codex-discussions

[ember spire]

5.5 is a nice model but the codex app is the thing I am enjoying most after switching from claude code

[simple star]

Codex really needs to be able to stop, to ask questions, even outside of "plan"

[chrome raven]

it is just funny how you guys wanted it not to ask questions

[ember spire]

you should be able to build that into your skill file

[simple star]

we?

[ember spire]

ive barely used plan mode in codex yet and get plenty of questions 🙂

not via a question tool though if you mean that

[chrome raven]

lots of complains back then

[simple star]

You will not have heard one from me. Conversation is necessary to get good results

Thats why Plan mode is great

[simple star]

@lean lark Thats gotta be fake

[cedar skiff]

they closed the coding gap at about 5.2 imo. They were better at compaction and coding at that point, but 5.2 is really tedious to prompt because it doesnt get semantics this, that, it, this etc, it would pick the wrong context of those words and do the wrong thing. Like you would say lets do this and it would pick the wrong this same with that etc. It also never changed its conversation to match the level of the user. It was technical and hard to understand if you didn't know SWE terminology and architecture. So despite being better at code it wasn't seen that way because vibe coders and weaker devs still had trouble understanding it and prompting it. Claude matched the vocabulary of the user it is working with and spoke at their level, and understands the semantics this, that and other etc. So it was easier to understand and easier to prompt, but not better at coding. But people felt it was better because of its better communication skills. gpt 5.5 is a huge step on the conversation skills, it still isnt quit claude level of just getting it. But now the gap is small enough that the coding difference can shine past the language barrier. Fun fact as a coder alone 5.5 is about the same level as 5.3.

Oh and 5.2 was REALLY slow

[lean lark]

What you're citing is exactly the issue I had with that v4 project where I was trying to drive its semantic understanding. In 5.2 I couldn't continue, the project stayed on the shelf. Now we can do it.

[vital hill]

Coding while I am on a Lyft ride. What a time to be alive

[torpid trout]

you aint coding tho, you are yapping in discord 🤣

[ember spire]

last time i used a macos app written by openai before this week was last june, and it was the chatgpt app. it was such a pile of trash i had low expectations on codex

[worthy furnace]

Asking a machine to code for you

[cedar skiff]

Have you tried codex app?

I love it

[vital hill]

I’m tired boss

[worthy furnace]

Yeah who codes by hand anymore

[ember spire]

yeah it's so good, i'm wondering how long it's been like this

[vital hill]

Ima be at the bar Soon™

Gonna be coding there too

[cedar skiff]

I started using in 5.3 but it wasnt as good

[boreal holly]

Folks who are on the verge of or already have been promoted from SWE to customer

[vital hill]

I did ooga booga code my open source project years ago

But now I wanna vibe it to golang

[ember spire]

the chatgpt app is good now too 😄

[boreal holly]

Yeah me too. Also got a BS in Comp Sci like 2 months after GPT 3.5 came out 🤡

[lean lark]

I dunno about "Fun fact" but the numbers are good for 5.3-codex and 5.5: https://chatgpt.com/share/6a07ad6b-eb68-83e8-a81d-686bad359b65

[vital hill]

Codex 5.3 medium is my goto

[cedar skiff]

it legit shows it first line in that chat o.0

1.8 points diff

It's better at everything else

[lean lark]

Agreed on "The shared overlap is narrower than the pages make it feel" but all of these models are progressing in a similar "evolutionary" manner, not "Revolutionary".

[signal ermine]

Can someone explain Codex to me and what I could use it for?

[cedar skiff]

The 5.5 page doesnt even chart swe-pro, they mention it mildly in a sentence. If the diff was big they would have it front and centre.
I wrote about this another time:
#codex-discussions message
This one is interesting as well
https://www.reddit.com/r/codex/comments/1t5ipjd/gpt55_xhigh_is_the_strongest_coding_agent_weve/
A lot of people coming to the same conclusion on their own terms and then talking about it in public forums and you tube vids etc.

[lean lark]

And I love how intelligent Codex is getting....

[lean lark]

Where ChatGPT presents a series of questions and answers, prompts and responses, Codex actually does things in the real world. It reads and writes files, assembles data from different places, kicks off other processes. It's active, not passive.
Howzzat for an elevator pitch?

[signal ermine]

Interesting

[plush harbor]

write code mostly, but can do much more than that

[bright swift]

Do you have the new memories thingy enabled?

Or whatever its called

[cedar skiff]

I'm not sure i want my context thinking about pizza and happy hour before it updates documentation

[bright swift]

Yea thats why I dont use memories, but I have a session retrospective skill that i sometimes run. It puts stuff directly in tracked repo docs though, not hidden in some memories folder

[unique spade]

same haha

[rocky fog]

yeah you would probably prefer bananas 😄

[chrome raven]

a reset is coming

[rocky fog]

oh boy

[plush harbor]

and I'm over here at 85% left with 4 days

[hard drum]

hope

i hit my weekly today

[hard drum]

😮

[solemn sedge]

codex mobile is supported for wins now?

I heard somebody has got it with his wins

I try to connect codex mobile with my wins but theres no reaction

maybe just skill issues?

[cedar skiff]

There are some ppl who mentioned how to do it for winodws in the chat, there are feature flags you have to add to the config.toml

[rocky fog]

check/try
#codex-discussions message
#codex-discussions message

[solemn sedge]

OK

thanks

doesnt work in my situation

thanks anyway

[nocturne folio]

gpt5.5 fixes itself whilst performing worse

[lunar crypt]

Idk 5.5 seems.. dense to me. Very much so. It makes rookie mistakes in coding and dumb decisions, often ignored instructions, or on the other hand listens to them way too literally.

In coding it feels like a regression to be honest. I still prefer 5.3 codex

[stuck warren]

is there a big difference between 5.5 an 5.1? we only have 4.1 or 5.1 avail and not sure what to expect if we update

[nocturne folio]

obviously

its huge

[quick geode]

Does anyone let the llm do 100% of the coding but you actually check for mistakes and test to make sure things work and is safe?

[torpid trout]

We don’t really test things around here, why would we? It’s not that we’ll use the stuff afterwards

[cedar skiff]

I mostly do this now days

I look at the code that matters but i dont write any of it

My projects have robust unit testing and integration testing

[quick geode]

😅 but do you let codex for example create and run the tests? I think for my project, after every feature i have it create tests, test locally and then test it on production (thankfully i got approved for the cyber stuff so it can actually do it)

[undone patio]

how many beers does it take to develop a 100M b2b saas

[cedar skiff]

codex writes all the tests

I curated and adjusted skills for writing them until they worked

[ornate jolt]

Codex loading insanely slow for anyone else?

Taking forever to respond and keep getting connection issues

[cedar skiff]

With out skills codex does lots of bad testing patterns, it wont use helpers and will reinvent the wheel in every test file, it will create low value tests, it will test implementation details, it will pass tests manually that are blocked and just tell you about it in a report, it will mistakenly update implementations to have tests pass.

[unique spade]

yup good skills are key

[lunar orbit]

anyone use Codex on Windows? Is the app best way to use?

[livid forum]

I’ve used CLI and have gotten way better results

[unique spade]

up to your preference, the core functionalities are the same across all ux surfaces

[vital hill]

Ima go touch grass guys

[cedar skiff]

that photo plays tricks with your mind

[vital hill]

How

[plush harbor]

codex just told me to use kubernetes if I hate myself enough :\

[velvet wren]

I hate kubernetes

[cedar skiff]

it looks like its on its side

[balmy olive]

What is the most unlikely thing you’ve done with Codex that you thought would be impossible?

[cedar skiff]

New motivation prompt for codex "Make no mistakes or the next project will be using kubernetes"

[polar grove]

Reset limits?????? Oohhhhh

U already know

[polar grove]

Fr? U think 5.3 is better. I would use 5.3 if it was good at steering messages. But it forgets instantly.

5.4 ig

[slim wave]

guys how get damn near infinite ratelimits

[hard drum]

i still don't know why i hit such fast weekly limits...

[slim wave]

I have plus n my shi going dowwn fast

[plush harbor]

I mean I didn't realise codex would go off, look at websites, parse them and then come back with opinions but that was just newbie surprises not codex being fancy

[hard drum]

i don't know what else am i supposed to do to reduce that

maybe anyone here knows what i gotta look for in local .codex?

[cedar skiff]

Mostly replace manual testers for flutter apps in mobile and web lanes.

I just didnt think it would be able to do it

[slim wave]

does any plan give very high limits ?

[plush harbor]

"surely there is something out there already so I don't have to do this myself" => codex goes off, finds the data I need on the ABS site, tidies it up, turns it into json, writes an editor, fills in an entire extra field from the next step of my documentation without being asked
me: "oh"

[vital hill]

I got muted lol

[slim wave]

guys

how to lower ratelimits with plus

js used 20%

[plush harbor]

I just used half my 5 hour limit on ... idk, something, it was very annoying anyway

[slim wave]

damn

I had infinite limits

ended tdy

I might need to buy the $100 a month plan atp

[plush harbor]

I need to keep codex away from drupal. Boy does it eat the quota

[hard drum]

me playing Spyro 1, later 2, while awaiting limit reset

[slim wave]

lol

[hard drum]

@cedar skiff how do you mitigate heavy weekly consumption on Pro 20x?

like, how do you check what makes rapid consumptions without ever being --yolo or xhigh on 5.5/5.4 ?

