fixed injection watcher

backend/app/api/routes/chat.py

@@ -48,8 +48,15 @@ async def send_message(
     if not service.get_session(session_id):
         raise HTTPException(status_code=404, detail="Session not found")
 
+    # Сохраняем user до стрима: иначе при обрыве SSE сообщение не попадает в БД.
+    service.save_user_message(session_id, payload.content)
+
     async def event_stream():
-        async for chunk in service.stream_response(session_id, payload.content):
+        async for chunk in service.stream_response(
+            session_id,
+            payload.content,
+            user_message_saved=True,
+        ):
             yield chunk
 
     return StreamingResponse(

backend/app/chat/service.py

@@ -172,13 +172,23 @@ class ChatService:
         self.db.refresh(message)
         return message
 
-    async def stream_response(self, session_id: int, user_text: str) -> AsyncIterator[str]:
+    def save_user_message(self, session_id: int, user_text: str) -> None:
+        self._save_message(session_id, "user", user_text)
+
+    async def stream_response(
+        self,
+        session_id: int,
+        user_text: str,
+        *,
+        user_message_saved: bool = False,
+    ) -> AsyncIterator[str]:
         session = self.get_session(session_id)
         if not session:
             yield self._sse("error", {"message": "Session not found"})
             return
 
-        self._save_message(session_id, "user", user_text)
+        if not user_message_saved:
+            self._save_message(session_id, "user", user_text)
         yield self._sse("status", {"phase": "preparing"})
         t0 = time.monotonic()
         messages = await asyncio.to_thread(_build_messages_for_session, session_id)
@@ -299,10 +309,15 @@ class ChatService:
             if not final_content and reasoning:
                 final_content = reasoning.strip()
             if not final_content and all_tool_notices:
-                # Notices уже ушли в SSE event: notice; здесь только финальный текст в БД.
                 final_content = "\n\n".join(all_tool_notices)
+                yield self._sse("token", {"content": final_content})
             if not final_content and tools_executed:
-                retry = await self.llm.complete(messages, tools=None, temperature=0.4)
+                retry = await self.llm.complete(
+                    messages,
+                    tools=None,
+                    temperature=0.4,
+                    visible_reply=True,
+                )
                 final_content = (retry.get("content") or "").strip()
                 if final_content:
                     yield self._sse("token", {"content": final_content})

backend/app/llm/client.py

@@ -171,6 +171,7 @@ class LLMClient:
         temperature: float = 0.7,
         model: str | None = None,
         for_extraction: bool = False,
+        visible_reply: bool = False,
     ) -> dict[str, Any]:
         use_tools = bool(tools) and self.tools_enabled and not for_extraction
         kwargs: dict[str, Any] = {
@@ -198,7 +199,7 @@ class LLMClient:
                 reasoning = str(value)
                 break
 
-        if not content and reasoning:
+        if not content and reasoning and not visible_reply:
             content = reasoning
 
         result: dict[str, Any] = {