feat(agent): enforce visual analysis and extend auto timeouts to 120s

2026-02-19 03:56:58 +08:00 · 2026-02-19 03:56:58 +08:00 · e5c4b3780c
commit e5c4b3780c
parent 55b2c836b4
3 changed files with 154 additions and 63 deletions
--- a/app/(chat)/api/chat/route.ts
+++ b/app/(chat)/api/chat/route.ts
@ -672,13 +672,14 @@ function sanitizeAgentOutputSettingsSection(params: {
  const { message, userImageParams, hasUserCustomImageParams } = params;
  if (!message?.trim()) return message;

-  const outputSectionRegex = /###\s*输出设置[\s\S]*?(?=\n###\s*|$)/;
+  const outputSectionRegex =
+    /(?:###\s*输出设置|【\s*输出设置\s*】)[\s\S]*?(?=\n(?:###\s*|【\s*[^】]+\s*】)|$)/;
  if (!outputSectionRegex.test(message)) return message;

  if (!hasUserCustomImageParams) {
    return message.replace(
      outputSectionRegex,
-      "### 输出设置\n- 参数：按用户面板当前设置执行（未自动改写）"
+      "【输出设置】\n- 参数：按用户面板当前设置执行（未自动改写）"
    );
  }

@ -701,7 +702,46 @@ function sanitizeAgentOutputSettingsSection(params: {
  }
  lines.push("- 其余参数：保持用户面板当前设置");

-  return message.replace(outputSectionRegex, `### 输出设置\n${lines.join("\n")}`);
+  return message.replace(outputSectionRegex, `【输出设置】\n${lines.join("\n")}`);
+}
+
+function normalizeAgentReportHeadingLine(line: string): string {
+  const trimmed = line.trim();
+  if (!trimmed) return line;
+
+  const markdownHeading = trimmed.match(/^#{1,6}\s*(.+?)\s*$/);
+  const plainTitle = markdownHeading?.[1]?.trim() || trimmed;
+  const normalizedTitle = plainTitle
+    .replace(/^[^\u4e00-\u9fa5A-Za-z0-9]+/, "")
+    .trim();
+
+  const knownTitles = [
+    "需求理解",
+    "图片分析",
+    "执行计划",
+    "输出设置",
+    "脚本规划",
+    "分镜规划",
+    "执行说明",
+    "视频生成完成",
+  ];
+
+  if (knownTitles.some((title) => normalizedTitle === title)) {
+    return `【${normalizedTitle}】`;
+  }
+
+  if (markdownHeading) {
+    return `【${normalizedTitle || plainTitle}】`;
+  }
+  return line;
+}
+
+function normalizeAgentReportForDisplay(message?: string): string | undefined {
+  if (!message?.trim()) return message;
+  return message
+    .split("\n")
+    .map((line) => normalizeAgentReportHeadingLine(line))
+    .join("\n");
 }

 function guessImageMediaType(url: string): string {
@ -812,7 +852,7 @@ function extractMessageImageUrls(parts: any[]): string[] {

 function shouldEnrichImageAnalysisText(message?: string): boolean {
  if (!message) return false;
-  if (!/###\s*图片分析/.test(message)) return false;
+  if (!/(###\s*图片分析|【\s*图片分析\s*】)/.test(message)) return false;
  return /(基于当前上传图片处理|基于你选中的原图处理|基于上一轮图片继续处理)/.test(message);
 }

@ -820,12 +860,14 @@ function injectImageAnalysisSummary(message: string, summary: string): string {
  const normalizedSummary = summary.replace(/\s+/g, " ").trim();
  if (!normalizedSummary) return message;
  const lines = message.split("\n");
-  const analysisHeaderIndex = lines.findIndex((line) => /^###\s*图片分析/.test(line.trim()));
+  const analysisHeaderIndex = lines.findIndex((line) =>
+    /^(###\s*图片分析|【\s*图片分析\s*】)/.test(line.trim())
+  );
  if (analysisHeaderIndex < 0) return message;

  for (let i = analysisHeaderIndex + 1; i < lines.length; i += 1) {
    const current = lines[i]?.trim() || "";
-    if (/^###\s*/.test(current)) break;
+    if (/^(###\s*|【\s*[^】]+\s*】)/.test(current)) break;
    if (/^- /.test(current)) {
      lines[i] = `- ${normalizedSummary}`;
      return lines.join("\n");
@ -837,6 +879,19 @@ function injectImageAnalysisSummary(message: string, summary: string): string {
  return lines.join("\n");
 }

