Accent CMS v0.26.0

Announcement: https://accentcms.dev/news/2026/09/accent-cms-0-26

Scripts are locked down everywhere Accent renders a page. Served pages carry a
per-response nonce, static builds carry script hashes, and the admin UI runs
without 'unsafe-eval', so an injected script no longer runs. Custom themes
with inline scripts or onclick=-style handlers need a one-line change per
script; the entries below say what it is and how to find what breaks first.

Components become a contract a theme can be checked against: canonical and
builtin components, manifests that describe their attributes, and an editor
that lists them. Plugins can declare their own Markdoc tags. Directory
previews with accent serve <dir> turn into navigable sites with search.
Releases add fully static Linux musl binaries for Alpine.

Three changes need action when you upgrade:

• Plugin API 0.2.0. A plugin built for 0.1.0 is refused at install and at
  load. Run accent plugin update <name> for each installed plugin.
• [shortcode] syntax is removed. Rewrite each one as a declared
  component, {% name %} ... {% /name %}.
• A broken internal link fails accent build. Pass --lax to report it
  as a warning instead.

The JSON REST API also moves in the static output layout; see Changed.

Added

• Served pages restrict scripts with a per-response nonce. accent serve
  now appends script-src 'self' 'nonce-...' to every page's
  Content-Security-Policy, with a new nonce on every response and no
  'unsafe-inline'. A script runs only when it comes from the site's own
  origin or carries nonce="{{ nonce }}", which every shipped theme now
  writes on each <script>; inline onclick=-style handlers are blocked.
  Custom themes with inline scripts or handlers need the same change: add
  the attribute and move handlers into a script, or set
  http_headers.security.script_policy.report_only: true to find what breaks
  first, or enabled: false to opt out. A <script> in page content is
  blocked on served pages. The admin UI and media keep the previous policy.

• The admin UI runs without 'unsafe-eval'. The admin now uses the
  Alpine.js CSP build and htmx with eval turned off, and the editor preview
  is a page of its own that carries the site's script policy, so a <script>
  in a draft is blocked in the preview as it will be when published.
  /_admin/ gets the strict script policy, report-only for now through the
  new http_headers.security.script_policy.report_only_paths (default
  ["/_admin/*"]); set it to [] to enforce. script_policy.exclude now
  defaults to []. A hotfix override of alpine.min.js must be the CSP build.

• Static builds restrict scripts too. A built page has no response to
  carry a nonce, so accent build records a sha256 hash of each inline
  script marked nonce="{{ nonce }}" and the origin of each marked script
  from another host, then writes them into the page CSP of its header
  artifacts; accent serve-static sends the same policy. Scripts in content,
  and theme scripts without the attribute, are blocked in static output. The
  build manifest format changed, so the first build after upgrading is a full
  rebuild.

• The admin lists what a theme's components are, and the editor inserts
  them. /_admin/components is generated from the schema registry: each
  component's origin, attributes and body kind, with a page per component
  showing the call to insert and the markup its theme renders. A switcher reads
  each theme in the theme directory. The page editor gains a Components
  palette under the body toolbar that inserts a call with its required
  attributes stubbed, into CodeMirror or the plain textarea.

• A component manifest declares what a list or object attribute holds.
  A list names its item type under items: and an object its fields under
  fields:, nested up to eight deep, and accent validate checks every item
  and field, naming the path to a value that does not match, such as
  links[1].label. A list without items: or an object without fields:
  is a manifest error. The documentation theme's new related component takes
  its links this way.

• Canonical components carry markup landmarks. Every conforming theme marks
  a canonical component's root with data-canonical and its parts with
  data-part, so a script or a stylesheet can find a callout's title or a
  tab's panel under any conforming theme. The default, site-dev and starter
  themes carry them.

• Plugins can declare Markdoc tags. A plugin lists its tags in
  [shortcodes] codes and ships a component manifest for each in its
  shortcodes/ directory, in the format a theme uses. Two settings decide what
  those tags may do: plugins.components.schemas (default true) validates
  content against the plugin's manifests, and plugins.components.render
  (default false) lets the plugin write the markup. With the defaults a
  plugin tag is validated and left in the page as written, and the log names
  the plugin and the setting. accent validate names a plugin that declares
  a tag and did not load, and accent validate --lax reports plugin component
  findings as warnings that do not fail the run. accent query components
  lists every tag content may use, where it came from, and what renders it.

• Canonical components, and a check that a theme implements them. Every
  theme implements the canonical components -- callout, accordion,
  button, card, tag, image, breadcrumb, tabs and tab, plus
  styles for the built-in table -- with fixed attribute names, so content
  that uses only them renders under any conforming theme. accent validate
  reports a theme that lacks one, with no content needed, and so does
  accent validate --templates. A page that uses a component the active theme
  does not declare is now told which theme and whether the component is
  canonical, in accent validate and in the build and serve logs. A theme
  that provides no components declares components: none in theme.yaml.
  accent validate --lax relaxes a theme's gaps and a page's missing canonical
  component, but not a name nothing declares, which may be a misspelling. It
  now combines with --templates.

• Builtin components Accent renders itself: download, video, diagram
  and search. {% download %} renders a file beside the page as a card
  with its name, type and size, the same card pdf_card() renders for a PDF.
  {% video %} plays a video beside the page with its WebVTT captions, and
  accent validate reports a video with no captions unless it says
  captions="none", which --lax relaxes. {% diagram %} renders a diagram
  with a caption and alt text through the same pipeline as a fenced block, from
  its body or a src file. {% search %} places the search box in a page.
  A file these components link to must be a type Accent serves; for any other
  type, such as .zip, add its extension to media.attachments.extensions.
  Accent now serves .vtt caption files as text/vtt.

