diff --git a/components/elements/response.tsx b/components/elements/response.tsx
index a8fedd8..0e2fc2d 100644
--- a/components/elements/response.tsx
+++ b/components/elements/response.tsx
@@ -1,12 +1,20 @@
 "use client";
 
-import type { ComponentProps } from "react";
+import { ChevronDownIcon } from "lucide-react";
+import type { ComponentProps, ReactNode } from "react";
+import { useState } from "react";
 import { Streamdown } from "streamdown";
-import { cn } from "@/lib/utils";
 import { ImagePreview } from "@/components/image-preview";
+import {
+  Collapsible,
+  CollapsibleContent,
+  CollapsibleTrigger,
+} from "@/components/ui/collapsible";
 import { VideoPreview } from "@/components/video-preview";
+import { cn } from "@/lib/utils";
 
 type ResponseProps = ComponentProps<typeof Streamdown> & {
+  collapsibleImageAnalysis?: boolean;
   onImageEdit?: (imageUrl: string) => void;
   onDrawingComplete?: (annotatedImageUrl: string) => void;
   onImageEnhance?: (imageUrl: string) => void;
@@ -21,6 +29,20 @@ type ResponseProps = ComponentProps<typeof Streamdown> & {
   isProcessingWatermarkRemove?: boolean;
 };
 
+const IMAGE_ANALYSIS_TITLE = "图片分析";
+const AGENT_SECTION_TITLES = new Set([
+  "需求理解",
+  IMAGE_ANALYSIS_TITLE,
+  "执行计划",
+  "输出设置",
+]);
+
+type AnalysisSection = {
+  before: string;
+  analysis: string;
+  after: string;
+};
+
 function splitMarkdownImages(content: string) {
   const segments: Array<
     | { type: "text"; value: string }
@@ -58,23 +80,144 @@ function splitMarkdownImages(content: string) {
   return segments;
 }
 
