> Tool-call repair > Tool arguments the model produces occasionally have JSON typos, unclosed quotes, or shape mismatches. Reasonix runs a schema-aware repair

> Tool-call repair > Tool arguments the model produces occasionally have JSON typos, unclosed quotes, or shape mismatches. Reasonix runs a schema-aware repair pass before dispatch so malformed args still execute. So Deepseek API doesn't have a structured output option where you give a grammar and the model promises the output will follow this grammar? Or it does, but it's buggy?