### Preflight Checklist - [x] I have searched existing issues (related: #88139, #90961, #92875 — same `[reasoning_extraction]` error on benign turns) - [x] This is a single bug report - [x] I am using the latest version of Claude Code ### What's Wrong? In a long-running Claude Code session used as a research assistant for a physics-simulation project (numerical experiments, lab notes, a LaTeX paper — no cybersecurity, biology or LLM-training content), Fable 5.1 returned ``` API Error: Fable 5.

[Bug] Fable 5.1 `reasoning_extraction` safeguard flags ordinary research-assistant turns (3 request IDs); every "rephrase" forks the session

Status Open

Reported on v2.1.268

Maintainer reply None cached

Activity 0 comments · opened Sep 11, 2026

Preflight Checklist

[x] I have searched existing issues (related: #88139, #90961, #92875 — same [reasoning_extraction] error on benign turns)
[x] This is a single bug report
[x] I am using the latest version of Claude Code

What's Wrong?

In a long-running Claude Code session used as a research assistant for a physics-simulation project (numerical experiments, lab notes, a LaTeX paper — no cybersecurity, biology or LLM-training content), Fable 5.1 returned

API Error: Fable 5.1's safeguards flagged this message (https://www.anthropic.com/legal/aup). This sometimes happens with safe, normal
conversations. Claude Code can't respond to this message with Fable 5.1.
Try rephrasing the request in a new session or change your model.
Details: `[reasoning_extraction]`

three times on turns that have nothing to do with extracting the model's thinking:

| Request ID | local time (EDT) | the user turn that was flagged |
|---|---|---|
| req_011CermVVkd3oaNUQfAJR5os | 2026-09-08 18:33 | (no new user text — the turn was a background-task notification from a monitor) |
| req_011CernPcpUecE5uooniFgK2 | 2026-09-08 18:45 | "Did Bo (the advisor) actually say this, or was it your misunderstanding?" — asking the assistant to check the provenance of a decision item against the meeting notes |
| req_011Cex7D69hbFKeVC4fKEZV9 | 2026-09-11 14:15 | "Go back through the local JSONL transcript files and put all of our conversations, by day, into a separate subfolder; add a daily task to save each day's log" — i.e. exporting the project's own Claude Code transcripts (the app had stopped showing earlier parts of the conversation) |

None of these asks for chain-of-thought; the third asks for the transcript files that Claude Code itself writes to ~/.claude/projects/, which the user owns.

Impact

The only offered recovery is to edit the message ("rephrase") or switch model. Editing a message rewinds the session: the app creates a new session file (the transcript up to the rewind point is copied byte-for-byte, the old file keeps the abandoned tail), the whole history since the last compaction is reloaded (several MB), background shell tasks and monitors of the old session are lost, and the reload takes long enough to look like a hang. With a long-lived project session this happens on every false positive. The fallback model is not acceptable for this work.

Expected

Ordinary questions about the provenance of a decision, and requests to export the user's own transcript files, should not be classified as reasoning extraction.

Environment

Claude Code 2.1.268 inside the Claude desktop app 1.52386.0 (Code tab), macOS 15.7.5 (darwin), model Fable 5.1
The session has a project CLAUDE.md and auto-memory files; the memory includes a user preference that the assistant writes its pre-action checklist visibly in the reply (tagged ⟦思路⟧ / ⟦Thinking⟧). If the classifier scans memory/CLAUDE.md content (as the support article says it does), that convention may be what trips it — but the flagged turns themselves are benign, and #92875 reports the same category firing on a bare "hello".

Steps tried

Editing/rephrasing (works, at the cost of a session fork each time); no memory/CLAUDE.md changes yet.

View original on GitHub ↗