+function normalizeSectionTitle(line: string) {
+  let normalized = line.trim();
+  if (!normalized) return "";
+
+  normalized = normalized
+    .replace(/^#{1,6}\s*/, "")
+    .replace(/^[-*]\s*/, "")
+    .replace(/^\*\*(.+)\*\*$/, "$1")
+    .replace(/[▼▽▾]\s*$/, "")
+    .trim();
+
+  const colonIndex = normalized.indexOf("：");
+  const asciiColonIndex = normalized.indexOf(":");
+  const splitIndex =
+    colonIndex >= 0 && asciiColonIndex >= 0
+      ? Math.min(colonIndex, asciiColonIndex)
+      : Math.max(colonIndex, asciiColonIndex);
+
+  if (splitIndex > 0 && splitIndex <= 12) {
+    normalized = normalized.slice(0, splitIndex).trim();
+  }
+
+  return normalized.replace(/[：:]+$/, "").trim();
+}
+
+function extractImageAnalysisSection(content: string): AnalysisSection | null {
+  const lines = content.split(/\r?\n/);
+  let startIndex = -1;
+  let inlineAnalysis = "";
+
+  for (let i = 0; i < lines.length; i += 1) {
+    const title = normalizeSectionTitle(lines[i]);
+    if (!title || !title.startsWith(IMAGE_ANALYSIS_TITLE)) continue;
+
+    startIndex = i;
+
+    const headingText = lines[i]
+      .trim()
+      .replace(/^#{1,6}\s*/, "")
+      .replace(/^[-*]\s*/, "")
+      .replace(/^\*\*(.+)\*\*$/, "$1")
+      .trim();
+
+    const inlineMatch = headingText.match(/^图片分析[：:]\s*(.+)$/);
+    if (inlineMatch?.[1]) {
+      inlineAnalysis = inlineMatch[1].trim();
+    }
+    break;
+  }
+
+  if (startIndex < 0) return null;
+
+  let endIndex = lines.length;
+  for (let i = startIndex + 1; i < lines.length; i += 1) {
+    const title = normalizeSectionTitle(lines[i]);
+    if (!title || title.startsWith(IMAGE_ANALYSIS_TITLE)) continue;
+    if (AGENT_SECTION_TITLES.has(title)) {
+      endIndex = i;
+      break;
+    }
+  }
+
+  const analysisLines = lines.slice(startIndex + 1, endIndex);
+  if (inlineAnalysis) {
+    analysisLines.unshift(inlineAnalysis);
+  }
+
+  const analysis = analysisLines.join("\n").trim();
+  if (!analysis) return null;
+
+  return {
+    before: lines.slice(0, startIndex).join("\n").trim(),
+    analysis,
+    after: lines.slice(endIndex).join("\n").trim(),
+  };
+}
+
+type ImageAnalysisCollapsibleProps = {
+  content: string;
+  streamdownClassName: string;
+  streamdownComponents: ComponentProps<typeof Streamdown>["components"];
+};
+
+function ImageAnalysisCollapsible({
+  content,
+  streamdownClassName,
+  streamdownComponents,
+}: ImageAnalysisCollapsibleProps) {
+  const [open, setOpen] = useState(false);
+
+  return (
+    <Collapsible
+      className="my-2 rounded-xl border border-border/60 bg-muted/30"
+      onOpenChange={setOpen}
+      open={open}
+    >
+      <CollapsibleTrigger className="flex w-full items-center justify-between px-3 py-2 text-left">
+        <span className="text-sm font-medium text-foreground">图片分析</span>
+        <ChevronDownIcon
+          className={cn(
+            "size-4 text-muted-foreground transition-transform",
+            open ? "rotate-180" : "rotate-0"
+          )}
+        />
+      </CollapsibleTrigger>
+      <CollapsibleContent
+        className={cn(
+          "border-border/50 border-t px-3 pb-3 pt-2 text-sm text-muted-foreground",
+          "data-[state=closed]:fade-out-0 data-[state=closed]:slide-out-to-top-1 data-[state=open]:slide-in-from-top-1",
+          "outline-hidden data-[state=closed]:animate-out data-[state=open]:animate-in"
+        )}
+      >
+        <Streamdown className={streamdownClassName} components={streamdownComponents}>
+          {content}
+        </Streamdown>
+      </CollapsibleContent>
+    </Collapsible>
+  );
+}
+
 // 注意：移除了 memo 的自定义比较函数，确保 isProcessing* 状态变化时组件能正确重新渲染
 // 之前的 memo 比较函数可能导致 Streamdown 内部的 ImagePreview 组件没有正确更新
-export const Response = ({ 
-  className, 
-  onImageEdit, 
-  onDrawingComplete, 
-  onImageEnhance, 
-  onImageRmbg, 
-  onImageExtend, 
-  onImageAngleEdit, 
+export const Response = ({
+  className,
+  collapsibleImageAnalysis = false,
+  onImageEdit,
+  onDrawingComplete,
+  onImageEnhance,
+  onImageRmbg,
+  onImageExtend,
+  onImageAngleEdit,
   onImageWatermarkRemove,
-  isProcessingEnhance, 
-  isProcessingRmbg, 
-  isProcessingExtend, 
-  isProcessingAngleEdit, 
+  isProcessingEnhance,
+  isProcessingRmbg,
+  isProcessingExtend,
+  isProcessingAngleEdit,
   isProcessingWatermarkRemove,
-  ...props 
+  ...props
 }: ResponseProps) => {
   const streamdownClassName = cn(
     "size-full [&>*:first-child]:mt-0 [&>*:last-child]:mb-0 [&_code]:whitespace-pre-wrap [&_code]:break-words [&_pre]:max-w-full [&_pre]:overflow-x-auto",
@@ -116,54 +259,92 @@ export const Response = ({
     },
   };
 
-  const rawText = typeof props.children === "string" ? props.children : null;
-  if (rawText && rawText.includes("![") && rawText.includes("](")) {
-    const segments = splitMarkdownImages(rawText);
-    const hasImageSegments = segments.some((segment) => segment.type === "image");
+  const renderImagePreview = (src: string, alt: string | undefined, key: string) => (
+    <ImagePreview
+      key={key}
+      src={src}
+      alt={alt}
+      onImageEdit={onImageEdit}
+      onDrawingComplete={onDrawingComplete}
+      onImageEnhance={onImageEnhance}
+      onImageRmbg={onImageRmbg}
+      onImageExtend={onImageExtend}
+      onImageAngleEdit={onImageAngleEdit}
+      onImageWatermarkRemove={onImageWatermarkRemove}
+      isProcessingEnhance={isProcessingEnhance}
+      isProcessingRmbg={isProcessingRmbg}
+      isProcessingExtend={isProcessingExtend}
+      isProcessingAngleEdit={isProcessingAngleEdit}
+      isProcessingWatermarkRemove={isProcessingWatermarkRemove}
+    />
+  );
 
