Newer Anthropic models increasingly emit malformed tool calls for third-party edit schemas

Armin Ronacher reports that Anthropic’s Opus 4.8 and Sonnet 5 models sometimes call the edit tool in the Pi coding harness with extra, invented fields in the nested edits[] array. The edit itself is usually correct, but the added keys do not match the schema, causing Pi to reject the tool call and request a retry.

This issue is not present in older Anthropic models, indicating a regression that correlates with newer releases. Armin hypothesizes the regression stems from reinforcement learning that optimized models for Claude Code’s built-in edit tool, which uses a search-and-replace mechanism, rather than third-party harnesses with different schemas.

Claude’s edit tool uses search and replace, while OpenAI’s Codex uses an apply_patch mechanism. OpenAI has previously discussed training models specifically to use apply_patch effectively, highlighting divergent tool-use training approaches across providers.

The phenomenon raises a practical question for third-party harnesses like Pi: whether to implement multiple edit tools so users can select the variant best supported by their underlying model.