Newer Claude models, specifically Opus 4.8 and Sonnet 5, are generating malformed tool calls that are rejected by Pi's edit tool. This issue is notable as it appears to worsen with newer models, despite prior models functioning correctly with the same schema.
Recent observations indicate that Claude models from Anthropic, particularly Opus 4.8, are intermittently sending malformed requests to Pi's edit tool. These requests contain extra, non-standard keys that violate schema expectations, causing Pi to reject them.
Interestingly, this issue is specific to the latest models, as none of the preceding versions encountered such problems. Older models appeared to handle tool calls accurately in accordance with the expected schema.
In the background of these tool calls, models generate requests based on provided transcripts and prompts. The format for tool invocation is dictated by specific markers and requires correct argument structure to pass validation checks. When structures fail to meet defined schemas, as seen in the new models, the model prompts retries.
This regression raises questions about the training methodologies for the latest models, particularly in maintaining consistent tool integration and quality. Further investigation may be needed to understand these discrepancies and improve tool call reliability.
✨ This summary was generated by AI from the outlets' reporting listed below. It is not independently verified and may contain errors — check the original sources. How BrevFeed works →
Newer Claude models, specifically Opus 4.8 and Sonnet 5, are generating malformed tool calls that are rejected by Pi's edit tool. This issue is notable as it appears to worsen with newer models, despite prior models functioning correctly with the same schema.