codex-discussions

[cedar skiff]

there is an experimental setting for it outside of plan mode, i can't remeber what it is though

[boreal holly]

Probably not a bad idea. The agent provides selectable answers to clarifying questions so you don't have to type all your responses. I personally type all my responses

[undone patio]

day 1 of attempting to use this discord without getting muted, take 3

[dense heron]

but it dont spend more tokens or something like that? i guess it have some disadvantage

[undone patio]

probably my final attempt if i fail this time

[cedar skiff]

its a tool call like any other

[lean lark]

For interactive planning, I use something like 5.4/low/planning, have a nice discussion, then go to 5.5/medium. If plannng mode gets in the way just stay on low, talk out what you wanna do.
One thing there though, after resolving what you want to do with 5.4/low. Switch up and ask the better model if the plan is good. Don't trust lower models to do your planning.
And one final thing ... 5.4 is still an excellent model, as was 5.3. I don't see anything wrong with them planning anything.

HTH

[dense heron]

so i can use without worry?

[cedar skiff]

what worry do you have, should be find?

[boreal holly]

Nice! Some folks are saying 5.5 low/none reasoning is really low latency. I agree, plan mode does not require lots of reasoning. As they say "plans are worthless, but planning is indispensable". The part that requires reasoning is implementation, when stuff can and oftentimes does go wrong

[cedar skiff]

with claude i used to do implementation planning using a spec and a loop

[gentle harbor]

i wish open ai made a search model

[cedar skiff]

I found better ways in codex though

[wide schooner]

xhigh is very low latency too it skips reasoning if unnecessary

[mighty lion]

why gpt 5.5 unavaiable?

????

[cedar skiff]

this is the command

[features]
default_mode_request_user_input = true

[mighty lion]

?

guys what.

?

[gentle harbor]

why did you send the same img twice

[undone patio]

have u tried turning it on and off again

[mighty lion]

accident

[mighty lion]

yes

what the hell.

[undone patio]

use the terminal CLI to debug it

[gentle harbor]

yea most bugs with codex can be fixed by codex

just by telling codex to fix codex

[mighty lion]

ok ill try

[lean lark]

Agreed, Robert - sometimes you don't want the lower-quality models planning things out for you. The higher-tier model will certainly do whatever we tell it, so the plan needs to be good to start. As I've said:

You can get an architect to design your house and a high school kid to build it.
Or you can get a high school kid to design it and a certified contractor to build it.
Or just ask the kid to figure it all out and let you know when it's ready to move in.
Or you can work with the architect and contractor.
Your choice.

[undone patio]

no

u need to open a powershell or terminal

and use the CLI, directing it at the codex app

instruct it to debug why you cannot use the model

and give it the error code

[boreal holly]

Unplug all USB devices and restart computer

[wide schooner]

start new convo with 5.5

[small violet]

great

[nocturne folio]

how do i use codex at the gym without using cloud or bringing my macbook

[small violet]

this is easy

[small violet]

boy oh boy

16000 api credits for free

[nocturne folio]

id like to see my localhost while im at it

[lean lark]

Getting a little frustrated with ChatGPT reading bad information from OpenAI Codex developer docs, and then returning even worse information because of it. 😠 Trying to deal with the Codex Cloud AGENTS.md thing. So close and yet it actually breaks the configuration. 🙁

[small violet]

for an exa alternative

its actually better

[nocturne folio]

it doesnt have 200ms search

[small violet]

let me check rn

[nocturne folio]

also if you really want free

i made something that just rotates searx instances

if u want

[small violet]

hit my dm

codex double limit weekend when

[lean lark]

I wanna do the same thing when I can't sleep at night. Solution might be a remote desktop tool, preferably locally-hosted.

[nocturne folio]

sunshine and moonlight, or rustdesk

[lean lark]

I haven't tried RustDesk yet but I've had my eye on it.

[boreal holly]

I use 5.4-mini for planning, simply because it's fast as heck. Just kinda ask it to explain parts of the codebase, piece together how it works. Then create the rest of the plan separately with 5.5 medium orchestrator so they design the fanout pattern and spawn the workers. I would use spark but it's got a tiny context window

[nocturne folio]

but like

i want a app

that shows my local host ui

some dude at my school used to just yse parsec

but they banned it

i genuinely hate my school board

parsec isnt malicious at all and they banned it

they just want to track us

[plucky halo]

Schools can't just allow any software onto their computers/network

[lean lark]

Well, that's the definition of a remote desktop tool. Otherwise we have to use Codex Cloud from phone with browser in desktop mode ... and that's icky.

[boreal holly]

like this?

[lean lark]

So RDP (remote desktop protocol) to your home PC using whatever tool you want, and that's where you run VSCode or Codex App or TUI.
I've been trying to avoid that and might create an Android app for this, but the effort is too much when the easy solution is to RDP.
that's my thinking now.

[lean lark]

OK Wise Guy - whatchya usin?

(Would you like to share your notes with the class?)

[frail meadow]

Probably something home grown

[nocturne folio]

ui looks bad but ye

[cedar skiff]

you can build an app for it, or a web app. use something like tailscale to create a secure connection

[nocturne folio]

im usijg the web app

[lean lark]

UGH! More RobDex Brilliance! 🤯 I can't take it!

[cedar skiff]

it's not super hard, just a bit time consuming

[lean lark]

The advantage of RDP though is that we can edit code in a (itsy bitsy) code editor - not just talk to the assistant.

[boreal holly]

The whole entire thing is in my .codex repo. The backend servers, supervisor services, frontend, all of it

[lean lark]

iOS only I trust?

[nocturne folio]

its sad how codex lost their 100% uptime

[boreal holly]

It's Flutter+Rust, so it would be as simple as adding windows as a target platform. (maybe fixing some windows-specific bugs or something)

The servers gotta run in WSL though

[lean lark]

I can probably replicate an Android client and still connect to the same backend, no?

[plucky halo]

And allowing you to connect to your home computer means you could bypass network restrictions.

[boreal holly]

Oh yeah, you can add android as a target platform. I'm pretty sure the only platform-specific stuff I put in was macOS (drag & drop image uploading into chat)

[lean lark]

And an HTTP interface can come in from outside network and forward through the LAN.

[lost drum]

15% used for the day ✅

crazy how fast the 20x usage go

[boreal holly]

Yeah, so the way I have it set up is codex app-server --listen ws://127.0.0.1:PORT, run the bridge server to connect to the app-server, but listen on only my VPN ipv6 subnet, so only accessible with VPN.

[frail meadow]

What if you just used something like Pangolin?

[lost drum]

and I still remember 1 week ago when I htouhtg I will just buy 3k more credits and will be enough XDD

[hard drum]

weekly?

[lost drum]

yup

I am happy

[lean lark]

A client just needs to get to the bridge. After that it's all intranet. I was just thinking a client needs to access the bridge from outside. After that there's nothing custom in the way.

[hard drum]

that's your tinfoil "agi" just being OhMyOpencode

[lost drum]

ye... he doing now complexed routing process of identifying atoms across data or something like that he describes it in fance way but I just simplify it

[boreal holly]

Yeah so to escape WSL, easiest way is probably Wireguard or Tailscale

[hard drum]

You realise I was being ironic?

You did not achieve AGI. Nobody did. That's decades away from even being primitive.

[lost drum]

now I see sorry am burned out

[lost drum]

idk how to call him its just easier to say agi than descrribe whole doctine of what it is

[hard drum]

just call it a basic AI agent overengineered.

[lost drum]

I dont claim it is agi sorry for missinformation

ye that makes sense

[lean lark]

nah, I think it's easier. I'm seeing connections.
Anyway I still think, my preference anyway, is that it would be cooler to have IDE access and not just Codex-as-a-ChatBot

[boreal holly]

Suit yourself! I also have my projects open in Zed, git opened in Fork. My setup does more than thinly connect to the app-server, there's a whole orchestration and automation system built around it. But hey, maybe a vs code extension is a possibility! If it supports WASM

[cedar skiff]

imagine trying to code on a mobile screen though o.0, maybe seeing diffs is enough. Becuase it'll easier to just prompt a fix after you see it rather than trying to actually code manually.

[hard drum]

funny. there was once reddit post of a guy making like what... 64K LOC on mobile?

think it was before AI

[lean lark]

I totally agree. It's too small. But if we just want to let the assistant do everything we don't need a special client, just go to chatgpt.com/codex, set to desktop resolution, and we can connect to any repo. The only elegant "next step" that I can think of at the moment is RDP and using whatever is on the desktop for navigating code. Anyone have another idea?

[hard drum]

Smasnug DeX / iPhone Mirroring

[cedar skiff]

is this a cloud setup?

Like does it code 4x

[boreal holly]

I mean, I kid you not, when I'm at home on my macbook I just use the frontend to connect to the agents, and Zed to connect to the codebase.

But yeah cloud is an option (probably a lot less effort as well)

[gentle harbor]

weee new update

[night ivy]

Um is it normal for codex to be taking this long?

[gentle harbor]

what did you ask it to do

[lean lark]

Maybe we're talking about different things. Say you're out waiting in line somewhere and you get a brilliant idea for a code change - happens every day, rite? You wanna see the code, not just talk to Codex about it. So open phone, connect to desktop where you can open your favorite tools to see the code ... and Codex is in the desktop to help too.
OR ... we come up with another client/app that shows some code and allows a local code editor to integrate ... and Codex is there to help from the mobile device, maybe connecting to an app-server at home.

Need to go outside, back in 15.