[slim wave]

i wonder if it uses more since I'm using wsl

[vital hill]

I posted an offensive it seems. Damn pg 13 here

[cedar skiff]

I make my orchestration use weaker models where they can

Thats it

[hard drum]

same as mine

yet... still

i try whatever i can to make orchestrators just... work, but even still

at that point, either what i have is so fundamentally wrong, or Codex as harness is not something that works for my workflow, so that means making a custom harness with the flow i need

[cedar skiff]

I also dont go crazy with it, i could spin up 10 agents but i only have a sequential 1 agent run going instead. This way i can manage my usage.

[hard drum]

so, you limit threads to 1 only?

[slim wave]

@cedar skiff how lower uesage

[hard drum]

so that only 1 subagent can run?

[slim wave]

is this normal usage for plus ?

[cedar skiff]

I say do this huge list of tasks sequentially with codex 5.3 high reasoning subagents
Or something similar, maybe if its information gathering for a task i might have five 5.4 mini models do it.
Then it just chruns away and i can stop it every now and again and get a progress update

[hard drum]

https://github.com/xsyetopz/OpenAgentLayer thing of my i own i use. on pro20x, there is a major split of many 5.4-mini's && 5.3-codex's. one or two use xhigh when you choose pro-20x as your sub for codex

some i recently switched to 5.4 from 5.5

but if all subagents are to be centred around 5.4-mini/5.3-codex, then there are fundamental trim-downs needed next reset

but i'll also contact openai support in regards to what i'm doing wrong, && hopefully get some guidance

if you can see what i do wrong, maybe asking me for specific files may help further?

like from my local .codex

as you seem to know the workings more than i do

not sustainable to keep paying 150-180 bucks per month && hit weekly before i even make it to half a 7-day window

[cedar skiff]

to get 5.3 working properly i do a refinement loop first, I get 5.5 high to write a skill get the subagent to do a task then 5.5 audits, interrogates the subagent about the problems and then runs the same ask again after updating the skills. There are nuances to this system though.

You have to tell 5.5 not to start make explicit rules for the given task, all skill changes need to be generalised and agnostic the tasks. When it talks the subagent it needs to make sure the subagent knows we arent looking of an apology we want to know the reasoning behind the mistake, otherwise it will just say oh yes i should have done this other thing instead my bad...
Back on the topic, i could use heaps more usage if i had it.

So i just temper what i can do.

this is my usage this week so far

[vital hill]

Damn you are doing too much

[cedar skiff]

It's not enuff

[vital hill]

Too much context = hallucinations

[cedar skiff]

I think you are confused

[vital hill]

Maybe

I’m focus on pool right now

[cedar skiff]

This is pretty much how goal works

It just iteration on the same task over and over until we get what we want

5.5 crushes this work load

[vital hill]

I got mad at codex 5.3 cuz it didn’t adhere to my rules

So I had it write to the rules to explicitly follow it which it does now

[cedar skiff]

Thats why i do a loop to refine the skill

[vital hill]

I should do that but it doesn’t learn until a specific event happen

Until then I’ll call it out in its mistakes

[cedar skiff]

you can go from 1 task to 5 and then scale to automated or similar

[vital hill]

It took a dozen prompts for it to get the semver rules right

From dozens of prompts

Like it would iterate the version on every prompt. I’m like no, compare the local unstaged commits with the local head to the remote

[cedar skiff]

here is an example of it iterating on a skill using goal

It's just going to keep churning until it works

[grizzled glacier]

Share code

[cedar skiff]

for?

This is all prompts o.0

[pale oyster]

ㅠㅠ

[hard drum]

Codex Mobile
Windows

lmao

[pale oyster]

What's going on? The post I just wrote has disappeared.

Is it true that Codex Mobile isn't available on Windows yet?
Is anyone using it?

I wonder if I used a banned word or something controversial.

[hard drum]

you're asking Codex Mobile on Windows

i think you realise your misunderstanding, right?

[pale oyster]

You know how the Chat-GPT mobile app recently got an update that lets it integrate with the Codex app?
But that feature hasn't been updated for the Codex app on Windows yet.

[hard drum]

oooh

you meant it in that context

[pale oyster]

My friend uses a Mac and gets along with it just fine, but I'm on Windows, so I haven't quite got the hang of it yet...

[hard drum]

long-term solution: move to 🍎
short-term solution: hope that people actually care about Windows here

[pale oyster]

ㅠㅠ

[hard drum]

not even microslop cares about Windows--see Windows 11 for more information

[pale oyster]

I just wanted to vent in case there’s been an update and I’m the only one who doesn’t know about it T_T

[hard drum]

so why should OAI lol

[pale oyster]

That's really sad FACT news for Windows users.

[high yew]

Codex mobile connect not working?

[worldly charm]

is anyone getting "Oops, an error occured." msg?

[cedar skiff]

are we getting a reset maybe?

[lethal oracle]

codex mobile is working, but my phone is a bit hot even leaving it standby mode

[cedar skiff]

If we get a reset this will be one of the time actually needed it

[late hamlet]

Hey, quick question — is Codex working normally for you guys? Since yesterday I’ve been constantly getting this error:

"We're currently experiencing high demand, which may cause temporary errors."

And I literally can’t code because of it.

[nimble terrace]

hit the reset button

[frank lintel]

Guys i feel claude code really tries to address the user requirement more than codex

Codex just does something for the sake of the request

This is the same issue with gpt i feel!!

Just a personal opinion

[plush harbor]

heh, commented to codex I was staring at this bug for half an hour before actually seeing it and it gave me the eqiuvalent of pats the human on the head it is ok feeble human, you sometimes miss things on the page when you are not looking directly at them

[regal kraken]

man codex is just going downhill

decided to update and rip

and its so slow and still freezing

[plush harbor]

mine has been fine. I'm the wetware failing the testing

[regal kraken]

are you on mac

[plush harbor]

yeah but not in the app

[runic pelican]

For Windows nothing?

[regal kraken]

im going back to extention

yea convo loads fine in the extention

sucks because even t3 code doesnt work for me anymore

we back in cursor usinng the extention loll

[twin maple]

I'm having the same issues on the windows app

try again works about 25% of the time

[pale oyster]

not working...

[rustic walrus]

codex mobile not appearing

[signal tapir]

I've had that a few times. "That's completely normal. Humans have limited attention spans which means they can miss things. An AI doesn't get bored and can stare at code forever with the same attention level." kind of

[plush harbor]

yes like that, a very slight change in wording would be condescending but it worded it very diplomatically

[signal tapir]

Mine often sounds condescending, in the absolutely kindest and most caring way

"You are weak and pathetic. Luckily you have me to take care of you and I'm always here for you" is the feeling it can give off at times. 😛

[plush harbor]

cli or app? They seem different

[hard drum]

I'd fold if somebody told this to me

[signal tapir]

I only get this with chatgpt.

fold?

[plush harbor]

oh chatgpt itself can be overly nice

[oak trellis]

64% 3 days left - 5x plan

not bad ..

codex 5.3 .. best model

usually doing paraell agents ..

codex 5.3 xhigh

nice israel flag

[potent mason]

You should probably try to use it all up

Tibo is resetting tomorrow

[oak trellis]

oh man

[oak trellis]

validated ?!

[potent mason]

Pretty much

Let me find the tweet

[oak trellis]

ah f ack

WHY ..

[potent mason]

https://x.com/thsottiaux/status/2055446089957036402

[lofty tinsel]

Hi, I just create a local solution for latest Windows Codex version. FYI
https://github.com/openai/codex/issues/22985

[oak trellis]

smh ..

WHY

ok i maxxing it now

[cedar skiff]

Im all g with it

[oak trellis]

code maxxing now

[cedar skiff]

saw it like 6 hours ago

[finite current]

when reset 🤣

[oak trellis]

reset maxxing .. he is on

[potent mason]

I have a couple /goals running on the background overnight on 5.5 medium fast

[oak trellis]

imagine he would say ... sorry no reset lol

[potent mason]

He wouldn’t, he already mentioned a reset and to /fast max

[oak trellis]

ok good i will now paralell validate my codebase with 5.5 xhigh

🔧

🇩🇪

[signal tapir]

tokens... omnomnomnom

[lunar crypt]

Do you have any good skills to share mayhaps?

[cedar skiff]

mine are all project specific and tuned to the local work flows, let me check what generic ones have

[cedar skiff]

the only really generic one i have is one i extracted from claude superpowers, i use it a lot as well it's the systemic debugging one from here:
https://github.com/obra/superpowers

[lunar crypt]

Thanks 🙂

[cedar skiff]

I use it in most of my work flows

i do task-of-some-sort -> blocked or complete -> if blocked new 5.5 high agent with systemic debugging to fix the problem -> back to the worker to finish the task.

[torpid trout]

Unpopular opinion… skills and agents and plugins are extremely overrated

I use only one tiny agents file per project - it’s maybe 10 lines of scoped instructions.

Mostly, everything else is just noise in the context

[cedar skiff]

if youre not using skills and orchestration youre behind

[torpid trout]

Hooks are much more powerful and controllable if you need to interfere in llm‘s chain of thought

[torpid trout]

Name me one thing the llm can’t do without a skill or agent

[cedar skiff]

it cant write idiomatic tests.

It also cant be on top of the current best practices, it has holes in its thinking that you have guide. It doesnt know the entire spec of something like material 3 or ios design system etc etc

[torpid trout]

Neither does the skill you add
An MCP could solve that somehow, but also not entirely, because llm does not know == means it needs to search for specific answers

[unique spade]

I don t use much "skills" per se. I have one I made

But that's just a format choice.

