Opus also does this and then writes comments in code or PR descriptions describing how it went wrong earlier in the session.
The biggest issue with LLMs is that they still suck at general contextual awareness and ability to judge what is appropriate.