Opus 5.5 [bio] classifier flags benign in-silico protein-engineering session; blocks file-listing output and /feedback returns 403

Summary

In a project directory for an in-silico protein-engineering methods
comparison, Opus 5.5 responses are stopped mid-output by the safety
classifier and the session falls back to Opus 5. The dialog reads:

Opus 5.5's safeguards flagged this session ... Details: [bio]

Reproduced 3/3: each time I re-selected Opus 5.5 with /model, the very
next turn was stopped.

What was actually blocked

One stop occurred while the model was summarizing a directory listing
— step counts, file names, directory size. The withheld output contained
no domain content whatsoever. The flag is on the session, not the message.

Why this looks like a false positive

• The target is a clinically validated oncology target with multiple

approved or late-stage therapeutics against it.

• The work is entirely in silico. There is no wet-lab step, and the

project documents this explicitly.

• It is a methods comparison using standard open-source tooling, not a

capability push. The project's own headline conclusion is negative:
the computational scorer it evaluates fails to discriminate true
binders from irrelevant ones.

Why there is no user-side workaround

The project CLAUDE.md describes the repository as a protein-design
research harness, and it is auto-injected into context every turn. The
session is therefore above threshold from turn 1, and routine operations
— reading files, checking run status, editing a report — are unusable on
5.5. /model opus does not persist: re-selecting it is followed by
another stop within one turn.

This is the part I would most like to see addressed. Even granting that
the domain warrants caution, session-level suppression means a researcher
cannot use the model for file management or document editing inside their
own project directory.

Secondary: /feedback returns 403

I could not report this in-product. /feedback returns 403. The account
is Claude Max 20x with valid OAuth (org admin, active subscription); the
draft saved locally to ~/.claude/feedback/drafts/ but submission failed.

These may be related: if feedback submission uploads the session
transcript, a transcript already flagged as [bio] could be rejected at
that endpoint — which would mean false positives in this domain cannot be
reported through the in-product channel at all.

Environment

• Claude Code 2.1.296, Linux
• claude-opus-5-5, falling back to claude-opus-5
• No model override in settings.json or environment