Jev Use
Give agents typed Jev decisions with explicit screening and escalation to their host.
Added to Jevfast
How it uses Jev
jev-use screens a proposed agent decision, then sends supported typed questions and the caller's state to a configured Jev backend. Jev can choose among caller-supplied actions, rate a value, or answer a check; the library validates and normalizes the typed verdict and can escalate uncertain or unsuitable decisions to the caller. The action names and meanings come from the caller's question definition.
What Jev decides
Choose the next run actionExample answers · not a recorded Jev response · Source ↗
Question 1 · next
YOUR APP
INSTRUCTION
Choose the next action from the options defined by the caller.
STATE
The caller supplies the current run state and a Choice question whose options are merge, rerun, and hold with their meanings. The selected action below is illustrative and does not claim a runtime answer.
JEV · CHOICE
- merge
- rerun
- hold
App workflow
- Screen the proposed decision
The dispatch layer checks whether the question is supported and appropriate for Jev and returns a typed escalation when it should not proceed.
- Send caller-defined questions
The Jev client packages the caller's state and typed question definitions for the configured backend; the caller controls the options and criteria.
- Validate and normalize
The judge validates the backend response and produces typed verdicts, confidence metadata, and escalation information.
- Return or escalate
The agent consumes a usable decision or receives the explicit reason to hand the choice back to the caller or another model.
Why it is interesting
It gives an agent a separate fast decision step for tasks that do not need generated prose. The screening and explicit escalation path let an application keep unsafe or uncertain decisions with its existing higher-level logic.