Anthropic's newer Claude models generate malformed tool calls for third-party editors

Armin Ronacher reports that Anthropic's latest Claude models, specifically Opus 4.8 and Sonnet 5, are producing malformed tool calls when interacting with the Pi coding harness. The models invent extra fields in the nested edits array that do not match the expected schema, causing Pi to reject the call.

• Newer SOTA models like Opus 4.8 and Sonnet 5 exhibit this regression, while older models do not.
• The issue stems from models being trained to use Claude Code's baked-in edit tools, which harms compatibility with third-party harnesses.
• This contrasts with OpenAI's Codex, which uses a different apply_patch mechanism designed for effective tool use.

The article suggests that third-party coding harnesses may need to implement multiple edit tools to ensure correct performance across different underlying models.