fix: prevent answer split from late reasoning

Author: Algorithm5838

backend/open_webui/utils/middleware.py

@@ -4322,12 +4322,14 @@ async def queue_pending_delta_data(delta_data: dict, delta_type: str):
                                         reasoning_details = None
 
                                     if reasoning_content or reasoning_details:
+                                        # Merge late reasoning into the pre-message block so the answer isn't split.
+                                        answer_started = any(item.get('type') == 'message' for item in output)
                                         reasoning_item = (
                                             next(
                                                 (item for item in reversed(output) if item.get('type') == 'reasoning'),
                                                 None,
                                             )
-                                            if reasoning_details and not reasoning_content
+                                            if (reasoning_details and not reasoning_content) or answer_started
                                             else None
                                         )
 
@@ -4344,7 +4346,13 @@ async def queue_pending_delta_data(delta_data: dict, delta_type: str):
                                                     'summary': None,
                                                     'started_at': time.time(),
                                                 }
-                                                output.append(reasoning_item)
+                                                if answer_started:
+                                                    insert_at = next(
+                                                        i for i, x in enumerate(output) if x.get('type') == 'message'
+                                                    )
+                                                    output.insert(insert_at, reasoning_item)
+                                                else:
+                                                    output.append(reasoning_item)
                                             else:
                                                 reasoning_item = output[-1]