dsh-prompt-optimizer

dsh-prompt-optimizer

English | 中文

One-click prompt polishing for the DSH composer: select nothing, just type a draft and press ✨ (or Alt+O) — the plugin rewrites it into a clearer, better-structured prompt. Zero-config by default: it follows the current session's model through the harness host services, so no API key is needed. A self-configured OpenAI-compatible endpoint is supported as an alternative.

Features

• One-click optimize — a ✨ button on the composer's right; Alt+O while the composer is focused does the same
• True streaming preview — real SSE over the harness webServer: every text-delta from llm.stream is pushed immediately and rendered token by token; reasoning is streamed first, so you watch the model think while it works
• Zero-config default — reuses the current session/agent default model (host agentDefaultModel + llm services), no API key required
• Custom endpoint mode — uncheck "follow session model" and plug in any OpenAI-compatible /chat/completions endpoint (base URL + key + model)
• Action row on completion — replace the draft in place, copy, re-optimize, or dismiss
• Bilingual UI — follows the DSH language (中文 / English) live, no reload
• Self-hosted config — settings persist in ~/.dsh/prompt-optimizer-config.json via a loopback RPC channel, independent of the host settings registry
• Dark-mode ready — all colors follow DSH theme variables; fixed brand blue + white text in deep-night mode
• Local-only credentials — the API key (custom mode only) lives in the local config file and goes only to the endpoint you configured

Screenshots

Requirements

• DSH with a web or desktop profile
• Node.js ^22.19.0 or >=24.0.0 (only needed to build from source)
• pnpm is recommended when installing into a profile

Install

The bundle entry (id: prompt-optimizer) is self-declared by this package's cordis.patch.yml — no manual patch file is needed.

From npm

dsh plugin --profile desktop add dsh-prompt-optimizer

For a web profile, use --profile web. Restart DSH (quit fully and reopen), then a ✨ button appears to the right of the composer.

From source (development)

git clone https://github.com/winditer/dsh-prompt-optimizer.git && cd dsh-prompt-optimizer
npm install
npm run build          # produces dist/client.js
dsh plugin --profile desktop add .    # links the workspace into the profile by package name

Or install by hand: in the target profile's package.json (e.g. ~/.dsh/profiles/desktop/package.json):

{
  "dependencies": {
    "dsh-prompt-optimizer": "link:/absolute/path/to/dsh-prompt-optimizer"
    // ...
  },
  "dsh": {
    "profile": {
      "bundles": [ /* ... */, "dsh-prompt-optimizer" ]
    }
  }
}

then pnpm install inside the profile directory and restart DSH.

Uninstall

Remove dsh-prompt-optimizer from the profile's dependencies and dsh.profile.bundles, clean up the installed package, and delete the config file ~/.dsh/prompt-optimizer-config.json if you no longer need it.

Usage

• Type a draft in the composer, click ✨ (or Alt+O) — the preview card appears over the composer
• While optimizing: reasoning text scrolls in secondary color first, then the polished prompt streams in token by token
• When done: 替换草稿 writes the result into the composer in place · 复制 copies it · 重新优化 re-runs · 放弃 dismisses
• The preview belongs to the session where you started it: switching sessions hides it, switching back restores it

Configuration

Open 设置 → 通用设置 → Prompt 优化:

Setting	Default	Meaning
使用当前会话模型	on	Follow the session/agent default model (zero-config). Off: enable the fields below
接口地址 (base URL)	https://api.deepseek.com	Any OpenAI-compatible /chat/completions endpoint
API Key	—	Your key for the custom route (ignored in follow mode)
模型名	deepseek-v4-flash	Model name (ignored in follow mode)

Settings are saved in ~/.dsh/prompt-optimizer-config.json (same directory as other DSH config; removed with the plugin).

Custom endpoints must support CORS and SSE streaming (official DeepSeek, OneAPI-style gateways work).

