I build small, sharp tools for people who run coding agents — the kind of tools that answer one question well instead of five questions badly.
中文:我做一些「小而准」的工具,主要解决智能体(AI coding agent)上下文与插件 生态里的具体问题。下面每个项目都写清了「现在能跑什么、还没做什么」。
| Project | What it does | Why it exists |
|---|---|---|
ai-context-doctor npx ai-context-doctor |
Tells you which files in a repo would flood your agent's context, and writes the ignore rules that stop it. Ships a DeepSeek Harness panel (dsh-aiignore). |
Agents read a 40 KB lockfile once and pay for it in every turn after. |
dsh-plugin-toolkit npx dsh-plugin-toolkit verify . |
Preflight for DeepSeek Harness plugins: pack the tarball, install it, import the entry the loader mounts, mount the panel — before you tag a release. Also init, dev, release. |
A plugin can install cleanly and still do nothing. I shipped five of those bugs; the checks now catch all five. |
| dsh-plugin-example | A ~70-line plugin you can read in one sitting and copy. | Every plugin tutorial starts with a 400-line skeleton. |
npx ai-context-doctor --json # what would flood the context
npx dsh-plugin-toolkit verify . # will this plugin actually install?- Every claim has a check behind it. If a README says "tested", there is a test file you can run; if it says "works with X", CI runs X.
- Write down the failure. Five plugins that installed and did nothing documents five real bugs — the exact error text, the root cause, and the check that catches each one now. Two of those were mine.
- Say what is not built yet. The READMEs separate what runs today from what is planned, because a README that promises more than the code is the most expensive kind of debt.
- Personal site: wjddidi-1.github.io
Public, and each README separates what runs from what was never finished:
| Project | What it is | State |
|---|---|---|
| Smart-Robotic-Car-with-LLM-Interaction | "Hey car, come here" → intent → motion, in a simulator | 31 tests, path drawn from the run |
| HRI_NaoHospital | NAO receptionist: expression recognition, dialogue, Dijkstra routing | planner tested; robot parts need NAOqi |
| NetworkFinal | Traffic classification with differential privacy | results table, including the worst ε |
| GML-Machine-Learning-On-Graphs | ASL recognition from skeleton graphs (GNN) | pipeline only; no accuracy claimed |
| chat-web | Small chat UI with a documented event protocol | the CORS bug it fixed is written down |
| TensorflowBasic · CPS · Tips-for-Github | Notes: TensorFlow basics, LTL specs, Git commands | cheat sheets, not tutorials |
Everything here is Apache-2.0. Issues and pull requests are welcome.


