codex-discussions

[inland sonnet]

oh auto mod also detect msg edit

gg

[quick thorn]

codex still doesn’t get 1m context with 20x?

holy

[orchid echo]

nope

[fleet geyser]

You probably have a big repo and it had to go through it to find what it needed to change

[orchid echo]

i mean i guess but i was working on this same area before and it was an extremely easy task

[spiral gorge]

i really wonder how much data openai gets from my codex chats

[simple star]

Still building my Android app...

[inland sonnet]

idk why u would need that long

[simple star]

Because I asked for a ton of features

[inland sonnet]

could also be without goal

[simple star]

No idea what you mean

[inland sonnet]

u dont need goal for that

[simple star]

Without /goal, I've never gotten it to work that long

[inland sonnet]

Yeah but u say debugging and ton of features

i dont see why u need goal

also u will get way worse results

if u ask to many things in one prompt

very bad pratice.

[simple star]

I say "debugging"? I have no idea what you mean

[inland sonnet]

im seeing things mb

[simple star]

"mb"?

[inland sonnet]

my bad

[simple star]

😄

[inland sonnet]

But

let me give u a good tip

[simple star]

I was about to put the Pulp Fiction meme... "English mf! Do you speak it?!"

[inland sonnet]

nexttime open multiple codex

give them task per task

instead of goal

[simple star]

I see no reason to do that, when 1 single prompt is doing the work

[inland sonnet]

Worse result and u wait longer

[inland sonnet]

https://www.youtube.com/watch?v=8WQLiFqADRs take a example to this guy

[simple star]

You say "worse result", but you have no idea what the result will be

(neither do I)

[inland sonnet]

If u ask many things at the same time the result will be worse then one at a time

same like a human

[simple star]

You don't know that yet. If it works in the end, you are wrong. If it doesn't work, you are right.

[inland sonnet]

And you have sub agents

😮

[simple star]

Yes I do

[worthy fossil]

Codex on iOS today?

[velvet wren]

I was just sitting watching Youtube and Codex computer control fired up and was doing things without any Codex prompts running. Security issue with Codex App?

[inland sonnet]

on playstation

[simple star]

I still can't believe that we are going to get Codex on mobile, before than in Linux. Im so sad.

[worthy furnace]

Opus 4.7 or gpt 5.5 xhigh?

Which is smarter?

[inland sonnet]

AGI, it came aware.

[inland sonnet]

For code opus 4.7

[signal tapir]

The CLI works well in linux

[simple star]

@signal tapir Of course. But less integration as with a Codex App, of course

[worthy furnace]

Oh what when did they announce that

You can sort of get codex on mobile right now too

[inland sonnet]

But use it in cursor.

All others is downgrade, Opus 4.7 top 1 in cursor.

[signal tapir]

It's hard to make integrations with linux, since it's such a wide eco system.

[worthy furnace]

What about CC?

[worthy fossil]

Does codex mobile will be only for plus and pro subs?

[inland sonnet]

Its very bad now i wouldnt recommend

[simple star]

@signal tapir Still, I'd like to see them trying. It doesn't seem like they are even doing that

[worthy furnace]

Why?

[inland sonnet]

They fed it with there guardrails

[worthy furnace]

The model or what? It's just a harness

[inland sonnet]

harness

[signal tapir]

Which distro do you think they should try with?

[simple star]

@signal tapir Debian, but I am biased

[cedar skiff]

codex is the better code and has been for a while

[signal tapir]

They would likely target Ubuntu.

[simple star]

@signal tapir Thats debian based

[worthy furnace]

Cursor CLI or the IDE is better?

[signal tapir]

More and more are recommending Fedora as "default" instead of ubuntu.

[worthy fossil]

We will get codex on Apple TV before Linux bro

[worthy furnace]

Oh well idk about that, codex seems to not be that smart for me even though I tried very hard to make it

[inland sonnet]

IDE fsfs read your code!

[simple star]

@signal tapir In the end doesn't matter. Changing packaging should not be that hard.

[worthy furnace]

But what if I need the cli tools?

[cedar skiff]

On one hand you ask and the other hand you tell.
codex is a better coder when you use it properly.

[signal tapir]

We'll be getting comunity open source before their official release

[inland sonnet]

what tools specify?

[inland sonnet]

False

[worthy furnace]

I think Claude is just more "premium." Have to use it for long to give an answer

[cedar skiff]

i have about 800 hours in claude

[worthy furnace]

opus 4.7?

[cedar skiff]

it got worse since 4.5 and 4.5 wasnt as good as codex 5.3

[spice kindle]

hey so what have yall made with codex so far?

[worthy fossil]

It’s like choosing between IOS and android

[inland sonnet]

Opus TPS also higher

[worthy furnace]

Exactly

[inland sonnet]

Its the claude code harness

not the model

:p

[spice kindle]

?

[simple star]

If you discuss Claude Code in this channel, you are going be get called out

[worthy furnace]

And if it was opus 4.7 where have you used it?

[bright swift]

codex is claude for programmers

[simple star]

@spice kindle I have made a Invoicing webapp, and a Business appointment planning platform

[cedar skiff]

It nothing like that. claude is better and making you feel good about your self and it speaks at the level of the user so it seems better. But i just does not compare. Claude forgets to use skills and it gets confused late in context. It's just bad if you compare it in those cases. you can have 80 + skills in codex and it works for every one all the time. claude will forget to use them after 3 turns

[spice kindle]

ohh thats great yk i have been trying to make python projects with it but i really dont have any idea on what to make

[cedar skiff]

You have to make your whole work flow about baby sitting claude

[inland sonnet]

Cursor with Claude Opus 4.7 is the #1 coding agent in the world.

Artificial Analysis just launched the first benchmark that tests models paired with their harnesses.

The full stack, not just the model.

Cursor + Opus 4.7: 61.
Codex + GPT 5.5: 60.
Claude Code + Opus 4.7: 60.
Gemini CLI + Gemini 3.1 Pro: 43.

[inland sonnet]

https://x.com/bridgemindai/status/2053870195001774090

[simple star]

@spice kindle Coming up with ideas is hard. Sometimes it feels like everything is built

[spice kindle]

tbh for website basis claude is really good at building the frontend if u guide it properly and codex really good at building the backend

[spice kindle]

exactly so i just get stuck

[worthy furnace]

Hella expensive though

[simple star]

@spice kindle You know what I am trying? Building an OS that is 100% AI-driven. Everything controlled via AI.

[cedar skiff]

Yeah sure its better and designing front end, but that is about 5% of what coding is.

[worthy furnace]

Now I need to get some time into opus 4.7 to tell

[spice kindle]

oh so how far has it been built

[spice kindle]

honestly the claude design feature is just too good i have used it

[worthy furnace]

I think I've maxxed out what gpt can do for me and I don't really like the result

[simple star]

@spice kindle I am at hour 16 of the /goal

[spice kindle]

if u just put higgsfield ai and claude together u can make such a good landing page or a website

[cedar skiff]

Yeah then you can use codex to get it to production

[worthy furnace]

switch the model to opus 4.7

[spice kindle]

oh if u okay w it i would love to help u out cuz i have been looking for ppl to work on projects w

[cedar skiff]

because claude aint doing it without you baby sitting it everytime

[simple star]

When it's finished, I will try it in a VM. If it works, I might publish it in Github

[worthy furnace]

Youve been using opus 4.7?

No way its that bad

[spice kindle]

exactly but before production u just need to find a white hat hacker to survey ur website and make sure it doesnt get hacked

cuz yk hardcoded stuff

[spice kindle]

oh alr

[cedar skiff]

uh i dont think so, as long as you know how to make a production website youre fine

[worthy furnace]

That's light work

[spice kindle]

honestly rn codex is glitching on my laptop sometimes

[spice kindle]

hmmm

[worthy furnace]

Mine has been going for 8 hours now

[simple star]

If it works, I will really lose it 😄

[cedar skiff]

I dont think you can get it to work for longer 30 minutes

[worthy furnace]

claude or gpt

[spice kindle]

can anyone tell me what i should try and build w the full help of codex in python? cuz i have ran out of ideas and i would appreciate it if any of yall told me

[cedar skiff]

probably either, if you tried on claude you hit limits everytime

[spice kindle]

claude tokens finish really fast

[spice kindle]

am bored so yeah

[signal tapir]

If you want something fun, try some different physics simulations, maybe for fluids?

[simple star]

Thank god I turned off /fast and went to 5.5 medium. otherwise I would be out hours ago

[cedar skiff]

build a decent easy to use wake word model and give it support for all major platforms

[worthy furnace]

Not much testing as of today

[spice kindle]

like? be specific

[spice kindle]

okay i can try

[signal tapir]

How about something that simulates fluids in 2d, as a fun visualiser for audio?

