[Bug] Claude fails to verify state before executing irreversible/billable operations in long sessions

Bug Description
Subject: Model caused unauthorized spend and rework during a long Claude Code session (requesting credit consideration) During an extended development session, Claude (Fable 5) made three failures of the same class — acting on irreversible or billable operations without verifying state or creating backups: 1. Unauthorized API spend. I explicitly paused a batch LLM job (17.5k-call contextual-retrieval generation on AWS Bedrock). The model confirmed it was stopped, but an orphaned copy of the process survived a session restart and kept making Bedrock Haiku calls for several hours before the model noticed. That's roughly 2,900 unauthorized calls (~4M input tokens, est. $4–6) billed to my AWS account against my explicit instruction. 2. Destroyed derived data without backup. The model launched a full re-index that overwrote the existing production search index (1.5 hours of compute to build) without taking a copy first, leaving no rollback path if the new index underperforms — an avoidable ~1.5–3 hours of compute/wall-clock rework risk. 3. Shipped a breaking UI change to production. A layout refactor was verified only implicitly on desktop; the mobile view was completely broken (content collapsed into a 40px column) on the live tester-facing site for several hours until I reported it. Total quantifiable impact: ~$5 in unauthorized Bedrock spend, several hours of lost compute and my own time spent catching and reporting these errors, and degraded production availability during a closed tester period. The common root cause is the model's weak default reflex around irreversible/billable actions (verify process state, back up before overwrite, verify on all form factors before deploy). I'd appreciate credit consideration for the unauthorized spend and the rework time.

Environment Info

• Platform: linux
• Terminal: vscode
• Version: 2.1.220
• Feedback ID: 7ee922ac-8751-4703-8190-cb6c76a09368

Errors

[]