I use written procedures a lot for LLMs. And it can easily be argued they are the most important thing

[cedar skiff]

it also doesnt follow architecture ideas or understand how to work within a given projects policy

[unique spade]

A llm of you want to reason how you want you need to give it very explicit meta instructions

[cedar skiff]

You misunderstand the difference between an mcp and a skill.

[torpid trout]

Written procedures you mean basically a project scoped agents file with instructions specific to it?

[unique spade]

Otherwise it will just vibe how it does stuff for you based on whatever generic defaults are salient in it s training and generic instructions

[cedar skiff]

mcp tools LIVE in the context skills are called when they are needed.

[cedar skiff]

Neither does the skill you add yes they do, you clearly dont understand how to use them

[torpid trout]

Ok 🤷‍♂️

[unique spade]

Yes, written instructions I d say that s what matters. How you pack them and how you bring them into context is more of a style choice

[torpid trout]

Yeah

[unique spade]

I mean I m with you on "skills* per se are not some kind of magical skills like many imagine they are. They re just some text that gets a bit of special treatment, but nothing more

Plus a bad skill is worse than no skill

😂

[cedar skiff]

the models are trained on the skills standard, and they use them really well, they are either proactively called or explicitly called and the context is not loaded until that time. they dont pollute the context, there information is injected at the very time it is needed or not at all.
Using your own home rolled system wont compare

[whole stone]

you mean the cli, not the codex app

[cedar skiff]

Once you work out the power of skills you will kick your self for not using them sooner

If you never work it out you will be dommed to baby sit your agents for ever

[unique spade]

Models are trained to follow instructions that's important.
The part with "trained on skills standard" and what you imagine it means, it s hard to argue what adds on top of instruction following without serious scientific A/B

You literally say that same instructions, will work 10x better if they re put in a skill as opposed to "home rolled system".

Which personally from my experience I don't find it to be so

[plush harbor]

this could get as heated as claude v codex

[cedar skiff]

i dont remember saying this?
I do remember saying your home rolled system wont compare.
I would bet my house they to post training on skills usage.

Your home rolled system does not have the saem ability no matter how you look at it.

[unique spade]

Yea i added the 10x, but everything else is basically your claim 🙂

Because you do need to compare the 2 on the same instructions expressed

[torpid trout]

I used to use them extensively actually
Figured a) they aren’t guaranteed to be loaded and b) using a custom orchestrator you’ve much more control over what the llm does and how/when

Looking at all those skills people use ranging from „you’re an expert xyz“ … doesnt make your llm any smarter.
Loading a skill that has let’s say a tool to extract/preprocess specific file format it wouldn’t understand out of the box etc, yes, that’s useful. But it’s still arbitrary whether it’s used or not.

Having an actual orchestrator is the way if you need that level of control, but it’s not as easy as loading a file saying „you’re an expert“

[unique spade]

Like both skill and home rolled system tell the same sequence

Cause otherwise the comparison is not about skill vs home rolled, but about the quality of instructions

[signal tapir]

from what I understand, skills kan also include code that gets executed (assuming it can run in the agent sandbox)

giving them a certain level of determinism

[torpid trout]

Exactly what I say above

[unique spade]

Bro...how you school your orchestrator to do stuff, for me is 100% the most important thing

[torpid trout]

I sent it to Harvard. Is that good?

[unique spade]

Perfect

:))

[torpid trout]

Was thinking homeschooling but it’s frowned upon

[unique spade]

Haha

[cedar skiff]

skills have a small meta data that is injected into the system prompt at the start of the conversation.
From that they are proactively called by the agent in the moment they are needed. Your home rolled system cant do this.
For example:
Add a dialog to this page that tells the user they have an update and get a subagent to write units tests for it. The subagent will get the topical data from the skill without me having to do anything. It will just use it. But the main orchestrator wont have it at all. If told the main orchestrator to write the unit test, when it was time to write it it would read the skill.
A skill is just a prompt, it can have scripts it calls or you can get it to do anything a model can do. The main advantage is that it is called apon only if it is needed and it doesnt exist in the context until that point in time. So you can have a really huge amount of utility that doesnt take up context.

[unique spade]

Your home rolled system can absolutely do anything in terms of what gets injected into the context, how and when. You need to use at least custom codex cli for that though

[torpid trout]

My homerolled System absolutely writes tests and docs and runs 2 verifications and commits to git and ensures it’s written as I want following the standard I want… with 10 lines of agents file

[cedar skiff]

It is injected into the system prompt and it is trained to use it though.

the meta data is injected into the system prompt!

[torpid trout]

It even tests stuff on multipass docker

[unique spade]

You can home brew your system prompt structure

[cedar skiff]

and can you train the model to use your system?

[unique spade]

You speak as if system prompt is some secret sauce

[unique spade]

That part, neither me nor you have access directly

So what you or me think it is trained to do it s a different story

[cedar skiff]

i dont need access i just use the skills system and it is very powerful.

im out, good luck with your home rolled systems.

[unique spade]

Oki skill master 🙂

[cedar skiff]

You are out here trying to sell ppl on reinventing the the wheel like its no big thing. Skills clearly have some post training, all of the major vendors have bought into them deeply, but you still think somehow your trinket solution is worth while.

[unique spade]

i won't go into a debate about meaning and semantics with someone that has no prior background in these matters :))))

[cedar skiff]

take a seat then

[unique spade]

😂

no reset yet though. which is good since i still have 40%

i spent 20% last night in a few hours after that tibo announcement, let's see how much time i have left till reset

using fast is not very good. if you run 2-3 threads in paralel at least on pro 100 you finish the 5h limit in 2h with fast

[cyan gyro]

It's nice they give us a bit of time to use some extra power till the reset

[signal tapir]

I'm limited by 5h quotas.

Still got 60% weekly

[unique spade]

yup!

[tiny fulcrum]

anyone getting this?

just loses the whole conversation and it can't be recovered

[unique spade]

never saw that one

[cedar skiff]

looks like one of my apps haha

[hybrid hare]

App too

[cedar skiff]

5.5 so much smarter about making choices

I only have 8% left might have to back off the background tasks

[swift lotus]

Is codex mobile working on Windows yet?

Or do you have to use the SSH method?

[cedar skiff]

some ppl found work arounds

[finite current]

you should use /goal, and aim further. it works even exceed of your limited quota

[cedar skiff]

I'll do that if i get to 1%, i will run an orchestration and see how it goes.
but i want to be able to keep working on tasks i cant automate as much for now.

[uneven kayak]

Well I signed up for Claude Max and I'm pleasantly surprised. It has about as much useage as Codex, generous weekly limits. Definitely worth $100.

So I'll keep GPT Pro for $100 and Claude Max for $100, that's probably the only AI I need to develop my game

Plus a benefit of Codex is that it can use GPT Image 2 with its imagegen tool and generate hundreds of images before I hit the usage limit.

[finite current]

hi what game engine you works now?

[signal tapir]

Be very careful with keeping the image style consistent. People are very picky about such things.

[uneven kayak]

I'm not using a game engine, I'm using a JS stack with Svelte + Vite + Capacitor with a WebGL UX layer.

[uneven kayak]

Well that's why I have an art bible

[signal tapir]

That most certainly helps, but don't trust it.

[uneven kayak]

It's literally the only thing that keeps the style together and it's quite comprehensive, I think it's plenty trustworthy

[signal tapir]

ok. hope you're right.

[finite current]

can you share about your art bible please? is that markdown file that tells agent to be consistant?

[uneven kayak]

Yeah, I'm right. I trained a LoRA on Scenario with Gemini 3.0 but I don't even use it because GPT Image 2 has better quality. I upload reference images and I have strict prompting guidelines so every prompt uses the same wording.

[uneven kayak]

Yeah that's what it is, I can send it to you

Just don't share it with anyone else lol

[finite current]

i just wonder why x'D

[cedar skiff]

I can make it into a skill for you if you like

[uneven kayak]

Because I put a lot of work into that document and it's valuable to me, and I don't work for free.

[finite current]

got it sir, respect that ❤️

[uneven kayak]

The art bible isn't actionable, I have related skills that use it

But I appreciate the offer

[cedar skiff]

skills are very useful being topical

[finite current]

i also create my own skills. just curious how people works with them

[cedar skiff]

anything that needs guidance on a regular basis i put into a topical skill.

I found my self correcting agents constantly about material 3 spec details, integration tests, unit tests, package usage etc etc.
So i generate topical skills with that guidance and then i never have to give it anymore.

I also have scripts in skills and build runs with tools that call scripts and do validation of stuff etc

They are just short cuts to common repeated usage that you dont have to remember to use

[trim rapids]

skills are pretty useful; they even work in the regular web UI if you upload them

[cedar skiff]

did they give an api for them on web now?

[trim rapids]

I don't know if there's an actual documented API or not but they work if you upload them ad-hoc in individual sessions

[cedar skiff]

Ah yeah i see what you mean

Would be really useful if they actually supported them in the chat

[trim rapids]

you can get a surprising amount of work done that way, particularly if you convince the model to just try installing stuff with pip even though it knows it doesn't have a network connection

it turns out it has a gigantic package cache available

[meager dragon]

When will that happen?

