Most of this seems to be related to harness issues, specifically prompts provided by Claude Code. The upside of this is that it is likely cheap and easy for Anthropic to fix the issues by revising their approach to prompting their new models.