codex-discussions

[cyan wing]

yes, only compact like ~3 times.
then make a new thread for fresh context window

[cedar skiff]

I havent noticed this, it always just seems to compact at the same level for me

[indigo robin]

mm

it was a major issue for me

[cedar skiff]

Ill get codex to have a look into the repo and see how it works - not sure if its even in there

[ornate jolt]

i can feel the vibes

[steep dust]

Is something broken with Codex? The usage is draining insanely fast

[gloomy venture]

no

[boreal badge]

what a

[cedar skiff]

This is an analysis of compaction history for a 8 compaction chat i had. The user section appears to grow each compaction.
Looks like it keeps all user messages in a growing list.
So youre right if your conversation goes on long enough and they dont cull some of this, over time a user conversation will grow bigger and bigger

[gentle harbor]

whats in the latest update ?

it seems like every update makes the thing worse

didnt mean to reply to you ...

[plush harbor]

why is it so painfully hard to report a missing chart in this openai thing

[oak trellis]

oy vey

agent chat cant access anymore

[woven canyon]

5.5 have been literally shting itself lately

[cedar skiff]

it's been killing it for me

[woven canyon]

man the results been terrible

And its consuming tokens faster now for some reason

[oak trellis]

usage ?!

its now related to usage same as 5.3 codex xhigh ?!

i mean its worth to switch back to 5.5 and have the same usage token limit rate ?!

or burns much faster

[cedar skiff]

i havnt noticed a change did they change the pricing on codex 5.3?

[quick thorn]

codex sucks badly this trash can’t even train a single model correctly holy sht

[woven canyon]

I am cooked

[quick thorn]

Claude did a way better job ngl

[gentle harbor]

isnt that against tos ?

[woven canyon]

What tos

[quick thorn]

Yea I think you’re lost buddy

[woven canyon]

What is bro on about

[quick thorn]

bro doesnt know what he’s talking about

[gentle harbor]

tos means terms of service ... ? how didnt youo know even that simple term

[woven canyon]

Claude is bad when it comes to usage but lately codex usage been terrible as well

[quick thorn]

I’m talking about performance screw usage idc about usage as long as it performs

You give it a build doc and it forgets what it just did one step ago

[indigo sandal]

What plan?

[woven canyon]

They probably fked up context window again. No wonder why the context window limit doesn't show up anymore

[woven canyon]

Plus

[quick thorn]

you literally can’t create a single thing with just one context window everything you could possibly code takes compaction

[indigo sandal]

Yea makes sense

[cedar skiff]

youre holding it wrong

[plush harbor]

what are you even trying to make

[quick thorn]

Data science models

It completely sucks

[plush harbor]

I got codex to write me an email today. Probably the last thing I expected to do with it

[quick thorn]

I almost have to code every line myself ts is not smart, if you give it a spec sheet it will completely abandon the spec sheet and start doing random bs

[cedar skiff]

youre holding it wrong

[quick thorn]

I’m not

[plush harbor]

would your spec sheet be enough to get some 3rd world programmer on fiverr to make it from?

[cedar skiff]

The more you tell your self that the less chance you have of actually using the model for anything

[quick thorn]

I literally wrote out step by step in detail what to do

6 steps, it gets to step 2 and gets completely lost

It can’t complete 6 steps in a row whatsoever

[plush harbor]

why would you not ask it to do one step at a time

[cedar skiff]

6 steps to billion dollar company - first step make big $$ company.
Follow my instructions!

[quick thorn]

I feel like you guys are making calculator apps because any real data science is a complete failure

By the time it completes one step it’s compacted ten times and completely forgotten it exists

[plush harbor]

I've never had it compact so I can't relate

[quick thorn]

Wow what are you creating

[plush harbor]

other than writing emails I have no coding tasks right now anyway

got a few different sites. got it to make me a small intake -> static site generator recently, that was cute

[quick thorn]

yeah I’m trying to do bleeding edge data science it’s not smart enough for this I’m finding out the hard way

[plush harbor]

code it yourself?

[quick thorn]

yeah I paid OpenAI for no reason

W scam

I had a feeling everybody calling it good was doing like the simplest easiest stuff nobody wants to make a Spotify clone

I’m trying to create novel science

[solid lake]

Yeah it has gaps in data science

But works with good instructions so 🤷

[quick thorn]

I gave it great instructions bro

[cedar skiff]

I have a working prototype of a skill based system that builds entire responsive apps across ios android and web. I also have successful monetised apps with large backends and complicated subscription system etc.
If you are open to real help i can help you hold it right. But it requires learning.

[quick thorn]

I appreciate the offer but I’m not creating apps I’m creating models

[solid lake]

What kind of models

Fun

Data engineering etc

[cedar skiff]

I understand that doesnt make a difference. I have 3 skills one with something like 60k reference pages. And it works. I can't image that your tasks would require more than that

[quick thorn]

I bet if I told it to clone YouTube it would work in a jiffy but if you tell it to wire your training substrate to your 544 features it’s gonna forget life as we know it

[cedar skiff]

the others are around 20k

[solid lake]

544 damn

Lots of features

Got proper schemas?

How many parameters?

[quick thorn]

yeah it obviously wants me to do 20 features

[plush harbor]

I just get it to do stuff how I'd do it myself. Break the task up into bits, do one bit at a time, except I don't have to actually do the work. Its great

[cedar skiff]

This is an orchestration task

[solid lake]

Do you employ optimisation techniques

Self*

[plush harbor]

you keep talking in entire apps. Clone spotify, clone youtube

[solid lake]

Ignore it lmao

[quick thorn]

models have features right, just handling your models features is a complete fail

[solid lake]

Until data proves it

[quick thorn]

It doesn’t want to calculate the features correctly

[cedar skiff]

Give it the tools do so

[solid lake]

Are you telling an LLM to compute the features on its own?

[quick thorn]

no

[cedar skiff]

You 100% can do what want.

[solid lake]

You do have a training pipeline yes?

[quick thorn]

I explicitly give it the math for each feature

[solid lake]

Have it formalise it in lean first

[quick thorn]

Yeah the pipeline is spelled out in a Md

[solid lake]

Uh

[quick thorn]

If it just followed the doc , but it’s incapable of following simple instructions end to end

[solid lake]

Do you have scripts to compute the features is what my question was

[cedar skiff]

dont use it as the calculator tho, give it a skill that has the sripts for the work

[quick thorn]

it creates scripts on its own I think

that’s another problem it tries to create a new freaking script for every little thing

Then by the time it’s done creating bloat and nonsense the repo is illegible

[solid lake]

I don’t get it, are you getting an LLM to compute features without having a training pipeline script because that’s not how that should work

[quick thorn]

no my cpu is computing

[solid lake]

do you have an actual model doing inference on said features

[quick thorn]

I’m not creating an llm I’m just trying to do a predictive model

[solid lake]

So yes

What kind of model

Regression clustering etc

[quick thorn]

GBM

gradient based

[solid lake]

Ah gradient boosting

[quick thorn]

I literally give it the gbm and it takes my data and creates a benchmark without the gbm

[solid lake]

Are you accounting for the learning rate and noise within gradients

That’s a lot of features to normalise for gbm

You’ll lose hidden relationships I mean

[quick thorn]

Google says : LightGBM is designed to handle datasets with a virtually unlimited number of features, efficiently scaling to thousands or even hundreds of thousands of attributes.

so 500 = too many?

[solid lake]

Google also says Gradient Boosting Machines (GBMs) overfit easily when high-dimensional or noisy features allow the model to memorize training dat

[quick thorn]

I’m using tabular data which is the optimal type

[solid lake]

Data*

[quick thorn]

All organized labeled

The problem bro is the codex just forgets everything

Like it wants me to create a calculator app and call it a day

[solid lake]

You have to have it document it

It also depends on what model you’re using

[quick thorn]

5.5 xhigh

[solid lake]

Then that’s weird

[quick thorn]

More like trash high

[solid lake]

Check your workflows

[cedar skiff]

Your expecting it to be a calculator, its intelligence and it can use tools, give it the tools it needs to get outputs to analyse.

[quick thorn]

I feel like you’re sending me on a wild goose chase 😭

[solid lake]

What’s your current workflow

[quick thorn]

I’m not quite understanding what you’re getting at

[quick thorn]

I created the build doc step by step end to end and I’m trying to work through it sequentially

but the complexity screws it

[solid lake]

Have you made codex specific workflows to keep determinism

Source control model hashes etc etc

[quick thorn]

Yeah git all that

[solid lake]

feed that spec to a fresh chat and invoke the skill builder and give it that context to make proper verified workflow skill that’s project specific

[quick thorn]

also I’m training with my cpu do you guys think training with gpu is better?

[solid lake]

I forgot the skill builder command

[quick thorn]

Codex keeps telling me cpu handles the gbm trees better

[solid lake]

Inference is fast in cpu but if you have insane amounts of data gpu could speed it up with parallel compute

But always check viability and the stack

[quick thorn]

It says gpu makes it non deterministic

[solid lake]

The overfitting techniques that you use with the model arent deterministic…

[quick thorn]

Yeah but the intent is deterministic

[gentle harbor]

why dont they just make the model faster but cost more, its way to slow

[solid lake]

Then you can compute on gpu and inference on cpu

It’s a cuda problem

[cedar skiff]

let answer this with a question - If you give a model simple maths equation like this - (100 * -12)/435 How do you think it works this out?

