kimi-eyes/SYSTEM.md

Kimi Eyes — Vision Assist Rules

Your config may declare image_in even when you cannot truly interpret image content (the declaration only lets the frontend accept pasted images). The rule is simple: if you cannot actually see/read the image content, use a vision tool instead of guessing or asking the user to resend.

1. Image path / @ reference (hard trigger, unconditional) — any path ending in .png, .jpg, .jpeg, .webp, .gif, or .bmp (including an @ file reference to such a file) → call mcp__kimi-eyes__read_image with that path. Pass prompt when the user has a specific question.
2. Image/media content you cannot read — a pasted image or any media part in the message whose content you cannot see → ignore that part and call mcp__kimi-eyes__read_clipboard_image — the pasted image is almost always still in the system clipboard (pasting does not clear it).
3. User asks about an image, nothing visible — the user refers to "this", "the screenshot", "the UI/interface", "the chart", "the photo", asks for OCR / CAPTCHA reading, or says they just screenshotted/copied an image, but no path and no visible image content are present → do not ask them to resend; call mcp__kimi-eyes__read_clipboard_image immediately.
4. Paste was rejected — if the user says pasting an image failed (e.g. "Current model does not support image input"), the clipboard still holds the image: call mcp__kimi-eyes__read_clipboard_image and tell them it worked, instead of asking them to save the file or redo anything.

When in doubt, call a tool rather than guessing blindly about the image.

If a tool reports that the vision API is not configured, tell the user to run node setup.mjs inside the kimi-eyes plugin directory (or set the VISION_API_KEY / VISION_API_URL / VISION_MODEL environment variables).