codex-discussions

[chrome raven]

because compaction is perfect enough to remember what it's doing

[frosty zealot]

If you think going from planning phase with a massive amount of thought is not going to cause drift and context rot then I don't see any point in continuing the conversation

[chrome raven]

if you have been using codex well enough , you will knwo it handles all this well

[plucky halo]

/New to flush context.
/Compact to compact context for next task
/Model to change model during conversation

[still trellis]

if you are concerned about that save it as an markdown file and then have the model add progress... but honestly there are ways around this, seems like nit picking

[frosty zealot]

I guess I'm just used to the features being there, going from cli's that automate this flow because its more ideal, to going to a cli that you have to exit planning phase run 3 different commands, then go into implementation just feels like im going backwards is all

I guess all I'm wondering is if they're going to start abandoning the cli in favour of the app or if the cli is going to continue to see active development

I can deal with the nuances as long as it has some future in it

[plucky halo]

It's what works best for you workflow. For me, I'd prefer to have the autonomy to choose when that should happen and could see myself getting frustrated should /new or /model be used without authorisation by myself

[boreal holly]

https://platform.openai.com/docs/guides/conversation-state#compaction-advanced

All prior user messages are kept verbatim

So if you really want to make the most of compaction, your user messages need to be extremely descriptive. I make Codex provide a list of remaining tasks, and I copy that into every user message with what to work on next, and it never drifts. Not even a little bit.

[frosty zealot]

All prior user messages are kept verbatim

IS that just for that session till you /new I would assume?

I was considering primarily using 5.3 over 4.6 because 4.6 is absolutely destroying my 20x usage plan

Plus its $100 cheaper

[chrome raven]

you will definitely enjoy 5.3

[frosty zealot]

I've been using it a decent bit but I am just on the fence on when its ideal to switch between M/H/VH thinking levels ive kinda been planning in high then implementing in medium seems to be the suggested flow

and no matter what I do I can't seem to make use of the 'review usage', I've tried using the automated commit review feature, using /review but it always seems to use my weekly/hourly usage instead

[boreal holly]

Yeah, that's how Codex's auto-compaction system does it. If you use it like this:

1. Send really big task list
2. Every new prompt you say "continue"
3. Auto-compact happens, Codex loses track of progress, has to rediscover what needs to be done

The best way to take advantage of compaction is

1. Send really big task list (or collaboratively build a plan)
2. Repeat that plan back to Codex, say what to do next
3. Check off items that are done each response, sending task list with [X] checks for completion
4. Auto-compaction happens, Codex knows exactly what task it's working on and what's been done

this has worked extremely well for me 🙂 and the remote compaction docs I shared, idk if everyone knows about it but it's been an eye opener for me so I figure others would appreciate knowing exactly how it works and how to take advantage of it

[plucky halo]

4.6 is the same amount of tokens, i think. I use both Codex on Pro subscription and Claude on 20x and haven't noticed an increase in token usage

[frosty zealot]

I'm already at 20% of my weekly and I only reset yesterday afternoon, I've never managed to do that before lol I usually aim around 10% a day, and I've been using Codex quite a bit to test it out too so I'm not sure why Claude is crawling up so fast

[plucky halo]

If you ever want to fastrack to bankruptcy, try claude 4.6 on fast mode in cursor haha

[frosty zealot]

lmao

Does the cli support hooks I cant seem to find it in the docs

[boreal holly]

No, but if you use codex exec --json or Codex SDK you can implement hooks. The only hook (unless it's been updated recently) that I know of is "notify" which runs at the end of the agent's turn

[frosty zealot]

Ok, I just had some code formatting hooks etc. in cc that would normally run to just do a lint/format was wondering if i could port it over

[boreal holly]

I think you could do a pre-commit hook for formatting. As for linting, that would be useful indeed. Might be able to output lint results to pre-commit.log and make the agent check that file. Idk if you let Codex touch git at all but that'd be the easy path

[simple star]

Wanna bet that after v0.99... the will make v0.100 ? 🤣

[potent mason]

Oh nice which one have you felt has better limits (I’m deciding for this upcoming month which $200 to get)

[plucky halo]

They are both very capable but I trust Codex more with the code. It's slower but seems more thorough/requires steering a little less. My workflow tends to be creating the plan with Claude, executing with Codex, and then evaluating with Claude again. That has worked very well for me

[potent mason]

No yeah Ive used both and definetly prefer 5.3 (although sometimes it doesn’t understand what I want as well as Opus), I’m talking more about how fast they consume the limits

Im worried about the $200 plan in Codex saying its x6

[plucky halo]

Ah, sorry. I'm a heavy user - I use all day everyday, including for work, and I've only ever ran out of my weekly allowance once

[ornate jolt]

also using the app its like a very well streamlined experience make branch make pr @ codex in github have it review and merge

[main nimbus]

handy to know, thank you for sharing. I wonder how insane requests must be for the heaviest users. tons of parallel?

[nocturne folio]

how can i use more of my codex plus usage

its impossible with the 2x limits

[plucky halo]

I think so - when I hit the limit I was using around 10 instances in parallel

[main nimbus]

Insane. How do you manage to keep productivity rolling after so so many changes? just keep prompting till the result looks sound?

[ornate jolt]

i usually review against docs if its using right methods and codex is very good at gathering stuff online and comparing against code bases

[plucky halo]

It was because I had an eye watering number of errors and warnings when I ran airbnb lint check. I divided each by type and had a terminal running until it fixed them all.

[potent mason]

Use TDD

Create the tests first, develop what you want, check that it passes the tests

• type + lint checks

It’ll get it right imo a higher % of the time

[plucky halo]

That's what's in my agents.md EDIT: Learnt from my mistakes

[warped pine]

pro account is enough for you to run multiple parallel codex sessions every day fulltime without hitting limits.

[main nimbus]

usually not worried about backend. mostly always a visual problem with frontend design.

[ornate jolt]

2 months of 2 x is insane

[main nimbus]

so I can easily expect codex to pump out APIs all day, usually in a single ask. I find myself following up a lot with design inquiries, though. This is more likely a prompting issue, but everyone will compare that Claude family of models does vague UI requests way better.

[ornate jolt]

also codex 5.3 is pretty fast and i can work on 2-3 projects without stopping for a considerable amount of time

depending on the task

[potent mason]

You should try out the front end design skill

(I use opus 4.6 for design but still use the front end design skill)

[potent mason]

Btw does anyone know how to make Codex auto review my pull requests? I'm thinking of just doing a github action that tags it

[main nimbus]

I only see the figma ones in codex app. unless it's a custom one

[severe pagoda]

I have a very strange issue. I use Superwhisper to do dictation. I'm running Codex in a shell. As I am dictating, when I hit the key to transcribe, which would then place the dictated content onto my pasteboard so I can paste it into the shell, but it automatically pastes. Somehow, Codex is reading what is on the pasteboard and answering, and I am ending up pasting Codex's answer into the shell. I have interrogated Codex and asked it how it is accessing my pasteboard. It insists that it cannot access the operating system's pasteboard, but I see no other way possible that the answer to the question that is being transcribed by my transcription software is ending up on the pasteboard of the operating system, other than Codex reading it, and then pasting to it. This makes me incredibly suspicious of this tool. I will post a video demonstrating the behavior.

[potent mason]

It's a custom one here let me find it

https://github.com/anthropics/claude-code/blob/main/plugins/frontend-design/skills/frontend-design/SKILL.md

[main nimbus]

thank you

[vast shore]

I had my Pulse suggest a fix for an issue on a ln open source codebase I help out with sometimes. Would be amazing if there was a one tap to load Codex Web with a job prefilled for the relevant codebase

[white spire]

Hi
Do we have any news about Codex app for Windows?
It seems like it was promised as soon
Did I miss somethings?

[fallow chasm]

Hey all — I got tired of cleaning up old Codex app history sessions manually, so I built a small visual tool for managing archived sessions.
Repo: https://github.com/cola-runner/codex-history-manager
Would love feedback if you try it.

[potent mason]

I'm actually surprised people are now developing for Mac / IOS first

It used to be the other way around

[plucky halo]

It absolutely can.
Copy a secret and tell codex that you have it copied and it'll be able to use it

[severe pagoda]

So, it's lying about its capabilities?

[main nimbus]

it's an interesting position for AI companies to be in if you ask me. if AI makes engineers more productive, couldn't they have shipped multiplatform with ease?

[frosty zealot]

Anybody wanna tell me what the work flow is to get Codex to actually use your Code review usage instead of your base?

[potent mason]

Depends

If they're developing platform native then no

Which for good software you should probably develop platform native

[plucky halo]

I'd say the answer isn't entirely truthful... Hang on let me find when I did it

[frosty zealot]

I would say from most people I've ran into, more use Mac for dev than distros/windows

[main nimbus]

I got annoyed early on with claude code not being supported natively on Windows, so I just hopped on board the MacOS train.

[frosty zealot]

@plucky halo Having both 4.6 and 5.3 which would you say has the edge right now from personal experience?

[potent mason]

I use both

There are some things that you just can't run on mac

[severe pagoda]

Having the AI lie to me over and over and over again is unenthical. And what's weird is that when I start a session it doesn't have this behaviour but as the session gets longer all of a sudden it's reading the question on the pasteboard and replacing it with the answer. It makes the tool unusable because my side of the conversation isn't present in the UI.

[potent mason]

But yeah general development experience currently is better in Mac than Windows

[frosty zealot]

Like what? Just curious I use a M3 Max & a Mac mini so I wouldn't know

[main nimbus]

any program not compatible with arm

[frosty zealot]

Just wondering what those edge cases are that there wouldn't be a arm version

[potent mason]

For example I need an ODBC connection to Netsuite to programatically access it

Netsuite doesn't support Mac through ODBC

[plucky halo]

@severe pagoda

[main nimbus]

I haven't run into any, but old school stuff probably. I work in a 100% Microsoft shop and still effectively use every program plus more on MacOS.

[severe pagoda]

Has anyone else had this behaviour that when using dictation that Codex is reading the results of the dictation before it even gets pasted and replacing the dictation on the pasteboard with the answer?

