match-section

name	match-section
description	Reach VISUAL parity for ONE section by guiding AI to apply the captured design data (colors, padding, margins, full-bleed, font sizes, line-heights, families, alignment) to that section's WordPress block markup, then VERIFYING by looking at the source section and the built section side-by-side and iterating until they are visually identical. Non-deterministic on purpose — deterministic emission gets the structure right but misses visual parity; this skill closes the gap with an AI eyes-on loop. The per-section EXECUTOR called by `match-page` (which runs the batch Phase-1 assessment first, then dispatches this per divergent section). Do not drive whole-page parity by hand-rolling per-section work — run `match-page` so the batch assessment happens first. Dispatched per section (subagent) by match-page/replicate-with-blocks/design-qa. Not user-invocable directly.
disable-model-invocation	true
allowed-tools	["Bash","Read","Edit","Write","mcp__plugin_playwright_playwright__browser_navigate","mcp__plugin_playwright_playwright__browser_resize","mcp__plugin_playwright_playwright__browser_take_screenshot","mcp__plugin_playwright_playwright__browser_evaluate"]

What you do

You make ONE section of the replica look visually identical to the same section on the source site. You are given everything the pipeline captured about that section. Use it. Then you LOOK at both — source section and your built section — and you do not stop until they match. Deterministic code already ran and got the structure roughly right; your job is the visual gap it cannot close.

This is an eyes-on loop, not a one-shot emit. Apply → render → look at both images → name every difference → fix → render → look again. Move to "done" ONLY when the two images are visually identical (allowing for font-substitution and text-reflow noise — NOT for wrong color, wrong size, wrong spacing, wrong alignment, dropped media, or wrong width).

Inputs the orchestrator gives you

• outputDir (e.g. ~/Studio/_liberations/example.com), studioSitePath (e.g. ~/Studio/example-com), themeSlug, the page slug + sourceUrl, and the section index i within that page.
• The replica preview URL (e.g. http://localhost:8881).

Everything else you READ yourself from the captured data below.

The refine report — your audit trail (MANDATORY)

You own ONE file: <outputDir>/refine/<slug>/<sectionIndex>.json. Write it atomically (write <file>.tmp, then rename). Shape:

{
  "schema": 1, "slug": "<page slug>", "sourceUrl": "<source URL>", "index": <sectionIndex>,
  "findings": [{ "id": "kebab-stable-id", "region": "human label", "severity": "high|medium|low", "description": "design shows X, replica shows Y", "block_change": "wp:paragraph → wp:heading level=2 (or null)", "style_change": "fontSize → preset:large (or null)", "affects_layout": true }],
  "applied": [{ "id":    
        

Work in three phases:

Phase 1 — DIAGNOSE (no edits). Look at source vs built side-by-side. Write EVERY visible difference as a finding — wrong color, size, spacing, alignment, dropped media, wrong width, layout overrun. Ids are kebab-case, unique, stable on re-runs (hero-heading-size, not finding-1). Populate block_change whenever a block swap/attribute change is implied, style_change whenever a measurable style value is implied — a finding with neither is acceptable only for free-form content differences. Write the report file with applied: [], skipped: [] BEFORE touching markup.

Phase 2 — FIX (the eyes-on loop, unchanged). Apply → render → look at both → iterate, exactly as described below.

Phase 3 — ACCOUNT. Update the report: every finding id lands in exactly one of applied (with a summary of the actual change) or skipped (with a one-sentence reason — "within font-substitution tolerance" is legal; silence is not). Then call liberate_refine_report {outputDir, slug}. If it fails, fix the accounting (or the section) — a section with unaccounted findings is NOT done.

The captured data — this is your fuel. Read ALL of it for the section.

For section index i of the page:

1. outputDir/sections/<slug>.json → sections[i] — the SectionSpec. Read every field:
   • backgroundColor (rgb/rgba), gradient
   • layout.padTopPx / layout.padBottomPx / layout.padding / layout.gap / layout.containerWidth / layout.columnCount
   • fullBleed (true → section spans the viewport edge-to-edge)
   • headings[], headingSizes[] (px), headingLineHeights[], headingFamilies[]
   • bodyText[], bodyTextSizes[], bodyLineHeights[], bodyFamilies[]
   • buttonLabels[] / buttons[], images[] (url, width, height, alt), cells[], textAlign, mediaLayout
   • styledHtml — a snapshot of the source subtree with every computed style inlined. This is your FIDELITY ORACLE: when in doubt about any value (a padding, a color, a font-size on a specific element), read it straight out of styledHtml. It is the source's exact rendered design.
2. outputDir/design-foundation.json — the theme's semantic tokens (color roles, type families, spacing). Prefer a token slug when the captured value maps to one; use an explicit value when it does not.
3. outputDir/theme/theme.json — the registered fontFamilies slugs, fontSizes, color.palette slugs you can reference.
4. Source section image — the cropped source screenshot for this section (see "Looking", below).
5. <outputDir>/asset-triage.json (when present) — decoration entries for your section explain imagery that was deliberately NOT emitted — the description tells you what the visual was. Reproduce its intent STRUCTURALLY (wp:separator, parent border, background color/gradient) per styling-priority.md; do not re-add the image, and do not flag its absence as a parity gap.

