强化图片Agent预分析并降低模板化回复

app/(chat)/api/chat/route.ts

@@ -699,12 +699,18 @@ export async function POST(request: Request) {
           !smartAutoFriendlyMessage;
 
         if (needsRicherAutoReply) {
-          smartAutoFriendlyMessage = buildAutoModeFallbackMessage({
-            text: autoTextForFallback,
-            hasImage,
-            fromHistory: smartAutoHistoryImages.length > 0,
-          });
-          console.log("[Smart Auto Mode] 🧩 Upgraded to richer conversational reply");
+          if (isImageTask) {
+            smartAutoFriendlyMessage =
+              autoGraphResult.analysis?.conversationalReply ||
+              "我先分析你这轮图片里的主体、风格和构图，再给你执行方案。";
+          } else {
+            smartAutoFriendlyMessage = buildAutoModeFallbackMessage({
+              text: autoTextForFallback,
+              hasImage,
+              fromHistory: smartAutoHistoryImages.length > 0,
+            });
+          }
+          console.log("[Smart Auto Mode] 🧩 Filled missing conversational reply");
         }
 
         if (autoGraphResult.optimizedPrompt) {
@@ -737,12 +743,8 @@ export async function POST(request: Request) {
         let bestModel = null;
         if (hasImage) {
           bestModel = allModels.find(m => m.model.supportsImageGen && !m.model.isSubModel);
-          // 回退时也保持结构化、可感知的自动模式反馈
-          smartAutoFriendlyMessage = buildAutoModeFallbackMessage({
-            text: autoTextForFallback,
-            hasImage: true,
-            fromHistory: autoFromHistoryForFallback || smartAutoHistoryImages.length > 0,
-          });
+          smartAutoFriendlyMessage =
+            "我先完成这一轮图片分析，再按你的要求执行修改并返回结果。";
         } else {
           smartAutoFriendlyMessage = buildAutoModeFallbackMessage({
             text: autoTextForFallback,

lib/ai/auto-agent-graph.ts

@@ -69,8 +69,11 @@ async function generateAgentConversation(
     const { getLanguageModelById } = await import("./dynamic-provider");
     const allModels = await getEnabledAiModelsWithProvider();
 
+    const preferredVisionModel = allModels.find(
+      (m) => m.provider.type === "alibaba" && m.model.supportsVision && m.model.isEnabled
+    );
     const candidateModel = state.hasImage
-      ? allModels.find((m) => m.model.supportsVision && m.model.isEnabled)
+      ? preferredVisionModel || allModels.find((m) => m.model.supportsVision && m.model.isEnabled)
       : allModels.find((m) => m.model.isEnabled && !m.model.isSubModel);
 
     if (!candidateModel) {
@@ -87,10 +90,42 @@ async function generateAgentConversation(
       .map((msg) => `${msg.role === "user" ? "用户" : "助手"}: ${compactText(msg.content, 120)}`)
       .join("\n");
 
+    const validImageUrls = state.imageUrls
+      .filter((url) => /^https?:\/\//i.test(url))
+      .slice(0, 4);
+
     const imageInfo = state.hasImage
       ? `用户本轮可用图片数量：${state.imageUrls.length || 1}。图片来源：${state.imageFromHistory ? "历史消息复用" : "本轮上传"}。`
       : "用户本轮没有上传图片。";
 
+    let visualAnalysis = "";
+    if (state.hasImage && validImageUrls.length > 0) {
+      try {
+        const visionPrompt = `请分析以下参考图片，给出可用于生图/修图执行前决策的结果。
+要求：
+1) 每张图输出 1 行，格式：图X：主体...；风格...；构图...；光影与色调...；材质/氛围...
+2) 只写客观可见信息，不要编造品牌、尺寸、镜头参数。
+3) 全部输出控制在 6 行内。`;
+
+        const visionInputs = validImageUrls.map((url) => ({ type: "image" as const, image: url }));
+        const { text: visionText } = await generateText({
+          model,
+          messages: [
+            {
+              role: "user" as const,
+              content: [{ type: "text" as const, text: visionPrompt }, ...visionInputs],
+            },
+          ],
+          temperature: 0.2,
+          maxTokens: 360,
+        } as any);
+
+        visualAnalysis = visionText.trim();
+      } catch (error) {
+        visualAnalysis = "";
+      }
+    }
+
     const systemPrompt = `你是一个真实执行中的 AI 生图/修图 Agent。
 请根据用户需求和本轮输入图片，给出“执行前”的分析汇报。
 
@@ -110,6 +145,7 @@ async function generateAgentConversation(
       `【用户原始需求】${state.text || "（无文本）"}`,
       `【意图分析】intent=${analysis.intent}; subIntent=${analysis.subIntent || "none"}; confidence=${analysis.confidence}`,
       `【图片信息】${imageInfo}`,
+      visualAnalysis ? `【视觉模型预分析】\n${visualAnalysis}` : "",
       state.context ? `【画布上下文】${compactText(state.context, 300)}` : "",
       history ? `【最近对话】\n${history}` : "",
       state.hasImage
@@ -119,10 +155,7 @@ async function generateAgentConversation(
       .filter(Boolean)
       .join("\n\n");
 
-    const imageInputs = state.imageUrls
-      .filter((url) => /^https?:\/\//i.test(url))
-      .slice(0, 4)
-      .map((url) => ({ type: "image" as const, image: url }));
+    const imageInputs = validImageUrls.map((url) => ({ type: "image" as const, image: url }));
 
     const requestPayload = state.hasImage && imageInputs.length > 0
       ? {