[Bug] Opus 5.5 false positive on harmless prompt causes persistent session failures via transcript contamination
Bug Description
Opus 5.5 false positive [reasoning_extraction] on a harmless prompt; the flagged phrase keeps killing sessions via the transcript
Claude Code 2.1.283 (VS Code + CLI), Windows 10, claude-opus-5-5, 2026-09-26.
Minimal repro (transcript attached to the Feedback ID below): a fresh session in an empty folder, one ordinary work prompt (asking for a draft in Russian before a Swedish message) → "Opus 5.5's safeguards flagged this message". Request req_011CfSF3qMFSbZEuGKXem4iG, Message msg_011CfSF3qpXdLsQJDWy2b9Py.
Impact: the phrase stays in the local .jsonl transcript, so any later session that reads it gets flagged too; several work sessions were lost this way. Sonnet 5 is not affected. Claude Code blocks redacting my own transcripts ("Session Transcript Tampering").
Requests: review the false positive; show what triggered the flag; allow recovering the session instead of a hard stop; provide a supported way to remove a flagged message from a transcript.
Environment Info
- Platform: win32
- Terminal: null
- Version: 2.1.283
- Feedback ID: f89f051b-0a05-4eb4-8441-4afb5e7d4eee
Errors
[]This issue has 2 comments on GitHub. Read the full discussion on GitHub ↗