feat(multimodal): 视觉分层路由（多模态检测→主/视觉/OCR 工具）+ 语音 STT→TTS 降级链（异步）

internal/runtime/multimodal/vision/vision.go

@@ -0,0 +1,79 @@
+// Package vision implements the layered image routing described in
+// DevDocs/design/06-多模态.md §0.
+//
+// A request that carries image content is first checked by DetectMultimodal.
+// The VisionRouter then picks one of the layered routes: the main model
+// directly (when it accepts images), a tool-shaped vision/OCR adapter (when the
+// main model self-assesses poorly or cannot see), or the OCR tool (when the
+// image is mainly text). Vision/OCR adapters are registered as sdk.Tool
+// capabilities through the capability registry, and their output is a tool
+// result, so the image itself never occupies main-model context.
+//
+// Images may come from a local path, an http(s) URL, or a base64/data: URI.
+package vision
+
+import (
+	"strings"
+
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/model"
+)
+
+// DetectMultimodal reports whether msgs contains any non-text content part,
+// i.e. whether multimodal handling is required. Today model only distinguishes
+// text from image content, so this is true exactly when a message carries an
+// image part; the check is written against "not text" so future content types
+// (audio) extend it without rework.
+func DetectMultimodal(msgs []model.Message) bool {
+	for _, m := range msgs {
+		for _, part := range m.Content {
+			if part.Type != model.ContentTypeText {
+				return true
+			}
+		}
+	}
+	return false
+}
+
+// ImageSource classifies where an image's bytes come from.
+type ImageSource string
+
+const (
+	// SourcePath is a local filesystem path.
+	SourcePath ImageSource = "path"
+	// SourceURL is an http(s):// URL.
+	SourceURL ImageSource = "url"
+	// SourceBase64 is a base64 payload, optionally wrapped in a data: URI.
+	SourceBase64 ImageSource = "base64"
+)
+
+// ImageRef is a single image reference plus its resolved source kind.
+type ImageRef struct {
+	// Source is the classified source kind.
+	Source ImageSource
+	// Data is the local path, URL, or base64/data: payload.
+	Data string
+}
+
+// ParseImageRef classifies a raw image source string into an ImageRef. Empty
+// input is rejected. Local paths that exist on disk are classified as paths;
+// http(s) URLs as URLs; every other value, including a data: URI, is treated as
+// a base64 payload. This keeps routing deterministic without guessing MIME.
+func ParseImageRef(src string) (ImageRef, error) {
+	if src == "" {
+		return ImageRef{}, ErrNoImage
+	}
+	if strings.HasPrefix(src, "http://") || strings.HasPrefix(src, "https://") {
+		return ImageRef{Source: SourceURL, Data: src}, nil
+	}
+	if strings.HasPrefix(src, "file://") {
+		return ImageRef{Source: SourcePath, Data: strings.TrimPrefix(src, "file://")}, nil
+	}
+	// A data: URI is base64 already; treat as base64, not path.
+	if strings.HasPrefix(src, "data:") {
+		return ImageRef{Source: SourceBase64, Data: src}, nil
+	}
+	if fileExists(src) {
+		return ImageRef{Source: SourcePath, Data: src}, nil
+	}
+	return ImageRef{Source: SourceBase64, Data: src}, nil
+}