamat.info/poops/poops.schema.json
  for a project that has not installed Poops yet, and can be attached by
  filename from VS Code's json.schemas setting instead of by editing the
  config. The $schema key is inert to Poops — it is recognised and otherwise
  ignored, and nothing was added to what Poops installs into your project.

  The schema is hand-written, so Poops' own test suite holds it to the code: it
  is validated against the draft-07 meta-schema, and poops.json plus every
  complete config example in the README and on the documentation site is
  validated against it, so an example that stops being valid config now fails
  the build.

• A key misspelt inside a block is now named too. The unknown-key warning
  stopped at the top level: "stlyes" was caught, "inn" in a styles entry was
  not. That entry compiled nothing, poops -b exited 0, and the file that never
  appeared was the only sign. Poops now checks every block it owns against the
  schema it ships, at startup:

  [info][warn] Unknown key "inn" in styles[0] — ignored. Valid: in, out, options
  

  Key names only, and only where Poops owns them: images belongs to
  poops-images, site is yours to name, and a companion's top-level block is
  left alone — an unrecognised key in any of those passes without comment.
  Types are not checked, so "minify": "yes" still reaches the compiler and
  fails there. exec keeps its own warning, which says the more useful thing —
  that the stage never runs. The schema read is the copy inside
  node_modules/poops, so where $schema points changes nothing.

  The walk is unknown-keys, a new
  dependency and the only one this release adds — zero dependencies of its own,
  written for this and published separately because poops-images and septic read
  the same file and want the same warning.

Changed

• A top-level key naming a package you depend on is no longer called
  unknown. poops.json is shared — septic
  reads a septic block out of the same file — but every build printed
  Unknown config key "septic" — ignored, a warning about something working
  exactly as designed. The name is now checked against your dependencies,
  devDependencies, peerDependencies and optionalDependencies first, and a
  match passes in silence. Declaring the package is enough; Poops still never
  loads it or reads its block. Nothing by that name declared, and the warning
  is unchanged — which is what still catches "stlyes".

  Your editor cannot see node_modules, so the schema cannot make that call: it
  allows an object under any name it does not know and rejects everything else,
  meaning "stlyes": [ … ] is flagged there but "srve": { … } is not. The CLI
  catches what the editor lets through.

Fixed

• llms alone now generates its files. Page entries were collected only
  when searchIndex, sitemap, nav or feed was configured, so a markup
  config whose only index feature was llms compiled the site and wrote
  nothing — no warning, no output. Pairing it with a sitemap was the
  workaround; there is nothing to pair it with now.

• llms-full.txt no longer carries a page's machinery. The corpus is built
  from the Markdown source, which is read before the template engine runs, so it
  held whatever the source held: template comments, tags and output expressions,
  and every inline <style> and <script> on the page. A page that wraps its
  body in a {% set body %}…{% endset %} capture and styles itself in a
  <style> block could hand an LLM more plumbing than prose. All of it is
  stripped now, and the prose a capture wrapped stays. Fenced blocks, inline
  code spans and {% raw %} bodies are left alone — a sample documenting
  template syntax is content. A feed's article HTML is built from the same
  stripped source.

What's Changed

• Bump the actions group across 1 directory with 4 updates by @dependabot[bot] in #52

Full Changelog: v2.0.0...v2.1.0

v2.0.0

Poops 2.0 has no new features. It is the release where the core loop — watch,
build, reload — stops carrying old dependencies and old spellings, and where the
compat shims that survived the whole 1.x line finally go.

Changed

• Poops now requires Node.js 22 or newer. Node 20 reached end of life in
  April 2026, and the dependencies the dev loop is built on — the file watcher,
  esbuild — assume a modern runtime anyway. CI tests on Node 22 and 24, on both
  Ubuntu and Windows.

• Live reload is now served by Poops itself, on the server port. The
  livereload package is gone, and with it a second port, an old websocket
  stack, and the snippet you had to paste into your templates.

  "livereload": true alongside "serve" is all it takes. Poops answers
  /__poops_reload as a server-sent events stream and appends a small client
  script to each HTML page it serves — appends it to the response, so nothing
  lands in your build output. Behaviour is unchanged otherwise: one reload per
  save once the build settles, CSS-only builds swap stylesheets in place
  instead of reloading, and the browser reconnects by itself after a restart.

