NewCUA-JEV
Choose and verify constrained computer actions across GUI and structured tools using Jev.
Added to Jevfast
NewHow it uses Jev
CUA-JEV turns observed computer state into legal action candidates. Jev selects one intent-and-channel action; code validates the option, guards its execution, runs a registered executor and independently verifies the resulting state before observing again.
What Jev decides
Read-file policy test context (illustrative)Example answers · not a recorded Jev response · Source ↗
Question 1 · action
YOUR APP
INSTRUCTION
Select one legal intent-and-execution action to advance the task. The test offers gui (click a button) and mcp (read with MCP).
STATE
Repository-owned test context: task “read”, subgoal “read file”, and two candidate IDs gui and mcp. The policy adds observation_id, source, nested state and serialized available_actions. This is a source-defined test request shape, not a live execution trace.
JEV · CHOICE
- gui
- mcp
App workflow
- Observe and enumerate
The task adapter reads structured state and creates grounded action candidates.
- Choose and guard
Jev returns a candidate ID; the guard checks freshness, identity, preconditions and permissions.
- Execute and verify
A registered executor runs the action. An independent verifier checks state before the loop repeats.
Why it is interesting
The framework chooses both what to do and which available computer interface should do it. Candidate generation, execution permissions and terminal verification remain in code. The reviewed release is scoped to bounded workflow adapters, not arbitrary Windows tasks.