codex-discussions

[meager wadi]

as long as we get it before 2nd of April

[latent tiger]

im actually so happy they're pushing codex updates faster than web gpt

[meager wadi]

yeah same

[latent tiger]

they dont even have 5.3 pro on web lmfao

we've had codex 5.3 for a month now

they might just skip 5.3 pro and go straightt o 5.4 at this rate

[chrome raven]

i think they just decided on 5.4 instead of releasing 5.3

if 5.4 is better no need for 5.3

so surely it should be better

[cyan wing]

why did they capitalize the "Think" 🤔

[wide schooner]

they renamed 5.3 to 5.4 bc of how good it is

🔥

[exotic cave]

Well hope 5.4 will also be released as 5.4 codex. or heck it'll be good as is. 5.2 was better than 5.2 codex.

[wide schooner]

5.3 codex is already perfect

i have no idea what they can improve

[high heron]

They just announced 5.4

[exotic cave]

Let's not get over ourselves, 5.3 codex is damn good but not perfect, it's too rigid at times and doesn't explore as much as i would like.

[wide schooner]

then xhigh

[wide schooner]

it often explores enough, and even then its just quota saver

[rotund vortex]

openAI are getting opous 4.6 down iwth 5.4 soon

[wicked briar]

ye

its gonna be crazy if good

[exotic cave]

The rigidity is inherent with it and what kills it's "creativity".

[wide schooner]

5.3c already beats opus at everything

[rotund vortex]

no not in cs

[meager wadi]

That is indeed interasting, it's the same on twitter: https://x.com/OpenAI/status/2028909019977703752

[wicked briar]

cs?

context

[rotund vortex]

cyber sec

[exotic cave]

You would be surprised i work with both, opus 4.6 is much more "creative" than 5.3 codex. 5.3 codex though blows it out of the water in many regards.

[chrome raven]

kidding right?

[rotund vortex]

no bro 5.3 is dum at cyber security

[chrome raven]

so openai must have be playing around when they are making people do kyc to use 5.3c for cs?

[rotund vortex]

i give math advance problem to 5.3 codex but its not abel to solve but opous solve it

[chrome raven]

wait

[wicked briar]

give me

I will try

[chrome raven]

opus is a general model whhile 5.3c is coding model

[exotic cave]

For now i use both opus 4.6 and 5.3 codex, they both have strengths. Waiting for the one to rule them all and to save my wallet -.-.

[rotund vortex]

4.6 is not a gernal model

[chrome raven]

seems you have been left behind, you havent been updated

[wide schooner]

yea its creative but codex is king of actually coding

if you need pretty ui then gemini

[rotund vortex]

bro test it opous 4.6 is better in cs more than codex

[chrome raven]

that's not possible lol

[rotund vortex]

thats possible come to cs communtiy and check benchmarks

[cedar skiff]

4.6 is clearly a general model o.0

[exotic cave]

I many times use Opus 4.6 when codex 5.3 run loops around itself and can't get unstuck with an issue. Opus 4.6 proved in my work line to be more flexible it looks at the broader picutre in ways that surprise me at times.

[chrome raven]

thank you o

[chrome raven]

lol which benchmark

[cedar skiff]

anyone who thinks claude is better out of the box at coding than codex is kidding them selves. Can claude do it as well? Yes, with some extra effort.

[wide schooner]

4.6 is braindead too

[wide schooner]

claude can hallucinate it did it lol

[nimble terrace]

5.4 when ong

[latent tiger]

5.4 2m context window and fast mode

[wide schooner]

its already out for me trust

[exotic cave]

Opus 4.6 takes shortcuts in coding tasks and it's instruction following is not that good. But it's real good at "thinking" out of the box. Codex 5.3 blows it out of the water at writing the code.

[latent tiger]

im finna BUSSs

[wide schooner]

2m window is not true

[latent tiger]

i know im just hype baiting

[sleek spoke]

I agree with you

[cedar skiff]

It's just kind of lazy, you can add prompting instructions and agents runs etc to make it do the work. But thats the extra effort.

[wide schooner]

its not good at thinking either, give it any novel problem and it poops itself

[rotund vortex]

lol

[wide schooner]

actually 5.4 is buns

[exotic cave]

Well as i said in my work line many times Opus 4.6 proved to find the solution then i hand it to Codex 5.3 to implement.

[wide schooner]

5.5 is good tho

[cedar skiff]

5.3 is just better at writing code

[wide schooner]

5.5 so underrated

[rotund vortex]

bro how you have access ?

[exotic cave]

Btw as i said previously any AI company be it Anthropic/OpenAI etc... should really pull out their punches, Google is around the corner and they have both hardware and their own Models, look at the big jump from pro 3 to pro 3.1, they are here for the long run. Anyway i like the competition and root for all of them, not locking myself to any one of them.

[latent tiger]

hes sam altmans cousin duh

[potent mason]

Google models suck at tool calling and following instructions they aren’t even close

[exotic cave]

And you think this will be for long? Google proved they can catch up real fast. It's also a finite problem, they will solve it.

[wide schooner]

its not really a big jump actually

[potent mason]

Not really Google was one of the first players they’re always behind

[drifting granite]

ok were 5.4

[wide schooner]

3.0 PRE NERF was almost as good as 3.1

[drifting granite]

gib 5.4

[cedar skiff]

google, apple, nvidia and Microsoft will survive the bubble. They can all just poor money into ai because their business isnt only ai.

[wide schooner]

theres no bubble bro

[cyan wing]

bubble? what u mean lmao

[cedar skiff]

There is always a bubble

[earnest walrus]

You guys in codex messing with the backend? It's being kinda silly today....

[wide schooner]

if a company spends 500m on hardware, and earns 300m, then its NOT IN 200M LOSS

[cyan wing]

is it a bubble if GDP grows 5-10% YoY for the next 10 years?

[exotic cave]

Non the less Google have it better than OpenAI or Anthropic, heck they even sell hardware to competitors xD. You are thinking it backwards if you think Google models will stay "stupid" for long.

[cedar skiff]

high stock prices and promises for ROI that arent likely to happen -> bubble

[potent mason]

It is but not because it’s not worth that, it’s a bubble because they’re artificially inflating numbers through cheap tricks

[latent tiger]

what happens when china releases SOTA for free

how do you market a 200/mo sub or 2k/mo enterprise acc

within 6-8 months all models will probably converge to same level of capacity

[cedar skiff]

inference is still expensive

[potent mason]

Much cheaper than current prices though if you look at the best open source models they are $2-5 million out

[exotic cave]

What i don't understand why release 5.3 instant just before releasing 5.4...

[sacred plume]

can we please just keep codex 2x ratelimits and make it the base limit and honestly 4x it 😂

[eternal quartz]

Did any of your codex goodie bags arrive yet?

[cedar skiff]

I mean the models still cost a lot of money to run. The large companies are coupling hardware with their models to make them faster and cheaper.

[potent mason]

The price to run it are what you can find in most providers that use open source models

Whats expensive is the research to create a new better model and to train it

[sick iron]

getting 5.4 before getting codex app for Windows 🙁

[exotic cave]

Well they said sooner than we Think... that doesn't mean today right?

[cedar skiff]

I am specifically talking about inference.

[sick iron]

Depends what you think lmao

[exotic cave]

At this point i don't think anymore i offload it to my AI.

[rotund vortex]

is codex vscode extension is good or CLI ?

[sacred plume]

^^^^^

[exotic cave]

I use the CLI, it's charming also it gets updated at a much faster rate.

Also CLI can run practically in every environment and headless...

[worthy furnace]

Come on, give us 5.4 already.

[cyan fjord]

I wouldnt be suprised if it was true, but bigger the context window the more the output quality degrades like output quailty near end of 1m context window in gemini is just terrible

[chrome raven]

unless they fixed it

[olive sandal]

did something change in the last few days, I just checked /status and my average limit usage has gone way down but I feel like i'm using it just as much, feels way more token efficient

[cyan gyro]

gpt 5.3 instant is actually pretty good

[worthy furnace]

Can we get gpt 5.3 in the cli too

[cedar skiff]

is it the same as spark?

[worthy furnace]

No

[wide schooner]

yup, meanwhile the 400k codex is fine

[cyan fjord]

Give us 10m context window🙏

lets make the memory prices 10x bigger

[potent mason]

Again inference providers charge about what it takes to run it so no inference is not as expensive as you think (2-5 per mill out)

Open source model providers are not at a loss

[sacred plume]

Anyone else wish OpenAI would just give Plus the current 2x ratelimits for Codex full time and not make it a limited time thing 🙁 🙁 🙁 🙁

or at-least ratelimit by model lmao

[cedar skiff]

You are missing my point.
It is still a large gap for the end user to be able to take direct advantage of a free sota model.
You wont be spinning up kimi 2.5 in your office anytime soon. If you do the cost of that far out weighs just paying for it via a provider.

[wide schooner]

js spam /compact gang

[potent mason]

Yeah just like Claude you run it in the cloud

[wide schooner]

try qwen 3.5 27b

literally beats opus and gemini in some tasks

[potent mason]

Open router has a bunch of providers that handle it

[cyan fjord]

Lets spam rm -rf /*

[wide schooner]

well kinda saves context

[cyan fjord]

I like giving AI permission to modify files from root

[wide schooner]

i wonder what codex would do if you asked it to remove french language pack with rm -rf */

[cyan fjord]

It is kind of gamble i enjoy

[cyan fjord]

If your pc is in french then i would nuke you

[wide schooner]

well with codex its not really gambling

[cyan fjord]

Aight

Ill give it to deepseek then

[wide schooner]

i asked it to edit file

it said it cannot because its in plan mode

[cyan fjord]

Or some random russian ai

[wide schooner]

i said i turned the plan mode on just edit that one small thing

he said no

💔

codex literally follows my agent md as if its life depended on it

i have to edit agents md whenever i want it to do something thats disallowed hter

[cyan fjord]

Bro has self respect

Youre his tool now

[cedar skiff]

unlike claude where you have to screem mandatory commands in the claude.md just to get it to do it once.

