From 484c1807e3fa1e2b1c4526f09bed4f4f7add120c Mon Sep 17 00:00:00 2001
From: Your Name <your.email@example.com>
Date: Thu, 19 Feb 2026 04:03:55 +0800
Subject: [PATCH] fix(agent): enforce concrete visual summary content

---
 app/(chat)/api/chat/route.ts |   5 +-
 lib/ai/smart-analyzer.ts     | 104 +++++++++++++++++++++++------------
 2 files changed, 74 insertions(+), 35 deletions(-)

diff --git a/app/(chat)/api/chat/route.ts b/app/(chat)/api/chat/route.ts
index 27f53945a..1ddd8fd7b 100644
--- a/app/(chat)/api/chat/route.ts
+++ b/app/(chat)/api/chat/route.ts
@@ -2233,7 +2233,10 @@ export async function POST(request: Request) {
                 imageSummary: summary.trim(),
               });
             smartAutoFriendlyMessage = ensureImageAnalysisSummarySection(baseMessage, summary.trim());
-            console.log("[Smart Auto Mode] ✅ Image analysis summary ready");
+            console.log(
+              "[Smart Auto Mode] ✅ Image analysis summary ready:",
+              summary.trim().slice(0, 160),
+            );
           } else if (smartAutoFriendlyMessage && shouldEnrichImageAnalysisText(smartAutoFriendlyMessage)) {
             console.warn("[Smart Auto Mode] ⚠️ Image summary unavailable, keep original analysis text");
           }
diff --git a/lib/ai/smart-analyzer.ts b/lib/ai/smart-analyzer.ts
index 6e0cf075b..8a76ec053 100644
--- a/lib/ai/smart-analyzer.ts
+++ b/lib/ai/smart-analyzer.ts
@@ -557,7 +557,7 @@ async function summarizeWithZhipuOfficialApi(params: {
             {
               role: "system",
               content:
-                "你是图像观察助手。请根据图片给出执行前说明，输出严格两行中文：第1行“画面：...”，第2行“关键要素：...”。不要输出其他内容。",
+                "你是图像观察助手。只描述图片可见内容，不要提执行动作或用户意图。输出严格一行中文。",
             },
             {
               role: "user",
@@ -581,9 +581,10 @@ async function summarizeWithZhipuOfficialApi(params: {
 
       const message = payload?.choices?.[0]?.message;
       const responseText = String(message?.content || message?.reasoning_content || "").trim();
-      if (responseText) {
+      const normalized = normalizeVisualSummaryText(responseText);
+      if (normalized && !isWeakVisualSummary(normalized)) {
         console.log(`[Smart Analyzer] Zhipu official summary model used: ${modelId}`);
-        return responseText;
+        return normalized;
       }
     } catch {
       continue;
@@ -620,12 +621,10 @@ export async function summarizeImageForExecutionReport(params: {
   const zhipuProvider = zhipuProviderFromAgent || zhipuProviderFallback;
 
   const promptText = [
-    text ? `用户需求：${text}` : "",
-    context ? `画布上下文：${context.slice(0, 220)}` : "",
-    "请给出当前画面的客观描述，便于执行前说明。",
-  ]
-    .filter(Boolean)
-    .join("\n");
+    "只根据图片中可见内容做客观观察，不要推测用户意图。",
+    "禁止出现“选中/待处理/按需求/处理/修改/转换”等执行词。",
+    "输出一行：画面：主体与关键元素；场景：环境；风格：视觉风格；光线：明暗与色温。",
+  ].join("\n");
 
   if (zhipuProvider?.apiKey && zhipuProvider?.baseUrl) {
     const summary = await summarizeWithZhipuOfficialApi({
@@ -647,33 +646,40 @@ export async function summarizeImageForExecutionReport(params: {
     const model = await getLanguageModelById(selectedVisionModel.model.id);
     if (!model) return null;
 
-    const fallbackVisionImages = await buildAnalyzerVisionImagesByMode({
-      imageUrls,
-      preferRemoteUrls: shouldPreferRemoteVisionUrls(selectedVisionModel),
-    });
-    if (fallbackVisionImages.length === 0) return null;
+    const preferRemote = shouldPreferRemoteVisionUrls(selectedVisionModel);
+    const modeAttempts = Array.from(new Set([preferRemote, !preferRemote]));
 
-    const { text: summaryText } = await generateText({
-      model,
-      messages: [
-        {
-          role: "user",
-          content: [
-            {
-              type: "text",
-              text: `${promptText}\n只输出一行，格式：画面要点：主体+场景+风格+光线（不超过45字）。`,
-            },
-            ...fallbackVisionImages,
-          ],
-        },
-      ],
-      temperature: 0.1,
-      maxTokens: 120,
-      abortSignal: AbortSignal.timeout(ANALYZER_TIMEOUT_MS),
-    } as any);
+    for (const preferRemoteUrls of modeAttempts) {
+      const fallbackVisionImages = await buildAnalyzerVisionImagesByMode({
+        imageUrls,
+        preferRemoteUrls,
+      });
+      if (fallbackVisionImages.length === 0) continue;
 
-    const normalized = summaryText?.replace(/\s+/g, " ").trim();
-    return normalized || null;
+      const { text: summaryText } = await generateText({
+        model,
+        messages: [
+          {
+            role: "user",
+            content: [
+              {
+                type: "text",
+                text: `${promptText}\n仅输出一行，不超过60字。`,
+              },
+              ...fallbackVisionImages,
+            ],
+          },
+        ],
+        temperature: 0.1,
+        maxTokens: 120,
+        abortSignal: AbortSignal.timeout(ANALYZER_TIMEOUT_MS),
+      } as any);
+
+      const normalized = normalizeVisualSummaryText(summaryText);
+      if (normalized && !isWeakVisualSummary(normalized)) {
+        return normalized;
+      }
+    }
   } catch (error) {
     console.warn("[Smart Analyzer] summarizeImageForExecutionReport fallback failed:", error);
   }
@@ -681,6 +687,36 @@ export async function summarizeImageForExecutionReport(params: {
   return null;
 }
 
+function normalizeVisualSummaryText(text?: string | null): string {
+  if (!text) return "";
+  return text
+    .replace(/\s+/g, " ")
+    .replace(/^[-*•]\s*/, "")
+    .replace(/^(画面要点|画面描述|图片分析|分析结果)[:：]\s*/i, "")
+    .trim();
+}
+
+function isWeakVisualSummary(summary: string): boolean {
+  const normalized = summary.replace(/\s+/g, " ").trim();
+  if (!normalized || normalized.length < 14) return true;
+
+  const weakPatterns = [
+    /待处理图/,
+    /选中.*位置/,
+    /当前上传图片/,
+    /基于.*原图处理/,
+    /按.*需求执行/,
+    /仅做.*调整/,
+    /不改动原有内容结构/,
+    /转换为.*场景/,
+  ];
+  if (weakPatterns.some((pattern) => pattern.test(normalized))) {
+    return true;
+  }
+
+  return false;
+}
+
 // 历史消息类型
 interface HistoryMessage {
   role: "user" | "assistant";