Annotations for Scrollable elements in a11y-dom hybrid (#463)

* include backendDOMNodeId

* skip ax nodeId if negative

* replace role with dom tag name if none or generic

* add xpath to AXNode type

* revert unnecessary changed lines

* revert more unnecessary changed lines

* changeset

* add getScrollableElementXpaths & expose it on the window

* call browser-side scrollable elems fn, inject into observe output

* changeset

* speedup

* prettier

* prune before updating roles

* take xpath out of AXnode type

* find scrollable elems

---------

Co-authored-by: Miguel <[email protected]>

.changeset/early-tables-type.md

@@ -0,0 +1,5 @@
+---
+"@browserbasehq/stagehand": patch
+---
+
+include 'Scrollable' annotations in a11y-dom hybrid

lib/a11y/utils.ts

@@ -241,6 +241,9 @@ export async function getAccessibilityTree(
   await page.enableCDP("Accessibility");
 
   try {
+    // Identify which elements are scrollable and get their backendNodeIds
+    const scrollableBackendIds = await findScrollableElementIds(page);
+
     // Fetch the full accessibility tree from Chrome DevTools Protocol
     const { nodes } = await page.sendCDP<{ nodes: AXNode[] }>(
       "Accessibility.getFullAXTree",
@@ -249,16 +252,28 @@ export async function getAccessibilityTree(
 
     // Transform into hierarchical structure
     const hierarchicalTree = await buildHierarchicalTree(
-      nodes.map((node) => ({
-        role: node.role?.value,
-        name: node.name?.value,
-        description: node.description?.value,
-        value: node.value?.value,
-        nodeId: node.nodeId,
-        backendDOMNodeId: node.backendDOMNodeId,
-        parentId: node.parentId,
-        childIds: node.childIds,
-      })),
+      nodes.map((node) => {
+        let roleValue = node.role?.value || "";
+
+        if (scrollableBackendIds.has(node.backendDOMNodeId)) {
+          if (roleValue === "generic" || roleValue === "none") {
+            roleValue = "scrollable";
+          } else {
+            roleValue = roleValue ? `scrollable, ${roleValue}` : "scrollable";
+          }
+        }
+
+        return {
+          role: roleValue,
+          name: node.name?.value,
+          description: node.description?.value,
+          value: node.value?.value,
+          nodeId: node.nodeId,
+          backendDOMNodeId: node.backendDOMNodeId,
+          parentId: node.parentId,
+          childIds: node.childIds,
+        };
+      }),
       page,
       logger,
     );
@@ -360,6 +375,69 @@ export async function getXPathByResolvedObjectId(
   return result.value || "";
 }
 
+/**
+ * `findScrollableElementIds` is a function that identifies elements in
+ * the browser that are deemed "scrollable". At a high level, it does the
+ * following:
+ * - Calls the browser-side `window.getScrollableElementXpaths()` function,
+ *   which returns a list of XPaths for scrollable containers.
+ * - Iterates over the returned list of XPaths, locating each element in the DOM
+ *   using `stagehandPage.sendCDP(...)`
+ *     - During each iteration, we call `Runtime.evaluate` to run `document.evaluate(...)`
+ *       with each XPath, obtaining a `RemoteObject` reference if it exists.
+ *     - Then, for each valid object reference, we call `DOM.describeNode` to retrieve
+ *       the element’s `backendNodeId`.
+ * - Collects all resulting `backendNodeId`s in a Set and returns them.
+ *
+ * @param stagehandPage - A StagehandPage instance with built-in CDP helpers.
+ * @returns A Promise that resolves to a Set of unique `backendNodeId`s corresponding
+ *          to scrollable elements in the DOM.
+ */
+export async function findScrollableElementIds(
+  stagehandPage: StagehandPage,
+): Promise<Set<number>> {
+  // get the xpaths of the scrollable elements
+  const xpaths = await stagehandPage.page.evaluate(() => {
+    return window.getScrollableElementXpaths();
+  });
+
+  const scrollableBackendIds = new Set<number>();
+
+  for (const xpath of xpaths) {
+    if (!xpath) continue;
+
+    // evaluate the XPath in the stagehandPage
+    const { result } = await stagehandPage.sendCDP<{
+      result?: { objectId?: string };
+    }>("Runtime.evaluate", {
+      expression: `
+        (function() {
+          const res = document.evaluate(${JSON.stringify(
+            xpath,
+          )}, document, null, XPathResult.FIRST_ORDERED_NODE_TYPE, null);
+          return res.singleNodeValue;
+        })();
+      `,
+      returnByValue: false,
+    });
+
+    // if we have an objectId, call DOM.describeNode to get backendNodeId
+    if (result?.objectId) {
+      const { node } = await stagehandPage.sendCDP<{
+        node?: { backendNodeId?: number };
+      }>("DOM.describeNode", {
+        objectId: result.objectId,
+      });
+
+      if (node?.backendNodeId) {
+        scrollableBackendIds.add(node.backendNodeId);
+      }
+    }
+  }
+
+  return scrollableBackendIds;
+}
+
 export async function performPlaywrightMethod(
   stagehandPage: Page,
   logger: (logLine: LogLine) => void,

lib/dom/global.d.ts

@@ -36,5 +36,6 @@ declare global {
       width: number;
       height: number;
     }>;
+    getScrollableElementXpaths: (topN?: number) => Promise<string[]>;
   }
 }

lib/dom/process.ts

@@ -11,18 +11,24 @@ export function isTextNode(node: Node): node is Text {
   return node.nodeType === Node.TEXT_NODE && Boolean(node.textContent?.trim());
 }
 