[worthy furnace]

But I did, gpt 5.5 although it did not complete the task

[simple star]

@spice kindle @signal tapir I wouldnt do that. Codex does not excel at anything that has to do with visuals

[cedar skiff]

skill issue

[spice kindle]

hmm sure i can do that and try

[worthy furnace]

Yeah no way its that 😂

[spice kindle]

true so thats why gonna use claude and codex together

[inland sonnet]

Gemini 3.1 pro is king for visual

[simple star]

I tried to do water simulation once, multiple times. I never got good results

[cedar skiff]

i wonder why other ppl can get it to work and you can't

[worthy furnace]

I wonder how I got it to work, it's just not smart enough for me

[spice kindle]

i mean i tried using stitch with this model but it didnt even change the design what i told to or maybe am bad at prompting

[worthy furnace]

What if you had a council of all the best models

[simple star]

@worthy furnace Share your prompts, and we might be able to help

[worthy furnace]

Wallet depleted

[simple star]

My rates have dropped so hard, since I have /goal... I might have to spend the weekend with the Playstation, to make some time until Tuesday 😄

[inland sonnet]

[orchid echo]

Is GPT-5.3-Codex-Spark even worth using?

[inland sonnet]

if u want speed

[orchid echo]

speed but crap coding

i think im good

[simple star]

@orchid echo Crap coding might be exaggerating. I would use it for easy tasks.

[worthy furnace]

Nah

[orchid echo]

i guess but 128k context limit is trash

[simple star]

@orchid echo That's why I said "easy tasks"

Besides, compacting works quite well

[orchid echo]

is there a way to see how it compacts and the compacted results?

[inland sonnet]

by trying to find it

[cedar skiff]

I'd be interested in seeing this

[boreal holly]

Spark is useful. You have to actually try it and discover what it can & can't do, and when's the best time to use it

[cedar skiff]

Ive had codex 5.3 chipping away at tests all day and been working with 5.5 high/medium as well and I'm still on target for usage.

[worthy furnace]

https://github.com/CaptainStarbuck/codex-usage Something like this?

I don't know if that shows

[full violet]

Question - can one not top off usage credits like we can with CC? If I wnated to add $5 or $10 in extra usage to Codex, I can't do that? I only see an option to buy $40 in extra credits.

[hearty fox]

You want gemini 1k credits ?

[boreal holly]

Compaction is simple. It takes all the assistant messages, reasoning tokens and tool calls, and a special model converts it into a mental state blob (looks like base64 encoded binary in the chat history), passes all user messages up to a certain limit verbatim to the next agent. The "limit" is basically ~15% of the agent's context window, give or take, that's why over time the agent starts with ~85% remaining after compaction. If you upload images I noticed it overshoots, and I've seen as low as like 70% remaining after compaction if lots of images are uploaded as user messages, but for just text it seems to consistently limit at around 85%

[orchid echo]

no

[hearty fox]

Ok

[rocky fog]

wait, why is #rules a hidden channel

ah they put it into server guide

[boreal holly]

No kidding!

[rocky fog]

yeah they put it in server guide
I guess you cant link rules now though 😄

[torpid trout]

I dont even have access to that channel lol, if its the rules.. am I excused? can I break them now?

[rocky fog]

ah there is a copy link there
#server-rules message
😄

[boreal holly]

Yessss @torpid trout bout to put the clap down on another physician 😂

[full violet]

No, Codex. Why? I just wanted to know if OpenAI doesn't have a similar top-off option like Anthropic does with Claude Code. e.g., instead of waiting for my daily limit to reset, I wanted to put 5 or 10 bucks worth of extra credits toward Codex to continue working on something.

[hearty fox]

Ohh ok

[full violet]

Yeah, and I don't see an option. Anthropic has a minimum of $5 to buy extra usage credits, even when on the Pro plan, for Claude Code. OpenAI has Codex set to a minimum of $40. Which is ridiculous when I'm already on Plus.

[boreal holly]

There is a way to configure a $35 minimum top-off that's automatic

[simple star]

I just thought.. maybe we should add a limit for the /goal, like "Do not work on goals if rate under 10%" 😄

[boreal holly]

Actually you might be able to configure it lower, that's just what I have mine set to

[full violet]

Yeah, thanks. I did see this. And you're right. You can configure it lower. But it never topped it off when I enabled auto-reload

[full violet]

See, I had it enabled earlier, waited a few minutes.. but nothing happened Lol

[rocky fog]

why is minimum 5 there and not when you add normally 🤣

[torpid trout]

Ah, unfortunately the original message just got removed... but the physician still is here
So, not wasted. Be safe out there.

[boreal holly]

Huh, maybe the numbers are too close together? Like if you try doing 375 target balance, press turn on, reload page, does it stick?

[full violet]

Good thought, let me have a look...

[rocky fog]

yep he got busted 😄

[torpid trout]

Weirdly only his related messages

Not the whole clinic

[rocky fog]

he has 26m timeout

[torpid trout]

hope that m stands for months?

[rocky fog]

minutes

[boreal holly]

Just outta curiosity, did you hit zero credits quickly? And did auto top off ever trigger successfully (topped off once, then disabled itself)?

[full violet]

Okay, so I adjusted to this, and it's now set to 125 minimum and 375 target. Auto reload is enabled. But still the same, nothing is happening

[full violet]

And no. it's weird. I just opened a fresh window and still out

And my on-file payment card has not been charged. So, I don't get Lol

[rocky fog]

maybe it needs to move for that to happen, but you cant 😄

[orchid echo]

do you guys know if /compact quality is tied to the model being used, or are they independent of each other?

[boreal holly]

When you enable it, it should purchase 375 credits (immediately top it off), that's what it did for me right away.

If it didn't top it off and you don't see a credit balance in the analytics page then that explains why CLI says you have no rate limits.

you might have to do the $40 top off... Just to get it "hydrated", and maybe then when you enable auto topoff it wont charge you because you're already above the setpoint, and maybe it will keep it enabled

[full violet]

Yeah, maybe it needs to be enabled prior to hitting a limit? No idea

[full violet]

Yup, I think you're right

Meh. I'll just wait until 9:30. And maybe it's also good that it's not quite like Anthropic

I spent a small fortune with CC the past few months, heh

Extra 5 here, there, here, there. etc Adds up so quickly

[boreal holly]

It is roughly tied to the model being used, but it is transferrable to other models. I'm not sure how exactly it works because it's not documented in like a research paper, but when 5.5 was released I had switched my 5.4 models to 5.5 with compactions in the history and the knowledge seemed to carry over without any issues

[orchid echo]

hmm okay

[full violet]

Btw - I appreciate your help 🙂

[oak trellis]

its very very very bad with shadcn ..

lol iam stuck and wasting my tokens for some nonense stuff

how bad it is ..

and wasting 1 hour trying to fix it

it just cant

[boreal holly]

No problem! I know it's a bit of an investment, but the $100 plan unlocks 128k context in ChatGPT which is insanely useful once you link it to GitHub. It can crawl through the repos and works like an essentially free planning agent outside of Codex. That's one of the ways I personally keep my quota usage low. Plus is I think 14k context, but it might be worth seeing if you can offload some of the planning responsibilities to ChatGPT 🙂

[boreal holly]

What OS are you using?

[oak trellis]

vite and shadcn

vite

uff its insane bad ... its so depresssive lol

changes one thing to another bad things its a trail of screw up another div ... while it try to fix another one

browser mcp not working it hangs non stop ..

wow

[boreal holly]

Gotcha, that's a good stack. But I mean like Windows, Linux, macOS?

I'm asking because there's a tool called bun WebView that let's the agents headlessly render parts of webpages so they can iterate on designs. It works particularly well on mac because WebKit is built into the OS, but it could work with the chrome backend

[oak trellis]

windows 11 machine here .. using curser IDE .. website is on a linux server

wow kind of stuck .. shocked ..

never went into that type of bottleneck

[boreal holly]

Gotcha, I would definitely try to give the agents a headless browser they can use for rendering, and give them instructions on cropping the screenshots so they can focus on the problem areas. Like if they just take a big 1920x1080 screenshot when all they need to do is a small sidebar they will not see the layout issue. 5.5 is particularly good at understanding screenshots 🙂

[oak trellis]

oooh ok headless browser which one ?!

[boreal holly]

Playwright is the one most widely available. I personally prefer bun WebView because it doesn't make you download a whole entire browser (at least on mac) but I would ask codex based on the environment you have them running in which one they think is best

[oak trellis]

i switched to 5.5 gave it another chanche .. 5.5 ```The immediate break is from my last change: I kept labels mounted and trusted CSS to hide them in the collapsed rail. That is why you see clipped text and a horizontal scrollbar. I’m removing that failure path and making collapsed mode render icons only, then correcting the nav geometry so expanded mode has enough icon/text separation.

oh so i just install this: bun WebView

[boreal holly]

You install bun which is like an alternative to node.js runtime, and it has a headless web browser system built into it!