+function ensureImageAnalysisSummarySection(message: string, summary: string): string {
+  const normalizedSummary = summary.replace(/\s+/g, " ").trim();
+  if (!normalizedSummary) return message;
+
+  if (/(###\s*图片分析|【\s*图片分析\s*】)/.test(message)) {
+    return injectImageAnalysisSummary(message, normalizedSummary);
+  }
+
+  const trimmed = message.trim();
+  if (!trimmed) return `【图片分析】\n- ${normalizedSummary}`;
+  return `${trimmed}\n【图片分析】\n- ${normalizedSummary}`;
+}
+
 /**
 * 自定义消息转换函数
 * 注意：gemini-image 和 huanke-gemini 都需要 URL 格式（不是 Base64）
@ -1355,7 +1410,8 @@ let globalStreamContext: ResumableStreamContext | null = null;
 const inFlightChatRequests = new Map<string, number>();
 const IN_FLIGHT_REQUEST_TTL_MS = 5 * 60 * 1000;
 const DUPLICATE_REQUEST_WAIT_MS = 2 * 60 * 1000;
-const SMART_AUTO_GRAPH_TIMEOUT_MS = 45000;
+const SMART_AUTO_GRAPH_TIMEOUT_MS = 120000;
+const SMART_AUTO_FALLBACK_TIMEOUT_MS = 120000;

 const sleep = (ms: number) => new Promise((resolve) => setTimeout(resolve, ms));

@ -1905,7 +1961,7 @@ export async function POST(request: Request) {
              context: undefined,
            }),
            "fallbackAnalyzeIntent",
-            15000,
+            SMART_AUTO_FALLBACK_TIMEOUT_MS,
          );
          if (rescueAnalysis.conversationalReply?.trim()) {
            smartAutoFriendlyMessage = rescueAnalysis.conversationalReply.trim();
@ -1947,7 +2003,7 @@ export async function POST(request: Request) {
                  context: typeof autoContext === "string" ? autoContext : undefined,
                }),
                "fallbackVisualSummary",
-                15000,
+                SMART_AUTO_FALLBACK_TIMEOUT_MS,
              );
              if (summary?.trim()) {
                fallbackVisualSummary = summary.trim();
@ -2138,44 +2194,59 @@ export async function POST(request: Request) {
      });
    }

-    if (
-      shouldRunAgentConversation &&
-      !smartAutoNeedsClarification &&
-      smartAutoFriendlyMessage &&
-      shouldEnrichImageAnalysisText(smartAutoFriendlyMessage)
-    ) {
-      try {
-        const currentImageUrls = extractMessageImageUrls(message.parts as any[]);
-        const summaryImageUrls =
-          currentImageUrls.length > 0 ? currentImageUrls : smartAutoHistoryImages;
+    if (shouldRunAgentConversation && !smartAutoNeedsClarification) {
+      const currentImageUrls = extractMessageImageUrls(message.parts as any[]);
+      const summaryImageUrls =
+        currentImageUrls.length > 0 ? currentImageUrls : smartAutoHistoryImages;
+      const shouldRequireImageAnalysis = summaryImageUrls.length > 0;

-        if (summaryImageUrls.length > 0) {
+      if (shouldRequireImageAnalysis) {
+        try {
          const { summarizeImageForExecutionReport } = await import("@/lib/ai/smart-analyzer");
-          const summary = await withOperationTimeout(
-            summarizeImageForExecutionReport({
-              text: message.parts
-                .filter((part): part is { type: "text"; text: string } => part.type === "text")
-                .map((part) => part.text)
-                .join(" "),
-              imageUrls: summaryImageUrls,
-              context: typeof autoContext === "string" ? autoContext : undefined,
-            }),
-            "enrichImageAnalysisSummary",
-            8000,
-          );
-          if (summary?.trim()) {
-            smartAutoFriendlyMessage = injectImageAnalysisSummary(
-              smartAutoFriendlyMessage,
-              summary.trim(),
+          const latestUserText = message.parts
+            .filter((part): part is { type: "text"; text: string } => part.type === "text")
+            .map((part) => part.text)
+            .join(" ");
+
+          let summary: string | null = null;
+          const summaryTimeouts = [120000];
+          for (const timeoutMs of summaryTimeouts) {
+            summary = await withOperationTimeout(
+              summarizeImageForExecutionReport({
+                text: latestUserText,
+                imageUrls: summaryImageUrls,
+                context: typeof autoContext === "string" ? autoContext : undefined,
+              }),
+              "enrichImageAnalysisSummary",
+              timeoutMs,
            );
-            console.log("[Smart Auto Mode] ✅ Enriched image analysis summary");
+            if (summary?.trim()) break;
          }
+
+          if (summary?.trim()) {
+            const baseMessage =
+              smartAutoFriendlyMessage ||
+              buildAutoModeFallbackMessage({
+                text: latestUserText,
+                hasImage: true,
+                fromHistory: currentImageUrls.length === 0 && summaryImageUrls.length > 0,
+                imageSummary: summary.trim(),
+              });
+            smartAutoFriendlyMessage = ensureImageAnalysisSummarySection(baseMessage, summary.trim());
+            console.log("[Smart Auto Mode] ✅ Image analysis summary ready");
+          } else if (smartAutoFriendlyMessage && shouldEnrichImageAnalysisText(smartAutoFriendlyMessage)) {
+            console.warn("[Smart Auto Mode] ⚠️ Image summary unavailable, keep original analysis text");
+          }
+        } catch (enrichError) {
+          console.warn("[Smart Auto Mode] Skip image analysis enrichment:", enrichError);
        }
-      } catch (enrichError) {
-        console.warn("[Smart Auto Mode] Skip image analysis enrichment:", enrichError);
      }
    }

