>I think, in most cases, GPT5-Codex finally is as good as a senior engineer for my specific use case.
This is beyond bananas to me given that I regularly see codex high and Gpt-5-high both fail to create basic react code slightly off the normal distribution.
That might say something about the understandability of the react framework/paradigm ;)
Quality varies a lot based on what you're doing, how you prompt it, how you orchestrate it, and how you babysit and correct it. I haven't seen anything I'd call senior, but I have seen it, for some classes of tasks, turn this particular engineer into many seniors. I still have to supply all the heavy lifting (here's the concurrency model, how you'll ensure exactly-once-delivery, particular functions and classes you definitely want, a few common pitfalls to avoid, etc), but then it can flesh out the details extremely well.
If you really want to see it fail at something easy, try to have write something that can use JSX but doesn't use React (Bun, Hono, etc). Seems like no amount of context management and detailed instructions will keep it from reaching for React-isms.
Do you mind if I ask what kind of React code you're working on? I've had good success using Codex for my frontend development, especially since all of my projects consistently rely on a pretty widely used and well documented component library. I realize that makes my use case fairly narrow, so I don't think I've discovered the limits you have.
Today I was trying to get it to temporarily shim in for development and consume the value of a redux store via merely putting a default in the reducer. Depending on that value, the application would present different state.
It failed to accomplish this and added a disgusting amount of defensive nonsense code in my saga, reducer and component to ensure the value was there. It took me a very short time to correct it but just watching it completely fail at this task was borderline absurd.
Thanks for the context! I feel the same way. When it fails it fails hard. This is why I'm extremely skeptical of any of the non-cli cloud solutions - as you observed, I think the failures compound and cascade if you don't stop them early, which requires a compelling interface and the ability to manually intervene very fast.
This is beyond bananas to me given that I regularly see codex high and Gpt-5-high both fail to create basic react code slightly off the normal distribution.