diff --git a/game-cloud/game-module-aigc/game-module-aigc-server/src/test/java/com/wanxiang/huijing/game/module/aigc/saa/SaaStreamHangReproTest.java b/game-cloud/game-module-aigc/game-module-aigc-server/src/test/java/com/wanxiang/huijing/game/module/aigc/saa/SaaStreamHangReproTest.java
new file mode 100644
index 00000000..273364e5
--- /dev/null
+++ b/game-cloud/game-module-aigc/game-module-aigc-server/src/test/java/com/wanxiang/huijing/game/module/aigc/saa/SaaStreamHangReproTest.java
@@ -0,0 +1,177 @@
+package com.wanxiang.huijing.game.module.aigc.saa;
+
+import org.junit.jupiter.api.Test;
+import org.junit.jupiter.api.condition.EnabledIfEnvironmentVariable;
+import org.springframework.ai.chat.messages.UserMessage;
+import org.springframework.ai.chat.model.ChatModel;
+import org.springframework.ai.chat.model.ChatResponse;
+import org.springframework.ai.chat.prompt.Prompt;
+import org.springframework.ai.openai.OpenAiChatModel;
+import org.springframework.ai.openai.api.OpenAiApi;
+import reactor.core.publisher.Flux;
+
+import java.time.Duration;
+import java.util.List;
+import java.util.concurrent.atomic.AtomicInteger;
+import java.util.concurrent.atomic.AtomicLong;
+import java.util.concurrent.atomic.AtomicReference;
+
+/**
+ * SAA openai+流式 跑 M3「静默挂死」最小复现（诊断脚手架，非产线门）。
+ *
+ * <p>逐一验诊断假设：① stream() 是否真流式（块逐步到达 vs 整体缓冲）；② thinking 开时 content 帧是否到达
+ * （还是全进 reasoning_content → MessageAggregator 等不到 content 永阻塞）；③ blockLast() 是否返回。
+ * 全程显式超时（不挂 900s），证据直打 stdout。
+ *
+ * <p>门控 {@code -Dsaa.stream.repro=1} + NEWAPI_KEY；不破普通构建。
+ */
+@EnabledIfEnvironmentVariable(named = "NEWAPI_KEY", matches = ".+")
+class SaaStreamHangReproTest {
+
+    private static final String BASE = System.getenv().getOrDefault("NEWAPI_BASE_URL", "http://100.64.0.8:3000");
+    private static final String MODEL = "MiniMax-M3";
+
+    @Test
+    void repro_stream_hang() throws Exception {
+        if (!"1".equals(System.getProperty("saa.stream.repro"))) {
+            System.out.println("[skip] 加 -Dsaa.stream.repro=1 才跑本诊断");
+            return;
+        }
+        String key = System.getenv("NEWAPI_KEY");
+        String base = SaaStudioGraph.stripV1Suffix(BASE);
+        System.out.println("[repro] base=" + base + " model=" + MODEL + " keyLen=" + (key == null ? 0 : key.length()));
+
+        // 生产同款：NO_PROXY OpenAiApi（读超时 600s）+ model() 对 M3 注入 thinking:adaptive+reasoning_split+48K
+        OpenAiApi api = SaaStudioGraph.openAiApiNoProxy(base, key, 600_000);
+        OpenAiChatModel m3 = SaaStudioGraph.model(api, MODEL, 0.0, 16000);  // maxTokens≥8000 → 生成角色配置（开 thinking）
+        // 真实规模 prompt：强制大输出（完整游戏源码），逼近 code 角色真实负载（48K maxTokens + 长 thinking + 长 content），
+        //   复现 e2e 大请求下的挂死（小 prompt 短答案不复现）。
+        String bigSystem = "你是一个 H5 小游戏代码生成器。产出一个完整、可运行的原生 JS 打砖块游戏单文件模块，"
+                + "包含完整的挡板物理、小球反弹、砖块矩阵、碰撞检测、计分、胜负判定、渲染循环。"
+                + "代码要详尽完整、注释充分、不省略任何函数体。只输出一个 ```js 代码块。";
+        Prompt prompt = new Prompt(List.of(
+                new org.springframework.ai.chat.messages.SystemMessage(bigSystem),
+                new UserMessage("实现打砖块：挡板接球把上方砖块全部消除即胜，掉球则负。要求代码完整详尽，不少于 300 行。")));
+
+        // ===== 实验1：直接订阅 stream()，逐帧打印到达时刻 + content/reasoning 长度（判真流式 + content 是否到达）=====
+        System.out.println("\n===== 实验1：OpenAiChatModel.stream() 逐帧 =====");
