fix: respect content edits and deleted blocks in output

Author: Algorithm5838

backend/open_webui/utils/middleware.py

@@ -98,6 +98,7 @@
     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 (
@@ -2164,16 +2165,52 @@ def process_messages_with_output(messages: list[dict]) -> list[dict]:
 
     For assistant messages with 'output' field, produces properly formatted
     OpenAI-style messages (tool_calls + tool results). Strips 'output' before LLM.
+    Respects content edits and dropped <details> blocks by filtering output items
+    against the stored content field before conversion.
     """
     processed = []
 
     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)
-            if output_messages:
-                processed.extend(output_messages)
-                continue
+            # normalize; guard against None or list-typed content
+            content = message.get('content', '')
+            if not isinstance(content, str):
+                content = ''
+
+            # Drop output items for <details> blocks removed from content
+            filtered_output = filter_output_by_content(message['output'], content)
+
+            # Split content around <details> blocks; assign each text segment to
+            # the corresponding message item in order → preserves pre/post-tool placement
+            text_segs = [
+                s.strip()
+                for s in re.split(r'<details\b[^>]*>.*?</details>', content, flags=re.S)
+            ]
+            seg_idx = 0
+            modified_output = []
+            for item in filtered_output:
+                if item.get('type') == 'message':
+                    text = text_segs[seg_idx] if seg_idx < len(text_segs) else ''
+                    seg_idx += 1
+                    modified_output.append({
+                        **item,
+                        'content': [{'type': 'output_text', 'text': text}],
+                    })
+                else:
+                    modified_output.append(item)
+
+            output_messages = convert_output_to_messages(modified_output, raw=True)
+            # Trailing segments not consumed by message items → append as plain text
+            trailing = '\n'.join(s for s in text_segs[seg_idx:] if s).strip()
+            if trailing:
+                output_messages.append({'role': 'assistant', 'content': trailing})
+            elif not output_messages:
+                # No structured output at all; fall back to plain text from content
+                plain = '\n'.join(s for s in text_segs if s).strip()
+                if plain:
+                    output_messages = [{'role': 'assistant', 'content': plain}]
+            processed.extend(output_messages)
+            continue
 
         # Strip 'output' field before adding (LLM shouldn't see it)
         clean_message = {k: v for k, v in message.items() if k != 'output'}

backend/open_webui/utils/misc.py

@@ -282,6 +282,72 @@ 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. Items with no ID are kept for backward
+    compatibility with content serialized before id= was added.
+    Items whose type is entirely absent from content are dropped wholesale.
+    Legacy: if a type is present with no id= attrs, keep the first N items
+    of that type where N = the number of blocks of that type remaining in
+    content, so individually deleted blocks are respected even without IDs.
+    """
+    if not isinstance(output, list):
+        return []
+    if not isinstance(content, str):
+        content = ''
+
+    present_ids = set(re.findall(r'<details\b[^>]*\bid="([^"]+)"', content))
+    present_types = set(re.findall(r'<details\b[^>]*\btype="([^"]+)"', content))
+
+    # 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',
+    }
+
+    # No <details> blocks: pass through if no structured item has an ID (pre-serialization legacy)
+    if not present_types:
+        if not any(item.get('call_id') or item.get('id')
+                   for item in output if DETAILS_TYPE.get(item.get('type', ''))):
+            return list(output)
+
+    # types with ≥1 id= in content; others treated as legacy (pre-id=)
+    types_with_ids = set(re.findall(r'<details\b[^>]*\btype="([^"]+)"[^>]*\bid="[^"]*"', content))
+    types_with_ids |= set(re.findall(r'<details\b[^>]*\bid="[^"]*"[^>]*\btype="([^"]+)"', content))
+
+    # Legacy (no id=): count remaining blocks per type for positional matching
+    type_block_counts = {}
+    for t in re.findall(r'<details\b[^>]*\btype="([^"]+)"', content):
+        type_block_counts[t] = type_block_counts.get(t, 0) + 1
+
+    filtered = []
+    seen_legacy = {}  # kept count per legacy type
+    for item in output:
+        details_type = DETAILS_TYPE.get(item.get('type', ''))
+        if details_type is None:
+            filtered.append(item)  # non-visual item (e.g. 'message'): always keep
+            continue
+        if details_type not in present_types:
+            continue  # entire type removed from content: drop
+        if details_type not in types_with_ids:
+            # Legacy (no id=): keep first N items, N = block count in content
+            n = type_block_counts.get(details_type, 0)
+            if seen_legacy.get(details_type, 0) < n:
+                filtered.append(item)
+                seen_legacy[details_type] = seen_legacy.get(details_type, 0) + 1
+            continue
+        item_id = item.get('call_id') or item.get('id', '')
+        if not item_id or item_id in present_ids:
+            filtered.append(item)  # no ID (old format) or ID still present: keep
+        # else: ID not found in content → user deleted this 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: