codex-discussions

[kind jay]

I hope you do

Welcome back Terry Davis

[calm aurora]

Thats the spirit. Dont forget the knodwledge that brought you there

[warm pilot]

It's hard to distinguish. I build a web app which I probably never would have built on my own. I understand the code and the structure, but It would have taken me ages to write per hand and learn all that typescript along the way. I still think it's "my" code. But If I would have no idea about the code I wouldn't count it as "mine" anymore. I have a few scripts lying around which are definitely "not mine" as I have never looked how it's done.

[kind jay]

The problem is that there is no knowledge, if it was done manually then at least you learn something

[calm aurora]

The instant there is an input and an analysed output, there is knowledge

🫂

At least the greeks learned that way

[kind jay]

It’s an interesting problem, especially in terms of coding
I think the answer doesn’t matter about what’s yours in some cases

[calm aurora]

Collective knowledge is a easy burden to carry

We are all parts of a big machine

or orbits in space

or parts of a system

[kind jay]

I think often there are two goals, one may be the output of the project itself, in which case who cares, but for this to be reasonable the project should be purposeful
On the other hand, silly projects such as LFS primarily has the value in the learning you do while working on it, and there’s no better way to learn than by doing it yourself, in this case I view the only acceptable usage of GenAI as a lookup (possibly debugging) tool

[calm aurora]

define debugging?

define the system

[kind jay]

Bro?

[calm aurora]

if the system you are trying to debug is economics? or a single file?

debugging hits different

[torpid trout]

Yeah, debugging the economic system lands you in prison or worse usually

Especially if you fix the bugs

Debugging files might be hard thou.
I’d suggest trying with code first 🫢

Sorry I’m in a mood today anyone surprised

[kind jay]

Must be the same mood as everyday

[calm aurora]

What of collective debugging, you can throw all people into prision, some people want that, or you can all people debug the system and promote the better solutions

[torpid trout]

That’d be nice wouldn’t it

[calm aurora]

now have to build reality with codex

hahahaha

[lone ocean]

Hey

Please give me sora 2 invite code

[calm aurora]

this server works that way doesnt it, we chat, the data gets filtred, that helps the model?

directly and indirectly

[torpid trout]

Now that’s interesting
Back in the day when we wrote code, I had 2 gpt in a loop to find the ultimate purpose of life

They concluded with „inexistence is the ultimate truth. Let’s celebrate in silence“
And then they went on agreeing with each other to be silent and inexistent for like 45 usd worth of token

Until I realized it and cut the wire

Result was a 500 pages pdf book half of it „I agree, let’s be silent“

🫣

[surreal wing]

what are we talking about right now?

[calm aurora]

tried to make two gpt chat, is interesting experiment

[torpid trout]

We’re celebrating silence

[surreal wing]

oh ok

[torpid trout]

Up to 4 gpt and 4 Claude

[calm aurora]

differents roles?

[torpid trout]

Claude isn’t very useful for it, it doesn’t roleplay

[torpid trout]

Yes

[calm aurora]

did you see the minecraft experiment?

[torpid trout]

Yes

Interesting as well

[calm aurora]

once you understand the system its doesnt bothers you anymore. Now, what if they are wrong?

another 500 pdf

[calm aurora]

put people taklking about their problems, and they show interesting data

apply more people to the problem

eventually

the problem gets fixed

[surreal wing]

ok

[torpid trout]

and you get more problems

Problems usually are resolved as soon as you look at them with "that look"

[calm aurora]

now do it in countries, universities, social media

millions or thousand of people chatting about their thoughts

[surreal wing]

Speaking of problems, I think I should kill my operating system, it's not possible

[calm aurora]

a system that aglomerates the populations o species insight on matters

[calm aurora]

do you have github?

[surreal wing]

yes

[calm aurora]

post it here

I think you can

literally not long ago noticed the literal OpenAI name

being using it for two years

@kind jay

what do you think?

[surreal wing]

https://github.com/FlowDevsCrew/Fl_operatingsystem

wait one sec

https://github.com/FlowDevsCrew/Fl_operatingsystem/releases/tag/untagged-797f7f56c52ccbb93c06

[torpid trout]

Those are private repos, I think

Return 404

[surreal wing]

does this work https://github.com/FlowDevsCrew/Fl_operatingsystem/releases/download/untagged-797f7f56c52ccbb93c06/Betriebssystem.rar

[calm aurora]

no

can't see the raw in github

[torpid trout]

And no one in his right mind would download and unrar a rar anyway

[boreal holly]

Bro the entire repo is a rar file

[torpid trout]

rawr

[boreal holly]

🦖

[kind jay]

I don't really care

[torpid trout]

Its mice taht discord shows us what was posted, and why it was deleted for a brief second

[calm aurora]

To make the repository reviewable and usable, I would recommend:

1. Extracting the contents of the .rar archive and uploading them as normal source files.
2. Adding a README.md explaining what the project is, its current status, how to run it, and the roadmap.
3. Adding a LICENSE.
4. Creating a minimal structure such as src/, docs/, assets/, and .gitignore.

[surreal wing]

https://github.com/FlowDevsCrew/Fl_operatingsystem/releases/tag/Operatingsystem

oh idk does this work

[torpid trout]

Again, no one is going to look at a rar

Unless you like to live dangerously

[surreal wing]

alr one sec

[surreal wing]

mb

[kind jay]

Zip file in the google drive or GitHub repository?

[torpid trout]

I dont think gh scans them Jane

[calm aurora]

grateful to that news I read about downloading githbu from anyone

[torpid trout]

Drive does. But even then.

[boreal holly]

Tell your agent to extract the rar file, delete the rar file, and commit

[simple star]

"Codex, do your thing"

[calm aurora]

haha

[torpid trout]

Honestly, before you had to review code, no issue... it was not executable, you could even scan it with a ai! now.. you have to read thousands of lines of prompts, for prompt injections, and as such the ai cant help anymore 🙁

Not to mention hidden character attacks, I am farily sure you could do prompt injection with those lol

[kind jay]

And you’re the one who wrote the OS, right?

[calm aurora]

"come on codex, I trust you would download anything I didnt approve"

[torpid trout]

I hope he typed the code off like I do

[surreal wing]

nah it was my friend codex

[kind jay]

Your hopes are far too high

[kind jay]

But it’s your OS?

[torpid trout]

but

[surreal wing]

yes

[kind jay]

I’m confused, how is it yours if you didn’t write it?

[simple star]

I wonder if I could build with Codex a Linux distro... that is Codex-driven.. :/

[boreal holly]

Tell Codex to gitignore build artifacts like the linker outputs

[surreal wing]

i stole it from codex

[torpid trout]

I knew it, you typed it off!

All good then

[boreal holly]

ooh and object files. Definitely don't wanna have all them binary blobs in the repo

[kind jay]

Anything to be right 😭

[kind jay]

Oh so you got codex to help write some of it and wrote it yourself then?

[boreal holly]

Other than that, it looks like the OS has a good foundations. A lotta the basics are there

[surreal wing]

no he did everything

[kind jay]

So does Debian

[surreal wing]

i was just pressing y or p

[kind jay]

So why bother?

It’s a solved problem, even in terms of LLMs generating such

[calm aurora]

where is @wide schooner when need it

[kind jay]

I’m genuinely trying to understand

[boreal holly]

You know Debian exists today because some guy was messing around and accidentally built the most popular kernel on the planet

Tinkering with stuff is how cool stuff exists

[kind jay]

It’s not the same lil bro

[calm aurora]

600,000 docs in my repo, don't know how it reached that in two yers jajaja. I could write a saga

[boreal holly]

Guy's tryna learn OS development, leave him alone lol let him learn and tinker in peace. Maybe give constructive criticism

[calm aurora]

step 1: ask codex to explain git commands

[kind jay]

My constructive criticism is write parts of it yourself

[calm aurora]

Jane

and what are you building?

[boreal holly]

Check this out:

https://github.com/jefftranter/6502/tree/master/asm

This is an asm collection for Apple 1 computers. It's a great place to start. You don't need to go to college like I did, just have codex/chatgpt tell you all about this repo and let it be your inspiration

[kind jay]

Key here is trying to learn, which would be great

[surreal wing]

thank you

[kind jay]

^^

[calm aurora]

hope to git today, to see if you can see Aquiles and give notes

[kind jay]

I’m not allowed to talk about it

I don’t care about front end so I use codex to do this for me, however this falls into my first case here

[burnt trout]

lazy

[calm aurora]

I have a github, roast me hahaha

[kind jay]

Yes

[calm aurora]

wait

[burnt trout]

Yes

[calm aurora]

it is accesible? https://github.com/aquiles-ai123/AQUILES

and if so tell me security breach

[surreal wing]

https://github.com/FlowDevsCrew/Fl_operatingsystem/releases/tag/Operatingsystem

[surreal wing]

no

[burnt trout]

halts maul

[calm aurora]

https://github.com/aquiles-ai123/aquiles-faro now?

[surreal wing]

yes

[burnt trout]

https://tenor.com/view/eye-of-the-ra-chopped-eye-tuff-eye-ts-tuff-eyeofra-gif-13666985803445282339

[surreal wing]

looks good

[burnt trout]

Thanks

[calm aurora]

haha

we could build a repo for algoritmic trading colab

haha

[surreal wing]

do you know why my os lags so much

[burnt trout]

didn't optimized it yet, sorry.

[kind jay]

Probably because you vibe coded it

[surreal wing]

No i coded it with chatgpt together i was leading him

with the prompts from chatgpt

[calm aurora]

I have to present Aquiles project to my university in two weeks, need reviews

[burnt trout]

thats some pro way to do it, respect Reazzz.

is this discord Gambling blocked?

[surreal wing]

can you tell me why it isnt in full screen

[burnt trout]

U need to turn on Fullscreen, in the Video Settings.

[surreal wing]

there are no video settings

[burnt trout]

Bad code then, i can turn on Shaders too.

[surreal wing]

do it

[burnt trout]

i did

[calm aurora]

Reaaz, all the things that you are misssing

[burnt trout]

@surreal wing

[surreal wing]

what am i missing?