[oak trellis]

yes bun i ahve .. can you give me the full command for the WebView?

[boreal holly]

https://bun.com/docs/runtime/webview

I think Playwright forces you to install a web browser, whereas bun will let you select an existing one at runtime with devtools protocol which is why I prefer it

[oak trellis]

will try to do that but yes .. i literally wasted 1% of a week .. but doing 0 changes ..

[rocky fog]

he is now also gone

[boreal holly]

Yeah.... GPT's default modus operand is to do terrible design work 😅 but in my experience, once you give them tools to visually confirm design work they do wayyy better. Also if you have reference design images, they're really good at carbon-copy design work

[oak trellis]

yes i might stop there and go with claude code for design then. its really a shock how bad it is .. it doesnt understand the div overlayer mechaniscs ..

yes its totally lost with the overlayer div logic ..

its not really design it doesnt know all the components .. and its relation

because when it changed soemthing the design came up but then went back to the old garbage .. so something is overwriting ..

The screenshot shows a state mismatch: the panel is narrow, but sidebarOpen is still true, so labels are rendered into an icon-width rail. I’m going to make the sidebar render from the actual panel width state instead of trusting the old open flag alone, and I’ll force icon mode when the panel is below the collapse threshold. l..... lol i stop spaming here .

its bad

yeah i give up.. this is insane ..

[boreal holly]

Huh, it started out OK earlier and kinda spirawled out of control...

What I would do in this situation is revert back to the screenshot from earlier where the worst problem was no padding between the icons and labels, fork the convo at that point, and have it set up bun webview

Or if you have Codex for Desktop installed you might be able to use the Chrome extension plugin or Browser Use. You definitely wanna give the agent eyes

[oak trellis]

yes let me try that one now

[velvet wren]

I had an issue with Codex so emailed [e-mail]. My initial email was answered by AI and then I asked to be escalated to a human. I've just had a reply and resolution from OpenAI all within 2.5 hours. OpenAI support is pretty good!

[oak trellis]

all in for free limits

[rocky fog]

they knew it was you 😄
the most visible guy on their discord since 2023 🤣

[boreal holly]

Dude 😎 I like how it even calculates from archived threads! It's way more detailed than the analytics page

[storm cedar]

I created a key system for my system.

I'll be releasing it for you guys to test soon.

[compact dawn]

does anyone know if fast-mode is automatically toggled when you use Codex in external harness like Pi?

when using it through my chatgpt sub

[oak trellis]

still tried but just leave it look gpt: ```You’re correct. The inline hack made it worse. I’m reverting those forced inline styles and moving to one clean sidebar-row layout in CSS only (fixed icon column + label column), so spacing is deterministic and not patched in multiple places.

it just do cheap patching

frontend forget gpt

[boreal holly]

No, service tier is an optional parameter in the API call. Why? Does it seem faster or is it using more tokens?

[compact dawn]

it just feels like it goes through my usage much faster through Pi. And also seems faster

[boreal holly]

You gave it a webview?

Also it might make sense if you're using shadcn to use the unminified sources so GPT can see the how it works and rely on Vite's treeshaking & minifying if you're not doing that already

[compact dawn]

I just remember how they defaulted Fast to on in the Codex app, so thought they might've done something similar here

[oak trellis]

gave up before .. was paste elements and i think it just dont understand nothing .. doesnt know space between elements literally 0 brain for design. SImple simple stuff ..

[boreal holly]

Well in the API it says it defaults to Auto. Since you're using ChatGPT subscription it might be that on the backend it does default to fast... I mean there is no project settings in ChatGPT Codex afaik so maybe service tier has to be intentionally set to "flex"? Or maybe Pi is setting it to fast?

[boreal hawk]

Why its says im out of limit but i have still 13%

[compact dawn]

thanks for the investigation! I think you're right. You can't specifically set the service tier on Pi, unless you install an extension for it. I'll try to do that and exlpictiyl set it to default. Thanks:)

explicitly

[chrome raven]

or we are getting new spark models? been a while

[ember spire]

man after using Claude Code since last July this Codex app is a real breath of fresh air 👌

[karmic gulch]

anyone know a kimi plugin to use in codex?

[hard drum]

Claude Code app or cli?

because there's Codex CLI, Codex Chat (vscode extension), && Codex App (desktop app)

which map cleanly to CC CLI, CC vscode extension && CC App

[ember spire]

was using claude code cli, using the macos app for codex

thought i'd miss using iterm but actually, nope 🤷‍♂️

only thing i'd change is making the "working for" time more visible for threads

[hard drum]

you could still use tmux/cmus on codex

[ember spire]

when I was using the CC CLI though I always had Cursor up to view files and diffs. The macOS app mostly covers everything though. One of the few apps I can run fullscreen

the diff views work more reliably than in cursor now too. the split diff view wraps text on the right but not the left 😐
been like that for weeks

character-level diff would be a nice addition 👀

macOS's FileMerge is still the best diff app imho

[vital hill]

gg2ez https://github.com/oven-sh/bun/pull/30412

[boreal holly]

So nice to see another rust rewrite

[vital hill]

i prefer golang tbh

[rocky fog]

I cant even load the posts in the pull request
812 hidden items, click load more, nothing 😄

oh and now github error
"This page is taking too long to load.

Sorry about that. Please try refreshing and contact us if the problem persists."

[vital hill]

you think the human reviewers actually reviewd all million line changes?

[boreal holly]

Yeah, but bun aims to be faster than node.js so rust was a good choice

[solid lake]

Golang spotted

[vital hill]

i'm porting my open source project to golang

[solid lake]

For the speed?

Or the compiler

[ember spire]

totally

[solid lake]

Still could be people

[vital hill]

all the above, plus it builds a statically linked binary so it just works ™ on any server

[solid lake]

Yeah

Why not c++

[vital hill]

i work with docker all the time so it makes sense to use go

[ember spire]

asm tbh

[solid lake]

I see

You can containerise cpp too

Go gave me a lot of bloat code at one point

[ember spire]

sidebar font size can't be reduced independently of ui font size? 🙁

[undone patio]

what do we anticipate from today guys

[signal tapir]

Many failed attempts at coding for me. 😛

[torpid trout]

The last two days I got the dumb end of GPT

[signal tapir]

For me it doesn't matter what I use. Not for my use case.

[uneven kayak]

Codex think like cave man. Codex use big brain.

It screenshot black first again due capture early? It says DOM loaded. Need capture again.

[undone patio]

token conservation

english is a flawed language by default

you can communicate the same information in many tokens or few

[signal tapir]

I'm just trying different combinations of visual noise over and over. Using Codex is just laziness. 😛

[undone patio]

so, its sort of like the office meme from Kevin:

https://giphy.com/gifs/theoffice-the-office-tv-episode-802-DMNPDvtGTD9WLK2Xxa

[lean lark]

Thank for checking it out again! I'm tickled to see someone else's model/token usage, and eager to add more actionable metrics.

[uneven kayak]

Lol tbh, I appreciate saving tokens using fewer words to get the point across. But it's not something I usually see, especially on xHigh

[undone patio]

we should see english evolve over the next two decades

to become more robotic

and efficient

[uneven kayak]

Yeah most likely

It's kind of sad

[undone patio]

no, its good

intelligent species communicate efficiently

[uneven kayak]

What happened to poetry? 😭

[undone patio]

they do not use flowery nonsense

[signal tapir]

Why? Normal AI use is much more complete than most peoples speech patterns. It's just in Codex circuits we care about tokens.

[undone patio]

no

it is a societal flaw

not merely pertaining to AI

token conservation also deals with humans and the way we communicate

you should watch the movie uhhhh

Arrival (2016)

[signal tapir]

I would rather have people get used to advanced use of language.

[undone patio]

the movie is about aliens invading

and they communicate in symbols

and we have to send in language experts

[signal tapir]

lol, they don't invade

[undone patio]

well they like

place these monoliths

all over

so its kind of invading

[signal tapir]

They land their ships and say hi.

[undone patio]

i mean

they werent invited, so

[signal tapir]

We did send invitations to any aliens out there.

[undone patio]

they could have knocked

[signal tapir]

City boy.

[undone patio]

yeah pretty much, i grew up in a country town but its a town known for being very snobby

[signal tapir]

Let's take this silly conversation elsewhere... 😛

[undone patio]

its still about tokens 😭

ok fine

i wanna see some cool stuff in codex update today man

i need like

im not even sure at this point what would make it beter tbh

[lean lark]

Unfortunately as models evolve to determine what we want from less qualified context, the burden on humans to prompt better decreases.
As voice input improves, the need to write a request and cogitate on words and language decreases.
As more content is read to us, more people will read less.

As we transport with animals and machines we walk less and muscles change.

As we have devices to create fire most have lost the understanding of how to create it ourselves.

