[Bug] Claude 3.5 Sonata safeguards over-blocking routine code review tasks on user's own repositories

Bug Description
Fable 5 safeguards have blocked every coding session I've attempted for 6 consecutive days, and my access expires tomorrow — I've effectively never been able to use the model. Latest example, today: three variants of a routine PR-review prompt on my own repo (smeanalytica/conversa, a WhatsApp SaaS for real estate agencies), all flagged. Prompts: (1) "please audit and review the prs i have on gh", (2) a fully specified version asking to summarize each PR, check diffs for logic errors, edge cases, and regressions around webhooks, appointment booking, and RLS policies, with merge verdicts, (3) same with repo path softened. No security research, no exploit work — standard code review on my own project. Suspected trigger: security-adjacent vocabulary (audit, RLS, safe to merge). Same tasks run fine on Opus 4.8. Request: recalibrate the classifier for routine code review, and given the model was unusable for my entire access window, consider extending the trial period.

Environment Info

• Platform: darwin
• Terminal: antigravity
• Version: 2.1.201
• Feedback ID: 7e51122f-9016-4ba7-8beb-2292ff93364e

Errors

[]