scriptoriumadmin/open-webui
1
0
Fork 0
mirror of https://github.com/open-webui/open-webui.git synced 2026-11-05 07:57:06 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 9

fix: stop replaying encrypted reasoning_details to a different model after a mid-chat model switch

Browse source 
Switching models mid-chat after a reasoning model responded breaks the chat permanently: stored reasoning_details of type reasoning.encrypted (or carrying a signature) are cryptographically bound to the model that produced them, and the new model's provider rejects every subsequent request, including retries and forks.

Each assistant message already records which model produced it. On request build, when that model differs from the current target, the bound entries are now dropped from its reasoning_details while plain text and summary entries are kept, so the new model still sees what the previous one was thinking. Same-model requests are untouched, since Anthropic-family models require their signed reasoning blocks back and replaying reasoning saves tokens.

The filter sits in process_messages_with_output, the single point every outbound message list passes through, so persisted chats, continue-generation and temporary chats are all covered. Filtering at request time rather than stripping at save keeps the stored data intact: switching back to the original model replays its reasoning again. Messages saved before this change have no model recorded and are left as-is.

Fixes #28240
This commit is contained in:
Classic298 2026-08-07 12:57:10 +02:00
parent 29541cbb52
commit d8ceb45baa
2 changed files with 40 additions and 15 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

53
backend/open_webui/utils/middleware.py
View file

@ -1980,6 +1980,9 @@ async def convert_url_images_to_base64(form_data, user=None):
return form_data
PERSISTED_MESSAGE_KEYS = ('id', 'role', 'content', 'output', 'files', 'contextSummary', 'usage', 'model')
async def load_messages_from_db(chat_id: str, message_id: str) -> Optional[list[dict]]:
"""
Load the message chain from DB up to message_id,
@ -1993,10 +1996,7 @@ async def load_messages_from_db(chat_id: str, message_id: str) -> Optional[list[
if not db_messages:
return None
return [
{k: v for k, v in msg.items() if k in ('id', 'role', 'content', 'output', 'files', 'contextSummary', 'usage')}
for msg in db_messages
]
return [{k: v for k, v in msg.items() if k in PERSISTED_MESSAGE_KEYS} for msg in db_messages]
def get_reasoning_format(model: dict) -> str | None:
@ -2016,8 +2016,32 @@ def get_reasoning_format(model: dict) -> str | None:
return None
def strip_bound_reasoning_details(output: list) -> list:
"""Drop model-bound reasoning_details entries; providers reject them replayed to another model (#28240)."""
stripped = []
for item in output:
details = item.get('reasoning_details')
if item.get('type') == 'reasoning' and details:
if not isinstance(details, list):
details = [details]
portable = [
detail
for detail in details
if not (
isinstance(detail, dict) and (detail.get('type') == 'reasoning.encrypted' or 'signature' in detail)
)
]
if len(portable) != len(details):
item = {k: v for k, v in item.items() if k != 'reasoning_details'}
if portable:
item['reasoning_details'] = portable
stripped.append(item)
return stripped
def process_messages_with_output(
messages: list[dict],
model_id: str,
reasoning_format: str | None = None,
) -> list[dict]:
"""
@ -2025,14 +2049,20 @@ def process_messages_with_output(
For assistant messages with 'output' field, produces properly formatted
OpenAI-style messages (tool_calls + tool results). Strips 'output' before LLM.
Model-bound reasoning_details from a different model are dropped.
"""
processed = []
for message in messages:
if message.get('role') == 'assistant' and message.get('output'):
output = message['output']
source_model = message.get('model')
if source_model and source_model != model_id and isinstance(output, list):
output = strip_bound_reasoning_details(output)
# Use output items for clean OpenAI-format messages
output_messages = convert_output_to_messages(
message['output'],
output,
raw=True,
reasoning_format=reasoning_format,
flatten_tool_images=True,
@ -2041,8 +2071,8 @@ def process_messages_with_output(
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'}
# Strip 'output' and 'model' before adding (LLM shouldn't see them)
clean_message = {k: v for k, v in message.items() if k not in ('output', 'model')}
processed.append(clean_message)
return processed
@ -2249,13 +2279,7 @@ async def process_chat_payload(request, form_data, user, metadata, model):
if assistant_message_id:
assistant_message = await Chats.get_message_by_id_and_message_id(chat_id, assistant_message_id)
if assistant_message and (assistant_message.get('content') or assistant_message.get('output')):
db_messages.append(
{
k: v
for k, v in assistant_message.items()
if k in ('id', 'role', 'content', 'output', 'files', 'contextSummary', 'usage')
}
)
db_messages.append({k: v for k, v in assistant_message.items() if k in PERSISTED_MESSAGE_KEYS})
system_message = get_system_message(form_data.get('messages', []))
form_data['messages'] = [system_message, *db_messages] if system_message else db_messages
@ -2323,6 +2347,7 @@ async def process_chat_payload(request, form_data, user, metadata, model):
# Process messages with OR-aligned output items for clean LLM messages
form_data['messages'] = process_messages_with_output(
form_data.get('messages', []),
model_id=model['id'],
reasoning_format=get_reasoning_format(model),
)
form_data['messages'] = sanitize_tool_pairs(form_data['messages'])

2
src/lib/components/chat/Chat.svelte
View file

@ -3220,7 +3220,7 @@
);
if (message.output && message.role === 'assistant') {
return { role: message.role, output: message.output };
return { role: message.role, model: message.model, output: message.output };
}
if (message.role === 'user' && imageFiles.length > 0) {