dsh-math-input
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.
Why you need it
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.
Install
Prerequisites
- DeepSeek Harness
>= 0.1.1-rc.2
- Node.js
>= 20.0.0
- Chrome or Edge (for WebGPU support and SharedArrayBuffer)
Install the plugin
# 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.
Uninstall
# 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.
Usage
Click the "+" button to the left of the input row to open the menu:
1. Handwriting input
- Click "Handwriting input" in the menu — a canvas modal pops up
- Draw a formula with mouse, touch, or stylus
- Stop for ~1.5 seconds (configurable in Settings) — the engine recognizes automatically
- The result renders as a KaTeX preview; the LaTeX source is editable below
- Click "Confirm and insert" — the formula enters the composer as
\[ ... \] and renders inline
- Click "Clear" to start over, "Undo" to remove the last stroke
2. Screenshot OCR
- Click "Screenshot and recognize" in the menu
- Ctrl+V to paste an image, or click "Upload file" to select a local image
- The engine extracts LaTeX from the image automatically
- Review the preview and click "Confirm and insert"
3. LaTeX syntax editor
- Click "LaTeX syntax editor" in the menu — a dock panel expands below the input row
- The palette offers Greek letter buttons (
\alpha, \beta, \pi, etc.) and template buttons (\frac{}{}, \sqrt{}, \sum_{}^{}, etc.)
- Click any button to insert the corresponding LaTeX code into the editor
- A live KaTeX preview renders on the right
- Click "Insert" — the formula enters the composer as
\[ ... \]
Inline rendering
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.
Settings
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.
Local testing
No need to push to GitHub or publish to npm — follow these five steps to verify locally.
Step 1: Build the project
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).
Step 2: Link the local package
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.
Step 3: Launch DSH
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.
Step 4: Verify functionality
In the DSH web UI, check each item:
- A "+" button appears to the left of the input row
- Clicking "+" opens a menu with three items: Handwriting, Screenshot, LaTeX editor
- Handwriting pad: modal renders, canvas accepts pointer drawing
- Settings: Settings → Math Input shows five controls
- Changing a setting persists across page reload
- Type
\[x^2\] in the composer — a KaTeX chip renders below
- LaTeX editor dock: toggles from the menu, palette inserts snippets, Insert writes
\[...\]
Step 5: Re-test after code changes
After modifying code:
# rebuild
npm run build
# stop DSH (Ctrl+C), restart
npx @deepseek-ai/dsh web --no-open
Notes & limitations
- Model download: first recognition triggers a ~7.2 MB download (encoder 3.4 MB + decoder 4.0 MB + vocab 4 KB); subsequent loads use IndexedDB cache.
- WebGPU auto-fallback: when
webgpu is selected but the browser doesn't support it (requires Chrome 113+), the engine automatically falls back to wasm.
- SharedArrayBuffer: ONNX Runtime Web uses multi-threaded WASM when
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.
- Handwritten vs printed: the CoMER model is trained on the CROHME handwritten math expression dataset and optimized for handwriting. Screenshot OCR of printed formulas may underperform.
- Browser compatibility: requires a modern browser with WebAssembly SIMD support. Chrome / Edge recommended; Firefox mostly works; Safari has limited support.
- Zero-token guarantee: the plugin never calls
ctx.llm — all recognition runs locally in the browser, with zero API costs.
Development
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.
Documentation
License
MIT
