Hey, thanks for the report and the request ID, we confirmed this on our end.
This is a known issue with the Composer models where they hallucinate CJK text and unrelated project content. Important note: this is not a data leak from another user. The model generates fictional content like project names, file paths, and commands that looks real but is not from anyone’s actual codebase. It’s hallucinated training data, not cross-session contamination.
We’ve seen the same pattern from several users recently:
- Composer 2 outputs foreign language with completely unrelated code
- Composer suddenly switching to Chinese
The team is aware and tracking this. Your request ID helps us gauge the scope.
Workarounds for now:
- Start a new chat if it happens
- Switch to a non-Composer model like Claude or GPT to avoid it entirely
Let me know if you keep running into this.