A team has 2,000 unit tests, almost all with mocked repositories and HTTP clients, yet bugs keep reaching production at the database and API boundaries. What is wrong with their strategy and how would you fix it?

Their tests only check that the code calls its mocks the way the author expected. They don't check that the code works with the real database and APIs. Each mock returns whatever the test told it to return. So these bugs all pass the tests:

• a wrong SQL query
• a missing migration
• a unique constraint that the data breaks
• a transaction that never commits
• a third-party API that changed its response

The fix is to move confidence to the boundaries. Add:

• integration tests that run repositories and queries against a real Postgres in Testcontainers (a library that starts real services in Docker for your tests)
• API tests that go through the real Nest pipeline, using supertest
• contract tests, or realistic recorded responses, for third-party APIs

Then delete or simplify unit tests that only repeat the implementation, such as "calls repo.save with X". Keep unit tests where there is real logic to check. In them, assert outcomes (what the code returns or changes), not calls.