sbuilder-mcp 0.30.0 → 0.31.0

package/CHANGELOG.md

@@ -6,6 +6,23 @@ All notable changes to this project are documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [0.31.0] - 2026-09-11
+
+### Added
+- sb_import and sb_import_site now turn a set of panels with a button row into a real `tab` element instead of flattening it, as long as every panel's label can be read off the page (via `aria-controls`, a shared `data-*` pairing, or position when the button row's clickable count matches the panel count); a panel set whose labels are filled in by script rather than present in the DOM is reported as `panel-set-without-labels` and left flattened rather than guessed at.
+
+## [0.30.1] - 2026-09-11
+
+### Changed
+- sb_import and sb_import_site's default `max_images` and `max_nodes` rise from 24/300 to 60/900, so a real shop homepage's full catalogue is no longer cut off mid-band.
+- A flex or grid container that wraps is now capped at 60 columns per row instead of 12, so a wide product shelf is captured as its own row instead of being coerced into a vertical stack.
+
+### Fixed
+- sb_import and sb_import_site no longer read images from `src` alone: a lazy-loaded `data-src`/`data-original`/`data-lazy-src`/`srcset` (widest candidate) is now read as the real photo, so a lazy-loaded product gallery is no longer imported as a handful of pictures with the rest reported as missing.
+- Content hidden inside a tab panel, carousel slide or filtered grid (`display:none` beside a visible sibling of the same kind) is now revealed and captured, so a homepage that lays its catalogue out in tabs no longer loses every tab but the one shown at load.
+- `<button>` elements are no longer skipped as form controls: a button with words in it is now captured as a call-to-action or link the same way an `<a>` is, so "Add to cart"-style buttons are no longer dropped from the import.
+- The section-candidate fallback now fires whenever the chosen sections cover under half of the page's non-chrome text, not only when they capture nothing at all, so a page whose only `<section>` elements are a breadcrumb no longer imports as that breadcrumb alone.
+
 ## [0.30.0] - 2026-09-10
 
 ### Added

package/CHANGELOG.vi.md

@@ -6,6 +6,23 @@ Mọi thay đổi đáng chú ý của dự án được ghi lại trong file n
 Định dạng dựa trên [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 và dự án tuân theo [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [0.31.0] - 2026-09-11
+
+### Added
+- sb_import và sb_import_site giờ biến một bộ panel có hàng nút thành một element `tab` thật thay vì làm phẳng nó, miễn là nhãn của mọi panel đọc được trên trang (qua `aria-controls`, một cặp giá trị `data-*` chung, hoặc theo vị trí khi số phần tử bấm được trong hàng nút khớp với số panel); một bộ panel có nhãn do script điền vào chứ không có sẵn trong DOM sẽ được báo là `panel-set-without-labels` và giữ nguyên dạng làm phẳng thay vì bị đoán mò.
+
+## [0.30.1] - 2026-09-11
+
+### Changed
+- Giá trị mặc định `max_images` và `max_nodes` của sb_import và sb_import_site tăng từ 24/300 lên 60/900, nên toàn bộ catalogue trên trang chủ một shop thật không còn bị cắt cụt giữa chừng một band.
+- Một container flex hoặc grid có wrap giờ được giới hạn 60 cột mỗi hàng thay vì 12, nên một kệ sản phẩm rộng được chụp lại đúng thành một hàng thay vì bị ép thành một cột dọc.
+
+### Fixed
+- sb_import và sb_import_site không còn chỉ đọc ảnh từ `src`: một ảnh lazy-load qua `data-src`/`data-original`/`data-lazy-src`/`srcset` (chọn ứng viên rộng nhất) giờ được đọc như ảnh thật, nên một thư viện ảnh sản phẩm lazy-load không còn bị nhập vào với vài tấm còn lại bị báo là thiếu.
+- Nội dung ẩn trong một tab panel, slide carousel hay lưới lọc (`display:none` cạnh một anh em cùng loại đang hiển thị) giờ được hiện ra và chụp lại, nên một trang chủ trình bày catalogue theo dạng tab không còn mất hết các tab trừ tab đang mở lúc tải.
