Refine agent text mode workflow

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

@@ -98,6 +98,11 @@ import type {
 } from "@/lib/ai/context-memory/types";
 import { ChatSDKError } from "@/lib/errors";
 import { extractImageUrlsFromText } from "@/lib/image-url";
+import {
+  buildPromptMemoryUpdate,
+  extractPromptMemoryFromWorkingState,
+  isPromptCarryForwardRequest,
+} from "@/lib/ai/prompt-memory";
 import { uploadFile } from "@/lib/storage/service";
 import { getVideoStoryboardConfig } from "@/lib/video-agent/config";
 import { runRuxaNineGridPipeline } from "@/lib/video-agent/pipeline/ruxa-nine-grid-pipeline";
@@ -3059,6 +3064,7 @@ export async function POST(request: Request) {
       | Awaited<ReturnType<typeof getConversationWorkingStateByChatId>>
       | null = null;
     let storyboardPreferenceMemory = null;
+    let reusablePromptMemory = null;
     let smartAutoHistoryMessagesFromDb: DBMessage[] = [];
     let preferredAgentModelId: string | undefined;
     let pendingGridConfirmation:
@@ -3186,6 +3192,7 @@ export async function POST(request: Request) {
       return null;
     });
     storyboardPreferenceMemory = extractStoryboardPreferenceMemory(requestWorkingState);
+    reusablePromptMemory = extractPromptMemoryFromWorkingState(requestWorkingState);
     pendingGridConfirmation =
       extractPendingGridConfirmationFromWorkingState(requestWorkingState);
 
@@ -3234,9 +3241,22 @@ export async function POST(request: Request) {
             chatId: id,
           });
         }
-        confirmedPromptForGeneration = extractConfirmedPromptFromWorkingState(requestWorkingState);
+        confirmedPromptForGeneration =
+          extractConfirmedPromptFromWorkingState(requestWorkingState) ||
+          reusablePromptMemory?.optimizedPrompt;
+        const hasImageInputForPromptCarryForward =
+          message.parts.some((part: any) =>
+            part?.type === "file" || part?.type === "image" || part?.type === "image_url"
+          );
+        const isPromptCarryForwardExecution =
+          Boolean(confirmedPromptForGeneration) &&
+          isPromptCarryForwardRequest({
+            text: requestTextContent,
+            hasImage: hasImageInputForPromptCarryForward,
+          });
         forcePromptDrivenImageGeneration =
-          isConfirmedPromptGenerationRequest(requestTextContent) &&
+          (isConfirmedPromptGenerationRequest(requestTextContent) ||
+            isPromptCarryForwardExecution) &&
           Boolean(confirmedPromptForGeneration);
         if (!requestResolvedMediaInput) {
           requestResolvedMediaInput = resolveChatMediaInput({
@@ -3457,7 +3477,10 @@ export async function POST(request: Request) {
             availableTools: [],
           };
           smartAutoFriendlyMessage =
-            smartAutoFriendlyMessage || "收到，我会按你刚确认的提示词直接生图。";
+            smartAutoFriendlyMessage ||
+            (isPromptCarryForwardExecution
+              ? "收到，我会使用上一轮优化后的提示词，结合你这轮图片继续执行。"
+              : "收到，我会按你刚确认的提示词直接生图。");
         }
 
         if (smartAutoPlan) {
@@ -4689,6 +4712,12 @@ export async function POST(request: Request) {
                 resolutionMode: "prompt_only",
                 confidence: 90,
                 optimizedPrompt: promptOnlyOptimizedPrompt.slice(0, 1000),
+                promptMemory: buildPromptMemoryUpdate({
+                  optimizedPrompt: promptOnlyOptimizedPrompt.slice(0, 1000),
+                  sourceIntent: "image_gen",
+                  sourceText: requestTextContent.slice(0, 500),
+                  reusable: true,
+                }),
                 imageSummary: promptOnlyImageSummary.slice(0, 1200),
                 sourceImageUrls: promptOnlySourceImageUrls,
               },

