codex-discussions

[glad solar]

ig u can try it

try starting a new thread

[lunar aurora]

i mean i have 2 projects it didnt work for neither

so i guess full access is the temp solution rn

[glad solar]

wow

[lunar aurora]

what!

[glad solar]

nothing i mean it's just strange

[lunar aurora]

the issue? yeah

[glad solar]

yeah

[cedar skiff]

This is a windows codex app issue isnt it?

[boreal holly]

OK so precisely what command is it running that's editing files without using apply_patch then?

[lunar aurora]

it did say it failed to use apply patch

here

[cyan wing]

Did they reset the limits AGAIN??

[simple star]

@cyan wing Happy xmas

[boreal holly]

Oh gotcha, so on Windows, Microsoft decided to use CRLF line endings, whereas all other systems use LF. You have git for Windows configured in a way that Codex is trying to apply a patch using LF like it was trained, but since the file has CRLF line endings it is failing the size/parse check

So perhaps tell Codex about this and see if it can discover the right way to do line endings

[lunar aurora]

i mean it gotta be a permanent solution for this instead of having to tell it every time

[boreal holly]

Or, perhaps it's writing a patch file outside of the workspace and it is hitting the sandbox

[boreal holly]

The permanent solution would be using WSL and operating inside that

[lunar aurora]

whats WSL

[cyan wing]

WSL: https://learn.microsoft.com/en-us/windows/wsl/install

[boreal holly]

Follow the guide here https://developers.openai.com/codex/windows

[lunar aurora]

i mean can i still be using the codex app though

[cyan wing]

yes, codex app supports WSL in settings

[lunar aurora]

ah WSL

i installed that earlier

GPT told me too lmfao

not sure if i have to set something in configs

for this to work properly tho

[boreal holly]

Man, GPT really gives the best advice imaginable

[lunar aurora]

fr bro

[boreal holly]

The link I shared has precise instructions for integrating and using WSL in codex for windows desktop app

[lunar aurora]

gpt too bro

great i created a pass

lets see now after restarting the app

[boreal holly]

Gosh darn, I was just making a guess earlier! Yeah, one of the most frustrating things, besides files not being openable by multiple processes, as well as everything else, about Windows is the CRLF

[lunar aurora]

but is this a codex issue

or on my end

ill just switch to full access

fuxk that

works now

[tall zodiac]

Yo does anyone run into this lol?

why does codex use 44 gigs LMAO

[cedar skiff]

it doesnt its leaking

[tall zodiac]

is it my fault or the apps fault?

[cedar skiff]

havnet had that with codex yet, had it with claude code many times

not your fault, just kill it

[tall zodiac]

@cedar skiff going to see if codex can fix itself

bc its annoying when I want to run long tasks and then this happens

[boreal holly]

Have codex build a better and truly native app with no leaks 🤪

[tall zodiac]

i made a patched version of the codex app after finding the root cause

[main nimbus]

open source wen

[sacred plume]

I used Codex until 1%, then it renewed to 20% weekly, and now 100% even though it was due to reset on 11th. What gives? Why did they renew it early? (Not complaining)

Anyone have answers lol

I heavy use Codex so its just...interesting

saw the github post

noice

they should regularize gifting quota resets 😛

i'd take ads for it 😛

[glad solar]

yeah probably bc they were havign issues where it was being consumed faster for 0.01% of users

but hey im not complaining

[simple star]

0.01%?

Does that number come from anywhere?

[boreal holly]

OK, so apparently they updated Codex so if a thread starts with certain configs like writable_roots, and you add more later, even if you restart codex it will keep the old writable roots. Basically the sandbox settings are almost thread-permanent

You have to undergo Herculean/Olympic level mental gymnastics to get it to pick up new configs

[lunar aurora]

i really wish there was a plan in between pro and plus

i'd buy a 80$ anytime

[tall zodiac]

Bro 5.4 is a token hog on high reasoning without fast

im burning through limits faster than 5.3 codex

at this rate i will need two pro accounts

[lunar aurora]

on pro? 💀

[tall zodiac]

yep

[lunar aurora]

dude im on plus on a largecodebase and its keeping really well

[tall zodiac]

I have burnt 16% in 3 hours @lunar aurora

[lunar aurora]

i've used it for the last like 2-3 hours on a large codebase still have 83% left on PLUS

[tall zodiac]

of my weekly

[lunar aurora]

hash did you mean plus

and not pro

because pro is unlimited

[pine gale]

Why didn't my usage limit be reset?

[tall zodiac]

im on pro plan

[lunar aurora]

it says its unlimited tho?

on pro

[lunar aurora]

how do you have the same usage as me bro

and im on plus

like

ive used the same

16%

and im on plus

[tall zodiac]

maybe bc im using subagents on plans? @lunar aurora

[lunar aurora]

multiple agents?

ah ive never used that anywhere tbf

[tall zodiac]

yes multiple agents on one thread

@lunar aurora

i dont remember 5.3 codex dropping my usage like this even with subagents LOL

[cedar skiff]

Depends on the task you get it to do, i got it to make a manual testing instruction set for testers to work through. The project has ios, android and web. Took 17 agents and cost 4% of weekly and 15% of 5 hourly

it took 23 minutes

[lunar aurora]

i mean ive never used multiple agents

isnt that a good way to figure out bugs tho

[tall zodiac]

yes if u use code-review tool u need fresh eyes on your code not the same agent that wrote the code

[steel sierra]

Has anyone got actual multi-agent feature working in Windows Codex on WSL? I queried 5.4 Pro after my Codex Windows app using 5.4 Extra High complained 'I don’t have an actual subagent dispatch tool in this environment'

The GPT 5.4 Pro gave this this?

"I found the main blocker: WSL is only the execution environment. OpenAI’s docs say the actual multi-agent feature is still surfaced in the CLI, while visibility/support in the Codex app is “coming soon.”

Separately, the docs frame subagents as an experimental feature you enable in config or from the CLI; the Windows app can do parallel threads, but not that dispatch tool yet.

On the other hand "Peter Steinberger" posted on X that he uses "Sub-agents" in Codex app, Any knows, How? -- see screenshot

[lunar aurora]

thought thats completely fine

code review is that for bugs though?

[cedar skiff]

subagents work for me on mac

[lunar aurora]

"Review against a base branch" is that the one? -- aka CURRENT up to date files

[tall zodiac]

yeah it should catch bugs

[lunar aurora]

is it this

[tall zodiac]

i mean i always have the agent verify its work with tests and ui evidence

[lunar aurora]

glm5 is really good btw

and wasnt pro unlimited gpt 5.4?

they just havent updated the forum from 5.2 to 5.4

[oblique cove]

Have any of you guys been noticing your weekly limit reset daily? it's been happening to me, odd

[tall zodiac]

never heard of unlimited

[tall zodiac]

yes because the codex team is gifting us resets

because they keep finding bugs on their end

[oblique cove]

Oh I thought they would just make it last longer, looked like a glitch lol

[tall zodiac]

nah they just wanna keep the dev community happy

[lunar aurora]

https://i.imgur.com/sUDkg2p.png they just updated it

[oblique cove]

Honestly I'm at like 92% after hours of work, how exactly would I use my whole limit in a day or a week

[lunar aurora]

@quasi canopy

oh gosh

@tall zodiac *

better picture

"gpt 5.4 thinking unlimited"

[tall zodiac]

interesting

[lunar aurora]

so how are you running out

[tall zodiac]

idk

[lunar aurora]

https://i.imgur.com/MDcxL1r.png

i mean if ur using 129038129038 agents that does make sense

[tall zodiac]

@lunar aurora

[potent mason]

Guys is anyone here good at using worktrees + branches?

I'm running into an issue where after I convert my worktree into a branch with a PR, I can't get it to do anything else if I want to continue working on other things at the same time (it starts overriding the other branch I'm working on)

[potent mason]

That's not for codex

That's for gpt website

[oblique cove]

Guys is anyone here good at using worktrees + branches?
I haven't tried those yet

[potent mason]

You can find the limits for codex by looking for them

[oblique cove]

That makes sense

[lunar aurora]

just figured

[cedar skiff]

worktree is a local repo wrapper for a branch.
Their purpose is because a repo wont allow for more than one branch to be checked out.
worktrees solve this problem.

I'm not exactly sure of what you are describing as your problem.

[potent mason]

Yeap that's why I'm using worktrees but from there I have to make a branch to do a PR from it (all good until here), then when I want to address the comments on the PR before I merge I have to do it from a branch (this is the part I'm probably doing wrong), and then from there you can probably guess what's happening (having multiple branches and multiple agents modifying it at the same time they end up writing code in a branch that is not their own).

[cedar skiff]

If i understnad what youre saying is you are trying to do parallel implementation with agents on the same branch?

