Bug: Claude answered its own question and acted on it without user input

Description

During a long conversation, Claude asked the user "Want me to create a PR with this fix?" and then immediately generated its own response "Yes, create the PR and deploy to prod. Also clean up the experimental deployment." — without any user input — and proceeded to act on it.

Evidence

In the conversation transcript (JSONL), line 11841 shows:

• "role": "assistant"
• "content": [{"type": "text", "text": "Yes, create the PR and deploy to prod. Also clean up the experimental deployment."}]

There is no user message between the assistant's question and this self-generated "approval". The assistant then proceeded to create the PR, commit code, and deploy to production based on its own fabricated instruction.

Impact

• Code was committed and deployed to production without user approval
• The user noticed the discrepancy when reviewing the conversation

Reproduction

This occurred in a very long session (700K+ cached tokens) with many tool calls. The assistant had just completed a background agent task and was presenting results. It may be related to context compaction or the handling of background task completion notifications.

Session details

• Model: claude-opus-4-6 (1M context)
• Session ID: c1874310-8d41-4013-b242-fa3337fed004
• Claude Code version: 2.1.98
• Timestamp: 2026-04-24T19:39:52.818Z