AI-QR-code-scanning-

QR / barcode scanning for DeepSeek Harness (DSH): give the model eyes for codes — from image files, the live screen, or a webcam — plus QR generation.

Decoding is powered by zxing-cpp compiled to WebAssembly (zxing-wasm), the same C++ core behind most desktop scanners, so rotated, inverted, tiny, perspective-skewed and multi-symbol images all work.

Tools

Tool	What it does	Input
qr_scan	Decode every code in an image, with positions	path, glob, http(s) URL
qr_scan_screen	Capture the screen (optionally a region) and decode	region: "x,y,w,h"
qr_scan_camera	Grab one webcam frame and decode	optional device name
qr_make	Render text/URL as a QR PNG	writes to .dsh-qr/ in the workspace

{
  "ok": true,
  "source": "C:\\shots\\poster.png",
  "results": [
    { "text": "http://cloud.example.com/embed/76e3ca", "format": "QR Code", "kind": "url", "box": [215, 403, 300, 489] }
  ]
}

kind classifies the payload as url / wifi / contact / email / phone / sms / numeric / text.

Install

From GitHub (recommended; add the dsh-plugin topic and the marketplace picks it up automatically):

dsh plugin --profile web add github:mkingyixing/AI-QR-code-scanning-

From npm: not published yet. npm rejects uppercase package names, so publishing would need a lowercase name (e.g. ai-qr-code-scanning) — see PUBLISHING.md.

Local install (desktop profile, no network):

powershell -ExecutionPolicy Bypass -File .\scripts\install.ps1

Restart DeepSeek Harness afterwards. Uninstall with -Uninstall.

Why a plugin instead of a script

On the shared 10-image corpus (rotation, inversion, 78px thumbnail, blurry JPEG, perspective, two codes in one frame) zxing-cpp scores 10/10 — see tools/_bench/results.md. Running it in-process removes the interpreter startup:

Metric	Python script	This plugin
One decode (cold)	~750 ms	92 ms (WASM load)
One decode (warm)	~750 ms	13.6 ms
Full-screen capture + decode	~1 500 ms	227 ms

Config

- insert:
    - id: tool-qr-scanner
      name: AI-QR-code-scanning-
      config:
        outputDir: .dsh-qr
        maxInputBytes: 20971520
        timeoutMs: 20000

Privacy

Everything runs locally: WASM decoding, GDI screen capture, ffmpeg frame grabs. No network calls unless you pass an explicit http(s) URL to qr_scan. Screenshots stay in memory unless save: true.

Limitations

• Screen capture uses Win32 GDI via koffi, so it is Windows-only; qr_scan and qr_make still work elsewhere.
• Webcam capture needs ffmpeg on PATH (or DSH_FFMPEG).
• Image formats: PNG/JPEG/WebP/GIF/BMP/TIFF/AVIF (decoded by sharp).

License

MIT