pi-vision-bridge

Transparent image support for text-only models in the Pi coding agent.

Why this exists: DeepSeek V4 (and other text-only models) can't see images. When you paste a screenshot or the agent reads an image file, the content is silently dropped. This extension bridges that gap — it forwards images to Gemini for description and injects the text so any model can "see" them.

Works with any text-only model, not just DeepSeek. Vision-capable models (Claude, GPT, Gemini) are skipped automatically.

What's included

Vision Bridge Extension

• Intercepts pasted images (Ctrl+V) and read tool calls on image files
• Forwards to Gemini 2.5 Flash for fast, cheap descriptions (~$0.001/call)
• UI-debugging-optimized prompt — calls out spacing, alignment, glitches
• In-memory cache avoids re-processing the same file
• /vision-proxy command toggles on/off at runtime
• PI_VISION_MODEL env var to override the vision model

Setup Skill (optional)

Guided skill for configuring text-only models in Pi — models.json, thinking levels, and compaction tuned for large context windows. Invoke with /skill:deepseek-setup.

Install

pi install pi-vision-bridge

Or locally:

pi install /path/to/pi-vision-bridge

Requirements

• GEMINI_API_KEY environment variable

Configuration

# Override the vision model (default: gemini-2.5-flash)
export PI_VISION_MODEL=gemini-2.5-flash-lite   # ultra-cheap
export PI_VISION_MODEL=gemini-3.5-flash         # latest gen

Usage

Once installed, it works automatically. When you're on a text-only model:

• Paste an image (Ctrl+V) → Gemini describes it, injected into conversation
• Agent reads an image file → same flow, cached per file path
• Switch to a vision model → extension skips itself, zero overhead

Toggle with /vision-proxy in your Pi session.

License

MIT