• Search results highlight your term on the page they open. Every
  result link now carries a text fragment built from the three longest
  words you typed, so the browser scrolls to the first match and marks it. No script runs
  on the page that opens; both bundled themes paint the mark yellow
  through ::target-text, since Chrome's own default is a pale lilac.
  Chrome, Edge, Safari, and Firefox support text fragments; any other
  browser opens the page as before.

• Rename a model field without editing every page. A field can list its
  old names under renamed_from: [teaser]. A page that still uses an old
  name loads with the new one, and the next save from the admin editor, the
  CLI, or an agent renames the key in the file where it stands, keeping its
  value and comment. accent validate reports each page still on an old
  name and ends with a count per rename. A page that sets both names keeps
  the new one's value and gets a warning; saving that field in the editor
  settles it, and the value you were looking at is the one that stays.

• Search and a light/dark toggle in the preview viewer. A directory
  preview (accent serve <dir>) now has a search box in the header, backed
  by the index the preview already built; press / to jump to it. Every
  preview follows the system colour scheme and has a sun/moon button to
  read in the other one. The choice persists across pages and previews,
  and selecting the system's scheme again goes back to following it. The
  viewer also shows the Accent favicon.

• nav_tree() template function. Returns the page hierarchy as a
  nested tree (the root page, then children all the way down, each node
  with url, title, and nav_title, in menu.order then URL order),
  built once per render on the server, so a theme renders a site-wide
  navigation with one {% for ... recursive %} loop instead of scanning
  all_pages per level. Nesting follows each page's hierarchy_url, now
  on every listing entry (all_pages, page.children, and the rest), so a
  page a permalink pattern or a frontmatter url: moved stays under its
  directory parent. Handles a deployment base path and attaches a page
  whose parent has no page to its nearest ancestor. The built-in viewer
  theme's sidebar now uses it.

• Directory previews honour .gitignore. The auto-index walk behind
  accent serve <dir> (and content.auto_index on any site) now consults
  every .gitignore from the content directory down, with git's rules
  (anchored and directory-only patterns, globs, negations, last match
  wins), so pointing serve at a repository serves what git tracks.
  content.auto_index.gitignore: false opts out; --write-config records
  the flag.

• Directory previews are navigable sites. accent serve <dir> now
  gives every directory that holds markdown a section page: a directory
  with a README.md, index.md, or default.md keeps it, and one without
  gets a generated page listing its sub-sections and pages, so a tree with
  no root README is reachable from /. The built-in viewer theme gains a
  sidebar tree expanded along the path to the current page, breadcrumbs,
  and a home link; single-file and URL previews are unchanged. Hidden
  entries and the node_modules and target directories are skipped. The
  same behaviour is available to any site through
  content.auto_index.enabled: true (with content.auto_index.ignore for
  the skip list), and theme.embedded: viewer lets a config-driven site
  use the built-in viewer theme without a themes directory on disk.

• accent serve <dir> --write-config [PATH] writes a config.yaml
  that reproduces the preview (content directory, automatic section pages,
  the viewer theme or the --theme-dir theme, address and port, server-side
  diagrams) and then serves as usual. A relative PATH is resolved against
  the served directory; an existing file is never overwritten. The written
  config also drives accent build and accent validate.

• Ad-hoc serving (accent serve <file.md> / accent serve <dir>) renders
  fenced Mermaid and Svgbob blocks server-side as inline SVG. The synthesized
  config forces diagrams.render_mode: server for local targets, since the
  embedded viewer theme ships no client-side diagram loader; the viewer theme
  now styles the rendered .diagram-wrapper (responsive sizing, dark-mode
  backdrop, captions). Remote URL targets keep their existing third-party
  content posture.

• Alpine Linux support: releases now include fully static
  x86_64-unknown-linux-musl and aarch64-unknown-linux-musl binaries with
  no glibc requirement (and no runtime requirements at all), smoke-tested on
  Alpine before publishing. install.sh detects musl-based systems and
  selects the musl build automatically; glibc systems keep the existing gnu
  builds and their 2.28 floor.

• install.sh --require-verify (or ACCENT_INSTALL_REQUIRE_VERIFY=1) makes
  every skipped verification a failure: a checksums file, signature, or
  signing key that will not download, or a machine without gpg or
  sha256sum, now stops the install instead of warning. The default stays
  fail-open for the interactive one-liner. The deploy workflows run the
  installer from the checkout with the flag set, so a deploy either verified
  both the checksum and the signature or did not install at all.

• api.rate_limit throttles the API instead of documenting an intention to.
  It counts requests per minute per client against the API routes only, answers
  429 with Retry-After, and reports X-RateLimit-Limit/-Remaining/-Reset
  on every API response. 0, the default, stays unlimited. Behind a proxy, set
  the new api.trust_forwarded_for so clients are counted individually.

Changed

• The component inspector reads the theme once per edit. An attribute
  edit read the theme directory and parsed every component manifest three
  times, on the async runtime, and validated the whole document to show one
  tag's findings. It now loads the registry once, off the runtime, and
  validates the tag's own node.

