[Feature Request] Add Jev support to dsh

[lenkophyll]

Add a Jev plugin to dsh, delegating bounded judgments in the agent loop to Jev, while the LLM is only responsible for planning, generation, and open-ended reasoning. The plugin registers a jev_ask tool that takes state + questions and returns structured answers. It just so happens that I recently saw the dsh development branch working on computer use and browser use—combined with Jev, this should work really well.

[nanami-0713]

This exists now — I built exactly this a few days ago: dsh-jev-decide (npm).

Same shape as your proposal: one registered agent tool, jev_decide, that takes state + a typed question and returns a structured, calibrated answer — the LLM keeps planning and generation, bounded judgments get delegated:

jev_decide({
  state: "…",                        // what to evaluate
  question: "…",                     // the judgment to make
  type: "noul" | "choice" | "score", // yes-no / pick-one / rubric score
  options: ["…"],                    // for choice
  levels: ["…"],                     // for score
})
// → { model, type, answer, confidence?, probabilities?, usage }

Install:

dsh plugin --profile web add dsh-jev-decide

Credentials resolve in order: plugin config → TYPESAFE_API_KEY env → the ~/.dsh/.credentials.yaml refs seam, so it reuses the same credential store as other providers.

Verified end-to-end today on both 0.1.2-rc.1 (pinned) and 0.1.6-alpha.2 (fresh boot via npx): tool registers, calls the TypeSafe API, returns calibrated probabilities.

Your computer-use/browser-use instinct is exactly where I think this shines — e.g. a choice-type "which of these three elements is the submit button" before committing a click, or a noul guardrail ("is this page state safe to act on?") without burning a reasoning paragraph on it. I also put together a small reproducible evidence repo, jev-resume-screening, that uses it for resume–JD screening with criteria iterated v1→v3 under regression tests — a worked example beyond routing/guardrail, if useful.

Happy to add a jev_ask alias if you prefer that name.