+- Phần tử `<button>` không còn bị bỏ qua như một control của form: một button có chữ bên trong giờ được chụp lại như một nút kêu gọi hành động hoặc liên kết, giống cách một `<a>` được xử lý, nên các nút kiểu "Thêm vào giỏ" không còn bị rớt khỏi bản nhập.
+- Cơ chế dự phòng chọn section giờ kích hoạt bất cứ khi nào các section được chọn chỉ phủ dưới một nửa nội dung chữ không phải chrome của trang, chứ không chỉ khi chúng không chụp được gì cả, nên một trang mà các phần tử `<section>` duy nhất là một breadcrumb sẽ không còn bị nhập vào chỉ như breadcrumb đó.
+
 ## [0.30.0] - 2026-09-10
 
 ### Added

package/dist/domains/site/importmap.js

@@ -167,6 +167,29 @@ function one(c, t) {
                 children: body,
             };
         }
+        case 'tab': {
+            const items = (c.children ?? [])
+                .map((item) => one(item, t))
+                .filter((n) => n !== null);
+            // Two is the floor for a tab: one panel with a button over it is a
+            // heading the visitor cannot dismiss.
+            if (items.length < 2)
+                return null;
+            return { type: 'tab', children: items };
+        }
+        case 'tab-item': {
+            const label = c.text?.trim();
+            const body = (c.children ?? [])
+                .map((k) => one(k, t))
+                .filter((n) => n !== null);
+            // NO LABEL, NO PANEL — and this is stricter than the accordion beside it
+            // on purpose. `accordion-content` seeds its own summary, so an unlabelled
+            // one still opens; `tab` builds its BUTTON ROW from these labels, so an
+            // unlabelled panel is one the visitor has no way to reach.
+            if (!label || !body.length)
+                return null;
+            return { type: 'tab-content', specials: { label }, children: body };
+        }
         case 'divider': {
             return { type: 'divider', style: { width: '100%' } };
         }

package/dist/tools/importpage.js

@@ -464,7 +464,13 @@ export function registerImportTools(server, ctx, session) {
             }
         }
         const tokens = tokenDoc ? tokensFromPage(tokenDoc.doc) : {};
-        const shots = await captureMany(plan.pages.map((p) => p.url), { maxImages: max_images ?? 24, maxNodes: max_nodes ?? 300 });
+        const shots = await captureMany(plan.pages.map((p) => p.url), 
+        // 900, not 300. The old ceiling was chosen against marketing pages;
+        // a real shop's homepage is a different quantity — ttgshop.vn measured
+        // 9,829px of catalogue across a dozen collection bands, and 300 cut it
+        // off in the middle of the third. A cap is here to stop a runaway page,
+        // not to decide how much of an ordinary one survives.
+        { maxImages: max_images ?? 60, maxNodes: max_nodes ?? 900 });
         const byUrl = new Map(shots.map((s) => [s.url, s]));
         // ONE UPLOAD PER IMAGE FOR THE WHOLE SITE, not per page. A logo, a payment
         // strip and a footer badge appear on every page of a real site, and

package/dist/vision/capture.js

@@ -74,7 +74,16 @@ function capturePage(limits) {
         // nothing, so a contact page says it had one. The CONTROLS stay ignored —
         // a stray input outside a form is chrome, and the fields of a form that IS
         // reported are counted there rather than walked into.
-        'NAV', 'INPUT', 'SELECT', 'TEXTAREA', 'BUTTON',
+        //
+        // BUTTON IS NOT HERE EITHER, AND USED TO BE — filed with the form controls,
+        // on the reasoning that fits `input` and does not fit it. A `<button>` with
+        // words in it is a CALL TO ACTION, which on a shop is the most important
+        // interactive thing on the page: MEASURED on ttgshop.vn, 27 of them dropped
+        // in one capture, every "Mua ngay" and "Thêm vào giỏ" among them. A button
+        // that really is a form control is unreachable anyway — the FORM branch
+        // returns without walking its children — so the only ones this reaches are
+        // the ones a reader would press.