  Migrating: delete the livereload.js snippet from your templates (a leftover
  one is harmless — it will 404 quietly), and drop livereload.port,
  livereload.exclude, livereload.extraExts and livereload.exts from your
  config. The last three had already stopped doing anything when the reload
  server stopped watching files in 1.5.1.

• The file watcher moved from chokidar 3 to chokidar 5. Two majors of
  watcher fixes, and one fewer legacy dependency tree under node_modules.
  Nothing in your config changes: watch has always been a list of directories,
  never globs, so chokidar 4 dropping glob support costs nothing here. Watch
  mode behaves as before — one rebuild per save, CSS still hot-swaps, deletions
  still remove their output.

• esbuild moved from 0.25 to 0.28, and the default target from es2019 to
  es2020. esbuild is pre-1.0, so its minor bumps shift output by design;
  absorbing that is what a major of ours is for. Expect small differences in
  your bundles — the CommonJS interop helper is more careful around a throwing
  module, and Symbol.for calls are annotated as side-effect free, which makes
  minified output slightly smaller.

  The new default target is the visible half: optional chaining (?.) and
  nullish coalescing (??) are ES2020, and compiling them down for browsers
  that have shipped them since 2020 only made bundles bigger. Entries that set
  their own target are untouched.

• timeDateFormat is now dateFormat. It sets the default format for the
  date filter; the old name said the same thing twice. Deprecated and still
  read through 2.x, gone in 3.0.

  Engines see the rename too: registerFilters({ timeDateFormat, markupOut })
  is now registerFilters({ dateFormat, markupOut }). This is a hard rename —
  it lands before the engine interface becomes public API in this same release,
  so a custom engine must update its parameter name for 2.0.

• output is now out in the markup sub-features. llms, nav, feed,
  searchIndex, sitemap and robots name their output file with out, the
  same word every entry in the config already uses:

  { "nav": { "out": "nav.json", "root": "docs" } }

  output is deprecated but still honoured through 2.x, warning once per
  feature, and stops working in 3.0. The string shorthand ("nav": "nav.json")
  is unaffected.

• Markup settings belong under markup.options. markup now has the shape
  every other entry has — in, out, and everything else in options:

  {
    "markup": {
      "in": "src/markup",
      "out": "dist",
      "options": {
        "engine": "nunjucks",
        "site": { "title": "My Site" },
        "searchIndex": "search-index.json"
      }
    }
  }

  1.x read engine, site, data, includePaths, timeDateFormat,
  collections, searchIndex, sitemap, llms, robots, feed, nav,
  baseURL and autoescape at the markup. level as well. That placement is
  deprecated but still honoured through 2.x — each stray key logs a warning
  naming its new home — and stops working in 3.0. Where both are set, options
  wins.

Added

• The markup engine interface is public API as of 2.0. markup.options.engine
  has always accepted any importable module, and
  poops-shopify ships a production
  engine against it — but nothing said the interface was stable, so a rename in
  a patch release could have broken it silently. It is documented in
  the engine API reference, semver applies to it from here on,
  and a contract test asserts both builtin engines still expose the shape — a
  failing test is a breaking change caught before it ships.

• serve.base defaults to the markup out directory. Nearly every config
  set it to the path it had just built into. Set it only when the server should
  serve somewhere else; an explicit value still wins, and a project with no
  markup still serves the working directory.

Removed

• The ssg config key. It has been an alias for reactor since the rename,
  and it was the last compatibility shim in the codebase. A config still using
  it now gets told what to call it instead:

  [info] Config key "ssg" is renamed to "reactor" in 2.0 — ignored.
  

Internal

• Dependency bumps. book-of-spells 1.3 → 1.4 and a postcss patch, both
  routine. Dev-only: neostandard 0.12 → 0.13, and sulphuris 2 → 4 for the
  example site's styles. ESLint stays on 9 — neostandard 0.13 still declares a
  peer of eslint@^9, so 10 has to wait for it.

Fixed

• "Edit this page on GitHub" links work for collection items on Windows. A
  collection item's filePath kept native separators while a regular page's was
  posix, so on Windows the link for a post came out as
  …/edit/main/src/posts\hello.md. Both are posix now.

• The image cache is read correctly across platforms. Its lookups all key on
  posix paths, so a cache written on Windows — committed next to the images it
  describes, or built in CI — matched nothing and galleries came up empty. Keys
  and variant paths are normalized on read. The output-directory containment
  check also no longer accepts a sibling directory whose name merely starts with
  the output dir's (dist-old for dist).

