diff --git a/deploy/profile-scripts/llm_client.py b/deploy/profile-scripts/llm_client.py
index 3c14d52b..20e5ea0f 100644
--- a/deploy/profile-scripts/llm_client.py
+++ b/deploy/profile-scripts/llm_client.py
@@ -25,6 +25,9 @@ import urllib.error
 # ── 常量：所有重评调用统一使用 ──
 # 2026-07-21 A/B 实测：flash 与 pro 在新 prompt 下质量差距微弱，flash 快 ~40%
 REASSESS_MODEL = "deepseek-v4-flash"
+# flash 对某些 prompt 会稳定返回空内容（688617 实测三连空），pro 能正常输出。
+# 空输出时自动升级到 pro 重试一次。
+FALLBACK_MODEL = "deepseek-v4-pro"
 
 # 主通道：OCG 上游直连（与 hermes providers.ocg-key6 同源）
 # key 运行时从 hermes config 读取（SSOT，不落盘到代码库）
@@ -145,6 +148,29 @@ def call_llm(prompt, model=None, max_tokens=4096, timeout=150,
                 ok, result = _post(ch_url, ch_headers, payload, timeout)
                 elapsed = time.monotonic() - t0
                 if ok:
+                    # ── 空输出升级：flash 对部分 prompt 稳定返回空（688617 实测），
+                    #    换 FALLBACK_MODEL(pro) 重试一次 ──
+                    if not result.strip() and model_name != FALLBACK_MODEL and OCG_HEADERS:
+                        print(f"  [LLM] {ch_name} {model_name} 空输出({elapsed:.1f}s)，"
+                              f"升级 {FALLBACK_MODEL} 重试...", flush=True)
+                        esc_payload = json.dumps({
+                            "model": FALLBACK_MODEL, "messages": messages,
+                            "max_tokens": max_tokens,
+                        }).encode()
+                        try:
+                            ok2, result2 = _post(OCG_URL, OCG_HEADERS, esc_payload, timeout)
+                            total_attempts += 1
+                            if ok2 and result2.strip():
+                                print(f"  [LLM] 升级 {FALLBACK_MODEL} 成功, 输出{len(result2)}字", flush=True)
+                                return {
+                                    "ok": True, "content": result2, "error": None,
+                                    "model": FALLBACK_MODEL,
+                                    "elapsed": time.monotonic() - t_start,
+                                    "attempts": total_attempts, "channel": ch_name + "+esc",
+                                }
+                            print(f"  [LLM] 升级 {FALLBACK_MODEL} 仍空/失败", flush=True)
+                        except Exception as e2:
+                            print(f"  [LLM] 升级 {FALLBACK_MODEL} 异常: {str(e2)[:100]}", flush=True)
                     print(f"  [LLM] {ch_name} 尝试{attempt+1}/{retries+1} 成功, "
                           f"{elapsed:.1f}s, 输出{len(result)}字, model={model_name}", flush=True)
                     return {