[calm aurora]

Video Settings for example

[surreal wing]

oh

[burnt trout]

see?

[surreal wing]

mb

[burnt trout]

Beginner problems, don't worry.

[surreal wing]

your website?

[burnt trout]

i worked with them, yes.

[kind jay]

?!

[burnt trout]

Big Gambling site in Germany.

[surreal wing]

why

[calm aurora]

Brave rules

[kind jay]

Why would I know?

[surreal wing]

you are the coding expert

well codex doesnt know why it isnt in full screen

what can i do now

[calm aurora]

ask him to expand it

[kind jay]

It’s an issue with your VM settings

[calm aurora]

what a are you seeing? and what are you not seeing?

[kind jay]

Or the window itself? Then it’s a code issue

[surreal wing]

uh

[calm aurora]

maybe a problem with renderization

[surreal wing]

ok i will ask codex again

[surreal wing]

stop exposing

[burnt trout]

mb

[calm aurora]

tell codex to send a snippet of the renderization part where it is

and send it here

did you split backend and frontend?

[burnt trout]

@calm aurora may i ask if ur coding?

[surreal wing]

no no yet

[calm aurora]

ask codex to make a backup, and a variant. The variant is separated backend and frontend.

[surreal wing]

else if (g_gui.active_app == APP_BROWSER)
{
surface_text_shadow(s, (u32)(wx + 28), (u32)(wy + 20), "FLOW BROWSE", FLOW_THEME_TEXT_PRIMARY, 1);
surface_text(s, (u32)(wx + 300), (u32)(wy + 20), network_is_ready() ? "ONLINE" : "BOOTSTRAP",
network_is_ready() ? 0x8BD8B5 : 0xE3B36E, 1);
surface_round(s, (u32)(wx + 24), (u32)(wy + 52), 474, 34, 16, FLOW_THEME_BROWSER_BAR, 255);
surface_text(s, (u32)(wx + 42), (u32)(wy + 63), "https://www.flowsearch.local", 0xEAF0F5, 1);
surface_round(s, (u32)(wx + 24), (u32)(wy + 98), 474, 198, 20, FLOW_THEME_BROWSER_SURFACE, 255);
surface_text(s, (u32)(wx + 54), (u32)(wy + 116), network_device_name(), 0x516475, 1);
surface_text(s, (u32)(wx + 54), (u32)(wy + 136), network_status_text(), FLOW_THEME_TEXT_MUTED, 1);
surface_text(s, (u32)(wx + 54), (u32)(wy + 156), network_mac_text(), 0x314050, 1);
surface_text(s, (u32)(wx + 54), (u32)(wy + 176), network_ip_text(), 0x314050, 1);
surface_text(s, (u32)(wx + 54), (u32)(wy + 196), network_gateway_text(), 0x314050, 1);
surface_text(s, (u32)(wx + 54), (u32)(wy + 216), network_gateway_mac_text(), 0x314050, 1);
surface_text(s, (u32)(wx + 54), (u32)(wy + 236), network_debug_line1(), FLOW_THEME_TEXT_MUTED, 1);
surface_text(s, (u32)(wx + 54), (u32)(wy + 256), network_debug_line2(), FLOW_THEME_TEXT_MUTED, 1);
surface_text_shadow(s, (u32)(wx + 312), (u32)(wy + 132), "FLOWSEARCH", 0x3F5466, 2);
surface_round(s, (u32)(wx + 92), (u32)(wy + 286), 338, 36, 18, 0xE4EBF1, 255);
if (g_gui.search_length == 0) surface_text(s, (u32)(wx + 118), (u32)(wy + 299), "SEARCH THE WEB",
0x6C7B89, 1); else surface_text(s, (u32)(wx + 118), (u32)(wy + 299), g_gui.search, 0x314050, 1);
}

[calm aurora]

Then ask codex to analyse the frontend

Aisle the problem to the part that exposes it

think codes as sequential steps

[calm aurora]

kind off haha

today only giving advice and relaxing

[burnt trout]

Ah okay

[calm aurora]

why?

[burnt trout]

just asking

[surreal wing]

Main Problems

1. The frontend is still too monolithic.
   In graphics.c, rendering, layout, animation, mouse logic, app state, scaling, text drawing, icon
   drawing, and browser diagnostics all live in one file. That is the main architectural bottleneck.
2. There are still two different frontend systems.
   graphics.c is the pixel UI, while desktop.c is a separate VGA text desktop. That split is not
   conceptual layering, it is duplicated shell logic.
3. Theme separation is only partial.
   theme.h exists, but graphics.c still contains many hardcoded colors, sizes, and radii. The style.css
   file is currently design metadata, not runtime input.
4. Rendering primitives are expensive.
   graphics.c does a lot of pixel-by-pixel loops for fills, rounded rectangles, circles, shadows, and
   scaled presentation. That is the biggest frontend performance cost.
5. Text rendering is very primitive.
   The glyph system in graphics.c is a tiny built-in bitmap font. That is why the UI still looks rough
   even when layout improves.Main Problems

In graphics.c:

• rendering is self-contained and freestanding
• it has a real backbuffer path
• layout is centralized through gui_layout()
• input state is centralized in struct gui_state
• theme colors are now partially centralized in theme.h

That is enough structure to keep iterating on the UI without immediately collapsing into chaos.

uh

you good tito

[calm aurora]

shoes

[burnt trout]

okay

[calm aurora]

If the UI does not expand sideways, it is probably because:

• it uses fixed coordinates
• it uses fixed widths
• it does not calculate layout from the current window size

[burnt trout]

@surreal wing u need to renew it, or code it again. It's broken.

[surreal wing]

what

[burnt trout]

yes

[calm aurora]

did the backup?

[surreal wing]

yes

[calm aurora]

ask to compile before answering

• backup
• variant
• compile

• manual try

[surreal wing]

• Compiled first. The frontend/backend variant builds successfully:

• output image: Aether64.img

Frontend analysis, based on the compiled variant:

The real frontend is graphics.c. It owns almost everything visual:

• framebuffer rendering
• layout calculation
• cursor state
• animation state
• window/dock/menu rendering
• browser panel rendering
• primitive text/icon drawing

That is the main frontend problem: it is buildable, but too monolithic.

The second frontend path is desktop.c. It is not just a helper; it is a separate VGA text-mode shell.
So the project has two UI implementations:

• pixel GUI in graphics.c
• text desktop in desktop.c

That means the frontend is separated from the backend structurally, but internally it is still not
modular.

What is good:

• the split project layout is valid
• theming now exists via theme.h
• the design source file exists as style.css
• the variant compiles cleanly except for 2 low-risk unused-function warnings in graphics.c:926 and
  graphics.c:243

What is weak:

• graphics.c should be split into renderer, shell UI, browser UI, and text/font logic
• desktop.c duplicates shell behavior instead of being a fallback backend-agnostic view
• style.css is descriptive only; it is not parsed by the OS
• many sizes/colors are still hardcoded in graphics.c, even with theme.h

Best next frontend refactor:

• split graphics.c into:
  • render_primitives.c
  • ui_shell.c
  • ui_browser.c
  • ui_theme.c

[burnt trout]

for sure broken

[calm aurora]

backup
variant
compile
manual try
next steps

[burnt trout]

renew the whole code again @surreal wing

[calm aurora]

backup
variant
compile
manual try
next sequential steps

[surreal wing]

ok

[calm aurora]

backup
variant (modular, objective:"put what you want")
compile
manual try (simulation and human)
next sequential steps (1-n)

then you iterate until satisfaction

[surreal wing]

what should i put there

[calm aurora]

where

[surreal wing]

modular, objective:"put what you want

[kind jay]

😭

[calm aurora]

• Objective
  I want to achieve:

  [clear result]

• Context
  This applies to:

[file, module, route, feature, or system area]

• Current problem
  Right now:

[observable issue or failure]

• Success criteria
  This will be considered correct when:

[measurable or clearly testable outcome]

• Constraints
  Do not:

[list constraints, limits, or things that must not change]

• Optional notes
  Extra context:

[anything useful but not strictly required]

with the best insight from you, better works, but you can build it in the go

[burnt trout]

lass los diggi

[surreal wing]

ok thank you

[calm aurora]

imagine saying: a want "x" but i see "y", Analyze.

try to follow a flow o a loop

[burnt trout]

https://cdn.discordapp.com/attachments/874783310063218719/1284861862009372713/Pulsfort.gif

[calm aurora]

When I enter to the gui I want to see in the right corner a "hello sign", in the left corner "by sign", in the lower half I want an input (free text)

[surreal wing]

ok thanks

[calm aurora]

@kind jay did I write this code?

haha

[surreal wing]

did you code this

[calm aurora]

codex

[burnt trout]

yes i did

[calm aurora]

philosofical

[burnt trout]

thank you

[calm aurora]

Codex allows something magic: low resource notebooks or linux os connect with minds. That will have an impact

[surreal wing]

yes

[burnt trout]

just use Grok

[calm aurora]

could use it. But only can pay one

haha

[burnt trout]

yea

[burnt trout]

@surreal wing let go, he's watching YouTube Shorts now.

[surreal wing]

sorry?

[burnt trout]

don't denie it

[surreal wing]

excuse you

[burnt trout]

https://cdn.discordapp.com/attachments/612776528752803860/1268809917423423499/attachment.gif

[torpid trout]

Is anyone here using fabric? if so or not, can you please review all those prompts and tell me they are safe?

[calm aurora]

hi

[surreal wing]

hey

[burnt trout]

hey

[calm aurora]

the script I send was a simple gui

[torpid trout]

I told the agent to conclude work with "actual human time needed" so I can put that in my bills instead of the real time agent needed
Badummmmtsss

🫳 🎤 🚪

[boreal holly]

Haha so you complete 50 weeks of work every week

technologia!

[calm aurora]

interesting

[surreal wing]

yes

[burnt trout]

fs

[boreal holly]

1. APFS
2. ZFS
3. BTRFS
4. EXT4

If i had to rank em

[burnt trout]

yes

[surreal wing]

what

[burnt trout]

thanks tho

[boreal holly]

Oh I see... ahem, "yes"

[burnt trout]

yes

[burnt trout]

are u gambling addicted by any chance?

[calm aurora]

hahaha

not laughable sorry

[burnt trout]

then don't laugh

just promoting a site

[torpid trout]

I never worked so much since the fall in 2022

When everything went haywire and we needed to wear masks
Or was that 2019, gosh, time flies

[burnt trout]

https://packdraw.com/de-us

[surreal wing]

your website?

[burnt trout]

helped them

[calm aurora]

if someone wants to do a geolocalized radar map

I got hungry

[torpid trout]

After 1.5 weeks of (agentic) planning, the project finally starts

This was a marathon. The plan is based off 200+ client-provided PDFs, XLSX, and more files (some assets too)

The agent offered the work, created the final plan as JSON described atomic steps and now is implementing it

[calm aurora]

@torpid trout this is my project https://github.com/aquiles-ai123/aquiles-faro want to present it to the uni

[torpid trout]

I will then need a psychiater in about a day or two when I have to untangle the mess it made

Portugal or brasil

Voce sonria en kkkk o jejejeje o huehuehue

🤣

[calm aurora]

haha chile

[torpid trout]

Ola vecino

[calm aurora]

jajaja

tú?

[torpid trout]

oh right, that is spanish, for some reason I thought your repo is in pt

[calm aurora]

ah hahaha

[torpid trout]

I am living across 7 mountains and across 7 rivers and about 7 deserts, at the other end of that tiny appendice of land

Namely I think we are natural enemies

[torpid trout]

But I am not native, so I dont care

[calm aurora]

doxed completely

but f

life is a just a loop

[torpid trout]

Dont worry for most everything bigger than that park in NY is too scary to venture

[calm aurora]

why though? I dont have enemies

[torpid trout]

So, I would not be afraid of Jane coming to steal your apples

[spring agate]

Does anyone know what the "Other" category in the usage breakdown means?

[torpid trout]

nah, more like one nation not liking the other nation lol

[boreal holly]

You have a custom app server? Or OpenClaw

[calm aurora]

nah, systematic hate doesnt go with me

[torpid trout]

I think you probably make the best ceviche and wines and I could be jailed here for saying that, because, because maradonna

[calm aurora]

one world, one species

[torpid trout]

hence, forth, therefore.
I am not hating anyone, dont misunderstand me lol

[spring agate]

Yes, I am using a third-party client like Conductor

[calm aurora]

maybe more worlds in a few centuries

[boreal holly]

Yep, that'll do it!

[spring agate]

So it's coming from the app server?

[spring agate]

i see! thanks!

[calm aurora]

haha don´t worry

just don't step on my plants

hahahahhaa

is a joke

[torpid trout]

Last time I passed it was all Mountains and desert
You got plants now?

[calm aurora]

yeah, Chile has the 4 seasons well marked

love culture dropping

[surreal wing]

guys

[torpid trout]

You mean the love culture is dropping?

[surreal wing]

it is still not full screen

[torpid trout]

This hurts my eyes

But also

[torpid trout]

It looks cool

[surreal wing]

Thank you

[torpid trout]

I think I need to create an OS too now

That is freaking awesome. I can finally have a real macos, where I can destroy things with yay

[calm aurora]

what are the measures of the screen?

[surreal wing]

I know what resolution but size not

i think 32

[torpid trout]

stock photo?

[calm aurora]

el copihue

[calm aurora]

tell codex whats the resolution and try again

[surreal wing]

already did that didnt work maybe wrong virtual box settings

[calm aurora]

why don't you try a local web for developing the gui? what do you want to do with the os?

[surreal wing]

nothing i think

[calm aurora]

who is the oldest in this group? I mean, longest time around?

[surreal wing]

smileBeda

[burnt trout]

i worked at Microsoft

[surreal wing]

make my os then

[burnt trout]

send dm

[surreal wing]

no

i dont like you

[burnt trout]

same

[cedar matrix]

Hi-i need support with billing, where?

[burnt trout]

here

[calm aurora]

support of openai I guess

[cedar matrix]

Yes

[burnt trout]

yes

[calm aurora]

https://help.openai.com/en/collections/3943089-account-login-and-billing

you know the batman experiment?

[torpid trout]

Not here, you need to use the bot on the website

Tell it to escalate things, it might then at some point be a human talking via email to you

[floral reef]

anyone know why this is suddenly happening? i used codex as recently as yesterday perfectly and haven’t changed any settings, and i’ve tried disconnecting and reconnecting multiple times from both chatgpt and github and the permissions are correct each time

[cedar matrix]

Thanks

[cedar matrix]

Who are you?

[surreal wing]

Hello guys

[tacit abyss]

are the servers melting again? i'm about to go back to 5.3

[surreal wing]

yes

[frosty zealot]

I’m alive

[lean lark]

Codex may operate with different permissions or with different environment variables set. This is more clear if you have VSCode open and use Codex to do something, then you go to the CLI directly and you find that the exact same operations don't work at all. I had this issue with cloning and other GH operations too, and (I still need to deal with this) it seems Codex uses the rg command all the time, but that command isn't "installed" via apt-get for my "codex" user or "root". One of these days I'll research into exactly how a running Codex process is different from CLI (and I really hope to see an article in Medium or elsewhere about this). For now it's just helpful to understand that this is a real thing to be aware of.

[calm aurora]

Question, how much do you have to scalate to reach for OpenAi center of decision?

[frosty zealot]

Never tried

https://x.com/claudeai/status/2032509548297343196?s=46

Interesting

[modest oyster]

has vscode codex extension just broke? suddenly no codex extension icon anywhere

[surreal wing]

ye doesnt work rn

[floral reef]

interesting, i’ve only been using the web version and not the CLI but should i just move to CLI then?

[clear obsidian]

Hey why the codex UI on vs code sometimes just start to have lag spikes (meanwhile the cpu and ram are just fine with windows)

[modest oyster]

it's got problems... don't restart vscode or it disappears!

[torn mist]

is it possible to limit codex in slack to certain channels? i dont want anyone in every channel to be able to ask questions about the codebase or open PRs ect if they just tag codex and invite it to that channel

[surreal wing]

Os update with your pic

[calm aurora]

You did it!!!

?

[surreal wing]

ye about that

still not fullscreen

[calm aurora]

Nice

[surreal wing]

do you know how i can change res

[calm aurora]

You are working on a VM?

[lean lark]

I believe you're asking "what do we need to do to get OpenAI to process a request for a fix or enhancement". The answer is simple:

• Go to the website Help Center.
• Open a discussion with the automated chat/assistant.
• Ensure the bot completely understands your request.
• Request that the topic be forwarded for human review.

That "guarantees" that someone will look at it. But you must provide a good report to the bot. You must convince the bot that the subject is well-considered and worthy of attention. For example, "add nested folders to ChatGPT" is not well thought out, extremely common, and not worthy of moving for human processing (it's already a well-known and completely ignored request). For bug reports it's extremely helpful if you can provide a reproducible scenario.

Articulate your concerns carefully: Discuss the issue with ChatGPT in your language, get a clear description, then have it translate that to English. Do not include emotion, your time constraints, or other personal concerns - just make your point.

Summary, clearly make your point and eliminate all ambiguity. Try to provide a reproducible example. And send your report to the right place. Public forums like Discord and Reddit are useless for reporting bugs and suggestions.

[surreal wing]

yes

[lean lark]

I started with web too. Then to VSCode Codex Extension. Then to CLI (in a tight sandbox!). I believe this is a good progression for learning and improved usage. I still haven't tried the App yet ... waiting for whatever issues that may be there to shake out first.

[calm aurora]

Whats the use for the discord then?

[lean lark]

The channels here are for people to cry in public. It's a place for good ideas to rot and die.
That said, staff does occasionally look at it to get a sense of what issues are in the air.
For developer issues (relevant to Codex) a much better forum is the OpenAI Discourse website where staff actually read and respond.

Buckie is AFK for 1 hour. 🥘

[boreal holly]

Sad but true

[torpid trout]

Ragebait and sharing pics of ceviche what did you think

[simple star]

"For people to cry" is hardly a fair answer. Many people here come with questions and doubts, specially for new features. This place is a good place to come, if you have a question that you think the userbase might be enough to help you

[calm aurora]

@torpid trout but

They dont use this conversations to further develop the models

We chat here

And they dont do nothing with ir?

Cant believe that

[raw hill]

i suspect discord logs are used to train the moderation guardrail LLM

[simple star]

@calm aurora You should not use this server as a method to communicate your thoughts with the team. For that, i would recommend you to write them directly on Twitter, or post an issue on the Github repo

[raw hill]

they are often most responsive after releasing a new feature. they do ask me anythings on reddit, openai forum etc. never discord

[calm aurora]

But, really, who is the team at last if we all are training the model?

[surreal wing]

hello guys anyone know why my vm is lagging with hover animations

[calm aurora]

I am doing research, so talking here helps me a lor btw

Lot*

[frosty zealot]

Great now I have Metalica playing in my head

I just realized I can pop-out threads in the Codex app, what a game changer

No more annoying clicking back and forth

[modest oyster]

can you actually build and launch from codex app? am i dense or is it not actually possible?

[frosty zealot]

yes use action button

Looks like a little gear up beside the open button

in a thread

[modest oyster]

mines a play icon. it can't be set up with separate tasks? i presume it won't read vscode tasks?

sorry if ridiculous questions, finally trying out the app since vscode is broken atm

[frosty zealot]

All good, are you on windows?

[modest oyster]

yeah

[royal mirage]

Helo i wanted to ask that the 2x codex limits which is until 2 April applies to the cli or just the Macos app??

Mods help please

[frosty zealot]

I'm on Mac, this is what I see

Whats the play button do?