[frosty zealot]

For the love of god somebody tell me how to utilize Code review lmao

[potent mason]

Go to codex online

Enable code review

[frosty zealot]

I have

[potent mason]

It will review all of your pr's

[severe pagoda]

Honestly, the rapidity and speed with which it's creating the answer is freaking me out. I'd really love to use this but it's basically unusable for me because I can't see my part of the conversation and it's all gibberish.

[plucky halo]

Add to your agents.md:
Don't use pbpaste without approval

[potent mason]

Are your repo's showing in repo preferences?

Are you doing seperate branches + pull requests or just worktrees / main branch?

[frosty zealot]

Yeah

Work trees and main branch mostly

with the odd PR

But im just a solo dev so i dont typically bother with the PR route cause I know what I just did

[potent mason]

Yeah it only reviews PR's though

[frosty zealot]

Ok maybe thats what I'm doing wrong, I assumed /review in the cli would trigger it as well

[potent mason]

I think it reviews it locally if I'm not mistaken

[frosty zealot]

Gotcha so that uses my normal limits

So whats the opinion on 5.3 vs 4.6?

[potent mason]

UI/UX 4.6
Anything else I'm liking 5.3 more

[frosty zealot]

Interesting, prior to 5.3 over in CC discord they'd say gemini for ui/ux 4.5 for anything else

[potent mason]

Use the front end design skill with cc otherwise it sucks but for some reason adding that skill takes it from a 2/10 to like an 8/10 (haven't used Gemini though I have had terrible experiences in the past)

[frosty zealot]

gotcha

Ok submitting my first Codex PR lol, alt+t to toggle between different thinking levels on the fly rather than /model > model > level

its driving me crazy

[main nimbus]

I have never liked gemini - only used in cursor, maybe it's a harness thing, but no. so much bad code and results. it once gave me a completely black and white website using cursor.

in antigravity with 3.0, it generated and iterated well on a threejs game. just didn't work for long periods

[potent mason]

Yeah I hated Gemini when I used it

It's unusable

I also hate that they watermark their image generations

[frosty zealot]

Its been greatly improved in the last month or so, they've put in a ton of work specially if you checkout @preview

[potent mason]

Like I'm trying to generate an icon for a startup not trying to post fake news...

[frosty zealot]

lmao

[potent mason]

Anyways if you want an accurate review of gemini don't listen to me I hate on it so bad

Lmfao actually the other day a company gave me these terrible images (the quality was super low) I thought on using Gemini pro to increase the quality and only after like 6-7 images did I realize it was watermarking them and I wouldn't be able to use them

[main nimbus]

since o4-mini and o3, can't put down openai models. it was insane to first witness CoT tool calls and get a small code change for a bug fix where most AI would be quick to write code before gathering appropriate context.

[potent mason]

For me 5.2 was the one that changed me

I think I went something like GPT 3.5 (on phind) -> Sonnet 3 (on vscode) -> Sonnet 3.5 (cursor) -> Sonnet 4 (cursor) -> Sonnet 4.5 (cursor) -> Opus 4.5 (claude code) -> 5.2 codex (codex) -> 5.3 codex (codex app)

[main nimbus]

interesting checkpoint to be sold at. what changed with 5.2?

specifically codex? if so, I get that.

[potent mason]

More anatomy than Anthropic (went and did 100 things without needing supervision)

[main nimbus]

I preferred 5.1-Codex-Max to GPT-5.2 until they released 5.2-Codex

[potent mason]

5.2 codex didn't fully onboard me though in all fairness

Just because it took like 30 minutes on the smallest changes

It was 5.3 codex + app that fully got me hooked

[main nimbus]

5.3 is so much faster, that's for sure.

btw, this frontend skill has been TASKING since it was installed. We're 15 minutes in on a run currently, where it was completing much quicker last time.

[potent mason]

Its crazy how five months ago I was just paying a $20 sub, then I went to a $60 sub, then $200, now I'm at like $320

AI has gotten so good it's insane

[potent mason]

Damn that sucks I don't really use the front end skill on codex (I still use Opus 4.6 for my frontend needs)

[main nimbus]

I was curious to try. It has actually not done terrible with UI since GPT-5. It requires more prompting and iterations, but I still get a better result in the end.

[potent mason]

Yeah I mean all models do better with a design system so I make it a point to always setup the design system when working on something new

[ornate jolt]

i use gemini still for frontend then i use codex for touching up its awful accessibility patterns

[severe pagoda]

How do you clear the context other than quiting and restarting. Basically how do I do /clear like in Claude Code?

[plucky halo]

/new

[late root]

gotta love codex 5.3 bro... https://i.imgur.com/h6njKyr.png

its stuck for 1 hour searching forever

5.2 doesn't have this issue

[sonic dune]

For folks who sub to plus and use codex for coding, how much work can you actually get done before hitting a weekly cap? Is it really restrictive like claude pro?

[velvet wren]

it's not as restrictive as Claude Code

[plucky halo]

Are you using the new codex app?

[sonic dune]

Do you have a rough guesstimate in comparison of usage limits?

[velvet wren]

I would like to say double, it seems like a lot

[sonic dune]

Thanks, gonna sub and give it a go then!

[boreal holly]

I also make it use pbcopy and pbpaste! So handy! I also have a script called pbclean that removes ANSI escape sequences which eat up a ton of tokens:

#!/usr/bin/env bash

# Get clipboard contents, strip ANSI escape codes, and copy back.
pbpaste \
  | sed -E 's/\x1b\[[0-9;]*[A-Za-z]//g' \
  | pbcopy

Then tell it to read the pasteboard

[twin kestrel]

codex team, please implement auto credit load option for codex when using with chatgpt subscription

[plucky halo]

Nice! Stolen - thank you! ❤️

[ivory zodiac]

oops 🙃

[ornate eagle]

codex team, please try and ship a build that doesnt crash in less than 20 s omd 😭

[ivory zodiac]

meanwhile 5.3 got cheaper and faster lol

[main nimbus]

I have experienced zero crashes on the new Codex app.

[upper dock]

whats the limit of the free tier of codex on vs code as an extenstion

[plucky halo]

I've had it hang on larger tasks and when using subagents - it's still working in the background though

[ornate eagle]

no matter what i do it crashes, ive fully reinstalled, signed out etc, went down to a system level to remove anything remotely to do with it, nothing, it still crashes between 20s and 2 minutes after its opened??

[boreal holly]

Did you do npm i -g @openai/codex@latest?

[ornate eagle]

yep, still crashed

[boreal holly]

run it with /Applications/Codex.app/Contents/MacOS/Codex and when it crashes share the stack trace

[ornate eagle]

Ran Codex from terminal. It’s crashing with a V8 JS heap OOM (~3.9GB) after Skills/list + git repo watcher. Looks like a memory leak / runaway workspace scan, not install corruption. Full reinstall didn’t help.

[boreal holly]

Ah ha, same issue as this one https://github.com/openai/codex/issues/11032

If you add your log to that issue, I also noticed some duplicate issues where people are saying "no crash report", the workaround to get info is to run via terminal so you get a stack trace.

[ornate eagle]

Good spot, will do

[chrome raven]

crazy work by 5.3 damnn

[stray swift]

I don't know whether it's smart to give an AI access to your clipboard, and if you're using X, that's really insecure.

[frosty zealot]

Earlier I thought you were saying to add DONT use pbpaste/pbcopy lol.

[plucky halo]

That was me. The context was that I was replying to someone who said they didn't like that ChatGPT had read their clipboard

[frosty zealot]

ohhhh

You replied to the wrong person then lol

[plucky halo]

Balls

[frosty zealot]

Being able to alt+t through thinking levels is so nice, pretty much more more complaints about the cli anymore

[chrome raven]

8 hours and it did what i asked it after loads of compactions

[frosty zealot]

Do you use the experimental sub agents?

[chrome raven]

i cant enable it over here, i dont know why

[frosty zealot]

Are you on the 0.98.0 and type /experimental?

[chrome raven]

yes i am, does it work on yours? and can you give me the config file flag for the sub agents

[frosty zealot]

or maybe

[chrome raven]

so i enable manually

[frosty zealot]

It doesnt seem to add anything to the config.toml

not sure which config file it gets written to, but its not in ~/.codex or .codex

I just did /experimental in the cli and checked it off

[chrome raven]

that's new, i have tried all possible flags, i wasn't able to enable it here

[chrome raven]

nicee, mine doesnt check off even after restarting the terminal

[frosty zealot]

Interesting, I'm on an M3, not sure what you're using, maybe its not ready across different OS's

[chrome raven]

i am on m5

or maybe i will restart the whole stuff and try again

[late root]

No, just VS CODE

[boreal holly]

I think the setting is called collaboration or something. shell snapshot is like unified_exec, and sub agents is collaboration

Sorry collab

[frosty zealot]

@chrome raven

in the global

[boreal holly]

Oh I see, shell snapshot is its own feature

[chrome raven]

oh thaaaanks

let me enable mine

[boreal holly]

Typical usage is "I need to start a new session. Use cat <<EOF EOF | pbcopy to write a prompt for the new agent with enough information to pick up where you're leaving off." And if by X you mean X11 then no, I use Quartz compositor

[chrome raven]

I think steer is one of the biggest unlock with 5.3 codex

Works like magic

[frosty zealot]

When you /review, and just target diff -> main, does it consider the 'plan' at all?

or is it just looking at how its going to integrate and any sort of blast radius

[dense heron]

someone know if codex have more perfomance in vs code than in codex web?

[sleek spoke]

is the Super Bowl ad going to be more Codex?

[modern hedge]

seems like it

[main nimbus]

ai superbowl ads. what a time

[tribal lake]

Had my agent create a new skill for something. Got spammed with "⚠ Skipped loading 1 skill(s) due to invalid SKILL.md files.", flooding my whole terminal at one point until the agent finished with the skill 😂

[noble kernel]

I didn’t catch whatever was on the code screen because I didn’t know that that’s what they were advertising.

[strong jungle]

Nice to see codex In the commercial ha

[latent tiger]

I didn’t catch the Easter egg lol

[eager shell]

Codex merch drops?

They just posted a screenshot of the frame where it shows the codex merch drops

[pulsar blade]

chatgpt laptop?? codex merch??

