Small decisions. Interesting possibilities.Submit contentSubmit
DeepSeek Harness Jev Plugin preview

DeepSeek Harness Jev Plugin

A community plugin adds typed Jev decisions to DeepSeek Harness skill selection, file ranking, supervision and tool-output handling.

Added to Jevfast

How it uses Jev

The plugin sends task context with one Noul judgment per skill or file path. It ranks the returned probabilities to select skills for the agent catalog or reorder existing glob results; the main agent still loads selected skills and reads files through its normal tools.

What Jev decides

File relevance judgment (illustrative request shape)Example answers · not a recorded Jev response · Source ↗
Question 1 · candidate-0
YOUR APP
INSTRUCTION

Is this returned path relevant to the task? The implementation appends the candidate path to this prompt.

STATE

Task context, glob pattern and search path (defaulting to "."). One Noul question is generated for each path already returned by the glob tool; the candidate path appears in its question instructions.

JEV · NOUL
YesNo
Skill usefulness judgment (illustrative request shape)Example answers · not a recorded Jev response · Source ↗
Question 1 · candidate-0
YOUR APP
INSTRUCTION

Would this skill help the task? The implementation appends the skill name and description to this prompt.

STATE

Current task context; one question is generated for each model-invocable skill, with that skill name and description in the prompt.

JEV · NOUL
YesNo

App workflow

  1. Check feature and inputs

    The plugin checks that the relevant feature is enabled and that the task context and candidate set meet its guards.

  2. Ask Jev to rank candidates

    It submits a typed Noul question per skill or returned path, then sorts answers by probability and applies the configured result limit.

  3. Apply the ranking

    Selected skill summaries supplement the session catalog; file ranking returns the same glob results in Jev-ranked order.

Why it is interesting

Jev is attached at existing harness extension points, so a separate decision layer can narrow catalogs and rank already-discovered paths without replacing the main model or rescanning the filesystem.