Electricitysheep/dsh-handbook

What Is This Project?

dsh-handbook (short for DeepSeek Harness 白皮书) is an open-source, community-driven manual for DeepSeek Harness (abbreviated dsh)—an official agent runtime that DeepSeek released on August 13, 2026. Think of dsh as a “Lego base” for building AI agents: it provides a pluggable foundation where you can add or swap components (called “plugins”) to customize how an AI assistant behaves, what tools it can use, and how it interacts with your system.

The handbook fills a gap in the official documentation: while DeepSeek’s docs focus on architecture, this project offers a from-zero-to-one path—installation, usage, plugin development, performance tuning, and real-world case studies—with every command tested on real hardware. It’s written primarily in Chinese, with English translations available for many chapters.

What Does the Handbook Contain?

The project is structured as a 15-chapter guide (plus appendices), organized into five phases:

• Phase 1 (Chapters 1–2): Introduction and 5-minute quick start—install dsh, run your first conversation, and understand the two modes (web UI and headless CLI).
• Phase 2 (Chapters 3–4): Deep dive into profile and the plugin system—how to mount plugins, extend functionality, and write your first plugin from scratch.
• Phase 3 (Chapters 5–6): Real-world scenarios and performance tuning—five example case studies, reasoning-effort strategies, and common pitfalls.
• Phase 4 (Chapters 7–9): Ecosystem and orchestration—how to integrate with MCP (a protocol for connecting external tools), use sub‑agents, and build deterministic workflows.
• Phase 5 (Chapters 10–15): Advanced topics—complex case studies, future outlook, honest limitations, security & sandboxing, cost/cache engineering, and a final ecosystem report.

Appendices include a glossary, a command cheat sheet, and a benchmark comparing dsh against other agents (like Claude Code, Codex, OpenCode, Gemini CLI, Kimi CLI) using the same model.

Who Is This Handbook For?

• Beginners who have never used dsh—follow the 3‑day learning path and you’ll go from zero to a working setup.
• Developers looking for cloneable plugin templates and a full configuration reference.
• Evaluators comparing different agent frameworks—the handbook includes a capability matrix and benchmark results.
• Performance tuners—learn about reasoning-effort levels, caching, and cost optimization (the handbook reports a 97% cache hit rate in its tests).
• Anyone wanting to understand the dsh ecosystem through real discussions, plugin examples, and community insights.

Key Features & Assets

The README highlights several practical tools:

• Quick start: One command (npx -y @deepseek-ai/dsh web) launches a web UI at http://127.0.0.1:3080.
• Cheat sheet and config reference—copy‑paste settings for common scenarios.
• Plugin template—a minimal example showing how to mount a plugin in two steps.
• FAQ covering five common questions (e.g., “Is dsh a model?” → no, it’s a runtime).
• Comparison matrix vs. five other agents, showing that dsh is open‑source, model‑agnostic, and has an official plugin system (a differentiator).
• Benchmark results: on a 5‑task test with the same model, dsh completed tasks in 130 seconds (vs. 70s for omp and 172s for opencode), all achieving 45/45 correct.

How to Use the Handbook

• Online reading: https://electricitysheep.github.io/dsh-handbook/
• PDF downloads: Both Chinese and English complete versions are available in the Releases section.
• Start reading: Jump to the 3‑day learning path or the 5‑minute quick start.

Community & Contribution

The handbook is a living document—it has gathered 280+ stars on GitHub, 195+ replies in the official discussions, and has an ongoing feedback pipeline to incorporate community Q&A. If you run a real case, you can submit it for inclusion (with attribution). The project also curates a list of recommended community plugins (e.g., git panels, memory engines, cost trackers) and actively links with other ecosystem projects.

License

Content is under CC BY-NC-SA 4.0; example code is MIT.

---

This explainer is based solely on the information provided in the project’s README.