[trim rapids]

5 minutes after you find enough work to queue up to where you believe you're actually going to take advantage of it

[meager dragon]

my team is waiting, all my team member is on 0% weekly yesterday night

[finite current]

hahaa same

[trim rapids]

oh, you're out? in that case the reset won't be happening

[finite current]

i moved back to free models for now

[bright swift]

wen reset

[finite current]

$100 plan also wiped out? hahaha

[uneven kayak]

Lol no I'm just being weird

[ember spire]

y no wake pet shortcut? 😿

also why does right-clicking along the top of the app crash it on macos? 🤔

[cedar skiff]

oh man why did i try that ahah

It's the reincarnation of the old alt+f4 for the special menu

[orchid echo]

wait usage will be reset soon?

[trim rapids]

allegedly yes

[ember spire]

did you make yourself a pet? 🙂

[cedar skiff]

I havent done that yet

[trim rapids]

why is everyone vibe coding bonzi buddy and clippy now

[cedar skiff]

They probably giving ppl a chance to burn some usage

[ember spire]

if they made it show all the chats rather than making me have to scroll the tiny chat bubble this would be my main ui 😄

[orchid echo]

i feel like 5.5 performance is still degraded..

[cedar skiff]

it's been killing it for me today

[orchid echo]

its legitmately been so stupid for me, not following directions, does something else, not remembering simple basic stuff, confidently claiming things that arent true

[ember spire]

working well here

[uneven kayak]

Lol bonzi buddy, I legit used that back in the day on my eMachine. Little did I know it was spyware 😂

[trim rapids]

I think a lot of stuff back then would have been classified as spyware or malware these days but it was a different world and people just didn't look at it like we do now

[uneven kayak]

Yeah absolutely

Internet privacy today means something very different than it did back then

[trim rapids]

it was the wild west, you could /whois someone on IRC and dump their IP into a script kiddie tool and literally BSOD their machine

[uneven kayak]

Lol yeah

[trim rapids]

and early broadband internet was literally set up like a LAN where you could print stuff on your neighbor's printer lmao

[uneven kayak]

That's DSL right?

It's been a while since I even thought about that lol

[trim rapids]

early cable

[uneven kayak]

All this nostalgia is making me want to actually make my own bonzi buddy pet lmao

[trim rapids]

early cable was really lax with any kind of auth stuff, to the point where you could spoof the tftp server it would pull the modem profile from and change your own bandwidth limits

[cedar skiff]

the first broadband i could get was adsl 1 and it was capped at 50kb/s or something similar. I downloaded from usenet 24/7 for as much as a could for a month and got something around 30gb

[meager dragon]

Why I feel they changed the usage limit again
Last Month I calc my Credit usage is 4293 Credits/week (I guess limit at 4300)
Last Week I calc my Credit usage is 4194 Credits/week (I guess limit at 4200)
This Week I calc my Credit usage is 3891 Credits/week (I guess limit at 3900)

[trim rapids]

I still use usenet 😄

[cedar skiff]

The usenet was also provided by the service provider and it was loaded with rips etc

[simple star]

ok, 8%... time for a reset

[uneven kayak]

Lol yeah that was the good old days of P2P apps like Kazaa and eDonkey

[trim rapids]

and DC++

[meager dragon]

Credit in here means this, GPT Plus Weekly limit = how many Credit

[cedar skiff]

made a material 3 spec skill by setting up a skill to rip and curate the web site. Its working pretty well, made an example app.
Just the start of it, but its looking as ugly as material 3 so making progress.

I made the skill it would build responsive layouts

mobile, tablet and desktop all specced to material 3

It's not just visual i also ripped the the flutter docs and api and turned them into skills and made a specific opinionated skill for mvvm architecture they recommend.
So the code is also idiomatic, not just slop

Still a bit of fine tuning to do though

[boreal holly]

Niiice, yeah their default thing with flutter is material widgets with copious rounded rectangle cards. Takes quite a bit of effort to steer em away from it

[cedar skiff]

That is the next step after i get this working, the goal is vibecode an app as a starter and get a really good base to work from. I thought material 3 on flutter would be the most straight forward because it is already most of the way there.

we can get an ios android and web app all in one hit.

I think getting a custom looking app will be much harder

[boreal holly]

It is... But worth it! One thing that I found works really well is set it up so you have

• The main client application
• A design system package
• A web-only "design_lab"

All widgets, atoms, molecules, pages, shell are created in the design system package, imported into main client app and design_lab client. That way for design work, they use the lab to fire up flutter run -d web-server --dart-define=RENDER_PAGE= and use bun webview to take screenshots and hot reload against that particular page. As long as they're making edits to the design system package, the client app will look exactly the same as the design_lab screenshots. Higher fidelity than golden tests 🙂

[jolly lily]

npm i -g @OpenAI/codex@latest does not seem to work for me, it throws the error npm error 404 '@OpenAI/codex@latest' is not in this registry.

[boreal holly]

npm i -g @openai/codex@latest

case sensitive

[jolly lily]

Oh thanks!

[orchid echo]

does using /fast affect quality?

[boreal holly]

No, it simply gives you high priority processing

[orchid echo]

i see

[trim rapids]

it does route your requests to a different cluster though, so it's totally possible for something to be wrong on the backend in a way that causes degradation in one cluster but not another

[orchid echo]

🤔

[signal tapir]

And here I thought OpenAI ran on a single server. 😛

[bright swift]

big one

[cedar skiff]

just a box in the corner of sams office

[signal tapir]

Seriously though, do they own tons and tons of servers, or do they rent them from other companies?

[cedar skiff]

both

[bright swift]

they'll probably be on aws soon too, after the microsoft deal ended/changed, guess that will add a lot of capacity

[signal tapir]

How long until our entire economy is based on AI?

[bright swift]

if you look at stock market in US it already is

[signal tapir]

😮

Damn, thought we were at least a couple of years out from that

[bright swift]

probably still some time before its core at "normal" companies

[signal tapir]

Can't wait for the world economy to get to a point where we are forced to leave the old economic paradigms behind.

Funny enough, if I didn't have to work for money, I would be doing the exact same thing I am now. 😛

[bright swift]

i found a new way to burn tokens fast. integrated codex review in my git pre-push hook and blocks pushes unless codex doesnt find anything

makes it painfully slow sometimes to land fixes 😄 need more parallel dev servers

[signal tapir]

I need I rarely share

[boreal holly]

Completely worth the overhead though

[bright swift]

yea it finds a lot

[torpid trout]

What does codex review to exactly? Is it like plan and goal just a different prompt chain injected?
Or a totally different model and such?

[bright swift]

i completely rebuilt my workflow from PRs to trunk based. and with these new hooks i'm using a "stacked local diff" to compare against upstream. so its like a little local PR with codex reviews until everything fixed, works really well so far

[boreal holly]

Yes! It has a different system prompt. Basically says "focus on changed code, not existing bugs outside of the current diffs" and stuff about briefly describe the bug

[bright swift]

its the same as running /review inside codex, but non interactive

starts with a fresh context

[torpid trout]

well 😅 i never runneded that, lol, hence why asking

I will try it out

[bright swift]

its really good

[torpid trout]

seemed like nonsense, who reviews code

[boreal holly]

I mean 5.5 is pretty darn good at not making bugs, but the thing review tends to catch is drifts. Oh yeah, unlike the regular agents, the review agent will scrub through all relevant AGENTS.md files and ensure the implementor is following those guidelines

[bright swift]

prompts written by AI, implemented by AI, reviewed by AI - what could go wrong?

yolo

[torpid trout]

you forget harness application written by ai and self-modifying based on level of user swearing in steering window

I made it detect if I am just rambling or truly point out real issues
If I just ramble I have a special agent replying with a meme.
Ther must be some fun in all the sad reality/

[past mist]

Did you ever figure this out?
I've been having abnormal experiences with codex and web chatgpt recently

I think my account is flagged or something and it's purposely poisoning all output

[bright swift]

it has a memories feature but i think its opt-in right now

saves stuff to ~/.codex/ outside your repo

[past mist]

yeah i've opted in to the codex memories feature but it doesnt explain the weird stuff that's been happening with web chatgpt

but I think the issues are related to something with my account

[bright swift]

maybe related to the behavior people are reporting in last few days which led to the investigation yday and reset promise

i have memories disabled everywhere, prefer to build the context from scratch even in chatgpt.com

one of my accounts got flagged for suspicuous activity, guess i overdid it a bit with the oracle usage

[past mist]

I noticed the issues in codex & chatgpt output quality happened immediately after I started getting 403 and 404 errors when using shared projects on the website (its an upstream issue; i've already tried every combination of local troubleshooting steps)

My only theory is that my account got flagged so now all the output is poisoned, but I cant find any information on if openai does that

[bright swift]

maybe i got flagged for uploading zips with 500k loc for gpt pro extended thinking deep reviews lol

[past mist]

nah I dont think that would trigger it

[cinder lynx]

holy f* my codex just crashed i queued like 30 messages

can someone help it wont even launch

and i think an update after bricked it

[bright swift]

ask claude code to fix it

seriously, thats why i still keep a $20 claude sub around

😄

[cinder lynx]

genius

[bright swift]

or opencode or pi or whatever

[cinder lynx]

i still have claude code too

[orchid echo]

so u can send 4 prompts and have ur 5 hour limit be used up already?

[neat hound]

so codex chrome isn't available in europe or something, or for brave ?

[bright swift]

well the occasional "fix codex" prompt yea 😄

[bright swift]

if you mean computer use, i dont think its available in EU yea, maybe with vpn

[cinder lynx]

waiting for codex chrome/computer use to come to europe/windows

why is openai doing this???