+        'NAV', 'INPUT', 'SELECT', 'TEXTAREA',
     ]);
     const forms = [];
     /** The provider and id behind an embed URL, or null if this platform has no element for it. */
@@ -136,6 +145,104 @@ function capturePage(limits) {
     // a throwaway tab that is closed straight after.
     for (const d of Array.from(document.querySelectorAll('details')))
         d.setAttribute('open', '');
+    // A TAB PANEL THAT IS NOT THE OPEN ONE MEASURES AS ZERO, and on a shop that
+    // is most of the page.
+    //
+    // Same principle as the `<details>` pass above and the same sentence decides
+    // it: the source's collapsed state is not content. MEASURED on ttgshop.vn —
+    // a homepage that lays its catalogue out in tabs (PC GAMING, WORKSTATION,
+    // AMD, MINI, …) — 73% of the page's own text was in panels carrying
+    // `display:none`, so the import kept one tab's products and silently dropped
+    // every other tab: product names, prices, discounts, stock. Nothing reported
+    // it, because a hidden element is indistinguishable from an absent one once
+    // it has been skipped.
+    //
+    // REVEALED BY THE SHAPE OF A PANEL SET, never by unhiding what is hidden. A
+    // `display:none` element with no visible SIBLING is a modal, a drop-down, an
+    // off-canvas menu or a mobile-only copy of a desktop bar — showing those is
+    // how an import grows a navigation drawer in the middle of a page. What marks
+    // a panel is that it sits beside a peer of the same kind that IS shown: one
+    // tab open, the rest waiting. So the reveal needs a visible sibling, real
+    // content of its own, and a parent that is not page chrome.
+    //
+    // The display value is COPIED from that visible sibling rather than forced to
+    // `block`: a flex row of cards revealed as a block would stack, and the walk
+    // reads `display` to decide what is a row.
+    const revealPanels = () => {
+        const parents = new Set();
+        for (const el of Array.from(document.querySelectorAll('[style*="display"], .hidden, [hidden], [aria-hidden]'))) {
+            const p = el.parentElement;
+            if (p)
+                parents.add(p);
+        }
+        // Any container can hold a panel set; the attribute scan above only finds
+        // the common spellings, so the sweep below is over every parent of more
+        // than one element, bounded by the document itself.
+        for (const el of Array.from(document.querySelectorAll('*'))) {
+            if (el.children.length > 1)
+                parents.add(el);
+        }
+        for (const parent of parents) {
+            if (inPageChrome(parent))
+                continue;
+            const kids = Array.from(parent.children);
+            if (kids.length < 2)
+                continue;
+            // A VISIBLE PEER IS NOT ENOUGH, and the first version of this stopped
+            // there — which reveals every hidden thing on the page, because anything
+            // in a content flow has visible siblings. A pinned test caught it: a
+            // `<p style="display:none">` beside a visible paragraph came back as
+            // content, and that `<p>` is hidden because its author hid it.
+            //
+            // A PANEL IS A CONTAINER THAT MATCHES ITS PEER. Same tag as a sibling
+            // that IS shown, and children of its own — which is what a tab body, a
+            // carousel track or a filtered grid always is, and what a hidden
+            // paragraph, a stray span and an empty slot never are.
+            const shownByTag = {};
+            const hidden = [];
+            for (const k of kids) {
+                const cs = getComputedStyle(k);
+                const r = k.getBoundingClientRect();
+                if (cs.display === 'none')
+                    hidden.push(k);
+                else if (!shownByTag[k.tagName] && r.width > 0 && r.height > 0) {
+                    shownByTag[k.tagName] = cs.display;
+                }
+            }
+            if (!hidden.length)
+                continue;
+            // STAMPED WHILE WE STILL KNOW. Once revealed, a panel is indistinguishable
+            // from an ordinary visible child, so the split is recorded here — and
+            // ONLY the ones that were actually hidden. A first version stamped every
+            // child that shared a tag with a visible peer, which swept in the BUTTON
+            // ROW itself: it is a `<div>` with children like the panels are, so the
+            // tab came out with its own controls as one of its panes.
