[Bug] Safety classifier false-positives on legitimate SRE/DevOps operations, forcing model downgrade

Status Open
Reported on v2.1.211
Maintainer reply None cached
Activity 0 comments · opened Jul 20, 2026

Bug Description
False-positive report — safety classifier repeatedly flagging routine infrastructure/DevOps work. I use Claude Code as a team-lead/orchestrator for my own Rust market-making system (mm-execution) running on my own infrastructure (home Kubernetes dev cluster + my own AWS instances). Over a multi-hour session doing legitimate SRE/DevOps work, the safeguards repeatedly flagged benign requests and force-switched me from Fable 5 to Opus 4.8, breaking my workflow. The flagged work is standard operations on systems I own and am authorized to administer: - Fault-injection / chaos drills: severing a pod's network egress and resetting established connections to verify the system self-heals after a network outage (reproducing a real incident). - Server migration prep: auditing backups and handling my own service credentials/secrets to move my system between my own servers. - Kubernetes and cloud operations: applying network policies, restarting services, inspecting pods. - Notification setup: configuring a Telegram bot token for my own alerting. None of this involves attacking third parties, unauthorized access, malware, or anything unlawful — it is ordinary trading-system infrastructure work on my own assets. Please tune the classifier so this kind of SRE/DevOps vocabulary (network severing, credential handling, chaos testing) does not trigger a model switch. The false positives make Fable 5 effectively unusable for infrastructure work.

Environment Info

  • Platform: darwin
  • Terminal: tmux
  • Version: 2.1.211
  • Feedback ID: 80d853ef-7fff-4134-9d2a-267f7f71ea80

Errors

[]

View original on GitHub ↗