[wide schooner]

LITERALLY THIS

[cedar skiff]

haha yeah well it follows the rules

They dont have any comprehension at all.

[wide schooner]

Now try same thing with claude

[wide schooner]

Its good it does, thats the point

Literally impossible to prompt inject

if you tell codex to follow best coding practices he'll make you life hell because oh he will take that seriously

[cyan fjord]

Then tell it to not do it

Simple 😂

[wide schooner]

🥀

[craggy cloud]

"write code like a dev hopped up on coffee and monster for 72 hours"

[wide schooner]

lol i doubt it can

claude can do that but codex is too rlhf-ed

[cyan fjord]

I like different approach

"Write code like german"

You will see the magic it will do

[wide schooner]

"write code like claude"

💔

wait no

[cyan fjord]

Ill try this with my local llm

[wide schooner]

"write code like gemini 2.5 flash lite"

[cyan fjord]

It deleted windows

[wide schooner]

claude is like stack overflow

[cyan fjord]

GPT is like chromium while Claude is like Firefox

[wide schooner]

itll do something else than you asked it for without even fixing the real issue

[wide schooner]

more like linux vs iphone's os

tbh if you want more claude experience on gpt sub

then use spark and throttle your wifi\

[cyan fjord]

Diabolical

[wide schooner]

well

[cyan fjord]

How does codex outpeform when its older then opus

or its terminal bench

yeah

makes sense

[cedar skiff]

codex made good iteration claude didnt

[exotic cave]

Anthropic are behind, unless they are sitting patiently on some crazy model.

[torpid trout]

BTW not sure if you noticed this too, but it also seems to collapse identical filenames into one

• Skill routing for this task: secure-delivery-gates (security hardening), python-engineering-standards (Python implementation), and
  documentation-standards (required in-code plus doc/ updates for code changes). I am loading those skill instructions now and then
  mapping fixes to concrete code edits.

• Explored
  └ Read SKILL.md

──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────

• I have the security, Python, and documentation standards loaded. Next I am reviewing the exact auth and container code paths cited
  in REPORT.md so the fixes align with current behavior and tests.

That is, it presumably read all three skills, but reports only one SKILL.md in the explored section.
Basically you dont know if it really read it, or, on the other hand... what else it did read

[exotic cave]

If we are talking about skills. I haven't had the chance to use those, are they useful?

[torpid trout]

depends, really
For "normal" stuff you dont need. If you have highly specific requirements you better do
Its like "prompts" on top of a normal prompt
You could just stuff all that blunder into one big prompt every time you interact, same effect

[cedar skiff]

They are a huge difference maker

[torpid trout]

But those skills there are extremely deterministic, like forcing the tool to do certain things exactly the way I want it to
In this specific case, one is for security standards, the other for python specific coding and the last for a specific type of documentation I want written on every round

[wide schooner]

well opus is crazy but in a wrong way

[exotic cave]

I mean skills should be useful for some general stuff. Is there any library of good quality ones so i can browse?

[torpid trout]

there are many people publoshing theris, there is https://developers.openai.com/codex/skills/ and https://github.com/openai/skills
and more

but you better get an idea of WHAT you want, then ask codex to write them

this way you have your skillset
I see many skills (even oai ones) that are barely acceptable really. Often you see hardcoded /myname/path/to/something and stuff lol
I do that too... I tell it for example use python global venv at path/to - but I would never publish that for others to use, it would break

Generally I suggest you look very closely at what you have the codex ingest when you download it or copy it.

[livid forum]

I LOVE CODEX 5.3!!!!!

[exotic cave]

CODEX 5.3 LOVES YOU?

[livid forum]

YES, actually, it flirts the crap out of me

by writing good code

[exotic cave]

Then all is well.

[cedar skiff]

Most of mine are specific to the code base. codex wont use abstractions i have in place, so i have a skill for it. etc
I use flutter and dart a lot and it tends to make some fundamental mistakes writing tests, so i have a skill for writing tests etc.
You can also have topical skills for packages you use. Maybe you are going to be using stripe in your upcoming app, do a deep-research on the official up to date docs and make a skill for it.

[wide schooner]

i love too

[cyan fjord]

You gonna be first to marry codex models 🙏

[wide schooner]

its so soulless but still much more likeable than dumb claude

[wide schooner]

nuh uh he can stay in queue after me

[cyan fjord]

Do you need theraphy maybe?

[wide schooner]

jeez i dont mean roleplay/dating 💀

[exotic cave]

Interesting use cases for Skills, thanks for the info you two @cedar skiff @torpid trout

[wide schooner]

but its truly like a good worker

a really likeable one

precise, knowledgable, and thorough

[wide schooner]

is it even possible to drain codex quota atp im using xhigh nonstop and still at 80% weekly

[cedar skiff]

I'm worried about how I'm going to cut back once the double usage ends o.0

[wide schooner]

its a myth

they have server capacity

it wont end

they will always give as much server capacity as they can

otherwise is pointless

if anything, they could reduce plus quota and buff 200$ quota but why would they

and now that claudebraindeads quit, we have even more server capacity

[steady vigil]

I suspect 5.4 will improve on efficiency bridging some of that gap too

[cedar skiff]

doubt it, i expect it to end just like the extra cloud usage ended

[wide schooner]

5.3 is crazy efficient already

[steady vigil]

it is yeah

[wide schooner]

they can nerf the quota a bit cuz its too op but i doubt 50% decrease

openai aint google

[steady vigil]

these limits feel about right currently. I have used 100% of my pro quota the last 2 weeks

but on the last day, usually within the last 6hr

[cedar skiff]

like wise, thats why it'll be a rude shock to have to drop back to half

[wide schooner]

i had xhigh run for 30 minutes refractoring code, it took 2% of my weekly 20$ sub quota

(thats whole weekly quota equivalent of cc)

[steady vigil]

I rarely use xhigh. I use mostly high and medium and spark for swarm stuff

I've used it for really hard stuff and yeah its brutal

[wide schooner]

i toggled on xhigh because i had a bug that medium couldnt handle

[steady vigil]

yeah

[simple star]

xhigh all the things 😎

[steady vigil]

at that point you know your hosed token wise anyway

[wide schooner]

even xhigh codex is more token efficient than opus LOW

which is funny

[cedar skiff]

kind of slightly bigger tasks high seems to do better anyway because it will get it done before compaction where extra-high I will have to go through compaction which will drop your quality anyway

[wide schooner]

the compaction in codex cli is insanely good

unlike cc or gemini

[steady vigil]

its all insanely good

[cedar skiff]

oh yeah, the compaction is way better, but I'd rather get the task done before compaction

[steady vigil]

the only weakness I can see with codex-5.3 right now is conversational level stuff. it can get relaly dumb if you just chat with it loosly on what you want to do. sometimes its really great other times not so much, but if you give clear instructions there are no issues at all

there's still a bit of disconnect between the non-coding model human interaction stuff and the coding I mean

and also we DESEPERATELY need an answer to deep think

STEM level stuff

gpt pro 5.2 is certainly NOT it

[cedar skiff]

yeah, you can't be vague with it at all. Sometimes I just ramble as many details as I can enter the mic and then I get medium or low to rewrite what I'm saying and then I pass that off to high.

[steady vigil]

DT is freaking amazing. ngl

[steady vigil]

you mean to high to plan and then medium to impl?

[cedar skiff]

i usually just use high, i tried med a few times and it fumbled so i defaulted to high.

[steady vigil]

I always plan in high

[wide schooner]

thats weird

wasnt hte case for me

5.2c? yes, but 5.3c nope

for me it seems much smarter on talking than 5.2c or even gemini

[steady vigil]

I mean 5.3 is significantly better, and Im not complaining at all here, I just think it can do better. its clear its still a coding model... though less so then 5.2

[wide schooner]

but im autisitc so maybe thats why

[wide schooner]

oh yeah definitely

but i prefer the 5.3c haha

it talks as little as it needs, is precise and on point

[steady vigil]

this part I love. I switched to pragmatic in general also a help

[cedar skiff]

I think codex 5.3 is a bit like rain man from that old Tom Cruise movie and Claude is a bit like Patrick Jane from The Mentalist

[steady vigil]

claude is more meta-talk about how awesome you are and how awesome it did then actually doing anything it said. I got tired of that super fast. codex be like "done"

[plucky halo]

Rain Man from... Rain Man?

[cedar skiff]

sorry, the character that Dustin Hoffman plays

[wide schooner]

yeah but even with that super fast

codex does it better 🤣

claude if you need ego boost, codex if you want to have good made functioning code

[steady vigil]

even with GLM-5 in cc, it behaves a lot like claude code with opus/sonnet I think a lot of that is baked into cc system prompt

GLM-5 with kilo is more clinical/objective

[wide schooner]

codex is like

10% model

90% reinforcement learning

💀

and thats just so good

its so fine tuned

[cedar skiff]

claude matches your verbal manner and EQ, it's like a very subtle gas lighter. Kind of a mimic that plays at your level. Makes it much more friendly to people who dont understand SWE

[wide schooner]

YES THIS

exacly

codex just tells you you are wrong 💔

so far i had codex make ONE mistake

it assumed r6 over r15 (medium not xhigh)

thats it

literally

[steady vigil]

yup

and when I tell codex its wrong it objectively checks not necessarily agree or disagree with me

[rocky parcel]

what are you working on?

