Currently, the Anthropic Computer Use implementations (Typescript, Python coming soon) rely on Playwright's page.screenshot() functionality and passes the output to the LLM for its next instructions. However, this screenshot doesn't include the full browser window i.e. the address bar. Because of this, Claude cannot see the navbar it types into when it uses cmd-l to focus it.
We should consider using a different mechanism to take screenshots for Claude. The original reference implementation uses gnome-screenshot.