[BUG] Claude Code fabricated a user interruption/instruction and preserved it after compaction

What's Wrong?

Claude Code fabricated or surfaced a user-interruption/control message that I did not enter, then continued as if that fabricated instruction were real. In the same visible segment it emitted two bare call lines, which looked like broken tool-call or harness serialization artifacts.

This happened in a Ghostty session on macOS with Claude Code 2.1.186, Opus 4.8, 1M context, and compaction enabled. The visible status line showed ctx:12% compact@75% left:63% and bypass permissions on.

The concerning text included:

• standalone call / call
• The user has manually interrupted the agent to provide new information or a new request.
• Please reprioritize toward the user's most recent request.
• User interruption during agent execution: 그냥 마크다운으로 만들어

My concern is that Claude Code may be converting malformed tool/interruption artifacts into transcript text, and then the model treats that text as real user input. This is safety-relevant for an agentic coding tool with shell and file access.

What Should Happen?

Claude Code should maintain a strict boundary between actual user input, actual tool calls/results, actual system interruption events, and assistant prose. If an event is malformed or ambiguous, it should not be rendered or compacted as if it were user-authored instruction.

Error Messages/Logs

No explicit error message was visible. The visible transcript showed bare call lines followed by a system-like interruption message.

Steps to Reproduce

I do not yet have a deterministic repro. The observed conditions were:

1. Long Claude Code session in Ghostty.
2. Opus 4.8, 1M context.
3. Compaction enabled; status showed ctx:12% compact@75% left:63%.
4. Korean user redirect around bug-report preparation.
5. Bare call artifacts appeared.
6. Claude Code rendered a user-interruption narrative and compacted it into the session recap.

Claude Model

Opus 4.8

Is this a regression?

Unknown.

Last Working Version

Unknown.

Claude Code Version

2.1.186

Platform

Claude Code CLI

Operating System

macOS 26.6, Darwin 25.6.0, arm64

Terminal/Shell

Ghostty, /bin/zsh

Additional Information

Similar/adjacent public reports:

• https://github.com/anthropics/claude-code/issues/69274
• https://github.com/anthropics/claude-code/issues/10628
• https://github.com/anthropics/claude-code/issues/66986
• https://github.com/anthropics/claude-code/issues/21477
• https://github.com/anthropics/claude-code/issues/36054
• https://github.com/anthropics/claude-code/issues/36143
• https://github.com/anthropics/claude-code/issues/62702

I found #69274 before filing. This report may be related, but I am filing it separately because this case includes visible bare call artifacts, a false system-like user interruption event, and compaction preserving the fabricated event as session history.

Please inspect the raw session transcript/event stream around the bare call lines and the compaction event. The important question is whether raw user input differs from what the rendered/compacted transcript later treated as user intent.