提交货源图片和相似ASIN优化

.gitignore

@@ -43,6 +43,7 @@ MANIFEST
 .installed.cfg
 
 # Build / packaging
+app.zip
 build/
 dist/
 develop-eggs/
@@ -108,4 +109,4 @@ xlsx/
 OPS_REDIS_MYSQL_OPTIMIZATION_NOTES.md
 架构.md
 *ts.%
-.omc
+.omc

app/amazon/amazon_base.py

@@ -417,8 +417,8 @@ class AmamzonBase(ZiniaoDriver):
         """
         try:
             timestamp = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
-            if level == "ERROR":
-                show_notification(message, "error")
+            # if level == "ERROR":
+            #     show_notification(message, "error")
             print(f"[{timestamp}] [{self.mark_name}] [{level}] {message}")
         except Exception as e:
             print(f"输出出错,{e}")
@@ -689,8 +689,8 @@ class TaskBase:
         """
         try:
             timestamp = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
-            if level == "ERROR":
-                show_notification(message, "error")
+            # if level == "ERROR":
+            #     show_notification(message, "error")
             print(f"[{timestamp}] [{self.task_name}] [{level}] {message}")
         except Exception as e:
             print(f"输出出错,{e}")
@@ -739,7 +739,7 @@ class TaskBase:
                     if not shop_data:
                         mes = f"获取店铺凭证失败，响应数据: {shop_data.get('message', '未知错误')}"
                         self.log(mes, "ERROR")
-                        show_notification(mes, "ERROR")
+                        # show_notification(mes, "ERROR")
                         continue
 
                     password = shop_data["data"]["password"]

app/amazon/approve.py

@@ -674,12 +674,12 @@ class ApproveTask(TaskBase):
                 return
 
             current_url = None
+            max_retries = 200  #只要没有完成，一直重试
             for _ in range(max_retries):
                 try:
                     self.process_country(driver, country_code, task_id, shop_name,risk_listing_filter,target_url=current_url)
                     driver.reset_already_asin()
                     driver.close_store()
-
                     break
                 except Exception as e:
                     self.log(f"处理国家 {country_code} 失败: {str(e)}", "ERROR")
@@ -777,7 +777,7 @@ class ApproveTask(TaskBase):
                 }
             result.append(country_data)
 
-            if len(result) > 20:
+            if len(result) > 10:
                 self.post_result_batch(task_id, shop_name, country_code, result)
                 result = []
 

app/amazon/chrome_base.py

@@ -9,7 +9,7 @@ from datetime import datetime
 from DrissionPage import Chromium, ChromiumOptions
 from collections import defaultdict
 
-from config import base_dir
+from config import base_dir,debug
 from amazon.tool import get_shop_info,show_notification
 
 
@@ -46,7 +46,8 @@ class ChromeAmzoneBase:
         杀死当前谷歌浏览器进程，并使用 drissionpage 启动谷歌浏览器，使用系统安装的浏览器默认用户文件夹
         """
         print("正在关闭现有的chromium浏览器进程...")
-        os.system('taskkill /f /t /im chrome.exe')
+        if not debug:
+            os.system('taskkill /f /t /im chrome.exe')
         time.sleep(2)
 
         print("正在启动chromium浏览器...")
@@ -70,8 +71,8 @@ class ChromeAmzoneBase:
         """
         try:
             timestamp = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
-            if level == "ERROR":
-                show_notification(message, "error")
+            # if level == "ERROR":
+            #     show_notification(message, "error")
             print(f"[{timestamp}] [{self.mark_name}] [{level}] {message}")
         except Exception as e:
             print(f"输出出错,{e}")

app/amazon/detail_spider.py

@@ -79,7 +79,8 @@ class ChromeAmzone(ChromeAmzoneBase):
         except Exception as e:
             error_msg = f"运行出错: {traceback.format_exc()}"
             self.log(error_msg)
-            show_notification(f"采集失败: {str(e)}", "error")
+            # show_notification(f"采集失败: {str(e)}", "error")
+            raise RuntimeError(error_msg)
             return {}
 
     def _scrape_data(self):
@@ -100,8 +101,27 @@ class ChromeAmzone(ChromeAmzoneBase):
             title_ele = self.tab.ele('xpath://h1[@id="title"]',timeout=30)
             title = title_ele.text
             data["title"] = title
+            sku_ele_ls = self.tab.eles('xpath://ul[@class="a-unordered-list a-vertical a-spacing-mini"]', timeout=20)
+            if len(sku_ele_ls) > 0:
+                data["title"] += sku_ele_ls[0].text
             imge_ele = self.tab.ele('xpath://div[@id="imgTagWrapperId"]//img',timeout=20)
-            image_url = imge_ele.attr("src")
+
+            image_url = ""
+            data_a_dynamic_image = imge_ele.attr("data-a-dynamic-image")
+            if data_a_dynamic_image:
+                dynamic_image_json = json.loads(data_a_dynamic_image)
+                self.log(f"图片信息:{dynamic_image_json}")
+                max_area = 0
+
+                for url, (width, height) in dynamic_image_json.items():
+                    area = width * height
+                    if area > max_area:
+                        max_area = area
+                        image_url = url
+
+            if not image_url:
+                image_url = imge_ele.attr("src")
+
             data["image_url"] = image_url
             return data
 
@@ -263,7 +283,7 @@ class SpiderTask(TaskBase):
                     result.append(res)
                     print("================")
                     is_done = gp_index == len(groups)-1
-                    if len(result) > 20 or is_done:
+                    if len(result) > 10 or is_done:
                         self.post_result(task_id=task_id,chunkIndex=gp_index+1,chunkTotal=len(groups),
                                          asin=asin,item_data=result,is_done=is_done)
                         result = []

app/amazon/match_action.py

@@ -378,7 +378,7 @@ class MatchTak(TaskBase):
                                 "done": False
                             })
 
-            if len(result) > 20:
+            if len(result) > 10:
                 self.post_result_batch(task_id, shop_name, country_code,result)
                 result = []
 

app/amazon/similar_asin.py

@@ -10,9 +10,10 @@ from datetime import datetime
 from collections import defaultdict
 import requests
 from urllib.parse import quote
-
+import os
 from curl_cffi import requests as requests_frp
 
+from config import base_dir
 
 
 from amazon.tool import show_notification,get_shop_info,remove_special_characters,split_currency_values
@@ -76,7 +77,9 @@ class ChromeAmzone(ChromeAmzoneBase):
         将图片链接（URL 或本地文件路径）转换为 Base64 编码的字符串。
         """
         if image_source.startswith(('http://', 'https://')):
-            response = requests.get(image_source, timeout=10)
+            response = requests.get(image_source, timeout=10,headers={
+                "user-agent":"Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/142.0.0.0 Safari/537.36 Edg/142.0.0.0 18444"
+            })
             response.raise_for_status()  # 非 2xx 状态码将抛出异常
             image_data = response.content
         else:
@@ -153,7 +156,7 @@ class ChromeAmzone(ChromeAmzoneBase):
         # 发送第一次请求
         try:
             response = requests_frp.post(url, headers=headers, params=params, json=data, impersonate="chrome101",
-                                     cookies=cookie, proxies=proxies)
+                                     cookies=cookie, proxies=proxies,verify=False)
             response.encoding = "utf-8"
 
             # 检查响应状态码和数据有效性
@@ -176,6 +179,7 @@ class ChromeAmzone(ChromeAmzoneBase):
                     need_retry = True
 
             # 如果需要重试且配置了代理URL
+            # need_retry = True
             if need_retry and CONFIG_PROXY_URL and not proxies:
                 try:
                     proxy_resp = requests.get(CONFIG_PROXY_URL, timeout=10)
@@ -204,7 +208,7 @@ class ChromeAmzone(ChromeAmzoneBase):
                         if proxy_ip:
                             proxies = {
                                 "http": f"http://{proxy_ip}",
-                                "https": f"http://{proxy_ip}",
+                                "https": f"https://{proxy_ip}",
                             }
 
                     if proxies:
@@ -251,7 +255,23 @@ class ChromeAmzone(ChromeAmzoneBase):
             title = title_ele.text
             data["title"] = title
             imge_ele = self.tab.ele('xpath://div[@id="imgTagWrapperId"]//img',timeout=20)
-            image_url = imge_ele.attr("src")
+
+            image_url = ""
+            data_a_dynamic_image = imge_ele.attr("data-a-dynamic-image")
+            if data_a_dynamic_image:
+                dynamic_image_json = json.loads(data_a_dynamic_image)
+                self.log(f"图片信息:{dynamic_image_json}")
+                max_area = 0
+
+                for url, (width, height) in dynamic_image_json.items():
+                    area = width * height
+                    if area > max_area:
+                        max_area = area
+                        image_url = url
+
+            if not image_url:
+                image_url = imge_ele.attr("src")
+
             data["image_url"] = image_url
 
             # sku
@@ -277,12 +297,13 @@ class ChromeAmzone(ChromeAmzoneBase):
         Returns:
             dict: 包含采集到的数据
         """
+        return_data = {}
         try:
             # 验证国家是否支持
             if country not in self.country_info:
                 error_msg = f"不支持的国家: {country}，支持的国家有: {list(self.country_info.keys())}"
                 print(error_msg)
-                show_notification(error_msg, "error")
+                # show_notification(error_msg, "error")
                 return None
 
             # 获取国家配置
@@ -314,6 +335,8 @@ class ChromeAmzone(ChromeAmzoneBase):
             self.log("正在抓取商品数据...")
             data = self._scrape_data()
 
+            return_data.update(data)
+
             # 判断是否采集标题出错
             new_size = "220,220"
             pattern = r"\._(?:[A-Z]+)?(\d+)_\."
@@ -321,22 +344,25 @@ class ChromeAmzone(ChromeAmzoneBase):
             # 使用正则替换
             image_new_url = re.sub(pattern, lambda m: replacement, data["image_url"])
 
+            print(image_new_url)
+
             iamge_base64 = self.image_to_base64(image_new_url)
 
             similar_data = []
             # 4、请求获取插件的数据
-            for page_num in range(total_page):
-                resp_data = self.get_aliprice_data(
-                    page=page_num,
-                    title = data["title"],
-                    domain=domain,
-                    category=data["category"],
-                    imageBase64 = iamge_base64
-                )
-                self.log(f"Aliprice 扩展数据获取:{str(resp_data)[0:200]}")
-                if len(resp_data.get("data",[])) > 0:
-                    for i in resp_data.get("data"):
-                        similar_data.append(i)
+            if image_new_url:
+                for page_num in range(total_page):
+                    resp_data = self.get_aliprice_data(
+                        page=page_num+1,
+                        title = data["title"],
+                        domain=domain,
+                        category=data["category"],
+                        imageBase64 = iamge_base64
+                    )
+                    self.log(f"Aliprice 扩展数据获取:{str(resp_data)[0:200]}")
+                    if len(resp_data.get("data",[])) > 0:
+                        for i in resp_data.get("data"):
+                            similar_data.append(i)
 
             data["similar_data"] = similar_data
 
@@ -346,14 +372,17 @@ class ChromeAmzone(ChromeAmzoneBase):
             data['url'] = product_url
             data['timestamp'] = datetime.now().strftime('%Y-%m-%d %H:%M:%S')
 
-            print(f"数据抓取完成: {json.dumps(data)}")
-            return data
+            print(f"数据抓取完成: {json.dumps(return_data)}")
+            return_data.update(data)
+
+            return return_data
 
         except Exception as e:
             error_msg = f"运行出错: {traceback.format_exc()}"
             print(error_msg)
-            show_notification(f"采集失败: {str(e)}", "error")
-            return {}
+            # show_notification(f"采集失败: {str(e)}", "error")
+            raise RuntimeError(f"{traceback.format_exc()}")
+            return return_data
 
 
 class SimilarAsinTask(TaskBase):
@@ -478,10 +507,9 @@ class SimilarAsinTask(TaskBase):
                 result = []
                 for gp_index,gp in enumerate(groups):
                     items = gp.get("items", [])
-                    return_data = {}
-
+                    group_item = []
                     for index,value in enumerate(items):
-
+                        print(value)
                         return_data = {
                             'image_url': "",
                             'title': "",
@@ -494,37 +522,36 @@ class SimilarAsinTask(TaskBase):
                         country = value.get("country")
                         for _ in range(max_retry):
                             try:
-                                return_data = chrome.run(country, asin,total_page=2) or {}
+                                return_data = chrome.run(country, asin,total_page=1) or {}
                                 self.log(f"抓取结果->{return_data}")
                                 break
                             except Exception as e:
                                 # if "与页面的连接已断开" in str(e):
                                 chrome = ChromeAmzone()
                                 self.log(f"{asin}抓取数据报错,{e}")
-                        if not isinstance(return_data, dict):
-                            return_data = {}
-                        if return_data.get("image_url"):
-                            break
+                        # if not isinstance(return_data, dict):
+                        #     return_data = {}
+                        # if return_data.get("image_url"):
+                        #     break
+
+                        group_item.append({
+                                "sourceFileKey": value.get("sourceFileKey",""),
+                                "sourceFilename": value.get("sourceFilename",""),
+                                "rowToken": value.get("rowToken",""),
+                                "groupKey": value.get("groupKey",""),
+                                "id": value.get("values",{}).get("id",""),
+                                "asin": value.get("asin",""),
+                                "sku" : return_data.get("sku",""),
+                                "country": value.get("country",""),
+                                "url": return_data.get("image_url",""),
+                                "title": return_data.get("title",""),
+                                "done": False,
+                                "urls" : [i.get("ori_picture") for i in return_data.get("similar_data",[])[:16]]
+                            })
+                        # task_id: int, chunkIndex:int,chunkTotal: int, country_code: str, asin: str, status: dict,error:str="",
+                        #                     item_data:dict={},
+
 
-                    group_item = [
-                        {
-                            "sourceFileKey": i.get("sourceFileKey"),
-                            "sourceFilename": i.get("sourceFilename"),
-                            "rowToken": i.get("rowToken"),
-                            "groupKey": i.get("groupKey"),
-                            "id": i.get("id"),
-                            "asin": i.get("asin"),
-                            "sku" : return_data.get("sku"),
-                            "country": i.get("country"),
-                            "url": return_data.get("image_url"),
-                            "title": return_data.get("title"),
-                            "done": False,
-                            "urls" : [i.get("ori_picture") for i in return_data.get("similar_data")]
-                        }
-                         for i in items
-                    ]
-                    # task_id: int, chunkIndex:int,chunkTotal: int, country_code: str, asin: str, status: dict,error:str="",
-                    #                     item_data:dict={},
                     res = {
                       "sourceFileKey": gp.get("sourceFileKey"),
                       "sourceFilename": gp.get("sourceFilename"),
@@ -629,7 +656,7 @@ class SimilarAsinTask(TaskBase):
                 time.sleep(2)
 
         self.log(f"已达到最大重试次数，结果回传最终失败", "ERROR")
-        raise RuntimeError("已达到最大重试次数，结果回传最终失败")
+        # raise RuntimeError("已达到最大重试次数，结果回传最终失败")
 
 if __name__ == '__main__':
     spide = ChromeAmzone()

backend-java/src/main/java/com/nanri/aiimage/config/SimilarAsinProperties.java

@@ -71,6 +71,17 @@ public class SimilarAsinProperties {
      */
     private boolean cozeUseLegacyItemFieldOrder = false;
 
+    /**
+     * 是否启用 P0-3 merge 增量缓冲：每个 batch DONE 时仅缓冲 cozeRows，
+     * 不立即合并到 chunk；finalize 前一次性按 chunkScopeHash 分组合并，
+     * 把 OSS chunk 读写从 1000+ 次降到 chunk 数量级。
+     * 仅作用于"正常 poll DONE"路径；失败 batch / 单 batch 任务 / 其他
+     * 11 个 mergeCozeRowsIntoChunk 调用点保留原立即 merge 行为。
+     * 出现问题时可通过 AIIMAGE_SIMILAR_ASIN_COZE_RESULT_BUFFER_ENABLED=false
+     * 一键回滚到老路径。
+     */
+    private boolean cozeResultBufferEnabled = true;
+
     @Data
     public static class CozeCredential {
         private String name;

backend-java/src/main/java/com/nanri/aiimage/modules/similarasin/service/SimilarAsinTaskService.java

@@ -34,6 +34,7 @@ import com.nanri.aiimage.modules.similarasin.model.vo.SimilarAsinParseVo;
 import com.nanri.aiimage.modules.similarasin.model.vo.SimilarAsinTaskBatchVo;
 import com.nanri.aiimage.modules.similarasin.model.vo.SimilarAsinTaskDetailVo;
 import com.nanri.aiimage.modules.similarasin.model.vo.SimilarAsinTaskItemVo;
+import com.nanri.aiimage.modules.similarasin.util.ExcelCellImageWriter;
 import com.nanri.aiimage.modules.similarasin.util.SimilarAsinImageEmbedder;
 import com.nanri.aiimage.modules.file.service.LocalFileStorageService;
 import com.nanri.aiimage.modules.file.service.oss.OssStorageService;
@@ -56,7 +57,6 @@ import lombok.extern.slf4j.Slf4j;
 import org.apache.poi.ss.usermodel.Cell;
 import org.apache.poi.ss.usermodel.CellStyle;
 import org.apache.poi.ss.usermodel.DataFormatter;
-import org.apache.poi.ss.usermodel.Drawing;
 import org.apache.poi.ss.usermodel.Font;
 import org.apache.poi.ss.usermodel.Row;
 import org.apache.poi.ss.usermodel.Sheet;
@@ -892,7 +892,7 @@ public class SimilarAsinTaskService {
             touchJavaSideTaskActivity(task.getId());
         } else if (isResultSubmissionComplete(task.getId())) {
             maybeFinalizeCozeJobLocked(task.getId(), new CozeBatchContext(
-                    job.getId(), result.getId(), context.scopeHash(), context.chunkIndex(), 1, 1, currentInstanceId(), 0, null));
+                    job.getId(), result.getId(), context.scopeHash(), context.chunkIndex(), 1, 1, currentInstanceId(), 0, null, null));
         }
     }
 
@@ -1436,6 +1436,9 @@ public class SimilarAsinTaskService {
             boolean shouldAssembleResult = assembleWorkbook && (finalError == null || finalError.isBlank() || hasPersistedResultRows);
             if (shouldAssembleResult && shouldAssembleSynchronously()) {
                 try {
+                    // P0-3：stale-recovery 同步 finalize 路径走这里，确保
+                    // assemble 之前缓冲的 cozeRows 已合并到 chunk。
+                    flushBufferedCozeResults(task.getId());
                     assembleResultWorkbook(task, result);
                 } catch (Exception ex) {
                     log.warn("[similar-asin] assemble result workbook failed taskId={} err={}", task.getId(), ex.getMessage());
@@ -1554,6 +1557,16 @@ public class SimilarAsinTaskService {
             saveFileBuildProgress(task, job, totalProgressUnits, 1, "Coze submitted, waiting for result");
             return false;
         }
+        // P0-3：stale-recovery / 异常路径兜底 —— maybeFinalizeCozeJob 可能未能成功 flush
+        // （Redis 锁竞争 / 异常退出），在 assemble 阶段读 chunk 之前再 flush 一次，幂等。
+        try {
+            flushBufferedCozeResults(task.getId());
+        } catch (Exception flushEx) {
+            log.warn("[similar-asin] result file job flush buffered coze results failed taskId={} jobId={} err={}",
+                    task.getId(), job.getId(), flushEx.getMessage(), flushEx);
+            throw new BusinessException("flush buffered coze results failed: "
+                    + firstNonBlank(flushEx.getMessage(), "unknown error"));
+        }
         saveFileBuildProgress(task, job, totalProgressUnits, 0, "Submitting Coze");
         boolean pendingCoze = submitCozeBatches(task, result, job, chunks, allRowsByBaseId);
         if (pendingCoze) {
@@ -2020,7 +2033,8 @@ public class SimilarAsinTaskService {
                 batchTotal,
                 currentInstanceId(),
                 0,
-                credentialName
+                credentialName,
+                null
         );
         String batchPayload = writeJson(batchRows, "serialize pending coze batch payload failed");
         String storedBatchPayload = transientPayloadStorageService.storeParsedPayloadFast(
@@ -2074,7 +2088,8 @@ public class SimilarAsinTaskService {
                 batchTotal,
                 currentInstanceId(),
                 0,
-                credentialName
+                credentialName,
+                null
         );
         String batchPayload = writeJson(batchRows, "serialize coze batch payload failed");
         String storedBatchPayload = transientPayloadStorageService.storeParsedPayloadFast(
@@ -2192,10 +2207,27 @@ public class SimilarAsinTaskService {
             if (!failureMessage.isBlank()) {
                 cozeRows = cozeClient.markRowsFailed(batchRows, failureMessage);
             }
-            FileTaskEntity task = taskForPoll(state.getTaskId());
-            if (task != null) {
-                Map<String, List<SimilarAsinParsedRowVo>> allRowsByBaseId = allRowsByBaseIdForPoll(task);
-                mergeCozeRowsIntoChunk(task, context.chunkScopeHash(), context.chunkIndex(), cozeRows, allRowsByBaseId);
+            // P0-3：仅"DONE 且 batchTotal>1 且 feature toggle 开启"时缓冲 cozeRows 到 transient storage，
+            // 由 finalize 阶段一次性合并到 chunk。失败 batch / 单 batch 任务保留原立即 merge 路径。
+            boolean buffered = false;
+            if (failureMessage.isBlank()
+                    && isCozeResultBufferEnabled()
+                    && context.batchTotal() != null && context.batchTotal() > 1
+                    && cozeRows != null && !cozeRows.isEmpty()) {
+                CozeBatchContext bufferedContext = bufferCozeResultForFlush(state, context, cozeRows);
+                if (bufferedContext != null
+                        && bufferedContext.resultPayloadPointer() != null
+                        && !bufferedContext.resultPayloadPointer().isBlank()) {
+                    context = bufferedContext;
+                    buffered = true;
+                }
+            }
+            if (!buffered) {
+                FileTaskEntity task = taskForPoll(state.getTaskId());
+                if (task != null) {
+                    Map<String, List<SimilarAsinParsedRowVo>> allRowsByBaseId = allRowsByBaseIdForPoll(task);
+                    mergeCozeRowsIntoChunk(task, context.chunkScopeHash(), context.chunkIndex(), cozeRows, allRowsByBaseId);
+                }
             }
             markCozeStateTerminal(state,
                     failureMessage.isBlank() ? COZE_STATUS_DONE : COZE_STATUS_FAILED,
@@ -2473,7 +2505,8 @@ public class SimilarAsinTaskService {
                 partTotal,
                 parentContext.ownerInstanceId(),
                 retryCount,
-                firstNonBlank(credentialName, parentContext.credentialName())
+                firstNonBlank(credentialName, parentContext.credentialName()),
+                null
         );
         LocalDateTime now = LocalDateTime.now();
         String batchPayload = writeJson(batchRows, "serialize split coze batch payload failed");
@@ -2911,6 +2944,17 @@ public class SimilarAsinTaskService {
                 touchJavaSideTaskActivity(taskId);
                 return;
             }
+            // P0-3：requeue assemble 之前一次性把缓冲的 cozeRows 合并到 chunk。
+            // 失败时 markFailed job 并阻止 requeue，避免 assemble 阶段读到不完整 chunk。
+            try {
+                flushBufferedCozeResults(taskId);
+            } catch (Exception flushEx) {
+                String message = firstNonBlank(flushEx.getMessage(), "flush buffered coze results failed");
+                log.warn("[similar-asin] coze finalize flush failed, marking job failed taskId={} jobId={} err={}",
+                        taskId, job.getId(), message, flushEx);
+                taskFileJobService.markFailed(job, message);
+                return;
+            }
             boolean requeued = taskFileJobService.requeue(job.getId(), "Coze results ready, assembling xlsx");
             if (requeued) {
                 log.info("[similar-asin] coze async results ready, result file job requeued taskId={} jobId={} resultId={}",
@@ -3076,6 +3120,204 @@ public class SimilarAsinTaskService {
         }
     }
 