+            let stamped = 0;
+            for (const k of hidden) {
+                if (!shownByTag[k.tagName] || k.children.length === 0)
+                    continue;
+                k.setAttribute('data-sb-panel', '');
+                stamped += 1;
+            }
+            if (stamped >= 1)
+                parent.setAttribute('data-sb-panelset', '');
+            for (const k of hidden) {
+                // Real content only — an empty slot or a script-filled placeholder is
+                // not a panel worth showing, and revealing it costs a skip either way.
+                // Low floor on purpose: a panel whose whole content is a price —
+                // "1.490.000 VNĐ", thirteen characters — is exactly the panel a shop
+                // hides, and a threshold tuned for prose skipped it.
+                const shownDisplay = shownByTag[k.tagName];
+                if (!shownDisplay || k.children.length === 0)
+                    continue;
+                const holds = clean(k.textContent).length > 2 || k.querySelectorAll('img').length > 0;
+                if (!holds)
+                    continue;
+                const was = k.getAttribute('style') ?? '';
+                k.setAttribute('style', `${was};display:${shownDisplay} !important`);
+            }
+        }
+    };
     /**
      * What the page CALLS this icon, in its own words. A candidate, never a verdict.
      *
@@ -145,6 +252,64 @@ function capturePage(limits) {
      * decides whether this platform has one by that name — nothing in the page can
      * answer that.
      */
+    /**
+     * THE REAL FILE, wherever the page put it — and on a shop that is almost never
+     * `src`.
+     *
+     * MEASURED on ttgshop.vn: 100 images, and 84 of them carry `data-src` with no
+     * `src` at all. Those 84 are the PRODUCT PHOTOS. Reading `src` alone imported
+     * a shop with a sixth of its pictures and reported the rest as
+     * `image-without-src`, which reads like the page's fault rather than ours.
+     *
+     * Lazy loading is not an edge case, it is how the web ships images: every
+     * mainstream loader (lazysizes, lozad, and most CMS themes) parks the URL in a
+     * data attribute and fills `src` only when the image nears the viewport — and
+     * a full-page screenshot does not scroll, which is the same reason `sb_look`
+     * walks the page before it fires.
+     *
+     * `srcset` is read for its LARGEST candidate rather than its first: the
+     * platform re-encodes what it is given, so handing it the 350w thumbnail when
+     * the page also offers 1400w throws away detail nothing can recover. A
+     * `data:` URI is refused wherever it appears — it is the 1x1 placeholder the
+     * loader is displaying until the real one arrives, which is precisely the
+     * placeholder this import must not ship.
+     */
+    const LAZY_SRC = ['data-src', 'data-original', 'data-lazy-src', 'data-lazy', 'data-echo', 'data-url'];
+    const widest = (srcset) => {
+        let best = '';
+        let bestW = -1;
+        for (const part of srcset.split(',')) {
+            const bits = part.trim().split(/\s+/);
+            if (!bits[0])
+                continue;
+            const w = /^([0-9]+)w$/.exec(bits[1] ?? '');
+            const n = w ? Number(w[1]) : 0;
+            if (n >= bestW) {
+                bestW = n;
+                best = bits[0];
+            }
+        }
+        return best;
+    };
+    const realSrc = (el) => {
+        const direct = el.getAttribute('src');
+        if (direct && !direct.startsWith('data:'))
+            return direct;
+        for (const a of LAZY_SRC) {
+            const v = el.getAttribute(a);
+            if (v && !v.startsWith('data:'))
+                return v;
+        }
+        for (const a of ['srcset', 'data-srcset']) {
+            const v = el.getAttribute(a);
+            if (v) {
+                const w = widest(v);
+                if (w && !w.startsWith('data:'))
+                    return w;
+            }
+        }
+        return '';
+    };
     const iconName = (el) => {
         const use = Array.from(el.querySelectorAll('use'))[0];
         const ref = use ? use.getAttribute('href') ?? use.getAttribute('xlink:href') : null;
@@ -308,6 +473,116 @@ function capturePage(limits) {
                 skip('hidden');
                 return [];
             }
+            // A TAB SET, WHEN — AND ONLY WHEN — ITS LABELS CAN BE READ.
