Playjev — filedcom

Playwright keeps execution, locators, events and assertions; Jev only selects among real nodes and a fixed vocabulary from compact YAML state, with no generated selectors, JavaScript or silent fallback.

Use check for yes-or-no state with a probability, choose among caller-defined options, rate on an ordered scale, and act to score targets, execute through Playwright and verify the outcome.

Values stay in caller-supplied fields while Jev selects controls; deterministic code performs fills, checks and option selection, including whole forms in one target-selection round trip.

You need Node 20 or newer, Chromium, and a TypeSafe key in the environment. The wrapped page still exposes the normal Playwright APIs alongside the semantic methods.

Outcome-based browser evaluations adapted from public Browserbase Stagehand fixtures, plus a fixed-answer smoke suite drawn from the official WebVoyager dataset. The local run covers labelled native checkboxes, a pointer-intercepting styled radio, hidden inputs with dynamic reveal, custom dropdowns, shadow DOM, cross-origin iframes, bulk iframe forms and multiple tabs: 8 of 8 passed, with the bulk iframe form completing in 0.91s against 7.41s for five separate actions. Those are observed timings, not guaranteed benchmarks.

The page is captured through accessibility and frame nodes, normalized into sparse YAML with numbered, scoped entries carrying page url, title, and per-node parent, type, text, actionable flag and state. Internal selectors and browser node ids never appear in that YAML; they stay in the local snapshot map used to resolve Jev's numbered choice, so Jev selects among real nodes without ever receiving a raw HTML tree.

Not yet: the README marks it experimental, says the API is actively evolving as it is tested against broader browser benchmarks, and that it currently requires Chromium over CDP plus TypeSafe Jev API access. The WebVoyager check is described as a small fixed-answer integration gate using official tasks, not a score on the full 643-task benchmark, so 4 of 4 live tasks across Cambridge Dictionary, arXiv, GitHub and Wolfram Alpha should not be read as a benchmark result.