@arhen/pi-core-vision

Install

Requires the pi coding agent — install it first: npm install -g @earendil-works/pi-coding-agent.

pi install git:github.com/arhen/pi-core-vision

or try without installing:

pi -e git:github.com/arhen/pi-core-vision

Transparent vision fallback for text-only models in pi.

Overrides the built-in read tool:

• text file → built-in behavior, untouched
• image + active model sees images → built-in behavior, untouched (pi's own resize + native attach)
• image + text-only model → pi resizes the image (Photon WASM), then the extension sends pi's resized output to a vision model and returns a compact text description

The model sees one result either way — no double reading, no new tool to learn. Text-only models (e.g. DeepSeek) can finally read screenshots, diagrams, and error messages.

Configure

Two modes — raw (any OpenAI-compatible endpoint) or registry (models from pi's own registry, auth via auth.json//login/env).

Set the vision model via /pi-vision command, env vars, or ~/.pi/pi-vision.json (JSON wins over env).

/pi-vision set baseUrl=https://api.openai.com/v1 apiKey=sk-... model=gpt-4o-mini
/pi-vision show          # current config (apiKey masked)
/pi-vision reset         # clear config file

Env vars:

export PI_VISION_BASE_URL="https://api.openai.com/v1"
export PI_VISION_API_KEY="sk-..."
export PI_VISION_MODEL="gpt-4o-mini"

~/.pi/pi-vision.json (extra options: prompt, maxTokens):

{
  "baseUrl": "https://api.openai.com/v1",
  "apiKey": "sk-...",
  "model": "gpt-4o-mini",
  "maxTokens": 1500
}

Registry mode

Use any model pi already knows — no duplicated credentials. The vision model must declare "input": ["text", "image"] in models.json, and auth resolves through pi's normal channels (stored credential in auth.json, /login, or provider apiKey).

{
  "provider": "anthropic",
  "model": "claude-sonnet-4-5",
  "maxTokens": 1500
}

/pi-vision set provider=anthropic model=claude-sonnet-4-5
/pi-vision set provider=kitchen model=gemma-4-26b-a4b-it

OpenAI-compatible providers (openai-completions) are called through the extension's own transport (retry, SSE-safe, cache); other APIs (anthropic-messages, google-generative-ai, custom) go through pi's provider machinery.

Any OpenAI-compatible endpoint works: OpenAI /v1, Google Gemini /v1beta/openai, Alibaba DashScope /compatible-mode/v1, Ollama /v1, LM Studio, vLLM. If your gateway streams SSE by default, the extension forces stream: false.

How it works

• Delegates every read to pi's own createReadToolDefinition — byte-identical built-in behavior (Photon resize to 2000px / 4.5MB, magic-byte mime detection, truncation).
• Checks ctx.model.input.includes("image") at call time. Vision-capable model → built-in result untouched. Text-only model + image → vision model describes pi's already-resized base64.
• Nested vision usage is reported back, so pi session stats stay accurate.
• Raw-file fallback (with 20MB guard) covers the case where pi's image processing is unavailable (e.g. BMP).

Benchmark

Single-run comparison vs community alternatives, text-only parent model (deepseek-v4-flash), same screenshot through the same kitchen gateway. See also pi-vision-handoff and pi-sense.

Healthy vision model (gemma-4-26b via kitchen)

| tool | e2e | flow | precision | |---|---|---|---| | pi-vision (this) | 32s | clean 1 read → description in result | 6/6 facts (title, theme, tabs, chat, layout) | | pi-vision-handoff | 44s | clean 1 read → context swap | 5/6 | | pi-sense | 122s | description in result | ~5/6 |

Vision API down (kitchen haiku 429 for the whole window)

| tool | result | |---|---| | pi-vision (pre-fix) | 200s+ hang — threw on 429 → parent retry-looped the read | | pi-vision (post-fix) | 19s — graceful [image: description unavailable] → model moves on | | pi-vision-handoff | 42s — graceful placeholder → OCR fallback | | pi-sense | 27s — graceful placeholder → OCR fallback |

Findings:

• Fastest end-to-end when vision is healthy, and the only tool that returns the full description inside the read result.
• Vision failures degrade gracefully (placeholder text, no hang) — shipped after the 429 incident above.
• Descriptions are framed as UNTRUSTED DATA (prompt-injection mitigation).

Caveats: single run per cell; gateway routing flakiness affects variance.

Development

bun src/self-check.ts    # logic self-checks (no pi needed, no API calls)

License

MIT