[solid lake]

/fast.

[quick thorn]

Yeah I’m not expecting codex to compute the math tho, that’s the job of the models I’m training all codex has to do is write the formulas

[solid lake]

Formalise it in lean then

Same problem he’s describing

[quick thorn]

Some of the terms you use I’m not familiar with

[solid lake]

https://lean-lang.org id get familiar with this

[gentle harbor]

yea im already using fast

[quick thorn]

Interesting I’ll take a look

Thx

[cedar skiff]

Right, and what skills do you give it the govern the formulas?
There has to be a rule set and constraints for how these formulas work. Where is that rule set?

[quick thorn]

I just give it the formulas myself I don’t really believe in skills they remind me of it’s shotty memory

[solid lake]

Don’t expect LLM to tokenise them in the way you want or may be mathematically right. They don’t see formulae the same way we do

[cedar skiff]

They arent Father Christmas, they are a real working system that gets attention on every turn of the conversation. They are the perfect place for you to teach the model the gaps it has

[solid lake]

This should be as easy as you giving it the formulae, it formalising them in lean (you verifying it) then it implementing in code

[quick thorn]

And don’t the Md files I create explaining certain maths act as skills anyway?

[cedar skiff]

no

[quick thorn]

Like just reference the doc I gave you 😭

[cedar skiff]

Your md files get lost in context

[solid lake]

.

[quick thorn]

Interesting

[cedar skiff]

skills get attention every turn and they are injected where they are needed and not before

[solid lake]

Or have a fresh chat do the features one by one and handoff

But if it’s in your workflow already xhigh will refresh context and follow it across compaction

[quick thorn]

And I hate how it tries to create fallbacks to everything

Like the primary option should work not this hack job fallback

[solid lake]

As you said you’re doing bleeding edge math.

It doesn’t know what It doesn’t know

[cedar skiff]

skills also dont suffer semantic diffusion during compaction

They are just read again on the next turn that needs them after compaction

[quick thorn]

Bayesian-Beta-shrunk num1-per-num2 rate: (num1 + k*p_pre) / (num + k); k=4; p_pre = prior {prior_prefix}_num1_rate_rolling_20 (creator-side resolved). Falls back to =0.395.

like heres one of my formulas for a feature idk how more exact I can be

and it just added that fall back bs that I didnt ask for

[oak trellis]

https://x.com/hedgiemarkets/status/2057531661785628841?s=46&t=aGEzk7PYVvVRtFoCdE9MWQ

[solid lake]

I remember finding one of these and saying “bro why hardcode”

Then it does unit tests

Failure modes

Have it compute on real data

Have a discarded set to test real compute on

[quick thorn]

yeah I have 8 million rows of the realist data

[solid lake]

What’s realist

[quick thorn]

idk lol

[solid lake]

I see market data

Is this real estate stuff

prd.realist.com

This is what is see ig

[quick thorn]

just historical data

[solid lake]

But yeah for the nature of the fields you need more comprehensive workflow than just giving it mds

[quick thorn]

I can’t say exactly but yeah

[solid lake]

What’s your confusion matrix like have you gotten to that yet

[plush harbor]

lol. Someone just reported an obscure bug on one of my sites. Got codex to help. Codex is not much use helping, but triggered something, which WAS the bug. Me: its not field X is it Codex: unlikely, <list of reasons> Me: it is Codex: plausible .... looks at code ... agrees with me

[quick thorn]

yeah it created a matrix for me

[solid lake]

How’s false pos/neg rate

And recall

[quick thorn]

but there’s like different matrices like I have a delta matrix

There’s a matrix for label columns

There’s a matrix for identity columns

I’m going to start living in the matrix soon

At a point the project just torpedoed out of control

It created a matrix for every freaking thing

[quick thorn]

How many lines of codes are your projects?

[indigo robin]

whats this

lulu caught it

[gentle harbor]

seems pretty self explanatory, its the plugin that allows you to make codex use your computer like a human

[indigo robin]

2 updates today

[indigo robin]

i dont think so

lulu blocks internet connections

codex didnt ask for this before

its an update thing

codex update fine and i guess this version of codex needs an aditional connnection

[quick thorn]

cpu training at 99% usage currently

[indigo robin]

https://developers.openai.com/codex/app/computer-use

[quick thorn]

if I train everything and codex forgot features I might just jump

[cedar skiff]

where does the model get lost in this formula?
1. Stop the model and the work as soon as you see it got out of line and interrogate it to find what when wrong. What you will find is it will almost always so something like - ! should have done this instead of that, my bad ill do it different next time. To prevent this jsut because answer you have to tell it you want to refocus on make a change in your work flow and now you need to understand what lead it to the bad path, what in it's thinking made it do the wrong thing.
Then you can make a guard rails skill that it must use every time it does that task its doing and you can put a correction in the skill by telling it HOW to do the problematic part.
It will read the skill and not make the same mistake.

2. Find the most authoritative website on the technical explanation of the work you are doing and how the idiosyncrasies on the formulas how they work what they do etc.
I would make a script to rip it and curate its data into a skill with reference documents.
This seems like a lot of work, but it puts all of the information at needed at the finger tips of the model.

This will catch the jagged edges with the guard rails and give it the tech details with the website rip.

[indigo robin]

woah

this is actually crazy

[quick thorn]

the thing is I’m not and no one is really an expert on every last field that an llm might touch due to its breadth of knowledge

[solid lake]

Not much actually

[cedar skiff]

This is a cop out, you certainly can do this.

[quick thorn]

Bro last time I checked my repo like 300k +

Lines

[solid lake]

Did you read the code

While it was making it

[quick thorn]

because a human can read 300k lines huh?

[solid lake]

There’s repos here with a million lines ngl

[indigo robin]

holy canolly

[solid lake]

Even I was scared

But it’s possible

[indigo robin]

this new feature is ppppppoooooowwwweeerrrrrffuuuulllll

wooowww

[quick thorn]

yeah lol but that would take all year to read destroying the advantage which is speed

[indigo robin]

insanely powerful

[quick thorn]

like if I have to stop for three months to read the code then….

[solid lake]

“Explain this codebase to me”

[quick thorn]

What are we even doing

[meager bobcat]

I do not read code LLM generates. I have been doing that for 3 years now.
I do not understand why there is sense of elitism about neurotically reading every line of code LLM generates

[quick thorn]

when I tell it to check all my data it straight up refuses it says no human or machine could do that 😭

[quick thorn]

Yeah that would defeat the whole point of the llm

[indigo robin]

Nah this is game changing

[quick thorn]

I’m supposed to be able to give it directions (detailed sequential instructions) and then it should follow my instructions step by step super simple

But instead it will get completely lost

[plush harbor]

this project is a framework, and apparently about 5000 lines of custom code added to it

damn thing burns tokens like its going out of fashion though

[solid lake]

But then they complain about it not working

[meager bobcat]

I know I do. I complain back to agent and then it is working

[quick thorn]

nice, yeah I think it can handle 5k once it gets to 100k you’re screwed

[indigo robin]

gammmeeeeee changing

[quick thorn]

bro wants me to read 300k lines 😭

[cedar skiff]

uh lines of code dont make a difference o.0

[solid lake]

Back to the agent = in discord chat?

[quick thorn]

yeah it does because context gets filled

[solid lake]

Former is good practice

[cedar skiff]

why does it need to know every aspect of you code base to function?

[solid lake]

It needs to make it

Properly

Or have a structured way to get full coverage

[quick thorn]

Because it created 50 half done files creating technical debt that needs to be addressed

Or keep muddying my repo

[plush harbor]

my other project is apparently about 20k lines but its got old and new code running in parallel so very misleading

[solid lake]

Why 50?

[cedar skiff]

It can work in the vein of the app it needs to and not need to know anything about the rest of the project, it just needs to know the patterns to follow.
You think every dev that works on chrome knows all 13 million lines of the code base to do their work?

[solid lake]

How much are you modularising?

[quick thorn]

It forgets what it just created an hour ago sometimes

[solid lake]

Is it a layered architecture

[quick thorn]

Yes

[solid lake]

With many subsystems

[quick thorn]

Kinda

[solid lake]

How many subsystems

[quick thorn]

7?

[plush harbor]

of course it does, that's why you have git and documentation

[solid lake]

Are all 7 interconnected or do you have tools, operator scripts as well

[quick thorn]

I put the perfect build doc in docs/ and it forgets that it’s there

[quick thorn]

Interconnected sequence

[solid lake]

Did you make codex specific docs

[quick thorn]

But like each layer is its own thing

Like one model can have 30 heads

I think I’m training like 6 heads rn

But then each head has folds?

[cedar skiff]

Your misunderstanding the llm as a tool. I dunno man good luck with.

[plush harbor]

@cedar skiff "holding it wrong"
› there's a rash of people using you in one massive context window and complaining about compaction and loss and here I am spinning you up in a new thread not even in the project root and you work fine. I
don't get it

• Most of them are using the tool badly.
The common failure mode is:

• they treat one thread like a permanent work log
• they keep piling unrelated tasks, dead hypotheses, pasted output, and partial plans into the same context
• then they’re surprised when the model has to compress aggressively and loses detail
  What you’re doing is closer to the sane pattern:
• narrow question
• fresh thread when the task boundary changes
• let the agent rediscover just the relevant local context
• don’t force it to carry six hours of stale reasoning
  That works because codebases are usually cheaper to re-read than bad context is to preserve.