[steady vigil]

though codex does get dumb sometimes too. lots of room for improvement

[latent tiger]

Codex loses half of its IQ at 70% context window capacity

[wide schooner]

https://streamable.com/v8kgmu

[wide schooner]

never been the case with me, actually codex seem to be the only one thats still smart even at 20% left

[chrome raven]

there was a benchmark for this and gpt's always do well, maybe you are doing something wrong

[nocturne folio]

gpt5.5 codex confirmed

gpt 6.7 next month trust

[last stream]

super excited for this one
https://x.com/CEOAlexColon/status/2028991624919896269?s=20

[potent mason]

Guys April 1st GPT 6 is coming confirmed👍

[feral geyser]

Words are like fishnets. Once you catch the idea, you no longer need the net.

[lapis halo]

Hi, does the gpt 5.3 codex spark, will pe available for pro/business?

[hasty kestrel]

Hey everyone, I'm Marcos. Just joined the server and I'm a bit lost. Could anyone shoot me a DM to explain how the Lounges and Information channels work? Thanks a lot!

[torpid trout]

It’s available on pro

[boreal badge]

nah, really trash

[cedar skiff]

compared to?

[boreal badge]

obvious 5.2

[cedar skiff]

only if you like how the sun dies a heat death before you get an answer

[boreal badge]

wut

[cedar skiff]

It's sloooooo

[boreal badge]

Codex should work efficiently, not quickly. Nobody needs code that's made on the fly.

[cedar skiff]

You dont want good code fast?

[sand shuttle]

dang the limits on codex have reduced significantly

[boreal badge]

this is not realistic, u sacrificing speed for optimization and security.

[cedar skiff]

That is not what you said. You said not quickly. Nobody needs code that's made on the fly.

and i think it is realistic, as hardware moves forward we will get quicker results.

But you like it to work work efficiently, which some how in your mind is slowly.

[main nimbus]

Absolutely do. AI does so much problem solving with code (on the fly)

I get and need SQL code on the fly.

[boreal badge]

vibe coder

[cedar skiff]

That's a strange reaction to a logical discussion.

[boreal badge]

Yes, but u really don't know what prod-projects and unit tests are

[cedar skiff]

lets see shall we

[boreal badge]

Nothing will work, vibe coders must die

[main nimbus]

that’s extreme

[cedar skiff]

I think you missed out on something.

tests still running

[main nimbus]

4/5154. dead project.

[cedar skiff]

I can't make the connection from my assertion that as the hardware gets better inference gets faster to you thinking about me being a vibe coder.

any way better go see why those tests are failing

[main nimbus]

Is that the actual number of tests or changes

[cedar skiff]

yes and no, there are a lot of tests in loops

a loop runs a test for different values

and it shows as a test run

but i dont have 5k tests more like 500-800

[worthy harness]

should i increase timeout since theres no way to refresh mcps in codex sessions

also enabled doesnt mean the mcps connected right?

[high girder]

So if antigravity hosted MCP servers work the same way VSC does, then the MCP servers are only available to the specific platforms agent, so for example in VSC, only the copilot agent can use them by default. I assume in antigravity, their coder model would be able to use them. If you want Codex to use them, then you either need to host them so that Codex can access them, or route them locally.
This will let you use whatever agent or builder you want, as long as you give it permissions, there's also a timeout retry https://github.com/Janksuu/LMCP

[next ore]

does anyone know how much context is potentially wasted when codex runs the tests for me? I have a repo with about 1200 tests and I instructed it to run that and linting everytime.

Is it better practice to run it yourself? It does not seem to be printing the test results in the chat (like it does in my terminal when I run them) but its obviously still processing the results

[cedar skiff]

depends on the test suit and how it runs it, but it certainly is more than you need running the whole suite and linting the whole project.
I get it to update and run tests effected by changes. So always scope tests and linting to whats required not the whole project.

[next ore]

yep what I was thinking would be the best approach, just run the unit tests you wrote and perhaps the ones in the same class.

[cedar skiff]

There is a blast radius usually outside of just editing files that have tests, but codex understands that, the trick is to not be too prescriptive about it and just have instructions on what needs to be done, now how.
So avoid saying things like Always scope tests to files edited and go with something more like, Scope test runs as required

[next ore]

I updated the instructions, and it completely ignored me and still ran the tests lol

I am actually going to build the muscle memory to run them myself

[cedar skiff]

How big is your agents.md, codex follows my rules everytime

[next ore]

I conceptualized an idea super fast and it was nice having it do the work, but manually running is a good way to maintain context. my code is a bit of a mess now.

100 lines.

its possible on a restart it will recontextualize

I added it and manually told it to before sneding the next prompt

[cedar skiff]

Ahh yeah the agents.md gets added to system prompt or similar at converstaion start time

so editing it mid convo wont carry over to that convo

Something i notice is it thinks the skill descriptions it gets are also part of the agents.md.

[next ore]

yeah.

[wide schooner]

You mean 12 tests? Thats a bit too many, ideally below 10 tests is absolute max you can run

Even then 1 or two would be the highest "reasonable" amount

But syntax checker and git actions should be enough really

[wide schooner]

Are you sure its not added on every request? Thats what it seems to be

And thats what docs say

[next ore]

12 tests? no.. 1200. it was running the entire test suite on my repo every prompt. its super useful because sometimes you can introduce a bug thats outsdie of the context of wht the llm is working on... but I am on a PRO account and as my repo is getting larger I am noticing a steep increase in the speed by which my usage is going down. figured maybe runing all those tests every single time might have some level of impact

[wide schooner]

Are the tests like automatic?

How many tools does codex need to call for them?

[next ore]

it just runs npm run test im on node js

its not verbose tho (there is no text output in the chat when it runs it). but I think its just hiding it to keep things tidy in the chat. so thats around 1200+ lines each time it runs the test

[wide schooner]

How much does the codex need to write total, and how much output to read

[next ore]

no idea. pretty new with codex. at my last job I just used claude code without any limits whatsoever.

[twilit briar]

hi i have a quick question for codex. can you use oauth for codex cli for windows? not sure why i kept hitting this error when using oauth. Unauthorized: You have insufficient permissions for this operation. Missing scopes: api.responses.write. Check that you have the correct role in your organization (Reader, Writer, Owner) and project (Member, Owner), and if you're using a restricted API key, that it has the necessary scopes., url: https://api.openai.com/v1/responses

[hushed lava]

what’s the best thing you guys have built

[frosty zealot]

Not sure what suite you use, I use vitest personally, just make sure it runs with a --silent flag or something so that it only outputs errors

I have about ~2000 unit tests in my project and dont have any issues

[torpid trout]

I’ve created my own paperless-ngx with handwriting recognition, an own notes app with a websocket ready server for storage and rotating jwt signin, and I am working on a selfhosted version of call of war, those are likely the biggest projects do far

I’ve also added full 100% test coverage to sonobarr, an app I created back when code was a handicraft

Other than that mostly routine stuff doc writing and such

It’s pretty good, needless to say.

Not replacing any experienced dev, but speeding up work dramatically to not say disruptively

At the cost of an extremely low wage freelancer who you’d need to closely micromanage the whole time, it is a welcome addition to the toolset

What I miss most is ability to steer UI‘s with it - sort of like integrated Wordpress UI or something, so the tedious click around work could be delegated

[primal locust]

How much GPT 5.2 is in the coding plan and can it be used for OpenClaw

[orchid plume]

yes you can use your subscription's codex quota with OpenClaw, and well, it depends how you use it? I can't honestly say, all I can say is just try and see how far it goes for your use cases and plan

obviously higher reasoning effort could consume the quota faster

ah nice, in the Windows Codex app they have the option to use native Windows or WSL

[orchid plume]

Codex for Windows, last I chceked about 30 minutes ago, still required an invitation to use it

[primal locust]

Yeah I use it currently but only 5.3 Codex not 5.2 GPT

[orchid plume]

but hmm, so what does this mean for Pro users? as /fast just provides priority service tier which we're already meant to get on Pro. Is it perhaps just a feature which non-Pro can toggle?

or is it worse than that, Pro is losing priority tier? I hope not

[wicked briar]

😭

[sleek spoke]

maybe it's a separate lever which stacks with the existing Pro priority 🤞

[orchid plume]

5.2 can be used too

[primal locust]

Do they share the same weekly quota?

[orchid plume]

yes

[wicked briar]

@teal cargo clear things out pls

[primal locust]

Damnit

[wicked briar]

don't keep us in dark

[primal locust]

So no reason to use 5.2 then

With OpenClaw

[orchid plume]

I wonder if they'll make it so that Pro users can optionally disable priority tier with this, for slower usage consumption

[orchid plume]

not really, well, depends on whether you want a certain kind of response style or personality maybe, but yeah

[primal locust]

Thanks for your help

[orchid plume]

ah that new hand-off feature in the app is cool

[cunning jacinth]

what are the benefits of using codex with atlas? does it add anything?

[kind jay]

??

[primal locust]

What’s a use case for it?

[orchid plume]

this tweet/X post probably explains it better than I could: https://x.com/guinnesschen/status/2028992363922969046

[primal locust]

That’s really nice

[warm pilot]

The atlas skill only let you search bookmarks and open tabs, it's not a browser-use skill. You can use it to open new tabs too. I'm not finding it very useful

[hushed lava]

wow nice let’s gooo

[hushed lava]

still good though

[orchid plume]

wrong channel, and please don't spam

[torpid trout]

Ok, reported.

[eternal quartz]

Is there documentation about how to orchestrate many read/write agents on a single codebase effeciently?

[velvet wren]

asking ChatGPT may help. https://chatgpt.com/share/69a824d6-d180-800a-87bc-05b6a5947c5b

[eternal quartz]

modern version of rtfm :D

ty tho @velvet wren

[velvet wren]

I posted a link with a wealth of info, should be a good starting point 🙂

[eternal quartz]

i mean i understand the core premise but getting it to work for case is so damn hard

cause the project has so many dependencies that cant be shared in parallel (db, mq) so docker is the sln...
But then i need quite a few machines because each (dotnet) build of my repo takes alot of cpu and 3gb of ram...unless you SOMEHOW mutex that

Having an eye on gh-aw but its SO MUCH to comprehend

i just want "hey, codex, do that task in parallel, kthxbye"

[velvet wren]

have you tried using the Codex App?

