Fishsb/dsh-prompt-enhancer
One-click prompt enhancement: an independent LLM call rewrites your rough draft in the composer, fully undoable.
Install
via GitHub · GitHub source
Installing from GitHub runs the project's build script, which pnpm blocks until you allowlist it — run the command once and pnpm prints the exact key to add under `allowBuilds` in ~/.dsh/profiles/web/pnpm-workspace.yaml.
What happened when we ran it
Installed cleanly when we ran it
Every command here is run in a throwaway container against a clean profile, and the result is whatever the harness recorded — not a guess from the source. Last run 6d ago.
Due diligence
What it is — DeepSeek Harness plugin: one-click enhancement of fuzzy prompt drafts via independent LLM call, with direct replacement and full undo support.
Who it is for — If you are inputting fuzzy prompt drafts in the DeepSeek Harness composer and need independent LLM optimization, install this plugin. Users who have already clear prompt drafts can skip installing this plugin.
Watch out — Sandbox test passed: installed in a fresh profile and registered with harness. No obvious issues found.
The verdict — I would install it because it supports true cancellation during the enhancement process.
Generated by grok-4.6, and a starting point rather than a verdict. Where it says a plugin installs or does not, that is from a real run in a clean profile — everything else is read off the repository. Trust the source over this.
dsh-prompt-enhancer is a DeepSeek Harness plugin that enhances composer drafts with a separate LLM call, and can also turn speech into draft text. DeepSeek Harness loads the plugin into the web profile, where an ✨ composer action reads the current non-empty draft, sends it to the selected model chain, and replaces the draft with the result. The original text remains undoable; enhancement can be repeated or cancelled. Five modes range from direct rewriting to generating a complete development specification. An optional memory setting carries an optimisation–revision chain into the next optimisation and clears it when a message is sent.
The plugin also adds a 🎤 recording action. Speech uses cloud Qwen3-ASR or an optionally downloaded, local SenseVoice engine, with voice-activity detection to stop after a pause. Recognition can be normalised through the base LLM and can trigger enhancement automatically. Recordings are held in memory rather than written to disk. The web client must provide inputActions.setDraft; clients without that contract disable voice input or append recognised text instead.
dsh-prompt-enhancer suits developers who want draft rewriting or voice entry within the composer. It is a poor fit when external LLM processing is unsuitable, when pnpm is unavailable on the local machine, or when a client does not support the documented draft-insertion contract. Its separate restart utility can also create a desktop shortcut or run a Node command to restart the dsh-web service.
DeepSeek Harness exposes the plugin settings under Settings → Model and Plugins:
| Area | Controls |
|---|---|
| Model configuration | Ordered optimisation model chain, model enablement, thinking toggle, inline connectivity tests, and voice engine settings |
| Optimisation parameters | Optimisation mode, memory switch, context budget, timeout, output limit, and template |
| Voice recognition | Cloud Qwen3-ASR or local SenseVoice, global shortcut, local model download, text normalisation, and automatic enhancement |
The local SenseVoice model is optional. Selecting the local engine and choosing Download model downloads the documented 228 MB model with progress reporting.
DeepSeek Harness manages installation, updates, and removal through its plugin commands. The README documents dsh plugin --profile web update dsh-prompt-enhancer and dsh plugin --profile web remove dsh-prompt-enhancer; removal takes effect after restarting DeepSeek Harness.
The independent restart executor can be called without the web page:
node "<DSH_HOME>\AppData\Local\dsh-prompt-enhancer\executor\0.1.11\lib\updater-host.cjs" --cli restart --service dsh-web --profile web
The plugin can also generate a desktop shortcut from its port-restart settings. The executor supports service restart with process-level fallback.
The ✨ action rewrites the current draft, while slash-command prefixes are preserved and only the body is optimised. The voice action detects a pause, recognises speech, optionally normalises it, and inserts the result into the draft. Enhancement history is read and written only while the memory switch is enabled; sending a message clears that chain.
DeepSeek Harness must already be installed, and pnpm must be available on PATH. Voice insertion depends on the client injecting inputActions.setDraft: the official web client satisfies this contract, while other clients may degrade or disable the feature. The local engine uses an optional model download rather than bundling the model with the plugin.
Written from the project's own documentation and kept in sync with it. Where the two disagree, the source is authoritative — read the README on GitHub
Same category
zhu1090093659
Plugin and skin collection for the DSH Web UI: task board, Git graph, right-side panel, remote mobile UI, pet, live token stats, and a skin center.
No one-line install — this plugin lives inside a larger repository and publishes no npm package.
ccch1mneyyy
Claude Code-style full-screen terminal UI: pixel-whale header, live status line, and streaming thought expansion.
Installed cleanly when we ran it
omdsh-dev
Full sidebar workbench with file rendering and editing, terminal, Git, and subagents; third-party plugins can register new tabs.
Installed cleanly when we ran it
omdsh-dev
Codex-style `@file` mentions: search workspace files in the composer and attach their contents to prompts.
Installed cleanly when we ran it
huiliyi37
A terminal UI (TUI) for DeepSeek Harness.
Installed cleanly when we ran it
omdsh-dev
Interactive UI components rendered inline in replies: layout, charts, forms, quizzes, mermaid, 3D scenes, and an action event loop back to the model.
Installed cleanly when we ran it
