Jev browser automation with Playwright and Computer Use

Two integration paths: browser tests and desktop actions

jevcumber builds candidates from Gherkin steps and page controls, asks Jev to select an action, element and input value, then executes through Playwright. Frozen mode replays a lockfile without Jev; semantic assertions have separate limitations.

Source: jevcumber · pinned source ↗

Does Computer Use mean Jev sees screenshots?

In the reviewed Jev-cu revision, Jev receives text candidates from the accessibility tree. Codex Computer Use reads and operates the interface. This does not establish screenshot understanding by Jev or transferable speed gains across applications.

Source: Jev-cu (Jev Use) · pinned source ↗

Build a loop whose outcome can be checked

Editorial design: read a page revision and visible controls, provide finite candidates plus “no suitable target,” select, check that the target still exists, execute once, then read the new state. Record the target and outcome rather than using one stale snapshot for several actions.

Synthetic example: “Save draft” and “Publish” must remain separate candidates. Selecting a control does not replace application checks for user authorization and the expected resulting state.

How should automation reliability be evaluated?

Test first-time resolution separately from lockfile replay. A passing replay shows that recorded actions still work; it does not prove the current model would choose correctly.

Suggested edge cases include duplicate button names, a blocking dialog, changes during a request, network timeouts and an already-complete task. Track completion, wrong actions and abstentions. Stop with an explicit reason when confidence or available targets are insufficient. Neither project was run by this site.

Explore these projects next

Continue reading