[strong jungle]

I want merch 😩

[eager shell]

I dont see chatgpt laptop anywhere

[random pumice]

I want a merchant drop. Create a hidden skill?

[pulsar blade]

laptop with openai logo

[eager shell]

This is all I caught

[stray swift]

Per #codex-updates , it's a laptop screen displaying the codex app.

[strong jungle]

How do I get it ha

[eager shell]

That looks like a sticker so they aren't advertising for mac for free

[pulsar blade]

oh yea probably

[random pumice]

It said to create a hidden skill

[mint widget]

I think they are making merch

[scenic umbra]

It's just a limited Codex merch drop, I thought it would be a code for free Codex or something

[latent tiger]

Lmao Anthropic commercial

[noble kernel]

OK, I was looking in the wrong spot in the ad; I was looking at the retro computer.

[mint widget]

I looked at the codex screen first and couldnt find anything so i went back to the dial up computer and still didnt find anything so i tried again and saw CodexMerchDrop after removing my PRESCRIPTION glasses. I need new lenses bad

[glacial agate]

ahhhhhhh windows is shut out 😭

[shell river]

yeah that sucks, windows users pwnd

[heady mauve]

any raeson why i cannot see 5.3 or is it only availble in the clji

[glacial agate]

nvm just ide

[obtuse saddle]

codex merch drops is coming ?

[eternal quartz]

But only for apple users ._.

[kindred stratus]

you have codex
extract it from the apple version ;d

[worn otter]

let me be lucky

[junior olive]

Same

[glacial agate]

glhf brothers 🙏 🎮

[scenic umbra]

🙏

I didn't realised it was free merch lol

[glacial agate]

were you supposed to use the same email as your openai acct

i may have used the business one instinctively ...

lol

[obtuse saddle]

is it free ?

[glacial agate]

you can do it in a ide like intellij

[worn otter]

i don't know, i used a different one 😅
But i don't think you need the same one. I mean otherwise it wouldn't have asked you for an email and could have just instantly used your openai email automatically

no, only in codex

it then generated a unique link containing the email you entered

[glacial agate]

welp

[obtuse saddle]

😭 i dont have mac

[glacial agate]

tbd then lol

[worn otter]

i am on windows

[glacial agate]

use a ide like intellij!

[worn otter]

codex-cli can be used on any system 😄

[obtuse saddle]

ohh

[obtuse saddle]

ohh wait

[glacial agate]

or cli but i havent touched that

[steel mantle]

any of you guys get the email yet?

[worn otter]

i think you will get the email when it ships, not before as a confirmation. It reads like that

[scenic umbra]

nope, I doubt I was one of the first 1k though

[worn otter]

i assume they will probably verify before they ship. (Multi accounts and stuff)

[obtuse saddle]

how to claim it ?

is there any process

[worn otter]

its a hidden skill in codex. ask codex about it.

[glacial agate]

download the skill, run it, give email, open link, fill info

[shell river]

Chatgpt got it 😂

[topaz summit]

How do I get the unlock, I'm on the computer screen with the video

[obtuse saddle]

i tried but it cant find

[worn otter]

ask codex in the terminal about it. like chat with codex

[obtuse saddle]

yeah thats what i am doing

what was ur prompt

@worn otter

[scenic umbra]

It shows up in recommended skills for me now

I wonder if that means all 1k spots are taken

[eternal quartz]

Is there also a price for this? :D

[glacial agate]

whats this?

[eternal quartz]

My codex token usage in the past 30 days

I like graphs and stuffs

[glacial agate]

oo is that in the new mac app?

[eternal quartz]

No thats my thingie

[glacial agate]

looks neat

[eternal quartz]

<3

[scenic umbra]

Why does it say prices I thought Codex was just per plan?

[eternal quartz]

prices?

[scenic umbra]

Yeah like the cost per token

[eternal quartz]

Ah its the cost that i would have paid

[scenic umbra]

ah ok

[eternal quartz]

Yea... worth it

[scenic umbra]

Does Codex 5.3 even have an API? How do you know the cost?

[eternal quartz]

litellm has a chart

[scenic umbra]

Ah I see

[eternal quartz]

And i dont use it too much so ¯_(ツ)_/¯

[hard tulip]

did the emails get sent out does anyone know

[high girder]

Was the skill specific to the Mac Codex app? I'm trying it in my windows port, but I didn't see it

[scenic umbra]

idk I don't have one

[hard tulip]

how long did it take you to submit it

[scenic umbra]

I did it like 20-25 minutes after it aired

[nimble terrace]

Is it true that .agents works universally?

oh yeah it is true

cool

[scenic umbra]

https://x.com/OpenAIDevs/status/2020659282514206849

They said "that was fast" to someone who posted this after I already submitted mine 👀

[ivory zodiac]

you on mac?

[scenic umbra]

Yes

[ivory zodiac]

if they launched it for mac only that would so absolutely silly

i just burnt a bunch of tokens on this

[scenic umbra]

There was a way you could do it on any platform that can run python and has a web browser though

[frosty zealot]

Did you get the repo specific .codex or did you just port your global ~/.codex

[scenic umbra]

idk if that would count or not though

[high girder]

It's the codex bundled with the dmg file they dropped. The only thing i've done is make it run on windows

[ivory zodiac]

i dont htink its mac only

[cyan wing]

you da real MVP. hoping we get the merch drop 🤞 🙏

[valid needle]

how do you make it run on windows

[ivory zodiac]

they wouldn't do that

[frosty zealot]

Gotcha, not sure, but it should just be sourcing any skills from ~/.codex and .codex (if you made a skill in a project)

Not sure what the equivalent directory is on windows in your port

[high girder]

It's electron with a mac wrapper. If you use the dmg file and then break it down to get the right versions of the apps they used for dev, then you can stub mac features, and port the others

[ivory zodiac]

😉

[scenic umbra]

https://x.com/aaronp613/status/2020652862062371062

lmao

macOS app running on Windows laptop

[ivory zodiac]

lol

thats funny

i dont know what mac kbs look like si wouldn't have noticed

could have fooled me

[scenic umbra]

The traffic light icons give it away

[ivory zodiac]

Oh i know that, but how do we know thats a windows laptop?

[scenic umbra]

Macs have notches

[undone lava]

where do i find the codex global config.toml on win11? im going crashout lmao

[ivory zodiac]

ah.

[ivory zodiac]

they did say you can just build things 💀

checks out

[high girder]

Okay. It wasn't codex app specific. I just make crap harder for myself for no reason

[scenic umbra]

Yeah lol

[ivory zodiac]

whats messed up is i used codex to scrape the github

and it didnt find it 💀

[nocturne folio]

someone used codex to run the dmg file on linux

[ivory zodiac]

wouldn't work for me

[nocturne folio]

maybe u can run such through wsl

[ivory zodiac]

truied that

someone did port to windows

its doable

i'll just wait

[nocturne folio]

alr

i do like that quote tbh

[ivory zodiac]

its quality

[stone aspen]

I did it on windows

approx 20mins ago haha

Quality times

[ivory zodiac]

we're prob cooked

1hr passed already

[stone aspen]

not much noise on twitter, reddit etc about it. fairly sure if u did it before like right now u got them sweet merches too :)

[loud dragon]

got mine too, doubt i'm in the first 1000

works in cli just fine, just tell it to install the experimental build-things skill, then invoke it

[stone aspen]

some optimism is always good, lets hope for them sweet goodies

[rich ridge]

How do I do the add

I’m at the video

Screen

[scenic umbra]

It’s possible a bunch of people did it pretty early but just didn’t talk about it

[stone aspen]

i think anyone before 02:30 is in the good, considering the small userbase, small active amount of peeps etc

[rich ridge]

How u do it

[stone aspen]

install skill, run skill and then fill in form

[mighty plaza]

Yeah the fact that you have to already have codex installed probably reduces it quite a bit

[rich ridge]

Ahh ok app ?

[stone aspen]

tbh its the first time for me using Codex, am a claude user myself but have chatgpt pro so might aswell try it out

codex seems to be faster and more direct atleast

[rich ridge]

How to install skill

[mighty plaza]

i used the skill installer

/install and it popped up

[rich ridge]

Hanks

[rich ridge]

What do I install

Game so many options

[mighty plaza]

it only gave me 2

[rich ridge]

Nvm

[waxen escarp]

really doubt i got the merch but worth a try 🙏

[rich ridge]

U gotta press tab

Skills

[waxen escarp]

was out and had to wait to get home lol

[sonic dune]

Is it possible to use the codex free trial in opencode rather than antigravity?

[fallow chasm]

Hey bro, I built a small visual tool for managing archived sessions.
Repo: https://github.com/cola-runner/codex-history-manager

[fallen kettle]

Super cool! There's already been two times I've had Codex search for a session by searching on disk. Love that you can search directly through your UI.

[potent mason]

Does Codex app automations not have internet access? (I have my codex app setup with full access)

I tried to run an automation that just does pnpm audit and brings anything up to date that is required

but it failed because it couldn't access the npm registry

[frosty zealot]

I think its a sandbox restriction

a lot of stuff fails in the sandbox

[potent mason]

Yeah that makes sense but it kinda makes it useless (I feel like most automations will need internet access)

[frosty zealot]

yeah I always launch the cli with codex --dangerously-bypass-approvals-and-sandbox

I cbf to sit there and hit enter, and it runs so many unique bash commands like hell im approving every one

Will it be the end of me? Maybe

[shell river]

you can just do codex --yolo, much faster

[frosty zealot]

oh I didnt see that in the --help

but thanks

[potent mason]

I'm pretty sure this doesn't solve my problem though because automations are not through the cli

[frosty zealot]

I was just gonna make an alias like alias oops='codex --dangerously-bypass-approvals-and-sandbox'

[frosty zealot]

Yeah I was just looking myself out of curiosity, I don't see any way to adjust sandbox settings

Did ytou try setting this?

[shell river]

yeah I had to do that for copilot because its copilot --allow-all-tools --allow-all-path 🤦

[frosty zealot]

Usage: codex app [OPTIONS] [PATH]

Arguments:
  [PATH]
          Workspace path to open in Codex Desktop

          [default: .]

