This handles browser automation inside cmux webviews using a snapshot and element reference workflow. You open a surface, verify the URL loaded, take an interactive snapshot to get fresh element refs like e1 or e2, then click or fill those refs. It supports common wait patterns for selectors, text, URL changes, and load states. The WKWebView foundation means no CDP tricks like network interception or viewport emulation, but the high-level commands cover most form filling and navigation tasks. One sharp edge: complex pages can throw js_error on snapshots, so you fall back to get text body or get html body. The skill includes templates for login flows and session state persistence, which matters when you need to stay authenticated across surfaces.
npx skills add https://github.com/manaflow-ai/cmux --skill cmux-browser