> 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?