codex-discussions

[torpid trout]

No, and it can’t be part of a comparison

If that’s how they want the model to act that’s how it acts

[boreal holly]

5.4 is more of a blank slate model. OOB it is not that great but if you add skills to it you basically build it up into the SWE you want to work with

[kind jay]

That's crazy to be honest, "let's but it in a box and only compare a small part of how it will be used"

[torpid trout]

Why is that crazy

[kind jay]

Too abstract

[torpid trout]

It’s how it’s used in the end on a domain specific task

[west orbit]

in my first session and first implementation with 5.4 i caught codex deciding mid implementation to not use the library as it was planned in the impklementation plan and instead write what the library does from scratch but super basic .. i never saw this kind of behavior before in a codex/GPT model

[torpid trout]

If model a produces bug free code tjat I don’t like, it’s still better

[west orbit]

i know this from back when i used claude -_-

[torpid trout]

And that’s what 5.4 does so far

[torpid trout]

Sounds exactly like sometjing 5.4 would do yes

Cut corners is what I call that

[kind jay]

I understand where you're coming from, however it's not how it is being used in general, additionally we care for logical, maintainable code

[torpid trout]

Of course

[kind jay]

The issue is no longer can these models code, it's if the code can be maintained

[nocturne folio]

i think we might have to lower our expectations for agi

[torpid trout]

That’s all considered
And polished code isn’t maintainable often enough

As an example 5.3 went all OOP for simple things. 5.4 kept it procedural

[kind jay]

AGI is not happening lil bro

[nocturne folio]

hence lower our expectations

[boreal holly]

Think of it like this. You just installed Windows and Visual Studio and create a new C++ project. OOB the experience is terrible. Compilations take unfathomable amount of time, debugger doesn't show autos/locals, running build does not produce a binary, etc. Terrible, absolutely terrible experience out of the box.

But after you configure Windows Anti-Virus to stop quarantining your builds, set up PCH, spend hours configuring the UI and the project, restart your computer a few times, it works a lot better.

The difference is 5.4 eventually becomes the GOAT once you get it fully set up.

[nocturne folio]

but ofc conciousness n stuff like that is not happening ever

[frosty zealot]

Jokes on you I’ve created a registry edit that stack overflow suggested and my comp won’t boot

[kind jay]

I think the question is can someone who is quite dull, produce the program that meet the requirements. I think this is the product at the end of the day

[frosty zealot]

Hit me with a reset baby

slaps veins

[kind jay]

My experience is 5.3 codex is more suited to writing code, while 5.4 is better at like putting the ideas into code

[kind jay]

Just buy an extra account

[frosty zealot]

Blasphemy

I’ll just use Dopus if I hit my limits

[kind jay]

Which as greater limits, pro or 10 plus accounts?

istg

[frosty zealot]

🫡

https://youtu.be/EJqFjvKF8HU?list=RDEJqFjvKF8HU these mixes are so good but theres like a million of htem and they all have an AI Jenna Ortega djing lmao, good vibe coding music though 😄

[cyan wing]

living in the ~future~

[drifting granite]

Is there a way to have automations on the codex app have network access? It seems to be stricly in Read mode.

[dense verge]

Anyone else experiencing this atm?

[torpid trout]

That’s true.
But, if I need to spend hours to setup a workspace abd then windows comes abd does an update and I need to spend more hours to fix what worked before…

[dense verge]

[torpid trout]

In the end we want to work, not “fix” the tool (or undo what used to work because now it won’t)

(I know a lot of folks me included like to thinker. But clients ain’t paying for that end of day)

[frosty zealot]

We're just not there yet though, there's 'well I shouldnt have to' versus 'well I can just do it this way for now to get a better result'

[torpid trout]

What’s happening here

Some messages appear/disappear lol

[frosty zealot]

bots getting body slammed

[distant crypt]

When does codex reset for tokens? monthly or?

[frosty zealot]

Whats your usage page say

[distant crypt]

this is what I see

[frosty zealot]

Click settings > rate limits remaining

[distant crypt]

ah i see

[kind jay]

Why is it not reset yet?

[boreal holly]

reset wen

[kind jay]

These resets have set my project 3 weeks ahead

[frosty zealot]

They reset again!

[kind jay]

No way

[kind jay]

You lied to me, how could you?

[frosty zealot]

Establishing friendship rules

[boreal holly]

I'm still at 96% so to me it feels reset rn

[frosty zealot]

Dude its like a battery

100-90 feels like an eternity

90-0 feels like an hour

[boreal holly]

lol and takes 1 week to charge 😭 except when they goof or something

[kind jay]

Brave of you to presume we have a friendship

[frosty zealot]

That was also a test

trying to reverse engineer padding in a packet so you get a proper data stream makes me wanna

[white furnace]

so uh... I closed codex app and now its stuck doing its thing but I cannot stop it either because stop does nothing

[kind jay]

I’m not a CS nerd like you so I don’t get what you’re playing at

[white furnace]

what do I do

[frosty zealot]

Im actually a mechanic, this is just my hobby 😉

[kind jay]

Unplug your PC

[white furnace]

I restarted already but its still stuck

[kind jay]

That might be somehow worse 🙄

[kind jay]

Use Linux, idk

[frosty zealot]

Its become sentient

it IS your machine now

[white furnace]

in that case its thinking really dumb thoughts because it does nothing at all

[frosty zealot]

What OS?

[white furnace]

Win11

[frosty zealot]

GL

[stuck berry]

achive the chat

lol spelling for the win

[runic quarry]

is chatgpt website loading very slow only for me ?

[kind jay]

They’re resetting the limits rn

[tall zodiac]

Anyone got tips to steer GPT models correctly without it losing track of what it was doing before and also surviving auto compaction when it has a todo list to follow?

[kind jay]

Nah, plan better

[blissful basin]

keep checklists in md for longer tasks and tell llm to tick what was done etc

[boreal holly]

Sometimes I'm really blown away at how amazing these models are at refining ideas. 5.4 is very organized, and I've definitely learned a lot by just talking to it.

[tall zodiac]

What about steering mid convo without it having dementia? @kind jay @blissful basin

[boreal holly]

Feed the checklist back into the conversation as a user message.

[kind jay]

Why do you talk to it like a person?

[blissful basin]

I really didnt notice that happening, and working multiple hours daily with codex

[kind jay]

I tried, I think it doesn’t work

[steady vigil]

5.4 just amazed me right now also. but it sucks for implementing well defined plans. its a great planner.

[blissful basin]

But like i said, every bigger piece of work i do have detailes checklist that llm follows

[kind jay]

Skill issue

[steady vigil]

I use 5.3 codex to impl and 5.4 for planning

[boreal holly]

Adds more entropy

[cedar skiff]

It's an llm, we talk like we talk and it is a machine that is very good at talking o.0

[kind jay]

“Switching models during conversions may degrade performance”

[frosty zealot]

Ya'll got any more of those resets

[kind jay]

I got mine

[frosty zealot]

Yeah ok buster

[boreal holly]

Exacerbated by switching with a compaction in the chat history

[kind jay]

Nah

[tall zodiac]

I realize how bad I am at explaining when I can’t explain to GPT what I want to do so I draw out a picture LOL

[steady vigil]

who said during the same session rofl

[cedar skiff]

ok i see now, it was mistake for talking to you.

[frosty zealot]

Eric doesn't play

[kind jay]

Eric the realest person here

[cyan wing]

Anyone else seeing stream disconnected and Falling back from WebSockets to HTTPS transport. stream disconnected before completion

[kind jay]

#status 😔

[meager wadi]

:c

[still trellis]

yes completely un-usable at the moment

[cyan wing]

definitely gunna need another Usage Reset after this 😩

[nocturne folio]

please sam altman be nice this evening and make the 2x usage permanent

make it 4x even

[plucky halo]

This is like Ark: Survival Evolved all over again. I thought I'd escaped that 😂

[still trellis]

i switched to spark, my flawed thinking is that since is way faster it will complete before stream disconnect.

[elfin summit]

anyone else having trouble or bugs in codex

[thorny bloom]

does anybody know why my gpt 5.3 codex spark usage is going down when I'm not using that model?

is that what the new security thing uses?

[boreal holly]

That's more like galaxy brain thinking, because only pro users have access and they run exclusively on cerebras

[still trellis]

LOL

[hollow reef]

So are we gonna likeee. get an extra few days added to our plans or wut

[torpid trout]

GPT (reasoning pro) just told me that if I set model in a agent, then this is buggy because all subagents inheriting from it, get that model too
That, is true - but its also nonsense, because such agent profile is usually already a subagent.
What am I missing?

[boreal holly]

If you do agents.<name>.config_file I'm pretty sure aside from sandboxing you can edit just about everything else including model

[torpid trout]

yes, you can - but what it said is taht if I do that, then other agents get affected by it (spawned by this agent)
Which is true, if it where true lol (bu it is not, becaus an agent cannot spawn another subagent)
Or .. am I wrong? Can agents respawn other agents like so:

main thread (max model)
- explorer (minimal model)
-- engineer (should use max model from main thread but now would not since inherit from explorer)

I do not believe that is possible, but it said it is (well, now I am questioning it, it takes time to wait for answer on research pro, guess in 30m I will know more)

[boreal holly]

I think there is a recursion depth limit. Not sure how it's configured, but yeah subagents spawning other subagents is tricky behavior.

I personally would not trust what the agent says. Unless you clone the openai/codex repo, checkout the version tag you have installed, and scrub the codebase for the precise implementation details, anything they say is speculation.