• accent serve and accent serve-static start with a clean screen. In a
  terminal, a box shows the version, edition, license, mode, and a link to the
  documentation, and a ready block shows the address and the time startup
  took. Timestamped log lines begin after it. NO_COLOR turns color off.
  Output to a file, a pipe, or a service is still timestamped from the first
  line and still includes Listening on .... Internal startup lines, such as
  the watcher's poll interval and the search backend, now log at debug level;
  RUST_LOG=accent=debug shows them.

• A broken internal link fails accent build; --lax reports it as a
  warning instead. The build already resolved every internal link against
  its output. It now fails on a link whose target it did not write, listing
  every such link with the source file to edit and, for a draft, in-review, or
  scheduled page, the reason it was not written. Before upgrading, run the
  current release with --strict-links to find them. That flag is now
  deprecated: it is accepted, prints a notice, and is removed in 0.27. External
  links are still not checked.

• accent validate reports a link to a draft or in-review page as an
  error. It was an info finding, while the build now fails on the same link.
  The message says to publish the page or remove the link, and
  accent validate --lax relaxes it along with the component findings; its
  summary now reads --lax: N findings reported as warnings.

• gix now comes from crates.io, moving 0.84.0 to 0.87.1. The
  zoosky/gitoxide fork carries no patches, so Accent no longer pins a commit
  on it; the pin had fallen 1110 commits behind upstream. The update brings
  upstream fixes such as fallible pack-memory reservations. The tree-editor
  feature no longer exists upstream and is dropped. Accent's git code needed no
  changes. Fetch and push still use the system git, because gitoxide has no
  push yet.

• The default and starter themes use the canonical names: alert is now
  callout, with type in place of variant, and details is now
  accordion. The default theme keeps infobox, which released documentation
  versions use. accent migrate shortcodes still rewrites [alert] and
  [details], because a component manifest can list the shortcode names it
  replaces under migrate.from, with any attributes it renamed.

• rustls 0.23.45 (from 0.23.43), for RUSTSEC-2026-0285: TLS 1.3 handshake
  messages were accepted across encryption level boundaries. It affects the
  development HTTPS server and accent serve --tls. The manifest now requires
  0.23.45, so the lockfile cannot move back below the fix.

• wasmtime 48.0.2 (from 48.0.1): code generated by bindgen! compiles on
  the latest Rust nightly, and Wasmtime vendors its cap-primitives
  dependency, so cap-std and cap-fs-ext leave the dependency tree.

• accent build renders pages in parallel. The render pass runs on all
  cores instead of one; building the 644-page documentation site falls from
  12.4s to 2.2s on a 28-thread machine. Output is byte-identical however many
  threads run -- pages are reduced in URL order -- and RAYON_NUM_THREADS
  caps the pool (1 renders serially). When more than one page fails, the
  build now reports every failure together in URL order rather than stopping
  at whichever one a worker reached first.

• The editor reads the compiled model schema. The admin editor's form,
  the typing of a saved or previewed form post, the inline field check, and
  the model browser's field rows now read the JSON Schema a document model
  compiles to, instead of each re-deriving the model on its own. Nothing an
  author sees or saves changes. Two keys join the schema accent model show --schema prints: x-accent-order, which lists a model's fields in the
  order they are declared, on the model and on every object field that
  declares fields (the printed properties are sorted), and
  x-accent-widget on list items and
  map values, so a list of string and a list of text no longer read the
  same.

• A quoted number is a string. YAML says 5 is an integer and "5" is
  a string; model validation used to parse the quoted form and accept it for
  an integer or number field. That made a value's type depend on who
  asked -- validation called order: "5" an integer, while as_i64 read it
  back as absent and a template comparing order == 5 saw a string.
  Validation now reads what YAML resolved, so order: "5" is a type
  mismatch reported as got string. Nothing Accent writes is affected: the
  admin form, accent content update-frontmatter --set, and accent new
  all already emit real YAML numbers for a declared numeric field. A
  hand-authored page that quotes a number on such a field will now report a
  type mismatch; remove the quotes. An integer still satisfies a number
  field.

• min_length and max_length count Unicode code points, not bytes. A
  model field's length constraints were measured in UTF-8 bytes, so a
  70-character Open Graph title containing three em dashes measured 76 and
  was refused against max_length: 70 -- and the message reported the byte
  count as though it were a length, so the author counted 70 and was told
  76. They now count code points, which is what JSON Schema's
  minLength/maxLength specify. For most text that is what you would
  call a character, but not always: an emoji built from several code points
  -- a flag, a skin-tone modifier, a family sequence -- counts as more than
  the one character it draws as, and so does a letter written with a
  combining accent. Content near a limit and not written in ASCII
  can change verdict: max_length becomes more permissive and min_length
  more strict, because code points never exceed bytes. No page in this
  repository's four sites changes verdict.

• Component bodies render as markdown, and the manifest says how each is
  spliced. A component body used to be passed through literally when the
  template put it inside a block-level element, and in infobox and
  details the first paragraph was literal while later ones rendered. The
  manifest's body: field now decides the splice: body: markdown (block
  content) arrives between blank lines, so the template's <div> closes
  before it and the whole body renders through the ordinary pipeline;
  body: inline (a label, such as a button's text) arrives flush. Every
  shipped component keeps body: markdown except button, which is now
  body: inline and inline: true, so it can be written on one line or in
  a sentence without a placement finding. A theme outside this repository whose label-bodied
  component declares body: markdown sees that label wrapped in a
  paragraph after this release: declare body: inline for it. accent validate --templates reports a body: inline body placed on a line that
  starts with a block-level tag, and an indented line after a
  body: markdown body.

