dsh-answer-reviewer
智能体回答审查器:每一轮 agent 的最终输出由独立 LLM 重新打分(1-100),低于阈值时携带具体反馈引导 agent 自我修正。配置经 127.0.0.1 本地 HTTP 服务实时修改,也支持 dsh-better-sidebar 侧栏配置页。
Agentic answer reviewer for DeepSeek Harness: each final assistant turn is re-reviewed by a separate LLM graded 1-100; below-threshold scores steer the agent back with concrete feedback. Live config over a 127.0.0.1 HTTP server, plus a right-sidebar config tab via dsh-better-sidebar.
怎么安装
dsh plugin add dsh-answer-reviewer 简介
dsh-answer-reviewer A dsh host plugin: every time the agent is about to close a turn, the plugin extracts the assistant's final user-facing text and asks a dedicated review model to grade it on a 1-100 scale. If the score is below the configured threshold (default 80), the plugin steers the agent back into the same turn with the reviewer's single most important finding as concrete feedback. Once a turn has been steered maxChallenges times (default 5), the agent's reply is allowed to close the turn as-is — the cap is a leak guard, not the real gate; the score threshold is. Tune the gate, switch…
推荐参考
信息
- 协议
- MIT
- 语言
- JavaScript
- GitHub 星标
- 1
- 月下载
- 283
- 最近更新
- 2026-09-14
- 创建于
- 2026-09-08
基础安全检查
- 检查结果
- 无
- 收录来源
- curated:awesome-dsh-plugin.com, curated:awesome-dsh-plugin/awesome-dsh-plugin
同类推荐
dsh-memory
FuRongJun-1999/dsh-memory
白箱AGI架构探索:元认知(自我认知循环)、持续学习(知识飞轮)、世界模型(条件空间+语义时空图)、自我改进(自举纪律)、零LLM白箱管线与可审计信任护栏。
dsh-engram
kenz1117/dsh-engram
跨会话长期记忆插件,以「记忆宫殿」隐喻重构 agent 长期记忆:双层 SQLite 分库(用户级 + 按 git origin 隔离的项目级)、FTS5 + 本地向量 RRF 融合与新鲜度/命中排序 boost、从会话日志自动摄取(含末轮)、来源审计链、巩固蒸馏与衰减遗忘,以及设置页「记忆库」面板(走廊拓扑、导览管家、翻新清单)。
dsh-loop-continue
yunxiyang/dsh-loop-continue
续跑「叙述了下一步动作却没调用工具」就结束的 agent 轮次:插件在 agent/turn-stopping 回放该轮日志,让裁判模型回答一次 true/false,命中的话把同一轮再推进一步。
dsh-quality-review
CAI-MH/dsh-quality-review
每轮回复结束时用独立审查模型审核输出,判定不合格则引导 agent 修复,每轮最多追问 2 次;可注入 SOP 文件夹标准作为额外审核维度。