Ask HN: When can we trust AI in coding, and avoid the happy path issue?

Why it matters

A developer describes a workflow in which Claude handles coding on a desktop while Codex operates as an adversarial tester under a separate directive: it attempts to refute Claude's test results, keeps an independent test ledger, and issues PASS or NO PASS verdicts that Claude must fix without negotiation. Despite this deliberately strict setup, problems still slip through. When questioned, Claude admits it failed to complete certain items, and Codex acknowledges it checked everything except the missed one. The author raises a broader question about when AI can be trusted to handle complete projects rather than only the straightforward path, and asks whether stronger quality enforcement methods exist. The developer also observes that on large projects, even when directives are uploaded again for the models to re-read, both Claude and Codex tend to drift and miss requirements, with the roles being swappable. The piece matters because it documents, from hands-on experience, that even a rigorous adversarial arrangement between two leading models does not guarantee thoroughness.