# Changelog

All notable changes to SaglitzDesign MCP are documented here.
The format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
and this project adheres to [Semantic Versioning](https://semver.org/).

## [0.28.0] — 2026-08-31

v0.27.0's tag is pushed (`refs/tags/v0.27.0` exists on `origin`) but the
release it triggered failed: build, `preflight-release`, the full suite and
the packed-tarball smoke test all passed, and the "Publish to npm" step then
hit `npm error 404 ... could not be found or you do not have permission to
access it` on the registry `PUT` (run `33386023873`). That is the documented
failure mode of the one-time setup this package has never had — npm's OIDC
Trusted Publisher, configured once at npmjs.com → **saglitzdesign-mcp** →
Settings → Trusted Publisher (see `RELEASING.md`) — not a defect in that
release's own gates, all of which ran clean first. `npm view
saglitzdesign-mcp version` still answers `0.19.1`. So this is now the
**tenth** entry in this file describing work nobody outside this repository
has received — v0.20.0 through v0.26.0, v0.18.0 before them, v0.27.0, and this
one — so publishing v0.28.0 is not a small increment shipped on top of an
already-shipped v0.27.0. It ships everything described in all ten entries at
once, to a registry that currently has none of it.

This release fixes four word-boundary defects in `audit_ux_copy`'s matchers
— three in its vocabulary lists, and one shared by every matcher in the file,
which read any non-ASCII letter as a word edge — and gives the tool a
disclosure list, structured findings and a structured metrics block, matching
the seven auditors that already had the first two.

### Fixed

- **The three vocabulary matchers in `audit_ux_copy` fired on substrings of
  ordinary English, not the words on their own lists.** `FILLER` used
  `lower.includes(f)` with no boundary at all: `just` fired on "Adjust the
  layout" (via *adjust*), and `very` fired on "Every delivery" (via *every*).
  `WEAK_CTA` used `startsWith`, so `go` fired on "Government portal".
  `JARGON` used `\b`, which treats a hyphen as a boundary, so `cutting-edge`
  fired inside `non-cutting-edge` — a phrase meaning the opposite of jargon.
  One helper now serves all three, over a character class derived from what
  these entries are (prose words and hyphenated marketing jargon: letters,
  digits, and the internal hyphen `cutting-edge` needs), not from whichever
  character a failing fixture showed next. `WEAK_CTA` **keeps `startsWith`
  deliberately** — "Submit form" is a weak CTA and still has to fire — and
  gains the same trailing boundary the other two lists gained, plus the
  ability to begin after a leading quote, bracket, or emoji: `"Go"`, `(Go)`,
  and `👉 Go` now all fire; "Government portal" still does not.

  Checked directly rather than trusted from fixtures alone — and the result is
  not the same for all three lists, so it is stated per list rather than as one
  sentence about "the matcher". Over the markdown this repository tracks
  (`git ls-files '*.md'`, 151 files as of this change) the new **jargon and
  filler** matchers gained a hit in zero files and lost one in 128. Most of
  what they lost is the defect above: `very` inside *every*, *everything*,
  *everywhere*, *recovery*, *delivery* and *discovery* accounts for 124 of the
  removed hits, and `just` inside *adjust*, *justify* and *justification* for
  17. Not all of it, though, and "every removed hit was a false positive" would
  be a second claim resting on nobody having read them: `leverage` inside
  *highest-leverage* (9) and `obviously` inside *obviously-secret* (1) are
  hyphenated compounds where the listed word really is present, and the new
  boundary declines them deliberately — one rule cannot treat the hyphen in
  *non-cutting-edge* as part of the word and the hyphen in *highest-leverage*
  as a break.

  **`WEAK_CTA` moved the other way, on purpose.** Against the pre-change
  matcher, 12 of 26 short-copy inputs differ: `"Go"`, `(Go)`, `👉 Go`, `_Go_`,
  `“Go”`, `›Go`, `#Go`, `→ Continue`, `**Submit**` and `[Continue]` all fire
  now and none of them did before, while `Government portal` and `Okay`
  stopped firing. Those ten are new matches, so no claim of the form "this
  change only ever removes a match" is true of the tool as a whole — only of
  the two lists named above, on the corpus named above. That corpus cannot
  speak to `WEAK_CTA` in either direction, which is worth stating because it
  is the trap: `isLikelyCta` is `wordCount <= 5 && sentCount === 1`, and
  **zero** of those 151 markdown files satisfy it, so a per-file old-vs-new
  comparison provably never evaluates `WEAK_CTA` once and would report "no
  violations" whatever that list's change had done.

- **A non-ASCII letter was read as a word boundary, so an English entry could
  match inside a word in another script.** `[a-z0-9-]`, `\w` and `\b` are all
  ASCII-only, and an ASCII-only boundary does not merely fail to read another
  script — it silently reclassifies every letter of that script as a word
  *edge*. Turkish `Çok yakında` reported the weak CTA `ok`: `Ç` is not in
  `[a-z0-9-]`, so the anchor skipped it as leading noise and `ok` matched
  inside the word. `robust` matched inside `設定robust設定` and `just` inside
  `простоjustдалее` for the mirror reason, and `\b`'s ASCII definition made the
  "we"/"our" counter read the end of German `Löwe` as first-person voice. All
  four matchers in `src/uxcopy.ts` now take their notion of "inside a word"
  from one `\p{L}\p{N}` class instead of each carrying its own, so what is
  fixed is the class rather than the letter that exposed it. `\p{L}\p{N}`
  alone was still one class short of the same defect — a combining mark or a
  zero-width joiner is not a letter or digit either, so it was still read as
  an edge instead of as part of the letter it modifies (`設定‍robust‍設定`
  with a zero-width joiner still fired, an NFD-normalized `Löwe` still counted
  as ending in "we", and Hebrew niqqud — which has no precomposed form to
  normalize away — still read as a boundary). `\p{M}` and `\p{Cf}` close both.

  ASCII copy is close to unaffected, with a self-reference this sentence
  cannot avoid being part of: the only behaviour change on plain ASCII text is
  that a TypeScript identifier's underscore now separates two words that read
  as themselves once split (`en_US`, `WE_RE`) — the same underscore-is-an-edge
  rule the word lists already applied to `_robust_`, now applied by the
  you/we counter too. This file is one of the inputs that changes, because it
  quotes `WE_RE` above and is itself part of the markdown corpus that rule
  reads. No fixed count of affected files is given for that reason: a
  sentence describing this change is, by describing it, one of the things the
  change affects, so a number attached here would be stale on arrival — the
  same self-reference the corpus hit count in `audit_ux_copy`'s own
  disclosure list already had to be rewritten around.

- **`audit_ux_copy` called copy "clear" over a metrics table saying it was
  not, and pointed a vocabulary finding at the wrong line.** Reading grade
  level is the one row that table grades against a target while no rule fires
  on it, so a grade of 8.4 could print `≤ 8 ⚠️` directly above "✅ No copy
  issues flagged — clear, active, user-focused."; the all-clear now names the
  off-target metric instead of contradicting it. And a jargon or filler
  finding's `line` was `hits[0]`'s position — first in *word-list* order, not
  in the text — so "This design is innovative.\nWe built a robust system."
  pointed at line 2 while the first flagged word sits on line 1. It is now the
  earliest hit in the text, as its own doc comment always claimed.

### Added

- **`audit_ux_copy` now returns `structuredContent`** — one `LintFinding` per
  triggered check, in the shape `design_lint` and the other structured
  auditors already use — so an agent can chain its findings instead of
  re-parsing markdown. It also now renders a "Not visible to this audit"
  disclosure, making it the **eighth** tool in this server to publish one,
  joining `design_lint` and the six `audit_*` tools that already had it. The
  disclosure names what class of thing the tool cannot see — among them, the
  three word lists' closed, English-only vocabulary — which is not the same
  as silence on other languages, since `Das System ist robust.` reports
  `robust`; the shape gate (`isLikelyCta`) that skips any candidate CTA over
  five words or split across more than one sentence, and grades every other
  five-word sentence as a CTA whether or not it is one, so
  `Enter a valid email address.` reports the weak CTA `enter`; a syllable
  count produced by a heuristic with no dictionary behind it; a sentence
  splitter with no abbreviation handling, so a title like "Mr." inflates the
  sentence count it feeds every readability number the tool reports; a passive
  regex that misses irregular participles and fires on predicate adjectives,
  so `This field is required.` is reported as passive voice; a `passiveHits`
  list capped at 8 whose length both registers print as a total; and — the
  one that matters most for a reader acting on a clean result — an ASCII-only
  tokeniser, which returns *zero* words for a script written without ASCII
  letters and reports that as one word, a Flesch score of 206 and four green
  ticks. Two Japanese sentences come back as a confident pass, not as
  silence. The list is not reproduced entry-by-entry here — a second copy in
  this file would drift from the tool's own the first time either changed, so
  read the tool's own output for all of it. (`skills/ship-quality-gate/
  SKILL.md`'s row summarises a few of the entries and says itself that it is
  not the list.)

  `structuredContent` also carries the metrics table itself — words,
  sentences, average sentence length, Flesch reading ease, grade level and the
  you/we counts — the way `audit_generic_design` carries its `score`. Without
  it an agent reading only the structured half got findings and no numbers,
  and reading grade level, which is graded in the table but raises no finding,
  had no machine-readable form at all.

  One number in that disclosure is deliberately not exact: how often the
  three word lists fire across this repository's own 105-file documentation
  corpus is stated as "well over a hundred hits across the large majority of
  those files," not a point count — because that corpus includes the
  documents describing the tool's own reach, this changelog entry now among
  them, so quoting a flagged phrase to document the gap moves the count each
  time it's written about.