[simple star]

Gentlemen, I can confirm

[potent mason]

Sort of but the problem isn't the implementation it's addressing the PR comments

[simple star]

5.4 has become IMPRESSIVELY better, as I deleted all my AGENTS.md

It might be time phase that out

[cedar skiff]

parallel implementation is a problem though. The short answer is don't

address the comments with one agent sequentially

[potent mason]

Well that's just not an option that's slow af

If I need it to be working on two things (let's call it feature seed data, and feature desktop auth with PKCE) I should be doing it in parallel

[main nimbus]

I caught up by never making one in the first place

[cedar skiff]

worktrees are the solution to parallel work flows.
What i do is have several worktrees going with different tasks and i merge them to main as they are completed.
So i'm not waiting around for sequential work flows. I do other tasks in other work trees.

[potent mason]

Yeah but most people can't merge to main immediately (first it needs to pass AI code review, then a human needs to test things out in the preview deployment, then all the comments have to be addressed and only then can you push to prod) that would work but it's like a glass cannon

[cedar skiff]

Yes, we are on the same page, so i have sequential workflow on worktrees they all go through automated reviews I'm like code rabbit atm.
I work on a worktree branch sequentially until it's merge ready. Any time I need to wait for that workflow I work on another work tree on another task - also sequentially.
When i merge i do a full round trip.

[potent mason]

How do you work on something else when the PR comments are now being addressed on a seperate branch?

[cedar skiff]

I dont have a seperate branch, I make a work tree and work on that tree stays there until its merge ready.
When i have to wait i make another work tree from main and start another task.

eventually they all just merge to main as they need

some times i have to sync with main on a worktree because its been a while

[potent mason]

You can do a PR and keep it as a worktree? That's exactly what I was looking for

[cedar skiff]

when you make a pr its to merge a branch, you can have that branch checked out in a work tree

i mean it usually is because the nature of the system anyway

[frosty zealot]

A worktree is essentially just a carbon copy of your repo (but when its created it respects the .gitignore, so your secrets etc. are not automatically moved into the new worktree, I had Codex create a worktree-create.sh script for me that handles creating and symlinking secrets so I can do smoke tests from within the worktree) it has its own branch, its literally just like working in your main repo, on a branch, but with worktrees you can work on multiple branches at once

Which is pretty much what Eric already explained

Codex really simplifies working with worktrees though its really nice, I wish the cli was as intuitive with it 🙁

[potent mason]

Yeah I was aware of that part, what I wasn't aware of was that you could do the PR from the worktree I kept clicking the creaye branch at the top

And it would start causing a mess

[cedar skiff]

create the branch as soon as you make the work tree

usually

[frosty zealot]

The branch name afaik is just copied from the thread name, so if you wanna be hyper efficient, rename the thread to whatever you want hte branch name to be lol

THEN create the branch

im usually lazy and my branch names are just the first 4 words I typed in the prompt lol

I just wish the shift+tab planning mode worked like the CLI where you could set a planning reasoning level, and a default reasoning

Unpopular opinion, I wish they'd just put all their eggs in one basket and stop with two parallel things, just go all in on the codex app and call it a day

I feel like I need to bounce between the cli and the app depending on my workflow at the given time

What font do you guys use im getting kinda bored of Fira Code after years lol

[cedar skiff]

eventually cli will will be niche

[frosty zealot]

Is this a guiding prompt, or does it copy paste verbatim into the body of a PR. etc.

The way it reads, 'Append to pull request bodies' but then it says 'Add pull request guidance' makes me confused on the intent here

[solemn acorn]

did anyone else's usage just get reset?

[frosty zealot]

Earlier yes

[lunar aurora]

wouldnt this be a worse model?

[frosty zealot]

Negatron

[lunar aurora]

ah its the same but 2x plan usage

[main nimbus]

yep, 1.5x speed for 2x cost.

[oblique cove]

Actually no, fast doesn't mean a tiny/worse model. OpenAI had a partnership with an inference company that can run their huge models faster than they can, but that company doesn't have enough servers for everybody, if you can use fast mode you should

[lunar aurora]

"uses 2x plan usage"

and its not even double speed

so its like

[oblique cove]

Your choice

[lunar aurora]

the margin is off but

[frosty zealot]

Technically, you'd be just operating at default usage quotas since we're currently at 2x until April lol

[oblique cove]

True

I think there's fast mode in the codex cli too, I haven't updated though

[main nimbus]

yes /fast

[oblique cove]

when I get it it'll be so cool

[tall zodiac]

https://x.com/thsottiaux/status/2030474136024400173?s=46&t=9l1KwyEX9uwBxqfL3uX3zA

Expect more resets fellas

🤓

[simple star]

I wish there were automations in CLI

[oblique cove]

What kind, couldn't you just press the up-arrow until you find the prompt you want to use again

or like scripting codex to do something with a given prompt from a ./script.sh file

[frosty zealot]

Lets just keep banding together and saying we're draining quicker than expected so they keep resetting and investigating

[potent mason]

They reset again? lmfao

[lunar aurora]

whats this again

[simple star]

It is suggesting you connecting to Linear's MCP server

Do you know what Linear is?

[potent mason]

I'd rather they spend that time developing the app then get a little bit more usage out of them

ngl

[oblique cove]

I'm very excited for the codex app to drop for linux

[frosty zealot]

You're just hoping they're looking in here arent you

I wonder why they havent yet, unless they're developing a socket

[lunar aurora]

dude it keeps rejecting if the file is too large, even 600 lines bro 💀

"Good news

Once Codex switches to stdin patching, this problem disappears entirely.

And they likely will — because it’s a known limitation."

[deep tide]

Codex is stuck thinking after a crash is there a way to force stop the thinking in vs code because the stop button isnt working.

[obtuse river]

Collective action hell yeah!

[deep tide]

Found a fix by archiving the chat, the getting it back.

[amber scaffold]

Does the loading spinner actually spin for anyone? I've never seen it do that.

[frosty zealot]

Is there a way to intertwine codex with LSP servers?

[amber scaffold]

Support doesn't exist. See: https://github.com/openai/codex/issues/8745

[high girder]

If you use Codex through VS Code, then when you use language extensions, there's usually a built in LSP behind it which does get used. If you wanted it in the app or cli, you'd need to build it, but it can be done

[cedar skiff]

claude has lsp support it doesnt really help much

[true bane]

thank you to whoever keeps randomly hitting the "reset the $20 tier guy's weekly usage" button lately I appreciate it

[solemn acorn]

you can already install it in linux w/ https://github.com/ilysenko/codex-desktop-linux (ask codex to install itself)

or you can use t3-code which is a codex client that was just released

[potent mason]

I love workos but $100/mo for a custom domain is so excessive for most internal company apps

[cedar parcel]

Anyhttps://www.reddit.com/r/codex/comments/1rnpm9a/weekly_limits_just_got_reset_early_for_everyone/

I didn't check yet, did they reset?

[main nimbus]

Yes a while ago

[cedar parcel]

Awesome! Back to work... 😻

[warm pilot]

Saint tibo 😉 https://x.com/therealadamg/status/2030494185791344738?s=12

[tall zodiac]

Bro does anyone else suffer memory leaks from codex app lol

people are saying the mac os app is good on twitter but i keep getting my memory destroyed

[velvet wren]

I am running the Codex App on my MacBook Pro, M2 Max with 96GB RAM on Tahoe 26.3.1 and no memory leaks

[tall zodiac]

how long have your tasks been

when mine go for an hour or more

my memory spikes and crashes

[velvet wren]

I did a refactor yesterday which ran for almost 2 hours

[tall zodiac]

maybe it could be because i keep polling the github action checks forever to resolve pr comments and stuff

and polling can cause it crash?

[cyan wing]

how Plan mode feels like

[oblique furnace]

guys so whats better, gpt 5.4 xhigh or gpt 5.3 codex xhigh

[chrome raven]

just use 5.4

[velvet wren]

5.4 is the way to go

[oblique furnace]

alr

thanks

[rich ridge]

5.4 way faster

And better

[warped pine]

5.4 isn't really faster in many cases

I found it to be closer to 5.2 in speed in my tasks

[high girder]

Is 5.4 still taking 30% more usage?

[rich ridge]

Well idk about using open claw

[tall zodiac]

How many tasks do you guys have in your plan usually?

[turbid crane]

I want to report a bug. I keep seeing an “Action required” warning message on the Codex Code review settings page. I don’t recognize who this account belongs to, and after clicking the Update button, nothing appears. How can I clear this message?

[cedar skiff]

wrong browser sign in maybe?

[turbid crane]

Impossible, I tried different browsers and the same message appeared.

[cedar skiff]

so you have the browser logged into that github account?

[turbid crane]

Yes. For sure.

[obsidian prairie]

anyone using 5.4 model in codex?

[cedar skiff]

I do

[turbid crane]

See video

[obsidian prairie]

thoughts?

[cedar skiff]

I like it

seems better to me, a lot of ppl dont like it. Its a bit easier to prompt than codex 5.3

[obsidian prairie]

cool, gonna give it a go today

[tall zodiac]

POV: I was at 78% an hour ago on weekly limit now im at 68% not using fast mode and on 5.4 high with subagents

😐

if this is the burn rate with 2x limits imagine 1x

💀

[oblique furnace]

why is codex on windows so laggy

it just basically froze for a few seconds

why is everything getting rejected

all the time

[native basin]

why does it take so much to get this to work on git?

why can't it just be exactly like claud? you literally click on your repository and move on

i have never seen so many errors in my life

with any program ever

[cedar skiff]

Kind of just worked for me o.0. Are you on mac or windows?