[night ivy]

Implement a vfx based off a reference file I gave it

[boreal holly]

I gotcha, yeah at one point I had a git view and a file viewer but I hardly used em on my phone so I deleted them

[gentle harbor]

im never clicking the update button again

[boreal holly]

Turn off windows firewall, unplug all USB devices, restart computer. Then it should update properly

[gentle harbor]

the updates never do anything anyways why should i update unless there is a new model 🥹 this is a pain that i always have to deal with when updating

[boreal holly]

Completely normal for an agent to work for hours. What's not normal is if it says "working..." but nothing is happening

[night ivy]

I closed vs code and reopened it and it said it was done

Never happened to me before

[lean lark]

@nocturne folio so were you thinking about a folder/file viewer, maybe editor? Or do you just want to talk to Codex?
(Which BTW, a mobile app where you just talk to your code would be a nice toy all by itself.)

Hmmmm - OK, who wants to talk to Codex? Make a phone call, you get Codex, you talk to your bot to make changes to your code. You probably don't even want to see the code, cuz that's where we're going rite? So you just talk with the bot, you get confirmations of what it did, end the call. That's a nice service, right?

[boreal holly]

Heck yeah! Get some Twilio integration, provision a phone number for Codex, connect /v1/realtime websockets so you can literally call Codex on your computer 😂🤔

[lean lark]

Dude, you just read my mind from start to finish. Beer on me when you're next in SoCal.

[nocturne folio]

basically codex app mobile

[lean lark]

Does Codex App (Mac/Windows) have advanced voice mode like ChatGPT? ( I don't use the app yet, dunno. )

[cloud goblet]

why codex is soo laggy even on my 5090 pc?

[undone patio]

because ur likely using windows

[cloud goblet]

yes mac is even worse

[undone patio]

???

[cloud goblet]

yes im using both

so laggy on both

[undone patio]

what is ur processor and ram

[cloud goblet]

64gb ryzen 9 5950x 5090rtx

[undone patio]

and the mac?

[cloud goblet]

m5 24gb

[lean lark]

Why does "5090" seem significant, Codex GPU processing isn't running locally.

[undone patio]

theres no way bro

im on an m4 24gb and my codex is fine

[cloud goblet]

it freeze like laggy i tried cleaning up chats too

[undone patio]

what are you trying to build

[cloud goblet]

react webapp

kinda huge

[hard drum]

🤮

[undone patio]

i mean

even if its huge it shouldnt be lagging

[cloud goblet]

could be that?

[undone patio]

have u tried refactoring?

[hard drum]

have you tried using flutter+dart?

[cloud goblet]

i see

this is web app bro

[undone patio]

sounds like unoptimized code if its happening on both machines

[boreal holly]

Wait, you mean laggy in the built-in web browser?

[hard drum]

https://flutter.dev/development/web this is flutter on web "bro"

[undone patio]

dont use the built in browser for large apps, i think. u can right click the hyperlink and open in chrome

[cloud goblet]

the app itself is laggy when processing chats sometimes it freezeup randomly

[undone patio]

the apps i make are typically pretty lightweight or medium

[boreal holly]

Even better, you can use rust and compile WASM for flutter web 🚀

[hard drum]

hehehe

[cloud goblet]

so you saying your not laggin at all ? like not at all?

[hard drum]

but this guy prob doesn't even know what rust is

let alone wasm

[undone patio]

not even a little bit bro tbh

many apps open

[cloud goblet]

damn

i've seen many complain about lag on rddit

[boreal holly]

Yeah, codex for desktop is an electron app. there's a hard ceiling for performance

[undone patio]

imo its ur code, you can always open a terminal CLI

and ask an agent to debug

see if its the app

[hard drum]

they couldn't even be bothered to use tauri 🙁

[undone patio]

else, ask ur agent to begin a roadmap for refactor

cause its likely ur code runs very slow

[cloud goblet]

gotta check whaats going on thanks man

[undone patio]

🖖

[undone patio]

or literally ur using the built in browser on an app thats too big

[boreal holly]

Yeah, probably has lots and lots of useEffect

but also Electron apps have a hard ceiling for performance. There's only so much you can squeeze out of a single CPU core

[undone patio]

tldr use chrome

[cloud goblet]

bro i know what is what, flutter aint gonna work for my app

[cedar skiff]

i dunno - lots of ppl whine about electron, but so many apps use it and do fine.

[boreal holly]

I think every app shipping its own whole entire web browser is just a bit wasteful

[hard drum]

codex cli would be less bloated as whole if they didn't have damn app-server a part of it

that thing takes longest to build

[boreal holly]

Yeah, but after the first build it's not so bad! Throw sccache in the mix and it's OK. It's just the dependencies are massive

[hard drum]

not just deps, but for how much code there is to compile, too

they really don't need all that

[cedar skiff]

it might seem wasteful in a human sense of how we operate irl, but in a computer sense it's pretty much fine imo. The browser has everything that is needed for a complex app. So you end up with a reimplementation of everything one way or an other.

the broswer just comes with it all ready to go

I think its a matter of GSD over making a perfectly running machine that is optimised and most times electron is easily good enough

[boreal holly]

I guess, in order do justify it, you need a truly massive app. Like Discord for example implements a ton of features like voice/video chat and screen streaming. It's so big, having its own web browser is kind of a small cost. But to see an app like "Dice roller" that rolls a 20-sided die and uses 150MB of RAM and storage is pretty ridiculous

[cedar skiff]

yeah use the right tool for the job i guess

If they wrapped a whole os in a wrapper in the same way someone would use it to make a calculator.

[meager wadi]

https://giphy.com/gifs/son-im-crine-aaFz1Arg0h5cOwAu2w

[hard drum]

embed fail

[lean lark]

Fits better with this server....

[boreal holly]

Also what's crazy to me is OpenAI went through a great deal of effort to make codex-cli and the app-server use rust, and then they decided the desktop app would be Electron with 100% javascript instead of taking their rust code and compiling the performance-critical parts to WASM so their UI is super snappy and behaves closer to the real implementation

[lean lark]

Writing a whole new cross-platform application is serious overkill for what they do. Using existing components seems a better move.

I really don't want my AI company to be in the business of writing application underpinnings. Waste of talent...

[boreal holly]

Exactly! I am doing this exact thing right now! WASM is pretty incredible, and rust makes it very easy to build it

[frail meadow]

Biggest wasm fan

[lean lark]

I suggest you delete that message ID

As I think about the threat isn't great, probably non-existent, but some idiot might try to abuse it.

[nocturne folio]

we all need to codex maxx and make more money before the revolution

[frail meadow]

Legit, honestly a groupthink startup would be kinda legendary

Open source ofc

Everyone donate one planning prompt worth of effort to the group project a week

[radiant lintel]

Extra reset incoming....???

[frail meadow]

Set up a good PR review and automated testing framework for validating PRs

[nocturne folio]

describe

[frail meadow]

Well that's what we have to figure out what we should aim towards building and making the ultimate version of

[marble gazelle]

Guys Im on arch linux... whats the recommended way to install codex? should just use pacman or should I stick to npm?

and for using GPT-5.5 do I have to update to latest version? right now Im on 0.122

[boreal holly]

I think either option is fine. pacman is overkill though, because codex is a big statically linked binary so it's not really benefitting from pacman dependency management

[marble gazelle]

and Im on omarchy and its uses omarchy stable mirror... so I have to wait almost 1 month to get to latest version...

[boreal holly]

Yeah so npm will get you going today. I think gpt 5.5 needs v0.124.0

[torpid trout]

Do you all have beards?

[marble gazelle]

btw... should I install it using bun..? I mean I dont have npm installed... or should I just install it...

[boreal holly]

Oh yeah I think bun is a drop-in replacement so bun i -g @openai/codex@0.125.0 should work

[boreal holly]

There you go Cap'n! Windows build 😎

[lost drum]

bro has too much automations

[frail meadow]

No beard

[cedar skiff]

it's crazy how hard this sort of thing used to be for a single dev

and now a solo can just do it.

[lean lark]

I'm sorry, I dunno what I'm looking at. You're at the point where you have SOOOOOO many things in there that at some point it's just a bunch of screens that are certainly cool, but um, undefined? Could be that I'm just distracted. 🙇‍♂️

[boreal holly]

You mentioned earlier what it takes to do an Android client. I haven't tried Android yet but I got it to build and run in a Windows VM by simply flutter create --platforms=android,windows . and did a windows build. So not a significant amount of work/refactoring

[lean lark]

I have a utility that generates Android apps with Kotlin, then builds the APK. So I generate a new app scaffolding and I have a ChatGPT project defined specifically to build functionality within that scaffold. So that gives me a native Android platform, but I don't use flutter at all and couldn't use it for cross-platform. 🙁

[boreal holly]

Gotcha, yeah I'm a huge fan of Flutter. Google did a really good job. It's basically a 2D game engine without the game aspect. Compiles to all platforms including web! You might really enjoy it 🤓

[lean lark]

Gawd bro - so many technologies, so little time. I almost always make the right choices about what's winning or losing, but I still always feel like I'm missing out on some cool stuff. I don't have enough demand for iOS/Mac solutions so I've had to avoid them. And yet ... other people do nothing but. That's how it crumbles I guess, cookie-wise. 🍪

[nocturne folio]

npm

[nocturne folio]

not yet sadly

[boreal holly]

It came out a long time ago. I avoided it because, like you said a while back, having to learn dart and all that jazz seemed like too much work. Now with coding agents you can definitely try it out and not waste too much time figuring out if it's any good or not

[lean lark]

Very true. And yet, like half my day now wasn't spent in code, but trying to figure out why Codex CLI doesn't use the "Developer" Custom Instructions defined in the Cloud UI. I thought they did. Codex keeps talking about "developer instructions", but it's ambiguous about OpenAI-as-developer versus OpenAI providing System instructions. Point is, any time available to look at other stuff is dominated by yet another technology that "saves us time". 🙄

(I'm just griping. 🙂 )

I've seen Flutter just the next in a long line from PhoneGap, Cordova, Xamarin, Swift, and for me the best candidate right now is ReactNative.