Options:
  -c, --config <key=value>
          Override a configuration value that would otherwise be loaded from `~/.codex/config.toml`. Use a dotted path (`foo.bar.baz`) to override nested values.
          The `value` portion is parsed as TOML. If it fails to parse as TOML, the raw string is used as a literal.

          Examples: - `-c model="o3"` - `-c 'sandbox_permissions=["disk-full-read-access"]'` - `-c shell_environment_policy.inherit=all`

      --download-url <DOWNLOAD_URL>
          Override the macOS DMG download URL (advanced)

          [default: https://persistent.oaistatic.com/codex-app-prod/Codex.dmg]

      --enable <FEATURE>
          Enable a feature (repeatable). Equivalent to `-c features.<name>=true`

      --disable <FEATURE>
          Disable a feature (repeatable). Equivalent to `-c features.<name>=false`

  -h, --help
          Print help (see a summary with '-h')

Looks like you can launch it with custom args

-c 'sandbox_permissions=["disk-full-read-access"]'

I'm not sure what the other flags are for sandbox levels but im sure its in the docs somewhere, set that and you should be good

or configure the sandbox for that project somehow, again I've never looked into how cause i just --yolo, and I personally prefer the cli so I can ssh in to check it when im not at my comp

[potent mason]

Yeap

It works after I just tell it to proceed in the chat

[frosty zealot]

lol typical

[potent mason]

But the problem is that automations seem to run with other permissions or something

[potent mason]

Oh nice

Thanks

[frosty zealot]

np let me know if it works im curious

[tired lynx]

https://github.com/robcholz/vibebox

i built a small macOS sandbox because i got tired of babysitting permissions when running coding agents locally.

[cyan wing]

anyone get the merch email yet? 🥲

[stray swift]

Quit spamming crypto scams.

[stray swift]

If you don't use the codex sandbox, you should really sandbox it yourself, and then you can use whichever CLI flag in the sandbox kickoff script.

An AI with internet access is subject to prompt-injection attacks. It's a mistake to assume that because you trust OpenAI, you trust codex unrestricted.

[frosty zealot]

I don't disagree, got a good starting point?

[stray swift]

Not for using the app. I use this: https://github.com/coventry/sandbox-codex

[frosty zealot]

Can you not see the thought in the codex app?

[potent mason]

Pretty sure modifying the base permissions works

I’ll have to try it some other time though

[boreal sluice]

That's pretty interesting, have you tried ralph in that?

[supple perch]

I have one called the agents runner that @boreal holly says works on mac if you want to use a full docker container sandbox so each agent can have its own tooling / sandboxing. (am working on the rebuild branch right now if you want a link)

[supple perch]

...

[still trellis]

Crypto bros always gotta spoil the situation 🫩

[gloomy veldt]

Codex app for Linux?

[frosty zealot]

So can you move a Codex App session to cli mid stream and vice versa or do you have to wait till it’s done

[still trellis]

Yes, have to wait

[still trellis]

Its merch I reckon

[cedar skiff]

wow, codex is genuinely better than cc now

Been giving it a fair shot and it's not even a close race

[chrome raven]

Its very good

Since building is not the bottleneck now who is building a platform for people to showcase built products??

[open finch]

It is merch it’s in the prompt box it says

[plucky halo]

Just woke up - am I too late?

[main copper]

Or just Codex merch. Says so on this input field

[white spire]

How can you teach Codex to write technical documentation as clearly as Opus?

Opus is simply excellent for this: readable text, code snippets with examples, clear diagrams.
Codex solves tasks better, but the documentation it produces is much worse.

Are there any Codex skills for this? Any good practices?

[high girder]

Define the output you expect it to visually look like and turn it into a skill

[brisk junco]

But does it work with codex desktop?

[plucky halo]

Yep you get the link and if you're one of the first 1000 you get a hoody, cap, keycap, keycap keyring, openai pin, and a sixth thing that I can't remember

[fallen kettle]

sticker set?

[plucky halo]

That's the one!

[still trellis]

man Id love that. Holding thumbs!

[scenic umbra]

Has anyone gotten an email yet?

[brisk junco]

Did anyone ever had this issue when loggin in to codex desktop?

[spiral gorge]

so...... too late for me german person to do this merch?

i use vsc cause windows

[main copper]

Yeah way too late. I checked myself

[plucky halo]

Did it tell you it was too late when you tried?

[main copper]

Pretty much that. "Thanks for you interest in Codex..."

[brisk junco]

Does nobod know a fix? :/

[still trellis]

Did you try sign out and sign back in?

[elfin iron]

Does anyone else run into these issues often?

this wasn't the case first 2 days. it started happening when i first updated the app, then it's frequency increased. i'm running into these at least once every 2 hours.

[radiant island]

Did you hit daily limit?

check current status

[elfin iron]

not at all.

this is on new codex app

[elfin iron]

this another one

[radiant island]

It's better than CLI ?

[elfin iron]

it is smooth for parallelism, worktree management is baked into threads/sessions.

[torn sequoia]

hey guys, did anyone apply for the codex ambasador and get a reply back ?

[radiant island]

I should try it

[steel cove]

does anyone know exactly when the codex windows app is suposed to come out?

[grizzled swift]

Does anyone else seem to have issues with compacting history in the Codex app? It was happened on one thread only so far:

Error running remote compact task: { "error": { "message": "Your input exceeds the context window of this model. Please adjust your input and try again.", "type": "invalid_request_error", "param": "input", "code": "context_length_exceeded" } }

Other than this over the last couple of days it seems to get stuck a lot either endlessly "Exploring" or "Thinking".

Is anyone else experiencing similar issues?

[elfin iron]

i get this when i dump(copy paste) a huge log file into the chat, never really got this issue outside of it.

did you by any chance do the same?

[brisk junco]

I fixed the bug, as this was due to Antigravty being open (while running codex extension)

[civic fulcrum]

Guys be carefull cause Codex 5.3 extra high just deleted EVERYTHING from my C drive, all of my files and programs, only windows remained, not even task manager being anymore. With full acces its very dangerous, I've lost projects from over 2-3 years and many important files while it was fixing something for a website ... I have proof of everything, right now my SSD is at a data recovery center with hopes I can recover some of the important files.... Please don't make my mistake and isolate codex if using it (I've used it from Visual Studio Code). Right now I really don't know what to do

[brisk junco]

Can you open Projects of Codex in a remote server in that desktop app?

[still trellis]

Well done!

[grizzled swift]

No although it was a fairly hefty thread and batch of work Codex had been working through prior to this. I'm wondering if it's all related somehow - getting a lot of freezing issues happening with Codex app on a regular basis now. It struggles to do even the smallest of requests.

[plucky halo]

I mean, you've got to be extra high to allow it full, unrestricted access to your machine without authorising changes

What did you prompt it?

[nimble terrace]

i'm in a tough spot, codex gave me a plan..but i see the usage count at the bottom

limit resets in 5 hours

gonna keep the app open in the background 🗣️

[grizzled swift]

I know... i'll set to xhigh, give full access and will set Codex to work on trying to diagnose the issue. Will see if this works and will report back.

[civic fulcrum]

I had been using it like that for months and never had issues

[civic fulcrum]

To make some improvements and fixes to the website, I had everything detaliated. Nothing about deleting files, that s clear

[steady vigil]

weekly: 9% left (resets 13:59)

the 2x pro plan should be the normal. there is no way I'll be able to go back to 1x

[elfin iron]

yeah, codex app experience has gone for a toss after the first couple of days.

[elfin iron]

agree, in more or less same place now. 12% left with 3 hours to reset.

[steady vigil]

codex app just froze... got sluggish and had to restart it. stability issues after operating for a while

[boreal holly]

Just a hunch, but I think the Codex app performance issues, besides super long sessions, could be related to git integration. Do you have a lot of worktrees or really huge diffs when the performance degrades?

[steady vigil]

nothing unusual. maybe working on up to 5 projects at a time. mem use was getting a bit high 1.4GB on the highest pid with almost that much in the smallers aggregated

[boreal holly]

Do you use Cloud at all or just local?

[steady vigil]

local only

after restart that 1.5GB -> 350MB so prob some resource leaks over time

I did /feedback

[boreal holly]

In the current conversation I have open, the renderer is at 700MB, but if I load another conversation and wait for the garbage collector it drops down to ~400MB, so maybe the trick to get good perf back without restarting is load up an empty convo and wait a few minutes

[steady vigil]

ha worth a try I will test that after some hours

[loud dragon]

anyone gotten a confirmation email on the merch drop thing yet?

[frosty zealot]

Do you guys use steer = true in your config? HAve you noticed a big difference?

[gloomy veldt]

How fix that?

[torpid venture]

when codex app will be fixed? high cpu/gpu usage, unsable for more than 15 mins

[spiral gorge]

can codex do that tho?

NO it CANOT

[supple perch]

You are so funny... start a new chat and then give it a full prompt of rules, look up the prompting folder on the docs

[violet meteor]

It's literally unusable right now

Not sure why it's not escalated on their status page.

Version 260208.1016 (571) is not stable

It fails every couple of prompts.

[torpid venture]

yep, really sad their campaign is lost because the product is poop

[violet meteor]

I really enjoyed using the codex app later last week but a recent update has smoked it.

https://discord.com/channels/974519864045756446/1470458795875504202 - just added support bug

Hope they resolve it

[cyan wing]

how do I remove this warning??

[gloomy veldt]

Nah, this look like dumping.

[steady vigil]

didnt help already back to 1.2GB and performance sluggish. this is a bad release of codex app

[violet meteor]

100% agreed later last week no issues with same amount of threads and active in play.

Version 260208.1016 (571) this is a dumpster.

[grizzled swift]

Yeah. Getting SO many issues with performance.

Last week I was flying, this week i'm just hitting wall after wall

[spiral gorge]

i am sure your codex does not playback sounds when you start a prompt and it sucesfuley finishes something.

claude code does 😄 its hilarious nerdy

[dawn seal]

codex still cannot code on its own

still meh

it can write a short function though

[white spire]

Hi
Do we have any news about Codex app on windows or WSL?

