browser-forms

Capture element refs with browser:content, inspect ambiguous fields with browser:detail, then use select_option, set_checked, type, and upload_file against those refs. Verify every file path exists before calling upload_file, and confirm submission with browser:state or a screenshot.

Prefer ref-based form actions because they are stable across layout changes. Use coordinates only when no stable ref exists or the UI is intentionally canvas-like, such as drawing surfaces or captcha-style widgets.

Check the result with browser:content or browser:state after submitting, or take an explicit browser:screenshot and load it with vision_load. Browser screenshots are not automatically added to model-visible history, so verification must be requested explicitly.

Uploads fail when the file path is guessed or does not exist on disk. Always verify that every path exists before calling upload_file, and target the file input by its ref rather than coordinates.

Yes, for visual inspection: use browser:screenshot plus vision_load when layout, canvas content, or hidden state matters. This lets the model see the rendered page, since screenshots are not loaded into history automatically.