fix: reject unchanged text edit results

app/api/canvas/text-edit/route.ts

@@ -225,6 +225,36 @@ function isHttpImageUrl(value: unknown) {
   }
 }
 
+function getComparableImageIdentity(value: unknown) {
+  const raw = normalizeGeneratedImageUrl(value);
+  if (!isHttpImageUrl(raw)) {
+    return "";
+  }
+
+  try {
+    const parsed = new URL(raw);
+    parsed.hash = "";
+    // Query params often include thumbnails/cache-busters. Same host+path still means
+    // the model returned the input asset instead of a newly generated image.
+    return `${parsed.protocol}//${parsed.host}${parsed.pathname}`.toLowerCase();
+  } catch {
+    return raw.split("#")[0]?.split("?")[0]?.trim().toLowerCase() || "";
+  }
+}
+
+function createExcludedImageIdentitySet(urls?: Iterable<string>) {
+  return new Set(
+    Array.from(urls || [])
+      .map(getComparableImageIdentity)
+      .filter(Boolean)
+  );
+}
+
+function isExcludedImageUrl(url: string, excludedIdentities: Set<string>) {
+  const identity = getComparableImageIdentity(url);
+  return Boolean(identity && excludedIdentities.has(identity));
+}
+
 function collectImageUrlsFromUnknown(value: unknown, output: Set<string>) {
   if (value === null || value === undefined) return;
 
@@ -261,25 +291,38 @@ function collectImageUrlsFromUnknown(value: unknown, output: Set<string>) {
   }
 }
 
-function pickGeneratedImageUrl(urls: Iterable<string>) {
+function pickGeneratedImageUrl(
+  urls: Iterable<string>,
+  excludedIdentities = new Set<string>()
+) {
   const candidates = Array.from(new Set(Array.from(urls).map(normalizeGeneratedImageUrl)))
-    .filter(isHttpImageUrl);
+    .filter(isHttpImageUrl)
+    .filter((url) => !isExcludedImageUrl(url, excludedIdentities));
   const imageLike = candidates.find((url) =>
     /\.(avif|gif|jpe?g|png|webp)(?:[?#]|$)/i.test(url)
   );
   return imageLike || candidates[0] || "";
 }
 
-function parseImageFromGenerationText(text: string): { originalUrl: string; thumbnailUrl: string } {
+function parseImageFromGenerationText(
+  text: string,
+  excludedSourceUrls: string[] = []
+): { originalUrl: string; thumbnailUrl: string } {
   const markdown = String(text || "").trim();
+  const excludedIdentities = createExcludedImageIdentitySet(excludedSourceUrls);
+  const seenUrls = new Set<string>();
   const strictMatch = markdown.match(/!\[thumb\|\|\|([^|]+)\|\|\|[^\]]*\]\(([^)]+)\)/);
   if (strictMatch) {
     const originalUrl = normalizeGeneratedImageUrl(strictMatch[1]);
     const thumbnailUrl = normalizeGeneratedImageUrl(strictMatch[2]);
-    if (isHttpImageUrl(originalUrl)) {
+    [originalUrl, thumbnailUrl].filter(isHttpImageUrl).forEach((url) => seenUrls.add(url));
+    if (isHttpImageUrl(originalUrl) && !isExcludedImageUrl(originalUrl, excludedIdentities)) {
       return {
         originalUrl,
-        thumbnailUrl: isHttpImageUrl(thumbnailUrl) ? thumbnailUrl : originalUrl,
+        thumbnailUrl:
+          isHttpImageUrl(thumbnailUrl) && !isExcludedImageUrl(thumbnailUrl, excludedIdentities)
+            ? thumbnailUrl
+            : originalUrl,
       };
     }
   }
@@ -291,10 +334,11 @@ function parseImageFromGenerationText(text: string): { originalUrl: string; thum
     const url = normalizeGeneratedImageUrl(markdownMatch[1]);
     if (isHttpImageUrl(url)) {
       markdownUrls.add(url);
+      seenUrls.add(url);
     }
   }
 