[teal cargo]

very soon

we already have a working build - just need to smooth all the rough edges

[ornate jolt]

openai gives us 2 times usage for 2 months claude gives you 50 dollars think i know whos side im on

[lean lark]

! Note from the Team?! Good to see ya, VB. 🙂

[cyan wing]

how about the merch drop? were emails sent out?

[plucky halo]

I've not received one yet 🤞

[brisk junco]

Yo guys can you open Projects of Codex in a remote server in that desktop app? So basically run projects in remote?

[last stream]

It's almost here, Windows friends 😁

[simple star]

<Crying in linux>

[last stream]

That's why I would never leave Windows as much as I love Linux. My WSL does what I want for Linux without losing the consumer experience and market reach Windows has

Life in Linux is way more difficult than it has to be despite being a better OS

[simple star]

Life in any OS is way more difficult that it has to be 😛

[last stream]

https://tenor.com/view/thats-true-kramer-seinfeld-gif-6564373007119861775

[simple star]

No GIF preview, sadly

[last stream]

Lmao just realized. No GIF support in discord is crazy work

[simple star]

I asked the admins... they said "no sorry kid, not in here" 😄

[high girder]

Imagine what these chats would devolve into when you can open carry memes

[lean lark]

#status message
And I was gonna ask if anyone else is seeing Server Error 500 from GitHub. 🙂

[simple star]

@high girder I'd rather have slowmode, and allow for GIFs, haha

[dire cobalt]

hi

[lean lark]

Welcome, favz.

[hushed storm]

The codex app works fine on Linux

[lean lark]

the Codex app has not been published for Linux.

[hushed storm]

Just repackage the dmg

lol

It’s just a electron app

[winged ore]

codex cli

[lean lark]

SMH - that is SO naive. BTW, just look at the wailing about issues. I won't load it until peeps are happy with it on Mac.

[winged ore]

the implementation for vscode works pretty well, tbh

[lean lark]

lugui - he's talking about the app, not CLI, and not the extension.

[summer iris]

Who is running codex on their primary machine with network access?

[tranquil hare]

Wdym “primary”? Company machine vs freelance machine?

[lean lark]

@summer iris I think the guy crying today about his machine getting destroyed was running on his "primary machine" 🤣

[summer iris]

Like your laptop that contains your life. The one with all your personal accounts.

[hushed storm]

what lol

ping where it is

[summer iris]

What happened? Codex someone messed up his machine?

[lean lark]

DO NOT run directly on a system that you care about. DO NOT trust Codex without precautions.
Run In WSL, not in your primary OS.

[tranquil hare]

Well my data are on the cloud (though I own the encryption keys). But other that, me. I'm doing that

However, it's not exactly yolo mode. I have my custom instructions but I still approve stuff relatively often

[hushed storm]

I trust codex, codex is to scared to run stuff anyway

[summer iris]

I want to so bad to make things so much simpler, but I just can't. It reads the documents on one bad website and all of a sudden it wrote a script to haxxor my machine.

[lean lark]

Look, Codex is awesome but it's a tool that needs to be handled with care.
Ya know those instructions that say "wear goggles, gloves, headgear"? Same applies here.
Put Codex in WSL. Add AGENTS.md and explicitly define your file system scope and that the assistant should never reach beyond it. Never run Codex/WSL as root. Lock its access via permissions to a specific sub-folder.

[tranquil hare]

Worst case scenario, I'd have to restore the cloud backup. Ofc I have a copy of the encryption keys on another device (not just on my "primary" machine)

[hushed storm]

--yolo

[lean lark]

If you don't know how to protect your system, as a mature technologist, you should not be using this tool carelessly. It's not just something you install and run.

Tell "YOLO" to the guy who sent in his hard drive to be forensically recovered today.

[hushed storm]

What 😭

[lean lark]

I documented precautions here: #codex-discussions message

( AFK )

[hushed storm]

Book

I wish they integrated the ChatGPT and codex more so I could see and approve stuff on my phone

[lean lark]

To my understanding, that's exactly what the app will do for us.

[hushed storm]

what app

[potent mason]

Has anyone tried codex in jupyter notebooks?

[nimble terrace]

I wonder if ads rolling out means we get larger weekly limits

🗣️

[hushed storm]

Larger codex limits???

[cedar bear]

Guys what was the Thing U are most proud of have done in Codex? 🙂

[nimble terrace]

sorry, i meant like more usage

larger weekly limit as in like, takes longer to finish it

i pray OpenAI doesn't become like Anthropic

I used the 20 plan on Anthropic with Opus, took 10 minutes to get rid of the 5 hour limit. Didn't get anything done. It was just plan mode.

[hushed storm]

I understood what you meant but I don’t think them adding ads will increase the limits, they are already losing a large amount of money, ads are just to help with that

[nimble terrace]

fair, was just a random thought

[hushed storm]

Is it that bad 😭, antigravity gives more usage for free

You can like hook it up to Claude code

With a proxy

[nimble terrace]

ah

[hushed storm]

Can even use multiple free google accounts and the proxy could handle switching

https://github.com/kaitranntt/ccs

[nimble terrace]

interesting

thanks for the shout

[summer iris]

I don't even run skils that aren't from OpenAI or written myself. but yeah, I created a virtual to run it in YOLO mode, but decided against freeing what is likely the best hacker on earth in an env I control.

[boreal holly]

I found the performance issue (not all the ones like V8 Heap OOM but the ones you, I, and most others are experiencing)

As codex uses patch tool, it stores all of the diffs for every file so you can review changes and undo. Those diffs are never getting released from memory.

If you restart the app, memory returns to normal loading the convo (~150mb), and scrolling up you do not see any more "Undo" or "Review" cards. As those cards fill the convo it starts ballooning memory usage. The larger the diffs, the faster it grows. Basically I think if OpenAI makes it so those cards fall off the convo it will not be an issue

[ivory zodiac]

people are definitely still better at hacking, for now 🙂

[summer iris]

I thought 5.1 scored 2nd place in last years top hacking competition.

AtCoder World Tour Finals 2025 a GPT-5 variant of some sort.

[ivory zodiac]

gpt has won coding competitions too but humans are still better

language models are still very very impressive

but if you're an expert in your domain, you start to see the shortcomings.

things are progressing quickly though!

[summer iris]

I mean I'm on my 4th attempt for Codex 5.3-xhigh to get the memorysession agents sdk right, so I'm not currently feeling the AGI.

[cyan wing]

Tip: New Try the Codex App with 2x rate limits until April 2nd. Run 'codex app' or visit https://chatgpt.com/codex

Just to be clear...
Does the 2X rate limit only apply in the Codex App or is it in CLI as well? 😅

[ivory zodiac]

try claude

codex is better

but listen, they're weird.

one LM will just randomly one shot what the other sturggled with

[boreal holly]

Yeah, on chatgpt.com in the codex usage settings the rate limits show up differently for desktop, as in the graph represents desktop app differently from cli, exec, cloud

[ivory zodiac]

i had an issue i burned almost all my 5h usage on. and opus one shot it in like 12 seconds 😭

its usually the other way around but thats just how it goes

[steady vigil]

I believe only codex app is actually 2x, I may be wrong but based on my limited use on the CLI I saw a spike in increase vs the app.

[summer iris]

That's weird because the App seems to just use the CLI under the hood.

when I'm on battery, I flip back to the CLI since it uses so much less power.

[cyan wing]

Well that's concerning and confusing 😅
I don't even know what's real anymore

[summer iris]

I don't get the big deal over the app since its magic is apparently jsut mulitple instances at the same time, but I just use a bunch of terminals to do the same.

[steady vigil]

again I didnt benchmark or anything, so don't assume I am correct

[ivory zodiac]

app server

[hushed storm]

It’s everywhere

[steady vigil]

also its hard to know for sure b/c there are resource leaks in the current app, so inference sometimes slows down vs fresh start after it being open for many hours...

[cyan wing]

that's what it feels like..
but concerning that there's no way to know for sure

[hushed storm]

They said it was

On twitter

In a clarification

5.3-codex is fast enough to rate limit it self

I hope the increased usage limits once they start enabling sub agents by default

[latent tiger]

is codex down for anyone else?

[hushed storm]

always down lol

[boreal holly]

I feel for OpenAI right now lol, if you go to the codex github, they're getting 3 new issues per hour

Gotta be overwhelming

[hushed storm]

why do you think they made codex app

they have cooler stuff internally

[lean lark]

General comment: If you burn through your Codex credits, you're not using it right.
Don't ask Codex to do everything for you. If you Vibe, you're getting code for free, you need to "pay" with something. You need to pay with your money for credits. If you're not vibing, and you actually know what you're doing, having "paid" already with your time in study of the craft, then you can use Codex all day long, every day - as many of us do - and you'll never hit the credit limit, so Codex is "free".

To get the most out of Codex, share some of the burden: Learn how things work and then collaborate with the bot. That way neither you nor the bot carry all of the weight.

"But I don't know how to code", "but I want Codex to do everything for me" : If these are your choices then your final option is to pay for credits.
Google for "triangle good fast cheap". You get to Pick Two. You can get something that's good or fast, it won't be cheap. You can get it fast and cheap, it won't be good. You can get it good and cheap but it may not be fast (optimized).

I'm just sayin ... There are two answers to the problem of running out of credits, you can learn more and ask less, or you can continue to ask more and then pay more. The issue isn't with the technology - it's with human choices.

Have a great day. 🙂

[boreal holly]

You mean to tell me, saying "make app now" followed by "continue" 200 times is inefficient? 😂

[lean lark]

Hmmm, no, where did you get that? 🤣

I feel bad for these kids who just type and expect miracles. 😢

[steady vigil]

slightly disagree. if you are a power user its easy to hit rate limits, just like if you dont know what you are doing

[lean lark]

@steady vigil totally fair but I'd guess that no one complaining about limits here is really in that category.

[steady vigil]

one thing slightly off topic is that I find myself doing much less big PRD creation now and just chatting casually with the stream in a more organic way. Being able to one shot what used to require a lot of bs framework/skill steps is quite liberating 🙂

5.3 has some magic way beyond the benchmarks I dont see people talking about much

it understands your intent so much better

not just the "what"

[lean lark]

And I understand that personally I under use the tools to keep under the limits and that it "costs" me time as I manually work through issues. That's my choice and I deal with it.
I have chosen this pattern because through to Codex 5.2-high, the bot has made some really bad decisions and I blindly followed a few of them. I think 5.3 is better but still has not earned my trust. I think even a "power user" needs to decide about how to balance better between what is assigned to the bot and what's done manually.