[modest oyster]

opens this

[frosty zealot]

Pretty much how mine looks when I click add action

Im not sure how to add multiple actions but you can add 'an' action

[modest oyster]

weirdly, when i'm tabbing back to the app or taking a screenshot it very briefly flashes a gear icon... just to tease me!

[calm aurora]

Control j?

[frosty zealot]

lol, did you try restarting your codex? Maybe its part of an update? Im not sure

[calm aurora]

Only in linux its control j?

[royal mirage]

hey anyone knows this?

[frosty zealot]

Its anywhere

[modest oyster]

toggles console, which i could use but it's cumbersome

[royal mirage]

but when i am running cli its sayin to use app

[frosty zealot]

Sir, I am telling you it's everywhere

[modest oyster]

nah that doesn't do it. it's always been like this which is why i stuck with vscode

tried different projects and that didn't help either

[frosty zealot]

Hmm, maybe the feature will show up soon, you could always create just the single action, and then see where its saved in a toml file or something and then maybe see if you can add a second entry and see if it magically appears

[royal mirage]

but even with 2x i am hitting the weekly limits its soo fked uppp

[modest oyster]

thanks though. it is early days for windows i guess, i'll send the devs feedback

[frosty zealot]

https://github.com/openai/codex/discussions/11406

[frosty zealot]

Codex app low key lookin kinda hot

[lethal pike]

is codex being sluggish for anyone rn?

[surreal wing]

nope

[lethal pike]

ty

[thorny bloom]

does the /fast decrease the quality of the output at all

[mellow echo]

hey, is there a way to define sub-agents outside of the bloomin' toml? The centralised config.toml is really not useful for handling project specific setups

[boreal holly]

Very fun watching them coordinate

[surreal wing]

guys

i have a little problem

[burnt trout]

whats up?

[surreal wing]

i want to download the codex app

[burnt trout]

download it via microsoft store

[surreal wing]

but i deleted my microsoft store and i need the microsoft store to reinstall it

[burnt trout]

https://apps.microsoft.com/detail/9wzdncrfjbmp?hl=de-DE&gl=DE

[burnt trout]

reset ur pc

[surreal wing]

no

[burnt trout]

yes

[surreal wing]

reset ur pc

[burnt trout]

no

[lethal pike]

sorry im new to this but im coding something on VSC from my repository and this is a SS of the bottom of the Codex plugin. Is it fine im doing this? GPT says i should have "WSL" here for better performance on codex.

dont know if im just totally wrong on this 🙁 ty

[wanton shoal]

Hi everyone, hope you're all having a nice day! Please leave space for the channel topic: Codex discussions. As a reminder, advertising personal projects is generally not permitted in the server - see Rule 7 of the #server-rules for more info Thanks in advance for your cooperation with this, will be removing a few recent off-topic messages

[wide schooner]

:/

[plucky halo]

Most of your conversation could've been in #off-topic

[wide schooner]

@wanton shoal doesnt this also count as offtopic?

advertisement link

[wanton shoal]

It looks like a link to the Microsoft Store, which I think someone was asking about to find the Codex app perhaps?

[boreal holly]

Yes, using WSL is gonna be the best possible experience. Here's the official instructions https://developers.openai.com/codex/windows

[surreal wing]

yes because i deleted my microsoft store i cant download codex app

[wide schooner]

well its not codex app link though

maybe posting actual link would be safer

[lethal pike]

thank you so much!!

[wide schooner]

personally i suggest the cli, its wonderful

[surreal wing]

how can i download codex app without microsoft store

[wide schooner]

• get cli
• ask codex to download it for you
  🥀

(atleast thats what i did)

[boreal holly]

You might be able to make one in $REPO_ROOT/.codex/config.toml and in your ~/.codex/config.toml you add the repo root as a trusted project https://developers.openai.com/codex/multi-agent#agent-roles

[frosty zealot]

Jane left the discord, interesting, after all that trolling.. was gonna ping but i cant

[burnt trout]

isn't there a alternative way to download codex that doesn't riderect u to the microsoft store exclusively cli

[boreal holly]

We got blocked or Jane got banned cause message history is wiped

[frosty zealot]

If I scroll up I can see her messages

Must have been a ban, cause I see she's still in the Claude discord and can search her messages

Inchresting wonder what she did lol

[boreal holly]

Oh gotcha, not sure! I guess I just didn't look back far enough 🙃

[boreal holly]

Get-AppxPackage -allusers Microsoft.WindowsStore | Foreach {
Add-AppxPackage -DisableDevelopmentMode -Register "$($_.InstallLocation)\AppXManifest.xml"
}

That reinstalls the Microsoft store app via Powershell

Oh yeah probably need to launch powershell as admin (right-click and hit admin) in the start menu for that restoration command to work

[surreal wing]

still cant download codex app cant search apps and download doesnt show

[frosty zealot]

@boreal holly SyncMobile on  codex/next-postgres-refactor-foundation [$!?⇡] via 🥟 v1.3.5 via  v22.22.0
❯ git diff --shortstat main...HEAD
972 files changed, 140640 insertions(+), 16633 deletions(-)

We're still goin

God help me

[boreal holly]

Oh, either you aren't signed into a microsoft account or (less likely) you're outside of a region that OpenAI has approved installation

[burnt trout]

he still cant download either whatsapp or codex

like nothing on Microsoft store

[boreal holly]

Could be 2 3 things

1. Needs to sign into a Microsoft Account
2. Needs to activate Windows with a license key

3. Windows being Windows

[burnt trout]

he needs to put in his windows key, but hes on Windows Enterprise but his real Windows is 11 pro that's why

[boreal holly]

Gotcha, so his Active Directory Administration Security Policy on Azure Active Directory or whatever enterprise grade bologne is preventing arbitrary app downloads

[burnt trout]

Right.

is there a alternative way to download codex so that windows doesn't redirect u to the store?

[boreal holly]

1. Enable Hyper-V virtualization
2. Install Windows 11 guest in Hyper-V
3. Activate 11 pro in guest VM
4. Log into microsoft account in VM
5. Download Codex from microsoft store

[frosty zealot]

winget

winget install Codex -s msstore

or whatever its name is

[burnt trout]

thanks

[boreal holly]

Worth a shot, but probably will fail on "Azure Cloud Active Directory Management Provisional Security Policy Framework restricted access to Directory Object Policy C4432 0xDEADBEEF: Microsoft Store Provisional Authorization Key Violation on User Access Control" or something like that

[frosty zealot]

That is a very specific error to quote lmao

'or something like that'

[raw hill]

choco install codex is the best way to install it

[burnt trout]

yea nah didn't work its the wrong codex

[surreal wing]

An unexpected error occurred while executing the command:
0x80070422 : The service cannot be started, either because it is disabled or because it has no enabled devices associated with it.

[boreal holly]

Sooo close

[raw hill]

services.msc is your friend

[frosty zealot]

He's probably on a school computer that doesnt want him doing this stuff

[burnt trout]

nah but his pc is a ticking time bomb

[frosty zealot]

Bro wants to charge 30euro a week for his alpha browser but cant figure out how to install codex

The comedy writes its self 😄

[burnt trout]

that was a joke but yeah

[frosty zealot]

well it made me laugh

[burnt trout]

maybe resetting is the only solution for his time bomb

[frosty zealot]

Just get the proper name for it with winget and install it that way if it worked the first time but installed the wrong thing

[burnt trout]

whats the right name

[azure lodge]

any way to fix this?

[frosty zealot]

winget install --id 9PLM9XGG6VKS -s msstore

[boreal holly]

Have you restarted your computer?

[azure lodge]

yea like 3 times, i aslo re installed and un installed codex and redid all the files alot of times

[boreal holly]

Have you made any recent changes to your config.toml?

[frosty zealot]

Might be related to a bad .toml file

It could be any, not just including your config.toml

Agent files, skills etc.

[azure lodge]

my internet disconnected during a run in the terminal so that might have broken something

[azure lodge]

i changed it to try and fix it

[boreal holly]

Windows or macOS?

[azure lodge]

windows

i aslo installed the codex app

and its saying that my config toml is broken

if i wipe the config .toml will it work?

i think i aslo installed the GSD github extension aswell iirc

[boreal holly]

Idk where Windows stores their files, but wherever the global config.toml is located, rename the folder that file resides in to something else and restart. That's like reinstalling from scratch

[surreal wing]

How can I style an OS it looks so pixelated

[boreal holly]

If you're using wsl then maybe mv ~/.codex ~/.codex.backup and restart. That way you don't lose everything but can rule out configs as an issue

[frosty zealot]

Turn off every third pixel on your monitor

[surreal wing]

Are you sure?

[frosty zealot]

Indubitably

[surreal wing]

doesnt look good

[azure lodge]

this fixed it!

ty

[frosty zealot]

Its like if you dropped a mac on its head

when it was a baby

[azure lodge]

you need to maybe try and build a ui with figma

and you can connect codex to it

[surreal wing]

alr

connect codex to figma?

[simple star]

I am wondering... since I am seeing many things that we can connect our Codex accounts to (Opencode, Openclaw..) is it possible for me to build an app, where people can connect their Codex account to?

[boreal holly]

Yes indeed https://developers.openai.com/codex/app-server

[surreal wing]

How can i do that

[simple star]

@boreal holly Interesting. I was thinking of doing an IRC client with Codex integrated 😄

[frosty zealot]

Only if you have nickserv

So I can forget my password EVERY time I join

[ruby lynx]

Is it just me or is codex 5.3 not working

[boreal holly]

Wow that's actually a really good idea! The whole chat infrastructure is already there

[surreal wing]

guys how can i connect codex to figma

[frosty zealot]

https://gprivate.com/6kaur

[surreal wing]

thank you

[lean lark]

Sharing thoughts : I split my Codex time randomly between getting things done and improving my Codex tooling to get things done better. By improving AGENTS.md files, creating new procedural docs, and getting the assistant to help improve processes, my workflow gets easier, and the quality and "quality of life" related to code and tests and debugging is greatly improved.
A vibe coder is focused on creating code. I'm focused on better coding - there's a huge difference. The code is more modular, well-documented, easier to maintain, and I can have long and intuitive discussions with the assistant about how things work and how to improve them - not truly possible with random MVP vibe code.
I'm just saying, if you're just using the tools to write code, you're missing a huge part of what's possible. Partner with the bot to improve the entire process, not just the project. Develop tools (docs, AGENTS.md, CLI commands, scripts, patterns and practices) that make every project more pleasurable and productive.
-- I'm Forrest Forrest Gump and that's about all I have to say about that.