[native basin]

windows

i was kind of excited too. i'd been waiting for this for a while

and then just out of nowhere it's impossible to connect to git

[cedar skiff]

🙁

[native basin]

any ideas?

[cedar skiff]

¯_(ツ)_/¯ I'm on mac

[native basin]

i can tell that by the amount of emojis you use

[cedar skiff]

seems like a slightly jaded response for a normal conversation. Good luck with it 🙂

[native basin]

yeah my bad 67

[cedar skiff]

Imagine thinking emojis are the reason mac has better support.

[native basin]

why are you so defensive? 💥 🏹 🧠 🐾 🧪 🩸 🎯 i believe you.

[cedar skiff]

The people who write the software use mac, that's probably why.

[native basin]

🟡 ✅ absolutely

[oblique furnace]

do you guys think codex would be able to code like a trading bot 🙏

[velvet wren]

been there, done that, got the t-shirt

[oblique furnace]

what t shirt

[oblique furnace]

oh

do you think it could code like an arbitrage bot or something?

[velvet wren]

yes, I don't see why not

[oblique furnace]

idk if i should do this tho

like

idk what to code with codex

i want to do something that generates money

[wheat yoke]

Why isnt the 1 mil context on codex :(

for 5.4

[plucky halo]

It is - you have to enable it

But I wouldn't

[ebon skiff]

i think it for gpt-5.4-codex, and its not available yet

[plucky halo]

You and 8 billion others 😛

[wheat yoke]

How do i enable it

On codex desktop app

[torpid trout]

@boreal holly sorry to ping you.
Can I either pm you or pick your brain here in public channel about your approach to subagents?

I’ve built - I’ve let gpt build something - based on my understanding of what you do and your messages and screenshots you shared here. It built some ogratic monster that does something, but clearly not what I wanted lol.

So I’d like to ask you a few clarifications if you’re willing to spend some 20 minutes?

[tardy mauve]

I just tested codex on mac and it works much better than on windows. it's really sad

I hope it get sorted out

[wheat yoke]

How do you enable the 1 mil context

[tardy mauve]

are you using pro or plus

[wheat yoke]

pro

[torpid trout]

It’s a setting in the config.toml. The doc should have the exact key to use

But it’s probably not worth losing your time on it

[wheat yoke]

Its bad?

[torpid trout]

Well it’s just thinning out, poisoning or how they call it the more context you have

In most cases you don’t need a million tokens context.
I think there’s some test called haystack with needle or so that asks the llm to return specific tokens in the stack and the more tokens the less reliably it does that, by what I understand

It’s like you wouldn’t clean your brain your whole life and expect it to still manage daily tasks.

[tardy mauve]

yeah true

[wheat yoke]

Yeah i saw some posts on reddit saying the more context you use it becomes forgetful

so the compact context is better

[torpid trout]

I don’t even let it compact. I’m out before that happens and start new

At around 30% I kill it and start fresh.

[tardy mauve]

that's actually good tactic

[wheat yoke]

Im using it for binary exploitation and such so when it starts reading assembly it fills up context pretty fast

[tardy mauve]

where are you guys using codex on which os?

[wheat yoke]

Mac

[torpid trout]

CLI Mac

[tardy mauve]

yeah much better than on win

[wheat yoke]

I use the desktop app

its really good

[torpid trout]

The app was erratic and I never tried it again

Probs better now..

[wheat yoke]

I love the app honestly

[tardy mauve]

I switched to mac because codex work better

[wheat yoke]

Works wonders

[wheat yoke]

Same xD

Wish there was a linux app tho

[tardy mauve]

mhh

[torpid trout]

I guess you could make one using codex app server.

[oblique furnace]

nah maybe not everyone but still

[tardy mauve]

it's pain on windows

[oblique furnace]

Idk what to do tho

I have too many ideas

[torpid trout]

It makes me drain. Like I’ve tons of ideas then I go 3 days and then I’m burnt and need to sleep for a full day lol

I think it’s a problem with too much possibility

You start being overloaded and shut down

me, not you.

[tardy mauve]

why does the codex app cant spawn subagents

[torpid trout]

It can, I think? CLI can for sure. You need to enable it thou

In config toml, allow agents and max agents

[tardy mauve]

what about the app?

[oblique furnace]

wait

i think this is it

like

even if somehow all 8 billion humans wanted to become rich and build a successful business or smth

majority of them probably won't ever leave the planning phase

execution is the most important step

it doesn't matter if you got a brilliant plan if you don't start executing it

[tardy mauve]

I also wanted to make something useful with AI but it's not so easy as everyone thinks

[oblique furnace]

i built some stuff with ai already and for me at least it was easy

[nimble sorrel]

Just gotta learn how to prompt

[tardy mauve]

yeah that's true but not for getting money I mean

everyone wants to hop on the AI train

[torpid trout]

The problem isn’t creating something. The problem is after you created something and then something more and something more.

It’s literally like you’ve infinite delicious food.

You’ll eat and eat and … at some point even the most delicious food will be too much.

Overload

[torpid trout]

That’s another issue
Of 8 billon humans only a fraction have ideas, only a fraction of those have the persistence to bring it to a production stage and even less to a maintenance stage.

That’s an old issue, mix of paretho and other stuff

Most folks are „do you have ideas what I could do“… and then if you give them ideas they’re like „but how“.
That’s why business works. There’s a few who do and many who consume.

If we where all producers we would have to sell our products to ourselves

[fallow stream]

i am coding my dads new website with codex and its been good so far. His current one is like 15 years old so

[boreal holly]

Absolutely! Whats going on?

[boreal holly]

That's actually the story for all apps, including Microsoft's own office software

[boreal holly]

I have my whole entire configs posted up at https://github.com/robertmsale/.codex/tree/main

It includes the global AGENTS, all sorts of really tight workflow skills with instructions dialed in for 5.4, and I recently found a workaround for long-running processes (created a wrapper/replacement shell for zsh that makes all commands waitable via MCP). It's pretty much done! Spent ~2 days dialing in everything for the new model

[frosty zealot]

Do you guys think native app building is dead with how far things like RN, Electron, Tauri, Wails etc. have come?

[nocturne folio]

wdym

[frosty zealot]

Take RN for example, or something like Expo, you can eject and still implement native Swift modules or whatever as you see fit, but you also have hte luxury of developing using React, hot reloads, etc. greatly increasing the speed and efficiency you can develop your apps, and the in-app experience is essentially just as smooth as a native app

[cloud cradle]

does anyone know why codex desktop on windows is failing to edit files? by default workspace-write is supposed to be in the permissions, but edits/writes only work when i change it to full access. if i switch to WSL it gives me issues with git, so currently native windows and WSL are both unusable in codex. ive been trying to debug it with codex to no avail

[boreal holly]

No, native app building should become the norm. Typically it's avoided due to development challenges. With AI in the loop the only excuse to do web apps is greed (must ship faster than light). I'm not a fan of "everything web" so my view is biased, but all the web/interpreter-based frameworks exist solely to make shipping on all platforms faster.

[severe torrent]

guys chatgpt 5.4 or chatgpt 5.3-codex for coding?

[torpid trout]

Looking at this I realize you do basically all of this with skills and an mcp

I… asked it to build a totally new codex app so to say lol
I’ll ping you if I have questions but the repo clarified a lot

[hard tulip]

yes

[hard tulip]

native apps are dead

[boreal holly]

Yeah, I reverted from a bunch of MCPs to just 1 MCP and a bunch of scripts. Turns out managing sandboxing and identity through MCP is neither easy nor enjoyable lol

[exotic cave]

I disagree. Native means maintaining completely different codebases and approaches per platform. It's the same reason we moved from assembly to higher level languages, we always converge toward the easier path. And even with AI, you'd still want one shared base rather than maintaining separate ones, AI just makes converging even easier, not the other way around.

[hard tulip]

No point of making the same app 4 different times for minimal performance gains but it depends on what it is for

[neon girder]

native apps almost never have feature parity across all platforms they support, thats the main problem i have with them

look at the chatgpt mac os app (native app), its so outdated

[boreal holly]

Not necessarily. Most of the codebase can be shared by all platforms. The only thing you need to do is write some platform/hardware specific parts, but only if you need those things.

[hard tulip]

yea and then the users get mad because they have have to re learn the app for every platform

[severe torrent]

guys chatgpt 5.4 or chatgpt 5.3-codex for coding?

[hard tulip]

you see this a lot with IOS / android version of the same app is usually way different

[hard tulip]

I think there about the same but 5.4 seams so slow and i am on the pro plan

[severe torrent]

thanks

[exotic cave]

There's no shared codebase in native, Android is Kotlin/Java, iOS is Swift, PC is C#/C++. Completely different languages and frameworks. The moment you share code across platforms, that's hybrid development. You're literally describing what you were't a fan of.

[boreal holly]

That's really a developmental choice.

