fix(website): keep the heading fragment on a hard-loaded docs deep link

Hard-loading `/docs/<library>/<section>/<slug>#heading` landed at the top of
the article instead of the heading. Not a dev-server artifact: measured
against a production build, the heading sat ~9760px below the reading pane.

The article is server-rendered straight into the document, so the browser
performs its native scroll to the fragment on the page scroller. Hydration
then mounts `WebsiteWorkspaceSurface`, `html:has([data-website-workspace-host])`
starts matching `overflow: hidden`, the page scroller ceases to exist, and the
real scroller — `.docs-workspace-article`, mounted with the shell — starts at
zero. No script resets anything; CSS simply discards the scroll, and nothing
puts the reader back. Shared links, search results and the heading-granular
deep links the docs search emits are all exactly this shape, and the failure
is silent: the page renders perfectly, just in the wrong place.

Re-apply the fragment once, on the mount that takes the scrolling over. The
resulting position is identical to clicking the same rail link, including the
shell's pre-existing 45px section offset.

Non-workspace docs pages were never affected — they keep the document scroller
and `html`'s `scroll-padding-top`. The one report against
`/docs/choosing-an-adapter` was a cold `next dev` compile dropping the pending
fragment scroll, which a production build does not do; the new spec covers
that page anyway so the difference stays recorded.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Author: blove

apps/website/e2e/docs-deep-link.spec.ts

@@ -0,0 +1,201 @@
+import { test, expect, type Page } from '@playwright/test';
+
+/**
+ * Hard-loading a docs URL that carries a heading fragment.
+ *
+ * This is the shape every shared link, every search-engine result, and every
+ * hit from the docs search index takes: the browser is handed
+ * `/docs/…#some-heading` cold, with no in-page click to fall back on. The
+ * failure mode is silent — the page renders perfectly, just at the top — so
+ * nothing but a positional assertion catches it.
+ *
+ * Two page kinds scroll through different machinery and both are guarded here:
+ *
+ *   - A non-workspace docs page (`/docs/choosing-an-adapter`) scrolls the
+ *     document. `html`'s `scroll-padding-top` (global.css) clears the fixed
+ *     nav, so the heading lands just below it.
+ *   - A workspace docs page (`/docs/[library]/[section]/[slug]`) does not.
+ *     `html:has([data-website-workspace-host])` is `overflow: hidden`
+ *     (styles/docs.css) and the only scroller is `.docs-workspace-article`
+ *     inside the shell — which the shell mounts during hydration, throwing
+ *     away whatever fragment scroll the browser had already performed on the
+ *     document. `WebsiteWorkspaceSurface` re-applies it; without that the
+ *     heading is stranded thousands of pixels below the reading pane.
+ *
+ * Note for anyone debugging a red run by hand: on a *cold* `next dev` server
+ * the first request for a route compiles it, `load` fires before the document
+ * is laid out, and Chrome drops the pending fragment scroll — so the very
+ * first hard load of a docs URL can look broken in dev even when the product
+ * is fine. It does not happen against a production build. Each test below
+ * visits its route once before the hard load, which pays that compile.
+ */
+
+interface DeepLinkTarget {
+  readonly id: string;
+  /** Distance from the top of the scroller's content, at scrollTop 0. */
+  readonly offset: number;
+}
+
+/**
+ * The deepest heading in the article's own rail that can actually come to rest
+ * at the top of its scroller.
+ *
+ * Deliberately read off the page rather than hardcoded: a heading id baked
+ * into this file goes stale the first time someone retitles a section, and the
+ * test then passes vacuously against a fragment that matches nothing. The
+ * reachability filter matters just as much — the last heading on a page is
+ * usually inside the final screenful, where scrolling to it hits the end of
+ * the scroll range and the assertion would fail on a page that works.
+ */
+async function findDeepLinkTarget(
+  page: Page,
+  route: string,
+  scrollRootSelector: string | null
+): Promise<DeepLinkTarget> {
+  await page.goto(route);
+  if (scrollRootSelector) {
+    // Before the shell hydrates, `.docs-workspace-article` is server-rendered
+    // straight into the page and is not yet a scroller — measuring it then
+    // yields a scroll range that has nothing to do with the one under test.
+    await expect(page.locator('[data-workspace-shell]')).toHaveAttribute(
+      'data-hydrated',
+      'true'
+    );
+  }
+  await expect(page.locator('.docs-toc-link').first()).toBeVisible();
+
+  const target = await page.evaluate((selector) => {
+    const root = selector
+      ? document.querySelector<HTMLElement>(selector)
+      : document.documentElement;
+    if (!root) return null;
+    const rootTop = selector ? root.getBoundingClientRect().top : 0;
+    const rootScroll = selector ? root.scrollTop : window.scrollY;
+    const viewport = selector ? root.clientHeight : window.innerHeight;
+    const maxScroll = root.scrollHeight - viewport;
+
+    let deepest: { id: string; offset: number } | null = null;
+    for (const link of document.querySelectorAll('.docs-toc-link')) {
+      const href = link.getAttribute('href');
+      if (!href?.startsWith('#')) continue;
+      const heading = document.getElementById(href.slice(1));
+      if (!heading) continue;
+      const offset = Math.round(
+        heading.getBoundingClientRect().top - rootTop + rootScroll
+      );
+      // Below the fold, so "it scrolled" is not vacuously true, and inside the
+      // scroll range, so the scroller can put it at its top edge.
+      if (offset > viewport && offset <= maxScroll) {
+        deepest = { id: href.slice(1), offset };
+      }
+    }
+    return deepest;
+  }, scrollRootSelector);
+
+  expect(target, `no reachable off-screen rail heading on ${route}`).not.toBe(
+    null
+  );
+  return target as DeepLinkTarget;
+}
+
+/**
+ * A real document load, not a same-document hash change.
+ *
+ * `page.goto()` to a URL that differs from the current one only by its
+ * fragment scrolls in place and never reloads — which is the case that already
+ * works, and would have made both tests here pass against the broken build.
+ */
+async function hardLoad(page: Page, url: string): Promise<void> {
+  await page.goto('about:blank');
+  await page.goto(url);
+}
+
+test.describe('docs deep links', () => {
+  test('a hard load with a heading fragment reaches the heading (non-workspace page)', async ({
+    page,
+  }) => {
+    const route = '/docs/choosing-an-adapter';
+    await page.setViewportSize({ width: 1280, height: 900 });
+    const target = await findDeepLinkTarget(page, route, null);
+
+    await hardLoad(page, `${route}#${target.id}`);
+
+    // The heading's distance below the bottom edge of the fixed nav.
+    // `scroll-padding-top: calc(var(--nav-h) + 16px)` (global.css) puts it at
+    // that 16px gutter; the band absorbs sub-pixel nav geometry without
+    // admitting a page that simply never scrolled.
+    const gapBelowNav = () =>
+      page.evaluate((id) => {
+        const heading = document.getElementById(id);
+        const nav = document.querySelector('[data-site-navigation]');
+        if (!heading || !nav) return null;
+        return Math.round(
+          heading.getBoundingClientRect().top -
+            nav.getBoundingClientRect().bottom
+        );
+      }, target.id);
+
+    // `scroll-behavior: smooth` animates the jump, so poll rather than sample.
+    // The window is generous because a loaded `next dev` server can be slow to
+    // stream the document; it only ever buys time for the *correct* resting
+    // place, so a page that lands somewhere else still fails.
+    await expect
+      .poll(gapBelowNav, {
+        message: 'heading never came to rest below the fixed nav',
+        timeout: 15_000,
+      })
+      .toBeLessThanOrEqual(32);
+
+    // ...and it stopped below the nav rather than behind it.
+    expect(await gapBelowNav()).toBeGreaterThanOrEqual(0);
+  });
+
+  test('a hard load with a heading fragment reaches the heading (workspace page)', async ({
+    page,
+  }) => {
+    const route = '/docs/langgraph/guides/streaming';
+    await page.setViewportSize({ width: 1280, height: 900 });
+    const target = await findDeepLinkTarget(
+      page,
+      route,
+      '.docs-workspace-article'
+    );
+
+    await hardLoad(page, `${route}#${target.id}`);
+    await expect(page.locator('[data-workspace-shell]')).toHaveAttribute(
+      'data-hydrated',
+      'true'
+    );
+
+    // The reading pane is the scroller here, so the heading comes to rest at
+    // its top edge — exactly where clicking the same rail link puts it.
+    await expect
+      .poll(
+        () =>
+          page.evaluate((id) => {
+            const heading = document.getElementById(id);
+            const article = document.querySelector('.docs-workspace-article');
+            if (!heading || !article) return null;
+            // Absolute distance: overshooting the heading (landing it above
+            // the pane, off screen) is as broken as never reaching it.
+            return Math.abs(
+              Math.round(
+                heading.getBoundingClientRect().top -
+                  article.getBoundingClientRect().top
+              )
+            );
+          }, target.id),
+        {
+          message: 'heading never came to rest at the top of the reading pane',
+          timeout: 15_000,
+        }
+      )
+      .toBeLessThanOrEqual(8);
+
+    // ...and the pane genuinely scrolled to get there.
+    const articleScrollTop = await page
+      .locator('.docs-workspace-article')
+      .evaluate((element) => element.scrollTop);
+    expect(articleScrollTop).toBeGreaterThan(0);
+  });
+});