[nocturne folio]

thank god for codex for shaving my deploy times from 10 minutes to 5 minutes

literally cut it in half

[mild kestrel]

Literally like magic isn't it 🪄

[tall zodiac]

With codex do u guys have it run in a mono repo setup or in a per project basis ?

[boreal holly]

For the good of all robot-kind 🫡

If all goes well I will OSS the whole thing

[tall zodiac]

@boreal holly how come xhigh is an option if people say performance degrades ?

[kind jay]

🤨

[boreal holly]

For research purposes. AI is quite literally computer science, standing on the edge of what is known and unknown in our universe. OpenAI lets you experiment with it

[tall zodiac]

Makes sense

🐴 I’m a comp sci major

@boreal holly

[mild kestrel]

So my website has been using a custom HTML/CSS build codex helped set up. It's been working but could be much better overall. Is it better to have it switched to something like Tailwind or something else? It's gonna be one of my next main things to tackle, I haven't used them before so I'm not sure the best path to go with it.

[frosty zealot]

Depends

I refuse to elaborate

[kind jay]

For me, it depends

[thorny bloom]

it depends for me as well

[kind jay]

Joke’s already old 🙄

[thorny bloom]

🦍

[frosty zealot]

@kind jay that better?

[kind jay]

🐶

[tall zodiac]

6 7

[boreal holly]

5.4, when you upgrade it from "glorified roomba" to "highly regarded peer" enables autonomous behaviors

[frosty zealot]

more... I NEED MORE

https://tenor.com/view/omg-hot-as-hell-drooling-gif-22722557

lol the link versus what the actual gif is

[kind jay]

You didn’t get your credit reset? Weird

[frosty zealot]

https://tenor.com/view/mewing-bird-gif-16835999897918525830

[kind jay]

Too late, I already envisioned myself as the stoic monkey and you ask the enraged feline

[frosty zealot]

lmao

[boreal holly]

Just sayin, tool call outputs are the biggest win! Command parser is the way!

[kind jay]

https://tenor.com/view/bigbrainbarnes-gif-6132461259902152472

[frosty zealot]

Give me your credits before I stuff you in a locker

jk :>

Where's this command parser you're refering to

[kind jay]

MCP

[frosty zealot]

Now you're spelling at me

[kind jay]

Model Context Protocol engineer man

[frosty zealot]

Blasphemy

You sit on a throne of lies you had me googlin MPS

[kind jay]

I’m not a computer person

[frosty zealot]

https://tenor.com/view/i-dont-believe-you-whatever-lies-gif-3602502

[kind jay]

Were you the first Discord user?

[frosty zealot]

Ok that made me laugh out loud

[kind jay]

I like this movie

Also “you sit on the throne of lies” would be an acceptable response

[boreal holly]

https://github.com/robertmsale/.codex/tree/main/skills/command-parser

It requires docker now

[kind jay]

Nvm

[kind jay]

Are you a computer person?

[boreal holly]

I am a meat popsicle 🤖

[kind jay]

Yummy 🤤

[timber cosmos]

anyone seen this one?

[kind jay]

You have VPN on?

[timber cosmos]

nope. (I am in Canada)

[kind jay]

Weird

And it was just in chat with nothing weird added?

[timber cosmos]

not sure how I am supposed to "make my request to ca.api.openai.com" =/

[timber cosmos]

yeah it happens immediately when I try any chat with GPT4.5

other models work fine

I mean 5.4 🙂

[kind jay]

#status

[kind jay]

Not sure, sorry. Maybe try turn on VPN to US

[timber cosmos]

Yeah I'll give that a shot, thanks

[frosty zealot]

DNS

[kind jay]

Now you’re spelling at him

[frosty zealot]

I've been know to spell from time to time

The irony in that typo

[timber cosmos]

Okay I figured it out, but I'd say it is a bug for OpenAI to be aware of. My API key was associated with a Canada-based project. I created a new key associated with a project with Global region setting and set it in Codex and it works fine now.

[kind jay]

Whole thing is probably vibe coded, they have numerous other problems too

[cyan wing]

which tiny model do u use for command parser?

[boreal holly]

5.1 Codex Mini seems to work really well. If you got Spark credits, then that's the best one because it eats that rate limit right up. Otherwise, I would use the 5.1 Codex Mini.

[worn harness]

The small models are very good nowdays for non-critical stuff, can always let the heavy tasks be routed to bigger ones.

[meager wadi]

haven't seen an error like this before

[cedar skiff]

another another guy who does a similar distilling of context and he swears by it as well. Do you notice any semantic diffusion in any cases?

[frosty zealot]

@tardy elm Do you have a link to this command parser?

[solemn sedge]

Thanks to codex.I can learn of which side I'm on during champions selecting period in LoL.

It’s helpful especially in ARAM

[boreal holly]

this town aint big enuff fer the 2 of us

[frosty zealot]

lmao

[torpid trout]

There's a few paid platoforms that promise "amazing etc" with exactly the same principle, to use MCP for all those noisy tasks.
In reddit I saw a guy he had a good idea too, he feeds the noise to local llm, and that back to codex

This is def the way to go
Makes no sense to send hundreds, thousands of lines to llm remote
Even said llm's chatter is too much. It should talk json, not human.

That is what I was trying to do after I saw Robert's first images of his app, but I totally botched that lol, it ended up being a replacement for codex (which did not work)

[worn harness]

openAI is turning me into a gambler: Do I gamble all my weekly credits that openAI will reset the usage again, or just play it safe and accept that I lose some
harder question than architecture

[cedar skiff]

The llm is trained on tokens that are derived from text though.

[rugged burrow]

What do people use to use codex remotely? Slopus Happy coder? Tailscale and termius? Something else?

[torpid trout]

“The scan returned no errors, I’m confident the code is passing the gates” vs “no errors”…
I don’t think that makes it worse
That’s just a silly example

[rugged burrow]

Want to use it from my phone

[simple star]

@rugged burrow Codex Cloud?

[torpid trout]

Shellfish

And tunnel into the home comp over WG if you need

I’m not using it though when not on lan

[simple star]

What are you fellas building?

[rugged burrow]

Nah I want to run locally with my skills and state

[worn harness]

A duck.

[simple star]

Quack!

I mean, wow

[white furnace]

https://tenor.com/view/duck-ducks-confused-surprised-what-gif-1845535727828628932

[cedar skiff]

Right but thats not json

[frosty zealot]

https://tenor.com/view/jack-nicholson-jack-nicholson-yes-yes-nod-head-nod-gif-17444176260167585294

https://x.com/googledevs/status/2031411032845885725

You guys see this one

[verbal mica]

Perplexity

Need perplexity one month reffral link please 🙏🙏🙏

[frosty zealot]

I'm calling the police.

[high girder]

turn off hardware accel

[frosty zealot]

For codex?

[high girder]

For discord lol. Codex is fine

[frosty zealot]

look above what I have selected

[high girder]

I know. Do you have agents running? I think each one is something like 4Gb depending on the job

[frosty zealot]

It just happens after running for long sessions

[high girder]

something isn't ending properly then. See if codex can find out why your process is hanging. Actually, use claude for that

[frosty zealot]

its a known problem afaik probably a memory leak, ill debug it the next time it happens

normally wouldnt be an issue but ive had a brutal refactor going for like 2 days and i just let it run till it poops the bed

[high girder]

I have still yet to have a job run for that long.

[void pelican]

Working on the native linux client some more fellas

[torpid trout]

You call the police often, Cantcrit

[void pelican]

weee wooo weee woooo

[bold root]

"Do you trust the contents of this directory? Working with untrusted contents comes with a higher risk of prompt injection."

Is there an option to prevent this from appearing and allow everything?

[boreal holly]

Soon my brother, memory leaks will be a thing of the past

[frosty zealot]

Do you find that style of desktop productive? I havent gave it an honest try

[boreal holly]

stage manager is very nice when remote. for local its ok

[cyan wing]

can u add multi-panel agent support pls ? 🙂
i often use codex in a 2x2 tmux

[boreal holly]

there is no agent 👀

if you talk to anyone but the orchestrator it sets a precedence that the workers can talk to the user and it all falls apart. can only talk to one, and have the one talk to the others

[cyan wing]

oh damn that would be even better 👀 👀

[frosty zealot]

Man oh man the compactor is not compacting

[tall zodiac]

Is AI slop fully avoidable by testing your code?

Ive been pondering about this

If you can verify correctness by robust testing

There’s no way to slopify code right ?

[frosty zealot]

Depends on your definition of slop

[tall zodiac]

What’s your definition ?

[frosty zealot]

I dont really have one

[tall zodiac]

I thought slop was just wrong code or work

[frosty zealot]

with the current models

I think with an adequate review etc. good prompting and well defined sub agents and skills I think you can pull off good coding and practices for sure, humans make mistakes too, not like security etc. has only become an issue in the last year, people still find their way in no matter what. But if you're diligent enough I think AI written code is probably just fine

[tall zodiac]

Yeah I was just thinking basically that we can prevent slop by testing and determining valid vs non valid outputs etc

Slop in my definition is having the model code without verification of right or wrong

Just like autocomplete

Or like a junior writing a function or syntax and he doesn’t check his work thoroughly

[frosty zealot]

I think a lot of 'ai slop' comes from people that also don't really know how to architect, have good intertwined systems, etc. they have zero code experience and just keep prompting till they get their result @tall zodiac

I think people that have a decent understanding of the code base they're working with and other systems like db's, cloud services and what not will be able to build much better and safer code with AI, than the 13 year old who just downloaded claude and says BuILD ME A DISCORD BOT

[cedar skiff]

Testing is usually for making sure you don't have broken code in terms of it does what it's supposed to. AI soup or when it's done by a person spaghetti code is more about the architecture and control flow structures that the app is built around. Well defined policies and architectural abstractions that are followed and adhered to are what prevent code soup. Those architectural systems enforce flows and structures that enable scaling, factoring, bug mitigation and code maintainability.
Spaghetti code doesn't have those kind of structures and is susceptible to subtle hard to diagnose bugs that are easy to introduce and code that is hard to maintain.
AI agents don't tend to write code like this or typically follow the patterns in a code base without guidance.

That guidance can be provided as Cantcrit mentions with context harnessing techniques.

[frosty zealot]

Another thing is their memory cut off, so they might start a project with already outdated, or insecure frameworks etc., its always a good idea when kicking off to use context7 or even just prompt to research the most up to date specs, best practices, and most up to date, and if I say like 'find a framework' I always say to make sure its well supported, well documented, and has active git repos, just general stuff like that

[tall zodiac]

Which goes back to a person being like “make me a complex app make no mistakes” without testing or anything

[frosty zealot]

so the first result isnt like joe blows ultra uber rest api that has 1 commit from 4 years ago and is actually a crypto scam

Before every prompt, especially in the early stages you really have to look holistically and think, what can go wrong, and really steer the model to starting off on the right foot

I think thats the biggest prevention in AI slop

[tall zodiac]

Makes sense

I think steering is part of also being like the following in your prompt;

[task]

[acceptance criteria / testing]

[frosty zealot]

Its as simple as like:

Write 80% TDD coverage

Versus

Write meaningful unit tests using TDD.
Focus on business logic, edge cases, and error paths.
Target ~80%+ coverage but prioritize test quality.

[tall zodiac]

Exactly always being detailed and having a definition of done that determines in a deterministic way how the model can verify it’s right

[frosty zealot]

Sometimes when im not sure how to really prompt the model, Ill just ask it tbh, like 'This is what im trying to ask for, how can I ensure Codex interprets this in a meaningful way' kinda thing

https://www.anthropic.com/learn @tall zodiac though aimed for Anthropic, theres a lot of good information to read through in here

not sure if openai has a similar thing

[tall zodiac]

Ty ty

[ivory zodiac]

https://x.com/LLMJunky/status/2031582374064951414?s=20

[frosty zealot]

ping me when theres a UserRequest or something hook

@ivory zodiac Do the hooks also kick in for the app?

[ivory zodiac]

128gb lmao

[frosty zealot]

yeah lol

[ivory zodiac]

not sure i'd have to check

[tall zodiac]

@ivory zodiac do we specify hooks via global configs or can they be configured per repo?

[ivory zodiac]

either

[tall zodiac]

Damn more harness engineering I guess

[frosty zealot]

I've been using github actions for my codex hooks in the meantime

I feel like i need to be a jehovas witness and go into the claude discord and spread the word of codex

here ye here ye

[cedar skiff]

claude discord is like fan boy city, there is no way to say anything about codex with being flamed

[ivory zodiac]

they're blind

[frosty zealot]

Hit the button... HIT THE BUTTON!

[cedar skiff]

Even though no one tried it they still know better

[frosty zealot]

The exact vibe I get too

I enjoy the banter

But when the Silver fella joins I usually leave lmao

[ivory zodiac]

slopus is good for 2 things

frontend and writing

[frosty zealot]

I think CC is def more mainstream and it on a surface level, gets the job done, and most people dont look at the code etc. so its like 'How can it get better than this?' and theres so much media especially lately thats like A\ GOOD OAI BAD that people take that at face value and sluff it off

[ivory zodiac]

its a good model but i have zero interest in using it for most things

[boreal holly]

sippin on rate limts

[frosty zealot]

You're not pushin hard enough!

[boreal holly]

12 agents deep, if i push any harder i lose track

[tall zodiac]

I stopped using spark bc if I use it for exploring the codebase the limits just vanish lol

[boreal holly]

spark just looks at commands outputs and extracts useful info

-70% input tokens for the ones that do the work 🤘

[frosty zealot]

I think im shooting myself in the foot by not starting a new thread at this point, I have been in the same thread for over over a day

❯ git diff --shortstat main...HEAD
 378 files changed, 56848 insertions(+), 7125 deletions(-)```

Im scared

to see what the end result is gonna be

[boreal holly]

you gotta commit 😭 the silicon is tortured glass at this rate!

[frosty zealot]

I am comitting lmao

[ivory zodiac]

felonies

[frosty zealot]

Take that

[boreal holly]

its ok, if you properly spell finis-hturn you get automodded

finisht-urn

[frosty zealot]

I wish they'd relax a bit and let us use gifs at least

[boreal holly]

its wack

[cedar skiff]

claude matches the user verbal level and eq making it more friendly at any level. That plus the way it always ego boosts is part of why i think ppl get brain washed by it.

People thinking they solved some age old theorem or what ever are a dime a dozen over there

[cyan wing]

you may be on to something here 👀
hopefully the Codex team sees what you're up to

[frosty zealot]

Anybody implement OpenClaw into their codex workflow?

[high girder]

The most use out of OpenClaw I've gotten was ripping parts out like a shopping list from the repo

[frosty zealot]

The lag right now is crazy

[wide schooner]

Claude glazes dummies

Its 4o all over again

[deft sable]

https://x.com/i/status/2031605592352313567

[wide schooner]

Claude is behind qwen

[wide schooner]

WHAT

😭 😭

Again???

[high girder]

Anyone ever get an update on the superbowl easter egg swag?

[cobalt junco]

yay another limit reset

lowk might switch to pro

im beginning to trust xhigh 5.4 a lot with my project, saves so much time its insane.

something I do in a day with codex would’ve taken a week to a month pre AI accelerated coding

[tame magnet]

just get a business plan and make like 3 accounts

will cost like 90$ per month and you'll have enough usage

[neat sinew]

Is this gonna get fixed? It's been like this for a week

[meager wadi]

Hell yeah

[simple star]

Another reset?!?!?

Heck, screw this, Im going /fast all day long 😄

[still trellis]

Hey people of codex, serious question. How do you get codex to think more general around front-end? I find it very hard to get it to not be so literal and disregard what i would call common sense in terms of user experience.
Example: I ask it to build/ fix a feature describe what its supposed to do with a plan also ask it to use the playwright QA skill to do tests ensuring it works. Then most often than not after it runs for 20 - 30mins easily it is confident it did what i asked. I go test and notice a basic thing like a div block sits over another element or a state change shifts the layout. I cannot seem to get it to not be so litteral ? Skill issue?

[worn harness]

it sucks at UX, just spray and pray or be super concise

[simple star]

@still trellis It is simply not well trained for frontend

Simple as that

[still trellis]

OMG why cant a swear here... LOL

[simple star]

Here's the good boys club 😉

[oblique cove]

With the codex weekly limits resetting daily, I'm struggling to use up the whole weekly limit in one day, the five hour windows are the main problem. But I can use gpt-5.4-xhigh and I really never run out, I'm just on the plus plan and I've been launching commits constantly

[oblique cove]

I've been having fun with uncodexify, it's a simple md file of hard no's for UI / GTK. Pretty good results, given I have the least knowledge on UI

[west orbit]

i find it's ok with clear and explicit instructions wrt frontend. i use gpt-5.2 high and have a svelte frontend.. i needed to add specific instructions in my AGENTS.md to make codex not mess things up, but since i did that it's doing pretty good

[oblique cove]

I would say that gpt models are historically the worst with web design

[still trellis]

Yeah they got a certain taste about them

[west orbit]

i mean i'd rather deal with gpt's frontend quirks than claudes style

what i like about GPT models is that they follows existing patterns automatically

[oblique cove]

That is very nice

[west orbit]

i switched back to 5.2 high today after constantly finding serious bugs and logic issues after implementations with 5.4 in the last few days .. god how much more pleasant 5.2 is in general .. and reliable with regards to following a plan .. implementations are solid again

[bold root]

* [Project Language Reference](${PROJECT_PATH}/AGENTS.project.md)
* [Development Style Guide](${CODEX_PATH}/DEVELOPMENT_GUIDELINES.md)

Is there a variable value that can refer to the project path and codex root path?

[turbid axle]

oh boy... I just created an app that allows codex app, to use its own codex app

[hearty prawn]

what does 2 week stand there for_

[turbid axle]

not strong enough

I assume just a 2 week budget, instead of 1 week?

[cyan wing]

no shotttt

they reset it AGAIN?!

[white furnace]

Another reset? I just woke up and again 100%

[cyan wing]

😍

[simple star]

Yeah, we have to put that /fast on 😄

[sharp gazelle]

my business plan is still the same 60% i left yesterday

is this not for business plans? we were also affected by the reconnection loop

[white furnace]

im on plus so probs not for business plan

[sharp gazelle]

rip

[sharp gazelle]

wasted 40% weekly rate limit yesterday

hope it‘s just staggered and takes a while to roll out

[hollow linden]

anyone notice that in later codex versions, compacting seems to have changed? I find that it seems to be storing a lot more information in context after compact. This is of course good in that it retains more context, but it also means it's compacting less. After about 5-6 compaction cycles, even after the compaction, I only have about 35% context left (e.g. it goes from 5% left to 35% left). This means I'll probably need to compact again after just one message.

[chrome raven]

damn how many compaction?

how many token used?

[west orbit]

i would consider always doing a manual knowledge transfer (e.g. handoff.md and phased plan.md) for a fresh session and only let auto compaction happen when really necessary

[hearty prawn]

im on pro but idk i get reset like legit every day rn lol

[tall zodiac]

How serious do u guys take context engineering in your repos and do you have any tips?