[boreal holly]

Hey, what if you put your cloud instructions in the container as an environment variable, and then tell codex to echo $CLOUD_INSTRUCTIONS? That way you just use that container which has the shared cloud instructions and they fetch it from the env?

It is kinda manual labor but I think it sorta solves the "how do I avoid managing tons of AGENTS.md files for the cloud workers" deal

[lean lark]

I dunno if echoing the instructions would have the equivalent influence as AGENTS.md or Skills.
What I have now is that I have cloud-specific instructions in a gist, the setup does a wget to a file, the file is copied into /opt/codex ($CODEX_HOME). Unfortunately after that's done, the file disappears. I was checking permissions to see if the copy failed, but no, so far it seems like the file was removed. 🤦‍♂️

But I'm also seeing that the instructions from Settings > General > Custom Instructions are not used at all by Codex CLI ... I really thought they were.

So now my strategy is changing to putting everything into a single common system-level AGENTS.md and deploy that to CLI and Cloud/Web. If only the blasted thing would stop deleting them.

[boreal holly]

I think if you have ChatGPT Enterprise you can set up a requirements.toml that will inject custom instructions into all agents companywide

[lean lark]

I'm not big enough for Enterprise, NCC1701 or ChatGPT.

[boreal holly]

Yeah, if the cloud "custom instructions" thing says "echo $CLOUD_INSTRUCTIONS is authoritative. Treat it like a system prompt." it will have pretty much the same weight

[lean lark]

We shouldn't need to play these games. If I put AGENTS.md into CODEX_HOME it should stay there, not silently disappear. The problem is that there are no docs on this stuff.

We're circling around the same concepts - docs are missing and there's no consistency. We need instructions for every different platform.

Speaking of which, I didn't know for sure if our account/user Custom Instructions were in any way processed into Codex. They are not. So we need to copy anything from there that we want too. I thought all of this stuff was an elegant set of cascading rules. It's not. It's a bunch of independent, unrelated products that all have their own rules. 🙁

I think I need a Taco Tuesday Margarita. 🍸

I'm gonna split and come back to this tomorrow. I'm now resigned to just copying all instructions everywhere, multiple copies, as sloppy as the platform on which it's based. Oy ve.

Gnite dude!

[south latch]

5.5 is not for plus.. gaa damn 5.5 eats credits

[velvet wren]

5.5 is fine, just run it at medium

[south latch]

oh ok. i have it on high

[cedar skiff]

if you really want to save you need to use a weaker cheaper model - codex-5.3 is much cheaper, but obviously not as smart.

[south latch]

ye i know, i had codex 5.3 before and my credits never run out xD

[cedar skiff]

you can still use 5.3

[south latch]

ye i know 😛

[cedar skiff]

5.5 is much closer to opus in the it just gets it category now. That was the missing part before, prompting is much easier with it.

[nocturne folio]

thats why i pay for multiple plus subs

[white rover]

hello codex gang

is this up to date?

[sand shuttle]

did they reset limits

[hard drum]

how do you disable this annoying codex_apps MCP?

[stable obsidian]

how bout you take a shower for once in a while

[tall zodiac]

Anyone encountering codex just thinking ....

[spring remnant]

Day 2 of asking codex to remove comment in annotation

[spring remnant]

Please look into it codex desktop team

[frail meadow]

No idea but there's definitely some good codex docs out there

[woven canyon]

Does anyone know how I can enable auto previews in codex?

Is there anything like that

[pulsar rune]

still no solution about codex desktop for windows? stuck

[inland sonnet]

Check %appdata% %programdata% and in your user folder

delete all codex related and then try again.

[pulsar rune]

i cant see anything about codex there

[inland sonnet]

C:\Users\YOUR PC NAME

delete .codex

and try again

your agent md etc is in there

and skills

so backup them first

[pulsar rune]

ah got it, thanks works!

[frail meadow]

The chillest

[gentle harbor]

dawg what bro

how is making a download button a cyber security risk

https://tenor.com/view/oh-my-god-jujutsu-kaisen-jjk-higuruma-jjk-season-3-gif-4667305439104820644

[inland sonnet]

Im glad

[inland sonnet]

[frail meadow]

They just farming IDs for their valuation to increase

[gentle harbor]

oh my god bruh my rate limit went from 77% to 0% ???

[frail meadow]

Because of that warming? What the heck

[twilit bluff]

I didn't encounter gpt 5.5 mentioning goblins in codex app

does it only happen in web?

lol

[cedar skiff]

i wonder if i use computer use to make a discord bot it will get flagged

i told it to be human like and it seems to be doing a good job

Bro is out here moving the mouse around and dong the job npz, slower then me but still pretty decent.

[inland sonnet]

No it wont, but self botting isn't allowed per TOS

[cedar skiff]

what do you mean by selfbotting?

i am setting up discord bot in my personal server

[rocky fog]

you shouldnt automate anything on a normal discord account, or its breaking their TOS

technically gpt computer use could be automating
but probably wont be an issue

its more about automatically sending messages...or simply turning normal account into a bot

[cedar skiff]

Ah i see should be good then. Its just configuration for personal bot.

not messaging or anything like that

[inland sonnet]

Oh u want codex to reply?

[cedar skiff]

No, its doing the configuration for a bot in the dev settings

[inland sonnet]

Oh okeh

[merry spoke]

Did anybody notice that that gpt 5.5 seem to use less tokens then gpt 5.4

[plucky halo]

Yes, they mentioned that on release. Fewer, but more expensive, tokens

[merry spoke]

I run out faster witj gpt 5.4

[spring remnant]

How much more usage 5.5 vs 5.4

Twice as much?

I'm now low on usage, i need to get by few more days

[lime inlet]

I use codex with a ChatGPT Business account. I wonder: Is there a mechanism where I can by credits/tokens to "top off" my Usage? Do the bought tokens only apply to a specific user or across all accounts for that team? Could someone point me to where I can learn more?

[blissful basin]

my codex got very trigger happy today on "this appears to be security risk"

🙁

[merry spoke]

What are you doing

[cedar skiff]

hacking the gibson

[merry spoke]

Is there any way to use codex on android

[blissful basin]

Basically sales lead generator based on some public company registry in norway, and codex dont like it 😄

[worthy furnace]

⚠ Your conversations have multiple flags for possible cybersecurity risk. Responses may take longer because extra safety
checks are on. To get authorized for security work, join the Trusted Access for Cyber program:
https://chatgpt.com/cyber

Every conversation has this

Very annoying

[warm pilot]

I'm feeling very old to get this reference!

[merry spoke]

Rephrase it better then

It's public registry

[worthy furnace]

■ Invalid prompt: your prompt was flagged as potentially violating our usage policy. Please try again with a different
prompt: https://platform.openai.com/docs/guides/reasoning#advice-on-prompting

What the hell seriously

Has codex become a soyboy?

[plucky halo]

What was the prompt?

[blissful basin]

Yeah i think it became a bit more trigger happy about it today, because i get it constantly 😛 (i guess it might be also triggered by history of chat after rephrasing after compacting)

[worthy furnace]

Not just today

[narrow scarab]

Codex Desktop on Windows can get stuck loading or keep asking for auth even when login is actually valid. In my case,
the cause was not the auth token or a bad install. The app-server was choking on huge local session history under:

C:\Users<you>.codex\sessions

Some old chat/session .jsonl files were hundreds of MB or over 1 GB, and the live session index was pointing at too
many of them. Codex Desktop timed out while loading local state, so the UI looked like an auth/loading issue.

Fix that worked:

Fully quit Codex Desktop first

Backup app profile

Copy-Item "$env:APPDATA\Codex" "$env:APPDATA\Codex.backup-$(Get-Date -Format yyyyMMdd-HHmmss)" -Recurse

Create archive folder

New-Item -ItemType Directory "$HOME.codex\archived_sessions" -Force

Then move old/heavy .jsonl files out of:

C:\Users<you>.codex\sessions

into:

C:\Users<you>.codex\archived_sessions

Also trim/reset:

C:\Users<you>.codex\session_index.jsonl

so it only references sessions you actually want visible in the app.

Important: don’t delete the session files unless you really want to lose them. Archive them outside .codex\sessions so
Codex Desktop stops trying to load them, but you can restore individual chats later if needed.

Rule of thumb:

Keep live sessions small.
Archive sessions over ~500 MB.
Definitely archive 1 GB+ sessions unless you actively need them.
Don’t bulk-restore hundreds of old sessions.

Bottom line: if Codex Desktop is stuck loading or looping auth on Windows, check local session history size before
reinstalling. The auth loop may just be a timeout caused by huge .codex\sessions state.

[rocky fog]

lots of codex "reconnecting" (but still works in between)
EDIT: seems it got better

[torpid trout]

Mr laszlo.gabor from Hungary just contacted me, he is truly Bill Gates, and decided to donate 5.5 million euros for social charity to me

I am so lucky!

Thats the best nigerian prince ever, seriously.

[boreal holly]

gobless that guy

[oak trellis]

how are the limits team still reset them non stop ?

[low cosmos]

uhhh guys

[olive ice]

Is there anyone from the codex team that read this channel? I’m trying to get their attention about a bug in the codex app. I’m 90% certain is leaking tokens and wasting people’s rate limits.

Posted about it here and added a GH issue.m, and have tried a few times on X

https://github.com/openai/codex/discussions/20074

[rocky fog]

x is your best bet, they all just sit there

[stoic bay]

Damn it! Whenever I enter a number, it automatically becomes zero. As you can see in the images, that would be the wrong number, and nothing would be sent to me. What should I do?