lib/ai/prompt-memory.ts

@@ -0,0 +1,92 @@
+import "server-only";
+
+export interface PromptMemoryEntry {
+  optimizedPrompt: string;
+  sourceIntent?: string;
+  sourceText?: string;
+  reusable?: boolean;
+}
+
+function normalizeText(value: unknown, max = 2000): string | undefined {
+  const normalized = String(value || "").trim();
+  if (!normalized) return undefined;
+  return normalized.slice(0, max);
+}
+
+export function extractPromptMemoryFromWorkingState(
+  workingState?: {
+    lastTurnSummary?: unknown;
+  } | null,
+): PromptMemoryEntry | null {
+  const summary = workingState?.lastTurnSummary;
+  if (!summary || typeof summary !== "object") return null;
+  const summaryRecord = summary as Record<string, unknown>;
+  const memory = summaryRecord.promptMemory;
+  if (memory && typeof memory === "object") {
+    const memoryRecord = memory as Record<string, unknown>;
+    const optimizedPrompt = normalizeText(memoryRecord.optimizedPrompt, 4000);
+    if (!optimizedPrompt) return null;
+    return {
+      optimizedPrompt,
+      sourceIntent: normalizeText(memoryRecord.sourceIntent, 64),
+      sourceText: normalizeText(memoryRecord.sourceText, 500),
+      reusable:
+        typeof memoryRecord.reusable === "boolean" ? memoryRecord.reusable : true,
+    };
+  }
+
+  const resolutionMode = normalizeText(summaryRecord.resolutionMode, 64);
+  const optimizedPrompt = normalizeText(summaryRecord.optimizedPrompt, 4000);
+  if (resolutionMode === "prompt_only" && optimizedPrompt) {
+    return {
+      optimizedPrompt,
+      sourceIntent: "image_gen",
+      sourceText: normalizeText(summaryRecord.text, 500),
+      reusable: true,
+    };
+  }
+  return null;
+}
+
+export function buildPromptMemoryUpdate(params: {
+  optimizedPrompt: string;
+  sourceIntent?: string;
+  sourceText?: string;
+  reusable?: boolean;
+}): PromptMemoryEntry {
+  return {
+    optimizedPrompt: normalizeText(params.optimizedPrompt, 4000) || "",
+    sourceIntent: normalizeText(params.sourceIntent, 64),
+    sourceText: normalizeText(params.sourceText, 500),
+    reusable: params.reusable !== false,
+  };
+}
+
+export function isPromptCarryForwardRequest(params: {
+  text: string;
+  hasImage: boolean;
+}): boolean {
+  const normalized = String(params.text || "").toLowerCase().replace(/\s+/g, "");
+  if (!normalized) return false;
+
+  const refersPreviousPrompt =
+    /根据(上一个|上一句|刚才|上面|之前|那个)?提示词|按(上一个|上一句|刚才|上面|之前|那个)?提示词|用(上一个|上一句|刚才|上面|之前|那个)?提示词|用刚才输出的提示词|按刚才输出的提示词|按上面的提示词|沿用上个提示词|继续用那个提示词/.test(
+      normalized,
+    );
+
+  if (!refersPreviousPrompt) return false;
+
+  const wantsExecution =
+    /生图|生成图片|出图|画出来|做图|继续生成|继续做|重新做|重绘|优化这张图|优化这个图|处理这张图|处理这个图|按这个做|照这个做|按此生成/.test(
+      normalized,
+    );
+
+  if (wantsExecution) return true;
+
+  if (params.hasImage) {
+    return /优化|处理|重绘|参考这张图|结合这张图|基于这张图/.test(normalized);
+  }
+
+  return false;
+}
+

lib/db/queries.ts

@@ -2517,7 +2517,12 @@ export async function createAgentPreset(params: {
         slots: params.slots,
       });
 
-      const requiredUploadCount = validatedSlots.some((slot) => slot.requireUpload !== false) ? 3 : 0;
+      const requiredUploadCount =
+        params.type === "text"
+          ? 0
+          : validatedSlots.some((slot) => slot.requireUpload !== false)
+            ? 3
+            : 0;
       const avatarAssetId = String(params.avatarAssetId || "").trim() || null;
       if (avatarAssetId) {
         await validateOwnedReferenceAssets({
@@ -2705,7 +2710,12 @@ export async function updateAgentPreset(params: {
         userId: params.userId,
         slots: normalizedSlotsInput,
       });
-      const requiredUploadCount = validatedSlots.some((slot) => slot.requireUpload !== false) ? 3 : 0;
+      const requiredUploadCount =
+        nextType === "text"
+          ? 0
+          : validatedSlots.some((slot) => slot.requireUpload !== false)
+            ? 3
+            : 0;
       const nextAvatarAssetId =
         params.avatarAssetId !== undefined
           ? String(params.avatarAssetId || "").trim() || null