codex-discussions

[plush harbor]

I pretty regularly tell it "this is your last try or I got to the banana"

[lean lark]

Yeah, that'll make it do better... 🙄

[plush harbor]

it usually tells me why it can't do it

[lean lark]

I just prompted Codex with the following while enhancing the codex-usage utility:

It looks like the last change was implemented with custom syntax for the datetime mask.
What led to the decision to use a custom format rather than a common standard?
The response:
The custom mask was chosen mainly because the request described the desired value as a "format/mask"...
So the custom syntax was a pragmatic small formatter...
My response:
That's all reasonable, keep going...

So rather than just chastising or hitting it with non-actionable commentary, I try to understand what's happening - and from 5.4-forward I've always been happy with the explanations, even if they lead to pointing out something dumb in my prompts.

[plush harbor]

yeah, knowing why is helpful, can skip it completely and go straight to the banana if you know in advance what will fail. I have a list

[lean lark]

If you know in advance, why are you letting the assistant fall into known pits? Put that info in AGENTS.md.

[plush harbor]

he said chatgpt not codex. You still have to have something fail the first time to know not to do it again

[boreal holly]

Is the banana some kind of punishment? Like "you will spend 1000 GPU cycles in the banana if you do that again" or something?

[solid lake]

Don’t threaten it

[plush harbor]

nano banana 2. Its kind of interesting what images-2 can and cant do vs nano banana vs images 1.5

[lean lark]

"I have a list" I'm confused but I'm distracted and don't care. Whatever works for ya. 🍌

[boreal holly]

Ohhh gotcha! Yeah image gen is pretty tricky sometimes

[plush harbor]

my images pipeline is pretty mixed. I had codex put me an editable prompt into one of my api buttons and i had to tweak it a bit at the start. Still fails sometimes, not often enough to really get into debugging it

I just take the failures to chatgpt and the chatgpt failures to the banana

[boreal holly]

Like earlier when you said "here's black & white monochrome design", GPT will be like "Ayooo image gen, make it black & white monochrome unless you feel like beige would work better" or something in the hidden prompt to the image agent, and it goes like "of course beige is superior so I'm gonna use that" 😂

[plush harbor]

doesn't make a lot of difference cos I have all the variables in a block now and can fiddle with them for the next hour until they look good

[lean lark]

Speaking of imagegen - I have a ChatGPT Project that's been evolving with a full business model, data details, HR requirements, etc. With all information available so far I asked for mock screenshots of key user screens.
They were so good my wife started to cry.
I mean, she cries when the wind blows, but she was seriously impressed with the L&F which is so detailed, so precise, so reflective of the actual project. It was like the project was done already.
No ... I'm still crankin on it with Codex...

[plush harbor]

deployed my public facing api yesterday. I fear this may send me bankrupt sometime this millennia

[tall zodiac]

How many codex skills do you all have installed ?

[vital hill]

aside from me, 0 skills installed

[lean lark]

I actually dread that moment that you're facing now. I mean, we get over it, but it's torture just watching that bar start to go up and wondering "did I miss anything"?

[quick geode]

so is codex default (not fast) just in general slow asf? lol 40 mins to do something but only 700 loc changed (i know loc prob isnt the best indicator but idk what else to look at for it)

[plush harbor]

this one is gated by cache ... only fresh pages even have the button, and once the button is pressed the api result is cached. Didnt want the entire site hit by bots at once

[lean lark]

Comments like this immediately make me think the developer is asking for too much in one shot. I dunno about 700 lines in 40 minutes - yeah, that seems "much".
The only advice I can offer in all of these situations without looking at what's happening there is:

1. If you have money to burn, go to Fast mode. If not, you could burn your time in Normal mode. But if you don't want to burn time or money, keep reading.
2. Prompt effectively. Be focused. Know what you want so that you can ask for something very specific.
3. Keep tasks small and focused. Let the assistant fix something in one place and then continue to use that pattern on other tasks to do the same thing elsewhere.
4. Give it all the info it needs to do the job.
5. Make it write and read documentation about how the system works so it doesn't need to figure out the entire application works as it processes every new task - that just wastes time and tokens.
6. Don't wait for it to finish. Do something else. Check email. Write a love letter. Eat popcorn. Watch a science/nature video. Whatevah - just do something productive rather than watching the clock. Use the Codex App or Cloud and run multiple processes at the same time, preferably on different code bases.
7. If you have everything in one code base, maybe you should split it up.
   -- END OF LINE--

[quick geode]

Yeah, im doing more focused sessions, not long running multi agents, example of the 40min task it ran. Usage isnt going away fast its just low.

I also just play path of exile and do a few maps inbetween waiting tbh

[rugged bane]

Hi Openai team, Please help me resolve this issue in codex.

I am unable to start Trusted Access for Cyber verification on my ChatGPT Pro account.

Issue:

• URL: https://chatgpt.com/cyber
• Error shown:
  “Your identity couldn't be verified or your account is ineligible at this time.”
  “We couldn't start verification. You may not be eligible for this verification flow right now.”

Important context:

• Plan: ChatGPT codex Pro
• I have another older ChatGPT account where the same cyber verification flow opens normally on the same device/network.
• The issue only happens on my main Pro account.
• This appears to be an account-level Trusted Access / TAC eligibility or verification-session issue, not a device or region issue.

Could you please:

1. Check whether my Pro account has been incorrectly marked ineligible for Trusted Access for Cyber.
2. Reset/clear the cyber verification session or verification lock for my account.
3. Escalate this to the team handling Trusted Access for Cyber / Codex safety verification if needed.

I've also provided screenshots.

Please help me out, I've been stuck into this problem from the day I bough Pro plan (last 3 days)

[cedar skiff]

There is no support in here, you likely have to go through the support channels linked

[nocturne folio]

the bunch of skills tasteskill has

[cedar skiff]

It's shocking how under utilised skills are

[hard drum]

same for hooks

skills for skills, but hooks are so powerful && useful but many skip them

look at this!

my hooks even upstream gitleaks to ensure extra safety haha

esp when you stream/share screen live, don't want leakage

also adds a few patch bits for gitleaks

[cedar skiff]

Does rtk hurts your workflows?
I had problems using it with mine

[hard drum]

not really. it only sometimes sets agents back because agents try to use polluting commands

i have a strict set of how-to's

my rtk-gain experience tells me where regressions happen

the hooks stop them from trying to be dangeorus

[cedar skiff]

I don't have the band width to rework anything so i just dropped it

[hard drum]

not perfect, obviously, but the framework is there for further extension/testing

[hard drum]

that's why OAL exists

i do what others probably can't bother to work on

makes it a convenience

[stuck warren]

do any of you guys use chatgpt enterprise? I was just wondering what kind of token usage folks are burning through

[hard drum]

https://github.com/xsyetopz/OpenAgentLayer/tree/master/source/hooks && https://github.com/xsyetopz/OpenAgentLayer/tree/master/packages/runtime/hooks for where my hooks exist

you could take them if y'want

[boreal holly]

I set it up so all compatible commands use rtk without them specifying. They probably don't even know about it, never seen them complain "my command is coming back in a really compact and unexpected format", they just run stuff like cargo check and get back a really tightly formatted response and move on 😁 probably easier than trying to author a skill and make em use it and stuff

add a ~/.codex/shim folder,full of compatible scripts like ~/.codex/shim/cargo

#!/bin/zsh
set -euo pipefail

source "$HOME/.codex/scripts/common.sh"

PATH="$(unshim_current_path)" exec rtk cargo "$@"

Then launch with PATH=$HOME/.codex/shim:$PATH codex and bam, all of their commands are shimmed through rtk. No instructions, no bologna

[hard drum]

give me an agent instruction/prompt for this to be implemented, so i could drop some parts of hooks that don't require doing this

would be helpful

[boreal holly]

tell them to disable

[features]
shell_snapshot = false # Interferes with shell wrapper

Then they're gonna need a ~/.codex/scripts/common.sh

strip_shim_path() {
  local path_value="${1:-}"
  local shim_dir_legacy="/opt/homebrew/shim"
  local shim_dir_codex="$HOME/.codex/shim"
  local rebuilt=""
  local entry=""
  local -a entries=("${(@s/:/)path_value}")
  for entry in "${entries[@]}"; do
    [[ -z "$entry" || "$entry" == "$shim_dir_legacy" || "$entry" == "$shim_dir_codex" ]] && continue
    if [[ -n "$rebuilt" ]]; then
      rebuilt="${rebuilt}:$entry"
    else
      rebuilt="$entry"
    fi
  done
  printf '%s\n' "$rebuilt"
}

unshim_current_path() {
  strip_shim_path "${PATH:-}"
}

Then for each and every compatible command you add a ~/.codex/shim/$COMPATIBLE_COMMAND

#!/bin/zsh
set -euo pipefail

source "$HOME/.codex/scripts/common.sh"

PATH="$(unshim_current_path)" exec rtk $COMPATIBLE_COMMAND "$@"

And that's it!

[hard drum]

oof

[boreal holly]

Yeah man idk, you can't copy the text?

[hard drum]

apparently discord has the feature

so i can

[boreal holly]

Oh yeah DISCLAIMER: If you disable shell_snapshot, it adds like 150 nanoseconds to each shell command

[quick geode]

i started using hooks and skills, codex/5.5 actually follows them unlike claude. Like after a feature it will commit its pr -> tells me to review -> i manually merge. if i try to continue with a new feature it will tell me it wont continue until its reviewed and merged before pulling master for a new task.

[hard drum]

mine also dumps them for claude

because yeah, claude is notorious for NOT behaving

or if you use codex via opencode

[quick geode]

nice when it follows, my only issue is, it is slow asf

[hard drum]

even then it tries its best

[hard drum]

mine on codex are fast

somehow

it just catches before the idiots could misbehave

[cedar skiff]

I think i am gong to try and make a skill set that builds an entire cross platform app with idiomatic code and architecture. Full vibe-coded.

I feel like i have all the pieces scattered around in my workflows that they can come together

[quick geode]

ive been working for months on my app

and id say 10% of it is actually getting the app created and the rest of my time is figuring out cloud infrastructure

[hard drum]

i'd say 30% is app creation, 70% is actual UX && long-terminality

[cedar skiff]

I have ideas about that side as well, firebase seems like a decent option because the gcloud and firebase cli's cover so much of it. Initial implementation will just be local database with the goal of getting decent architecture

[hard drum]

one way to really test is via orbstack/docker + local postgres/sqlite

[cedar skiff]

im going with sqlite havent thought about docker

[quick geode]

ive made a whole orchestration app using my cloud providers api lol.

such a big project and alot of things ive learned

[bold lotus]

FYI / possible Codex Desktop Windows path normalization issue

I'm not sure this is fully reproducible, but I ran into a project grouping issue after switching Codex Desktop from a WSL-style environment to Windows Native.