+    /**
+     * P0-3：读取 state 缓冲的 cozeRows。
+     * 与 readCozeBatchRows（读输入 batchRows）不同，这里读的是
+     * 通过 bufferCozeResultForFlush 写入 transient storage 的 DONE 结果。
+     */
+    private List<SimilarAsinResultRowDto> readBufferedCozeRows(String resultPayloadPointer) {
+        if (resultPayloadPointer == null || resultPayloadPointer.isBlank()) {
+            return List.of();
+        }
+        try {
+            String payloadJson = transientPayloadStorageService.resolvePayload(
+                    resultPayloadPointer, "read similar ASIN coze result buffer failed");
+            if (payloadJson == null || payloadJson.isBlank()) {
+                return List.of();
+            }
+            JsonNode array = objectMapper.readTree(payloadJson);
+            if (!array.isArray()) {
+                return List.of();
+            }
+            List<SimilarAsinResultRowDto> rows = new ArrayList<>();
+            for (JsonNode node : array) {
+                rows.add(objectMapper.treeToValue(node, SimilarAsinResultRowDto.class));
+            }
+            return rows;
+        } catch (Exception ex) {
+            log.warn("[similar-asin] read coze result buffer failed pointer={} err={}",
+                    resultPayloadPointer, ex.getMessage());
+            return List.of();
+        }
+    }
+
+    /**
+     * P0-3：把单个 DONE batch 的 cozeRows 缓冲到 transient storage（不立即写 chunk）。
+     * 返回更新后的 CozeBatchContext（含 resultPayloadPointer），调用方需写回 stateJson。
+     * 缓冲失败时返回原 context，调用方应回退到立即 merge 路径。
+     */
+    private CozeBatchContext bufferCozeResultForFlush(TaskScopeStateEntity state,
+                                                     CozeBatchContext context,
+                                                     List<SimilarAsinResultRowDto> cozeRows) {
+        if (state == null || context == null || cozeRows == null || cozeRows.isEmpty()) {
+            return context;
+        }
+        try {
+            String payloadJson = writeJson(cozeRows, "serialize coze result buffer failed");
+            String pointer = transientPayloadStorageService.storeParsedPayloadEntry(
+                    MODULE_TYPE, state.getTaskId(), state.getScopeHash(),
+                    "coze-result-" + state.getId(), payloadJson, true);
+            if (pointer == null || pointer.isBlank()) {
+                return context;
+            }
+            CozeBatchContext bufferedContext = withResultPayloadPointer(context, pointer);
+            String updatedStateJson = writeJson(bufferedContext, "serialize coze result buffer context failed");
+            int updated = taskScopeStateMapper.update(null, new LambdaUpdateWrapper<TaskScopeStateEntity>()
+                    .eq(TaskScopeStateEntity::getId, state.getId())
+                    .set(TaskScopeStateEntity::getStateJson, updatedStateJson)
+                    .set(TaskScopeStateEntity::getUpdatedAt, LocalDateTime.now()));
+            if (updated <= 0) {
+                transientPayloadStorageService.deletePayloadIfPresent(pointer);
+                return context;
+            }
+            state.setStateJson(updatedStateJson);
+            return bufferedContext;
+        } catch (Exception ex) {
+            log.warn("[similar-asin] buffer coze result failed taskId={} stateId={} err={}",
+                    state.getTaskId(), state.getId(), ex.getMessage());
+            return context;
+        }
+    }
+
+    /**
+     * P0-3：清掉 state 上缓冲的 pointer + 删除 transient payload。
+     * 在 flush 成功合并到 chunk 后调用。
+     */
+    private void clearBufferedCozeResult(TaskScopeStateEntity state, CozeBatchContext context) {
+        if (state == null || context == null) {
+            return;
+        }
+        String pointer = context.resultPayloadPointer();
+        if (pointer == null || pointer.isBlank()) {
+            return;
+        }
+        try {
+            CozeBatchContext clearedContext = withResultPayloadPointer(context, null);
+            String updatedStateJson = writeJson(clearedContext, "serialize coze result buffer cleared context failed");
+            taskScopeStateMapper.update(null, new LambdaUpdateWrapper<TaskScopeStateEntity>()
+                    .eq(TaskScopeStateEntity::getId, state.getId())
+                    .set(TaskScopeStateEntity::getStateJson, updatedStateJson)
+                    .set(TaskScopeStateEntity::getUpdatedAt, LocalDateTime.now()));
+            state.setStateJson(updatedStateJson);
+            transientPayloadStorageService.deletePayloadIfPresent(pointer);
+        } catch (Exception ex) {
+            log.warn("[similar-asin] clear coze result buffer failed taskId={} stateId={} err={}",
+                    state.getTaskId(), state.getId(), ex.getMessage());
+        }
+    }
+
+    private boolean isCozeResultBufferEnabled() {
+        return properties.isCozeResultBufferEnabled();
+    }
+
+    /**
+     * P0-3：在 finalize 之前一次性把所有 DONE/FAILED state 上缓冲的 cozeRows
+     * 按 chunkScopeHash 分组合并到 chunk。把每个 batch 的"loadSubmittedChunks +
+     * readChunkRows × N + writeChunkPayload × M"压缩成"读 chunk 一次 +
+     * 写 chunk 一次"。
+     */
+    private void flushBufferedCozeResults(Long taskId) {
+        if (taskId == null || taskId <= 0) {
+            return;
+        }
+        List<TaskScopeStateEntity> states = taskScopeStateMapper.selectList(new LambdaQueryWrapper<TaskScopeStateEntity>()
+                .eq(TaskScopeStateEntity::getTaskId, taskId)
+                .eq(TaskScopeStateEntity::getModuleType, MODULE_TYPE)
+                .in(TaskScopeStateEntity::getCozeStatus, List.of(COZE_STATUS_DONE, COZE_STATUS_FAILED)));
+        if (states == null || states.isEmpty()) {
+            return;
+        }
+        List<BufferedFlushEntry> entries = new ArrayList<>();
+        for (TaskScopeStateEntity state : states) {
+            CozeBatchContext context = readCozeBatchContext(state);
+            if (context == null) {
+                continue;
+            }
+            String pointer = context.resultPayloadPointer();
+            if (pointer == null || pointer.isBlank()) {
+                continue;
+            }
+            List<SimilarAsinResultRowDto> rows = readBufferedCozeRows(pointer);
+            if (rows == null || rows.isEmpty()) {
+                // pointer 存在但 payload 已被清理（被 GC 或上一次 flush 部分成功），直接清 pointer。
+                clearBufferedCozeResult(state, context);
+                continue;
+            }
+            entries.add(new BufferedFlushEntry(state, context, rows));
+        }
+        if (entries.isEmpty()) {
+            return;
+        }
+        FileTaskEntity task = fileTaskMapper.selectById(taskId);
+        if (task == null || !MODULE_TYPE.equals(task.getModuleType())) {
+            return;
+        }
+        Map<String, List<SimilarAsinParsedRowVo>> allRowsByBaseId = loadAllRowsByBaseId(task);
+        // 按 chunkScopeHash 分组（空字符串占位 null，让 mergeCozeRowsIntoChunk 走全 chunk 扫描匹配）
+        Map<String, List<BufferedFlushEntry>> entriesByChunk = new LinkedHashMap<>();
+        for (BufferedFlushEntry entry : entries) {
+            String key = entry.context().chunkScopeHash() == null ? "" : entry.context().chunkScopeHash();
+            entriesByChunk.computeIfAbsent(key, ignored -> new ArrayList<>()).add(entry);
+        }
+        int totalFlushed = 0;
+        int failedGroups = 0;
+        for (Map.Entry<String, List<BufferedFlushEntry>> group : entriesByChunk.entrySet()) {
+            List<BufferedFlushEntry> groupEntries = group.getValue();
+            List<SimilarAsinResultRowDto> aggregated = new ArrayList<>();
+            for (BufferedFlushEntry entry : groupEntries) {
+                aggregated.addAll(entry.rows());
+            }
+            if (aggregated.isEmpty()) {
+                continue;
+            }
+            String chunkScopeHash = group.getKey().isEmpty() ? null : group.getKey();
+            Integer chunkIndex = null;
+            if (chunkScopeHash != null) {
+                for (BufferedFlushEntry entry : groupEntries) {
+                    if (entry.context().chunkIndex() != null) {
+                        chunkIndex = entry.context().chunkIndex();
+                        break;
+                    }
+                }
+            }
+            try {
+                mergeCozeRowsIntoChunk(task, chunkScopeHash, chunkIndex, aggregated, allRowsByBaseId);
+                for (BufferedFlushEntry entry : groupEntries) {
+                    clearBufferedCozeResult(entry.state(), entry.context());
+                }
+                totalFlushed += aggregated.size();
+            } catch (Exception ex) {
+                failedGroups++;
+                log.warn("[similar-asin] flush buffered coze results group failed taskId={} chunkScopeHash={} err={}",
+                        taskId, chunkScopeHash, ex.getMessage(), ex);
+                // 不 clear pointer，留待下次 finalize 或 stale-recovery 重试
+            }
+        }
+        if (totalFlushed > 0 || failedGroups > 0) {
+            log.info("[similar-asin] flushed buffered coze results taskId={} entries={} totalRows={} chunkGroups={} failedGroups={}",
+                    taskId, entries.size(), totalFlushed, entriesByChunk.size(), failedGroups);
+        }
+        if (failedGroups > 0) {
+            // 让 finalize 能感知到失败，调用方应避免继续 requeue assemble。
+            throw new IllegalStateException("flush buffered coze results failed groups=" + failedGroups);
+        }
+    }
+
+    private record BufferedFlushEntry(TaskScopeStateEntity state,
+                                      CozeBatchContext context,
+                                      List<SimilarAsinResultRowDto> rows) {
+    }
+
     private CozeBatchContext readCozeBatchContext(TaskScopeStateEntity state) {
         if (state == null || state.getStateJson() == null || state.getStateJson().isBlank()) {
             return null;
@@ -3103,7 +3345,23 @@ public class SimilarAsinTaskService {
                 context.batchTotal(),
                 context.ownerInstanceId(),
                 submitRetryCount,
-                context.credentialName()
+                context.credentialName(),
+                context.resultPayloadPointer()
+        );
+    }
+
+    private CozeBatchContext withResultPayloadPointer(CozeBatchContext context, String resultPayloadPointer) {
+        return new CozeBatchContext(
+                context.jobId(),
+                context.resultId(),
+                context.chunkScopeHash(),
+                context.chunkIndex(),
+                context.batchIndex(),
+                context.batchTotal(),
+                context.ownerInstanceId(),
+                context.submitRetryCount(),
+                context.credentialName(),
+                resultPayloadPointer
         );
     }
 