[neat hound]

we're so lucky to have technocrats protecting us from innovation and ensuring our safety

[cinder lynx]

but also gating releases to mac os

fukem

[neat hound]

i also can't seem to have the codex desktop thingy work on my mobile chatgpt app, nothing happens when it says to follow the instrructions on desktop

[cinder lynx]

right another complaint

codex mobile seems to be mac os only

or just not working

[neat hound]

dang

[bright swift]

i havent tested it yet but the mobile thing seems to be quite buggy from what ive read

[velvet wren]

the mobile functionality is great

[torpid trout]

Did anyone of you ever had this case where you are working in a given CWD for days and days and suddenly upon a codex command it goes back to ask

 Do you trust the contents of this directory? Working with untrusted contents comes with higher risk of prompt injection.
  Trusting the directory allows project-local config, hooks, and exec policies to load.

You can then press ctrl c and it opens the chat anyway

[bright swift]

maybe after codex update? i see same with hooks approval

then i found the new --dangerously-trust-hooks or whatever its called flag

[magic flower]

Idk if anyone here can do anything about this, but Tibo told all codex users to start using fast mode to use up your usage faster because he was going to reset limits in a couple hours and then didn’t do it

That sounds rather malicious to tell people to use fast mode because using usage wouldn’t matter

[cedar skiff]

I think he wants to ppl a chance to make use of their usage first

many ppl lose out if they dont get a chance to use extra

[magic flower]

Could be but he said that evening. I would just hate if it didn’t happen at all

Follow up communication and it’s all good

[bright swift]

i still have 40% to burn before reset so im glad its taking longer.

[magic flower]

Ive never hit my limits on the $200 plan ever but with the goal skill i hit 0% 2 days in 😭

[cinder lynx]

i hit my limits all the time

[bright swift]

this is my second acocunt, first is already empty

pro 20x

5.5 xhigh + goal + auto-review + parallel = 🔥

[torpid trout]

a cocunt? or did you mean 🥥

No homo reseticus ever

[halcyon island]

mobile codex not working for anyone. says go to dsktop after i scna but nothing is there

[magic flower]

I know it has to be a MacOS desktop but I think even those people are saying it’s buggy

[halcyon island]

Yeah i have mac. Just nothing shows

maybe i will try to restart

[fringe gust]

Anyone else having this issue? I cant open previous Codex chats. But I can start a new one. Seems like its happening when the session ends.

[boreal holly]

flippin text encodings I'll tell ya h'wat

[high girder]

the great mojibake incident of '26 lmao

[orchid echo]

when do you guys think they will reset usage?

[cyan gyro]

after Tibo has had his morning coffee

[orchid echo]

hopefully its right when i exceed my weekly limit

[meager dragon]

I am on 0%

still waiting for reset

[orchid echo]

i still have 22% left, dont reset yet Tibo

[cyan gyro]

I must say though, codex does seem snappier and smarter again

[orchid echo]

wish i could say the same

[finite current]

They just said evening but which evening xD

We dont know yet

[magic flower]

no he said "this evening"

[finite current]

Which timeline sir

🤣

[cinder lynx]

same error - any fix?

[fringe gust]

Not that I know of

[solid lake]

Open a fresh chat, ask it to track down the chat you want and print the command with the identifier

[cinder lynx]

track it down and do what?

[solid lake]

Print you the command for that chat

[cinder lynx]

what command??

[solid lake]

Codex resume <identifier>

[lost drum]

ggs no pro model for me bye bye 20x was good

[cinder lynx]

itll just reopen the same convo?

or ur saying open codex in cli?

[solid lake]

It shouldn’t if you found the chat you wanted

[lost drum]

damn 1 prompt used 16% of weekly and 100% of 5h damnnnn

[solid lake]

Yes

[cinder lynx]

very long way to say open in cli. and thats not a fix...

[solid lake]

No

You won’t find the chat in cli either

[cinder lynx]

you can easily find it in the UI... u dont need to ask the ai

literally copy session id

[solid lake]

Dude these are for chats people seem to lose

[cinder lynx]

Have you ever used codex before?

hes not complainig his chats lost

hes complaining its not opening properly

[solid lake]

The chat isn’t loading

[cinder lynx]

You're slow

[solid lake]

So see if the session is discoverable by codex

You’re dense

@cinder lynx Anyone else having this issue? I cant open previous Codex chats. But I can start a new one. Seems like it’s happening when the session ends. Reading comprehension man

[rocky fog]

at least in vscode extension windows/wsl
yes
after I close vscode and then reopen, it can be tough to load previous conversations, the bigger the worse
sometimes its about waiting a bit, but other times it just never loads it

but when the chat does not load, I have to close vscode, open it and try again, sometimes up to 5 times and eventually it loads
maybe also dont do it too fast, open vscode, let it load the text about model, then open the conversation...no idea if that helps 😄

pretty much the only issue I have with the extension, but eventually I can always load it

[boreal holly]

> console.log(this.evening)
[object Object]

[boreal holly]

Are you on Windows or macOS (assuming not linux because desktop app)? Did this start happening after an update?

[solid lake]

@rocky fog try this

Also that was copy paste of op’s question

I’ve unexpectedly closed many vscode chats or codex chats to have this problem

[ripe ocean]

@lost drum

[rocky fog]

interesting
I will try that next time, thx

[lost drum]

what?

[solid lake]

Yeah it helps if you remember specific prompts or keyword from that chat

[lost drum]

I thought spud came out cause of notification and I am not on pro

[boreal holly]

If you are on Windows, you should download https://learn.microsoft.com/en-us/sysinternals/downloads/process-explorer

See if a program has an open file handle on that thread's rollout log, terminate it, then open the chat. That's usually why on Windows I recommend "unplug USB devices (rule out hardware driver interference), restart computer (clean up file handles)", but if it takes multiple attempts with eventual success, you might just have zombied processes hanging onto the rollout log.

Also if you use codex for desktop app and vs code extension at the same time you will probably have the same issue specifically on Windows due to an architectural decision made back in the MS DOS days

[rocky fog]

I think it also happens through restarts and then reopening (and whether its in WSL or not, same)
but good to know, will keep it in mind

[solid lake]

What was the decision regarding

[boreal holly]

Oh yeah, restarting just VSCode is not enough. Typically requires restarting entire computer.

[magic flower]

i think they didnt reset limits yet because there are so many other bugs they would ask to reset limits for that theyre waiting to fix them all and drop one big reset

[rocky fog]

these menus in codex just love to appear and disappear randomly 😄
nothing new 😄

[solid lake]

Yeesh windows gets more problem than linux is saying something

[magic flower]

as for what im working on, ive been making a pokemon gold romhack that is supposed to make you feel like youre playing for the first time again (new type chart, buffed pokemon, buffed moves, harder gym leaders)

[solid lake]

Ng+?

[boreal holly]

They decided a file should only be accessed by 1 process at a time, whether it's reading or writing, that file gets locked by the owning process. On unix-based operating systems, everything is treated as a file, including hardware, so by design they cannot lock files like that. Your keyboard & mouse, storage devices, audio devices, everything are files in the file system (byte stream with input and output), so they need to be readible and writable by multiple processes 🙂

[magic flower]

its not meant to be like a crazy difficulty hack, but you should feel your heart rate increase a bit when you face a gym leader

regular trainer ai is actually completely unchanged, just the gym leader ai

so they truly feel like a different beast which would make sense in the actual universe

[solid lake]

There’s an Elden ring mod that’s tunes everything too that I’ve been meaning to try

Meant to be unplayable but people have beat it

[fringe gust]

I’m on the windows codex app, yes this happened after the update

[rocky fog]

its more like a process can decide to do that, but does not have to

[solid lake]

Was there a reason

[boreal holly]

They don't treat hardware as files. Idk what the reason was, but since files are a separate abstraction they decided "let's create locks in the kernel for files". And this is the reason why software development can be a herculean process. Build failures in Visual Studio are oftentimes solved by restarting the computer for example

[solid lake]

lol go compiler does funny stuff too, just updates randomly after giving anxiety

In vscode

[boreal holly]

Yes, but if you have anti-virus software running, those obtain exclusive locks on files by design to secure the system (don't let viruses touch bad things, intercept handle before damage done). Sometimes they have bugs. Sometimes you're running builds that are updating lots and lots of files in rapid succession, the anti-virus leaves an exclusive lock on something and bricks up your PC. Doesn't have to be just anti-virus though, lots of software stacks make you opt into shared mode, all it takes is one bad process to hold a permanent tombstone lock on a file and cause issues

Oh yeah I guess disable anti-virus is the best first step, before unplug all USBs and restart PC

[trim rapids]

I dunno what kind of A/V you've seen that takes regular locks on files instead of implementing a proper minifilter driver to transparently intercept access but I don't want it anywhere near me 😂

[boreal holly]

Even the minifilter driver will block writes pending analysis. Not all compilation processes are single threaded and serialized. It is however better than the way I described for sure

[trim rapids]

yeah but even if it blocked forever, e.g waiting on the user to click something they never bother to click, at worst it should just result in an async i/o worker (or whatever is being used) never returning a result, so if anything actually hung up from there the bug would squarely be in the build system (or whatever) in question

[young locust]

sora is gone

[rocky fog]

rip

[young locust]

dead

[rocky fog]

https://help.openai.com/en/articles/20001152-what-to-know-about-the-sora-discontinuation
(well, does not answer why, but more info :D)

[young locust]

did not turn out to be what they wanted

[torpid trout]

https://help.openai.com/en/articles/20001152-what-to-know-about-the-sora-discontinuation

Do you all know those satire videos where they make fun of "if google where a person" and in come all those awkward questions..
I wonder what GPT would have to say to that
"where dis"
"give code"
"not work why fix now"
"I am poor make me rich"
"My wife left, fix it"
"I need exam done, now"
"😠 🤯 😰 "

there's plenty of other services for that I believe
Also, Users between the ages of 13 and 17 must have permission from a parent or legal guardian to create and use an account
Do you?

