I suspect this issue is triggered when users have their own custom instruction files (like agents.md or claude.md) that overlap with Traycer-specific instructions. The sheer volume of duplicated instructions is likely causing context overload.
When this heavy context is passed back and forth between sub-agents, the friction between the user's custom instructions and the app's default instructions might be causing the AI to overthink or hesitate.
On a positive note, the overall workflow feels noticeably faster in Traycer compared to the Claude Code app. I'm not sure if it's just my perception, but even the API response times feel quicker through Traycer which is honestly a bit unbelievable. Great work on that front!