+            //
+            // The platform's `tab` synthesizes its whole button row from each
+            // `tab-content` child's `specials.label` (render/nodes/tab/html.go), so a
+            // tab with no labels is a stack of panels wearing a control nobody can
+            // use. That makes "are there labels" the right question AND the right
+            // discriminator: MEASURED on ttgshop.vn, the panel-set stamp also lands on
+            // a `grid-cols-5` product grid that happens to hide one card, and a label
+            // row is exactly what that grid does not have.
+            //
+            // Three pairings, strongest first, because a wrong label is worse than
+            // none — it puts one collection's name over another's products:
+            //   1. ARIA. `[role=tab][aria-controls=<panel id>]` says it outright.
+            //   2. A SHARED data-* VALUE. `<button data-id="7">` beside
+            //      `<div data-id="7">` is how most hand-rolled tab scripts wire
+            //      themselves, and it survives reordering.
+            //   3. POSITION, and only from a row that holds exactly as many
+            //      clickable items as there are panels — anything else is a guess
+            //      dressed as a match.
+            //
+            // Every panel must come out with a label or this is not a tab. ttgshop.vn
+            // itself does NOT become one: its tab buttons are empty in the DOM and
+            // filled by script, and its section titles are images. Deriving a label
+            // from the `data-url` slug was available and refused — "hot-sale" is not
+            // what the page says, and a Vietnamese slug comes back stripped of its
+            // diacritics, which is the invented-copy defect this repo reports on
+            // everybody else's seeds.
+            const tabSetOf = (host) => {
+                const kids = Array.from(host.children);
+                const revealed = kids.filter((c) => c.getAttribute('data-sb-panel') !== null);
+                if (revealed.length === 0)
+                    return null;
+                // THE BUTTON ROW IS A SIBLING OF THE PANELS, and it has to be taken out
+                // of the set before anything is counted — it is a container of the same
+                // shape, and leaving it in makes the panel count wrong and the pairing
+                // meaningless. It is the child that holds the clickable words.
+                // INNERMOST ONLY. `<li><button>Một</button></li>` is the ordinary way a
+                // tab bar is written, and counting both gives four clickables for two
+                // panels — which fails the count test that position pairing depends on,
+                // so a perfectly labelled tab bar read as unlabelled.
+                const clicksIn = (c) => {
+                    const all = Array.from(c.querySelectorAll('button, a, li, [role="tab"]')).filter((b) => clean(b.textContent).length > 0);
+                    return all.filter((b) => !all.some((o) => o !== b && b.contains(o)));
+                };
+                let bar = null;
+                for (const c of kids) {
+                    if (c.getAttribute('data-sb-panel') !== null)
+                        continue;
+                    if (clicksIn(c).length >= 2) {
+                        bar = c;
+                        break;
+                    }
+                }
+                const tagOf = revealed[0].tagName;
+                const panels = kids.filter((c) => c !== bar && c.tagName === tagOf && c.children.length > 0);
+                if (panels.length < 2)
+                    return null;
+                const clickable = bar ? clicksIn(bar) : [];
+                const labels = [];
+                for (let i = 0; i < panels.length; i += 1) {
+                    const panel = panels[i];
+                    const id = panel.getAttribute('id');
+                    let label = '';
+                    if (id) {
+                        for (const b of Array.from(host.querySelectorAll(`[aria-controls="${id}"]`))) {
+                            const t = clean(b.textContent);
+                            if (t) {
+                                label = t;
+                                break;
+                            }
+                        }
+                    }
+                    if (!label) {
+                        for (const attr of Array.from(panel.attributes ?? [])) {
+                            if (attr.name.indexOf('data-') !== 0 || attr.name === 'data-sb-panel' || !attr.value)
+                                continue;
+                            for (const b of clickable) {
+                                if (b.getAttribute(attr.name) === attr.value) {
+                                    label = clean(b.textContent);
+                                    break;