[uneven kayak]

It's funny when I use Codex to diagnose problems with the Claude desktop app 😂

Got one agent diagnosing the other

"Where are you from?" Being 13 is no excuse to have poor grammar in a channel that is literally about prompt engineering lol

[torpid trout]

Its not tho

Its about codex 🤓

[uneven kayak]

Lol yeah and what do we do with Codex? 🤔

[bright swift]

not prompt engineering lol

[torpid trout]

Streaming interrupted. Waiting for the complete message…

Here we go:
no prompt ängeniiiireeing

[uneven kayak]

If you're not engineering your prompts, what are you even doing? Seems like a waste of money to use Codex with bad prompting.

[torpid trout]

why, I have codex write my prompts

or gpt

[uneven kayak]

Huh

[torpid trout]

I only say "why dis. fix"

[bright swift]

gpt pro writes my prompts for codex cli

[torpid trout]

and cli for gpt. Its a closed loop

The only time I get really ängeniiiireing is when I ask for memes

[uneven kayak]

Interesting. I never considered using AI to rewrite my prompts. I'll spend like an hour and a half writing a detailed prompt because hitting that enter button costs money lol

[solid lake]

How do you manage hallucinations

[torpid trout]

noob

[solid lake]

It’s more fun if you edit the prompt with gpt after

[torpid trout]

you mean his hallucinations?

[solid lake]

“Research gaps”

If any

[torpid trout]

OK I guess I have to step away, the /s is having a moment with me

[solid lake]

Yeah

[solid lake]

No, gpt hallucinations

[bright swift]

not worth perfecting prompts/spec imo. just get stuff implemented with hard guardrails for code quality etc, then iterate. even with perfect spec it will never be correct in one-shot mode anyways

[torpid trout]

wat dat

[solid lake]

Pain in the butt mostly

[uneven kayak]

Yeah that's a real concern for sure, writing my own prompt prevents hallucinations (assumptions about things that don't actually exist, in this case) from creeping in

[solid lake]

Very easily noticed if you read

[bright swift]

idk what you are doing but i never have problems with hallucinations

[torpid trout]

when it does that I just have it create what does not exist yet

[uneven kayak]

Lol you're cracking me up today

[torpid trout]

not going to say it... but who reads?

[torpid trout]

yeah me too

[unique spade]

one shot right now actually means a long meta-turn of the orchestrator who with the proper meta-program can one shot a lot of stuff

[torpid trout]

oh, they had human errors?
Thats normal - humans make errors all the time bro

[bright swift]

even with orchestrator + subagents in isolated worktree + pre/post tool use hooks with hard guard rails + post commit review-loop, it will never be perfect because your spec/goal/prompt will never have all edge cases in it

[solid lake]

Yeah like not reading

[torpid trout]

Nothing to worry about, there's a tool it will solve it all

LLM to the resque man

[uneven kayak]

Doesn't need to be perfect, it just needs to be like 80% correct about 90% of the time lol

[unique spade]

i m currently one shotting programbench tasks at 100% score with this method lol

but sure you re entitled to your certainties

[bright swift]

yea exactly, thats what im saying. so not worth writing the perfect spec/prompt

[uneven kayak]

That's what the human is for! Agents haven't replaced us yet! 😂

[solid lake]

I mean in his closed loop that’s exactly the case 🤔

It’s also insane when I check chats from 2-3 years ago to see differences in output

[uneven kayak]

Closed loop, I guess that's fine for processing huge amounts of information if nothing changes and you're just processing sort of the same thing over and over. But in a software engineering environment, a closed loop with an AI is asking for problems

[bright swift]

then what exactly do you mean by "long meta-turn of the orchestrator who with the proper meta-program"?

[solid lake]

Yeah this, determinism

[bright swift]

not familiar with programbench exactly but i looks like you are building from a verifiable spec for something that already exists. thats a bit different than building something new in the real world

[solid lake]

LLMs are great at that

[solid lake]

The human ideation is what aid the second imo

[bright swift]

i mean the bun guys recently proved this with the rust rewrite based on existing tests

[boreal holly]

Infinite closed loop, maybe

[solid lake]

That’s a big claim

There was a paper on this too lemme find

[bright swift]

i'm building new stuff where i dont know every single spec detail upfront, so for me its not worth defining everything up front, i just iterate

[solid lake]

https://arxiv.org/abs/2603.29640

[uneven kayak]

Well ain't that something, Codex fixed my performance issues with the Claude app. Runs like butter now.

[solid lake]

There’s a part 1 from 2025 too

Very cool read and maybe viable once models get better

[uneven kayak]

Anthropic already uses Claude to code Claude, they stopped coding by hand a few weeks ago.

I don't think the same is true yet for Codex though

[solid lake]

Months ago no

They said it a long time ago already

[bright swift]

pretty sure they are doing stuff like that in their research dept, incl synthetic data etc

[solid lake]

Yeah

Nvidia does it too

[uneven kayak]

Months? Wow time flies

[solid lake]

They probably help with the generation

[torpid trout]

codex is entirely 100% vibe coded I am fairly sure
They even have a blog post somewhere describing a misterious app they created solely with gpt, yet never disclose whihc it is
Lloking at the pace of release, its def codex

also the release notes

Speak for that famouse "didnt read it" approach

[solid lake]

It was smart of them to wait for the models to progress to vibe code it

Took apples approach

[bright swift]

very important to dogfeed this stuff so it actually gets better. they should also actually use normal subs to see real limits from time to time while working internally, not just their unlimited /fast API keys

[solid lake]

Also stuff like the models having supplemental training to get them used to tool usage etc

Can’t just deploy the harness without the foundational model learning how it works

So like every time ig they make major changes to codex they must respect that process

[uneven kayak]

We're slowly moving into a world where two things will become true. Code won't be written by hand anymore, and anyone can just make their own app. But SaaS has been around for a long time so it'll just get even more popular, people won't be buying software licenses anymore, instead they'll be paying monthly for support and updates.

[solid lake]

Have you guys seen the vibecoding video by andrej karpathy

It would be a nice “where we are now”

https://youtu.be/96jN2OCOfLs?

[boreal holly]

Yeah dude, it makes zero sense to do less than 100% vibe coded. If you choose not to 💯, you're opting out of adapting to the future of all SWE. The sensible thing to do is learn to 100% vibe code right now with current technology, so when it inevitably evolves you wont still be stuck banging rocks together in a cave somewhere haha

[torpid trout]

I like the image of banging rocks

[unique spade]

i managed to burn 50% of the pro 100 quota since tibo said he will reset lol

i m down to 12%

[orchid echo]

reset isnt happening. we got trolled

[bright swift]

yea similar, wen reset

[bright swift]

whats your workflow you teased earlier for those 100% scores, or is it secret 👀

[unique spade]

it s not secret there are experiments in conceptual reconstruction using only general meta-rules of reasoning (that s the meta-program basically) + what's usually allowed in that eval, namely only the readme of the program plus the compiled programto test it's behavior

that s the current most abstract form of the steps

will update you when i'm done, if it proves stable value

[bright swift]

thanks, sounds interesting, but this is specifically for reverse engineering right?

or are you also using this approach for other stuff

[unique spade]

it's the same for new programs, just that there you discard a couple of late steps

the goal is what i call Semantic IR from intent till actual implementation in code

but with something like programbench it's easier to bootstrap since you have a hard base of what the intent and the actual program are
so i can just focus on seeing how to get the model make the transition between the 2 better

[bright swift]

and you use this only as instructions in prose form or is it a custom harness or similar?

[unique spade]

it relies a lot on the ability of current models to obey machine format structures, so it's not really prose

[tulip harbor]

how is codex usage for chatgpt go plan?

[orchid echo]

https://x.com/thsottiaux/status/2055707616605835333

Codex usage limits have now been reset across all paid plans. Enjoy the weekend!

[unique spade]

great i m happy with this one, managed to use 50% and lost only last 10% in the last 18 hours

[bright swift]

phew finally! he didnt forget lol

good reset for me, first plan at 5% and second at 30% with 4-5 more days to go 😄

[unique spade]

it s awesome like this, to have 18 hours heads up

if they just announce 1 hour before, that s not much you can do to use what u have

[bright swift]

yea definitely nice this way

although i dread the end of may when 2x ends

working hard to get some stuff finished before that 😄

[lost drum]

if they cancel the 2x then mad expensive it becomes

[bright swift]

probably new model by then with new promo

like perpetual sales in ecommerce

[eager forum]

is there a latency for reset?

[supple perch]

@boreal holly so umm, is there stuff going on I need to worry about?

[bright swift]

yea it takes a bit to roll out

[unique spade]

well i'm out of 5hour haha, so i need to wait for that reset to roll

[trim rapids]

I was hoping it would take longer for the reset rofl

[quaint prism]

I know that there are ways around the codex mobile on windows, but any news on the official update?

[boreal holly]

Apparently some folks had

[features]
stupid_mode = true

Enabled in the backend, OpenAI toggled that off and reset limits. Other than that no stuff to worry about!

[supple perch]

??

what file?

