内容 CC BY-NC-SA 4.0、示例代码 MIT——15 章中文优先的 DeepSeek Harness（dsh）从 0 到 1 手册，可在线阅读或下载 PDF；它讲解 dsh 而非安装进 dsh（生态层）。
DSH 适配
生态相关
作者声明
安全审计
未审计
最后核验
2026-08-21
许可证
作者未说明
01它能帮你完成什么？
Go from zero to productive with DeepSeek Harness: install, daily use, plugin development and performance tuning
A 15-chapter handbook with copy-paste commands verified locally, a 3-day learning path with daily goals and acceptance criteria, a one-page cheatsheet, and a cloneable plugin template with a full config reference
Developers and newcomers to DeepSeek Harness (dsh) who need a structured Chinese-first path that the official architecture docs don't provide
Compare DeepSeek Harness with mainstream agents and shortlist with real benchmarks
A capability matrix vs 5 mainstream agents (Claude Code, OpenAI Codex, OpenCode, Gemini CLI, Kimi CLI) plus same-model benchmarks on deepseek-v4-flash: omp 70s / dsh 130s / opencode 172s, 45/45 tasks correct
Teams evaluating which agent harness to adopt
Read the dsh ecosystem: pain points, capability gaps and the plugin landscape
Chapter 15 ecosystem report cross-validating an inventory of 1804 plugin repos with 780 discussion threads: 5 key conclusions and 6 capability gaps
Plugin developers and ecosystem participants deciding what to build or adopt on dsh
02如何接入 DeepSeek Harness？
前置条件
- Node.js ≥ 22 (required to run dsh itself, which the handbook's chapters teach)
安装步骤
- 01
Read the handbook online at https://electricitysheep.github.io/dsh-handbook/ or download the Chinese / English PDF from GitHub Releases
- 02
To follow the chapters, install and start the dsh Web UI: `npx -y @deepseek-ai/dsh web`, then open http://127.0.0.1:3080
$ npx -y @deepseek-ai/dsh web
- 03
For one-shot tasks (scripts/CI), run `dsh --profile headless "task"`
$ dsh --profile headless "task"
验证接入成功
- Open http://127.0.0.1:3080 in the browser and start a conversation (dsh Web UI)
03DSH 适配与能力边界
Chinese-first DeepSeek Harness (dsh) handbook / whitepaper — a 15-chapter documentation project (tutorials, benchmarks, ecosystem observation) that teaches dsh; it is read online or as PDF, not installed into dsh
15-chapter zero-to-one handbook
a new DeepSeek Harness user→install, quickstart, profiles & plugin system, plugin development, real cases, performance tuning, MCP / subagents / workflows, security & sandbox, cache & cost — every chapter with runnable commands
Cloneable plugin template & config reference
a developer writing their first dsh plugin→a cloneable template (pure functions + waterfall + tests) and a settings.yaml / cordis.patch.yml config reference — mounting a plugin takes 2 edits (add dependency + add insert line)
following the template edits package.json and cordis.patch.yml inside the dsh profile and runs pnpm installBenchmarks & agent comparison
a buyer or selector comparing agent harnesses→capability matrix vs 5 mainstream agents and same-model benchmark data (5 tasks × 3 rounds, median): omp 70s / dsh 130s / opencode 172s, 45/45 correct
Ecosystem observation center
1804 plugin repos + 780 discussion threads→Chapter 15 panorama report: 5 conclusions (Windows as top pain point, community filling official gaps, sandbox tooling scarcity, serialization bug family, cost transparency) and 6 capability gaps
04适合谁？何时不该用？
适合
- Developers and newcomers to DeepSeek Harness (dsh) who need a structured Chinese-first path that the official architecture docs don't provide
- Teams evaluating which agent harness to adopt
- Plugin developers and ecosystem participants deciding what to build or adopt on dsh
不适合
- The handbook targets dsh 0.1.0-rc.6 (pre-release); the README warns production use needs careful evaluation, and chapter commands can go stale as rc versions iterate with breaking changes.
- The handbook documents Windows as dsh's #1 pain point (Chinese paths — 15+ threads with the same root cause — koffi, ports, child processes); its own sample environment is Windows 11 + Node 24 with a documented Node-version red line.
05兼容性、维护与安全提示
- The handbook targets dsh 0.1.0-rc.6 (pre-release); the README warns production use needs careful evaluation, and chapter commands can go stale as rc versions iterate with breaking changes.
- The handbook documents Windows as dsh's #1 pain point (Chinese paths — 15+ threads with the same root cause — koffi, ports, child processes); its own sample environment is Windows 11 + Node 24 with a documented Node-version red line.
- Handbook content is CC BY-NC-SA 4.0 — attribution, non-commercial, share-alike — so reuse in commercial products is not permitted; only the example code is MIT.
CC BY-NC-SA 4.0 (content) / MIT (examples) · actively maintained (v1.8.0 released 2026-08-17)
06常见问题
dsh-handbook 是装进 DeepSeek Harness 的插件吗？
不是——它是文档：一本 15 章的手册，可在 electricitysheep.github.io/dsh-handbook 在线阅读，或从 GitHub Releases 下载中英文 PDF。它教你使用 dsh，本身不安装进 dsh。
这本手册覆盖哪个 dsh 版本？
dsh 0.1.0-rc.6（预发布阶段），模型为 DeepSeek-V4-Flash-0731。README 提醒生产环境请谨慎评估，且 rc 版本迭代快，章节命令可能失效。
手册里的命令真的能直接跑吗？
能——每章都提供作者在本机实测验证过的可复制命令，例如 `npx -y @deepseek-ai/dsh web`（需 Node.js ≥ 22）启动 Web UI 于 http://127.0.0.1:3080，`dsh --profile headless "任务"` 适合脚本/CI 的一次性任务。
只有中文版吗？
不是——中文优先、英文同步：除了中文 15 章完整版 PDF，还提供英文 README（README.en.md）和 83 页英文 PDF（约 16 万字符）。
内容可以二次使用吗？
手册内容为 CC BY-NC-SA 4.0（署名、非商业性使用、相同方式分享），仅示例代码为 MIT。
07相关的 DSH 工作流
archify
作者 tt-a1i
为编码智能体生成美观可验证的架构图、时序图与数据流图，输出自包含 HTML，支持动效与清晰导出。
openviking
作者 volcengine
为 AI 智能体打造的自进化上下文数据库，统一智能体记忆、知识 RAG 与技能。
nocobase
作者 nocobase
开源的 AI + 无代码应用搭建平台：AI 在久经生产验证的底层基础设施与所见即所得的无代码界面之上工作，帮你快速构建 CRM、ERP 等业务系统，兼顾速度与可靠性。
learn-harness-engineering
作者 walkinglabs
Harness 工程新手教程，从 0 到 1 系统学习智能体工作流框架。
08数据与来源
**DeepSeek Harness 中文手册 × 生态观察中心**——从 0 到 1 玩转 dsh，跟着 780 帖讨论区看懂生态 · 中文 · [English](./README.en.md)
但官方文档以架构说明为主，**缺少一条从零上手的路径**。**这本白皮书补上这条路**：从"什么是 Agent 运行时"讲起，到安装、使用、开发插件、性能调优——每一章都有可复制、可运行的命令，全部在本机实测验证。
页面基于项目公开文档、仓库元数据和 DSH Plugins 的结构化解析生成；最后核验于 2026-08-21。发现错误？提交更正。
最佳 DeepSeek Harness 插件
从全目录挑出的 12 个值得优先安装的插件，覆盖各个分类。