• Plugin API 0.2.0: the json value is UTF-8 bytes. The one
  open-shaped value in the plugin contract (frontmatter, filter values and
  results, model defaults, media metadata, data-store values) crosses the
  Component-Model boundary as the bytes of one JSON document instead of a
  flat arena of typed nodes, so a Rust plugin reads it with serde_json
  and a JavaScript plugin with JSON.parse. The package is
  accent:plugin@0.2.0 and the host's supported API range is now
  0.2.0..=0.2.0: a plugin built for 0.1.0 is refused at install and at
  load, and the load-time message names the fix, accent plugin update <name>. The hub resolver now falls back to the newest listed version
  this accent can run when the latest build targets a newer host, instead
  of refusing outright. The first-party plugins and both accent plugin new templates build on 0.2.0; their hub builds ship with the release.
  A payload that is not a JSON document is a plugin diagnostic (a template
  error for a filter, a warning for a model default or media metadata
  entry, an invalid-request for a data-store write), never a silently
  invented value.

• Derived caches now carry a schema fingerprint. Rendered pages,
  diagrams, processed media variants, and the incremental build manifest
  are keyed by one digest of the engine build, the document models (what
  each declares, so a comment or a reformatted model file changes nothing),
  the loaded plugins and their configuration, theme.yaml, the media,
  code, diagrams, and markdown config sections, and the site URL and
  base path, so an edit to any of them invalidates every derived artifact
  by construction. In accent serve, a jpeg_quality change followed by
  SIGHUP now reaches the next derivation (the processor used to keep the
  value it started with); processed-media variants live in one directory
  per fingerprint under the cache directory, and the directories of
  previous fingerprints are deleted when the value moves and at startup.
  The diagram cache key now carries the block options each renderer
  reads: width and height for Svgbob, the caption for the client-side
  markup, all of them for a plugin renderer. The first accent build
  after upgrading is a full rebuild: the manifest format moved to version
  3 and its hashes are SHA-256, which do not change with the Rust release;
  after that, a model edit rebuilds every page where it used to rebuild
  none, and so does a build under a different --base-url, which used to
  skip every page and leave the old prefix in every link.

• The template context is now derived from its Rust types. What a
  template can read is exactly the context structs, serialized: no
  hand-maintained mapping sits between them any more, so a new field
  reaches templates the moment it exists. Nothing a template reads
  changed -- snapshots of the context every page of the three sites in
  the repository receives were taken before the change and are identical
  after it -- and every internal URL still arrives joined with the
  deployment base path. A JSON Schema of the whole context is committed
  as a test snapshot, so a release's diff of that file is the list of
  changes to the theme contract.

• Editing a page's frontmatter no longer reformats it. An admin save,
  an MCP write, or accent content update-frontmatter now rewrites only the
  fields that actually changed: comments, blank lines, quoting style, and
  the order you put the keys in are preserved exactly as you wrote them.
  Saving a page you have not changed writes nothing at all. Previously every
  write rebuilt the whole block and re-emitted it, so the first save through
  the admin UI discarded every comment in a page's frontmatter.

• Duplicate frontmatter keys are now a parse error. Frontmatter parses
  through the same YAML policy as config and theme files (one parser for
  the whole tree; the second YAML engine is gone from the dependency
  graph). The previous frontmatter parser silently kept the last
  occurrence of a repeated key, so a page carrying, say, two title:
  keys -- or 1: beside "1":, which normalise to the same key -- built
  without complaint and dropped one value. Such a page now fails to parse;
  the fix is to delete the losing key. Keys are compared after
  normalisation, so the colliding pair is not always literally identical.

• A UTF-8 byte-order mark at the start of a markdown file no longer hides
  its frontmatter: the block is parsed normally instead of leaking into
  the rendered body.

• Content write operations (accent content update-frontmatter, body
  edits through the admin editor, MCP mutations) refuse a file that opens
  a --- frontmatter block it never closes, instead of silently building
  a second block in front of it. Such files still render (the
  unterminated block is treated as body, as before); only editing them
  through Accent asks you to fix the block first.

• Breaking (static output layout): the JSON REST API is built where it is
  served. accent build wrote the API to _api/v1/pages/blog/post/index.json
  while accent serve answered /api/v1/pages/blog/post -- a different prefix,
  a different path, and a file no static host serves for a directory request, so
  a frontend developed against the dev server could not read the built output.
  Both surfaces now derive every path from api.prefix (which neither read
  before), and every resource is written and answered at a .json URL:
  /api/v1/pages/blog/post.json resolves against accent serve,
  accent serve-static and a bare CDN alike. The old layout is gone; a
  consumer reading _api/v1/.../index.json moves to the new paths. The
  payloads were reunited too -- the built page detail had been dropping custom frontmatter
  entirely, writing null where the served one omits a field, and publishing no
  file for archived pages the API answers with 200.

Removed

• Breaking: [shortcode] syntax is gone. Square-bracket shortcodes were
  expanded before the markdown parser ran; that stage no longer exists, and a
  page still holding the syntax renders the brackets as text. Declared
  components ({% name %} ... {% /name %}) replace them and add what the old
  form could not: a manifest saying what each component takes, so a misspelled
  attribute is reported by accent validate instead of rendering as nothing.

  accent migrate shortcodes rewrites a site and has not been removed with
  the syntax it reads, so upgrading late costs a command rather than a manual
  rewrite. Run it without --write first; it reports what it would change and
  refuses anything it cannot verify.