+                                }
+                            }
+                            if (label)
+                                break;
+                        }
+                    }
+                    if (!label && clickable.length === panels.length)
+                        label = clean(clickable[i].textContent);
+                    if (!label)
+                        return null;
+                    labels.push(label);
+                }
+                const children = [];
+                for (let i = 0; i < panels.length; i += 1) {
+                    const body = walk(panels[i]);
+                    if (body.length === 0)
+                        continue;
+                    children.push({ kind: 'tab-item', text: labels[i], children: body });
+                }
+                if (children.length < 2)
+                    return null;
+                taken.nodes++;
+                return { kind: 'tab', children };
+            };
+            if (el.getAttribute('data-sb-panelset') !== null) {
+                const asTabs = tabSetOf(el);
+                if (asTabs)
+                    return [asTabs];
+                skip('panel-set-without-labels');
+            }
             if (tag === 'DETAILS') {
                 const kids2 = Array.from(el.children);
                 const summary = kids2.filter((c) => c.tagName === 'SUMMARY')[0];
@@ -385,8 +660,8 @@ function capturePage(limits) {
                 return [{ kind: 'heading', level: Number(tag.slice(1)), text }];
             }
             if (tag === 'IMG') {
-                const src = el.getAttribute('src');
-                if (!src || src.startsWith('data:')) {
+                const src = realSrc(el);
+                if (!src) {
                     skip('image-without-src');
                     return [];
                 }
@@ -398,6 +673,31 @@ function capturePage(limits) {
                 taken.nodes++;
                 return [{ kind: 'image', src: abs(src), alt: clean(el.getAttribute('alt')) }];
             }
+            if (tag === 'BUTTON') {
+                // Read exactly as an `<a>` is, minus the href: painted is a call to
+                // action, unpainted is a plain control, and the variant is what stops
+                // every one of them arriving as a pink pill — the defect the anchor
+                // branch below records from a documentation sidebar.
+                const text = clean(el.textContent);
+                // A BUTTON WITH MARKUP INSIDE IT IS A CARD, not a label — the same
+                // reading the anchor branch below already applies, and taking the text
+                // whole instead COST coverage rather than gaining it: a product tile
+                // wrapped in a `<button>` came back as one run-on string with the name,
+                // the price and the discount glued together, replacing three nodes that
+                // had been captured separately. Measured: 89.7% down to 82.7% on
+                // ttgshop.vn, recovered by descending first.
+                if (el.children.length > 0) {
+                    const inner = walkChildren(el);
+                    if (inner.length > 0)
+                        return inner;
+                }
+                if (!text) {
+                    skip('button-without-words');
+                    return [];
+                }
+                taken.nodes++;
+                return [{ kind: 'button', variant: looksLikeButton(el) ? 'cta' : 'link', text }];
+            }
             if (tag === 'A') {
                 const text = clean(el.textContent);
                 // A LINK THAT IS NOT A BUTTON IS STILL A LINK. It used to contribute
@@ -526,7 +826,20 @@ function capturePage(limits) {
                 // wrapping it — so the honest translation is the stack it already reads
                 // as. Twelve is above any real row seen here and far below a content
                 // grid.
-                const ROW_MAX = 12;
+                // THE CAP IS ABOUT SLIVERS, so it belongs on the containers that make
+                // them. A NOWRAP row divides one width by its column count — that is
+                // how 279 columns became 279 slivers — but a container that WRAPS never
+                // does: it lays out as many as fit and starts a new line, which is the
+                // same thing a twenty-card product shelf wants and exactly what the
+                // source was already showing.
+                //
+                // Twelve on a wrapping shelf was a rule coercing the layout rather than
+                // reading it: a shop's collection band came back as a vertical stack of
+                // cards because it had more than a dozen. The bound that remains is the
+                // whole-import node budget, which is the honest place for "this page is
+                // enormous".
