dsh-math-input
DeepSeek Harness 的零 token 离线数学输入：手写识别、截图 OCR、LaTeX 编辑器和内联 LaTeX 渲染。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:xiaxi626/dsh-math-input#489f50a2219d34164af85de3d8b7782db3d18358Math Input 以 dsh-math-input 发布，当前版本为 0.1.0。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
English | 中文
A zero-token, fully offline math input plugin for DeepSeek Harness (DSH). Handwrite formulas, OCR them from screenshots, or edit LaTeX directly — all recognized in your browser, no API key, no token cost.
DSH's composer is plain text. If you want the model to reason about a formula, you type LaTeX by hand — slow and error-prone. This plugin adds three input methods plus inline rendering:
| Input method | Best for | Engine |
|---|---|---|
| Handwriting | Have a stylus / mouse, want to write a formula fast | CoMER neural network (ONNX Runtime Web) |
| Screenshot OCR | Already have a formula image (PDF screenshot, etc.) | Same engine, image-to-tensor recognition |
| LaTeX editor | Know LaTeX syntax, want a symbol palette | No recognition needed, direct input |
| Inline rendering | Type \[ ... \] in the composer, auto-renders | KaTeX |
All recognition runs via ONNX Runtime Web (WASM / WebGPU). The model is 7.2 MB, downloaded once and cached in IndexedDB. The plugin never calls ctx.llm — your token bill stays at zero.
>= 0.1.1-rc.2>= 20.0.0# from GitHub
dsh plugin --profile web add github:<owner>/dsh-math-input
# without global dsh CLI
npx @deepseek-ai/dsh plugin --profile web add dsh-math-input
Restart the DSH profile after install (stop and re-run dsh web). A "+" button appears to the left of the input row — that means the install succeeded.
# remove from DSH profile
dsh plugin --profile web remove dsh-math-input
# or without global CLI
npx @deepseek-ai/dsh plugin --profile web remove dsh-math-input
Restart the DSH profile. The "+" button and all input windows will be removed.
The model cache (IndexedDB database
math-handwrite-models) is not automatically cleared. To clean it up manually, go to browser DevTools → Application → IndexedDB and delete the database.
Click the "+" button to the left of the input row to open the menu:
\[ ... \] and renders inline\alpha, \beta, \pi, etc.) and template buttons (\frac{}{}, \sqrt{}, \sum_{}^{}, etc.)\[ ... \]Any \[ ... \] you type (or paste) in the composer renders inline as a formula. For example, typing \[x^2 + y^2 = r^2\] renders the equation directly.
Open Settings → Math Input to configure:
| Setting | Description | Options |
|---|---|---|
| Recognition mode | Limits the recognition vocabulary | auto (all), number (digits & operators), expression (math expressions) |
| Beam width | Quality vs speed trade-off | 1 (fastest), 2, 3 (best quality) |
| Execution provider | ONNX inference backend | wasm (universal), webgpu (needs Chrome 113+, typically 2–5x faster) |
| Stroke debounce | Idle time before auto-recognition | 0.3 – 10 seconds |
| Interface language | Override UI language | zh / en, empty follows DSH language |
Settings persist on the Host side and survive page reloads.
No need to push to GitHub or publish to npm — follow these five steps to verify locally.
In the dsh-math-input/ directory:
cd dsh-math-input
npm install
npm run build # generates lib/ directory
To check types without building, run
npm run typecheck(tsc only, no output).
This plugin has both a Host entry and a Client bundle (dsh.client in
package.json). The Client is discovered through node_modules, so the
package must be linked first — an overlay alone won't load the UI.
Windows (Git Bash / MINGW):
PROJECT="$(cygpath -m ~/Downloads/dsh-math-input)" # ← your path
npx @deepseek-ai/dsh plugin --profile web add "file:$PROJECT"
macOS / Linux:
PROJECT="$(pwd)" # ← run from the repo root
npx @deepseek-ai/dsh plugin --profile web add "file:$PROJECT"
This creates a persistent link in the profile's node_modules. After code
changes, just npm run build and restart — no reinstall needed.
npx @deepseek-ai/dsh web --no-open
Open http://127.0.0.1:3080.
Host-only testing without linking: if you only need to test Host-side code (settings, typert manifest) without the Client UI, use an overlay patch instead. See Local testing guide for details.
In the DSH web UI, check each item:
\[x^2\] in the composer — a KaTeX chip renders below\[...\]After modifying code:
# rebuild
npm run build
# stop DSH (Ctrl+C), restart
npx @deepseek-ai/dsh web --no-open
webgpu is selected but the browser doesn't support it (requires Chrome 113+), the engine automatically falls back to wasm.Cross-Origin-Opener-Policy: same-origin and Cross-Origin-Embedder-Policy: require-corp headers are present. Without them, it falls back to single-threaded — recognition still works but is slower.ctx.llm — all recognition runs locally in the browser, with zero API costs.npm run typecheck # tsc --noEmit (strict)
npm run lint # ESLint 9
npm test # node:test + tsx
npm run build # tsdown + tsc → lib/
CI runs across Node 20 / 22 / 24: typecheck, lint, unit tests, build, and a check that committed lib/ matches a fresh build.
MIT