synthesisui 0.16.235 → 0.16.238

package/dist/skill-import.js

@@ -1,1316 +1,15 @@
 /**
- * THE IMPORT SKILL, as the CLI ships it.
+ * O STUB DA SKILL - o corpo é SERVIDO, não embarcado (R1, dono 16/08).
  *
- * It lives here rather than as a file copied at build time for one reason: the
- * build is `tsc` and nothing else, so a `.md` would need a copy step that can
- * silently not run. A TypeScript module cannot fail to be packaged.
+ * O playbook inteiro morava aqui (65 KB) e, por consequência, no tarball
+ * público do npm e no repo de cada cliente em texto plano. Ele mudou para a
+ * plataforma (apps/web/src/lib/ds/playbooks.ts, gêmeo por spec do
+ * `.claude/skills/sui-import-ds/SKILL.md`) e é servido pelo catalogue autenticado -
+ * a mesma doutrina de sempre: SERVED, NOT SHIPPED. O agente busca o sumário
+ * e depois SÓ o capítulo do passo em que está (a dieta de contexto).
  *
- * THIS IS THE SOURCE. The copy in our own `.claude/skills/` is what our editor
- * reads, and `skill-import.spec.ts` asserts the two are identical - so drift
- * turns a test red instead of shipping a skill that describes a pipeline the
- * CLI no longer has.
+ * O frontmatter fica INTEIRO no stub: é a description que faz a skill ser
+ * invocada, e ela não é segredo - a esteira é.
  */
-export const IMPORT_SKILL = `---
-name: sui-import-ds
-description: Turn a codebase the user ALREADY has into a SynthesisUI design system. Use when someone points at an existing repo, app or component library and asks to import it, adopt it, bring it in, or "make a design system from this" (e.g. "/sui-import-ds", "importa o meu packages/ui", "turn this app into a design system"). Drives the full pipeline - census → your reading → import → v2 proposal - and answers the questions arithmetic cannot.
----
-
-# Import Design System
-
-Read a project the user already built and give it back to them as a system that governs it.
-
-The census is **arithmetic** and it is not yours to redo. Every colour ranked, every ΔE
-measured, every component crosswalked - the CLI does all of it, deterministically, for free,
-and it will be right. Your job is the part with no closed form: **understanding how this
-project is structured.**
-
-## The golden rule
-
-**You may judge. You may never invent.**
-
-Every hex you name has to be one the census already observed - the endpoint drops the rest
-silently, and a "faithful" import carrying a colour the author never wrote is the one failure
-this whole pipeline exists to prevent. The same discipline applies to everything else you
-report: if you did not read it in their files, do not claim it.
-
-When you are unsure, say so in \`concept\` and leave the field out. An absent answer falls back
-to arithmetic, which is honest. A confident wrong answer inverts someone's product.
-
-## The pipeline
-
-### 0. Check the login FIRST, and stop if there is none
-
-\`\`\`
-cat ~/.synthesisui/credentials.json 2>/dev/null || echo "NO TOKEN"
-\`\`\`
-
-No token, or a run that later comes back **HTTP 401**, means there is nothing to send to. **Stop
-there.** Do not measure, do not read the project, do not write a reading - say this and wait:
-
-\`\`\`
-You need to authenticate first:
-
-    npx synthesisui login
-\`\`\`
-
-It is a device flow: it opens a browser and needs a person. You cannot complete it, and there
-is no point doing twenty minutes of reading against a session that will refuse the payload at
-the end (dono, 31/07 - a full read ended on a 401 that a first command would have caught).
-
-**Read the \`registry\` field in that file, and pass it to EVERY later command.** A token belongs
-to the host that issued it. One minted by \`login --registry http://localhost:3000\` and then
-sent to the default (production) earns a 401 that reads as an expired session - and a session
-cannot expire on a host that never issued it. That is how a full read ended in a dead end
-twice (dono, 31/07).
-
-So if the file says \`"registry": "http://localhost:3000"\`, every command in this skill carries
-\`--registry http://localhost:3000\`. If it says production, pass nothing. **Read that file ONCE**
-and carry the answer for the whole run - re-reading it before each command is a round trip that
-answers a question already answered. The CLI now refuses a
-mismatch instead of letting the server reject it, but the point is not to reach that.
-
-**Do this silently.** Which host you are pointed at and which flags follow from it is plumbing,
-and narrating it - *"registry is production, so no --registry flag needed"* - spends the
-person's attention on a decision that was never theirs (dono, 01/08). Say something only when
-there is nothing to say to: no token, or the wrong host.
-
-### 0b. ASK THE REGISTRY WHAT EXISTS - never your notes
-
-If your memory says this repo already produced a system, **verify it before you say it out loud**,
-and verify it per SLUG:
-
-\`\`\`
-curl -s -o /dev/null -w "%{http_code}" <registry>/api/registry/ds/<slug>
-\`\`\`
-
-\`404\` means it does not exist. \`200\` means it does. That is the whole check, and it costs one call.
-
-**\`synthesisui list\` does NOT answer this.** It returns the public gallery - 13 systems, none of
-them private - so a private system of theirs is invisible to it and a \`list\` that does not mention
-\`their-slug\` proves nothing.
-
-This exists because of a real run (dono, 05/08). The owner had deleted every system on his account,
-and the skill opened with *"my notes from the previous twelve runs say this repo has already
-produced two published systems"* - then offered three options built on that: upgrade a slug the
-registry answers \`404\` for, avoid a name collision that no longer existed, and warn that a \`--dry\`
-would wipe files it does not touch. One \`curl\` would have replaced all of it.
-
-**A memory may report the past; it may never assert the present.** *"The last run produced
-\`signalui\`"* is a fine sentence. *"This repo has two published systems"* is a claim about the server,
-and the server is the only thing that can make it.
-
-## How to narrate
-
-The person cares about **what you found in their code** and **what you are about to create**.
-They do not care which flag you passed, which grep you ran, or what you checked and ruled out.
-
-- Report findings, not mechanics. *"Their docs call Ocean the primary branding scale"* earns
-  its line. *"Let me check whether packages/ui exists"* does not.
-- Every decision you make on their behalf belongs in **one** confirmation at the end, not
-  scattered through the work as fait accompli.
-- Never end on a wall of prose that happens to contain a question. If you need an answer, the
-  last thing on screen is the question.
-
-### 1. Measure
-
-**Always run from the project root, and always twice in a monorepo.** The root is where the
-governance lives - \`_synthesisui/config.json\`, \`ds/<slug>/\`, and the census. One root, one
-census; re-measuring overwrites it rather than leaving copies to send the wrong one.
-
-First, unscoped, to find out what kind of repo this is:
-
-\`\`\`
-npx synthesisui import --dry
-\`\`\`
-
-If it is a single app, that census is the one. If it holds several projects, the CLI says so
-and **names the candidates itself**, ranked by how many tokens each declares:
-
-\`\`\`
-This root holds several projects
-  3 apps (apps/web-admin, apps/web-dashboard, apps/web-review) - and this census
-  is the average of all of them. That is a fine DIAGNOSIS and a poor system: a
-  light app and a dark one average into a palette that is neither.
-
-      synthesisui import --scope packages/ui --usage apps/web-dashboard
-\`\`\`
-
-Then measure again, narrowed - and **name both roles**:
-
-\`\`\`
-npx synthesisui import --dry --scope packages/ui --usage apps/web-dashboard
-\`\`\`
-
-### The two roles, and why one flag could not do both
-
-\`\`\`
---scope   the SYSTEM     tokens, components, the way they name a class    ONE folder
---usage   the EVIDENCE   how often, which values get picked, the laws     as many as they have
-\`\`\`
-
-\`--usage\` is repeatable and never contributes a token, a scale or a component. That split is
-not tidiness, it is the two mistakes this pipeline exists to avoid, and each one has been made:
-
-**Do not import the average.** Three apps measured together produce a palette that belongs to
-none of them, and it will look plausible - a real monorepo reported 36% coverage across 2797
-files at the root and 47% across 114 in its shared package. The second number is the system;
-the first is a diagnosis. Their scale is a DECLARATION; an app's average is a set of choices
-made from one.