+                const wraps = grid || cs.flexWrap === 'wrap';
+                const ROW_MAX = wraps ? 60 : 12;
                 if (lays && row && kids.length >= 2 && kids.length <= ROW_MAX) {
                     taken.nodes++;
                     return [{
@@ -535,7 +848,7 @@ function capturePage(limits) {
                             // A GRID ALWAYS WRAPS — that is what a grid IS — and `flexWrap` reads
                             // `nowrap` on one because the property does not apply. Carrying that
                             // literally gave the columns nowhere to go at any width.
-                            wrap: grid || cs.flexWrap === 'wrap',
+                            wrap: wraps,
                             children: kids,
                         }];
                 }
@@ -561,6 +874,12 @@ function capturePage(limits) {
         };
         return walk(root);
     };
+    // AFTER `inPageChrome` exists, and that ordering is load-bearing: the reveal
+    // asks whether a container is page chrome, and a `const` arrow read before
+    // its own definition throws inside `evaluate`, which kills the whole capture.
+    // The same class as the closure trap this file already carries — a name that
+    // exists but is not yet initialised.
+    revealPanels();
     // SECTION CANDIDATES, widest first: a page that marks its bands up
     // semantically is read that way, and one that does not falls back to the
     // top-level children of its main content, which is what a hand-written page
@@ -614,15 +933,119 @@ function capturePage(limits) {
         return acc;
     };
     let sections = build(candidates);
-    // THE FALLBACK HAS TO FIRE ON AN EMPTY RESULT, not only on an empty candidate
-    // LIST. A page can offer `<section>` elements that hold nothing this platform
-    // renders — a wrapper around a canvas, a slot filled by script later — and the
-    // old order took "we found candidates" as "we found content", so the whole
-    // page came back empty. Measured: tailwindcss.com kept 0 of 6,004 characters
-    // while reporting one skipped empty section.
-    if (sections.length === 0) {
+    // THE FALLBACK HAS TO FIRE ON A DERISORY RESULT, not only on an empty one —
+    // and this entry has now been widened TWICE, each time by a real page.
+    //
+    // First it fired only on an empty candidate LIST, so a page offering
+    // `<section>`s that hold nothing this platform renders came back empty
+    // (tailwindcss.com: 0 of 6,004 characters, one skipped empty section). That
+    // was fixed by firing on an empty RESULT.
+    //
+    // An empty result is still the wrong test, because it treats "we captured
+    // something" as "we captured the page". MEASURED on ttgshop.vn, a shop with
+    // 2,856 divs, 602 paragraphs, 110 headings and 100 images: the whole page
+    // lives in `div.homepage`, and the document's only two `<section>` elements
+    // are a BREADCRUMB and one more. The selector privileges `<section>`
+    // absolutely, so the breadcrumb WAS the result — one section, one text node,
+    // 34 of 4,993 characters — and because that is not empty, nothing fell back.
+    // A shop imported as its own breadcrumb, with no error at any step.
+    //
+    // So the question is coverage, against the denominator this file already
+    // records as the honest one: page chrome is skipped ON PURPOSE and must not
+    // count against the result, or every correct import of a nav-heavy site would
+    // look like a failure. Below half of the non-chrome text, the candidate set
+    // was simply the wrong reading of the page, and the body walk is tried and
+    // kept only if it does better — so a page where the sections really are the
+    // content pays one comparison and keeps its own answer.
+    const textOf = (nodes) => {
+        let n = 0;
+        const walk = (c) => {
+            if (typeof c.text === 'string')
+                n += c.text.length;
+            for (const k of c.children ?? [])
+                walk(k);
+        };
+        for (const c of nodes)
+            walk(c);
+        return n;
+    };
+    let chromeChars = 0;
+    for (const el of Array.from(document.querySelectorAll('header, nav, footer'))) {
+        if (inPageChrome(el))
+            chromeChars += (el.innerText ?? '').length;
+    }
+    const contentChars = Math.max(0, (document.body.innerText ?? '').length - chromeChars);
+    // NO SIZE FLOOR ON THE CHECK. An earlier version only asked the question on
+    // pages with more than 400 characters, which is the guard you write when you
+    // fear a fallback firing too often — but the fallback cannot do harm here: it
+    // BUILDS the alternative and keeps it only if it captured more, so the worst
+    // case is one wasted walk on a page that was already right. A small page is
+    // also exactly where one stray band is the whole import.
