Overview

dsh-prompt-polish

Prompt polish ✨ — a DeepSeek Harness WebUI plugin that adds a one-click button to rewrite rough input into a structured, professional prompt, written back into the input box.

README / EN

Package documentation

dsh-prompt-polish

A DeepSeek Harness WebUI plugin that adds a one-click Prompt Polish ✨ button to the input toolbar. Rewrite a rough, fragmented request into a structured, professional prompt — written straight back into the input box.

What it does

You type a rough request into the input box.
Click ✨ 润色 at the right end of the toolbar.
The plugin calls your configured LLM (DeepSeek by default) to rewrite the request into a clear, actionable prompt.
The polished text replaces your draft. Review it, then send normally.

Polish is non-destructive and reviewable — nothing is sent automatically.

Install

If your profile uses dshmarket, install from the marketplace UI. Otherwise:

dsh plugin --profile web add dsh-prompt-polish

Then restart the WebUI so the bundle is loaded.

Building from source (e.g. while developing): the published package ships pre-built lib/ artifacts. If you must build yourself, run pnpm install && pnpm build (uses tsdown).

Configure

The plugin is mounted via its cordis.patch.yml. Override defaults by supplying config in the patch, or pass provider/model per request from the client:

- insert:
    - id: prompt-polish
      name: 'dsh-prompt-polish'
      config:
        provider: openrouter
        model: deepseek/deepseek-v4-flash-0731
        maxTokens: 2000

Setting	Default	Description
`provider`	(from request or unset)	LLM provider to use for polish
`model`	(from request or unset)	Model id to use for polish
`maxTokens`	`2000`	Output token cap per polish pass

If neither the config nor the request supplies provider/model, the request is rejected (no model route available) — configure one of the two so polish is always routable.

Endpoint

POST /dsh-prompt-polish/polish (WebUI server, same-origin only)

{ "text": "..." }

Response:

{ "ok": true, "text": "the rewritten prompt", "truncated": false, "provider": "...", "model": "..." }

text — the polished prompt.
truncated — true when the model hit the token cap; the text is still usable and is written back into the box.
On failure: { "error": "..." } with 4xx.

The endpoint enforces same-origin (Origin must match Host) to prevent cross-site abuse.

Architecture

Server (src/index.ts) — registers the HTTP route, streams the LLM via ctx.llm.stream + BlockAssembler, and returns the rewritten text. Peer deps: @deepseek-ai/cordis, @deepseek-ai/dsh-llm.
Client (src/client/index.tsx) — registers a button on the conversation.input.right slot, fetches the endpoint, and writes back via inputActions.setDraft. CSS is self-contained (one injected <style> tag) so the browser bundle builds with a plain tsdown config.

Both halves build with tsdown. The client bundle is emitted in DSH's window.__ModuleLoader__.load({ id, factory }) format and externalizes react / react/jsx-runtime to the browser module table.

License

MIT. See LICENSE.

dsh-prompt-polish