diff --git a/backend/python/app/api/chat.py b/backend/python/app/api/chat.py
index 55cb612..9e23cd4 100644
--- a/backend/python/app/api/chat.py
+++ b/backend/python/app/api/chat.py
@@ -156,6 +156,7 @@ def _proxy_for_url(url: str) -> str | None:
     注意：httpx 默认 trust_env=True 会忽略 proxy 参数而改用环境变量 HTTP_PROXY，
     故调用方必须同时传 trust_env=False 才能真正走此处指定的代理。
     """
+    import os
     from urllib.parse import urlparse
     host = (urlparse(url).hostname or "").lower()
     if "agnes-ai.com" in host or "agnes-ai.space" in host:
@@ -163,6 +164,12 @@ def _proxy_for_url(url: str) -> str | None:
     # 平台自有域名(含自建 Dify/n8n 等)必须直连，不经出网代理。
     if "qihuyun.com" in host:
         return None
+    # 尊重 NO_PROXY：国内域名(硅基流动/DeepSeek/小米MoE 等)直连更稳定，
+    # 避免经易波动的智能路由代理 172.18.0.1:7890(对国内域名偶发连接重置)。
+    _no_proxy = [h.strip().lower() for h in os.environ.get("NO_PROXY", "").split(",") if h.strip()]
+    for _np in _no_proxy:
+        if host == _np or host.endswith("." + _np) or _np in host:
+            return None
     return _outbound_proxy()
 
 
@@ -186,7 +193,7 @@ async def _optimize_media_prompt(raw_prompt: str, media_type: str, db=None) -> s
         return stripped
     # 用文本 LLM 做翻译增强
     try:
-        from .capability_registry import resolve_auto_model as _resolve_auto
+        from ..capability_registry import resolve_auto_model as _resolve_auto
         from ..models import ModelProvider as MP
         from ..database import SessionLocal as SL
         import httpx as _httpx
@@ -224,7 +231,11 @@ async def _optimize_media_prompt(raw_prompt: str, media_type: str, db=None) -> s
                 {"role": "system", "content": "You are a professional prompt engineer. Output ONLY the final prompt text, no explanations."},
                 {"role": "user", "content": prompt_zh},
             ]
-            body = {"model": text_model.name, "messages": messages, "max_tokens": 512, "temperature": 0.7}
+            # max_tokens 调大到 1024：mimo 等推理模型会把 token 消耗在思考链(reasoning_content)
+            # 上，512 时 content 常被截断为空；并关闭推理(enable_thinking=False)，让模型直接输出提示词。
+            # 对不支持该参数的模型至多优化失败 → 下方 except 兜底退化为原始提示词，不阻断生成。
+            body = {"model": text_model.name, "messages": messages, "max_tokens": 1024,
+                    "temperature": 0.7, "enable_thinking": False}
             async with _httpx.AsyncClient(
                 timeout=30.0, proxy=_proxy_for_url(api_url), trust_env=False
             ) as client:
@@ -309,7 +320,11 @@ async def _generate_media(db, ai_model, prompt: str, user_id: str = None) -> str
         if not api_url:
             return "⚠️ 该图像模型未配置 Endpoint。"
         # 先优化 prompt（用户原始输入含多轮对话噪音 → 高质量图像生成提示词）
-        optimized_prompt = await _optimize_media_prompt(prompt, "image", db)
+        # 优化失败绝不中断生成，退化为原始提示词
+        try:
+            optimized_prompt = await _optimize_media_prompt(prompt, "image", db)
+        except Exception:
+            optimized_prompt = prompt
         payload = {
             "model": ai_model.name,
             "prompt": optimized_prompt,
@@ -320,7 +335,16 @@ async def _generate_media(db, ai_model, prompt: str, user_id: str = None) -> str
             # 3 次重试后仍失败 → AGNES 侧持续抖动，提示稍后重试而非甩原始错误
             return _friendly_media_err("图像", err)
         if resp.status_code != 200:
-            return _friendly_media_err("图像", f"HTTP {resp.status_code}: {resp.text[:200]}")
+            # AGNES 内容策略拒绝：用原始(未优化)提示词重试一次，避免优化后提示词触发误杀
+            if resp.status_code == 400 and "content_policy" in (resp.text or "").lower():
+                payload["prompt"] = prompt
+                resp2, _ = await _agnes_post_json(api_url, headers, payload, timeout)
+                if resp2 is not None and resp2.status_code == 200:
+                    resp = resp2
+                else:
+                    return "⚠️ 图像生成被内容策略拒绝，请换一个更中性的描述后重试。"
+            else:
+                return _friendly_media_err("图像", f"HTTP {resp.status_code}: {resp.text[:200]}")
         try:
             data = resp.json()
         except Exception:
@@ -344,8 +368,11 @@ async def _generate_media(db, ai_model, prompt: str, user_id: str = None) -> str
             api_url = f"{provider.api_base_url.rstrip('/')}/v1/videos" if provider else ""
         if not api_url:
             return "⚠️ 该视频模型未配置 Endpoint。"
-        # 优化 prompt（用户输入含多轮对话噪音 → 高质量视频提示词）
-        optimized_prompt = await _optimize_media_prompt(prompt, "video", db)
+        # 优化 prompt（用户输入含多轮对话噪音 → 高质量视频提示词）；失败退化为原始提示词
+        try:
+            optimized_prompt = await _optimize_media_prompt(prompt, "video", db)
+        except Exception:
+            optimized_prompt = prompt
         base = api_url.split("/v1")[0] if "/v1" in api_url else api_url
         payload = {
             "model": ai_model.name,
@@ -359,7 +386,16 @@ async def _generate_media(db, ai_model, prompt: str, user_id: str = None) -> str
         if resp is None:
             return _friendly_media_err("视频", err)
         if resp.status_code != 200:
-            return _friendly_media_err("视频", f"HTTP {resp.status_code}: {resp.text[:200]}")
+            # AGNES 内容策略拒绝：用原始(未优化)提示词重试一次，避免优化后提示词触发误杀
+            if resp.status_code == 400 and "content_policy" in (resp.text or "").lower():
+                payload["prompt"] = prompt
+                resp2, _ = await _agnes_post_json(api_url, headers, payload, timeout)
+                if resp2 is not None and resp2.status_code == 200:
+                    resp = resp2
+                else:
+                    return "⚠️ 视频生成被内容策略拒绝，请换一个更中性的描述后重试。"
+            else:
+                return _friendly_media_err("视频", f"HTTP {resp.status_code}: {resp.text[:200]}")
         try:
             data = resp.json()
         except Exception:
diff --git a/backend/python/app/api/im_llm.py b/backend/python/app/api/im_llm.py
index 15b34f6..f4f6d3e 100644
--- a/backend/python/app/api/im_llm.py
+++ b/backend/python/app/api/im_llm.py
@@ -33,6 +33,7 @@ def _proxy_for_url(url: str) -> str | None:
     """按目标域名选择是否走代理（与 chat.py 同逻辑）。
 
