November 2026 Summaries

No summary generated yet.

Playwright MCP version 0.0.84 can indefinitely hang after a successful navigation when its post-action “settle” delay relies on a page-side setTimeout that cannot run, particularly in Chromium on pages served with a Content-Security-Policy sandbox directive such as raw.githubusercontent.com, or when the Playwright page clock is paused. The issue affects tools sharing the same completion path, including browser_evaluate, browser_click, browser_press_key, and browser_run_code_unsafe, while browser_navigate generally still returns. Testing reported that Firefox handled the CSP-sandbox case normally, although paused clocks reportedly affect all major engines, and setting the settle timeout to zero does not resolve the problem because it still schedules a timer. As of November 1, 2026, an approved but unmerged fix proposed adding a Node-side deadline so completion no longer depends solely on page timers. Until a release includes that change, suggested mitigations include applying client-side per-call deadlines, navigating away after a stalled call, using Firefox for sandboxed pages, resuming paused clocks before further MCP actions, and retrieving static content through direct HTTP requests or scraping tools rather than a browser session.

Dynamic “load more” and infinite-scroll pages often defeat static scraping because JavaScript fetches content after the initial HTML response, so extraction should first look for stable paginated URLs or authorized JSON/XHR endpoints that can be requested directly within site permissions and rate limits. When browser interaction is necessary, it should use bounded click, wait, or scroll sequences, extract structured results, deduplicate records by stable IDs or absolute URLs, and stop at verified end markers or after a batch produces no new identifiers, while treating the latter as uncertain evidence of completeness. Context.dev Scrape supports one to five ordered actions per request and does not preserve ordinary page state between requests, although paid browser profiles can retain authenticated sessions; therefore, it is best suited to limited interactions rather than adaptive, long-running browser loops. Reliable workflows validate parsed output, handle partial responses and transient errors carefully, enforce limits on pages, scrolls, attempts, and deadlines, respect rate limits and Retry-After headers, and label datasets incomplete whenever a cap is reached without a confirmed endpoint. The text also compares Context.dev with ScrapingBee’s selector-driven interaction options and Microlink’s more customizable Puppeteer-based loops, while emphasizing that explicit pagination or authorized data feeds provide the strongest basis for complete collection.