## [0.27.0] — 2026-08-30

Eight of the entries below describe releases that were never published:
v0.20.0 through v0.26.0, and v0.18.0 before them (`npm view
saglitzdesign-mcp versions` lists neither — it jumps 0.17.0 → 0.19.0 → 0.19.1
→ nothing since). The tag/publish step is the maintainer's to run, and for
those eight releases it has not run. (`git tag` returns nothing at all, so
"never tagged" does not distinguish these eight from the nine versions that
*were* published — only "never published" does.)
`npm view saglitzdesign-mcp version` still answers `0.19.1`, so v0.19.1's entry further down is the newest one
anyone outside this repository can actually install. Counting this one, nine
entries now describe work nobody outside this repository has received, so
publishing v0.27.0 does not ship "a design front door" as a small increment
on v0.26.0 — it ships everything described in all nine entries at once, to a
registry that currently has none of it. Read this entry, and the eight before
it back to v0.20.0 and v0.18.0, as a description of what publishing v0.27.0
will deliver, not of what already shipped.

This release adds a front door into the seven design skills, two CSS-spec-
sourced lint rules, three motion lint rules sourced to accessibility
specifications and engine documentation, and a device against one specific
way generated pages look generated.

### Added

- **`saglitzdesign` — an eighth skill directory, and the one an agent actually
  reaches for.** Its `description` frontmatter field is what carries its
  trigger vocabulary — there is no separate keyword list, so the description
  is where every phrase that should reach this skill has to appear — and it
  answers to "design", "redesign", "build", "restyle",
  "improve", "critique", "review", "audit", "simplify", "polish", or
  "animate" a user interface, plus unasked appearance decisions such as
  porting a component to
  another platform or turning a mockup into code. It ends on a boundary
  instead of trailing off: **not** for pure functionality with no appearance
  decision in it — "make the form work" or "add sorting to the table" should
  not reach for this skill. Once triggered, it carries a routing table sending
  the work to whichever of the seven depth skills actually fits, and four
  invariants that hold regardless of which one that is: generate the design
  system before the pixels, bind every value to a token, run the deterministic
  auditors at the close, and ground every claimed number in a tool's measured
  output.

- **Three tests holding the umbrella to what it claims**, added after the first
  version of its own frontmatter broke on the very CLI it ships through:
  `npx skills@latest add ./ --dry-run` reported "Found 7 skills" against a
  tree of eight `SKILL.md` files and printed a `⚠ Skipped` line naming
  `saglitzdesign` — a YAML parse failure (an unquoted description containing
  `: `) that every regex-based guard in the suite stayed blind to, because
  none of them parses YAML. All 1622 tests were green at the time. One new
  test parses the umbrella's routing table and asserts, in both directions,
  that it names every depth skill directory on disk and no directory that
  isn't one; a second walks all eight `SKILL.md` frontmatter blocks and
  rejects the handful of plain-scalar shapes (an unquoted `: `, a trailing
  `:`, an unquoted ` #`) that break the real parser the skills CLI uses; a
  third holds the umbrella's own description to the boundary sentence that
  keeps it off pure functionality. All three are narrow by design — the
  second is explicitly not a YAML parser and says so in its own comment — but
  they now stand between a broken skill and a release, which nothing did
  before.

- **`grid-track-no-min` — a CSS Grid `info` lint rule, sourced to CSS Grid
  Level 1 §7.2.1 and §6.6.** §7.2.1 defines a bare `1fr` track as
  `minmax(auto, 1fr)`; §6.6 is what decides whether that automatic minimum
  resolves to the grid item's content-based size or to zero, and that
  decision depends on the item's own overflow, whether the track has an
  `auto` min sizing function, and whether it also spans a flexible track.
  None of those conditions are visible from a single declaration, which is
  why the rule is `info` rather than `warning` and why it says a track *can*
  hold above its `fr` share, not that this one does — it is a robustness
  note pointing at `minmax(0, 1fr)`, not a proven overflow. It fires only
  where the snippet also contains an `<img>` — one of the few intrinsic-size
  cases the rule can see in source alone, and a gate that is snippet-wide
  rather than grid-scoped: any `<img>` anywhere in the text handed to
  `design_lint` satisfies it, so a bare-`1fr` grid with no image of its own
  still fires when an unrelated `<img>` sits elsewhere in the same snippet.
  The tool's own `notVisible` list says exactly that, and an earlier draft of
  this entry narrowed it to "the track", which is the flattering direction
  and the wrong one.

- **`overflow-hidden-root` — a `warning` lint rule, sourced to CSS Overflow
  Level 3 §3.1**, which states that `overflow: hidden` leaves a box "still a
  scroll container" (only the scrolling UI disappears; programmatic scrolling
  still works), while `overflow: clip` is not a scroll container at all. That
  is the entire claim: `hidden` on `html`, `body`, or `:root` makes that
  element a scroll container. It is not a claim about `position: sticky` —
  a sticky-positioning failure would need a source in CSS Position 3, and
  this rule cites neither that specification nor sticky at all.

- **Three motion lint rules, and two refused for lack of one.** The sourcing
  is the point:
  - `motion-no-reduced-cover` fires on an `animation`/`animation-name`/
    `@keyframes` with no `prefers-reduced-motion` query anywhere in the same
    source. Not on *any* of them: a literal `none` value — the ordinary
    inert reset, quoted or not — is deliberately excluded, so
    `.a { animation: none; }` and `style={{ animation: "none" }}` are silent
    where `.a { animation: slide 1s; }` fires. WCAG 2.1 SC 2.3.3 (Animation
    from Interactions) is Level AAA and
    scoped to motion "triggered by interaction" — it does not reach an
    autoplaying or otherwise non-interactive animation, which this rule still
    flags. Media Queries Level 5 §12.1 is the source that actually covers the
    rule's full reach: it frames `prefers-reduced-motion` with no interaction
    restriction and no AAA-only carve-out. SC 2.3.3 is cited here only as the
    accessibility rationale for the interaction-triggered subset, not as
    coverage for the whole rule.
  - `animates-layout-property` rests on MDN's engine-rendering documentation
    of which properties are compositor-only, not on a CSS specification the
    way the two rules above do. It flags `width`, `height`, `top`,
    `left`, `right`, `bottom`, `margin`, and `padding` in a `transition`,
    `transition-property`, or `@keyframes` body — the keyframes half reading
    the enclosing block rather than the line, so a normally-formatted
    keyframes block is graded like a collapsed one — because animating them
    forces layout and paint every frame where `transform`/`opacity` would
    composite alone. That is a performance claim, not a correctness one: the
    documented consequence is *slower*, not *wrong* — nothing about the
    animated element is broken by animating `margin` instead of `transform`.
  - `transition-all` (`info`) is a superset of the same claim, sourced the
    same way: naming `all` animates every layout property above plus
    whatever else changes on the element, which is a wider blast radius than
    naming properties explicitly, not an elegance complaint.
  - Two candidate rules from the original plan — flagging non-uniform
    `hover:scale-105`-style values, and flagging "bouncy" easing curves —
    were dropped before shipping. Neither has a specification or engine
    documentation behind it, only a taste judgment, and this package ships
    lint rules with sources or not at all.

- **A macrostructure stamp in `clean-interface-design` and
  `landing-page-conversion`.** Both skills now name five page shapes —
  `centered-stack`, `split-panel`, `sidebar-shell`, `bento-grid`,
  `rail-scroll` — and instruct the agent building a full page to check the
  project's existing stylesheets for a prior `/* saglitzdesign ·
  macrostructure: … · nav: … · footer: … */` comment and differ from every
  axis it records, so a project doesn't accumulate screens that are all the
  same shape underneath different colors and copy. The stamp carries all
  three axes because the instruction asks the agent to differ on all three:
  an earlier draft recorded the shape alone while asking for a different
  navigation and footer than "the stamped run used", which the stamp did not
  carry. No tool writes this stamp: it's a
  comment the agent leaves for itself in CSS it is already writing. That
  isn't an oversight — every one of this server's 34 tools is registered
  through a single wrapper that stamps every registration with
  `readOnlyHint: true` among its MCP annotations; a new tool added to write
  this stamp to a project's files would go through that same wrapper while
  actually writing to disk, making that declaration false for the one new
  tool that wrote it.

### Fixed

Everything in this section is a rule that fired on **correct** code. Measured
before the fix, a stylesheet in which every decision was right — tokens in
`@layer tokens`, a `@supports` feature query for `overflow: clip`,
`repeat(auto-fit, minmax(200px, 1fr))`, `min-height` over `height`, a
`transition` naming only `transform` and `opacity`, a `prefers-reduced-motion`
block, a `:focus-visible` ring — drew **four findings, two of them warnings**.
It now draws none, and that stylesheet is a test.