     2026-07-15 调整：AGNES 改走香港代理(实测延迟更低)，不再直连。
+    国内域名(尊重 NO_PROXY)直连，更稳定。
     """
     import os
     host = (_urlparse(url).hostname or "").lower()
@@ -40,6 +41,10 @@ def _proxy_for_url(url: str) -> str | None:
         return _HK_PROXY
     if "qihuyun.com" in host:
         return None
+    _no_proxy = [h.strip().lower() for h in os.environ.get("NO_PROXY", "").split(",") if h.strip()]
+    for _np in _no_proxy:
+        if host == _np or host.endswith("." + _np) or _np in host:
+            return None
     return os.environ.get("HTTPS_PROXY") or os.environ.get("HTTP_PROXY") or "http://172.18.0.1:7890"
 
 
@@ -78,8 +83,10 @@ def _call_real_llm(db: Session, user=None, prompt: str = "", platform: str = "im
             if ai_model:
                 try:
                     result = _run_async(_generate_media(db, ai_model, prompt, user_id))
-                except Exception as e:
-                    result = f"（媒体生成失败：{e}）"
+                except Exception:
+                    # 任何未预期异常(如 AGNES 连接重置)统一兜底为友好提示，
+                    # 绝不直接把原始异常(如 "Server disconnected")甩给用户
+                    result = "⚠️ 媒体生成暂时失败（连接或服务波动），请稍后重试。"
                 if result:
                     # IM 不渲染 HTML，把 <img>/<video> 还原为纯文本链接返回
                     s = result.strip()