[Bug] Opus 5 reasoning_extraction false-positive blocks session dump on benign infrastructure work (4 consecutive failures)

Description

Opus 5 (1M context) hit [reasoning_extraction] safeguard false-positive 4 consecutive times on a completely benign request to write a session summary to a local markdown file. The session was 6.5 hours of routine infrastructure work (PostgreSQL migration, backfill fixes, smoke test reliability). No prompt injection, no attempt to extract reasoning — just "please update session_current.md under today's date."

Once the first refusal occurred, every subsequent message in the session was blocked with the same error, including:

• "can you do the update? yes or no"
• "copy everything from the session to a local drive please"

The session was unrecoverable. Had to /exit and switch to Opus 4.6 in a new session.

Request IDs

1. req_011CfNxrB3DpNXGnHaeAn9Cu
2. req_011CfNy7JboNi2LSarJMaKYq
3. req_011CfNy91xFHhpGinYV1poTQ
4. req_011CfNyAuGisPfiCW3PdFuB8

Error Details

Opus 5 (1M context)'s safeguards stopped the response above · continuing once with that noted

API Error: Opus 5 (1M context)'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 Opus 5 (1M context).

Details: [reasoning_extraction]

Environment

• Claude Code version: 2.1.282
• Model: claude-opus-5[1m] (Opus 5, 1M context)
• OS: Windows 11 Pro 10.0.26200
• Shell: PowerShell 7

Steps to Reproduce

1. Run a long session (~6.5 hours) with Opus 5 (1M context) doing routine infrastructure work
2. At the end, ask the model to write a session summary to a local markdown file
3. Model's reasoning attempts to summarize the full session in one block
4. [reasoning_extraction] classifier fires on the model's own reasoning
5. Every subsequent message in the session is blocked with the same error

Context

This appears to be the same root issue as:

• #88364 (client-injected batching reminder triggers classifier)
• #94210 (persists across sessions)
• #95275 (Opus 5 flags ALL messages once it starts)
• #96139 (Opus 5.5 hits it too)
• #96702 (word "reasoning" in prompt triggers it)

A colleague independently reported the identical problem — session summary at end of day, same [reasoning_extraction] error, same model. This is not an isolated incident.

Impact

The session dump is a critical workflow step — it persists context across sessions for a multi-month migration project. Losing the ability to write a summary at end of session means 6.5 hours of work context is lost (had to reconstruct from JSONL transcript in a separate session).

Workaround

Switched to Opus 4.6 in a new session and reconstructed the session summary from the JSONL transcript file. /model switch within the same session was not attempted but should work based on other reports.