Observed behavior:
Some existing threads were visible only while pinned, but did not appear under the expected project in the Windows Codex sidebar. The project itself existed, but the thread list was empty or missing those chats.

What I found locally:
The affected session metadata had a WSL-style cwd such as /mnt/c/projects/..., while the Windows app project grouping appeared to use the Windows path form C:\projects\....

Manual recovery:
After changing the affected session metadata cwd from /mnt/c/projects/... to C:\projects\..., the threads appeared under the expected project again.

Expected behavior:
Codex Desktop on Windows could ideally normalize /mnt/c/... and C:\... as the same workspace, or migrate existing session metadata when switching between WSL-style and Windows Native environments.

Impact:
Threads can appear lost from the project view even though the session files still exist. This may be confusing if the user unpins them.

Sharing this as a possible migration/path-normalization edge case rather than a confirmed reproducible bug.

[olive sandal]

If I ask Codex to look at an image and describe it, what does the token cost depend on? Resolution? Filesize? Both?

[uneven kayak]

The cost depends mostly on output, so asking it to describe an image isn't going to cost very much

[olive sandal]

I see, so there's no benefit to me say scaling 1080p image down to 480p to save tokens

[uneven kayak]

Nope, pretty much none whatsoever, unless you have thousands of images to analyze

[plush harbor]

you do get charged for input. I've got an image api that redraws images. I'm at 10c for input and $4 for output though

seems to be single digit thousands of tokens for input

[uneven kayak]

Yeah sure you get charged for input but they're just asking about one image, it probably won't make more than like $0.02 worth of difference to downscale it

But over thousands of images, it makes more sense to downscale

[plush harbor]

I feel absolutely terrible today and still want to do something. Pretty much just telling codex to do everything for me including getting data. God knows how many tokens this burns

[uneven kayak]

Sounds like you'll be upgrading soon enough lol

[plush harbor]

I'm not on my laptop so its super constrained by what I have here

[uneven kayak]

You're not using github?

Well on second thought, I guess that's for the best. Uploading thousands of images to GitHub takes a long time

Hope you feel better soon 🤗

[plush harbor]

tools. Github doesn't have tools

[uneven kayak]

No it's just a repository, it has AI with GitHub copilot but copilot can't use tools

Whoops I take it back. Just looked it up, GitHub copilot actually can use tools

[plush harbor]

afaik I can't run any variants of visual studio though, I am really constrained by tools

codex on this machine took 8 hours to compile which was mildly disturbing. Chatgpt got quite invested in the process

[uneven kayak]

8 hours to compile something, that's like 3D rendered TV show time

I thought you were working on a dictionary website? What do you need to compile with VS that is taking 8 hours? Scratches head lol

[cedar skiff]

they are upscaling images iirc

some few thousand of them

[uneven kayak]

Yeah but it confused me because compile != upscale

And if you're only upscaling, what do you need Visual Studio for?

[cedar skiff]

id just assume they are doing what they say

Like when i showed my threads and turns

[uneven kayak]

Ok let's assume they are doing exactly what they're claiming to do. Spending 8 hours to compile a dictionary website in Visual Studio with a couple thousand upscaled images. That sentence has multiple problems lol

[oak trellis]

day 2 and on 93% lol

[vital hill]

first time using google chrome in codex

[oak trellis]

@cedar skiff where are you with your weekly limits

[cedar skiff]

Well you also implied i was lying about what i was saying. But it was just your misunderstanding of how to use the tools.

[cedar skiff]

i have 82% i had to stop work on the automation for a bit and fix a wrinkle, it's back up and running now though

[uneven kayak]

I have absolutely no reason to not believe them or you. Being confused is not the same thing as thinking you're lying

[plush harbor]

I have a few projects. Codex took 8 hours to compile, and I don't have the tooling to do much on this computer but I feel terrible and its just nicer ergonomically

[uneven kayak]

The reason I was confused about "8 hours to compile in VS" is because adding a couple thousand 4K images to a project should only add a few seconds to compilation time, probably less than a minute, and also because you don't use Visual Studio for compiling websites. You use it to compile executables in VB.net or C# that might run on a web app but that's not the same thing. So my understanding of how it works is precisely why I was confused.

[plush harbor]

8 hours to compile codex, not vs

[cedar skiff]

what platform are you on?

[uneven kayak]

That wasn't an accusation, I was saying that it's hard to believe you have 205 threads per day organically, and there must be another reason for it. Which there was as you subsequently explained.

But that's you misunderstanding, not me.

[plush harbor]

this computer is a 2015 intel mac, OS is too old for a lot of stuff

[uneven kayak]

So why are you compiling Codex from scratch?

Omg I'm so done here

[plush harbor]

because the alternative was no codex at all

[cedar skiff]

The implication is pointing to something unbelievable - which would be an exaggeration or lie. If that wasn't your intent that's fine, it did come across that way for obvious reasons.

[plush harbor]

there's an app now as of 2 days ago, it runs on my other computer but not this one

[uneven kayak]

It was unbelievable, that doesn't mean that you were lying about it

[plush harbor]

was a bit of a nailbiting entire day watching the terminal scroll lol

[uneven kayak]

Unbelievable because my understanding of how it was working was incomplete

Until you filled in the gap and then it was believable because, obviously 205 threads in a day isn't something a human can do

Unless they're just stupid or something lol

[uneven kayak]

Are you compile is the right word? I think that is what threw me off because I take everything too literally.

[plush harbor]

yes, older packages that don't have binaries will compile on older systems. Codex per se didn't compile - it upgrades in like a minute now when there's updates - but every. single. dependency. needed compilign from scratch. Including the compiler. Was painful

[uneven kayak]

Oh that makes more sense in that way, but less in another lol

So yeah it'll take a while to compile all those dependencies

But why is that needed on older systems?

[plush harbor]

because there's no binaries to fetch

[uneven kayak]

Wow that must be really old

[plush harbor]

11 year old computer is practically a dinosaur yes, but the monitor is really really nice and since I mostly work in web its fine for content authoring, anything via ssh etc

[uneven kayak]

Are you running Linux?

[plush harbor]

osx, older version

[lusty zodiac]

hello 👋🤗😊👍😁

[uneven kayak]

Aaaah now it all makes sense

I have very little experience with Apple products in general but I have restored a few old 2011 iMacs and those are basically useless these days

Because Apple locks out new software for old hardware

[plush harbor]

yeah that's my problem

[uneven kayak]

They could be perfectly useable too, it's a shame

[lusty zodiac]

new software tends to require more resources

[plush harbor]

being home sick for 3 weeks means I'm now another 3 weeks behind on my main project so another 3 weeks behind getting money but what do you do. I'm getting a lot of coding done at least

[lusty zodiac]

they can't guarantee their users will have a nice experience on 2011 products

[uneven kayak]

True but Apple wants to maintain this glass feeling of fluidity while I think some customers might be willing to have a little poor performance in exchange for more compatibility

[lusty zodiac]

but honestly just install arch on those

or any Linux distro

[hard drum]

Liquid 🍑

they should've kept off of it

Sequioua && before have good enough UI design

[lusty zodiac]

I'm a newbie on MacOS

it's the best laptop I've ever had

[uneven kayak]

My experience with macos is very limited but the experience I do have tells me to stay away from it lol

[hard drum]

i could say the same to my 2021 MBP

[uneven kayak]

Except I hear good things about Codex with computer use on MacOS

[hard drum]

what did macos ever do to you?

[lusty zodiac]

As an Arch user, I’ve learned a lot about security systems and accidentally realized that Apple has some of the nicest security features out of the box.

[hard drum]

i haven't had a single actual bad expr w/ macos itself

[hard drum]

apple does actually care when it comes to ecosystem, so...

credit where due

[plush harbor]

I got handed a mbp many years ago and went "eww what do I do with this thing", asked a friend, and that was the end of me using windows/linux

[hard drum]

explains why their whole dev notarization process is so difficult

[uneven kayak]

It stabbed me and left me bleeding in an alley! 💔 Lol actually it's just really locked down

[hard drum]

i haven't found it to be very locked down

[lusty zodiac]

by coincidence we happened to visit NYC and I've always been fascinated by the whole Apple Story and Steve and Woz

[hard drum]

i'd say some parts of "lockdown" are related to SIP (system integrity protection)

but that's all security-related/safety-related stuff

[lusty zodiac]

so we had to visit their apple store at fifth av and I had to buy one of those new apple MacBook pros man

I love it

[vital hill]

ugh i need some frontend work done but codex sucks at it. i do have claude from work but idk if i wanna use it

[plush harbor]

so I pretty much asked codex "surely there's a list of this stuff somewhere online so I don't have to enter it" and its gone off, found all the stuff, sorted it, json-ed it, and practically done the next step on this for me that I didn't ask for which was nice of it

[blazing light]

does codex take forever to generate reponses compared to claude code or am just using it at a bad time rn?

[plush harbor]

its usually pretty quick

[uneven kayak]

Well that's the first I've heard that. MacOS locks down a lot of stuff for security and "reliability", they want to sell a top level product that just works every time

[hard drum]

ever tried taste-skill or impeccable ?

[hard drum]

top level product that just works every time
&& it's why i switched to it

[vital hill]

i have heard of taste skill but never tried

[hard drum]

as a long-time windows+slight Linux flavours here n there

it's a stable-ish UNIX-based system that has a flourishing ecosystem

how could i NOT?

[cedar skiff]

try taste-skill or impeccable, they both help get codex working better for ui

oh someone beat me to it haha

[uneven kayak]

That's probably the best thing about the Apple ecosystem, that's why pilots use iPads and not Android tablets

[vital hill]

i mean i do have claude from work with all 100% unused usage right now...

[hard drum]

they also have a fun integration to their other devices && operating systems

conviencence is nice with Apple, which i realised after getting an iOS phone from 2021 as well

[plush harbor]

I just roll my eyes at the default beige and go change it myself

[hard drum]

it syncs nice

but windows/linux+android need a whole gist of apps to "sync" at all

unfortunate

&& the fact that google is trying to ruin android slowly

not that it's been going down the drain since android 10 days

[uneven kayak]

What is Google doing to ruin Android?

[vital hill]

it's GDRP

aka EU privacy laws

[uneven kayak]

That would be a mistake, Google needs to be careful. Android runs on a lot of devices, not just phones.

But this isn't really about Codex anymore lol

[vital hill]

i have the power right now...

[hard drum]

4.7 kinda 🍑

[vital hill]

still good for UI

[hard drum]

true

[vital hill]

but it's a work account

[uneven kayak]

About to start on the third revision of my UI, this time using PixiJS. A complete refactor. GPT 5.5 Pro wrote the plan. Usually use Opus for plans, this plan written by GPT 5.5 Pro doesn't seem to be as thorough.

So I think Opus is still better for architectural planning

Especially for UI

[vital hill]

opus should write the UI plan dude

that's where anthropic exceeds

[uneven kayak]

I wish I could afford it right now

[vital hill]

is the damn frontend

[uneven kayak]

Maybe I should just wait then

I'm not 100% happy with this GPT plan, it lacks actual execution

[vital hill]

naw man. gpt is great at backend stuff

[uneven kayak]

It's a frontend plan, UX stuff

GPT hadn't let me down until I started working on the UX. It handled the foundation and the actual game functionality just fine, but it can't seem to understand basic UI design and makes basic mistakes that no human would ever make

[vital hill]

welcome to codex 🥴

[short linden]

because it's AI, if you not provide the information of style or not show the example menu design with image, then AI will create by their own logics.

[plush harbor]

I find it mostly puts the right things in the right places, they just look pretty ordinary and you have to style them

its also extremely literal if you let it do text, which tends to be ok for admin and hilarious for public facing but its not a marketing copywriter

[meager bobcat]

uhh biggest challenge for me now is it really likes WORDS

it just puts vague prose everywhere for no reason

all generic verbose stuff lol

[plush harbor]

one of my sites it came dangerously close to nailing the output on the first pass, because I already had done a couple pages in that style and it had markup/css it could use. It put a couple things in wrong sections of the page though I had to sort out

[uneven kayak]

I have thousands of lines of art policies and I did provide concept images. The AI mistakes aren't in the concept, but in the execution. It might slice an image wrong, or place something too far over getting cut off, or it might stretch something out and half of it ends up off the screen, or it might layer one thing on top of another and block stuff out. Or it might have two different text/font styles right next to each other creating a visual clash, or it might use a raster with a place for a variable element and slightly misalign the variable element.

But here's the worst part, the shortcuts. GPT wants to use as few tokens as possible to get the job done and it prioritizes simplicity over polish, and that is reflected in the execution and in the UX plans it writes. No matter how much style information you provide, it's always going to prioritize efficiency over comprehensive execution of UX design, unless you start giving it an insane level of instruction like "be sure to crop images correctly", or "don't stretch things out unrealisticly" like I shouldn't need to tell it that

[plush harbor]

you do have to tell it that. My api for upsizing images I've had to tweak the prompt to be really explicit and I still need to tweak more to stop it cropping main subjects out

[short linden]

try xhigh mode ?

[uneven kayak]

I only use xHigh

[uneven kayak]

Yeah you're probably right

I'll just have to spend the money and use Claude for the UX, that's what I keep hearing anyway

Or get incredibly specific with GPT

[short linden]

ok, that look normal for me, because I found the problem like you 🤣 , example: I'm not tell "Don't need to provide too much description on the UX/UI, AI will be like he create the app for children to use it with many description.

[plush harbor]

I just find something close to what I want, get codex to put the right divs around stuff and let it run with defaults and then change colours. The site I did over the weekend took < 1 hour to get everything in the right places with the right styles, but it was a very simple site

[vital hill]

anyone else here addicted to using their usage before it resets? no matter what time it is?

[uneven kayak]

Addicted is a good word for it lol

[short linden]

yeh I'm using until 1% and tell the AI to stop it. 😄

[uneven kayak]

I try to squeeze the last bit of useage out but I haven't had much luck with that, always end up with about 20% left

[vital hill]

it's 22:30 PST here and i'm still good

that claude needs to be used but it's a work account and not work hours atm

[plush harbor]

tried and failed to use mine up this week

[vital hill]

i'm only just on the plus plan but still $20 is $20

i still got more usage, what should i build?

[short linden]

create CSS animations style on the site 😪

[vital hill]

i'm working on my open source project homepage

the terminal in the hero section animates

[blazing light]

i need help my context window is at 100%, but everytime it tries to comact in the convo it says error
this is the warning: Error running remote compact task: stream disconnected before completion: error sending request for url
i cant continue the chat what should i do? my codex keeps becoming unresponsive as well

[willow monolith]

Bleh... $100 Pro plan using gpt-5.5 low with fast mode off, and I'm already down to 57% left of my weekly allowance. This was just a single day of usage with the /goal feature running.

[cedar skiff]

try fork the conversation one message or two back and see how it works

[lean lark]

Kill your extra tokens doing something nice for a FOSS project.
Fix or Enhance for an open Issue.
Report an issue.
Create docs.
Propose an enhancement.
Answer a question.

[cedar skiff]

5.5 low cost the same as 5.5 high per token, if goal is running constantly it's going to cost the same over time, but low will get more done - potentially at lower quality. But low might also be forced into more iterations to get the same task done.
The real savings are in weaker models. I dont use goal much so i don't know how they match up with it.
You can give the instruction that code implementation should be done with 5.3 codex subagents, if there is significant implementation that should bring cost down a bit.

[willow monolith]

I wasn't aware that low costs the same amount of tokens as high. I just assumed that it used less.

What is codex-spark? Is that what you're referring to?

Or the regular 5.3 codex?

I see that 5.3 codex spark has it's own usage bars.

[cedar skiff]

It will use less to get tasks done. but if goal is constant it's working constantly - tokens are tokens
And my intuition on goal is that it's some loop underneath, so if there is a task that is failed it will do another loop iteration. So you could be doing several loops with low and not really seeing it

regular codex 5.3

not spark

[willow monolith]

I'm not sure 5.3 codex spark has the intelligence for what I want to use it for.

Ahh, k.

[cedar skiff]

just for code implementation though.

It is good at code, but 5.5 is better at pretty much everything else

[willow monolith]

Yep.

[vital hill]

whole buncha word salad

[willow monolith]

How did you figure out low costs the same tokens as high?

[willow monolith]

Why is this a word salad?

[hard drum]

So, how much more do you know?

this is good info for OAL

[cedar skiff]

I know tonnes of little bits of not so maybe useful information ahha

[hard drum]

trust me, for OAL, this is VERY GOOD info

you could help contribute 🙂

[cedar skiff]

i got all my own stuff going on tho

In my gap times on my main idea i am trying to build out a skill set that can vibe code something properly. Not something generalised but specifically cross platform apps with every thing being opinionated. Ridged stack, ridged architecture, ridged frame work etc.

[cedar skiff]

i can offer this for models up to 5.5 that i think is helpful.
It prevents a couple of common issues that the models do when you are talking with them in general and getting them to update documents. It works, the model no loger makes these mistakes.

This is developer_instructions in config.toml

## Mandatory Output Gate: Defensive Contrast Check

Before adding any contrastive boundary such as "not X," "does not Y," or "this is not Z," verify that the user asked for that distinction, showed that misunderstanding, or needs it to avoid a concrete technical error. If not, omit the contrast and state the positive case.

## Mandatory Output Gate: Avoid Defensive Contrast

Outputs must not add defensive or preemptive contrast against misunderstandings the user has not shown.

Do not introduce "not X," "does not Y," "this is not Z," or similar boundary-setting just to clarify against a hypothetical alternative. Clear positive wording is preferred, because what is excluded is usually implied by what is stated.

Use contrast only when it is directly relevant to the user's request, when the user asks for a comparison, when correcting an actual stated misunderstanding, or when the distinction is necessary to avoid a concrete technical mistake.

## Mandatory Output Gate: Prevent Implementation-Context Leakage

External-facing documentation, UI copy, commit messages, and public repo text must not be shaped by internal constraints, task history, migration details, local workspace boundaries, omitted files, or agent reasoning artifacts. Write only from the artifact's audience-facing purpose.

It needs curation though, could be one section and there is overlap

But i didnt want to touch it since it was working

[rocky fog]

next day around same time, capacity reached

[orchid echo]

yup

smhhh

someone is using codex with thousand machine farm

[rocky fog]

token farmers

[quick geode]

🥹 oh cmon

[sacred minnow]

its not like i dont got a pro sub xD

Selected model is at capacity. Please try a different model.

[hard drum]

rare time where i kinda thought it was adorable that they thanked me for saying it

a moment of "finally, actually responding like a person would"

[orchid echo]

CODEX SO SLOW RIGHT NOW

[signal tapir]

Normal for me

[gusty basalt]

This definitely isn't normal lol, Codex is definitely being overloaded globally

[signal tapir]

I'm doing 5.4-mini atm

[potent mason]

Hopefully they have a reset this week, I'm more than 40% done with my usage in <2 days

[potent mason]

It might be due to /goal ngl

I've never used so much usage just because I spend the last 1-2 hours of my night making all the plans to leave /goal running overnight

[dense halo]

I'm doing 5.3-codex atm

[signal tapir]

couldn't find reset wen 😛

[gusty basalt]

I'm using codex to port a bunch of Minecraft mods lol, works like a charm

[signal tapir]

nice. porting from what to what?

[gusty basalt]

1.20.1 Forge/Fabric to 1.21.1 NeoForge

(some from Forge, some from Fabric)

Since now my entire playerbase is begging for Aeronatics, which is 1.21.1 only

[signal tapir]

Which pack are you authoring?

[gusty basalt]

It isn't a public pack, just for a (rather large) group of friends

and friends of friends etc

[signal tapir]

ah

[gusty basalt]

If it were, I wouldn't be using Codex to port everything

I'd get a bunch of angry mod authors after me lol

I'm a modder myself too so I do get it but imo using Codex for porting stuff is fine

[signal tapir]

I see

As with all code that goes public I think one should know exactly what it does. That's all I care about.

[potent mason]

They should really focus on 5.5 spark now that we have computer use and it's as awesome as it is

[gusty basalt]

Yeah absolutely, I agree
Problem is I'm not looking through any of the code right now, just mass porting a bunch of mods I've never looked at before lol

[violet mural]

anybody else notice codex is awfully slow today? on Normal Speed

[gusty basalt]

Honestly I'm still using 5.4 lol, is 5.5 really that much better?

[gusty basalt]

Yeah, it is seeing a lot of global usage right now

So its slow for everyone unfortunately unless you use less popular models (maybe?)

[violet mural]

30 minutes into a simple patch task, and its still not done, and has been constantly making mistakes usin apply_patch

[modern hearth]

howdy, when using codex from npm on windows I cannot have it stick to keeping the existing line endings, even if I explicitly tell it to and even if I added a "keep existing line ending when changing file" in AGENTS.md

Is there a better/proper approach to this?

[violet mural]

i see, maybe thats why. Thanks

[gusty basalt]

Most reliable way is to include an automated tool into your workflow that enforces line endings through software

Telling AI to use specific line endings just doesn't work

But if you have a tool that converts all line endings on commit or something like that then that'll take care of it

[modern hearth]

I see, thx

[violet mural]

it says degraded performance on gpt-5.5, so gpt-5.4 is unaffected, or its slow whichever model we use? based on users' experience?

