@JFPuget Had the same issue. For me it showed up when codex tried to reuse its auth credentials for a "custom" codex harness instead of generating a new set of credentials. Once I killed the other harness codex started to work again.
@ConstanceWaing 20 x .2 is 4.00. You mean a full day coding for you using claude was only 4 dollars? Are you just not using it that much, that is ridiculously cheap. On my vibe coded side project im hitting maybe 100 dollars a day, and even at work where I just have it do research it is still 10
@_xjdr As a note, I have noticed also 5.5 has been trying to "connect" separate features to each other and been introducing issues into the code. Its like it believes everything needs to affect everything, its kind of weird. Not sure if related.
@NicolasZu This is the sort of shit that gets the company to reduce limits all the way around. Yes, just burn lots of compute/power just to make sure you got the "best deal". You are like the fatass at a buffet that goes back for 7 plates.
@thsottiaux Why do I keep having to disable the damn github mcp tools, those shouldnt even be on by default. Also pushes out tools from the MCP server I am trying to use.
@hari__prasadd@ClementDelangue What, how? I mean unless the labs are going out of there way to double check all your code for vulnerabilities and the submitting PRs for it how does that make you LESS vulnerable? It just exposes your code to attackers. This makes no sense and why are you just going with it?
@leothecurious Dont you think the model makers can see your benchmark once you have run the model against it? They can just train on their user's data, you dont have to openly publish it.
@jianxliao One addendum to this, highly recommend if making your own cli commands to separate the cli into a client server architecture to allow for state/sessions. Otherwise things like playwright aren't going to work.
@polynoamial There is no such thing as uncontaminated after the first run. The model providers are assuredly training on their user's sessions (even if they don't admit to it) and once a model has been used on that problem, the company has the material to train on it.