• A style edit hot-swaps the stylesheet the page actually links. With
  minify on, the reload chain was told about site.css while the page linked
  site.min.css — no stylesheet matched, so every style edit reloaded the whole
  page instead of swapping the CSS. Both spellings are now reported (and with
  justMinified, only the minified one, since the other is deleted).

Full Changelog: v1.9.8...v2.0.0

v1.9.8

A fence could say more than its language, but only the language survived to the HTML — anything marking a fence for a later stage had to be an HTML comment next to it in the Markdown. The rest of the info string now lands on the code element as classes and data attributes, and all six places that render a code block finally agree on what one looks like.

Added

• The rest of a fence's info string becomes classes and data- attributes.
  1.9.7 stopped meta words leaking into the language class, which fixed the
  highlighting bug but threw the words away. So a fence could be labelled and
  nothing downstream could see the label — the marker for "this block is a live
  demo" had to live beside the fence as an HTML comment, one thing to keep in
  sync with another.

  Everything after the language now rides along: a bare word becomes a class, a
  key=value token becomes a data attribute.

  ```html preview tab=options widths=375,768
  <my-element></my-element>
  ```

  <pre><code class="hljs language-html preview" data-tab="options" data-widths="375,768">…</code></pre>

  A post-markup exec script then matches code.preview and reads the
  settings off the element, with nothing in the Markdown but the fence itself.
  Values are single tokens — no quotes, no spaces — which keeps the parser a
  split; anything needing a sentence belongs in the prose around the fence,
  not in its opening line.

  A bare word is a class rather than a valueless attribute, since a class is
  what the consumer selects on. When you do want the attribute — a flag read
  with hasAttribute instead of off classList — write the key with nothing
  after the =:

  ```html preview expanded=
  <my-element></my-element>
  ```

  <pre><code class="hljs language-html preview" data-expanded="">…</code></pre>

  Backwards compatible in the strict sense: a single-token info string renders
  exactly as before, which is every fence on every site today. Both the
  {% highlight %} tag and the highlight filter take the
  same info string.

Fixed

• All six renderers agree on what a code block is. The same three lines were
  copy-pasted across the Markdown renderer, the standalone highlight renderer,
  and the filter and tag in each engine — and 1.9.7 only fixed two of them. The
  engine copies still interpolated the whole info string into the class
  attribute, so they disagreed with the Markdown renderer about a fence's output
  and were right about language-html preview purely by accident. One
  codeBlock helper now emits every code block, so there is one place left to
  get wrong.

Full Changelog: v1.9.7...v1.9.8

v1.9.7

Added

• site.jsonld sets a JSON-LD default for the whole site. The jsonld
  filter picks BlogPosting for a dated page and WebPage for everything else,
  and a jsonld object in front matter overrode any of it. That escape hatch was
  per page, which is the wrong shape when the whole site is one type: a docs site
  is TechArticle, a knowledge base is FAQPage, a product catalogue is
  Product. The only way to say so was the same four lines of front matter in
  every file — and the failure mode is quiet, since a page that misses them still
  emits valid JSON-LD, just the generic type.

  The same object now works in your site data:

  {% raw %}{
    "markup": {
      "site": {
        "jsonld": { "@type": "TechArticle" }
      }
    }
  }{% endraw %}

  Every page renders as TechArticle — no front matter, nothing to forget. It
  isn't limited to @type; any key you would have set per page works site-wide,
  which is where the fields that genuinely don't vary belong:

  {% raw %}"jsonld": {
    "@type": "TechArticle",
    "license": "https://opensource.org/licenses/MIT",
    "isAccessibleForFree": true
  }{% endraw %}

  Precedence is defaults → site.jsonld → page.jsonld, so a single page still
  opts out of the site-wide type while keeping the rest of it — a FAQPage
  inside a TechArticle site keeps the license and only replaces @type:

  ---
  title: Frequently asked questions
  jsonld:
    "@type": FAQPage
  ---

  Two things it deliberately doesn't do. A site-wide @type beats the
  auto-detected BlogPosting as well — it's a default you set, not one poops
  guessed — so on a site that mixes docs with a blog, set the type per page
  rather than site-wide, or the posts stop being articles. And it merges into
  the page's own block only: the WebSite block on the homepage and the
  auto-appended BreadcrumbList on nested pages are structural, and a site-wide
  @type has no business rewriting them.