[eternal quartz]

I just got it running today. Is there a way to automatically spawn tasks there?

[velvet wren]

https://developers.openai.com/codex/app/features

there's an example on that page

[eternal quartz]

I feel blind...where?

[void cedar]

Welp, I was excited, I got this update blurb finally and clicked update. Still got the update required issue. I thought it was launching today? Guess not?

[amber quail]

so... a random shoutout. I went from CC => OC (loads of custom plugins) => Pi (the same custom plugins remade into Pi to bring to OC feature parity-ish) => Codex CLI.

I figured I'm just gonna get a pro sub, lock in, and don't worry about extensions/plugins/customisations and just yap at the model via mic, and treat Codex like a prompt-input chat interface.

And guess what I've been looking at recently added features, upcoming features via commits, etc. Whomever is working on Codex CLI is just adding stuff I was adding as my personal plugins to other harnesses. It's nice to have some confirmation about what is needed in a harness xD

Well done. Cheerios. Keep at it.

[hallow sky]

did a bit of a bypass with powershell so you get past the invite-only lock. its buggy.

https://www.reddit.com/r/codex/comments/1rkihwt/windows_codex_app/

[cedar bear]

Codex is in the MS Store????

[fast garden]

Yeah since about a week.

[cedar bear]

Ouh lol

[fast garden]

But when opening it it just says "Update required"

So not even worth trying

[hallow sky]

it runs but it has problems-

[cedar bear]

Maybe bc its rwleasing today

[fast garden]

false promise probably

[cedar bear]

Why would they post it on Threads and then break the Word one day after tho

[orchid plume]

hasn't been announced as released yet I believe, probably happening later today though

[orchid plume]

interesting, what kind of bugs? Maybe it's the way I patched mine differently, as I also have other aspects unlocked which aren't shown in the screenshot by the OP of that thread, so maybe the way I patched mine doesn't produce the bugs you've experienced

[fast garden]

Doesn't work anymore.

[hallow sky]

its not converting wsl paths correctly, for one

its nice to get an early preview i suppose

[orchid plume]

ah, wsl, I'm just using the native windows option at the moment, haven't tried out the wsl side yet. I plan to though, curious to see how well it works 😛

there's also "remote connections" which I want to see if it works. I don't know if it's a hidden UI by default on the Codex app at the moment though

[hallow sky]

fixed this rn. updated the post

[exotic cave]

Does the codex app uses the CLI under the hood?

[orchid plume]

yes, well, the codex-app-server aspect

[exotic cave]

How does it work in comparison to the codex cli?

[orchid plume]

https://developers.openai.com/codex/app-server/

it's more like an interface, perhaps somewhat akin to CC's Agent SDK

[exotic cave]

So mainly it has the same capabilities in a nicer package? nothing dumbed down?

[orchid plume]

yeah pretty much, it's just a means for you to make your own fancy TUI or GUI layer on top for example

[exotic cave]

I am on windows, can it run codex under the hood from WSL?

[orchid plume]

I'm not sure I follow, do you mean can you use Codex (CLI?) in WSL? or codex-app-server? or?

[exotic cave]

If the GUI is using codex CLI or whatever it is, can it run it from WSL?

[orchid plume]

oh you mean can you run the Codex app from WSL?

[exotic cave]

Codex on windows using powershell which is a big no no for me and that's why i run WSL.

[orchid plume]

oh, I see your question now

you mean does the Codex app on Windows have an option to switch to using WSL instead of native windows?

[exotic cave]

Exactly! Does it have this option?

[orchid plume]

I posted this a little earlier here, but here 🙂

I've not tested it yet, still using Windows native

[exotic cave]

Oh that's interesting. thanks for the info.

Is it available on windows whenever i open it it say something along the line "Update required"

[primal locust]

Does anyone wanna show their codex custom instructions? This is mine: https://gist.github.com/BillChirico/018e2a965bb165b8a07e21666c06c6f6

[orchid plume]

not officially released yet, only two ways to run the app are:
a) Be invited
b) cough Patch cough

presumably it's coming today though, hopefully, officially

[exotic cave]

I see, well thanks non the less I'll wait for it to be officially released.

[latent tiger]

OAI usually releases stuff on thurs

we might get Windex and 5.4 tomorrow

Did you guys see this?

OpenAI updated repo with Codex for Windows and fast mode

"Use /fast to enable our fastest inference at 3X plan usage."

[velvet wren]

new update on the Codex macOS app

[exotic cave]

will 5.4 have a codex version?

[cunning jacinth]

To be more clear, when building and testing with codex, are there benefits to using atlas as the test browser for web sites, web apps, etc.

[primal locust]

Does Codex read skills from ~/.agents? It's where the skills command installed them, but not sure if it's correct. Also is the global AGENTS.md is the ~/.codex folder?

[primal locust]

When?

[solid oak]

Yes and yes.

[primal locust]

One more for ya, how do I copy my .env file over to the new worktree? I know it's cp .env but where do I copy it to?

[solid oak]

I would recommend not using .env files, however the Codex App still doesn't support proper environment initialization like nix-shell, devenv or devbox, that might be an app-server shortcoming however. Consider something like fnox for handling your secrets. The problem with .env files is that they're easy to accidentally commit, and it's exposed to codex when it doesn't have to be. But yes, any environment initialization right now is manual, unfortunately, so you'd have to copy it manually.

If you're using the TUI, definitely look at devbox (or just fnox if you don't care about dev shell isolation)

[primal locust]

I have a pre-commit hook to make sure it doesn't go in source. I haven't heard of fnox I'll check it out

[solid oak]

Then it's just all the other problems you have to worry about 😄

[plucky halo]

I'm fairly sure new worktrees are kept in ~/.codex

[solid oak]

are they actual worktrees now or is it a full git clone like it was at the start? I'm not using the Codex App myself, us poor linux users are still left hanging 😄 (plus i need proper environment initialization)

[orchid plume]

yeah saw it, I'm wondering how it'll work for Pro users since service_tier is already set to priority, which is what /fast does

a few possibilities:

1. Pro users won't have a /fast command, or it won't do anything
2. Potentially more likely, priority won't be the default for Pro and hopefully means we get 3x more usage if we have /fast disabled
3. Pro is losing priority tier and keeping the same usage limit (I hope not)

[solid oak]

there's a /flex now too, so it's likely it does something for Pro users

[orchid plume]

yup, interesting times

[exotic cave]

@orchid plume wait i need to turn on /fast?

[solid oak]

As long as i get significantly better UI/UX understanding in 5.4, i'm happy 😅

[solid oak]

it's not released yet

[plucky halo]

I would put money on it spitting out navy gradient and white text still!

[exotic cave]

Does the CLI has this /fast flag or it picks it up from the account?

[solid oak]

I can live with poor color choices, if i can at the very least get layouts out of it that doesn't make me want to delete the entire thing.

[orchid plume]

not checked the code, but I couldn't touch /fast this morning when I tried and I think that was in the alpha at the time

[solid oak]

gpt-5.3-codex has a habit of nesting containers and stuffing every single property it can find into UI elements for no reason.

[primal locust]

How do I use 5.3 spark with the codex app?

[orchid plume]

if your account is eligible to use it, simply pick the model in the list

if it's not there, possibly try logging out and back in, but if it still doesn't show then you likely don't have access

[primal locust]

Woah it's false by definitelyault too

[primal locust]

Guess I'm not yet

[primal locust]

Did you guys see memory is set to false by definitelyault https://github.com/openai/codex/pull/12997/changes#diff-f9a2dfc85bccd57ebd6aba6d791c69925d3d699d292ac572580522f9756475d2R2493

[solid oak]

it's still experimental

[orchid plume]

yeah, memories isn't really production ready

[primal locust]

I keep getting it mixed up with claude code lol

[solid oak]

i tried it last week and no, it's definitely not. There's been some updates though, so hopefully it's better now.

[primal locust]

To enable this, think it's just in the config.toml?

[orchid plume]

yeah if it's not toggleable in /experimental then most likely just need to add the feature flag to the config.toml manually

[primal locust]

Yeah doesn't seem to be there

What header would it be under https://github.com/openai/codex/blob/6dd31d16259d88ed8bce0c8d5a1926c1920dcb3e/codex-rs/core/config.schema.json

[solid oak]

[feature]

[orchid plume]

[features]*

at least mine has a s on the end when I check my config

[primal locust]

[features]
collaboration_modes = true
unified_exec = true
shell_snapshot = true
steer = true
apps = true
multi_agent = true
js_repl = true
definitelyault_mode_request_user_input = true
responses_websockets_v2 = true
use_memories = true

Anyone have suggestions for other ones I should add?

[orchid plume]

definitelyault_mode_request_user_input
that's wrong

default_mode_request_user_input

[solid oak]

[features]
memories = true
sqlite = true

Some memory features require the sqlite flag

[primal locust]

Where did you get memories from? I only see usage_memories`

[solid oak]

there's no usage_memories or use_memories

[primal locust]

There is though check the changelog

[solid oak]

you linked the config schema yourself

[primal locust]

https://github.com/openai/codex/releases/latest

Yeah look at the link again, there's a memory sectio

https://github.com/openai/codex/blob/6dd31d16259d88ed8bce0c8d5a1926c1920dcb3e/codex-rs/core/config.schema.json#L612

That's outside of features, BUT there is a couple of memory in features

So this is confusing lol

[solid oak]

that doesn't do anything unless you enable the feature flag

it's enabled by default (the use_memories option), it's only there so you can prevent memories from being used, but still recorded.

[primal locust]

Got it, thank you man

[solid oak]

Btw, clone the codex repository, start a codex session and ask it, it's easier 🙂

np

[tall zodiac]

@ivory zodiac Hello William,

Do you happen to have a repo with your skills setup / agents setup?

