codex-discussions

[lean lark]

Use your AGENTS.md files to tell Codex how to do things. Be firm and explicit.
Tell it how you prefer your code: modular, fully comment code, small single-purpose functions.
Tell it to refer to documentation before it does anything else to see how the project works. (This saves it from looking around a lot and saves a lot of time and code duplication.)
Tell it to update docs on every update with details about new options, functionality, and how things work. (This makes the system self-supporting - it documents everything and then checks the docs to see how things work.)
Tell it how you like your code formatted, variables and namespace/class members named.
Tell it to lint the code before it runs and after, so that it knows what it broke - and fix it before it completes any task.
Tell it to build, test, and make sure the code works as-expected.
YOU are in charge. It doesn't write badly structured code. You're not providing it with instructions for how you want it structured yet.
I hope that helps.
Welcome!!

I happen to be looking at my top-level Codex instructions, which has most of my directives that I note above. I'll publish this, my server-level AGENTS.md file, and one of my general-purpose project-level AGENTS.md files. I'm in code now but can get this in my repo later tonight or in the morning (USA).

[frail meadow]

If you're building web applications, using playwright tests with video recording turned on helps with review speed

[lean lark]

Yup, that's for project-level instructions. I tell it to write PW scripts for new tasks and re-use existing scripts to verify no regressions.

Note again ... it writes the tests and then it uses them ... no slop code.

[frail meadow]

I think the models have gotten a lot better about not skewing the tests which has been nice

[lean lark]

Yeah - I have had to go back and remove some nanny-text from instructions because SOTA models don't need it ... some of it. 😉

[nocturne folio]

3 hours now

[tiny schooner]

Thanks. This is my problem. I don't have one.

[nocturne folio]

can codex run on this

[high girder]

Where's the other one?

[nocturne folio]

wdym

[high girder]

Wasn't there another one you had in...that other place.

[cedar skiff]

i had a 12 hour run over night last night, 5.5 orchestration is the best yet, the only reason it stopped is becuase of a memory leak in codex app.

[nocturne folio]

yea diff image

[tiny schooner]

I can't believe ChatGPT Plus is a lot more valuable than the Claude counterpart. I get a lot more usage, in a higher quality.

Hope I could upgrade to Pro 5x. Seems to be worth it.

[cedar skiff]

welcome to the light side

[dusk thorn]

https://x.com/thsottiaux/status/2048958572562710550?s=20

here we go gang

[cedar skiff]

oh wow! these guys are really rampig it up

[eager forum]

what plan on long-running tasks ? Max?

[cedar skiff]

x5 at least, and then you would want to use a weaker model for the implementation

[eager forum]

thats why my task always shows you've hit 5h limit 😢

when will business plan upgrade to 5x or more

[lost drum]

But web is good to use gpt 5.5 pro extended ot 5.5 heavy. Like you cant make codex do complecexed deepresearches with this models

[twilit bluff]

https://giphy.com/gifs/80s-mtv-romantics-sChf4Eo55W8x2

embed fail, lmao

[dusk thorn]

TIBO HIT THE BUTTON

PLEASE

[cedar skiff]

ripped off ... i had 50% left and had a planned workflow for it.

[meager wadi]

why ss when you can send a link...

https://x.com/thsottiaux/status/2048997818673537399

[maiden garnet]

Because any time I've sent a non-openAI link in this server I've been put in timeout LOL

Never know what channel accepts what here

[meager wadi]

hmmm i see

[dusk thorn]

IT WORKED

[twilit bluff]

ohh that's why my weekly usage is back to 100% lol

[dusk thorn]

DUDE IT WAS MY ASKINGS #codex-discussions message

TIBO

HES IN MY WALLS

your welcome gang (it 100% wasnt me)

[maiden garnet]

Can you tell him to do it more often and feed him while he's there

[cedar skiff]

my usage hasnt moved because my long running task was already running

[dusk thorn]

he gets angry when i dont turn the fan on gets kinda hot in there with his computer

[cedar skiff]

I still miss out though

[lost drum]

100% baby

glad I had only 3% left

[tribal widget]

Hi Guys a Question: does Gpt 5.4 consume less Credits then Gpt 5.5 in Codex App?

[cedar skiff]

yes

[lost drum]

but worse ouptut depends what u buildin

[tribal widget]

Thanks

[tribal widget]

Just easy stuff. 5.4 is enough for this

[dusk thorn]

my usage was gonna reset at 4 pm today, but i guess it comes many hours early!

[lost drum]

it makes faster decisions in my experience and your pormtp is beeing treated like he rlly wants to help you and not jsut "make the task done"

[cedar skiff]

5.3 is is enough for the easy stuff as well

and its cheaper again

[tribal widget]

Can i choose it in Codex app too?

[lost drum]

5.3 spark too its fast but I think dumb

[cedar skiff]

yeah

[tribal widget]

I saw only 5.4 and 5.5

[tribal widget]

Ok will try

[tribal widget]

5.3 Spark consumes less then Gpt 5.4 i guess?

[tribal widget]

Ok thank you

[lost drum]

idk that spark is werid I used it ealeir but idk how to porpperly use it

[tribal widget]

And 5.4 mini too?

[dusk thorn]

5 million reset soon??????

[tribal widget]

I will try Frist 5.4 mini and if its not enough then 5.4

[lost drum]

wdm? so do we spend usage as fast as possible?

[dusk thorn]

um

maybe

dont trust me entirely, but if you trust tibo

https://x.com/thsottiaux/status/2049001692725731829?s=20

[lost drum]

wen reset cause if again in 3 days and I wont spend every piece of usage then ggs

[dusk thorn]

waiting for the reset to show up in my codex rn

prob soon

[lost drum]

it will be hard for me to spend 100% of 20x tho I did it last week but I got to a point where I just test it

[dusk thorn]

what

[lost drum]

I got reset

[dusk thorn]

how

did you do that

100% of the 20x plan

[lost drum]

ye I been to 3%

look

[dusk thorn]

your the "max usage user" when openai is calculating plan costs 😭

[lost drum]

XDDD

• I havent had codex use any subagents and I worked on one project only

[keen bay]

hi

hello

@lost drum

[lost drum]

Hello

[keen bay]

Are you hiring?

😡

[lost drum]

Maybe in future but now I dont even have money for myself bro

[keen bay]

why you had to generalize

[lost drum]

wdm

[keen bay]

I aint nigerian

do you have a job?

[lost drum]

1000 of them

[lost drum]

thats why haha

[dusk thorn]

holy crap

[keen bay]

@dusk thorn

are you hiring?

[dusk thorn]

no

im broke

[lost drum]

ye I purchased pro plan to not hire

[keen bay]

One out of three people is millionire

[dusk thorn]

same fam

[lost drum]

in zibwabwe

[keen bay]

So one of us is a millionire

[dusk thorn]

in silicon valley yeah

[dusk thorn]

i wish

[keen bay]

Anyone else online message

@lost drum

[lost drum]

what you want bro?

[keen bay]

Shut up

Lol

[dusk thorn]

yeah borys shut up

[lost drum]

keep it as codex discussion or go to off topic and if you want to get hired go to youtube start posting and positoins yourself on market so ppl will ask you and not you them

i fyou have talent they will come to you

[dusk thorn]

apply to 5 jobs, 20 dollars a day, and say you got $100 pro plan and then you pay for your plan in one day

🤯

it has arrived

[keen bay]

this math is beyond my brain cells

[lost drum]

or run ads on x selling weavy workflows of AIGC

[lost drum]

spend to time

[keen bay]

@dusk thorn are you a student or unemployed?

@lost drum are you a student or unemployed?

[lost drum]

are you in to fiverr or just getting started🥷

[keen bay]

i am unemplyed

[dusk thorn]

what even is codex discussions at 1 am

[keen bay]

Clue: We live in garbage

[dusk thorn]

india

[lost drum]

USA

[dusk thorn]

where 😭

wait i remember this convo

[cedar bear]

Lol

[lost drum]

quantum is a thing

[keen bay]

@cedar bear

Are you hiring?

[cedar bear]

Bro for what should i hire

[lost drum]

@keen bay do the indian method where you go on yt live and cut fish

[cedar bear]

Defo left

[keen bay]

social media manager, digital marketing, sales, project manager

[cedar bear]

No

[keen bay]

software engineering

[dusk thorn]

you will not be saved in the robot revolution

[keen bay]

data analysis

data science

[cedar bear]

Bro is everything in one

[keen bay]

massager, therapist

[dusk thorn]

ironbot is a irl gpt wrapper

[keen bay]

happy massage

[dusk thorn]

nevermind

[keen bay]

happy ending massage

[cedar bear]

💀

[dusk thorn]

im lmao rn

[cedar bear]

Bro thats #codex-discussions U know that???? Right?

[dusk thorn]

can codex give happy ending massage

[keen bay]

fake taxi driver, fake dad, fake plumber?

[dusk thorn]

canceling my sub, hiring ironbot

[cedar bear]

Must be the Meta

[lost drum]

sin

[keen bay]

Maybe you are hiring a stepmom?