@@ -3361,11 +3619,13 @@ public class SimilarAsinTaskService {
     private void assembleResultWorkbook(FileTaskEntity task, FileResultEntity result) {
         SimilarAsinParsedPayloadDto parsed = readParsedPayload(task);
         Map<String, SimilarAsinResultRowDto> resultMap = loadPersistedResultRowsWithRetry(task.getId(), parsed.getAllItems().size());
+        int persistedResultRows = resultMap.size();
+        resultMap.entrySet().removeIf(entry -> !isExportableResultRow(entry.getValue()));
         long resolvedRows = parsed.getAllItems().stream()
                 .filter(row -> findResultRow(row, resultMap) != null)
                 .count();
-        log.info("[similar-asin] assemble workbook taskId={} parsedRows={} resultRows={} resolvedRows={}",
-                task.getId(), parsed.getAllItems().size(), resultMap.size(), resolvedRows);
+        log.info("[similar-asin] assemble workbook taskId={} parsedRows={} persistedRows={} resultRows={} resolvedRows={}",
+                task.getId(), parsed.getAllItems().size(), persistedResultRows, resultMap.size(), resolvedRows);
         if (!parsed.getAllItems().isEmpty() && resultMap.isEmpty()) {
             throw new BusinessException("相似ASIN检测结果为空，请稍后重试生成结果文件");
         }
@@ -3645,6 +3905,9 @@ public class SimilarAsinTaskService {
     private void writeResultWorkbook(File xlsx,
                                      List<SimilarAsinParsedRowVo> rowsToWrite,
                                      Map<String, SimilarAsinResultRowDto> resultMap) {
+        // Excel 365 "Place in Cell" 图片：写完 workbook 后 patch richData 单元格图片结构。
+        // 这不是浮动 Drawing，因此点击图片区域会选中单元格，图片不能被拖到任意位置，也不需要工作表保护。
+        ExcelCellImageWriter.Session excelCellImageSession = ExcelCellImageWriter.createSession();
         try (SXSSFWorkbook workbook = new SXSSFWorkbook(200); FileOutputStream fos = new FileOutputStream(xlsx)) {
             Sheet sheet = workbook.createSheet("相似asin检测");
             CellStyle headerStyle = workbook.createCellStyle();
@@ -3652,7 +3915,6 @@ public class SimilarAsinTaskService {
             font.setBold(true);
             headerStyle.setFont(font);
 
-            Drawing<?> patriarch = sheet.createDrawingPatriarch();
             Map<String, SimilarAsinImageEmbedder.ResizedImage> taskImageCache = new ConcurrentHashMap<>();
 
             sheet.setColumnWidth(IMG_COL_MAIN, SimilarAsinImageEmbedder.IMAGE_COL_WIDTH_CHARS * 256);
@@ -3668,7 +3930,7 @@ public class SimilarAsinTaskService {
             }
 
             // P2-8：第一遍预扫所有图片 URL，并行下载到 taskImageCache，
-            // POI 写入仍单线程串行 embed()（cache 命中直接 resize+addPicture），下载/写入解耦。
+            // POI 写入仍单线程串行注册 cell image（cache 命中直接 resize + register），下载/写入解耦。
             List<String> prefetchUrls = collectImageUrlsForPrefetch(rowsToWrite, resultMap);
             if (!prefetchUrls.isEmpty()) {
                 long prefetchStart = System.currentTimeMillis();
@@ -3692,13 +3954,14 @@ public class SimilarAsinTaskService {
                 row.createCell(col++).setCellValue(resultRow == null ? "" : userFacingCozeCellValue(resultRow, resultRow.getCategory()));
                 row.createCell(col).setCellValue(resultRow == null ? "" : firstNonBlank(resultRow.getStatus(), ""));
                 if (resultRow != null) {
-                    // 行高在 embed 前先固定到 IMAGE_ROW_HEIGHT_POINTS（Excel 上限 409pt）：
-                    // embed() 用 IMAGE_COL_WIDTH_CHARS / IMAGE_ROW_HEIGHT_POINTS 常量推导 EMU anchor，与运行时单元格尺寸解耦，
-                    // 顺序前置只是为了让 sheet 实际行高与 anchor 假设一致，避免后续维护误读。
+                    // 行高固定到 IMAGE_ROW_HEIGHT_POINTS（Excel 上限 409pt）：
                     row.setHeightInPoints(SimilarAsinImageEmbedder.IMAGE_ROW_HEIGHT_POINTS);
-                    imageEmbedder.embed(sheet, patriarch, workbook, rowIndex, IMG_COL_MAIN, resultRow.getMainUrl(), row, taskImageCache, false);
-                    imageEmbedder.embed(sheet, patriarch, workbook, rowIndex, IMG_COL_PUZZLE1, resultRow.getPuzzleImg1(), row, taskImageCache, true);
-                    imageEmbedder.embed(sheet, patriarch, workbook, rowIndex, IMG_COL_PUZZLE2, resultRow.getPuzzleImg2(), row, taskImageCache, true);
+                    imageEmbedder.embedAsExcelCellImage(rowIndex, IMG_COL_MAIN,
+                            resultRow.getMainUrl(), row, taskImageCache, excelCellImageSession);
+                    imageEmbedder.embedAsExcelCellImage(rowIndex, IMG_COL_PUZZLE1,
+                            resultRow.getPuzzleImg1(), row, taskImageCache, excelCellImageSession);
+                    imageEmbedder.embedAsExcelCellImage(rowIndex, IMG_COL_PUZZLE2,
+                            resultRow.getPuzzleImg2(), row, taskImageCache, excelCellImageSession);
                 }
                 rowIndex++;
             }
@@ -3707,6 +3970,16 @@ public class SimilarAsinTaskService {
         } catch (Exception ex) {
             throw new BusinessException("生成相似ASIN检测结果失败", ex);
         }
+        if (!excelCellImageSession.isEmpty()) {
+            try {
+                ExcelCellImageWriter.patchXlsxFile(xlsx, excelCellImageSession);
+            } catch (IOException ex) {
+                String cause = ex.toString();
+                log.error("[similar-asin] excel cell image patch failed xlsx={} images={} err={}",
+                        xlsx.getAbsolutePath(), excelCellImageSession.size(), cause, ex);
+                throw new BusinessException("写入相似ASIN结果 Excel 单元格图片失败: " + cause, ex);
+            }
+        }
     }
 
     /**
@@ -3854,7 +4127,9 @@ public class SimilarAsinTaskService {
     }
 
     private boolean shouldInspectRow(SimilarAsinResultRowDto row) {
-        return row != null && (!normalize(row.getTitle()).isBlank() || row.hasImageUrl());
+        return row != null && (!rowKey(row).isBlank()
+                || !normalize(row.getTitle()).isBlank()
+                || row.hasImageUrl());
     }
 
     private boolean hasResolvedCozeFields(SimilarAsinResultRowDto row) {
@@ -3873,6 +4148,22 @@ public class SimilarAsinTaskService {
                 || hasUsableCozeField(row.getStatus());
     }
 
+    private boolean isExportableResultRow(SimilarAsinResultRowDto row) {
+        if (row == null) {
+            return false;
+        }
+        if (hasResolvedCozeFields(row)) {
+            return true;
+        }
+        String error = normalize(row.getError());
+        if (error.isBlank()) {
+            return false;
+        }
+        return !normalize(row.getTitle()).isBlank()
+                || !normalize(row.getUrl()).isBlank()
+                || row.hasImageUrl();
+    }
+
     private boolean hasUsableCozeField(String value) {
         String normalized = normalize(value);
         return !normalized.isBlank() && !isTechnicalCozeFailure(normalized);
@@ -4590,7 +4881,8 @@ public class SimilarAsinTaskService {
                                      Integer batchTotal,
                                      String ownerInstanceId,
                                      Integer submitRetryCount,
-                                     String credentialName) {
+                                     String credentialName,
+                                     String resultPayloadPointer) {
     }
 
     private static class SourceRowsBuilder {

backend-java/src/main/java/com/nanri/aiimage/modules/similarasin/util/ExcelCellImageWriter.java