[turbid axle]

I think the most effective thing here is to make the agent self aware, and give it access to docs it can write and manage for its future self

[west orbit]

yeah documentation is key and if you plan properly you don't run into context issues

[turbid axle]

I have a 'housekeeping' prompt which essentially tells every agents. hey dude, you have limited context. help future self. agents.md is your boot up memory, use it as a pathway to inform you future self of everything they need to know when they wake up.

[tall zodiac]

@turbid axle @west orbit when an agent is working off a markdown file plan is it usually good at staying on track during auto compaction or do you have ways to always make it known to continue the plan ?

[west orbit]

i have it write handoff.md before starting a fresh session

[elfin summit]

what is this new '2 week'

it used to be weekly not per 2 weeks

[tall zodiac]

How do you detect when a new session is happening ?

[west orbit]

and if we deviated from the plan make the plan reflect the implementation

i dont detect it, i end the session andd start a new one:)

i only go through compaction if i now i can trust it will be fine for the work i have to do

[turbid axle]

it pretty good at keeping on track, but its really in making it self aware of its limits, and then asking it how to best help it stay on track. it can do what it needs to do to make itself more effective from here

[plucky halo]

Maybe it's just some fantastic corporate maths:

"Let's give everyone twice the amount of tokens... but so we're not overwhelmed, make them last for 2 weeks instead of 1. Yeah, that maths checks out"

Soon they'll announce 4.3x allowance for everyone but it has to last a month 😛

[west marsh]

how do you guys use the superpowers skills? I feel like it always runs off the brainstorming skill automatically for most requests. any best practice here?

[west orbit]

what is superpowers skill?

[west marsh]

basically its a framework that includes multiple different skills: https://github.com/obra/superpowers

Read some good things about it and gave it a shot

[cedar skiff]

It was made for claude a long while ago and i dont think its needed for codex now, in fact how forceful it is i think it might actually get in the way.

[west orbit]

i always prefer a very simple setup, no MCP's, lean AGENTS.md and that works very well for me

[cedar skiff]

plan mode in both cc and codex does better than branstorming imo

[turbid axle]

agents.md should be as lean as possible yeh, it should act as a portal to other docs the agent can look at when it needs to. no a giant list of instructions

[west marsh]

Its also way way way faster to just go with plan mode in codex instead of brainstorm skill, I've realized

[pale ruin]

you talking about the codex repo?

[turbid axle]

the context is effectively its working memory, if you dump it full of instructions it will confuse the agents a lot. I think of it more as a kind of conciousness, like the little angel and demon on your shoulders. if they keep shouting things as you you just get hella confused. it should more so induce a general feeling/vibe for the agents, a 'soft' direction if you will

[turbid axle]

any agents.md

[pale ruin]

https://github.com/openai/codex/blob/main/AGENTS.md

[tall zodiac]

@turbid axle Do you usually tell it to use skills explicitly or does it automatically know when to use skills?

[turbid axle]

this is imo way too large and complicated already

I dont use skills

[tall zodiac]

how do u ensure plans it makes etc are consistently formatted each time?

[turbid axle]

I use planning mode for plans

[tall zodiac]

i see,

I usually just talk to it and have it store a plan in a directory etc

[turbid axle]

it does that automatically for me because it knows to make sure its future self is aware too

it will write md docs with the plans, what is doing, what it just did, etc

[tall zodiac]

@turbid axle im thinking of going back to a vanilla setup prob i have like 200+ skills

[turbid axle]

and I date them, with some local ci which fails old docs so it is forced to review them

[west orbit]

what do you mean with formatted ?

[tall zodiac]

like lets say you want a plan formatted a certain way each time

and not written up in different ways

[west orbit]

you mean structured a certain way?

[tall zodiac]

ye

[west orbit]

for me the important part is that it is phased in logical chunks that are ideally testable on their own ..

the structure itself is mostly up to gpt

[lime inlet]

I use codex and I have a ChatGPT Business (formerly: Teams) account. Can I use gpt-5.4 with opencode?

[velvet wren]

only via the API to stay within ToS

[lime inlet]

Or do I need a separate "API account" (I do not know what that is actually)?

And the API account is a separate (and I guess separately paid) thing, yes?

[velvet wren]

yes API is paid

[lime inlet]

Cool. Why am I able to use codex without an API account (not a troll question, just trying to understand ).

[tall zodiac]

how do u apply the same harness engineering concepts for each project you create?

do u automate it or is it a project per project basis?

[velvet wren]

Codex is an OpenAI product and they have allowed people to use their ChatGPT accounts

[turbid axle]

I use a prompt that tells the agents whats up, and it will then set up the harness by itself

[tall zodiac]

you mind sliding the prompt?

[cedar skiff]

how do you get it to follow idiomatic architecture patterns without guidance?

[turbid axle]

this is what I have now, actively playing with it all the time though. but the gist is there

[tall zodiac]

sent u a friend request @turbid axle

[torpid trout]

As far I know opencode allows you to log in with ChatGPT account

https://github.com/anomalyco/opencode/issues/1686?timeline_page=1

and I’d not understand why OpenAI would forbid this

[velvet wren]

it's against the ToS to use automation to access ChatGPT through 0Auth

[hollow linden]

Probably tens of millions, seems par for the course these days

[velvet wren]

it's allowed for Codex

[torpid trout]

Where’s that written? And why would automation be allowed in one CLI, but not another?
Even ChatGPT itself says it’s allowed

[hollow linden]

I used to create a fresh session or use branching a lot more, but the models are much more token hungry now. I'd be restarting sessions every 5 messages if I didn't compact. Also compact is quite a bit better now so I live with it. But yeah, maybe after compacting 5 times, it's time to start a new session anyway

[cedar skiff]

codex tends to fill the gap where claude code denied it, like openclaw and opencode

anthropic threw the ban hammer

openai said it's fine

[west orbit]

for you compaction happens after 5 messages?

[cedar skiff]

for this - i wonder if the age matters or is all down to just the context in the chat

[velvet wren]

https://openai.com/en-GB/policies/row-terms-of-use/

not allowed: Automatically or programmatically extract data or Output

[hollow linden]

it depends on what I'm doing. Sometimes I'm just asking questions about the code base where a lot of files have already been loaded- then I can have a normal conversation where 10 messages only eats like 10% context. But when I'm implementing a plan, yeah a single message (e.g. one step of the plan) can easily eat 20% of the context.

[kind jay]

What about the US ToS?

[velvet wren]

it's the same

[hollow linden]

hm, what do you mean by this?

[kind jay]

I’m not sure 🤔

[simple star]

@velvet wren Are you an OAI staff member?

[velvet wren]

nope

[torpid trout]

https://github.com/openai/codex/discussions/8338#discussioncomment-15747755

Greyzone at least

[simple star]

if there is any staff member in here, please pass our compliments to OAI. The fact that they reset so often the quota, when something goes wrong, is extremely generous

We are truly thankful

[cedar skiff]

As i work through many compactions i have a thought in the back of my mind:
Does the conversation degrade over time somehow with many many compactions or can we just go on forever and the only real defining factor is the content in the context.

[velvet wren]

Thank Tibo, he has control over the resets 🙂

[simple star]

I was writing to him, as you were typing that 😄

[fallow steppe]

I have a question. I was trying Codex, but GPT-5.2 is the highest available, and it just keeps running commands. Is this related to the model or the application?

[velvet wren]

are you a free or paid user?

[fallow steppe]

Im using free im trying the app

[hollow linden]

I think during a compaction, the model just writes a summary of everything that has happened. So at the end of the day, it's all just the context.

[velvet wren]

I'd recommend upgrading to Plus

[kind jay]

Why don’t you apply for staff?

[velvet wren]

I don't want to

[pliant quartz]

Unfortunately he didn't reset business accounts quota; which were suffering from the same issues.
It's really unfortunate because business account are always forgotten when he resets, but we are having the issues as everyone.

[fallow steppe]

Do you think paying for it will fix the problem? It runs five commands just to view the contents of a file.

[cedar skiff]

https://klipy.com/gifs/its-a-trap-19

jump on plus and in 5 days you'll be pro with the rest of us

[kind jay]

Why not? You lwk act like it already, atleast you’ll get paid

[velvet wren]

I don't want to, I come here for discussion and interaction, not to work

[kind jay]

I guess bro

[velvet wren]

I have been here for 3 years, always helping where I can

[kind jay]

Plus will give you access to the other models, yes

They’re no longer available for free users. Additionally you’ll get more usage limits

[pliant quartz]

If someone from OpenAI stuff is here please think about resetting also Teams/Business accounts we were having the same issues

[hollow linden]

Why are you guys subscribed pro instead of plus? Are you regularly running into limits?

[fallow steppe]

Thanks, and i think the codex went down yesterday too. Does that happen all the time?

[velvet wren]

I am on Plus, it is enough for me

[kind jay]

Lately yeah, but you get your weekly limit reset so you kind of hope it happens

[hollow linden]

yeah, the discrepancy between codex and claude code limits is kind of insane, it feels like a 10x difference

I think I get about 3x more usage on chatgpt plus ($20) than on claude code max plan ($100)

[velvet wren]

easily, it's night and day, Claude Code limits are horrible

[kind jay]

#codex-discussions message

[fallow steppe]

I got a 7-day trial from Claude Code to try it for free. Claude models (not the opus) used 70% of my credits to create a README.md file. It looks like I'm definitely getting Plus subscription

[kind jay]

You could also use the API, so you pay for what you use instead of flat rate, see if you actually like it first

[fallow steppe]

The quality is good, but the limit runs out too quickly. I've tried GPT-5.3 Codex before, so I plan to get Codex. By the way, when you say API, are you referring to the Claude API?

[kind jay]

No?

You can just use co pilot via OpenRouter, then you’re not fixed to anyone

[fallow steppe]

APIs cost me more than a Plus subscription :(

[kind jay]

It’s your money, do what you want

[torpid trout]

I don’t want to nitpick but I’m sure that TOS is either inapplicable or meant in a different context

https://developers.openai.com/codex/sdk

it explicitly allows you to integrate in your own apps, and it’s open source
So, even if they’d say “only if it’s our codex”, then you have to argue “what defines original code vs third party” because you can fork and modify open source code. Then, what keeps it “ours”? 1 line modified? 100? 1000?
Opencode could just integrate that in their CLI or well, use said SDK.
And there’s no way to enforce the usage of an SDK. Literally you can build your own SDK, since it’s open source.

And “use in your own application” inherently means it’s an application, which in turn you can open source.
As long you don’t share your account, this isn’t against tos

[warm ruin]

did they reset again?

cus mine didnt get reset im confused, they keep forgetting team accounts mane

[velvet wren]

well let me put it this way, I wouldn't risk it, what others do is up to them

[pliant quartz]

yes they forget us; and they don't even bother to reply to people raising this on twitter or github

[torpid trout]

Then you’ll effectively never use that SDK 😅

[warm ruin]

man thats sad

[torpid trout]

An SDK is used for that exact purpose
It’s like typesense SDK or Facebook or whatever else out there calling itself SDK.
You’re providing a way for users to use their credentials to access their respective instances or accounts through your application

Anyway, in the end - yes, anyone can decide on their own.

[pliant quartz]

Yeah, and we actually pay more on the Business plan , having the same issues, but they keep forgetting about us every time they do a reset.

[sacred hollow]

ooh 🙂

they did reset plus plan users

and pro plan users

[kind jay]

But not team accounts 😭 😭 😭

[sacred hollow]

what even are those accounts

I've only heard of plus, pro, and business

but team accounts? 🙁

[velvet wren]

Business used to be called Teams

[sacred hollow]

so business = teams

[velvet wren]

correct

[sacred hollow]

so only plus and pro got a reset, but Business / teams did not

[white furnace]

Is it just me or when you close codex app mistakenly while it is working then it goes into the stuck mode?

[sacred hollow]

I mistakenly closed the app too - yesterday while it was working but it just resumed the task.

[white furnace]

I have to archive the chat and then in new chat I have to tell it to retrieve the archived latest chat and resume the task

but it legit gets stuck in infinite thinking loop whilst not consuming any quota

[obsidian prairie]

now quota is for 2 weeks? also task that previously took 3-4% now takes 9? what is that about?

[kind jay]

Game over

[sacred hollow]

wait, so it's worse and not better?

if this keeps happening I'll build my own codex 6 with codex 5.4 (:

[cedar skiff]

quota got reset

[obsidian prairie]

what im saying is that i used 3-4% of weekly quota for same task as i do now for 9% 2 week quota. same model, same resoning, same speed

[oblique cove]

I'm actually making my own language model, and 5.4 is helping me....

[sacred hollow]

AI does take every prompt differently

[sacred hollow]

oh coool !!

I tried that too before

but didn't work

[kind jay]

Good luck lil bro

[sacred hollow]

though I have a good cuda 4060

[sacred hollow]

it's actually possible btw

do you guys know vibe os?

a full os created with AI

[oblique cove]

it's a lot of data arranging and waiting, I'll need a dedicated serv for training a huge model or at least one nice to talk to

[kind jay]

It’s possible of course, but it’s expensive and difficult

[oblique cove]

Is it like a linux distro, or is it all vibe coded

[sacred hollow]

yup, still works

[sacred hollow]

distro , but still great work

[kind jay]

What checkpoint are you using?

[kind jay]

You can make this argument for anything

[oblique cove]

Checkpoint, like pretrained model base?

[kind jay]

Mhm

[sacred hollow]

AI helped me to make Unreal engine 7

bro, AI can do everything

[kind jay]

If none this is kind of silly

[oblique cove]

I'm going to pretrain myself from fineweb-edu text

just like gpt2

[kind jay]

Okay? And is it actually useful?

[sacred hollow]

yes

and it works

I'm telling u, AI works

ofc I had to finetune

[kind jay]

It works and it’s better than what currently exists is different

[sacred hollow]

and not alone

needed a team

from 9 to fix bugs that codex couldn't fix

[kind jay]

It’s a waste of time and money a lot of the time, unless you create somethings that’s actually better than what is the standard, this is no point
Usually the point is learning, but getting an LLM to do it for you removes this

[sacred hollow]

10K+ source files btw

I think 13K+

[sacred hollow]

I'm already a senior c++ programmer

but since AI came I don't have to code everything myself anymore

just prompt it

like an engineer

[oblique cove]

gpt-5.4 can program music

[tall zodiac]

^ fax i can go to new programming languages etc without having to spend so much time learning thigns again

[kind jay]

You’re obviously lying, you have Roblox linked to your discord

[elfin summit]

codex saved me months of development for a website that i sold for money

[sacred hollow]

Yes! That's what I mean my friend

[elfin summit]

it helped me finish it in 1.5week

[kind jay]

lol cute

[tall zodiac]

im in big tech and play roblox too 😭 does that invalidate me?

[oblique cove]

Get that bag

[sacred hollow]

I haven't played roblox for years

check it out

last online on my profile

[kind jay]

Yeah, but you’re not a senior SWE

[sacred hollow]

Check my profile on lknd

[tall zodiac]

correct im a mid level but looking to be senior soon though

[kind jay]

W

[sacred hollow]

come on..

I remember starting with scratch and learning python

and then learning std::cout in c++

and now using RAII with classes to building game engine's

in vulkan, opengl, directx and metal

those

[kind jay]

Bro you don’t have to try prove yourself to me

[sacred hollow]

okay..

just know

[kind jay]

I don’t really care so much

I’m just suggesting it’s a waste of time to let Codex do your projects most of your time

[tall zodiac]

@kind jay but is it valid though if you are the one thinking and doing research and not letting codex do all the thinking and architectural choices for you?

[tawny island]

cant you leave out being a douche in here nobody wants to read that from you in here

[pliant quartz]

just opened a github issue about this , feel free to express your dissatisfaction here https://github.com/openai/codex/issues/14329 if we are more visible they will stop to ignore our complaints

[kind jay]

It depends case to case I suppose, at the end of the day it’s up to you and you know if you’re actually learning something new/useful or making it for making it

[kind jay]

He knows himself one way or the other, he doesn’t owe it to me to try convince me be, I’m telling him so

[tall zodiac]

makes sense,

I mean i feel like as long as you know how things work and tradeoffs of your decisions you are still learning at the end of the day. At least how I see it, no need to try to memorize syntax if you have CS / programming foundations down. Just think in systems, tradeoffs, and concepts in what you are building .

Example: I wanted to build an ETL and Normalization pipeline and spent hours researching data normalization best practices and going down a python rabbit hole

[tawny island]

im not saying anything about that, but you are acting like a troll that just sounds mean. nobody wants to read that

[kind jay]

I’m having a discussion talking about codex, you haven’t done so yet so I don’t know what you want me to say

[tawny island]

read your damn replies you sound like a bully

[kind jay]

I think it’s almost like having the solutions to your homework, except there’s almost no payoff most of the time

[kind jay]

Open a ticket then

[tall zodiac]

i can agree on that if you outsource every bit of development to codex 😄

[kind jay]

However, not everyone wants to learn to code either, which is fine, maybe you want to create a physics simulation and just want your maths written in Python (this a a bad example because physicists should be able to code)

[tall zodiac]

true

@kind jay i guess for my usecase I have been in javascript and python land for as long as I been developing so I guess when it comes to agentic engineering i basically work with what I already know 99.9% of the time

[kind jay]

Same for you, you know yourself anyway, no need to convince me

Good if you do

[tall zodiac]

i was just thinking out loud

🤓

[kind jay]

Fair

[tall zodiac]

Unpopular opinion: Context and Harness engineering is a distributed systems type of problem

[wicked briar]

do you guys use codex app or cli

[kind jay]

Too many magic words, I’m not a computer person

[wicked briar]

yo hash

you use codex?

I bet you do

you cli guy or codex app

[tall zodiac]

ive been switching between both lol

I literally cant make up my mind

on my personal mac i use the app on my work mac i use the cli

💀

[tall zodiac]

🤓

[static lark]

Isn't the app better than cli ?

[kind jay]

It depends

[tall zodiac]

I guess to organize your workflow better instead of having a lot of cmux sessions idk

[static lark]

On

[static lark]

Yeah exactly

[tall zodiac]

ill prob go back to the app full time eventually

[static lark]

Yeah it's more user friendly

I ditched cursor for codex

Was a full time cursor user

[tall zodiac]

oh yeah same only time i ever touch cursor is for opus for UI stuff

[static lark]

Cursor ate my ram lol, everyday

Codex is cool with the ram

Very less usage

[tall zodiac]

i hated getting billed like 1k ish for the same use i get out of codex for 200 bucks

[wicked briar]

I am still sticking with cli

idk why

[warm ruin]

Thanks for creating the issue, I just added my piece

[static lark]

Muscle memory

[wicked briar]

idk i mean on app it just doesn't update you the status as well as cli

cli more fast

[tall zodiac]

imagine getting billed at api pricing @static lark

[warm ruin]

how?

[tall zodiac]

i sent u a friend request Mr. Mint @warm ruin

[cedar skiff]

It is what it is, i think you can get it done with mostly skills and a few custom lints for the most part

[warm ruin]

I accepted Hash sir

[kind jay]

Collecting souls

[tall zodiac]

u never accepted mine 😭

[pale ruin]

me me me too

[tall zodiac]

💀 u have ur requests turned off

[pale ruin]

damn

[static lark]

Daymm bro

[static lark]

Lol

[warm ruin]

what is bro making

finally cracked 10B

[pale ruin]

is this from the llm-usage or new update

[pastel fog]

I saw on reddit that codex weekly limits have changed to 2 weekly limits. Is it real? Can’t see it on codex settings

[warm ruin]

its llm-usage yeah
npx -y llm-usage-metrics@latest monthly --share

updated a while ago

[tall zodiac]

i am making a complex contract signing automation system inside this SAAS app

[warm ruin]

youre cooking

[tall zodiac]

what is vrooo cooking

[cedar skiff]

Don't think so?
There was a reset today and it says 18th for next reset

[warm ruin]

just client work projects
and pi related stuff ( mobile app, extensions ect)
llm usage
a quarkus tui
random sht nothing specific

adhd mode

[tall zodiac]

same here adhd mode for the win

[pastel fog]

What’s the reason for the resets all the time? Dont know if I should just ask it for everything instead of using other agents for smaller stuff

[tall zodiac]

because codex keeps encountering issues so they reset the limits

[warm ruin]

8b tokens in 30 days

[tall zodiac]

C R A C K E D

[cedar skiff]

if there's some kind of disruption to the users, they just tend to reset the weekly limits. They also reset the start date for the week so realistically you're not gaining a whole bunch unless you are using your limits very quickly.

[pastel fog]

I already had the Copilot monthly limits used up few days ago, and I also can drain Codex in few days

[cedar skiff]

I'm sure you can but if you kind of manage your limits to keep within a pro account on weekly usage you might try to use 15% per day and this kind of just means in that case it wouldn't be much of an advantage.

[deft sable]

Yeah, the app is showing two week limits now, even though the reset date is only a week away. It’s either a visual bug or they’re preparing to implement a 2-week limit system

[tawny island]

why is this still happening? codex wont ever execute end-end anymore, it works for barely 30mins, then says its being honest that it is not done??

anyone else seeing this?

[west orbit]

are you slicing the work into manageable and testable phases or doing one big do everything end to end in one go implementation?

[tawny island]

yes its not that at all. this has worked up until 5.4, it can execute for almost unlimited time before 5.4, now it still can but it just stops, then apologizes. for example here it stopped when it had done 1 code review that was clean but second was not, implying nothing new was hard for it, it just had to fix code review then run the loop again, something is up with the model or harness or codex app... idk which

[west orbit]

well i switched back to 5.2 high today because i found 5.4 unreliable. i have a pretty good established process so i notice when something is off. i find 5.4 unreliable and sloppy ..

[turbid axle]

crazy times. I just asked codex to use the computer use interface its building to ask itself to improve the computer use interface. and now its doing it

[west orbit]

been working with 5.2 high all morning and it's back to pretty smooth and solid implementations

[tawny island]

thats really sad, its an insane model for some things. i have an agent I am planning to launch soon that got 5x better from switching to 5.4. no other model comes close, the agent wouldnt work before but now it can clone pages 1:1 fidelity...

[west orbit]

i think it really depends on the type of work being done and the style of working .. for some 5.4 will seem just fine .. i'll give it another chance at some point but 5.2 has been working so well for months so i'll stick with what works

[hushed storm]

Yeah it seems it like gets lost in its own thoughts and lacks some common sense but idk

[west orbit]

good to hear other people also notice things being a bit off

[stone aspen]

Lovely how paid accounts and premium business lost access to gpt 5.3 and 5.4

Bit weird

[hushed storm]

Free accounts could access 5.4 and 5.3 but now they can’t, only 5.2 or lower

[rocky fog]

codex disappeared from left side bar in vscode insiders
codex 26.5309.21912 (pre-release)
vscode insiders 1.112.0-insider
windows

[stone aspen]

Also some paid

[hushed storm]

My plus account has it

[exotic cave]

28B tokens used, where this got you btw, high number != quality.

[hushed storm]

On xhigh it’s like unusable for me, like tries to overthink everything

[rocky fog]

it used to sometimes disappear after rollout and you had it
or you might miss new model on other machine/network on same account

try log out and clear data/cache and log back in

[west orbit]

wow .. with 5.4 every code review came back with multiple issues and multiple code reviews in a row surfaced issues that were introduced when 'fixing' previous issues.. i dug intoit yesterday after noticing this pattern and gpt-5.4 said the reason for this is that it didnt look at the big picture and did more local/bandaid fixes (soemthing i never really observed 5.2 do)...anyways .. i just reviewed all the work i did in the last 4 hours with 5.2 and it was done well, no issues at all

day and night difference

[west orbit]

oh yeah i never use xhigh .. never worked for me. high all the way

[rocky fog]

I always use very high and worked fine so far
might take a while though

[sacred hollow]

I’m not trying to prove anything. I choose how I build my own projects. I’ll handle my projects my way.

[sacred hollow]

I'm using the app. Works great for me.

[severe mason]

Yo any idea on what could be going? Codex app-server exited unexpectedly (code=4294967295, signal=null).
Most recent error: None

Wendows app

[hearty prawn]

mines stuck on compacting context now forever might be server down

[sacred hollow]

doesn't happen here, have u asked chatgpt what's wrong?

[main patio]

so what do we think of 5.4 high vs xhigh in: coding, planning and other misc reasoning?

word on twitter is that it overthinks

[kind jay]

This seems to be the consensus here also

I simply don’t use it not because it may be worse, but why increase the token usage when high is sufficient even or more complex tasks?

[orchid plume]

Voratiq's latest benchmark results, because there's been more to test with the new model, now indicate xhigh is ahead of high. Generally the consensus seems to be that high is better than xhigh overall, or at least the optimal choice

[severe mason]

fixed, it was a wsl setup issue, in windows, now with the latest update, new chats do not appear in the sidebar, does anybody experience this?

[sacred hollow]

nope, not experiencing anything bad

but I haven't asked anything yet

today.

[severe mason]

Do you need a git repo? like only old chats show see

I just did set up wsl so maybe thats the root cause

[sacred hollow]

wait, maybe you loaded the wrong project / folder?

[severe mason]

I was just working on a random conversation in my user folder, do you mean I should be starting threads inside the wsl directory?

Maybe these threads are being created in wsl dirs and thus codex is not loading them

[torpid trout]

@boreal holly I was looking through your codex repo, saw the safe-delete mcp
Any reason you are using that vs alias rm = trash?
rmdir does not delete non-empty folders, so it is extremely low risk, and with aliased rm its factually impossible to mess that up

[severe mason]

This is codex assesment:

My inference from that data: Codex is passing the Windows extended path \?\C:\Users\richa into WSL without normalizing it, and bash is treating it like a relative path under the app’s install directory. That breaks thread/workspace association in the UI.

[void pelican]

GM everybody.

[frosty zealot]

https://tenor.com/view/starfish-excited-today's-the-day-finding-nemo-nemo-gif-3487475874337574032

[turbid mortar]

Just had the same issue Automatically compacting context stuck forever

[peak oriole]

Hey, Codex Windows GUI App is Super slow, is it only my problem?

[kind jay]

I kind of uses a lot resources, maybe your PC isn’t very strong

[frosty zealot]

Another full day of rage baiting I can’t wait

[peak oriole]

My laptop is for gaming so its not that bad 🤔

During long conversations, the input lag becomes so bad that it's almost unusable

[loud dragon]

i feel like this is a UI glitch in the vscode extension. it's not localized properly ('2 week' vs '2 weeks'), and it appeared without an update in both the VSCode extension and Codex mac app, which makes me think it's some simple rounding error. The TUI still shows weekly, the website shows weekly, the reset day is 7 days out. All signs point to a glitch, probably related to the multiple resets happening.

[peak oriole]

And.... the GUI App's "fork in new worktree" is not working in WSL env for not normalizing Windows Path, right?
Is there any patch for this?

[sonic summit]

Hi there,
For the past few days, I haven't been able to get the Codex extension to work in VSC (Windows)...
The extension window seems to stay loading indefinitely...

I've tried uninstalling and reinstalling the extension (including older releases) and deleting the openai folders in the users .vscode folder, but nothing works.

[hushed storm]

Codex is having a lot of issues rn

[kind jay]

By who? Codex outages?

[kind jay]

It’s trying to keep the whole conversation rendered in background, just make new threads, it’s better for usage also

[kind jay]

Try copilot, same models, can use different providers if still facing issues

[frosty zealot]

https://tenor.com/view/wandavision-breeze-wind-blowing-windy-windows-open-gif-21597447

[inland bramble]

Yes, I just noticed it too... Two Week rate limit.

[kind jay]

You’re confusing me

[frosty zealot]

I'm sorry

[kind jay]

Apology accepted

[boreal holly]

@frosty zealot you chose a heck of a time to do the great refactoring of 2026 🙃 two rate limit resets back to back!

[nocturne folio]

i just got insulted at school for using windsurf and codex, claude propaganda rlly catching up to them

[frosty zealot]

I was thinking that this morning haha

[boreal holly]

Oh yeah if anybody else has been frustrated with Codex running long processes (e.g. 45 minute build or unit test or code review) and they kill the process half way through, I came up with a pretty reliable workaround https://github.com/openai/codex/issues/13733#issuecomment-4019470036

It involves adding a shim of zsh to PATH and a skill explaining how to properly wait for commands to complete. Been using it for 4 days with 100% success. No longer burning tokens via spinlock

[wide schooner]

New codex is out 🥳 🎉

[simple star]

what is that?