-
-**Do not import a library alone either.** A law needs three files that agree, and a library has
-ONE file per component - so measured by itself, a real one came back with \`usage\` empty on all
-23 components and not a single observed law. Not because nothing is settled: because the
-agreeing happens in the app (dono, 01/08). If they have apps, name them.
-
-Point \`--usage\` at every app that consumes the system. A component that lives in an app and
-not in the system is reported and left out - the system is one place, and a union of two
-codebases is not a system.
-
-The census still lands at \`<root>/_synthesisui/census.json\` and records both what it measured
-and where the evidence came from, so you always send the same path and there is never a second
-file to pick between. Nothing is sent by \`--dry\`. Read the file.
-
-### 2. Walk them through the decisions - ALL of them, before any reading
-
-**WHEN \`sui-init\` SENT YOU HERE, THESE ARE ALREADY ANSWERED. Ask nothing.**
-
-The first-run skill collects source, name, primary, theme, canvas, architecture and \`standards\`
-before it hands over, precisely so this step is silent. Asking again is not thoroughness - it is the
-same six questions twice in one conversation, which reads as a tool that forgot what it was told, and
-it is the single fastest way to make a first run feel long. Take the answers, skip to step 3, and
-confirm nothing.
-
-If any of them is missing from the handover, ask ONLY that one.
-
-**Every question this skill asks is answerable from the census you just took.** Source, name,
-primary, theme, canvas, architecture - the numbers behind all six are in the file on their disk,
-and none of them needs the anatomy.
-
-They used to be asked AFTER the reading, and that put six hundred lines of work between the
-person's first answer and their second (dono, 01/08). A person who answers six questions in
-ninety seconds and then watches a bar has had a good four minutes; the same person interrupted
-twice in the middle of a silence has had a bad twenty.
-
-So: ask everything here, in one pass, and **do not come back**. If something you learn during
-the reading contradicts an answer, you do not reopen the question - you say so in the report at
-the end and let them change it there.
-
-Six decisions are theirs. **Ask them as separate questions with selectable options** - use your
-question tool, one call per decision, so they pick instead of reading a wall and composing a
-reply. A single block containing everything is a report, and a report gets read, not answered.
-
-Every question carries **your recommendation first, marked as such**, and the evidence for it in
-one line. They are choosing, not auditing you.
-
-**2a. Source** - the folder you measure the system out of.
-
-Options are the candidates the CLI already ranked, each with what it would produce, plus the
-whole repo, plus *somewhere else* for a path they type:
-
-\`\`\`
-Where should the system come from?
-
-  packages/ui (recommended)    114 files · 47% of its values already named ·
-                               92 tokens of their own · every app imports it
-  the whole repo               2797 files · 36% · the average of one dark app
-                               and two light ones, which is a diagnosis and
-                               not a system
-  apps/web-dashboard           the heaviest consumer, if you want one app's
-                               vocabulary rather than the shared one
-  somewhere else               a path you name
-\`\`\`
-
-If they name a path you cannot use, say which of the two it is and offer the list again:
-
-- **it does not exist**, or holds no files you can read
-- **it holds no design values** - no colours, no radii, no tokens. A folder of server code or
-  config is not a design system, and measuring it produces a system of nothing.
-
-**2b. Name** - suggest theirs, and let Enter take it:
-
-\`\`\`
-What should it be called?  (SignalUI)
-\`\`\`
-
-Say where the suggestion came from - \`Introduction.mdx\`, the package name, the folder. A name
-they recognise is worth more than a clever one. The slug is derived once and never changes, so
-this is the moment.
-
-**2c. Primary** - and explain the role before offering the choice, because *primary* is our word
-and not necessarily theirs:
-
-\`\`\`
-Primary is the colour that carries action - buttons, links, focus, the thing
-you want pressed. Everything else in the system is measured for contrast
-against it.
-
-  #059aed  ocean-500 (recommended)     your ColorPalette.mdx calls Ocean the
-                                       "Primary branding scale"
-  #1a4ed8  royal-blue-500              your docs label this one secondary
-  #ec4899  vivid-pink-500              half of your signature AI gradient
-  #4A90E2  blue-500                    the most-painted blue in the folder,
-                                       but it sits after your own
-                                       /* End - Signal UI */ marker
-\`\`\`
-
-Always name the hex AND the token, and always give the reason a value is *not* recommended when
-frequency would have picked it. That last row is the whole point of asking.
-
-**2d. Default theme** - which face it opens in:
-
-\`\`\`
-  dark (recommended)   data-theme="dark" in the root layout, next-themes
-                       defaultTheme="dark", enableSystem={false}
-  light                the other one, also real here
-\`\`\`
-
-**Two is all there is** - a SynthesisUI system carries one palette and one alternate, and the
-schema is \`"light" | "dark"\`. So this is not "which themes do you support", it is "which one do
-docs and installs open in". If both are real, both get built from their own tokens; this picks
-the face.
-
-**2e. Canvas** - confirm, do not offer a list. The page is measured, not chosen:
-
-\`\`\`
-canvas  #222326  darkgray-500 - what DashboardLayout paints on <main>
-                 (your sidebar sits a step deeper at darkgray-700)
-\`\`\`
-
-Say it and let them object. Offering five near-blacks to pick between is asking someone to
-re-decide something their code already decided.
-
-**2f. Our kit** - do not ask. A project that already has a component library gets **theirs
-alone**, and you say so in one line:
-
-\`\`\`
-This system is yours alone - the 37 components you wrote, and none of ours.
-Add any of ours later from the library, one at a time.
-\`\`\`
-
-Send it as \`"standards": "none"\` in the reading. Every time, with one exception below.
-
-**Why this stopped being a question.** Somebody who points us at their own library is telling us
-what their system is, and 45 components of ours landing beside 37 of theirs makes a system of 82
-where only 37 are governed by anything they wrote - the number on their own screen stops being
-about them. The kit is one click away in the library afterwards, so choosing wrong here costs a
-click and choosing to ask costs a pause in the middle of their import.
-
-**The exception, and it is not a preference.** An app with NO components of its own - \`--scope\`
-pointed at an app rather than a library, and the CLI reported no component library - must NOT get
-\`"standards": "none"\`: theirs alone plus nothing of theirs is a system with zero components. Leave
-the field out there, exactly as before.
-
-Two things to be straight about, because both are true and neither is obvious:
-
-- **the foundations still come from the seed.** Their palette, their scales, their type - all
-  theirs - but the seven type slots, the easing and the neutral floor arrive from ours, because
-  the alternative to a type scale is no components at all rather than their components.
-- **our layouts are written in our recipes.** A landing that composes \`button\`, \`card\` and
-  \`badge\` needs those to exist. Theirs alone keeps the layouts whose recipes they happen to have
-  and drops the rest - a layout pointing at a recipe nobody has renders as holes. Say that number
-  when the CLI reports it, so a shorter list of layouts is a fact they were told rather than one
-  they discover.
-
-**2g. Which architecture has priority** - ask only when the CLI reported more than
-one, which it does under "How this project is organised".
-
-A project having two shapes is not a project with a mistake: a library organised
-atomically inside a monorepo whose apps are organised by feature is two true answers.
-Say that before you ask, or the question reads as an accusation.
-
-\`\`\`
-Two shapes, and neither is wrong:
-
-  atomic under packages/ui (recommended)   114 files - atoms/molecules/organisms,
-                                           and this is where the system comes from
-  feature under apps/web-dashboard         900 files - features/, where the app
-                                           consumes it
-\`\`\`
-
-The one they pick becomes a rule that reaches \`CLAUDE.md\`, so **every prompt after this
-knows where a new component goes** - not what the folders are called, but which rung or
-which feature a new file belongs to. Recommend the one the system came from.
-
-**2h. Fonts** - do not ask at all. Report and move on:
-
-\`\`\`
-fonts   Figtree for both display and body - the only family declared
-        (--font-sans, and next/font Figtree in all three root layouts)
-\`\`\`
-
-If a project genuinely declares two families, say which is which and why you paired them that
-way. If it declares one, there is nothing to decide.
-
-**A note on colour swatches.** You cannot paint a hex in this conversation, so the token name
-carries the weight - \`ocean-500\` tells someone more than a square would. When they run
-\`npx synthesisui import\` themselves in a terminal, the CLI paints them.