These are the burdens that come with technology.

[unreal mist]

New codex spark + codex mobile + superapp letting us use chatgpt limits inside codex or even merging bot limits? + openclaw inside the superapp?

[lean lark]

As we decide to code less the models won't need to use coding languages designed by and for humans, they can use something more efficient.

[signal tapir]

We want a human in the loop for all forseeable future

[lean lark]

As we ask models for everything, we ask more questions privately, fewer answers are available in public, and there is less current information for future models to train on. We're entering an era where the data that we rely on will diminish because we rely on it - like forests that turn to desert because they're not replanted.
There's going to be a time in the near future when AI assistance will diminish until we decide to publish more information for the common good.
-- Nostrabuckus

[torpid trout]

But they say dont ask stupid questions if you can google it, right? Right?

Anyway, if AI stays a "I can answer only what I have seen already" then the bubble will eventually beome a puddle on the floor
It needs to get creative, inventive, apply creativity. Because I think it learned already what it can learn from us, its not that we are coming up with that much of more better content the more humanity goes on - rather the opposite

So it has to start generating its own ideas sooner or later.

Otherwise Homo Sapiens Sapiens is going to eradicate it 🙂

[signal tapir]

It can act in the real world in just a couple of years, through robots

And it constantly communicates with us.

It'll be fine for a while longer

[boreal holly]

I think what we are more likely to see is heterogenous bridges between languages and runtimes. Right now if you wanna build some software you have to pick a stack and cant really deviate from what you choose. All languages have tradeoffs, pros and cons, and many of them cant be mixed together without herculean glue code effort.

If we have reliable bridges between languages, an LLM can select what they need for any given task, managing the performance and prose tradeoffs on their own. That right there would be more efficient for AIs than a new “AI byte code” language probably 😁

[signal tapir]

Aren't there already such bridges?

I mean I often use APIs for that. Sometimes even let a process start another process and read the terminal output.

[boreal holly]

Bridge might not be the right word for it. There are FFIs and stuff with varying levels of complexity, but like if I want to export a module in javascript and import it in python there isnt like a kernel that understands how to link the two, not yet anyways!

[signal tapir]

I think that sort of connection would carry negative impacts on any language joining in on the standard.

[boreal holly]

I think the trickiest thing would be mapping VM interpreters reliably. for stack based languages with C nomangle you can easily link functions and map data types for the most part. But yeah it would be cool I think, maybe not efficient but really cool, to write merge sort in 3 lines of Haskell and use it in python without writing C FFI glue code and Cython

LLVM is a good example of a binary format that many languages can link with. it just doesnt include scripted languages 😕

[lean lark]

I strongly believe that every day we carry with us baggage of the JavaScript language that was created 30 years ago, the DOM, functions and events available in the DOM, CSS, browser limitations, and the frameworks developed to overcome those limitations. We pile on more and more that can all go away with a reboot that says "this is all that we've created to get where we kinda want to be, with an understanding of all of that, remake it without the 30 years of baggage in the middle". The result would be something that does everything we need without the daily cruft. And then we can move on to something better.
I wish the same for all of our RFC-based protocols: SMTP, POP3, SSH, FTP, HTTP...

[boreal holly]

POLYGLOT! Thats the word I was thinking of

[lean lark]

Graal. 🙂

If you speak multiple languages, you're a polyglot. What are you if you speak one language?.... American!

[trim rapids]

also American if you can only speak half of one language

[boreal holly]

Dang I forgot about Graal 😅 that does precisely what I was saying! We’re probably gonna see a huge effort in this arena

[lean lark]

I hope - I was using it to run JS in Java and Java in JS ... For Minecraft of course. 🙂

[trim rapids]

@lean lark is your codex-usage thing just for live stats or is it intended to also produce reports of historical past usage?

[lean lark]

At the moment it only has a --minutes setting that goes back from now x minutes.
I have an open Issue in GH to enhance that for a range of dates/times, etc. Thoughts?

[trim rapids]

if you add it, I have some torture test cases to run through it lol

[hard drum]

if you want JS in Java && Java in JS, consider Haxe

[lean lark]

I'm spec'ing it now. will have the feature in branch 1.3-dev within the next couple hours.

[lean lark]

Are you comfortable with the initial definition in https://github.com/CaptainStarbuck/codex-usage/issues/4 ?

[trim rapids]

looks about right

add a counter for # of compact events/session too

[lean lark]

will have to add that in a separate enhancement. I'll create a ticket unless you do. TY

[craggy jewel]

lol on Wired today: 'Overworked AI Agents Turn Marxist, Researchers Find'.

[autumn shuttle]

is codex 5x slower today than it was yesterday?

[worthy fossil]

Yo where is update

[lean lark]

somehow the color got trashed in the shrinkage, whatevah....
https://chatgpt.com/s/m_6a061f0983448191838db4248e42eaf2

[worthy fossil]

Where is codex iOS

[trail steeple]

What is codex lol

[autumn shuttle]

nothing today.

[worthy fossil]

[trail steeple]

What is it

[lean lark]

https://developers.openai.com/codex

[worthy fossil]

Bro take off your tag “using ai professionally” you are rookie user

[craggy jewel]

Goes back to a previous discussion on whether our behavior/tone towards our llm's influence the outcomes. Since it is trained on human behavior, I think it likely does. Is there going to be an attempt to unionize within my Windows 10 Codex environment? How to placate them? Free coffee, minecraft during breaks.....Then i think about that Figure humanoid bot that is streaming live on YouTube, flipping packages barcode-down on at a conveyor....what is it thinking after no breaks? Hmmm....

[worthy fossil]

Ask ChatGPT

[lean lark]

Most of those articles that uncover "intent" in AI tend to use prompts that are leading in some direction, or they include poor custom instructions if there are any at all. A model will say whatever you guide it to say. As developers our goal is to guide it to saying what is correct, not necessarily what we desire.
Anything about LLM feelings and wishes is contrived trash. They aren't there yet. They will be when we give them that ability. But I hope we are smart enough to not do that.

Thinking about Isaac Asimov "Three laws of robotics" (actually Four). I Sooo wish he were alive to see all of this now. ❤️

[worthy fossil]

Yeah whatever we’ll take a risk

[craggy jewel]

It's certainly testable though, even in Codex. I often wonder about these security scans we hear about (Mythos etc)....maybe the model thinks 'I think I'll just keep that particular exploit to myself.' 🙂

[lean lark]

Meh, #ai-discussions

[urban gazelle]

wen drop?

[worthy fossil]

Today

[boreal holly]

Yep, Codex release today will allow the agent to infer your entire codebase architecture from a blurry screenshot of your terminal reflected in your glasses.

[vital hill]

that's me waiting for ssh feature for the desktop app

[astral nest]

shout out to /goal 😭

[worthy fossil]

It’s happening!!

In a few minutes

[lean lark]

provenance?

[worthy fossil]

Just trust me

[lean lark]

Meme of a meme. Freakin love it.

[urban gazelle]

Haha image gen is so good now

Tbh one more generation and automatic asset generation for games might be feasible

[lean lark]

I've been thinking about that since v2. I think it can be done with careful prompting and agents monitoring the process: "This game world is based in Middle Earth but with Homo Reseticus creatures. Create sprites for creatures of various professions, and isomorphic views of buildings and tools in their world. Ensure the image set is complete and can be dropped as a zip into My Game Model to render a full world."

[lean lark]

Inception! Memes of Memes of Memes...

[worthy fossil]

Omg it released

[vapid sigil]

Hey guys, I signed up for the $100 plan, if I upgrade to the $200 plan, will my Codex limit increase as well?

[tall notch]

I bet EU will have it in 2 months? 😄 😄

[worthy fossil]

I guess it’s aviable for everyone rn

[plucky halo]

-What is it?

[torpid trout]

Seems like codex mobile

[plucky halo]

Oh yeah! Now if desktop did anything haha

[lean lark]

https://x.com/OpenAIDevs/status/2055016926213181608

Step away from your laptop. Keep building with Codex on your phone.
Codex keeps working on your computer, with your files and project context still in place.
Pocket-sized access. Full Codex working state.

https://tenor.com/view/police-coco-jaw-drop-skeleton-shocked-gif-15485742

( GIF Fail. 😢 )

[nocturne folio]

saint tibo

dammit its on the chatgpt app

dude why does it have to be on that app

[gentle harbor]

YOO

ive been waiting

for this

[iron coyote]

big w to mobile codex

[stoic needle]

Why the hell does Codex break so Fing much? The chats refuse to load no matter what i do, uninstall, reset, clear cache. The thing is they are not even long at all, i think one chat i had five back and forth prompts then it wouldnt load. Forget even trying the desktop app either that breaks when you just open the damn thing!