[dusk thorn]

its a new irl model bro

[dusk thorn]

💀

[lost drum]

no sinning here

[cedar bear]

Does it reset rate limits like Codex?

[dusk thorn]

gotta be a way

what the f is the codex team doing genuinely

[keen bay]

I think we should beat people who are employed

[lost drum]

I think they unemployed and just work as iinfluencers cause gpt it self upgrading autonomousely

[dusk thorn]

codex still can get feature

not all models

[dusk thorn]

terrible

[cedar bear]

Guys whats one thing U still miss in Codex

[lost drum]

I can hire if you want

[keen bay]

yes

[lost drum]

5.5 codex

pro

[keen bay]

@lost drum

You are an AI professional

Who uses AI professionally

So you must be building a business

That is making 1 million dollars every hour

Do you want to expand your customer base to india?

@lost drum

[dusk thorn]

think very carefully borys

[keen bay]

why

[dusk thorn]

https://x.com/jxnlco/status/2049009055411425522?s=20

[lost drum]

I am back

[lost drum]

yes but I am starting out ask me in 1y bro

[keen bay]

you can hire me

for cheap

very cheap

trust me 🙂

😄

almost free

[lost drum]

I would pay good if I grow ecom brands cause there wiell be tasks that jsut cant be automated but I havent discovered yet the gaps where thigns get manual and annoying so yeah

[dusk thorn]

new method?

[keen bay]

I can work for free

please

hire me

[lost drum]

TOS alert

[dusk thorn]

openai tos?

[lost drum]

I think if you abuse it then yes

[dusk thorn]

no

cuz guess what

even Greg Brockman approves

and hes co founder of openai

[keen bay]

@lost drum accept my friend request

[haughty plaza]

Anyone knows of chatgpt business trial account to get that somehow?
i used to have trial account in the past but would like to try the new models before buying full subscription anyone still has a working link referral or antying for a month of free trial? totally fine with a referral link if that would benefit you also

[oak trellis]

codex with gpt 5.4 is just the best at the moment .. nothing comes close to it not even 5.5

[dense path]

Yeah 5.5 is much smarter and 5.4 feels dumb against it

[oak trellis]

i thought opposite 5.4 is smarter than 5.5 ..

how is the ratelimit you feel its same 5.5 and 5.4?

[dusk thorn]

5.5 is way better than 5.4 what are you on

[cedar skiff]

5.5 burns faster

it is better though, makes some of the same mistakes all the time, but does much better over all.

[dense path]

No it sadly burns faster

[oak trellis]

oh ok

i barley use it 5.4, it is awesome for me .. coming from opus .. so the luxury level of 5.5 not needed right now hahaha

opus is insane bad ..

[dense path]

hahaha I see what you mean

[oak trellis]

wasted 2 days backfill for my data because it just don't read .. even if i tell it to read. it assumes .. like no other model lol

[dense path]

I do really like GPT Models more mostly, but for frontend I think it has another side of the story

[oak trellis]

yes, cant comment on frontend haha

but there is a lot of google stuff i heard with combination .. design.md stuff etc

[dense path]

true

[oak trellis]

so excited to get into frontend so many new ideas and community stuff ..

i love that from gpt: ```I’m not going to promise fake numbers. I’m checking what the fastest operational shape should be on this table, then I’ll tell you what is realistic and what is fantasy.

opus: yes it will be fast.. lol

opus is a pleaser ..

[signal tapir]

Whew! I got a reset almost a full day ahead of schedule. 😁

[heavy rapids]

is that on pro or plus?

[signal tapir]

Teams

The ironic part is that I'm currently taking a break from coding to read up before I continue. 😛

[lost drum]

me too but I feel like I should no matter what use codex even so it runs in bg on some optional projects like Gym app for personal use or other stuff haha

[signal tapir]

I would just end up without credits when I needed them

[gentle harbor]

are you fr mine wouldve reset in 14 hours

why couldnt they just wait

[oak trellis]

man about to say .. wxxf my limits again ..

and yeah nothing is for FREE.. remember bailout is carried by the dollar holder aka world .. so they will get their money back later ..

who is winning are the early investors .. when IPO comes they get their insane X back ..

i win tiny now to get my project done 🙂

hooking us up now .. pain comes later ..

[winged depot]

Nice. Another 35% usage stolen before my weekly reset. I'm starting to think this is intentional.

[cedar skiff]

Yeah i'm getting the saem problem

[undone otter]

gg max out at day 1

[lost drum]

Pro 20x💀?

Cause I been with codex for whole week on 20x plan and left 3% before reset

I wonder how could you spend it in 1 day lol

[lost drum]

ye its intentional they calculate overall profit from doing reset (gotta be the ones that spend it to 0% before reset)

I might turn on speed 1.5x cause no way I will again use it all

lrady builded 80% of what i wanted

[oak trellis]

wait .. so if they do reset .. they change the reset timer itself ?

[undone otter]

no i was joking that it would be best to max out day 1 so every reset is huge for you but i never reach my limit with my pro plan and a work all day with it its the stupid days off with my girl

[winged depot]

Correct. Whenever Tibo "presses the button", it ends the week early for subscribers, taking away their remaining usage.

[oak trellis]

oh .. wait but you have then 100% again ?! ah wait lte me think my brain is rotten right now from blue screen lol

[lost drum]

thats why you use everything each day 15% and you good

[undone otter]

and low tier user get new limits so they get more positiv reaktions for that cose they get to zerro in 1 day everytime

[undone otter]

lol no you would get less if its reseting at day 5 you just get scamt for 2 days

[oak trellis]

ok damn .. That means you lost the 40 surplus you saved for Monday–Thursday planning.

[cedar skiff]

shouldnt have to

[oak trellis]

They’re complaining that an early reset may replace your current quota cycle instead of adding extra quota, which hurts strategic users.

correct ?

[lost drum]

ye it happend for me but I had 70% to spend on last 2 days and they resetted

[cedar skiff]

correct

[oak trellis]

damn . ..

chatgpt: ```My honest take

If true, users are right to complain. A reset should usually be:

top-up bonus, not erase and restart accounting.```

[lost drum]

even went to support for refund but they told me that I wont get it or additiaonal usage NO

I mean refund the usage I was scammed for not the hwole plan

but it happens

better to have codex than not have it so I am not complaining at all

[cedar skiff]

how i work i peak away at work during the week getting stuff done, and i plan a big task to work through with subagents in the last few days. This week it was integration tests....

[oak trellis]

oooh ok

ah yes and then they push the reset further away

[cedar skiff]

i did get a 12hour run last night

[cedar skiff]

i had something like 50% usage left, so essentially they stole 35% usage off me

[oak trellis]

wow

[lost drum]

ye I said that they are doing it intentionally they calculate it they now that most people tend to not spend eveyrthing at once and they leave they usage on last 2 days so thats when they reset

conspiracy theory tho

[oak trellis]

100% on purpose .. public they get fomo .. internal they cut usage ..

[lost drum]

Ye and they praise tibo for resets but most get hurt by it XD

[oak trellis]

lol chagpt: ```My blunt take

Power users should stop theorizing and start tracking:

reset timestamps
quota before/after
weekly usable total

Three weeks of data beats 300 Discord messages.```

[cedar skiff]

its not theory

it happened many times

[oak trellis]

i know .. it defends itself

always does ..

it tells me sometimes stop be emotional

was like w..f

[lost drum]

ye dont get that "shiny object syndrome"

[oak trellis]

but the limit stuff is really annoying .. so its always something .. was prepared to move complete to codex .. still will do it .. but getting emotianal prepared that there always come something new

[lost drum]

ye it was not 70 % sorry it was freaking 95% lol

[undone otter]

If they would care about the usage limits and wanted to give the community free usages, they could have just added 100% on top of the current limit, and you can use the additional 100% as long as you want or for two weeks after your normal limits are empty.

[winged depot]

This is unironically how we win against Altman.

[lost drum]

15% daily and forget

[undone otter]

win aagainst him easy thing just spam pro use in the aapp

[lost drum]

nahhhh better to buy 20x cheaper it is I think

[oak trellis]

that is so crazy now .. didnt thought about that .. so shad..y

[lost drum]

than to buy 20 plus acc

[lost drum]

I think its called "increase profit margin, practices"

[oak trellis]

man.. its so off because most user would say oh wow we get limits reset ..

[lost drum]

and then they tell you that you are wrong and spread "propaganda"

[inland sonnet]

💀

[lost drum]

that math those not make sense

[inland sonnet]

He prob means if you fully use it the 5 hours after each other

[lost drum]

You get whole weekly usage increased by 20x cause the 5h usage uses the weejly usage?????

[lost drum]

But bro the 5 hour uses the weekly usage I think

[inland sonnet]

It counts yes

[lost drum]

wdm

I never useg whole 5h on 20x plan but I spend whole weekly cause the 5h limit is only there to not let blow your whole usage in 1 day bro

[inland sonnet]

That only makes sense if you always fully use your 5h usage

right?

[lucid mason]

And that's why my Codex quota is all used up by the mid of the week and I keep switching back and forth between Codex (VS Code) and ChatGPT. It's not super convenient, but at least I don't have to care about quotas.

[lost drum]

BRO for example you have 100% of weekly usage and you use whole 100% of the 5 hour limit then the weekly decreases by idk maybe 5% or something

so the 5 hour usage those not matter at all if you spend whole weekly usage

[lucid mason]

Unfortunately OAI noticed by now and made chats reach their limit earlier. And to push even further the download links of the last chat message in a length limited chat don't work anymore. 🤷‍♂️

[lost drum]

thats whay I am sayin so its cheaper to buy 20x pro plan than than 20 plus accounts cause pro plan is 200$ and 20 plus acc is 400$ for same usage

[worthy furnace]

The usage limits became laughable

[inland sonnet]

me having plus for 3 dollars 😒

[worthy furnace]

5 hours more like 10 minutes at max

Make a $1000 / month unlimited tier

I'd buy

[lost drum]

spytools serveces would abouse it haha

[inland sonnet]

since a agent can run 24/7 now and abuse

[worthy furnace]

It is unlimited

[inland sonnet]

its inpossible to provide unlimited

[worthy furnace]

Or at least remove this annoying 5 hour limit

[lost drum]

larp

[inland sonnet]

usage speedrun @lost drum

[lost drum]

Yes sir

[inland sonnet]

also every prompt 100 subagents @lost drum

[lost drum]

nahhhhh I never used it on pro plan

[worthy furnace]

Definetly not getting the full 5 hour usage lmao

Its a scam

My 5 hour legit runs out in max 30 minutes of work

Thats not enough to do anything

[lost drum]

pro plan?

[worthy furnace]

Business I think but yeah, I should just buy Pro

[lost drum]

ye actually not possible unless you do 10 repos at once

[worthy furnace]

But I need a workspace

[lost drum]

bro for last week I used xhigh model in pro plan and never reached 5h usage even when I let it run overnitgh autonomousley

[worthy furnace]

Damn

[lost drum]

you cant have 20x on 100$ one

[worthy furnace]

Is pro really that good?

[lost drum]

you got 5x

I see

[worthy furnace]

Whats your weekly at?

And when did you buy

[lost drum]

ohhh I see

[worthy furnace]

Damn

How much was the 20x pro?

Good deal

[lost drum]

we can compare the usage but idk how to check how many usage I spend in tokens this whole week

no

[worthy furnace]

I am in eu too, where can I buy it?

For cheaper

[inland sonnet]

invite me aswell

[velvet wren]

sounds like a scam

[inland sonnet]

its an external api are u sure they are correctly reflecting same open ai limits?

[worthy furnace]

How do they achive more codex for cheaper

There is a promo for regural subs rn too

[inland sonnet]

how u can be sure its same 20x?

u can't know..

[worthy furnace]

How much is the free trial?

I mean how much credit you get lmao

Pro 20x on openai is £89
/ month for me

So its barely cheaper in usd

Idk why you would choose them over openai

[torpid trout]

I called this out 2 days ago. He was patient. Or not.
This is scam and surely against about every tos and whatever out there

You cannot buy OAI subs legally from my other seller than … drumroll! OAI.

Now, can we please do what should have been done 2 days ago?
Ban this user, he is advertising and promoting and facilitating resale of discounted codex subscriptions.

[digital thunder]

<:book_icon:1363314738255364126> Rule 7: No self-promotion, soliciting, or advertising.

-# Do not post or direct message any members of this server to promote non-OpenAI services, products, or projects. Exemptions: OpenAI API projects in #api-projects & GPTs in #custom-gpts.

[worthy furnace]

Yeah, scam

Not even discounted

[torpid trout]

I see the annoyance is gone now.
Good riddance

[inland sonnet]

It's a external API, makes no sense they make money on this unless they are a company that got credit from openai

and trying to make a extra buck.

with reselling-

Saw this with claude aswell

[torpid trout]

It doesn’t matter.
It’s not allowed to proxy or else reverse engineer the ChatGPT or codex service unless api is used.
And api would be so massively more costly you could not resell it for cheaper

[slender plinth]

my weekly got resetted?

yall aswell

[torpid trout]

Jane, why did you unfriend me? 😭

[inland sonnet]

Yes like i said, mostly the case they got credit from openai and try to resell.

[vital imp]

are their any promo's triggerable that give a trial or discounted pro? (I know it's something OAI would do because ive gotten months and months of free/discounted plus and business trial)

[inland sonnet]

No there is not, i believe they did a 2x standard pro rate promo before but same price.

[inland sonnet]

They do trials on plus but thats region based.

[lost drum]

where is he he banned for good?

[inland sonnet]

He broke rules so prob yes.

[torpid trout]

All his Messages deleted. I sure hope this obvious scammer and his buddy are now permablocked.

Problem is they probably will be back with socks

[lost drum]

would never trust such service tho

[torpid trout]

Id never trust such users in the first place
It was obvious from the start. But mods here are too lenient

[vital imp]

ok, yea im seeing 2x codex usage until March 31st. Im building an tool for the jobless (because i am involuntairly jobless, and like 4 close friends/family are now suddenly too so i was trying to save a few bucks lol )

[inland sonnet]

Understandble

[lost drum]

what those it do?

[inland sonnet]

double normal pro usage

[lost drum]

yes yes but the stuff he buildin

[vital imp]

Here's the poor summarization gpt-5.5 gave during my planning phase i did inside gpt-5.5 (screenshot because i sent to a friend via imessage and they wont read text if they have to scroll lol)

So the app makes two practical offers: help me get a better job now, and help me preserve the career evidence that makes the next resume stronger.

[lost drum]

would need to use local llms to not spread sensetive info I mean idk how it works I just heard its a thing am not dev at all so no clue

[vital imp]

theres more behind the scenes trying to game "ATS Systems", but my friend is mean whenever i share technical stuff so the summary leaves it out

[vital imp]

This isn't a commercial app, its for friends and family, and yes it can use local, im just using my free API tokens during testing, and also friends/family do not care about privacy at all

[lost drum]

ohhh I see

[vital imp]

I do this for love of the game and utilitarian purposes. I absolutely hate the shovelware slopfest devbro crap. Like i will not be making a macOS dictation app to charge $10/mo for on-device ASR lol

[oak trellis]

hmm anyone have issues code .. its so slow its not normal

so annoying slow right now

[vital imp]

sometimes its slow, sometimes its not. Unless you are paying for full chooch service tier or fast mode, you are low on the capacity triage

very little you can do on your end when service is down or slow unless you've managed to specifically mess up your networking for only codex and chatgpt

[lost drum]

not for me

[inland sonnet]

how much reasoning do you use?

that matters too.

[oak trellis]

always high but was way better 2 hours ago

i guess europe woke up

eats gpu

[inland sonnet]

It also depends on task

[oak trellis]

same task since a while

[lost drum]

sorry I deployed agi

[oak trellis]

i read today that in the talmud or one of their books they wait for the AI -> AGI to be alive .. its their messiah ...

sam is their minion

[lost drum]

I ahve the whole calendar

I can send whole deepresearch about it XD

[inland sonnet]

gpt 6 eating it all internally

[oak trellis]

send me

[oak trellis]

oh ..

[lost drum]

sure

[inland sonnet]

back to gpt 5.4 maybe faster!

[oak trellis]

using it right now

[inland sonnet]

if still to slow mini!

[oak trellis]

will try that .. thought about fast .. but nah its 1.5

now its back fast

[nocturne folio]

how can i use more tokens

dude

[oak trellis]

its ultra fast now

[chrome raven]

It would be so nice to have 5.5 pro on codex and this one is not as slow as other pro

[oak trellis]

ultra fast ..i guess they turn on compute somewhere lol

[chrome raven]

Probably Tibo said something about some compute turn up this week

[glacial ginkgo]

This morning (~7:00 AM ET), I started Codex App for Windows, it ran, I hit the top bar button to upgrade, and then it disappeared and never came back. I expect Apps to restart on their own or at least provide some user feedback that the upgrade ran and completed or failed. This GUI is getting more flaky every day. The open source Rust TUI has been far more reliable.

I have no way to measure things because it is Windows, but I suspect the Microsoft Store as the distribution platform is a source of a lot of the unreliably behavior regarding updates.

[signal tapir]

Development these days seems to be focused on fast results, not quality. Especially not user experience quality.

[unique spade]

speed issues? just tried to start work with 5.5 but it seems suck on thinking

[oak trellis]

5.4 had now its fast

didnt tested 5.5

[signal tapir]

stop it, and then restart vs code (if you are using that, otherwise I have no clue)

I noticed that the stuck bug only fixes itself cleanly if you manage to stop it before restart.