  Seven builtins were pure markup and are now components the default theme
  ships and your theme can redefine: infobox, alert, tabs, tab,
  button, details, edition. Five reached into parts of Accent a theme
  template cannot and have no content syntax any more -- [diagram], [pdf],
  [figure], [iframe] and [search] -- each keeping another way in: a fenced
  ```mermaid block or diagram(), pdf_card(), markdown image syntax
  or a raw <figure>, a plain <iframe> element, and search_form() in a
  template.

  [alert] was a deprecated alias for [infobox] that resolved ahead of a
  theme's own alert template, so a theme shipping one never saw it render.
  alert is now whatever your theme declares; write {% infobox %} for the
  titled callout.

Fixed

• Admin pages are never cached. Every page and fragment under /_admin/
  now carries Cache-Control: no-store. Before, with browser reload off, they
  carried the public page policy (public, max-age=300, s-maxage=3600), so a
  CDN or proxy in front of Accent could store an editor's page and serve it to
  others. http_headers.cache cannot override it.

• Force save works after an edit conflict. The button in the conflict
  dialog sent nothing, because it aimed at the save bar the dialog had
  replaced. It now saves, and a second conflict opens a fresh dialog in the
  same place.

• accent hotfix apply writes hotfix.yaml completely. The overlay's
  provenance file ended mid-line, so every git diff of it reported
  \ No newline at end of file and line-oriented tooling miscounted the
  last entry. The file is also written atomically now, through a temporary
  file that is renamed over it, so an interrupted run leaves the previous
  manifest intact instead of a truncated one. (b219)

• The developer site scripts bind loopback, and the specs site commits
  no credential. The shared worktree selector behind every scripts/*.sh
  site script forced 0.0.0.0, so every developer site was reachable from
  the local network whatever its config said, and specs/config.yaml
  paired that with an admin UI behind a token committed in plain text.
  Each site's config now decides its own bind, and every config that mounts
  an admin UI binds loopback; --address 0.0.0.0 after the selector is the
  opt-in. The specs config and the admin-dev config defer their admin token
  to the environment, the specs config its Basic Auth password too, and the
  site scripts generate what is missing per worktree and print it. An
  unexpanded ${VAR} placeholder no longer counts as an admin token, and
  the login page and the startup log name the variable to set. (b203)

• CI rejects a duplicate spec identifier. Two sessions filing specs on
  one day both listed the directory, saw the same gap, and claimed it, and
  nothing rejected the second claim until a hand renumber PR. A check on
  every spec change now fails a PR whose merge result holds two files under
  one identifier or a spec under the wrong family directory, and its
  --branches mode does the pre-filing scan across unmerged branches. (b308)

• /_health/git reports the sync status without SMTP. The probe
  answered {"status":"disabled"} unless mail notifications were
  configured and compiled in, so a git-backed publisher without an SMTP
  server had no HTTP-reachable sync status. Every fetch, from a webhook
  delivery or from polling, now records its outcome in the server state,
  and the probe reports it, the boot refresh and the last error text
  included; disabled means only that git.enabled is off, and a server
  that has not fetched since startup says unknown rather than ok.
  (b282, b309)

• The service preflight tests no longer race for a released port. The
  fixture picked a port by binding and releasing it, and the address check
  bound it again, so under the parallel test runner another test's
  ephemeral listener could land in between and add an "in use" finding
  to a test about something else. The probe is a parameter now: production
  binds, and a test that is not about the address passes one that always
  says free. (b283)

• git.sync.poll_interval polls. The field was declared, documented
  as the fallback for a forge that cannot reach the server, and read by
  nothing: a publisher without a working webhook stayed on the commit it
  booted on. A polling task now runs the webhook's fetch every
  poll_interval minutes on a publisher serving in production, under the
  same lock and with the same outcome handling, so a push reaches the site
  within the interval. Dev servers and authors never poll. (b201)

• accent new --field writes values the page loads back as typed. A
  list item was written raw, so tags=Blog: a post,rust scaffolded a
  mapping where a string was meant and the page failed to load, and a
  leading # vanished as a comment. Every scalar and every list item now
  goes through the YAML serializer; a list item takes the model's declared
  item type, so an integer list stays numeric; empty segments from a
  trailing comma are dropped; and a multi-line item continues under its
  dash. (b220)

• Two builds of the same content produce the same search index bytes.
  _search/docfind_bg.wasm differed between builds of identical content,
  which invalidated it at every deploy and blinded byte-level build
  comparisons: the keyword extractor ordered equal-score keywords by hash,
  so the per-document keyword budget kept a different subset each run.
  The DocFind fork now orders them by score, then by where they first
  appear in the text, so two builds of the same content give the same
  search index and the same rendered files; only the build manifest's
  timestamp differs between them. (b226)

• A scheduled page is written on the day it becomes visible, a page
  reaching its unpublish date is rendered again as archived, and a page
  that is a draft again loses its output. An incremental build decided
  what to render from content and template hashes, which a passing date
  never moves, so the page was listed by the sitemap, the feeds, and the
  search index but had no file behind it. The build manifest now records
  each page's effective status on the build date and rebuilds a page whose
  status moved, with no edit; a page that is no longer visible has its
  HTML, markdown companion, paginated continuations, and API detail file
  removed, which --dry-run announces; and every build-side pass reads one
  build date. (b256)