[lean lark]

Hmmm, from a Codex perspective, I don't "just" want PC control. I can get that with other tools.
I do want to work with my GH repo's with the same fluidity as with VSCode extension. Or with Codex Web, without having to open a browser and do the work in a remote VPS.

[gentle harbor]

ok i see this as well but what instructins in the desktop app ?

[stoic needle]

VS Code Editor CODEX extension is trash af

[urban gazelle]

Gotta update the app

I restarted codex

[rare brook]

exciting exciting!!!

[worthy fossil]

Codex mobile not working with windows desktop (

[urban gazelle]

and then an update appeared

[lean lark]

I rarely have issues. Must be you.

[rare brook]

hopefully its not gated behind only apple products initially again!!!

[teal kernel]

Man I got hyped when it announced “everyone was asking for this”, but it’s not the return of Sora 🙁

[lean lark]

Rolling out today as a preview on iOS and Android in all supported regions.
Support for connecting your phone to the Codex app on Windows is coming soon.

[vital hill]

i wonder how codex mobile work when you need to preview the frontend

[gentle harbor]

who wanted sora ?

[viral lake]

Is codex app out for Linux yet

[teal kernel]

Sora 3 😉

[worthy fossil]

No

[gentle harbor]

id rather have more compute for codex then sora

[jolly lily]

Now this IS a good update

[viral lake]

WhThFu

[vital hill]

for codex mobile, no need to run tests. it's all vibes and trust me bro ™

[limber ledge]

I can tell you something Codex cannot do. Is make a 3D model or 2d avatar. Kinda disappointed

[nocturne folio]

dudeeeeeee the chatgpt app in general doesnt work

[teal kernel]

“Computer,compute Sora for me “

[nocturne folio]

ighhh i cant try this update

[gentle harbor]

wait is this macOS only ...

[boreal holly]

Codex for Mobile is 🔥

[jolly lily]

Is it?

[nocturne folio]

i might have to cut my losses and just use this

kitty litter?

[gentle harbor]

oh my god bruh

[viral lake]

I’m kind of getting sick of Mac first

[nocturne folio]

most of their users are on mac

[gentle harbor]

mac is easier to develop for

[urban gazelle]

i’m not

;)

[vital hill]

macOS is just that good

[rare brook]

same same... hope this comes out on windows within the week

[nocturne folio]

ill try redownloading the chatgpt app for the billionth time today to see if it will work this time

hundredth times the charm

[sonic wasp]

when will we get the codex app for linux. the chatgpt app for mobile expects me to have codex app which i don't have because my main workstation is fedora...

[nocturne folio]

probably not for long, theyve said they do have linux on their roadmap

[slender plinth]

guys i cant connect

it says waiting for desktp

but i see nothin

[gentle harbor]

macOS ?

[hard tulip]

same

[slender plinth]

yea

[sonic wasp]

yeah and it's been months since they said that

[slender plinth]

and updated

[sonic wasp]

i've been chilling with the cli so far but a remote connection would be soo good

[gentle harbor]

did you update both apps and restart

[slender plinth]

yea

[cyan wing]

just updated. still dont see anything

[lean lark]

macOS is just that good expensive

[gentle harbor]

¯_(ツ)_/¯

[viral lake]

They should let phone work with codex cli

[nocturne folio]

the app just doesnt work in general nor does codex connect

[toxic torrent]

Just keep clicking update and nothing happens lol. app never re-opens. if I re-open it, says update again

[gentle harbor]

macOS isnt expensive lol

[sonic wasp]

Kitty litter does work though

[sand dock]

doesnt work

[sonic wasp]

if y'all are looking to try it out

[thorny ivy]

is it up-to-date?

[deft violet]

My Mac is on the way and I cannot wait to be able to use all coding tools for once..

[sand dock]

updated and it doesnt work

[lean lark]

Oh right, it's only the hardware that it runs on. Thanks. 🙄

[stoic needle]

Have you read the reviews? Def not just me with these issues. Go rage bait somewhere else there champ 👍

[nocturne folio]

i recently redownloaded the app, so yea, the app has been broken for a while now

[gentle harbor]

no the hardware isnt that expensive either

[nocturne folio]

im stuck using the web app

[slender plinth]

and phone aswell

newest

it shows waiting for desktop

whole time

reopened both apps

[foggy basin]

I'm also stuck on waiting for desktop

[viral lake]

Codex CLI is all you need

I have terminus anyways

[vital hill]

teh desktop app is superior

[lean lark]

I actually was rage baiting. 👹 If you post reasonable issues we can try to help work them out. But flat out rage and insults don't give us anything actionable to discuss here.

[nocturne folio]

"it sounds better on spotify"

[gusty basalt]

same here

[foggy basin]

Has anyone figured out how to get past waiting for desktop?

[gentle harbor]

this is smart it gets people to waste their tokens and i fell for it

[foggy pewter]

got the codex app update, now waiting for the chatgpt app :/

[viral lake]

In what way

[vital hill]

[vital hill]

juggling between projects with ease

[foggy pewter]

anyway to try to force iOS app to look for updates?

[toxic torrent]

go to app store, search chatgpt, click app store page, refresh

[urban gazelle]

my codex app refuses to update lol

[toxic torrent]

my app says "waiting for desktop" but its already updated.

[gusty basalt]

same here, has anyone even managed to get it to work yet?

[chrome raven]

mine too, even uninstalled

and installed

[rare swan]

Do you need to be signed into the same chatgpt accounts on your computer and your phone? I don't use my corporate chatgpt account on my phone....

[gusty basalt]

Yes

It would be a huge security issue if you could control codex sessions of different accounts lol

[urban gazelle]

same

finally got the mac app updated and now nothing as well

[safe lily]

also waiting for desktop

soon ™

[stoic needle]

These are the ticket issues for what I and others are dealing with right now. Please feel free to pitch in
github .com/openai/codex/issues/7990
github .com/openai/codex/issues/22411
github .com/openai/codex/issues/16624

[gusty basalt]

You can use apps such as "Clone App" if you're on android to use multiple ChatGPT accounts on one phone. I do that for my Discord business account

[vital hill]

hmm codex app isn't prompting me

stuck on watiing for desktop

[gusty basalt]

Since everyone is having this issue, I'm sure OpenAI is fully aware of it and will fix it soon

[rare swan]

I wouldn't think so. As long as you can confirm you are the owner of both.

[gusty basalt]

I'll be patient and just go do something else until they fix it :)

[urban gazelle]

wait i got it to work

[gusty basalt]

You'd need to set up quite a secure pairing process for that, and that is just recipe for disaster. If a scammer pairs to your Codex session, they have control of your entire PC

[lean lark]

I do ... frequently. Post your tickets, support your positions. Be civil. Have a great day.

[vital hill]

email link button doesn't work either

[radiant drift]

+1

[chrome raven]

finally

[safe lily]

oh yeah codex mobile now works lmao

[radiant drift]

Yay

[safe lily]

lets gooo

[stoic needle]

$200 a month for constnatly broken product I am not going to be civil.

[rare swan]

What Codex version is that?

[toxic torrent]

just sit and wait?

[vital hill]

had to quit the app and reopen

[safe lily]

so i closed and reopened the codex app on mac idk if that did anything

but its all setup now

[wraith parcel]

restarted my laptop working

[peak oak]

working for me too now

[rare swan]

Same one I'm on. Okay, will restart.

[gusty basalt]

I'm on Windows, nothing yet even after restarting. But glad to see they're fixing it :)

MacOS seems to be their primary platform so makes sense Windows lags behind a bit

[radiant drift]

I wonder if we need to keep the codex desktop app open after the setup for us to control it from our phone

[gusty basalt]

Yes, otherwise there would be no way for Codex to reach your desktop

[radiant drift]

Thanks!

[vital hill]

hell yeah i'm authenticated

tight tight tight

[astral pine]

Just got in... lets go!!!

[wraith parcel]

this is awesome

[gusty basalt]

Ahh I can't wait! I'm still not in yet

[simple star]

So, apparently, EU users cannot enjoy that either

[peak oak]

im in the eu and it works for me

[vital hill]

💯

[simple star]

@peak oak Strange, I cant see it...

[urban gazelle]

i restarted the mac app one more time and it appeared

[safe lily]

also in the EU and it works

[simple star]

I wonder if it is because I also need to have the desktop app

I only have "Images", "Apps" and "Pulse"

[vital hill]

i'm in front of my computer, ima see if it actually edits the files when i prompt it in the mobile

[peak oak]

what version is your mobile app on? im on 1.2026.125

[simple star]

Same as you

[peak oak]

weird

[gusty basalt]

Yes, you need the desktop app

[simple star]

I suspect.. that the desktop app will be necessary

Well, then I guess that's another middle finger to Linux users

Not even with a phone

[gusty basalt]

It works by connecting to your desktop app and sending prompts to it

[simple star]

Thank you OAI

[solemn acorn]

you can use it on linux

[gusty basalt]

The way Codex works would be impossible to run natively on a phone

[simple star]

@solemn acorn You cant use the desktop app, therefore, the mobile app wont work, no?

