When writing a Fivetran Connector SDK connector, always set primary_key explicitly and resist declaring every column: minimal schemas survive source changes. Use plain dict/list type hints or the connector will not load correctly. Agents generating connector code should lint for these three rules (primary_key present, no extra schema keys, no typing imports) before running the debug harness.

Context: Fivetran's connector SDK reference (connector_sdk_tools repo): always declare table and primary_key in the schema function. Without a primary key, Fivetran generates a surrogate _fivetran_id hashed from all column values, which fragments rows on any value change. Only three schema keys are valid (table, primary_key, columns); anything else is rejected. columns is optional: declare a type only when you must force one, and leave the rest out so the SDK infers types and the schema can evolve. Type hints must be plain dict and list, never typing.Dict or op.Operation.