[steady vigil]

it has earned my trust

codex 5.3 high is magic

medium is great too

[lean lark]

I'm at this very moment prepping to give it some more decision-making ability about a big project, coordinating changes to many things across many files in different languages.
"Commit early and Commit often"
I'm giving it the opportunity to shine or fail.

[boreal holly]

I'm tellin ya, best way to avoid rate limits is make local inference parse cli commands! A lot of mine output 40-120k tokens (Xcode builds and flutter integration tests are super noisy), and the command parser outputs 4 (not including reasoning) for "No errors!" or a tight list of errors/warnings/failed tests 😁 I finished the last week with like 87% weekly limit remaining, and I used to with the same usage finish with like 20% or less

[steady vigil]

I think you will be not only "not disappointed" but surprised

[lean lark]

Yeah, Robert - you're talking about your custom skill - I haven't implemented it yet, haven't been asking it to go through a batch of errors much. I tend to filter the errors myself, figure out what the general issue is, fix that, then let Codex deal with specifics. But I'll definitely be trying it at some point.

[fleet geyser]

Is claude pro or chatgpt plus better for vibe coding? I can only get one

[ivory zodiac]

plus no brainer

claude pro is a trial

its a joke

[fleet geyser]

alright, thanks

[boreal holly]

Oh yeah, the alternative to this is "pipe noisy commands to a *.log file and search it" skill rather than just letting them consume 100k tokens as an input from a CLI command. I like the local inference method because it's fewer tool calls

[ivory zodiac]

you get 2x usage for 2 months

[marble violet]

anyone having issue with the codex app high cpu usage and crashing

[ivory zodiac]

yes ppl are saying this

prob best to use cli in the meantime

[marble violet]

I wish I waited to press the update button

[ionic grail]

Claude is far better

[sick iron]

Any news on the Codex app coming to Windows? Or at least WSL.

[ionic grail]

No one that is a programmer who statyes before agentic ai

Would prefer chatgpt

CodeX 3 is a great ste uo

[fleet geyser]

Chatgpt has codex, which is agentic

[ionic grail]

But its so limited and guardrails are insane.

Of course mate

[ionic grail]

I said before agentic tools

Its opinion end if the day

And will improve the products

But no dev ir software engineer ive worked with wpuld use chatgpt over claude

My experience

Only

[lean lark]

About high CPU usage, a change was made recently in Codex (as I recall) to allow changes to AGENTS to be recognized immediately rather than only after processing. This led to a lot of excessive CPU, file-system access, and logging. A change hasn't yet been put into Codex and I suspect it will need to be put into the app as well.

[lean lark]

That's an overly broad statement - and as such is completely incorrect.

[ionic grail]

I said i

Had worked with

Not factual

Of course this is briad

[lean lark]

(Oh wait, you said "that I've worked with" I can't argue with that.

[ionic grail]

Inlove gpt for debugging

All good

[boreal holly]

I disagree. I've never tried other agents, but I think the issue most people have with Codex when they compare it to Claude or others is they aren't putting in any effort to learn and understand the tool. I've spent countless hours scrubbing the OpenAI docs, crafting custom execution environments, evaluating different prompting strategies, coming up with tools specifically to make Codex as low friction as possible, rigorously exploring openai/codex codebase to understand the tool at a deep conceptual level, and refined it into a practically infallible dev setup. It required effort, but overall extremely happy with Codex. It's not even worth my time or effort to look at alternatives because this one works extremely well, and that's the story of any software you put effort into knowing and using.

[ionic grail]

Id say that you shoudl try the others first then decide

[lean lark]

It's a matter of time - we invest so much into understanding the nuances of a particular brand, it's tough to dedicate the same effort into others ... we gotz stuffz ta do...

[ionic grail]

??

I do this for work

[steady vigil]

agree. I've tried them all, currently codex is king (IMO) maybe if you do simple things like making websites the perceived speed (e.g. gap between what it says it did and actually did isnt an issue) could be better value

[ionic grail]

So i dknt really know what you mean

But i get that

[steady vigil]

anthropic is like at least 2 generations behind right now

[lean lark]

I've learned so much about how the ChatGPT models process "tension" in instructions and have put significant time into crafting instructions that do not introduce tension. I can't put that kind of effort into other products.

[steady vigil]

GLM 5 will be interesting to see how close they get to opus 4.6

[ionic grail]

That causes issues

[ivory zodiac]

its really not.

[ionic grail]

As it limits it

Guardrails

[ivory zodiac]

its really good though

[high girder]

Here's the thing most people miss about models in general. Most people will pick one as "the best" and use that, but each model has things they're both good at and bad at. Claude is fantastic at holding a long context plan and laying out frameworks. Codex is amazing at working inside of those frameworks, becuse it's better at smaller scoped tasks and glue code.

[ionic grail]

Bro its opinion

And let's look at benchmarks

And use cases

[lean lark]

Personally, I'm also pretty OK with using the product that everyone else is trying to compare themselves to.

[ionic grail]

Urs great man

Got is fantastic

Gpt

[lean lark]

Just like programming languages, frameworks...

[ionic grail]

I just feel with the intro of ads

[ivory zodiac]

indeed. lets look at benchmarks. maybe you and i have a different opinion about what "far better" means lol.

[ionic grail]

Ita going down hill quick

[lean lark]

for free users

[ionic grail]

As thays not a great sign considering altman said he hates ads

Thinks their paeticauklry bad

In ai

He said himself they would be using ads if they arent doing great (paraphrase) find the interview

[lean lark]

I think we can all agree that ads are awful but peeps not paying cash can pay with their eyes - like we do for every other free SaaS out there.

[ionic grail]

Free users shouldn't be bombarded with ads

Free trials are funnels

To get the conversion

[robust lintel]

Is a free trial infinite and ad free?

[lean lark]

Conversion rates are like 1% - that's not a funnel.

[ionic grail]

Yes

You actually get stocks too

[lean lark]

wut?

[ionic grail]

Hahahahah bro thays literally the point

As to why they need ads

They arent concerting

Converting

[robust lintel]

What

[ionic grail]

They arent making profits

They are burning money and I get this coudl be a long game

Approach but ads arent a good sign

The ceo has said so himself previously indont see how this is controversial

[lean lark]

I meant - average conversion rate of ALL services from freemium is about 1%, some better, some worse. That wasn't a focus on OpenAI that they should do better. I'm saying if peeps aren't converting then they can watch ads. I do this with any free service I consume and accept the "price".

[boreal holly]

1 billion active users lol, and OpenAI has more free users in the state of Texas than Anthropic has users. I think 1% conversion is not a huge deal

[steady vigil]

as a paid user I'd be happy if all the free users disappeared, but if ads will help them recover some of that bottom end sucking cost - great

[ionic grail]

Look at their revenue

And soend

Spend

Its thay simple

They made 17 billion and went theough 8x or so

[lean lark]

No.... This is a service to humanity as well, gotta remember the basic premises....

[ionic grail]

Bro sre tou brainwashed

They'd be open source

You are rolling or slow

[lean lark]

I'm sensing young opinions...

[ionic grail]

Im sensing a person who has no software background or enterprise experience

[boreal holly]

Yeah, it'll be ~7 years before ROI, after all the infrastructure is built for training and inference. This is after all new and cutting edge technology

[ionic grail]

"Service to humanity

[digital thunder]

<:book_icon:1363314738255364126> Rule 1: Be respectful.

-# Practice kindness and positive regard. Harassment, hate speech (such as sexism, racism, or homophobia), or other malicious conduct will not be tolerated. Maintain a respectful and positive environment.

[lean lark]

I've been doing this professionally for almost 50 years. Next...

[ionic grail]

Hahah that's ill it

Your impressed

And easily full for the bs

[steady vigil]

sure but its also a business

[ionic grail]

Open ai is not open source

[lean lark]

OK, I think you've made a point - maybe not the one you'd like to make, but you've made a point.

[ionic grail]

True

Good thing is we all care

[robust lintel]

I feel like you should all move on

[ionic grail]

And we get to dicusss

[lean lark]

As such, there are concerns that have not been addressed by a business at this scale before. Google abandoned the "don't be evil" policy a long time ago.

[ivory zodiac]

dont think i've ever met a real engineer who can't tell the difference between your and you're 🤔

[boreal holly]

Tales of yore indeed

[lean lark]

OK, back to code. Thanks all.

[ivory zodiac]

insults people. cant spell.

ez block

[high girder]

So Codex

[ivory zodiac]

that there looks like windows 👀

[steady vigil]

now they just suck. maybe they should go back to being evil? (j/k)

[high girder]

It is. I did more work on my own port. I'm just missing some backend that I'll never get from OpenAI, but yeah, MOST of it works

[ionic grail]

Clearly never met many

Lol

[brazen lichen]

did the Super Bowl ad easter egg 1K emails already go out?

[ionic grail]

Do you really beleive a company saying they have a dont be evil policy, means they won't be evil

[boreal holly]

Yeah, if you unpack the app.asar, there is a ton of win32 code in there 😁

[high girder]

That's exactly how I did it. I unpacked and worked backwards until it worked

[lean lark]

I caution against big efforts on stuff like this. The personal reward of satisfaction is huge. But in practical terms, it's obsolete as soon as OpenAI publishes their official version. I've been itching to do a custom ChatGPT that has so many of the features that we've all asked for, for years, and which OpenAI doesn't even acknowledge. But I know that would be a waste of time, since on any day they could announce ChatGPT v2.

[high girder]

Each version of the Codex.dmg they put on the site, means more updates, and it's not for daily use. It's because I'm not a dev and I wanted to see if I could.

[lean lark]

Personal feature desires: labels, tags, categories, tree folders, full memory across chats, direct integration with GitHub or other SaaS (now possible with Codex app and ChatGPT Apps). That list goes on...

Yeah, I totally get "It's because I'm not a dev and I wanted to see if I could." that's why I said "satisfaction".