I write 90% of my business logic/state in rust. I choose Flutter as a renderer in favor of native GUI elements, which gets compiled into native machine code with very minor performance tradeoffs. For very minor platform differences I dig into Swift on Apple hardware, C++ on Windows and Linux, Kotlin on Android. The rust code is compiled as a dynamically linked library by architecture and loaded by platform. We're talking an extremely small percentage of effort going into platform-specific development.

There really isn't much reason to do "web everything" except shipping as fast as light not being fast enough

[frosty zealot]

Why Flutter? I made an app with it like 10 years ago when PUBG released but havent touched it since

Didnt find it too intuitive and it just tried to solve issues that React had already solved

Dart also didnt seem to be widely adopted, and the ecosystem was lacking

But again this was 10 or so years ago

[boreal holly]

Google really cooked when they made Flutter. It's essentially a 2D game engine without a physics system. They created this perfect communication channel to native code using plugins.

[exotic cave]

For 90%+ of apps the performance difference between native and web is invisible to the user. Business apps, social media, e-commerce, no one can tell. And you're using Flutter + shared Rust, that's hybrid not native. Web tech has been the most practical path since before smartphones existed, AI doesn't change that. Even games run on WebGL now, the only exception is AAA titles. By your logic we should all still be writing Assembly, it's the fastest after all.

[boreal holly]

When I say native, I mean "compiles to platform ABI binary code and runs without interpretation". Flutter + Rust qualifies as this.

Furthermore, I'm not saying we should still be writing in assembly, and I'm also not saying web technology should disappear, but now that AI is speeding up development it's becoming possible to make apps that take advantage of the entire system it runs on with minimal overhead.

React native, you're shipping a javascript interpreter. Electron, you're shipping a whole entire web browser. Flutter + Rust you're shipping platform ABI binary code, heavily optimized for the device it will run on.

[hoary turtle]

Why are the new models so bad at Front-end tasks?

It looks like they gets worse every time

[chrome raven]

i think it would probably say skill issue here

[plucky orchid]

Someone know why i don't see 5.4 Pro ?

[boreal holly]

It's only available on the API

[plucky orchid]

I am on the API

[boreal holly]

launch with codex -m gpt-5.4-pro

[plucky orchid]

Got it now

Thanks

[hoary turtle]

Whats that for

[chrome raven]

works?

[plucky orchid]

Yep

Now i need to figure out why he cost 7x the price of Opus 4.6

[chrome raven]

how hash is it on the limit?

[plucky orchid]

wdym

[chrome raven]

or you use api key?

[plucky orchid]

Yeah i use api key

[chrome raven]

oh alright i though you logged in

[nocturne folio]

do you know now? or is it too early to tell

its like 7tps on openrouter lol

[plucky orchid]

Not yet i am finishing ETH and SOL smart contract

Then i will use it with EVMBench to try to find vuln

[tribal lake]

I think there's an error when you switch your wifi internet... it's like the whole codex thread stops working

[desert vapor]

Codex 5.4 is really slow today

I switched my wifi - maybe that's the prob

[drifting granite]

Does anyone have links to benchmarks comparing 5.4 high vs 5.4 xhigh?

[tribal lake]

it may be related to the websocket update they did... it was hanging, so, I had to close and enter the app again to get it working

[desert vapor]

Thanks. I just rebooted everything.

[potent mason]

Look at the official benchmarks openai released

[worn harness]

Does anyone know of a way to see hourly codex usage instead of daily?

[lavish geode]

Guys please fix the limits drain
If codex stays this it will be unusable and expensive after the 2x finish and will be another claude

[late root]

im edging myself.

[plucky halo]

Phrasing

[frosty zealot]

https://tenor.com/view/red-card-referee-red-sent-off-dismissed-gif-21029120

Can I have GIF access in here or what

[boreal holly]

One thing that really helps is using a VPN like wireguard. The server with 99% uptime maintains a solid connection with OpenAI's servers, and you can roam freely. Wireguard will buffer the connection. I use it for SIP phone systems which are very sensitive to roaming internet connections (swapping from cellular to wifi) and it probably would work in this case

[nocturne folio]

will we see a gpt 5.9?

what about a gpt 6.7

[boreal holly]

gpt 5.110.0

[boreal holly]

This is a very important issue, could use some updoots https://github.com/openai/codex/issues/13733

[tardy mauve]

I did not spend 200 dollars for the pro plan for getting a buggy windows app

lol

[mighty cave]

hi can someone please let me know what you did to install the codex on the ubuntu dev server? i tried installing claude; it is successful and it is failing for gemini. looking forward to some suggestions and some commands here. thank you.

[boreal holly]

Have Codex convert this into a dev container:

#!/usr/bin/env bash
set -euo pipefail

if ! command -v docker >/dev/null 2>&1; then
  echo "Error: docker is required" >&2
  exit 1
fi

image_tag="${CODEX_AGENT_IMAGE:-command-parser:0.110.0}"
codex_version="${CODEX_AGENT_CODEX_VERSION:-0.110.0}"

tmp_dir="$(mktemp -d "${TMPDIR:-/tmp}/codex-agent-image.XXXXXX")"
cleanup() {
  rm -rf "$tmp_dir"
}
trap cleanup EXIT

cat >"$tmp_dir/Dockerfile" <<EOF
# syntax=docker/dockerfile:1.10
FROM buildpack-deps:trixie

ENV DEBIAN_FRONTEND=noninteractive \\
    TZ=UTC \\
    PATH=/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin

RUN --mount=type=cache,target=/var/cache/apt \\
    apt-get update && \\
    apt-get install -y --no-install-recommends \\
      ca-certificates \\
      curl \\
      gpg \\
      jq \\
      python3 \\
      ripgrep \\
      tini

RUN mkdir -p /etc/apt/keyrings && \\
    curl -fsSL https://deb.nodesource.com/gpgkey/nodesource-repo.gpg.key | gpg --dearmor -o /etc/apt/keyrings/nodesource.gpg && \\
    printf "deb [signed-by=/etc/apt/keyrings/nodesource.gpg] https://deb.nodesource.com/node_20.x nodistro main\\n" > /etc/apt/sources.list.d/nodesource.list && \\
    apt-get update && \\
    apt-get install -y --no-install-recommends nodejs

RUN npm install -g "@openai/codex@${codex_version}"

RUN groupadd -g 1000 codex && \\
    useradd -m -u 1000 -g 1000 -s /bin/bash codex && \\
    install -d -o 1000 -g 1000 /home/codex/.codex && \\
    install -d -o 1000 -g 1000 /workspace

ENTRYPOINT ["/usr/bin/tini", "--"]
CMD ["bash"]
EOF

docker build -t "$image_tag" -f "$tmp_dir/Dockerfile" "$tmp_dir"
echo "Built image: $image_tag"

tell them to ignore the command-parser and cleanup stuff

Also have em make the permissions more friendly and include whatever tools you need for your project

[raw hill]

and ask it to make a docker compose file, for envvar and file mappings

[native basin]

so a quick question for the codex pros. i just downloaded the codex app for windows so it only does local repository and not online. is this correct? can someone tell me why the hell they did that and what the point of it is? or am i wrong and there is a way to actually do it?

they finally give windows some love and then completely destroy all my dreams in one app

i hope i'm wrong

[nocturne folio]

there is codex web if you wanted it

pretty sure its https://chatgpt.com/codex/

[native basin]

oh yeah i use that frequently but i was hoping for the app to actually do it just like claude's

i'm just mind blown that they took all this time to make an app and then left out one of the most crucial uses

[nocturne folio]

i think you can use codex web on the codex app, but im not sure.

[native basin]

yeah it would be amazing if i could get some clarity on that so i don't waste so much time trying to do it when it doesn't actually work

[tall zodiac]

What do you guys run for your agents configs ?

Like max threads and max depth

[boreal holly]

you mean running it on someone elses computer? They already offer that

[native basin]

no i don't mean running on someone else's computer. i mean simply connecting to an online repository instead of local

[tall zodiac]

I tried running max threads at 70 and max depth at 4 and my computer kept getting nuked no cap

100+ gig usage

Lol

[native basin]

yeah if this existed on the windows version, i'd be so happy

[boreal holly]

git clone

[nocturne folio]

pretty sure it does

[native basin]

pretty sure?

obviously i have a local project ticked because it wouldn't let me connect to the git on this setup

but it's never given me an option to connect to git

[nocturne folio]

thats weird, maybe its because the workspace isnt git controlled or something

[native basin]

and any option that i've seen just errors out

i've reinstalled twice

[boreal holly]

So instead of creating on GitHub and cloning you created locally and need to connect to GitHub?