[glacial ginkgo]

I am working on a product demo sequence for my sales team. I have my knowledge graph database setup as an MCP server and I was going to drive the UX and storyboard through Codex App on Windows. The reliability issues are concerning as I am not confident sales could deal with all the restarting/rebooting without getting frustrated. For GPT+MCP as a stack, does anyone have a better end user tool to recommend? Maybe LMStudio?

[oak trellis]

something is off .. ```The previous wrapper still mangled SQL quotes. I’m simplifying it to plain heredoc commands now so we get the actual live numbers without shell gymnastics.

[stoic bay]

I have an issue: when I sign in to Codex, it asks me to add my phone number. When I enter my number, it only adds a 0 and no SMS message is sent. I don’t know if this is an issue unique to me or if others have experienced it.

[unique spade]

yea was on codex desktop, had update pending, updated, restarted fully the app and it works

[signal tapir]

I recommend not having a tool like modern LLMs as part of a live toolchain. Instead have results generated and cached before hand, and make sure a human is in the loop checking output before it enters the next step of the live toolchain.

[livid relic]

you could potentially set up an open webui instance?

[livid relic]

pretty sure the whole point of the product is the LLM lol

[glacial ginkgo]

Part of my value proposition is keeping the database local to the client with only the RAG results flowing to hosted LLM context windows. So, I am trying not to host or use networking for the database or MCP connection. I am using MCP stdio instead of HTTP. ChatGPT App for Windows won't support MCP running locally.

[stoic bay]

I have an issue: when I sign in to Codex, it asks me to add my phone number. When I enter my number, it only adds a 0 and no SMS message is sent. I don’t know if this is an issue unique to me or if others have experienced it.

[signal tapir]

Then the people using it are going to have to be ready for some rough edges methinks.

that's not saying you shouldn't make it as useful as possible of course. Just that if restarts and the likes re needed, they should be made ready for it so they don't get frustrated

[unique spade]

no idea, i'd switch to 2fa by authenthicator app is more robust than sms

[royal cloud]

Anyone tried the codex spark

[oak trellis]

its fast and extrem dump now

ultra g

arbage

[stoic bay]

How you switched to 2fa

[elfin iron]

this is getting out of hand guys, how to disable this thing

[velvet wren]

Windows?

[finite sun]

I say just let it

[unique spade]

must be in gpt settings somewhere, i think i always setup 2fa auth in the first place not sms

[hard drum]

That’s macOS

[elfin iron]

yeah macos, i now disabled it from settings

[velvet wren]

I allowed that

[small violet]

antigravity trash

does claude have something like codex app?

[stoic bay]

But you tried to sign in codex and you used 2fa instead of phone number verification???

[small violet]

@stoic bay

[velvet wren]

ask in the Claude Discord server?

[stoic bay]

Yes they have claude app

[small violet]

nah thats not my question

do u use gpt for frontend ?

[small violet]

or just answer

some people

lol

[finite sun]

They don’t have an app like codex you’ll have to use Claude code

[small violet]

claude code is the cli ?

[stoic bay]

Claude is still better in fronted specially claude design

[small violet]

and its not free

[finite sun]

Yes think you can use from the web and app also

[stoic bay]

Yes and in the app also

[stoic bay]

Yes it is for the paid users

[small violet]

damn

[finite sun]

They were running some “test”and scrapped Claude code included in the $20 plan not sure if it’s back

[small violet]

i sent one prompt using 4.6 thinking on antigravity and it said ive reached my quota

[small violet]

bro what

[stoic bay]

Yes that's real now antigravity is trash

[neon girder]

this happens to me if chronicles is enabled while i'm in a google meet
i usually just disable chronicles temporarily while i'm having meetings

[finite sun]

Either way Codex is better, I don’t get told to go to sleep or to do something else before it agrees to continue helping me cough opus cough

[small violet]

lol

i think ill just use 5.5 for revamp

should i use x high

[stoic bay]

I have a question, you have access to 5.5 in free plan???

[finite sun]

You’re doing frontend work you said?

[small violet]

no plus

[velvet wren]

I wouldn't use Codex for front end stuff

[small violet]

yeah

my website looks bad

i want to revamp

[small violet]

what do i use then

[finite sun]

Use xhigh for the first turn then turn it down to high to make edits maybe

[small violet]

the only models i have access to is only gemini and gpt tbh

and qwen family

[finite sun]

But maybe there’s better options that’s not really my domain

[small violet]

what do u use to see benchmarks

[finite sun]

(,: oh brother I don’t check those

[small violet]

damn

[finite sun]

I’ve heard you can really tune up any model with the right skills and such. The harness matters a lot

I think you can use Claude code with any model you want I could be wrong! But I think I seen a guide on that somewhere. Not sure if it’s locally run models specifically or if you can do it through ollama cloud also

[signal tapir]

You have to do some shenanigans to make claude code work with models it's not made to use

[small violet]

genuinely why is mythos there

[signal tapir]

hype?

[small violet]

😭

is deepseek api free

[finite sun]

Bro yea all that is just hype all the stuff I seen on X I’m sure you can get a really good look for your project with the right set up

[small violet]

yeah def

i think ill just have to work with gpt for now tbh

[finite sun]

Lots of stuff out there you could find

[boreal holly]

It works perfectly fine for frontend. Not difficult at all to be successful with GPT imho

[velvet wren]

how do I stop it making the same brown tinted web sites? with massive text layouts and image panels?

[boreal holly]

You gotta give it a reference design, give it screenshot access to the browser and ability to hot reload. It will work tirelessly to make any frontend precisely match the reference image

[velvet wren]

ok, I am going to try that right now

[rocky ice]

I'm curious about that how you guys manage many sessions/project? Is there any product or open source tools could help me manage different, cross platform, cross agents projects?

[hard drum]

I'VE BEEN SAAAAAAAVED

[torpid trout]

This stitch thing is pretty cool idea. And free, I see? Thats nuts, at best.

[tall zodiac]

Do you guys use image gen inside codex app or ChatGPT.com ? I find that the codex app makes my images look deep fried 💀

[velvet wren]

I make images using the macOS ChatGPT Desktop app

[rocky ice]

What is that App?

[hard drum]

CodexBar

[tall zodiac]

so codex or does ChatGPT have a separate app?

[velvet wren]

ChatGPT has an app for macOS

[torpid trout]

Codex has system prompts tailored to code, and as far I saw, no access to imagegen2, which is why its image gen is useless in comparison to chatgpt

use chatgpt or api.

[velvet wren]

@tall zodiac

[cloud goblet]

codex in browser has a issue it doesn't stick with 100% it always zoom in 120% where do i report this ?

[tall zodiac]

How do people end up using it for games and stuff in codex? @torpid trout i keep seeing tweets

[torpid trout]

how do you mean

Games are not images

[tall zodiac]

@torpid trout i see people use image gen to make sprites for games then have codex apply them to their game

[torpid trout]

oooh

that's a freaking icon. codex can generate images, no issue

[jovial finch]

Hey guys, thinking of resubscribed to the Plus plan for Codex, is it any good now? Back when I had a plan it used to hallucinate a lot and produce tons of unncesary code and delete my working stuff...

[torpid trout]

But, its quality is not the same as chatgpt /image gen

At least, in my experience.
And, given that codex has baked in coding system prompts it is just the wrong priming alltogether

[velvet wren]

I am on Plus and happy

[torpid trout]

I am actually developing a game (a remake of Call Of War) and all my assets in it are chatgpt > vectorize > then load into code
Code is done with codex
I could imagine that sprites would be somethign I let ai do but at this early stages I did not even go into optimization, its all standalone assets as of yet

Still I would not use codex for that, I think

[jovial finch]

Thank you for the feedback, I used to have the Copilot plan but now they removed models from the regular 10 euro monthly and moved it to the 40 euro one.

Not sure if we're allowed to talk about other companies here, but I have another subscription but the session limits are very hard so was thinking to return to codex, how's the quality?

[torpid trout]

Another thing I use chatgpt for is for example personas, etc, and then feed those to the "specialist", which is codex

[jovial finch]

and is the 5.5 (or other latest model) available for codex?

[tall zodiac]

I see , the only reason i was thinking of trying to use image gen inside codex is because codex had my codebase context and can search files etc @torpid trout

[velvet wren]

I am using GPT-5.5 in Codex with a Plus plan, it's good

[torpid trout]

Yeah, I do get the point
I think if it is just for "icon like" stuff, it is entirely ok. But do not expect the creativity and precision of the chatgpt
At least in my trials, odex image gen was like dalle1

[jovial finch]

Thank you for the the feedback, much appreciated, I'll give it a try :))

[velvet wren]

I one shotted this poker machine app using GPT-5.5 in Codex

[jovial finch]

lol, top 10 uses of vibe coding xD

Can't blame you here, that was one of my first projects too... xD

And it does look that it has improved by a lot in recent times

[rocky ice]

It's really good, I love codex with GPT-5-5, it could help me finish task without any management, and sometimes I feel it's better than claude max 20x

[jovial finch]

oh wow really? Because I'm between these two plans, (or the 5x instead of 20x) so wanted some feedback.