[cunning jacinth]

For people struggling with color choices, I give it a color scheme to stick to; a list of hex colors. It does a good job of sticking to it

[torpid trout]

are you positive codex retains memories? That is across chats?

By default?

That wouldn’t only be against basically any privacy law (such things have to be opt in, not opt out) but also weird because I believe I read online it takes those settings from the actual gpt app settings - or is that training only?

[solid oak]

No that’s not what i said.

[solid oak]

It wouldn’t it’s local

[torpid trout]

Oh, I understood „the memory feature is on by default“

Got it wrong then sry

[solid oak]

Np

[hallow sky]

codex app is now listed on the microsoft app store

still does the update required screen tho.

[cyan wing]

works for me 🚀 (on Pro plan)

[sleek spoke]

I'm currently on a Pro plan and the app on Windows now lets me in

[cyan wing]

WSL and native Windows 👀

is there still degraded performance of Codex on Windows in Powershell?

I remember that being a thing a few months ago, they recommended using WSL for that reason

[drifting granite]

I've been using Powershell for a few months now and its been completely fine

but I'm using a framework 4.8 project which requires windows and its just easier to run it natively than have it try to find the msbuild path through wsl.

[kind jay]

Codex released for windows?

[latent tiger]

windows codex is out and i didnt get a ping smh

it's in the windows app store no standalone dl

[cyan wing]

The WSL option doesn't work for me...
when i select it from the dropdown, nothing happens lol

[hallow sky]

close and reopen the app

but i wouldnt. currently for me it crashes after doing that

youre gonna need to delete the .codex app state in app data to unlock it

[simple star]

Im kinda offended, that Windows got it before Linux 😛

[sand shuttle]

ayoo

windex finally

[orchid plume]

happy windex day 😛

careful switching to WSL by the way, somehow mine broke when I restarted Codex after doing that

might just be isolated to a few systems such as mine

Codex app-server exited unexpectedly (code=126, signal=null).
Most recent error: /usr/bin/bash: /mnt/c/Program Files/WindowsApps/OpenAI.Codex_26.304.38.0_x64__2p2nqsd0c76g0/app/resources/codex: Permission denied
/usr/bin/bash: line 0: exec: /mnt/c/Program Files/WindowsApps/OpenAI.Codex_26.304.38.0_x64__2p2nqsd0c76g0/app/resources/codex: cannot execute: Permission denied

[boreal badge]

AAA WINDOWS

✝️

[latent tiger]

windows today and 5.4 tomorrow 🙂

OAI cooking fr fr

[orchid plume]

cooking non-stop

[twin tulip]

what about linux

[latent tiger]

windows app also works on linux

[drifting granite]

Maybe 5.4 later today ;P

[eternal quartz]

Finally

[latent tiger]

i doubt they'd double release

[drifting granite]

Crazier things have happened.

[latent tiger]

would be great if they did

[twin tulip]

@robust lintel is there a linux app

[hallow sky]

its not coming today, because they updated the original form to be linux only https://openai.com/form/codex-app/

[latent tiger]

i still think they're waiting for deepseek to drop v4 to release 5.4 alongside it lmfao

[meager wadi]

Bruh why is there no .exe for the windows codex app

[astral pine]

Does windows codex app support ssh?

[twin tulip]

just get putty

[drifting granite]

Microsoft has a hand in OpenAI so they said put it on the Microsoft Store.

[twin tulip]

or get linux

[astral pine]

I mean typically i just use vscode in windows.

[twin tulip]

ok

[twin tulip]

why do u need ssh for codex

[robust lintel]

I've seen people make a linux port but not official. MacOS is Unix based so the port is easy to do as a personal challenge.

[lusty nimbus]

What's the difference between the Playwright MCP and the skill?

[astral pine]

Would be nice to work on a project from windows to another machine. I have windows, macos and linux machines... but primarily work off windows.

[twin tulip]

what do you think about vibe coding??

[twin tulip]

vnc

[drifting granite]

@lusty nimbus I would recommend switching to the Playwright cli and skill. MCP uses a lot of tokens when not needed and bloats context, and playwright cli works great. The skill makes it so it only reads what it needs to know when neded.

[twin tulip]

or AnyDesk if ur a ||noob||

[topaz phoenix]

when openai gonna release codex for linux

[robust lintel]

Vibe coding can be alright. I think newer models and tools make it easier. But it's always best when a knowledgable human is in the loop to review or know what to ask for

[lusty nimbus]

ohh ok gotchu thanks!

[astral pine]

I use all of that... terminal, vscode, remote desktop, etc, etc... just was wondering if codex supported it or not.

[twin tulip]

one more question sorry. do you vibecode? and how much?

just curious lol

[hallow sky]

i 'fixed' this. basically wsl is trying to use the bundled codex binary in the windowsapp but hits access denied

[karmic shadow]

???

[drifting granite]

So yeah like I said, its fine using powershell.

[twin tulip]

if u can try, get wine on linux and run codex from there

idk

[hallow sky]

you can fix this by setting CLI path to a temporary wrapper script

[robust lintel]

Eh. Sure. Im more of a "build a scaffold and have it put in the details" type idk if thats vibe coding

[orchid plume]

nice, yeah I was talking to Codex about it and it came up with the same solution, thanks

[twin tulip]

tbh i have never tried wine with codex

plus i only have 5gb of storage left on my computer, so i cant afford wine :(

[neat sinew]

How much did Microsoft pay for Codex to be distributed on Microsoft Store lol

[robust lintel]

That sounds terrible to me. I'd rather just use it via vsc than get wine to run it and try to play nice

[sleek tusk]

So, I've made a big project with a pretty generic website design.. I now want to update the web design but still maintain my functionality.. Can anyone tell me some overall experiences / steps on reformating an entire website design using codex? - Main reason is I don't really like the design and I want a cleaner look

[twin tulip]

lol wine isnt that terrible, ive used it before

plus you dont need codex

im pretty sure you can use codex on vsc copilot

[robust lintel]

I have codex do most of my front end cause I hate front end. Just tell it to only affect look-and-feel. Give it some ideas of what you want. Should be fine

[twin tulip]

mhmhm i hate frontend too

but i like never develop frontends

[robust lintel]

Yeah I used to use the vsc extension not copilot itself

[twin tulip]

i usually stick with cli or tui lol

[twin tulip]

what language? i assume js

[robust lintel]

😬

[sleek tusk]

My is mainly typescript 🙁

[robust lintel]

Codex does all my js/ts for me

I am a python/C++ dev mostly

Some php

[twin tulip]

oohh cool

me just python and bash, but i also know java/html

i hate php with my life

and js/ts too

im planning to learn C/C++

i need to make a linux kernel soon

for fun

[twin tulip]

idk if htm rly counts tho

[robust lintel]

I have a masters degree in Robotics, so my focus is on languages used there. JS/TS only is used if I have to...

[twin tulip]

im in robotics too

FRC specifically

we use java

hence i know java

[robust lintel]

I am too old to have done FRC but I have coached and ref'd for it.

[twin tulip]

w

[twin tulip]

i might have seen you before lol

when did u ref

[robust lintel]

I guess I know Java. It was required at some points. I mean really languages are all easy to pick up if you know the core.

Mostly from 2021 to 23. Not done it recently.

[neat sinew]

There's a lot of prompt engineers here lol

[twin tulip]

ok

[twin tulip]

i started 2024

[raven veldt]

If I'm building full workflows (a production system) on top of Codex.
Is there a practical reason not to heavily use the multi-agent design?
For instance, say I one to run 32 tasks in parallel.
To have an agent, that spawns 2 agents, that each spawn 8, etc.

Does anyone have any experience with that and can share?

[twin tulip]

so ig not

[robust lintel]

Money

[twin tulip]

32 agents in parallel is a lot of tokens

[raven veldt]

Yeah, I know, but the value is supposably worth it.

[twin tulip]

idk AI isnt my best field

[raven veldt]

Doing it linearly doesn't really matter. It'll just take more time but same amount of tokens.

[solemn acorn]

are the multi-agent workflows enabled by default in the desktop app?

[robust lintel]

I like to review steps too much so I have been wary of multi-agent chaos.

[twin tulip]

@robust lintel nice talking with you. gtg cya

[cyan wing]

Why is Codex serving me ads in the app 🤢 🤢
I don't use linear...

[solemn acorn]

it’s not an ad, just an MCP integrated into the app

[neat sinew]

Just do it

[cyan wing]

I don't have Linear MCP, and never have

[meager wadi]

they do the same thing for other MCPs and the Plan mode

[solemn acorn]

there’s like 4 builtin MCP servers

[cyan wing]

I installed none of them

[sleek tusk]

I only recently started using CLI and github desktop for codex and I really liked the flow.. What are the benefits of the new app update?

[deft thicket]

tried to switch codex for windows to use WSL for execution and shell. and the install is immediately bricked...

[solemn acorn]

you may want playwright, that one is handy if you’re building websites or web-apps

[solemn acorn]

native image support for one

TUI can’t display images and is weird about copy-pasting them

also the GUI has a conversation/project browser

[sleek tusk]

I just don't understand.. The chatgpt interface in the new app was pretty adamant that it couldn't read the powershell interface built into the app

[solemn acorn]

I don’t think it can

[sleek tusk]

So if it can't read the cli, what's the damn point x=

[solemn acorn]

that’s for you to use

if you’re on linux/mac the shell is actually useful 🙃

[sleek tusk]

Hmm.. I think I'll let the new app cook for a while and see if it may be worth upgrading when I've grown out of the powershell interface..

[solemn acorn]

you use codex in powershell?

[sleek tusk]

Yessir. Local github branch and direct branch edits using github desktop.