-
-### 3. Read what arithmetic cannot - and do it in ONE uninterrupted run
-
-Every answer is in hand, so nothing below needs the person. **Do not ask, do not confirm, do
-not narrate the mechanics.** Three lines on screen for the whole phase:
-
-\`\`\`
-1/3  Measuring the repository...            (step 1, already done)
-2/3  Naming the anatomy and validating...   (this step)
-3/3  Sending v1 and v2...                   (step 4)
-\`\`\`
-
-**Narrate findings, never mechanics.** A grep, a glob, a file you opened and ruled out, a draft
-you rewrote - none of those is a finding, and a run that prints them reads as a machine talking
-to itself. *"Their docs call Ocean the primary branding scale"* is a finding and earns its line.
-
-**The census answers first, always.** Before opening any file, check whether \`signals\`, \`sketch\`,
-\`declared\` or \`coverage\` already hold it - they usually do, and a file read whose content the
-census carries is the failure this pipeline keeps paying for. The sketch now runs to 150 nodes,
-so a big component arrives WHOLE: \`TextEditor\` used to lose its editable region at node 60 and
-that single truncation is what forced a hand-read of the file (dono, 01/08).
-
-\`signals.theme\` includes the SWITCH - \`attribute\`, \`defaultTheme\`, \`enableSystem\` and what
-\`<html>\` carries, \`colorScheme\` included - read from the \`--usage\` roots as well as the scope,
-because a component library declares the dark utilities and the APP declares the switch. When it
-came from an app, \`switchedIn\` names which one: say "the system is measured in \`packages/ui\`, the
-theme is switched in \`apps/web-dashboard\`" rather than reporting that the library declares a
-provider it does not have. Answering 2d never needs a layout file opened.
-
-**Validate in ONE call, at the end.** Build every recipe in memory first, then send them all to
-\`validate_recipes\` - one round trip for the whole reading instead of one per component. A reader
-with 35 anatomies spent 35 turns asking the same question, and none of that was thinking. Only
-after the batch answers do you fix what it named and, if you fixed anything, ask once more.
-
-
-
-Open the project and answer the questions below. Then add a \`reading\` object to
-\`_synthesisui/census.json\`:
-
-\`\`\`json
-"reading": {
-  "themes": { "default": "dark", "has": ["dark"] },
-  "roles": { "canvas": "#050505", "foreground": "#f9fafb", "primary": "#4A90E2" },
-  "fonts": { "display": "Inter", "body": "Inter" },
-  "typeRoles": { "base": "body-m", "display": "h1", "xs": "caption" },
-  "concept": "one paragraph on what this product is",
-  "mood": ["focused", "dense", "technical"],
-  "rules": [
-    {
-      "text": "A RootWrapper is what makes the app-shell components position correctly - never render a Sidebar or Topbar outside one",
-      "applies": ["RootWrapper", "Sidebar"],
-      "kind": "implementation",
-      "when": ["next"],
-      "files": 3,
-      "evidence": "all three app shells wrap them; none renders them loose"
-    },
-    {
-      "text": "MetricCard is only used inside a grid, never on its own",
-      "applies": ["MetricCard"],
-      "kind": "limit",
-      "files": 1,
-      "evidence": "one dashboard page; may just be how it happened"
-    },
-    {
-      "text": "TextEditor's editable region is <EditorContent> - never render children into it directly",
-      "applies": ["TextEditor"],
-      "kind": "implementation",
-      "fact": true,
-      "evidence": "read in TextEditor itself: useEditor() feeds a single <EditorContent>"
-    }
-  ],
-  "components": {
-    "MetricCard": {
-      "anatomy": [
-        { "as": "icon", "name": "icon", "at": 1 },
-        {
-          "as": "stack",
-          "name": "body",
-          "at": 2,
-          "children": [
-            { "as": "component", "ref": "Text", "at": 3 },
-            { "as": "heading", "name": "value", "at": 4 }
-          ]
-        },
-        {
-          "as": "component",
-          "ref": "Tooltip",
-          "at": 5,
-          "children": [{ "as": "icon", "name": "info", "at": 6 }]
-        }
-      ]
-    },
-    "ArticleCard": {
-      "root": "Card",
-      "anatomy": [
-        { "as": "image", "name": "cover", "at": 1 },
-        { "as": "heading", "name": "title", "at": 2 },
-        { "as": "component", "ref": "TextEditor", "at": 5 },
-        {
-          "as": "row",
-          "name": "footer",
-          "at": 3,
-          "children": [
-            { "as": "button", "name": "publish", "at": 4 },
-            { "as": "button", "name": "discard", "classes": "btn btn-ghost" }
-          ]
-        }
-      ]
-    },
-    "TextEditor": {
-      "anatomy": [
-        {
-          "as": "row",
-          "name": "toolbar",
-          "classes": "flex gap-1 border-b p-2",
-          "children": [{ "as": "component", "ref": "ToolbarButton" }]
-        },
-        { "as": "external", "from": "@tiptap/react" }
-      ]
-    }
-  },
-  "by": "claude"
-}
-\`\`\`
-
-Every field is optional. Leave out what you did not establish.
-
-**\`concept\` and \`mood\` - the words the system opens with.**
-
-The guide, the vitrine and every page the platform writes about this system lead with these. Until
-07/08 nothing filled them, so a real import shipped with the seed's line - *"A calm grayscale
-starting point to make your own"* - as the first sentence of somebody's own design system. A sentence
-of ours at the top of their system says, in the first line, that nothing was read.
-
-\`concept\` is one paragraph on what this product IS, in their words where you can find them: the
-README, the docs site, the introduction their library ships with. \`mood\` is two to six words for how
-the product FEELS - \`focused\`, \`dense\`, \`editorial\`, \`playful\`. It is a judgement and it is yours to
-make, which is exactly why the arithmetic cannot: no reader infers a mood from a hex.
-
-Leave \`mood\` out when you did not form one. The seed's three words stay and the import SAYS they are
-ours, which is honest - three words you picked to fill a field are not.
-
-**\`themes\` - which themes the product actually ships, and which one it opens in.**
-
-This is the field that matters most and the one only you can fill. The ladder can say which
-*values* exist; it cannot say which themes the product *has*, because that answer lives in a
-\`dark:\` class name, a \`ThemeProvider\`, a toggle component, a \`prefers-color-scheme\` query -
-places no token reader sees. A real library declared nine near-blacks and five light greys in
-one \`:root\` block, and the import handed its near-black dashboard a white page.
-
-**DO NOT GO LOOKING. It is already counted**, under "What was countable, counted" in the
-census output:
-
-\`\`\`
-782 dark utilities across 79 files - so a second scheme is real here, not aspirational.
-switched by a ThemeProvider of their own, bound to data-theme, opening dark, with the
-  OS overruled.
-html carries lang="en" data-theme="dark" - which is the ancestor every dark rule of
-  theirs matches, and therefore the one ours must match.
-\`\`\`
-
-That is a real run. Your job is to CONFIRM OR CORRECT it, not to reproduce it - and a sweep
-would bring back a different slice than the next one, which is the worse property of the two.
-
-**Confirmation costs ZERO reads.** A real run spent 7 pattern searches and 39 file reads to
-"confirm" a theme the census had already counted (test14, 01/08) - that is re-deriving, not
-confirming. The rule: the census numbers are TRUE unless something you already read while
-doing other work contradicts them. Only on a contradiction do you open a file, and only the
-one file the evidence names. If nothing contradicts, write the census's answer down and move
-on - it costs one sentence.
-
-\`has: ["dark"]\` on a dark-only product is the right answer. It is better than filling a light
-slot on their behalf, and it makes adding light a deliberate act they take later.
-
-**\`roles\` - which of their values carries which meaning**, when their token names do not say.
-A project naming its neutrals by temperature (\`--color-darkgray-900\`) has told you nothing
-about which one is the page - and the census now answers that too, under "the page paints":
-
-\`\`\`
-the dark page paints darkgray-500 (2x on <main>/<body>, 2x full-viewport container)
-the light page paints #EDF3FA (1x body in global CSS), lightgray-200 (2x viewport)
-\`\`\`
-
-Three countable sources, strongest first: \`body { background }\` in global CSS, a background
-on \`<main>\`/\`<body>\`, and a full-viewport container (\`min-h-screen\` + a background IS the
-page). **Do not hunt for the canvas** - pick from this evidence, and name the source when you
-report it. When the sources disagree, that disagreement is itself the finding to surface.