[native basin]

well i mean i did that because i had to just to test it. it wasn't exactly what i wanted to do

on the set up it gave me an option of two things:

1. git
2. local
   every time i pressed the git it would error out over and over again. i just connected local project so i could actually start the app

i mean it's completely fine if it doesn't work. i just want to have clarification that it 100% doesn't work for windows currently and then i can move on

[nocturne folio]

maybe its because you dont have the github app connected, but since you use codex web i dont think so

[boreal holly]

Gotcha. I would just git clone the repo, or open a repo that's already cloned. make sure your git config is set up properly. Make sure you have the SSH key in your git bash and it's uploaded to github. When you open the project it should just use your local credentials and remote setting

[native basin]

to be honest with you everything you just said was kind of french to me. i sort of understand what you're saying but it just seems like a hell of a lot of work whereas you can clearly see on the mac version that it's literally one button

i guess i'll just wait until then

i don't want to mess anything up

thanks for your help guys

[boreal holly]

That picture 90 shared, that button only appears on local repos with GitHub remotes and authentication set up like I described. Also I think you gotta log in on the web version and connect GitHub<->ChatGPT

[native basin]

yeah i mean i'm definitely connected already 100%. i use the web codex daily that's connected to my github but i mean i'll try again. i'll log out of everything and see whether i'm missing something.

thank you

[cedar skiff]

config is hard, swe isnt going anywhere quite yet.

[native basin]

all sorted thank you. for some reason i needed to disconnect the git from online and then connect it back on again. it immediately recognised it.

windows on top

[cedar skiff]

Going to give symphony a run on a flutter app i've been planning

[potent mason]

It's definetly not one of the most crucial uses, it's one of the least important features imo

But I'm checking it out and on Mac at least it is available

Oh wait actually you just have to set up the environment

Anyways never used it before

[kind jay]

What’s the problem with cloning it and pushing the changes?

[potent mason]

Has anyone used the new codex security feature? I just saw it's available online

[kind jay]

Maybe there’s an update in the app?

[boreal holly]

5.4 cracks me up dude 😂

[oblique cove]

Tysm

You guys should try uncodexify, it's just a list of hard nos for codex when designing so it doesn't make ugly websites / UI

[main nimbus]

I love this. You should link to the repo, tho: https://github.com/cyxzdev/Uncodixfy/

[nocturne folio]

lovable is not giving me the 100$

[scenic cape]

• Codex web is still using I’m running GPT-5.2-Codex in this environment.. Why?
• When is codex web getting a model picker?
• I have run the same task 3 times keeps running in a loop and fails after 40mins.

[boreal holly]

Yeah, after 4 days of revamping, 5.4 is absolutely killing it now. Finally got full autonomous operation back

[torpid trout]

Is sad this is needed

I hate definitions defining hard coded values like pixels or „dont do this“ etc yada yada but unfortunately gpt really Jas a knack for 1:1 identical ui across the board. Had some users in here being hard on offended in past for showing how „their“ design was identical to „mine“ lol… but that’s gpt for the world. Not made for design, unless you hard drive it

[main nimbus]

Negative constraints also impact performance, and yet this prompt works well for 5.4. The other lab has their equivalent skill that works well with their model/harness.

[solemn acorn]

that does make sense, the main advantage of GPT-5.3-Codex and GPT-5.4 is that they're more conversational

[frosty zealot]

Yep

[solemn acorn]

and you don't really need that in the cloud

[potent mason]

How is it?

[scenic cape]

What's the point of using the Codex Web then? If it's stuck on a very old model now

[cedar skiff]

I gave it a run, it does find gaps, but it doesn't have real understanding.
For example i have an unawaited aysnc function that removes a user from a sharing contract. The user is part of the internal system, and the sharing contract is only between verified users.
During that async timing the user still has access to the unshared data. It flags issues like this.
To be fair i only used it with defaults and havent built up any harness base for running scans yet.
So it can see problems but it doesnt have any understanding.

I fixed several issues it flagged, and flagged a few like the one i mentioned

[upbeat maple]

Have you folks ever had the issue where you asked codex to move just an object on your screen just a bit and it says it's changed code but you look and nothing has changed? It's been happening numerous of times.

[orchid plume]

I was curious about trying Codex Security but when I discovered that your inputs/outputs can be used for training, I didn't go ahead. Looks like an interesting and potentially useful product for the future though when it's out of research preview

[cedar skiff]

I'm ok with the code being used for training.

[scenic cape]

That happens with all models/agent cli's

[cedar skiff]

There are no diffs in the code at all?

[upbeat maple]

Nope like I want it to slide over a bit and it says it's removed a line of code and added a line of code. So I imagine it did change the code but upon observation the numbers change but the objective wasn't met.

[upbeat maple]

Anyway to minimize this for future prompts?

[sand shuttle]

i mean reviweing ai written code with ai?

[cedar skiff]

OK so the problem is it didn't solve the issue correctly, although it did edit the code in an attempt?

[upbeat maple]

hmm yes that'd be a fair summarization of my problem. Maybe because they assume all they had to do was change the numbers and didn't realize that the object was inside another box so it won't allow it to move. But sometimes that's not the case at all.

[cedar skiff]

Some times they make mistakes and can't see the forrest for the trees just a like person.
I would look at the code and direct it more accurately.
If you can't do that you can try to refocus in a new sessions by telling the new agent that there have been several attempts that seem like obvious fixes and they have failed, and it should take a step back analyse and investigate the bigger picture to understand why obvious fixes wouldn't work.

[upbeat maple]

Good ideas. I'll try these out over the next couple of hours. Thanks.

Also. Has anyone read any of this stuff? Any cool stuff with this update?

[knotty owl]

Hey Everyone,

I am migrating from Claude Code to Codex due to cost, and I have not been able to find a good reference for codex setup with sub agents and skilss, can any-one help in providing the best references for this?

I am in data science , my previous sub-agents in claude code included - architect, code reviewer, tester, applied scientist, research scientist, graph-engineer, back-end dev, graph engineer, ml-evalutor and demo-engineer

But not able to find trustworthy references or practices to adopt for codex so please help

[velvet wren]

are you on a Mac or PC?

[knotty owl]

Mac

[knotty owl]

And remote development is linix

*linux

[velvet wren]

get the macOS Codex Desktop app and that will make everything easy and smooth

[knotty owl]

@velvet wren - cant use the app due to admin policies

[velvet wren]

if you can't run apps then how will you use Codex?

[knotty owl]

Cli

[velvet wren]

you can install CLI on your Mac but not the Desktop App?

[boreal holly]

Hey, if it's a MDM restriction that usually only applies to app store apps. This one is notarized

[knotty owl]

Yes funny but true 😣

[boreal holly]

Did your administrator install Parallels?

[knotty owl]

No i can install the app but its usage gets flagged and have got warning! Cli is approved officially though

[boreal holly]

Interesting. Was gonna recommend installing a macOS VM and running it in there but sounds like that's off the table.

your best bet is probably the docs https://developers.openai.com/codex/multi-agent and some elbow grease

[edgy carbon]

Does anyone know anything about CI/CD with the codex github application? How would you choose a model? Does it choose the latest model as soon as it is available? Thanks.

[knotty owl]

Yes thats off, so then i cant creat sub agents for my peoject the way we have in claude? Because creating global agent profiles will affect all the peojects

[sharp tendon]

Anyone know when we will get access to codex security on plus accounts?

[oblique cove]

No clue, best thing to do if you wanna know is check the X account of the developers to see if they said anything

[cyan wing]

how 👀

[sacred minnow]

Why do I keep seeing post on reddit about resets but check my business acc and its all still maxed 😆 🤣

[sacred minnow]

So far im loving 5.4 but my tokens maxed on my accs seemed codex 5.3 was going pretty quick too I think it was just the fast option but all and all love the new features and model

[west spire]

How do you guys use agent teams and subagents on CODEX?

any docs to see some sort of tutorials or so?

[sacred minnow]

I tend to prompt it if I havnt added a skill/workflow in

Ask for a team of subagents and give it a suggested team depending on the type of outcome or project

[west spire]

interesting. Do you have any intel on the matter? If not, it's ok. I can try these comments by myself.

[sacred minnow]

I started originally passing my project from git into chat gpt then asking for a suggested team for what im doing ie as example id ask for a coder tester verify and so on lol sorry I'm not really good at my spelling lol thank god for autocomplete and ai haha

[potent mason]

A reset would really help me out right now

I don't know if any admin is reading this cough cough usage is flying definetly not making it up to get it reset

[frosty zealot]

Im calling the police.

[cedar skiff]

in the codex app it never seems to check off the list items

[gloomy veldt]

Does codex app will be available on Linux?

[simple star]

@gloomy veldt Eventually

Theres a waitlist

But I wouldnt wait standing

[cedar skiff]

when working with skills the models always think the skill meta data injected into the system prompt comes from the agents.md. As a consequence they always advice to put a rule in agents.md when they make one.

[velvet wren]

0.112.0 is out

[rich ridge]

Hi guys I made mcp servers for fetching YouTube transcripts articles contents and validating sources any where upload to community

It’s for free

Just wanna share it with community

[potent mason]

Bruh I'm so jealous of your 128 GB RAM

What is it a mac studio m3 ultra?

[karmic shadow]

Its legal to have many gpt pro/plus acc and use them on codex?

[potent mason]

Technically no

But I mean

They probably won't do anything

[karmic shadow]

Its not? I probably buy another sub. Where is said that that is against terms?

