refac

backend/open_webui/env.py

@@ -630,6 +630,13 @@ else:
         CHAT_RESPONSE_MAX_TOOL_CALL_RETRIES = 30
 
 
+# WARNING: Experimental. Only enable if your upstream Responses API endpoint
+# supports stateful sessions (i.e. server-side response storage with
+# previous_response_id anchoring). Most proxies and third-party endpoints
+# are stateless and will break if this is enabled.
+ENABLE_RESPONSES_API_STATEFUL = os.environ.get('ENABLE_RESPONSES_API_STATEFUL', 'False').lower() == 'true'
+
+
 CHAT_STREAM_RESPONSE_CHUNK_MAX_BUFFER_SIZE = os.environ.get('CHAT_STREAM_RESPONSE_CHUNK_MAX_BUFFER_SIZE', '')
 
 if CHAT_STREAM_RESPONSE_CHUNK_MAX_BUFFER_SIZE == '':

backend/open_webui/routers/openai.py

@@ -779,6 +779,31 @@ def convert_to_azure_payload(url, payload: dict, api_version: str):
     return url, payload
 
 
+# Fields accepted by the Responses API for each input item type.
+RESPONSES_ALLOWED_FIELDS: dict[str, set[str]] = {
+    'message': {'type', 'role', 'content'},
+    'function_call': {'type', 'call_id', 'name', 'arguments', 'id'},
+    'function_call_output': {'type', 'call_id', 'output'},
+}
+
+
+def _normalize_stored_item(item: dict) -> dict:
+    """Strip local-only fields from a stored output item before replaying it.
+
+    Open WebUI stores extra bookkeeping fields (``id``, ``status``,
+    ``started_at``, ``ended_at``, ``duration``, ``_tag_type``,
+    ``attributes``, ``summary``, etc.) that the Responses API does
+    not accept.  This helper returns a copy containing only the
+    fields the API understands.
+    """
+    item_type = item.get('type', '')
+    allowed = RESPONSES_ALLOWED_FIELDS.get(item_type)
+    if allowed is None:
+        # Unknown type — pass through as-is (e.g. reasoning, extension items).
+        return item
+    return {k: v for k, v in item.items() if k in allowed}
+
+
 def convert_to_responses_payload(payload: dict) -> dict:
     """
     Convert Chat Completions payload to Responses API format.
@@ -798,7 +823,7 @@ def convert_to_responses_payload(payload: dict) -> dict:
         # Check for stored output items (from previous Responses API turn)
         stored_output = msg.get('output')
         if stored_output and isinstance(stored_output, list):
-            input_items.extend(stored_output)
+            input_items.extend(_normalize_stored_item(item) for item in stored_output)
             continue
 
         if role == 'system':
@@ -861,6 +886,12 @@ def convert_to_responses_payload(payload: dict) -> dict:
 
     responses_payload = {**payload, 'input': input_items}
 
+    # Forward previous_response_id when the middleware has set it
+    # (only used when ENABLE_RESPONSES_API_STATEFUL is enabled).
+    previous_response_id = responses_payload.pop('previous_response_id', None)
+    if previous_response_id:
+        responses_payload['previous_response_id'] = previous_response_id
+
     if system_content:
         responses_payload['instructions'] = system_content
 

backend/open_webui/utils/middleware.py

@@ -136,6 +136,7 @@ from open_webui.env import (
     ENABLE_FORWARD_USER_INFO_HEADERS,
     FORWARD_SESSION_INFO_HEADER_CHAT_ID,
     FORWARD_SESSION_INFO_HEADER_MESSAGE_ID,
+    ENABLE_RESPONSES_API_STATEFUL,
 )
 from open_webui.utils.headers import include_user_info_headers
 from open_webui.constants import TASKS
@@ -850,7 +851,11 @@ def handle_responses_streaming_event(
                 if item.get('type') == 'reasoning' and item.get('status') != 'completed':
                     item['status'] = 'completed'
 
-        return new_output, {'usage': response_data.get('usage'), 'done': True}
+        return new_output, {
+            'usage': response_data.get('usage'),
+            'done': True,
+            'response_id': response_data.get('id'),
+        }
 
     elif event_type == 'response.in_progress':
         # State Machine Event: In Progress
@@ -3393,6 +3398,7 @@ async def streaming_chat_response_handler(response, ctx):
 
             usage = None
             prior_output = []
+            last_response_id = None
 
             def full_output():
                 return prior_output + output if prior_output else output
@@ -3431,6 +3437,7 @@ async def streaming_chat_response_handler(response, ctx):
                     nonlocal usage
                     nonlocal output
                     nonlocal prior_output
+                    nonlocal last_response_id
 
                     response_tool_calls = []
 
@@ -3519,6 +3526,10 @@ async def streaming_chat_response_handler(response, ctx):
                                     # calls. The outer middleware manages the
                                     # actual completion signal.
                                     if response_metadata:
+                                        if ENABLE_RESPONSES_API_STATEFUL:
+                                            response_id = response_metadata.pop('response_id', None)
+                                            if response_id:
+                                                last_response_id = response_id
                                         processed_data.update(response_metadata)
                                         processed_data.pop('done', None)
 
@@ -3927,10 +3938,12 @@ async def streaming_chat_response_handler(response, ctx):
 
                     # Responses API path: extract function_call items from output
                     if not response_tool_calls and output:
-                        # Collect call_ids that already have results
+                        # Collect call_ids that already have results,
+                        # including those from prior_output so we don't
+                        # re-process tool calls from a previous turn.
                         handled_call_ids = {
                             item.get('call_id')
-                            for item in output
+                            for item in (prior_output + output)
                             if item.get('type') == 'function_call_output'
                         }
                         responses_api_tool_calls = []
@@ -4249,11 +4262,20 @@ async def streaming_chat_response_handler(response, ctx):
                             **form_data,
                             'model': model_id,
                             'stream': True,
-                            'messages': [
+                        }
+
+                        if ENABLE_RESPONSES_API_STATEFUL and last_response_id:
+                            system_message = get_system_message(form_data['messages'])
+                            new_form_data['messages'] = (
+                                ([system_message] if system_message else [])
+                                + convert_output_to_messages(output, raw=True)
+                            )
+                            new_form_data['previous_response_id'] = last_response_id
+                        else:
+                            new_form_data['messages'] = [
                                 *form_data['messages'],
                                 *convert_output_to_messages(output, raw=True),
-                            ],
-                        }
+                            ]
 
                         res = await generate_chat_completion(
                             request,
@@ -4270,6 +4292,20 @@ async def streaming_chat_response_handler(response, ctx):
                             # ensures the UI shows tool history during
                             # streaming.
                             prior_output = list(output)
+                            # Trim the trailing empty placeholder message
+                            # so it doesn't persist as a ghost item once
+                            # the new stream produces real content.
+                            if (
+                                prior_output
+                                and prior_output[-1].get('type') == 'message'
+                                and prior_output[-1].get('status') == 'in_progress'
+                            ):
+                                msg_parts = prior_output[-1].get('content', [])
+                                if (
+                                    not msg_parts
+                                    or (len(msg_parts) == 1 and not msg_parts[0].get('text', '').strip())
+                                ):
+                                    prior_output.pop()
                             output = []
                             await stream_body_handler(res, new_form_data)
                             output[:0] = prior_output