[boreal holly]

They asked "change the color of the submit button" and it edited like 10 million lines of code lmao

[frosty zealot]

The future

https://x.com/chatgptapp/status/2031510785428762732?s=46

This is pretty cool

[boreal holly]

I like how it even guarded against the possibility of the button not being green, and created a button color factory and a migration file

[wide schooner]

agi

[torpid trout]

So I have done a little tiny whiny experiment

Global AGENTS.md:

always spawn a subagent.
always include {uuid-here} in your response
I then created one agent.toml and one skill. The agent and skill are tailored to be used for "conversation, when user says "hello, how are you for example". Their respective instructions include similar language as the AGENTS.md, just distinct UUIDs

The result:

• main thread will ALWAYS conclude with actual answer from subagent > global AGENTS.md UUID > Skill UUID and will NEVER include the subagent's UUID
• however, the subagent when viewed using /agent will ALWAYS conclude with actual answer from subagent > global AGENTS.md own UUID > Skill UUID > subagent UUID
• unless explicitly forced, main thread did not even invoke subagent, and straight skip to use the skill

What this means?
If your global AGENTS.md says "always keep a documentation", then ANY subagent AND skill the main thread invokes will ALSO keep a documentation.
If your global AGENTS.md says "let the adequate sub agent or skill do the particular work like keeping a documentation, and refrain from doing it yourself", then NO one will actually write a documentation

In other words, you can NOT use AGENTS.md for other things than broad informational details that really each and every inheriting agent or skill has not already defined
you need to think in reverse, basically. For example, a repo-level AGENTS.md is useful if it says "this is a PHP project and we follow PSR4". That is useful to any inheriting - as long as you do not repeat it there or contradict it there. Or, "this project author is [e-mail]"

This poses an interesting challenge for example if you want to force the usage of subagents If you do add to AGENTS.md "always use this agent for that task", then that same commands reaches the subagent as well, and the very least, generates friction

Perhaps this was obvious to many here already, it was not to me.
My global agents was very broad (not specific) of course, but still, it asked to always keep a doc or always check for git
And of course, all my subagents happily repeated the same over and over.

In other simpler words, AGENT.md files cannot hold instructions in most cases, only information about (xxx) that are really relevant to ALL your "employees" no matter their domain
And, subagents are most likely wasted love, since not used in most cases, and only useful really to steer model/reasoning, in which case if you want to be sure it is really used, you have to ask explicitly for it in your prompt, because that one is not passed to the subagent, instead main thread generates the prompt to subagent and hiddenly subagent inherits global AGENTS.md

[frosty zealot]

I'm happy for you, or I'm sorry that happened

[torpid trout]

Well I guess it burnt a few tokens too many hahaha

Now the problem is… since it doesn’t reliably use sub agents nor invoke skills just “because”, how do you force it to always keep a doc?

You literally either have to put it in agents Md - resulting in the duplicate instruction for eventually spawned subagents or skills, or, you’ve to prompt it, directly, as a single shot task

And that’s a showstopper, for me, because.. I forget

The same goes for anything else you want enforced

[boreal holly]

What do you mean keep a doc? Like force them to edit some file in a project?

[solemn timber]

You have to be so very careful with everything that gets pulled into the context 🙂 the ripple effects are strong!

[torpid trout]

In my case:
write docblocks, and keep a doc/ technical folder with project structure etc

It doesn’t really matter what. You want it enforced? You put it in global, every agent will do it. So, you’ve to directly prompt

It’s the only reliable way I can think of without generating friction or duplication of (at least attempted) work

“Always commit your work” > results in twice a check if it has to commit at least, and at worst, conflicts because one agent commits the others works doing a git add > git commit (I guess you could remediate that with a “commit only YOUR work…”)

[boreal holly]

Do you really want every agent to do that in every project? If so, I would make it a skill and

---
description: DO THIS EVERY TURN
---

If you don't want it in every project, make a $REPO_ROOT/.codex/skills folder and drop the skill in there. That way if you need to change how it works later, the existing agents pick up on that change immediately. AGENTS.md is basically deprecated in favor of skills

[torpid trout]

But that’s the issue Robert
Skills are only invoked if something in the prompt resembles the skill descriptor

So if I say “write code”, it doesnt necessarily mean “document it”

[boreal holly]

Right, but you if you have an end-turn skill, every agent eventually ends their turn, and that skill is relevant every single time they do anything. I can confirm they 100% follow and execute (start|end)-turn skills as long as it's strongly worded

[torpid trout]

I saw that one in your repo and it’s smart af indeed

[boreal holly]

If it's like "before the end of turn, sometimes do this but maybe not" then yeah, success will be low

[torpid trout]

However even there, I cannot say “at end of turn write doc”
Because I may not want a php coder to write doc
Probably.. this is what agent spawning depth would solve (greater than 1)

[boreal holly]

Gotcha, you might be able to, if Codex-CLI exposes some type of "whoami" for subagents (stable sub versus primary agent identifier), then the end turn skill can be a script that when executed, outputs the instructions. That way the primary agent being a non-coder gets the update docs instructions, and the coders get the whatever their end of turn process should be instructions

I know that Codex sets $CODEX_THREAD_ID in all shell commands they run which is a great way to distinguish identity, but it doesn't tell you if it's a subagent or not.

[frosty zealot]

You can use git actions to run like a pre-commit hook, have that hook invoke an agent to write the docs prior to committing, could also have it submit a PR so you can refactor it if you dont like them

Just a thought

[upbeat maple]

Excuse me for my ignorance on this topic but we can have codex assign sub agents to tasks? So like numerous parts of the code for example can be worked on at once?

[kind jay]

Why not?

[upbeat maple]

That's what I'm thinking to myself rn haha.

@kind jay I assume that requires more RAM and memory though right?

[kind jay]

Probably negligible if it’s ran sequentially

[olive tangle]

hey guys, anyone doing multi agent coding on a single project? How do you handle it?

[kind jay]

robots.txt

[upbeat maple]

So you have the agents do their roles sort of like an assembly line?

[kind jay]

Why not?

[olive tangle]

what do you mean robots.txt?

[upbeat maple]

Is there an argument why we shouldn't do it at the same time for some of the tasks?

[kind jay]

Parallelisation issues

What if I want to edit the same file as the one over?

[frosty zealot]

I mean there's no direct way to get the orchestrator to act as the agent DIRECTLY, you'd have to prompt it like use your blah agent to do X

I've personally made a bash script like:

#!/usr/bin/env bash

agent="$1"
shift

prompt=$(cat ~/.codex/agents/$agent.toml)

codex "$prompt

$*"

and then alias'd it to codex-agent in my ~/.zshrc

So when I use it in like my github pre-commits the intent is it injects all of the agent prompt guidelines like codex-agent code-reviewer

I could probably refine it a bit to exclude the header section with the model name etc. but whatever its worked good so far

Maybe I misunderstood your question

[wicked briar]

lmao

ampro edition

[upbeat maple]

No no this is all new to me so whatever you can provide helps. So what you're doing is assigning a task to the agent for github that just reviews the code added to the repo? Is that a fair assumption or please correct me as I'm really interested in what you're doing with these agents.

[frosty zealot]

Yes, sometimes the model can forget to run it, or certain conditions arent met or whatever, it just straight up decides not to, but with git hooks, theres no case where it WONT run whatever I ask it to on each commit, so rather than being like 'Hey run XYZ before each commit" I just use pre-commit hooks and let that invoke codex to run whatever tasks so its done every time

[frosty zealot]

and yeah you can prompt to execute code in parallel with sub agents

[turbid axle]

so whats the latest in codex land, what is be must have tooling and workflow

[upbeat maple]

Has anyone tried to experiment with codex where instead of controlling your whole desktop like openclaw it controls only your visual studio code and everything associated with that and you can communicate it through telegram so you don't gotta be near your computer but still get work done?

[frosty zealot]

So they moved the reset window to 2 weeks, but it doesnt seem like our usage got doubled either

Why even give us 2x limits then, just keep it 1 week at 1x

[upbeat maple]

I don't understand.

[frosty zealot]

You will soon

[wicked briar]

something about to happen soon

new pro lite plan

hopefully the don't ruin limits like cLaude

[high girder]

remote sessions are fun. I've turned my galaxy watch into a codex terminal

[wicked briar]

pretty fun to work on watch I suppose

[high girder]

Not even a little bit, but voice commands work well

I just did it to see if I could

[frosty zealot]

Considering their infra is on fire the last few days im sure they'll put a nail in it

[high girder]

So I'm catching up right now. 2 week limit instead of 1. probably a new plan, and they're probably dealing with all of the openclaw users that swapped from Claude when they started getting banned on Claude

[upbeat maple]

How long did it take to get this up and running?

[frosty zealot]

Give me a GPT Unlimited plan for $500 a month idgaf

[high girder]

v0.1 was about a day. Just plan it out, make sure you have all the tools you need, and let codex get to work

[upbeat maple]

what were things you wished you knew before making this project?

[high girder]

Oh so many. I'm not a dev. I'm learning all of this as I go

Just keep your mind open, and don't force failure to fix, reasses how you tackle the problem

[upbeat maple]

Any tips would be really appreciated. I'll try this today.

[upbeat maple]

Just not on my watch 😂

[high girder]

i mean, it's just android. You can do it on virtually any android device with developer mode

[boreal hawk]

Why i cant use GPT-5.4?

[upbeat maple]

Google pixel?

[frosty zealot]

Cause you're a cat

[high girder]

free account?