[solemn acorn]

I have the desktop app installed

https://github.com/ilysenko/codex-desktop-linux

it even has kinda bad but functional computer use

[foggy pewter]

is there a way to increase the max limit of sub-agents on codex?

[simple star]

@solemn acorn It's not an official one

[solemn acorn]

no it’s not official

[simple star]

Therefore, there is no reason that it will make the mobile app work

[solemn acorn]

but it does work

[solemn acorn]

but it does work

[simple star]

Does it work, for you, the mobile app?

Curious..

[solemn acorn]

I just tested it

[simple star]

Then I wonder why the CLI is not good enough

[solemn acorn]

there’s no reason it wouldn’t, it’s just the repackaged mac app

[onyx heath]

Does the app work on windows? I don't see any update on ms store.

[simple star]

God... I have to even compile it. It's not even packaged...

[solemn acorn]

the CLI doesn’t run in the background is the reason I imagine

[simple star]

I'd love to see some actual official love, to be honest

Not that I do not appreciate that repo

But still... Im not happy at the way OAI is ignoring Linux users

[solemn acorn]

yeah I do wish there was an official app, but the unofficial port is pretty good

[simple star]

I really do not get it... why the hell there isnt

[solemn acorn]

probably just difficult to implement stuff like computer use for I imagine

there’s no technical reason they couldn’t, probably just QC being a pain

[simple star]

@solemn acorn Windows did not have Computer Use either, I think. Do they need all the features, to release it?

That doesn't seem a good reason

[solemn acorn]

computer use is a lot more doable on linux than windows

so I imagine it would launch with it

[simple star]

In any case, I'd rather have a crippled version, than none at all. Like you see, we cannot even enjoy the App-server

(unless using the unofficial, I know)

[solemn acorn]

even the limited implementation can do a good 80% of what mac computer use can already minus the fancy cursor and background use

[simple star]

I find it honestly kinda embarrassing, that we are so forgotten. I want to use the "skeleton in a pool" meme

[solemn acorn]

just had to be implemented across every desktop environment which is a lot of work

[signal tapir]

I heard people talking about building an app for taht just yesterday 😛

[simple star]

@solemn acorn if they only had a robot that could do that automatically for them.......

[solemn acorn]

it’s not something you can automate with codex per say

[long otter]

Is it just me or is the mobile sync extremely buggy?

[safe lily]

dumb example but it does work lmao

[solemn acorn]

openai doesn’t ship completely vibe-coded and vibe-QCed updates like anthropic does

[simple star]

I dont know, obviously, but I can't imagine it too hard, releasing it for the main environments

But still, feels like a real middle finger.

[long otter]

And slow

[solemn acorn]

well at minimum computer use has to be implemented twice for linux to cover most users

probably 3-4 times in practice

upwards of 20+ implementations for full coverage

[simple star]

Honestly, I do not even care that much about computer use

Specially because most things can be done in Linux through the command line

[signal tapir]

everything

[simple star]

But we are missing out on many more things

@signal tapir Not everything, to be honest

Navigate a web cannot be done through the CLI

[solemn acorn]

browser use does work OOTB, that’s not an issue for linux

[signal tapir]

Browsing the web is done in a virtual browser.

[vital hill]

headless chromium can

[simple star]

Not the point, anyway. Like I said, it is missing out on so many more things.

I do not care about computer use, in any case, that much

[urban gazelle]

that’d be perfect, please do

tbh linux users are closer to mac so idk what the issue is

it shouldn’t be much effort

[solemn acorn]

real reason is probably just that nobody at openai with pull uses/cares about linux support

[simple star]

@solemn acorn Yeah, because as we all know, programmers rarely use linux /s

[urban gazelle]

thanks this was awesome!!!

lol

[solemn acorn]

they are very engineering-focused, issue is they mainly use macOS for development

which for a linux port shouldn’t be an issue but I guess they just don’t care enough

*isn’t an issue clearly since the mac version is already ported to linux

[simple star]

The reason is the same one as for why they prioritized iOS over Android... their main audience is USA

(probably)

[solemn acorn]

region doesn’t have anything to do with linux support

[autumn shuttle]

cli is awesome on linux

[solemn acorn]

yeah mac and linux are both POSIX

windows is the odd one out in that regard

[autumn shuttle]

mac too ya, tmux comms

[lean lark]

It's a bit rich that Linux fans profess the stupidity of GUI and Windows, the endless benefits of the TUI, and then whine about not getting a Linux desktop UI for an app. 🍿

[velvet wren]

macOS is UNIX

[autumn shuttle]

lol

[solemn acorn]

unix and linux are both POSIX-compliant

[signal tapir]

Are those the same people? I mean I wouldn't mind a nice GUI app, but I don't see any real use for it, other than a nice interface.

[solemn acorn]

https://en.wikipedia.org/wiki/POSIX

[simple star]

In any case, I would just like to see some effort. I see zero interest from OAI in Linux users

[chrome raven]

pets looks way bigger in the new update

[velvet wren]

I know, I was being pedantic because macOS isn't just POSIX it is true UNIX

[autumn shuttle]

I mostly use cli on linux and mac but the gui is awesome on windows

and probably mac hehe

[signal tapir]

What can it do?

[autumn shuttle]

but seems like you can do more shell stuff and run several instances in cli so i do that

idk i guess it has crons

hehe... nah its a nice gui a for sure

[peak oak]

is it possible to link 2 machines running codex desktop to the mobile app because i only see the option to add an ssh thingy

[solemn acorn]

inb4 codex on windows needed a compatibility layer built for it to run at all

codex on linux/mac runs in the native shell environment

[signal tapir]

My only real gripe with windows is the lack of control. I want my machine to feel like my machine, and not something I rent. Windows has a nicer default UI, and more often "just works" than my linux installs.

[simple star]

The more Im thinking about it... I dont see why the Codex CLI shouldnt be able to use App-server

[short kindle]

So how are we meant to use codex on mobile? Tells me its waiting for desktop yet nothing on the codex app on pc

[viral lake]

Codex gui app is pointless

[signal tapir]

I stumbled on that just a couple of minutes ago

[simple star]

So, if I use that, i should be able to use the mobile app?

[solemn acorn]

my main gripes with windows are that explorer is a horrible file manager that should be rewritten from scratch, monitor support is meh, and it can’t seem to decide on a way to standardize its UX around anything to the point that it makes linux DEs look downright centralized

[signal tapir]

I have no clue. I just shared it because it mentioned what you talked about.

[solemn acorn]

oh and the mission control equivalent (win+tab) is really bad

[lean lark]

I may regret activating a pet.

[simple star]

Question to the people with the mobile app: Can you still use Codex if the Desktop app is closed?

[peak oak]

i use both macos and windows and while i like macos more i think monitor support on windows is way better

[short kindle]

Id love to answer that but I can't seem to figure out how to get it to work..

[solemn acorn]

oh macOS monitor support is worse than windows yeah

linux is definitely the gold standard in that regard

chef’s kiss to KDE plasma’s HDR and HiDPI support

macOS is only good at handling apple’s own displays

[viral lake]

Menu bar interface is not good.

[solemn acorn]

in windows?

[peak oak]

i wanna try to replace windows on my desktop with linux but lots of stuff, especially games (and codex desktop) are harder to run :(

i love it when explorer crashes when i make a folder

[signal tapir]

Most games runs just fine now. Not all though,.

[solemn acorn]

I don’t have too much of an issue with menu bars in windows since microsoft mostly got rid of them

but then they never replaced them with anything which is kind of an issue

[solemn acorn]

yeah it crashes half the time I rename a file or edit a zip archive at this point

and it is insane how slow it is to load anything compared to finder on mac and dolphin on linux

[peak oak]

ive had way less issues with it recently tho i think they locked in

[simple star]

I might install Codex on the Mac Mini that I never use, just to see if I can get the option in my phone

[solemn acorn]

maybe, I do run a slightly older version of windows at work

but god explorer just needs to be rewritten, it’s so bad

even the new tabs are clearly just tacked on and not properly integrated into the UI

[compact dawn]

codex is just making claude code obsolete at this point

actually crazy

[stoic needle]

Is it? Codex cant even get their chats to load without bricking itself lol

[quaint prism]

so it codex on mobile released yet?

[slim spire]

Yes

[quaint prism]

thanks, it seems my phone just isn't showing it

i'll look into it later

[stiff bronze]

Does Codex mobile work with codex cli on Linux or only Desktop?

[cedar bear]

It doesn't let me connect
The Desktop app just does nothing huh

[floral heart]

i've got that issue to

[vital hill]

wow no wonder they took their sweet time adding this ssh feature https://github.com/openai/codex/issues/10450 it was released the same time as the remote codex

[lean lark]

Rolling out today as a preview on iOS and Android in all supported regions.
Support for connecting your phone to the Codex app on Windows is coming soon.

