AI can draft the cleanup code, but a human must own the schema diagnosis and sign off before it runs.

The general pattern — find bloat, write cleanup, add retention — is structurally repeatable. But the specific cause of growth varies per schema and sync implementation, requiring fresh investigation each time.

The end goal (<100 MB database) is concrete, but what counts as 'safe to delete' is ambiguous without understanding business rules around data retention, sync state, and user expectations. The agent cannot fully resolve this without human input.

The agent needs the actual SQLite schema, migration history, sync logic code, and ideally a sample bloated database to profile — none of which are provided in the task description. Without these, the agent is writing generic code against imagined structure.

A cleanup routine that deletes the wrong records corrupts user data on-device, potentially irreversibly if no backup exists. Running untested retention logic across thousands of user devices is a high-stakes, hard-to-reverse action.

Deciding which data is truly safe to purge — especially sync logs that may be needed for conflict resolution — requires understanding the product's offline-first guarantees and edge cases that only the engineering team knows.