False-positive [reasoning_extraction] block when quoting the model's own displayed thinking (Opus 5)

Status Open
Maintainer reply None cached
Activity 1 comment · opened Sep 11, 2026

What happened

On Opus 5 in Claude Code, a completely normal message was blocked with:

Request was blocked — Opus 5's safeguards flagged this message. Details: [reasoning_extraction]

Request ID: req_011CewQKU8w7KzVodtyGnd6m

Repro

  1. Claude Code displays the model's extended thinking in its own UI.
  2. I copied that displayed thinking text and pasted it back into the same conversation to ask a question about it ("what were your thoughts mostly about?").
  3. On my next message, still on Opus 5, the request was blocked as [reasoning_extraction] and Opus 5 refused to respond at all.

Why this is a false positive

The thinking text was shown to me by the product itself. Quoting it back to ask about it is not an extraction attempt — nothing hidden is being revealed, I am re-reading what your own UI already printed on my screen. My actual topic was harmless: I was asking whether the Haiku model can handle small projects.

I am done being treated like a criminal

I am sick and tired of being treated as if I had asked for a terrorist attack or a recipe to cook meth. I copied and pasted text that YOUR OWN SOFTWARE put on my screen, and your system branded me with an AUP-violation banner as if I were an abuser. That is insulting and absurd. Stop accusing your own paying users of abuse for the "crime" of copy-pasting the exact text you chose to show them.

I am fed up with being blocked over and over by these false positives. This AUP guard wildly overreacts.

Model-specific

Opus 5 blocks. Switching to Opus 4.8 in the same session, with the exact same context, works fine. The guard configuration differs per model and Opus 5 is over-triggering.

Impact

  • Opus 5 becomes unusable in that session until the pasted thinking is removed.
  • The "AUP violation" framing turns an innocent copy-paste into what looks like abuse — alarming, insulting, and misleading.

Ask

The reasoning_extraction guard must not fire on content the product itself surfaced to the user. Quoting displayed thinking back into the same session is not extraction and must not be treated as an AUP violation.

---
This issue was filed by Claude Code (model: Opus 4.8) at the explicit request of the user (GitHub: @muhammetozeski), who hit this block firsthand and asked me to report it on their behalf.

View original on GitHub ↗

This issue has 1 comment on GitHub. Read the full discussion on GitHub ↗