[Bug] Anthropic API Error: Requests blocked by safety systems due to semantic traces from past security work
Bug Description
I keep getting kicked off fable, for no apparent reason. There is the possibility that past work (involving security hardening of my app), triggered defenses. A subagent lane decided to try red-teaming against my local database to ensure it's safe. This however happened he other day, which means the trigger is caused not by current work but activated possibly through semantic traces, picking up on past messages - yet disabling continuing work which has nothign to do with security.
I suggest a window is used to only trigger on recent/current work. This would allow the user to use opus for a few messages before switching back to fable for non-security related tasks.
Environment Info
- Platform: darwin
- Terminal: Orca
- Version: 2.1.240
- Feedback ID: 48573dbd-13bd-476d-b9a9-82a0273b9273
This issue has 1 comment on GitHub. Read the full discussion on GitHub ↗