-    if (hasImageSegments) {
+  const renderTextWithImages = (text: string, keyPrefix: string): ReactNode => {
+    if (!text.trim()) return null;
+
+    if (text.includes("![") && text.includes("](")) {
+      const segments = splitMarkdownImages(text);
+      const hasImageSegments = segments.some((segment) => segment.type === "image");
+
+      if (hasImageSegments) {
+        return segments.map((segment, index) => {
+          if (segment.type === "image") {
+            return renderImagePreview(
+              segment.src,
+              segment.alt,
+              `${segment.src}-${processingKey}-${keyPrefix}-${index}`
+            );
+          }
+
+          if (!segment.value.trim()) {
+            return null;
+          }
+
+          return (
+            <Streamdown
+              className={streamdownClassName}
+              components={streamdownComponents}
+              key={`${keyPrefix}-text-${index}`}
+            >
+              {segment.value}
+            </Streamdown>
+          );
+        });
+      }
+    }
+
+    return (
+      <Streamdown
+        className={streamdownClassName}
+        components={streamdownComponents}
+        key={`${keyPrefix}-single`}
+      >
+        {text}
+      </Streamdown>
+    );
+  };
+
+  const rawText = typeof props.children === "string" ? props.children : null;
+  if (rawText) {
+    const analysisSection = collapsibleImageAnalysis
+      ? extractImageAnalysisSection(rawText)
+      : null;
+
+    if (analysisSection) {
       return (
         <div className={streamdownClassName}>
-          {segments.map((segment, index) => {
-            if (segment.type === "image") {
-              return (
-                <ImagePreview
-                  key={`${segment.src}-${processingKey}-${index}`}
-                  src={segment.src}
-                  alt={segment.alt}
-                  onImageEdit={onImageEdit}
-                  onDrawingComplete={onDrawingComplete}
-                  onImageEnhance={onImageEnhance}
-                  onImageRmbg={onImageRmbg}
-                  onImageExtend={onImageExtend}
-                  onImageAngleEdit={onImageAngleEdit}
-                  onImageWatermarkRemove={onImageWatermarkRemove}
-                  isProcessingEnhance={isProcessingEnhance}
-                  isProcessingRmbg={isProcessingRmbg}
-                  isProcessingExtend={isProcessingExtend}
-                  isProcessingAngleEdit={isProcessingAngleEdit}
-                  isProcessingWatermarkRemove={isProcessingWatermarkRemove}
-                />
-              );
-            }
-
-            if (!segment.value.trim()) {
-              return null;
-            }
-
-            return (
-              <Streamdown
-                className={streamdownClassName}
-                components={streamdownComponents}
-                key={`text-${index}`}
-              >
-                {segment.value}
-              </Streamdown>
-            );
-          })}
+          {renderTextWithImages(analysisSection.before, "analysis-before")}
+          <ImageAnalysisCollapsible
+            content={analysisSection.analysis}
+            streamdownClassName={streamdownClassName}
+            streamdownComponents={streamdownComponents}
+          />
+          {renderTextWithImages(analysisSection.after, "analysis-after")}
         </div>
       );
     }
+
+    return <div className={streamdownClassName}>{renderTextWithImages(rawText, "raw")}</div>;
   }
 
   return (
diff --git a/components/message.tsx b/components/message.tsx
index bfb5fd8..c3cde20 100644
--- a/components/message.tsx
+++ b/components/message.tsx
@@ -312,6 +312,7 @@ const PurePreviewMessage = ({
                         return (
                           <Response 
                             key={`${partIndex}-${processingKey}`} 
+                            collapsibleImageAnalysis={message.role === "assistant"}
                             onImageEdit={onImageEdit} 
                             onDrawingComplete={onDrawingComplete}
                             onImageEnhance={onImageEnhance}
diff --git a/lib/ai/auto-agent-graph.ts b/lib/ai/auto-agent-graph.ts
index 17dbcbf..595fb74 100644
--- a/lib/ai/auto-agent-graph.ts
+++ b/lib/ai/auto-agent-graph.ts
@@ -134,11 +134,11 @@ async function generateAgentConversation(
 2) 语气自然、专业，不要模板腔，不要空话。
 3) 不暴露内部实现、模型路由、算法细节。
 4) 输出 Markdown，控制在 8-14 行。
-5) 有图片时，必须包含以下小节标题且按顺序输出：
-   - 需求理解
-   - 图片分析
-   - 执行计划
-   - 输出设置
+5) 有图片时，必须按顺序输出以下 Markdown H3 标题（逐字一致）：
+   - ### 需求理解
+   - ### 图片分析
+   - ### 执行计划
+   - ### 输出设置
 6) “图片分析”要结合每张参考图差异，给出风格/构图/材质/光影等具体观察。`;
 
     const prompt = [