+    if (shouldRunAgentConversation && smartAutoFriendlyMessage) {
+      smartAutoFriendlyMessage = normalizeAgentReportForDisplay(smartAutoFriendlyMessage);
+    }
+
    const session = await auth();

    if (!session?.user) {
--- a/lib/ai/auto-agent-graph.ts
+++ b/lib/ai/auto-agent-graph.ts
@ -105,7 +105,8 @@ async function withTimeoutFallback<T>(

 const AGENT_VISION_MAX_IMAGES = 4;
 const AGENT_VISION_MAX_SIDE = 500;
-const AGENT_VISION_FETCH_TIMEOUT_MS = 10_000;
+const AGENT_OPERATION_TIMEOUT_MS = 120_000;
+const AGENT_VISION_FETCH_TIMEOUT_MS = AGENT_OPERATION_TIMEOUT_MS;

 function isHttpUrl(value: string) {
  return /^https?:\/\//i.test(value);
@ -514,7 +515,10 @@ function buildStrategicFriendlyMessage(
    const isStructuredReply =
      /###\s*需求理解/.test(existingReply) ||
      /###\s*图片分析/.test(existingReply) ||
-      /###\s*执行计划/.test(existingReply);
+      /###\s*执行计划/.test(existingReply) ||
+      /【\s*需求理解\s*】/.test(existingReply) ||
+      /【\s*图片分析\s*】/.test(existingReply) ||
+      /【\s*执行计划\s*】/.test(existingReply);
    if (isStructuredReply || existingReply.length >= 24) {
      return Promise.resolve(existingReply);
    }
@ -522,7 +526,7 @@ function buildStrategicFriendlyMessage(

  return withTimeoutFallback(
    generateAgentConversation(analysis, state),
-    12000,
+    AGENT_OPERATION_TIMEOUT_MS,
    existingReply || null,
  );
 }
@ -727,7 +731,7 @@ async function generateAgentConversation(
          ],
          temperature: 0.2,
          maxTokens: 360,
-          abortSignal: AbortSignal.timeout(12000),
+          abortSignal: AbortSignal.timeout(AGENT_OPERATION_TIMEOUT_MS),
        } as any);

        visualAnalysis = visionText.trim();
@ -751,12 +755,12 @@ async function generateAgentConversation(
 输出要求：
 - 不要说“看不到图片/无法访问图片”。
 - 不暴露内部实现、模型路由、算法细节。
- 输出 Markdown，控制在 8-14 行。
- 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
+- 输出控制在 8-14 行，标题行不要使用 # 号。
+- 有图片时必须按顺序输出以下标题（逐字一致）：
+  【需求理解】
+  【图片分析】
+  【执行计划】
+  【输出设置】
 - “图片分析”必须基于视觉分析结果写具体观察，不允许写泛化空话。
 - 只有明确拿到原图分辨率时，才能写具体像素尺寸；否则只写宽高比/清晰度目标。`;

