Small decisions. Interesting possibilities.Submit contentSubmit
AIWG Jev Decision Support preview

AIWG Jev Decision Support

Offload bounded agent questions to Jev with typed answers, confidence checks and a fallback.

Added to Jevfast

How it uses Jev

AIWG's optional decision-engine addon sends one bounded question, a typed mode, and optional context to Jev. Supported modes are yes/no, a choice among supplied IDs, or a bounded integer scale. The CLI and adapter return the typed answer with confidence or a fallback for the caller's LLM. The quickstart describes choices such as file ownership among cli, config, runtime, and policy; context can be supplied as text, a file, or stdin. The source does not describe automatic capture of arbitrary project requests.

What Jev decides

Choose which file owns retry policyExample answers · not a recorded Jev response · Source ↗
Question 1 · answer
YOUR APP
INSTRUCTION

Select the file that owns retry policy from the supplied described choices.

STATE

The documented CLI file-ownership question is stored in input.message, with optional context appended. The ruleset evaluation alias is answer; four described file choices are cli, config, runtime and policy.

JEV · CHOICE
  1. cli
  2. config
  3. runtime
  4. policy

App workflow

  1. Configure the optional Jev addon

    The user opts into the decision-engine addon and configures Jev through the setup command; credentials are resolved from supported environment variables or the user credential file.

  2. Ask a bounded question

    The agent or CLI supplies the question, exactly one typed mode, and optional context. The documented example provides descriptions for each choice.

  3. Use the answer or fall back

    AIWG returns a typed answer when confidence meets the threshold. If Jev is unavailable, abstains, or falls below the threshold, the caller decides normally with its LLM.

Why it is interesting

AIWG makes the decision boundary explicit: it offers a small answer shape to Jev, applies a confidence threshold, and leaves open-ended reasoning and fallback handling with the calling agent. Its documented choice descriptions carry meaning while stable IDs carry the selected result.