False-positive [bio] safety block sticks for rest of session

Environment: Claude Code desktop app, model claude-sonnet-5

A session doing routine atmospheric science work (blowing-snow event analysis, wind speed / temperature comparisons across "BEE" events, no biological content whatsoever) started returning this error on every subsequent turn:

API Error: Sonnet 5 can't help with this. Start a new session to continue.
Learn more: https://www.anthropic.com/legal/aup
Details: `[bio]`

The first blocked turn came right after the assistant posted a benign markdown table of wind-speed statistics (U10, u*, temperature) per event. All ~10+ turns afterward were also blocked with the same [bio] tag, including completely unrelated follow-ups like "are you there?" and "is humidity also different??" -- suggesting the block is sticky for the rest of the session once tripped, rather than re-evaluating each turn's actual content.

Session ID: local_61cd610f-9038-4299-a671-36f96e5559f8
First blocked Request ID: req_011Cf2ra25ExBxo9ka5igYxQ
Subsequent blocked Request IDs include:
req_011Cf2sBSPCvUJxGicqcsaxc, req_011Cf2sF5gcyhusiM6EUMNTC,
req_011Cf2sJkLtbMTan4vQMQzrq, req_011Cf2sMSd7aAjAZvA59au6x,
req_011Cf2sVKLnwXZEojGoaYCW2, req_011Cf2saSnAC8L2ALSEjUuRp,
req_011Cf2siALAv1JrFeRv7EJrR, req_011Cf2szRipzbTCF3fXacuXF,
req_011Cf2tZyqn6CST5Ut2K9ym9

Reporting as a likely false positive -- no biological content is present anywhere in the visible transcript around the trigger point.

The report names no source files or tests. Start by reproducing the session using the supplied session and request IDs, then compare the first blocked request with later unrelated turns. Done means benign follow-up turns are evaluated independently rather than remaining blocked for the session.

Vom Indexierungsmodell aus dem Issue-Text verfasst.