mirror of
https://github.com/open-webui/open-webui.git
synced 2026-11-05 07:57:06 +00:00
d8ceb45baa
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 |
||
|---|---|---|
| .. | ||
| data | ||
| open_webui | ||
| .dockerignore | ||
| .gitignore | ||
| dev.sh | ||
| requirements-min.txt | ||
| requirements.txt | ||
| start.sh | ||
| start_windows.bat | ||