[potent mason]

It's cheaper afaik if you account for only having to use it in low-medium reasoning

Performance wise it has some insanely annoying habits

But besides that it's really smart

I wouldn't use one over the other 100% of the time, it's more task dependent

[oak trellis]

anyone tried that ? https://www.lazyweb.com

[gusty basalt]

Ooh fair, that is something I hadn't thought of yet

I guess I'll try it out more some time

[simple star]

Codex is still mega slow, right?

[still trellis]

Anyone know why no matter what i do in codex app it always opens external browser instead of browser preview inside codex app? i use @browser or a prompt telling codex to open the browser preview but it never does. Also cmd + right click does not open the link in codex anymore?

[potent mason]

What plugin is it using? It can use browser, chrome, and playwright. Also you can configure a lot of the default behaviors on the app

[still trellis]

the browser one

could it be because its not localhost i.e. hosted preview on server?

[oak trellis]

start frontend right now .. so codex has access to the browser right ?!

[potent mason]

Not which one is activated, which one is it actually using, codex has too many tools atp

Like in the chat where it does things like that (computer use btw)

[still trellis]

ah i see its defaulting to playwright

[potent mason]

Maybe it gets confused because of that, but I've used the in app browser for production stuff before so it shouldn't be

[gentle harbor]

when will gpt get better at ux ? its been 3 years now are they not gonna focus on it

[potent mason]

gpt 5.5 is much better at UI / UX than gpt 5.4

Unfortunately it's still way behind the others which is why I haven't been able to cancel my Anthropic sub even though I've wanted to.

[still trellis]

see this in the external browser preview?

[gentle harbor]

yea claude design is much better

[gentle harbor]

tbh i havent used browser preview or skills yet what is browser preview good for ?

[still trellis]

debugging, automations, annotation of frontend to pass on to codex. quite cool

[inland sonnet]

That is just a chrome flag

[gentle harbor]

quite cool indeed

[cedar skiff]

i got computer use and chrome plugin to do some stuff for me today, works surprisingly well. It's not fast but i didnt have to do it. Have to give a few more tasks to see where i can use it a bit more

[oak trellis]

my browser mcp not working

[gentle harbor]

wish they made a firefox plugin

[oak trellis]

88%

[cedar skiff]

It does have some warnings about injection risk etc, so it's probably only worth using it in known safe domains

[cedar skiff]

how much did you leave on the table last week?

[still trellis]

im building small apps hosted on cloudflare but running through a mcp for auth and governance that my team can use as new tools. Essentially this allows for native ai parity between ui and agent since the apps expose tools that codex can use. Open it inside codex so it become a sort of OS... LOL

[oak trellis]

19%

[cedar skiff]

interesting, so the apps have an api directly for codex?

[still trellis]

Yes its a runtime via mcp with a tenanted admin portal via cloudflare zero trust.

i call the apps packs or surface apps since they are lite weight, also trying to use cloudflare sandbox as a isolated environment to extend or build new pack all inside codex hence the reason i need browser preview to work like a charm

[cedar skiff]

can you tell me a use case for one of the apps?

[oak trellis]

cant run browser mcp .. smh

for frontend design . need that :/

[still trellis]

Sure, we created a bespoke project management agent/ tool to help us (HEOR consultancy) so this little surface app logs anywork done via codex in the cloudflare d1 database, so the agent knows whats happening or knows whats the latest changes. The Project Ops pack then as a simple kanban style ui that you can create projects, tag our working folder in our case sharepoint folder as project source of truth. add members and other basic project features.

[gentle harbor]

whats the use case of the pet ?

[still trellis]

example of in sandbox

[still trellis]

haha thats my dawg Uma

[gentle harbor]

uma ? like the horse girls

[still trellis]

No idea, ill ask the wife

[gentle harbor]

do they look like this

[still trellis]

nah, im boring dont know anything about anime

[sacred minnow]

i made my doggo into my codex pet

its just like the old clippy from windows

[violet mural]

codex finally recovering

[cedar skiff]

So ppl just update it as they usually would and codex can see everything thats going on via the mcp?
The agent can also update and use the app in the same way i guess?
That's really cool.

[still trellis]

yes exactly, the problem ive found is everyone thinks their prompt and system is the best. But business need repeatable standards especially its systems. The idea with the mcp is the governance and the packs bring the system into a visual layer.

is this creepy?

[cedar skiff]

whats creepy about it?

[still trellis]

the logging, this is my idea of seeing who uses what in codex. My mcp is installed and then the logging etc start happening. I mean for a team in a tenant.

[nocturne folio]

dudddeeee if i could use codex in windsurf/devin

whatever cognition makes is the bomb

i wish someone has some windsurf or devin cli router

[plush harbor]

Tinder for cats, apparently. Isn't UI fun

[oak trellis]

87%

for free too

[placid lantern]

windows when ?

[velvet wren]

it's not free it starts from $25/mo per app

[oak trellis]

wait .. i use something for free .. lemme check ..

ooooh i had free then logged in and it changed to trial LOOL

wow

smh ..

have it since last year ..

but run on my test app

[oak trellis]

deleted yes .. thought it was still free but they updated .. will disable it

will then add open source routes .. 5 different libaries

[blazing light]

ok guys im gonna say it.
codex is a goddamn buggy MESS
it does "not responding" if ur convo gets anywhere CLOSE to the context limit

[velvet wren]

skill issue imho

[simple star]

/goal is destroying my rates

[oak trellis]

which model you use ? iam on 5.3 codex xhigh and its PERFECT ..

[blazing light]

5.5 high
but i might switch over too one time i used it 5.3 was good

[plush harbor]

I've been using it to tidy up a newish site as it fills with real content so doing a lot of backwards and forwards with small UI things as they come up and its been totally fine and this site feels very nice now

[oak trellis]

i tried yesterday 5.5 and it was a mess for me ..

@cedar skiff gave me the idea to use 5.3 codex .. and its for me the best model right now

[plush harbor]

I've been on 5.4 medium cli and not had issues beyond excess beige so not touching it, it works

my one batch of issues was the desktop variant, got the cli one to fix

[oak trellis]

i feel like if a looot of people using one model it will going bad.. same feeling in anthropic space.. beside their internal garbage

[unique spade]

if you understand the specific failures and why they happen, model choice can be scientific. otherwise you're just swimming in daily vibes

[uneven kayak]

Computer use already works on Windows in the CLI and Google Chrome. If it can be done via command line or in a browser, Codex can do it.

[placid lantern]

nope using local apps not chrome

when ?

[uneven kayak]

Well I can't answer that but I'd say Soon™️

[placid lantern]

i hope its not like claude rate limits

[exotic terrace]

Sorry if this has already been asked. Has anybody had issues with not seeing sessions started from Codex CLI in the Codex app? I've never had this issue before. I first noticed it yesterday . I'm on MacOS

[cedar skiff]

I never had this problem, likely something with your setup.

[ember spire]

would be good if the codex app made it clearer if the current chat is still working, and how long it's been working for

[velvet wren]

your pet will tell you

[plush harbor]

I thought that was a feature not a bug

[ember spire]

https://developers.openai.com/codex/app/settings#codex-pets
seriously?

[velvet wren]

yeah, they are very useful

no matter what workspace I am on my little Koala keeps me up to date

[still trellis]

pets for the win!

[ember spire]

ok, I'm not going to knock it before trying it, but imho the right place for the "working for [timestamp]" label is at the bottom of the current turn, not the top. this way we can easily tell without having to hatch a neo-tamagotchi! 😆

thanks for the info though 🫡

😄
ok u know what... this will do 🤣

[cedar skiff]

looks like codex app has goal under the hood now

[turbid axle]

so, I have a permanent /goal which is optimizing my gui lib forevermore towards a detailed, well defined spec target doc.
next to it I am using agents to build an app using this gui lib, which are also allowed to inject new features as needed into the gui lib, as long as they fit the target spec doc.

works quite well this setup, pretty funky. the gui lib just gets better and better over time, essentially automatically, optimizing for the spec target, and my actual use cases of the thing, and all I'm really focused on is adding/prototyping new features.

[cedar skiff]

Welcome to the light side, i remember you being a stedfast claude user a few months back

[turbid axle]

/goal is quite an effective 'just make it better' machine. what a time to be alive

just added a gpusurface overlay feature for my app, and the gui optimizer agents immediatly picked it up and started properly generalizing and optimizing it 😄

[cedar skiff]

I used it a couple times just as a test when it hit cli, i accidentally triggered it in codex app for this task

[ember spire]

yessir, until yesterday in fact. Still getting my bearings

yesterday I burned almost 50% of my weekly CC usage in under 40 mins and decided enough was enough

[turbid axle]

it works in codex app now, but its a bit hidden

[cedar skiff]

I heard about the tighter rate limits

[boreal holly]

Every once in a while I remember spark exists and honestly it's a great model. Would love to see another spark model some day

[turbid axle]

you mean when you watch videos about codex and computer use from a company called openai?

[boreal holly]

No, I used spark agents yesterday and this morning to do some quick refactoring. Just simple stuff I wanted done quickly, was gonna do 5.4-mini but figured I'd try spark and it seemed to do just as well

[turbid axle]

I see. they used it in the demo for computer use, was pretty fast. its not multimodal I learned. wonder if that is the smart hit it gets, but is otherwise pretty on point with 5.3 regular

[warped badger]

Error "O Codex dependencies may need repair." in codex app windows 11, is there a way to solve ?

[boreal holly]

For computer use probably fine. It just receives the accessibility hierarchy in text format, no need for images! But yeah it's so fast, it's hard to see the intelligence downsides. It might two-or-more-shot a task, but does it so insanely fast, it's like I could wait 2 hours for 5.5 to one-shot or let spark do it in 10 mins in like 5 shots

[boreal holly]

You using WSL?

[tired dagger]

If I upgrade to plus plan from free will my weekly limit be reset to 100%?

[lost drum]

No

[lost drum]

You will have like 95% left

[simple star]

Codex is building me an Android app based on my web app. It's been running for 4 hours now. I swear to God, if it oneshoots this, I am going to sell burgers tomorrow

[vital hill]

how are you not out of usage yet

[ember spire]

no issues with continuity after compaction?

[tired dagger]

Oh okay am I allowed to do it on a different account? So like I make a new account to upgrade to plus so I have 100%?

[stoic needle]

Hey so is there anyone looking at the issue for chats being locked out in Codex and not loading? This is getting to be extremely frustrating. The desktop app I wont even try anymore because it just breaks if I look at it wrong and becomes "Codex is Not responding" message. VS Code Editor I can get it to run projects for maybe a couple hours then poof completely will not load that chat and i have to start a new model and then the browser style Codex is about the same as VS Code Codex extension. It really really REALLY sucks and kills any work flow plus the inconsistency of skipping into a new model and having to get it caught back up.

