fix(agent): keep thinking status separate

app/agent/callback/__init__.py

@@ -98,18 +98,27 @@ class StreamingHandler:
             self._thinking_started_at = time.monotonic()
             self._thinking_active = True
             self._thinking_status_text = self._format_thinking_status(0)
-            if self._can_stream():
-                self._buffer = self._thinking_status_text
+            if self._should_buffer_thinking_status():
+                self._buffer = f"{self._thinking_status_text}\n\n"
 
     def thinking_finished(self) -> None:
         """结束模型思考状态，移除尚未被正文替换的状态占位文本。"""
         with self._lock:
             if not self._thinking_active:
                 return
-            if self._buffer == self._thinking_status_text:
-                self._buffer = ""
-            self._thinking_active = False
-            self._thinking_status_text = ""
+            self._remove_thinking_status_locked()
+
+    def _should_buffer_thinking_status(self) -> bool:
+        """判断是否把思考状态写入消息渠道的可编辑正文。"""
+        return self._can_stream()
+
+    def _remove_thinking_status_locked(self) -> None:
+        """从正文缓冲区移除思考状态块，同时结束本轮计时。"""
+        status_text = self._thinking_status_text
+        if status_text and self._buffer.startswith(status_text):
+            self._buffer = self._buffer[len(status_text) :].lstrip("\n")
+        self._thinking_active = False
+        self._thinking_status_text = ""
 
     def _format_thinking_status(self, elapsed_seconds: int) -> str:
         """生成消息渠道可直接展示的思考状态文本。"""
@@ -118,11 +127,13 @@ class StreamingHandler:
     def _refresh_thinking_status(self) -> None:
         """更新可编辑消息中的思考计时文本，保持正文尚未生成时也有反馈。"""
         with self._lock:
-            if not self._thinking_active or not self._can_stream() or self._buffer != self._thinking_status_text:
+            if not self._thinking_active or not self._should_buffer_thinking_status():
                 return
             elapsed = int(max(0.0, time.monotonic() - (self._thinking_started_at or time.monotonic())))
-            self._thinking_status_text = self._format_thinking_status(elapsed)
-            self._buffer = self._thinking_status_text
+            next_status_text = self._format_thinking_status(elapsed)
+            if self._buffer.startswith(self._thinking_status_text):
+                self._buffer = next_status_text + self._buffer[len(self._thinking_status_text) :]
+                self._thinking_status_text = next_status_text
 
     def set_dispatch_policy(self, allow_dispatch_without_context: bool = False) -> None:
         """
@@ -139,10 +150,7 @@ class StreamingHandler:
         with self._lock:
             emitted = token or ""
             if emitted and self._thinking_active:
-                if self._buffer == self._thinking_status_text:
-                    self._buffer = ""
-                self._thinking_active = False
-                self._thinking_status_text = ""
+                self._remove_thinking_status_locked()
 
             if self._pending_tool_stats:
                 if self._live_tool_summary:

app/agent/web.py

@@ -19,6 +19,10 @@ class _WebAgentStreamingHandlerMixin:
     _lock: Lock
     _pending_tool_stats: dict[str, int]
 
+    def _should_buffer_thinking_status(self) -> bool:
+        """WebAgent 通过结构化 SSE 展示思考状态，不把它写入正文。"""
+        return False
+
     def __init__(
         self,
         on_emit: Callable[[str], None],

tests/test_agent_streaming_lifecycle.py

@@ -5,6 +5,7 @@ import asyncio
 import pytest
 
 from app.agent.callback import StreamingHandler
+from app.agent.web import _get_web_agent_streaming_handler_type
 from app.schemas.types import NotificationChannel
 
 
@@ -16,6 +17,7 @@ def test_thinking_status_is_replaced_by_first_answer_token() -> None:
     handler.thinking_started()
     assert handler._thinking_active is True
     assert "思考中" in handler._buffer
+    assert handler._buffer.endswith("\n\n")
 
     handler.emit("最终答案")
 
@@ -23,6 +25,38 @@ def test_thinking_status_is_replaced_by_first_answer_token() -> None:
     assert handler._buffer == "最终答案"
 
 
+def test_thinking_status_stays_separate_from_message_channel_tool_summary() -> None:
+    """消息渠道的思考状态应与工具统计保持独立段落，并在正文到达时移除。"""
+    handler = StreamingHandler()
+    handler._can_stream = lambda: True
+    handler._streaming_enabled = True
+
+    handler.thinking_started()
+    handler.record_tool_call("read_file", tool_kwargs={"file_path": "README.md"})
+
+    assert handler._buffer.startswith("🤔 思考中 · 已用时 0 秒\n\n")
+    assert "\n\n（读取了 1 个文件）\n\n" in handler._buffer
+
+    handler.emit("最终答案")
+
+    assert "思考中" not in handler._buffer
+    assert handler._buffer.endswith("\n\n最终答案")
+
+
+def test_web_agent_thinking_status_is_only_a_structured_event() -> None:
+    """WebAgent 的思考状态不能进入正文，只能通过 SSE 结构化事件展示。"""
+    events: list[dict[str, object]] = []
+    handler = _get_web_agent_streaming_handler_type()(lambda _text: None, events.append)
+
+    handler.thinking_started()
+    handler.thinking_finished()
+
+    assert handler._buffer == ""
+    assert events[0]["type"] == "thinking"
+    assert events[0]["status"] == "running"
+    assert events[1] == {"type": "thinking", "status": "done"}
+
+
 @pytest.mark.asyncio
 async def test_repeated_streaming_start_retains_previous_flush_owner() -> None:
     """重复启动必须先等待旧 flush owner 结束，再发布新一轮上下文。"""