MIT 许可的项目制 Harness 工程课程(14 讲、8 项目、15 种语言),提供可直接套用的模板、harness-creator 技能与 DeepSeek harness 设计拆解——属生态学习资源,不作为插件装入 dsh。
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
MIT
01它能帮你完成什么?
Learn how to build the environment that makes AI coding agents complete real engineering tasks reliably
14 lectures and 8 hands-on projects (around one evolving Electron knowledge-base app) covering the five harness subsystems: instructions, state, verification, scope, and session lifecycle
Engineers already using coding agents who want better stability and quality, and tech leads who need to understand how environment design affects agent performance
Improve an existing coding agent immediately with copy-ready harness templates
Starter templates (AGENTS.md, init.sh, feature_list.json, session progress log) dropped into your own repository so every agent session starts from the same state
Developers running coding agents — DeepSeek Harness (dsh), Claude Code, Codex or similar — on real projects who want more stable multi-session runs
Scaffold and audit a production-grade harness in your own repository
The harness-creator skill scaffolds AGENTS.md, feature lists, init.sh and verification workflows in minutes; audit-harness.sh checks a repo against all five harness subsystems and exits 0 when all CRITICAL items pass
Teams adopting harness engineering who want a quick scaffold and a repeatable audit for their agent workspace
02如何接入 DeepSeek Harness?
前置条件
- Familiarity with the terminal, git, and local development environments (course prerequisite)
- Node.js / npm to run the VitePress documentation site locally
- At least one coding agent (Claude Code, Codex, or another IDE/CLI agent supporting file editing, command execution, and multi-step tasks) to complete the projects
安装步骤
- 01
Clone the repository, or read the course on the documentation website (walkinglabs.github.io/learn-harness-engineering)
- 02
Run `npm install` to install the VitePress tooling
$ npm install
- 03
Run `npm run docs:dev` and open the local URL that VitePress outputs in your browser
$ npm run docs:dev
验证接入成功
作者未说明
03DSH 适配与能力边界
Ecosystem learning resource: a course teaching harness engineering methodology (including a breakdown of DeepSeek's harness design) — read it online or locally, then apply its templates and methods to your own coding-agent workflows; it is not installed inside dsh.
Project-based harness engineering course
14 lectures + 8 hands-on projects + learning path in the repo's docs→a systematic method for the five harness subsystems (instructions, state, verification, scope, lifecycle) applied to real repositories
Copy-ready Resource Library
templates such as AGENTS.md, feature_list.json and init.sh→drop-in harness files you can use in your own repositories today
you copy template files into your own repositoryharness-creator skill + audit-harness.sh
your own project repository→a scaffolded production-grade harness (AGENTS.md, feature lists, init.sh, verification workflows); an audit that exits 0 when all CRITICAL items pass
harness-creator scaffolds harness files into your projectthe one-liner `curl ... | bash` downloads and executes a remote shell script; you can clone the repo and run `bash tools/audit-harness.sh` locally instead15-language content and PDF coursebooks
course sources translated into 15 languages→localized course pages; PDF coursebooks generated via `npm run pdf:build`
`npm run pdf:build` writes output files to artifacts/pdfs/
04适合谁?何时不该用?
适合
- Engineers already using coding agents who want better stability and quality, and tech leads who need to understand how environment design affects agent performance
- Developers running coding agents — DeepSeek Harness (dsh), Claude Code, Codex or similar — on real projects who want more stable multi-session runs
- Teams adopting harness engineering who want a quick scaffold and a repeatable audit for their agent workspace
不适合
- The hands-on projects require at least one coding agent (Claude Code, Codex, or another IDE/CLI agent with file editing, command execution and multi-step tasks). Without one you can read the course but cannot complete the projects as intended.
- The README's quick way to run audit-harness.sh pipes a remote script into bash (curl ... | bash), which executes network-fetched code on your machine. If you prefer not to, clone the repo and run `bash tools/audit-harness.sh /path/to/your/repo` locally instead.
05兼容性、维护与安全提示
- The hands-on projects require at least one coding agent (Claude Code, Codex, or another IDE/CLI agent with file editing, command execution and multi-step tasks). Without one you can read the course but cannot complete the projects as intended.
- The README's quick way to run audit-harness.sh pipes a remote script into bash (curl ... | bash), which executes network-fetched code on your machine. If you prefer not to, clone the repo and run `bash tools/audit-harness.sh /path/to/your/repo` locally instead.
- The course is actively evolving content — new lectures and projects were added in July and August 2026 — and has no tagged release, so course structure and materials may change between checkouts.
MIT · actively maintained (last push 2026-08-19)
06常见问题
这是装进 DeepSeek Harness 里的插件吗?
不是。它是一个生态学习资源——一个课程仓库。你可以在线阅读,或本地执行 `npm install` 后用 `npm run docs:dev` 预览,再把课程方法和模板(AGENTS.md、init.sh、feature_list.json)套用到自己的仓库。课程本身不绑定特定代理,还包含对 DeepSeek 构建 harness 方式的拆解(“一切皆插件”、能力接缝、事件管线)。
开始做项目前需要准备什么?
需要熟悉终端、git 和本地开发环境,并且至少有一个编码代理——Claude Code、Codex,或任何支持文件编辑、命令执行和多步任务的 IDE/CLI 代理。没有这类工具也能读课程内容,但无法按要求完成实战项目。
课程里的 harness 模板怎么配合我自己的编码代理使用?
从资源库把入门模板放进项目根目录即可:AGENTS.md 作为代理的操作手册,加上 init.sh、feature_list.json 和会话进度日志。README 说明进度文件与代理无关:Codex、OpenHands、Antigravity 等编码代理只要在根指令里要求启动时读取、交接前更新,就能使用同一份文件。
harness-creator 技能和 audit-harness.sh 是什么?
harness-creator 是一个可复用的代理技能,能在几分钟内为你自己的项目搭建生产级 harness(AGENTS.md、功能清单、init.sh、验证流程)。audit-harness.sh 是零依赖的 shell 审计脚本,按五大子系统(L03–L12)检查现有仓库,所有 CRITICAL 项通过时退出码为 0,无需 Node.js。
课程有中文或日文版本吗?
有。课程提供 15 种语言,包括简体中文、繁体中文和日文,全部翻译完整覆盖;还可以用 `npm run pdf:build` 在本地生成 PDF 教材。
07相关的 DSH 工作流
dsh-agent-teams
作者 nanmicoder
DeepSeek Harness 的 Agent Teams 多智能体协作插件,支持多个 AI Agent 组成团队,协同完成复杂任务,实现任务分配、并行执行、成员通信与团队协作。 AgentTeams plugin for DeepSeek Harness
dsh-blender-plugin
作者 sixtysevenlf
DSH × Blender 直连插件 v3.0.6:15 个模型工具让 AI 驱动 GUI Blender(视口 / 自定义视角 / 无头 worker / 作业层 / 事务回滚 / 配方库),34 family / 228 op 契约层;npm 安装 + addon zip,GUI 优先、无头备用;dsh-TUI 选中 chip / @ 提及 / /blender,随包建模 skill,支持 WSL / Windows / macOS。
dsh-free-search
作者 dddmuc
为DeepSeek Harness提供的免费网络搜索服务商:即使没有API密钥也可使用。Free web search provider for DeepSeek Harness:usable even without an API key.
dsh-bili-scroll
作者 c1overzz
DSH 插件:每天替你刷一轮 B站,读字幕挑出值得的、写一份日报,值得的 @ 你
08数据与来源
Learn Harness Engineering is a course dedicated to the engineering of AI coding agents.
[How DeepSeek builds its harness](docs/en/harness-designs/deepseek/index.md) — "everything is a plugin," capability seam…
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。