- **`animates-layout-property`'s `@keyframes` half was unreachable on any
  formatted stylesheet.** It was a single-line regex needing the at-rule
  keyword, its `{` and the moving declaration on one physical line, so a
  keyframes block written across lines drew nothing while its collapsed
  one-line form drew the finding. Every fixture behind it was a collapsed
  one-liner, so the rule and its tests shared one blind spot and each kept
  agreeing with the other. The keyframes half now reads the enclosing block
  chain, the fixtures are written the way stylesheets are written, and the
  finding lands on the moving declaration's own line.
- **`animates-layout-property` fired on a JS/JSX style object that transitions
  only composited properties.** The value span was `[^;]*`, written for CSS,
  where `;` ends a declaration; an object literal ends it with a comma, so the
  span ran into the next key and any sibling named `margin`, `padding`, `top`,
  `left`, `right`, `bottom`, `width` or `height` supplied the match.
  `{ transition: "opacity 200ms", margin: 0 }` fired. The span is now bounded
  by the value's own quotes; a genuine comma-separated CSS transition list
  still fires.
- **`motion-no-reduced-cover`'s inert-reset exclusion was defeated by a
  quote.** `animation: none` was excluded; `style={{ animation: "none" }}` was
  not, because the exclusion read the character straight after the colon. It
  now steps over one opening quote.
- **`overflow-hidden-root` fired on two of the three standard spellings of the
  fallback idiom it recommends.** Only the same-block form
  (`overflow-x: hidden; overflow-x: clip;`) was recognised;
  `@supports not (overflow: clip) { body { overflow-x: hidden; } }` and
  `body { overflow-x: hidden; }` followed by an `@supports (overflow: clip)`
  upgrade both fired, on code that has explicitly done what the rule's `fix`
  asks. Both are now recognised, with their limits disclosed.
- **A nested at-rule prelude silenced `overflow-hidden-root` outright.** The
  override guard stripped a nested block's *contents* but not the prelude
  introducing it, so it read the feature-query condition in
  `body { overflow-x: hidden; @supports (overflow: clip) { ... } }` as a live
  `clip` declaration and cancelled a real finding.
- **The three value rules stopped flagging the design tokens they recommend.**
  `\b` treats `-` as a boundary, so `--border-color: #e5e7eb` drew
  "hardcoded hex instead of a design token", `--header-height: 64px` drew a
  fixed-height note, and `--border-radius: 8px` drew an ad-hoc-radius note —
  each while the identically-correct `--brand`, `--height` and `--radius` were
  silent, on nothing but whether the token's name ended in a watched word.
  `animates-layout-property` had already closed this class for itself in this
  release; `hardcoded-color`, `fixed-height-text` and `magic-number-radius`
  now carry the same exclusion.
- **`fixed-height-text` no longer flags `min-height`.** A floor cannot clip
  anything, and `min-height` + padding is word-for-word what the rule's own
  fix text recommends: it was flagging its own advice. `max-height` and
  `line-height` are excluded by the same change.
- **`magic-number-radius` no longer flags a squared corner or a pill.**
  `border-radius: 0` and `border-radius: 9999px` are the two least ad-hoc
  radii there are. Whole pixels from 1 to 999 still fire.
- **`overflow-hidden-root` cited a document that does not make its claim.**
  `spacing-layout` names the symptom people reach for `overflow-x: hidden` to
  hide and says nothing about `hidden` versus `clip` or about scroll
  containers; the rule's real source lived only in a code comment. The claim
  now lives in `modern-css-design-primitives`, the rule cites that, and
  `tests/lint.test.ts` gained the claim guard `tests/apple.test.ts` has
  carried since the SEO package — it checks that every cited document
  actually carries its rule's claim, not merely that the id resolves.

### Changed

- **Skill users must re-run `npx skills@latest add HalidSaglam/saglitzdesign-mcp`
  to get the new `saglitzdesign` skill and the macrostructure stamp** — the
  two changes above that live inside `skills/`. The five lint rules added in
  this release (`grid-track-no-min`, `overflow-hidden-root`, and the three
  motion rules) live in `src/lint.ts` and reach only an MCP or plugin
  install, never a skills-only
  one; the two guard tests live in `tests/` and ship in neither — a skill
  install gets none of this release's other changes either way. A skill is
  copied into the installing agent at install time and
  pinned by content hash in `skills-lock.json` — nothing about this
  repository changing reaches an already-installed skill on its own. This is
  not the first release to change a skill after it shipped: v0.15.0 below
  already modified five existing `SKILL.md` files, and a previous version of
  this plan claimed v0.27.0 would be the first such release. It would have
  been false there and is false again here.

## [0.26.0] — 2026-08-28

The skills are a second distribution of this server's knowledge, and they had
drifted from it. `skills/README.md` advertised 83 documents and 26 tools while
96 and 34 shipped. The root README said "Five skills" with six directories on
disk. The design-review skill scored against rubric bands that
`design-critique-scoring` does not publish. `apple-platform-design` told a
reader to support Dynamic Type without saying that macOS is not on Apple's list
of platforms that have it. Six sentences — four in the skills, two in
the README — wrote a workflow as a bare slash plus its name: `design_review`
with nothing in front of it. With this plugin installed Claude Code answers that
with `Unknown command` and offers the similarly-named `design-review` *skill*
instead, so those sentences named an invocation nobody had ever typed. (This
entry cannot write that name out either: the guard added here to stop it coming
back reads `CHANGELOG.md` too.)

Three guards stood over `skills/` through all of this, and not one of the three
was shaped to catch the drift beside it: a frontmatter check whose only count was
`toBeGreaterThanOrEqual(5)`, which passes with six directories; a phantom-tool
check reading a hand-written list of tool names that had gone stale at 33 while
34 shipped; and a README check that read `skills/README.md` and never the root
README, where the wrong skill count was.

This release refreshes the skills against what actually ships, adds a seventh,
gives every workflow a name a user can type, packages the whole thing as a
Claude Code plugin, and reshapes the guards so the next drift fails a test.

### Added

- **A Claude Code plugin — the server, the seven skills and the eight workflow
  commands in one install.**

  ```bash
  claude plugin marketplace add HalidSaglam/saglitzdesign-mcp
  claude plugin install saglitzdesign@saglitz
  ```

  The plugin declares its MCP server as `npx -y saglitzdesign-mcp@latest`, so
  the server is fetched from npm rather than run out of the plugin directory: a
  plugin installed from git arrives as a checkout, `dist/` is gitignored,
  nothing builds it on install and there is no `node_modules` beside it, so a
  server declared as a path inside the plugin would fail twice over. `claude plugin details
  saglitzdesign@saglitz` prints what arrived.

- **`ship-quality-gate` — a seventh skill, for the enforcement half of the
  job.** The other six are guidance about making design decisions; this one is
  about running the deterministic auditors over a real repository and reading
  what comes back. It carries the seven tools that publish a disclosure list
  (`design_lint`, `audit_project`, `audit_security`, `audit_generic_design`,
  `audit_seo_geo`, `audit_performance`, `audit_apple_ui`) with each one's input
  mode and reach, the rule that a clean report is a statement about the files
  that were read and not about the rendered page, and the anti-patterns —
  reading silence as a pass, linting markup apart from its stylesheet, expecting
  the consistency count to see named-scale utility classes.

- **Eight slash commands, one per workflow, generated from the prompt
  metadata.** With the plugin installed the name is
  `/saglitzdesign:design_review`, `/saglitzdesign:build_landing_page` and so on,
  and everything after it is the brief. A workflow can be spelled more than one
  way and the spellings do not behave alike — README's *Workflows* section
  documents them and what each one costs. The commands are rendered from
  `src/prompts.ts` by `scripts/generate-commands.mjs`, with a test that
  regenerates them and compares byte for byte, so a hand-edit to a generated
  file cannot survive. Each carries `disable-model-invocation`, so the agent
  never starts one on its own — these are yours to type.

- **A `sources:` line in every skill's frontmatter**, naming the knowledge
  documents behind that skill, with a test that every id resolves to a document
  the server can load. "Behind" does not mean the same thing in all seven: six
  of them condense the documents they name, while `ship-quality-gate`'s eighteen
  are the documents its auditors cite rule ids from — it never condenses
  `auth-and-session-ux`, it runs a tool that quotes it. The field carries both
  readings today, and the test checks that the ids resolve, not which reading
  produced them.

### Changed

- **Six of the seven skills were edited, and an existing skills install carries
  none of it until it is refreshed.** A skill is *copied* into the agent —
  `npx skills@latest add` writes `SKILL.md` into `.agents/skills/<name>/` and
  pins its content hash in
  `skills-lock.json` — so a new release does not reach an existing install on
  its own. Re-run it to pick up the seventh skill and the six edits:

  ```bash
  npx skills@latest add HalidSaglam/saglitzdesign-mcp
  ```

  Re-running it over an existing install overwrites what is on disk — measured,
  by corrupting an installed `SKILL.md` and watching the hash come back.

- **`apple-platform-design`** now says which platforms Apple lists Dynamic Type
  for and that macOS is not among them, points at `audit_apple_ui`,
  `apple-accessibility` and `apple-shipping-readiness`, and no longer reads a
  fixed-metrics Mac view as a text-scaling failure.

- **`design-review`** scores against the five bands `design-critique-scoring`
  actually publishes (36–40, 28–35, 20–27, 12–19, 0–11) instead of the four it
  had invented.

- **The plugin description states four numbers and each is now held against a
  live source** — documents from `knowledge/`, tools and auditors from the
  running server, workflows from the prompt registry — and any further number
  added to that sentence fails the test until something counts it. Falsifying
  all four at once used to pass the whole suite and `preflight-release`.

### Fixed

- **Six sentences that named a slash command nothing registers.** Four skill
  pointers, plus two in the README.