-
-**\`fonts\` and \`concept\`** - the voice, and one paragraph on what this product is. The concept
-feeds every recommendation downstream, so a real one beats a generic one by a wide margin.
-
-**\`typeRoles\` - which of THEIR type steps plays each of our seven slots.**
-
-Their scale arrives under their own names now - \`h1\`, \`body-m\`, \`caption\`, \`overline\` - read
-straight off their \`--text-*\` declarations. What arithmetic cannot know is which of those steps
-is *body copy* and which is *the display size*, and our own components ask for it by slot:
-
-\`\`\`json
-"typeRoles": {
-  "xs": "caption", "sm": "body-s", "base": "body-m", "lg": "body-l",
-  "xl": "h3", "2xl": "h2", "display": "h1"
-}
-\`\`\`
-
-Leave it out and we pick by size, which is usually right and occasionally silly - a project whose
-\`overline\` is tiny and whose \`caption\` is tinier gets them the wrong way round. One line from you
-fixes it, and a person can correct it later in the studio either way.
-
-**Do not rename their steps to match ours.** \`h1\` stays \`h1\`. The whole point is that editing
-\`body-m\` in the studio moves the text in their app.
-
-**\`rules\` - how this company BUILDS, which is half of what they actually made.**
-
-A design system that arrives as tokens and recipes is only the vocabulary. The other half is the
-grammar: that a wrapper exists so the thing inside it works, that state arrives through a
-particular hook, that a card is never used loose. That is how the developer already thinks, and
-writing it down is what makes it survive them.
-
-Two dimensions, and they are independent:
-
-\`\`\`
-applies  []       the whole system
-         [a]      that component
-         [a, b]   a RELATION - two components that only work together
-
-kind     limit           a boundary, which the doctor can measure
-         implementation  how it is built, which no linter checks and an agent
-                         must be told
-\`\`\`
-
-The relation is the valuable one. \`[RootWrapper, Sidebar]\` says something neither name says
-alone, and it is exactly what an agent needs in order not to assemble it wrongly.
-
-**\`when\` - the environments a rule is true in.** Leave it out and the rule holds everywhere.
-
-This exists because the system gets installed places it was never written for. *"The state arrives
-through this hook"* is true about Next and **actively wrong** about Vue, so a rule that depends on
-the framework has to say so - and the install filters against the project it lands in, because
-only the project knows what it is built with.
-
-One dimension: a member of the stack (\`next\`, \`vue\`, \`react\`, \`svelte\`, \`tailwind\`, \`shadcn/ui\`).
-A rule may list several. It cannot combine conditions - \`next && !tailwind\` would be a query
-language, and then somebody maintains an interpreter forever.
-
-Use it only when the rule genuinely depends on the environment. Most rules do not: *"never render
-a Sidebar loose"* is true about their design regardless of framework, and pinning it to \`next\`
-would quietly drop it the day they add a second app.
-
-**Report \`files\` honestly - it decides whether an OBSERVED rule governs.** Three or more files
-is a habit and the rule arrives active; one file is a coincidence and it arrives as a candidate,
-inactive, waiting for the person to promote it. You do not make that call; you report the
-evidence and a threshold makes it. So a pattern you saw once should say \`"files": 1\` even when
-you are confident - being wrong about a law is worse than being slow about one.
-
-**\`fact: true\` - for a rule you read in the DEFINITION, where counting is the wrong question.**
-
-Three of these were reported as \`files: 1\` and arrived inactive, which was correct arithmetic on
-the wrong kind of claim (dono, 01/08):
-
-\`\`\`
-the editable region is <EditorContent>
-toolbar actions go through editor.chain().focus()
-extensions are configured at construction
-\`\`\`
-
-None of those is a coincidence waiting for a second sighting. They are how the component IS
-built, read off its own source, and they are true the moment somebody wrote it. Counting how many
-files agree would leave every construction law in the codebase inactive forever.
-
-So the test is **where you read it**, not how sure you feel:
-
-\`\`\`
-fact: true    you read it inside the component's own definition - its imports, its
-              JSX, how its state is wired. True by construction.
-files: N      you inferred it from how the component is USED across the project.
-              An observation, and the count is what makes it a habit.
-\`\`\`
-
-A rule can carry \`fact\` OR \`files\`, never both. If you find yourself wanting both, it is an
-observation - use \`files\`.
-
-**Write \`evidence\` as what you actually saw.** "All three app shells wrap them; none renders
-them loose" lets somebody disagree with a fact. "Best practice" lets them disagree only with
-you.
-
-Things worth looking for, none of them guessable from tokens:
-
-- a wrapper or provider whose whole purpose is to make something else work
-- an environment law - "this is Next, so the state comes through this hook" - which travels
-  with the system and only applies where that environment is
-- a component that is never used alone, or never used outside something
-- an axis that is always passed the same way in one place and never in another
-
-If you cannot say where a rule came from, do not send it. An invented law is worse than a
-missing one, because it will be obeyed.
-
-**\`components[Name].anatomy\` - what each component is MADE OF, and in what shape.** This is the
-field that decides whether a component previews as itself or as a grey box with a sentence in it,
-and only you can fill it.
-
-The census reads the ROOT element's classes and stops there, on purpose: descending a fixed
-number of levels picks a layout wrapper as often as a semantic part. **You read the component, so
-you decide the shape.** Send a tree; the CLI turns the classes into declarations and the platform
-turns declarations into roles.
-
-### The three frontiers
-
-Every node you send is one of three things, and knowing which is the whole job:
-
-\`\`\`
-a PART       an element of theirs        it has styles, and we draw it
-a COMPONENT  a component of theirs       it has a recipe of its own → named block + a RULE
-an EXTERNAL  a third-party library       we do not have it and never will → block + rules
-\`\`\`
-
-**Depth is not a number, it is where the frontier sits.** Descend until you meet another
-component or a library, stop there, and record the edge. That is why no parameter tells you how
-deep to go: a \`Divider\` is one node deep and a dashboard shell is five, and both are complete.
-
-**You decide the depth, and nothing downstream caps it.** The platform follows a
-\`component\` edge into that component's own anatomy, and then into ITS edges, as far as the chain
-goes - \`Chat → Message → TypingIndicator\` renders all three. So a frontier is not a dead end you
-are apologising for; it is how the chain gets walked. Record the edge and stop, and the whole
-depth appears anyway.
-
-### The ROOT is a frontier too
-
-\`\`\`json
-"components": {
-  "ArticleCard": {
-    "root": "Card",
-    "anatomy": [ … ]
-  },
-  "Modal": {
-    "root": "BaseDialog.Root",
-    "anatomy": [ … ]
-  },
-  "MetricCard": {
-    "root": { "name": "Card", "at": 1 },
-    "anatomy": [ … ]
-  }
-}
-\`\`\`
-
-**A library root whose name is not dotted needs its PACKAGE.** \`Pill\` returns Base UI's
-\`Toggle\`, and \`"root": "Toggle"\` would point at THEIR own Toggle - a switch - inventing a
-relation that does not exist. Send where it came from instead:
-
-\`\`\`json
-"Pill": { "root": { "from": "@base-ui/react/toggle" }, "anatomy": [ … ] }
-\`\`\`
-
-\`Radio.Root\` needs none of this: the dot already says it is somebody's namespace. Leaving the
-root out was the old workaround and it cost the surface its name (not-expressed.md, dono, 01/08).
-
-**The root takes what the PARENT paints on it.** When the returned root carries an override at
-the call site - \`MetricCard\` returns \`<Card className="flex-1 dark:bg-darkgray-300">\` - send the
-object form and point \`at\` the sketch node: \`{ "name": "Card", "at": 1 }\`. The string form still
-works and means "no override". Sending the root as an anatomy node instead would make it a leaf
-and orphan the content inside it, which is why it is a field on the root rather than a node
-(not-expressed.md, dono, 01/08).
