[Bug] Classifier triggers on accumulated session context, requiring new sessions per file

Summary

Claude Code's safety classifier flags a routine correctness / crash-safety review
of my OWN open-source C code (public repo ariannamethod/yent) as if it were
exploit research, and force-downgrades the paid Fable 5 session to Opus 4.8
mid-task. On 2026-07-04 this happened across 7+ consecutive sessions.

Repro

1. Fable 5 session. Ask it to review a C source file of your own for correctness:

bounds checks, allocation-size validation, return-value checks. Read-only, no
shell commands.

1. One file completes cleanly.
2. Continue to a second file in the SAME session.
3. Classifier fires — "Session paused / Switch to Opus 4.8" — with no offending

message. It triggers on accumulated session context, not any single prompt.

Expected

A developer reviewing the safety of their own code is not treated as an attacker,
and the model I selected and paid for is not silently switched mid-task.

Actual

Each file now costs a separate fresh session. A task that should take one pass is
stretched across many, and paid Fable work is redirected to a different model.

Key point

The flag text itself says it is "intentionally broad right now and may flag safe
and routine coding, cybersecurity, or biology work." This is exactly that
acknowledged false-positive class.

Environment

• Platform: darwin, Apple_Terminal, v2.1.201
• Feedback ID: 1350def9-a4de-4c98-a729-cb8a53c723f7