• accent cache clear empties the processed-media cache too, and a
  deleted or renamed media file loses its variants. The command cleared
  only the diagram cache, and nothing invalidated processed variants when
  the admin deleted or renamed their source or when a media file changed
  under hot reload, so stale variants stayed on disk until capacity
  eviction. The command now clears both caches and reports each, with
  --diagrams and --media to limit it; the admin operations drop the
  source's variants; hot reload watches the shared media directory too and
  drops the variants of a changed file by name; a config reload forgets
  the entries whose files a clear removed. (b238)

• Generated documentation no longer references internal spec IDs. 218
  doc comments across 121 source files named a feature, bug, research,
  implementation, architecture, or epic tracker ID, which cargo doc and
  accent docs rendered for readers who cannot resolve it. Every one is
  reworded to name the thing instead, the ID moved to a plain code comment
  beside the item, and a check in CI and both local gates keeps the count
  at zero. (b224)

• accent build --dry-run writes nothing, and --clean --dry-run deletes
  nothing. A dry run copied the theme assets, compiled the styles, and
  wrote the fingerprinted copies into the output before printing that no
  files would be written, because the rebuild plan needs the fingerprint
  manifest and the manifest needs the compiled CSS on disk; with --clean
  it also removed the output directory first. The dry run now skips the
  output preparation and does the asset work in a temporary directory that
  goes away with the plan, so the output directory is neither created,
  changed, nor removed. (b305)

• A build under another --base-path into the same output directory
  rewrites its redirect stubs. The permalink, .md extension, and
  versionless redirect stubs refused to overwrite a file that differed from
  what they would write, which is right for another pass's output and wrong
  for their own stub from the previous build whose target moved, so a prefix
  change left every stub unprefixed and the conformance check failed the
  build. The build manifest now records the stubs a build owns; the next
  build rewrites those, and only those, when their targets move, adopts a
  stub it finds byte for byte as its own, and removes a recorded stub it no
  longer emits instead of leaving it to dangle. Anything else at a recorded
  path stays. (b254)

• A rebuild into the same output directory no longer adds a fingerprinted
  copy of every theme asset per build. The fingerprint pass walks the
  output theme/assets/, where compiled styles live, so from the second
  build on it hashed its own copies and wrote editions.<hash>.<hash>.css,
  one more segment per build, and nothing removed them. A file shaped like a
  copy of a sibling that exists beside it is now a copy, never a source: the
  current one is left in place, and a stale one, an old hash or a doubled
  name, is removed once the build has succeeded, so the directory holds one
  copy per asset however often you build into it. A file the theme ships
  under a hash-shaped name beside something the build does not fingerprint
  is left alone. The emitted HTML was never affected. (b255)

• An incremental build after a stylesheet edit renders every page again
  when fingerprinting is on. The fingerprint filter writes the hashed
  URL of an asset into every page that links it, and no page hash saw the
  asset behind it, so a build without --clean after a stylesheet edit
  rendered only the pages whose content changed and left every other page
  linking the previous hash. The build manifest now records a digest of the
  fingerprint manifest, and a change to it is a full-rebuild axis, like the
  theme and schema hashes. (b304)

• An enum drop-down shows the file's value even when it is not a declared
  one. A component attribute or frontmatter field holding a value outside
  its declared set displayed the first declared option, while the finding
  beside it named the real value. The value is now a selected option of its
  own.

• Ignored directories and the site's own config file are no longer served
  or published. On a site whose content directory is the tree root, as
  accent serve <dir> --write-config writes it, every image, HTML, JSON, or
  YAML file under node_modules/ or under a path the tree's .gitignore
  hides was a public URL, and so was the site's config.yaml. With
  content.auto_index on, what the loader hides is now not a static asset
  either, at the bare page URL, under /content-media/, or in the build
  output; and the loaded config file is never served or published in any
  layout. The preview and the deployed site answer the same media URLs.

• accent build refuses an output directory inside the content
  directory. A site whose content.directory is ., which
  accent serve <dir> --write-config writes, built to the default output
  inside its own tree. The page-local media pass then copied the output as
  page media, then the copy, until the path was longer than the filesystem
  accepts. The build now refuses such an output directory before it writes
  or cleans anything, naming both directories, and --clean refuses an
  output directory that contains the content directory, since cleaning it
  would delete the site. The same guard covers every directory the build
  reads: an output inside a mount source, the media directory, the theme's
  assets/, or the .well-known directory is refused too, since the next
  build would read it back, and with --clean the output directory may not
  be, contain, or lie inside any input directory, the theme, models, data,
  and plugin directories included.

• accent validate no longer reports "Can't nest" warnings for valid
  Markdown. A page with an HTML comment and a tag reported "Can't nest
  'text' in 'document'" at line 1, and bold, italic or link text that wraps
  onto a second source line reported "Can't nest 'softbreak' in 'strong'".
  The documentation site went from 608 warnings to none. The first came from
  accent-proust, fixed in 0.12.1; for the second, Accent allows line breaks
  inside emphasis and links when it validates.

• A directory preview no longer polls build output and .git.
  accent serve <dir> watched the whole directory, so serving a repository