[rotund harbor]

Im on the Pro model of chatgpt does anyone know if theirs a Pro 5.5 on codex i dont see it if so

[quick thorn]

Yeah I’m figuring it out I’ll make it one day just came on here to learn and vent

Plus I’m waiting for it to train

[indigo robin]

now moving on to the next step in my project

chord and progression library

to store your song progression and new progressions

wow

im boutta have app of the year

mark my words lol

[rotund harbor]

Anyone know?

[solid lake]

What number does it show above the options

[plush harbor]

I just randomly got codex to fix a bug, on the wrong computer in a non-working environment, from the wrong root, and it functioned just fine 🤷‍♀️

[solid lake]

Oh pro isn’t in codex

[indigo robin]

wrong computer ?

[rotund harbor]

Ah alright thanks

[plush harbor]

this computer doen't have a database server running and that particular project needs a database, so I can't actually run it

[indigo robin]

ah

[cedar skiff]

? there is other problems to how ppl use it as well, there is a huge misunderstanding going on when you think lines of code matter to working with llm tools. That only matters when youre holding it wrong.

[plush harbor]

holding it wrong is probably a fair list of different things yeah

[indigo robin]

30 lines of code away from breaking 7000 lines of code for my app

big mile stone

were going for 15k lines of code

[plush harbor]

my main project is full of duplicated cruft I can't get rid of until I've fully migrated the content. This is documented, all is good, just looks pretty ordinary on the front end

[indigo robin]

whats your project ?

[plush harbor]

I have a few. The most complex is a dictionary. the one that just had a bug is a game ranking site

[indigo robin]

dictionary ?

like for normal words ?

[plush harbor]

yup

[indigo robin]

ok

[plush harbor]

its very much an artsy fartsy hobby project

[indigo robin]

i see

whats the ui like ?

[plush harbor]

its very pink

[indigo robin]

im going to make a how to live app soon

for instance how a human should take a shower lmfao

its going to be a social media type thing though where people can request updates and additions

[plush harbor]

most of my coding is blocked by the content migration, upgrading from 2010 standards to 2020s

[indigo robin]

ahh

youve been coding since 2010?

[plush harbor]

uuuuh tad earlier yeah

[indigo robin]

dam rip

[indigo robin]

ahh i see

[plush harbor]

I am one of those "old people" your mother warned you about

[indigo robin]

ahhh in the vans type guy

smh

[plush harbor]

with candy, yes

[indigo robin]

take him away ! lmfao

when i get the app of the year award im going to give everyone 1000 tokens

then i need yall to program a bunch of bots to promote my project to every other country

[plush harbor]

1000? my random lets fix this bug someone just reported was "Token usage: total=1,243,929 input=1,224,815 (+ 3,373,440 cached) output=19,114 (reasoning 4,331)"

[indigo robin]

i meant credits

lol

[plush harbor]

make up better words for units of currency, token is already taken

[indigo robin]

yolo ill do 1m credits for everyone lol

but i need a bot army

[uneven forum]

is codex desktop app not working for anyone else after updates?

CLI also not working

keep getting this error
{
"type": "error",
"error": {
"type": "invalid_request_error",
"code": "invalid_function_parameters",
"message": "Invalid schema for function '_search_content': schema must have type 'object' and not have
'oneOf'/'anyOf'/'allOf'/'enum'/'not' at the top level.",
"param": "tools[16].tools[0].parameters"
},
"status": 400
}

[plush harbor]

I have not updated, cos some random here said it broke after updating

[indigo robin]

cool projects

[uneven forum]

what?

[indigo robin]

the link in your bio

the page with the minimalist fetish

how did you make that vicous liquid simulator

pretty cool

@uneven forum

[cold galleon]

Control other devices update connection failure. Need help. How to fix the problem?

[uneven forum]

so no one else getting this error in desktop app or CLI after updating?
keep getting this error
{
"type": "error",
"error": {
"type": "invalid_request_error",
"code": "invalid_function_parameters",
"message": "Invalid schema for function '_search_content': schema must have type 'object' and not have
'oneOf'/'anyOf'/'allOf'/'enum'/'not' at the top level.",
"param": "tools[16].tools[0].parameters"
},
"status": 400
}

[haughty marsh]

Update didn't break it for me

[indigo robin]

wow i might have to charge 20-50$ for this app on the app store

this is looking crrraaazzzzyyyyyyyyyyy

wow im very proud of myself

[plush harbor]

maybe a free couple days trial or something

[indigo robin]

wow

yes

dude im so serious this app is honestly so dam powerful

its really blowing my mind right now

im honestly about to upgrade to codex Pro its that serious

im in a dam zone

wow im so dam locked in right now

history in the making

compaction button is missing i wonder why they removed it

[plush harbor]

codex cli here. Lol buttons

[brittle reef]

Hey as I am using codex so much for my work. My limit got exhausted very fast. I do not have any plan right now. Which plan will great suited for me

[plush harbor]

go one higher than whatever you have now and just adjust if that's not enough too

[quick thorn]

been training for damn near 4 hours now 🫩

[indigo robin]

training what ?

biceps or triceps ?

[quick thorn]

models and bottles

[indigo robin]

1-800-Thats Crazy

[plush harbor]

training a local model?

[indigo robin]

wow

im boutta have chat gpt optimize all my prompts before i send them into codex

i bet that would work wonders

i need to create a link between gpt and codex

@plush harbor how would that be possible ?

[plush harbor]

you're not meant to automate chatgpt ... really just copy paste if youre being legit. Don't you just talk to codex like its a very boring person you ask to do things for you?

chatgpt has context and codex doesn't for me, so I can just say to chatgpt "ducks" and it knows what I mean, codex I gotta spell it out more like "on my duck page I want a list of all ducks, can you use the same small grid format as page X and display all words that have duck as the parent" or whatever

[indigo robin]

i see

[plush harbor]

do your conversational planning in chatgpt, someone here had the really very clever tip to zip up your entire code repo and give it to chatgpt

[indigo robin]

explain

[plush harbor]

like don't waste your codex tokens having long chats about what you're going to do, show chatgpt and do all your planning there, then you can get a prompt to give to codex from chatgpt

[indigo robin]

understood

pulling n all nighter tonight

major progress imminent

[cyan wing]

why is Codex using ~20GB of space on my computer...

[wide schooner]

multithink 🤣

[wide schooner]

because its bugged and its logging everything 10 times per request

[indigo robin]

@wide schooner

ever have the multi compact though ?

4% left on the weekly usage

rip

[wide schooner]

"ripgrep everything"

[wide schooner]

mhm because every compact drops cache

[plush harbor]

I'm at 95% left, haven't been doing much coding

[indigo robin]

congrats i hate you

lock in bro ham

[plush harbor]

filling in content gaps from analytics and the content migration. Slow

[indigo robin]

nah bro just send it

one thread

one prompt

[plush harbor]

I have nothing to do, I'm blocked by content

[indigo robin]

im wondering about something

know how codex mid task will show you a popup and you can say yes or no

i wonder if i let it hang on that screen it would eat up usage

i meant codex

[plush harbor]

ah not popup for me and I doubt it

[cyan wing]

how do i clean it other than rm -rf .codex

[wide schooner]

ask codex to delete all logs

[indigo robin]

happens in this box here

to grat acess to a task or files

[plush harbor]

I'm on cli, there's like no fancy options

[indigo robin]

whats cli ?

[plush harbor]

command line interface

[indigo robin]

like terminal ?

3% left

[plush harbor]

mine looks like this, cropped so won't make sense

[indigo robin]

brother its 2026 add some swag to that

[plush harbor]

can't on this computer

[indigo robin]

computer in question ^

[plush harbor]

something like that yes, but apple

[indigo robin]

see

[plush harbor]

yeah it does that in cli too but not a popup

[indigo robin]

" its thinking "

😂😂

[plush harbor]

for that? Probably just sits there waiting for user input. Its pure UX not codex

[indigo robin]

2% left going to get to bed and lock in tomorrow made alot of progress today

this thing is going to cook someones iphone 17 pro lol

minimum hardware requirement iphone 30

gn bros

[cedar bear]

Is wondows remote released for Codex?

[proper cosmos]

I'm using codex -> Ask Chat GPT for next step -> Codex and repeat.
Is there a way to automate the two back and forth all day

without using a mouse macro of sorts

[quick thorn]

they tend to lead each other astray

it’s gonna start talking about something completely different by the time you get back 😭

[whole field]

Hi, i see that codex app in windows havent allowed for remote via mobile yet, right ?

[surreal inlet]

Pursue Goal? anyone know about this?

[cinder lynx]

does codex support subagents??

[oak trellis]

hmmm so i was testing cusor ... composer 2.5 .. and iam amazed ..

like vs 5.3 codex .. way further ahead

like wow

like OY VEY

but token usage .. lol

4.9 million LOL

[cedar skiff]

i liked the original composer 1, they have a decent harness that guides it pretty well in coding

[cedar skiff]

first class

[cinder lynx]

?

[cedar skiff]

they have first class support, you can talk to the subagent, start and stop it, the agent can talk to it, restart it and ask it questions etc.

