Spaces:
Running
Running
| """The family descriptions the planner agent and the router use. | |
| FAMILY_DEFS defines each visuomotor family for the planner agent (miniwebagent.planner_agent): its prompt lists them, and | |
| its delegate action picks one. FAMILY_HINTS is the short form the oracle router's done-gating note quotes | |
| (miniwebagent.evaluate.router_note). The screenshot planner and the subtask routers that also lived here were removed on | |
| 2026-10-06 (archived in MiniWeb-archive/cleanup-20261006/removed_legacy_routing.py). | |
| """ | |
| from __future__ import annotations | |
| FAMILY_HINTS = { | |
| "Drag & gesture": "dragging, sliders, drawing or signing on a canvas, reordering by drag, panning/zooming a map, scrubbing media", | |
| "Discrete selection": "clicking choices: filters, dropdowns, sorting, date ranges, ratings, reactions, toggles (follow, save), deleting rows, playing media", | |
| "Text entry": "typing free text: search, messages, comments, spreadsheet cells, calculators, translation, code, typed signatures", | |
| "Form transaction": "filling and submitting a multi-field form: create, edit, book, pay, check out, sign in, share, cancel, configure, compare", | |
| "Navigation": "moving to a page: following menus or links, routes/directions, joining a meeting, carrying information to another site", | |
| "Out-of-page I/O": "moving data in or out of the page: upload, export/download, copy to clipboard, save a file, image editing, a code from another app", | |
| "Reasoning base": "reading the page to report an answer: find, count, compare or compute values", | |
| } | |
| # The family definitions (user, 2026-09-30: "a prompted planner ... explain and define each family in the prompt"). | |
| # Written once, before any WebArena run with them, and not tuned on WebArena. | |
| FAMILY_DEFS = { | |
| "Navigation": "getting to the page where the work happens: opening a menu, tab, section or link, opening an item's " | |
| "own page from a list, going back. The step is finished when that page is open.", | |
| "Text entry": "typing into one text field and submitting it: a search box, a message or comment, a cell, a calculator. " | |
| "Searching for an item by name, number or keyword is Text entry.", | |
| "Discrete selection": "choosing among options already on the page: dropdowns, checkboxes, radio buttons, sort orders, " | |
| "filter chips, date ranges, ratings, reactions, toggles (follow, save, like), deleting a row, " | |
| "playing media.", | |
| "Form transaction": "filling in several fields of one form and submitting it to create or change something: posting, " | |
| "creating, editing, booking, paying, checking out, signing in, sharing, configuring.", | |
| "Drag & gesture": "an action that needs the pointer dragged: a slider handle, drag-and-drop, reordering by drag, " | |
| "drawing or signing, panning or zooming a map by drag. Only when the task needs a drag.", | |
| "Out-of-page I/O": "moving data into or out of the page: exporting or downloading a file, uploading a file, copying to " | |
| "the clipboard, saving or printing.", | |
| "Reasoning base": "reading what is on the screen and answering: finding, counting, comparing or computing values and " | |
| "reporting the result. It does not change anything on the site; it may scroll to read.", | |
| } | |