[BUG] system task reminders / split prompts are being presented in a way that's indistinguishable from user input, causing the model to attribute system-generated content to the user

[hugopalma17]

Preflight Checklist

• I have searched existing issues and this hasn't been reported yet
• This is a single bug report (please file separate reports for different bugs)
• I am using the latest version of Claude Code

What's Wrong?

• Model: claude-opus-4-6
• Bug: System task notifications are being processed as user-originated messages.
  Agent attributed a system-generated prompt to the user and acted on it with false confidence.

Claude code CLI on macOS v2.1.32

What Should Happen?

Probably just an issue of the 4.6 model roll out behavior change.

Error Messages/Logs

N/A.

Steps to Reproduce

• Reproduction: The task reminder system that says "The task tools haven't been used recently..."
  appears to be triggering model behavior as if it were user instruction. In this case the agent claimed
  the user said "report to hq" when no such user message existed.

Claude Model

Opus 4.6

Is this a regression?

Yes, this worked in a previous version

Last Working Version

No response

Claude Code Version

2.1.32

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

iTerm2

Additional Information

No response

[github-actions]

Found 3 possible duplicate issues:

1. [Bug] Hallucination: False User Confirmation from Task Notification and Compaction Events - why are <task-notification>s user content?? #20920
2. [BUG] Background task killed notifications cause assistant context confusion #21700
3. [BUG] Phantom user messages - commands executed without user input #14269

This issue will be automatically closed as a duplicate in 3 days.

• If your issue is a duplicate, please close it and 👍 the existing issue instead
• To prevent auto-closure, add a comment or 👎 this comment

🤖 Generated with Claude Code

[ljmatkins]

This bug is causing real issues - when a background notification comes in, it frequently makes Claude think we somehow responded to its question, and then it proceeds to barrel forward without user input.

[github-actions]

Closing for now — inactive for too long. Please open a new issue if this is still relevant.