Jev Browser — Ying-Kai-Liao

Page reading moves into a single ~300 ms System One request that answers which element, which action, which value, and whether the step is done, blocked, erroring, or irreversible in one round. Code still describes the page and Playwright still performs the action. The calling LLM reads page content only when it chooses to take over, which is where most of the token saving comes from.

Jev never writes text; it answers only with probability distributions over options you supply. Everything free-form comes from the caller as candidates, which is why browser_do takes a values argument next to the goal. A value that is not among the candidates cannot appear, and every other decision is one of yes/no, pick-one, or a rating.

The two misses were counting (add until 3, flagged likely_done) and verifying a sort. The README also names three classes of work to write steps around: ordered sub-goals inside one step, open-ended goals like scroll to load more, and judgments that compare many values. Split those per step or verify them yourself with browser_check or the numbered element list.

Across roughly 200 rounds it flagged exactly one irreversible control, the place-order button on saucedemo. When the next action, or a confirm dialog it opened and dismissed, looks irreversible (order, pay, send, delete), browser_do returns needs_confirmation with a pending record, and you re-call with allow_irreversible only when you actually want that action taken.

Because a snapshot is the act-on-the-page surface, not a reading surface. browser_read extracts the whole document, cuts it at headings and sections, and has Jev rate each block against your question in batches of forty, returning only the blocks that help inside a character budget. On a long article it returned about twelve thousand of one hundred seventy-eight thousand characters in a few seconds, and below an eight-thousand-character floor it returns the page whole without calling Jev at all.

Done, likely-done when the page looks finished but Jev is unsure, needs-login, needs-confirmation for an irreversible next step or dialog, error when the page shows one after the last action, stuck or max-actions with no progress, ambiguous when the target confidence is low, and blocked for a captcha or denial page. Each one comes with supporting fields the caller is meant to act on.

Substantially, using pure formatting options that cost no model call. On a measured Wikipedia page, a default snapshot of about 27,000 characters dropped to under 10,000, a filtered snapshot reached roughly 800, and a diff taken after scrolling reached about 50. The options narrow what is rendered rather than changing what is collected.