-function getMainScrollableElement(): HTMLElement {
+/**
+ * Finds and returns a list of scrollable elements on the page,
+ * ordered from the element with the largest scrollHeight to the smallest.
+ *
+ * @param topN Optional maximum number of scrollable elements to return.
+ *             If not provided, all found scrollable elements are returned.
+ * @returns An array of HTMLElements sorted by descending scrollHeight.
+ */
+export function getScrollableElements(topN?: number): HTMLElement[] {
+  // Get the root <html> element
   const docEl = document.documentElement;
-  let mainScrollable: HTMLElement = docEl;
-
-  // 1) Compute how “scrollable” the root <html> is
-  //    i.e. total scrollHeight - visible clientHeight
-  const rootScrollDiff = docEl.scrollHeight - docEl.clientHeight;
 
-  // Keep track of the “largest” scroll diff found so far.
-  let maxScrollDiff = rootScrollDiff;
+  // 1) Initialize an array to hold all scrollable elements.
+  //    Always include the root <html> element as a fallback.
+  const scrollableElements: HTMLElement[] = [docEl];
 
-  // 2) Scan all elements to find if any <div> has a larger scrollable diff
+  // 2) Scan all elements to find potential scrollable containers.
+  //    A candidate must have a scrollable overflow style and extra scrollable content.
   const allElements = document.querySelectorAll<HTMLElement>("*");
   for (const elem of allElements) {
     const style = window.getComputedStyle(elem);
@@ -33,25 +39,44 @@ function getMainScrollableElement(): HTMLElement {
 
     if (isPotentiallyScrollable) {
       const candidateScrollDiff = elem.scrollHeight - elem.clientHeight;
-      // Only pick this <div> if it has strictly more vertical “scrollable distance” than our current best
-      if (candidateScrollDiff > maxScrollDiff) {
-        maxScrollDiff = candidateScrollDiff;
-        mainScrollable = elem;
+      // Only consider this element if it actually has extra scrollable content
+      // and it can truly scroll.
+      if (candidateScrollDiff > 0 && canElementScroll(elem)) {
+        scrollableElements.push(elem);
       }
     }
   }
 
-  // 3) Verify the chosen element truly scrolls
-  if (mainScrollable !== docEl) {
-    if (!canElementScroll(mainScrollable)) {
-      console.log(
-        "Stagehand (Browser Process): Unable to scroll candidate. Fallback to <html>.",
-      );
-      mainScrollable = docEl;
-    }
+  // 3) Sort the scrollable elements from largest scrollHeight to smallest.
+  scrollableElements.sort((a, b) => b.scrollHeight - a.scrollHeight);
+
+  // 4) If a topN limit is specified, return only the first topN elements.
+  if (topN !== undefined) {
+    return scrollableElements.slice(0, topN);
   }
 
-  return mainScrollable;
+  // Return all found scrollable elements if no limit is provided.
+  return scrollableElements;
+}
+
+/**
+ * Calls getScrollableElements, then for each element calls generateXPaths,
+ * and returns the first XPath for each.
+ *
+ * @param topN (optional) integer limit on how many scrollable elements to process
+ * @returns string[] list of XPaths (1 for each scrollable element)
+ */
+export async function getScrollableElementXpaths(
+  topN?: number,
+): Promise<string[]> {
+  const scrollableElems = getScrollableElements(topN);
+  const xpaths = [];
+  for (const elem of scrollableElems) {
+    const allXPaths = await generateXPaths(elem);
+    const firstXPath = allXPaths?.[0] || "";
+    xpaths.push(firstXPath);
+  }
+  return xpaths;
 }
 
 export async function processDom(chunksSeen: Array<number>) {
@@ -80,7 +105,8 @@ export async function processDom(chunksSeen: Array<number>) {
 export async function processAllOfDom() {
   console.log("Stagehand (Browser Process): Processing all of DOM");
 
-  const mainScrollable = getMainScrollableElement();
+  const mainScrollableElements = getScrollableElements(1);
+  const mainScrollable = mainScrollableElements[0];
 
   const container =
     mainScrollable === document.documentElement
@@ -481,7 +507,7 @@ window.restoreDOM = restoreDOM;
 window.createTextBoundingBoxes = createTextBoundingBoxes;
 window.getElementBoundingBoxes = getElementBoundingBoxes;
 window.createStagehandContainer = createStagehandContainer;
-
+window.getScrollableElementXpaths = getScrollableElementXpaths;
 const leafElementDenyList = ["SVG", "IFRAME", "SCRIPT", "STYLE", "LINK"];
 
 const interactiveElementTypes = [