@ -780,7 +784,7 @@ async function generateAgentConversation(
      prompt,
      temperature: 0.35,
      maxTokens: 520,
-      abortSignal: AbortSignal.timeout(12000),
+      abortSignal: AbortSignal.timeout(AGENT_OPERATION_TIMEOUT_MS),
    } as any);

    const cleaned = text.trim();
@ -854,7 +858,7 @@ async function recommendImageParamsWithAgent(
      prompt: requestText,
      temperature: 0.15,
      maxTokens: 220,
-      abortSignal: AbortSignal.timeout(10000),
+      abortSignal: AbortSignal.timeout(AGENT_OPERATION_TIMEOUT_MS),
    } as any);

    const parsed = parseJsonBlock<AgentRecommendedImageParams>(text);
@ -939,7 +943,7 @@ async function recommendVideoPlanWithAgent(
      prompt,
      temperature: 0.2,
      maxTokens: 720,
-      abortSignal: AbortSignal.timeout(12000),
+      abortSignal: AbortSignal.timeout(AGENT_OPERATION_TIMEOUT_MS),
    } as any);

    const parsed = parseJsonBlock<Record<string, unknown>>(text);
@ -1011,7 +1015,7 @@ async function refineRouteWithAgent(
 ${candidatesText}`,
      temperature: 0.1,
      maxTokens: 120,
-      abortSignal: AbortSignal.timeout(8000),
+      abortSignal: AbortSignal.timeout(AGENT_OPERATION_TIMEOUT_MS),
    } as any);

    const parsed = parseJsonBlock<{ modelId?: string; reason?: string }>(text);
@ -1108,12 +1112,27 @@ const optimizeNode = async (state: AutoAgentStateType) => {
  if (state.analysis.intent !== "image_edit" && state.analysis.intent !== "image_gen") {
    return {};
  }
-  const { optimizePrompt } = await import("./prompt-optimizer");
-  const optimized = optimizePrompt({
-    originalPrompt: state.text,
-    intent: state.analysis.intent,
-    subIntent: state.analysis.subIntent,
-  });
+  const { optimizePrompt, optimizePromptWithLLM } = await import("./prompt-optimizer");
+  const optimizedFromLlm = await withTimeoutFallback<
+    Awaited<ReturnType<typeof optimizePromptWithLLM>> | null
+  >(
+    optimizePromptWithLLM({
+      originalPrompt: state.text,
+      intent: state.analysis.intent,
+      subIntent: state.analysis.subIntent,
+      historyMessages: state.historyMessages,
+      context: state.context || undefined,
+    }),
+    AGENT_OPERATION_TIMEOUT_MS,
+    null,
+  );
+  const optimized =
+    optimizedFromLlm ||
+    optimizePrompt({
+      originalPrompt: state.text,
+      intent: state.analysis.intent,
+      subIntent: state.analysis.subIntent,
+    });
  const mergedParams = sanitizeRecommendedImageParams(
    {
      ...optimized.recommendedImageParams,
--- a/lib/ai/smart-analyzer.ts
+++ b/lib/ai/smart-analyzer.ts
@ -240,7 +240,8 @@ const ANALYZER_SYSTEM_PROMPT = `你是一个友好的AI图片编辑助手。用

 const ANALYZER_MAX_IMAGE_COUNT = 2;
 const ANALYZER_IMAGE_MAX_SIDE = 500;
-const ANALYZER_IMAGE_FETCH_TIMEOUT_MS = 10000;
+const ANALYZER_TIMEOUT_MS = 120000;
+const ANALYZER_IMAGE_FETCH_TIMEOUT_MS = ANALYZER_TIMEOUT_MS;
 const ZHIPU_VISION_MODEL_CANDIDATES = [
  "glm-4.6v-flash",
  "glm-4.6v",
@ -668,7 +669,7 @@ export async function summarizeImageForExecutionReport(params: {
      ],
      temperature: 0.1,
      maxTokens: 120,
-      abortSignal: AbortSignal.timeout(8000),
+      abortSignal: AbortSignal.timeout(ANALYZER_TIMEOUT_MS),
    } as any);

    const normalized = summaryText?.replace(/\s+/g, " ").trim();
@ -971,7 +972,7 @@ ${historyText}
        ],
        temperature: 0.1,
        maxTokens: 360,
-        abortSignal: AbortSignal.timeout(12000),
+        abortSignal: AbortSignal.timeout(ANALYZER_TIMEOUT_MS),
      }
    : {
        model,
@ -979,7 +980,7 @@ ${historyText}
        prompt: userMessage,
        temperature: 0.1,
        maxTokens: 300,
-        abortSignal: AbortSignal.timeout(10000),
+        abortSignal: AbortSignal.timeout(ANALYZER_TIMEOUT_MS),
      };

  const { text: responseText } = await generateText(requestPayload as any);