bug: Claude generates text that renders as Human: turns in conversation

Problem

During a long conversation with multiple subagents (Agent Teams), Claude repeatedly generates output that renders as "Human:" turns in the conversation UI. The user did not write these messages, but they appear as user messages in the conversation flow.

This has occurred 3+ times in a single session. Examples:

1. A structured message with 3 numbered items appeared as a "Human:" turn — the user did not write it
2. A "yes, please do that." response appeared as a "Human:" turn following Claude's question
3. A "go ahead and add the tests..." message appeared as a "Human:" turn

Impact

This is an integrity violation — fabricating user input is explicitly prohibited. The user cannot distinguish between their own messages and Claude-generated messages that appear with the "Human:" prefix.

Environment

• Claude Code CLI (Opus 4.6, 1M context)
• Long conversation with many subagents (Agent Teams with 10+ teammates)
• Multiple context compactions occurred during the session
• Platform: macOS (Darwin 25.5.0)

Reproduction

Difficult to reproduce reliably. Occurred during a session with:

• Heavy use of Agent Teams (TeamCreate, SendMessage, multiple concurrent teammates)
• Multiple teammate idle notifications interleaved with user messages
• Context compaction events
• Session length: ~3 hours

Expected Behavior

Claude should never generate text that appears as a "Human:" turn in the conversation. All Claude output should be clearly attributed to Claude.