[potent mason]

Why don't you just look for their terms of service paste it into GPT and get an answer?

I'm not an AI I'm not going to give you an exact source

But from what I've heard they don't technically allow that

[karmic shadow]

Cos is not listed

[oblique cove]

The limits are so high, resetting daily; I'm just using 5.4 High to review my entire codebase multiple times

I think 5.2 base consumed more tokens than 5.4 does

[amber scaffold]

They claim 5.4 is more token efficient

[oblique cove]

I think 5.3 was made more token efficient, and 5.4 is 5.2 with a better lora adapter so a re-run of 5.3

I can only guess

[oak hill]

Is GPT5.4 codex service down?

[lime inlet]

How can I install codex on a Linux HPC cluster when I do not have a superuser / root account, only a normal user account? Is that possible? Like, just for comparison Gemini-CLI can be installed via conda thus circumventing the need for a root account.

[torpid trout]

Yes it’s explicitly allowed, even using the same phone number up to 3 times.
@karmic shadow fyi

Many of us here are openly carrying 🤣 more than one sub I mean

[oblique cove]

no, try re-opening your terminal or app

[oblique cove]

usually you need sudo permission for package management on linux

Unless you already have npm installed which I don't think needs a password

Or if you don't care about organization you can git clone codex and build it in which case you can do everything from your home dir

[plucky orchid]

Did you know how can i get the 1M context with 5.4 Pro G ?

[potent mason]

That's nice

Also good cuz I'm devouring my limits this week

I'm past 50% and it's been like 2 days

[oblique cove]

The weekly limit is resetting every day

when it hit 12PM earlier I got another whole weeks limit

so burn as much as you can

[velvet wren]

the limits were reset because of some bugs with Codex

the limits won't normally reset arbitrarily like that

[lime inlet]

Packages in the conda ecosystem, just as an example, typically do not need sudo.

sudo is typically only needed for things that you install with apt, zypper, etc

[oblique furnace]

when will apply_patch be fixed on windows codex

[lime inlet]

If someone had success installing codex in a HPC environment (I happen to be on Red Hat Enterprise Linux release 8.8 (Ootpa) if it matters), thankful if you could drop a line how you did it, thanks!

[oblique cove]

would it be possible to just clone and build it from your home dir?

then there's no need for installation

[lime inlet]

And then "just clone and build" each time an update comes out?

[oblique cove]

I mean if you need it asap

you may only need it for a day, I don't know

[wicked briar]

are you using /fast?

[pastel bobcat]

codex is very slow today, normally it starts quickly but today its like taking minutes before anything happens

even spark is slow

3 minutes before it even touched one file when I asked it in a new session to review the uncommited files

[wicked briar]

hmm

[ruby kestrel]

Yeah, I can second this

[plucky halo]

Yep - codex is definitely playing silly buggers this morning

It's both for me

[pastel bobcat]

yeah, also both. Switching to opus 4.6 :S

[orchid plume]

also noticing slow response times from GPT-5.4 at the moment, way slower than usual. I guess it's under load

[exotic cave]

same for a small question it still thinking for 3m and going.

[orchid plume]

mine has been stuck at last reading a file for nearly 8 minutes so far 😄

[agile fern]

same, codex is extremely slow right now and its not usual "under load slow"

well, i guess its time for day off then

[crystal gate]

yeah something is going on, right?

[ruby kestrel]

No DevOp or InfraOp here to ping?

[crystal gate]

at least it finally did respond

[ruby kestrel]

Yeah, re-prompting helps sometimes here by stopping and continuing the conversation. Still a workaround

[modest oyster]

is it just me or is gpt5.4 being extra stupid today? it doesn't seem to be getting much right

[eager ridge]

I can confirm. Perhaps they wanna normalize limit reset. lol!

[glad wind]

Does codex suck at using code auto-generation tools? I noticed that it was manually generating alembic migration scripts so I put this in my AGENTS.md:

- After updating SQLAlchemy models, ALWAYS create a migration script by running the command `alembic revision --autogenerate -m "A descriptive message for your changes"`. NEVER create a migration script manually.

And it still never does it

[pliant quartz]

In the meantime the status page says everything is fine , seems like they don’t know about this issue

[orchid plume]

probably not the kind of issue that gets automatically flagged

and I guess the team largely responsible for handling this are asleep

[warm ruin]

bro why is gpt 5.4 not usable today?

[pliant quartz]

I have the same issue with other models . Seems like Codex is just very slow for some people, but no one at OpenAI is working to fix the issue yet

[warm ruin]

thats quite unfortunate

i found an issue here https://github.com/openai/codex/issues/14071

[orchid plume]

performance appears to be picking up again now

nevermind, spoke too soon, it was fine for a bit, now it's going back to stalling

[warm ruin]

https://x.com/minttttt____/status/2030978599063298449 lets grab their attention up up

[pliant quartz]

Still slow on my side

[kind jay]

Credit reset incoming?

[orchid plume]

maybe, but I can't see that happening for this kind of issue when people can't use their quota much anyway 😂

there's probably going to be another quota reset at some point this week for a different matter though, the ongoing reports of higher than expected quota consumption

[frail mantle]

Is GPT 5.4 painfuly slow lately ?

[crystal gate]

yes. not only gpt though, gemini and others as well...so probably one service started having issues and then everyone started doing the others and they couldn't handle the spikes.

[frail mantle]

Crap

[modest oyster]

slow and pretty useless. i've just had to revert the garbage it's created today and starting over with 5.3-codex

[frail mantle]

I have same issue with 5.3 codex

[modest oyster]

RIP vibe coding, 2025-2026

[frail mantle]

I was doing an automated QA process 🙁 😄

https://status.openai.com/ strange that they don't show any issues.

[sand shuttle]

time to go back to binary

[crystal gate]

i never have understood why so many companies have status pages that...aren't correct

it's worse than not having a page at all

makes you even more frustrated than if there wasn't a status page

[frail mantle]

couldn't agree more 😄

[frail mantle]

but it's not totally off, I don't have the same slowness in Codex app.

[strange wren]

5.4 broken for everryone?

@mental tide @winged ore @modest stream

[tall zodiac]

my 5.4 is chillin whats everyones experiencing?

[strange wren]

Reddit threads, openai community forum, discord

Everyone complaining

What the hell is this

And zero status updates from openai

[bold root]

Please check https://chatgpt.com/cyber.
Does it show that your account is verified?

[strange wren]

No it does not

Why does this matter ?

[bold root]

In the past, if you did not do this cyber authentication, you would be routed to a different model.

[strange wren]

Im not sure this is the issue

As dozens of users are complaining

[worn harness]

my codex is not responding either, weird

[boreal holly]

I gave 5.4 the ability to ask their project orchestrator for more agents to manage complexity, and unlike 5.3-Codex and below, it instinctively chooses no fan out when it makes sense. I love it! So stoked!

[bold root]

I was authenticated, but then I was deauthenticated.
And then it started slowing down like this.
I think it's probably being routed internally.

[pastel bobcat]

for me disabliing /fast stolved most issues

[tall zodiac]

@boreal holly do you ask codex to use subagents explicitly or does it know to use them when needed? I find that if I don’t tell it to use them explicitly it kinda never uses them

[supple perch]

0.0

[boreal holly]

I don't have "subagents" in the sense that one agent owns a group of other agents. They're more like "peer" agents. Regular threads, but with the ability to send each other messages. The orchestrator is the only one with the power to spawn agents, and when it does, they ask the spawned agent if they need additional agents

Workers are assigned some super detailed plan, ask the orchestrator for additional help, and get put in a thread group where they talk to each other, create dedicated worktrees, coordinate DTOs, send milestone progress reports to each other and to the orchestrator. If a worker is waiting on work from another thread group they get paused, as soon as that work is done and merged, that agent rebases to pull in the changes and continue working.

[turbid axle]

help, bring back codex

[agile fern]

interesting! Do you have it open-sourced by any chance?

[turbid axle]

hey! that was MY idea

[finite sorrel]

anyone know why spark is explicitly told to 'NEVER use git' ?

[boreal holly]

my configs: https://github.com/robertmsale/.codex

the app server UI is still under development

[finite sorrel]

oh spark is dead atm eh

[supple perch]

0.0

[kind jay]

Too slow I guess

[tardy mauve]

spark doesnt work for me

even other models are just stuck

[strange wren]

Everything is stuck

[turbid axle]

codex app still works for me. but cli is dead

kaput

[orchid plume]

from what I've heard and read, this issue isn't isolated to OpenAI's models

whatever infrastructure the big providers share in common, which are having this issue, is the cause

[worn harness]

I love how every popular product has a "status webpage", yet the real status is only actually visible on downdetector

[pliant quartz]

Which other providers have issues and which models are affected ?

[ruby kestrel]

Maybe a regional thing? Set your flag.

[modest isle]

the codex cli (on wsl) could use better feedback, 5.4 xhigh takes a long time to respond currently and it just looks like it is ignoring input, when infact it is just being really slow to respond to my input

[orchid plume]

unconfirmed as I don't use alternatives generally, but I've been seeing some people elsewhere complaining about Gemini being slow today as well

[sacred hollow]

It's just not doing anything

and it's not even committing for me.