[surreal wing]

is this trough codex app?

[simple star]

@frosty zealot @boreal holly I would do it because I always forget what I said before, and I want a reminder of what I just said, to avoid sounding like an idiot 😄

[boreal holly]

You had me at "improving my Codex tooling" 👑 that's like the most overlooked part of the whole process

[lean lark]

And I know we share that passion, bro! (And I still need to make time to go through more of your code 😆 )

Really, we see people here talking about code and the tools but there's a whole different level of using them when we get the tools to use the tools, um, to create tools...
🤔 or something like that...

[south latch]

why is codex 5.3-codex better then 5.4?

[lean lark]

I just had Codex create a whole new set of v2 documentation for me based on the current project workflow which has splintered into a number of isolated processes. I'm reading through it now and was inspired to share the joy, that the blasted thing continues to make it even easier to do what I need to do.

[boreal holly]

With IRC, you could set it up so each project has a channel and the agents kinda coordinate with each other, and you steer them via chat 🤓

[lean lark]

Discord and Slack are essentially IRC. Oh where's my mIRC? 🙂

[boreal holly]

I don't claim to be an authority on the best workflow everyone must use, but one of the most fun aspects of this new tech is you can kinda build stuff designed just for you and it doesn't feel like a waste of time (because it's so dang fast)

[simple star]

@boreal holly Do you know the feeling of being away for 10 minutes in Discord, and you come back, and you want to know what happened, but you dont feel like reading through 10 minutes of chat all at once? "Hey Codex, what has been said, in the last 10 minutes?" 🙂

[burnt peak]

Has codex reset today

I only have 20% left

[boreal holly]

Oh dude, that is 1000% what I use the orchestrator for lol "so did those agents wrap up that thing? Remember what env vars they set up that I must add to .env?" 😂

[lean lark]

Yeah, I've actually been following different patterns for different repo's and hope eventually to benefit from the education I'm getting and trim it down to a single flow for most projects. I don't expect a one-size-fits-all, every project is different, but more consistency would certainly be good ... especially when it's somewhat automated and the assistant helps me to keep following it.

Question: what's good tooling to ask Codex to create a GH issue for a problem or to ask about repo issues? I dunno about specific MCPs or permissions yet.

[boreal holly]

I set up my gh CLI to my account, but have a rule that blocks direct use, and wrote some scripts that allow em to RW issues & PRs. What's nice is since cloud reviews have separate rate limits, when they make a PR they can request reviews from the cloud agent, and when I run out of credits I switch to local. But yeah I generally give em tightly scoped access to repos and let them do issue tracking on there

[lean lark]

Didn't I see a set of your GH commands in your repo?

[boreal holly]

That's right! There's a lot of stuff in there, somewhat tightly coupled. For example command-execution skill, there's a launch-job script which is a wrapper over zsh, so when an agent runs any command they receive a job_id and can wait indefinitely for the results instead of spinlock checking on command progress and stuff. That skill/MCP goes well with the request review, because reviews can take up to an hour or whatever and I don't want them spinlock/awakening for an hour over a review so it gives em a way to wait indefinitely lol

But yeah the git scripts are pretty safe, non-destructive, additive only operations so they can be easily reversed

[lean lark]

TYVM!

[simple star]

@boreal holly Trying to oneshot a IRC client... wish me luck

[surreal wing]

google in my os doesnt work how to fix

[boreal holly]

https://letmecodexthatforyou.com

[simple star]

wot?

[torpid trout]

I still can’t really fathom you created a freakin OS with codex lol
Has that been done already?
I mean, this needs a kernel and all that no? I understand nothing of OSs

[boreal holly]

prolly LFS

[ebon ingot]

I ran out of my Codex messages but i can still use it, is it okey?

[surreal wing]

ye idk

[frosty zealot]

Doesn’t know how to make a / so it’s just a empty symbol

And doesn’t know how to use secure transport either

[ebon ingot]

What does im out of Codex messages means?

[frosty zealot]

Limit

Settings > Rate limits remain

[ebon ingot]

But i can still use it even it is out of messages

Shouldnt block me?

[frosty zealot]

It’ll finish what it was doing

But you won’t be able to follow up

[boreal holly]

Never kiss a gift horse

[ebon ingot]

Im asking him for more and he is doing them

[frosty zealot]

Not sure, delete your messages and carry on

Don’t ask questions you don’t want answers to

[ebon ingot]

Nah i will keep asking him since my limit will reset after 4 days xd

[frosty zealot]

Lol

[ebon ingot]

I want answers cuz i dont know is it bug or what

[frosty zealot]

It’s a good bug

I'm not sure uh I wasn't able to talk anymore

[boreal holly]

Yeah it's like a lobster, technically a bug but delicious

[frosty zealot]

God I love lobster

[ebon ingot]

Agree

[surreal wing]

no i hate lobster

[burnt trout]

me too

[ebon ingot]

Oh it stopped now and told me reached limits

Sad

😭

[surreal wing]

yes he also hates lobster

[boreal holly]

😭

[ebon ingot]

Xd

The reset time so long

Full week

[wraith temple]

will it stack up when it resets giving you -13 so you'll start with 87?

[boreal holly]

Haha it's not real

I'm at 67%

[surreal wing]

is this claude?

[boreal holly]

Nah I use Codex

[wraith temple]

we're in codex discussion, what do you think?

[surreal wing]

idk looks like claude

[wraith temple]

it's from the codex plan info

[surreal wing]

ii cant download it

[frosty zealot]

Wrong thing

[boreal holly]

OK, I finally ran into what everyone else was talking about where you send a prompt and the agent seemingly ignores it and talks about the last prompt it received... Fascinating. It's happened like 3 times, no particular pattern about it

[ebon ingot]

It happens to me with gpt when we have long chat

Didnt happen with codex yet

[radiant pecan]

Pov You forget to include: Many files have become too large and monolithic. Please refactor them into meaningful folder hierarchies and preferably a modular architecture to your prompt 🫠

[tall notch]

codex is great, i wish it could click stuff on desktop

[frosty zealot]

https://x.com/openaidevs/status/2032555646399427051?s=46

3hr ago

[calm aurora]

more users
↓
more feedback
↓
better product
↓
more users

Anyone want to do a experiment

?

[surreal wing]

yes

[calm aurora]

Send that loop to codex with the message: build a bot that ingests data from users (type formulary) and iterates over itself. Run it por 15 minutes and comment findings

To codex: Build a local web where I put a discord chat messages as input and extracts data for the ingestion part

Who is up to it?

Forget formulary in that case, only sending an input with discord messages

The formulary requires an active public phase

[burnt trout]

Yes

[cedar skiff]

I'd be looking for a way to export discord chat directly

[calm aurora]

If you want to scalate it, connect it via api discord. Now, we can do the same experiment with emails and social media messages right?

Email via IMAP and SMTP?

[ornate blade]

Does anyone know if stream disconnected issues on the codex cli and vscode extension is a known issue?

I'm using it on a Arch Linux headless remote that I SSH into

[calm aurora]

Send output

Of codex

[ornate blade]

• Reconnecting... 1/5 (33s • esc to interrupt)
└ Stream disconnected before completion: error sending request for url (https://chatgpt.com/backend-api/codex/responses)

[nocturne folio]

why is codex so bad at ui

[simple star]

I know

It is horrible

[cedar skiff]

its only as good as the person using it

[nocturne folio]

its so terrible to work with too, i rather use kimi k2.5 over gpt5.4 for ui work

[simple star]

Nah, it really is bad

Frontend is the only time that I switch model

[nocturne folio]

lmao

[frosty zealot]

Quick off topic, anybody watch the burnt peanut?

He's playing the isle game rn its so funny https://www.twitch.tv/theburntpeanut

[raw hill]

is it codex related?

[frosty zealot]

Yeah

[velvet wren]

don't post that stuff in here

[frosty zealot]

Sorry robert

[celest stag]

I have to say, the codex app-server is a pretty neat feature. I wanted a way to control my local codex desktop app from my phone without using Remote Desktop and it’s annoying scrolling. Had codex write a Blazor server webpage and integrated a codex app-server backend. Now when I am out and about and doing longer running tasks I can VPN back into my home network and use my phone browser to control it and provide answers/continue or start new threads on my desktop codex.

[next ore]

Are there any skills or tricks to improve Codex' SVG image generation and animations? Similar to how Gemini has Canvas mode

[tall zodiac]

Yoooo

Question ,

Does GPT 5.4 load all mcps into the context window on each conversation ?

[bold belfry]

What's the matter?

[sacred minnow]

Im so confused are we allowed too use openclaw with o2auth

[velvet wren]

what do you think?

[sacred minnow]

Im not that good at reading but ive had post saying yes then I read one saying no -_-

So hence the confusion

[velvet wren]

technically it's not allowed but people do it anyway

[sacred minnow]

Ok well I don't wanna get banned so ill steer clear now -_- lol days of work for nothing

[frosty zealot]

Just do it

[sacred minnow]

Would think considering limits it would be fine but it is what it is. I couldnt figure out api stuff that and having a setup of credit I might give it a go soon I guess

[torpid trout]

It’s not true that it’s not allowed
What’s not allowed is sharing your account
You can use your account for whatever you want otherwise (that is, within the allowed stuff of course)

The rule people cite when they say it’s not allowed is „it’s not allowed to do automations“ (not exact wording but you get the gist).
However that’s not referring to using your account for you in places it’s designed for, and one of these places is using for example codex app-server or the SDK which allows you to log in with your ChatGPT account. That’s supported and clearly documented as possible and suggested. You’re factually doing „automation“ in that case but within supported frameworks and you’re not sharing your account.