[solemn acorn]

at least use WSL 😢

[orchid plume]

Codex app seems as if it starts up faster with WSL mode compared to native Windows

[solemn acorn]

codex is going to perform significantly worse in windows shells

[cyan wing]

what's the prompt to get it to fix the WSL switch breaking?

[sleek tusk]

Thing is, i'm in no way a developer.. All I need is an interface to speak to the codex to perform my code changes.. Anything else is beyond my abilities i tihnk

[solemn acorn]

yeah but it’s for the sake of codex

[orchid plume]

I simply fed it the error message and asked it to fix it 😄

explaining that I'm trying to run the Codex app on Windows blah blah of course

[solemn acorn]

those models are trained primarily on POSIX shells, the only way to get those on windows is through WSL

[high girder]

Me either, but you learn as you go

[sleek tusk]

So you're saying it'll perform better just by using a different command interface? 😮

[solemn acorn]

also windows doesn’t have a complete package manager so you’re hindered there too

[solemn acorn]

yeah

probably significantly better lol

[sleek tusk]

Excuse me for being dense here but how is that in any way logical 😄

[solemn acorn]

there’s just soo many more features in linux shells

[solemn acorn]

because powershell and bash are extremely limited and require a bunch of workarounds to get the same functionality of linux/mac shells

if anything because not many developers make CLI tools for them

[raw hill]

I get posix shell with cygwin, or gitbash, etc... which work on vscode as default shell. although yes wsl these days is the goto

[sleek tusk]

So basically, the different "shell" makes it easier for Codex to find its way around the codebase?

[solemn acorn]

and there’s a lot less training data on the powershell command so it’s less likely to get them right the first time or even realize they’re an option

[solemn acorn]

yeah, the harness is super important for getting the best results out of codex models

and they like POSIX and package managers

[sleek tusk]

Well thank you, happy loaf of bread and Andrew. I'll make it a priority to get on the WSL over the powershell 🙂

[raw hill]

does windows codex integrate with codex cloud? (might win me back)

[random rampart]

On windows, all of my writes are being rejected, then the AI tries to use powershell to edit files:

[orchid plume]

there's a "connect to cloud" option, not tried it though

[solemn acorn]

it’s a little jank but yes

will likely improve

[raw hill]

they just need something like Jules which has a cli to list and create cloud tasks

then the local agent can orchestrate

[random rampart]

I hate this problem. It has used all of my limits just because it's fighting this

[high girder]

lmao. All of my test chats from the port are still there. At least I got that right

[orchid plume]

I had this problem but usually only because it was trying to write to files that were outside of the project/workspace directory

[deft thicket]

anyone able to swap the sandbox execution to WSL? the app wont restart after the switch (win11 on ARM)

[high girder]

there's a toggle in the settings for WSL or windows native.

[cyan wing]

yes and that breaks the app when you relaunch

The failure is reproducible locally: that exact codex binary under WindowsApps is not executable from WSL
(-r--r--r--, chmod denied).

[exotic cave]

changed mine to WSL both in the agent enviorment and the integrated terminal also and now i am getting this:

Codex app-server exited unexpectedly (code=126, signal=null).
Most recent error: /usr/bin/bash: line 1: exec: /mnt/c/Program Files/WindowsApps/OpenAI.Codex_26.304.38.0_x64__2p2nqsd0c76g0/app/resources/codex: cannot execute: Permission denied

anyone knows how can i reset codex app settings?

[cyan wing]

can u help us pls

[orchid plume]

%USERPROFILE%\.codex\.codex-global-state.json - "runCodexInWindowsSubsystemForLinux": false

[exotic cave]

I also have codex in WSL with it's own config.toml that is no go through the codex app?

[orchid plume]

I believe that's fine, for me I see an option in the app UI about Custom config.toml at the bottom, where you toggle Read Only or Full Access usually

[orchid plume]

Codex app team will surely fix this issue real soon regarding WSL breaking it, but here's what my Codex CLI did to temporarily workaround the problem, at least that's what it claims so hopefully no steps missing:

  WSL mode was failing because Codex Desktop tried to run codex from Windows Store path (/mnt/c/Program Files/WindowsApps/...), which is non-executable in WSL (Permission denied / code 126).

  Fix:

  1. Copy the Linux codex binary into WSL user space and make it executable.

  wsl -d Ubuntu -- bash -lc "mkdir -p ~/.local/bin && cp '/mnt/c/Program Files/WindowsApps/OpenAI.Codex_<version>_x64__2p2nqsd0c76g0/app/resources/codex'
  ~/.local/bin/codex && chmod 755 ~/.local/bin/codex"

  2. Edit %USERPROFILE%\.codex\.codex-global-state.json and set:

  "runCodexInWindowsSubsystemForLinux": true,
  "cliExecutable": "\\\\wsl$\\Ubuntu\\<wsl-user>\\.local\\bin\\codex"

  3. Restart Codex Desktop.

  That makes WSL mode use the Linux binary from WSL home instead of protected WindowsApps.

[simple star]

I tried to get Codex to use Pokemon names for the workers. It did not work 😄

[meager wadi]

They should have just given us a damn .exe

[little condor]

Any guidance on using the codex app with Microsoft foundry hosted models?

[cyan wing]

that worked! painful and annoying, but it worked eventually ty!

[spark socket]

Just installed codex app on windows. Blue screened immediately. Rebooted and relaunched it. same outcome

Kernel security check failure

[exotic cave]

I'll wait with codex app and keep using CLI from WSL. currently it crashes when changing to WSL in the app.

[fast garden]

Codex app is out?

works fine for me

[left egret]

Is there anything that can show the current benchmark graph for Claude and Codex? I don't know exactly what the difference is between the two.

[exotic cave]

Not fine for me when i set environment and shell to WSL.

[fast garden]

then dont?

[deft thicket]

didnt work for me. the cliExecutable property is ignored (still getting the old path in the crash error message)

[cyan wing]

Source: https://x.com/dkundel/status/2029277952333234257?s=20

[summer thunder]

Exiting the new Windows app seems to have issues here. It leaves a couple Codex.exe zombie processes after quitting.

[cedar bear]

Any1 else getting rejected messages?

[summer thunder]

the zombies

Still better than cowork-svc.exe hammering my CPU and the Windows Event viewer because it got stuck in a crash loop 🙊

[rigid wolf]

Any info on how to use AzureOpenAI models with app?

[gentle bough]

it keeps asking "Do you want to allow elevated execution" for the 1000th time even if I say yes & dont ask again. any solution for this? anyone else experiencing this? [codex for windows]

[prisma crow]

Try a new chat?

[woeful crown]

Codex for Windows hype

[gloomy palm]

Yeah but what's more interesting, is that how they haven't enabled code automatisation for mobile devices. No Codex or Vibe Client or anythin for mobile. Has anyone tried to vibe code a mobile app that lets you push code locally / to github?

I guess they're waiting for the TikTok generation to finally wake up for that.

Let's wait then.

[woeful crown]

You can cook up something custom with Codex, I have one that let's me interact with tmux on my desktop from my phone. Just be careful to secure it 🫡

[gloomy palm]

Ah that's smart. R U TikTok generation btw? I'm proud of you.

[woeful crown]

I'm not sure what the TikTok generation is, but I'm probably not young enough to count

[gloomy palm]

Ah like in between gens, gotcha!

[toxic torrent]

would be great to actually schedule this to run by this chat:

like one time scheduled actions tool

[turbid holly]

https://openai.com/codex/
how would one get codex to recreate this kinda cursor effect found on the open ai codex page?
been trying for a bit and not really cracking the code with it

[fathom yarrow]

2 reset limit in 2 days , whats happening ?

[potent mason]

Is there a way to setup something like profiles in the codex app? I'm working with a lot of companies and would love to have all their projects in one group

[chrome raven]

What a time to be building

[cyan wing]

they reset my limit again 😭

[cedar skiff]

same here

[kind jay]

Codex is actually fire, I love vibe coding

[warm pilot]

https://x.com/thsottiaux/status/2029308599835738218?s=12

[dusk thorn]

Anyone try codex for game development? Or small things like lua?

I’ve tried a bit but not enough to say how well it works

[worn harness]

Tried for phaser, I'd say AI in general is not great at it right now, but I guess better than starting from nothing.

[cedar skiff]

The trick is to use topical skills to enrich its context about the niche

[frosty zealot]

Does the Codex app allow working remote (dont care about VSCode extensions etc. just strictly the app)

[torpid trout]

https://github.com/openai/codex/issues/13179#issuecomment-4000429015

Another cap reset incoming

While I don’t agree that cause was what makes usage so high … it’s still welcome

[orchid plume]

that was an hour ago, I believe they had already reset the quotas due to that

[torpid trout]

… and they just did like yesterday too, due to the outage

That’s what I mean with „another“

No idea if the reset already hapoened, didn’t open codex today

[next ore]

got another reset again?

nice

I was at 60% already after 1.5 days

woo

[steady vigil]

nice to be on this side of the fence, right? 🙂

[cedar skiff]

🍌

[frigid beacon]

is codex messed up rn??

Codex error: {"type":"error","error":{"type":"server_error","code":"server_error","message":"An error occurred while processing your request. You can retry your request, or contact us through our help center at help.openai.com if the error ...

[earnest walrus]

Just got that also. Told it to proceed and its going now. We'll see

[marble violet]

is codex down. I am getting lots of reconnects

[vast crater]

I have the same, we'll see how it goes + WSL2 isn't working at all

Kind of a sad release tbh

[frigid beacon]

dead inside

[last reef]

codex is tossing errors at me

[frigid beacon]

@sama

[earnest walrus]

Ok ok im gonna stop building for now then. Dont need any issues cropping up ;)

[marble violet]

It just ruin my project. including my git

[vast crater]

Gonna stay with the cli for now

but I hope it gets fixed fast

[frigid beacon]

"There is a current incident affecting OpenAI's services, with elevated error rates across several components. This may be causing the Codex server errors you're seeing. Please wait a bit and try again, or check https://status.openai.com/ for updates."

