👁️ DeepSeek Vision Bridge (deepseek-vis)

Vision for text-only LLMs (DeepSeek etc). Recovers images pasted into agent harness sessions directly from harness storage — no clipboard — converts them to ASCII layouts, OCRs text (macOS Apple Vision / Tesseract), overlays text at spatial coordinates.

Requirements

• Python 3.8+ + Pillow (python3 -m pip install Pillow)
• macOS: Xcode CLT for Apple Vision (xcode-select --install), optional
• Linux/Windows: Tesseract OCR

Install

git clone https://github.com/ashishsubedi/deepseek-vis.git
cd deepseek-vis
./install.sh             # interactive
./install.sh --target opencode   # or pi / all

Installs CLI to ~/.local/bin and skill to:

• OpenCode → ~/.config/opencode/skills/deepseek-vis
• Pi → ~/.pi/agent/skills/deepseek-vis

Other harnesses: copy SKILL.md, src/, bin/ into the skills dir manually.

Usage

deepseek-vis                           # auto-harvest pasted images from active harness
deepseek-vis clipboard                 # harvest directly from OS clipboard
deepseek-vis --harness <opencode|pi>   # specify explicit harness
deepseek-vis a.png b.png               # specific image files

Option	Description
-w <INT>	ASCII width (default 90; 120 complex UIs, 60 icons)
-m <mode>	hybrid (default) / ascii / ocr / overlay
--json	raw JSON
--no-edge	disable edge-sharpening

Files

SKILL.md          # skill definition (auto-loaded by harnesses)
install.sh        # installer (opencode, pi)
bin/deepseek-vis  # CLI entry
src/
  main.py       # CLI + source resolution (harness > file paths)
  harvest.py    # recover pasted images from harness storage
  ascii.py      # image → ASCII layout
  ocr.py        # Apple Vision / Tesseract OCR
  overlay.py    # OCR text overlaid on ASCII grid
  mac_ocr.swift # macOS Apple Vision helper (compile-on-demand)

Tests

python3 -m unittest discover tests

Uninstall

rm ~/.local/bin/deepseek-vis
rm -rf ~/.config/opencode/skills/deepseek-vis ~/.pi/agent/skills/deepseek-vis

License

MIT