diff --git a/src/main/java/com/wok/supportbot/app/AssistantApp.java b/src/main/java/com/wok/supportbot/app/AssistantApp.java
index faa297c..a3a4010 100644
--- a/src/main/java/com/wok/supportbot/app/AssistantApp.java
+++ b/src/main/java/com/wok/supportbot/app/AssistantApp.java
@@ -44,7 +44,6 @@ import java.util.List;
 import java.util.Map;
 import java.util.UUID;
 import java.util.concurrent.CopyOnWriteArrayList;
-import java.util.concurrent.atomic.AtomicBoolean;
 import java.util.concurrent.atomic.AtomicInteger;
 import java.util.concurrent.atomic.AtomicReference;
 
@@ -519,11 +518,9 @@ public class AssistantApp {
                 });
         // 聚合所有分片用于埋点（在 doFinally 时取完整回复文本）
         StringBuilder aggregated = new StringBuilder();
-        // 首片标记：第一片 delta 需带 role=assistant，后续片仅含 content
-        AtomicBoolean first = new AtomicBoolean(true);
         return preserveTrailingWhitespace(rawStream)
                 .doOnNext(aggregated::append)
-                .map(chunk -> buildOpenAiChunk(completionId, model, created, chunk, first.getAndSet(false), null))
+                .map(chunk -> buildOpenAiChunk(completionId, model, created, chunk, false, null))
                 .doOnComplete(() -> aiCircuitBreaker.recordSuccess(AI_CIRCUIT_KEY))
                 .doOnError(e -> {
                     aiCircuitBreaker.recordFailure(AI_CIRCUIT_KEY);
@@ -543,13 +540,17 @@ public class AssistantApp {
                                     usage != null ? usage.getTotalTokens() : null,
                                     events));
                 })
+                // 首片（仅 role=assistant、无 content）在流订阅时立即发出，确保 SSE 响应头/首字节及时 flush。
+                // 推理模型（如 doubao-seed）思考阶段 delta.content 为空、被 preserveTrailingWhitespace 吞掉，
+                // 若不提前发首片，思考阶段将无任何字节输出，前端等待首字节会触发 60s 超时。
+                .startWith(buildOpenAiChunk(completionId, model, created, "", true, null))
                 // 流正常结束时追加 finish_reason=stop 的 chunk 与 [DONE]
                 .concatWith(Flux.just(
                         buildOpenAiChunk(completionId, model, created, "", false, "stop"),
                         "[DONE]"))
-                // 错误兜底：脱敏错误信息，避免泄露内部细节；role 是否出现取决于此前是否已发出过内容片
+                // 错误兜底：脱敏错误信息，避免泄露内部细节（首片 role 已提前发出，此处不再带 role）
                 .onErrorResume(e -> openAiFallbackStream(completionId, model, created,
-                        "抱歉，AI 服务调用失败：" + maskError(e.getMessage()), first.get()));
+                        "抱歉，AI 服务调用失败：" + maskError(e.getMessage()), false));
     }
 
     /**
diff --git a/src/main/java/com/wok/supportbot/config/ChatModelFactory.java b/src/main/java/com/wok/supportbot/config/ChatModelFactory.java
index 124c6de..15438cc 100644
--- a/src/main/java/com/wok/supportbot/config/ChatModelFactory.java
+++ b/src/main/java/com/wok/supportbot/config/ChatModelFactory.java
@@ -66,8 +66,8 @@ public class ChatModelFactory {
 
     /** AI 模型调用连接超时 */
     private static final Duration AI_CONNECT_TIMEOUT = Duration.ofSeconds(10);
-    /** AI 模型调用读取超时，含流式生成时间 */
-    private static final Duration AI_READ_TIMEOUT = Duration.ofSeconds(120);
+    /** AI 模型调用读取超时，含流式生成时间（推理模型思考+生成耗时较长，放宽到 300s） */
+    private static final Duration AI_READ_TIMEOUT = Duration.ofSeconds(300);
 
     /**
      * 构建带超时配置的 RestClient.Builder，供 DashScope 和 OpenAI 兼容路径共用。