Overview & result
Jev Desktop adds a bounded decision loop to Codex Computer Use for browser tabs and native macOS apps. Codex understands the goal, prepares text, defines the allowed actions and verifies the result; Jev selects an operation and its compatible target in one request; Computer Use performs the click, fill, scroll or shortcut. It reuses the existing Codex Computer Use runtime.
From the original post
TypeSafe Jev action selection inside Codex Computer Use
How Jev fits in the loop
- Codex defines the goal, the prepared text and the list of allowed actions.
- Computer Use observes the approved interface.
- One TypeSafe request predicts the operation and a target for every operation; only the target for the chosen operation is validated and executed.
- Codex verifies the final result and handles recovery.
How to reproduce
- Inspect the original showcase and source material at https://github.com/yikangy873-gif/jev-desktop
- Verify the bounded prompt and input candidate schema configured for Jev
- Benchmark decision latency and classification accuracy against baseline models
Why this build matters
It keeps a short action loop inside the existing Computer Use runtime, so the main agent is not consulted after every click, while Codex keeps control of scope and permissions.
Limitations
- Actual speed depends on app launch, page loading and accessibility observation latency.
- Canvas interfaces, arbitrary native menus and inaccessible custom widgets are not automatically supported.
- Every executable control must be explicitly allowed by the calling Codex task.