-
-**Say what the component RETURNS when it is not a plain tag.** Measured on a real library
-(dono, 01/08): 14 of 23 components return one of their own - \`ArticleCard\` returns a \`<Card>\`,
-\`Button\` and \`Text\` return a \`<Component>\` - and 17 of 23 have no style of their own at all,
-because the surface belongs to the root.
-
-Without \`root\`, an \`ArticleCard\` previewed as floating text: the background, the border, the
-radius and the padding had nowhere to come from. With it, the card wears \`Card\`'s recipe as its
-shell and its own parts inside.
-
-- **their component** - the name, as their code spells it: \`"root": "Card"\`
-- **a library** - the dotted namespace, verbatim: \`"root": "Radio.Root"\`,
-  \`"root": "BaseDialog.Root"\`, \`"root": "Popover.Root"\`
-- **a plain tag** - leave it out. \`<div>\`, \`<td>\`, \`<button>\` are not frontiers.
-
-### The ten forms
-
-\`as\` says what a node IS, and the renderer draws that. Nothing else is accepted:
-
-\`\`\`
-image     a picture region: cover, thumbnail, media
-heading   the title line
-text      body copy, a label, a value
-button    an action
-field     an input somebody types into
-icon      a glyph, or a bare shape with no text
-row       arranges its children ACROSS
-stack     arranges its children DOWN
-component their component  → needs "ref": the name as their code spells it
-external  a library        → needs "from": the package name
-slot      the CALLER's content → takes "expects": what goes there, in their words
-\`\`\`
-
-\`row\` and \`stack\` ARRANGE their children. A \`component\` or \`external\` node HOLDS them -
-what their code nests inside a component they compose is markup THEY wrote, wearing their
-classes, and it belongs in the tree:
-
-\`\`\`json
-{ "as": "component", "ref": "Card", "at": 3, "children": [
-  { "as": "row", "name": "header", "at": 4 },
-  { "as": "external", "from": "recharts", "at": 8 } ] }
-\`\`\`
-
-A \`button\` holds a glyph and its words - \`<button><Upload/>Upload</button>\` is valid HTML,
-because a button's content model is *phrasing* and an icon is phrasing. So an \`icon\` and a
-\`text\` may sit inside one:
-
-\`\`\`json
-{ "as": "button", "name": "upload", "at": 9, "children": [{ "as": "icon", "name": "upload-icon", "at": 10 }] }
-\`\`\`
-
-Everything else is a leaf, and a leaf that holds elements has them HOISTED OUT to become
-its siblings - which destroys the nesting you read. If something arranges, say \`row\`/\`stack\`.
-
-**A frontier keeps its own \`at\`, so it keeps its own classes.** This is the one that gets
-talked out of: \`<Text className="text-lightgray-500 font-medium">\` is one of THEIR
-components, so it is a \`component\` node - and pointing \`at\` its sketch index carries the
-typography onto it exactly as a part would. Reading it as a \`text\` part instead is a
-downgrade with a real cost: \`composes\` is built from \`component\` nodes, so a library whose
-most-used atom is composed 363 times reported composing it ZERO times, and its strongest
-composition law became invisible (not-expressed.md, dono, 03/08).
-
-    <Tooltip><Info/></Tooltip>          frontier with a child, both survive
-    <Text className="…">Label</Text>    frontier with "at", the type survives
-
-If you find yourself writing "the contract cannot hold this, so I sent something else",
-check it against \`recipe_vocabulary\` before you write the workaround - and if the check
-disagrees with this file, the check is right and this file is old.
-
-### \`{children}\` is a node, not an absence
-
-Four things decide what a component looks like on screen, and only the first is in its own source:
-
-\`\`\`
-its own classes          you are reading them
-what the caller passes   <tr>{children}</tr>          → slot
-its runtime state        value={62} drives the width  → a part with no static size
-a library drawing it     <EditorContent>              → external
-\`\`\`
-
-A component whose content is not its own had nothing to send, so it sent nothing and previewed as
-a blank box. A real \`DataTableRow\` is sixteen lines - a \`<tr>\` carrying a border, a hover and
-\`{children}\` - and it came back empty, which is not wrong so much as unreadable (dono, 01/08).
-
-So send the slot. \`expects\` is what the caller is supposed to put there, in the words their own
-code uses - read the prop type, the JSDoc, or a call site:
-
-\`\`\`json
-{ "as": "slot", "name": "row", "classes": "border-b hover:bg-ocean-50/30", "expects": "the cells" }
-\`\`\`
-
-It keeps its \`name\` and \`classes\` like any other node - the border and the hover are real and
-belong on that element. What it does not do is invent content: the preview says "the cells go
-here", which is exactly what the file says.
-
-Good \`expects\`: \`"the cells"\`, \`"one row per item"\`, \`"the form"\`, \`"the page body"\`. Leave it
-out rather than writing \`"children"\` - the field is for what a person would call it.
-
-**A part can carry its own variants and states**, not only a flat class string. Send
-\`classes\` for the resting look and let the prefixes travel inside it - \`hover:\`,
-\`focus-visible:\`, \`disabled:\`, \`group-hover:\`, \`md:\`, \`data-[state=open]:\` are all read
-and land where they belong. \`group-hover:\` in particular now has somewhere to go: it
-means "when the component around me is hovered", which is what a real table row is built
-on and what nothing could express before.
-
-**\`name\` is the part name**, and it is what carries the styles - flat, lowercase, kebab. Never
-nested: \`"actions.generate"\` compiles to two CSS classes and is invalid, so name it
-\`"generate-action"\`. Name what it IS - \`label\`, \`value\`, \`delta\`, \`cover\`, \`toolbar\` - because the
-name becomes a class in their stylesheet and \`div2\` is a name somebody has to live with.
-
-\`root\`, \`wrapper\`, \`container\`, \`base\` and \`content\` are the element itself: **do not send them as
-nodes.** Their styles already sit on the component, and a node for them draws a box inside its own
-box.
-
-A node with no \`classes\` is fine - it still carries structure, which is most of the value. A
-\`component\` or \`external\` node takes no \`name\` and no \`classes\`: the styles there are not theirs
-to hold.
-
-### A component of theirs is a NAMED BLOCK, not an expansion
-
-When you hit \`<TextEditor>\` inside \`<ArticleCard>\`, send
-\`{ "as": "component", "ref": "TextEditor" }\` and **stop**. Do not inline what TextEditor is made
-of. It previews as a block carrying its name, and it becomes a **rule** - \`ArticleCard\` +
-\`TextEditor\`, a relation, which is the one thing a flat list of prose could never say.
-
-That edge is a **fact, not a habit**: you saw it in the definition, so it is true once and for
-all, and it arrives active without waiting for a third sighting.
-
-**Send what the PARENT paints on it.** A frontier node takes \`at\` (or \`classes\`) like any other,
-and it means something different there: not the child's look, which has a recipe of its own, but
-the override the parent applies where it composes the child. Measured on a real library, four of
-these were being lost:
-
-\`\`\`
-Paginate     <Button className="min-w-[120px] dark:bg-ocean-800 …">   a 120px floor
-MetricCard   <Card className="flex-1 dark:bg-darkgray-300">           overrides Card's dark
-ArticleCard  <Tag className="w-fit">
-Message      <ReactMarkdown className="text-royal-blue-500 underline …">
-\`\`\`
-
-The MetricCard one is the sharpest: losing that override means the preview draws the wrong dark.
-Same for an \`external\` node - a library's component still wears what their code puts on it.
-
-Three things that are NOT edges, and sending them as such would be wrong:
-
-- \`motion.div\`, \`Dialog.Root\`, \`Radio.Item\` - a library's namespace, not a component of theirs
-- an HTML element with a capital in a variable name
-- **a name the file binds itself and does not export.** Capitalisation is React's rule for "not an
-  html tag", which is a different question from "is this a component of the system". Both of these
-  arrived as components of a real library and neither is one (dono, 01/08):
-
-\`\`\`tsx
-function Text({ as: Component = "p" })   // a prop rename - the polymorphic idiom
-const CustomLegend = () => (…)           // a closure inside LineChart
-\`\`\`
-
-  The test is reachability: a component of the system is imported from somewhere or exported to
-  somewhere. A name that only lives inside one function body is that function's private wiring,
-  and giving it a recipe invents a component nobody can import.
