[Bug] False positive reasoning_extraction flag when searching local Claude Code session transcripts

Bug Description
False positive report — [reasoning_extraction] flagged twice in a normal Claude Code session. Request IDs: req_011CeEVsnV4PiFuaxeTSCkis and req_011CeEV1XegwFKg6VKabbMrE (2026-08-20). Context: an ordinary software session (ESP32 course pages, fixing link-minting bugs in my own recording tool). Mid-session I asked Claude to find an incident from earlier today in my OWN local Claude Code session transcripts (~/.claude/projects/*.jsonl). Grepping those files put prior assistant output into context, and the distillation classifier apparently read that as thinking-extraction. I was retrieving my own work history on my own machine — a documented Claude Code feature area (session transcripts), not an attempt to extract model reasoning. Impact: Fable 5 errored out twice mid-task; nothing in the conversation was about biology, cybersecurity, or LLM internals. Suggestion: whitelist or down-weight reads of the user's own ~/.claude/projects transcripts in Claude Code — searching your own session history is routine and currently indistinguishable from an attack to this classifier.

Environment Info

• Platform: darwin
• Terminal: Apple_Terminal
• Version: 2.1.236
• Feedback ID: 7c24655c-5ab5-44fb-97c8-c873537807af

Errors

[]