  polled every file under target/, node_modules, and .git twice a
  second, and reloaded the page whenever a build or a git command wrote
  there. The watcher now skips what the preview skips: hidden entries, the
  content.auto_index.ignore names, and gitignored paths. The same applies
  to a plain accent serve of a --write-config preview. accent serve <file> watches only the file's own directory.

• accent build leaves media URLs in code samples alone. A page that
  showed a processed-image URL such as /media/hero.jpg?w=960 in a code block
  or inline code got a Media source not found warning for each query. When
  the site also had real processed images, the build rewrote the sample to a
  generated filename such as /media/hero-960w.jpg. Code samples now build as
  written, and a missing source image is reported once, naming a page that
  uses it.

• Pagination works on static hosts. A paginated collection linked its
  pages as /blog/page:2, but the build wrote them to blog/page/2/, so on a
  deployed static site every link past the first page was a 404. Pages now
  live at /blog/page/2 under both accent serve and accent build, and
  accent serve redirects the old /blog/page:2 form permanently. A
  collection on the home page linked //page:2, which browsers read as
  another host; it now links /page/2.

• Builtin components under a base path linked to the prefix twice. With
  site.base_path: /repo, {% download %}, {% video %} and {% search %}
  wrote /repo/repo/.... The prefix is now applied once, including a video's
  poster.

• accent init --docs includes the files its pages link to. The docs link
  a sample script under media/downloads/, which scaffolds did not include.

• diagram() shows a caption once when the diagram renders in the
  browser. With diagrams.render_mode: client, or hybrid in dev, a
  caption rendered twice: once in the browser renderer's figure and once in
  the wrapper's.

• Site search no longer throws on a short query, and a new query
  restarts keyboard selection. Typing a single character logged a
  ReferenceError from the search script, and after arrowing through
  results, typing more kept the old selection, so Enter could open a
  result other than the highlighted one. Affects every theme that uses
  search_form().

• search_form() renders nothing while search is disabled. With
  search.enabled: false, a theme that called it unconditionally rendered
  a search box that could not work, and accent build output carried a
  script tag for a _search/search.js the build never wrote.

• A future_only or past_only bound on an object sub-field is
  enforced. A temporal constraint one level down -- event.when: { type: date, past_only: true } -- was only reported when something else
  on the same field already had a finding, so a page whose one problem was
  a date on the wrong side of now validated clean. accent validate,
  accent build, the admin save, and the publish gate all report it now.
  A page relying on the gap reports a new warning, or a new error under
  content.validation.mode: strict; the fix is to correct the date or drop
  the bound.

• A URL built by a template helper can no longer break out of its
  attribute. media(), url(), cdn_url() and pdf_card() mark their
  result safe so a URL's own delimiters survive into the page, but part of
  that result is author input: a " in a component attribute or a
  frontmatter field closed the attribute and opened another, so content
  could add a live event handler to a theme's <img> or <a>. Those
  helpers now percent-encode the characters that may not appear in a URI
  and that end an attribute. No ordinary URL changes, and an
  already-encoded path is left alone.

• A component's link works on a sub-path deployment. A component
  template writing href="{{ attrs.href }}" emits the autoescaped
  / spelling of a root-absolute link; the base-path rewriter read
  only the literal spelling and left it unprefixed, which the build's
  conformance check then failed. Both spellings are now rewritten.

• A page that opens with a component gets a readable description.
  Component tag syntax reached the derived lead -- and so the meta
  description, the social-card text, the llms.txt summary and the search
  snippet -- and both search indexes tokenised attribute names as content
  terms. Tag syntax is now stripped alongside raw HTML, keeping the tag's
  body, and code samples showing a tag are preserved.

• Editing a component manifest hot-reloads. A shortcodes/*.yaml edit
  was classified as an uninteresting change, so the dev server kept serving
  the cached page with the old declaration while the component's template
  half reloaded normally.

• Loading a content directory that holds build output no longer takes
  minutes. The content scan indexed every non-markdown file as a media
  asset and, for upload duplicate detection, read and hashed each one; a
  repository root with 145 GB of build artifacts under auxiliary target
  directories took over five minutes to load on every command. The hash is
  now computed only for a file an upload could produce (within
  admin.media.max_upload_mb, MIME type on admin.media.allowed_mime_prefixes);
  everything else stays indexed without one. accent serve <dir> also
  prints a scanning line before the load.

• Pages that name a template the viewer theme lacks render instead of
  returning 404. Under the embedded viewer theme (accent serve <dir>,
  accent serve <file>, theme.embedded: viewer) a page with an explicit
  template: from another theme returned 404 in serve and aborted the
  build; serve, build, and the admin preview now share one rule and fall
  back to the viewer's default layout. A theme on disk keeps the loud
  error for a missing explicit template.

• Every string round-trips through the YAML writer and reader, block
  scalars included. Four shapes the writer spelled in a form the reader
  read back differently are fixed in the YAML library: a value that is
  exactly one newline read back empty; a value ending in two newlines
  gained a newline on every save when another field followed it; a value
  whose first line is spaces only produced a file the reader refused; and
  a list item whose first line starts with a space or a tab read back with
  two extra spaces. The round-trip property tests that found them now run
  with no exclusions (b212).

• A duplicate key in config.yaml, or in a page's frontmatter, is
