pi-aux-vision

中文

Registers a describe_image native tool for Pi: the main model (e.g. deepseek-v4-flash, which has no image input) decides when to call it, passes an image path and a specific question, and the extension routes the request through pi's official pipeline to a configured vision model. The result comes back as a tool_result in the conversation context.

Install

pi install npm:pi-aux-vision

Or to try it without installing:

pi -e npm:pi-aux-vision

Manual install: drop the pi-aux-vision/ directory under ~/.pi/agent/extensions/, then run /reload in Pi.

How it works

• On startup, reads ~/.pi/agent/aux-vision.json; with no config, auto-discovers the first available (authenticated, image-capable) vision model and writes it to the config. Falls back automatically when the configured model becomes unavailable.
• Auth, protocol serialization, and retries all go through pi's official pipeline (ctx.modelRegistry.complete), supporting google-generative-ai, openai-completions, and anthropic-messages protocols.
• Every result begins with an exhaustive transcription base (image type, verbatim transcription of all visible text, layout/order), then answers the question and ends with a completeness attestation — so a main model without image input can reason from the base, not just from a narrow answer. When the output hits the token cap, the tool prepends an explicit truncation notice instead of silently returning a partial base.
• Image limit is 10 MB (the intersection of the three providers' limits); oversized images are compressed with pi's official resizeImage before failing.
• After the first describe_image call or /vision test of a session (success or failure), the TUI footer shows vision: provider/model until the session ends — dim vision: prefix, accent model name, and a ! in error color after a failed call. New sessions start hidden; toggle with showInFooter.

Compatibility

Verified against pi 0.86.0 (2026-09-19). The peer dependency range stays "*"; the source type-checks against the current SDK via npm run typecheck.

Configuration

~/.pi/agent/aux-vision.json:

{
  "enabled": true,
  "provider": "google",
  "model": "gemini-2.5-flash",
  "maxOutputTokens": 8192,
  "maxRetries": 2,
  "maxRetryDelayMs": 5000,
  "showInFooter": true
}

• maxOutputTokens: output token cap for the vision model. Every result begins with an exhaustive transcription base (image type + verbatim transcription of all visible text), so dense screenshots need a larger budget; when the cap is hit, the tool prepends an explicit truncation notice instead of silently returning a partial base
• maxRetries: retry count (initial + N attempts; 4xx is not retried, handled by the official pipeline)
• maxRetryDelayMs: backoff ceiling, in milliseconds
• showInFooter: show the vision: provider/model status in the TUI footer after the first describe_image call or /vision test of a session (default true)

Commands

Command	Description
/vision status	Current provider/model, protocol, enabled state, footer switch
/vision set <provider> <model>	Set a vision model explicitly and enable it, writes to config
/vision list	List available (authenticated) image models, mark the current one
/vision enable / /vision disable	Toggle; when disabled, describe_image is hidden from the main model
/vision test [path]	Verify the full pipeline with an auto-generated test image; optional custom path

Tool

describe_image(image_path, question) — read from disk → encode → single vision-model call → text result.

• image_path: absolute path or path relative to the working directory; supports png / jpeg / gif / webp / bmp
• question: a specific question, e.g. "extract the stack trace shown in line 4 of the error message" or "why is the button shifted 10px to the right?"

The vision model answers in the language of the question. Errors (file not found, unsupported format, call failure) return as structured text for the main model to handle; no confirmation dialogs.

Test

npm test

or directly:

node .test/build.js && node .test/test-run.mjs

Type compatibility against the current pi SDK is verified with npm run typecheck.

.test/ bundles the extension modules with mocked pi dependencies via esbuild and covers config read/write, model discovery, describe_image success/failure paths, test-image generation, and the footer state machine + wiring.

License

MIT