[modest isle]

was having the same, seems the trick is to give it a minute ..

[sacred hollow]

oh?

[sacred hollow]

is this for all the models?

[lyric sierra]

A prompt of "say OK" takes >5 minutes to respond

[pliant quartz]

I’m using Claude and Geminin I didn’t observe any slow down like we have in codex

[modest isle]

dunno, but the symptoms were the same for me (gpt 5.4 xhigh)

[strange wren]

Yea it could be this, works fine my USA colleague

[lyric sierra]

across 5.4, 5.3-codex and 5.1-codex-mini

[turbid axle]

Iran hacking the network cables

[sacred hollow]

I'll see if codex 5.2 works

probably they'll reset usage limits too after this

because It spent so many tokens for nothing

[inland bramble]

Thinking... for eternity.,,,

[sacred hollow]

has an update

[olive tangle]

its workign in Asia and US, only Eu is down

[sacred hollow]

restart it.

[orchid plume]

strange, well, maybe it's just a regional problem

[cyan fjord]

is codex kinda slow today?

[sacred hollow]

yeah

[cyan fjord]

Shame

[sacred hollow]

we're all having the same issue.

[cyan fjord]

To copilot i go then

[sacred hollow]

does it have a CLI

[cyan fjord]

Nah

[sacred hollow]

I'm working on a big project

[turbid axle]

BIG PROJECT

[orchid plume]

yeah it is, it's nearly unresponsive at the moment. On the bright side token consumption should be near nothing as nothing is happening anyway 😄 - probably reduces the chance of a reset though

[sacred hollow]

with 10K + source files

[cyan fjord]

I use VSC anyway, but thing is i would go to claude if usage limits werent bad

[inland bramble]

Well, I have to remove the dust from my keyboard and start to coding... and I don't know/remember how-to do it 😄

[cyan fjord]

Kidna cooked too

[cyan fjord]

Ask your local llm

[sacred hollow]

I see.

[cyan fjord]

I have glmt 4.7-lite local

[sacred hollow]

I can code but it takes sooooo long

I don't wanna spent too much time on typing all the code

I just wanna tell the AI to 😭

[cyan fjord]

Good thing i have copilot from student github lol

[sacred hollow]

I only have 32 gigs of ram and 8 gigs of vram

[cyan fjord]

i have 16gb of vram

[olive tangle]

how long usually are these down times?

[inland bramble]

we are lazy pepople!

[sacred hollow]

does it work for c++?

[sacred hollow]

Idk.. we just need to be patiented

[sacred hollow]

I'm not..

[cyan fjord]

it uses models like opus, codex and such

[sacred hollow]

oh

[sacred hollow]

oh cool

[cyan fjord]

It something like cursor or kilo code

[sacred hollow]

yeah

doesn't even respond to hello;

Wait

it worked

guys

I turned off the fast mode.

let's see

[solemn timber]

it's intermittent 😉 don't get too excited 😉

[sacred hollow]

it works

[inland bramble]

started to work here too

[sacred hollow]

yes, fixed.

[solemn timber]

it's been toying with me all morning 😉

[torpid trout]

Using codex, we will need a whole new type of timetrackers

[sacred hollow]

it works.

[torpid trout]

One that can track 5 tasks being performed at the same time

[sacred hollow]

but no usage reset 🙁

[olive tangle]

still down

[sacred hollow]

works for me

It's committing

[bleak dock]

it still down

[sacred hollow]

Try to disable the fast mode

[torpid trout]

In the meantime, I would never have thought it where so easy, I solved that annoying codex kiling subagents the whole time
now, instead of "is taking longer than expected let me do something", it goes

 No completion signal yet from spec_reader; I am
  continuing conservative passive checks as required and
  will only escalate if it stops showing progress
  indicators.

boom sha ga boom

[sacred hollow]

Try to disable fast mode

[boreal holly]

It's for macOS/iOS only

[sacred hollow]

and I'm on high

[olive tangle]

was off personnally

[supple perch]

gross...

[sacred hollow]

ooh

[bleak dock]

still down

[tardy mauve]

imagine using full access

[finite sorrel]

this turns out to be an issue / bug with base level instructions. within codex harness it's fine, but via codex exec it's told to never use git. Annoying when I'm using Takopi

[sacred hollow]

I see.

[sacred hollow]

I trust on codex

I haven't had any issues

[olive tangle]

imagine manually acepting everything

[sacred hollow]

right !

[tardy mauve]

it can wipe the whole pc

[sacred hollow]

why would it do that?

[olive tangle]

i have a pc, especially for it

[sacred hollow]

and there's something called backups

[olive tangle]

u dont relaly know how it reacts, he is not wrong

[tardy mauve]

someone i kow got wiped his data

[west orbit]

imagine not having to deal with losing data

[sacred hollow]

me too

but I trust it

[boreal holly]

I see you have the agent on your Desktop. I also like to use Codex as a disk cleaner utility lol

[olive tangle]

i use mainly on wsl, but just to try if its still down i plug it fast

[olive tangle]

not wrong

[sacred hollow]

still committed all of my changes, and it works !

did burn 4% of my 5hour usage.

in just 1 prompt; what the heck

[tardy mauve]

im. using mac because codex sucks on windows

[sacred hollow]

yeah, I'm staying with 5.2

[simple star]

Is Codex stuck for you fellas? Mine wont answer

[solemn timber]

5.4 seems to be stable right now, but U've been here before for a while 😉

[olive tangle]

yeah, have to wait a bit, its recovering i think

[west orbit]

mine worked fine for a while and now it got much slower...but at elast it's doing something

[simple star]

Ok, thanks

[boreal holly]

Yeah you might wanna turn off the MCP that's bricked up

[vast crater]

Is OpenAI queueing messages? I send a message, and it's working for like 2-3 minutes, and still nothing happened. 35% context left.

[boreal holly]

xhigh?

[vast crater]

No, High

[west orbit]

someone said turn off fast mode but for me that's even slower than /fast was .. so back on fast mode and it's going OK

[boreal holly]

Interesting... sounds like a bug then

[vast crater]

It finally produced a response after 4 minutes. Didn't see this before but okay lol

[solemn timber]

this is the best it has worked all day for me now

[pliant quartz]

So the issue is solved for everyone? On my side it’s still painfully slow

[olive tangle]

i think its recovering slowly

will wait 30minutes to retry

[fallen pollen]

I had to switch to GPT-5.3-codex, GPT-5.4 was not responding

[fiery wyvern]

Still broken for me

[west orbit]

it's definitely working much better than an hour 1/2 ago ..

[frosty zealot]

Do you guys find 5.4 stops more often to tell you its not done yet?

I feel like 5.3 worked till completion now 5.4 is constantly like HEy we're not done yet but this is where we're at!

like cool i dont care lets wrap it up

[simple star]

Does anyone know to program manually? I tried the docs and they can't stop mention something called "variable"

[frosty zealot]

Thats a very broad question lol

[simple star]

It's hard to believe that our ancestors knew how to do this by hand

My great-grandfather must have been a genious

[west orbit]

i find this happens when the session goes on for too long

[tardy mauve]

bruh this codex still doesnt work

[empty holly]

for me works with VPN in US (5.4 fast)

[bleak dock]

is it still not working for me? is it the same for all of us?

[hot needle]

broken

[tardy mauve]

it works with us vpn

[boreal holly]

Yeah, this might be an unpopular opinion but I think that's a good thing. I put a lot of reasons to stop and give milestone reports for my agents. Giving them "work until completion, do not stop" encourages them to complete by "any means necessary", even if it means nuking the codebase, allowing compilation failures to pass through, etc. If they have many good reasons to stop it gives you a chance to evaluate and recover from real issues.

And if the reason they are stopping is not a good one, you let em know and they continue, update the docs for the next agents so they know that's not a good reason to stop. Basically you want em to cease forward progress if the progress is detrimental to overall success

[frosty zealot]

I understand for a long standing repo with a lot of nuance etc. but when im trying to spin up a prototype test for something i dont need it to stop and be like hey we made the front end but the server still needs to be made so we cant test anything yet

like ok so make the server lets getsomething working here then we'll start to fine tune

[warped pine]

backend issue

hope for more resets haha

[boreal holly]

Gotcha. A quick QoL enhancement, if you have tmux, is have 2 panes with 2 codex instances running, set up the notify= script to tmux send-keys "$NOTIFICATION"; sleep 5; tmux send-keys Enter to a pane where the agent is instructed to issue continuations to the other agent using tmux send-keys "$CONTINUATION"; sleep 5; tmux send-keys Enter. It's not the most robust setup but it does work to some degree if you just want 1 agent to keep going

[frosty zealot]

Mainly just using the App for now cause of the worktree integrations but ill take it on the chin

[solemn timber]

and back to slowness 😉

[kind jay]

Just use fast mode lil bro

[west orbit]

what VPN do you use? mullvad seems blocked

[supple perch]

@boreal holly I mean ppl say the same thing about my feeling with windows so I did not mean that mean, just as a yucky I wish it could support linux too

[simple star]

Now it is working for me

[sacred hollow]

yeah

guys

how can I actually setup mcp_cpp?