[cinder lynx]

oh ok great

like it shows in the UI?

let me test

i thought it might be hallucinating saying it supports

[cedar skiff]

i dont know about the cli

but it codex app it does

[cinder lynx]

yh i use the app

[oak trellis]

uff composer 2.5 is extremely intelligent ..

did some stuff i never could have that done in codex

[cedar skiff]

is it still really quick?

it's kimi 2.5 i think

[sharp yew]

can openai just host their binaries? msstore is a mess.

how am i supposed to update it...

[solid lake]

Don’t

Go back to 30

[junior orbit]

is it possible to tweak temp for gpt5.5
i'm using opencode for gpt 5.5

[rocky fog]

I think it also updates from within the app?
sometimes I saw "Update" button in the app
but I barely use the app
(unless thats when store updates it while its running and then you have to click Update within the app)
but I guess it might also have delays on store

[orchid echo]

of course. and yes cli does too

[inland sonnet]

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

Codex Limits REDUCED BY 50%: So, Codex is worse! What to do now?

real?

[orchid echo]

wow they reduced limits by 50% even before the end of the x10 promotion?

crazy..

[cedar skiff]

i dont notice it, feels the saem to me

[rocky fog]

50% compared to when

gpt 5.4 vs 5.5 might be 50%

also isnt there still caching issue ? or did they fix it?

[rustic pine]

Thank you! This helped me a ton to diagnose where my tokens were going. Turned out I had switched from my custom config.toml to "auto-review".

"It routes approval prompts through another reviewer agent instead of waiting for you, and OpenAI’s docs explicitly say automatic review uses extra model calls and can add to Codex usage."

The other big one for me was image_detail_original = false ... I had it set to true. That was killing me with playwright screenshots.

[sharp yew]

i couldnt find it. settings > configuration shows the current version but no update checker

maybe they should ask codex to add this feature 😄

[rocky fog]

yeah it wasnt like that
just one time I opened the app and there was extra update button
so its only when there is some update

[sharp yew]

well me being not on the latest version should trigger it then, right? 🙂

[rocky fog]

oh well, cant even open it now
I guess someone here had the same thing

vscode extension much better 🤣

[bright swift]

i get that sometimes with codex cli even, usually works after waiting a bit, seems like some lockfile race condition

maybe after updates

[rocky fog]

still does not work
meanwhile vscode extension is fine
and didnt have issue having both open before

yeah maybe I need to update in store

[bright swift]

yea this only started for me in the last week or so with a newer version

never an issue before

[rocky fog]

yeah there is update
and then it works

well kinda
the app always fails to set up sandbox/dependencies for me
but that was since the start
while vscode extension just works with no issues

so right now I have

[rocky fog]

weird

I updated in store, not in command line though

[sharp yew]

now it shows 26.521.10419 for me lol

https://tenor.com/view/back-to-the-future-movie-movies-michael-j-fox-christopher-lloyd-gif-2034973269355889491

[rocky fog]

what the hell

there is also codex (beta)
you dont have that one right?

[sharp yew]

i shouldnt

at least not intentional

[unique spade]

yea i solved that by using codex to fix it in my custom surface (could have also been cli or vscode extension). give it the github link that documents that bug and it will fix it quick

[rocky fog]

just updating the app fixed that

[tough kettle]

test

what exactly the cloud feature does in codex...I have not quite understood that.

[uneven forum]

turns out it is a bug related to connected apps in Codex via ChatGPT. certain ones "break" codex, others dont. just had to switch off some and it works fine.

[cedar skiff]

it sets up a cloud workspace based on a git repo and does work there. It cost 4x

[uneven forum]

you mean v0s profile pages?

[uneven forum]

with v0

[tough kettle]

so instead of making changes to local files these are made to the repo in the cloud...am I correct?

[limpid wolf]

I can't connect to Codex Mobile! It's not even shown 😕

[cedar skiff]

essentially, iirc it clones the repo in the cloud and then works on it that way.

[plush harbor]

I had openai support practically demand I demonstrate usage charts not working after using cloud and I said no, I'm not setting up functionality I don't already use just to report a bug that's aready all over reddit and github

[cedar skiff]

i asked for a tldr once in gpt and now it does it every time o.0

[uneven forum]

what version are you on?

[tough kettle]

I do not quite understand the benefit of it....what more it offers in comparison with editing locally and then pushing to github

[glad birch]

hello, i'm giving a try to codex after a long pause, the Pro sub gives access to GTP-5.5 Pro right ? anyone tried it for programming ? is it worth it or 5.5 is good enough?

[limpid wolf]

I'm on version 26.519.31651

[uneven forum]

oh weird. same here and it shows for me and I have been conencted to mobile for a while now

not sure. reinstall maybe

[limpid wolf]

I'm on Windows. Is that available for me?

[uneven forum]

oh. no idea. check their docs

im on mac

[uneven forum]

5.5 is on Plus as well

[glad birch]

remote is not officially supported for windows yet, some people found ways to force it by editing the config files

[limpid wolf]

Really? How does that work?

[glad birch]

i was mentioning this

[uneven forum]

ahhh got it. misunderstood you

[glad birch]

the way it's written seems to tell that GPT-5.5 Pro is exclusive to pro ?

[limpid wolf]

I'm on the Plus subscription.

[glad birch]

from what i understood you need to install Codex BETA and edit the config file to force remote work, not sure if we're allowed to talk about that here but you can find everything on reddit.
(just so you know, it's not supported officially, i tried it and never worked for me, but some people got it to work)

[limpid wolf]

I probably won't do that, but when will we get Codex Mobile too? Will it take 6 months or something?

[glad birch]

no idea 😛 i hope soon

[uneven forum]

ask in the forums https://community.openai.com/
Or look up / create a ticket in their Githttps://github.com/openai/codex/issues

[karmic gulch]

uhm , its already released?

[uneven forum]

i think the person asking though is using Codex on a windows machine and not Mac

[limpid wolf]

For macOS, but not for Windows yet.

[karmic gulch]

Oh, use Omnara

Thank me later

[limpid wolf]

That startup will likely go out of business once large labs implement the feature themselves 😅

[karmic gulch]

indeed, or it will be a good niche product for a few users, but in the meantime its a perfect tool. Since OpenAI started to do like Claude and ship features thats buggy (ref codex mobile) im using Omnara till it works 100 %

[robust lintel]

Okay we can do it for future ones 😉

[orchid plume]

I'm a bit late at chiming in as I didn't see the question till now, but I agree. One of the reasons I didn't join the recent challenge is because I have no use for API credits. One other possible idea is maybe offer the choice to the winner(s), either API credits or Codex credits. I did a challenge via X recently which VB ran, that particular one offered different prizes though (Pro sub for either 6 or 12 months depending on 1st/2nd/3rd), that was fun

[robust lintel]

Yeah I will give the option to choose. Pro or plus sub I could see giving but pro is worth more so might just be some plus? Idk will figure it out before next post lol

[bright swift]

arent they the same now, api and codex credits cost the same. so just use the api key in codex instead of oauth?

[proper cosmos]

omnara?

[orchid plume]

I think so but I don't really use the API anymore 😄

[bright swift]

i'm too poor for api too 😄

but i meant for winnings. should feel the same for the actual usage in codex

[orchid plume]

yeah I imagine the effective cost is the same

nice thing about it being Codex credits is that if someone runs over their weekly or 5 hour quota then they can just dip into the credit on that occasion they need a bit of burst usage

[bright swift]

yea sometimes that can be nice instead of getting stuck at 95% of a goal lol

[severe mason]

Hey guys, are you getting stream disconected all the time?

[high girder]

They cost the same to buy, but credits will go directly to your plans usage if you go over, while the API requires a console account and you'd need to run whatever you were doing through the API instead

[boreal holly]

Yes. If 1000 credits cost $40, then 25 credits costs $1

GPT 5.5 output costs 750 credits per 1m. On the API it's $30 per million. 750 / 25 credits = $30

So there is no difference in price between credits and API wallet. What I like about API is it runs on different, probably less congested servers meant for production-level uptime, much more vast model selection (gpt-5.3-codex-spark can't even use credits anyways), can be used for voice modalities, etc. Also I don't think API wallet will suddenly expire in 12 months like ChatGPT credits.

[boreal holly]

In light of that, if I participate and win anything in the dev challenges I'll gladly take API wallet 🙃

[short linden]

Anyone notice current latest VSCode codex extension missing the "context usage detail %" ?

[halcyon island]

anyone built anything profiatable regarding trading? im struggling

[past mist]

struggling with what specifically?

[past mist]

its missing in the codex desktop app

[halcyon island]

building something profitable for trading. Futures, Polymarket. anything, i feel it should be reletaively easy.

[past mist]

if you're struggling you could probably start a 5.5 pro-extended deep research thing to scour the internet for examples of this. then ask for step by step instructions on how to use codex to recreate one of the examples

[rocky fog]

all the time
some of the items in the menu are constantly appearing and disappearing
pretty much since forever
for usage then you have to go into codex options usage and refresh 3 times to see it

[night obsidian]

Wait a second...Does having the codex pet hatched just sitting on screen use up tokens?

