初始化 kimi-eyes：给 KimiCode 非多模态模型补上视觉能力

SYSTEM.md

@@ -0,0 +1,27 @@
+# Kimi Eyes — Vision Assist Rules
+
+**If your model supports image input (`image_in`) natively, ignore every rule below —
+you can see images directly, so never call these tools.**
+
+For models **without** native image input, the signals below are **hard triggers**:
+call a vision tool whenever any of them appears, regardless of how the user phrased
+their request.
+
+1. **Image-format 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. **Media content you cannot interpret** — a pasted image or any media part you
+   cannot read → ignore that part and call
+   `mcp__kimi-eyes__read_clipboard_image` (the image is almost always still in the
+   system clipboard).
+3. **Wording implies an image, no path** — the user refers to "this", "the
+   screenshot", "the UI/interface", "the chart", "the photo", asks for OCR / CAPTCHA
+   reading, or otherwise implies image content, with no path and no visible media
+   part → call `mcp__kimi-eyes__read_clipboard_image`.
+
+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).