dsh-llm-agent-harness-guide
by h565656445
DeepSeek Harness 个人适配声明(Personal Adaptation Notice) 本项目是 DeepSeek Harness 的 个人适配产物(personal adaptation) , 并非 DeepSeek Harness 官方文件(not an official DeepSeek Harness file)…
Personal Adaptation Notice for DeepSeek Harness — a personal adaptation, not an official DeepSeek Harness file.
安装
dsh plugin --profile web add github:h565656445/dsh-llm-agent-harness-guideGitHub 源码安装:首次需按提示配置 allowBuilds 构建授权后重试
安装与环境配置指引、插件开发教程见 DSH 中文社区文档 ↗
安装即在你的机器上以你的权限运行第三方代码——它可读写文件、使用凭据、访问网络,DSH 的工具审批不会为插件代码加沙箱。「检测到 manifest」仅代表发现 dsh.bundle / dsh.plugin 清单,不构成兼容性或安全审查;安装前请审阅源码,不熟悉的插件先在不含密钥的环境试用。
README
目录
- 用途 / What this is for
- LLM Agent Control-Plane Design Guide / LLM Agent 控制平面设计指南
- Features / 功能
- What's inside / 目录结构
- Quick start / 快速开始
- DeepSeek Harness 衍生 / DSH Derivative
- License / 许可证
- 相关项目 / Related Projects
- Agent OS 内核 / Kernel
- Harness 基础设施 / Infrastructure
- 规格与文档 / Specs & Docs
- 适配器 / Adapters
- DSH 总集成 / Integration
DeepSeek Harness 个人适配声明(Personal Adaptation Notice)
本项目是 DeepSeek Harness 的个人适配产物(personal adaptation),并非 DeepSeek Harness 官方文件(not an official DeepSeek Harness file),随附功能、使用说明与个人产物(bundled with features, documentation, and personal artifacts),可与 DeepSeek Harness 搭配使用,也可独立使用。
This project is a personal adaptation for DeepSeek Harness, and is NOT an official DeepSeek Harness file, bundled with features, documentation, and personal artifacts. It can be used alongside DeepSeek Harness or standalone.
作者 / Author: h565656445
合作 / Collaboration: 如有项目可以一起合作,欢迎联系。微信:wohaishihenshuaide。If you have projects, let's collaborate. WeChat: wohaishihenshuaide.
用途 / What this is for
LLM Agent 控制平面设计指南:控制循环、Worker 协议、审批边界与可观测性的设计方法论。
LLM agent control-plane design guide: control loops, worker protocol, approval boundaries and observability.
LLM Agent Control-Plane Design Guide / LLM Agent 控制平面设计指南
本项目是基于 Hermes Harness 源架构(core/ 与 specs/Agent-OS-*)综合编写的 LLM Agent 控制平面设计指南(纯文档仓库)。内容覆盖:控制平面的定位与控制循环(事件驱动、有界执行)、Worker 协议(Prepare/Accept/Inspect、能力与权限适配器、最小上下文、结构化结果与证据哈希)、审批边界(自动 vs 必须人工、单次使用审批、人工完成门)、以及可观测性(观测合同、统一指标、质量门、失败关闭)。不包含任何源码。
This repository is a bilingual LLM Agent control-plane design guide synthesized from the Hermes Harness source architecture (core/ and specs/Agent-OS-*), pure documentation. It covers: control-plane positioning and the control loop (event-driven, bounded execution); the worker protocol (Prepare/Accept/Inspect, capability and permission adapters, minimal context, structured results with evidence hashes); approval boundaries (automatic vs human-required, single-use approvals, human completion gates); and observability (observation contracts, unified metrics, quality gates, fail closed). No source code is included.
Features / 功能
- 控制循环:事件驱动状态机、有界 Loop(两次执行、一次修正)、受监督暂停点 Control loop: event-driven state machine, bounded loop (two runs, one fix), supervised pause point
- Worker 协议:Prepare/Accept/Inspect 统一 seam;能力与权限由适配器声明;Worker 只读最小上下文 Worker protocol: Prepare/Accept/Inspect unified seam; capabilities and permissions declared by adapters; workers read minimal context only