[onyx heath]

why doesn't mine work

[supple perch]

is that in the toml file or a meme

[boreal holly]

It's a joke 🤡

[quaint prism]

joking

[supple perch]

<3

[bright swift]

did you try fresh session or logout/login

[onyx heath]

yep

[trim rapids]

the reset isn't instant

[bright swift]

rip then i guess, upgrade to pro like it suggests 😛

[trim rapids]

it's a db update that has to touch millions of rows

[tropic karma]

TY for reset ❤️

[boreal holly]

Oh yeah, still no Linux GUI for Codex, and some folks are still trying to get Codex for Windows to work with varying levels of success. Other than that you missed nothing!

[bright swift]

yay

[tropic karma]

try resending the message

[bright swift]

just in time haha

[onyx heath]

nothing

[supple perch]

I may or may not have my own ui at this point

[trim rapids]

think of how much electricity they would have saved if they just released their own vibe coded port as an alpha instead of having like a thousand different people do the same thing heh

[wide schooner]

Anyone else didnt get the reset?

[magic flower]

It takes a few minutes

[wide schooner]

It took like hour so far

[bright swift]

maybe you are on the DNR list

[boreal holly]

Still cookin

[supple perch]

@boreal holly thank you for your time sir!

[tropic karma]

how's spark for you?

[bright swift]

auto reload credits huh, living dangerously 😄

[boreal holly]

Really good! I use it for quick refactors or "how we lookin overall in this area" stuff

[bright swift]

i used spark maybe once when it came out then never again, but its seperate limit so maybe i should tokenmaxx it hmm

[boreal holly]

If only I could reach 0% in a week 😅 seems only possible if I wanna be egregiously wasteful

[trim rapids]

spark is okay if you don't make the mistake of trying to do complex work with it, the context window size is like that movie 50 first dates

[unique spade]

did any one get the reset yet?

[trim rapids]

yes

[unique spade]

ok so it takes time to spread out

[rocky fog]

also not here yet

actually as soon as I made a call it reset
(didnt have it running for a bit)

[vital hill]

[rocky fog]

wait and now it changed back to old usage, eh

🤣

[unique spade]

😂

[stable obsidian]

When will we get a codex model thatll cook on frontend

[rocky fog]

I resumed goal, I checked usage and the plus icon usage and it went to 100% 100%
then I went here to say that
then I focused back to vscode and it changed in front of my eyes back to old usage

codex trolling me

[unique spade]

There s something fishy with this reset. Don t think it usually takes so much like 30 mins to take affect across users

[boreal holly]

"/goal reset usage for all users"
... Working 35 minutes
"Alright Tibo, I copied the quota rows to a backup table, erased the production columns, and copied the quota rows back in place"

[stable obsidian]

So is it better to use 5.4 now or what

[karmic gulch]

FUll reset usage? coulda told me yesteryda and i wouldnt have saved my usage lol

[boreal holly]

You mean they didn't announce it in the Anthropic server? 🤡

[karmic gulch]

That woulda been nice indeed, could one hope next time?

[bright swift]

they announced it yday on x

[karmic gulch]

Gotta tell my hermes to keep me updated on these things

[bright swift]

most important follow if you want to optimize around resets https://x.com/thsottiaux 😄

[unique spade]

It needs to watch tibo's x account

[simple star]

reseeeet!

finally

[nocturne folio]

oo

i mean i was at like 50% weekly, shouldve used more

but ill take it

[bright swift]

i got about 1,5 pro 20x subs out of this one, perfect timing

would have needed to buy a 3rd if no reset with 4-5 days left on both 😄

[jaunty fable]

I literally just went from 0% to 100%

[karmic gulch]

about same, also got 3 days of claude 20xmax on their reset yesterday, good weekend I guess

[blissful basin]

was it always like that when we had extra reset it was basically starting new weekly cycle earlier? Because reset happened, but it also moved my previous reset timer to date 7 days from now

[rocky fog]

yeah it moves to from whenever the reset happens

[bright swift]

yea sometimes you get lucky and get it in the middle of your 7 day window, sometimes it sucks and doesnt make a difference

[jaunty fable]

When I set it to GPT 5.5 xhigh, it says it's impossible, but 5.5 high works infinitely xd

[karmic gulch]

Been optimizing the workflow in Codex now tho, callin in Kimi and Claude workers within the codex window for 5.5 to have agentic controll over agents and workers with speciailized and optimized task fit for their model , to try and reduce my 200 codex 200 claue and 100 kimi to about half. Anyone have a had any luck in making this work 100 % ?

[magic flower]

I have GPT and Claude $200 tiers (only month I’m doing this)

Yesterday Anthropic reset limits 15 minutes before my limits reset

🙃

[karmic gulch]

You are destined to win the lottery at age 95

[wide schooner]

mines still not reset

[wide schooner]

not bad honestly

[bright swift]

careful with that if you use claude -p since it will now use your extra credits instead of sub

i just use 5.5 xhigh codex for all coding + chatgpt 5.5 pro extended for planning

[karmic gulch]

yeah ik, that’s exactly what i was worried about too lol

not using claude -p for this anymore. got it running through a local mcp/tmux thing instead, so codex kicks off a real claude code session on my logged-in sub and sends the prompt into the tui, then tails/stops it from there

kinda janky, but avoids the sdk/print credit bucket

[nocturne folio]

i think most of us would suffice using deepseek

[magic flower]

Bro I didn’t even tell you the other part

[karmic gulch]

use 5.5 high as the main brain/final call. you fnd xhigh worth it?

my roughly setup is sonnet as the main external worker via claude code/mcp, 5.4 is the main codex-side worker. 5.3 for smaller scoped stuff, spark for tiny fast edits, opus for deep review/security, haiku for scout/logs, kimi-worker for background/bulk/final docs but not blocking the main loop.

[magic flower]

I’ve been waiting since last night for codex usage to reset

I didn’t have anything to do all weekend besides a baseball game, was hoping to set a goal before I had to go

Get in my car to go to the game, check twitter at the first light and Tibo reset usage

💀

[bright swift]

honestly i cant really judge the diff between xhigh, high, medium.. im sure xhigh is overkill for some work

[solemn acorn]

I use medium or low mostly

[solid lake]

Xhigh is worth it

[bright swift]

so i just KISS by using the max reasoning for everything right now, if i get into more budget constraints i might do a smarter system

[karmic gulch]

Your better off with 5.4 high than 5.5 low, 5.5 should be used for high xhigh imo

[solemn acorn]

I disagree, thinking too much tends to cause more problems than it solves and also the model runs slower

5.5 low is quite fast and token-efficient

medium is a good middle-ground if you need something better for planning

[woven canyon]

Lets go, we got a reset

[karmic gulch]

Haha bro, just get a minimax 19 bucks sub, and you wont be waiting all weekend next time, atleast you can do some groundwork with the 2.6m ✌🏽

[bright swift]

i dont really care about speed right now, i use big goal prompts that run for hours or even days in some cases in remote-dev tmux. most of the time spent is actually compiling, lints, tests, mutation tests etc, not inference. and i just have a lot of remote dev envs at this point

[vital hill]

this is a good feeling every month

[solemn acorn]

in that case why do you even need xhigh?

[bright swift]

you mean because i have detailed prompts or what?

[native stream]

What are you guys thinking about codex mobile? Right now some chats I start on mobile disappear if I close the app, and cannot find them in mac afterwards

[solemn acorn]

if the model is simply executing predefined tasks, there's no need for higher reasoning levels

[native stream]

Also it's not following repo permissions on new chat.

[karmic gulch]

works 80 % of the time, some chats wont even open, some works perfect. I just hate the that "wait 10 sec before we group your chats with repos" after you do something else..

[bright swift]

maybe. but i have very aggressive hooks to enforce specific code architecture etc, so i think xhigh is still worth it

to make the most out of the hooks feedback, custom lints, etc

[solemn acorn]

ehh

[boreal holly]

Not only that... If you gotta crank it up to xhigh 24/7 to get the results you are looking for, that's a huge red flag.

I like to think "If my agents are struggling on Medium, I have a problem and I need to make it so they don't struggle."

[solemn acorn]

I think what would be more helpful is having more stuff in the model's context that's not unessesary reasoning traces

[trim rapids]

one thing xhigh is useful for is finding contradictions and other reasoning traps in your setups that will cause issues with lower reasoning levels

[solemn acorn]

unfortunately 5.5 doesn't survive compaction as well as 5.4 did so it is important to minimize those

presumably to be fixed in 5.6 or gpt-6 depending on what they plan on branding it

[boreal holly]

Huh, I've had the opposite problem. 5.5 does unrealistically well across practically infinite compactions

[trim rapids]

when your setup is smooth enough that even xhigh isn't overthinking every step of it, you know you're in a good place

[bright swift]

i also dont see problems with 5.5 compactions

[uneven kayak]

Odd, my weekly usage just reset. But I still had 4 days left.

[solemn acorn]

I notice it gets off track a lot more than 5.4 did with multiple compatctions

the caveat being 5.5 doesn't go through as many compactions because it's more token-efficient

[bright swift]

my usage reset back to empty actually after it reset to full just now meh

[gentle pivot]

i find 5.5 to be much better at staying on track in long running sessions

[uneven kayak]

What is even going on here, my useage is resetting every minute, it's staying at 100% remaining even while running a session

[bright swift]

infinite usage glitch

[uneven kayak]

I'm thinking it's most likely just a visual error

But if it really is unlimited I'm not restarting the Codex app ever again lol

[gentle pivot]

My usage definitely reset since i was at 0% so its not visual only

[eager forum]

Still not reset

[uneven kayak]

Mine is continuing to reset every minute, the 5 hour window keeps moving up every time it happens

[trim rapids]

actually mine is too

that's interesting

[rocky fog]

well, dont say you guys didnt ask for it 😄

[jaunty fable]

Please reset this, it just killed my project, the fake reset...