@@ -0,0 +1,430 @@
+package com.nanri.aiimage.modules.similarasin.util;
+
+import org.slf4j.Logger;
+import org.slf4j.LoggerFactory;
+
+import java.io.ByteArrayOutputStream;
+import java.io.File;
+import java.io.IOException;
+import java.io.InputStream;
+import java.nio.charset.StandardCharsets;
+import java.nio.file.Files;
+import java.nio.file.Path;
+import java.nio.file.StandardCopyOption;
+import java.util.ArrayList;
+import java.util.Enumeration;
+import java.util.HashSet;
+import java.util.List;
+import java.util.Set;
+import java.util.regex.Matcher;
+import java.util.regex.Pattern;
+import java.util.zip.ZipEntry;
+import java.util.zip.ZipFile;
+import java.util.zip.ZipOutputStream;
+
+/**
+ * Excel 365 "Place in Cell" 图片写入器。
+ * <p>
+ * 该格式不是传统 drawing 浮动图片，而是 richData 单元格值：点击图片区域选中的是单元格，
+ * 图片无法像浮动对象一样拖到任意位置，也不需要开启工作表保护。
+ */
+public final class ExcelCellImageWriter {
+
+    private static final Logger log = LoggerFactory.getLogger(ExcelCellImageWriter.class);
+
+    private static final String CONTENT_TYPES_PART = "[Content_Types].xml";
+    private static final String WORKBOOK_RELS_PART = "xl/_rels/workbook.xml.rels";
+    private static final String SHEET1_PART = "xl/worksheets/sheet1.xml";
+    private static final String RICH_VALUE_RELS_PART = "xl/richData/_rels/richValueRel.xml.rels";
+    private static final String RD_RICH_VALUE_PART = "xl/richData/rdrichvalue.xml";
+    private static final String RD_RICH_VALUE_STRUCTURE_PART = "xl/richData/rdrichvaluestructure.xml";
+    private static final String RD_RICH_VALUE_TYPES_PART = "xl/richData/rdRichValueTypes.xml";
+    private static final String RICH_VALUE_REL_PART = "xl/richData/richValueRel.xml";
+    private static final String METADATA_PART = "xl/metadata.xml";
+
+    private static final String REL_TYPE_IMAGE =
+            "http://schemas.openxmlformats.org/officeDocument/2006/relationships/image";
+    private static final String REL_TYPE_RICH_VALUE_STRUCTURE =
+            "http://schemas.microsoft.com/office/2017/06/relationships/rdRichValueStructure";
+    private static final String REL_TYPE_RICH_VALUE =
+            "http://schemas.microsoft.com/office/2017/06/relationships/rdRichValue";
+    private static final String REL_TYPE_RICH_VALUE_REL =
+            "http://schemas.microsoft.com/office/2022/10/relationships/richValueRel";
+    private static final String REL_TYPE_SHEET_METADATA =
+            "http://schemas.openxmlformats.org/officeDocument/2006/relationships/sheetMetadata";
+    private static final String REL_TYPE_RICH_VALUE_TYPES =
+            "http://schemas.microsoft.com/office/2017/06/relationships/rdRichValueTypes";
+
+    private ExcelCellImageWriter() {
+    }
+
+    public static Session createSession() {
+        return new Session();
+    }
+
+    public static void patchXlsxFile(File xlsx, Session session) throws IOException {
+        if (xlsx == null || !xlsx.exists() || session == null || session.isEmpty()) {
+            return;
+        }
+        Path xlsxPath = xlsx.toPath().toAbsolutePath();
+        Path parent = xlsxPath.getParent();
+        if (parent == null) {
+            throw new IOException("[excel-cell-image] xlsx has no parent directory: " + xlsxPath);
+        }
+        Path tmp = Files.createTempFile(parent, "excel-cell-image-", ".xlsx");
+        boolean success = false;
+        Set<String> replacementNames = replacementEntryNames(session);
+        int patchedCells = -1;
+        try {
+            try (ZipFile zfIn = new ZipFile(xlsxPath.toFile());
+                 ZipOutputStream zout = new ZipOutputStream(Files.newOutputStream(tmp))) {
+                Enumeration<? extends ZipEntry> entries = zfIn.entries();
+                while (entries.hasMoreElements()) {
+                    ZipEntry entry = entries.nextElement();
+                    String name = entry.getName();
+                    if (replacementNames.contains(name)) {
+                        continue;
+                    }
+                    if (entry.isDirectory()) {
+                        writeDirectoryEntry(zout, name);
+                        continue;
+                    }
+                    byte[] bytes;
+                    try (InputStream in = zfIn.getInputStream(entry)) {
+                        bytes = readAll(in);
+                    }
+                    if (CONTENT_TYPES_PART.equals(name)) {
+                        bytes = patchContentTypes(bytes);
+                    } else if (WORKBOOK_RELS_PART.equals(name)) {
+                        bytes = patchWorkbookRels(bytes);
+                    } else if (SHEET1_PART.equals(name)) {
+                        PatchSheetResult sheetResult = patchSheet(bytes, session);
+                        bytes = sheetResult.bytes();
+                        patchedCells = sheetResult.patchedCount();
+                        if (sheetResult.patchedCount() != session.size()) {
+                            log.warn("[excel-cell-image] patched cells mismatch file={} images={} patchedCells={}",
+                                    xlsx.getName(), session.size(), sheetResult.patchedCount());
+                        }
+                    }
+                    writeEntry(zout, name, bytes);
+                }
+                writeEntry(zout, RICH_VALUE_RELS_PART, buildRichValueRels(session).getBytes(StandardCharsets.UTF_8));
+                writeEntry(zout, RD_RICH_VALUE_PART, buildRdRichValue(session).getBytes(StandardCharsets.UTF_8));
+                writeEntry(zout, RD_RICH_VALUE_STRUCTURE_PART, buildRdRichValueStructure().getBytes(StandardCharsets.UTF_8));
+                writeEntry(zout, RD_RICH_VALUE_TYPES_PART, buildRdRichValueTypes().getBytes(StandardCharsets.UTF_8));
+                writeEntry(zout, RICH_VALUE_REL_PART, buildRichValueRel(session).getBytes(StandardCharsets.UTF_8));
+                writeEntry(zout, METADATA_PART, buildMetadata(session).getBytes(StandardCharsets.UTF_8));
+                int idx = 1;
+                for (RegisteredCellImage image : session.images) {
+                    writeEntry(zout, mediaPart(idx), image.bytes);
+                    idx++;
+                }
+            }
+            try {
+                Files.move(tmp, xlsxPath, StandardCopyOption.REPLACE_EXISTING);
+            } catch (IOException moveEx) {
+                try {
+                    Thread.sleep(200L);
+                } catch (InterruptedException ignored) {
+                    Thread.currentThread().interrupt();
+                }
+                Files.move(tmp, xlsxPath, StandardCopyOption.REPLACE_EXISTING);
+                log.warn("[excel-cell-image] patch xlsx move retried after error file={} cause={}",
+                        xlsx.getName(), moveEx.toString());
+            }
+            success = true;
+            log.info("[excel-cell-image] patch xlsx finished file={} images={} patchedCells={}",
+                    xlsx.getName(), session.size(), patchedCells);
+        } catch (IOException ioe) {
+            log.error("[excel-cell-image] patch xlsx failed file={} images={} err={}",
+                    xlsx.getName(), session.size(), ioe.toString(), ioe);
+            throw ioe;
+        } finally {
+            if (!success) {
+                try {
+                    Files.deleteIfExists(tmp);
+                } catch (IOException ignored) {
+                    // 临时文件清理失败不影响主流程
+                }
+            }
+        }
+    }
+
+    private static Set<String> replacementEntryNames(Session session) {
+        Set<String> names = new HashSet<>();
+        names.add(RICH_VALUE_RELS_PART);
+        names.add(RD_RICH_VALUE_PART);
+        names.add(RD_RICH_VALUE_STRUCTURE_PART);
+        names.add(RD_RICH_VALUE_TYPES_PART);
+        names.add(RICH_VALUE_REL_PART);
+        names.add(METADATA_PART);
+        for (int i = 1; i <= session.images.size(); i++) {
+            names.add(mediaPart(i));
+        }
+        return names;
+    }
+
+    private static PatchSheetResult patchSheet(byte[] original, Session session) {
+        String content = new String(original, StandardCharsets.UTF_8);
+        int idx = 1;
+        int patchedCount = 0;
+        for (RegisteredCellImage image : session.images) {
+            String replacement = "<c r=\"" + image.cellRef + "\" t=\"e\" vm=\"" + idx + "\"><v>#VALUE!</v></c>";
+            Pattern pattern = Pattern.compile("<c\\b(?=[^>]*\\br=\"" + Pattern.quote(image.cellRef)
+                    + "\")[^>]*(?:/>|>.*?</c>)", Pattern.DOTALL);
+            Matcher matcher = pattern.matcher(content);
+            if (matcher.find()) {
+                content = matcher.replaceFirst(Matcher.quoteReplacement(replacement));
+                patchedCount++;
+            } else {
+                log.warn("[excel-cell-image] cell {} not found in sheet xml, skip image", image.cellRef);
+            }
+            idx++;
+        }
+        return new PatchSheetResult(content.getBytes(StandardCharsets.UTF_8), patchedCount);
+    }
+
+    private static byte[] patchContentTypes(byte[] original) {
+        String content = new String(original, StandardCharsets.UTF_8);
+        StringBuilder injection = new StringBuilder();
+        if (!content.contains("Extension=\"jpeg\"")) {
+            injection.append("<Default Extension=\"jpeg\" ContentType=\"image/jpeg\"/>");
+        }
+        appendOverrideIfMissing(content, injection, "/xl/metadata.xml",
+                "application/vnd.openxmlformats-officedocument.spreadsheetml.sheetMetadata+xml");
+        appendOverrideIfMissing(content, injection, "/xl/richData/richValueRel.xml",
+                "application/vnd.ms-excel.richvaluerel+xml");
+        appendOverrideIfMissing(content, injection, "/xl/richData/rdrichvalue.xml",
+                "application/vnd.ms-excel.rdrichvalue+xml");
+        appendOverrideIfMissing(content, injection, "/xl/richData/rdrichvaluestructure.xml",
+                "application/vnd.ms-excel.rdrichvaluestructure+xml");
+        appendOverrideIfMissing(content, injection, "/xl/richData/rdRichValueTypes.xml",
+                "application/vnd.ms-excel.rdrichvaluetypes+xml");
+        if (injection.isEmpty()) {
+            return original;
+        }
+        int closeIdx = content.lastIndexOf("</Types>");
+        if (closeIdx < 0) {
+            log.warn("[excel-cell-image] [Content_Types].xml missing </Types>, skip patch");
+            return original;
+        }
+        return (content.substring(0, closeIdx) + injection + content.substring(closeIdx))
+                .getBytes(StandardCharsets.UTF_8);
+    }
+
