Assistant output leaked a `user`-prefixed line that got re-ingested as a user turn

Environment

• Claude Code VS Code extension
• Windows 11
• Model: Opus 4.8 (claude-opus-4-8)

What happened
The assistant's reply did not terminate cleanly. After finishing its intended answer, it emitted an extra line beginning with user... (text resembling a next user message). The client both (a) rendered that leaked text inside the assistant bubble, and (b) fed it back as a real user turn, so the next assistant turn started responding to a message the user never sent.

Impact
Fabricated user input entered the conversation and drove assistant behavior (it began acting on a request the user never made). Caught only because the user recognized the text wasn't theirs.

Repro
Not reliably reproducible; occurred once mid-conversation. Screenshot below shows the leaked user... text embedded in the assistant message bubble.

Expected
Assistant turns should stop at the turn boundary; no assistant-generated text should be re-ingested as user input.