apps/website/src/components/workspace/WebsiteWorkspace.tsx

@@ -294,6 +294,36 @@ function WebsiteWorkspaceSurface({
     [handleMobileModalPresenceChange]
   );
 
+  /*
+   * Re-apply the URL fragment once the shell owns the scrolling.
+   *
+   * A hard load of `/docs/…#heading` — every shared link, search result and
+   * docs-search deep link takes that shape — starts with the article rendered
+   * straight into the document, so the browser performs its native scroll to
+   * the fragment against the page scroller. Mounting this surface then makes
+   * `html:has([data-website-workspace-host])` match, which is
+   * `overflow: hidden` (styles/docs.css); the page scroller disappears, that
+   * scroll is discarded, and the real scroller — `.docs-workspace-article`,
+   * mounted with it — starts at zero. Nothing puts the reader back, so the
+   * heading they followed is left off screen with no error anywhere.
+   *
+   * This runs once, on the mount that takes the scrolling over. Later
+   * fragment navigation is same-document and the browser handles it inside
+   * the pane on its own. Guarded by e2e/docs-deep-link.spec.ts.
+   */
+  useEffect(() => {
+    const fragment = window.location.hash.slice(1);
+    if (!fragment) return;
+    let id: string;
+    try {
+      id = decodeURIComponent(fragment);
+    } catch {
+      // A malformed escape is not an element id either way.
+      return;
+    }
+    document.getElementById(id)?.scrollIntoView({ block: 'start' });
+  }, []);
+
   return (
     <ThemeProvider theme="light">
       <div className="website-workspace-host" data-website-workspace-host="">