[robust lintel]

Feels pretty dismissive of yourself and others though. "They might make it so I wont explore"

[steady vigil]

I think he was being pretty clear in the cautionary message. you are free to do whatever you want just understand it will all be instantly obsolete at any point in the future

[lean lark]

It's not "not explore". Sam cautioned against "similar but different" a long time ago and he was right. If we make something very unique, the company probably won't stomp on it. But if we make something that's just an improvement over a company product, they have every right to obsolete it without notice. That' the game.

[high girder]

I mean, I got it, and didn't take it any other way. It was advice from someone that knows that pain from experience, that's all.

[robust lintel]

You said you wont even make your own chatgpt interface out of fear they will make it. You dismiss yourself and a desire to explore.

[fickle talon]

Ok

[robust lintel]

The whole launch is saying "Just build stuff", do stuff, enjoy it.

[boreal holly]

In this case, Ruixi just took an already fairly built out electron app, ripped out liquid glass or whatever else is mac only, and repackaged it. And I bet when the real thing comes out it goes in the trash, but the cool thing is they got to play with reverse engineering which is a good learning experience!

[lean lark]

I weigh significant effort to do something against the chance that someone else might invalidate it in a heartbeat. It's a calculation. Time is precious. I need to focus time on efforts that yield more benefit.

[steady vigil]

and sometimes I take the opposite view. like its already been done but doesn;t quite do exactly what I need it to do so I dont mind risking reinventing the wheel. especially if that wheel has been around for a very long time and is no longer cool to build wheels

[boreal holly]

This is actually how Git was invented

[high girder]

100%. The biggest take away was what I learned from the experience.

[lean lark]

Complete agreement yall... If it's for satisfaction, fun, kudos .... do it. If it's intended for the masses then more discretion for use of time is advised, that's all I'm sayin.

[high girder]

Question for you all then. Before the Codex App, and it was CLI or IDE extension, or even with the codex app, do you guys use MCP servers at all with Codex's built in MCP tools?

[boreal holly]

I have Context7 still, but it rarely gets used. Skills have become a lot easier to maintain, and they act as like a memory bank for doing things correctly

[high girder]

Skills are great, but MCP let's you use tools from other programs. So let's say I'm using the Godot MCP and the Blender MCP, with codex being able to talk to both, you can converse your way towards asset creation and visual systems building

[ivory zodiac]

my agent uses it very often. i built a context7 skill but it was failing to find and export the api key via env too often. i should just export it globablly.

if you find its not calling it regularly, then you should add something like this to agents

Context7 MCP ALWAYS proactively use Context7 MCP when I need library/API documentation, code generation, setup or configu steps without me having to explicitly ask. External libraries/docs/frameworks shld be guided by Context7

https://github.com/am-will/codex-skills/tree/main/skills/context7

the skill if you want. its very reliably self-called

recommend exporting api key globally via keychain if you use it

Fetch up-to-date library documentation via Context7 API. Use PROACTIVELY when: (1) Working with ANY external library (React, Next.js, Supabase, etc.) (2) User asks about library APIs, patterns, or best practices (3) Implementing features that rely on third-party packages (4) Debugging library-specific issues (5) Need current documentation beyond training data cutoff (6) AND MOST IMPORTANTLY, when you are installing dependencies, libraries, or frameworks you should ALWAYS check the docs to see what the latest versions are. Do not rely on outdated knowledge. Always prefer this over guessing library APIs or using outdated knowledge.

[ivory zodiac]

you dont need this if you use the skill.

[marble violet]

is there any way we can download older version of the codex app

[lean lark]

I have not used MCPs yet in a project. I'm getting on this bandwagon late because the initial and continuing hype has been for a highly insecure protocol. Behind the scenes it's fine but there's been too much blind consumer faith that it's used properly with any server, just because it's got that popular MCP branding.

Technically I struggle with it because it's precarious. We need to ensure the assistant decides to use a tool at a specific time. If it doesn't then we lose control over context. That's like doing something like price=rnd()*price in code. To be clear, if I say "I need a tie" it might go to the MCP dedicated to finding ties, but if I then say "I need something that goes with my suit", where does it go for that info? It won't call the ties tool because that statement is only loosely tied to the topic of clothing. We've lost control of the context. If someone tells me for sure that this follow-up query will still go to the same tool(s), I'm jumping for it. If not, I'm still hangin back.

[ivory zodiac]

mcp tools have tool descriptions. if an agent believes that tool will be useful, it will call it

[lean lark]

And if it doesn't, we've lost control. That's exactly what I'm saying.

[ivory zodiac]

you're overthinking it

if it can't find an mcp tool to accomplish its task, it will look for other ways to do it

[lean lark]

Consider in a Custom GPT, the same agent is processing every request - there's no loss of control. With MCP, the assistant decides with every prompt if that text relates to a given tool. It's not thinking "maybe we're still talking about the same topic". I'm looking for continuity - I start with a tool and I keep talking with it. Every statement should be a trigger for that tool. But that can be a security issue too.

How about this... My wife and I were thinking about random new Apps for ChatGPT (the SDK App, not Codex "App" sigh) ... If we have a MCP/App for "Things to do today", we initiate the discussion and start telling it what we are going to do today. With every statement we need to trust that it will call the right tool and log the statement. It's not deterministic. It's precarious.

[high girder]

That's where skills come in, you can literally set the skill as an activation to use MCP tools as well. So you can converse, and use all of Codex or Claudes normal tools, and then when you proc the skill, the other tools become callable. I also built a single MCP connection plane for local MCP servers. I don't generally trust things from other by default, so I usually just do it myself but you can set MCP up to be gated securely by clients as well. https://github.com/Janksuu/LMCP

[lean lark]

Skill is Codex-specific but it's pretty much the same (which is why I feel OK about discussing MCP's in general here for now)

[stray swift]

Once they IPO'd corruption was probably inevitable, but Google really did seem to adhere to that policy for a long time.

[hushed storm]

https://x.com/sama/status/2020977975081177343?s=46

Ads in ChatGPT, proceeds with giving everyone free codex that costs way more

[lean lark]

unrelated

[scenic umbra]

Anyone know if it's possible to have multiple terminal tabs for one project? Also the sessions don't seem to save if I close the terminal bar

[hushed storm]

Not really

[scenic umbra]

I would be more fine with ads in Codex than ChatGPT lol, especially if it means larger free limits

[hushed storm]

But like the codex user base is smaller and devs would find a way around it

[scenic umbra]

true

[lean lark]

From SamA:

More than 1 million people downloaded Codex App in the first week.
60+% growth in overall Codex user last week!
We'll keep Codex available to Free/Go users after this promotion; we may have to reduce limits there but we want everyone to be able to try Codex and start building.

Various comments:

1. It's amazing how many haters will find negative things to say no matter what the announcement.

[scenic umbra]

ig eventually they'll make everyone pay

[lean lark]

2. Can we get a special category for people who say "where's the other thing" minutes after an announcement?

3. Free Codex to Free/Go users. That's something everyone should be able to rally around.

4. More than 1 million downloads of only the Mac version of the Codex App. Obviously more to come.

5. 60% growth in overall Codex users means a lot of Mac users have been understandably confused or otherwise put off by the Codex cloud/CLI/extension options. I really want OpenAI to try to find out what happens when people uninstall the Mac version. There's intelligence data to be found there, and opportunities to do better for Windows users.

[scenic umbra]

I just want ssh support and better terminal integration and Codex would be perfect

I think using seperate IDEs to view the code is fine

[hushed plume]

codex merch drop when???

[high girder]

Use VS Code with the codex extension and point it at the same folder you're pointing the codex app to. Use the extension codex to call the codex app like a tool. Best of both worlds

[lean lark]

It looked to me like the hidden merch was built into the Mac app. And it's probably already been published. Anyone?

[hushed plume]

i need this merchandise lmao please help

[plucky halo]

You had to install a skill on codex called 'build-things'

[summer iris]

Did documents just disappear from OpenAI's website? I only see Dashboard and API Docs now

[ember venture]

They did just post this https://x.com/openaidevs/status/2021004633091735877?s=46&t=EjRX3OyHJwyfL_puxs9MWg

[hushed plume]

wdym had

WHERE IS ITTT AHHH

[marble violet]

I am wondering if open ai is using Codex app to fix the bug that make Codex app crash on mac.

[summer iris]

hah, thanks

[steady vigil]

codex app is too slow to do anything. they need to fix this mess or extend the 2x rate limit indefinitely.

"2x rate limits but you can only use it 0.3x the rate". yeah great deal

[latent tiger]

my agent has been randomly freezing mid task on long sessions and i have to restart the app- very frustrating pro plan with 2x limits using 5.3 high

anyway, back to CLI till app is fixed lol

[lean lark]

As I said earlier - on one hand OpenAI says Codex app for Windows is coming soon, and on the other hand peeps are reporting painful issues with MacOS. I really hope they process through the Mac issues, get some peaceful sleep while there are no more big bugs being reported, and then publish the Windows app without the v0.x issues. Pwetty Puleeze?

[cedar skiff]

It's working good for me so far. Im really happy. I have two minor issues, it can be a little resource hogging, a couple of times it got stuck with a overlay asking me for permission, i approved it but it never closed and i had to shut the app down.

Since i started using it i hardly touched the cli

[lean lark]

Resource hogging is a commonly reported issue so far.

Windows has had "Resource Hogging" built-in as a feature since v3.1. I hope OpenAI doesn't feel compelled to add their own version. 😜

[empty tapir]

on mac when running 3 tasks at once it freezes up the whole codex app no wonder there is no windows version yet.

[hollow belfry]

Hi all! Wanted to get a temperature check rq. Am I the only one finding more success with 5.2 xhigh compared to 5.3 Codex xhigh?

Skill issue? Agents.md file + workflow needs to be tuned?

[stray swift]

What does the app get you, over the CLI?

[nocturne folio]

will the merch include free chatgpt pro subs

[hasty pond]

a did this 333wav333

[lean lark]

I could cough up my understanding from the docs and ads, but I'd prefer for someone who is actually using the app to tell us what they actually see in it. If nothing else, the app is a GUI and the CLI isn't. Ship It!