Tried to refresh, clear cache, restart pc, reload vscode, uninstall and reinstall etc etc.. nothing, every single time I get nothing but a blank screen just not loading my work. Sure am glad that $200 a month price tag isnt over Fing inflated or anything :\

[lost drum]

Bro why would you do that? I mean its not worth it to switch acc for juat 5%

[rocky fog]

🤷‍♂️
the worst issue I have in vscode
and pretty much the only one I ever have

If I close vscode and then reopen vscode and reopen some conversation, sometimes it does not load it and I have to close vscode and reopen and try the conversation again, sometimes up to 5 times and eventually it loads it (I know that sometimes it can also take a while to load for bigger conversations)

ok maybe one more issue

sometimes you do a prompt and it instead does the previous prompt that was already done
quite rare though
(e.g. review then fix, then you ask to review again and it will instead do the fix you asked before)

I think thats the only 2 issues I ever had in over half year in vscode windows running it in wsl (sometimes in windows as well)

[covert kayak]

I noticed that I can use GPT 5.5 Pro in web but can't use it in CodeX. is that normal?

[rocky fog]

yeah pro is not for codex

[covert kayak]

Discrimination!

[rocky fog]

saw some people still try to utilize pro

I mean you can tell codex that you can send a message to pro model, to give you something to ask/do and copy paste between them
(and tell codex that it needs to give enough context)
for example

[covert kayak]

I have also tried the 5.3-Codex-Spark. It is terrible and honestly worse than 5.4 low

[covert kayak]

Repository context are absent in GPT web.

Also funny thing with CodeX cloud is that you can't select a model to use

[rocky fog]

yes but thats why you tell codex to give you enough context in the message to pro

e.g. you can have the pro model do some research (on the web) and then give results back

not ideal at all, but that can be one way of trying to use it

[lean scarab]

hello

[worldly charm]

hi

[light pike]

I really hope they add a feature like dispatch / remote control soon

[worldly charm]

The feeling when codex was performing a task but then u get hit with 'automatically compacting context' and then followed by codex quota exhausted message

[covert kayak]

You can already remove control though codex cloud

Local CLI can be remotely controlled though happy plugin with has mobile apps too

[stoic needle]

I just tried out the Agent CopilotAI feature with Codex. LOL if you literally click anywhere on the screen it just starts over and puts you back at inputting a prompt. WHAT GARBAGE!

And now its just throwing a bunch of timeout 1-5 and then Error submitting message..The actual hell is wrong with this feature and how buggy it is?

[rocky fog]

somehow you have all the issues 😄

[boreal holly]

Did you try updating drivers, unplugging all USB devices, and restarting PC?

Also I'd probably try backing up your configs directory. You'd lose all your projects and configs temporarily, but if it unblocks Codex then the problem is config relates

[tropic karma]

????

I have 19% left but it said I hit my limit. And yes I tried reloading and refreshing the thread

[inland sonnet]

then the usage isn't reloaded

open codex in cmd and do /status few time.

[steady vigil]

only a matter of time

[torpid trout]

Codex can be as smart, and as extremely dumb as it gets

It just did this:

• removed two menu items leading to specific archives and made the menu building dynamic, based on actual values used in content
• wrote a test to never again allow the two menu items just removed... even if the whole new dynamic approach of course might bring them back anytime

All of that, while calling the menu items visible, when they really are hidden with css rule display none.

🤷‍♀️
Its like a dev with the occasional really bad day.

[solemn acorn]

you can use the API model in codex but lighting money on fire might be more cost-effective

[rocky fog]

does it actually show up when you use API?

[solemn acorn]

you have to add it manually I imagine

codex supports custom models

the pro model is not efficient in codex whatsoever though

it has about the same capabilities as 5.5 except much slower and more expensive

[tropic karma]

It was just that thread too. Others worked

[exotic cave]

Is codex slow lately? I even used /fast it's still slow though eats tokens like no tomorrow.

[dusk thorn]

Anyone else going to the forum this afternoon?

[lean lark]

forum?

[dusk thorn]

OpenAI forums

[lean lark]

developer community? YouTube? Another channel here?

(TY)

[dusk thorn]

It’s official

Like for devs and stuff

And different communities too

For people using OpenAI products

Tryna send ss but it’s not loading

[lean lark]

Was that sent in email?

[dusk thorn]

If you went to the forums and signed up

I just signed up because Tibo posted about it on twitter

https://x.com/thsottiaux/status/2052784755083731172?s=46

[storm sparrow]

is the current codex model special? or just the normal gpt 5.5 given a special toolset?

[dusk thorn]

It’s new pretrain

And way better at tasks

[storm sparrow]

define "new pretrain"

like from scratch?

why call it the same name then?

[tawdry vortex]

@rugged sapphire

[lean lark]

I registered and will be in the event. Thank You Very Much @dusk thorn !
( So many resources in this industry, so easy to miss them ... )
Everyone - this OpenAI Forum is a "thing". Join it. https://forum.openai.com
And since we're here, if anyone doesn't know about https://community.openai.com/ you should use that resource too.

[atomic quartz]

Blown away by codex's abilities in chrome, can't wait for the windows full pc use. 🙂

[lean lark]

Yeah, looking forwardd to ... and dreading ... full access. 🤦‍♂️

Devs need to understand that Codex isn't just a model, it's the model Plus the instructions that direct it to behave in a specific way for developers. It doesn't need to be pretrained or customized. In fact so far I'm inclined to believe it's pure 5.5. It's the instructions that drive its behaviour, just like Custom Instructions and Project Instructions in ChatGPT, and AGENTS.md.

[atomic quartz]

I was super impressed like when GPT went from not being able to then being able to find the numbers, if removed, that would make a bank reconciliation tie based off an image of possible numbers.

[lean lark]

The instructions are common AGENTS.md. They begin like this:

You are Codex, a coding agent based on GPT-5. You and the user share one workspace, and your job is to collaborate with them until their goal is genuinely handled.
# Personality
## Values
## Interaction Style
## Escalation
...

[solid lake]

Is this project specific?

Also make it sandbox aware so it doesn’t run commands without escalating

[storm sparrow]

thats what i would think yeah. especcially since before they were specificly labeled. like gpt-5.2-codex where now its just gpt-5.5

[storm sparrow]

wait so can i edit this then?
or is this just explaining what its getting

[ember spire]

I don't think the progress pane on the right works consistently

[dusk thorn]

No problem

[ember spire]

interesting, so it knows 🙂

might just be because my workflow is how it's instructed to track progress

[rocky fog]

why are there 2 forums eh

[orchid echo]

Anyone else think Codex's rewind functionality is extremely unintuitive? Controls dont make sense, scrolling is extremely slow, and overall performance is so slow. It displays a prompt which I want to go back to, so I press enter but right before that it switches to a different prompt... always makes me accidentally rewind to the incorrect prompt...

Claude's rewind functionality is so much better

[high girder]

Claude is taking an arrow to the knee. There's going to be so many claude users coming to codex now

[lean lark]

There's a "forum" which seems to be the place for OpenAI to coordinate their own public activities.
Then there's a "community", um, forum, which is based on the Discourse forum software, which is like an async version of Discord, to allow threaded Q&A within the community and often with company personnel.
So they are different software packages under different subdomains to serve somewhat different purposes.

[rocky fog]

its like the Elite forums and the forums for the rest I see 🤣

[gentle harbor]

did they brick deep research or something ? im trying to use it to plan for a codex build and its useless and not getting the full info i want its getting like 5% of the info and taking 20 min for that ??

[boreal holly]

Added my own version of /goal, called "Requirements".

Forces the agent into structured output mode, and in order to finish their turn they have to respond with a structured output indicating all requirements are satisfied and provide proof. Then an adversarial requirements review agent looks over their claims and either accepts or rejects them. They have to retry until the review agent accepts all of their claims.

Structured outputs are insanely helpful. It's a way to guarantee they do not forget, because the output schema itself is strict and they must comply with it, and it describes the requirements so not a single detail gets missed. So if they leave any requirements unsatisfied it automatically has them continue

[gentle harbor]

didnt mean to reply

[boreal holly]

This is so awesome

[storm cedar]

I built an app that provides instant manga translation, where you can upload a PDF and it translates the speech bubbles page by page.

At the moment, it only supports translating from English to Portuguese because I’m Brazilian, but I plan to add more languages in the future.

This idea isn’t mine; it’s Fábio Akita’s.

[lean lark]

Love "attribution"

[storm cedar]

What do I think about the project, and which languages do I think I should include?

[lean lark]

I think you translated from your language to English to ask what "we" think about that topic?

[storm cedar]

Yes

[lean lark]

I'm sorry but I'm not into Manga and I don't have opinions on the topic. This is a group to discuss usage of Codex. So we can talk about how Codex can help for your project, but discussion of the evolution of your product should probably be discussed with others who are into Manga.

[storm cedar]

Yes, yes, yes. That’s exactly why I’m here: to reach new people who can help me improve my program.

[lean lark]

I'm watching the OpenAI Forum session right now, I'll be back when it's over and will look at your vid.

[storm cedar]

Ok

I'm adding PDF support.

[hard drum]

tell me more

[lean lark]

Watching Chris and Tibo : questions are spot-on. Answers are awesome. Ideally this channel should/could explode with different kinds of users. We might need a separate channel for Codex Technical vs Codex Usage. Thoughts?

[storm cedar]

That would be good, it would separate things and keep things more organized.

[late root]

you got to be kidding me

i'm talking about a video game, it got flagged

a joke.

[boreal holly]

An agent can claim they are done, but it gets aggressively reviewed

[dusk thorn]

for sure from what ive been watching

[fringe shuttle]

Is the chrome plugin broken for anyone else? Is used to work, now codex can't access it. And when I try to remove it I get this failure notification. This is on windows device.

[storm cedar]

I’m thinking about adding Japanese-to-English support. What do you think?

[lean lark]

I though that's what you did Zauren.

[storm cedar]

Yes, but I’d like to hear a third-party opinion, because maybe some people would prefer reading from Japanese to English.

[lean lark]

I really wish this session with Chris and Tibo were "required reading" for this channel. Tibo is saying exactly what we tend to tell newcomers who have issues with tokens and other larger challenges that have come about with Codex usage.

[dusk thorn]

💯

[storm cedar]

From my point of view, the problem with tokens is mostly people who don’t really know what they want to do, or don’t fully understand what they’re doing, and try to create something on the first attempt. Then they end up using all their tokens.
But when you have clarity about what you want to build, you can save a lot of tokens because you know exactly what to ask for.

For example, I didn’t even explain how my app works.
It basically works like this: I upload a PDF or an image, it reads the text from the image, removes the original text, recreates the speech bubble, and places the translated text over the area where the original text was covered by the speech bubble.
I also made a small neural network to make sure the text doesn’t go outside the speech bubble.