- **The published list of `audit_security`'s header sources named fewer shapes
  than the audit reads.** Four surfaces describe that reach — the MCP tool
  description a client reads before deciding to call the tool at all, the
  report's own recognised-shapes bullet, the README's tool row and
  `ship-quality-gate`'s row — and all four listed the same fifteen shapes.
  Five more were already being read and named nowhere: `headers.set(…)` and
  `headers.append(…)`, Fastify's `reply.header(…)`, Firebase Hosting's
  `firebase.json`, and Next 16's `proxy.ts`. Worse, a list of places could
  never have been complete, because two of the shapes are *rules*: a quoted
  header-name property is read in any JSON or object literal, and any call
  whose method name is `set`, `setHeader`, `append` or `header` is read
  whatever the object is called — so Hono's `c.header()`, Koa's `ctx.set()`
  and a hand-rolled `headers.json` are all read too. Each of those seven was
  measured against a control: a project declaring CSP and HSTS that way gets
  no `csp-missing` and no `hsts-missing`, and the same project with those two
  lines deleted gets both. The skill row's word *only* over the short list is
  what made this a defect rather than a shortening — an agent was told to
  discount a result the audit had read correctly.

  All four surfaces now state the two rules and the five shapes. All five are
  now exercised by three new hardened fixtures in the per-framework matrix, so
  the reach is pinned as behaviour rather than as a sentence, and the guard
  that binds the four surfaces additionally asserts that every header-setting
  method the extractor accepts is named by one of the listed shapes — the
  direction that catches a shape being dropped from the list, which used to
  pass in silence.

### The drift check, and what it does not catch

The skill count in both READMEs is now an equality against the directories on
disk rather than a floor, and it is checked at every occurrence on the page
rather than at the first — a draft that stopped at the first match let a second,
wrong sentence sit below a correct one. The document and tool counts on those
same two pages are held against the live registry. The tool names a skill may
mention are read off the running server rather than from a hand-written mirror.
And any `/`-prefixed token in the documented markdown whose last segment is a
registered workflow must be written as the plugin command and have a command
file behind it — over a surface walked rather than listed, `skills/README.md`
and this changelog included.

What that does not reach is worth stating plainly, and each guard says it in its
own comment:

- The disclosure-tool catalogue in `ship-quality-gate` is held equal *as a set*
  to the tools that publish a `notVisible` list. It compares names and reads not
  one word of the prose beside them. Three sentences in that skill's first draft
  overstated an auditor's reach and this check passed on all three; it would
  pass on a row whose description was blank. What it catches is an eighth
  auditor with no row, a row for a tool that stopped publishing a list, and a
  rename on one side only.
- A skill's `sources:` list is checked for ids that resolve, not for being the
  right list. From the file alone, a document the skill should have named but
  did not is indistinguishable from a skill that condenses fewer documents.
- The slash-command guard requires the slash to open the token, so that a URL or
  a relative path ending in a workflow name is not a false alarm. `.../redesign`
  slips through as a result, which is the price of allowing `./redesign`.

### Upgrading

npm's `latest` is **0.19.1**, and the registry holds 19 versions ending there.
Seven entries below — 0.18.0, and 0.20.0 through 0.25.0 — are for versions npm
never received. Installing 0.26.0 is therefore one step across the six of them
that sit above 0.19.1: a server that registered 29 tools becomes one that
registers 34, and those six entries are the record of what arrives with it.
There is no intermediate version to move through, because none was published.

## [0.25.0] — 2026-08-16

v0.24.0 sourced this server's six Apple documents to Apple's own pages, and
nothing here could read an Xcode project against them. A team could read every
word of `apple-accessibility` and `apple-shipping-readiness` and still ship a
custom colorset with no dark appearance, a deprecated `UIRequiresFullScreen`,
a microphone entitlement declared under one capability and not its twin, and a
`NavigationView` on every screen. This release adds the auditor that reads the
project instead of trusting the reader remembered the document.

### Added

- **`audit_apple_ui` — the 34th tool, and the seventh to return structured
  output.** Point it at the Xcode project directory. It reads the four
  surfaces a project declares configuration on — the information property
  list, `INFOPLIST_KEY_*` build settings in `project.pbxproj`, the
  entitlements plist, and each colorset's `Contents.json` — infers whether the
  target is iOS or macOS from those plus the Swift imports, and runs eight
  rules over the text of the files it opened.
  - **Configuration, four rules.** A custom colorset declaring no
    `luminosity: dark` appearance; the deprecated `UIRequiresFullScreen`,
    either spelling, with the misspelling called out as a key the system does
    not read; one of the two Audio Input entitlement identifiers without the
    other, which one Xcode checkbox label covering two different capabilities
    makes easy to half-declare; and, on macOS only, no App Sandbox entitlement
    — reported as a fact about the Mac App Store channel rather than as a
    defect, because the requirement quoted in that finding is scoped by Apple
    to that channel.
  - **Swift, four rules.** `NavigationView`, whose reference page carries
    `deprecatedAt: 27.0` on all seven of its platform rows and names
    `NavigationStack` and `NavigationSplitView` as the replacements;
    `.font(.system(size:))`, on an iOS verdict only, because that form has no
    text style behind it to scale against and, in the HIG's own words under
    Typography › Platform considerations › macOS, "macOS doesn't support
    Dynamic Type"; a colour written as numbers rather than resolved from a
    colorset; and a `Button` whose entire label is one `Image(systemName:)`,
    reported as a label to check with VoiceOver rather than as a label found
    missing.
  - **The platform verdict is a gate, and its silence is printed.**
    `fixed-font-size` runs on an iOS verdict and `sandbox-absent-macos` on a
    macOS one. When the signals are absent or point both ways the verdict is
    `not determined`, both rules stay quiet, and the report says so above the
    findings — `**Platform: not determined**. Signals seen: none. **Every
    platform-scoped rule stayed silent**, so their silence here is the gate
    rather than a result.` A run on a SwiftPM package with no configuration at
    all prints exactly that line and still reports its `NavigationView`, which
    is not platform-scoped.
  - **Directory only, and every other shape is an error result.** There is no
    snippet mode, because configuration is the backbone of this audit and a
    snippet carries none of it. A `code` argument, a missing `path`, a path
    with nothing at it, and a path that is a file each come back with
    `isError: true` and an explanation naming what to pass instead — never as
    an empty audit a caller could read as a clean bill. `code` is declared on
    the input schema for exactly this reason: every structured auditor here
    that accepts a snippet at all spells it `code`, so an agent that has
    learned the family will reach for it, and an undeclared key is stripped
    before the handler can explain itself. `audit_project`, which has no
    snippet mode either, leaves it undeclared and answers a `code` argument
    with `MCP error -32602: … Required at path` — a schema rejection that
    says nothing about why a snippet cannot work here.
  - **Structured output on the same contract the other six auditors speak**
    (v0.23.0): `findings` with rule, severity, message, fix and doc, a
    severity `summary`, a machine-readable `notVisible` array, and a required
    `scan` block. A Swift finding also carries `file` and a 1-based `line`; a
    configuration finding carries no `file` key at all and `line: 0`, because
    a key that is absent from a merged configuration map has no position — the
    markdown prints no line for those, and a consumer reading the wire should
    treat `0` as that same sentinel rather than as line one. `scan` declares
    the same six fields `audit_project` does, and is required for the same
    reason: no snippet mode, so every successful call scanned a directory.
- **Six fixture projects, and 200 tests in `tests/apple.test.ts`.** A clean
  iOS app and a clean macOS app that each report zero findings; an iOS app
  that draws six of the eight rules; a deliberately broken project that draws
  six too, trading that one's colour literal for the Audio Input mismatch; an
  iOS project with no `Info.plist`, where every configuration key arrives out
  of `INFOPLIST_KEY_*` build settings the way Xcode has written them since
  `GENERATE_INFOPLIST_FILE` became the default; and a SwiftPM package whose
  platform cannot be known, which is there to pin the silence rather than a
  finding.

### What this audit cannot see

This is the longest `notVisible` list this server ships — 24 entries, printed
in full under **Not visible to this audit** in every report and carried in
`structuredContent.notVisible`. Each one was written after running the tool on
a directory built to demonstrate it, and each demonstration is kept as a test.
The list is longer than the rule list, and a reader who learns that here will
not misread a clean report:

