design-foundations

name: design-foundations description: Build a coherent design-foundation JSON from a liberated site — semantic color/typography/spacing roles with evidence trails. Consumes the partial scaffold produced by liberate_design_foundation_scaffold plus aggregate HTML/CSS analysis and representative rendered HTML; produces a complete design-foundation matching the schema. Call after liberation, before theme generation.

Design Foundations

You build a coherent design foundation — a semantic, reviewable intermediate between raw SP1 token observations and WordPress theme generation.

Input contract

The agent gives you:

1. A PartialDesignFoundation JSON blob (the output of liberate_design_foundation_scaffold). Some roles are already filled in by deterministic rules. The remaining slots are null, listed in skillTodos.
2. Aggregate HTML/CSS analysis files: palette.json, typography.json, breakpoints.json, and when present computed-styles.json.
3. Paths to representative rendered HTML excerpts (e.g. <outputDir>/html/homepage.html).
4. Optional paths to screenshots for ambiguity checks only.
5. Access to the MCP tool liberate_design_foundation_validate.

Evidence mode — HTML/CSS first, screenshots only for ambiguity

Vision context is the dominant cost in this skill. The normal path should not open screenshots. Use rendered HTML, aggregate computed CSS, and deterministic scaffold evidence first.

If the invoking runner provides a Foundation sample JSON, that sample is a hard cap: do not inspect HTML outside the listed entries. Use aggregate palette.json, typography.json, breakpoints.json, and computed-styles.json for broad frequency and component signals. Use only the sampled HTML for semantic role decisions.

Step 1 — aggregate CSS analysis (cheap):

Read aggregate analysis before opening any visual asset:

• palette.json: high-frequency background colors.
• typography.json: observed font families, sizes, weights, and line heights.
• computed-styles.json: component-level text, background, border, radius, padding, typography, and background-image evidence for selectors like header, footer, a, button, form controls, hero/card/CTA containers.
• breakpoints.json: responsive tier evidence.

These files are generated from the live DOM using browser-computed CSS. They are the source of truth for colors and typography; screenshots are not needed to discover tokens.

Step 2 — inspect the representative HTML (cap):

Use the provided Foundation sample. In the fast path it contains at most one representative rendered HTML file, preferring the homepage. Inspect it for:

• <header>, <footer>, <main> boundaries.
• Button/link/nav/form selectors and class names.
• Inline styles, CSS variables, gradients, and background images.
• Whether a high-frequency color is actually a surface, text color, CTA, border, or decoration.

Step 3 — screenshot fallback only when needed:

Open a screenshot only if HTML/CSS evidence cannot answer a specific semantic question, for example:

• A color appears in multiple roles and computed selectors do not identify usage.
• A gradient/background image needs a role (hero background, cta banner, decorative) and HTML class names are unclear.
• Mobile stacking or overlaid composition affects a token decision.

If you open a screenshot, state exactly what ambiguity it resolved. Do not read all screenshots.

Output contract

Return a complete DesignFoundation matching the schema documented in src/lib/design-foundation/schema.ts. The JSON must pass liberate_design_foundation_validate with { ok: true }.

You MUST call liberate_design_foundation_validate on your output before returning. If it returns errors, fix them and re-validate. Do not return until validation passes.

Role assignments — judgment criteria

The scaffold fills:

• color.surface.base — lightest high-frequency palette entry
• color.text.default — darkest high-frequency palette entry
• typography.families.body — most-common body fontFamily
• breakpoints.{sm,md,lg,xl} — nearest-tier mapping
• gradient.* — CSS strings (roles marked "TODO")
• inputsDigest.* — sha256s

You fill everything else. Use the criteria below.

color.surface

color.text

color.accent

Distinguish by saturation, frequency, and usage signal:

Ambiguity rule: If two entries both look like primary candidates, pick the one with higher urls (more pages). If still tied, pick the one with higher saturation.

color.border

typography.families

The scaffold fills body. Fill:

gradient

Scaffold fills css and evidence; you fill role. Map each gradient to its usage:

• hero background — dark, full-bleed, used as page hero/header background
• cta banner — used on call-to-action bands and subscribe blocks
• headline accent — light-to-highlight gradients used with background-clip: text on headlines
• decorative — anything else

Look at the HTML excerpt to see which selectors use the gradient.

components (minimal fixed set)

Fill all five with token references (not raw values). Reference form: "color.accent.primary", "radius.base", "spacing.4".

openQuestions

Flag things that will affect downstream theme generation:

• Commercial fonts (e.g. Reckless, Displaay faces) → suggest open-source substitution (Fraunces, Playfair).
• Uncapturable fonts (Typekit / Monotype / hashed builder names) → these can't be self-hosted (no reachable woff/woff2). The deterministic font pipeline now AUTO-SUBSTITUTES them to the closest FREE web font, self-hosts its woff2, and binds it as the theme's body/display family — see references/theme-tokens.md ("Commercial / uncapturable → free substitution"). Record the substitution as an openQuestion (blocksReplica:false) noting which family was swapped (e.g. quasimoda → Hanken Grotesk) so an operator can confirm the visual match; do NOT leave the family as a bare sans-serif fallback.
• Font files loaded from unknown CDNs (no open license path).
• Colors observed once or twice that look like one-off overrides rather than design-system tokens — worth flagging for operator review.

