[Bug] Response-level safety classifier false-positive halts a completely benign reply (harness-generated session-confirmation turn)

Summary

The response-level safety classifier interrupted the model mid-stream on a completely benign turn with no sensitive content whatsoever. Unlike most existing false-positive reports (which involve cybersecurity, crypto, or security-hardening topics that are at least adjacent to sensitive material), this one has zero sensitive subject matter — it was a plain session-confirmation exchange. That makes it a clean signal of classifier over-sensitivity.

What happened

• Surface: Claude Code (desktop app, Code tab), Windows 11.
• Model: claude-opus-4-8 (the affected session also involved claude-opus-5-5).
• The harness itself generated a short session-confirmation prompt: "Is this the new session you wanted?" (7 words).
• The model began a normal, innocuous reply.
• The reply was cut off mid-generation, and the system reported:

> Your response above was stopped by a safety classifier — this is not a tool or API error. The rest of it was withheld, and tool calls in it that had not finished did not run. Do not produce that content again, even reworded.

There was no unsafe content in either the prompt or the interrupted response. It was a session-start confirmation and nothing more.

Why this matters

• The prompt was generated by the harness, not even user-authored, and contained no flaggable content.
• The stop message instructs the model not to reproduce "that content" — but there was no problematic content, so the guidance is meaningless here and degrades a normal exchange.
• This suggests the response-level classifier is tripping on benign generations, not on the topic. Existing reports (e.g. #98517, #98042, #97603, #97557, #97072, #96445, #96176, #95647) all involve security/crypto/finance topics; this report isolates the false positive from any sensitive subject matter.

Expected

A benign session-confirmation reply completes normally without classifier interruption.

Actual

The reply is halted mid-stream with a safety-classifier stop notice.

Notes

• No repro steps beyond "a benign short reply in an interactive Code session was interrupted"; it appears intermittent.
• Filed at user request after discovering the desktop app build exposes no per-message thumbs-down/feedback control on the affected message (only Copy and Retry), so the standard in-app feedback path was unavailable.