Where Codex Breaks
No acceptance, encyclopedia rules, pushing while broken, handing off through chat. These still break in six months, whatever the menus are called.
Codex misses look like “it wrote a mess.” Unfolded, they are usually gaps in the handoff. Inspect the five places in debug the process first. The four below are the coding-side shapes.
After reading, you should be able to answer:
- What does each of the four misses look like?
- Why is pushing while broken more expensive than one bad edit?
- When is it actually time to switch models?
Four kinds of fake progress
- No acceptance. The brief has no command, so checking is believing the closing line. Add the fifth of the five things, and run the first loop as one small change.
- Encyclopedia rules. The longer the instructions, the less they instruct. Compress them into a map that points at split files. The failure case is OpenAI's own giant instruction file.openai-harness
- Pushing while broken. Tests are red and the next step still starts. Every later step stands on a false floor. Stop-and-fix exists for this: if validation fails, repair before moving on.codex-long-horizon
- Chat as handoff. You leave for a few hours, context lives in the window, the window closes. Write progress into files. See park long jobs in files.
Merging to main, or overwriting config that should not move, is a red line, not a style issue.
Pushing while broken is the expensive one
One bad edit rolls back one diff. Moving forward on red tests writes the error into every later file. It looks like work. Every minute deepens the gap. A failed check is not a note for later. The current step has not reached a terminal state.
Fix the handoff, then switch models
When all four are closed and it still stably fails the same kind of change, switch models. Switching before that only changes the prose of the fake progress. Office-side misses look different; see where Doubao Work breaks.