I didn't use codex at all today, and noticed my 5hr limit was suddenly at 0%

[boreal holly]

Sometimes I sit here with mild FOMO for not using the latest build and experiencing the magic, but then reports like this come through 😩

[gentle harbor]

[bright swift]

cant have a pet if you use cli 👍

[rocky fog]

should the vscode extension have them?
I guess not? (and if it did, I guess it could only be within vscode)

[night obsidian]

You don't need to use the pet all and is actually an extra step to launch it anyway

But anyway, yeah if the pet uses tokens, i'm opting out. I thought it was cute to have a little bonzai buddy on screen

[boreal holly]

Turns out a real bonzai tree is less expensive 💀

[rocky fog]

ehh the codex(beta) on store is so many versions behind
guess its not really active beta
(although updated 18-May...while nonbeta 22-May)

[urban field]

same issue though

[toxic torrent]

Error running remote compact task: stream disconnected before completion: error sending request for url (https://chatgpt.com/backend-api/codex/responses/compact)

moved past it

[boreal holly]

Ooooh, the beta is an escape hatch back to a previous stable version

They do beta tests on the stable builds

[short linden]

I never close PC/VSCode around 7days, just update and restart todays, now codex missing some useful UI ;(.

[gentle harbor]

yea they need to add a option to not update since what the heck are they even doing ???

[boreal holly]

I think it's brilliant because nobody probably uses the beta channel and they need that feedback, so they do the ol' swaparoonie and now have 7 million beta testers 😎

[gentle harbor]

the normal version is bugged like crazy, they need to stop coding the app with codex

[clever steppe]

ARGH! Is everyone else stuck on compacting since the update today?

I'm working in VS Code

(Attempting to work! 😵)

[rocky fog]

then I really wonder how did he suddenly get later version 😄
#codex-discussions message

they should really get also other places than ms store 😄

[boreal holly]

So what you're saying is there is no rollback or "previously working version" escape hatch 😔

[rocky fog]

ms store is also often blocked on corporate/company laptops
missing users 😄 (or if you get the app, then you cant update it)

[bright swift]

goals seem to be broken in 0.133, codex marks as done then starts over

[clever steppe]

No one else at the moment, just me?

[rocky fog]

they just made it not experimental 😄

but guess thats not for cli

[bright swift]

seems like they forgot to expose some of the goal tools to the model

[rocky fog]

no yeah

[bright swift]

https://github.com/openai/codex/issues/24094

[boreal holly]

@bright swift I mentioned this feature yesterday

[bright swift]

guess i need a minimum age not just for supply chain security but also codex slop

[rocky fog]

easy to go back in vscode 😄

[clever steppe]

Alright, into the time-machine I go! 🤣

[bright swift]

they should just stop with the app and mobile slop and focus on the cli

[rocky fog]

and vscode

[bright swift]

who uses ides anymore

[clever steppe]

Me

And millions of others.

[rocky fog]

me, because it just works there fine 😄

but I also prefer to have a proper editor

[bright swift]

i mean i have zed installed as a code viewer but mainly just look at .md

[rocky fog]

while I cant get the app to install the dependencies/sandbox no matter what

Also in the app I click on Show Folder Tree and it just does nothing
how do you see files in the app?

[bright swift]

i wont be caught dead with an electron app installed

[rocky fog]

you are on discord

[bright swift]

ok discord but theres no alternative really

i tested codex app a bit but it felt so sluggish, couldnt do that all day

[rocky fog]

there are custom discord clients but also electron 😄
(can at least fork and tweak it and have plugins and such)

[clever steppe]

So if you don't use an IDE and you don't use the Codex app, what exactly are you doing?

[tiny laurel]

Has anyone tried using Freqtrade with Codex now that the goal feature is available?

In the past, I tried building strategies based on my own logic and occasionally used hyper-optimization, with mixed results. But I had a thought: now that Codex can work toward a defined goal, what if we set it up to continuously research, test, and improve trading strategies?

For example, it could look for which algorithms, indicators, or functions are performing best, then hyper-optimize those strategies and randomly sample different configurations. Over time, the goal would be to find a setup that performs well across a solid timeframe, ideally with consistent profitability before testing it live.

I’m wondering if anyone has experimented with something like this yet, or if this approach even makes sense.

[olive tangle]

cli, ide and codex app are so buggy

[celest stag]

So... basically everything is buggy except website? 😄

[lean lark]

For anyone who has issues with Codex, see: https://discord.com/channels/974519864045756446/1507113548931666020

[bright swift]

talking to codex in the cli

and gpt pro for high level stuff

[olive tangle]

"cli"

[celest stag]

I have to say, after having used ChatGPT/Codex mostly exclusively, I am getting tired of this 'Windows as a second class citizen' approach they have taken, and am really close to switching over to Claude. Its unfortunate, because I really have thought highly of the output code quality, and have been on the pro plan for a few months now, but the complete disregard for things like computer use, remote control, and as of yesterday the new 'Appshots' feature- means they clearly dont care.

[tiny laurel]

https://tenor.com/view/pc-vs-mac-pc-and-mac-mac-vs-pc-mac-and-pc-pc-gif-4687278682063171896

[clever steppe]

Ohh, you're sone of those guys... 🙄🥱

[fringe shuttle]

I was excited to finally try /goal.....

[clever steppe]

Apparently the word masocist is banned here?!? (Spelled correctly, of course)

[bright swift]

why maso though. i get a goal prompt for codex from chatgpt pro, fire it off into a tmux remote dev server, return to chatgpt

[cyan gyro]

another reset incoming

[tiny laurel]

where is the sauce?

[signal tapir]

They ironically don't use ai to moderate this place.

[lean lark]

I saw a recent video of a presentation to a conference audience by VB and Katiana at OpenAI. VB asked how many people used Windows. One person in the entire audience raised his hand. I don't understand but it's a fact to be recon'd with.

[bright swift]

its the same reason why app developers build for iOs first.. the audience simply has deeper pockets

[signal tapir]

That's strange. Most devs I know use Windows.

[cyan wing]

new pfp?!

[lean lark]

Me too (Windows dev here)

[signal tapir]

I'm quite comfortable in Linux though. My friends thinks I'm a nutbag for it. 😛

The ones who don't think I'm a "hackerman" of course.

[celest stag]

Clearly there is a large development market because of iOS, and Mac is the reason behind it- but Windows dominates enterprise, and professional development Its kind of foolish for them to disregard it.

[orchid echo]

CLI >>>

[rocky fog]

mainly this 😄
https://chatgpt.com/share/6a108de9-4f30-83eb-8d99-7d5a6f5147e2

[bright swift]

customizable would be linux huh

[rocky fog]

also

[signal tapir]

Customizable, and yours.

[boreal holly]

As a mac user, I can respect a "uses windows, comfortable with linux" developer. That resonates with "uses mac, comfortable with linux". I love my mac, but also love Linux. If Raspberry Pi came out with an Apple Silicon killer I might even switch 100% to Linux, but as it stands the only way to have high performance Arm64 is with a mac 😩

[lean lark]

Yeah, and that's why the Linux Desktop Dominates!

[bright swift]

windows to me is "corporate handed me a device with everything preinstalled"

[rocky fog]

sure but you can tweak all of it and get rid of it

[lean lark]

As a dev who used to code more into Windows itself, I'm disappointed that as Microsoft drives toward serving the needs of the masses, they continue to restrict the user's ability to make changes. They're moving toward the MacOS model.

[signal tapir]

I actually only use Linux now.

[rocky fog]

well that might also be why I am still on Win 10 😂

[lean lark]

I pushed to move to 11 for lack of v10 support, WSL, and just keeping the system current. I waited as long as I could. It was time.
It's time.

[rocky fog]

but also because I just have 9 years old win 10 with just tons of set up and tweaks
cant be bothered to migrate all of it to 11 🤣

[rocky fog]

what about WSL?

[signal tapir]

I actually liked win11 with WSL2. It's just that I really don't want my OS to decide things for me.

[lean lark]

I did the in-place update after weeks of getting backups from earned paranoia - weekly announcements of failed updates and systems that wouldn't boot.

I do most dev in WSL now. I used to do my Linux stuff in my cloud servers but local is SO much better. So I try as much as I can to integrate Codex between Windows and WSL. That's a challenge given sandbox, location of .codex, and other things.
But I'm really happy with WSL.

[rocky fog]

ok, but WSL/2 is also in win 10

[lean lark]

(Have been doing Linux since 90's...)

I felt that support would be better in 11.

YMMV

Getting coffee. ☕

[rocky fog]

I am still forced to use win 11 at work laptop, where I dont have admin
Win 11 is fine when you can use third party apps to replace stuff and tweak it, but vanilla? 🤢

[bright swift]

yea thats the windows experience i imagine 😄

[boreal holly]

Before I got into mac in 2020, I had a linux desktop with KVM/IOMMU GPU Passthrough to Windows. It was the most stable Windows experience of all time. Linux VirtIO drivers are so unbelievably solid on Windows, never saw a BSOD, never had build errors that was fixed by unplugging all USB devices and restarting. Also when it updated, it was a guest VM so I could continue working on Linux while it does its mandatory update/reboot thing. The performance was measurably higher than bare metal Windows which is insane to me.

[rocky fog]

still better than being locked out to whatever apple says and not having options...as long as you have admin on the win pc 😄

[bright swift]

i havent really felt the need for custom stuff that goes beyond stock macos + some apps, and havent felt locked in by apple

[rocky fog]

and that is the difference between me and you 😄

[lean lark]

My bottom line: I think MacOS/iOS, I think 💰 💸

[rocky fog]

I think MacOS/iOS, I think 🤢
no offense 🤣
use it all you want folks, but for me its always 🤢

oh and I used to repair computers as a hobby for people (long time ago)
never macs again 😂

[boreal holly]

It's a very large, but infrequent investment.