Couple months ago Claude was REALLY good, and it still is, but the rate limiting is hard and I think I'll give gpt a try once more

[torpid trout]

Codex PRO 200 usd plan is unkillable man
I mean, speically with 5.5, the token waste is so minimal, you barely notice something is cheweing away for hours in your machine

Claude is better at agent mangament, there is no doubt at that. Codex does not know what agent mangament is. But, it does not matter realluy

[lost drum]

I can agree

its rip off for openai

[tall zodiac]

@velvet wren how the heck do u change the thinking mode in chatgpt desktop?

[rocky ice]

Year, if you want to use it more, you must pay $200

[cyan gyro]

why do you need subagents if the codex main agent is killing everything

[torpid trout]

That is what I was saying yes. However, they can be and are useful, and codex botches that, totally

[velvet wren]

it's in the menu

[lost drum]

200$ and you get usage worht of 3k$

[tall zodiac]

in the menu?!??!

@velvet wren

[boreal hawk]

How can i make codex to be good in designing

[jovial finch]

ahh understood, because this summer I'm tempted to delve deeper into agent orchestration with my home lab server, but that would require apis and I don't want to burn through credits fast.

But at the moment looking for just a cheap subscription to help with developing web apps, front end and backend, Claude adheres to promts really well, couple months back Codex would hallucinate and only cuase issues so I switched away from it, now considering to give it another shot.

Or would you say there is another plan possibly similar to github copilot, or perhaps directly go via API to open ai?

[hard drum]

Did they gimp 5.5 already? Starting to see a reason why the web chat version needs this in memory:

Does not want unsolicited advice, guidance, reassurance, suggested next steps, coping strategies, interpretations framed as what they should do, or suggested wording/scripts in emotional, interpersonal, imagined, dream, memory, trauma, or hypothetical scenarios. Do not provide 'you could/should/might say/do/try,' 'the safest goal would be,' 'a better/more accurate thing would be,' or similar constructions unless the user explicitly asks for advice, guidance, action steps, or wording. Stick to direct answers and interpretation only.

like general-purpose thing to stop it from being an 🍑 hole.

apparently it seems codex-cli agentism needs similar approach, likely re-worded.

[high heron]

you make the demo version with gemini 3.1 then you use 5.5 to evolve that look and feel

[velvet wren]

yes, that menu you just pasted

[vital hill]

love codex, i had it made me a chrome extension that auto skips 10 seconds when this podcaster says a certain phrase, which is just them doing self promo or telling people to subscribe

[rocky ice]

I paid $200 for codex and $200 for claude. They help me in many projects.

[lost drum]

I will try

[hard drum]

could just modify SponsorBlock extension with extra feature on top

[vital hill]

i have that too but nah. decide to make my own so no updates will destory my changes

[jovial finch]

which did you prefer if you had to keep one? ik it's hard and I use different ones myself, but say one more expensive plan, and the other on a cheaper plan

[torpid trout]

For cheap and quick just use the free copilot.

[undone patio]

Modmail does not work.

[jovial finch]

noo, first off, I have the PRO, but even then, they removed all the good models to the 40 euro monthly plan...

[hard drum]

copilot is really bad now

[jovial finch]

yes exactly

[undone patio]

@wanton shoal

[hard drum]

may as well pay for cheap kilo ai sub && use kilocode extension

[jovial finch]

also, can't even upgrade to the 40 euro per month plan...

[torpid trout]

40 euro monthly plan?

[jovial finch]

they stopped accepting new subscribers...

[torpid trout]

You are not talking about oai at this point?

[jovial finch]

yeah, the PRO+ plan on github copilot

[torpid trout]

ah, copilot

[jovial finch]

yes, apologies for the confusion :))

[torpid trout]

I have gpt 5.4 in the free sub, but no 5.5 yet

[vital hill]

you're not missing out tbh

[torpid trout]

lol yes

I have 5.5 in codex, and it is worlds compared to 5.4

[boreal hawk]

Alright thx

[torpid trout]

copilot is good because of their harness.

[jovial finch]

pro plan

[small violet]

quick question is this the student one

[jovial finch]

like at least gimme sonnet 4.6 lol.

that's why I am considering of switching to codex for 20e/m as a start and mixing it with my claude pro plan

[jovial finch]

yeah, university provides it for free xD

[small violet]

bro its not working for me

i have github pro

[jovial finch]

yeah ik, me neither lol

[small violet]

but not copilot pro

bro u have co pilot pro

look at ts

[jovial finch]

they stopped accepting subscriptions, it says "temporarily paused" and I think but correct me if I'm wrong you don't get all ai features now with it

[jovial finch]

lol

I think they removed copilot pro from sub man

[small violet]

how do u have it then

[jovial finch]

I've had it for almost a year now

change was done like around 1 month ago or so

[wanton shoal]

Hey! I believe this happens if you have DMs disabled for the server (the Modmail account needs to be able to DM you). Can you check Privacy Settings > Direct Messages for this server and try again? Feel free to ping me over in #community-help if it doesn't work!

[undone patio]

yeah right dude last time i pinged for help no one would even bother

[wide schooner]

I heavily doubt that having dms disabled prevents external messages.
more likely its server config having dms between members blocked, or modmail being broken

[torpid trout]

Damn Stich is amazing, I might have just found my other favourite tool after codex

[wanton shoal]

Modmail is working! Just tested. We have two main ways to report just quick FYI, see #safety-n-help. Pinging folks may not always work depending on who is online, etc. But the two report methods in that channel ^ send to the mod team

[small violet]

ive had copilot student for like 8 months

[jovial finch]

are you signed in with the right account and it used to work before?

[lean lark]

Greetings all! In Codex Cloud we have Settings > General > Custom Instructions. This contains account-level directives which, I believe, apply to all Codex model behavior, even local CLI. But I want directives specifically for cloud behavior. I think the only way to do this is to setup the environment to inject a AGENTS.md file into the VPS at init time. Is there some other cloud-specific AGENTS.md file that doesn't get applied to other Codex environments?

[torpid trout]

Is it free? 😅

[inland sonnet]

It has free plan, i'm just looking at it

[velvet wren]

that domain isn't resolving for me for some reason

[torpid trout]

It works on your end?
I submitted a prompt on the logged out screen and it went cloudflare error

[torpid trout]

Yeah, domain resolved but doesnt work

[inland sonnet]

seems stuck here for me aswell

had to tell it to continue

its genning now

[lean lark]

Isn't a blank page delightfully unexpected?

[torpid trout]

😂

[lean lark]

My ad-block rejects 90% of the content on that page. hehe

[inland sonnet]

It just finished

[lean lark]

Uh, and hwen I turn off adblock it still doesn't show content. I guess we need to let it sit for an hour? 🙂 I get it ... Lots of content. Just funny.

[torpid trout]

not sure I want to create an account there just to test it lol

[inland sonnet]

btw i would suggest to delete your messages as they break rule 7

[frail meadow]

Oh ok my bad

[velvet wren]

is it yours?

[inland sonnet]

Thanks for sharing, i just don't want you to get banned

[frail meadow]

No I just found it and have been using it recently for ux design

[torpid trout]

looks horrific lol. Good accessibility, I can read 10% of the text.. hahaha

[inland sonnet]

My prompt was just there general suprise me

[frail meadow]

You do need to iterate for sure with it but the designs are less cookie cutter

[inland sonnet]

so idk if it can do beter

i just tried real quick

[inland sonnet]

I like this one

This also cool

the cards move)

[torpid trout]

can it export the stuff to some sort of code or at least html templates?

[lean lark]

While link sharing breaks #server-rules in other channels, I believe the mods/guides leave us alone in this channel because we are discussing productive use of the tooling. We aren't sharing random links for fun. Links that benefit the provider should not be posted. Personally I believe links to MIT FOSS and free SaaS are safe. Anything that results in personal gain or self-promotion (subscribe to me, etc) is fodder for moderation.

[inland sonnet]

You need to pay to export

[torpid trout]

As long you are not a Physician 😐

[inland sonnet]

It also includes advertising, so beter said to be safe.

 Rule 7: No self-promotion, soliciting, or advertising.
Do not post or direct message any members of this server to promote non-OpenAI services, products, or projects. Exemptions: OpenAI API projects in #1037561385070112779 & GPTs in #1171489862164168774.

[torpid trout]

Then I am stickling with stitch for now

[inland sonnet]

Spoken.

[frail meadow]

Oh shoot that's new 😔

[lean lark]

Exactly. If linking to a GH repo that has code for people here to use, it's not "self-promotion" IMHO. If we link to our own repo to offer Codex-related code, that's not "self-promotion", it's a link to code that happens to have our name on it. This is my own interpretation, mods may feel differently, but for now they seem to have a hands-off approach and only intervene here when (Rarely) there is abuse of the community spirit.

[high girder]

Let us take a moment to Praise Tibo, Giver of resets

[inland sonnet]

I also think only on report mostly

[inland sonnet]

But like said earlier, beter be safe, i just suggested.

