Jev CU
A Codex computer-use skill that asks Jev to choose actions from accessible on-screen controls.
Added to Jevfast
How it uses Jev
The loop reads accessible interface text and indexed controls, then sends the goal, app, bounded context, recent actions, constraints, and candidate labels to Jev. Jev chooses a target and action, judges whether the goal is visibly done, and flags whether the action needs confirmation. Codex CUA performs reading and execution; local policy gates sensitive actions.
What Jev decides
Illustrative Jev requestExample answers · not a recorded Jev response · Source ↗
Question 1 · action
YOUR APP
INSTRUCTION
What is the next action type?
STATE
Goal, app/context, candidate elements, recent actions and constraints.
JEV · CHOICE
- click_element
- click_at
- drag
- set_value
- type_text
- press_key
- scroll
- wait
- ask_user
App workflow
- Observe the interface
The CUA driver reads accessible controls and a small window/focus context for the selected app.
- Prepare a bounded request
Code clips candidate controls, includes the goal and recent actions, and constructs the target, action, done, and risk questions.
- Apply the policy gate
Local policy evaluates the chosen action and risk signal; sensitive operations pause for user confirmation.
- Execute or preview
The loop runs the selected CUA step or records it in dry-run, then observes the interface again up to the step limit.
Why it is interesting
It makes the action, target, completion, and risk judgments explicit while sending text rather than screenshots. The default is dry-run, and policy plus app allow-lists constrain consequential actions.