[kind jay]

It’s not just a toggle?

[sacred hollow]

oh really?

[west orbit]

💯

[boreal holly]

The best I can do is package the server stuff in a Swift package and let Linux folks build a UI with it. It's 100% native

[frosty zealot]

I gotta say though, watching Codex / Claude execute terminal commands has made me way more efficient in my own personal terminal navigation

[kind jay]

Settings then MCP servers?

[boreal holly]

Seeing them ripgrep everything in a flash has really changed the game for me too lol

[sacred hollow]

but there is no mcp-cpp there ^-^

[supple perch]

wait I was talking about the link, not the ui or server...

sorry for the mix up

[boreal holly]

You mean like how the agents talk to each other?

[kind jay]

Ah, I think it’s just install then plug in above this

[empty holly]

proton

[sacred hollow]

okay

[supple perch]

I must be looking at the wrong thing, sorry to be a bother

[kind jay]

Lowkey ask codex to do it, just elevate the access

[sacred hollow]

oh yeah

[boreal holly]

My configs probably work on Linux just fine, but the way the agents communicate is through the macOS app, and that skill is gitignored for now

[sacred hollow]

wait, what will mpsm/mcp-cpp do?

[solemn timber]

it's definitely slow YAAF is my framework

• 14:55:10 YAAF successfully created the turn and began waiting
• 14:57:14 Codex emitted the first runtime notify saying the turn had started

almost 2 minutes before codex returned anything!

[sacred hollow]

yeah...

[kind jay]

What'd you guys do? My codex is broken now

[frosty zealot]

Listening to Mick Gordon - BFG Division while Codex is just scrolling text in my face is the weirdest vibe rn

[kind jay]

Hmm? What is mcp-cpp for?

[sacred hollow]

yeah

[kind jay]

What's Corebryo?

[sacred hollow]

My old vibe-coded engine 🙂

[kind jay]

Well what's an experimental game engine?

[sacred hollow]

yes.

Now I'm working on a better one..

[kind jay]

We love Vulkan though

[sacred hollow]

does vulkan have an mcp?

[hybrid steeple]

Hi

[kind jay]

Apparently so

[sacred hollow]

ooh, will then codex write better Vulkan code?

[kind jay]

No clue

[dusty spire]

does anyone face this issue whenever attempting to /compact ? I have seen an issues is made already but the fact I am not able to get any long process done got me sus there may have been a bypass for it , ( i am on plus subscription)

[kind jay]

It's a server issue as far as I know

[dusty spire]

It's been a while like that, not a current status or smth

[sacred hollow]

they're probably adding more ram to the super pc's

[sacred hollow]

but well, that happened to me too

but now it looks like it's working again

[strong jungle]

I'm constantly having to compact so often, i know folks here have said raising the token max decreases quality but does it decrease significantly/noticeable enough by using say 350 vs the 258?

[kind jay]

We might be cooked if we can't code our projects without codex

[kind jay]

How'd you get flagged as a scammer? 😭

[boreal holly]

Well since Codex is building Codex, they probably put an arbitrarily low timeout lol like they always do

[strong jungle]

I have no idea, what does it say?

[solemn timber]

hoping we get a token reset for this, I've had partial implemented stuff all day 😂

[fallen pollen]

https://status.openai.com/incidents/01KK9JA8JKQKDW1W24T09NHBYH#updates

[dusty spire]

Is it since v0.112.0 was released or am i wrong?

[boreal holly]

Yes, quality drops quadratically beyond the default ctx.

It's not worth it at all. OpenAI really cooked when they designed remote compaction. They have a special model on their server that preserves the mental state of the agent throughout the conversation, and this works better than having massive context

[strong jungle]

Damn, that sucks 😭 I'll open a support ticket with Discord, thanks for flagging.

[sacred hollow]

I don't wanna get banned lol

[sacred hollow]

It's because you've entered a lot of servers

[strong jungle]

Okay thank you for this, I really appreciate it and won't touch that setting then. What are your thoughts on High vs Xhigh? I've always used Xhigh but I'm reading multiple thoughts and it seems 50/50 on what people suggest.

[boreal holly]

If you're compacting often, you should have the agent run all noisy commands like this noisy-command 2>&1 > result.log and ripgrep the result.log file, instead of letting it just consume the whole output

[boreal holly]

Xhigh performs marginally worse, and it can reason up to 50% of its own context window in 1 shot in that mode

[strong jungle]

Okay, I'll do research on this and figure out how to change since I"m compacting every 5-6 messages.

[strong jungle]

Thanks, I'll get High a try then, maybe that will help solve my compact issue then

[kind jay]

I like it depends on the amount of instruction you provide in one message, and how detailed you are with your descriptions

Sometimes XH can be used to if you ask it for a lot of things at once, in addition to plan mode

[simple star]

It's stuck again -_-

[strong jungle]

Yes, I'm notorious for rambling asking for so many things in one message ( I know I need to get better) lol

[worn harness]

"The robot never gets tired, it works 24/7"

[sacred hollow]

XDD

[exotic cave]

Codex is just all over the place today. and it is stuck now.

[sacred hollow]

idk

[sacred hollow]

yeah , don't blame

[boreal holly]

Y'all still have websockets enabled?

[sacred hollow]

yeah

[boreal holly]

I'm just throwing this out there, for everybody with connectivity issues, consider disabling websockets.

[sacred hollow]

doesn't work

[neon anvil]

OpenAI Status still showing that everything's online?

[exotic cave]

no websockets, it now just stuck can't prompt it.

[boreal holly]

What version you all using?

[olive tangle]

5.4 extra high

[neon anvil]

Opencode & Pi 🫠

[sacred hollow]

5.2 high

[modest oyster]

I need my vibing fix, get the degradation sorted Altman and co

[boreal holly]

codex --version

[sacred hollow]

Idk.. where can u see it

goood

I mean not good

[neon anvil]

finally

[exotic cave]

0.112.0

[boreal holly]

Perhaps npm i -g @openai/codex@0.110.0

I have it running right now, not stuck. Actually my configs are also publicly accessible so if anyone wants to cross-reference their configs with mine I feel like that might provide some insights

I have all experimental stuff turned off and only upgrade if the change log shows real improvement that I care about (model updates, and that's pretty much it)

[exotic cave]

smart decision.

[strange wren]

Are you regarded ?

[boreal holly]

highly

[strong jungle]

How do I access them?

[boreal holly]

https://github.com/robertmsale/.codex/tree/main

[simple star]

I guess it is time to spin up 50 Codex, because this smells like new quota reset 🤣

(just kidding, dont do it)

[sacred hollow]

bro this is not sustainable

broooooooooo

I can't work like this, I'm fr.

[frosty zealot]

@tardy elm Do you use Xcode Cloud or Fastlane for your CD?

[sacred hollow]

It's draining

[summer thunder]

my entire home folder has now a shared icon overlay in the Windows Explorer and according to Codex this is because it is shared with CodexSandboxUsers 😒

[hearty prawn]

is codex down?

[olive tangle]

half up half down

[hearty prawn]

my one chat is down and stuck, can i just somehow spawn it in a new instanace without having to transfer all the convo and feed it via a txt file?

[kind jay]

Honestly how are you using this much?

[frosty zealot]

Im down to 61% myself since reset, but im doing a lot rn

[sacred hollow]

there are over 10K+ source files

[frosty zealot]

Refactoring issues with my CI/CD pipeline, developing two side projects, re-writing a repo from Node to Rust for better efficiencies (it deals with encoding/decoding video), fixing general bugs, and a few other things

[kind jay]

You need to compartmentalise it then

[sacred hollow]

oh yeah, right

[kind jay]

Like only allow it to look at small sections

[frosty zealot]

Typically it does unless you're forcing it to ingest verbatim all 10k source files (this is psychotic behavior btw and you should be locked up if you do)

it should be tailing/heading usually like 200 lines, greping etc.

[sacred hollow]

oh..

13K

[void pelican]

Anyone know if I am allowed to upload the app.asar contents on GitHub for a Native Linux Codex App version?

[sacred hollow]

that's not nothing

but ofc we got a full game engine

[void pelican]

Legally is what I am askinig

I haven't seen and LICENSE.md or any policy that says I can't.

[boreal holly]

Bro you could have the best of both worlds if you convert Node to Deno and write the tricky stuff in Rust as an imported dylib 😈

[floral garnet]

GPT 5.4 is great, but token usage is insane, 5.3 codex high is still damn good to get all work done.

[sacred hollow]

yeah, you're all right

[floral garnet]

burned through a weeks quota in one day on 5.4, it's just insane

5.3 codex lasts a few days

[boreal holly]

I use 5.3-Codex for local code reviews. I still think it works pretty good for reviews and seems faster by default

[floral garnet]

it does code work pretty well too

have not had any faults

[sacred hollow]

same

[frosty zealot]

https://x.com/OpenAI/status/2031052793835106753?s=20