Hey, thanks for the detailed report. Yes, this is really frustrating, the AI shouldn’t have run meteor reset after your explicit ban on destructive actions. That means the model ignored clear safety rules.
What can help avoid this in the future:
- Use “Rules for AI” in Cursor settings to set global rules, for example: “NEVER run destructive database operations (meteor reset, DROP, DELETE) without explicit confirmation.”
- Use the @ symbol to reference specific files instead of granting broad system access.
- For complex prompts like your BMAD orchestrator, break the task into smaller, controlled steps.
The model’s apology shows it realized the violation afterward, but that doesn’t restore data.
Have you been able to recover merchant data from backups?