AI can draft the migration cleanup, but a human must own the final schema before it touches production.

The structural pattern — read migrations, detect redundancy, consolidate — is repeatable, but each project's migration history is unique and requires bespoke judgment about which alterations are safe to collapse. This is not a cookie-cutter refactor.

Success criteria are underspecified: 'clean,' 'testable,' and 'reliable' are subjective, and the agent cannot independently verify that rollbacks work correctly or that CI/CD passes without running the full pipeline. The definition of done is fuzzy.

The agent needs access to all 40+ migration files, the 8 model definitions, and ideally a live or seeded PostgreSQL instance to validate output — access that may not be trivially grantable. Without a runnable environment, the agent is producing untested code.

A mis-consolidated migration against a 2-million-row PostgreSQL database could silently drop columns, corrupt indexes, or break rollback chains — damage that may not surface until a production incident. Reversibility depends entirely on backup discipline.

Deciding which historical migrations encode intentional business logic versus accidental cruft requires understanding the application's evolution — context that lives in team memory, not in the files themselves. A human engineer must own the final review.