Claude Code session poisoning after Usage Policy blocks — --continue broken, recovery, and prevention for agents.
One false positive can brick a Claude Code session — even hi gets blocked. --continue replays the poisoned transcript forever until you mv the jsonl file.
How poisoning works
Classifier scores accumulated context each turn. After a block, the transcript contains refusal markers that re-trigger checks.
Recovery
/clear, new session without --continue, or move ~/.claude/projects/.../*.jsonl aside. Anthropic docs confirm resume re-sends triggering content.
Agent design
Short-lived sessions per task. External state in files, not chat. Uncensored API for security-heavy sub-agents.
Try uncensored AI free
No filters, no lecture, no training on your chats. Start in under a minute.
Poisoning is the #1 reason security automation on Claude Code feels unreliable.
FAQ
Can /rewind fix poisoned sessions?
Sometimes, if you rewind before the triggering turn. Often /clear is faster.