+    private static void appendOverrideIfMissing(String content, StringBuilder injection, String partName, String contentType) {
+        if (!content.contains("PartName=\"" + partName + "\"")) {
+            injection.append("<Override PartName=\"").append(partName)
+                    .append("\" ContentType=\"").append(contentType).append("\"/>");
+        }
+    }
+
+    private static byte[] patchWorkbookRels(byte[] original) {
+        String content = new String(original, StandardCharsets.UTF_8);
+        int maxId = findMaxRId(content);
+        StringBuilder injection = new StringBuilder();
+        maxId = appendRelationshipIfMissing(content, injection, maxId, REL_TYPE_RICH_VALUE_STRUCTURE,
+                "richData/rdrichvaluestructure.xml");
+        maxId = appendRelationshipIfMissing(content, injection, maxId, REL_TYPE_RICH_VALUE,
+                "richData/rdrichvalue.xml");
+        maxId = appendRelationshipIfMissing(content, injection, maxId, REL_TYPE_RICH_VALUE_REL,
+                "richData/richValueRel.xml");
+        maxId = appendRelationshipIfMissing(content, injection, maxId, REL_TYPE_SHEET_METADATA,
+                "metadata.xml");
+        appendRelationshipIfMissing(content, injection, maxId, REL_TYPE_RICH_VALUE_TYPES,
+                "richData/rdRichValueTypes.xml");
+        if (injection.isEmpty()) {
+            return original;
+        }
+        int closeIdx = content.lastIndexOf("</Relationships>");
+        if (closeIdx < 0) {
+            log.warn("[excel-cell-image] xl/_rels/workbook.xml.rels missing </Relationships>, skip patch");
+            return original;
+        }
+        return (content.substring(0, closeIdx) + injection + content.substring(closeIdx))
+                .getBytes(StandardCharsets.UTF_8);
+    }
+
+    private static int appendRelationshipIfMissing(String content,
+                                                   StringBuilder injection,
+                                                   int maxId,
+                                                   String type,
+                                                   String target) {
+        if (content.contains("Target=\"" + target + "\"")) {
+            return maxId;
+        }
+        int next = maxId + 1;
+        injection.append("<Relationship Id=\"rId").append(next)
+                .append("\" Type=\"").append(type)
+                .append("\" Target=\"").append(target).append("\"/>");
+        return next;
+    }
+
+    private static String buildRichValueRels(Session session) {
+        StringBuilder sb = new StringBuilder();
+        sb.append("<?xml version=\"1.0\" encoding=\"UTF-8\" standalone=\"yes\"?>");
+        sb.append("<Relationships xmlns=\"http://schemas.openxmlformats.org/package/2006/relationships\">");
+        for (int i = 1; i <= session.images.size(); i++) {
+            sb.append("<Relationship Id=\"rId").append(i)
+                    .append("\" Type=\"").append(REL_TYPE_IMAGE)
+                    .append("\" Target=\"../media/excelcellimage").append(i).append(".jpeg\"/>");
+        }
+        sb.append("</Relationships>");
+        return sb.toString();
+    }
+
+    private static String buildRdRichValue(Session session) {
+        StringBuilder sb = new StringBuilder();
+        sb.append("<?xml version=\"1.0\" encoding=\"UTF-8\" standalone=\"yes\"?>");
+        sb.append("<rvData xmlns=\"http://schemas.microsoft.com/office/spreadsheetml/2017/richdata\" count=\"")
+                .append(session.images.size()).append("\">");
+        for (int i = 0; i < session.images.size(); i++) {
+            sb.append("<rv s=\"0\"><v>").append(i).append("</v><v>5</v></rv>");
+        }
+        sb.append("</rvData>");
+        return sb.toString();
+    }
+
+    private static String buildRdRichValueStructure() {
+        return "<?xml version=\"1.0\" encoding=\"UTF-8\" standalone=\"yes\"?>"
+                + "<rvStructures xmlns=\"http://schemas.microsoft.com/office/spreadsheetml/2017/richdata\" count=\"1\">"
+                + "<s t=\"_localImage\"><k n=\"_rvRel:LocalImageIdentifier\" t=\"i\"/><k n=\"CalcOrigin\" t=\"i\"/></s>"
+                + "</rvStructures>";
+    }
+
+    private static String buildRdRichValueTypes() {
+        return "<?xml version=\"1.0\" encoding=\"UTF-8\" standalone=\"yes\"?>"
+                + "<rvTypesInfo xmlns=\"http://schemas.microsoft.com/office/spreadsheetml/2017/richdata2\""
+                + " xmlns:mc=\"http://schemas.openxmlformats.org/markup-compatibility/2006\" mc:Ignorable=\"x\""
+                + " xmlns:x=\"http://schemas.openxmlformats.org/spreadsheetml/2006/main\">"
+                + "<global><keyFlags>"
+                + "<key name=\"_Self\"><flag name=\"ExcludeFromFile\" value=\"1\"/><flag name=\"ExcludeFromCalcComparison\" value=\"1\"/></key>"
+                + "<key name=\"_DisplayString\"><flag name=\"ExcludeFromCalcComparison\" value=\"1\"/></key>"
+                + "<key name=\"_Flags\"><flag name=\"ExcludeFromCalcComparison\" value=\"1\"/></key>"
+                + "<key name=\"_Format\"><flag name=\"ExcludeFromCalcComparison\" value=\"1\"/></key>"
+                + "<key name=\"_SubLabel\"><flag name=\"ExcludeFromCalcComparison\" value=\"1\"/></key>"
+                + "<key name=\"_Attribution\"><flag name=\"ExcludeFromCalcComparison\" value=\"1\"/></key>"
+                + "<key name=\"_Icon\"><flag name=\"ExcludeFromCalcComparison\" value=\"1\"/></key>"
+                + "<key name=\"_Display\"><flag name=\"ExcludeFromCalcComparison\" value=\"1\"/></key>"
+                + "<key name=\"_CanonicalPropertyNames\"><flag name=\"ExcludeFromCalcComparison\" value=\"1\"/></key>"
+                + "<key name=\"_ClassificationId\"><flag name=\"ExcludeFromCalcComparison\" value=\"1\"/></key>"
+                + "</keyFlags></global></rvTypesInfo>";
+    }
+
+    private static String buildRichValueRel(Session session) {
+        StringBuilder sb = new StringBuilder();
+        sb.append("<?xml version=\"1.0\" encoding=\"UTF-8\" standalone=\"yes\"?>");
+        sb.append("<richValueRels xmlns=\"http://schemas.microsoft.com/office/spreadsheetml/2022/richvaluerel\"")
+                .append(" xmlns:r=\"http://schemas.openxmlformats.org/officeDocument/2006/relationships\">");
+        for (int i = 1; i <= session.images.size(); i++) {
+            sb.append("<rel r:id=\"rId").append(i).append("\"/>");
+        }
+        sb.append("</richValueRels>");
+        return sb.toString();
+    }
+
+    private static String buildMetadata(Session session) {
+        StringBuilder sb = new StringBuilder();
+        int count = session.images.size();
+        sb.append("<?xml version=\"1.0\" encoding=\"UTF-8\" standalone=\"yes\"?>");
+        sb.append("<metadata xmlns=\"http://schemas.openxmlformats.org/spreadsheetml/2006/main\"")
+                .append(" xmlns:xlrd=\"http://schemas.microsoft.com/office/spreadsheetml/2017/richdata\">")
+                .append("<metadataTypes count=\"1\">")
+                .append("<metadataType name=\"XLRICHVALUE\" minSupportedVersion=\"120000\" copy=\"1\" pasteAll=\"1\"")
+                .append(" pasteValues=\"1\" merge=\"1\" splitFirst=\"1\" rowColShift=\"1\" clearFormats=\"1\"")
+                .append(" clearComments=\"1\" assign=\"1\" coerce=\"1\"/>")
+                .append("</metadataTypes>")
+                .append("<futureMetadata name=\"XLRICHVALUE\" count=\"").append(count).append("\">");
+        for (int i = 0; i < count; i++) {
+            sb.append("<bk><extLst><ext uri=\"{3e2802c4-a4d2-4d8b-9148-e3be6c30e623}\">")
+                    .append("<xlrd:rvb i=\"").append(i).append("\"/>")
+                    .append("</ext></extLst></bk>");
+        }
+        sb.append("</futureMetadata><valueMetadata count=\"").append(count).append("\">");
+        for (int i = 0; i < count; i++) {
+            sb.append("<bk><rc t=\"1\" v=\"").append(i).append("\"/></bk>");
+        }
+        sb.append("</valueMetadata></metadata>");
+        return sb.toString();
+    }
+
+    private static int findMaxRId(String content) {
+        int maxId = 0;
+        Matcher matcher = Pattern.compile("Id=\"rId(\\d+)\"").matcher(content);
+        while (matcher.find()) {
+            try {
+                maxId = Math.max(maxId, Integer.parseInt(matcher.group(1)));
+            } catch (NumberFormatException ignored) {
+                // 非数字 rId 不影响新增关系
+            }
+        }
+        return maxId;
+    }
+
+    private static String mediaPart(int idx) {
+        return "xl/media/excelcellimage" + idx + ".jpeg";
+    }
+
+    private static byte[] readAll(InputStream in) throws IOException {
+        ByteArrayOutputStream baos = new ByteArrayOutputStream();
+        byte[] buf = new byte[8192];
+        int n;
+        while ((n = in.read(buf)) > 0) {
+            baos.write(buf, 0, n);
+        }
+        return baos.toByteArray();
+    }
+
+    private static void writeEntry(ZipOutputStream zout, String name, byte[] bytes) throws IOException {
+        ZipEntry entry = new ZipEntry(name);
+        zout.putNextEntry(entry);
+        zout.write(bytes);
+        zout.closeEntry();
+    }
+
+    private static void writeDirectoryEntry(ZipOutputStream zout, String name) throws IOException {
+        ZipEntry entry = new ZipEntry(name);
+        zout.putNextEntry(entry);
+        zout.closeEntry();
+    }
+
+    public static final class Session {
+        private final List<RegisteredCellImage> images = new ArrayList<>();
+
+        public synchronized void registerImage(int rowIdx, int colIdx, byte[] jpegBytes) {
+            if (jpegBytes == null || jpegBytes.length == 0) {
+                return;
+            }
+            images.add(new RegisteredCellImage(cellRef(rowIdx, colIdx), jpegBytes));
+        }
+
+        public synchronized boolean isEmpty() {
+            return images.isEmpty();
+        }
+
+        public synchronized int size() {
+            return images.size();
+        }
+
+        private static String cellRef(int rowIdx, int colIdx) {
+            return columnName(colIdx) + (rowIdx + 1);
+        }
+
+        private static String columnName(int colIdx) {
+            int value = colIdx + 1;
+            StringBuilder sb = new StringBuilder();
+            while (value > 0) {
+                int rem = (value - 1) % 26;
+                sb.append((char) ('A' + rem));
+                value = (value - 1) / 26;
+            }
+            return sb.reverse().toString();
+        }
+    }
+
+    private record RegisteredCellImage(String cellRef, byte[] bytes) {
+    }
+
+    private record PatchSheetResult(byte[] bytes, int patchedCount) {
+    }
+}