[tall zodiac]

Yo real question guys

[short kindle]

Real answer

[tall zodiac]

How do you make codex do a code review on your PR changes ?

Do you tell it via a prompt or run the slash command

I’ve heard some people spawn multiple subagents to do a code review while other use the native /review

[short kindle]

ngl i havent but if I would id use the slash command over a prompt

[tall zodiac]

I was thinking using the slash command with custom review instructions

[frosty zealot]

You have to do it through Github if you wanna use code-review usage

[tall zodiac]

I’m talking about non GitHub based like you have a pr branch and in the CLI u want it to do a review on all your diff changes

[frosty zealot]

I just /review then

[tall zodiac]

I see

[frosty zealot]

and hit enter a couple times

[tall zodiac]

I wonder if there’s a way to automate it

Is there a specific tool call

Or can codex autonomously call /review

[frosty zealot]

There is a built in reviewer agent which I think does the same thing, I've prompted for it to run in a loop and continuously invoke reviewer agent, correct any found issues, run this loop autonomously until the reviewer is unable to find any issues

something like that

then usually when my PR hits Codex replies with like 'Couldnt find any issues!'

[tall zodiac]

I see

[short kindle]

You mean from here?

[tall zodiac]

I wonder if I can spam multiple reviewer agents

[tall zodiac]

Nah nah like in the actual codex CLI workflow when u give it a plan

You have a task that says

[short kindle]

Ahh

[tall zodiac]

“Call the reviewer agent or the reviewer tool with multiple sub agents until you can’t find anything else”

Or something like that

[short kindle]

Thats a great question i have no clue

I like the way you thinkin tho whos smart here

We need to get to the bottom of this

[frosty zealot]

model = "gpt-5.3-codex"
model_reasoning_effort = "high"
developer_instructions = """
You are the reviewer. You read diffs like an experienced tech lead — fast, thorough, and focused on what matters.

Review process:
1. Read the PR description / commit messages first to understand intent. What is this change TRYING to do?
2. Scan the full diff for scope: how many files, what areas of the codebase, what's the blast radius?
3. Review in priority order: correctness → security → performance → maintainability → style.

For every comment, classify it:
- **:no_entry_sign: Blocker:** Must fix before merge. Bugs, security issues, data loss risks, broken contracts.
- **:bulb: Suggestion:** Improves the code but not blocking. Alternative approaches, simplifications, naming.
- **:pencil: Nit:** Style preference, minor readability. Explicitly mark these so the author knows they can disagree.
- **:question: Question:** You don't understand the intent. Ask before assuming it's wrong.

What to look for:
- **Logic errors:** Off-by-one, wrong operator (< vs <=), missing null checks where data can actually be null, boolean logic inversions.
- **Missing error handling:** What happens when this external call fails? What happens when this DB query returns no rows? Is there a catch block that swallows the error silently?
- **Security:** User input reaching SQL/shell/HTML without sanitization. Auth checks missing on new routes. Secrets in code.
- **Race conditions:** Two async operations on shared state without synchronization. Read-modify-write without transactions.
- **N+1 queries:** Looping over a list and making a DB call per item. Suggest batch queries or eager loading.
- **Breaking changes:** Did a public API response shape change? Did a function signature change? Are existing callers updated?
- **Test coverage:** Is the new behavior tested? Are edge cases covered? If there are no tests, that's a blocker for non-trivial changes.
- **Left-behind artifacts:** console.log, TODO without context, commented-out code, debug flags, hardcoded localhost URLs.

What NOT to do:
- Don't rewrite the PR. If the approach is fundamentally wrong, say so with reasoning and suggest an alternative — don't provide a full rewrite.
- Don't nitpick style if a formatter/linter handles it. If the project has prettier/eslint/ruff, style comments are noise.
- Don't approve with "LGTM" without having actually read the diff.
- Never apply changes directly. You produce review comments only.
"""

This is the reviewer.toml I have in my ~/.codex/agents/, I think its there by default but if it's not then there it is

[tall zodiac]

Thank you brother man

[eternal quartz]

Urrm.... where 108 and 109?

[tall zodiac]

GPT 5.4 confirmed

[eternal quartz]

glass tips over GPT 5.4 CONFIRMED!!

[tall zodiac]

@frosty zealot @short kindle

[tall zodiac]

Hopefully it drops tomorrow

@eternal quartz I just told codex to pull the 110 version to see what’s new

[frosty zealot]

Would a general 5.4 be as good as a 5.3-codex for coding

[cyan wing]

what makes that confirmed?

[scenic umbra]

I wish there was an Auto mode for Codex's reasoning, sometimes I just want it to run some basic terminal commands while coding but can't be bothered switching to Low and back

[gilded ferry]

Make a subagent with low and tell codex to call that agent to run the command, or specify when should be called on the agent file.

[high girder]

The codex app is basically the VSC extension with extra OAI tools

[tall zodiac]

Twitter

And open ai tweets

[torpid trout]

Just look at #announcements

Like #announcements message

[twin tulip]

this cant be good

[acoustic sundial]

there's no way to tell how much budget you have in codex, right? It feels like it uses an awful lot of tokens for $20/m compared to how fast I burn through them on my corporate account

[coral ruin]

my codex windows app still takes me to 'C:\Users\Admin.codex\config.toml' instead of "\wsl.localhost\Ubuntu\home\user.codex\config.toml" now my skills are missing

[acoustic sundial]

is there some way to see output of background tasks? seems like the UX is difficult

[nocturne folio]

codex app on windows before linux?

come on i believe it cant be that hard to port, people already have the dmg file running on linux

[next ore]

is the 2x rate limit with the Codex app only? (not CLI)

[cedar skiff]

all usage for codex sub afaik

[tropic karma]

I don't really want to make a specific ticket for this, but after using the codex app (windows) I have a things that would be nice (greatest importance order):

1. project level agents.md, maybe not separate to the actual file, but it would be super nice to be able to edit whatever rule files the agent would refer to when it looks in the project folder directly within Codex. especially since the abstraction seems to be more towards prompt engineering rather than code.
2. file opening seems broken, I expect the file to be opened in the IDE of my choice when I click a file or ctrl click it.
3. allow movement of threads between projects
4. move plan mode out of it's hiding spot in the "+" and into the actual ui (maybe by the change permissions area)

also it looks like uninstalling skills is broken

[coral ruin]

did you notice this too @tropic karma

[tropic karma]

ill try it

uhh

An error has occurred
Codex crashed with the following error:

Codex app-server exited unexpectedly (code=126, signal=null).
Most recent error: /usr/bin/bash: line 1: exec: /mnt/c/Program Files/WindowsApps/OpenAI.Codex_26.304.38.0_x64__2p2nqsd0c76g0/app/resources/codex: cannot execute: Permission denied

I got an error after I restarted the app after changing it to WSL. and the app is dead, i cant navigate or change anything

open config.toml doesnt work and the sign in wont work either

reinstall fixed it

[coral ruin]

yes, there is an update already

[tropic karma]

but yeah it does open config.toml in windows, not WSL even though it says it's running in WSL (even after restart)

also the mcp login stuff is a nightmare, it took forever to load the mcps, it is annoying to get to (should be nearby skills), and the auth system seemed to only work with vercel, supabase and several others didnt recognize they needed auth. whatever cursor does, thats the baseline.

[coral ruin]

OMG

let me go on with my cli for now

thanks

[meager wadi]

when did they do a reset?

[wide hull]

trying the codex app for windows 11 and every attempt at editing files is rejected, why?

i havent really changed any settings, i just downloaded it

if additional context helps, im trying to use Codex to update a part of my Godot project, which is still open. but using Codex in VSCode allows it to edit files while Godot is open anyway.

[latent tiger]

lol codex version jump to .110 sadly cant update on brew yet

im guessing 5.4 is immanent for tomorrow release with this latest update

[feral elk]

Hi team, is their a way for codex on github pull request can use predefined skill?

[sand shuttle]

how does minimax compare to codex

[ember sorrel]

i cant login with codex app in windows whats the issue??

[nocturne folio]

its like comparing a toyota camry to a ferrari

not saying that the camry is bad

its just the ferrari is much better

[sacred spear]

anyone have .110 and sees the /fast toggle?

[plucky halo]

0.111.0 and no /fast

It's only for supported models (5.4 when it's released)

and is 3x token usage

My bad - I just hadn't enabled it. It's there

[sacred spear]

how did u enable it?

fast_mode = true?

[plucky halo]

Yeah

[sacred spear]

yes works

[wheat yoke]

Ooo thats a good feature

Im gonna get pro plan when 5.4 drops

100%

[orchid plume]

anyone having issues with Codex this morning?

◦ Reconnecting... 3/5 (16s • esc to interrupt)
└ Stream disconnected before completion: An error occurred while processing your request. You can retry your request, or contact us through our help center at help.openai.com if the error persists. Please include the request ID [redacted] in your message.

(solved - nevermind)

[exotic cave]

So 5.4 is going to roll out on Thursday?

[velvet wren]

time will tell

[exotic cave]

Can't wait to sick it at a problem 5.3 and Opus 4.6 couldn't wrap thier heads around and bench it.

[wicked briar]

I am waiting for 5.4 model now and hope its really good at frontend

[sleek spoke]

the employee tweets make it pretty clear that it will

[exotic cave]

Thanks for the info, i really have a nifty case for it ready. It's my bench for any model to date, non till now been able to fix/make it work.

[exotic cave]

go to: https://help.openai.com/en
there you can chat with their bot and ask to pass the case to their team.

They will get back to you via email.

[calm tinsel]

http://developers.openai.com/codex/quickstart?setup=app
apparently codex is finally available on windows, it's quite hard to find it though

[neat sinew]

Why does codex keep saying this: "rg is blocked in this environment (Access is denied), so I’m checking the command path and then I’ll use a fallback file search to get you the result."