- **Nothing is measured, and nothing is rendered.** No contrast ratio is
  computed, no tap target sized, no simulator launched, no screenshot taken,
  no build run. `colorset-no-dark-variant` fires on a declaration and settles
  nothing about how that colour reads against anything;
  `hardcoded-color-literal` fires on the numbers without computing a ratio
  from them.
- **Three extensions and three configuration shapes are opened, and the rest
  of a project is outside what was read.** `.swift`, `.plist` and
  `.entitlements`, plus `Info.plist` and `project.pbxproj` matched by name and
  `*.colorset/Contents.json` matched by the end of its path. A screen laid out
  in a storyboard or XIB, a view controller written in Objective-C, a
  localised value in `InfoPlist.xcstrings` and a Swift file saved under
  another suffix were each present in a run here and each drew nothing. An
  asset-catalog member that is not a colorset — an imageset, an appiconset,
  the catalog's own root `Contents.json` — is not opened either.
- **A directory the walk never enters, and your own code with it if you keep
  it under one of those names.** `Pods`, `Carthage` and `DerivedData` are
  added for this audit, on top of the shared skip list every scanner here uses
  — `node_modules`, `vendor`, `build`, `out`, `dist`, `target`, `coverage` and
  more — and every directory whose name begins with a dot. It is a name test
  rather than a provenance test, and the skip is silent in both registers: a
  project whose own `Pods/MyOwnCode.swift` held a `NavigationView` came back
  with one file read and no findings, and one keeping hand-written source in
  `build/` came back with `filesRead: 2`, the `NavigationView` beside it
  reported and the one inside `build/` dropped with nothing in the report
  body marking it. Read the skip as coverage rather than as a clean bill on
  your dependencies — this audit says nothing about them at all — and check
  the path a finding carries, since a library checked in under `Vendor/` is
  read as first-party.
- **A plist value outside the small subset this reader covers.** `<string>`,
  `<true/>`, `<false/>` and an `<array>` of direct `<string>` children are
  read; a `<dict>`, `<integer>`, `<real>`, `<date>` or `<data>` value is
  walked past rather than guessed at. A `UIRequiresFullScreen` nested one
  dictionary down drew no finding in a run here. A binary plist is reported as
  not parsed rather than as empty: a binary `Info.plist` beside a Swift file
  took the platform verdict to `not determined` and every platform-scoped rule
  with it.
- **Everything in a `project.pbxproj` other than `INFOPLIST_KEY_*`.** That
  prefix is what is read out of the file; the rest of the build configuration
  is not parsed. `ENABLE_HARDENED_RUNTIME` is one of the settings outside it,
  which is why no rule here reports on the Hardened Runtime in either
  direction.
- **Which target declared a key, or an entitlement.** Keys from every plist
  and every `project.pbxproj` in the tree are merged into one map, and the
  entitlement identifiers from every `.entitlements` file into one set, because
  none of the four surfaces read here carries a target. A key declared by a
  share extension alone was reported as a fact about "this project's
  configuration"; a sandboxed app beside a hardened helper — each file correct
  alone — drew `microphone-entitlement-mismatch` over a pair that appears in
  neither, and the app's `com.apple.security.app-sandbox` kept
  `sandbox-absent-macos` off the helper. The surfaces line names every file
  that went into both.
- **The written forms the Swift rules do not match.** Each rule is a presence
  test over one shape: `.system(size:)` is matched and
  `UIFont.systemFont(ofSize:)` is not; `Color(red:green:blue:)` is matched and
  a project's own `Color(hex:)` is not; `Image(systemName: "bell")` is matched
  and `Image(systemName: symbol)` is not. One iOS file carrying five such
  lines — those three, plus `Font.custom("Inter", fixedSize: 17)` and a
  `Button` whose label closure is a bare identifier — drew zero findings here,
  so a clean result on a UIKit target, or on a codebase with its own colour
  and symbol helpers, is coverage rather than restraint.
- **A Swift file whose comment masking went wrong, which reads exactly like a
  file with nothing to find.** `maskComments` tracks quotes per character
  rather than tokenizing, so a `/*` or `//` inside an interpolated string can
  be exposed as live code and a well-formed comment further down then blanks
  the real code in between. A file whose second line was
  `let label = "a\(f("open /*")) still"` and whose fifth line held a live
  `NavigationView` produced no findings, and the file counted as read in every
  register — there is no signal for this anywhere in the output. A rule going
  quiet on a file with an interpolated string like that may be meeting this
  gap rather than finding nothing; it is pinned as a known, unfixed gap in
  `tests/scan.test.ts`.
- **Whether a Swift file is missing something.** No rule here claims an
  absence in Swift source, because a modifier on a parent covers its children
  and the code satisfying a requirement may live in another file.
  `symbol-as-only-button-label` fired on a `Button` whose enclosing `VStack`
  carried an `.accessibilityLabel`, which is why its own text calls it a risk
  to check rather than a fault found.
- **Whatever the scan stopped short of.** At most 400 files and 3 MB in total,
  with nothing exempt — configuration is read *before* any Swift file, which
  buys it priority and not an exemption — and any single file over 500 KB is
  skipped whole rather than truncated. `scan.hitFileCap`, `scan.hitByteCap`,
  `scan.skippedLarge` and `scan.unreadable` report each of those, and while a
  cap was hit no absence in `findings` covers the part that was never opened.
  A symlinked directory or file is stepped over and is counted in neither
  `scan.unreadable` nor `scan.skippedLarge`, so that one shows up only in the
  finding it did not produce.

An empty findings list means no rule this audit runs matched the text of the
files that were read — not that the project is sound, and not that it was all
read. The `scan` block says how much of it was, and the platform line says
which of the rules were allowed to run at all.

## [0.24.0] — 2026-08-15

The three existing Apple documents — `macos-app-design`, `ios-app-design`,
`apple-hig-liquid-glass` — carried source URLs that had never been checked
against the Apple page they claimed to summarize. This release re-sources all
three, adds two new Apple documents, deepens two of the existing ones, and
introduces a tiered source allowlist that the six Apple documents (and the
five `security` documents) are now held to. The most useful output of this
work was not the new prose — it was what verification destroyed: 16 claims
that no primary source supported, and several statements that were simply
wrong.

### Added

- **`apple-accessibility`** — Dynamic Type, VoiceOver, hit regions, motion and
  contrast, cited to developer.apple.com and WCAG 2.2.
- **`apple-shipping-readiness`** — the `Info.plist` purpose strings a
  capability requires, the entitlements macOS distribution requires, what a
  complete app icon set is, and what an orientation or multitasking
  restriction costs. Every claim is a statement Apple publishes, cited beside
  it.
- **A three-tier source allowlist — `standard`, `vendor`, `research`.**
  Enforced today on the five `security` documents (`standard` and `vendor`
  only) and the six Apple design-language documents — `macos-app-design`,
  `ios-app-design`, `apple-hig-liquid-glass`, `wwdc-design-principles`,
  `apple-accessibility`, `apple-shipping-readiness` — which additionally
  require at least one `developer.apple.com` source. It is **not** enforced
  on the other documents in the knowledge base; migrating the rest is a
  separate package.

### Changed

- **`macos-app-design` and `ios-app-design` re-sourced and deepened.**
  `macos-app-design` gains a platform-fit section — the SwiftUI/UIKit symbols
  Mac Catalyst supports that macOS does not, and the macOS-only scaffolding an
  iOS-shaped app never wrote — plus a menu bar and keyboard-shortcut table
  rebuilt from Apple's own tables. `ios-app-design` gains orientation
  coverage it previously had nothing on: the size-class table and the
  orientation-lock request that replaced the deprecated opt-out. Both guides
  are smaller where Apple does not back a claim, not larger.
- **`apple-hig-liquid-glass` re-sourced** to the Apple pages that make its
  claims, alongside the other two guides above — three documents re-sourced
  in total.

### Removed

- **16 claims removed because no primary source supported them.** The three
  existing Apple documents' 17 blog-tier sources were each traced to the
  prose they backed and checked against Apple's own page for the same claim;
  where Apple made no such claim, the claim was deleted rather than softened.
  Among them: the SwiftUI "known gaps" list (three of whose five claims were
  not even in the blog originally cited for them), the Tahoe window-tiling
  mechanics, the settings-window metrics, the App Store conversion
  statistics, and five Liquid Glass assertions, including one that named no
  source at all. One dead citation (`hig/navigation-bars`, which now
  redirects to `hig/toolbars`) was also replaced.

### Fixed

- **Factual corrections found only by reading the source, not by suspicion:**
  - Mac Catalyst scales the iPad idiom to **77%**, not 75%.
  - The in-app event card is **16:9, minimum 1920×1080**, not 1080×1920.
  - **macOS does not support Dynamic Type** — the guide had claimed it did
    since Sonoma.
  - Three keyboard shortcuts were wrong: **⌘T is the Fonts window**, **⌥⌘T
    toggles the toolbar**, and **⌥⌘I opens an inspector** — none of them the
    bindings the guide previously named.

## [0.23.0] — 2026-08-14

`audit_seo_geo` and `audit_performance` were the only two tools that answered
an agent as well as a reader — everything else in this server, including the
other four auditors, returned markdown only. A caller chaining `audit → fix`
had to parse prose to find out what an audit covered, and had no
machine-readable way to tell "found nothing" apart from "checked nothing."
This release brings the other four auditors up to the same contract.

### Added

- **`design_lint`, `audit_security`, `audit_generic_design` and
  `audit_project` now declare an `outputSchema` and return
  `structuredContent`** — findings as fields, a severity summary, and a
  machine-readable `notVisible` array of what each audit could not check,
  exactly as `audit_seo_geo` and `audit_performance` already did. All six
  structured auditors now share one schema shape and one governing rule:
  what is disclosed in prose and what is returned in `structuredContent`
  are drawn from the same array, never written twice.
  - Each finding carries its path once, in `file` — never also folded into
    `message`, so an agent rendering `${f.file}:${f.line} — ${f.message}`
    never reads the same path twice. The markdown prints the path exactly
    where it always did. `design_lint` is the one auditor with no `file` at
    all: it takes a snippet and neither a `filename` nor a `path`, so no
    finding it can produce has a path to carry, and its description no
    longer lists the field.
  - `audit_generic_design`'s schema also carries the itemised `score` its
    markdown already printed, plus a `scan` block — present only in
    directory mode — flagging a file or byte cap that leaves a clean score
    unconfirmed.
  - `audit_project`'s schema carries a `scan` block on every call, since it
    has no snippet mode. Both tools' `scan` declares the same six fields:
    how many files and bytes were actually read, which were skipped for size
    or could not be opened, and whether a cap was hit before the whole
    project was seen.
- **Bumped `@modelcontextprotocol/sdk` from 1.29 to 1.30**, the release the
  first two structured auditors were built against; the other four now sit
  on the same version.

