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/router.go

@@ -0,0 +1,128 @@
+package vision
+
+// Assessment is the model's own confidence in reading an image.
+type Assessment string
+
+const (
+	// AssessmentUnknown means the model reported no self-assessment.
+	AssessmentUnknown Assessment = "unknown"
+	// AssessmentGood means the model is confident in its read.
+	AssessmentGood Assessment = "good"
+	// AssessmentPoor means the model judged its recognition as poor.
+	AssessmentPoor Assessment = "poor"
+)
+
+// ModelInfo describes the main model's vision capability.
+type ModelInfo struct {
+	// Name identifies the model.
+	Name string
+	// SupportsVision reports whether the model accepts image input.
+	SupportsVision bool
+	// SelfAssessment is the model's own confidence for the current image.
+	// It drives the layer-2 tool route when SupportsVision is true.
+	SelfAssessment Assessment
+}
+
+// DecisionKind is the chosen layered route.
+type DecisionKind string
+
+const (
+	// KindMain routes the image to the main model with the image in context.
+	KindMain DecisionKind = "main"
+	// KindVisionTool routes to a dedicated vision tool, returning only the tool
+	// result so the image never occupies main-model context.
+	KindVisionTool DecisionKind = "vision_tool"
+	// KindOCRTool routes to an OCR tool, returning only the extracted text.
+	KindOCRTool DecisionKind = "ocr_tool"
+	// KindUnsupported means no route can process the image.
+	KindUnsupported DecisionKind = "unsupported"
+)
+
+// Decision is the outcome of a routing call.
+type Decision struct {
+	// Kind is the chosen route.
+	Kind DecisionKind
+	// Tool names the vision or OCR tool to invoke, set when Kind is a tool route.
+	Tool string
+	// Reason is a human-readable explanation for observability and logs.
+	Reason string
+}
+
+// RouteRequest carries routing hints from the caller on top of the model info.
+type RouteRequest struct {
+	// TextPrimary reports that the user indicated the image is mainly text, so
+	// an OCR tool is preferred over a general vision model.
+	TextPrimary bool
+	// MainSelfAssessedPoor is set when the main model has already judged its own
+	// read of the image as poor, requesting a vision/OCR tool route.
+	MainSelfAssessedPoor bool
+}
+
+// RouterOption configures a Router.
+type RouterOption func(*Router)
+
+// WithVisionTool binds the name of the registered vision-understanding tool.
+func WithVisionTool(name string) RouterOption {
+	return func(r *Router) {
+		if name != "" {
+			r.visionTool = name
+		}
+	}
+}
+
+// WithOCRTool binds the name of the registered OCR tool.
+func WithOCRTool(name string) RouterOption {
+	return func(r *Router) {
+		if name != "" {
+			r.ocrTool = name
+		}
+	}
+}
+
+// Router implements the layered vision routing policy. It does no I/O; the
+// caller uses the returned Decision to dispatch the request.
+type Router struct {
+	visionTool string
+	ocrTool    string
+}
+
+// NewRouter builds a Router. Unknown (empty) tool names leave the corresponding
+// route unavailable, so the router degrades to KindMain or KindUnsupported.
+func NewRouter(opts ...RouterOption) *Router {
+	r := &Router{}
+	for _, o := range opts {
+		o(r)
+	}
+	return r
+}
+
+// Route picks a layered route for an image request given the target model info
+// and caller hints. The order is: the main model directly when it can see and
+// has not self-assessed poorly (layer 1); otherwise a vision or OCR tool whose
+// result replaces the image in context (layers 2 and 3); otherwise
+// KindUnsupported when nothing can process the image.
+func (r *Router) Route(target ModelInfo, req RouteRequest) Decision {
+	poor := target.SelfAssessment == AssessmentPoor || req.MainSelfAssessedPoor
+	// Layer 1: the main model accepts images and is confident, so use it first.
+	if target.SupportsVision && !poor {
+		return Decision{
+			Kind:   KindMain,
+			Reason: "main model supports vision and is confident; use it directly",
+		}
+	}
+	// OCR is preferred when the user says the image is mostly text.
+	if req.TextPrimary && r.ocrTool != "" {
+		return Decision{Kind: KindOCRTool, Tool: r.ocrTool, Reason: "image is mainly text; route to OCR"}
+	}
+	if r.visionTool != "" {
+		return Decision{Kind: KindVisionTool, Tool: r.visionTool, Reason: "route to dedicated vision tool"}
+	}
+	if r.ocrTool != "" {
+		return Decision{Kind: KindOCRTool, Tool: r.ocrTool, Reason: "route to OCR tool"}
+	}
+	// No tool is bound. A sighted main model remains the only option.
+	if target.SupportsVision {
+		return Decision{Kind: KindMain, Reason: "no vision/OCR tool bound; keep image with main model"}
+	}
+	return Decision{Kind: KindUnsupported, Reason: "main model cannot see and no vision/OCR tool is bound"}
+}