diff --git a/backend/python/app/api/im_bot_user.py b/backend/python/app/api/im_bot_user.py
index a0dff98..e6fb53c 100644
--- a/backend/python/app/api/im_bot_user.py
+++ b/backend/python/app/api/im_bot_user.py
@@ -131,7 +131,11 @@ def _recent_history(user_id: str, limit: int = 6) -> str:
         for r in reversed(rows):
             dd = _row_to_dict(r)
             role = "用户" if dd["role"] == "user" else "七虎AI"
-            pairs.append(f"{role}：{dd['content']}")
+            content = dd.get("content") or ""
+            # #973：每条历史最多 500 字符，防止超长图片 URL/完整清单撑爆多轮 prompt
+            if len(content) > 500:
+                content = content[:500] + "…"
+            pairs.append(f"{role}：{content}")
         return "\n".join(pairs)
     except Exception:
         return ""
@@ -654,11 +658,8 @@ def _wechat_receiver_loop(account_id: int):
                     ok, ret, _ = _wechat_send_media(baseurl, bot_token, sender, ctx, media_type, media_url)
                     # 对话历史里存可渲染的媒体标签（经平台中转，AI 聊天 WeXin Bot 窗与历史页均可内联显示）
                     relay_url = _platform_relay_url(media_url)
-                    reply_for_log = (
-                        f'<img src="{relay_url}" alt="生成图片"/>'
-                        if media_type == "image"
-                        else f'<video src="{relay_url}" controls></video>'
-                    )
+                    # #973：对话历史只存简短占位符，不存超长 relay URL（含 AWS token，会撑爆多轮 prompt）
+                    reply_for_log = "[图片]" if media_type == "image" else "[视频]"
                 else:
                     reply_text = reply if isinstance(reply, str) else str(reply)
                     ok, ret, _ = _wechat_send_message(baseurl, bot_token, sender, ctx, reply_text)
diff --git a/backend/python/app/api/im_llm.py b/backend/python/app/api/im_llm.py
index 7dcf46a..ceda381 100644
--- a/backend/python/app/api/im_llm.py
+++ b/backend/python/app/api/im_llm.py
@@ -277,10 +277,11 @@ def _call_real_llm(db: Session, user=None, prompt: str = "", platform: str = "im
             f"mcp_tools={len(mcp_function_definitions)}"
         )
 
-        # 平台知识库注入（对齐 Web Chat——微信用户提问"平台有哪些技能"也能得到完整清单）
+        # 平台知识库注入（#973：IM 侧精简——只注入静态文档，不注入 104 动态清单，
+        # 避免 prompt 过长(6892字符)导致推理模型响应慢(63s)且间歇性返回空 content）
         kb_context = ""
         try:
-            pk = get_platform_knowledge(db=db)
+            pk = get_platform_knowledge(db=None)
             if pk:
                 kb_context = f"【平台知识库——优先使用以下知识回答用户问题】\n{pk}"
         except Exception:
@@ -420,6 +421,9 @@ def _call_real_llm(db: Session, user=None, prompt: str = "", platform: str = "im
                         if round_num >= max_rounds - 1:
                             break
                         continue
+                    # #973：空 content 且未到最后一轮 → 重试（推理模型间歇性返回空）
+                    if not ai_content and round_num < max_rounds - 1:
+                        continue
                     break
 
                 # 检查本轮调用是否成功