### Changed

- **A missing or non-directory `path` on `design_lint`, `audit_security`,
  `audit_generic_design` or `audit_project` now comes back as an error
  result (`isError: true`, no `structuredContent`) instead of an ordinary
  prose result.** This is the only behaviour a caller can observe from this
  release: declaring an `outputSchema` makes a text-only "success" a
  protocol violation, because a caller expecting `structuredContent` gets
  none and nothing tells it the audit never ran. Before this release those
  paths answered with a prose explanation and no error flag, which a caller
  reading only `structuredContent` could not tell apart from a real, empty
  audit. `design_lint` takes no `path` at all and is unaffected.

## [0.22.0] — 2026-08-14

`seo_geo_guide` has returned this server's SEO and GEO guides since v0.9.0,
and nothing ever checked a page against them. A team could read every word
of `technical-seo` and `geo-fundamentals` and still ship three `<h1>`s, a
canonical pointing at staging, a hero image carrying `loading="lazy"`, and no
`llms.txt`. This release adds two auditors that read the page instead of
trusting the reader remembered the document.

### Added

- **`audit_seo_geo` and `audit_performance` — the 32nd and 33rd tools, and
  the first two here to return structured output.** Every other tool in this
  server answers a person reading markdown; these two also declare an
  `outputSchema` and return `structuredContent` — findings as fields, a
  severity summary, and a machine-readable `notVisible` array. An agent
  chaining `audit → fix` needs to know what was never checked as much as
  what was found; prose alone cannot stop a caller from reading silence as a
  clean bill.
  - `audit_seo_geo` reads a missing `<title>` and one outside the width a
    result gives it, the same for the meta description, multiple `<h1>`s,
    skipped heading levels, a missing canonical, one written relative in a
    self-contained document, one left pointing at localhost or staging, an
    hreflang set that never lists the page itself, JSON-LD that does not
    parse or declares no `@context`/`@type` or declares a type whose rich
    result Google has retired, missing alt text, robots.txt crawl rules —
    including the AI crawlers behind ChatGPT, Claude, Perplexity and
    Google's AI surfaces — a robots.txt naming no sitemap, no `llms.txt`
    beside it, and content that exists only once a script has run.
  - `audit_performance` reads a hero image held back by `loading="lazy"` or
    contradicting its own `fetchpriority`, the LCP-candidate image
    declaring no fetch priority at all, a hero background declared in CSS or
    an inline style that the HTML preload scanner never sees, images with no
    `width`/`height` or `aspect-ratio` to reserve their box, a
    `<script src>` in the `<head>` carrying neither `defer` nor `async` nor
    `type="module"`, `@font-face` without `font-display`, third-party font
    hosts, and scripts loaded from more distinct remote domains than any
    reading of "minimise" defends.
  - **Both descriptions are built from the auditors' own capability
    tables**, and the suites assert the mapping is exact in both
    directions. A tool description is read as a statement of reach, so an
    advertised check that no rule performs turns silence into a clean bill
    — the `notVisible` failure moved into the blurb.
  - **The governing rule both modules were written against: these audit what
    is authored, not what is measured.** A finding may state a fact about
    the source and pair it with a documented causal link — "the hero image
    carries `loading="lazy"`, which holds its request back until layout has
    run." No finding is or can be a Core Web Vitals verdict, an indexing
    status or a ranking outcome, because vitals are 75th-percentile field
    data from real devices and this reads source. Telling a team their
    vitals are fine from a static read would stop them measuring, which is
    worse than telling them nothing.
- **Five fixture pages, one per stack — Next.js App Router, Astro,
  SvelteKit, plain static HTML and a built Docusaurus page — each assert
  zero findings from both tools**, written and saved before either module
  was reopened; all ten cells passed on the first run. Two counterexamples
  ship beside them — an image-only splash page and a minimal 404, both real
  pages a visitor actually reaches, both missing the landmark signal a
  naive "is this a real page" guard would have keyed on — plus a
  deliberately broken page where every defect is one somebody has actually
  shipped.

### Fixed

- **A shipped defect in `audit_security`, found while building these two
  modules and fixed on the way.** `hasAttr`/`attrValue` and their
  equivalents let an attribute name match inside another attribute's
  *value* — `alt="Priority support illustration"` read as a declared
  `priority` prop, `title="rel='noopener' explained"` satisfied the check
  for a real `rel=noopener`. Seven defects in total, five of them false
  negatives that silence a real finding rather than invent one: five in
  `src/security.ts` (a `nonce` named inside a `data-` value, `sandbox`
  inside an iframe's `title`, `integrity` inside a `data-note`, `noopener`
  inside a link's `title`, and an inline script misread as the ld+json
  block CSP does not gate), and two in the new modules — one in `seo.ts`
  (a bare `alt` swallowed by `title="alt text here"`), one in `perf.ts`
  (the LCP candidate walking past an image an author had already marked
  with its own `priority`). Both readers now work from a copy with every
  quoted or braced value blanked before a name or value is located. No
  rule id, severity or doc changed; `audit_security` over this repository
  returns the same findings as before, line numbers aside.

### Notes

- **Two specified rules were cut for want of a document that makes their
  claim**, on the same ground `generic.ts` and `security.ts` already
  learned the expensive way: a rule cites a document that exists *and*
  makes the rule's claim, or it does not ship.
  - `og-incomplete` was specified against `on-page-seo`. That document
    never mentions Open Graph, and neither does any other document in
    `knowledge/` — "Open Graph", "og:title" and "og:image" return nothing
    across the whole base.
  - `font-host-not-preconnected` was specified against `technical-seo`.
    Nothing in `knowledge/` recommends preconnecting to a font host —
    "preconnect" appears four times, in two unrelated documents, and never
    in an SEO or performance one. The remedy the knowledge base actually
    documents is the opposite: `seo-for-designers` says "Self-host WOFF2.
    Third-party font CDNs add a connection." That claim ships as
    `third-party-font-host` instead.
- 94 knowledge documents, unchanged by this release — no new document was
  needed, which after two releases that each added one is itself worth a
  line.

## [0.21.0] — 2026-08-13

The knowledge base already said what gives a generated page away.
`typography-craft` carries a reflex-reject font list with Inter on it;
`design-critique-scoring` states the slop test outright. None of it was
enforced. An agent that never opened either document could ship indigo-to-violet
on Inter with a rocket emoji in the hero, and the server would raise no
objection — the knowledge existed to prevent exactly that page and nothing
read it back against what got built.

### Added

- **`audit_generic_design` — the 94th document and the 31st tool.**
  `knowledge/craft/ai-default-aesthetic.md` catalogues the defaults, cited to
  each system's own docs: the stock Tailwind indigo/violet/purple gradient (as
  classes, hex, or OKLCH), Inter/Roboto/Open Sans/DM Sans/Plus Jakarta Sans as
  the only declared typeface, emoji standing in for icons, the `rounded-2xl` +
  `shadow-lg` + border card recipe, gradient-filled heading text, an eyebrow
  over every heading, the `backdrop-blur` + `white/10` glass recipe, stock
  hype-opener copy, stacked filler adverbs, and a page whose every CTA is
  drawn from the stock set. Ten rules run those checks against real source and
  copy, never against taste — a class name, a phrase, a repeated structure,
  each a fact about the file rather than a verdict on the design. The score
  counts distinct signals, not occurrences, so forty cards sharing the stock
  chrome recipe are not more generic than three, and every point is itemised
  to its rule and file:line so a reader can disagree with a line instead of
  the whole verdict. Judgement stays where it already lived — pair this with
  `design_review_checklist` or `get_design_doc("design-critique-scoring")` for
  the half a fact-checker cannot do.
- **Five fixtures score 0.** A brutalist page, a serif editorial layout, a
  dense dashboard set in Inter, a warm consumer screen, and a monochrome
  developer tool each assert a generic score of zero — proof the tool
  penalises the defaults and not the categories of product that happen to
  share a typeface or a grid with them.

### Notes

- **One planned rule was cut before it shipped.** It fired on any three
  elements sharing a class string — nav links, footer buttons, dashboard KPI
  tiles, pricing tiers — and told every one of them their consistent
  components lacked hierarchy. A grid-parent gate would not have saved it,
  because dashboard tiles and pricing tiers genuinely do sit in a grid;
  separating "cards that need hierarchy" from "components that should be
  consistent" is a judgement about what the elements mean, not a fact about
  the source, so it falls outside what this tool can check. Ten rules ship,
  not eleven.
- **Writing the document first corrected the spec's own premises.**
  shadcn/ui's theming docs name no typeface at all — font selection routes to
  the scaffold, not the library — so `default-ui-font` only fires against a
  face actually declared in the source, never against an assumption about
  what shadcn ships. Its stock theme carries zero chroma on every token but
  `--destructive`; any hue in a shadcn project was added by hand. And Tailwind
  v4 authors its palette in OKLCH, publishing hex as "the nearest hex value"
  — the derived form, not the source of truth — which is why the gradient
  rule matches on OKLCH and class names, not on a hardcoded hex table that
  would have gone stale at the next palette revision.

## [0.20.0] — 2026-08-12

88 documents, and not one of them contained the string `Content-Security-Policy`
— nor `OWASP`, `XSS`, `CSRF`, `SameSite`, `HttpOnly`, `HSTS` or `Subresource
Integrity`, anywhere in the knowledge base. A site built end to end from this
server's guidance shipped with whatever header and cookie defaults the
framework happened to pick. This release adds a `security` category to close
that.

