Box health with Jev
Checks whether a box's agent is ready for a message and, if you allow it, presses the key that frees one stuck on a menu or a one-time screen.
- Reading time
- 3 min read
- Who
- Owner
How it works
A box is only useful while its agent sits at its input line, ready for a message. After a restart, an update or someone poking at the terminal it can be somewhere else: a settings menu left open, a command palette, release notes waiting for Enter, a sign-in screen, a crash back to the shell. Each of those strands the next chat request, scheduled run or answered question, and nobody is there to notice.
What DevBench recognises by itself costs nothing and sends nothing anywhere: an agent that is working is left alone (Escape would cancel its turn), a signed-out box goes to the account machinery (a key cannot fix a login), a usage limit goes to failover, and a box with no session is restarted. Every other screen goes to Jev, a decision model from TypeSafe AI. Jev reads the screen and answers two questions from fixed lists: what is on the screen, and which one key would bring the agent back. It cannot write text, so a screen can never talk it into typing into your box. Its answer is held to the few keys allowed for that state and to a confidence threshold; below it, nothing is pressed and the box is left for a person.
| Setting | What happens |
|---|---|
| Off | The default. Only DevBench's own checks run, and nothing is sent anywhere. |
| Check and log only | Jev judges the screen and the verdict is logged. Nothing is pressed. |
| Check and fix | Jev's key is pressed, and the box is checked again after each key, three rounds at most. |
Checks run after a CLI update restarts a box, when someone presses Check this box above a workspace's terminal, and every ten minutes across running boxes. The sweep only acts on a screen that has not changed since the previous sweep, so a menu somebody has just opened is left to them.
What leaves the box. With box health on, the text on a box's terminal screen is sent to
TypeSafe AI. Obvious secrets are masked first: API keys, tokens, private keys, lines like
password=. DevBench keeps a record of each check (the state, the confidence, what was pressed,
and a hash of the screen, not the screen itself). Turning it on is your organization's consent to
that, which is why it is off until an owner chooses.
If the tile says Not available on this DevBench, the server has no key for Jev and only DevBench's own checks run.
Steps
Open Settings → Boxes and find Box health.
Choose Check and log only first.
Watch Recent checks for a while: which boxes were stuck, and what Jev would have pressed.
Switch to Check and fix once the log looks right.
On a workspace, press Check this box above the terminal to check one box now.
The answer says whether the agent is ready, what was pressed, or what needs a person.
Choose Off to stop sending screens to TypeSafe AI.
Worth knowing
Terminals show secrets. Masking catches the obvious ones, not every one, which is why box health is off until an owner turns it on.
A signed-out box is never fixed with keys. It goes to the account machinery (a fresher login, the next account), and to an owner's sign-in when nothing else works.
Only an owner changes the setting. Anyone on a workspace can press Check this box, and it works with box health off too: DevBench's own checks answer it, and a box with no agent session is restarted.