diff --git a/app/services/grok/services/chat.py b/app/services/grok/services/chat.py
index eec1e8ae8..e7b92e9d3 100644
--- a/app/services/grok/services/chat.py
+++ b/app/services/grok/services/chat.py
@@ -537,8 +537,8 @@ def __init__(
         self.fingerprint: str = ""
         self.rollout_id: str = ""
         self.think_opened: bool = False
-        self.think_closed_once: bool = False
         self.image_think_active: bool = False
+        self._content_started: bool = False
         self.role_sent: bool = False
         self.filter_tags = get_config("app.filter_tags")
         self.tool_usage_enabled = (
@@ -816,7 +816,6 @@ async def process(self, response: AsyncIterable[bytes]) -> AsyncGenerator[str, N
                     if self.image_think_active and self.think_opened:
                         yield self._sse("\n</think>\n")
                         self.think_opened = False
-                        self.think_closed_once = True
                     self.image_think_active = False
                     for url in proc_base._collect_images(mr):
                         parts = url.split("/")
@@ -860,15 +859,26 @@ async def process(self, response: AsyncIterable[bytes]) -> AsyncGenerator[str, N
                 if (token := resp.get("token")) is not None:
                     if not token:
                         continue
-                    if is_thinking and self.think_closed_once and not self.image_think_active:
-                        continue
                     filtered = self._filter_token(token)
                     if not filtered:
                         continue
+                    # 判断是否在 Agent 思考/处理阶段：
+                    #   - isThinking=true → 归入 think
+                    #   - 有 messageStepId → Agent 处理中，归入 think
+                    #   - image_think_active → 图片生成中
+                    #   正式内容开始后，丢弃中途插入的思考（Grok 官网也隐藏了这部分）
+                    has_step_id = bool(resp.get("messageStepId"))
                     in_think = (
-                        (is_thinking and not self.think_closed_once)
+                        is_thinking
+                        or has_step_id
                         or self.image_think_active
                     )
+                    # 正式内容已开始后，丢弃中途插入的 Agent 思考（1-2 句内部注释，无用户价值）
+                    if self._content_started and in_think and not self.image_think_active:
+                        continue
+                    # 空 token 不关闭 think 块（搜索结果间的空 token 不算正式内容）
+                    if not in_think and not filtered.strip():
+                        continue
                     if in_think:
                         if not self.show_think:
                             continue
@@ -879,7 +889,7 @@ async def process(self, response: AsyncIterable[bytes]) -> AsyncGenerator[str, N
                         if self.think_opened:
                             yield self._sse("\n</think>\n")
                             self.think_opened = False
-                            self.think_closed_once = True
+                            self._content_started = True
 
                     if in_think:
                         self._record_content(filtered)
@@ -901,7 +911,6 @@ async def process(self, response: AsyncIterable[bytes]) -> AsyncGenerator[str, N
 
             if self.think_opened:
                 yield self._sse("</think>\n")
-                self.think_closed_once = True
 
             if self._tool_stream_enabled:
                 for kind, payload in self._flush_tool_stream():
@@ -1022,6 +1031,8 @@ async def process(self, response: AsyncIterable[bytes]) -> dict[str, Any]:
         response_id = ""
         fingerprint = ""
         content = ""
+        # 兜底收集非 thinking 且无 messageStepId 的最终内容 token
+        fallback_tokens: list[str] = []
         idle_timeout = get_config("chat.stream_timeout")
 
         try:
@@ -1041,6 +1052,13 @@ async def process(self, response: AsyncIterable[bytes]) -> dict[str, Any]:
                 if (llm := resp.get("llmInfo")) and not fingerprint:
                     fingerprint = llm.get("modelHash", "")
 
+                # 收集非 thinking 且无 messageStepId 的 token（最终内容兜底）
+                is_thinking = bool(resp.get("isThinking"))
+                has_step_id = bool(resp.get("messageStepId"))
+                if not is_thinking and not has_step_id:
+                    if tok := resp.get("token"):
+                        fallback_tokens.append(tok)
+
                 if mr := resp.get("modelResponse"):
                     response_id = mr.get("responseId", "")
                     content = mr.get("message", "")
@@ -1140,6 +1158,10 @@ def _render_card(match: re.Match) -> str:
         finally:
             await self.close()
 
+        # modelResponse.message 为空时（多智能体模型），用兜底 token 拼接
+        if not content and fallback_tokens:
+            content = "".join(fallback_tokens)
+
         content = self._filter_content(content)
 
         # Parse for tool calls if tools were provided
diff --git a/app/services/reverse/app_chat.py b/app/services/reverse/app_chat.py
index 526059f0d..fdb486967 100644
--- a/app/services/reverse/app_chat.py
+++ b/app/services/reverse/app_chat.py
@@ -102,6 +102,19 @@ def _resolve_custom_personality() -> Optional[str]:
             return None
         return value
 
+    # modelMode → modeId 映射（Grok Web 新 API 格式）
+    # 基于浏览器前端 JS 逆向和 API 全量测试验证：
+    #   付费 SuperGrok 的多智能体模式需要 modeId 字段才能正常响应
+    #   有 modeId 时不发 modelName/modelMode（浏览器前端逻辑）
+    _MODE_ID_MAP = {
+        "MODEL_MODE_FAST": "fast",
+        "MODEL_MODE_EXPERT": "expert",
+        "MODEL_MODE_HEAVY": "heavy",
+        "MODEL_MODE_GROK_420": "expert",
+        "MODEL_MODE_GROK_4_1_THINKING": "expert",
+        "MODEL_MODE_GROK_4_1_MINI_THINKING": "expert",
+    }
+
     @staticmethod
     def build_payload(
         message: str,
@@ -152,8 +165,13 @@ def build_payload(
             "toolOverrides": tool_overrides or {},
         }
 
-        if model == "grok-420":
-            payload["enable420"] = True
+        # 优先使用 modeId（Grok 新 API 格式，付费号多智能体模式必需）
+        # 有 modeId 时移除 modelName/modelMode（浏览器前端逻辑）
+        mode_id = AppChatReverse._MODE_ID_MAP.get(mode)
+        if mode_id:
+            payload["modeId"] = mode_id
+            payload.pop("modelName", None)
+            payload.pop("modelMode", None)
 
         custom_personality = AppChatReverse._resolve_custom_personality()
         if custom_personality is not None: