From Zero to One: Build and Deploy a DeepSeek Harness Prompt Optimizer Plugin

DeepSeek Harness (DSH) overview

DSH is an extensible AI agent framework that separates Host (server‑side capabilities) and Client (browser UI) via a plugin architecture based on Slot, Remote, and Typert RPC mechanisms. All components—including models, tools, strategies, storage, and UI—are treated as interchangeable plugins.

Prompt‑optimization plugin

The plugin dsh‑ai‑prompt‑optimizer adds a “✨ AI 优化” button next to the chat input. When clicked, it reads the current draft, invokes the default model configured in DSH, receives a structured prompt that fills in task goal, context, constraints, and expected output format, and inserts the result back into the input box without sending automatically.

Development steps

Install DSH globally: npm install -g @deepseek-ai/dsh and start the web UI with dsh web.

Create a JavaScript plugin project with two files. Host.js implements server‑side logic: receives the draft, calls the model via ctx.llm.stream(...), and returns the optimized prompt.

Inject a React button into the chat UI using a Slot on the input box’s right side.

Register a type‑safe RPC method with Typert Remote, e.g. ctx.tools.register('promptOptimizer', handler).

Iterative refinement

Version 1 generated prompts by simple string concatenation.

Version 2 called the AI model directly via ctx.llm.stream.

Version 3 fixed a compatibility error (“unsupported provider request field at /temperature”) by removing the unsupported temperature field from the request payload.

Example RPC request

{"type":"client-request","rpcId":"6b77f21b-c3ad-46f9-89d3-8d817ee74420","method":"promptOptimizer/optimize","payload":{"args":{"request":{"draft":"今天的天气不错，我想做个系统"}}}}

Example RPC response

{"type":"server-response","rpcId":"6b77f21b-c3ad-46f9-89d3-8d817ee74420","result":{"ok":true,"value":{"ok":true,"prompt":"xxxxxxxxxx"}}}

Open‑source repository

Source code: https://github.com/wuchubuzai2018/dsh-ai-prompt-optimizer

Publishing to npm

Create an npm account and generate an access token.

Add the token to a local .npmrc file (do not commit this file).

Run npm login and then npm publish. Ensure the version follows semantic versioning.

Installation methods

From npm: dsh plugin --profile web add dsh-ai-prompt-optimizer From local source: clone the repo, run pnpm install, then dsh plugin --profile web add . Directly from GitHub:

dsh plugin --profile web add github:<owner>/<repo>

Uninstall

Remove the plugin with dsh plugin --profile web remove dsh-ai-prompt-optimizer.

Key observations

The plugin demonstrates the full DSH plugin lifecycle: architecture basics, coding, debugging, packaging, and distribution.

Using ctx.llm.stream enables streaming model calls without hard‑coding API keys.

Typert Remote provides type‑safe communication between the browser and Host.

Slot injection allows UI extensions without modifying core DSH code.