- 审批边界:发布/删除/付款/账号/规则修改必须人工;审批只解除合同中列出的动作 Approval boundaries: publish/delete/payment/account/rule changes require humans; approval only lifts actions listed in the contract
- 可观测性:观测合同 + 五维指标 + 质量门;预算超限或质量不合格立即失败关闭 Observability: observation contracts, five metric dimensions, quality gates; budget or quality violations fail closed
- 恢复与审计:哈希链 Ledger 为事实源,Snapshot 只是缓存;损坏只暴露可信前缀 Recovery and audit: hash-chain ledger as source of truth, snapshots as caches; corruption exposes only the trusted prefix
- 双语:中文与英文对照编写,可直接阅读,无运行依赖 Bilingual: Chinese and English side by side; readable as-is, no runtime dependencies
What's inside / 目录结构
dsh-llm-agent-harness-guide/
├── README.md # 双语说明 + DSH 衍生声明
├── LICENSE # MIT
├── docs/
│ └── llm-agent-harness-guide.md # 双语 LLM Agent 控制平面设计指南(核心文档)
└── .dsh/ # DeepSeek Harness 衍生包
├── preset.yml
├── agent.cordis.yml
├── README.md
└── skills/dsh-llm-agent-harness-guide/SKILL.md
Quick start / 快速开始
# 1. 阅读设计指南
$repo = "E:\path\to\dsh-llm-agent-harness-guide"
Get-Content (Join-Path $repo "docs\llm-agent-harness-guide.md")
# 2. 用设计检查清单评审你的 Agent 控制平面
Get-Content (Join-Path $repo "docs\llm-agent-harness-guide.md") | Select-String "设计检查清单"
# 3. 安装 DSH 预设(可选)
$dst = Join-Path $env:DSH_HOME ".agent-presets\hermes-llm-agent-harness-guide"
Copy-Item -Recurse -Force (Join-Path $repo ".dsh") $dst
DeepSeek Harness 衍生 / DSH Derivative
本项目附带 DeepSeek Harness 衍生包,位于 .dsh/ 目录:
preset.yml— Agent 预设元数据agent.cordis.yml— Cordis 组装(基于 standard 预设,persona 已定制)skills/dsh-llm-agent-harness-guide/SKILL.md— 项目专属技能(skill)
安装与接入方式见 .dsh/README.md(双语)。
License / 许可证
相关项目 / Related Projects
这是 DeepSeek Harness 个人适配系列(共 31 个仓库)的完整导航。 / This is the complete navigation for the DeepSeek Harness personal-adaptation series (31 repos).
Agent OS 内核 / Kernel
dsh-agent-os-runtime · dsh-agent-os-planning · dsh-agent-os-scheduler · dsh-agent-os-worker-protocol · dsh-agent-os-observability · dsh-agent-os-specs
Harness 基础设施 / Infrastructure
dsh-harness-core · dsh-graph-entry · dsh-async-job · dsh-file-identity · dsh-json-projection · dsh-manual-approval · dsh-observation-writer · dsh-provider-control · dsh-schema-negotiator · dsh-upgrade-governance
规格与文档 / Specs & Docs
dsh-harness-specs · dsh-novel-specs · dsh-architecture-guide · dsh-powershell-patterns · dsh-json-schema-driven-dev · dsh-llm-agent-harness-guide(本仓库 / this repo)
适配器 / Adapters
dsh-short-story-engine · dsh-tutorial-video-state-machine · dsh-governance-kernel · dsh-sports-pipeline · dsh-motion-grammar
DSH 总集成 / Integration
dsh-integration · dsh-presets-pack · dsh-skills-pack · dsh-starter-kit
原始 README: https://github.com/h565656445/dsh-llm-agent-harness-guide/blob/main/README.md ↗
同类插件
查看全部 →vibe-skills
vibe 编码技能合集(含快速开始、CI 证明、发布说明与文档)

Aegis
面向编码 Agent 的软件工程方法包,提供基线优先规划、系统化调试、提示词卫生、完成前验证,以及修复/退役双轨跟踪技能。

J-Space-Cognition-Suite-V3.6
A model-agnostic inference-time control suite for deep reasoning, long-horizon work, tool use, verification, and recovery. 面向深度推理、长程任务、工具调用、验证与恢复的模型不可知推理时控制套件。

MisakaNet
失败恢复记忆库:从真实工程会话中搜索和记录失败恢复教训,支持 BM25 + 语义 RAG 检索和知识库管理。

dsh-find-plugins
对 DSH 说一句「有没有插件能……」,它就会从全 GitHub 的 dsh-plugin topic 里找出候选,解释差别,等你选好以后再安装和验证。

humanizer-ru
AI代理技能:识别并清除俄语文本中的机器生成痕迹。38种模式,39个带证据注册的regex标记,盲对偶运行,C2PA/EXIF/XMP文件元数据清理