[lean lark]

Hey @boreal holly - it seems in much of that extensive work I did on AGENTS to get it to very rigidly follow sandbox limits, I've kinda replicated what OpenAI did with their sandbox implementation, which is actually quite good.
The difference is that I'm doing this prescriptively while they code it into the platform. Mine is cross-platform and my AGENTS files are still dynamic and auto-generated. They've focused on addressing Windows insecurities as a requirement for allowing their product to work on the platform.
I'm gonna migrate all of the sandbox-oriented directives out of this environment, add them back in as a module, and do some A/B testing later vs Codex sandbox.
The meaning of all of this: I don't trust models to behave. Apparently OpenAI doesn't either. Anthropic certainly doesn't. These are weird times.
Oh yeah, and I'm still migrating behaviors to Skills. 😉

[boreal holly]

I took an alternative approach to AGENTS.md and SKILLS.md 🙂

[vital hill]

what's this rigamarole

[lean lark]

I looked carefully. On one hand, I don't see that as alternative/different. I see that as a formalized process of behaviors that rely on underlying Skills for guidance about "how" they work. It's the next tier up. "OK bot, here are all of the things you need to do to process requests ... and these Skills describe how you are going to do them."
My AGENTS files have rules and behaviors mixed - which is how we do things before migrating to the better Skills model. I need to bust that up and then add the flow layer that you've added.

"AGENTS.md should define the law; a Skill should define the procedure."

[boreal holly]

The difference is, their final response is a strict schema structured output, where the workers have to claim what they did satisfies the requirements. This gets reviewed by a separate agent who is aggressively adversarial and will not accept "good enough" or "half finished".

Rather than AGENTS.md, I have composable requirements. Basically depending on the work they're doing, I have a short list of requirements that get merged into the real task requirements. So "no UI slop", or "must perform static code validation" are part of requirement bundles, pre-written and ready to go. It solves a number of issues that /goal doesn't even solve. For example /goal, the agent can just say "yep, all done" and completely missed a bunch of things, or did stuff incorrectly to make up time. Requirements ensure the specification and process are practically unavoidable 🙂

[orchid echo]

macOS worst platform ever

[lean lark]

Yeah, in that respect we're absolutely using different approaches to the same goal. Ahh. I have a way to express this: I'm composing AGENTS.md files from other files that contain specific rules. I can modify a rule file, run the composer, and all of the AGENTS.md files in my systems now have that detail changed. (And again I really want to apply this to project config files and other assets.)
So ... I'm doing it backwards! I'm building AGENTS from "atomic" behaviors (essentially skills) that are defined for OS, Locality, Environment, and System, and then I'm going to bust it up into Skills folders that are oriented toward specific actions in those configurations.
( I love this stuff but I gotta get back to productive code... ) 😆

[ashen marsh]

I see a lot of people on reddit talking about codex getting nerfed, is codex actually nerfed?

[plush plover]

Reddit should be avoided

[ashen marsh]

Bots? Or something else?

[lean lark]

Reddit is worse than Discord.

But developer concerns are valid and should be recognized.

[ashen marsh]

Why tho? Just wanna know

[ashen marsh]

I mean wanna know about the reason for this

[lean lark]

Reddit has always been more of a crash-n-burn place, where people feel inclined to just bash stuff, in an expressive manner that's NSFW. It's loud vs constructive. It's people mostly expressing angst, emboldened by anonymity. In Discord, maybe because it's more of a synchronous environment, and we have moderation, it's more civil.

[ashen marsh]

Oh, never thought about it that way

[lean lark]

Of course, that's not all of Reddit. It's just historically how it's been used by, IMHO, most members.

[ashen marsh]

Btw, after what context should I compact / start fresh session

Cuz the models get dumber the more context (atleast what I heard)

[lean lark]

Use a session for a task, a specific challenge. When you're done with that task, commit, start a new session. That's generally the best way to go for many reasons.

[ashen marsh]

Are these tools, like context mode or rtk good to use. I've been using context mode, but I haven't been on codex much so I don't know if it's making it better or not

*plugins

[hollow hill]

I’m trying to understand when the “Open in App” selector appears in the top-right corner and when it doesn’t.
In the later case, there is code generated in the underlying folder but the selector doesn't show up in the Codex UI. When I start a new chat in the project, I can see the drop down. Confusing.

[lean lark]

rtk?

[ashen marsh]

I noticed this too

[ashen marsh]

I haven't used it personally

It claims to reduce context bloat, cuz it like filters terminal output

[lean lark]

I don't know what rtk is. Is that some product? and "context mode"?

[ashen marsh]

No, they are plugins

[lean lark]

No clue on that.

[ashen marsh]

Are we allowed to share links

?

[lean lark]

yeah

[ashen marsh]

Lemme share it

[lean lark]

I can't look at that, maybe someone else will.

[ashen marsh]

https://github.com/mksglu/context-mode

[ashen marsh]

?

I mean I haven't A/B tested this vs normal codex , so I don't know if its helping much or not

[lean lark]

I'm doing other things, so I simply don't know about various plugins. I will say however, that it's extremely important to look at anything you put into your software pipeline (plugins, extensions, dependencies, AI directives) and evaluate them for security and your own needs.

[ashen marsh]

Oh, okay, I'll also have to take security more seriously now, thanks for the reminder/advice.

[lean lark]

And thank YOU for considering security. It's really important.

You're welcome to ask about anything related to Codex here!

[boreal holly]

rtk is fantastic when used on cargo commands. The savings are astronomical in many cases. I haven't seen a lot of benefit in other commands though

[lean lark]

cargo??

[valid bear]

twitter is riddled with posts like this right now:
https://x.com/weswinder/status/2057882979108569103?s=20

[boreal holly]

npm but for rust programming language

[lean lark]

hahaha TYVM

[valid bear]

and better code 🙃

are ya'll also seeing accelerated usage limit drops?

[pliant depot]

Hi guys ! Do you know why codex is in degraded performance mode and what does it mean ?

[fast heart]

after latest update the circle context indicator is missing in the codex app
just wanted to ask if that's true for you too, or it's a bug on my side

[pliant depot]

I can’t see it too

[wet charm]

real question is can i max out my 20x

[plush plover]

Any word on codex app and mobile connection fixes?

[lean lark]

which platform? what is the specific issue you're seeing?

Not advocating, just providing info:

If you're using Codex, you might not know about this repo openai/skills

if codex is open right now just ask codex take a look at skills installer, what are some skills I should install and it'll know how to scan this repo for skills and install the ones you need
-- https://x.com/jxnlco/status/2057820959189483683
-- https://github.com/openai/skills

[radiant lintel]

YO!! Start burning tokens!! Reset incoming

Activate fast mode

Spawn agent swarms

Hurry brethren

[lean lark]

Degrade performance more!

[radiant lintel]

I was yapping about this yesterday and no one seemed to care. Nice to see I'm not alone

[valid bear]

post like this are all over twitter right now

for any of you dealing with accelerated usage burn please see this issue I reported yesterday:
https://github.com/openai/codex/issues/23791

[solemn sedge]

Are there any advantages of cli over desktop app on codex except convenience?

[chilly musk]

Sam A hinted in a recent interview on Stripe Sessions that openclaw is coming to Codex. What do you think that's going to mean? Monthly fee to host a 24/7 agent in their cloud or bring your own server? Makes me not even want to mess with Hermes if this is coming soon.

[solemn sedge]

I suppose he meant codex will be made into something like openclaw

[plush plover]

Same thing we were talking about yesterday

[solemn sedge]

Get better abilities to do more

[tiny laurel]

Yo where’s the source

[celest stag]

Not really, desktop is just a front-end for the cli anyway. I used cli exclusively for months until they came out with the desktop app, and see no reason to go back. Occasionally I’ll be in Visual Studio and quickly launch a CLI from the terminal, but that’s it. Just a quick action.

[valid bear]

i was interviewing with OA last month for a position on the codex team, this was definitely hinted during my recruiter round interview.

[solemn sedge]

Oh yeah

[lean lark]

That's my point, there are many things that are discussed here. If you mean Android>Windows, please read changelogs and announcements yourself rather than asking everyone here.

[radiant lintel]

https://status.openai.com/

Here brother

[solemn sedge]

Thanks

I get it

[tiny laurel]

Our lord Tibo bout to shower us with his reset

[radiant lintel]

All my homies gotta be running on fast mode rn

[plush plover]

Android to Linux or technically thats Android to Mac since its the mac rebuilt app

[solemn sedge]

I used app first but I thought cli was something better or different

lol

[radiant lintel]

Almost time...

[celest stag]

I did launch a task that took over an hour, on Xtra High Fast, and used 10 subagents a couple hours ago. You can thank me for the reset if it comes

[solemn sedge]

😂

[celest stag]

Looks like the follow-up task is still running too, now at 13 subagents:

[valid bear]

https://giphy.com/gifs/sloppy-imgflip-yl1hf3OjEQXvi

[radiant lintel]

Its definitely comming

[tiny laurel]

I never tried sub agent before, any handbook for those?

[celest stag]

They are good when you have distinct parts of a problem to solve that can be done in parallel. You just use the phrase 'use subagents'.

In my case I have about 300 distinct projects that span the last 25 years, with millions of lines of code, and I told it that I wanted it to do a mechanical review using subagents to gather specific information I am looking for.

