arc-cua
A computer-use action layer that lets an agent hand bounded desktop subtasks to a fast decision model for execution.
Added to Jevfast
How it uses Jev
The caller supplies a goal, inputs, verification conditions, constraints, and an action limit. The runtime observes the desktop, builds a legal action space, asks Jev to choose, guards against stale state, executes, and repeats until a terminal condition.
What Jev decides
Check one subtask verification criterionExample answers · not a recorded Jev response · Source ↗
Question 1 · verification 0
YOUR APP
INSTRUCTION
Assess only the supplied criterion against current desktop state. Planned or attempted actions are not proof; select UNKNOWN if evidence is insufficient.
STATE
The bounded subtask supplies verification criteria; the representative question evaluates one criterion against current observed desktop state. The three answer options are fixed.
JEV · CHOICE
- SATISFIED
- NOT_SATISFIED
- UNKNOWN
App workflow
- Submit bounded subtask
A planner supplies the goal, input values, verification conditions, constraints, and maximum action count.
- Observe and decide
The executor reads the desktop, builds the legal action space, and asks Jev for the next action.
- Guard and execute
Freshness checks ensure the target still matches current state before the UI action runs.
- Verify or return
The loop repeats until the subtask completes or reaches a limit; the executor returns status and actions taken.
Why it is interesting
It moves repeated low-level UI decisions into a bounded executor while preserving high-level planning outside and checking each action against current desktop state.