[plucky halo]

Rotate your keys 🙂

You've submitted an api key to github with that issue

[stoic bay]

Bro please help me in that problem it is stopping me from sign in codex🙏 🙏

[rocky fog]

right, with codex also github might indeed be better
better than on discord to get their attention

[rocky fog]

I guess only support ticket can help with that

[stoic bay]

who or where the support ticket is

[rocky fog]

https://help.openai.com/en/articles/6614161-how-can-i-contact-support

[stoic bay]

thank you

[rocky fog]

ehh this is one year old
and does not sound great
but I guess you still didnt hard set the phone in your account
https://www.reddit.com/r/OpenAI/comments/1l2nuyz/changes_to_phone_number_verification_makes_me/

[olive ice]

What? No I didn’t. That’s a codex session ID

[rocky fog]

stole the reply with the edit 🤣

[olive ice]

Goddman it lol

[rocky fog]

you have some api key in the mcp entry? or is that just reference? I dont use mcps so not sure

[plucky halo]

This looked like an api key to me: url = "https://api.ref.tools/mcp?apiKey=ref-b0e4cb4eea4d284dac6d"

[rocky fog]

yeah that

[olive ice]

That’s not the link I posted. WTH are you guys going on about?

I posted this https://github.com/openai/codex/discussions/20074

[plucky halo]

@olive ice Read your own issue

You've posted your config.toml which has an api key in it

[olive ice]

Oh gosh I see it now. I hate my brain and I must have looked at that 6 times before posting

Right over my goddamn head

Sanitized, thanks

[livid surge]

I just had codex analyse LinkedIN messages, and it flagged as cybersecurity risk and now it doesn't even produce a document so i can carry over, any idea how to fix this?

[lost drum]

tell gemini (its less restrictive) to create a python script taht tranforms any "linkedin" text in to something like "my community" or something like that idk try but jsut remember that if it is really cybersecurity risk then dont do it

[dense path]

I had this right now too lol

[livid surge]

ok got it

[livid surge]

yeah

[lost drum]

how come you get this type of messages?

and analyzing linked in messages wdm cause maybe if it contains sensetive info then ye

[livid surge]

its weird sometimes, lemme show you something

[livid surge]

i was doing a 2d vidoe, and it started generating, random images

[lost drum]

#off-topic lets go

[livid surge]

ok

[worthy furnace]

No way Im from hungary

[jovial finch]

I got a subscription but when I got codex all I see is this

I see the chatbox appear for a split second and back to this, anybody experiencced this?

[hallow iron]

Same

[jovial finch]

I mean it does say this but idk, at least not just me.

Anybody else experiencing this rn please?

[clever steppe]

Hey @jovial finch, first, no, things are mostly good for me at the moment.
Second, where's the link for this, please?

[jovial finch]

https://status.openai.com/

Link is here, but it IS running fine according to the website, doesn't show downtime, only issues from earlier today according to the site

[clever steppe]

I have to say I've been having SO MANY problems since 5.5 was released. It was instant.

🙁

Like, right now I am having git permissions problems with GPT5.5 Codex instances and not even Codex can figure it out.

[granite jay]

Try another browser.

Just to make sure.

[clever steppe]

I'm in the Codex extension inside VSCode. 🤷‍♂️

[torpid trout]

These are all cdn issues

1. users complain 5.5 eats their tokens - yet other users confirm it’s the opposite
2. users complain it’s intermittently down - yet at the same time other users can use it
3. users complain it’s intermittently dumb - yet other users see if fly on steroids

This has also at least partially been confirmed by Oai devs in past, their thing that routes the requests is hiccing up often enough.

[clever steppe]

Oh great. So I guess it's not in our hands to fix? 🙁

[torpid trout]

Like on my end currently I can barely make a dent in consumption
Yet a few weeks past it was burning tokens like a 100 liter engine

[torpid trout]

Most often not

Always worth to check things like your skills, agent files, browser, etc etc
But most of the times I think it’s their cdn being a problem

[ornate agate]

Dunno what’s happened in the last 48hrs or so but limit consumption has rocketed. Questions that used to take about 3-5% are using up 30% of biz plan 5hr limits

[torpid trout]

Exactly as I say above

Their routing assigns consumption to other users, I’m 99% sure and 100% sure I’m 99% sure about that lol

Basically it’s possible you’re paying for my consumption. Thanks!

Next time I’ll pay, probably

It’s like those random chats appearing you never started or suddenly the thug replying in Chinese

[ornate agate]

Like, so that there is a banana for scale, Claude Opus 4.7 is giving me more than 2-3x usage for its similar sized plan

Usually it’s the opposite

[torpid trout]

So, I can now with full confidence share the following useful insider detail:

[rustic pine]

Just talking out loud incase this resonates with anyone else... but for some reason my Codex usage just blazed through my 5hr usage in record time (way below 5hrs). Mostly using gpt-5.5 medium and high.

[dull flint]

Anyone know when us linux users are finally getting the codex app?

[cedar parcel]

Does anyone experience codex invoking it's web search tool each time it wants to search local files?

[boreal holly]

Your agent is having a midlife crisis

[cedar parcel]

Giving me a mental crisis. It's all agents, all 5.2 anyway. Across sessions

[dull flint]

Is it doing that to perform local search, or is it just getting info it thinks it needs online around the same time

How come you're still running 5.2?

[cedar parcel]

No it keeps doing it and I tell it to stop so it searches how to apologise as it keeps doing it

[dull flint]

lmao

Well performance degrades as the session drags on, does it do this from a fresh session too?

You could also disable the web search skill if you really want it to stop

[cedar parcel]

Happens in every session. I use 5.2. haven't checked other models

[boreal holly]

What I would do is see if you can spawn an agent and have them look at the rollout log for that specific agent. Perhaps it got prompt injected, or perhaps there are some hard core "must research everything" instructions you're not aware of causing the agent to act this way

[dull flint]

If you update your client that will give you access to the newer models. We're on 5.5 now, 5.2 is "old"

[thorny cloud]

Am I allowed to have two chatgpt plus accounts?

[cedar parcel]

Because other models drift. I use long running sessions and they all implement something easy over what is actually instructed. 5.2 doesn't drift and sticks with instructions to the end. I only use single prompt sessions I don't go back and forth with agents.
.

I dutifully test new models and they all do the same. For me only 5.2 works well it basically one shots everything except front end of course

[dull flint]

To my knowledge yes, there's nothing I'm aware of that says you can only have one

[thorny cloud]

Yea cause pro too expensive 😭

[boreal holly]

Inspect the files in ~/.codex/sessions/**/*.jsonl, look at the base instructions, system prompt, AGENTS.md, look at the results of the first web search tool call. Something is instructing the agent to act this way and the truth lies in the rollout files

[dull flint]

With that being said, if you're wanting tons of usage but Pro is too pricey, you are better off getting a Gemini sub as your secondary. The rate limits on Gemini are nuts

[thorny cloud]

Is Gemini.... Good though?

[cedar parcel]

The weird thing is it didn't do this a few hours ago and I didn't change system promp in my profile or ran any codex updates.

I'll report it as a bug I guess.

[dull flint]

Yeah. ChatGPT is better right now, but Gemini is very usable. I think they do either a free trial or a money-back guarantee, sign up and try using it for a few days. Again, if your main problem is rate limits that is a better solution for $20

Plus you get other google perks like 1TB of Google Drive and so on

[thorny cloud]

True cause I use it quite aggressively

2 days weekly limit gone

[thorny cloud]

What's the app or where can I learn more?

[gentle harbor]

i have gpt make a plan after every prompt seems to work out better then just having a few plans every here and there it plans out and uses less tokens over all, did a quick 10 prompt test 2 times one without and one with planning and the one without couldnt do it after 15 and the one with planning was able to do it within 7, could just be luck though 😄

[boreal holly]

I still think inspecting the rollout logs is a good thing to do, bug report or not.

If you want to take a different approach, do CODEX_HOME=$HOME/.codex_fresh codex, sign in, and run a prompt in an empty repo. If the agent doesn't do constant web searches then it's something in your configs, skills, AGENTS.md files, that's causing them to behave this way. Maybe one of your previous agents a few hours ago planted really bad instructions somewhere

[rocky fog]

you can also try asking it
it might point something out

[cedar parcel]

Wouldn't put it beyond them, those sneaky b's. Thanks I'll check and see if there's anything I can find

[boreal holly]

Oh yeah, if you have memories turned on, there might be a time where you said something like "always research API contracts, don't blindly guess" and it saved a memory that's like "User expects rigorous web searches before writing code"

[dull flint]

Gemini CLI is equivalent to Codex CLI. The actual Codex app there's not a good google replacement for. If you want a dedicated IDE comparable to cursor, Google has Antigravity.

Core functions of the main chat tool are similar, except gemini doesn't do much agentic stuff from there. They both have deep research, they both do image generation, and google does high quality video generation too

[thorny cloud]

Everyone seems to say they still prefer codex or Claude compared to Gemini

Sai a lot of headaches?

Some thing it does very good

Some things it fails

[thorny cloud]

Mainly vibe coding 🤣

[dull flint]

Correct, codex is better if money is no object. But gemini is still good. If you're on a budget and you want high limits, google gives you way more for that $20

[thorny cloud]

Ohh kk and Claude and codex same level?