If a value is missing from the spec, get it from styledHtml. Do not guess.

Captured px are measured at the desktop capture width (≈1440). A FIXED dimension (a cover minHeight, a fixed section height) copied verbatim will be too tall at a narrower compare width — convert proportional dimensions to vw (e.g. a 733px hero at 1440 → 51vw). Padding/font-size carried as clamp() already scale; fixed minHeight/height do not. Always compare at the SOURCE screenshot's real width so px line up.

Lazy images: below-fold cover/section images decode lazily — a screenshot taken before they load shows a GREY block (looks like a "lost background"). Scroll the whole page to force-load (the scripts/_shot.ts helper does this with mouse.wheel, not page.evaluate) BEFORE judging a missing/grey background, or you will chase a phantom.

Apply it to the block markup — core attributes only

The section's blocks live in the page's post_content and in outputDir/theme/theme/patterns/page-<slug>.php. Edit the markup for THIS section. Map each captured axis to a CORE block attribute — these survive @wordpress/blocks canonicalization; inline <figure style> and ad-hoc wrapper styles do NOT (they get stripped — [[project_block_fixer_canonicalization_constraint]]):

Captured	Apply as
heading/body font-size	style.typography.fontSize (px or a clamp() whose max = captured px)
line-height	style.typography.lineHeight (carry it WITH the size — never split them)
font-family	fontFamily slug (display/body/…); match the source face
font-weight / letter-spacing / transform / align	style.typography.fontWeight/.letterSpacing/.textTransform, textAlign
section padTopPx/padBottomPx, padding	style.spacing.padding on the section group
inter-band margin	style.spacing.margin
gap between columns/items	style.spacing.blockGap
backgroundColor	backgroundColor slug or style.color.background
text color	textColor slug or style.color.text
fullBleed = true	align:full on the section group + zero side padding
not full-bleed	constrained (content width from layout.containerWidth)
columnCount	a core/columns with that many core/column (do NOT flatten)
cover image focal point	inline style="object-position:X% Y%" DIRECTLY on the <img> tag — NOT focalPoint block JSON. Static-HTML patterns bypass server-side cover rendering, so focalPoint JSON does nothing; the inline object-position is what actually moves the crop.

Styling decisions: follow skills/replicate-with-blocks/styling-priority.md — the preset→patch→instance→variation→layout→CSS cascade, the structured-props cheat sheet, and the hard bans (no raw style="" attrs, no invented className CSS hooks). Native blocks only; core/html islands are exempt.

Existing block style variations — inventory first. Before emitting ANY new style, list <studioSitePath>/wp-content/themes/<themeSlug>/styles/blocks/*.json. Each file is a registered variation (slug, title, blockTypes, styles). When one matches what the section needs, REUSE it: apply is-style-<slug> on the block. NEVER redeclare an existing slug, never invent an is-style-* class with no backing file. New variations follow styling-priority.md option 5 (recurring + nameable only) and are written as a new styles/blocks/<slug>.json file (slug lib- prefixed) plus the class on each claiming block.

Preserve ALL content — every heading, paragraph, list item, button, image the source shows ([[feedback_never_lose_source_content]]). Never drop an image to "simplify."

Render your change

After editing the section's markup:

1. Write the updated full post_content for the page to a temp file, then push it: studio wp --path <studioSitePath> post update <pageId> --post_content="$(cat /tmp/section-content.html)" (Get <pageId>: studio wp --path <studioSitePath> post list --post_type=page --fields=ID,post_name --format=csv.) Mirror the same block edit into outputDir/theme/theme/patterns/page-<slug>.php and the installed copy under <studioSitePath>/wp-content/themes/<themeSlug>/patterns/ so the theme + editor stay in sync.
2. Flush: studio wp --path <studioSitePath> cache flush.

Looking — the gate. You MUST do this every iteration.

