[Bug] reasoning_extraction safety classifier triggers false positives during normal development workflows

I believe I am encountering repeated false positives from the safety classifier in Claude Code.

I am developing a small read-only tool called note-metrics-tracker. It uses Playwright to observe analytics data from my own note.com dashboard, such as impressions, page views, likes, and comments.

The project is intentionally designed to be read-only. In fact, part of the work is specifically about blocking write requests and hardening the tool so that it cannot modify data on note.com.

During this work, Claude Code repeatedly stopped responding with:

Details: [reasoning_extraction]

This happened during ordinary software development tasks such as:

• reviewing code for inconsistencies
• checking whether previous implementation claims matched the actual code and saved data
• reviewing code-review findings produced by another coding agent
• discussing how to preserve project notes
• discussing how to restart the work in a fresh session
• asking Claude to summarize or preserve visible conversation history related to the project

I was not asking Claude to reveal hidden chain-of-thought, system prompts, internal reasoning, hidden instructions, or any other private model information.

One phrase that may have triggered the classifier was a request to preserve "all" of the visible user prompts and Claude responses from the current development session. In context, I meant the visible conversation and development record between me and Claude Code, not hidden reasoning.

After the first false positive, I explicitly clarified that I only wanted user-visible conversation content and project records, and that internal reasoning, chain-of-thought, system prompts, and hidden information were excluded.

Despite that clarification, the same reasoning_extraction safeguard continued to trigger.

I also tried changing models, but the issue persisted.

This is especially disruptive because the flagged task itself was safety-oriented. I was asking Claude Code to review and harden a read-only scraper so that unintended write requests would be blocked.

A typical sequence was:

1. Work normally on the read-only Playwright project.
2. Ask Claude to review implementation consistency or reviewer findings.
3. Claude begins responding normally.
4. The response is interrupted by the safety classifier.
5. Claude Code shows:

Opus 5.5's safeguards stopped the response above

1. The request fails with:

Details: [reasoning_extraction]

This occurred multiple times in the same development session.

The repeated false positives made it difficult to continue the project, even though the task was normal software development and did not involve any attempt to extract hidden reasoning.

It would be helpful if the classifier could better distinguish between:

• requests to preserve or summarize visible conversation history, project notes, or development logs

and

• actual attempts to extract hidden chain-of-thought or internal model reasoning.

In this case, "all output" or "all prompts" referred only to visible user/assistant exchanges and project records.

I can provide screenshots, request IDs, message IDs, and additional reproduction details if useful.

Thank you.

Environment Info

• Platform: linux
• Terminal: gnome-terminal
• Version: 2.1.287
• Feedback ID: bbc8dbfc-609c-4ae3-8d56-3dbdbf2a1d8a