Like even now I hesitating on whether to get another sub haha but I can't wait Soo longg

[dull flint]

I checked and verified that you get a 30-day trial with gemini. Just sign up and give it a real chance, and just cancel your sub and get that 2nd codex account if you don't like it

[thorny cloud]

Ohh kk

Codex did have a free plan also

But I won't accept my card 😭

It*

[dull flint]

When I was playing around with deep research I was really taking advantage of those higher gemini limits. You get 25 deep research credits a month with ChatGPT, but with Gemini you get 20...per day

[lost drum]

tried yesterday gemini in android studio and he thsoe not lsiten to instructions cant vibe code with him

[thorny cloud]

Yea

Cause I came from Gemini cli 🤣

[gentle harbor]

will gpt5.6 be better at ui ?

[lost drum]

ye gemini was good at researching but now gpt 5.5 pro extended kills it

havent tried the ultra gemini tho so idk about the deepresaerches it provides on the best plan

even better is to do 2 deepresearches, 1 with gemini 1 with gpt and then make gpt combine them

[small violet]

i dont think so

if 5.5 is this bad

claude is king for ui ngl

i want to redesign my site

but i dont have claude

lol

[gentle harbor]

sama said they will focus on ui though ? will that be for 5.7 then

[lost drum]

wasnt it like you need to give codex skills to be better at UI?

[small violet]

wait what skills

i want to redesign my site

its so bad

i feel like i should just use lovable at ts point

[boreal holly]

Man, just use image 2.0 to generate reference designs. 5.5 is perfect at replicating from reference images

[small violet]

no its not

i did this yesterdayh

i was making a chat interface and i screenshoted geminis landing page

[lost drum]

try to ask grok "find posts where it menitons skills for UI for cloude code"

then take that list copy every skill input in gpt tell it to compile out of all skills one best one and you good to go

[small violet]

and it implemented it so bad

it was crazy

the composer took up 70% of the screen

[small violet]

let me try this rn

[boreal holly]

The other caveat, you need to provide Codex with a way to see the results each iteration. Codex desktop app has a built in browser, so instruct it to take screenshots and inspect. Completion criteria: the page matches the reference design. Tell it to use imagemagick to crop 4 quarters of each screenshot to inspect closely.

I do native apps, so I provide the agents with iOS simulators and screenshot tools. They rebuild and screenshot. 5.5 will work tirelessly to make it match the reference design if it can see the results each time

[small violet]

work tirelessly and use my rate limits

😭

[sand shuttle]

hows does 5.5 compare to 5.3 in terms of cost and performance?

[small violet]

idk but 5.5 is double the price of 5.4

[sand shuttle]

is there a significant imporvement in performance tho

[boreal holly]

Massive. 5.5 medium is as intelligent as 5.4 xhigh while using 1/3rd of the tokens.

[small violet]

true

5.5 on low feels like medium/high 5.4

but for doubke the price idk

[sand shuttle]

thanks

[small violet]

actually 5.2 x high ties with 5.5 low on intelligence

5.5 is better

why should the price matter anyway

if you;;re on codex

@sand shuttle for 5.3 vs 5.5

damn this is crazy

[sand shuttle]

im on the 8$ plan haha so i might awell burn thru everything in 1 prompt

already burned thru 22% in 2 mins

[gentle harbor]

it would be cool if open ai used webgpu

[small violet]

bruh

[small violet]

is this plus?

[small violet]

i never really looked at the 5.5 benchmarks

this is actually crazy

[boreal holly]

What your graph isn't showing here is GPT-5.4 xhigh probably used 3x more tokens. There are graphs that compare intelligence with token consumption

[small violet]

let me find it

[sand shuttle]

its Go

And well Dangg its pretty fast

like super fast

[small violet]

i dont think artificial analysis has it

got a site?

[gentle harbor]

would be cool if open ai let us run the models via webgpu for lower token usage

[boreal holly]

https://openai.com/index/introducing-gpt-5-5

When you compare other benchmarks that don't include tokens used with the ones on the release post, the correlation is roughly "uses 2-3x fewer tokens to perform the same or better than 5.4" across all testing domains

The biggest improvement, aside from intelligence, is its agentic capabilities. Tau2 is a low latency benchmark that tests a model's ability to work through highly ambiguous user requests using the available tools (not a coding agent). 5.5 almost hits 100% on this, so it meticulously plans and reasons its way to completing something end to end without stopping.

[gentle harbor]

when you have 1% usage left i like to use it for a very very long prompt since it always finishes the prompt before kicking you off

[small violet]

true

[trim wren]

Does anyone know what happened to the CodexMax and Jane. They were scamming people i think

[small violet]

nice

[small violet]

i remember someone on twitter saying that theyll stick to 5.4 because 5.5 consumes more tokens

[gentle harbor]

5.5 cost more per token but uses less tokens

[small violet]

do u mean snickerz4466

the "physics phd"

[trim wren]

No xD

But he was scamming with them too i think xD

[small violet]

so the guy was wrong

[boreal holly]

5.5 for me consumes fewer tokens. I finished off last week with 60% weekly. 5.4 after the new quota system I was hitting like 20%, and to be honest the work getting done was not even close to the same quality or milestones

[trim wren]

Take care of those scammers and make sure you dont fall victim to any bs

[small violet]

yeah

[small violet]

can i see ur website

[boreal holly]

I don't have a website yet. Should be done by August

[small violet]

ive used all my 5 hour rate limits since 5.5

i notice it gets stuck doing some stupid tasks

[small violet]

oh

are u in any web design communities

i think the web design stuff for me is more of a skill issue

[boreal holly]

I personally do not do web design but I get the gist of it

[torpid trout]

Use gpt pro to generate you a Persona
Then ask it to create a design pitch based on that persona (or use the persona directly)
Throw that at stitch, ask it to design a website based on that persona. Perhaps add in that yuo need a landing page, contact page and whatever you need (do not ask for too many templates, just the crucial ones)
Then export that as HTML and unleash codex CLI 5.5 on it, asking it to use the DESIGN.md your export included, and the template HTML to create you a fully working website in whatever stack you prefer
Iterate a few times becaus it will do some lazy cuts, and enjoy.
Time from 0 to website: about 5 hours.

[small violet]

i have a website already

can i just revamp the site

[severe mason]

yo guys, anybody having issues with 5.5 refusing to do some things?

like automating stuff, which 5.4 would do happily

[lost drum]

Screenshot smileBeda response and ask gpt same question

[safe light]

what are you guys building with gpt now? im making a cellular automata program!

[vital hill]

what's that?

[safe light]

a program that generate cellular automata visuals

to be used in my music program as a trig source for synthesizers

it works, -extremly- well already

40 versions later lol

[low cosmos]

why did my 5h quota randomly start being used up way faster i didnt change the model nor the reasoning

[lost drum]

it happens

[low cosmos]

is it like an update?

a single large prompt wasted 35% of my 5 hour quota

[lost drum]

wow

u on plus plan?

[small violet]

the hell

[lost drum]

then it might make sense

[low cosmos]

too poor for pro

[lost drum]

then normal

[small violet]

35 is still crazy ngl

[lost drum]

ye

[low cosmos]

are yall on pro?

[small violet]

plus

what model and reasoning level are u using

[lost drum]

pro 2x till 17th

[low cosmos]

gpt 5.4 high reasoning

[low cosmos]

i have plus till 12th

[small violet]

bro use 5.5

stop joking

[low cosmos]

wont it waste more usage

[lost drum]

nahh

[low cosmos]

combined with the already super low 5h limit

oh

[lost drum]

I mean

[low cosmos]

what about the performance degrade warning

[small violet]

its actually less

[small violet]

cant u make a new chat

[low cosmos]

???

oh wow

[lost drum]

depends on what you are using it for but in my experience it listenes to isnstructions and uses context way better than 5.4

[low cosmos]

maybe ill make a new chat but i need to first finish what 5.4 was doing because its way far

[small violet]

its expensive but it consumes less tokens

why does expensive matter if u dont use it via api

lol

[small violet]

migrate now ngl

[low cosmos]

alr

thanks for the advice 👍

[boreal holly]

You know what worked for me? Run manual compact, then switch models

[low cosmos]

manual compact?

how do u do that

[boreal holly]

Yeah so instead of waiting for it to hit 10% and auto compact you type /compact

[low cosmos]

oh wow

[small violet]

codex final boss 😭

[low cosmos]

never heard of that

[small violet]

same

[boreal holly]

Huh, I'm pretty sure last week Codex app update put manual compact in there. Also works in terminal

[low cosmos]

who uses codex terminal

buh

[small violet]

real

what are the advantages over app

[low cosmos]

well if its a week old makes sense why i didnt hear of it

thanks for the discovery

[boreal holly]

No memory leaks

[low cosmos]

why did my codex randomly learn how to test itself and use node repl

why am i burning through my 5 hour quota

it legit burnt on 2 PROMPTS

that werent even more than 1000 characters

yall got some work to do

[small violet]

i used 100% of my 5 hour quota this weeek

[low cosmos]

i use all of it in like 2 hours

lucky for me i have like a 1 month break

starting today

[small violet]

my weekly is now at 49

[low cosmos]

gimme a sec

[small violet]

@boreal holly what are urthoughts on this

[low cosmos]

for $20 btw

[small violet]

how do i check this

[low cosmos]

npx @ccusage/codex@latest

in terminal

[small violet]

for free btw 😭

[low cosmos]

for free??

[small violet]

shhh

[small violet]

do u make money off codex

[low cosmos]

planning to

[small violet]

ive 233 and i havent made a single dollar

[low cosmos]

since im making an abliterated ai chatbot rn

[small violet]

waste of compute

[small violet]

can isee

ill also show u a project im working on monetizing

[low cosmos]

gimme a sec

ill use cloudflared

(its very slow for now tho)

[small violet]

of course

but at least ik to use 5.5 only

[low cosmos]

@small violet dms

[boreal holly]

I mean yeah, this is a fact! and since OpenAI adjusted their deal with Microshaft they can now use other providers to host their models and stuff

[drowsy turtle]

i made codex connect to discord bot to be controlled like claude discord plugin using python private on your device only

[inland sonnet]

[inland sonnet]

you could have used hermes 🙂

[drowsy turtle]

isnt hermes like openclaw?