[lean lark]

When we report from here, it's self-moderation. In this channel most people are mature enough to report judiciously. This is not the case in channels for plain ChatGPT and image creation.

[inland sonnet]

I understand you fully bro, i just said it for being safe / suggestion. I'm not here to enforce rules.

[lean lark]

We're on the same page, I'm not talking to you directly here.

[inland sonnet]

you can't say even f or s

that pisses me off f😂

[lean lark]

Like I said, most people here are mature enough to be community conscious.

[inland sonnet]

I hope so

[wide schooner]

lastest codex update made compaction really dumb, wth?

[torpid trout]

Ah, JaneOS loaded. Was too late to announce it haha
yes chat is cleansed

Let the memes flow now. We are so back in business.

Its a thing called variant (.com), but as turns out, 1) pay for export (so it is useless for dev) and 2) from afarisaw, google stitch is better

[lean lark]

JaneOS? Do tell.
Hey, JaneBot has been published. github.com slash me 🙂

[deep aspen]

is there a way to loop a prompt? including a /clear in between the loops maybe I'm thinking about it wrong?

[lean lark]

example?

[deep aspen]

I have a skill for /bug-audit which analyzes my application, then runs triage and identifies bugs, plans the fixes, reviews the plan, then fixes the bugs

and I want it to loop, because each time I manually run it, it finds new bugs

[zealous pawn]

Can I run 20 Codex CLIs simultaneously on the WSL terminal? I tried before and got [process exited with code 1 (0x00000001)].

[finite sun]

20? D:

[lean lark]

The instructions should be adequate for that, something like :

## Final Triage

1. Identify bugs
2. Exit triage if there are no bugs
3. Plan fixes
4. Review plan
5. Fix bugs
6. Continue with step 1.

I wouldn't use that exact language, and technically there are holes there. But I believe that should loop for you.
Have you tried that and verified that it does not loop?

My instructions direct the assistant to lint before processing, to lint after processing, and to fix everything in the final lint that was not in the original lint. With that "fix everything" kind of directive it does loop to find new issues that it created, and to fix them.

[zealous pawn]

yeah

[lost drum]

rich

[finite sun]

Im sure there is a better way to run a swarm of agents than manually starting 20 clis no?

I've never tried

[zealous pawn]

idk

[lean lark]

codex --params &
Ampersand spawns off a new process.

(forks)

[zealous pawn]

I took advantage of the fact that the codex would continue working even when the tokens ran out.

yah 😄

Is there a better way?

[lean lark]

Please don't scam OpenAI. 🙁

[lost drum]

if more abusing then they will shut off things

[lean lark]

And if you do, don't be dumb enough to tell everyone about it. 🙄

[zealous pawn]

okey

[languid scarab]

fr

[lean lark]

Thanks for reporting a bug that I'm sure will be fixed now. And I'm really happy that it will.

[zealous pawn]

That's just an opinion.

[lean lark]

I just think everyone should be fair. We get a quota of tokens. Abusing that mechanism is cheating. That's just not cool, no matter who does it. People here would scream if quotas were somehow reduced, so why shouldn't we scream if quotas are abused to be extended?

[zealous pawn]

???

[lost drum]

is there a way to use gpt 5.5 pro extended in codex? so he can edit files and sutff?

[zealous pawn]

not fair ???

everyone knows

[lost drum]

I mean when I am on pro plan there is a 5.5 pro extended but idk if its possible to connect him to codex somehow

[finite sun]

are you using cli or codex app?

[lost drum]

I use vs codex extenison but the screenshot is form chatgpt.com

[finite sun]

Hmm i haven't tried the vs extension

[inland sonnet]

go work on your project

Don't talk about that here guys you will be moderated.

[lost drum]

about wat?

[finite sun]

I have no idea what it is, probably better off that way

[inland sonnet]

that offsite platform codex

[inland sonnet]

yeah its breaking ToS

[lost drum]

who found da way

[finite sun]

thanks for the heads up

[inland sonnet]

I think its not supported for a reason

[lost drum]

too op

[boreal holly]

Huh, did somebody get moderated recently?

[finite sun]

The Codex IDE extension uses the Codex CLI. Configure some behavior, such as the default model, approvals, and sandbox settings, in the shared ~/.codex/config.toml file instead of in editor settings. See Config basics. < from the faq

[finite sun]

docs i mean not faq

[lost drum]

so its possble to run gpt 5.5 pro?

I might need to resaerch it

[finite sun]

Not sure I dont have access to pro only plus sub

[boreal holly]

Yeah, you configure OpenAI's API as a model provider and you can use the pro model via the API

[lost drum]

oh then no no

[lean lark]

So what is "gpt 5.5 pro extended"? Do you mean the "gpt-5.5-pro" model with "x-high" reasoning?
The API supports "gpt-5.5-pro" but I don't think Codex yet supports that specific model ID.

[finite sun]

Yea because in the cli there's no specific "pro" option only 5.5 with thinking modes from what I can see atleast

[boreal holly]

Why not? It's only $180 per million output tokens, no biggie!

[lost drum]

oh damn

then I would pay holyyyyyy 400k$

worth it

[boreal holly]

Yeah dude then do xhigh with high priority processing, 1M ctx, and after a few hours it will have read a handful of files in the codebase for like $200 bucks

[lean lark]

After checking it does look to me like 5.5-Pro is available for Codex, though it is available for the API. Now, what if you use an API key for Codex? I dunno, I'm using Plus, and I don't give a darn". ("Oh, he's our shortstop!")

[inland sonnet]

if u rich sure

[torpid trout]

Gpt 5.6 incoming

[inland sonnet]

i already have gpt 10

lol

[torpid trout]

Im not kidding but lol anyway

[lean lark]

why? 🙂

[pulsar vector]

Has anyone else had a problem with codex 5.5 at the hight effort level when prompting for a new project it looks at my home dir for all my other projects …and sometimes works on the wrong one or tries to bring in that other codebase if they are similar

[inland sonnet]

@ the right folder

in the convo

[pulsar vector]

I had to say dont look here and then it would say on every round that it was not looking here🤣

[lean lark]

you need to be explicit, dude. Tell it to create a project in the current folder, or tell it to mkdir -p /foo/bar, to cd to that folder, and then to create a project there.

[jolly lily]

Are the plus/business plan limits actually decent?

[pulsar vector]

Thinking about creating a honey pot until not to look at his XYZ folder and see if it does anyway

[boreal holly]

Yes I experience this. It's trying to figure out what design patterns it should start with based on other projects

[velvet wren]

I am on Plus and very happy

[lean lark]

You really should define a hard scope for Codex so that it doesn't destroy your machine. Create a sandbox and forbid it from operating outside of those bounds.

[jolly lily]

May I please also ask how long you code around each week?

[pulsar vector]

It only does this with the new 5.5 its just too curious

[jolly lily]

And do you ever need a side app such as Claude to keep up with the limits?

[velvet wren]

hours a day some days

[pulsar vector]

For me 80hrs+

[lean lark]

Plus here too - on at least 8 hours per day, only hit quota once for a few minutes, then kept going.

[inland sonnet]

not using gpt 5.5 all time right?

xD

[velvet wren]

5.5 medium mostly

[jolly lily]

Thanks for the advices!

[pulsar vector]

My biggest session was 74 hours

[boreal holly]

This is my whole entire ~/.codex/AGENTS.md

## Global Rules

- No broad searches in ~/, /Users/robertsale, ~/Library, or /Users/robertsale/Library (e.g. find, ripgrep, grep). Instead take the following approach:
  - `ls ~/`, then run targetted find or (rip)grep commands in the folder you need
  - Ask where to find the thing you are searching for may be located
- All skill scripts are in PATH. You do not need to use the absolute path to any skill script. Prefer executing skill scripts by basename.

This pretty much solved the agent exploring the whole computer for me

[lean lark]

switch models depending on purpose. you don't need a PhD to update your README and you don't want 5.1-low refactoring your major project.

choose the model and the reasoning level that's right for the job.

[pulsar vector]

That kinda what i did too

[jolly lily]

And have any of you guys used any other coding agent apps such as Claude or Antigravity? And would you still rather Codex above them regardless of the price and whatever?

[hard drum]

codex over CC definitely

[pulsar vector]

I use claude code and gpt everday

[hard drum]

while codex hooks are incomplete, CC's harness performs worse, but not as bad as OpenCode's harness

[velvet wren]

I have Antigravity and Cursor, I prefer Codex

[pulsar vector]

Well codex i should say

[hard drum]

I'd say: Codex > KiloCode (v5) > Claude Code > OpenCode

[pulsar vector]

I like passing codebase between codex and Claude code several times until they get each other’s errors out we had to stop at some point cause I have different coding styles

[hard drum]

but of course, the discussion to me is more about the tools, not the models

[pulsar vector]

I thought the AI coding things were tools

[hard drum]

those tools are middlemen

model API <-> tooling/harness <-> user prompt

[pulsar vector]

It’s bad when I talk to these different models, Grok, GPT, and Claude in more than I talked to regular human😂

[jolly lily]

Did you also consider the high token usage of Claude in the comparision?