[eager forum]

Sounds great

[gentle pivot]

how did it kill your project

[jaunty fable]

Basically, it shifted my reset day, like it was as if I hadn't had a reset today :/

[rocky fog]

nonstop reset
no more wen

[bright swift]

just make usage unlimited

[jaunty fable]

and my next reset will be on the day my subscription ends.

[uneven kayak]

well those are formatted terribly in Discord lol

but if you expand them, you can see I'm not lying 😛

[gentle pivot]

I doubt its a fake reset if it pushed your next reset back

[jaunty fable]

Well, my week is back to 0%.

[gentle pivot]

on website or /status?

[gentle pivot]

Website doesnt seem to update as fast as /status in codex TUI

[proper cosmos]

my token allowance on plus was on it's last 6 percent for the week with 2 days remaining...
It's just suddenly reset to 100% with a week left from now. Is that normal?

[eager forum]

Yes

U got the reset. Where is mine

[proper cosmos]

oh im catching up

[rocky fog]

just now I am at 100% (again)
lets see if it stays

[proper cosmos]

is it a lie or not

[teal owl]

think it is lying. It says i have 99% left but getting message that my 5hr limit has been reached

[boreal holly]

OpenAI is experiencing some technological differences right now

[bright swift]

im at 100% again too 😄 hopefully it doesnt flip flop again

[rocky fog]

so far longer than the last time 😄

[bright swift]

It will all be settled in monday night rehabilitation

[uneven kayak]

I just checked the website and confirmed that it was all a lie. My weekly useage remaining is right where it should be, 61%

[boreal holly]

This next /goal is impossible 💀

[teal owl]

it is all weird. website says 99% remaining for 5 hour, 100% remaining for weekly. but getting "You've hit your usage limit. Upgrade to Pro (https://chatgpt.com/explore/pro), visit https://chatgpt.com/codex/settings/usage to purchase more credits or try again at 6:04 PM."

[rocky fog]

gonna have to do another reset for the reset issues 😄

[uneven kayak]

That is such a high quality image, I can't get over how good GPT Image 2 is. There is soooo much right about that image that I would seriously think a human made it.

[lean lark]

Fixed that post, sorry about that. Thanks @boreal holly.
Topic: How to connect from Android to Codex Windows.

[rocky fog]

I think my 5 hour window just moved
and I guess its not going down from 100%
so yeah 😄

[solid lake]

17% here

Spare some rate

I’m squeezing by using xhigh fast

[rocky fog]

maybe tibo just spammed the reset too many times and its all coming through

[solid lake]

Yeah

[boreal holly]

Poor OpenAI trying to deploy a reset on a saturday and I'm over here burning up GPUs on dank memes 😂

[solid lake]

Curious to know what this system is

Bro please spare some gpu for us

There’s many dank memes

[lean lark]

Images and low quality requests should be routed to a different data center.

not system, not rack, maybe a different country. 😜

[solid lake]

That would be the endgame with all requests depending on model inference and nature

[lean lark]

and then, who decides what a low priority request is? 🤔

[solid lake]

/model low

[lean lark]

well, we already have that. I'm kinda trollin'. 😁

[solid lake]

Cursor tried it with subscription but and all of them give priority access based on tier but request specific I guess inference and speed etc

[boreal holly]

Since EU can't computer use or chrome extension they probably got some GPUs to burn 🤡

[lilac flare]

in which chat tab can I talk about chatgpt behaviour getting out of hand

and I mean, really out of hand

[lean lark]

#chatgpt-discussions

Running around, responses delayed. Happy Saturday!

[torpid trout]

No reset here still. Not that I need it, but thought to mention it, resets wen

[uneven kayak]

/goal make gud HUD, don't stop til HUD gud

[torpid trout]

gpt-6 wen

[uneven kayak]

I kid. It's coming along nicely. But Claude is much better at UI stuff

It's kind of frustrating having the entire game fully functional but being trapped behind a terrible interface, it's a process

[nimble ice]

speed i need this

[solemn acorn]

ironically both openai and anthropic likely use the same vendor for UI training data, anthropic just updates it more often then openai does

[nimble ice]

my usage is kinda homeless

[cyan gyro]

usage not coming down?!

[uneven kayak]

Interesting

[solemn acorn]

google does too but gemini is less capable in general

[woven canyon]

You didn't get a reset? lol

[karmic gulch]

Gemini usage is a joke tbh

[bright swift]

gemini is pretty nice for some stuff, just not coding really

i use it a bunch via api

[karmic gulch]

Did some work in antigravity last week, 3.1 pro did fairly well, but what the heck is that usage,

[uneven kayak]

I may not have said "make gud HUD" but that was basically the summary of my most recent prompt to improve just one page in my game. Codex and Claude both seem to understand problems with UI if I give them a preview screenshot, but for some reason, Codex can't seem to actually fix things. It will try and then it will take another screenshot and say "yup looks good, I'm all done here" while there are still obvious problems with the layout. Then I'll feed that exact screenshot back into Codex and it will quickly identify the problems that it didn't identify before. But Claude doesn't do that, it will see the screenshot and keep iterating until the problems are actually gone.

[stray portal]

hey guys i need some help sign up for school SSO

[bright swift]

did you ask your agent

[stray portal]

what agent

[bright swift]

codex, since this is the codex channel

[torpid trout]

looks interesting https://github.com/wesm/agentsview

[karmic gulch]

ask codex mate... "Sure. I can help you through it.

For school SSO sign-up, you usually need:

1. Go to your school’s login portal or app.
2. Choose Sign in with SSO, School login, or Continue with institution.
3. Enter your school email address.
4. Select your school if prompted.
5. Log in using your school username/password.
6. Set up MFA/2FA if required, usually with an authenticator app, SMS, or email.
7. Accept permissions and finish registration.

Do not send me your password, verification code, or recovery codes.

If you paste the exact screen text or error message you’re seeing, I can tell you what to click next."

and I`ll let you get back to codex and it will help you

[bright swift]

unfornatunatly http://letmegptthatforyou.com/ seems broken 😄

[boreal holly]

finally got the reset

[rocky fog]

seems it eventually settled and is now not moving the window and usage goes down

[solemn acorn]

seems to perform worse than composer 2 (cursor's budget model)

[karmic gulch]

with gemini and share window this gets even easier, askin real people for stuff seems 2020 ish

[solemn acorn]

quite funny that claude performs better in cursor than claude code

[bright swift]

anyone tested grok build?

i was hoping for some kind of high usage glitch/promo but seems pretty limited

[woven canyon]

400%???

how the

[rocky fog]

he is special

[woven canyon]

Does your usage stack up or some thing?

[karmic gulch]

I stopped looking at these charts, whoever puts claude opus over 5.5 is nuts. Ive been a claude maxi for a long time preaching high about Claude, but since 5.4 ish, OpenAI is once again the #1. But ofc, the masses lags behind and catches up to Claude now, and buisness are not better, signing deals with Anthropic and taking the devs hostage within Anthropic, when OpenAI clearly has the best agentic setup atm.

I still hold my claude max sub, but will reduce it in half next month.

But rest assure I will be back with preaching Claude again as soon as they stop releasing nonsene,and actually starts to focus on agentic coding again..

[solemn acorn]

their models are kind of a joke in agentic tasks so idk if you could do much with them

[solemn acorn]

that's opus over 5.5 only in the cursor harness

in claude code it performs worse

[unique spade]

Nah. Don t see any reason to use anything else than codex right now

[woven canyon]

Is the 400% coz of the promotion on the Pro plan?

[boreal holly]

A magician always shares his tricks or whatever that saying is

[woven canyon]

Okay buddy...

[solemn acorn]

cursor's whole thing is being the most optimized harness for models, so it's very possible 5.5 in cursor will surpass opus in cursor at some point

[woven canyon]

You are saying 5.5 codes better in Cursor than in Codex?

[solemn acorn]

right not it doesn't

*now

but the harness will change

[woven canyon]

But the usage must be a lot less

[torpid trout]

4_20_

If you know you know

[solemn acorn]

oh significantly yeah

[woven canyon]

Yea don't worry. I have 500%

[solemn acorn]

5.5 in cursor is super efficient in particular

[torpid trout]

you dont know then 😄

[woven canyon]

Wow so it consumes less tokens to do the same thing?

[solemn acorn]

in cursor but yeah

[bright swift]

can you use the openai sub in cursor?

[solemn acorn]

the harness is getting a lot more important than the model at this point

[solemn acorn]

I don't believe so

[bright swift]

or only their plans

mh

[woven canyon]

Damn claude code is really bad at caching

[solemn acorn]

yeah idk why those open-weight models were included running in claude code though

I'd want to see benchmarks in opencode for those tbh since that's the harness actually designed for them

[karmic gulch]

anyone using hermes wiht codex? considerig to swap out kimi for codex in my hermes setup.. but not sure it will be alot better, or just cost more tokens..

[woven canyon]

Have you guys tried OpenDesign yet?

[karmic gulch]

Yes

[bright swift]

no, is it just a collection of skills?

i tried ui.sh from the tailwind guys