diff --git a/mateclaw-server/src/main/java/vip/mate/agent/AgentGraphBuilder.java b/mateclaw-server/src/main/java/vip/mate/agent/AgentGraphBuilder.java
index 2ccc1a2d..cc22af1d 100644
--- a/mateclaw-server/src/main/java/vip/mate/agent/AgentGraphBuilder.java
+++ b/mateclaw-server/src/main/java/vip/mate/agent/AgentGraphBuilder.java
@@ -1066,7 +1066,8 @@ public class AgentGraphBuilder {
         String completionsPath = resolveOpenAiCompletionsPath(baseUrl, kwargs);
         RestClient.Builder restClientBuilder = applyHttpTimeouts(
                 restClientBuilderProvider.getIfAvailable(RestClient::builder));
-        WebClient.Builder webClientBuilder = webClientBuilderProvider.getIfAvailable(WebClient::builder);
+        WebClient.Builder webClientBuilder = applyHttpTimeoutsToWebClient(
+                webClientBuilderProvider.getIfAvailable(WebClient::builder));
 
         // Spring AI OpenAiApi 构造函数会先 set User-Agent 为 "spring-ai"，再 addAll 我们的 headers，
         // 导致自定义 User-Agent 被追加而非覆盖。因此对需要伪装客户端身份的 provider（如 kimi-code），
@@ -1431,6 +1432,31 @@ public class AgentGraphBuilder {
         return builder.requestFactory(rf);
     }
 
+    /**
+     * Apply equivalent timeouts to the WebClient that backs OpenAI-compatible
+     * STREAMING calls (chat completions with {@code stream:true}). The
+     * RestClient version above only protects synchronous HTTP — without this,
+     * the streaming code path uses the default {@code WebClient} which has
+     * neither connect nor read timeout, so a stalled provider can hang the
+     * call forever (observed: a single volcengine-plan request held the agent
+     * thread for 9+ minutes with no error, until the user manually pressed
+     * Stop). That kept the failover chain idle because nothing threw.
+     * <p>
+     * Uses {@link JdkClientHttpConnector} with the same {@link HttpClient} we
+     * already use for the RestClient so the dependency surface stays clean
+     * (reactor-netty is not on this project's classpath — Spring's webflux
+     * starter is excluded by design).
+     */
+    private WebClient.Builder applyHttpTimeoutsToWebClient(WebClient.Builder builder) {
+        HttpClient httpClient = HttpClient.newBuilder()
+                .connectTimeout(Duration.ofSeconds(10))
+                .build();
+        org.springframework.http.client.reactive.JdkClientHttpConnector connector =
+                new org.springframework.http.client.reactive.JdkClientHttpConnector(httpClient);
+        connector.setReadTimeout(Duration.ofSeconds(180));
+        return builder.clientConnector(connector);
+    }
+
     /**
      * 从 generateKwargs.headers 中提取需要强制覆盖的 headers。
      * 用于通过 RestClient/WebClient 拦截器绕过 Spring AI OpenAiApi 的默认 User-Agent。
diff --git a/mateclaw-server/src/main/java/vip/mate/agent/chatmodel/AgentAnthropicChatModelBuilder.java b/mateclaw-server/src/main/java/vip/mate/agent/chatmodel/AgentAnthropicChatModelBuilder.java
index 4aeb7dd1..10c53de2 100644
--- a/mateclaw-server/src/main/java/vip/mate/agent/chatmodel/AgentAnthropicChatModelBuilder.java
+++ b/mateclaw-server/src/main/java/vip/mate/agent/chatmodel/AgentAnthropicChatModelBuilder.java
@@ -86,7 +86,8 @@ public class AgentAnthropicChatModelBuilder implements ChatModelBuilder {
         String baseUrl = provider.getBaseUrl();
         RestClient.Builder restClientBuilder = applyHttpTimeouts(
                 restClientBuilderProvider.getIfAvailable(RestClient::builder));
-        WebClient.Builder webClientBuilder = webClientBuilderProvider.getIfAvailable(WebClient::builder);
+        WebClient.Builder webClientBuilder = applyHttpTimeoutsToWebClient(
+                webClientBuilderProvider.getIfAvailable(WebClient::builder));
 
         AnthropicApi.Builder builder = AnthropicApi.builder()
                 .apiKey(apiKey.trim())
@@ -200,4 +201,25 @@ public class AgentAnthropicChatModelBuilder implements ChatModelBuilder {
         rf.setReadTimeout(Duration.ofSeconds(180));
         return builder.requestFactory(rf);
     }
+
+    /**
+     * Streaming counterpart of {@link #applyHttpTimeouts(RestClient.Builder)}.
+     * Without this, Spring AI's AnthropicApi would back its streaming chat
+     * call by a default WebClient with neither connect nor read timeout — a
+     * stalled provider could hang the agent thread indefinitely while the
+     * failover chain idles (no exception = no signal).
+     * <p>
+     * Mirrors AgentGraphBuilder.applyHttpTimeoutsToWebClient: same JDK
+     * HttpClient + JdkClientHttpConnector path, so the dependency surface
+     * doesn't pull in reactor-netty (excluded by this project's pom).
+     */
+    static WebClient.Builder applyHttpTimeoutsToWebClient(WebClient.Builder builder) {
+        HttpClient httpClient = HttpClient.newBuilder()
+                .connectTimeout(Duration.ofSeconds(10))
+                .build();
+        org.springframework.http.client.reactive.JdkClientHttpConnector connector =
+                new org.springframework.http.client.reactive.JdkClientHttpConnector(httpClient);
+        connector.setReadTimeout(Duration.ofSeconds(180));
+        return builder.clientConnector(connector);
+    }
 }