Greptile Summary

This PR fixes a critical limitation in Stagehand where scrolling actions could not be performed inside iframes. The changes address two specific technical issues:

1. Execution Context Problem: The original implementation used page.evaluate for all scrolling operations, which meant scrolling could only occur at the page level and not within iframe contexts. The fix refactors three key scrolling functions (scrollToNextChunk, scrollToPreviousChunk, and scrollElementToPercentage) in actHandlerUtils.ts to use locator.evaluate instead, leveraging the locator's inherent iframe-awareness.

2. Same Process Iframe (SPIF) Context Resolution: For SPIFs that share the same Chrome DevTools Protocol (CDP) session as the root document, the code wasn't properly specifying execution contexts when resolving XPaths for scrollable elements. This caused resolveObjectIdForXPath to only search in the root document context, missing elements inside SPIFs. The solution introduces a new getFrameExecutionContextId function in utils.ts that creates isolated worlds for SPIFs, ensuring XPath evaluation happens in the correct execution context.

The changes integrate seamlessly with Stagehand's existing iframe handling architecture by building upon Playwright's locator system rather than introducing new iframe traversal logic. A new evaluation test iframe_scroll has been added to validate the functionality, testing that the act method can successfully scroll within a same-process iframe and verify the scroll position.

Confidence score: 3/5

• This PR addresses real functionality gaps but has several implementation concerns that could cause issues
• Score reflects solid architectural improvements but potential reliability issues in the test implementation and some complex CDP context handling
• Pay close attention to evals/tasks/iframe_scroll.ts for brittle frame access patterns and lib/a11y/utils.ts for CDP context management complexity

_{4 files reviewed, 3 comments}

_{Edit Code Review Bot Settings | Greptile}