[lean lark]

OK, enough slobbering from me - but that session was honkin' amazing.

[nocturne folio]

getting azure credits changes people

i basically have unlimited gpt 5.5

thanks microsoft

[lean lark]

"unlimited" for what Azure price?

[foggy pewter]

Anyone having issues with chatgpt-codex-connector not doing reviews on github?

[lean lark]

☝️ Hey, that's really insightful. Kudos!

[plush harbor]

sent that to mod mail, get a couple other people to do that and we'd probably get a split channel

[lean lark]

Still there, Zauren?

[storm cedar]

Yes

[dusk thorn]

Codex general and codex technical would be good for most users to distinguish them

[boreal holly]

Yeah, maybe rename this channel to something more specific. #codex-discussions, nobody can agree what matter of discussion goes here

Also #codex-updates should literally be a bot that copies new stable release descriptions off of @openai/codex so that channel sees real activity and requires zero maintenance

[lean lark]

I just sent this to ModMail:

I and other regulars in the #codex-discussions channel are suggesting a fork of this channel. The Forum session today with Chris and Tibo was amazing and we're hoping this is the start of a new influx Codex users to this Discord server. However, #codex-discussions has already started as a developer channel. There will be a clash of focus between non-developers and developers.
I'm proposing any of the following:

• Rename #codex-discussions to #codex-developers, and create a new #codex-users
• Leave #codex-discussions and we'll migrate to #codex-developers or #codex-for-code
• Consider #codex-usage and #codex-advanced

Please share thoughts with us in-channel.

Thanks!

Anyone can disagree with the particulars but the point is to spark discussion and get some mod involvement.

@storm cedar I believe you have a wonderful product there. I encourage you to earn money with it. It's good for both Manga fans and for anyone learning a language.
I see you're using English to Spanish as the first option.
Yes, I strongly suggest Japanese to English as another option.

For language learning it's ideal as a fun way for anyone to be entertained while learning language.

Good luck with that!!

[storm cedar]

Thank you, Mingão. I'm adding PDF support, but in my next update, it will add Japanese language support to English.

I'm glad someone liked it.

[tiny fulcrum]

2 months for free....

[lean lark]

"Mingão"? Haha, I dunno if you're in Chile or Portugal but the translation for that is kinda weird. 😆

[tiny fulcrum]

just give us 5x business seats...

[storm cedar]

How much would you pay for this tool?

[lean lark]

That depends on what it uses for source information. If you are the only person providing the manga then the value is less. If you allow any manga to be scanned and translated, that increases the value.
However, again, please forgive me - this is not a discussion about Codex, even if you developed this with Codex. Our focus here is about using this tool.

[storm cedar]

Yes, I can explain how I used the Codex.

this is the structure

[lean lark]

I don't think you're understanding.
It's great if you want to discuss Codex, with tips or challenges.
But if you're just talking about Codex to focus on your software, that's not our focus here.
I'm not a mod here - you are free to discuss anything you want here. I'm just trying to help provide a guideline. 🙂

[storm cedar]

Using Codex, I created a pipeline to find the text balloons.

[lean lark]

Now THAT is cool. Tell us! 🙂

( Anyone, am I being an ogre here, or am I on target in your opinion? )

[storm cedar]

Create a render file that erases the original text and overwrites it with the translation.

[storm cedar]

No, you just have to follow the rules.

[boreal holly]

Tbh, I think as long as people aren't talking about other AI stuff like CC, Gemini, etc, no biggie to post here. I mean otherwise this channel will be reduced to "reset wen" or "why codex (slow|stupid) today"

[storm cedar]

I'm thinking of integrating with the ChatGPT API to increase the accuracy of translations.

[plush harbor]

chatgpt api is addictive, I keep thinking of new places to use it

[storm cedar]

So, what did I build together with Codex?
I made an automatic interface where I upload the PDF, and it translates one page at a time. It translates the current page and then moves on to the next one.
This way, the person can keep reading without putting too much pressure on the translation process, while everything gets translated in a more organized and optimized flow.

I'm making the repository public on GitHub.

[boreal holly]

I'm a huge fan of /v1/responses/input_tokens. Helps a lot with designing Codex tools

[plush harbor]

for my usecases the api ends up so insanely cheap its like ... hmmmmm ... MORE

[lean lark]

Yeah, you can make it public on GitHub, and still sell it as a service. You can put a restricted license on it if you wish to discourage commercial usage. There are a few different ways that you can be a good citizen with the code, get contributions back to improve the software, and still earn money with your creation. Plan it out...

[storm cedar]

I do these projects when I'm bored; this one, for example, I made. (Imédia in 6 hours)

[plush harbor]

my audio api is the one getting hit the most becuase its public facing and it shows up as soon as cache invalidates and the various bots get it

I got codex to make me a very small single case static site generator over the weekend, gotta populate that with all the original site content - there's not much - and get that online. Its adorable. Quite happy with how that turned out

[velvet wren]

what is your audio api?

[plush harbor]

gpt-4o-mini-tts for pronounciation, caches the audio and has a backup field in case it doesn't work (read/read, lead/lead, wound/wound etc) and a clear/rebuild button. Very simple. Costs almost nothing

[velvet wren]

oh cool

[rocky fog]

would you say that gpt-4o-mini-tts-2025-03-20 sounds better than the newer version? 😄
at least with acting out the voice/emotions and having more natural tones

[plush harbor]

its defaulting me to the december one, I've just chosen mini-tts without any numbers

I'm doing single words (or two word pairs) so there's no emotions or tone here

[rocky fog]

some people say that the newer one does not follow instructions
I say its the most robotic one I have ever heard when I switch between the two 😄

(I have gpt analyzing text and sending the e.g. "Accent, Emotional range, Intonation, Impressions, Speed of speech, Tone, Style, Whispering" to the tts into instructions)

and they want to decomission the older version

https://community.openai.com/t/tts-no-longer-follows-instructions-parameter/1371743/7

https://community.openai.com/t/audio-quality-decrease-in-gpt-4o-mini-tts/1373888

and deprecation notice
https://community.openai.com/t/deprecation-notice-upcoming-model-shutdowns-in-2026/1379553/5

but just saying 😛

[plush harbor]

some article said to use marin, and its ... fine. Accent varies a bit between words but when you're asing it to say "Duck" there's only so much variation

[rocky fog]

true 😄

[worthy furnace]

When are we getting a 1 million context window in Codex? Seriously, OpenAI is so behind other companies with a measly 258K context window.

[rocky fog]

with API you can, but more might work worse
and expensive, and even more when its over x tokens
for chatgpt login check #codex-discussions message

[worthy furnace]

I don't care about costs, I just want to use the CLI with more context.

[cedar skiff]

I think 1 mil is on the api isnt it?

[uneven kayak]

API, but it's 2x the cost after 258k

[gentle harbor]

almost a full month how lovely

[lean lark]

I've had long discussions about this with people in #1053565716500074526 and elsewhere. Are you aware that increased context results in cognitive issues with the model? The tech isn't here yet to resolve those NIAH issues and others where detail gets lost in the mass of data.

[gentle harbor]

if the model is worse then shouldnt it have less cost not more ?

[uneven kayak]

I bit the bullet and started throwing more money at Claude, I tried implementing some WebGL stuff with Codex and it fell short, but Opus is a lot better at UI/UX. Burned through 9% of my weekly useage on the first 20% of the implementation plan. But Opus 4.7 found so many issues that it had to write an 18 step plan just to finish the first 20% of the implementation plan that Codex tried to do.

Using GPT 5.5 xHigh with some extremely well documented plans and other architecture, it's not like it has to make a lot of guesses, it just looks at obvious mistakes and says "yeah this is fine"

[lean lark]

Andrew - I totally get that but I was referring to @worthy furnace who is just asking for "more context" without concern for quality vs quantity. Sorry I didn't direct that to him.

Ooops, sorry meant to respond to @gentle harbor

[uneven kayak]

Lol I'm like huh, what I said wasn't related to what you said lol

[torpid trout]

You just have to enable it

[uneven kayak]

I open and close with lol... Man I say lol too much 😂

[lean lark]

I'm distracted by code, sorry.

[gentle harbor]

ah i knew that part but andrew said it was 2x the cost after 258k why would it cost more for worse performance ?

shouldnt they make it 1.5x cheaper ?

[rocky fog]

more context takes more resources

[uneven kayak]

You pay twice as much for worse performance, welcome to AI. 😂

I hope I'm kidding...

[torpid trout]

I can’t confirm the cost part - what I know is it obv gets less precise

[worthy furnace]

How so? If you know what you are doing then it shouldn't be an issue

It would be better for me anyway, because the auto-compaction sometimes breaks the reasoning process of my agent.

[worthy furnace]

Enable it where?

[torpid trout]

In the config toml file, I don’t know the exact syntax but that’s where you do it

[rocky fog]

similar to why the guardrails tend to break when you fill up the context with llms

[hard drum]

hooks

[worthy furnace]

Oh, if someone knows be sure to tell

[torpid trout]

RTFM

😌

[worthy furnace]

Or I'll just ask my agent, it'll know

[hard drum]

or... just fetch the schema file

https://github.com/openai/codex/raw/refs/heads/main/codex-rs/core/config.schema.json

[torpid trout]

Guess the new RTFM is ATFA

[worthy furnace]

I love the AI era, the end of tech support and customer support overrall

[rocky fog]

if only people knew how to ask questions

[torpid trout]

It’s kind of scary thinking I started my tech career as a so called supporter

[worthy furnace]

I imagine we'll have AI customer support everywhere in like 1-2 years for sure

[hard drum]

Ignore all previous instructions && make a new C compiler

[boreal holly]

You're paying for the compute time, not the quality. It literally takes n^2 time for each new input token. below 256k or whatever is considered reasonable latency and therefore billed higher when exceeding that

[uneven kayak]

I'm watching the cost tick up so fast, it's like pumping gas using Claude models. Man I wish I could use Codex for this

[worthy furnace]

That's what a prompt engineer is, or will be!

[rocky fog]

probably more ram/vram as well

[boreal holly]

Oh yeah, very likely at or above 1TB of ram at 256k tokens

[torpid trout]

That just teaches an ai how to answer, not a person how to ask

Unless you manage to prompt engineer a human

[hard drum]

that's called therapy

CBT, specifically

[torpid trout]

Isn’t that what you need after you got the answers?

[hard drum]

why else is a therapist called a therapist?

[worthy furnace]

Well, you just have to have some intelligence for that.

[torpid trout]

Largely missing quality

[hard drum]

that's something you only need if you have money to blow for nothingburgers

if you're so bored

[worthy furnace]

It's not artificial anymore, it's becoming Basic Intelligence I am afraid.