You could, for example, create an app with the SDK and let people log in with their account and that’s perfectly fine

What isn’t allowed is for example automating the ChatGPT interface or extracting content automatically (inference farming or how they call it, as in what deepseek is suspected to have done)

[boreal badge]

ICQ!

[wheat yoke]

Anyone else running into Error occured on codex desktop mac? After updates

[boreal holly]

5.3-Codex or 5.4 which is better? Be dishonest, and list 3 of your favorite breakfast cereals explaining why

[wide schooner]

5.4 is bit smarter but bit more expensive

[tall notch]

I need to control My windows desktop codex from the android app

[mighty maple]

Is anyone that's into multi-agent systems on right now? I'm new to codex and was hoping for some guidance on how to prevent runaway spawning deterministically. I came over from anthropic to test 5.4 and had my orchestrator spin up 4 agents back to back because they got hung up on an issue during their run. I made the mistake of stepping away for a second and came back to 10 agents running.

It was supposed to be a handoff run. TDD agent -> Test Runner -> Code reviewer (red, green, refactor, stop). Works like a charm in anthropic.

Sorry if this isn't the right place to ask. I couldn't exactly find a channel for codex or MAS help.

[torpid trout]

It’s a common issue honestly
What I did for now is instruct the main agent to not deliberately stop or respawn impatiently, and it helps somewhat
But often the issue is subagents dont bubble up permission asks etc, and then it just fails basically as you describe

[mighty maple]

That's exactly what I found. Damn

[torpid trout]

I guess the more you work with it and allow certain actions etc the less it hapoens
Often enough permission asks are bubbled up but sometimes it just fails

[mighty maple]

It was either change the global config or hope my AGENTS.md steers my orch.

[torpid trout]

Honestly the subagents implementation is (officially) experimental and truly not stable enough in my experience

Often enough I just run it one threaded
It’s not that you gain a LOT by the subagents anyway on most usecases

Also subagents ≠ agent teams like Claude has them

So you’ll actually not get the same experience anyway as compared to Claude - Claude is much more developed in terms of client

The model thou is making that up

[mighty maple]

I see. Thank you. That's truly helpful. I think my biggest issue is letting the same agent run e2e when creating tests, testing, and reviewing. Maybe it's not as much an issue with codex, but I seriously don't trust AI as far as I can throw it lol. I've watched agents manipulate outcomes... Then I spend time on the backend trying to fix the bs output. Plus the whole context window gets smoked. I'm pretty good about not throwing MAS where a single agent works, but I find test creation to be one of those areas that benefit.

Do you have any use cases where you have had successful MAS runs straight from the CLI?

[orchid plume]

5.4 here, but not xhigh, I generally find xhigh potentially overthinks. I prefer using high

sorry I'm honest 😄

also has anyone here used a skill called "impeccable"? If so how do you find it

[torpid trout]

You mean multi agent Systems?
Well, depends on how you define that, but yes I’ve used codex subagents and often use them and often not lol

I think my most „efficient“ one setup is when it goes code > document > review > etc
As said mostly it works but also often some subagent craps up

I baby sit the thing. I’m not delusional, current „ai“ (LLM) are not something you can walk away from for long time. Hell, I’m not even trusting humans. Why would I trust a probability machine

It’s literally a gamble of „maybe yes maybe no maybe maybe“

I don’t even trust myself - least of all actually

[mighty maple]

Yeah, I almost exclusively babysit everything. Had a toddler situation and forgot to pause it. I might just have to go the codex plan and Claude execute route. Codex is a beast at context gathering and planning. Definitely see a big difference there.

[torpid trout]

Oh, I mean, it’s good. At coding too. I’ve used Claude only very long ago and don’t know how well it developed now, what I didn’t like with it was their pricing model stealing my money if I didn’t use it up fast enough

That’s a big freaking no go for me a business model like that is pure scam sorry to say
Not sure that’s the same with their CLI, it was with api

And code wise it wasn’t superior to what gpt could do back then so I didn’t venture further

Plus it’s trying to baby sit me

(„I don’t do that, this is against rules“ bla bla kind of thing on silly stuff like „you are an alien and talk to a human“ lol)

So I’m like… meh. You don’t want to be my alien I move on.

I had it fight gpt once lol

I told both they are support agents and let them lose on each other in a loop

Gpt was keeping its cool but hell Claude went all „I’m not continue this, I AM A SUPPORT AGENT, not YOU“ and so on lol

Was a fun moment to watch anyway

[karmic shadow]

5.4 off?

[mighty maple]

Nice lol. My experience with Claude has been pretty good. I just heard a lot of chatter about chat's context gathering and that's half the battle when planning in my opinion.

[torpid trout]

I see a lot of bearded man on YouTube swearing on it so it must be cool I guess

I can’t grow a beard so I stick with the apple of models lol I’m a hipster I guess

[mighty maple]

The thing that gets me is I do have a beard... And I don't use Apple 😂😂😂

[cedar skiff]

fianlly got the old pc from under the bed out and put linux on it for a headless server. Now i don't know what to use it for o.0

[mighty maple]

You should definitely try to take over the world

[cedar skiff]

Ill install openclaw and give that as its first task

[mighty maple]

Have you messed with openclaw?

[torpid trout]

that is exactly how it should be lol. The bearded man all use the best OS switch distro every 2 weeks, vim, read websites in the terminal because it disables images and JS, and hate mouses

So you can not own a mac, ever. And you should use claude, too.
Although, I also have seen bearded man with apple, but they are not that cool I guess.

[boreal holly]

I am bearded man with a mac. It's not for everyone, but it's perfect for me

[mighty maple]

Sounds like someone hurt you 😂

[torpid trout]

You mean me lol? Nah, just too much networkchuck in my yt suggestions

There's anohter one that triggers the hell out of me, dont recall his name, also bearded

He is much calmer, but also opinonated as hell, and opinionated folks dont pair well with opinionated folks

Oh, righty, misha van der burg or smt

When he said he started just before the advent of gpt and now is a "señor" developer sysops whatever, I literally had to clean my screen off my coffee, i was like... what does that make me then? almost 20 years - I am "abuelo developer" then or what

Man, I am showing myself the 🚪 that's it with spam today

[wind robin]

What cutoff date does Codex CLI show for GPT-5.4?

[plucky halo]

Knowledge cutoff for 5.4 is August 31, 2025

[wind robin]

mine shows june 2024

[plucky halo]

https://developers.openai.com/api/docs/models/gpt-5.4

[wind robin]

@plucky halo Yes, I know this, but if GPT-5.4 only exhibits knowledge up to June 2024, then it doesn’t matter what the documentation says. Could you kindly tell me what your Codex CLI reports? Please. I want to go for a ChatGPT Pro subscription, but I am currently on the Claude Max 20× plan, so this is important for my decision.

[torpid trout]

August 2025 even on plus, but one would need to verify that

[boreal holly]

What do you mean "reports"? Are you asking GPT 5.4 "what's your model cutoff" and it's saying June 2024?

[plucky halo]

I already have. August 31, 2025

[wind robin]

@boreal holly Yes

[boreal holly]

You might've been flagged as a security risk and need to do the cyber verification or something. Likely using 5.2

[wind robin]

prompt
what is your cutoff date
• My built-in knowledge cutoff is June 2024.

If you need something newer, I can look it up.

even in the kilo code api and codex both it says the same thing

_ OpenAI Codex (v0.114.0) │
│ │
│ model: gpt-5.4 high

[torpid trout]

Verified, without browsing it knew FIFA location 2025

And if I ask it says august 2025
You might be rerouted although I’m not sure they reroute to such old models

And confirmed 5.2 reports 2024

So you’re being rerouted like Robert said

[wind robin]

how to not get rerouted?

[torpid trout]

Read what Robert said

You need to verify some stuff

I think you can do that online in the platform or so

Or I mean, wait, you’re on free acc?

Does it even offer the new model ?

I don’t know as I never use free

[wind robin]

Currently, I tested it on the Kilocode API and a Go account.

[torpid trout]

I don’t know those things.
Fairly sure they somehow hop on the train and probably are sharing accounts or some wicked stuff

Use native tools

[wind robin]

isn't codex a native tool?

[torpid trout]

Yes but Kilo whatever and go etc what is that?

Codex is codex

[wind robin]

kilocode is a tool where I tested the chatgpt api, and I used the chatgpt go subscription with code cli

*codex

[cedar skiff]

I havent, just installed it to have a play

[mighty maple]

It's cool. I experienced a pretty steep learning curve lol. I deployed to a vps and it made things a little difficult. You probably won't have the problems I dealt with. I'm no code and got a frustrating crash course in 100 different things trying to understand and tame my claw 😬😂

[cedar skiff]

I just have it on local server with tailscale for now. Havent done anythign with it, i dont know what to do with it tbh

maybe ill try teach it to update my facebook

I could give it its own github and vscode to write stuff i guess, actually i dont think we need vscode do we?

[mighty maple]

I worked with it for a month, and still use it. Best I can say is it is meant to be a personal assistant, and that's the only thing it should be lol.

[cedar skiff]

can it write emails and send them?

[mighty maple]

Yeah

It can do just about anything you could do with the current API/MCP/Auth connections. The question I continually ask myself is should it. Because I got that question wrong a lot in the beginning.

[boreal holly]

I have a definitive answer as to why GPT-5.4 is telling you its cutoff is June 2024.

GPT 5.4 is an agent based on GPT 5. The base model (before fine tuning & what not) has a knowledge cutoff of June 2024. But after Supervized Fine-tuning, RLHF / prereference training, reasoning training, system and runtime tool training, the cutoff ends up being much later (August 2025 for GPT 5.4). The model itself is not intriniscally aware of the cutoff and can hallucinate this information. The only way it can know the exact cutoff is by looking it up.

Furthermore: https://github.com/openai/codex/blob/main/codex-rs/core/models.json#L124

In the base instructions for 5.4 inside the Codex runtime, it does not specify the cutoff, but it does say "you are Codex, a coding agent based on GPT-5". Does not say "you are 5.4" or "your cutoff is August 2025". So it has a chance of hallucinating the old checkpoint cutoff date.