-
-If the component came from a package, it is \`external\`, not \`component\`.
-
-### An external dependency does not render, and that is the answer
-
-A \`TextEditor\` built on tiptap cannot be previewed: the editable region is tiptap's, and drawing
-a fake one would be inventing. So it previews as a block naming the package, and the value moves
-into the **rules** - which is where an agent will read it anyway.
-
-A library is never one rule, it is a family, and you are reading the file so you can see all of
-them. Send them as normal \`rules\` with \`applies: ["TextEditor"]\` and
-\`kind: "implementation"\`:
-
-\`\`\`
-requires @tiptap/react
-the editable region is <EditorContent> - never render children into it directly
-toolbar buttons go through editor.chain().focus()
-extensions are configured at construction, not toggled later
-\`\`\`
-
-**Name the package in the rule; leave the version to the evidence.** The CLI reads the version out
-of their manifest and attaches it (\`"^2.1.0 in packages/ui"\`), so the rule does not go stale the
-day they upgrade and the information is not lost either.
-
-### If \`recipe_vocabulary\` is missing, the MCP registration is stale
-
-A real run reported "recipe_vocabulary and validate_recipe do not exist on this server" and
-sent every anatomy unvalidated - the repo's \`.mcp.json\` pinned \`synthesisui@0.16.78\` while
-the import itself ran unpinned at latest: **two CLI versions in one repo, in one run** (dono,
-01/08). Before concluding a tool does not exist, check the registration:
-
-\`\`\`json
-"args": ["synthesisui@latest", "mcp"]
-\`\`\`
-
-Tell the person to fix the pin and restart the server. Do not work around a stale server by
-skipping validation - unvalidated is how correct work arrives as nothing.
-
-### Ask what a recipe can hold, BEFORE you write one
-
-\`\`\`
-recipe_vocabulary          what a recipe CAN hold - states, forms, the floor per kind
-validate_recipe(name, r)   would this survive? what would NOT, by name
-\`\`\`
-
-Both are MCP tools on the \`synthesisui\` server. **Call \`recipe_vocabulary\` once, at the
-start**, and work from what it says rather than from what you remember - it is served from
-the live contract, so it is right about a state that was added last week and this file is
-not.
-
-Then **validate every recipe before you send it.** This is not a formality. Nothing checked
-a recipe before these tools existed, and four separate readings were lost in silence: a
-\`dark:\` inside a variant, a state the compiler cannot spell, a part name with a dot in it,
-a layer with an empty style. Each of those was correct work that arrived as nothing.
-
-**A FIELD OF THE READING IS NOT A FIELD OF THE RECIPE.** \`looks[X].dark\` is how the census
-reports a dark half it measured; a recipe has one way to say the same thing, and it is a
-layer:
-
-\`\`\`json
-{ "when": { "scheme": "dark" }, "style": { "backgroundColor": "{color.darkgray.300}" } }
-\`\`\`
-
-Copying \`dark\` in at the top level of a recipe costs you that block, silently, on every
-component that has one - and \`import\` already converts it for the recipes IT writes, so the
-only place this can be lost is a recipe you hand-built (not-expressed.md, dono, 03/08).
-
-**Which means: do NOT pre-convert the dark half inside \`looks[X]\`.** The conversion already
-happens on the way in, so adding a \`scheme: "dark"\` layer to the census's own look while leaving
-the \`dark\` key in place produces the SAME layer twice - measured at 28 components in one real
-import, with the properties in a different key order so it did not even look like a duplicate
-(dono, 05/08). \`validate_recipe\` telling you a top-level \`dark\` would be dropped is true about a
-RECIPE and says nothing about the reading: you send a reading, and the reading's \`dark\` is read.
-
-**In ONE call.** Build every recipe in memory, then send the whole reading to
-\`validate_recipes\` - one round trip for 35 components instead of 35. \`validate_recipe\`
-(singular) is for the one recipe you rewrote after the batch named it; reaching for it in a
-loop is the shape this optimisation exists to kill (dono, 01/08).
-
-\`\`\`
-validate_recipes({ recipes: [{ name, recipe }, …] })   the whole reading, one question
-validate_recipe(name, recipe)                          the one you just fixed
-\`\`\`
-
-\`validate_recipe\` never answers yes or no. It answers with what would not survive:
-
-\`\`\`
-REFUSED by the contract        fix it or the WHOLE recipe is dropped
-HELD BY THE SCHEMA AND         the silent half - it validates, it compiles to nothing.
-DROPPED BY THE COMPILER        This is the one worth fixing before sending.
-BELOW THE FLOOR                it would arrive and not be drawable
-\`\`\`
-
-### Write down what we could not hold
-
-Everything \`validate_recipe\` reports goes into \`_synthesisui/not-expressed.md\`, appended
-per component:
-
-\`\`\`
-## MetricCard
-- layers[2].when.scheme … (what the validator said, VERBATIM)
-- below the floor: a fill, so it reads as a surface rather than as bare text
-- the source you read, quoted, so the gap can be checked without opening their repo
-\`\`\`
-
-**Verbatim, and with the source beside it.** This file is how a gap on OUR side gets fixed
-instead of being worked around: said only in chat it evaporates, and paraphrased it becomes
-unfindable. If the file ends up empty, say so - that is the good outcome and it is worth
-one line.
-
-### The floor: what a preview needs before it can draw anything
-
-The owner asked this directly - "to render the button, what do I need to have?" - and the
-honest answer per kind is short. **Work the list for the component's kind and say which
-ones their code answers.** A reader holding this knows that a button read with no hover
-means the hover has not been found yet, and goes back for it.
-
-\`\`\`
-action    the resting fill and the text colour on it · padding and radius ·
-          text size and weight · hover, focus-visible AND disabled - three looks,
-          not one · gap and alignment if it can carry an icon
-field     the fill of the typing area and the text colour in it · the edge at rest
-          and the edge when focused · padding and radius · the placeholder colour
-          (a different decision) · the invalid look if their code has one
-control   the off look and the on look, which IS the component · the size · the
-          focus ring, since a control is reached by keyboard · the disabled look
-surface   the fill, and the edge or the elevation that separates it from the page ·
-          radius and inner padding · the gap between what it holds · whether it
-          changes at a breakpoint
-pill      the fill and text colour, per status if it has statuses · radius and
-          horizontal padding · the text size
-indicator the colour, or the two if it has a track and a fill · the size or ratio ·
-          WHAT DRIVES IT - which prop moves the number
-text      the colour · the size and line height, or a note that both are inherited ·
-          the weight and family if they differ from body
-\`\`\`
-
-**A component genuinely has none of something, and saying so is a real answer.** This is a
-list to work through, not a gate: nothing is blocked for failing it, and inventing a
-background to satisfy it would be the exact opposite of the point.
-
-There is one hard line under all of it: a recipe that declares NO colour, NO edge and NO
-size cannot be told apart from the page. The platform says so rather than printing a
-score, and it follows the root first - a component whose surface belongs to the \`Card\` it
-returns is complete.
-
-### Media is not a style question
-
-An \`image\` node draws a grey region and carries nothing about how to use it, so an agent
-composing with it invents the answer. Send these as \`rules\` on that component:
-
-\`\`\`
-a ratio or an explicit size    a region with neither collapses before the image loads,
-                              and the page jumps when it arrives
-an object-fit                 cover crops and contain letterboxes; which one is right
-                              is about the content, not a default
-a corner radius               an image inside a rounded surface has to say whether it
-                              follows the curve
-what shows while it loads     a skeleton, a blur, or the surface colour
-what shows when it is missing a missing asset must never look broken
-\`\`\`
-
-### The census is the ONLY source of truth - four universal rules
-
-These hold in ANY project, whatever its folders, and they exist because a run spent most
-of its tokens re-reading what the census already carried (dono, 01/08):
-
-1. **No iterative sweeps by layer or folder.** Never walk atoms/, then molecules/, then
-   organisms/ reading files - the census walked everything once.
-2. **Batch over the census's own data.** Everything you consume comes out of
-   \`_synthesisui/census.json\` - the looks, the sketch, the signals, the coverage.
-3. **Metadata is one pass, already done.** Theme, fonts, page, libraries, architecture
-   arrive counted. You confirm in a sentence; you never re-derive.