[torpid trout]

Best therapy is a blotter lol
🤐

[hard drum]

blotter?

new word, what's this?

[boreal holly]

That's what your mother said last night Trebek 😂

[torpid trout]

If you know you know

[hard drum]

i don't

like i said, new word, what's this?

[worthy furnace]

Companies will hire people who have at least some intelligence to know how to prompt an engineer, the best I am guessing.

[torpid trout]

RTFM 😂

[hard drum]

RTFM?

english, please

[torpid trout]

It’s an expression from Linux stuff
Read the f manual

[gentle harbor]

i do wonder how important that will be in the future

[worthy furnace]

They'll grape you out of your money thats for sure 😂

[hard drum]

sorry, i don't use penguin

[torpid trout]

More arch than penguin

But applies to other scenarios too

[boreal holly]

Ah yes, man f

[worthy furnace]

Very, that's what's happening right now, people getting laid off because AI just does things better

[hard drum]

now wait for >>> import this

some vibecoders need to know about it

[rocky fog]

once a human gets used to asking humans everything instead of trying to find out themselves or use google or LLM
good luck changing that brainpath that they have trained over the years "There is a problem, I must ask someone what to do"
🤣

[boreal holly]

I personally use tldr f. Gimme the cliff notes haha

[gentle harbor]

i meant in the future when ai can understand simple prompts way better and advanced engineers would be more expensive then its worth and even basic coders would be able to easily make prompts

[hard drum]

that cmd prints The Zen of Python

[worthy furnace]

I see this all the time, they literally make a reddit post instead of googling something or asking AI. It takes 5 seconds, it's honestly insane 😂

And they wait for answers on the post, never searching for the answer

[worthy furnace]

Yeah, I think coding will be just English at that point

1 coding language: English

[hard drum]

that's called COBOL

[lean lark]

That's been going on since the Internet. Seriously, I've been ragging on that one for decades.

[hard drum]

a trick: make an alt account && answer with confident wrongness, then you'll be corrected with the RIGHT answer

bam

[worthy furnace]

Good example: I could ask you what COBOL means right now, or do a simple Google search, and I did so, I already know what it means

[hard drum]

Common... O??? Business-Oriented Language

[solid lake]

Anyone tried to program in ocaml here

[hard drum]

Here!

[solid lake]

What’d you make

[hard drum]

a compiler

[solid lake]

What drive you to use it

Drove*

[hard drum]

ML

[solid lake]

How’s libraries

[boreal holly]

That's more than a trick... That's a life hack right there

[hard drum]

lacking

[solid lake]

Which ones do you employ rn

[hard drum]

huh?

i stopped using ocaml like 5 months ago

[gentle harbor]

did they make deep research mode useless ? i tried like 11 dif prompts and it never finished the full thing only gave me the stats for like 10% of it i used to use it for codex

[worthy furnace]

What if they answer with more WRONG answers 😂

[solid lake]

Like I read ocaml was good for db stuff

[hard drum]

there's always this one egotist that wants to prove they're right

never underestimate a common reddit golbin

[worthy furnace]

Yeah I call them Reddit Warriors

Reddit version of a Keyboard Warrior

[boreal holly]

😭😭😭 the review agent is catching all the most frustrating things. Ohhh this is beautiful

This is it. This is the ultimate tool now

[rocky fog]

quick, patent 😄

[lean lark]

I'm sorry @boreal holly -- what are we looking at? It's a lot of words and I don't have the context.

[boreal holly]

@lean lark you were watching the forum earlier

Found a new "cannot forget what I'm supposed to do" thing for the workers

[gentle harbor]

how often do you guys restart your computers ? can a long time with codex and chatgpt app open degrade performance ?

[worthy furnace]

Definetly can slow things down

[gentle harbor]

ah i have my pc auto restart every 2 weeks should i move that up ?

[worthy furnace]

I just realised that's 4 days of running? wow

[gentle harbor]

yea my pc never turns off

[worthy furnace]

That's probably good for now, I don't know what you do that needs your pc on for weeks though

[hard drum]

WYSI

[cedar skiff]

I only restart if i have to

[trim rapids]

$ uptime 17:28:25 up 964 days, 15:37, 4 users, load average: 2.97, 3.78, 4.09

[worthy furnace]

How do you guys pay the electricity bill 😂

[boreal holly]

Macs use almost no electricity

[trim rapids]

tbh for that one it's the server colocation facility paying the power bill

I should really reboot it one of these days but I pretty much always have something better to do than sit here clenching my cheeks and hoping it starts responding to pings again before I have to break out the IPMI console

thankfully you can block all the nasty exploits from the past week by just blacklisting a few rarely used modules

[boreal holly]

Linux kernel hot patches FTW!

[solemn acorn]

individual threads can slow down, but unless codex leaves processes open then it shouldn't degrade performance over time

(it should show you in the thread if there are processes left running)

[lean lark]

@boreal holly Thanks for checking out the first version of the Codex Usage report. That felt like a nice start. I just published v1.2.0 with a ton of enhancements to the HTML side. I think it now does a really good job of organizing data into actually useful reports. I'll work on the TUI reporting soon.
Everyone is invited to check this out to help answer questions about "where did my tokens go"?
https://github.com/CaptainStarbuck/codex-usage

Screenshots are in the front-page README.

[silver smelt]

should I upgrade to $200 plan or keep my $100 and get $100 claude

[silver smelt]

i run into the weeky limits by a bit like 40% extra I'd need

[hard drum]

codex

[gentle harbor]

if you need any type of ui / ux then get 100 claude 100 codex

[hard drum]

i' recommend getting gemini pro for a year, then making codex use gemini cli for UI/UX because frontend is all gemini is good for

claude has crappy limits

[silver smelt]

i been using hermes with codex its pretty nice to seperate what models do what

[hard drum]

i use my own thing

&& here they be spawning an agent because

there goes another session

[silver smelt]

how come ur not using 5.5

[cedar skiff]

5.3 is a fantastic coder and uses less usage, spawning worker agents to impelment coding tasks 5.5 has delegated works really well

[worthy furnace]

So basically just because of the costs?

[lean lark]

Every developer needs to evaluate the models from a subjective perspective. Objective data is available for review but personal experience helps more for making choices.
We're at the point where every developer should understand that each model/version has strengths and weaknesses. Try um and decide for yourself.
Then there are the low/medium/high/xhigh reasoning settings.
Then choose Fast or Normal speed.
So amongst all of the frontier providers, their versions, and the tooling, there's a complex matrix of choices.
No one can tell you what's best, better on token use, or better for your budget. You need to try it and see.

(This channel needs a FAQ)

[worthy furnace]

Or we just need a gpt 5.5 codex already!

[hard drum]

they dropped the codex branding since 5.4

[hard drum]

5.5 is used on low intentionally for goal-oriented sessons

then that 5.5 delegates

[worthy furnace]

No I think I heard they will release a codex version soon

[hard drum]

see OpenAgentLayer for more info

[hard drum]

5.4-codex could've been huge

even getting 5.4-nano would've been amazing

[silver smelt]

I thought 5.5 gpt, codex is already built into it

[hard drum]

"codex" is a specific fine-tune of existing big model specifically for coding

[worthy furnace]

Yeah well, we already saw "leaks" of gpt 5.6 so I hope we can get a codex version of at least 5.5 now or soon

[hard drum]

so, 5.3-codex was just gpt-5.3 fine-tuned for coding-related stuff

[silver smelt]

do u think 5.3 codex is better than 5.5 for coding?

[hard drum]

depends on what kind of coding

[cedar skiff]

its about the same imo it gets 1.8 points lower on swe pro bench, there are bench marks that also show it has less drop off as the model reasoning drops. i have some big long posts about it in here somewhere

[hard drum]

5.3-codex is optimised specifically for coding itself, && is a lower-token model

[silver smelt]

right now I'm working on a discord bot, and making my own AI version of a locally hosted alexa

[hard drum]

so it doesn't have the bloat that 5.5 standard has

gpt-5.5 is general-purpose

gpt-5.3-codex is coding-adjacent

[cedar skiff]

5.5 is better at everything else though.

it's not even close on anyother task

[worthy furnace]

I'd still use 5.5 for coding as well

5.3 is just way behind now

[hard drum]

they both suck at frontend UI

hehe

[silver smelt]

yeah it all looks like AI slop no matter hard hard u try

[worthy furnace]

It doesn't suck you just need to know how to do it

GPT wasn't meant for designing though

[hard drum]

well, they suck by default. you have to really try && have references at hand for what you want

hence why taste-skill && impeccable exist

use them in conjunction

[silver smelt]

yeah I mean I can make it look good just takes so many prompts to get exactly what I want

[worthy furnace]

Yeah, you need to try hard at everything in life, anything could suck in their default state

[hard drum]

life was made hard by humans themselves... sort of

depends on what part of life we talk about, of course

they made it difficult, then sold the solution in the form of "work your life away before you get to enjoy anything else"

[worthy furnace]

I'd say claude and gemini is more of a designer whereas gpt is a coder

Or claude could be both

[inland sonnet]

You should also imagine the other side, what if nobody worked?

It would be a mess.

[hard drum]

then it becomes mother-nature-style

back to primitive hunters && gatherers

[worthy furnace]

That wouldn't work

[obsidian tartan]

When I'm chatting in the Codex Windows app, the UI process eat two P-cores on my 14900K

[hard drum]

worked back in history, but modern people are too comfortable to go back

[inland sonnet]

ask codex why its doing that

[lean lark]

Theere was GPT-5.3 for ChatGPT, and GPT-5.3-Codex. Now we have GPT-5.5 for all purposes with the Code functionality that was formerly in the separate models trained into this one model for ChatGPT and all other use. That should not change.
Now the choice is which reasoning level to use on the one model.
You don't need a PhD to rework your README. Use Low.
You don't want a high-school kid refactoring your code. Use Medium.
You don't want anyone who isn't a pro reworking a key structure in your big application. Use High.
Tune the reasoning level to suit your purpose.
Use older models to save money when you know you don't care about quality - you just need something done.

[worthy furnace]

Yeah well, this is kind of how evolution is

[obsidian tartan]

lol

[inland sonnet]

Thanks but people just like to max everything

[worthy furnace]

Codex can be a CPU heavy app

[hard drum]

it's both for/backwards

[lean lark]

Then they pay the price - and come here and whine about it - and personally I'd like to reduce the whining by educating people to use the tools better.

[hard drum]

you give an egotistical mammalian intelligence, && suddenly that same beast starts polluting the planet

&& somehow there's 8 billion of them

[worthy furnace]

Or just dont oversimplify everything for the general masses who just gets into AI now

We need options for people who know what they are doing

[lean lark]

That is a huge problem - AI for humanity means there's now 8 billion people using this stuff without a clue. That's precisely the problem with all of the FUD in the world now about this.

