A green test suite felt like an objective finish line until previews kept revealing what passing tests could not: completion is a human decision, not a CI event.
A local request-minimizing boundary kept live session cookies out of the model chat while Codex still helped build the extension: 45 cookies to one in one documented run, 41 to one in a separate capture.
Automated Playwright screenshots were genuinely useful; the friction of reaching them from a phone is what made live previews the primary review surface.