Architecture

Two halves, one package:

• Host half — lib/index.js. Persists config over a loopback RPC channel (/dsh-prompt-optimizer, get/set), and registers an HTTP JSON API at /dsh-prompt-optimizer/api through the harness webServer service. Runs the session-default optimization via llm.stream; background streams live in an in-memory Map cleared on unload.
• Client half — src/*.ts, bundled to dist/client.js (esbuild, wrapped in __ModuleLoader__.load({ id: "dsh-prompt-optimizer", … }); the id must equal the installed package name). Renders into the conversation.input.right button, conversation.input.overlay preview card and settings.general.item row; talks to the host with fetch POSTs.

Host API

All endpoints are POST /dsh-prompt-optimizer/api/<method>; every response is { ok: true, value } or { ok: false, error }.

Method	Body	Returns
sessionModel	{}	{ provider, model, reasoningEffort? } — the session's default model
optimize.stream	{ provider, model, text, system?, reasoningEffort? }	text/event-stream — event: reasoning frames first, then event: delta per token, event: done at the end
optimize.start	{ provider, model, text, system?, reasoningEffort? }	{ taskId } — background accumulation (fallback path)
optimize.poll	{ taskId }	{ done, text, error? } — accumulated text while streaming
optimize.abort	{ taskId }	{ ok }

Protocol details: only POST is accepted (405 otherwise); the body is JSON with a 1 MB cap; unknown methods return 404.

Security notes

• Default route sends no credentials — it reuses the harness's configured provider.
• The custom-mode API key stays local (~/.dsh/prompt-optimizer-config.json), and only goes to the endpoint you configured.
• Optimizations appear only in the preview card; the polished text reaches a session only if you press 替换草稿.

Development

npm run build   # esbuild: src/index.ts → dist/client.js (__ModuleLoader__ bundle)
npm test        # node runner over tests/entry.ts (state machines, channels, SSE parser)

Project layout

src/index.ts          Client entry — slot wiring, RPC/config glue, host probes
src/OptimizeButton.tsx / PreviewCard.tsx / SettingsRow.tsx
src/optimizer.ts      Config defaults, system prompts, OpenAI-compatible fetch/SSE client
src/session-optimizer.ts  Host channel: sessionModel + SSE stream + fallback poll
src/preview-state.ts  Preview card state machine (pure reducer)
src/preview-bus.ts    Module-level event bus shared by button / card / orchestration
lib/index.js          Host half — config persistence + HTTP JSON API (makeHandler + createApiRoute)
dist/client.js        Built client bundle (__ModuleLoader__ format, load id = dsh-prompt-optimizer)
cordis.patch.yml      Bundle entry declaration (insert: { id: prompt-optimizer, name: dsh-prompt-optimizer })
scripts/build.mjs     Build script (esbuild + bundle wrapper)
tests/entry.ts        Unit + integration tests (61)
assets/               Screenshot

Gotchas (learned the hard way)

• Bundle id must equal the package name — arrive() throws bundle loaded without registering <id> otherwise. scripts/build.mjs hardcodes the correct id.

• Profile bundles don't get the cordis timer service — use plain browser setInterval/setTimeout (disposed in React effect cleanup), exactly like the sibling dsh-elf bundle.

• Do not use session.create/fork for generation — a background session never executes (the renderer's fabricated ids are silently rejected, forked sub-sessions don't trigger the model), which manifested as "optimizing forever". Drive the model from the host half via llm.stream instead.

• Do not run the streaming protocol over connection.rpc.call — on desktop the renderer's rpc.call hangs on the second call within one flow (verified: sessionModel ok, next call never arrives). Host channels go over HTTP (webServer).

• Prefer link: over file: when installing a workspace copy — file: copies files, so edits/rebuilds go stale.

• Client changes go live on page refresh; host changes require a full DSH restart.

• A broken build script silently keeps the old bundle — npm run build must print ✓ Built; if it only prints a Node version banner, the script is failing (a past regression left a stale dist/client.js that looked "current").

• Fresh publishes can trip the profile's minimumReleaseAge policy — if the profile enforces pnpm's release-age supply-chain check, a version published less than ~24 h ago fails dsh plugin … add <pkg> with ERR_PNPM_MINIMUM_RELEASE_AGE_VIOLATION. Add the exact name@version to minimumReleaseAgeExclude in the profile's pnpm-workspace.yaml (and keep the entry current when you release a new version):

  minimumReleaseAgeExclude:
    - dsh-prompt-optimizer@2.0.0
  

License

MIT