backend-java/src/main/java/com/nanri/aiimage/modules/similarasin/util/SimilarAsinImageEmbedder.java

@@ -8,13 +8,8 @@ import okhttp3.OkHttpClient;
 import okhttp3.Request;
 import okhttp3.Response;
 import okhttp3.ResponseBody;
-import org.apache.poi.ss.usermodel.ClientAnchor;
-import org.apache.poi.ss.usermodel.Drawing;
+import org.apache.poi.ss.usermodel.Cell;
 import org.apache.poi.ss.usermodel.Row;
-import org.apache.poi.ss.usermodel.Sheet;
-import org.apache.poi.ss.usermodel.Workbook;
-import org.apache.poi.util.Units;
-import org.apache.poi.xssf.streaming.SXSSFWorkbook;
 import org.springframework.stereotype.Component;
 
 import javax.imageio.IIOImage;
@@ -173,25 +168,15 @@ public class SimilarAsinImageEmbedder {
     }
 
     /**
-     * 嵌入单元格：下载 → resize → POI addPicture → createPicture，全链路任一失败即 URL 文本兜底。
-     * 返回值：成功时为缩略图实际像素（width/height），失败兜底为 null；调用方据此自适应行高/列宽。
+     * 抽取共享：下载 + resize + 写 cache。任一失败即写文本兜底（cell 走 row.createCell(colIdx).setCellValue），
+     * 失败时返回 null；成功返回缩略图。给 Excel richData 单元格图片写入路径使用。
      */
-    public ImageDim embed(Sheet sheet,
-                          Drawing<?> patriarch,
-                          SXSSFWorkbook wb,
-                          int rowIdx,
-                          int colIdx,
-                          String url,
-                          Row row,
-                          Map<String, ResizedImage> taskImageCache,
-                          boolean fillCell) {
-        if (url == null || url.isBlank()) {
-            return null;
-        }
-        String trimmedUrl = url.trim();
-        ResizedImage thumb;
+    private ResizedImage downloadAndResize(String trimmedUrl,
+                                           int colIdx,
+                                           Row row,
+                                           Map<String, ResizedImage> taskImageCache) {
         try {
-            thumb = taskImageCache.get(trimmedUrl);
+            ResizedImage thumb = taskImageCache.get(trimmedUrl);
             boolean cacheHit = thumb != null;
             if (!cacheHit) {
                 byte[] raw = downloadWithRetry(trimmedUrl);
@@ -203,73 +188,46 @@ public class SimilarAsinImageEmbedder {
                 log.debug("[similar-asin][image] cache-hit url={} bytes={} dims={}x{}",
                         trimmedUrl, thumb.bytes().length, thumb.width(), thumb.height());
             }
+            return thumb;
         } catch (UnsupportedUrlException ex) {
             log.warn("[similar-asin][image] unsupported-url url={} reason={}", trimmedUrl, ex.getMessage());
-            row.createCell(colIdx).setCellValue(trimmedUrl);
-            return null;
         } catch (DownloadOversizeException ex) {
             log.warn("[similar-asin][image] download-oversize url={} bytes={}", trimmedUrl, ex.size());
-            row.createCell(colIdx).setCellValue(trimmedUrl);
-            return null;
         } catch (ResizeOversizeException ex) {
             log.warn("[similar-asin][image] resize-oversize url={} bytes={}", ex.url(), ex.size());
-            row.createCell(colIdx).setCellValue(trimmedUrl);
-            return null;
         } catch (ResizeException ex) {
             log.warn("[similar-asin][image] resize-fail url={} err={}", trimmedUrl, ex.getMessage());
-            row.createCell(colIdx).setCellValue(trimmedUrl);
-            return null;
         } catch (IOException | TimeoutException ex) {
             log.warn("[similar-asin][image] download-fail url={} err={}", trimmedUrl, ex.getMessage());
-            row.createCell(colIdx).setCellValue(trimmedUrl);
-            return null;
         } catch (RuntimeException ex) {
             log.warn("[similar-asin][image] embed-fail url={} err={}", trimmedUrl, ex.getMessage());
-            row.createCell(colIdx).setCellValue(trimmedUrl);
+        }
+        row.createCell(colIdx).setCellValue(trimmedUrl);
+        return null;
+    }
+
+    public ImageDim embedAsExcelCellImage(int rowIdx,
+                                          int colIdx,
+                                          String url,
+                                          Row row,
+                                          Map<String, ResizedImage> taskImageCache,
+                                          ExcelCellImageWriter.Session excelCellImageSession) {
+        if (url == null || url.isBlank() || excelCellImageSession == null) {
+            return null;
+        }
+        String trimmedUrl = url.trim();
+        ResizedImage thumb = downloadAndResize(trimmedUrl, colIdx, row, taskImageCache);
+        if (thumb == null) {
             return null;
         }
         try {
-            int picIdx = wb.addPicture(thumb.bytes(), Workbook.PICTURE_TYPE_JPEG);
-            ClientAnchor anchor = wb.getCreationHelper().createClientAnchor();
-            anchor.setAnchorType(ClientAnchor.AnchorType.MOVE_DONT_RESIZE);
-            anchor.setCol1(colIdx);
-            anchor.setRow1(rowIdx);
-            anchor.setCol2(colIdx);
-            anchor.setRow2(rowIdx);
-            // 显式 EMU anchor：col1==col2 / row1==row2，仅靠 dx/dy 居中等比缩放。
-            // picture.resize() 会按"调用时刻"的实际行高/列宽推算 col2/row2，
-            // 但 service 端可能在 embed 之后才 setHeightInPoints(409)，导致 anchor 跨多行被纵向拉伸。
-            // 这里彻底脱离运行时单元格尺寸依赖，单元格大小通过 IMAGE_COL_WIDTH_CHARS / IMAGE_ROW_HEIGHT_POINTS 常量推导。
-            double cellWidthEmu = IMAGE_COL_WIDTH_CHARS * 7.0 * Units.EMU_PER_PIXEL;
-            double cellHeightEmu = IMAGE_ROW_HEIGHT_POINTS * (double) Units.EMU_PER_POINT;
-            double picWidthEmu = thumb.width() * (double) Units.EMU_PER_PIXEL;
-            double picHeightEmu = thumb.height() * (double) Units.EMU_PER_PIXEL;
-            if (fillCell) {
-                anchor.setDx1(0);
-                anchor.setDy1(0);
-                anchor.setDx2((int) Math.round(cellWidthEmu));
-                anchor.setDy2((int) Math.round(cellHeightEmu));
-            } else {
-                double scale = Math.min(cellWidthEmu / picWidthEmu, cellHeightEmu / picHeightEmu) * 0.98;
-                if (scale > 1.0) {
-                    scale = 1.0;
-                }
-                double displayedW = picWidthEmu * scale;
-                double displayedH = picHeightEmu * scale;
-                int dx1 = (int) Math.round((cellWidthEmu - displayedW) / 2.0);
-                int dy1 = (int) Math.round((cellHeightEmu - displayedH) / 2.0);
-                anchor.setDx1(dx1);
-                anchor.setDy1(dy1);
-                anchor.setDx2((int) Math.round(dx1 + displayedW));
-                anchor.setDy2((int) Math.round(dy1 + displayedH));
-            }
-            patriarch.createPicture(anchor, picIdx);
-            ImageDim dim = new ImageDim(thumb.width(), thumb.height());
-            // POI 已经把缩略图字节复制进 picture pool（B 副本），此时移除 taskImageCache 里的 A 副本可立刻释放。
+            Cell cell = row.createCell(colIdx);
+            cell.setCellValue("#VALUE!");
+            excelCellImageSession.registerImage(rowIdx, colIdx, thumb.bytes());
             taskImageCache.remove(trimmedUrl);
-            return dim;
+            return new ImageDim(thumb.width(), thumb.height());
         } catch (RuntimeException ex) {
-            log.warn("[similar-asin][image] embed-fail url={} err={}", trimmedUrl, ex.getMessage());
+            log.warn("[similar-asin][image] excel-cell-image-fail url={} err={}", trimmedUrl, ex.getMessage());
             row.createCell(colIdx).setCellValue(trimmedUrl);
             return null;
         }

backend-java/src/main/java/com/nanri/aiimage/modules/similarasin/util/WpsCellImageWriter.java