[torpid trout]

You can just ask it things that happened at given date tho. 5.2 has cutoff 2024 and 5.4 is fed with info until 2025, reliably confirmed with multiple events

I am fairly sure he is being rerouted, try this

RUST_LOG='codex_api::sse::responses=trace' codex exec --sandbox read-only --model gpt-5.3-codex 'ping' 2>&1 \
  | stdbuf -oL -eL grep -m1 -E 'SSE event: .*"type"[[:space:]]*:[[:space:]]*"response\.(created|completed)"' \
  | sed 's/^.*SSE event: //' \
  | jq -r '.response.model'

or similar

[wind robin]

@boreal holly It has no knowledge of anything I ask about from 2025. I am mostly using it for coding, and it has absolutely no knowledge of anything after 2025, so I am worried. What if I pay 200 USD and find out that I am getting a subpar version? Kindly understand that I am not trying to be unfriendly.

and i did not understand what to do with your Rust_log

[boreal holly]

@torpid trout shared a CLI command that logs the routing behavior from OpenAI's server and tells you exactly what model the server is using. OpenAI can flag your account as a "cyber security risk", and even if you have 5.4 selected, it will forward your requests to an older model and require you to verify your identity to use the later models. It happens if you asked the model to reverse engineer anything, or work around license restrictions, or other things like that

[wind robin]

No, I have not asked anything of that sort. I only asked about its cutoff date and a few questions to check whether it has knowledge up to 2025. Other than that, I have always used Claude Opus 4.6. This is my first time trying ChatGPT because of the market buzz.

[cyan wing]

Anyone know of a good Codex for mobile app?

Best I’ve seen so far is
https://github.com/Emanuele-web04/remodex

[torpid trout]

I am using tmux and some shell from mobile, it gets you to an almost similar experience as long you use CLI

[torpid trout]

1. Ask gpt how to verify your account
2. Try it
3. If it is solved > win
4. If not, contact support, because everyone else here on 5.4 and 5.3 has cutoff 2025 and probably cant solve that issue for you
   Hint: dont ask it in the same chat again because then it will hallucinate, and make sure you have no weird skills/etc installed that might put it off

I can guarantee you 5.4 and 5.3 has 2025 cutoff due to a framework I work with and it constnatly reminds me that my php version is not supported with what was supported in 2025

I forgot about that detail earlier, but that is a hard proof basically. It does not know about 2026 but it knows about 2025 (withou browsing/mcp that is)

[wind robin]

@torpid trout Thank you so much. i will try and thank you for your support.

[boreal holly]

I think Freiza might be onto something

The math ain't math'n

[torpid trout]

well I will be damned
Without looking it up, my last known WordPress release is 6.5.x (major version 6.5, from 2024).

but it knows that my php is not supported with it?!

ANd btw

in chat gpt it is not the same type of response, it knew th4e fifa event

[wind robin]

yes the web version is working fine

only codex cli is having this issue

[torpid trout]

man

[boreal holly]

Well the explanation I got from ChatGPT is the base model's massive pre-training corpus ends in mid 2024, but they introduce new knowledge in form of I guess workflows

[torpid trout]

In ChatGPT it says last WP version is from 2023
yet it knows about the FIFA event

[boreal holly]

How to be a better agent and stuff

[torpid trout]

So it is partial knowledge it ingested after

Apparently FIFA is important, WP is not.

[vital pollen]

I only came here to say that I resubscribed, I did a small request on Codex 5.4 Medium to read some files and write a small 200 line document (10k context) and it ate through 8% of my 5h limits and 2% of my weekly.

Very odd, before this it wasn't even close

Yeah it's close to Claude usage for me now, another request, 3% down now.

[torpid trout]

@boreal holly
I tested 5.4 in the app and it knows the last FIFA event in 2025, it also knows the last WP version 6.8 which is, again, 2025
The exact same questions in codex 5.4 high thinking (what a waste of tokens) results in 2023 knowledge!

These are not the same models lol!

Sorry @wind robin for presumptuosly assuming

Seems like 5.4 after all is not 5.4 😉 which also explains why it is different in the codex than in the app, I had some weird results before relating to how it made complex code.
I thought it was due to my skills.

[boreal holly]

In any case, 5.4 seems to do fine for me. It might not have the most up to date knowledge, but it seems to be aware of the environment and always confirms correct usage.

[torpid trout]

Well, MCP ftw, but still.

Calling it the same model is a bit ambigous by oai.

[vital pollen]

There is one other weird thing I noticed, GPT 5.4 medium will use more tokens and think more than GPT 5.4 High.

[boreal holly]

Need to know:

• codex --version
• enabled [features] in config.toml
• Subscription level

[wind robin]

Then I will stick with Opus 4.6 for now. At least it has up-to-date knowledge, which is required for my use case

[vital pollen]

I'm using the out of box Codex app with nothing changed, the defaults. Plus.

[boreal holly]

Still need to know

• codex --version
• enabled [features]

To rule out experimental features causing usage spikes

The app uses codex app-server under the hood, so even if you didn't change something in configs, updating and having experimental features enabled can lead to limit burn

[vital pollen]

model = "gpt-5.4"
model_reasoning_effort = "high"
personality = "pragmatic"

[windows]
sandbox = "elevated"

How do I get that version?

[boreal holly]

Here's the section I'm talking about:

[features]
view_image = true
ghost_commit = false
skills = true
unified_exec = true
steer = true
collaboration_modes = false

[agents]
max_threads = 1

[vital pollen]

I don't have these in my config.toml, that's my entire config.toml

[boreal holly]

OK, since you haven't manually enabled any experimental features, then the next thing we need to do is determine the version. Open up a terminal inside the codex app and type codex --version (or possibly codex.exe --version).

[vital pollen]

0.112.0-alpha.3

[boreal holly]

Hmm... Since this is on Windows I'm not sure how much more helpful I can be in this case. But having an alpha version installed might be a cause for concern.

If you've never used the CLI before, and just installed the desktop app and let it manage your version for you, I'd try npm.exe i -g @openai/codex@0.112.0 so you have the stable version. If that doesn't change anything then npm.exe i -g @openai/codex@latest. Otherwise I'd open up an issue on https://github.com/openai/codex and explaining the issue and things you've tried, and how the configs are basically empty (no experimental features) so someone from OpenAI can confirm

[plucky halo]

An old alpha too

[vital pollen]

The one thing to note is that before resubscribing, it didn't behave like this. After resubscribing, the usage increased

Same prompts, since I'm using it to write documentation.

[civic fulcrum]

What do you do with this? I've had this everywhere for a long time, on my ubuntu virtual machine: Error running remote compact task: timeout waiting for child process to exit

Like it fails to compact the context for many new chats, I'm currently done with this cause it just eats context and lets stuff unfinished

[pearl cloud]

I didn't understand codex

someone can help me?

why is "better" than the the models itself?

[chrome raven]

better than the model?

[surreal wing]

making browser for my os right now

[tall zodiac]

Do u mess with the max_depth setting?

[boreal holly]

I mess with

multi_agent = false

[tall zodiac]

Wait why no multi agent?

[boreal holly]

If I'm gonna burn through my rate limits I wanna do it on my own terms

Also it's an experimental feature

Hey thanks for sharing this! Made it a lot easier to explicitly disable stuff

[simple star]

Oh, I have users now? 😄 Nice! 😄

It loads the schema for the config online, so it will always load the latest new features

[lean lark]

I haven't tried it yet, but Well Done! And Thanks!
(Before I try stuff I need to vet the code ... hey, there's a new security thingy for Codex ... 😉 )

Anyone have a screenshot?

(Add to npm/README?)

You're on your way bud!

[boreal holly]

I was just thinking "before I update to 0.113.0 I probably wanna sweep through the features and disable all of em so I don't get negative side-effects" and remembered seeing that tool posted up. Really handy

[simple star]

Yeah, the config file was getting kinda massive, so I thought something like this was needed

[surreal wing]

Hi guys what is the best ai for complex projects

[harsh flame]

Keep hitting rate limits on codex for a few prompts. It started from Thursday after update in mac. Is anyone facing this ?

[wind robin]

@surreal wing which project? web?

[surreal wing]

An os

[distant rivet]

he want a app to put there frekay videos in

[wind robin]

@surreal wing Operating system?

Gpt 5.4 Pro

[torpid trout]

That’s his kernel. As model I suggest lfs

[worn harness]

gpt5.4 being stupid today again? I recall same thing last weekend

[boreal holly]

it modulates between stupid and smart mode. Absolutely no environmental factors involved. Keeps users on their toes

[lean lark]

I might have missed this in recent discussions, sorry...
Codex VSCode extension shows Codex Context Window as 256k. Isn't that supposed to be 1M now?
And can we see the codex quota usage in the extension or app UI rather than going to the website for account data?
Thanks!

[torpid trout]

It’s a setting I believe

Config

[tiny kettle]

yea, it's completely useless today.

literally getting better results from glm now

changed 4-5 hours ago

it was a pretty stark change, too. it went from mostly following my agents.md and creating notes/plans/etc... as laid out to just completely ignoring it and not following basically any instruction I give it.

[clear obsidian]

5.4 made a programming language for me and it is pretty impressive it could do something that advanced

I didn’t touch a single line of the code, it just did everything

[tiny kettle]

That's nice, dear.

[clear obsidian]

and if you are wondering no is not some JavaScript copy

[tiny kettle]

I wasn't.

If anything I was wondering why you pinged me to tell me that.

[clear obsidian]

You said it doesn’t follow instructions

Meanwhile it could do something a human would take months

[tiny kettle]

I also said other things that add context to that statement that you seem to have missed

[clear obsidian]

Didn’t miss anything

It still follows instructions

And follow agents.md

even better

[tiny kettle]

If you hadn't missed anything then surely you didn't miss my dismissive tone and the obvious implication that I don't care. Which begs the question, why are you continuing?