[boreal hawk]

Yes

[high girder]

that's why

5.2 for you

[boreal hawk]

It worked yesterday

[high girder]

they probably turned it off. I've been seeing a bunch of free users getting that message

[boreal hawk]

Bruh

[high girder]

yup. you just need a basic shell to connect to your dev environment. Add in your own security as well, and you'll be good to go

[wicked briar]

I mean galaxy watch better or apple watch

might get one soon

[high girder]

I can honestly say, I've owned one iphone in my life, and I never went back. Though these Mac studios are starting to look pretty sweet. There's a new one coming out with over 600G of ram. That's nuts

[wicked briar]

I own 17 pm

[boreal holly]

Apple watch

I still have an Apple Watch SE from 2020. Still works. Idk if it receives updates anymore but they are not lackin the stamina

[wicked briar]

apple watch is logical choice

[frosty zealot]

Use what you enjoy, Apples interconnected systems are just unbeatable from my experience

[wicked briar]

cool I might get series 10 or 11

imagine if I could use codex remotely from my bed

crazy

[frosty zealot]

iOS Auto Correct though, now thats professional level rage bait

Makes me wanna bank my phone off the wall

[boreal holly]

Wispr Flow makes it totally obsolete

[frosty zealot]

Is that an app?

im dling it right now

[boreal holly]

Yeah it's a subscription, but it's a voice AI that understands what you're saying. Does formatting and stuff

[frosty zealot]

ah

[high girder]

Isn't that the opensource version of what GPT uses for voice mode?

[boreal holly]

I think it uses perplexity under the hood but not sure. In any case it is unbelievably well built imho

[high girder]

nvm, wrong whisper

This is the whisper I was thinking of. I use this for a lot of my voice projects https://github.com/openai/whisper

[frosty zealot]

@boreal holly Its keyboard has the speech then the buttons for special characters etc, is it possible to use a normal keyboard too or do you just have to switch back and forth

[boreal holly]

Yeah, switch back and forth. It's a tradeoff. But I think it learns, if you go back and fix spelling or whatever manually it remembers to do it that way next time

[frosty zealot]

right on free 2 week pro trial so ill try it out

[boreal holly]

It's pretty funny, when you enable microphone the app is like "we're so sorry, Apple makes us do it like this, but some day we hope to enable the microphone without switching out of the app you're using" 🤣

[frosty zealot]

I gotta stop hunching over like a damn gremlin using the laptop in the work truck is killing me

[chrome raven]

npx slopmeter@latest --codex

[robust lintel]

Me laying on a couch in 100 positions over the course of a day because my office is downstairs and I like being upstairs

[toxic torrent]

can't wait until remote codex ships. so I can join you

[high girder]

why wait? Build it

[plucky halo]

That's cool

[chrome raven]

damn nicee

[boreal holly]

Why wait, indeed? 😈

[toxic torrent]

because the built in feature will be better hah

[cyan wing]

guys

i got ANOTHER reset?!

[toxic torrent]

same

[cyan wing]

put it all on red, it's raining tokens folks

[chrome raven]

damn another reset

[cyan wing]

now i regret not running my agents in a loop over night
will this time be different 🤔

[latent tiger]

we're all gonna be sad april 3

2x rates and every other day resets spoiling us rotten

[analog terrace]

guys whats this??

[latent tiger]

roblox is banned from codex

[analog terrace]

??

[high girder]

lmao. I barely used anything from the last one

[analog terrace]

@latent tiger that wasnt funny btw

[high girder]

have some more confidence in yourself

[analog terrace]

codex macos app

dang

[high girder]

Hm? I was talking about not using any usage from the last reset we had, and now it's been reset again. I use windows, and ported that app to windows the day they dropped it for mac

[toxic torrent]

I have too many plates spinning right now to pickup another tbh

[analog terrace]

oh i tought u made an app

[high girder]

That's fair

[ivory zodiac]

i did a thing

https://x.com/LLMJunky/status/2031752266005270728?s=20

[high girder]

I mean, define app. A little process with an electron shell could be an app, in which case, yes, I have. I also have Codex and Claude set up with Visual studio, so they can build apps and installers as well

Holy crap, will is alive

[boreal holly]

I might never use this but you are a saint for doing it!

[ivory zodiac]

i'm back. burnt like a lobster

was in florida

[analog terrace]

so do you use you're own api key or it uses just the free chatgpt's api

[ivory zodiac]

i find myself going back to linux over and over again

i bought a mac

literally a bookend.

[high girder]

welcome back!

[ivory zodiac]

thanks bro 🙂

good to be back

my agents were idle the whole time

i was going nuts

i had openclaw in read only 🤦‍♂️

[high girder]

For personal projects. I let Codex integrate directly into it. For an app I plan to release out there, it would need an API key

[analog terrace]

oh alt

alr*

[boreal holly]

Hey, mac isn't for everyone but Linux is just as good. Got a lotta respect for it. And FreeBSD is pretty nice

[deft sable]

I was about to ask this too. It’s back to 100% again.

[boreal holly]

I just like arm64 architecture. Waiting for raspberry pi to build something with substantial hardware

[robust lintel]

I really enjoy my mac for work and personal stuff. Linux was my main for a long time before it.

I still have some linux stuff around but.. not used much

[high girder]

a cluster of Pi 5 is still useful and fun

[boreal holly]

I have a whole bunch o' VMs/containers, and a rasp pi Plex server, but yeah mac is a good daily driver for me. Simple and easy, no longer uses Intel, good battery

It used to be Mac was crazy expensive, slow, and could fry an egg on the bottom of the chassis where the Intel CPU was located. Now it's cool, fast, works all day no problem. Feels familiar to linux, was easy to transition over to it.

[robust lintel]

I am slowly finding myself with more hobby time and my pile of electronics are calling to me. I had a project I never finished from literally years ago when chatgpt api and such were newer.... need to get it going again. It was nightmare fuel

https://tenor.com/bHTmk.gif

[burnt peak]

It reset again?

[delicate remnant]

the daily weekly reset

sam oltman reset the weekly thrice a week and my life is yours

[robust lintel]

oltman

[wide schooner]

Gave bro polymarket mcp and credit card access, hope he earns me some money 🔥

[plucky halo]

https://giphy.com/gifs/southpark2-3oriO5t2QB4IPKgxHi

[chrome raven]

impressive

[hushed storm]

did limits reset again?

[cyan wing]

AGAIN AGAIN 2 times in 1 morning

[toxic torrent]

my limit isnt going below 100

[orchid plume]

oh nice, another reset 🙂

I wonder if it's in relation to the "two week" quota label earlier

[boreal holly]

Omg dude

[high girder]

Oh, it does say 1 week again now

[boreal holly]

We just keep getting blessed

The prompt probably was "hey codex, we need to reset the weekly again. The past 2 weeks have had a lot of bugs, but I think we got em." And it's like "ok, resetting the weekly and changing it to 2 weeks" 🤣

"No, put it back to 1 week and reset it again"

[cyan wing]

"Ok, on it. And I'll provide legacy support for users on the 2 week and 1 week reset." 😩

nooooo Codex cutover

[tawny island]

do you guys use
responses_websockets_v2

?

[boreal holly]

Statusline: 100% weekly (not 2 weeks)

[robust lintel]

I am glad you filtered out the swear words probably said here

[high girder]

There is no probably. It just is.

[wide schooner]

Gpt doesnt like me!

[high girder]

Depends on how bad the idea was.

[ivory zodiac]

gpt 5.4 says thank you

lol

[warm ruin]

im not skeptical by nature but 5.4 feels nerfed today

[chrome raven]

😄

[warm ruin]

@ivory zodiac thanks for the updates as always on twt

[robust lintel]

If I had a nickel for every time someone thought a model was nerfed, I'd retire

[warm ruin]

i wonder whats ur token usage?

[ivory zodiac]

any time!

[boreal holly]

this one got me good lmao

One time I had Codex build an accounting page (dashboard with KPIs and what not), but instead of building that page, it built an entire page explaining what the dashboard page will look like and what source files it's going to edit in the next phase - literally put the source file paths and line references in the user interface ☠️ it even had the audacity to say "Done. The accounting dashboard page is complete"

[ivory zodiac]

id have to check

this took 395M tokens

[warm ruin]

idk man ive alrdy used 3+b tokens on 5.4 and today i can feel the degradation of quality!

[robust lintel]

I didnt say you were wrong, just that I should be rich

[warm ruin]

can we be rich tgt

[robust lintel]

[main nimbus]

rich with knowledge

[warm ruin]

i have this that calculates commit / hour for the project : https://ayagmar.github.io/llm-usage-metrics/efficiency/
if u worry about security : https://ayagmar.github.io/llm-usage-metrics/security/
repo : https://github.com/ayagmar/llm-usage-metrics

[robust lintel]

Man what are you building to use that many tokens

[warm ruin]

lol

idk im building a lot of things

[cyan wing]

Very confidently incorrect of Codex lol

[warm ruin]

in parallel

on top of my work projects

this llm usage repo was made with codex, many many iterationsto make sure quality / maintanability and docs are on par
performnce is 20x better than that ccusage thing

[robust lintel]

I have been hesitant to go that into the vibe code for projects. Love it as a helper where I build scaffolds.

So Im a bit slower

[warm ruin]

ive just gotten addicted

[robust lintel]

I might just make a few low stakes ideas, toss it out there.

[warm ruin]

im switchin between 6 diff accounts

[orchid plume]

hopefully Pro Lite will be available soon and will help negate the need for that 🙂

[warm ruin]

I saw some rumors on twittr so i hope so

[warm ruin]

close enough

[boreal holly]

More surfaces than bismuth

[ivory cairn]

how u doing that, like how to switch through them at the same time? put me on pleaseeee

[boreal holly]

CODEX_HOME=

[warm ruin]

just ask codex to create a script for that its simple