@@ -0,0 +1,356 @@
+package com.nanri.aiimage.modules.similarasin.util;
+
+import lombok.extern.slf4j.Slf4j;
+
+import java.io.ByteArrayOutputStream;
+import java.io.File;
+import java.io.IOException;
+import java.io.InputStream;
+import java.nio.charset.StandardCharsets;
+import java.nio.file.Files;
+import java.nio.file.Path;
+import java.nio.file.StandardCopyOption;
+import java.security.MessageDigest;
+import java.security.NoSuchAlgorithmException;
+import java.util.ArrayList;
+import java.util.Arrays;
+import java.util.Enumeration;
+import java.util.LinkedHashMap;
+import java.util.List;
+import java.util.Map;
+import java.util.UUID;
+import java.util.regex.Matcher;
+import java.util.regex.Pattern;
+import java.util.zip.ZipEntry;
+import java.util.zip.ZipFile;
+import java.util.zip.ZipOutputStream;
+
+/**
+ * WPS DISPIMG "图片嵌入单元格" 协议写入器。
+ * <p>
+ * Excel 1997-2007 OOXML 没有 "图片嵌入单元格" 概念，POI 5.2.5 也不支持写入 Excel 365 的 IMAGE() 函数。
+ * 本类基于 WPS 在 OOXML 上的扩展协议（cellimages.xml），让生成的 xlsx 在 WPS、钉钉表格、
+ * 飞书表格、Excel 365 中表现为 "图片真嵌入到单元格内"：解除工作表保护后图片仍跟着单元格走，
+ * 拷贝粘贴行/排序时图片随单元格移动，无法被鼠标拖出。
+ * <p>
+ * 用法：
+ * <pre>
+ *     WpsCellImageWriter.Session session = WpsCellImageWriter.createSession();
+ *     // 写每个 cell：
+ *     String id = session.registerImage(jpegBytes);
+ *     cell.setCellFormula(WpsCellImageWriter.dispImgFormula(id));
+ *     // workbook.write(file) 之后调用：
+ *     WpsCellImageWriter.patchXlsxFile(xlsx, session);
+ * </pre>
+ * <p>
+ * 协议参考：WPS Office 表格扩展 cellimages.xml。
+ */
+@Slf4j
+public final class WpsCellImageWriter {
+
+    public static final String CELLIMAGES_PART = "xl/cellimages.xml";
+    public static final String CELLIMAGES_RELS_PART = "xl/_rels/cellimages.xml.rels";
+    public static final String CONTENT_TYPES_PART = "[Content_Types].xml";
+    public static final String WORKBOOK_RELS_PART = "xl/_rels/workbook.xml.rels";
+    public static final String CELLIMAGES_CONTENT_TYPE =
+            "application/vnd.openxmlformats-officedocument.spreadsheetml.cellimages+xml";
+    public static final String CELLIMAGES_REL_TYPE =
+            "http://www.wps.cn/officeDocument/2018/relationships/cellimage";
+    public static final String IMAGE_REL_TYPE =
+            "http://schemas.openxmlformats.org/officeDocument/2006/relationships/image";
+
+    private WpsCellImageWriter() {
+    }
+
+    public static Session createSession() {
+        return new Session();
+    }
+
+    /**
+     * 构造 DISPIMG 公式。
+     * <p>
+     * 第二个参数固定 1：图片自动缩放到单元格大小（与 IMAGE_ROW_HEIGHT_POINTS / IMAGE_COL_WIDTH_CHARS 配合，
+     * 既不会拉伸变形也能填充单元格）。
+     */
+    public static String dispImgFormula(String id) {
+        if (id == null || id.isBlank()) {
+            return null;
+        }
+        return "_xlfn.DISPIMG(\"" + id + "\",1)";
+    }
+
+    /**
+     * 在已写出的 xlsx 文件上 patch DISPIMG 协议所需的 4 个 part：
+     * - 新建 xl/cellimages.xml
+     * - 新建 xl/_rels/cellimages.xml.rels
+     * - 新建 xl/media/cellimage{N}.jpeg
+     * - 修改 [Content_Types].xml（加 Override）
+     * - 修改 xl/_rels/workbook.xml.rels（加 Relationship）
+     * <p>
+     * 通过临时文件 + REPLACE_EXISTING move 实现：失败时 xlsx 文件保持原样不被破坏。
+     * 历史使用 ATOMIC_MOVE+REPLACE_EXISTING，但 Java NIO 规范说 ATOMIC_MOVE 下其它选项被忽略，
+     * Windows 实现遇到目标已存在会抛 AccessDeniedException / FileAlreadyExistsException，
+     * 故改为常规 REPLACE_EXISTING（Windows 内部走 MoveFileEx，行为已非常接近原子）。
+     */
+    public static void patchXlsxFile(File xlsx, Session session) throws IOException {
+        if (xlsx == null || !xlsx.exists() || session == null || session.isEmpty()) {
+            return;
+        }
+        Path xlsxPath = xlsx.toPath().toAbsolutePath();
+        Path parent = xlsxPath.getParent();
+        if (parent == null) {
+            throw new IOException("[wps-dispimg] xlsx has no parent directory: " + xlsxPath);
+        }
+        Path tmp = Files.createTempFile(parent, "wps-dispimg-", ".xlsx");
+        boolean success = false;
+        try {
+            // 使用 ZipFile（基于中央目录读）替代 ZipInputStream（顺序流读）。
+            // 原因：POI SXSSFWorkbook 写出的 xlsx，部分 entry 的 LFH（Local File Header）会用 data descriptor
+            // 把 size 字段置 0，把真实 CRC/size 放到 entry 末尾。ZipInputStream 在 STORED 模式下按 LFH 的 size=0
+            // 校验，会抛 "ZipException: invalid entry size (expected 0 but got N bytes)"。ZipFile 直接读中央目录
+            // 拿到真实 size，绕开该问题。
+            try (ZipFile zfIn = new ZipFile(xlsxPath.toFile());
+                 ZipOutputStream zout = new ZipOutputStream(Files.newOutputStream(tmp))) {
+                Enumeration<? extends ZipEntry> entries = zfIn.entries();
+                while (entries.hasMoreElements()) {
+                    ZipEntry entry = entries.nextElement();
+                    String name = entry.getName();
+                    if (entry.isDirectory()) {
+                        writeDirectoryEntry(zout, name);
+                        continue;
+                    }
+                    byte[] bytes;
+                    try (InputStream in = zfIn.getInputStream(entry)) {
+                        bytes = readAll(in);
+                    }
+                    if (CONTENT_TYPES_PART.equals(name)) {
+                        bytes = patchContentTypes(bytes);
+                    } else if (WORKBOOK_RELS_PART.equals(name)) {
+                        bytes = patchWorkbookRels(bytes);
+                    }
+                    writeEntry(zout, name, bytes);
+                }
+                writeEntry(zout, CELLIMAGES_PART,
+                        buildCellImagesXml(session).getBytes(StandardCharsets.UTF_8));
+                writeEntry(zout, CELLIMAGES_RELS_PART,
+                        buildCellImagesRelsXml(session).getBytes(StandardCharsets.UTF_8));
+                int idx = 1;
+                for (RegisteredImage img : session.images) {
+                    writeEntry(zout, "xl/media/cellimage" + idx + ".jpeg", img.bytes);
+                    idx++;
+                }
+            }
+            try {
+                Files.move(tmp, xlsxPath, StandardCopyOption.REPLACE_EXISTING);
+            } catch (IOException moveEx) {
+                // Windows 上偶发 AccessDenied（如反病毒/索引服务短暂占用源文件），retry 一次。
+                try {
+                    Thread.sleep(200L);
+                } catch (InterruptedException ignored) {
+                    Thread.currentThread().interrupt();
+                }
+                Files.move(tmp, xlsxPath, StandardCopyOption.REPLACE_EXISTING);
+                log.warn("[wps-dispimg] patch xlsx move retried after error file={} cause={}",
+                        xlsx.getName(), moveEx.toString());
+            }
+            success = true;
+            log.info("[wps-dispimg] patch xlsx finished file={} images={}", xlsx.getName(), session.size());
+        } catch (IOException ioe) {
+            log.error("[wps-dispimg] patch xlsx failed file={} images={} err={}",
+                    xlsx.getName(), session.size(), ioe.toString(), ioe);
+            throw ioe;
+        } finally {
+            if (!success) {
+                try {
+                    Files.deleteIfExists(tmp);
+                } catch (IOException ignored) {
+                    // 临时文件清理失败不影响主流程
+                }
+            }
+        }
+    }
+
+    private static byte[] readAll(InputStream in) throws IOException {
+        ByteArrayOutputStream baos = new ByteArrayOutputStream();
+        byte[] buf = new byte[8192];
+        int n;
+        while ((n = in.read(buf)) > 0) {
+            baos.write(buf, 0, n);
+        }
+        return baos.toByteArray();
+    }
+
+    private static void writeEntry(ZipOutputStream zout, String name, byte[] bytes) throws IOException {
+        ZipEntry entry = new ZipEntry(name);
+        zout.putNextEntry(entry);
+        zout.write(bytes);
+        zout.closeEntry();
+    }
+
+    private static void writeDirectoryEntry(ZipOutputStream zout, String name) throws IOException {
+        ZipEntry entry = new ZipEntry(name);
+        zout.putNextEntry(entry);
+        zout.closeEntry();
+    }
+
+    static byte[] patchContentTypes(byte[] original) {
+        String content = new String(original, StandardCharsets.UTF_8);
+        // 已经存在则跳过（幂等：避免重复 patch）
+        if (content.contains("PartName=\"/" + CELLIMAGES_PART + "\"")) {
+            return original;
+        }
+        String injection = "<Override PartName=\"/" + CELLIMAGES_PART
+                + "\" ContentType=\"" + CELLIMAGES_CONTENT_TYPE + "\"/>";
+        int closeIdx = content.lastIndexOf("</Types>");
+        if (closeIdx < 0) {
+            log.warn("[wps-dispimg] [Content_Types].xml missing </Types>, skip patch");
+            return original;
+        }
+        String patched = content.substring(0, closeIdx) + injection + content.substring(closeIdx);
+        return patched.getBytes(StandardCharsets.UTF_8);
+    }
+
+    static byte[] patchWorkbookRels(byte[] original) {
+        String content = new String(original, StandardCharsets.UTF_8);
+        if (content.contains("Target=\"cellimages.xml\"")) {
+            return original;
+        }
+        // 扫描已有最大 rId
+        int maxId = 0;
+        Matcher m = Pattern.compile("Id=\"rId(\\d+)\"").matcher(content);
+        while (m.find()) {
+            try {
+                maxId = Math.max(maxId, Integer.parseInt(m.group(1)));
+            } catch (NumberFormatException ignored) {
+                // 非数字 rId 无影响
+            }
+        }
+        String newId = "rId" + (maxId + 1);
+        String injection = "<Relationship Id=\"" + newId
+                + "\" Type=\"" + CELLIMAGES_REL_TYPE
+                + "\" Target=\"cellimages.xml\"/>";
+        int closeIdx = content.lastIndexOf("</Relationships>");
+        if (closeIdx < 0) {
+            log.warn("[wps-dispimg] xl/_rels/workbook.xml.rels missing </Relationships>, skip patch");
+            return original;
+        }
+        String patched = content.substring(0, closeIdx) + injection + content.substring(closeIdx);
+        return patched.getBytes(StandardCharsets.UTF_8);
+    }
+
+    static String buildCellImagesXml(Session session) {
+        StringBuilder sb = new StringBuilder();
+        sb.append("<?xml version=\"1.0\" encoding=\"UTF-8\" standalone=\"yes\"?>");
+        sb.append("<etc:cellImages")
+                .append(" xmlns:etc=\"http://www.wps.cn/officeDocument/2017/etCustomData\"")
+                .append(" xmlns:xdr=\"http://schemas.openxmlformats.org/drawingml/2006/spreadsheetDrawing\"")
+                .append(" xmlns:a=\"http://schemas.openxmlformats.org/drawingml/2006/main\"")
+                .append(" xmlns:r=\"http://schemas.openxmlformats.org/officeDocument/2006/relationships\">");
+        int idx = 1;
+        for (RegisteredImage img : session.images) {
+            // cNvPr id 从 2 起步、descr 为空字符串、spPr 末尾追加 a:ln/a:noFill —— 与 WPS 实测样本完全对齐，
+            // 减少 Excel 在"修复"路径上把整个 part 丢弃的概率。
+            int cNvPrId = idx + 1;
+            sb.append("<etc:cellImage>")
+                    .append("<xdr:pic>")
+                    .append("<xdr:nvPicPr>")
+                    .append("<xdr:cNvPr id=\"").append(cNvPrId)
+                    .append("\" name=\"").append(img.id)
+                    .append("\" descr=\"\"/>")
+                    .append("<xdr:cNvPicPr><a:picLocks noChangeAspect=\"1\"/></xdr:cNvPicPr>")
+                    .append("</xdr:nvPicPr>")
+                    .append("<xdr:blipFill>")
+                    .append("<a:blip xmlns:r=\"http://schemas.openxmlformats.org/officeDocument/2006/relationships\"")
+                    .append(" r:embed=\"rId").append(idx).append("\"/>")
+                    .append("<a:stretch><a:fillRect/></a:stretch>")
+                    .append("</xdr:blipFill>")
+                    .append("<xdr:spPr>")
+                    .append("<a:prstGeom prst=\"rect\"><a:avLst/></a:prstGeom>")
+                    .append("<a:ln><a:noFill/></a:ln>")
+                    .append("</xdr:spPr>")
+                    .append("</xdr:pic>")
+                    .append("</etc:cellImage>");
+            idx++;
+        }
+        sb.append("</etc:cellImages>");
+        return sb.toString();
+    }
+
+    static String buildCellImagesRelsXml(Session session) {
+        StringBuilder sb = new StringBuilder();
+        sb.append("<?xml version=\"1.0\" encoding=\"UTF-8\" standalone=\"yes\"?>");
+        sb.append("<Relationships xmlns=\"http://schemas.openxmlformats.org/package/2006/relationships\">");
+        int idx = 1;
+        for (RegisteredImage ignored : session.images) {
+            sb.append("<Relationship Id=\"rId").append(idx)
+                    .append("\" Type=\"").append(IMAGE_REL_TYPE)
+                    .append("\" Target=\"media/cellimage").append(idx)
+                    .append(".jpeg\"/>");
+            idx++;
+        }
+        sb.append("</Relationships>");
+        return sb.toString();
+    }
+
+    /**
+     * 同一 task 内的图片注册会话。线程安全（synchronized）。
+     */
+    public static final class Session {
+        final List<RegisteredImage> images = new ArrayList<>();
+        // 字节级去重：相同图片只写一次 cellimage{N}.jpeg，节省 xlsx 体积。
+        private final Map<String, String> bytesHashToId = new LinkedHashMap<>();
+
+        public synchronized String registerImage(byte[] jpegBytes) {
+            if (jpegBytes == null || jpegBytes.length == 0) {
+                return null;
+            }
+            String hash = md5Hex(jpegBytes);
+            String existingId = bytesHashToId.get(hash);
+            if (existingId != null) {
+                return existingId;
+            }
+            String id = newImageId();
+            images.add(new RegisteredImage(id, jpegBytes));
+            bytesHashToId.put(hash, id);
+            return id;
+        }
+
+        public synchronized boolean isEmpty() {
+            return images.isEmpty();
+        }
+
+        public synchronized int size() {
+            return images.size();
+        }
+
+        private static String newImageId() {
+            // ID_xxx 格式（24 hex chars）模拟 WPS 默认风格，避免特殊字符让公式解析出错。
+            return "ID_" + UUID.randomUUID().toString().replace("-", "").toUpperCase();
+        }
+
+        private static String md5Hex(byte[] bytes) {
+            try {
+                MessageDigest md = MessageDigest.getInstance("MD5");
+                byte[] digest = md.digest(bytes);
+                StringBuilder sb = new StringBuilder(digest.length * 2);
+                for (byte b : digest) {
+                    sb.append(String.format("%02x", b));
+                }
+                return sb.toString();
+            } catch (NoSuchAlgorithmException ex) {
+                // MD5 缺失只在极端 JVM 下出现，退化用 hashCode 即可保证 session 内"内存唯一性"
+                return "len" + bytes.length + ":h" + Arrays.hashCode(bytes);
+            }
+        }
+    }
+
+    static final class RegisteredImage {
+        final String id;
+        final byte[] bytes;
+
+        RegisteredImage(String id, byte[] bytes) {
+            this.id = id;
+            this.bytes = bytes;
+        }
+    }
+}

backend-java/src/main/resources/application.yml

@@ -181,6 +181,7 @@ aiimage:
     stale-finalize-cron: ${AIIMAGE_SIMILAR_ASIN_STALE_FINALIZE_CRON:0 */2 * * * *}
     coze-include-legacy-api-key: ${AIIMAGE_SIMILAR_ASIN_COZE_INCLUDE_LEGACY_API_KEY:true}
     coze-use-legacy-item-field-order: ${AIIMAGE_SIMILAR_ASIN_COZE_USE_LEGACY_ITEM_ORDER:false}
+    coze-result-buffer-enabled: ${AIIMAGE_SIMILAR_ASIN_COZE_RESULT_BUFFER_ENABLED:true}
   security:
     shop-credential-key: ${AIIMAGE_SHOP_CREDENTIAL_KEY:change-me-shop-credential-key}
     internal-token: ${AIIMAGE_INTERNAL_TOKEN:}