+        long t0 = System.currentTimeMillis();
+        AtomicInteger frames = new AtomicInteger();
+        AtomicInteger contentFrames = new AtomicInteger();
+        AtomicLong contentChars = new AtomicLong();
+        AtomicReference<Long> firstAt = new AtomicReference<>(null);
+        AtomicLong lastFrameAt = new AtomicLong(0);
+        AtomicLong maxGap = new AtomicLong(0);
+        try {
+            Flux<ChatResponse> flux = m3.stream(prompt);
+            flux.doOnNext(cr -> {
+                long at = System.currentTimeMillis() - t0;
+                long gap = at - lastFrameAt.get();
+                if (lastFrameAt.get() > 0 && gap > maxGap.get()) {
+                    maxGap.set(gap);
+                }
+                lastFrameAt.set(at);
+                if (firstAt.get() == null) {
+                    firstAt.set(at);
+                    System.out.println("  [首帧 @" + at + "ms]");
+                }
+                int idx = frames.incrementAndGet();
+                String txt = "";
+                try {
+                    if (cr.getResult() != null && cr.getResult().getOutput() != null) {
+                        txt = cr.getResult().getOutput().getText();
+                    }
+                } catch (Exception ignore) {
+                }
+                if (txt != null && !txt.isEmpty()) {
+                    contentFrames.incrementAndGet();
+                    contentChars.addAndGet(txt.length());
+                }
+                if (idx <= 6 || (txt != null && !txt.isEmpty() && contentFrames.get() <= 3)) {
+                    System.out.println("  帧#" + idx + " @" + at + "ms content="
+                            + (txt == null || txt.isEmpty() ? "无" : "有[" + txt.length() + "]"));
+                }
+            }).blockLast(Duration.ofSeconds(300));
+            System.out.println("  [stream 完成 @" + (System.currentTimeMillis() - t0) + "ms] 总帧=" + frames.get()
+                    + " content帧=" + contentFrames.get() + "(" + contentChars.get() + "字) 首帧=" + firstAt.get()
+                    + "ms 最大帧间隔=" + maxGap.get() + "ms");
+        } catch (Exception e) {
+            System.out.println("  [stream 异常 @" + (System.currentTimeMillis() - t0) + "ms] "
+                    + e.getClass().getName() + ": " + e.getMessage()
+                    + "  已收帧=" + frames.get() + " content帧=" + contentFrames.get());
+        }
+
+        // ===== 实验2：生产真链复刻 = StreamingCallChatModel.call() 套「180s 单次超时 + cancel(true) + ≤3 重试」 =====
+        //   这是 SaaStudioNodes.callWithTimeoutAndRetry 的等价复刻（生产 generate 节点真用）。
+        //   假设：M3 thinking 阶段 >180s（实验1 已证 content 首帧 @~203s）→ 每次 get(180s) 必 TimeoutException →
+        //   cancel(true) 中断 blockLast 线程，但底层 JDK HttpClient SSE 连接不被中断取消（连接泄漏）→ 反复重试 → 静默卡死。
+        System.out.println("\n===== 实验2：生产真链复刻（StreamingCallChatModel.call + 180s超时×3重试 + cancel）=====");
+        ChatModel wrapped = SaaStudioGraph.openAiFactory(api).create(MODEL, 0.0, 16000);
+        long t1 = System.currentTimeMillis();
+        AtomicReference<ChatResponse> got = new AtomicReference<>();
+        AtomicReference<Throwable> err = new AtomicReference<>();
+        // 整体放后台线程 + join 上限 620s（3×180s + 退避 + 余量），看是否在合理时间内「返回 / 抛超时 / 仍卡死」。
+        Thread th = new Thread(() -> {
+            try {
+                got.set(prodLikeCallWithTimeoutAndRetry(wrapped, prompt));
