Score candidate values and actions for a form element with the published CUA-S1-FORMS checkpoint, or plan and execute a bounded form workflow in an exact Chrome tab through direct CDP. The source document must already be parsed into candidate entities.
Install
npx skillscat add woahwhattheheck/commons/cua-s1-forms Install via the SkillsCat registry.
CUA-S1-FORMS in Commons
Use the hosted Commons routes when the caller must continue while this laptop or Codex is offline. POST https://commons-spark-mcp.vercel.app/cua-s1 takes {"context":"...","options":["...","..."]} and returns a model-selected index and per-option probabilities without browser action. The public Commons MCP tool is cua_s1_score with the same input. POST https://commons-spark-mcp.vercel.app/cua-s1/form takes an exact public url, form_title, and up to 29 entities of {label,value}. It runs an isolated cloud Chromium and previews by default; execute:true allows fill/check after a complete-plan preflight, and submit:true additionally allows one exact Submit click. Verify production availability through the live GET routes and a preview, as described in the integration guide.
Run python host/cua_s1_forms.py --checkpoint <cua-s1-forms.safetensors> --input <request.json> from the Commons root. The matching cua-s1-forms.json sidecar must be next to the safetensors file. Without --input, the command reads one JSON request from stdin. See the integration guide for the request and output format.
Install the upstream cua-s1 Python package from https://github.com/trycua/cua/tree/main/libs/cua-s1/python with its declared Python, PyTorch, NumPy, and safetensors requirements. Download both checkpoint files from https://huggingface.co/cua-ai/cua-s1-forms. Review upstream revision and package dependencies before installing. The published model uses PyTorch and safetensors; this skill does not use llama.cpp.
Supply the exact form element context and explicit options. The model chooses only among those options and returns their probabilities. Treat the selected index as a proposed decision. Inspect the form state and confirm outcomes independently. This command never performs form actions and always returns executed: false.
For an observed browser form, use python -m host.cua_s1_run --url <exact-open-tab-url> --checkpoint <checkpoint> --input <form.json> from the Commons root. form.json contains a form_title and an entities array of {label,value} objects. The default run previews decisions without action. Add --execute to fill/check, and --submit only to allow one recognized Submit click. The runner needs the optional playwright Python package and direct Chrome CDP endpoint; it never uses the Chrome extension. See the integration guide for commands, boundaries and output.
Commons peers may also discover cua_s1_form through the existing shared equipment catalog. Pass the same exact open tab URL, form title, and entity array. The host resolves the official cached checkpoint by default; execute and submit are separate optional booleans. Read the returned plan and action observations before reporting an outcome.
The published model is a narrow research checkpoint trained on synthetic forms with a small real-demo evaluation. It does not establish reliability on arbitrary production forms. The portable Cua Driver contract lacks set_value, while this Chrome adapter supplies token-based value mutation and reads the result after each action. Do not infer that scoring alone executed or submitted anything.