dsh-prompt-boost-pro — enhance the draft before you send it
English | 中文
A sparkle button in the DeepSeek Harness (DSH) composer tool row. Click it, pick a mode, and the
model rewrites your draft into a better prompt — shown as an original-versus-enhanced preview that
you accept, re-run in the other mode, or discard. Nothing touches your draft until you accept.
Inspired by Trae CN's "optimize input": type a rough idea, get a directly executable instruction.
Unlike a one-shot rewrite, this plugin keeps you in control — you always see both versions first, and
after accepting, the same button becomes an undo until you change or send the text.
Features
- Two modes, picked per use — Structured reorganises the draft for the task at hand (goal,
constraints, expected output, acceptance), Light only disambiguates, fixes wording and tightens;
both keep your intent and your language, and neither invents requirements you did not write.
- It never interrogates you. The prompt policy forbids handing back information the agent can look
up itself (repo layout, stack, scope, previous conclusions). At most one
[待补充: …] placeholder
survives, and only for a decision that is genuinely yours.
- Comparison preview, never a silent overwrite — original and enhanced side by side; accept,
re-run in the other mode, copy, or discard. If you edited the draft while the preview was open, it
says so and makes the overwrite explicit.
- Undo that knows when it is over — after accepting, the button reads "undo" only while the draft
still equals the enhanced text; sending or editing the draft returns it to normal instead of offering
an undo that cannot work.
- Truncation is surfaced, not swallowed — output budget adapts to the draft (4000 → 8192 tokens) and
a capped answer is still returned, flagged, with a hint instead of an error.
- Model route resolved four ways — plugin config → the model currently shown in the picker → the
session's last request → the deployment default. A brand-new session works on the first click.
- Dismissals everywhere — click the icon again, click anywhere outside, or press
Esc; the preview
also closes with Esc. The button stays clickable while its menu is open, so the menu can never trap
you.
- Bilingual UI (zh / en), theme-native (real DSH design tokens), no tracking, no config files
written.
Requirements
- DSH web GUI (
dsh web), harness 0.1.2-alpha.1 or newer; the current line (0.1.5-rc.x) is tested.
- Node
^22.19.0 || >=24.0.0 (the host half runs inside the harness process).
- A session whose model is routable — the default model counts, so no setup is needed.
Install
# from the profile that runs your web GUI (usually "web")
dsh plugin --profile web add github:xiaoshengliang2002/dsh-prompt-boost-pro
# or from a released tarball
dsh plugin --profile web add https://github.com/xiaoshengliang2002/dsh-prompt-boost-pro/releases/latest/download/dsh-prompt-boost-pro.tgz
# or from a local checkout (symlinked; rebuild shows up without reinstalling)
dsh plugin --profile web add link:/path/to/dsh-prompt-boost-pro
Then restart dsh web and reload the page. The lib/ bundles are committed, so installing from source
needs no build step and no pnpm build approval.
Uninstall: dsh plugin --profile web remove dsh-prompt-boost-pro.
Where the button lives
+ | permission control | ✨ enhance | summon-expert entry | … | model | context meter | send
It registers conversation.input.left with order: -10, so it sits immediately after the permission
control. Move it after the summon-expert entry by changing order to 20 in src/client/index.ts.
How it works
| Half | Role |
|---|
Host (lib/index.js) | Registers POST /api/prompt-boost-pro/enhance on the shared API channel (same-origin, cookie-authenticated), resolves the model route, calls ctx.llm.stream() once, cleans the text and answers JSON. |
Browser (lib/client.js) | Registers two slots: the tool-row button with its mode menu, and the preview dialog in the composer overlay. Keeps one state machine per session and writes the draft through inputActions.setDraft only on accept. |
Wire contract, error codes and the prompt policy are documented in
docs/design.md; the awesome-dsh-plugin checklist is in
docs/awesome-dsh-plugin-checklist.md.
Development
The repository is a normal npm package; lib/ is build output that is committed on purpose (so source
installs are build-free).
npm run typecheck # tsc --noEmit, 0 errors
npm run build # esbuild → lib/index.js + lib/client.js
npm test # 4 offline suites: host 28, client 28, UI dismissal/undo 40, prompt policy 27
npm run release:pack # sync docs + build + npm pack
npm run release:verify # unpack the tarball and assert 31 contract/listing conditions
No network is required: node scripts/link-dsh.mjs links the packages of a local DSH checkout into
node_modules (esbuild, typescript, react, @types) so the build and tests run offline.
preview/index.html is a clickable preview of the composer UI built from DSH's real design tokens and
this plugin's real stylesheet; preview/shot-preview.cjs renders the state screenshots into
preview/shots/ (declared for the storefront via screenshots.json).
Version history
- 0.1.2 — reasoning-effort passthrough. The client sends the model picker's current reasoning
effort (low/high/max) with each request and the host passes it to the LLM call; deployments can pin
one via the new
reasoningEffort config key. Fixes 400 {"code":"1210"} on models that always
think (e.g. GLM-5.3-Flash), whose default effort is "off" and gets rejected by the upstream API.
- 0.1.1 — upgrade-window compatibility. The host registers the legacy
POST /api/prompt-boost/enhance
path alongside the current one, and the client falls back to it when the current path answers 404/405 —
so a half-updated pair (page refreshed but server not restarted, or the reverse) keeps working instead of
failing with "unexpected response (HTTP 404)".
- 0.1.0 — first release. Two enhancement modes (structured / light); original-versus-enhanced
preview before applying; undo that expires with the draft; click-outside/
Esc dismissal; adaptive
output budget with a truncated flag instead of dropping a capped answer; four-level model route
resolution so a brand-new session works on the first click; bilingual UI and a custom sparkle icon.
123 offline assertions plus 34 packaging checks. Wire path: POST /api/prompt-boost-pro/enhance; the
button sits after the permission control.
License
MIT
