Repository navigation
Do not send reasoning traces and tool calls made during reasoning when switching from reasoning model to non reasoning model #3035
Description
Activity
This issue might be a duplicate of existing issues. Please check:
- Add a way to delete message parts and messages in the sdk #3033: Add a way to delete message parts and messages in the sdk - addresses the same concern about tool calls and reasoning parts confusing models when switching contexts. The author mentions wanting to remove tool parts from plan messages when switching from GPT5 to Sonnet.
Feel free to ignore if none of these address your specific case.
Metadata
Metadata
Assignees
Labels
No labels
Reasoning traces and tool calls seem to confuse non reasoning models a lot. For example if you switch from GPT5 to Sonnet 4.5 the model becomes unusable because it gets lost in all the reasoning context given by the previous model
Probably we should strip the reasoning parts after the message completes even if no model switch is made imo
Instead only the actual final message should be kept