Pi Gemini Image Bridge

v0.6.0 - Pi extension that relays tool-result images to Gemini-family models as user-message attachments across Responses proxies.

Purpose

Pi's read tool already produces native image blocks in toolResult messages, and that path works for Claude, GPT, and native Google Gemini. The remaining gap is the proxy boundary: a Responses proxy such as CLIProxyAPI translates top-level user input_image parts correctly but can drop images nested in function_call_output items. This extension closes exactly that gap and does nothing else.

The target setup uses @router-for-me/pi-cliproxyapi-provider alongside this extension. The CPA provider extension discovers models, registers the cliproxyapi-codex-responses API, and sends inference through CPA's Codex Responses endpoint. No models.json entry is required. This project only adapts image placement; it does not register CPA models, credentials, or endpoints.

Install

pi install npm:@router-for-me/pi-cliproxyapi-provider
pi install npm:pi-gemini-image-bridge

For a local checkout, add the package to Pi's settings.json:

{
  "packages": [
    {
      "source": "/path/to/pi-gemini-image-bridge",
      "extensions": ["+src/index.ts"]
    }
  ]
}

Requires Pi >= 0.84.1, Node.js >= 23.6.0, and a configured CPA provider extension. Use /login CLIProxyAPI (or /login cliproxyapi) in Pi to configure the CPA connection.

Behavior

Before each request, a request-scoped context handler runs for Gemini models that declare image input support on the allowlisted Responses APIs:

• openai-responses
• openai-codex-responses
• cliproxyapi-codex-responses

Images from consecutive toolResult messages are moved, in order, into one transient user attachment message. Tool text, names, IDs, and error state remain intact. The session history and model turn count are unchanged; the relayed view exists only for the current request. Applying it twice produces the same result.

Everything else is Pi-native. Clipboard pastes, read execution, image normalization (size and downscaling), rewind, resume, and model switching all follow Pi's built-in behavior; this extension never rewrites user input.

Verify

npm test
npm run typecheck

For a manual check, ask a CPA-proxied Gemini model to read an image file and confirm the model describes its contents.

v0.6.0

• Removed the clipboard path-to-attachment pipeline, self-contained [Image #N: filename] labels, transcript label compaction, rewind/model-switch rehydration, and the WASM adaptive encoding pipeline. Pi's native read image support replaces the whole input track; image size handling follows Pi's native normalization. The extension now only performs the Responses tool-image relay.

Limits

• The relay is allowlisted and does not modify native Google Gemini or OpenAI Completions traffic. Unknown models or APIs are left untouched.
• Image data itself is never re-encoded, resized, or budgeted; whatever Pi's native pipeline produced is what gets relayed.

See architecture/index.md for module contracts and data flow, and architecture/decisions.md for product decisions.

License

MIT