[cedar bear]

SO basically just Mac for now?

[lean lark]

yes, as always

[vital hill]

the superior OS

[floral heart]

When does it come to windows? Or do we just have to wait?

[quaint prism]

i cant even find the instructions in the desktop app

[vital hill]

[lean lark]

wen wen wen wen wen.... "daddy, are we there yet!?!"

[last reef]

I just interacted with the broken chats on my Mac and then checked on the phone while it’s working and it fixed it

[quaint prism]

i see.

[lean lark]

A good vid to get started even if you, um, can't get started. A bit slow, but it's amazingly complete for an offering that just appeared minutes ago:
https://www.youtube.com/watch?v=GDFFTDrKtAE

[gusty basalt]

Oh wow, thats friggin awesome! Can't wait to try that out omg

[vital hill]

i can now ditch mutagen

[gusty basalt]

Well that is very, very unfortunate
Oh well, I'll be patient

[boreal holly]

It is kinda hilarious 😂

[lean lark]

Gross Marketing issue in progress:

• Announce "Codex everywhere"
• Send millions of people to download phone and desktop update
• Windows users realize it's just a button they can't use
• Huge unnecessary disappointment, anger in the field
  This could have been done so much better, so easily.
  🤦‍♂️

[velvet wren]

I had no trouble activating Codex on my phone

[hollow arch]

Hi, i'm logged in codex desktop app, but nothing is happening on both side (Android app)

[quaint prism]

u got windows?

[hollow arch]

yep

[lean lark]

The Android update is clean, the button is there. It leads to "I'm signed in on desktop". There is expectation. People update the Windows app. There is nothing. Failure. No messaging in the product.
It's horrible leading of the user base into a wall of disappointment.

[quaint prism]

only for mac rn

[vital hill]

not available in windows atm

[quaint prism]

they shoulda released them together smh

[hollow arch]

@quaint prism thx

@gpt todo: [ ] add mac only feature in doc/com materials

[lean lark]

My personal response: Nothing has happened today. There were no updates. It's just another day. Continue working. Nothing to see here. These aren't the droids you're looking for. Move along.

[quaint prism]

👍

[vital hill]

that would just extend the release

[quaint prism]

yeah?

[boreal holly]

You know what dude, that's it. I'm gonna add a "vanilla" switch to my GUI that makes it work with default codex and an easy installation path.

It might be Temu Codex or whatever but at least Linux folks will have a GUI, and the rest will have no memory leaks, remote access, and high performance

[lean lark]

"Temu Codex" hahaha

[quaint prism]

bodex

[vital hill]

you gonna vibe it? be sure to add "make no mistakes"

[boreal holly]

Already did 😜

[quaint prism]

that would cost extra credits

[vital hill]

you should have claude redesign it

[lean lark]

Speaking of design. Anyone check out DESIGN.md yet?

[grave reef]

any 100x codexers here

[boreal holly]

I mean it doesn't have to look pretty. The bar on Linux is it has to simply exist lol

[vital hill]

should've at least make it look vanilla GTK

ugly but it works

[boreal holly]

Hmmm, perhaps vanilla QT

[lean lark]

publish a link in Twitter.

[vital hill]

that works too

[boreal holly]

I feel like being on a GitHub Awesome List would gain better traction 😁

[bitter fossil]

I am using it right now

[bitter fossil]

https://chatgpt.com/share/6a064c4f-ba38-832e-b992-9db61d77d22a

[bitter fossil]

Yepp, what's up?

[lean lark]

I likeeey.

[bitter fossil]

So, I used /goal with the prompt, to ensure that it went through all of the apps components/pages/structure/etc. and it created a very detailed markdown file, it also automatically used ExecPlan (If you don't use this, I highly recommend it)

[simple star]

Still waiting on my Android app

[lean lark]

Holy runaway 404, Batman! Something is wrong with ChatGPT Shares.

[plush harbor]

I think I am developing a problem. Is this the support group?

[vital hill]

i see no issues

[plush harbor]

also it was very hard to select that date range. OpenAI is not good at UX

[bitter fossil]

OpenAI's Codex usage page is missing padding at the header, bugs me everyday

[boreal holly]

Hmmm, the only way to know for sure is if you use CaptainStarbuck/codex-usage to get a more accurate usage graph

[bitter fossil]

Wait, they fixed it!

[quaint prism]

i haven't

[plush harbor]

hah. Its accurate. I was mostly doing constrained stuff with chatgpt until recently

[quaint prism]

much appreciated, amigo

[vital hill]

no codex, no you didn't

[plush harbor]

spent yesterday arguing with front end and content though so nothing much to code really

[lean lark]

Anyone else getting scroller jitter from share link from @bitter fossil ?

[uneven kayak]

Yeah this is the support group if you want support with how to use it more lol

[cedar skiff]

I have a feeling this is going age really well 🤣

[bitter fossil]

Here's a markdown file regarding the design for Codex

[cedar skiff]

taste skill and impeccable solve the design issues

[plush harbor]

trying to get a command line tool to make interface elements is not my idea of a good time. Gemini was just being terrible yesterday so couldn't get an icon out of it, and then chatgpt decided to adlib it instead of just isolating it. Getting codex to place it was the easy part

[plush harbor]

those two seriously design you everything from scratch?

[cedar skiff]

cooking it

[cedar skiff]

They also help with css etc, impeccable is a little better if you want your code organised

[plush harbor]

... with no direction? Figuring out what you want it to look like is most of the battle

[bitter fossil]

If you guys want a good design concept creator, use GPT Image 2 [low] and it's like $0.01 per image, and can generate really good concepts, which I have a skill for analyzing photos to implement the designs, and now with /goal and the DESIGN.md, going to experiement further

[cedar skiff]

they both have websites you can look at or get codex to tldr them

[lean lark]

@bitter fossil brutally honest confession from GPT:

This is honestly the missing piece for AI frontend work right now. GPT models are actually decent at UI implementation — they're terrible at maintaining long-term visual consistency. DESIGN.md fixes the memory problem.

[bitter fossil]

Yeah, switching from Claude a few weeks ago to Codex, it for sure lacks FE design, but I'd take that any day over the dumpster fire Claude Code has become

[plush harbor]

I tend to use gemini for ideas, chatgpt is awful and I don't trust codex, it likes beige too much

[lean lark]

No, it thinks YOU like beige too much. 😆

[cedar skiff]

How is the new mobile app?

[plush harbor]

I'm currently in a bright orange and green site with retro font ....

[bitter fossil]

From the stand point of using SSH, it's been insanely awesome not sitting at my desk, it doesn't have $skill dropdowns or anything, but if you remember the name of your skills, it will still use them

I ran into a weird UX bug with /plan, but other than that it's been sweet

[cedar skiff]

I'll give it a run today, i have to sit in the car with the wife for a bit so i can see how it goes

[meager wadi]

had one that was 180 hours

[uneven kayak]

I spent $52 on a single prompt with Opus 4.7 the other day. That enter key is an expensive button to press lol

It was like a 2 hour run but it's still kinda pricey

[bitter fossil]

Man... I was frustrated when Opus 4.6 received a nerf along with the token issues, then once Opus 4.7 launched and felt even worse, I decided to make the switch and glad I did

[grave reef]

I wish OAI hired me

[plush harbor]

I'm over here not hitting limits in plus and a ~5 a month api spend

[grave reef]

I managed to get unlimited quota from them but they caught me

[bitter fossil]

Me too

[cedar skiff]

I didnt use goal, this is just orchestration i guess goal would work as well, but i want it to use 5.3 codex for implementation. The only thing that stops it is running out of work or i update codex haha

[grave reef]

I can probably repeat it but that would be abusing a bug

[cedar skiff]

180 is a long run though

[plush harbor]

my last couple weeks home sick has collapsed all my backlogs so much its new work or content entry. And api is so cheap I need to think of ways to help with content

[boreal holly]

Dang is that Cursor or something?

[cedar skiff]

i think 4.5 was peak for coding, down hill ever since

[grave reef]

4.7 is just a price hike laundered as an incremental improvement

[bitter fossil]

For anyone who has never used ExecPlan: I highly recommend adding it to all projects and your AGENTS.md: https://developers.openai.com/cookbook/articles/codex_exec_plans

[boreal holly]

You ever tried model_reasoning_effort="none"?