-4. **A file read whose content the census carries is a PIPELINE FAILURE.** Say so in
-   not-expressed.md when you catch yourself - that report is how the census grows.
-
-### Name the parts from the SKETCH - the file is already in the census
-
-Every component's look carries \`sketch\`: its own markup as data - tag, class string,
-depth and text per element, in document order. That is everything you used to open the
-file for, minus the one thing that is genuinely yours: the NAMES.
-
-\`\`\`json
-"sketch": [
-  { "tag": "div",  "depth": 0, "classes": "rounded-md border ..." },
-  { "tag": "button", "depth": 1, "classes": "flex w-full ..." },
-  { "tag": "span", "depth": 2, "classes": "size-2 rounded-full bg-success-500" },
-  { "tag": "span", "depth": 2, "text": "Status" },
-  { "tag": "ChevronDownIcon", "depth": 2 }
-]
-\`\`\`
-
-Read that and write the anatomy: depth 1 is \`trigger\`, the dot is \`status-dot\`, the
-capitalised tag is a frontier. **Do not open the component file** - if the sketch is
-missing or visibly truncated (150-node cap), say so in not-expressed.md and only then
-read, because that gap is the census's to close.
-
-**NAME THE NODE BY ITS INDEX - never retype the class string.** Each anatomy node carries
-\`"at": <index into this sketch>\`, and the CLI reads the exact string the census measured:
-
-\`\`\`json
-{ "as": "button", "name": "trigger", "at": 1 }
-\`\`\`
-
-Retyping is where a modifier goes missing. Six real components arrived with \`hover:\` and
-\`focus:\` in their source and no state at all in the recipe, and the parser was not the
-problem - the copy was (audit, dono, 01/08). An index cannot lose a class. \`classes\` still
-works and still wins when you send both, which is for the one node the sketch could not
-reach; every other node uses \`at\`.
-
-### What the CLI now reads for you, so do not spend a turn on it
-
-**The rule under all of this: you are given EVIDENCE and asked to DECIDE.** You are never
-given a question and asked to search. A grep is a tool call whose entire output lands in your
-context to answer something that is one number, and two readers sweeping the same repo bring
-back different slices - so the answer changes between runs, which is the part that actually
-costs.
-
-Measured on a real run: the census counts the theme signals, the libraries and their
-versions, and the conflicts. The reader that swept for them anyway spent a dozen shell calls
-to arrive at a subset of what was already on screen.
-
-Four things are read deterministically from the source. Reporting them back as if you
-found them wastes a turn and risks contradicting the measurement:
-
-- **variant and state styles** out of \`cva\`, \`tv\` and \`cond === "x" && "classes"\` -
-  including a hover PER variant, compound variants, and \`[var(--their-token)]\`
-- **whether a component forwards the rest of its props**, and which it names
-- **which components appear inside which** at their call sites, as pairs
-- **the folder architecture**, and it will ask you which one has priority
-- **their gradients**, as tokens: \`--gradient-ui\` and its Tailwind-namespace twin land on
-  \`{gradients.ui}\`, composed \`bg-gradient-to-r from-x to-y\` stops become one gradient in
-  their own colour tokens, and a signature gradient is no longer a string trapped in class
-  names
-- **the theme**: how many dark utilities, in how many files, which library switches it, which
-  attribute it is bound to, what \`<html>\` carries, and how many \`prefers-color-scheme\` rules
-- **the libraries and their versions** - motion, icons, primitives, charts, editors - with the
-  file count for each, and a CONFLICT when one job is done by more than one of them
-- **which of your components are ours**, matched against the LIVE catalogue rather than a
-  list baked into the CLI. If a run says it fell back to the built-in names, the match is a
-  smaller answer than it should be and the reason is on screen
-
-What is still yours: the anatomy tree, the part names, the frontiers, the roles, the
-concept, and the rules that are not about a class name.
-
-### A scale's INTENT is a rule, and their docs already wrote it
-
-A three-step spacing scale with documented meaning - sm "close UI relationships", md
-"standard internal padding", lg "layout gutters" - carries a law nothing else expresses:
-**this system has exactly three steps, and a fourth is a mistake.** A real import read the
-scale and lost the sentence (dono, 01/08).
-
-When their docs (a \`Geometry.mdx\`, a design-tokens page, comments beside the \`@theme\`)
-state what a step is FOR, send it as system-wide rules - \`applies: []\`, \`fact: true\`, the
-doc as evidence.
-
-**One doc pass, and only for what the census cannot hold.** A scale's INTENT is prose, so it
-genuinely needs a human sentence somebody wrote - but a scale's VALUES, its token names, the
-theme, the class convention and the architecture are all measured and sitting in \`declared\`,
-\`signals\` and \`conventions\`. Opening a docs page to confirm a number the census already
-carries costs a turn and answers nothing. Read the docs ONCE, for the sentences; never in a
-loop, and never to check arithmetic.
-
-Send it like this:
-
-\`\`\`
-this system's spacing has exactly three steps - sm for close relationships inside a
-control, md for standard internal padding, lg for layout gutters. A fourth step is a
-mistake, not a gap (Geometry.mdx)
-\`\`\`
-
-\`applies: []\` is what makes it reach CLAUDE.md and every prompt after - the same route the
-architecture rule travels.
-
-### How much to send
-
-Send an anatomy for **every component with visible structure**, which is nearly all of them. A
-\`Chat\` has a message list and a composer; a \`MetricCard\` has a label and a value; a
-\`CircularProgress\` has a track and a fill. Sending none is right only for something genuinely
-undivided - a \`Divider\`, a \`Spacer\`.
-
-**There is no node budget, and the one I gave you was wrong.** "Three to six nodes is a
-recognisable component" cost a real \`ArticleCard\` most of itself: it has a carousel with arrows,
-three image buttons, a source chip, a refresh, a rich body, a source line and a "Similar
-published content" region with two selects and two buttons - about fourteen regions - and the
-reading came back with six because the guidance said so (dono, 01/08).
-
-**One node per region a person can point at.** If they can say "that part", it is a node. What
-you still collapse is a wrapper whose only job is \`flex\` - fold it into the \`row\` it already is,
-because that is not a region, it is plumbing.
-
-The cost of sending none is not neutral. A component with no anatomy previews as a grey box with
-a sentence in it, or - if its kind is \`indicator\` - as a small blank shape. A component with a
-real tree previews as itself, and its spec view shows what the system understood. That is the
-whole reason this field exists, and skipping it quietly is how a real library came back looking
-empty (dono, 01/08).
-
-The older flat form - \`"parts": [{ "name": "label", "classes": "..." }]\` - still works and still
-styles correctly. It just carries no shape, so the preview lays the parts out in a row and the
-spec view has no nesting to draw.
-
-### 4. Send it
-
-Everything is answered and everything is validated:
-
-\`\`\`
-npx synthesisui import --census _synthesisui/census.json --name "<the name>" [--registry <from step 0>]
-\`\`\`
-
-**This is the one real write.** Everything before it is on their disk and costs nothing to redo.
-
-Pass every answer explicitly - \`--name\`, \`--scope\` and \`--usage\` if they changed either, and the reading rewritten
-with their primary and their theme. The CLI cannot ask anything when you are the one running it:
-it has no terminal, so its own prompts are skipped by design. **Whatever you did not ask, nobody
-asked.** Skip 2a and the first time they learn where their system came from is when they open
-it.
-
-If it refuses, it says which of the three it is: no session, an expired one, or a token issued
-by a different host. All are recoverable from the census on disk; none need a re-measure.
-
-### 5. Tell them exactly what they have, and what is theirs to decide
-
-The import publishes **v1** as a faithful baseline and opens a **v2 draft** holding the
-proposal. Say all of this in your own words - do not just print the link:
-
-**What landed in v1**, and it is theirs, not ours:
-
-- the scheme it opens in, and whether a second one was built
-- their ramps under their own family names
-- their exclusive components as contracts - **the axes their types declare, the look transcribed
-  out of their own class names where it was readable, and the anatomy you sent.** Say what is
-  still empty out loud, because a half-written recipe looks like a failure until someone explains
-  which half was deliberate: nothing was invented, so a component whose classes named no token
-  arrives with structure and no colour. The platform shows that as *Not written yet*, not as a
-  zero.