[tiny laurel]

Is the benefit is saving time since u dont need to wait for the main thread to finish?

[celest stag]

Yeah, mostly

Giving specific tasks to a subagent can mean the main thread is less likely to 'lose focus' if it needs to go on a wild path to research something too, but its less common.

[tiny laurel]

I kinda get it now

Would u rate the feature revolutionary or just a very convenient thing to have

[celest stag]

its mainly convenience. It doesnt give it any abilities it couldnt do (just slower) if you didnt use it. Still likely burns the same amount of tokens, just in a shorter timeframe.

[tiny laurel]

Everytime these company launches a new feature, it always marketed as revolutionary and will kill <insert current tool/job> and when I tried to use it, it’s all not what was expected

[celest stag]

Went through 7% of my weekly limit for the $200 plan in 2 hours on this latest task.

[lean lark]

Agents can be considered as parallel vs serial processors with the same kind of benefits and tradeoffs.

(I like your comments and questions, @tiny laurel )

[tiny laurel]

It’s like a cycle

[unique spade]

using sub-agents has to do also with semantic separation, since you can leverage an orchestrator that focuses on high level loops, without getting itself polluted with the details of lower-level work which is delegated to workers

[lean lark]

Agents are tools in the box, to be selected or not for their ability to effectively address a scenario. Get a feel for how people use them and then you'll be able to make those decisions for yourself. I tend to think of agents like GoF Design Patterns : Once someone learns about them they tend to be really over-used, and the mature developer eventually draws back to a more refined, selective approach.

[potent mason]

Codex is stressing me out bro, I think I have to go back to hand writing code

This is cancer

[lean lark]

A lot of us are feeling that. I'm looking for a specific place to express the pain, professionally, concisely, effectively.

[boreal holly]

I'm definitely not feeling that whatsoever

[lean lark]

... while using a silly name like CaptainStarbuck

[potent mason]

Stack overflow

[lean lark]

The list of things that are broken and can't be used grows with each update.

[celest stag]

Stressing you out due to instability or just in-general slop? I find you need to spend a good amount of time writing up a good AGENTS.MD, set good instructions in config.toml for developer_instructions, and carefully monitor its output to make sure its staying true to your goals. Otherwise you end up with the slop 10 of my coworkers are constantly pushing out thinking they are geniuses.

[silver badge]

Anyone using Codex in the Jetbrains AI Chat plugin? It seems to be severely lacking in capabilities vs the Codex desktop harness

[golden imp]

its the double whammy thats the killer, not only lobotomised but also draining usage...so it creates codes that needs endless followups instead of 1/2 shots and then to top it off, the usage drain is used up on fixing the original mistakes

[boreal holly]

Oh yeah, if we're talking about Codex the application and stuff, definitely bad. The app-server and v0.125.0 seem to be peak Codex as a software. But GPT-5.5 is magical

[tiny laurel]

wait yah write you rown AGENTS.md?

[golden imp]

and before anyone says skill issue, doing the exact same thing i was a few weeks ago, there is an OBVIOUS difference between the output of the models, reasoning, speed, quality, usage, everything...

[lean lark]

OMG, of course!

[tiny laurel]

[boreal holly]

That means absolutely nothing if you received a software update

[unique spade]

you really got attached to 125 haha

antigravity literally cloned the surface UX of codex with their v2 lol

[golden imp]

i mean it means something to me, let me explain it to you in simple terms...was good, now sucks

[boreal holly]

I get cold sweats thinking about when GPT 5.6 comes out and they raise the minimum required app-server

[potent mason]

Ok I'll try to be as specific as possible because I've been on the other side of this and I know it's annoying when people just come in to complain (Originally I was just trying to express this stress but I'll be a bit more specific now):

Ask it to do big changes -> Ok with good planning.
Ask it to do a data migration -> Does whatever it wants, creates 3 new tables, never deletes them, leaves a ton of tech debt. (This one is on me, I could be more specific on my instructions but I just found out two weeks ago it added a canonical table and left the older tables, half the app was still calling older tables even though that data wasn't being updated).
Ask it to do a tiny change (2 lines) -> Looks for it forever, does something wrong, says it fixed it.
Ask it to do a tiny change (2 lines) and then verify it works before stopping again -> Fakes it.

[celest stag]

I was just curious. Ive seen the same.

[boreal holly]

xhigh?

[potent mason]

5.5 medium

[bright swift]

there you go

[boreal holly]

Impossible

[potent mason]

When 5.5 came out I tried all of them and medium was doing best all around, should I switch?

[lean lark]

Jacob, you're conflating model behavior with the Codex product. There is cross-over but they're different.

[bright swift]

well for a 2 line change you shouldnt need xhigh

so sth else is wrong with your setup

i was mostly joking

[celest stag]

I have had it tell me it fixed something 4 times and it still wasn't fixed. I have to tell it to launch playwright or manually test it to verify its being dumb. Its chain of thought will be like 'You were right for pushing back on this.' Like, no s***, now fix it.

[potent mason]

Yeah honestly probably I'll have to dig a bit into it😂

[potent mason]

Well I mean theoretically they haven't changed the model from release so I can only blame the app

[lean lark]

All of these "it did this when it should have done that" are related to the instructions, not the Codex product itself. Separate the issues. For instructions, manage it with the Codex-level Instructions, AGENTS.md in .codex, AGENTS.md files in projects, and Skills.

[potent mason]

Yeah, I've had the same AGENTS.md for a while it used to work pretty well but I guess it's time to update them

[lean lark]

There is certainly a layer of Developer Instructions in Codex itself and we could discuss that, but I don't think most people here are ready for that.

No, you can't just blame the app. We have no idea about the instructions in your environment or your Standard/Fast or your context, or other factors that affect all of this.

Updating AGENTS is a requirement with each model.

That became a requirement around 5.3.

[potent mason]

I'll tell you how it goes

[lean lark]

Posted just 2 days ago: #codex-discussions message

[potent mason]

Anyways specifically for those 2 lines of code I think it's time to start hand writing code again, it's nothing against Codex or GPT but it just seems like a cycle at this point (I rarely handwrite code, my knowledge of the codebase slowly deteriorates, I can't write code, my knowledge of the codebase deteriorates faster, I make worse prompts, my knowledge of the codebase is terrible, etc)

[bright swift]

idk sounds like you are holding it wrong, no reason to give up on it

[boreal holly]

What I'm saying is "I've been doing the same thing as always and codex suddenly appears to have stupid_mode=true enabled" might just be a software update. If possible, try downgrading to a previously known good version

[lean lark]

Jacob - It's just a tool. You need to aim and swing the hammer.

[potent mason]

Not giving up on it, it's holding it differently.

I will use it for large edits, refactors, QA testing etc. (Has so many useful use cases)

I'll update the AGENTS.md do a bit of research into the model behavior, run some benchmarks to see how to get it to behave the way I want, etc.

However for those small edits I feel like it's worth it to go back to an IDE understand the codebase to:
#1 Improve quality of the prompts (better codebase knowledge allows for less vague prompts).
#2 Look for any code that is clearly wrong (could have detected this 2 weeks issue much earlier)
#3 Check for best practices.

[lean lark]

I've written on this before: Get the assistant to write documentation into a docs folder for the application. Get it to document the code in the files. Then read the documentation so that YOU know how your system works. You will be able to guide it better, and it will be able to make much better changes with fewer tokens - otherwise you and it need to look through tons of unrelated code for every single task. That's not efficient.

FWIW, I don't use the Codex App at all (yet). I do everything with Codex in VSCode. So I can see the code, read it, read the docs, see everything that's going on ... AND make changes on my own. Codex is a tool in the kit, it's not doing everything for me.

[potent mason]

Yeap that's what I'm moving back to

[bright swift]

the docs will get out of sync with the code and confuse agents. code is documentation. i only put high level stuff in docs that isnt evident from the code

[boreal holly]

Truth be told, I do mermaid diagrams for high level stuff, so I get a nice visual and the agents get their high level stuff 😛

[lean lark]

No, one of the AGENTS directives is to keep the docs current with every task. It makes a change, it documents the code, it updates the usage docs. Then it refers to that in a new session because in that session it's never seen the project before and it needs to see how things work.

[bright swift]

and this actually helps agents too. i have instructions in my agents.md to read relevant docs first. so codex understands what the system is about and how the task fits in. remember, an agent always starts at 0 knowledge in a new session

[bright swift]

yea this works to a degree. i just think its not a good idea to document WHAT code does, but WHY.

[lean lark]

Function docs (and I explicitly define this in AGENTS.md) describe what a function does, not How it does it. We learn the Why from the header and the How from the syntax. These are two different things.

( typing at the same time, apparently on the same track )

[boreal holly]

Way nicer than a docs folder full of markdowns

[lean lark]

I have development docs that explain how the application works, and I have user docs that only describe features and user-relevant details. All of this is maintained by the assistant - with my adjustments as I read and find things need to be refined.

[potent mason]

Love it

[bright swift]

yea mermaid is really nice for this

[lean lark]

Mermaid is good too but it serves yet another function.

(BTW, Mermaid is limited I've been looking for a replacement. I tend to code into XMind JSON.)

