codex-discussions

[boreal holly]

Basically I have all these model and reasoning effort adjustments throughout the app, but I could get rid of it all and hard code 5.5 medium

[deft gyro]

And feels faster

[hard drum]

5% weekly left... i am doomed

next reset for me is when i will focus A LOT on my upcoming v4 as harness overlay

about time i took it more seriously

then make a stripped-down codex-cli build, perhaps

[boreal holly]

I'm telling you, the biggest scam is all the new features they keep adding!

unified_exec=true: the agent wakes up every 5 seconds to check on a long running process. After 20 seconds, gets impatient and terminates it. Then spends countless hours trying to patch away a problem that doesn't exist. Burns all the tokens on earth.

unified_exec=false + no timeout: command runs in perpetuity, the agent doesn't wake up. Sure if it's more than 10 minutes you lose the input cache bonus, but at least they didn't spend countless hours wreaking havoc on non-issues. Also if it hangs you simply interrupt and tell em it hung for real. No biggie.

[frosty zealot]

This is only relevant to using sub agents I assume

[unique spade]

well, you can always tweak what is broken. the beauty of oss 🙂

[frosty zealot]

What if you unified_exec=false without a custom build

[hard drum]

maybe if codex's codebase wasn't too big

[boreal holly]

It's relevant if you want them working autonomously. If you wanna walk away and know the agents aren't patching away non-issues then totally relevant

[unique spade]

30 mins once in a while is not too big to get the bahvior aligned with what you want

[boreal holly]

The agent gets to decide the timeout. Between 5 seconds and 10 minutes. So they typically set the timeout for 5 secs, after it fails they upgrade to 10, then 15, etc until it reaches 10 mins and then they give up. There's a hard 10 minute cap, but they can choose less lol it also wastes tokens because they run the command like 20 times and see partial outputs, timeout errors, and investigate stuff.

[hard drum]

i'd still want to strip it down to what's actually worth keeping

i don't need 699K lines of mess

that thing is maybe at best 150K of goodies

[unique spade]

right, i agree with that. seems a legit goal for a custom build tbh

another thing is they refactor the whole thing every week 😂

[hard drum]

it's obviously too vibed without actual checks

but somehow works really well for what it is

also the config is kinda confusing in itself, so having a CC-like config might be better because it's easier to manage lol

[boreal holly]

They added some deps that require you to set an env var to force cargo to fetch submodules with the CLI in order to build

[unique spade]

yea but if you like minimalism you want a mini robust core that does what it needs to do and you don't carry the rest that in your own case you never really need

[undone patio]

anyone found a good solution for sound effect generation yet?

[unique spade]

but all in all is fun tinkering with it, at least for me

esp since you keep vanilla too and can always use that too

btw what's the most use 3rd party open source harness that you can use with your gpt subscription?

opencode or?

never tried any yet

[bright swift]

pi.dev

[hard drum]

opencode.

[unique spade]

oki so pi.dev and opencode

[bright swift]

opencode is buggy mess

[hard drum]

true, but it is mentionable.

[unique spade]

don t plan to use them, but to inspect them and borrow what i need

[hard drum]

KiloCode v5 is the "opencode" replacement tbh

not v7. don't get that

[bright swift]

i'd prefer pi for that

[unique spade]

cool thx, will look into them

[bright swift]

you can just use the base version and build on it, dont even need to fork

its super minimal

[unique spade]

nice

[lament lily]

https://github.com/google-gemini/gemini-cli/compare/v0.39.0...v0.39.1

[unique spade]

i just love to use codex as my operator

[bright swift]

one thing about pi that is cool is its very self aware

you can just talk to it about improving itself

without extra prompting

[lost drum]

I dont undestand you can just dowload it right?

Sorry I never used github

[inland sonnet]

u just type git clone link/.git

xd

[lost drum]

So why he doing it this way haha

[inland sonnet]

then ai will run the cmd

[lost drum]

And why repos have file like LICENSE?

[unique spade]

i don t want to install pi

i want to inspect it

[lost drum]

But you would download whole zip and send it to him?