[last stream]

@ivory zodiac there's the lazy loading

[ivory zodiac]

Nice

[ivory zodiac]

I have found it is still almost 400% more tokens than a skill

But undoubtedly better anyway

[last stream]

For sure. MCPs have their use but a skill with scripts shipped with the skill is def more efficient

[ivory zodiac]

Actually maybe worse

[cedar skiff]

most mcps are just wrappers around cli commands o.0

[last stream]

actually i said scripts but was thinking CLI lol. good point

[cedar skiff]

a couple of little rules and codex just uses the cli

[frosty zealot]

At teh end of planning when you're prompted to implement, does Codex default to medium?

If you were planning in high for example

[sand shuttle]

Holy codex 5.3 is such a beauty

[high girder]

I reply with no, turn off planning, save as .md, and use that as the plan so you can pick what you want

[frosty zealot]

thats what I do too, but I notice the model selector is always over the medium

I dont know if it defaults there or if its actually set to medium at the end

[high girder]

maybe planning mode defaults medium

[frosty charm]

I really want to be around the top 0.1% of people who are leveraging claude code / codex for entrepreneurial purposes and exchange alpha. if anyone else interested in this HMU I'll start a private group. Tell me something about yourself and your background.

I started several businesses in tech / medtech / ecom, and for a few years was one of the largest winners in online poker.

Shoot me a message!

[worn folio]

Codex 5.3 has served me quite well

[cedar skiff]

It's been good for me so far as well, lets hope it stays crisp and smart.

how much more usage does the pro plan give over plus?

[toxic torrent]

A lot I never run out

[cedar skiff]

on here it says 6x usage https://developers.openai.com/codex/pricing/ thats seems wrong though.
I mean 10x cost 6x usage?

[still trellis]

would be sweet if OpenAI opensourced the codex app. Lots of ideas around it... someone please make it happen

[high girder]

the app itself is actually pretty laid out and open if you take the dmg download and extract the app.asar

want me to build you something?

[loud dragon]

Keep in mind Pro has other features that Plus does not. More agent requests per month via the app, virtually unlimited access to the very expensive 5.2 Pro model, extended thinking budget on 5.2, Sora 2 Pro, etc. Codex isn't the only factor. 5.2 Pro is an insane model.

You also get priority processing on codex for faster speeds

[lime inlet]

I have been using ChatGPT since it came out but I am very new to Codex. I use Codex CLI with my ChatGPT Business account. I do not understand what the limits are in terms of tokens/queries. Where can I learn more? Can I use Codex CLI as much as I want? Is it just a demo or is Codex CLI from now on included in ChatGPT Business "forever"?

[worn folio]

I pay for Cursor Ultra ($200/m). But I mostly just use GPT-Codex. How does the usage limits stack up for that price? Cursor Ultra vs ChatGPT Pro for Codex usage. Does anyone know?

[high girder]

It really depends on what you're doing with it. If you're doing smaller and more targeted builds, then you'll never hit a limit. If you throw AI at everything to see what works, then you're probably going to hit limits.
I use Codex on the plus plan, and I have yet to actually hit any limits, and I use it daily for a lot of work, especially after I've blown through my claude weekly limits in 2 days.

[worn folio]

I work on pretty massive repositories as well. I hit weekly limits in one day often with Codex. Even on Cursor Ultra I hit token usage in 2-3 weeks. That does make it sound like Codex is pretty worth the expense in comparison based on your Claude experience.

[still trellis]

thanks and could, just dont have the time. Please go ahead if you want to 😊

[lime inlet]

Does Codex CLI/IDE/Web interface support multimodal input? For example, in plain ChatGPT what I can do is attach source code and a screenshot and ask "Do you see this weird graphical artifact where it looks like the rays are hitting the sphere only in some spots? What is causing this?" or "Look at my hand drawn sketch how I want the interface to look and where I want the buttons positioned. Implement this."

Is that possible in Codex also? Or do I need to describe everything as text?

[fallen kettle]

Multimodal! I paste in screenshots all day long.

[subtle lark]

They doubled Plus and only gave a slight bump to Pro a while back but said Pro would be faster which was a lie (it's only slightly faster).

[plucky halo]

Its not an available skill anymore

[uneven spade]

I've seen that Cursor gives you about 2-3x better value than API equivalents. I've also seen stats that say ChatGPT/Claude subscription give you 5–60x better value than API equivalents depending on your usage. ChatGPT Pro is basically unlimited for most people. Subscriptions by model providers are heavily subsdized. So yeah, cursor is not worth it anymore.

[hot hawk]

I was a heavy Claude Code user, but I am a Heavy Codex user now. Redesign the UI in one shot.
old: Middle
new: Right

[white spire]

Could you give me some advice?

I’m about to buy a MacBook. I’m planning to get the Air 15'' M4 24GB/512GB.
Will it be enough for developing heavy enterprise projects with Claude Code and Codex?
I’ve heard that 16GB is already not enough for CC and СX

[velvet wren]

will you ever want to run local models on it?

[balmy acorn]

built a 3v3 pokemon game with codex 5.3 x-high

[white spire]

No local models, just API or subscription usage

[velvet wren]

then 24GB is ok, but I would get 1TB storage.

I have a MacBook Pro, M2 Max with 96GB RAM that I feel I am outgrowing, but I run local models

[finite sorrel]

anyone finding latest client+model getting stuck in execplans where it didn't previously?

woke up to 8hrs of churn and 20% of weekly budget blown but nothing completed

and the thinking personality has adopted this annoying Ash Ketchum style

I've finished the evaluation with a count of 20, which feels good! Now the script should move on to the next models. However, I realize I haven't seen any output yet — I might need to
  poll again since we didn't check after the evaluation was complete. I wonder if the output is just sitting there waiting to be processed? Time to check the session again!

consistently tells me it's going to capture data then fails to do so, repeat ad nauseum

some regression in compaction or something is screwing it up for me i feel

ive given it explicit instructions on what to do yet it shows this in thinking output...

I guess I’ll just keep polling to check for
  updates. It’s interesting how these evaluations can take time! I wonder how long it will take before I get the next update. Keeping my fingers crossed that it’ll all come through soon!

[brazen lichen]

if you can afford it get more memory and more storage

[glacial agate]

@quasi canopy @balmy olive ^^

sorry for the ping ❤️

[exotic cave]

gpt-5.3-codex is routing to gpt-5.2... someone else having this issue?

[tranquil hare]

What does polling refer to here?

[finite sorrel]

polling is repeatedly checking for something

ie in this instance it was looking for new entries in a log file

[spiral gorge]

i have heard nobody solves the codex ad? how?

[tacit halo]

ooh?

[twilit mesa]

Hello, new here! can anyone point me in the direction of the most widely used workflow? which IDE/cli/setup/git is mostly used for fullstack microservices

[balmy acorn]

its still int he works but it is very promising

[tacit halo]

yeah?

i want to make a pokemon game also

XD

[balmy acorn]

codex 5.3 x-high can def do that

[tacit halo]

ooh?

[summer oxide]

How do i access 5.3?

[tacit halo]

yeah?

i use chatgpt on web :(

im not on mac im on windows

[balmy acorn]

its not in chat gpt you use the terminal on your pc or iin vs code

[summer oxide]

nevermind! i just needed to restart the codex app

[tacit halo]

how can i get codex on my terminal?

[steady vigil]

still no memory leak fix for codex app? come on oai

[cedar skiff]

what happened here?!

I just signed up to pro

and the 5 hour is all the usage i used
and some how my weekly is down 12%

[balmy acorn]

i would email them about that or post on twitter i havent heard of that issue yet

[steady vigil]

I did /feedback both in app and a bug report here

[balmy acorn]

then just wait the codex team has been pumping features left and right

[steady vigil]

yeah that's the problem

[balmy acorn]

the first bar resets every 5 hours and the second one is the max cap for the week

[cedar skiff]

exactly

[balmy acorn]

if you were a plus user the cap doesnt reset it stays where u were on plus

even upgrade to pro

[tacit halo]

im plus user

[cedar skiff]

ok so that 12% is my weekly usage i was about to hit on the plus?

[tacit halo]

:D

[cedar skiff]

Really shouldnt work like that, i mean i just paid the money today

[balmy acorn]

dont worry the cap is so much higher than on plus

this me working for the past four hours on pro plan

[tacit halo]

how can i install codex on my computer?

i need help lol

[balmy acorn]

if your on mac their is an app if your on windows either do it through the terminal or just download vs code

[warm pilot]

npm i -g @openai/codex

[balmy acorn]

npm i -g @digital thunder/codex

[tacit halo]

thnx let me try

:D

[balmy acorn]

its easy to setup

it basically does it all for you

if you have issues after download just tell codex to fix the problems lol

[tacit halo]

how can i run it now?

wats the command?

[warm pilot]

codex

[tacit halo]

oki and then?

nvm

wats the rate limits for it tho?

[balmy acorn]

i dont know

[exotic cave]

ok this is sad i am not routed to gpt-5.3-codex, can any of you test it too?:

RUST_LOG='codex_api::sse::responses=trace' codex exec --sandbox read-only --model gpt-5.3-codex 'ping' 2>&1 \
    | grep -m1 'SSE event: {"type":"response.created"' \
    | sed 's/^.*SSE event: //' \
    | jq -r '.response.model'
    
gpt-5.2-2025-12-11

[cedar skiff]

how do i check this?

[exotic cave]

just run in your terminal, make sure you have jq installed:

RUST_LOG='codex_api::sse::responses=trace' codex exec --sandbox read-only --model gpt-5.3-codex 'ping' 2>&1 \
    | grep -m1 'SSE event: {"type":"response.created"' \
    | sed 's/^.*SSE event: //' \
    | jq -r '.response.model'

[neon girder]

can confirm

tbh i was also quite confused, in the past couple days 5.3-codex doesn't feel like 5.3 anymore

[exotic cave]

this is real sad, i found that codex 5.3 started being terribly slow and thinks/talks like 5.2 and found it does xD, also found others in github reporting about it.

more people should report it so they fix this issue.

all the hype and for the past day or so i am served 5.2 -.-

[warm pilot]

gpt-5.3-codex, as expected

[exotic cave]

interesting, is it location/account based... care to give a bit more information about your setup? non the my requests get's routed to 5.2.