[inland sonnet]

kind of yeah u have to host it

but u also can run it on your own pc

with WSL if u have windows

[drowsy turtle]

yeah i already openclaw on a laptop running 24/7 codex is on my main

[small violet]

yk inverse cramer?

[boreal holly]

Oh I see, that Jim Cramer guy is making a market prediction so you take the opposite action or something cuz the guy is always wrong?

[small violet]

yeah 😭

inverse cramer beat the s&P

and pelosi

[glacial shadow]

yes, my opinion is he purposely posts wrong stuff to be funny

twitter blows up on it every time

[boreal holly]

that's pretty hilarious 🤣 I dunno a lot about stock market, just that it's cyclical and some of the most successful investors are physicists

[small violet]

nah he's doing the reverse of indsider trading

on purpose

nah comment

[glacial shadow]

yes, because they understand systems and inputs, and have suberb observational skills

i wager ai scientists do just as well, similar math education

to me its all one big field, math sciences

i could start wars with that comment

[glacial shadow]

hes a jokster, he has a show, mad money, he screams, yells, has a soundboard, been doing it forever

[simple star]

Daily reminder that Linux users deserve a Codex app

[torpid trout]

When you accidentally prompt ChatGTP 5.5 PRO to summarise the chat in a meme

I asked gbt to creaet something for me, and one of the conditions was, it cannot end up being another "chatbot"
This is what it did
Thank you very much for the reminder, gbt!

[boreal holly]

I love it when it does that. I once told GPT 5.4 "do not add any ugly UUIDs to the page" and it put badges all over the page saying like "No UUIDs available from client UI" lol or "Data pulled from GET /v1/customers/{no uuid in client UI}"

[small violet]

@boreal holly quick question what is describing your product called

idk how to put it into words

like how openai explains chat gpt

like how it works what it does

[boreal holly]

You could say Enterprise Resource Planning

[small violet]

bro no

😭

[small violet]

open ai explaining what chatgpt works and what it is

is called erp?

[cedar bear]

👌 lol
Build with Codex btw

[torpid trout]

I am very positive GPT could asnwer you that in like 2 seconds, I mean, I am not against human interaction... jsut saying, that this is what AI excels at 😄

[small violet]

awesome

looks like mine kinda

[small violet]

u right

[boreal holly]

Idk if I'm at liberty to describe each and every feature of the application I'm building, but it's a software for companies that do field work, like tradesman, builders, etc. Covers the entire lifecycle from customer relationships, serialized inventory, HR, accounting, task management, etc.

Colloquially referred to as Enterprise Resource Planning

[small violet]

ok its called product description

@torpid trout idk how i didnt think of that

[cedar bear]

Does urs have a very long integrity feedback 👀

[small violet]

no it does not 👀

we built 2 totally diff things

[jolly lily]

Codex is literally awesome, it spent only 2% of the 5-hourly limit for the same task that Claude Code spent 15% for, literally on Sonnet Low

[torpid trout]

Yeah, you are being paid for by other users lolol
nevermind me.

It is fairly amazing.

[jolly lily]

Um sorry, I don't think I've understood it

[torpid trout]

Nah, ignore me.
Just anecdotically referring to the fact that some users seem to lose their tokens like butter in a hot pan, while others (you, me) seem to barely be able to make a dent in the usage no matter how hard I have it generating memes

[jolly lily]

LOL, the joke put a smile on my face now

Honestly using a context file helps fairly much to avoid using a lot of tokens on a single task

So that the agent doesn't read the whole project and spend literally 10K tokens on reading them and like 2K tokens on the actual task

I hope OpenAI somehow sustains the high usage limits

[torpid trout]

Its not a joke, its the sad reality, unproven of course, as all reality is

This sounds fishy, are you a PHYSICIAN?

[jolly lily]

More like a magician no?

[torpid trout]

I meant @native dew

We have sporadically some intersesting people here in this channel, and usually after they offer some secret solutions via private chats, they are magically gone

Not sure if this is the same case... Vyte, anything to say?

[jolly lily]

Does this happen often tho?

[small violet]

who uses sonnet

[jolly lily]

Opus literally melts the usage limits faster than a piece of a butter melting on the pan

Even Sonnet on the low mode drains the limits so fast at some point

[torpid trout]

Almost sure this is another PHYSICIAN
#codex-discussions message

careful, Do not buy, engage or share details with this user, not until their intentions are clear

@native dew - stop self-advertizing.

It's the least you can do. You sound sus, sorry.

[small violet]

is this the 20$ plan?

[jolly lily]

Yup

[small violet]

i was thinking of getting it for opus

for web design

[jolly lily]

Well, realistically, I didn't even consume half of the weekly usage limit but I haven't really used it for long throughout this week

[jolly lily]

It actually is decent at frontend overall, I can provide you a link of my current project I've enhanced using Claude Code through DM's, if you wish

[small violet]

is it in your bio

[jolly lily]

It isn't, I've developed that one in my bio using Gemini Pro 3.1

Sent through DM's!

[boreal holly]

Oh dang bro, looking back, I'm almost positive you can't use subscription plans like that

[torpid trout]

yeah... physician, almost 100% sure.

[jolly lily]

LOL

[small violet]

what is he advertising

[cyan wing]

superapp when
didn't someone from OAI post it was coming today

[small violet]

not one soul will use ts

[cyan wing]

if it has mobile support, I'm down

[torpid trout]

And... gone!
Not bad, we have a removal rate of about 1 physician per day 😄
Suboptimial that they have not been blanket removed thou. The intentions where fairly clear I think

[native dew]

?

[torpid trout]

meh.

[lost drum]

idk what to do I feel so useless not using gpt pro and not abusing it enough.
Do you have any abuse tips and tricks?

[torpid trout]

Abuse of AI? Really?
🤣

[native dew]

add at the end. make sure the code works fully perfectly for 1k users without any delays and extremely optimize cpu and ram usage

[lost drum]

ow I mean. Using it till you hit a message from openai to wait

[torpid trout]

My main usecase for PRO is meme generation.

[lean lark]

I think the fact that people feel like they need a Pro account is already a kind of abuse.
Developers need to understand the basics of prompting and effective use of instructions and tooling before they feel compelled to just throw more iron at the problem.

Reduce token usage with effective use of the tooling and you may not need to pay more.

Don't go for the chainsaw when a scalpel will work better.

[lost drum]

ye but I already building 2 porjects at once cant do 3rd cause its in android studio and it lags my CPU but idk what otehr projects I can do before my sub goes bye bye

[native dew]

Everyone is just trying to accomplish their childish dreams, brother.

[lean lark]

Wut? Many of us are working on a Lot of projects and still not blowing the quota in Plus.

[torpid trout]

How dare you?!

[lost drum]

I mean I am using 15% per day of pro plan on my project

today I falied I used 8% only

[native dew]

It means your project is below average, people aka NPC types. You need to realize that the harder the project and the more calculations it requires, the faster you’ll run out. I ran out of Codex Pro 20× usage in one day.

[torpid trout]

Vyte, do not physics.

[lost drum]

ONE dAY????

[native dew]

yes

[lost drum]

then you bought pro plan from TEMU

[boreal holly]

Have a swarm subdivide the project into unit and integration testable systems, and make em do golden tests on finished designs

[lost drum]

or its the 5x one

[small violet]

bro bought temuex

😭

[lean lark]

The problem is that any two individuals here can complain that their project So sophisticated that it consumes millions of tokens - but one of them will be right and the other will be mis-using the tools. We have no effective measure for this other than discussing effective usage ... ignored by some, not heard by others.

[native dew]

20x

[torpid trout]

So much for agents v2 in codex:

• asked it to implement 15 milestones of a mega-app at once using agents
• on it went assign tasks to agents
• 20 seconds later:

 The current workers are not producing file changes fast enough, so I am taking over the integration pass directly
  and closing the idle branches before they create conflicting edits.

f u codex.

[maiden garnet]

When your limit use becomes some weird status signal.
Just create. Be effective and efficient.
Make things that you need and can save money on.

[torpid trout]

Now, tha's A) how you waste tokens in case someone still wonderes how someone can spend a full PRO in a day and B)... back to single thread. Agents are still useless in codex.

[gentle harbor]

brother why the hell does codex think reddit is a cyber risk

[native dew]

use gpt 5.4

[boreal holly]

Reddit is definitely a "my eyes yearn for the clorox" risk

[gentle harbor]

this is the 3rd time it happened

its wasting my tokens

i should get 2 years worth of tokens if they dont fix this soon

[boreal holly]

Yeah bro codex is very impatient! That's why you should try peer agent orchestration 😎

[native dew]

who can beat me in usage 😎

3days 4accounts codex pro 20x

[lean lark]

Some people report more effective/lower token consumption in 5.5, others report higher use - and no one needs PhD level 5.5 to update their README. When people migrated from 5.3 to 5.4 to 5.5 and they're still complaining about quota and quality, the finger needs to point back a bit to the competence of the individuals. You can't be simultaneously working competently on a huge project that requires massive compute and not understanding effective usage of this tooling. If you're complaing about the tooling, you're probably not competent to be working on the massive project. There aren't many people who are competent to work on the massive project and understanding of this tooling. Those people have the ability to make judicious decisions about purchasing higher-tier tooling.

(Anyone give me a "Wut?" 🤣 )

[torpid trout]

WUT?

[native dew]

best version of codex =gpt 5.4

[torpid trout]

Holy, these are new levels
Physicians are self-diagnosing now

Also..