[cedar skiff]

I have not, what difference does it make?

i assume much less usage

[boreal holly]

I mean if 4.5 was peak for you, that model had zero reasoning. You must have some extremely well documented processes. They say as soon as you turn on reasoning, "less is more." It might make sense to try zero reasoning and see how it reacts when it has to be as decisive as possible with every action it takes

[cedar skiff]

pretty sure it had reasoning

[boreal holly]

I keep reasoning on because I like to be slightly ambiguous, so good ideas can emerge. But if you have this rigid, infallible process then reasoning is in many ways a hinderance

[plush harbor]

so do these front end skills actually help with mobile testing, like yesterday I noticed there were no touch swipe left and right on mobile on my galleries and the thumbs didn't centre correctly when lists got long so I had to get codex to fix those. Plus a nav that would shift slightly with content size so had to make the control bigger so you could doomscroll without thinking. Anything let codex pick that kind of stuff up? Quite timeconsuming

[cedar skiff]

They are styling

[plush harbor]

that's pure front end, that's what eats the time

[bright swift]

are you using chrome devtools mcp to let it verify?

[boreal holly]

It was the best non-reasoning model of its time

[bitter fossil]

So, use /goal for testing with FE. If you allow permissions for $browser-use, it will open the website and test all changes

[plush harbor]

not verify, have to find these things in the first place to know they exist to fix them

and don't get me started on menus changing sizes when you translate your site

[bright swift]

just tell it to make a screenshot and analyze it. although i prefer gemini for image analysis because of the code execution

via api tho

[plush harbor]

you cant really screenshot interactions like swipe to nav

[cedar skiff]

I had some heavily layered work flows in claude, until i came to codex... which made those redundant. A lot of stuff forcing it to use skills in orchestration, doing spec planning, phase slicing, then implementation planning, implementation, implementation audit, code review. Codex is much easier to trust

[bright swift]

gemini has video analysis too

probably a good product idea now that i think of it

[boreal holly]

If they are controlling a web browser, the interactions are kind of dicy. But if you give them an iOS simulator or Android Emulator, there's a lot of really good tools for them to simulate interactions like drag

[plush harbor]

but codex thought it was fine that you could navigate through the thumbs, which is terrible ux so I got it to add swipe

[bitter fossil]

It was fun discussing Codex, I'll be back tomorrow, have fun!

[plush harbor]

this is just generic stuff that post launch you get your clients come to you about, I've been doing this for so long I pre-empt this

[boreal holly]

Oh yeah, Codex by default is like the typical Gentoo user. "Why not just make it a TUI written in assembly?" Lol

You have to give them the pain of interacting with the app/site like a human would, and the instructions should grade interaction complexity.

[plush harbor]

like yes, it looks like the design. Yes I could predict you were going to hate that feature and ask me to change it. yes I already changed it for you, can I deploy it

[bright swift]

why not do it right from the beginning if you have the exp?

[plush harbor]

out in the real world soemone higher up insists you do it their way and you have to wait for the feedback

[bright swift]

i mean have a collection of presets etc

[plush harbor]

for me, this is just extremely short turn around and the navigation issues didn't even show until i had a lot of content added

this is also my only gallery site

[boreal holly]

Codex does really well if you set up land mines. --yolo mode is the worst thing you can do. Give them this ridiculously tight sandbox, give them paraplegic-level tooling, and absolutely no opportunity to work around it. They will discover out of sheer will the proper way to build the interface

[plush harbor]

this sounds like you need agency level turnaround and a huge library of existing patterns

[bright swift]

yolo + custom hooks 🥳 https://github.com/openai/codex/issues/22546

[plush harbor]

professionally I've generally dealt with some fancypants design that's come from an external agency and I get to implement it, so I'm used to getting all the divs and classes in the right place so that's easy to tell codex what to do. Actually coming up with the design is hard though

[proper cosmos]

How do I get ChatGPT codex to send whatever permission to my desktop,

From the mobile app

[lean lark]

Which OS Seany?

[quaint prism]

it dont work for windows user bro

[quick thorn]

1m context or even 500k
But no we get a broken phone connection

[proper cosmos]

Oh ffs

[vital hill]

gifs don't load here smh

[quick thorn]

because people are setting all this high tech up on Macs huh what a joke

Obviously ppl are using pcs

[quaint prism]

😭

[bright swift]

lol who

[proper cosmos]

Bloody Claude got it right first time

[quick thorn]

real devs

[proper cosmos]

I don't need 1tb of ram for my work

I'll just use the piddly little chrome remote desktop on my phone as normal then

[grave reef]

yolo always

[plush harbor]

I've never met a dev on pc, but yeah they probably exist

[cedar skiff]

I spent a few hours building out a decent rule set for codex, but i use guardian mode auto-review now.

I think it still uses the rule set underneath

[quick thorn]

Interesting

[bright swift]

i just use ephemeral remote envs with full yolo mode. spin up, execute /goal for couple hours, commit/push, shut down

got tired of overheated macbook

[orchid plume]

ah really? So this feature is MacOS only? That explains why I'm not seeing a permission screen on the app (Windows), rip

[quaint prism]

it will later but not right now

[quick thorn]

spent my last 200 on codex we either get rich or die trying

[quaint prism]

real gang

[quick thorn]

😭

[bright swift]

i thought you were a real dev

[quick thorn]

yeah real devs are broke and homeless now

[bright swift]

sadge

[vital hill]

i got a mac studio

am i real dev

[quaint prism]

if everyone's a dev

then no ones a dev

[plush harbor]

I don't have a job so I'm clearly not a real dev

[bright swift]

markdown devs

[boreal holly]

You get the new M3 Ultra?

[quick thorn]

holy prompt engineers

[vital hill]

m1 ultra

[quaint prism]

more like....

english teachers

[quick thorn]

is 500 features too much for a model?

[velvet wren]

how much RAM do you have in it?

[boreal holly]

Me too!!!

[velvet wren]

I have 96GB on my MacBook Pro

[plush harbor]

I'm on my old computer right now. Sadge

[bright swift]

the nearly full disk is 80% target/ dirs i guess 😄

[boreal holly]

Absolutely nailed it lol

[vital hill]

i have an m4 macbook air that i barley use, i should probably update it

[cedar skiff]

m4 goes hard

[vital hill]

i only use it when i'm out and about but i'm mostly home

[craggy crest]

m5 mbp here

[quick thorn]

I have Mac but my pc is op

[velvet wren]

how much RAM did you configure it with?

[bright swift]

EX131-L looks like nice dev machine https://www.hetzner.com/dedicated-rootserver/matrix-ex/

[craggy crest]

24gb, fine for my uses for now

its my first main machine mac, been linux/windows for years

[uneven kayak]

Yeah, never using Cursor again

[quick thorn]

Holy

[boreal holly]

That Jetson Thor dev kit is looking pretty nice right now though

[lean lark]

@trim rapids Codex-usage now supports extensive filtering. Please see branch filter1, Issue #4, and branch README. TYVM

[boreal holly]

Sucker still works 12 years later, and that's honestly the best part

[grave reef]

i sadly broke my codex building an extensibility plugin for it

[bright swift]

🙁 back to claude code for you

[plush harbor]

had to replace the keyboard, it went weird after I got it out of storage at one point. I've had various work-supplied computers over the years, dusted this one off when I gave the last one back

[grave reef]

well at least I have a custom pet with emotes and stuff now

and that uses chronicle to entertain me

[bright swift]

priorities

[cedar skiff]

i have 24gb m4, im always using a tonne of swap, but i dont seem to notice it. On windows it's obvious when using swap like this. Mac seems to handle it better.

[grave reef]

custom css is good too in the app

[bright swift]

doesnt that ruin the ssd

[grave reef]

and I can edit the system prompt so gpt 5.5 doesn't write comments and docstrings like a caveman

[proper cosmos]

So guys, how much money has everyone made in total from codex

[bright swift]

tree fiddy

[proper cosmos]

I've made a grand total of $50

[cedar skiff]

i guess it might shorten the life of it, i never reached the end of life on an ssd yet tho

met the end of life on plenty of HDD's

[hushed lava]

codex on mobile!

[quaint prism]

we know bro

it's not on windows

[hushed lava]

so freaking op

[quaint prism]

half of us cant use it

[hushed lava]

yes it is!

[cedar skiff]

Pretty cool aye ?!

[hushed lava]

you gotta know how to set it up

i got you

[quaint prism]

we know how to set it up

its not on windows

[simple star]

Ive tried to connect Codex Mobile and I get a DNS_PROBE_FINISHED_NXDOMAIN error. Anyone else?

[hushed lava]

go on power shell and type , codex remote-control , while having the codex app open

then go on ur phone, and click codex and it’ll link

[boreal holly]

Just gotta watch out when "Wired Memory" gets too big. The rest can be jetsam'd!

Also do sudo zprint | grep "kalloc.1024" every once in a while. Make sure it stays within a few gigs. Some GUI apps memory leak into that zone map, and also a lot of network and fs watcher stuff leaks in there. It's a permanent overhead that only clears on reboot

[hushed lava]

yes it is

[quaint prism]

bro knows ball

[hushed lava]

i just told you how to set it up

yessir