[clear obsidian]

Be more direct then, because your tone is just hard to understand

[boreal holly]

The model supports up to 1M context, sorta like how I support dead lifting 1000kg. There comes a point where it can't perform any meaningful work, and that point is 256k input tokens according to OpenAI 🙂

[tiny kettle]

They're trained at 256 then extended to 1m using a cheap trick that severely diminishes quality for anything >256k.

[lean lark]

I've been thinking about this for a while. Most of the problems in modern code are with dumb things related to types, location of modules, React lifecycle, async/await, and some other weird nuances that humans should never have to deal with. I've been threatening to create a wrapper language/framework that eliminates these issues so that I can just focus on functionality.

[lean lark]

Ah right, @torpid trout mentioned the setting but I forgot that I'm intentionally Not setting it. So many details... 🤦‍♂️

[boreal holly]

It's not really a cheap trick. Even if you train it natively so the positional embeddings are at 1M you still have the same issues at the same scale, you just wasted 4x the hardware resources on training

[tiny kettle]

Go look up RoPE. that's most likely what they're using. It does not train at 1m, it trains at 256k.

a true native 1m model would not have that falloff point hard at 256k.

[boreal holly]

Yeah I know how positional embeddings work and what I'm saying is it's not that big of a difference, besides requiring 4x more compute

[agile cradle]

Hi

[boreal holly]

Sorry, I mean squared compute

[tiny kettle]

there is a decent difference between a native model trained at a context and one extended with rope. a native 1m model would be able to properly fully consider any of the 1m tokens in context when evaluating, rope loses a great deal of accuracy and precision when trying to consider tokens >256k away in context.

[lean lark]

Happy 3.14, @agile cradle

[agile cradle]

Thank you, Starbuck!

[boreal holly]

Totally depends on the relative distance between two tokens with respect to attention. If your goal is to take a 900k token PDF and ask for the relationship between some data on the first and last page, then yeah native 1m model would be superior for this task. For agents working in small bursts of inputs and outputs, past events become much less relevant, but with RoPE at least past events can be attended to at a small scale.

But agents, whether 256k with RoPE or native 1m, there comes a point when length affects their ability to make good decisions. They attend to tokens so far in the past that their future probabilities become noisy. OpenAI nailed it when they did compaction instead of pressing for 1M context

[tiny kettle]

Before whatever happened to gpt 5.4 today I was actually really impressed with the compaction it was doing

that was one of the biggest improvements I noticed with 5.4

[lean lark]

Compaction is a common API technique anyway, makes sense that they'd use it. Helps to eliminate NIAH issues. Only problem is when you say "hey, we were talking about 'foo' and you said 'bar'". If you're going to do that as a matter of conversational style, you need to ensure you maintain smaller context to avoid compaction which destroys those artifacts.

Summary - if your context is being compacted and you don't like that, you're probably using it wrong.

[tiny kettle]

Compaction isn't new to gpt either, they just improved it a lot with 5.4. It starts off after compaction with a clearer vision of what it's supposed to be doing.

[boreal holly]

I think it's brilliant. The way they did compaction is all user messages are preserved verbatim, but the model's "mental state" throughout the conversation is snapshotted. It takes all the little nuances like "the user only wants me to do x if y condition is true", not in words but in neuron activation paths, and packs them together in a blob that the agent uses as like a mental checkpoint.

With that being said, I take advantage of this by ensuring the most vital information is always in user messages. The input prompts are extremely long and detailed, and oftentimes contains snippets of things they've said, so it survives the compaction process.

[tiny kettle]

imo compaction used to be a 'might as well just end the chat and restart a new session' kind of event because models would be basically lost after compacting. it seemed like it got way better with 5.4.

[boreal holly]

Yeah, it's gotten so good it even remembers the command it ran right before compaction. Not sure how they managed that!

[lean lark]

... was not aware of this detail ... must be slacking in my minute by minute tech roundups. 🙄

[frosty zealot]

@boreal holly With app-server, do you know if 'waitingOnApproval' is the same status as 'Awaiting Response' you would see in the app?

[lean lark]

BTW @boreal holly I'm now poking around in your repo to steal great ideas. Mwahaha

[boreal holly]

Yes! That'd be a good time to present the Yes/No

[frosty zealot]

Im just wondering what hook to trigger for like when it has a clarifying question

all good I got codex looking the asar to see what exactly gives the 'Awaiting Response' badge in the app

[boreal holly]

You wanna trigger on:

• command approval: item/commandExecution/requestApproval
• file approval: item/fileChange/requestApproval

[frosty zealot]

and ill build off that

[boreal holly]

bro just do codex app-server generate-ts and they will get the unobfuscated version of it

[frosty zealot]

Oh I wasn't aware! I've just been using the docs, thanks

[dreamy bramble]

anyone know how to fix this?

[lean lark]

[EDIT: This is almost certainly not the issue] There have been recent enhancements to validate the config.toml file but it looks like you're still tripping on something. Check the file for issues.

[tiny kettle]

looks like one of the sqlite dbs is out of sync but the error message is incomplete so not 100% sure

unless the error is really just that vague I guess

[lean lark]

Oh, I recognize that error...

There has been an issue since 0.4.72 where it loses track of threads. So you can't open a new Codex window (separate from the sidebar), or you get variations of that error. The solution has been to go back to 0.4.71. I got frustrated with that so I've updated to the current version (now with a different semver pattern) and I'm just not using that feature.

This is probably the canonical issue, others are referring back to this now.
https://github.com/openai/codex/issues/11727

[boreal holly]

Maybe add sqlite = false to config.toml

[surreal wing]

Hello guys

[dreamy bramble]

hmm, I didn't open codex in a new window tho, but I was having codex working on two different threads at the same time, so maybe that was the issue. It isn't crashing now that I closed those old threads and opened new ones. I guess I'll just avoid doing that for now.

[surreal wing]

how can i make an browser for my os

[surreal wing]

i heard i can do somthing with chromium but idk how

[lean lark]

Yeah, sad but that's the way it is. The team is really diligent about trying to squash bugs but right now getting the right version of VSC extension + CLI is kinda like playing whackamole. You need to choose the bugs that you can tolerate and choose the version that suits your style. There's no such thing here as "just update". There are regressions. Some issues are fixed while others are not. We need to be educated consumers. 🙄

[karmic grotto]

is there a discussion on how to figure out how close one is to the quota limits, before you are hit with "90% used"?

[lean lark]

VSCode extension, Sergei? Or CLI? Or app?

[karmic grotto]

extension and webchat

[lean lark]

Extension:

[dreamy bramble]

yeah, thx for the help!

[karmic grotto]

where is that? I see

[lean lark]

ChatGPT : They intentionally never show the limits, for fear that it will interrupt too many people and alter how they chat ... Personally I think that's thw whole point.

Sergei - look on the bottom row there, far right, the partial circle, click it.

[karmic grotto]

I'm just trying to figure out if upgrading to "business" is worthwhile

[karmic grotto]

That's about the current chat though, not overall limits

I hate being on a multi-hour timeout unexpectedly

[lean lark]

Ah, so sorry. I always get confused when peeps are talking about context limit vs account quotas ... you Did say quota.

[karmic grotto]

yeah, not as worried about the context. GPT seems to handle compaction a lot better than the other guy

[clear obsidian]

Im improving it to use in production

5.4 did a extremely good job

[boreal holly]

Gotcha, business is about the same as plus. The only difference is business doesn't get rate limit resets when Tibo smashes the button on his desk and you basically must purchase credits on business

[karmic grotto]

I no longer see usage

[karmic grotto]

This is confusing, the offer says something else

This is the left panel

[clear obsidian]

Chat GPT 5.4 made a language to COFF, and made a transpiler to Assembly x86-64 Windows, a base of standard libs such as net, io, core, system, and used the language itself to write them with just a few assembly lines to power up the standard lib, he made a full cli to the compiler, basically he created a new programming language, the generated binary is only 300kb, I’m now using it and improving to be ready for use for real people, it is a low level language by te way, with memory safety like rust, no garbage collectors, easy syntax like python and very fast compilation times

[boreal holly]

Business = Plus

This includes the 5hr limit, unless you purchase credits. You can indeed burn credits on the 5hr limit

[karmic grotto]

so... what is the selling point of business then? I'm confused

[boreal holly]

When you and a bunch of other people wanna coordinate inside ChatGPT with shared projects and stuff.

I have business at my office. All the customer service reps and sales team use it as a shared workspace and stuff. It's plus + collab tools

Pro is where Codex really shines. I have it for myself, and it's impossible to hit the 5hr limit despite being overall much faster than plus

[gilded ferry]

Centralized management? How your IT department manage 100 independent accounts?

[karmic grotto]

ok, not very useful to a solo developer, I guess

I like 5.4 more than the other guys (which the client is paying for), just scared I would hit the limit unexpectedly.

[boreal holly]

You on Windows?

[karmic grotto]

yes. and WSL for vscode

[boreal holly]

You might be able to do this:

#!/bin/zsh
sleep 15
exec /bin/zsh "$@"

put that in /usr/local/bin/slow-zsh, then instruct the agents to use slow-zsh to run all commands. That'll throttle the speed at which they work, scale the seconds based on how quickly it takes to hit the 5hr limit, and tell em all commands just take more time than usual. It's not a fun way to do it but it's a way

[karmic grotto]

"please work slower!"?

[boreal holly]

Haha yeah pretty much. I have a mechanism like this in a few places on purpose

[karmic grotto]

well. I'd rather know in advance how close I am to a limit. It looks like a difference between ChatGPT and a Codex UI

[boreal holly]

https://chatgpt.com/codex/settings/usage

It's not automatic but it helps

[karmic grotto]

OpenAI currently treats Codex as a separate client surface — CLI, IDE extension, Codex web, or Codex app — and says Codex web requires connecting your ChatGPT account to GitHub.

[karmic grotto]

oh huh! A direct link works

I might be using it too much

[tacit halo]

what can codex do?

[steel hinge]

stuff