That's exactly my understanding as well. This is, essentially, the LLM hallucinating user messages nested inside its outputs. FWIWI I've seen Gemini do this frequently (especially on long agent loops).