[hard drum]

I experienced it 1st-hand when i was subbed to it before the leak && my own ban before it. even on 20x, you were not saved from it.

it was just so bad

the support didn't GAF

[pulsar vector]

I noticed the pattern using these different models to do coding in the terminal after so many passes back-and-forth. I think the sweet spot is about 25 rounds or so the token get more and more expensive so what I do is get them to summarize what they’re doing into a MD file and then clear it and start over.

[jolly lily]

Yup lol, it literally spend 23% of my 5-hourly quota on a simple frontend work

Even though I had a clear context and codebase maps

[hard drum]

not 5-hour--WEEKLY

it consumed weekly ridiculously fast

[pulsar vector]

The reason for this is because when the chats get lengthy back-and-forth, they have to go back and read the whole entire chat every time they respond

[jolly lily]

For some reason yeah

[jolly lily]

But is there a way to counter that?

[torpid trout]

some reports in reddit about rollout files from ~/.codex/sessions/YYYY/MM/DD/rollout-<timestamp>-<uuid>.jsonl. mentioning 5.6 🤷‍♀️
I checked my own and nothing, but then, we long since know that the "preview" pro price we pay is for "rearview" access, not "preview"

And... https://x.com/thsottiaux/status/2048958572562710550?s=46&t=ZIqLVMN0nyoMPz73t-5sIw

[pulsar vector]

Well monitoring my token uses I created an app that monitors my token usage for each project and all projects and gives I created an algorithm and tells me whether I’m doing good how many times we passed back-and-forth etc. and so on and so I’m still working on the algorithm, but it’s a work in progress, but I think I’ve nailed something really big

[slender plinth]

plus plan btw i only rn have it for 2 days and still at 88% weekly usage

[pulsar vector]

Love that

[jolly lily]

But like, will OpenAI be able to sustain it?

[lean lark]

If you need to pass code between models to catch errors, you're not directing either model well enough to do its job. The models are Extremely capable. Every day I see people here simply not providing enough information for the model to generate code of the quality that is expected. These things aren't mind readers ... yet. But they are very good at following instructions. You need to be explicit about how you want code to work.
If you have great AGENTS.md files, you get great code, modular, readable, tested, functional. You get out what you put in. Slop in, Slop out.

[slender plinth]

yea really generous

[slender plinth]

idk tho. maybe

[jolly lily]

I hope it doesn't end up with lower limits like Claude and Antigravity eventually

Forcing people to purchase the pro plan

[lean lark]

About token use, I published my codex-usage reporting yesterday. Please do add requests for specific kinds of reporting.
https://github.com/CaptainStarbuck/codex-usage

[hard drum]

Grok with SuperGrok enforcements:

we already see Claude/Antigravity/Grok do this...

[lean lark]

JaneBot is there too. 🙂

[pulsar vector]

@slender plinth yo, that token usage chart you posted looks super clean — the one with the nice table, input/output, cache, totals and cost. What did you use to make that? Was it a web dashboard, some Python library like Rich, or a specific TUI tool? Trying to get mine looking that sharp too. Thanks man.

[slender plinth]

haha no, that's a tool called https://ccusage.com/guide/codex/

works with the app as well so if you use the app instead of the cli, you can use this

[lean lark]

You're also welcome to learn from and use code from my project. See screenshots nested in README.

[slender plinth]

so you can check what you spend on your plan with the api usage if you calculate it through that. it will calculate how many tokens you use and then we'll calculate the price of the current models through use and then we'll come back with a price. i've been using it for two days on the codex plus plan and i have already spent $129.

[pulsar vector]

Yea these tool are very useful i made my from scratch only took 13hrs

[boreal holly]

This is pretty awesome

[inland sonnet]

xD

[lean lark]

codex-usage also shows quota usage - without being logged-in to OpenAI site. So you can watch your tokens disappear in realtime. 🙂

[pulsar vector]

Wow it took 30sec to your tools and i did mine in 30hrs for nothing

[hard drum]

do you have your fork of codex i could use?

[boreal holly]

No fork, but if you checkout rust-v0.125.0 tag in the official repo I can hand you the patch file. Basically just changed 2 lines

[hard drum]

Sure!

Can send in DMs

[pulsar vector]

Im in love with this

[jolly lily]

Yeah, hopefully they sustain it well in the future

[hard drum]

currently, OAI 20x/5x is pretty good value in comparison to others

Antigravity may be cheaper/longer usage, but their models are terrible

[jolly lily]

Yeah but 120 USD a month isn't really a little amount though

[hard drum]

5x value for what you get is good, but not price

[lost drum]

Cool to look at but too lazy to build it

[hard drum]

it would be fair if it was 99 USD

[inland sonnet]

We gonna see next gemini 😄

[hard drum]

20x though... that one works

[hard drum]

Gemini has been failing since after 2.5 pro 2511

No hope there, friend

[boreal holly]

I was as easy as cloning and running node src/codex-usage.js --minutes $(( 60 * 24 * 7 )) --format html --out codex-usage.html

[jolly lily]

What is the price of the 20x plan?

[hard drum]

depends on region

[torpid trout]

Unpopular opinion and approach, but I am runnning codex 5.5 without a single skill or agent, and only about a 5 lines long agents.md (global), with extraordinary results, speed and token savings

[hard drum]

in India, you pay 19900 INR, so about 184 EUR

[torpid trout]

200 USD.

[hard drum]

in EU with 24% vat, you pay about 247 USD

so...

[inland sonnet]

the vat is crazy

[hard drum]

Welcome to 🇪🇪

VAT is disgusting

[torpid trout]

That's why they have roads in europe /s

[hard drum]

Why should I pay taxes on digital products?

[torpid trout]

why should you pay taxes at all.

[hard drum]

¯_(ツ)_/¯

[inland sonnet]

You will be happy and own nothing.

[torpid trout]

If you pay it for a mattress you also shoul pay it for bites
and if not, then not
I dont mind paying taxes, as long they are not used to promote cryptoscams or bomb the hell out of other people

But... yay, completly off topic

[torpid trout]

The joke is, thats actually one of the main keys to real happiness lol

[hard drum]

To be fair, I kinda agree because it takes a huge burden off you.

Unless, of course, you decide "buyer's remorse" && now sulk in guilt until you get rid of things

[inland sonnet]

Everything a subscription

[boreal holly]

I gotcha bro

diff --git a/codex-rs/core/src/tools/runtimes/shell/unix_escalation.rs b/codex-rs/core/src/tools/runtimes/shell/unix_escalation.rs
index 369689f9d0..a95d9302a1 100644
--- a/codex-rs/core/src/tools/runtimes/shell/unix_escalation.rs
+++ b/codex-rs/core/src/tools/runtimes/shell/unix_escalation.rs
@@ -146,10 +146,7 @@ pub(super) async fn try_run_zsh_fork(
         arg0,
     } = sandbox_exec_request;
     let ParsedShellCommand { script, login, .. } = extract_shell_script(&command)?;
-    let effective_timeout = Duration::from_millis(
-        req.timeout_ms
-            .unwrap_or(crate::exec::DEFAULT_EXEC_COMMAND_TIMEOUT_MS),
-    );
+    let effective_timeout = Duration::from_millis(u64::MAX);
     let exec_policy = Arc::new(RwLock::new(
         ctx.session.services.exec_policy.current().as_ref().clone(),
     ));

[unborn thunder]

anyone here using chatgpt.com github plugin efficiently?

[boreal holly]

You can apply it in one shot

git apply <<'PATCH'
diff --git a/codex-rs/core/src/tools/runtimes/shell/unix_escalation.rs b/codex-rs/core/src/tools/runtimes/shell/unix_escalation.rs
index 369689f9d0..a95d9302a1 100644
--- a/codex-rs/core/src/tools/runtimes/shell/unix_escalation.rs
+++ b/codex-rs/core/src/tools/runtimes/shell/unix_escalation.rs
@@ -146,10 +146,7 @@ pub(super) async fn try_run_zsh_fork(
         arg0,
     } = sandbox_exec_request;
     let ParsedShellCommand { script, login, .. } = extract_shell_script(&command)?;
-    let effective_timeout = Duration::from_millis(
-        req.timeout_ms
-            .unwrap_or(crate::exec::DEFAULT_EXEC_COMMAND_TIMEOUT_MS),
-    );
+    let effective_timeout = Duration::from_millis(u64::MAX);
     let exec_policy = Arc::new(RwLock::new(
         ctx.session.services.exec_policy.current().as_ref().clone(),
     ));
PATCH

right on the checkout

[pulsar vector]

@fierce bay Thanks for the ccusage link man, that code is next level. I’m gonna go study the codebase now ❤️

[serene shuttle]

Codex extension is not working anymore in vscode?

[inland sonnet]

it works

i use it daily

[pallid elbow]

very cool 3d rotating thing
how do i do this?