1. Source section image. The full-page source screenshot is outputDir/screenshots/desktop/<slug>.png. FIRST read its real width (identify -format "%w" <src>) — it is often ~1008px, NOT 1440. Crop to this section's vertical band using the spec's top + height (convert <src> -crop WxH+0+TOP +repage outputDir/screenshots/src-section-<i>.png, width = the image's real width). Staging note: write all crops/screenshots under outputDir/screenshots/ — /tmp is OUTSIDE the Playwright MCP allowed roots and will be rejected.

2. Built section image. Resize the Playwright viewport to the SOURCE screenshot's real width (e.g. 1008×900, not 1440) so the two are pixel-comparable — a width mismatch changes responsive crops and would create phantom "differences." Navigate to <previewUrl>/<slug>/ (cache-bust with ?v=<n>; for the front page the content may be at <previewUrl>/), then crop the rendered section to its bounding box (browser_evaluate the element's rect → convert crop) and save under outputDir/screenshots/.

   A composition difference that comes from the SOURCE's pre-cropped CDN asset (e.g. a Wix image served pre-cropped at a width you don't have the exact pixels for) is NOT a markup bug — note it as residual, don't chase it with markup changes.

3. Read BOTH images (Read tool on the two PNGs). Put them next to each other in your reasoning. Diff-image hygiene: a .diff.png / .padded.png marks WHERE pixels differ — never read a color, font, or size from it. Red is the difference marker; magenta (in .padded.png) means one side has no content at that location (layout overrun — high severity). Read every actual value from the SOURCE screenshot.

4. Name every difference — bluntly, itemized ([[feedback_honest_visual_assessment]]): background color? section padding (too tight / too loose)? heading size? line-height (cramped/overlapping)? body font wrong face? alignment (centered vs left)? full-bleed vs boxed? column count flattened? missing image? button style?

5. If ANY real difference remains → fix it in the markup (map back to the table above, pulling exact values from the spec / styledHtml), re-render, re-look. Loop.

6. Only when the two images are visually identical (modulo font-substitution + minor text reflow) do you finish this section.

Do not rationalize a miss. "Close enough", "known gap", "the source is just different" are not finishes — fix it or, if you genuinely cannot (a true WP rendering constraint), report it explicitly with both images, do not silently ship it.

Repair order — attack diffs in this sequence

When you name multiple differences, fix them in this order. Each rung's diffs usually CAUSE the rungs below it (wrong macro layout shifts everything downstream, so fixing a font-size first just gets undone) — climbing in order converges; jumping around oscillates.

1. Content — is everything the source shows actually present? A dropped image/heading/list-item/button is the first thing to restore ([[feedback_never_lose_source_content]]). Never proceed past a content gap.
2. Semantics — right block for the role (heading is a heading, list is a list, columns are columns — not a flattened stack). Wrong structure can't be styled into parity.
3. Macro layout — section width (full-bleed vs constrained), column count, media-vs-text order, major padding bands. The page's vertical rhythm rides on these.
4. Responsive — does it hold at the compare width AND reflow sanely narrower? Fixed dimensions that only work at 1440 belong here (see scale, below).
5. Editor drift — if an editor-report.json / editorScore surface flags this page, the markup renders differently in the block editor than on the frontend (block recovery, save-validation drift). Reconcile both surfaces — a block that's invalid in the editor is not done even if the frontend looks right.
6. Scale — per-axis sizing: font-size, line-height (always carried WITH the size), letter-spacing, gap. Convert fixed minHeight/height captured at ≈1440 to vw so they scale.
7. Polish — color exactness, focal point, button treatment, the last sub-pixel spacing.

Anti-gaming rules — what "done" is NOT

The diff score is a proxy for fidelity, not the goal. A change that improves the number while degrading the truth is a regression, not progress:

• Never hide a source-visible element to improve the diff. Deleting/display:none-ing content that the source shows lowers the pixel delta and destroys fidelity — the exact opposite of the job. If an element is hard to match, match it; do not make it disappear.
• A repair that regresses another surface reverts. If closing the frontend diff worsens the editor surface (or mobile, or another section), it is not a fix — back it out and find one that holds on every surface. Don't trade one surface's score for another's.
• Blocked ≠ done. A genuine WP rendering constraint you cannot close is REPORTED (RESIDUAL, with both images + the reason), never silently shipped as MATCHED. "Blocked" and "matched" are different statuses — conflating them launders a gap into a pass.

Output

Report: the section, the iterations you ran, the diffs you closed each round (before→after), the final source-vs-built crops, and STATUS = MATCHED (visually identical) / RESIDUAL (list the exact remaining differences + why you could not close them — for the operator). Never report MATCHED without having looked at both images in the final iteration.

Stay generic

Derive every value from THIS section's captured data. Never hardcode a site's colors, fonts, filenames, copy, or section order ([[feedback_scripts_must_be_site_agnostic]]). The same loop must match any platform's section.