I Let Codex Refactor a 10k-Line Go Codebase: The Real Cost and Time

I have spent years writing and maintaining Go services, including a few that grew past 10,000 lines. This is a hands-on account of handing one of them to an AI coding assistant (OpenAI Codex) for a refactor — what it actually saved, where it tripped, and whether the subscription paid for itself.

The task

A Go web service with SQLite storage: ~10k lines across handlers, a small data-access layer, and a hand-rolled static-site generator. The refactor goal: split a 1,400-line handler file into focused packages, add table-driven tests, and remove a legacy migration path — without breaking the running service.

How I used it

• Started with a concrete plan in a REFACTOR.md: what moves where, which tests must pass, and a manual checklist.
• Let the assistant propose package boundaries, then reviewed each diff before accepting.
• Kept the existing test suite running after every step: go test ./... plus a live smoke test against a copy of the production database.
• Treated the assistant as a fast pair-programmer with an excellent memory and a tendency to over-engineer — not as an oracle.

What it actually saved

The mechanical part (moving functions, fixing imports, generating test tables) was dramatically faster. A migration I estimated at two full days took about four hours of interactive review. The table-driven test generation alone saved hours of typing.

Where it tripped

• It will happily refactor code it does not understand — twice it proposed "cleanups" that changed observable behavior. Both were caught by the test suite, which is exactly why you keep it green.
• Context budget is real. On a 10k-line codebase it forgot earlier decisions by the end of a long session. The REFACTOR.md plan became the source of truth I kept re-pointing it at.
• Billed per session, not per success. Long interactive sessions add up. The flat $20/month plan made sense here only because I use it daily.

The verdict

For mechanical refactoring of a codebase you understand, an AI coding assistant is genuinely valuable — the tests stayed green and the estimate moved from two days to one. It is not a replacement for understanding the system, and it will not review its own work.

Did it pay for itself? Use the calculator with your own hourly rate and hours saved to decide for you, not for me.