NewCodexQA Jev Browser
A browser automation skill that uses page-observed controls and Jev choices to carry out and verify user goals.
Added to Jevfast
NewHow it uses Jev
The goal and current indexed action space form the state. Choice questions select an operation and conditionally available target; text choices come from phrases in the goal. The actor validates the selection against the observed set, acts, then checks the page.
What Jev decides
Choose the next browser actionExample answers · not a recorded Jev response · Source ↗
Question 1 · operation
YOUR APP
INSTRUCTION
Choose the next supported operation for the goal from the current page action space.
STATE
Representative source-derived page offering clickable, editable, select, and readable controls, with the current goal and recent action history. The operation Choice offers these nine supported IDs; real page observations filter unavailable actions and produce separate dynamic target questions.
JEV · CHOICE
- CLICK
- TYPE
- SELECT
- SCROLL_DOWN
- SCROLL_UP
- WAIT
- READ
- DONE
- BLOCKED
App workflow
- Observe the page
Playwright builds an indexed action space from visible controls and their supported operations.
- Choose and validate
Jev chooses an operation and relevant target or goal-provided text; the actor rejects anything outside the current action space.
- Act and check
The actor performs the action and checks the page again. Visible evidence determines whether the goal condition passes.
- Replay successful cases
Case generation records a passing path that can be replayed and verified without Jev.
Why it is interesting
It constrains decisions to controls discovered in the page and checks visible outcomes after acting, while successful generated cases can replay without a decision model.