+    if (sections.length === 0 || (contentChars > 0 && textOf(sections) < contentChars * 0.5)) {
+        // THE CONTENT ROOT, not `body` — otherwise the whole page comes back as ONE
+        // band and every arrangement the source had is gone. Measured on
+        // ttgshop.vn: `body` yields `div.homepage`, one candidate holding 9,829px of
+        // shop, so the import produced a single section with 467 text nodes in it.
+        //
+        // A real page wraps its content two or three deep before the bands start
+        // (`body > div.homepage > div.container > div.section-collection…`), so the
+        // root is found by DESCENDING while one child still holds nearly all the
+        // text. That child is a wrapper by definition — it has siblings that are
+        // scripts, chrome and empty slots — and its children are the bands.
+        // Bounded, and it stops the moment the text spreads out, which is exactly
+        // when the bands have been reached.
+        const textLen = (el) => (el.innerText ?? el.textContent ?? '').length;
+        let root = document.querySelectorAll('main')[0] ?? document.body;
+        for (let depth = 0; depth < 6; depth += 1) {
+            const kids = Array.from(root.children).filter((k) => !inPageChrome(k));
+            if (kids.length === 0)
+                break;
+            const total = textLen(root);
+            let biggest = kids[0];
+            for (const k of kids)
+                if (textLen(k) > textLen(biggest))
+                    biggest = k;
+            if (total < 200 || textLen(biggest) < total * 0.7)
+                break;
+            root = biggest;
+        }
+        // BOTH READINGS, AND THE ONE THAT COVERS MORE WINS. Descending finds the
+        // band split on a page that wraps its content; it LOSES material on a page
+        // whose content is spread across two wrappers, because everything outside
+        // the biggest one is left behind — measured here as 89.7% falling to 82.7%.
+        // Neither rule is right for every page and the comparison costs one walk,
+        // so the page decides rather than the heuristic.
+        // A SPECULATIVE BUILD MUST LEAVE NO TRACE, and the first version of this
+        // left three. `build` is not pure — it pushes every `<form>` it meets onto
+        // the shared list, counts every skip, and spends the image and node budget
+        // — so trying a second reading REPORTED a one-form page as having three and
+        // trebled every skip count. Caught by two tests that were already there,
+        // which is the argument for asserting side effects and not only results.
+        const snap = () => ({
+            forms: forms.length,
+            skipped: { ...skipped },
+            taken: { ...taken },
+        });
+        const restore = (was) => {
+            forms.length = was.forms;
+            for (const k of Object.keys(skipped))
+                delete skipped[k];
+            Object.assign(skipped, was.skipped);
+            taken.images = was.taken.images;
+            taken.nodes = was.taken.nodes;
+        };
+        const before = snap();
+        let bestState = snap();
         const main = document.querySelectorAll('main')[0] ?? document.body;
-        sections = build(Array.from(main.children));
+        for (const from of [root, main]) {
+            restore(before);
+            const wider = build(Array.from(from.children));
+            if (textOf(wider) > textOf(sections)) {
+                sections = wider;
+                bestState = snap();
+            }
+        }
+        restore(bestState);
     }
     const link = Array.from(document.querySelectorAll('link[rel="canonical"]'))[0];
     const canonical = link ? (link.getAttribute('href') ?? '') : '';

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "sbuilder-mcp",
-  "version": "0.30.0",
+  "version": "0.31.0",
   "description": "MCP server that designs and operates a Store Builder site — pages, data, theme and publish — through the platform's own API and live-edit protocol.",
   "mcpName": "io.github.vuluu2k/sbuilder-mcp",
   "type": "module",