The desktop agent, rebuilt one lesson at a time

What it does

This is not WorkBuddy’s source — it’s a clean-room teaching blueprint that rebuilds a desktop AI assistant from scratch across 24 chapters. Each chapter adds exactly one mechanism (agent loop, tool dispatch, permission hooks, sidecar, memory, context compaction, audit sandbox) with its own code.py and diagram. There’s also a stdlib-only “Mini WorkBuddy” so you can trace a full request path without any of the furniture.

The interesting bit

The framing is the value: “the model is the brain, the harness is the operating system.” The whole curriculum is organized around three fundamental tensions — finite context vs infinite information, autonomy vs safety, model cost vs task complexity — and every chapter is an answer to one of them. That’s a rarer teaching structure than “here’s how to call an API.”

Key highlights

• 24 chapters, 28 diagrams, each chapter’s code inherits the previous and adds one mechanism (documented in a progression-contract.md)
• Runs fully offline — the complete harness tour works without any API key
• Multi-provider adapters: DeepSeek, OpenAI, Anthropic
• Covers the unglamorous parts real tutorials skip: session resume, JSONL event logs, output externalization, permission hooks, audit hash chains
• Includes a candid code-quality review comparing itself against the sibling learn-claude-code project

Caveats

• Primary language is Chinese; an English README exists but the course content itself is Chinese-first
• It’s a tutorial, not a product — you’ll learn the architecture, not ship a WorkBuddy clone out of the box

Verdict

If you’ve built a CLI agent and hit the wall where desktop engineering complexity multiplies, this is a well-structured map through that wall. If you wanted production-ready assistant code, this isn’t that — and it says so up front.