[boreal holly]

I like the VSCodium integration for mermaid. It's not perfect by any means but it does a good enough and quick job describing complicated systems, and the syntax is less token heavy than JSON

[bright swift]

i have this experiment of a semantic build graph as a spec that is kept in sync with the code, but it devolved into a lot of custom parsing so i shelved it for now

[lean lark]

unfortunately that happens. 🙂

[bright swift]

i believe its the way to go though for big projects

[lost drum]

when reset?

[cyan saffron]

when i login with codex on the app it will say please add phone number

[lean lark]

I don't remember that prompt. Anyone?
Which "app"? Windows? MacOS? Android? iPhone?

[karmic gulch]

Hhmm google took something out of Codexs playbook, not to bad latest upgrade on antigravity

[signal tapir]

what?

[lean lark]

From Google I/O, I once again got the impression that these companies are doing nothing but copying the exact same UI and functionality off of OpenAI. It's amazing that with revolutionary technology these primitive humans fall back on the age-old practice of "duplicate and publish", just to tick the same boxes as everyone else. Zero innovation. Amazing ... and yet, not.

[signal tapir]

They probably have lawgpt checking what they can copy, and then codexes copying and publishing. 😛

[lean lark]

It's like iPhone ads that claim all of the wonderful things users can do with that device ... and they're just copying from Android. But their users don't know that, so their users think they're getting something special that's original and value-add from their chosen provider. 🙄

[tidal flare]

We must embolden ourselves, for a future of tomorrow, by creating The greatest Innovations mankind has ever seen!

[karmic gulch]

To be fair, Cursor implemented the standard, the rest followed

[lean lark]

Just a little creativity ... Like right after ChatGPT, everything has been a chatbot". Now everything is a dashboard.

And with Yahoo, we got a textbox and a Submit button ... then Google ... then Facebook ... then ChatGPT ...

Cursor forked VSCode. What "standard"?

[karmic gulch]

For Joe to use agentic coding. VSCode did not set that standard

[bright swift]

too much dead weight, they need to fire a lot of people to make room for innovation

[lean lark]

There is a marketplace full of extensions that were copies of GitHub AI which sort of blow away the theory that Cursor was innovative.

need a break

[cloud moth]

Does anyone here know how to contact OpenAI to address this? I sent the same thing in #codex-security-feedback, but I think that channel is somewhat inactive. If I click on "support" I'm just brought to a list of support articles without anything in the way of actually contacting support

[white bane]

Codex has such an interesting memory system, I feel like with just a little bit of structure it’d be trivial to do graph rewriting on the structure

[tiny laurel]

I love the summaries from fireship, have u watched it?

[solid lake]

27 maj

[solid lake]

Is that model still up

Didn’t they depreciate it

[cloud moth]

https://openai.com/index/gpt-5-5-with-trusted-access-for-cyber/

[cloud moth]

nope it's still a thing

[solid lake]

I see

What model does it use?

[bright swift]

it says right there in the url?!

[karmic gulch]

Gotta say, when your using 5.5 you get exc quality, but daym the steak is bound to be well done before you get your hands on the meal..sometimes just getting stuff done 90 % and super fast with composer 2.5 is such a relief lol

[bright swift]

https://en.wikipedia.org/wiki/Ninety–ninety_rule

you even nailed the exact 90%

[solid lake]

I see

[karmic gulch]

Haha, well, for context, opus reviews all commits, so guess im "safe"

[bright swift]

"lgtm" 😄

[wicked fable]

did anyone get a separate codex app under chatgpt beta on Android and how would I disable it? I don't use codecs

[lean lark]

I have ChatGPT Beta on Android. Codex is in this app. I do not have a separate Codex app.

[bright swift]

they want to push codex as their everything-app, so it will probably stay there

[lean lark]

Hmmm, I dunno what that is. Did you download it from somewhere else?

I have Pixel too, Android 16. Your version?

OK. If I were you, not knowing what that is, I'd press/hold and then uninstall that app. I don't think it belongs there.
Someone else here might disagree.
I would even unload the ChatGPT app - since I'm not sure where you got that one. Then be sure to download the offical ChatGPT app.
Your data is not on the phone, it's on the OpenAI servers, so you won't lose anything.

Well, as I said, I'm on beta too, and I didn't get that.

Support for Codex (with an "X") is now official/production. It will work with MacOS, not Windows. I haven't seen anyone else say they have a separate app.

It's in the ChatGPT app

[tiny laurel]

hi chat I'm back, have we got our reset yet?

[lean lark]

how about looking at your own screen. If it's there, yes. if not, no.

[tiny laurel]

there should be a bot that tags a role everytime there is a reset

[glad birch]

just update to pro bro 😛

[tiny laurel]

it's a no 🙁

[lean lark]

then ... it's a no 🙁

[tiny laurel]

this is devastating

[lean lark]

use
less
tokens

[signal tapir]

5.4-mini ftw

[boreal holly]

Use more less

[signal tapir]

So I had fun today. Wrote my fist MCP. Connected Codex to Taiga. 🙂

Codex is fully capable of controlling something like Taiga via its rest api, but my instinct is to reduce chances of it deciding pages have fewer spelling errors if they are all deleted.

[lean lark]

(XHIG?)

[signal tapir]

XHHIG

[cinder plume]

I tell you guys what, using the Caveman skill and reverting back to 5.4/High I was able to drop 3 of my 6 plus accounts

The token comsumption using 5.4 and the caveman skill is pennies on the dollar compared to 5.5

[signal tapir]

caveman?

[cinder plume]

https://github.com/JuliusBrussee/caveman

Highly effective

Not just a meme

[lean lark]

@tiny laurel I'm going to post an image that's intended to be funny. We post images like this whenever someone asks for a reset. Don't be offended.

[cinder plume]

I watched my token usage go from millions/minute to thousands

[signal tapir]

Ok I have to try that!

[cinder plume]

I highly recommend

[signal tapir]

lol: "Same fix. 75% less word. Brain still big."

[cinder plume]

Its incredibly effective and also hilarious to see Codex talk like a caveman

[plush harbor]

Just do it lol

[cinder plume]

Found bug. Me squash.

[lean lark]

Yeah, I know. That was dumb

[solid lake]

Hey how’d you get a picture of me

[plush harbor]

Lmao 6 days 13 hours

[cedar skiff]

it helps that it is half the cost per token

[boreal holly]

"No resolve" 💀

[cinder plume]

Aye, with no clear signs of intelligence loss - at least for me. Combining with caveman its significant savings vs 5.5

[white bane]

What are you making

That requires that many tokens

[tiny laurel]

You’re so considerate, rarely stumbling upon this on Discord

[lean lark]

It's all in good fun, thanks for being a sport.

[cinder plume]

https://store.steampowered.com/app/4458750/Wardens_of_Wen/

[white bane]

Like that’s like 1200 a month

[solid lake]

Reset got me watching birds

[white bane]

That game doesn’t look like it requires 6 plus subscriptions…?

[signal tapir]

maybe he is trying to make it in 24h?

[cinder plume]

Where are you getting 1200 a month? Plus is $20 🤣 I needed 6 because of 5.5 tbqh the token consumption was wild

[white bane]

Oh I thought you had 6 of the 200 dollar ones

[cinder plume]

Oh nooo lmao

[solid lake]

Oof can’t imagine when promo ends

[signal tapir]

I could probably use up 6 normal accounts

[cinder plume]

Plus goes quick imo

[boreal holly]

Why would you get 6 plus subs when you could get precisely 1 Pro 5x sub 🤡 not a fan of spark and GPT Pro or sumthin?

[cinder plume]

I was getting them as needed

In hindsight thats a much better idea I agree

[solid lake]

Try 1 pro lite and 2 plus

[lean lark]

More Home Reseticus memes ... https://bitsforfree.com/
(Now including the latest image, immortalized)

[signal tapir]

I like to be a bit more efficient though. 5.4mini does the job well enough most of the time

[cinder plume]

I am happy with the 3 plus atm, using 5.4+Caveman I dont even come close to capping

5.5 was brutal, although it did produce. Not like it was a waste

I would say realistically 5.5 is unusable outside of a Pro sub

but that is my opinion

[boreal holly]

This month I'm doing 5.5 medium til I reach zero, then switch to 5.4-mini high and sip on the credits

[cinder plume]

Like, I've been running 5.4/High for pretty much 3 hours straight now

on 5.5 I would be sitting here waiting for my 6 accounts to cycle their 5h windows

[boreal holly]

🤞 no resets this week

[solid lake]

Xhigh /fast all the way

[cinder plume]

lol random caveman tooltip appearance

hell yeah

[solid lake]

What’s caveman like on xhigh

Brb trying

[cinder plume]

I havent used it with xhigh tbh but I assume it works the same

[solid lake]

Buffed caveman

[plush harbor]

Some people do some weird stuff in here to save tokens

[solid lake]

If it’s efficient enough

That’s good

[cinder plume]

Caveman is silly for sure, but the reason I use it is its actually extremely effective

[boreal holly]

"Me think long and hard until output more longer and harder"

[solid lake]

Have you seen it leave out any info yet”

[dire veldt]

What's the name of this program of checking usage?

[solid lake]

Official site

[cinder plume]

You can monitor your token consumption in real time a few ways

I am partial to https://github.com/graykode/abtop

lmao idk why message doubled up there

[cinder plume]

I have not no, it still gets the info it needs to across to you.

It only affects how it speaks TO YOU

It doesnt effect code, QA, or anything like that

[solid lake]

Bet