+            } catch (Throwable t) {
+                err.set(t);
+            }
+        }, "repro-prodlike");
+        th.setDaemon(true);
+        th.start();
+        th.join(620_000);
+        long dt = System.currentTimeMillis() - t1;
+        if (th.isAlive()) {
+            System.out.println("  [!!! 生产真链 620s 未返回 = 静默卡死复现 @" + dt + "ms] 主线程仍 alive");
+            for (StackTraceElement ste : th.getStackTrace()) {
+                System.out.println("      at " + ste);
+            }
+            // 同时 dump 进程内所有 saa-llm-call 线程，看泄漏的 blockLast 线程是否堆积
+            int leaked = 0;
+            for (Thread t : Thread.getAllStackTraces().keySet()) {
+                if (t.getName().contains("repro-llm") || t.getName().contains("saa-llm")) {
+                    leaked++;
+                }
+            }
+            System.out.println("      泄漏的 LLM 调用线程数=" + leaked + "（>1 = cancel 未真停底层流，连接泄漏）");
+        } else if (err.get() != null) {
+            System.out.println("  [生产真链 抛异常 @" + dt + "ms（约=N×180s）] " + err.get().getClass().getName()
+                    + ": " + err.get().getMessage() + "  → 生产把它归 llm_error，但耗时已逼近/超 900s 预算");
+        } else {
+            ChatResponse cr = got.get();
+            String txt = (cr == null || cr.getResult() == null) ? null : cr.getResult().getOutput().getText();
+            System.out.println("  [生产真链 返回 @" + dt + "ms] contentLen=" + (txt == null ? 0 : txt.length()));
+        }
+        System.out.println("\n===== 复现结束 =====");
+    }
+
+    /** 生产 SaaStudioNodes.callWithTimeoutAndRetry 的等价复刻（180s 单次超时 + cancel(true) + ≤3 重试）。 */
+    private static ChatResponse prodLikeCallWithTimeoutAndRetry(ChatModel model, Prompt prompt) throws Exception {
+        java.util.concurrent.ExecutorService pool = java.util.concurrent.Executors.newCachedThreadPool(r -> {
+            Thread t = new Thread(r, "repro-llm-call");
+            t.setDaemon(true);
+            return t;
+        });
+        Exception lastEx = null;
+        for (int attempt = 0; attempt < 3; attempt++) {
+            long a0 = System.currentTimeMillis();
+            java.util.concurrent.Future<ChatResponse> f = pool.submit(() -> model.call(prompt));
+            try {
+                ChatResponse r = f.get(180, java.util.concurrent.TimeUnit.SECONDS);
+                System.out.println("    尝试#" + (attempt + 1) + " 成功 @" + (System.currentTimeMillis() - a0) + "ms");
+                return r;
+            } catch (java.util.concurrent.TimeoutException te) {
+                boolean cancelled = f.cancel(true);
+                System.out.println("    尝试#" + (attempt + 1) + " 180s 超时 → cancel(true) 返回=" + cancelled
+                        + "（cancel=true 仅意味发了中断信号，不等于底层 HTTP 流真停）");
+                lastEx = new RuntimeException("LLM 调用超时(180s)", te);
+            } catch (java.util.concurrent.ExecutionException ee) {
+                System.out.println("    尝试#" + (attempt + 1) + " 异常 @" + (System.currentTimeMillis() - a0) + "ms: " + ee.getCause());
+                lastEx = (ee.getCause() instanceof Exception) ? (Exception) ee.getCause() : new RuntimeException(ee.getCause());
+            }
+            if (attempt < 2) Thread.sleep(2000);
+        }
+        throw lastEx != null ? lastEx : new RuntimeException("重试 3 次后失败");
+    }
+}