Blind DB Debugging

Overview

Diagnose database issues (missing or incorrect data) via SQL row counts without ever seeing actual data values. The agent generates hypotheses and queries; the user runs them and reports only row counts back.

Phase 0 -- Gather Context

Before writing any query:

1. Understand the expected output -- what the user sees (or should see) in the final app or product. Describe it at the product level, not the database level.
2. Trace back through the code -- follow the call chain from the UI or API response back to the database: view -> controller/resolver -> service -> query. Identify which queries, joins, or filters are responsible for producing that output.
3. Ask only what the code cannot answer:
   • Scope -- one record, a subset, or all records?
   • When it started -- after a deploy, migration, data import, or unknown?
   • A working example (optional) -- one ID or key that behaves correctly; enables control group queries.

Do not skip this phase. Hypotheses must be grounded in the actual code path, not guesses about the schema.