Experimental Jev browser

Use this skill only when the user explicitly requests a Jev browser experiment. The regular browser skill remains the default and owns harness, driver, identity, authorization, focus, and cleanup policy. Jev adds a bounded decision loop for an owned page in Orca or an isolated agent-browser session; it does not select or create the browser identity.

This suspended experiment requires the utils-agent plugin and its browser skill. The links below resolve in the shared marketplace layout. If your client caches plugins separately, locate utils-agent:browser in its skill catalog and resolve policy references from that skill directory. Read that policy before running Jev; keep shared policy and helpers in utils-agent.

Run a trial

1. Read the regular browser skill and its policy. Follow their rules for the chosen harness and driver. Use Jev only with Orca or standalone agent-browser; if the regular skill selects another driver, continue with that skill instead.
2. Read the Jev trial reference. It defines job fields, runtime and focus requirements, handoffs, and qualification limits. Choose dynamic discovery for authorized, reversible work with ordinary controls. Choose explicit controls when the action list must be narrower. Keep consequential writes, credential entry, login, CAPTCHA, unsupported widgets, and visual judgment in the parent browser workflow.
3. Give the runner an owned page, exact allowed origins, a bounded goal, and caller-authorized ordinary values. Supply independent final checks whenever the outcome has a machine-readable predicate. For desktop Orca, also meet the reference's foreground agreement and focus checks; isolated headless sessions use their qualified background mode.
4. Run run-jev with one JSON job on stdin. Inspect the JSON result, then verify the page's actual state. Report success only when independent checks establish the requested outcome. A handoff returns control to the regular browser workflow on the same page; reconcile any uncertain action before retrying it.

Keep the regular browser skill as the default until the experiment meets its documented evaluation gates.