-  const markdownUrl = pickGeneratedImageUrl(markdownUrls);
+  const markdownUrl = pickGeneratedImageUrl(markdownUrls, excludedIdentities);
   if (markdownUrl) {
     return {
       originalUrl: markdownUrl,
@@ -308,8 +352,9 @@ function parseImageFromGenerationText(text: string): { originalUrl: string; thum
   } catch {
     collectImageUrlsFromUnknown(markdown, collectedUrls);
   }
+  collectedUrls.forEach((url) => seenUrls.add(url));
 
-  const url = pickGeneratedImageUrl(collectedUrls);
+  const url = pickGeneratedImageUrl(collectedUrls, excludedIdentities);
   if (url) {
     return {
       originalUrl: url,
@@ -317,6 +362,10 @@ function parseImageFromGenerationText(text: string): { originalUrl: string; thum
     };
   }
 
+  if (Array.from(seenUrls).some((url) => isExcludedImageUrl(url, excludedIdentities))) {
+    throw new Error("模型返回的是输入原图，未完成文字替换");
+  }
+
   throw new Error("未解析到生成图片");
 }
 
@@ -331,6 +380,7 @@ function buildTextEditPrompt(
     "2. 必须保持原图排版、字重、笔画风格与字体形状尽可能一致，仅替换指定文字内容，其他视觉元素不变。",
     "3. 输出图严格保持原图宽高比例，不裁切、不拉伸、不加边框。",
     "4. 若目标文案更长，优先微调字距或字号，确保自然贴合原位置。",
+    "5. 必须输出一张新生成的图片，不要直接返回输入原图或源图链接；如果无法完成替换，请明确失败，不要用原图冒充结果。",
   ];
 
   if (replacements.length > 0) {
@@ -358,6 +408,7 @@ function buildTextEditPrompt(
 async function generateTextEditImageWithFailover(params: {
   candidates: TextEditModelCandidate[];
   imageUrl: string;
+  sourceImageUrls: string[];
   prompt: string;
 }) {
   let lastError: unknown = null;
@@ -399,7 +450,7 @@ async function generateTextEditImageWithFailover(params: {
         throw new Error(resultText);
       }
 
-      const parsed = parseImageFromGenerationText(resultText);
+      const parsed = parseImageFromGenerationText(resultText, params.sourceImageUrls);
       return {
         ...parsed,
         model: selected.model.displayName,
@@ -492,6 +543,7 @@ export async function POST(request: Request) {
     const { originalUrl, thumbnailUrl, model } = await generateTextEditImageWithFailover({
       candidates,
       imageUrl: generationSourceImageUrl,
+      sourceImageUrls: [imageUrl, generationSourceImageUrl],
       prompt,
     });
 

components/agents/AgentImageTextEditDialog.tsx

@@ -101,6 +101,27 @@ function isHttpImageUrl(value?: string | null) {
   }
 }
 
+function getComparableImageIdentity(value?: string | null) {
+  const raw = String(value || "").trim();
+  if (!isHttpImageUrl(raw)) {
+    return "";
+  }
+
+  try {
+    const parsed = new URL(raw);
+    parsed.hash = "";
+    return `${parsed.protocol}//${parsed.host}${parsed.pathname}`.toLowerCase();
+  } catch {
+    return raw.split("#")[0]?.split("?")[0]?.trim().toLowerCase() || "";
+  }
+}
+
+function isSameImageAsset(left?: string | null, right?: string | null) {
+  const leftIdentity = getComparableImageIdentity(left);
+  const rightIdentity = getComparableImageIdentity(right);
+  return Boolean(leftIdentity && rightIdentity && leftIdentity === rightIdentity);
+}
+
 export function AgentImageTextEditDialog({
   open,
   imageUrl,
@@ -380,6 +401,10 @@ export function AgentImageTextEditDialog({
         throw new Error(result.error || "文字替换失败");
       }
 
+      if (isSameImageAsset(result.data.originalUrl, normalizedImageUrl)) {
+        throw new Error("模型返回的仍是原图，未完成文字替换，请换一条文案或稍后重试");
+      }
+
       toast.success("文字替换已完成");
       setDeletedTextItems([]);
       setDetectedSourceUrl(result.data.originalUrl);