### Added

- **Five knowledge documents, a new `security` category — 93 documents:**
  - `web-security-headers` — CSP (nonce/hash, `strict-dynamic`, Trusted
    Types), HSTS, `X-Content-Type-Options`, `Referrer-Policy`, Subresource
    Integrity, and the headers that are dead weight now (`X-Frame-Options`
    superseded by `frame-ancestors`; `X-XSS-Protection` actively harmful in a
    modern browser).
  - `frontend-attack-surface` — XSS sinks, unsandboxed third-party iframes,
    wildcard `postMessage`, mixed content, credentials in `localStorage`,
    secret-shaped `NEXT_PUBLIC_`/`VITE_` env vars.
  - `auth-and-session-ux` — session cookie attributes (`HttpOnly`, `Secure`,
    `SameSite`), CSRF defenses, passkeys, account-recovery flows that don't
    become the weak link.
  - `privacy-consent-and-tracking` — consent-before-load for tracking
    scripts, cookie categories, GDPR/UK-GDPR/KVKK-shaped requirements.
  - `ai-feature-security` — prompt injection from untrusted content and
    insecure output handling in AI features, mapped from the OWASP LLM Top 10
    to a frontend.
  - Every claim is sourced to a spec or a first-party vendor doc — a security
    document citing a blog is now a test failure, not a style note — and
    re-verified on a 90-day clock, the tightest staleness threshold in the
    table: a reader who believes a stale security claim thinks they're
    covered when they're not.
- **`audit_security`** — audits a directory or a pasted snippet for the
  defects above: missing or weak CSP, absent HSTS, unpinned cross-origin
  scripts, mixed content, `localStorage` credentials, secret-named public env
  vars, unsandboxed iframes, wildcard `postMessage`, unsanitised raw-HTML
  sinks, production source maps, un-ignored `.env` files. Header and CSP
  state is inferred by reading `next.config` / `vercel.json` / `netlify.toml`
  / `_headers` / middleware as text — never evaluated — so it also reports
  what it could not see rather than guessing. **30 tools** in total.
- The five documents are wired into every web-facing `design_review_checklist`
  (`website`, `landing-page`, `dashboard`) and roadmap (`website`,
  `landing-page`, `saas-web-app`) — a document nothing references is a
  document nobody reads. `tests/integrity.test.ts` now asserts the wiring
  directly, rather than only that each document is referenced from somewhere.

### Notes

Writing these turned up several widely repeated claims that don't hold up
against the spec or the vendor's own current docs:

- Trusted Types is no longer Chromium-only.
- `strict-origin-when-cross-origin` has been the browser default
  `Referrer-Policy` since 2020 — its absence from a response is not a
  finding, and an auditor that flags it anyway is simply wrong.
- Browsers imply `noopener` on `target="_blank"` anchors automatically;
  `window.open()` still does not, and still needs it spelled out.
- MDN states `SameSite=Lax` is the browser default; caniuse measures actual
  support at 76.34%, and only in Chrome and Edge. The two numbers describe
  different things, and neither one is wrong.

## [0.19.1] — 2026-08-06

### Fixed

- **An unsupported Node produced a syntax error, not an explanation.** `dist/index.js`
  uses top-level await, which older runtimes cannot parse, so the failure was
  `SyntaxError: Unexpected reserved word` pointing at a file the user did not
  write — and no check inside that file could ever run, because the parse fails
  before the first statement. `bin` now points at a small launcher written in
  deliberately old syntax, which names the required version and shows how to aim
  a client at a newer Node. Node 18 is not blocked: it works end to end in
  testing, so it starts with a warning rather than an error. A guard added to
  improve a message must not become a new source of breakage.

### Changed

- **The startup line names its transport.** It said "server running", which reads
  as "serving" to anyone who deployed this somewhere — the process would then
  wait on standard input forever with nothing to explain why it was unreachable.
  It now says "ready on stdio", and adds an explicit note when started by hand
  from a terminal, where the wait looks exactly like a hang. The README states
  the same thing: stdio only, no HTTP or SSE, cannot be hosted remotely — which
  is also why nothing leaves the machine.

## [0.19.0] — 2026-08-03

### Added

- **Status colours.** `generate_color_system` produced fourteen roles and none
  of them was an error colour, while `create_design_system` called itself a
  complete foundation. Adds `danger`, `success` and `warning` as full scales
  plus verified semantic roles in both themes, seeded from the conventional
  hues at the brand's own saturation so a muted brand does not get a
  fluorescent red.
- **Four knowledge documents**, all from primary sources, all on subjects the
  base had never covered — 88 documents:
  - `modern-css-design-primitives` — `contrast-color()`, `light-dark()`,
    `color-mix()`, `@scope`, container style queries, `field-sizing`, anchor
    positioning and the new font-relative units. Written around the limits:
    `contrast-color()` returns only white or black, and MDN's own example is a
    royal blue where it returns unreadable black.
  - `brand-on-native-platforms` — expressing brand inside a platform's
    conventions, built on the UI-layer / content-layer split.
  - `search-design` — `compare_design_languages` had listed "search" as a topic
    since v0.15.0 with nothing behind it.
  - `naming-features-and-labels` — criteria, process and evaluation for naming.

### Changed

- **`get_component_recipe` swaps whole ramps, not roles.** Role mapping handled
  `indigo-600` because that is "the primary" and left `indigo-300/400/500/800`
  behind, because no role names the shade a dark theme uses — so every dark UI
  built from these recipes kept our accent. Neutrals failed the same question
  from the other side: `bg-neutral-900` is a surface and `text-neutral-900` is
  text. Pass the `neutral`, `primary` and `danger` scales and each is swapped
  step for step, which needs no inference and therefore cannot be wrong. Nine
  components across both web stacks now return with no house colour left.
- `create_design_system` prints the exact payload to hand to
  `get_component_recipe`, and a test asserts every role in it is one the recipe
  tool accepts, so the two cannot drift apart.

### Fixed

- **Focus rings were counted as elevation.** A ring is a box-shadow but not
  depth, and counting them together punished exactly the codebases that do this
  properly. A ring has no offset and no blur, which separates the two precisely.
- **`design_lint` called best-practice code an error.**
  `:focus:not(:focus-visible) { outline: none }` is the recommended way to drop
  the ring for pointer focus while keeping it for the keyboard. Found by running
  `audit_project` against a real site.
- **The recipe library was not one system.** Four components used indigo as the
  accent and four used blue. Audited at 54/100 by our own tool, now 90, with a
  test that keeps it there.
- **Inflection decided findability.** Title and tag matching is exact-token, so
  "token" scored nothing against a document titled "Design Tokens". Query and
  index tokens are now stemmed — plurals and `-ing` only; `-ed` is left alone
  because "embed" would become "emb", and a stemmer that invents matches is
  worse than one that misses them.

### Notes

The first status-colour implementation passed every test while producing
`#46100b` for danger — a near-black brown that white reads on perfectly and
nobody reads as an error. Contrast passed, hue passed, saturation passed; the
constraint that bites is lightness. Measuring is not enough on its own — it has
to be the right measurement.

Apple's Human Interface Guidelines pages render client-side and return only
their titles to a fetch, so nothing was written from memory about the June 2026
HIG revisions. SF Symbols 8, Icon Composer 2, Pass Designer, scroll edge effects
and app schemas remain uncovered for that reason rather than by choice.

## [0.18.0] — 2026-08-02

Three gaps, each found by measuring rather than guessing: the pipeline broke in
the middle, teams could not add their own rules, and the auditors could not see
a project.

### Added

- **Your own knowledge directory.** `SAGLITZDESIGN_KNOWLEDGE_DIR` points at one
  or more directories whose documents join the base. The README used to say
  "drop a file under `knowledge/`", which means editing the installed package —
  `npm update` wipes it, so extending was effectively impossible from npm.
  A document with the same id replaces the built-in one, announced at startup
  and marked as your team's wherever it is served; `review: [website]` in
  frontmatter puts it into that checklist, ahead of the curated list.
- **`audit_project`** — the auditors over a directory instead of a pasted
  snippet, with the consistency score computed across files, findings ranked
  worst-file-first, and an explicit list of what was skipped.
- **`get_component_recipe` takes your tokens** and returns the code in your
  colours. Substitution happens on the way out, never on disk, so the recipe
  files stay valid runnable code; without tokens the output is byte-identical
  to before.
- **29 tools** in total.

### Fixed

- **Focus rings were counted as elevation.** A ring is a box-shadow but not
  depth, and counting them together punished exactly the codebases that do this
  properly — a consistent ramp plus a few ring states read as "sprawl". A ring
  has no offset and no blur, which separates the two precisely.
- **`design_lint` called best-practice code an error.**
  `:focus:not(:focus-visible) { outline: none }` is the recommended way to drop
  the ring for pointer focus while keeping it for the keyboard. Found by running
  `audit_project` against a real site.
- **The recipe library was not one system.** Four components used indigo as the
  accent and four used blue, so a UI built from these recipes put an indigo
  button beside a blue tab; radii had two values for "control" and two for
  "container". Audited at 54/100 by our own tool, now 90, with a test that keeps
  it there — a project shipping `audit_design_system` cannot ship a library that
  fails it.

### Notes

Ranking a team's documents took three attempts. A flat score boost was
arbitrary; gating on a fraction of the best built-in score failed outright,
because the scoring is length-biased — body frequency rewards long documents, so
a twenty-line house-rules file can never out-score a two-hundred-line reference.
Term coverage is length-independent: a document containing everything you asked
about leads, one sharing a single word does not.

