Fill scripts

A skill fills its page with JavaScript that runs inside the page. There are two scripts, and both follow the same rules:

• Plain ES2019, no dependencies. No ?., ??, ||=, numeric separators, globalThis, replaceAll, .at(), Promise.allSettled, or classes with private fields. generate.fill_script.find_forbidden_syntax checks it, and CI runs node --check on the bench skill's fill.js.
• Compact JSON only. Page text in results is truncated and treated as data.

The inline setter (fast path)

generate/inline_setter.py builds a small setter that the skill carries in full in references/fill.md, with the page's selector map, so one javascript_tool call fills the page and nothing else is read first. Measured on the bench skill, this took one fill from 10,869 to 3,877 tokens compared with pointing at fill.js.

It reports one status per key: ok, partial, rejected, missing, escalate, unknown.

Each part of the setter costs a reader on every fill, so render_engine includes only what the page needs:

Part	Included when
Views walk (_VIEWS)	The page has views (steps, tabs, dialogs)
Open steps (_OPEN_FORM, _STEP_TARGET)	The form is behind a click
Layout check (_VERIFY)	The page has a measured layout
locate(key)	Some field must be done by hand

Writing values the way a framework sees them

Both scripts write values through the element prototype's own setter (put in the setter, setValue in fill.js) and click checkboxes rather than setting checked. A framework-controlled input ignores a direct el.value = …, while a naive read-back still says ok. test_setter_fills_inputs_a_framework_controls (a browser test) rebuilds that trap and fails on a direct write.

Open steps

An open step is a CSS selector or text:Label. One piece of JavaScript resolves them, _STEP_TARGET: the probe runs it through build_open_script, and the skill runs it in openForm, which clicks only while the form's first steady field is hidden, because an open step is often a toggle.

Views

_VIEWS reaches each part with stepTarget, waits for its fields, fills it, leaves it, and reports view and stopped (with the page's [role=alert] and field error texts). run(values, {"view": name}) stops in a part. The probe reaches each view with the same call (build_probe_call), filling earlier parts with example values.

The layout check

The check is optional and exists only with a probe-measured layout. The probe measures each control's box from its final selector at a 1280×900 viewport. choose_landmarks picks 3 to 8 landmark fields, one per row, from fields that are in no view and always shown; build_geometry writes them relative to the first. At fill time the setter reports layout.state:

State	Meaning	Writes?
match	Same layout	Yes
resized	Same structure at another viewport	Yes
blocked	Something covers the form	No
changed	Not the expected layout	No

A failed check writes nothing, but it must never stop the AI inspecting the page and filling with {"verify": false}. Stored positions only ever verify; anything clicked is measured live with locate(key).

fill.js (fallback)

scripts/fill.js is generate/fill_engine.js plus a generated field map at the top. The engine is the same in every skill (ENGINE_VERSION = 1); only the map differs. The AI reads it only when the setter answers escalate for some keys.

• Size is measured in bytes: MAX_SCRIPT_BYTES = 48000. The field map is written one entry per line. A size rule is never met by making lines longer.
• It walks the fields in fill order, waits until each exists and is visible (which handles conditional fields), applies the field's strategy, waits for dependent options to load, checks compiled rules locally, reads each value back, and collects errors from aria-invalid, role=alert and the text next to each field.
• It refuses to run when location.hostname differs from the skill's app-url hostname.
• validate.spec.check_fill_script checks its size and that the engine is untouched and ES2019.

Testing the scripts

uv run pytest -m browser starts the bench server and runs the scripts in Chromium for every bench case: fill, follow-up, server rejection, wrong host, missing field. It checks what the server received, and that the probe created zero records. See Testing.