fix: respect content edits for assistant messages with output

Author: Algorithm5838

backend/open_webui/utils/middleware.py

@@ -97,6 +97,7 @@
     convert_logit_bias_input_to_json,
     get_content_from_message,
     convert_output_to_messages,
+    filter_output_by_content,
     strip_empty_content_blocks,
 )
 from open_webui.utils.tools import (
@@ -475,9 +476,9 @@ def serialize_output(output: list) -> str:
             )
 
             if status == 'completed' or duration is not None or not is_last_item:
-                content = f'{content}<details type="reasoning" done="true" duration="{duration or 0}">\n<summary>Thought for {duration or 0} seconds</summary>\n{display}\n</details>\n'
+                content = f'{content}<details type="reasoning" done="true" id="{item.get("id", "")}" duration="{duration or 0}">\n<summary>Thought for {duration or 0} seconds</summary>\n{display}\n</details>\n'
             else:
-                content = f'{content}<details type="reasoning" done="false">\n<summary>Thinking…</summary>\n{display}\n</details>\n'
+                content = f'{content}<details type="reasoning" done="false" id="{item.get("id", "")}">\n<summary>Thinking…</summary>\n{display}\n</details>\n'
 
         elif item_type == 'open_webui:code_interpreter':
             content_stripped, original_whitespace = split_content_and_whitespace(content)
@@ -513,9 +514,9 @@ def serialize_output(output: list) -> str:
                 output_attr = f' output="{html.escape(output_json)}"'
 
             if status == 'completed' or duration is not None or not is_last_item:
-                content += f'<details type="code_interpreter" done="true" duration="{duration or 0}"{output_attr}>\n<summary>Analyzed</summary>\n{display}\n</details>\n'
+                content += f'<details type="code_interpreter" done="true" id="{item.get("id", "")}" duration="{duration or 0}"{output_attr}>\n<summary>Analyzed</summary>\n{display}\n</details>\n'
             else:
-                content += f'<details type="code_interpreter" done="false"{output_attr}>\n<summary>Analyzing…</summary>\n{display}\n</details>\n'
+                content += f'<details type="code_interpreter" done="false" id="{item.get("id", "")}"{output_attr}>\n<summary>Analyzing…</summary>\n{display}\n</details>\n'
 
     return content.strip()
 
@@ -2040,11 +2041,40 @@ def process_messages_with_output(messages: list[dict]) -> list[dict]:
 
     for message in messages:
         if message.get('role') == 'assistant' and message.get('output'):
-            # Use output items for clean OpenAI-format messages
-            output_messages = convert_output_to_messages(message['output'], raw=True)
+            # Drop output items for <details> blocks removed from content
+            output = filter_output_by_content(
+                message['output'], message.get('content', '')
+            )
+
+            # Use content for text (respects edits), output for structured items
+            content = re.sub(
+                r'<details\b[^>]*>.*?</details>', '',
+                message.get('content', ''), flags=re.S,
+            ).strip()
+            non_message_items = [
+                i for i in output if i.get('type') != 'message'
+            ]
+            output_messages = convert_output_to_messages(non_message_items, raw=True)
+
             if output_messages:
+                # Prepend edited text to first assistant message
+                for om in output_messages:
+                    if om.get('role') == 'assistant':
+                        om['content'] = (
+                            (content + '\n' + om['content']).strip()
+                            if om.get('content')
+                            else content
+                        )
+                        content = ''
+                        break
+                if content:
+                    output_messages.insert(
+                        0, {'role': 'assistant', 'content': content}
+                    )
                 processed.extend(output_messages)
-                continue
+            elif content:
+                processed.append({'role': 'assistant', 'content': content})
+            continue
 
         # Strip 'output' field before adding (LLM shouldn't see it)
         clean_message = {k: v for k, v in message.items() if k != 'output'}
@@ -2635,10 +2665,6 @@ async def tool_function(**kwargs):
             }
         )
 
-    # Strip empty text content blocks from multimodal messages
-    # to prevent errors from providers like Gemini and Claude
-    form_data['messages'] = strip_empty_content_blocks(form_data.get('messages', []))
-
     return form_data, metadata, events
 
 

backend/open_webui/utils/misc.py

@@ -258,6 +258,59 @@ def flush_pending():
     return messages
 
 
+def filter_output_by_content(output: list, content: str) -> list:
+    """
+    Drop output items whose <details> block was removed from content.
+    Matches by id attribute for all details-backed types.
+    For legacy content that lacks id= on a given type, all items of
+    that type are kept (safe default).
+    """
+    # All IDs present in remaining <details> blocks
+    present_ids = set(re.findall(r'<details\s[^>]*\bid="([^"]+)"', content))
+
+    # Which <details> types exist in content at all
+    types_present = set(
+        m.group(1) for m in re.finditer(r'<details\s[^>]*\btype="(\w+)"', content)
+    )
+
+    # Of those, which carry id= attributes (post-fix content).
+    # Legacy content without id= on a given type → keep all items of that type.
+    types_with_ids = set()
+    for m in re.finditer(r'<details\s[^>]*\btype="(\w+)"[^>]*\bid=', content):
+        types_with_ids.add(m.group(1))
+    for m in re.finditer(r'<details\s[^>]*\bid=[^>]*\btype="(\w+)"', content):
+        types_with_ids.add(m.group(1))
+
+    # Map output item type → <details> type attribute value
+    DETAILS_TYPE = {
+        'function_call': 'tool_calls',
+        'function_call_output': 'tool_calls',
+        'reasoning': 'reasoning',
+        'open_webui:code_interpreter': 'code_interpreter',
+    }
+
+    filtered = []
+    for item in output:
+        item_type = item.get('type', '')
+        details_type = DETAILS_TYPE.get(item_type)
+
+        if details_type is None:
+            # Not a details-backed type (e.g. message) — pass through
+            filtered.append(item)
+        elif details_type not in types_present:
+            pass  # type completely gone from content — drop
+        elif details_type not in types_with_ids:
+            # Legacy content — has <details> but no id= — keep all
+            filtered.append(item)
+        else:
+            item_id = item.get('call_id') or item.get('id', '')
+            if item_id in present_ids:
+                filtered.append(item)
+            # else: user deleted the <details> block — drop
+
+    return filtered
+
+
 def get_last_user_message(messages: list[dict]) -> Optional[str]:
     message = get_last_user_message_item(messages)
     if message is None: