A desktop-agent framework with an optional Jev skill that selects controls and actions from native accessibility data.
Jev selects a target and action and estimates presence and risk; local policy decides whether to execute.
Keeps the full UI tree out of the main Agent context while exposing the selected action.
The separate Jev scripts were inspected; desktop behavior and end-to-end latency were not tested here.
Inspect the optional single-step act.mjs path for selecting a native application control.
Material analysed：
act.mjs reserves a no-match outcome and caps ordinary targets at 254 in this revision. The script reports truncation and points readers toward --root for focused observation.
Inspect the supporting source ↗act.mjs can replace an unsupported verb with the first supported fallback, beyond type/set-value equivalence. Capability compatibility is not proof of preserved intent: review the corrected command before --execute; the existing risk answers are not a fresh assessment of that replacement.
Inspect the supporting source ↗policy.mjs builds one target table per advertised operation and consumes the chosen operation’s target. The single-step act.mjs truncation and reconciliation findings must not be generalized to every agent-desktop path.
Inspect the supporting source ↗Choose it for accessibility-backed native controls. Before broad automation, inspect the smaller subtree containing the intended control and keep execution disabled while checking the proposed command.
Read MAX_OPTIONS, reconcile and the execute branch in scripts/jev/act.mjs. Follow a target without the requested action capability and confirm the emitted correction; distinguish the proposal output from the --execute branch.