Conversation permanently blocked after credential-related context — rewind doesn't help
Bug Description
When Claude Code guides the user through setting up a credentials file (e.g., creating a JSON config for an API service), and the conversation context touches on passwords/credentials, the safety guardrail system permanently blocks the entire conversation. Every subsequent message gets blocked, regardless of content.
Steps to Reproduce
- Claude Code suggests creating a credentials config file (in my case,
jqdata_credentials.jsonfor a Chinese stock data SDK) - Claude creates the file template and instructs the user to fill in username/password locally
- User sends a follow-up message (e.g., "cancel using this service, delete the credentials file")
- Message is blocked by safety guardrails
- User rewinds the conversation and tries again with different wording
- Still blocked — any message, no matter how innocuous, gets blocked
- Even messages like "continue" or completely unrelated requests are blocked
- The conversation is permanently unusable; the only option is to start a new session
Expected Behavior
- Block a single problematic message if necessary, not the entire conversation
- Allow the user to rewind and continue with different input
- At minimum, provide a clear explanation of what triggered the block
- Never block messages that have nothing to do with credentials (e.g., "cancel using this service")
Actual Behavior
The safety classifier appears to be contaminated by the conversation history. Once credential-related tokens exist in the context, every subsequent message is flagged as high-risk regardless of its actual content. Rewinding does not help — the context pollution persists.
Impact
- Paying customer unable to use the product they paid for
- Hours of conversation context and work progress lost
- User was following Claude's own instructions when the block was triggered — Claude suggested the credential setup flow, then blocked the user for participating in it
- Forces starting a completely new session, losing all prior context
Environment
- Claude Code CLI
- macOS
- Max plan (paying customer)
Additional Context
The irony: Claude itself initiated the credential file creation flow. It created the template, told the user how to fill it in locally, and even offered to open a text editor. The user never sent any actual passwords in the chat. The mere topic of credentials in the conversation history was enough to permanently lock it.
This issue has 4 comments on GitHub. Read the full discussion on GitHub ↗