  reported with its path and position (site.name: duplicate key "name" at line 3, column 3) instead of the bare key name.

• A frontmatter field name containing [, ], or * is written as key
  text. The lossless editor handed every path to the YAML library's query
  grammar, which reads [n] as a sequence index and * as a wildcard: a
  --set a[x]=1 wrote the key a, and a --remove a[x] deleted a. The
  grammar gained a quoted form for such a segment, and the editor spells
  every segment through it, so such a field is added, changed, removed, and
  renamed under its own name like any other (b234).

• A ${VAR} placeholder under a config key containing ., [, ], or *
  (site.meta names such as og.type, a versioning bucket like
  v0.26-dev) is expanded in place like any other. Such a key used to send
  the whole file down the tree route, which rewrote every number-like
  string in it (01234 loaded as 1234) and moved a type error's position
  out of the file, and only a debug line said so. The one site that still
  takes that route, a placeholder inside an anchored value, is logged as a
  warning naming it (b235).

• Config values keep their spelling when environment placeholders are
  expanded. A string value that YAML would read as a number, such as an
  SMTP password 01234, was rewritten to 1234 every time config.yaml
  loaded, because ${VAR} expansion re-emitted the parsed tree before the
  typed read; the server then authenticated with the wrong secret and
  reported nothing. Expansion now edits the file text in place, so every
  value outside a placeholder reaches the config exactly as written, and a
  type error after an expanded line reports the line number in your file
  rather than one in an internal buffer (b227).

• A frontmatter field name containing a non-printable character (a line
  break, DEL, a C1 control, or the non-characters U+FFFE and U+FFFF; tab
  is allowed and written as an escape) is refused by name before anything
  is read or written: the MCP update_frontmatter and create_page tools
  return an invalid-params error, accent content update-frontmatter --set and --remove exit with the key named, and every content API
  write -- the admin save and the content agent included -- refuses the
  operation. Previously an agent-supplied key with a newline in it reached
  the writer, and before the YAML emitter fix collapsed the page's
  frontmatter into a single string (b212).

• A page whose frontmatter cannot be parsed no longer vanishes silently.
  The scan now logs a warning naming the file and the error (serve and
  build both show it), accent build still ships every other page, and
  accent validate reports the skipped file as an error and exits 1 --
  previously it printed "No issues found" over the dropped page.

• Theme templates HTML-escape {{ ... }} expressions again. The explicit
  auto-escape callback introduced with the Markdoc bridge returned
  AutoEscape::None for every non-component template, believing that was
  MiniJinja's default for *.html.jinja names; the default in fact strips
  .jinja and applies HTML escaping, so authored values had always been
  escaped -- and briefly were not (b223).

• Site protection (protection.basic_auth, protection.ip_allowlist) on a
  server started without --production is enforced whether or not
  dev.browser_reload is on. The free-in-dev allowance keyed off the
  browser auto-refresh setting, so turning that off on an unlicensed dev
  server silently served every path open, with one WARN per request that
  blamed licensing. The gate now follows the --production flag alone;
  under --production a Standard or Pro license still decides.

• A Basic Auth username or password that is still an unexpanded ${VAR}
  placeholder at startup refuses to start, naming the user entry and the
  variable. Before, an unset SITE_AUTH_PASSWORD protected the site with
  the literal string ${SITE_AUTH_PASSWORD} -- a value written in the
  config file. Empty credentials are rejected the same way;
  ${VAR:-default} keeps working and authenticates with the default.

• The HTTPS token for git-backed content no longer appears in the git
  child's command line. It reaches the credential helper through the
  child's GIT_REMOTE_TOKEN environment variable; argv carries only the
  variable's name, as the module documentation had claimed all along.

• The Linux release binaries run on Debian 10, Ubuntu 20.04, RHEL 8,
  Amazon Linux 2023 and newer again. v0.25.0 needed glibc 2.39 and refused
  to start on Debian 12 and Ubuntu 22.04, because both Linux targets
  inherited the glibc of whatever image built them (ubuntu-latest and the
  cross-rs main image both moved to Ubuntu 24.04). Both targets are now
  linked against a pinned glibc 2.28 floor, the release build checks the
  produced binary against that floor, and a smoke job runs each Linux
  archive on a glibc 2.28 base before anything is published. The aarch64
  Linux binary also embeds the dependency SBOM now, like every other target.

• install.sh runs the binary once before installing it. A binary that
  cannot start on this system is no longer installed under an
  "Installation complete!" message: the installer exits 1 with the loader's
  error, the glibc version the release needs and the one the system has,
  and leaves any earlier installation untouched.

• Hot reload watches document models. Editing a model in the project
  models/ directory, in the theme's models/, or in a plugin's models/,
  or a _model.yaml under content/, during accent serve now rebuilds
  the content index and reloads the browser, so validation, the admin
  schema browser, and the editor field views show the new schema on the
  next request. Before, the edit stayed invisible until an unrelated
  content or template save, an admin page save, SIGHUP, or a restart. A
  file the watcher cannot classify (an editor swap file, a README) no
  longer reloads the browser either.

• api.allowed_origins is honoured. The legacy api.cors layer applied
  permissive CORS regardless of the list, so the documented way to restrict
  origins in production did nothing. ["*"], the default, is unchanged.

Verifying your download

curl -fsSLO https://raw.githubusercontent.com/AccentCMS/accent/main/release-signing-key.asc
gpg --import release-signing-key.asc
gpg --verify checksums-v0.26.0.txt.asc checksums-v0.26.0.txt
sha256sum -c --ignore-missing checksums-v0.26.0.txt