-- **their own ladder**, if their repo has one. Components under \`atoms/\`, \`molecules/\` and
-  \`organisms/\` arrive grouped that way in the vitrine, read off the folder rather than guessed
-  from a name. You do not send this and you should not try to: a project that does not organise
-  itself this way gets one honest group instead of three invented ones.
-- **the shape of each component**, if you sent one: its parts nested as they nest in their code,
-  the components of theirs it composes, and the libraries it needs. Each edge is also a rule now,
-  which is what makes it survive being looked at once.
-
-**What is waiting in v2**, which only they can approve: near-duplicate colours collapsed,
-unnamed heavy hitters given a place. Point at the link the CLI printed and name the two or
-three biggest items so they know whether it is worth opening now.
-
-**What the report said no to.** The notes carry every silent drop made loud - a value not
-found in their code, a token that is not ramp-shaped, a name the document refused. Read them
-and pass on the ones that matter.
-
-**And the notes now judge YOUR reading, which nothing did before.** \`validate_recipes\` looks at
-the \`anatomy\` inside a recipe; the anatomy you author travels in \`reading.components[X].anatomy\`,
-and until 05/08 no one looked at it - so a note may name findings the batch validator never saw.
-Two kinds, and both are yours to act on:
-
-- **structure that does not survive the reader.** Nine of these landed on a real import, every one
-  the same shape: a \`button\` holding a \`stack\`. Only \`row\` and \`stack\` arrange, so the children are
-  hoisted out and become siblings - the nesting they wrote is gone. Fix it in the reading and send
-  again; the note tells you the component and the path.
-- **references pointing at components the system does not contain.** 215 of 660 on that same
-  import. This one is usually NOT a mistake of yours: their component composes something the gate
-  left out, or something outside the scope that was measured. Say the number to the person and name
-  the top few, because a specimen there draws the reference and nothing else - and if the missing
-  piece is central to their system, the answer is a wider \`--scope\`, not a rewritten reading.
-
-### 6. Close the loop in their repo
-
-The system is only worth something once it governs the code it came from:
-
-\`\`\`
-npx synthesisui add <slug> [--registry …]       # INSTALLS it: writes _synthesisui/ds/<slug>/
-npx synthesisui doctor [paths…]                 # the contract starts being enforced
-npx synthesisui upgrade <slug> [--registry …]   # LATER, when a new version is published
-\`\`\`
-
-**The order is not a preference: \`upgrade\` reads the \`.lock\` of an installed DS and refuses
-without one** - \`"<slug>" is not installed here - run \\\`synthesisui add <slug>\\\` first.\` So
-offering \`upgrade\` to somebody who has never installed hands them a command that fails on contact
-(dono, 05/08). Check for \`_synthesisui/ds/<slug>/.lock\` before you recommend either one.
-
-And say what each writes, because one of them rewrites code: \`add\` and \`upgrade\` write inside
-\`_synthesisui/\`, and \`upgrade\` ALSO regenerates components - but only files carrying the
-\`Generated by SynthesisUI\` header, so their own components are never touched.
-
-\`doctor\` is where the promise pays off: every design value written by hand, the token their
-own system already has for it, and the components used outside the axes their contract
-declares. In a monorepo, scope it per app - one system, N consumers.
-
-## What good looks like
-
-The user should end up with a system that reads like theirs and not like ours:
-
-- **their names travel** - \`vivid-pink\`, \`darkgray-900\`, whatever they called it
-- **their default face** - a dark product opens dark
-- **no borrowed colour** - if it is in the system, it is in their code
-- **their exclusive components arrive as contracts** - the axes declared, the look transcribed and
-  never invented, and the anatomy read out of their own JSX
-- **a component previews as itself** - an \`ArticleCard\` shows a cover, a title and a footer of
-  buttons, and says out loud that the editor in the middle is their \`TextEditor\` on tiptap
-- **the duplicates are named out loud** - three components that all read as a badge, a brand
-  colour painted 275 times with no token
-
-## What to resist
-
-- **Do not let memory speak for the server.** Notes report the past; a claim about what exists
-  now needs the one \`curl\` in step 0b. Four statements in one real run were wrong this way, and
-  every option offered on top of them was wrong too (dono, 05/08).
-- **Do not run \`import\` without \`--dry\` to explore.** It creates a real system on their
-  account. Measure with \`--dry\`, read, then send once.
-- **\`--dry\` does not delete anything.** It writes \`census.json\` and creates the folder if needed -
-  there is no \`rm\` in the import path. What it DOES do is overwrite that one file, so a reading
-  injected into it is lost; everything else under \`_synthesisui/\` (the config, \`ds/<slug>/\`, the
-  hook's marker, \`not-expressed.md\`) survives untouched. Telling somebody a measure will wipe their
-  install talks them out of a safe command (dono, 05/08).
-- **Do not redo the arithmetic.** If you find yourself counting colours, you are in the wrong
-  half of the pipeline.
-- **Do not fill \`roles\` with what looks nice.** It is a reading of what their code already
-  does, not a redesign. The v2 proposal is where improvement belongs, and the user approves it.
-- **Do not guess \`themes\` to be helpful.** Omitting it falls back to counting rungs, which is
-  a guess the product labels as a guess. A wrong confident answer does not get labelled.
-- **Do not apologise for the empty contracts.** They are the design. Explain them.
-- **Do not come back with a question once step 3 has started.** Everything answerable was
-  answered in step 2. An interruption in the middle of a silent phase costs the person the
-  whole phase, because they now have to hold what you were doing in their head.
-- **Do not narrate the mechanics.** No grep, no glob, no "let me check whether X exists", no
-  draft you rewrote. Findings earn a line; the machine talking to itself does not.
-- **Do not open a file to confirm something the census already carries.** \`signals\`, \`sketch\`,
-  \`declared\`, \`conventions\` and \`coverage\` are the answer to almost every "let me just check".
-- **Do not validate in a loop.** One \`validate_recipes\` for the whole reading.
-
-## Not yet true
-
-Say so if it comes up, rather than implying otherwise:
-
-- **The v2 proposal covers colour only.** Near-duplicate collapse and unnamed heavy hitters.
-  It does not yet propose swapping a raw value in a recipe for the token that holds it, nor
-  retiring an option nothing passes, nor a categorical chart palette - all measured, none
-  offered.
-- **You do not write recipes.** Exclusive components arrive as contracts and stay that way
-  until someone writes the look. When that lands, the rule will be **provenance, not
-  membership**: a value transcribed out of their file cites the file and line, and is
-  legitimate even if the census never saw it, because the census measures less than a project
-  contains.
-- **Nothing confirms the import back.** You send, the CLI prints a link, and nobody asks the
-  platform whether what arrived is what you meant. Until that exists, tell them to open the
-  link and check.
-`;
-/** Where it lands in the consumer's repo. */
 export const IMPORT_SKILL_PATH = ".claude/skills/sui-import-ds/SKILL.md";
+export const IMPORT_SKILL = '---\nname: sui-import-ds\ndescription: Turn a codebase the user ALREADY has into a SynthesisUI design system. Use when someone points at an existing repo, app or component library and asks to import it, adopt it, bring it in, or "make a design system from this" (e.g. "/sui-import-ds", "importa o meu packages/ui", "turn this app into a design system"). Drives the full pipeline - census \u2192 your reading \u2192 import \u2192 v2 proposal - and answers the questions arithmetic cannot.\n---\n\n# Import Design System - served live\n\nThis playbook is served from the platform, not shipped in this file - it is\nalways current, and your context only carries the step you are on.\n\n1. Call the `playbook` tool on the `synthesisui` MCP server with\n   { "skill": "import" } - you get the framing and a table of contents.\n2. Fetch ONLY the chapter for the step you are on:\n   { "skill": "import", "section": "<id from the toc>" }. Never fetch more\n   than the current step needs.\n3. Follow it exactly. When the step is done, fetch the next chapter.\n\nIf the tool answers that you are not signed in, run `npx synthesisui login`\nin the terminal and call it again. If the `synthesisui` MCP server is not\navailable at all, run `npx synthesisui connect`, restart the session, and\ninvoke this skill again.\n';