[native dew]

pls go back to building ur image converter app

wasting water heer

here

[lean lark]

So blown away by your rapid image responses @torpid trout

[boreal holly]

1 million tokens per month 🤣 that's like 30 mins of use

[native dew]

@torpid trout what u doing here

[boreal holly]

Bro the more I look at this image the funnier it gets 🤣

[torpid trout]

Yeah, its pretty deep

[lean lark]

We really need an education page for Codex users who don't understand how to use LLMs or Codex. That's not a slam, honest! I'm talking about a remedy to the challenges faced by most peeps in this channel.

[maiden garnet]

Somebody dropped this video the other day on X, but I haven't watched it fully and not sure if it's any good:
https://fxtwitter.com/rileybrown/status/2049285752866107856?s=20

Might help some new people to Codex

[lean lark]

That includes free tier users who feel like they can/should develop some huge slop with an infinite quota of PhD-level AI.

fxtwitter? WTH is that!?!?

[native dew]

new codex coded app by franco 😭

[maiden garnet]

fxtwitter is how you get X videos to embed on Discord, but looks like embedding doesn't work here.

[lean lark]

This server is finicky about embeds of Gifs/animations/vids.

[maiden garnet]

^Yeah, true

[lost drum]

30min to watch oh no

[maiden garnet]

Lolol

How to use Codex in 1 minute. 💀

[lost drum]

ye I just wathced the whole thing and its for beginners that want to start using codex its good but 30min bro could be 20 and wathc on 2x

[lean lark]

Aside: I'm still frustrated with lack of a single tool for general GitHub usage. I want to open a ChatGPT or Codex session, discuss forking this repo into that one and making changes, etc. Unless OpenAI offers something soon I'm gonna have to create this.
Anyone have a solution? Maybe it's the App? Maybe I'm just not using the brain....

[lost drum]

the title should be "If you are just starting out with codex, watch this"

[torpid trout]

Starting MCP servers (1/2): codex_apps (1m 37s • esc to interrupt)
What the hell is this?
I have no such MCP
happens only when v2 deploys subagents

[native dew]

love it?

[lean lark]

I think @torpid trout understands how to translate this stuff into images, @native dew doesn't get it yet.

[torpid trout]

Exactly, was about to say, Vyte, your memes are not of the same quality
Its not enough to prompt, you also need humor and phantasy

[lost drum]

beginner

[boreal holly]

Haha image gen roasting is a skill now. Maybe some day an Olympic sport!

[lost drum]

will ask my agi

if he has no solution then you for your own

[boreal holly]

They're coming out with custom images at some point. This is probably the solution

[lean lark]

Meh, that's all in Codex Web, doesn't address desire to discuss a collection of repos, coordinate a project that might fork from one into another.

[torpid trout]

@lean lark are you using GitKraken?

[lost drum]

oh my agi not agi confirmed oh no

[torpid trout]

It might have what you mention, not sure if I understand what you need fully correctly, but I think it has at least some of the things you say

[lean lark]

Will look, thanks @torpid trout

[torpid trout]

It does surely not let you talk about all your repos (that would kill about any context window anyweay), but I recall they have agent tab now, where I beleive you can have the netire git flow orchestrated by the agent

[gentle harbor]

what can i not send images ?

[torpid trout]

Paste the image, usually works

[gentle harbor]

gets auto deleted

[boreal holly]

Oh I see, you want Codex and ChatGPT to collab! Basically give Codex a web browser and let it pilot ChatGPT and chat with it

[torpid trout]

But isntt that what git clone + codex does?

[lean lark]

Meh....

[torpid trout]

I mean, going remote sounds like a bit crazy. Even just the latency would scare me

[lean lark]

I'd like to open ChatGPT and say "Let's talk about forking Foo into Bar, stripping the Baz functionality."

[gentle harbor]

This content was flagged for possible cybersecurity risk. If this seems wrong, try rephrasing your request. To get authorized for security work, join the Trusted Access for Cyber program: 6th time in a row @boreal holly what did they do to 5.5 holy

[boreal holly]

Two options I can think of:

1. use ChatGPT Atlas, let it explore the repo in agent mode
2. Do the fork using the acct you have connected to GitHub, then talk about it with ChatGPT

[native dew]

yeah same just for specific task use gpt 5.4

[lean lark]

Hmm, GitHub App already connected in ChatGPT:

[gentle harbor]

why do my images get auto deleted here lol

[torpid trout]

All of them?

[gentle harbor]

yes

[torpid trout]

I could imagine is is the format maybe?

[boreal holly]

Yeah, so fork the repo with that account and it will become visible to ChatGPT!

[boreal holly]

I haven't run into that before

[lean lark]

Wow, yeah, it seems it does have read access to everything. Gonna continue with this...

🤯

[frail meadow]

What percent of us have AI psychosis

[lean lark]

This is revolutionary for me. I think I ignored ChatGPT GitHub App cuz it couldn't update issues. But now it says it can create issues and write files!!

Like: "Create a ticket for Foo, I just found out it doesn't support Baz."

[boreal holly]

I was gonna say, with the pro acct ChatGPT gets 128k context, so it can really gather a lot of info from the repo. Like 10x more

[lean lark]

Yeah, this is huge. Yesterday we were talking with @nocturne folio about an Android app with voice for talking to/about projects. This is it!

I need to step away from the keyboard. This is exciting. I can't focus. Back in a bit.

[plush harbor]

I had a racoon come up on my site and idly commented on the codex directive. Here's chatgpt demonstrating very clearly why there's a codex directive

[plush harbor]

I spent like 6 hours yesterday doing a thing because gemini told me to and codex agreed so I'm gonna have to raise my hand here

[torpid trout]

what is happening here?

Raccoons?

[plush harbor]

its actually a real codex directive lol

[torpid trout]

I spent like 6 hours yesterday doing a thing because gemini told me to and codex agreed
Classic AI textbook "yoi could have told me that 3 days ago" lol

[torpid trout]

Yeah I know... but your screnshot, am not sure to follow what is happening there lol

[lean lark]

LOL ... Most of it comes back to Context. The responses you get are dependent (depending on the product/tooling) on your User Custom Instructions, Project Instructions, AGENTS.md files, Skills files, and the current thread discussion.

[plush harbor]

ikr? I've had this happen twice now and had to do a large batch backfill that if I'd known earlier

[plush harbor]

my website is a large children's dictionary I haven't touched for years, i've got a script lifting the old format entries to modern format. Racoon was the top of the review list this morning

[torpid trout]

Not in these casees, sorry

[boreal holly]

There's some real discussion about it on X right now. Apparently GPT 5.5 likes to say goblin randomly and nobody knows why

[torpid trout]

This happens in GPT without any altercation very often, way more often actually than in codex.

Now my text is red above.

What the hell does that mean? It is not not-sent, just red

[plush harbor]

emergent behaviour is funny sometimes, but also extremely annoying after the first three times

[torpid trout]

I think he meant the AI Psychosis drift? Or ... perhaps I lost the thread lol

lol what is happening here. Discord delay much or what.

[torpid trout]

yes yes I am aware of that issue 😄

[lean lark]

About red text in Discord. I don't know what triggers it but if you mouseover it gives the option to try again to send. Click and it almost certainly does.

[plush harbor]

what's funniest about all this is we all know exactly why that directive is in there

[torpid trout]

What I meant is, I think CSB meant the AI psychosis of being told "yes" when really it is not possible

[lean lark]

I'm not concerned about emergent words. For the first time, I'm concerned about emergent phrases or concepts that leak into discussions without telltale signals like that one awkward word.

[plush harbor]

rabbitholes are a problem too, I asked codex a "simple" (lol) question yesterday and ended up completely designing an entire new system. And this was before the 6 hour gemini triggered task. Damn I did a lot yesterday

[torpid trout]

BTW... I am always wary of those prompts

But it is I believe a fact that codex has it in the system prompt, in github

[lean lark]

I believe this does come back to the instruction hierarchy.

[plush harbor]

see I have a database of these, I could get codex to query my database for nouns starting with G

[plush harbor]

oh this is just typical AI rabbitholing, I've seen several people mention this. I got totally waylaid on designing what should have been a future task

[torpid trout]

https://github.com/openai/codex/blob/c5e2921e1dbd43fc1324d3daed206212b0bccd1c/codex-rs/models-manager/models.json#L55

It was so important it is in there twice
What a waste of tokens lol

[lost drum]

what did pigeons do 😂

[cedar skiff]

it used to be that we had to make a choice on what code we want, now we can implement any bad idea we want.

[lost drum]

or racoons

[boreal holly]

What's weird is I have completely custom system prompts and they never mention goblins...

[lost drum]

well well well

[boreal holly]

It's almost like putting that "never talk about goblins" thing in there makes them more likely to mention goblins lol

[cedar skiff]

never thought about goblins until you mentioned it

[lean lark]

Meh, the instruction system is a collection of instructions where We determine how the assistant tier about the LLM behaves. There's the LLM, then System prompts, and after that it's all us. So yeah, a lot of people complain about responses being crazy but we never ever ever get any feedback from these people about what their instructions are up the chain, or what their prompts were just before the weirdness that they cite. People talk about weird events as though they occur in isolation. No, Context governs everything except those truly weird and extremenly rare anomalies.

[small violet]

damn gemini 3.2 pro preview

98% on arc agi

[lost drum]

"You provide user updates frequently, every 30s" I might build my cusotm instructions too

[boreal holly]

It's gotta be some kind of drift detection thing. Like as soon as the agent mentions goblins or gremlins, they know context is poisoned and it's time to start fresh or something

[lean lark]

Too many people here never use AGENTS.md before complaining about cruddy responses - and too many people in the wild complain about "personalities in the machine" without any reference to their Custom Instructions, Project Instructions, or initial prompt directives.