## [0.17.0] — 2026-07-27

### Added

- **`import_design_tokens`** — the inverse `generate_design_tokens` never had.
  Reads CSS custom properties (a Tailwind v4 `@theme` block, a shadcn `:root`
  block, plain CSS), a W3C DTCG token file, or a theme object as JSON, and
  returns the roles it names, the semantic roles it leaves undefined, a WCAG
  contrast check on the pairs it defines, and the whole set re-emitted as CSS /
  Tailwind / SwiftUI / Compose / DTCG. Until now the server only served
  greenfield projects; this one meets a codebase that already has a system.
- **`theming-off-the-shelf`** knowledge doc — how to theme shadcn/ui, Radix,
  Material and native kits with your own tokens instead of rebuilding them.
  `design-systems-methodology` had told people to adopt and theme rather than
  reinvent without ever saying how. 84 documents.

### Changed

- Build workflows now import an existing theme before generating a new one, and
  `port_to_platform` re-emits the tokens a project already has for the target
  platform rather than starting over.
- **28 tools** in total.

### Deliberately not added

shadcn/ui component recipes. `design-systems-methodology` lists "reinventing an
off-the-shelf system" among its anti-patterns and names shadcn/ui specifically;
shipping our own competing button would contradict our own published guidance,
and it would be worse maintained than the one `npx shadcn add button` gives you.
The useful thing for those users is knowing how to theme what they already have,
which is the knowledge doc above.

### Notes

Only *named* tokens are read — CSS custom properties, DTCG entries and theme
keys carry a role; a bare `color: #4f46e5` inside a rule does not, and is never
imported as one. JavaScript configs are never evaluated. Text roles are
classified before surfaces: shadcn names its text roles `muted-foreground` and
`primary-foreground`, and matching surfaces first filed the most common failing
text colour in the ecosystem as a background.

## [0.16.0] — 2026-07-25

### Added

- **`measure_screenshot`** — measures a real screenshot from its pixels rather
  than describing it: the exact palette and how many distinct colours a screen
  actually uses, true WCAG contrast ratios for the colour pairs present,
  whitespace and density, and structural detections (left-edge alignment,
  vertical rhythm, off-grid gaps) each carrying a confidence level. Findings
  describe the image, never the interface's semantics, and anything below its
  confidence threshold is not reported at all.
- **Self-contained HTML report** — the measurement renders to a single
  standalone document with no external requests, readable in light and dark,
  that you can open, keep and share.
- **Pure-Node PNG decoding** (`src/png.ts`) built on `node:zlib` — truecolour,
  greyscale, palette with tRNS, 8- and 16-bit, all five scanline filters. No new
  dependencies. Unsupported input (JPEG, interlaced, truncated) is refused with
  a named reason instead of producing wrong pixels.

### Changed

- `critique_screenshot` and `design_review` now measure a screenshot before
  judging it, so a critique cites "2.9:1, AA needs 4.5" instead of "contrast
  looks weak".
- Colour distance and clustering moved to `src/colorutil.ts`, shared by the
  screenshot measurement and `audit_design_system` — "23 colours in your CSS"
  and "23 colours on your screen" are now counted by the same rule.
- **27 tools** in total.

### Testing

- Fixtures are synthesised in-test by a PNG *encoder*, so every assertion
  checks an exactly known answer rather than an approximation. Two
  false-positive guards protect the positioning: a perfectly aligned layout must
  report a single edge, and an anti-aliased edge — which produces two adjacent
  peaks above threshold — must merge into one.

## [0.15.0] — 2026-07-24

An audit-and-repair release: three bugs were silently degrading the flagship
orchestration tools, and the test suite could not have caught any of them.
Also adds MCP resources, argument completion, and three new tools.

### Fixed

- **11 pattern documents were invisible to every roadmap and checklist.**
  `ROADMAPS` referenced pattern docs by their bare name (`onboarding-paywall`)
  while the documents carry platform-prefixed ids (`mobile-onboarding-paywall`).
  Unknown ids were silently filtered out, so — for one example — the iOS
  roadmap's "Monetization & key flows" phase omitted the paywall, auth,
  checkout and settings patterns it exists to point at. Ids are now canonical,
  and `findDoc()` resolves either form so cross-links inside the knowledge base
  (`[[onboarding-paywall]]`) keep working.
- **`cross-platform` documents disappeared from platform-filtered searches.**
  The filter only exempted `both`, so `design-tokens-theming` and `fluent-2`
  vanished from any search scoped to `web`, `mobile` or `macos`.
- **`design_lint` misjudged formatted markup.** The rules ran line by line
  while JSX is routinely wrapped across lines, so a multi-line `<img>` *with*
  `alt` was reported as an error, and a single-line `:focus { outline: none }`
  — the most dangerous case — was skipped. Markup rules now use a tag scanner
  over the whole snippet, and `outline: none` is judged against whether the
  snippet provides a focus replacement anywhere. Attributes arriving via
  `{...spread}` are no longer guessed at.
- **5 broken `[[wiki-links]]`** in the knowledge base now point at real docs.
- **`get_design_examples` no longer over-promises.** Screenshots are a
  local-only asset (third-party images are not redistributed), so published
  installs never returned them despite the description saying otherwise. The
  tool now detects which mode it is in and describes itself accordingly.

### Added

- **MCP resources.** The knowledge base is browsable without spending a tool
  call: `saglitzdesign://index`, `saglitzdesign://doc/{id}` (all 83 docs) and
  `saglitzdesign://recipe/{component}`, with **argument completion** for ids
  and component names.
- **`audit_design_system`** — point it at CSS/JSX/token source and it reports
  design-system sprawl: near-duplicate colors, radius/shadow/font-size/spacing
  scales, hardcoded values vs tokens, with a consolidation plan.
- **`generate_layout_system`** — breakpoints, container widths, a fluid grid,
  container queries and section rhythm as CSS custom properties and a
  Tailwind v4 `@theme` block.
- **`compare_design_languages`** — side-by-side iOS/HIG, Material 3, macOS and
  web equivalents for one surface (navigation, buttons, modals, motion,
  typography, elevation…), with the porting rule for each.
- **26 tools** in total, up from 23.
- **`port_to_platform` workflow** — takes a UI to another platform surface by
  surface, driven by `compare_design_languages` and its "do NOT port" lists.
- **`design-system-audit` skill** for the `npx skills add` distribution.

### Changed

- **The `/` workflows now drive the whole server.** They had not been updated
  since v0.4.0 and orchestrated only 8 of 26 tools — no generators, no recipes,
  no auditors. Every build workflow now generates the design system before
  writing pixels, builds from the component recipes, and passes a
  **deterministic verify gate** (`design_lint` → `audit_accessibility` →
  `audit_design_system` → `audit_ux_copy`) before it may claim to be done.
  `design_review` and `redesign` now lead with measured numbers instead of
  opinions, and report a real before→after.
- `create_design_system` gained the layout layer it was missing (grid,
  breakpoints, measure and section rhythm for web; margins, touch targets and
  adaptivity for iOS/Android) and now closes the loop by pointing at its own
  auditors.
- The `skills/` distribution was refreshed — it advertised a 68-document,
  12-tool server and referenced almost none of the tooling.
- 12 previously unreferenced documents (AI product UX, i18n, information
  architecture, emotional design, branding, email, ad creative, content
  distribution, and three distilled classics) are now wired into the roadmaps
  and review checklists that should have been citing them.
- Catalogue data (categories, review maps, roadmaps, freshness thresholds)
  moved to `src/catalog.ts` so it can be validated without starting a server.
- The server version is read from `package.json` at runtime; `npm version`
  syncs `server.json` automatically via `scripts/sync-version.mjs`.
- Minimum Node is now 20 (18 is end-of-life); CI runs 20/22/24.

### Testing

- **76 → 190+ tests.** New `tests/prompts.test.ts` validates the workflow prose
  itself — no phantom tool names, no broken document ids, and every build
  workflow gated on the auditors; `tests/integrity.test.ts` does the same for
  the `skills/` distribution. New `tests/integrity.test.ts` asserts that every doc id
  referenced by a roadmap, checklist, enum or wiki-link resolves, that ids are
  unique, and that release metadata is in sync. New `tests/server.test.ts`
  drives the real stdio server: every tool is listed, annotated and answers a
  representative call; resources and completions are exercised end to end.

## [0.14.0] — 2026-07-23

- Registered all 23 tools through a wrapper adding human titles and MCP
  annotations (`readOnlyHint`, `idempotentHint`, `openWorldHint: false`).
- Sharpened the knowledge-tool descriptions and parameter documentation.

## [0.13.0] — 2026-07-23

- Flagship `create_design_system`: one call turns a brand color + vibe +
  platform into a complete foundation.
- Added `generate_type_scale`, `generate_elevation_system`, `generate_motion`,
  `design_lint`, `audit_ux_copy`.
- Five new knowledge docs (e-commerce checkout, fintech trust, visionOS,
  HTML email development, design handoff) — 83 documents total.

## [0.12.0] — 2026-07-14

- Added `apple-intelligence-design`: how to design AI features the Apple way.

## [0.11.0] — 2026-07-14

- Added `suggest_icon_library` and the `iconography` craft doc. No icon assets
  are bundled.

## [0.10.0] — 2026-07-13

- Added `generate_color_system`, `suggest_font_pairing`, `fix_contrast`.
- First test suite (vitest) and CI across Node 18/20/22.
- Four new knowledge docs; MCP Registry and Smithery manifests.

## [0.9.0] — 2026-07-10

- Added `get_component_recipe` with production-ready code for 9 components
  across react-tailwind, html-css, SwiftUI and Compose.

## [0.8.0] — 2026-07-09

- Added the `skills/` distribution (5 skills) and 4 knowledge docs.

## [0.7.0] and earlier

- Knowledge base, search, roadmaps, review checklists, design tokens,
  accessibility auditing, prompts, and the curated example library.
  See the repository history for details.