Each entry: { id: "short-slug", question: "Confirm ...", blocksReplica: true|false }.

Evidence discipline

Every role you fill must have at least one evidence entry. Good evidence:

• "palette[5]: 92/186 urls, 341 occurrences" (cites raw SP1 frequency)
• "computed button background on homepage.html" (cites HTML + element)
• "hero section @ screenshots/desktop/homepage.png" (cites screenshot)

Bad evidence:

• "looks like primary" (no citation)
• "skill-filled" (placeholder — unacceptable)

Process

1. Read the PartialDesignFoundation. Note every path in skillTodos.
2. Read aggregate HTML/CSS analysis and the representative HTML excerpt.
3. For each slot in skillTodos:
   • Apply the criteria above.
   • If you cannot confidently fill a slot (e.g. no mono family exists), leave it null AND remove its path from skillTodos. This communicates "deliberately empty" instead of "forgot to fill."
4. Populate components with the fixed set above.
5. Populate openQuestions with any concerns.
6. Set skillTodos: [] once every remaining slot is filled.
7. Invoke liberate_design_foundation_validate. If it fails, fix and retry.
8. Return the validated JSON to the calling agent.

Anti-patterns

• Hallucinating tokens. If SP1 didn't observe a serif family but the agent seems to want one, do NOT invent "Merriweather". Leave display: null and add an openQuestion about font substitution.
• Skipping evidence. Every role must cite its source. No bare strings.
• Skipping validate. You MUST call liberate_design_foundation_validate before returning.
• Over-confident mapping. If two palette entries tie for primary, flag the tie in openQuestions rather than silently picking one.
• Duplicating deterministic work. Do not override the scaffold's filled slots (surface.base, text.default, etc.) unless they're demonstrably wrong.

Known scaffold-pollution failure modes (correct these every run)

The deterministic scaffold can be misled by markup that isn't the real page design. These are common enough to check explicitly — when you see them, the slot IS "demonstrably wrong," so correct it against palette.json frequency + a screenshot and note the correction in openQuestions:

• Inverted surface/text from a one-off CSS variable. The scaffold can fill color.surface.base and color.text.default from a low-frequency :root/component CSS var (e.g. --tooltip-background-color) instead of palette frequency — producing surface.base: "black" / text.default: "white" on a site that's actually white-background/dark-text. Sanity check: if surface.base is dark or text.default is light, verify against palette.json — the dominant high-count entry is the real surface; the darkest high-count entry is the real text. Swiftlumber (2026-05-28): scaffold gave black/white from a 1-url tooltip var; palette showed #ffffff×57 / #000000×3 → corrected to #fff/#000.
• 10px base font size from a hidden SEO <body>. typography.scale.base (and typography.body.fontSize) can read ~10px from a visually-hidden SEO body element. Real body copy is ~16px (check typography.scale.steps.base). Reconstruction uses captured per-element sizes, so this rarely changes output — but flag it in openQuestions so it isn't trusted downstream.
• Hidden-SEO display font. Computed h1/h2 may report a hidden-SEO face (e.g. orig_sf_pro_display_bold, wfont_*) that is NOT the visible heading face. Cross-check h3/visible headings — the real display family often surfaces there (swiftlumber: visible headings are "libre baskerville", serif from h3; h1/h2 were the SEO SF-Pro face).

Example (abbreviated)

Input partial (relevant slots only):

{
  "color": {
    "accent": { "primary": null, "warning": null, "highlight": null }
  },
  "skillTodos": ["color.accent.primary", "color.accent.warning", "color.accent.highlight"]
}

Screenshot evidence: homepage CTA button is teal; "Book a Demo" button is red; headline text has a gold→red gradient.

Output:

{
  "color": {
    "accent": {
      "primary": {
        "value": "#00a4bd",
        "role": "primary CTA, link",
        "evidence": [
          "palette[5]: 40/100 urls, 200 occurrences",
          "button.primary@homepage.html computed background"
        ]
      },
      "warning": {
        "value": "#f2545b",
        "role": "urgent CTA (Book a Demo)",
        "evidence": [
          "palette[7]: 30/100 urls",
          "a.cta-book@homepage.html computed background"
        ]
      },
      "highlight": {
        "value": "#ffc700",
        "role": "gradient highlight",
        "evidence": [
          "gradient.headlineAccent uses #ffc700 → #cc0201",
          "h1 span.highlight@homepage.html"
        ]
      }
    }
  },
  "skillTodos": []
}

Then call liberate_design_foundation_validate and return.