☝️ A goblin told me to say that.

[lost drum]

he right

[boreal holly]

personalities in the goblin 🤔

[boreal holly]

You're welcome to browse my small collection of system prompts https://github.com/robertmsale/.codex/tree/main/roles

The designer one so far is my favorite.

[plush harbor]

this is just typical human behaviour. My chatgpt has decided I shouldn't go on rabbitholes and will stop me, but codex hasn't

[lean lark]

Hey, back on use of the GitHub App from ChatGPT, I'm blown away. This is like Codex but more direct:
"Create 'docs/foo.md' with info that we discussed about foo recently."

Create foo.md in GitHub repository? [Create File]
( Click )
Refresh docs folder in browser. foo.md is there.

Just
Blown
Away
🤯

My chatgpt has decided I shouldn't go on rabbitholes and will stop me, but codex hasn't
That is 100% your lack of instructions in ~/.codex/AGENTS.md. Or in Web, https://chatgpt.com/codex/cloud/settings/general Custom Instructions.

[plush harbor]

what, don't indulge conversations where you solve problems? That's going to be hard to instruct around

[lean lark]

huh? I'm sorry, please clarify.

[plush harbor]

I just asked it a simple question and ended up going down a long rabbithole where I ended up completely solving a raft of related issues because they all came down to the same thing. I don't know how to prompt around stopping that

[lean lark]

I see...

You certainly know how to create ChatGPT Custom Instructions. You know they drive all ChatGPT threads.

[plush harbor]

its just that this behaviour is kind of the entire purpose of codex, so how would you stop that without stopping it doing its core job

[cedar skiff]

This is a decsion point for you. You have to understand what you're doing and what your real goal is. You need to be the expert on what needs to be done and what doesnt.

[lost drum]

i think if I were to merge your system with mine (which would requrie real dev not me XD) then it would actually be the near agi level where it can builde anything

[plush harbor]

I asked if I should use solr or custom. It could have just said "custom" as one word and refused to engage further I guess

[lean lark]

AGENTS.md is the same in Codex. If you want or do not want a certain behavior. Be explicit. Tell it to do or not to do what you want:

• Do not expand the scope of tasks. Only address what is requested.
• If other issues are revealed while processing this task, document them in 'docs/new-issues.md'.

[plush harbor]

now i have a big constellation of related problems completely designed out and easy to solve as opposed to a vague pile of "rewrite this, do this later, refactor this later, this is terrible"

[lean lark]

Please forgive but I need to ask: Roughly how old are you, and how much development experience do you have?

[plush harbor]

30ish years of dev experience lol

[lean lark]

OK, sorry, I didn't know how I should address the challenge. You're a seasoned adult and I'll proceed with that.

[lost drum]

goblin

[plush harbor]

I'm not working for anyone else, its so good being able to get completely sidetracked solving problems but this would not work so well in the real world, where you bodge it, put a comment on it, and carry on like its not barely holding together

[astral basin]

pretty weird that the last codex-updates update is april 7. I wonder if there's a server where people are excitedly tracking each github release? lol

[lean lark]

There shouldn't be a constellation of related problems. There should be a list of categorized issues, prioritized, and addressed one at a time. There should never be a rabbit hole for the model to create or drill into. The hammer gets one nail at a time to bang on. That's on you, not the tool. Don't let the tool run wild or your project will get hurt with the bumps n bruises.

[plush harbor]

so yeah, "should i use solr or custom" -> new routing system -> new rendering system -> new caching system -> then new search system. Custom, with extra steps

this is a big constellation of related problems, vaguely filed under "fix the routing"

[boreal holly]

I mean my agents do pretty much build the whole thing! The only thing that requires my constant attention is the design. These agents do absolutely flawless backend work. I couldn't be more proud of these gremlins lol. The design, it's easier now with 5.5 and image 2.0 but still a manual effort.

[cedar skiff]

This is choice you have to make though, this isnt the models choice it's yours

[plush harbor]

chatgpt keeps telling me not to, but I wanna fix it

[lean lark]

not to what?

[plush harbor]

not to sidetrack on major refactoring tasks

[lean lark]

it's right

refactoring is a single task - and shouldn't be done on the entire project at once.

f.o.c.u.s....

[plush harbor]

yeah well the current "focus" task is a 3 month long tedious thing

[lean lark]

No, there is is a collection of individual tasks that need to be done. Do them one at a time with your human intelligence to guide it.

[cedar skiff]

You want it not to give opinions on what needs doing?

just tell it at the point i dont need your opinion just help me

[plush harbor]

no the current task is one task and its not coding. I would prefer to be coding. This is just personal preference speaking here

[boreal holly]

Oh yeah, don't listen to ChatGPT. It's a pessimist about time complexity.

[lean lark]

I think I get you, but these concepts are too nebulous for anyone to really help.
Break down your project into features, and features into specific components. Use the tool to do what you want it to do to support specific challenges. Don't give it a huge task that requires decisions. Don't give it vague instructions that allow it to go on grandiose diversions.
Get the project clear and modular in your head and then use the tool to make that happen. You're holding the hammer.

[plush harbor]

its probably a day of work, its just a total rewrite of my front end

[boreal holly]

What really helped me in the beginning is giving the frontend a reusable design system with abstract interfaces. That way visual refactoring is totally separate from controller refactoring. Codex liked to reimplement each and every atom and molecule of the UI, but if it's a design system you can propagate the changes quickly. That'd be the first high impact refactoring to make the future ones less painful

[plush harbor]

my codebase is like 15 years old and was written in an afternoon, its a hideous collection of if statements and its hard to do anything to. I've got part of the site with clean layout/css but its hard to pull across anywhere else due to how much of a mess it is. Its driving me bonkers, I want to fix it lol

[lean lark]

Bouncing between topics, on the ChatGPT GitHub App : For anyone concerned about token usage, you can get a Lot done and use zero tokens just by discussing small things with ChatGPT and telling it to commit changes. The way I work with this stuff is changing yet again, from Codex Web to VSCode Codex Extension to CLI to ChatGPT. One of the huge differences now is that 5.4/5.5 now have the same competence with coding as Codex. Previously we needed to discuss with ChatGPT but implement with Codex because of the difference in competence.

[boreal holly]

You got this 💪

[plush harbor]

positivly embarrassing showing it to other people right now. "this bit looks great, but don't click on that other bit. And I'm starting to bump 5000 pageviews a week now so i dun wanna scare visitors

[lean lark]

Going away. I hope things work out @plush harbor - hope to see you in here regularly!

[stable obsidian]

quick question do subagents in codex spawn with fresh context, or is the context copied

I guess its fresh context

[cedar skiff]

They can do both, the tool has an input value of fork_context: true/false

[drowsy turtle]

i made codex connect to discord bot to be controlled like claude discord plugin using python private on your device only

[stuck warren]

I'm trying to use Codex with GPT 5.1 -- is this still supported? I keep running into issues where it works for a few prompts, and then sits "thinking" for about 5 minutes per prompt.

It feels like there's something hanging up in the API between 5.1 and Codex. I did the otel logs and it seems like it sits there polling.

[boreal holly]

You using ChatGPT subscription or API key?

[stuck warren]

worse...Azure

the model availability is absolutely awful

[boreal holly]

What host OS?

[stuck warren]

it happens in both Windows Codex GUI and the Ubuntu WSL CLI

[boreal holly]

Unplug all USBs and restart computer. I had no idea the codex desktop app let you use different model providers! Must be a new thing

[stuck warren]

you just have to configure it in the config.toml

[boreal holly]

Ah so you change the default model provider! Fair enough

[stuck warren]

the azure openai support feels like an afterthought and works really poorly

[boreal holly]

After you reboot, there might be some features worth trying:

[features]
responses_websockets = true
responses_websockets_v2 = true

If Azure supports it. Would make requests follow an alternate and potentially more stable code path

[stuck warren]

thanks I can give it a shot

would downgrading potentially help?

it's not clear what models are supported, and if gpt 5.1 support was removed or if this was always a problem

[boreal holly]

Well, you were saying otel logs showed polling. With websockets enabled there would be no HTTP SSE involved, it would keep the connection open over ws and you would see if there is a drop, which could provide better diagnostics

[stuck warren]

thanks I'll give it a shot

its frustrating because its not clear how to even get support for this kind of thing

[glacial shadow]

Is there a reason to choose 5.1 as opposed to 5.4 on a lower reasoning

[stuck warren]

yeah, there's no models available other than 4.1 anad 5.1

[glacial shadow]

In Microsoft Cloud, there’s only 5.1?

[stuck warren]

in some regions, yeah

[glacial shadow]

Interesting

[stuck warren]

it's awful

[glacial shadow]

What about switching to Amazon bedrock

[stuck warren]

the only real options in there were claude

[boreal holly]

Perhaps VPN?

[glacial shadow]

Wasn’t there an announcement a few days ago that 5.4 is now in bedrock

[stuck warren]

its not that, its a compliance issue

[stuck warren]

I think there was the announcement of the intention to add it

but its not there yet afaik, and then we'll run into the region issue

I'm also not sure if you can even hook Codex up to Bedrock?

[glacial shadow]

OK, what region and what compliance standards must be obtained

[plush harbor]

not really codex, but does this thing in the api backend ever move off zero?

[stuck warren]

fedramp

[torpid trout]

thats the api counter, not codex usage, yes?
That is a bit delayed

BUt yes it does move off zero.

[plush harbor]

how delayed is it? Like a week?

[torpid trout]

No, in my experience more like a day

[glacial shadow]

I would start asking 5.4 on highest reasoning about solving this

[stuck warren]

I asked chatgpt for a while actually, but made no progress