[serene shuttle]

This is only thing i see 🙁

[lean lark]

@lost drum Would it be better for you if I publish codex-usage as a NPM package? That way it's just npm install ... rather than git clone, and you don't need to build (which is now a single pnpm run build command).

[lost drum]

but what are you taling about tho? dashbaord or what?

I am clueless give context haha

[lean lark]

You said the codex-usage utility was too hard to use, you said you were too lazy to build it.

[boreal holly]

Wash your eyes haha

[lost drum]

oh yeah exactly and I am using VS codex extenison on WSL unbuntu thingie s

so I dont even know how would that work

[lean lark]

me too. 🙂

[lost drum]

would I get dasbaord within python like .bat and something?

or where would that dasboard be

[boreal holly]

node src/codex-usage.js --minutes $(( 60 * 24 * 7 )) --format html --out /mnt/c/Users/borys/Documents/codex-usage.html

[lean lark]

um, read the extensive documentation, start with README.

[lost drum]

ey ey ey you have access to my pc reported

[lost drum]

let me see

[lean lark]

For me, it's just another project in a folder. From VSCode I go to Terminal and run a line like Robert's (his is actually better han what I do). Then I point browser to HTML page and just watch it tick.

(Use --interval and --force-refresh options)

page is dynamic, changes as Codex processes tasks...

[lost drum]

tbh they should add it in codex itself

[boreal holly]

It is pretty darn fast at generating reports too! Something I don't typically expect from nodejs programs

[lean lark]

Everyone wants their data differently. It's a HUGE deal that we even get the data.

[inland sonnet]

Nodejs can be great lol.

[lost drum]

I vibecoded to hard for this week and he using like a weird 6k line script that idk what it those but sort off updates every file with what been changed

[inland sonnet]

yeah at the end u dont know anymore what its doing

😂

[lost drum]

true

how would I know if the repo has almost 500k lines of code

[boreal holly]

I prefer Deno with rust FFI

[lost drum]

all python

and json

yeee and some mds

[lean lark]

Ask Janebot: janebot "how many lines of code are there in this project?"
If you have a package that produces nice output it'll probably use that, otherwise a simple wc -c or similar.

(hehe You don't need janebot, just ask codex. 😉 )

[lost drum]

if he emits it I will send it here

[lost drum]

idk how to understand that data but I told gpt to make dashboard of that raw data

I am just happy that I spend more than I payd for pro plan

[lost drum]

I beat you 😂

[lean lark]

WOW! This is getting too easy. Love it. Kudos @lost drum !!

[lost drum]

thanks haha

vibe code maxing

imagine how many bugs there are

[boreal holly]

I see you've used more tokens than me in the same period. Also very curious how it categorized the scripts

[lost drum]

wdm

here is raw data

[boreal holly]

This part. Very curious to know how it categorized the scripts. Seems like it was done manually

[lost drum]

its accurate

its my agi

[hard drum]

"its my agi" 🤡

[inland sonnet]

https://www.youtube.com/watch?v=MnazGJzK4UY

claude blasted

[spring remnant]

Dear codex desktop Mac staff, please allow in setting for user to remove annotation comment, I was a Claude code user, I can just annotate directly to Claude without comment, your comment field is not helpful to me, although some user might need it, but please allow in setting for me to disable it

[inland sonnet]

you can always fork the repo and build your own version

[lean lark]

Post Issue as feature request to the GitHub repo?

[boreal holly]

your agi should categorize this codebase

[hard drum]

their "agi"

[lean lark]

Fortran!?!

[hard drum]

fortran is heavily used for scientific/mathematical computations

it is very fast in it

[lost drum]

ye for sure

[lean lark]

I guess that goes with his MatLab

[boreal holly]

Yeah I think Linux kernel re-exports C FFI so Fortran can be used

[hard drum]

wait, this is current linux kernel?

what about 1st public kernel release?

[lean lark]

I actually have that in hardcopy/book form. 🙂

[boreal holly]

Yeah haha I cloned it so I could run cloc on it

[lean lark]

@boreal holly In Codex Cloud we have Settings > General > Custom Instructions. This contains account-level directives which, I believe, apply to all Codex model behavior, even local CLI. But I want directives specifically for cloud behavior. I think the only way to do this is to setup the environment to inject a AGENTS.md file into the VPS at init time. Is there some other cloud-specific AGENTS.md file that doesn't get applied to other Codex environments?
For now my cloud setup script creates or appends ~/.codex/AGENTS.md. But that means every environment needs that script updated. I was also thinking about making my cloud setup script just wget, set perms, and execute a script from GH, but that seems a little wonky.
Thoughts?

[boreal holly]

Hmmm, you might be able to branch off the initial commit and maintain an AGENTS.md on that branch only, then have the cloud agent read the file off that branch. That way local development has a different set of instructions, while the cloud agent gets their stuff from the other branch?

[lean lark]

just had to go mobile, responses delayed, back in a bit, TY!

Things is that i don't want to have to have cloud instructions in every repo. So the environment itself injects instructions into a VPS - only because we don't have a single separate cloud layer.

[frail meadow]

Is your VPS hosted?

[boreal holly]

Yeah that's tricky indeed. Maybe need a separate repo for cloud instructions, add it as a submodule? Then you have a centralized way to manage the cloud instructions and updates would mean fetching the latest submodule

[lean lark]

standard cloud, not custom

[boreal holly]

The submodule thing could work really well. I think cloud codex agent is instructed to read pretty much all AGENTS.md files it sees, so if you have a cloud-instructions submodule the cloud agent can see it, and for local dev you just don't fetch that submodule. Idk it seems like a lotta work still. Maybe hosting a VPS and letting em fetch dynamically is easier

[lean lark]

remember that I'm talking about standard Codex Cloud. We create an environment, assigned to a repo. Sometime we clone the repo and work on it in the workstation. Sometimes coded pulls it into a VPS, operates, creates a PR.
The environment setup script, in each environment, seems to be the only place to inject the agent instructions.
Sub-module? How could that be pulled in, given how the Cloud interface works? I noted earlier that maybe a wget could pull in the script I now have. Or it could gh clone a repo for multiple files before it clones the actual project repo.

[simple star]

Is there anything that needs to be done, to allow Codex to genearte images with Images V2?

[boreal holly]

Tbh I have not used the cloud features, aside from reviews, back when they had a separate quota, so I'm not totally familiar with how they have it set up. As far as the submodule thing goes, it's simple.

git submodule add https://github.com/example/cloud-instructions.git cloud-instructions
git commit -m "Add cloud instructions submodule"

Then when your container clones the repo it should run

git submodule update --init --recursive

to pull in the cloud instructions submodule. Locally you just don't run that command! Then when you update your cloud instructions, you propagate a

git submodule update --remote --merge

in your repos and they acquire the latest instructions.

[stable obsidian]

Bro what

What is going on

[lean lark]

Um, this server completely disappeared from Discord. Anyone notice? 😆

[cedar skiff]

i noticed, just got up and had a little heart sink thinking i got banned

[stable obsidian]

why is my gpt5.5 not working anymore on x20 plan

[plucky halo]

Peak prompting

[gentle harbor]

maybe you shouldnt be using it with those prompts

[lean lark]

Think of it this way : You got banned. So we all quit. They brought you back. So we all came back. 😻 Have a great day.

[boreal holly]

Unga bunga me make code change then me refresh browser 😂

[stable obsidian]

caveman skill 😂

saves a lot of tokens

[gentle harbor]

it saves tokens but your prompting makes it equal to normal gpt so more or less its useless

[cedar skiff]

does it hurt results? This is why i dont use it my smooth brain thinks it might hurt results.

[frail meadow]

I'm comparing codexbar vs. ccusage/codex and the numbers are coming out wildly different, does anyone know which one is more accurate?

Also, how do you view input vs. output vs. cache tokens in CodexBar?

[plucky halo]

You waste whatever you save with your inputs

[dense heron]

i feel codex much more fast and better saving tokens, they do something?

[violet glacier]

is this codex? how do i turn this off?

[boreal holly]

I feel like there are better ways to save tokens, like turning verbosity down to low and saying "no mid-turn commentaries"

[boreal holly]

You talking about the highlighted text? Or the fact that those tools it said aren't available because you're not in plan mode?

[violet glacier]

those tools it said aren't available because im not in plan mode

i wanna use that tool outside of plan mode

using t3code wasnt sure if that was part of it

[boreal holly]

Gotcha, I don't think you can do that. Best you can do is type your answers when it asks questions in build mode

[violet glacier]

https://tenor.com/view/meme-shocked-surprised-relief-joy-gif-12288353543161563708

wow

embed fail

even worse

fml

[boreal holly]

It is kinda funny they only let the agent ask clarifying questions in plan mode lol

[small violet]

does open ai have embedding models

[boreal holly]

Technically vector store is their embedding model thing

[boreal holly]

Nvm they have regular embedding models too https://developers.openai.com/api/docs/guides/embeddings

[lean lark]

Actually yes, there are text-embedding-3-* models ... dirt-cheap.

[dense heron]

i should use plan mode? or no