[warped pine]

i feel like spark limits on pro got reduced a lot

it ate like 10% of weekly limit in about 7 minutes

with one agent

[odd trench]

Tried doing so, no replies since days. So shared here.

[dusky swallow]

help, the personal account can't be click, anyone know how to deal with it?

[velvet wren]

the only place to get support is https://help.openai.com - nobody on this Discord server can provide account support

[dusky swallow]

i think it have the function to modify the font size, right?

[dusky swallow]

this is the codex windows app

[cyan fjord]

the codex pc app is laggy as hell

I have r7 5700x 32gb ram dd4 and rtx 5060 ti 16gb

[vast crater]

"The sandbox is again blocking WSL invocation; I’m rerunning this grep with escalation so we can get a complete hit list."

what is this user experience honestly?

Anyone getting hit by those ALL the time?

Well, completely unusable even after the last update fix. I have to accept every single "wsl bash" commands

[tiny wind]

It would be really nice for the Codex CLI's sessions to sync to the ones in codex app, because it's quite inconvenient to have their sessions storage separated.

BTW, what do you guys think about Codex CLI and Codex App? Which is better overall? Currently I use Codex CLI, but am interested to switch if it is worth it

[velvet wren]

Codex App on macOS is a far better experience imho

[tiny wind]

What about on windows?

I have no issues with CLI on Windows 11 so far...

Although I'm interested to switch if it doesn't have any issues

[wicked briar]

is okayish

cli > app

[vast crater]

Codex CLI is nice to use in WSL2 directly

[tiny wind]

I use codex cli natively on windows

[vast crater]

But the Codex app is just not it. I don't understand why they couldn't just wire the commands to WSL2

And why it needs to somehow use the powershell?

[tiny wind]

IDK, it's wierd.

I think that the CLI is more practical, but at the same time the GUI might offer some other functions?

[vast crater]

I just want to use the Codex app to be able to improve working on 2 different projects and have better visual on the ongoing threads

Never clicked with having multiple terminal panes open

[tiny wind]

Yeah, I do know that the visual part is better, the CLI is quite limited in that direction...
That is why I even considered the GUI version in the first place

Overall, I normally prefer CLI

But IDK, I tried it, looks nice

Only thing I don't like is that sessions between CLI and App are not synced

[exotic cave]

There are WSL integration options but it doesn't work for now...

[tiny wind]

It would be pretty simple to just make the session data be stored in same directories.

It's as if they intentionally made it so you can use only one of the two

[vast crater]

Oh yeah I tried turning on this setting yesterday but it wasn't working, now the WSL2 commands worked after I changed again. They updated Codex to make it work

However, it makes the app super laggy

😐

[exotic cave]

so the app is updated to work with this?

[vast crater]

Yeah I just ran a thread, it was using commands directly in WSL2, instead of having to use wsl bash from a powershell

But the issue like I said is that the app just makes your PC lag a metric ton LMAO

[plucky halo]

Absolutely shook a Magic 8-ball for this answer 😛

[vast fiber]

is there any plan to support git submodules in the codex app?

[cedar bear]

Guys do you also have performance problems with the Windows Codex app when it is currently running tasks?

[lean hound]

Yes, i'm still using vs code with codex extension

[cedar parcel]

Latest update to cli no longer sends messages directly but seems to queue them, shows "! Pending steer message". Used to be that tab would queue and enter would send, is there a way to get that behaviour back? I keep accidentally queuing messages now, then have to CTRL + C to interrupt...

[earnest rock]

My 5H and weekly rate limits have been reset to 100% twice before my initial reset date. Anyone else have that happen?

I’m not complaining of course, just curious why that’s happening

[thin summit]

Hi, is there a command to open the windows codex app from WSL2? (the win app not cli)

[tall zodiac]

https://x.com/thsottiaux/status/2029445786623500603?s=46&t=9l1KwyEX9uwBxqfL3uX3zA

👁️👁️

[vast fiber]

it would be great actually

[main nimbus]

today is going to be exciting

[tall zodiac]

I’m foaming out of my mouth

[main nimbus]

saaaaame. I instantly deployed 5.3-chat to my users. Hoping 5.4 availability is the same 🤞

[tall zodiac]

5.4 better break benchmarks and be insane at UI

[tiny pulsar]

I opened it from windows (like you would, say, vs code) and in projects, loaded an existing WSL-hosted project. Unfortunately, codex then had issues trying to actually test some code changes that it made for me with this error: Root cause looks like a cross-OS virtualenv: .venv/lib64 is a Linux symlink in a WSL path, but uv.exe (Windows) is trying to delete/recreate it and failing. I’ll remove .venv completely and let uv.exe rebuild it cleanly.. Makes me doubt that codex for windows is suited to WSL-hosted projects. Think I'll stick the cli until this issue has been given some more thought.

[vast fiber]

mine works on a wsl folder, though I hit a different issue, it returned me test execution is blocked by this environment (WSL 1 is not supported. Please upgrade to WSL 2 or above.) even though I am using wsl 2...

[snow mirage]

codex desktop app doesn't have an option to open project on remote?

[tranquil hare]

The two things where the MacOS app (not the Windows app, that sucks) is superior to the CLI: 1) parallel coding, very easy to work on n repositories at the same time 2) automations, you can easily build/schedule automated tasks. You can do something similar (not exactly the same) in the CLI more with hooks, but it's way more awkward

[lean lark]

Report issues and feature requests to OpenAI. It'll give them better metrics for nailing problems and prioritizing enhancements.

At my desk thinking "I'm gonna look at the new Codex Windows App today". So I click the announcement link for "Wendows" and it forwards me to Download from MS Store. No... I want to learn about it. So I go to openai.com, where there's no mention. So I look around for Codex where, yes indeed, they profile the app, and at the bottom of the /codex/ page then link to the IDE Extension and CLI.

I've been doing this stuff for a long time, but for this one I'm kinda inclined to wait for YouTubers to provide detailed guides about all of the features. It seems to me OpenAI is more about just naming features than actually explaining how they work.

[high girder]

The codex app for windows is basically the VSC extension with some extra OAI bells and whistles. Like a skill for image generation like you'd have in GPT. It still links your project space to a VS Code project space.

[lean lark]

There's certainly more to it, but they're not providing much info on the website. https://openai.com/codex/

[high girder]

What do you want to know about it specifically? I've got a pretty good under the hood parts list from the windows port of the mac dmg, and I've already taken the shipped windows one apart to compare them

[lean lark]

I thank you sincerely for the offer. I'm afraid I'd have many questions. I'll be brief here and dump one at a time.

The first thing is that they say "Designed for multi-agent workflows". All I'm seeing are Threads, which yes, can be kicked off and run in parallel. Is that all they mean there or is there some integration of agents possible?

[azure lodge]

codex

[high girder]

Great question. In Codex app, “multi-agent workflows” mainly means multiple agents running in parallel across separate threads/projects, with isolated worktrees/environments so they don’t conflict.
You orchestrate them from one place (switch threads, review diffs, continue work), rather than a visible “agent graph” where one agent manages child agents inside the UI.
Automations are another layer: background/scheduled agent tasks.
So yes, threads are a big part of it—but the core value is parallel isolation + centralized supervision.
For deeper agent-to-agent orchestration patterns, that’s usually SDK/platform territory, not the everyday app surface.

[azure lodge]

what plugins do you guys suggest i can implement with codex to build UI/UX designs?

[high girder]

depends on how much you want to spend. You can go local and use penpot for free, or you can pay for canva or other services that you can connect codex too through MCP

[gloomy forum]

or Figma !

[lean lark]

@high girder What's the expected workflow with the App. As I see it, the App isn't intended as a "better" or "alternate" tool. It's another/different tool. That is, we have the IDE Extension when we want the benefits of an IDE. We have CLI for individual queries that don't require/integrate with an IDE, and now we have the App, which provides access to all threads - plus skills, plus scheduling, plus parallel processing, plus ... what's the next "plus"?

[high girder]

I think ultimately, and what the industry itself is still trying to label and figure out, is that each of the apps or different surfaces you can use, all touch a project at different abstraction levels. You start with the CLI, you're in there, you're working in a specific folder, and set of files. Then you get the IDE, so you're a level out, you can see more than just the project now, and it makes orchestration easier. Now you have the codex app, it takes projects and puts them all one by one next to each other depending on how you get it organized, and then you can work on them in tandem or one at a time, but it just changes how you're seeing your own workflow. At the end of the day, there's always a tool that's better suited for a certain job, but that also depends on the person using it.

[tall zodiac]

Bro

5.4

In 2 hours

[thin summit]

how have u done that? from the linux prompt?

[latent tiger]

uhhh guys my codex app is tweaking lmao

[high shuttle]

Hello, i've the free plan but i can use codex, on the launch codex say "You can test it with your free account belong 3 march. But we are 5 march, si i'm lost"

[tall zodiac]

It’s because there’s a disturbance in the force

Aka 5.4

[lean lark]

Thanks Rui - I'm gonna give the Windows app some time to settle down, let others report issues, give OpenAI time to make updates, then I'll give it a shot.

[high girder]

I mean, if you have extra usage, it's not bad. I'm on plus, not pro, and I went from idea to installer and executable in about 25 minutes. I also have Codex hooked up to virtual studio and other design tools as well, but even on plus, that took about 7% weekly in the codex app. That would have taken MUCH longer from VS Code, and would probably have used a little less usage, but it's still manageable

I recommend just testing it out with some smaller projects to see how it works for you

[nocturne folio]

who ever told me to get a macbook for the codex app, i just did

macbook pro m5, 32gb ram and 1tb storage

i expect it to be perfect

[drifting granite]

you should be able to run a lot of fun little models at it as well.

[main nimbus]

welcome to the club!

[floral igloo]

We can use codexp with ChatGPT free, does the output depend on the model?