deepseek-harness-jev:给 DSH 加 Jev 决策插件
开源 MIT 插件,在技能排序、agent 监督、工具输出准入、审批升级等环节插入 TypeSafe Jev 判断,主模型照常规划调用工具。

Agent oversight doesn’t have to mean rewriting your harness. deepseek-harness-jev is a native DeepSeek Harness (DSH) plugin for builders who want TypeSafe Jev judgments at selected points in an agent workflow. It helps you add optional decision support around selection, supervision, tool output, approvals, and session review while the main model continues to plan, answer, and call native tools. Key features: • Skill + file ranking – ranks existing skill summaries and glob-path results before the agent uses them • Agent supervision – adds optional drift reminders, completion checks, goal checks, and instruction guidance • Tool-output admission – can filter unneeded progress or repeated notices while keeping an original-output recovery reference • Workspace approval – can assist with eligible native single-operation escalation requests; other cases return to human approval • Stage navigation – classifies recorded completed model steps and links consecutive stages to their trajectory evidence It’s open-source (MIT license). Link in the reply 👇