The catalog is now organized by task type

Every dataset now carries one of five task types:

Solve — write code from a specification Debug — find and fix a defect in provided code Optimize — improve a working implementation along a cost axis Explain — trace or justify the behavior of code that runs Review — judge correctness against a claim

The streaming endpoint accepts a sample_type query parameter. Pass any of the five slugs to get only that slice. Omit it for every row.

The response includes an X-Sample-Type header echoing the filter. If the dataset has no rows of the requested type, the endpoint returns 404.

Nothing about existing keys or existing datasets changed. Same keys work. Same rows stream. They are now filterable.