Fishsb/dsh-prompt-enhancer

A DeepSeek Harness (DSH) plugin with two core capabilities:

• ✨ Prompt enhancement — rewrite composer drafts in place with one click, fully undoable
• 💬 Voice recognition — speech-to-text that stops on silence, with cloud / local dual engines; the transcript is filled into the draft

✨ Two core features

1. Prompt enhancement (✨)

The ✨ button in the composer toolbar triggers an independent LLM call and rewrites the current draft in place; keep refining, undo anytime, or cancel while enhancing.

• One-click enhance — the ✨ button triggers an independent LLM call and replaces the draft; continue refining, undo anytime, cancel while enhancing
• 5 optimization modes — Basic (direct) / Lite (previous-round context) / Standard (rules + retrieval) / Expert (task analysis + full retrieval) / One-click Publish (complete dev-spec generator)
• Memory switch — when on, pre-send rounds (optimize → edit → re-optimize) accumulate into a memory chain the next round replays and senses your edit direction; sending the message clears it; when off, nothing is read or written
• Model chain — try multiple models in order, reorder, toggle thinking, run inline connectivity tests

2. Voice recognition (💬)

The 🎤 record button beside the composer starts listening; the transcript (cloud Qwen3-ASR / local offline SenseVoice dual engines) → optional refinement (de-filler) → filled into the draft → can be enhanced with one click. Stops automatically on silence (VAD), audio stays in memory only and is never written to disk.

• Dual engines — cloud Qwen3-ASR / local offline SenseVoice (framework + optional download, slim release)
• Stops on silence — VAD silence detection, no manual stop needed (auto-stops after at most 60 seconds)
• Hotkey wake — recordable global hotkey, tap / long-press dual trigger
• Transcript refinement — optionally refine the transcript through the base LLM to remove filler words
• Auto-enhance — when on, filling the draft after recognition automatically triggers prompt enhancement

🔧 Other capabilities

• 🌐 i18n — follows the DSH interface language (中文 / English)

🚀 Install

dsh plugin --profile web add github:Fishsb/dsh-prompt-enhancer#v3.4.0

Restart DSH (dsh web) after installing — the ✨ button appears in the composer toolbar.

ℹ️ Version note: the latest tag v3.4.0 (2026-09-19) already contains the ✨ official slot-contract fix (Issue #8 / #10) and the cloud ASR fix (Issue #9); the command above installs that tag. If you previously installed #main under the old note, re-running it pins you to a released version. Note: from v3.4.0 the in-plugin restart capability is removed (restart DSH manually after an update) — see release notes.

Requires DeepSeek Harness installed locally and pnpm in PATH.

Client compatibility (voice recognition): 🎤 voice input relies on the client injecting inputActions.setDraft (the official web client satisfies this); third-party clients implementing the same contract can load it, and when the capability set differs, voice input degrades gracefully (no insert capability → transcript appended to the end of the draft; no injection at all → 🎤 disabled with a notice). The local offline engine uses a "framework + optional download" model: the plugin ships without and does not auto-download models; go to Settings → Model configuration → 💬 Voice recognition → set engine to "Local" → in the "Local model" area click Download model (SenseVoice 228MB, with progress), and it takes effect automatically when done. See compatibility notes (client dependency boundary and slot contract).

Composer toolbar (✨/🎤) client contract: the composer buttons and the error strip mount on the session-scoped slots conversation.input.right / conversation.input.dock. The official renderer (@deepseek-ai/dsh-client-ui-renderer ≥ 0.1.2-rc.1; web and DSH Desktop share the same source) injects a sessionId prop, useSession/useInput selector hooks, and an inputActions prop into slot entries (it never provides props.session / props.input); the plugin reads state through that contract since that fix (Issue #8 / #10, commit 0197ae7, released with v3.4.0) and keeps a fallback for older hosts that provide the props.session / props.input shape. Third-party client renderers must implement the same contract (sessionId plus the hooks and actions above) for ✨/🎤 to render.

Update / remove:

dsh plugin --profile web update dsh-prompt-enhancer
dsh plugin --profile web remove dsh-prompt-enhancer

After remove, restart DSH to fully unload the running instance.

After an update is installed, restart DSH manually for it to take effect (the plugin no longer restarts DSH for you).

📦 Library notes

Core logic lives in standalone Node modules, reusable from other scripts: lib/updater-host.cjs (update executor: download / verify / install / rollback), lib/platform-service.cjs (cross-platform service management), lib/sys.cjs (env & paths). See each module's header comments.

🎯 Usage (prompt enhancement)

1. Type any non-empty text (slash commands keep their prefix; only the body is optimized)
2. Click the ✨ button
3. Wait for the independent LLM call; the draft is replaced with the enhanced version
4. Not satisfied? Click Undo to restore the original

📸 See it work

Voice recognition (the 🎤 record button, stops on silence):

Voice recognition settings (engine switch / hotkey wake / model download / refinement):

⚙️ Configuration

Settings → "Models & plugins":

Tab	Description
Model configuration	Configure the optimization model chain: tried in order, reorderable; the Voice recognition section (engine switch / hotkey wake / local model download / refinement)
Optimization parameters	Mode / memory switch / context budget / timeout & output limits / templates

📚 Docs

• Releases
• CHANGELOG
• Compatibility notes

Privacy: the plugin records or reports nothing; enhanced results come from an external LLM — verify before sending.

🌐 Downloading voice models on restricted networks

Local models are hosted on Hugging Face. If your network cannot reach it directly (common in mainland China):

1. Use a local proxy: keep your proxy client running (the system proxy switch can stay off); add the top-level field "download": { "proxy": "http://127.0.0.1:10808" } (or socks5://…) to the config file %DSH_HOME%\dsh-prompt-enhancer.config.json; after saving, click download again to go through the proxy;
2. Place the files manually: download the model files from hf-mirror.com into %DSH_HOME%\dsh-prompt-enhancer-asr\models\<model id>\ (sense-voice needs model.int8.onnx + tokens.txt); refresh the settings page and it is detected as installed;
3. Downloads have built-in resume support and automatic multi-source switching (HuggingFace ↔ hf-mirror), so an occasional interruption can simply be retried to resume.