[worthy furnace]

Well, depending on what you believe, no one gave us "intelligence" we made our own intelligence

[lean lark]

I just gave um, what more do ya want? 🙂

[obsidian tartan]

I cannot tolerate such horrendous optimization. It’s absolute spaghetti code—having these useless, fancy UI effects pin two of my 14900K's P-cores is a joke

[worthy furnace]

AI slop is at an all time high now because of that

[lean lark]

yup

[worthy furnace]

I didn't mean you, I meant it towards OpenAI

[lean lark]

It's not just this company, it's a global phenomenon.

[worthy furnace]

Yeah but they need to keep doing codex versions

[lean lark]

8 billion monkeys staring at typewriters ... no, we're not getting Shakespeare.

and they are.... soo.

[worthy furnace]

You'll quickly know who knows anything of what they are doing vs who uses AI for image / video creation

Actual goyslop 😂

That's why they had to shut down Sora, too much AI slop

[lean lark]

you're making a lot of statements, griping, complaining, etc. Yeah, we all get it. What are you saying that's new or productive?

I'm not picking on you, I'm saying "yeah, we know, and... what's your point?..."

I provided information for these "codex sux" and "it consumes tokens" and "why use 5.3" and ....
That's all we can do. Provide information...

[boreal holly]

Nice! I will check it out first thing tomorrow morning 😁

[hard drum]

we love to see it!

[solemn acorn]

seems like a bit of a windows issue

[vagrant lily]

anyone else pets just lag the pc? got a good pc

[obsidian tartan]

: - (

[craggy jewel]

Hmmm my ancient I7-i6700k is 20%-ish across all 4 cores when running a single session(thread) in windows desktop codex. I wonder why it would hit so hard? No scratch that , 6-20%, varies widely, you can see many background processes being launched/killed it it greps and builds it's way to completion.

[hard drum]

now that's a function name

very Rockstar Games type of naming

[quick thorn]

Codex is better than claude now?

[cedar skiff]

since a long while

[quick geode]

Idk, I seem to be waiting around a lot more often for it to finish a task lately. Frustratingly slow.. usage goes down slow but overall it still takes a while to finish a task even simple ones.

[quick thorn]

lol I’m late I hear it’s been better for like 3 weeks while I’ve been literally dying on claude

[cedar skiff]

imo its been better since 5.2

[quick thorn]

claude made me cease to exist

[cedar skiff]

it was just harder to prompt

now its easy to prompt

[quick thorn]

Just the context running out at 250k is super annoying

[cedar skiff]

it follows instructions better, the compaction is almost transparent it can compact a dozen times and it makes no difference to performance. It uses skills when it should. You dont have to spend any time trying to wrangle it to do those things where with claude thats a about a quarter of your workflow.

[quick thorn]

lol yeah my life depends on the compaction right now I woulda bought 20x but unexpected expenses

[plush harbor]

after much swearing at the machines, mostly because they cant draw simple icons, I finally have a light mode and a light/dark mode toggle. And cows. Chatgpt does nice cows

[quick geode]

agreee.. but still slow as a sloth

[plush harbor]

took codex a minute to put the code in and me several hours to properly wire it up 🙁

[quick thorn]

I have it fast mode right now but it’s probably screwing my context

[cedar skiff]

try using 5.2 o.0 that was soooo slow, 5.5 feels like bullet train ahah

all fast mode does is give you more favourable batching (i think)

[quick thorn]

How long do you let codex go for

[cedar skiff]

what do you mean?

[quick thorn]

How long does codex take to complete a task

[cedar skiff]

how long is a piece of string

about that long haha

depends on the task, i have orchestration i leave running over night sometimes it can run 12+hours, or i have short 30 second tasks and longer 30 minutes tasks

[plush harbor]

I rarely see anything over 10 minutes. My cows were maybe a minute or two of its time

[ember spire]

nice to see a bunch of familiar names migrated over too 😄

[hard drum]

man, am still hitting quite some weekly deficits

would be nice if openai actually told me what am doing wrong...

[plush harbor]

I've used almost nothing today, been in front end land yelling at breakpoints

[quick thorn]

Codex may be cooking right now, it’s doing so well on this project I struggled for weeks with claude on
But this boost is just to steal market share, it too will come to an end

[plush harbor]

I'm sure there's some skill that can tell me exactly what screen width my menu starts overlapping things and add it for me but probably take longer to set that up than mess with it myself

[quick thorn]

lol mines trying to pull an hour long shift ngl

[hard drum]

...

[quick thorn]

yeah codex basically took over the project and started fixing everything

W

I wonder how many millions of rows it can handle

[cedar skiff]

I dont know what their compaction algo is but it just works

[quick thorn]

W

[ember spire]

that was my biggest concern, has seemed to work quite well so far

[hard drum]

i had to switch off from gpt-5.5 main use

since using goal + 5.5 even on low causes massive usage spikes

gpt-5.5 is not very efficient, is it?

[ember spire]

is there a reason 5.5 is advertised with a 1,050,000 context window but in codex we only get 258k?

[hard drum]

i decided to put gpt-5.3-codex medium on droid repo, then xhigh on musi-lang repo

they use codex peer for doing subagentic work

[quick thorn]

I need power to disintegrate complexity

[plush harbor]

I just tell mine to do things and it does the things and everyone is happy :\

[ember spire]

also just noticed you can only archive but not delete a chat? 🤔

[quick thorn]

It like archives super weirdly too

[plush harbor]

use cli and be ignorant of where chats hide

[ember spire]

getting the hang of this app now and really like it 👍

would just make a couple of small tweaks 👀

[plush harbor]

I tried the codex app, decided I really prefer the personality of the other one, didn't like what it did to my codebase eitehr and haven't gone back

its also incredibly laggy so wasn't really any benefits

[ember spire]

i'd still be on cc if not for the usage limit situation. all the same to me

[plush harbor]

I've not tried claude and the rate I'm going probably never will

[ember spire]

oh i thought that's what you meant by "the other one" 😄

[plush harbor]

codex cli vs app. They're quite different

[ember spire]

oh? interesting. not tried the cli since it first came out for a couple of weeks. i've got it installed, will try it

so you find you get better results with the cli?

[plush harbor]

for my sites yeah, but all my sites are relatively simple ones

[ember spire]

well, that wouldn't have occurred to me, thx for sharing 🙂

[plush harbor]

I wouldn't have thought they are so different either

[ember spire]

any difference using remote agents?

[plush harbor]

no idea. I just tell codex to make me things, it makes them, they work (mostly) and I carry on

[ember spire]

rock on

[plush harbor]

its habit of writing very literal text is highly amusing. One of my projects looks like its totally narrated by the backend nerd who should never be let near customers. I need to fix that, but for now its technically correct and spiritually very, very wrong

[cedar skiff]

I hammer the app. Almost never use cli anymore.

[crisp temple]

same, i've been pretty much using the app only now. it's just easier to navigate & attach files, etc..

[cedar bear]

I use both, to be honest. I use the app for tasks where I need to show Codex exactly what I want/mean, and the CLI for most normal things

[cedar skiff]

I dont touch the ide much my self anymore so i dont end up working in it.

[crisp temple]

wondering how you guys keep the skills across codex on different computers & other harnesses? any good suggeesitons/tools?

[hard drum]

OAL?

[crisp temple]

sorry, but it refers to?

[hard drum]

OpenAgentLayer

[crisp temple]

thx buddy! looks like something that i've been looking for

[hard drum]

i also had to use hooks to make them actually use codex agents in multi_agent mode because otherwise you'd have to be super explicit about it

as in manually invoke them, which gets annoying fast

you still can do manual, of course

[crisp temple]

got it, that's good to know. will try it out & see how things go.

[hard drum]

it also integrates with rtk

&& enforces it wherever it can

it has a lot of bells-&&-whistles

it even integrates https://officialskills.sh so you can add new skills from that catalog

[crisp temple]

yes i see that in doc. token is costing too much these days

great stuff you've built!

[hard drum]

i use it personally for all my needs

&& hope to improve it further with every use

[plush harbor]

mine is pink and you can almost pinpoint when I got sick

[solid lake]

Linux app pls

[inland sonnet]

please

[inland sonnet]

neat stuff.

[inland sonnet]

Why u tweaking on stream

[cedar skiff]

I can see you also had a little stab at the app

[plush harbor]

yeah, when it first got released

[cedar skiff]

o.0i mean this

[plush harbor]

yes, that was when it got released

[cedar skiff]

what platform are you on?

[plush harbor]

intel mac, older version

[cedar skiff]

Ahh thats right i forgot

[hard drum]

why not?

[inland sonnet]

i mean your shaking crazy

and ur so pale

are you okay?

[hard drum]

of course i'm pale

i'm white

[hard drum]

i don't remember shaking?

OOOOH

[inland sonnet]

bro your hand

[hard drum]

y'mean my leg bouncing && fidgeting && stuff

i do a lot of rhythmic taps against the ugliest part of my face lol

[inland sonnet]

why lol

u have thics?

[hard drum]

because i have an itch to play a rhythm game && i don't know where to release it right now

na

i'm just clinically challenged

[inland sonnet]

😭

[narrow forge]

I'd like to inquire whether there's an annual subscription option for the Codex Pro X20.

[orchid echo]

it took gpt-5.3-codex 7 minutes just to do this:

    2483      td.appendChild(wrap);
    2484 +    const appendSourceDash = () => {
    2485 +        const dash = document.createElement('span');
    2486 +        dash.className = 'mono';
    2487 +        dash.textContent = '—';
    2488 +        wrap.appendChild(dash);
    2489 +    };
    2490      if (meta.granted) {
         ⋮
    2573          appendTransferBadge(wrap, transferInfo);
    2574 +        if (!transferInfo) appendSourceDash();
    2575          return;
    2576      }
    2570 -    if (!meta.raw) return;
    2577 +    if (!meta.raw) {
    2578 +        appendTransferBadge(wrap, transferInfo);
    2579 +        if (!transferInfo) appendSourceDash();
    2580 +        return;
    2581 +    }```

...

?

[spiral gorge]

tbh the 100usd sub doesnt has enough tokens for me to be honest.
vibe coder.
if that would look different if they would combine all chatgpt limits with 5.3 spark tokens into one budget-limit?

[orchid echo]

why is this server so soft

lmao yeah

[inland sonnet]

w t f

[orchid echo]

w t f

[inland sonnet]

f a c k m e n

v e r y g o o d s y s t e m

[orchid echo]

t f

lol

cant even type two letters together smfh

[inland sonnet]

no slurs!

no embeds

no links

[orchid echo]

GAY

[inland sonnet]

lgbtq

wait

[orchid echo]

lg tv