brauliogusmao stops Mike from swallowing its own errors
When the AI fails mid-answer, users now learn why instead of staring at a generic "stream error."
Anyone who has used an AI assistant knows the dead-end moment: the answer stops, and the screen says something useless like "error" with no clue what happened. brauliogusmao traced that failure in this fork. Previously, when the underlying AI service threw a problem, the real explanation was thrown away and replaced with a hardcoded generic message - and on top of that, the interface sometimes failed silently, showing nothing at all.
The fix digs out the actual reason and surfaces it. A second pass goes further, sorting failures into named categories - out of credits, bad credentials, rate-limited, request too long, service overloaded - so the assistant can show a human-readable message tuned to each one rather than a catch-all. The handle owner flags the matching logic as something to harden later, since it leans on the exact wording vendors use.
Spotted something wrong? Or know the PR text has fresher detail than the writeup above?