  Same shallow merge as the front-matter object, for the same reason — nested
  schema is rare, and the escape hatch is meant for whole-key replacement.

Full Changelog: v1.9.6...v1.9.7

v1.9.6

Added

• out can be a template, for styles and scripts alike. v1.9.5 named a
  glob-matched index.* after its directory, which fixed the case a component
  library actually hits. It did nothing for anything else: point a glob at
  src/elements/*/theme.scss and every match still falls back to its own
  basename, so the components overwrite each other in the output directory and
  the last one to build wins. There was no way to say what you wanted the files
  called.

  Now there is. An out carrying {% raw %}{{dir}}{% endraw %} or
  {% raw %}{{name}}{% endraw %} resolves per entry point instead of naming one
  shared destination:

  {% raw %}{
    "styles":  { "in": "src/elements/*/theme.scss", "out": "dist/css/{{dir}}-{{name}}.css" },
    "scripts": { "in": "src/elements/*/widget.ts",  "out": "dist/js/{{dir}}-{{name}}.js" }
  }{% endraw %}

  src/elements/accordion/theme.scss  →  dist/css/accordion-theme.css
  src/elements/tabs/theme.scss       →  dist/css/tabs-theme.css
  src/elements/accordion/widget.ts   →  dist/js/accordion-widget.js
  src/elements/tabs/widget.ts        →  dist/js/tabs-widget.js
  

  {% raw %}{{dir}}{% endraw %} is the match's directory relative to the glob's
  static prefix — the same name an index.* entry gets, so the two rules
  agree on what a component is called, and widening the glob keeps the segments
  it no longer pins down. {% raw %}{{name}}{% endraw %} is the basename
  without its extension. Whitespace inside the braces is fine
  ({% raw %}{{ dir }}{% endraw %}), matching the banner templates.

  Tokens work in directory segments too, so the flat layout isn't the only one
  available:

  {% raw %}{ "in": "src/elements/*/theme.scss", "out": "dist/css/{{dir}}/theme.css" }{% endraw %}

  A literal entry fills {% raw %}{{dir}}{% endraw %} with its own directory
  name, which keeps mixed arrays of globs and plain paths working:

  {% raw %}"out": "dist/css/{{dir}}.css"{% endraw %}
  
  src/scss/main.scss                 →  dist/css/scss.css
  src/elements/accordion/index.scss  →  dist/css/accordion.css
  

  For scripts the template's extension is honoured as well, which is the cheap
  way to ship one format per entry point rather than one bundle per format:

  {% raw %}{ "in": "src/elements/*/index.ts", "out": "dist/esm/{{dir}}.mjs",
    "options": { "format": "esm" } }{% endraw %}

  A template wins over the index.* rename — you named these outputs, and
  renaming them behind your back is the thing globs were already doing wrong.
  It's also exempt from the "more than one entry file needs a directory out"
  guard, since a template already resolves to a different file per entry rather
  than to one file everything overwrites. What it does not do is invent
  uniqueness: {% raw %}"out": "dist/{{name}}.css"{% endraw %} across component
  directories collides exactly like a plain directory out would, and its
  scripts equivalent fails the build outright with esbuild's "Two output files
  share the same path". That's what {% raw %}{{dir}}{% endraw %} is there
  for.

Fixed

• Live CSS reload no longer guesses the output path. The watch chain fed
  livereload a path derived from the config — a directory out joined with the
  basename of in. For a single named entry point that guess was right. For a
  glob it was the pattern's own basename, so src/elements/*/index.scss
  reported dist/css/index.css, a file that was never written; a templated
  out reported the template verbatim, braces and all. The livereload client
  looks for a loaded stylesheet matching the path it is handed and, finding
  none, falls back to reloading the whole page — so editing a component's Sass
  flashed a full reload instead of swapping the stylesheet in place, and any
  scroll position or open state went with it.

  The styles compiler now records the files it wrote and the watch chain
  reloads exactly those. One entry, one glob or twenty templated outputs, the
  reported paths are the ones on disk, because nothing derives them a second
  time.

• The watcher's output zones understand templates. A compiler writing into
  a watched directory must not retrigger itself, which the watcher prevents by
  zoning each task's out. A templated out is only a fixed path up to its
  first token, so dist/{{dir}}/theme.css zoned a directory literally named
  {{dir}}, protecting nothing. Zones are now taken from the static prefix.

Full Changelog: v1.9.5...v1.9.6