[unique spade]

so i need the github repo on my pc so codex can look into it from inside it

[lost drum]

Would be faster I think

[nocturne folio]

fufufu

[nocturne folio]

cuz of legal protection

[lost drum]

But u can just delete that file

[nocturne folio]

doesnt work that way

[unique spade]

nah i prefer the cleanliness of a cloned repo

since i make that folder a project and can fork it if i want, copy parts of it elsewhere,use the benfits of git in short

[nocturne folio]

you have the right to download their ip, but you still abide to their rules

if they catch you regrading these rules then they could take legal action

[lost drum]

But if its hosted on your decice then how would they know?

[torpid trout]

are you adding more zeros to your name every time you type here?

[unique spade]

because you acknolwedge the content of that license when you download it

[nocturne folio]

they dont know, they will never know

[unique spade]

if a license doesn t allow you to use that repo for commercial purposes for example, and you do, and you get sued by the repo owner...

[boreal holly]

https://letmegpt.com/?q=And why repos have file like LICENSE%3F

[nocturne folio]

i just bought the domain grapeee.com, and i like the like word extending thingy

[lost drum]

But how would they know if You use partiarly their system and edit it and use it within other thigns like inspo type of thing

[torpid trout]

holy moly website

[torpid trout]

it's not about them knowing, its about them telling you.

the police also does not know you steal, until they do, and then, "non knowledge of the law does not make you immune to the law"
That is another way of sayin "you have been warned bro"

[unique spade]

that depends on a lot of factors, it's not a question that can be answered in abstract.
that's why the entire branch of legal system around Ip is for, to settle issues like this if it becomes so important that one of the party files a legal complaint

[lost drum]

🫡

[lost drum]

What if you vibe code and someone tells you that its his system lol

[unique spade]

he would need to sue you and prove it

[lost drum]

I see I see

[bright swift]

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

[frosty zealot]

Enjoy the entire spongebob movie as a gif

[bright swift]

thanks exactly what i was looking for

[finite sun]

Cinema

[bright swift]

i made this session retrospective skill and now codex is telling me to send clearer prompts. this isnt going the way i thought it would

rip dreams of self-learning

[finite sun]

Yesterday I had to start codex in YOLO mode for it to push to Git for me me without using the connector. I’m thinking with 5.5 out maybe they changed security settings? Just me or anyone else?

[bright swift]

no idea, tell codex to check the changelog

but one tip: that github connector burns a lot of tokens. codex is very good with git and gh cli, just let it use them directly

[frosty zealot]

A lot of that stuff can also just be done manually to save on tokens, or create scripts and use them in hooks

[boreal holly]

Haha good ol dave hasslehoff

[finite sun]

To this day I’ve never reached my limit on codex (,: luckily

[bright swift]

sweet summer child

you will soon 😄

[kind jay]

Unc

[bright swift]

hows your of going

[frosty zealot]

You wish that was you

[kind jay]

Hm? Wdym

That’s my cat

[frosty zealot]

Oh sorry

Looks a lot like mine

Its refactor time

[boreal holly]

# ~/.codex/rules/10-git-rules.rule
prefix_rule(
    pattern = ["git", "push", "--force"],
    decision = "forbidden",
    justification = "Don't do that"
)
prefix_rule(
    pattern = ["git", "gh"],
    decision = "allow",
)

[frosty zealot]

https://github.com/openai/codex/issues/15485

Looking for more thumbs up on this issue

Need every one to do it

[kind jay]

You woke her up

[frosty zealot]

lmao

[bright swift]

i need thumbs up too pls https://github.com/openai/codex/issues/18969

[frosty zealot]

why did you thumbs down the guy saying he would work on it lmao

[bright swift]

because he didnt read the contributing rules

they dont accept prs, its invite only

and i can prompt codex myself to implement it

so just low value comment, likely automated

[kind jay]

She keeps following me into my computer room, probably because it’s warm

[frosty zealot]

Oh I assumed he was allowed, looks similar to another guy that implemented something I posted on before

[lost drum]

Nice setup

[kind jay]

Ty

[lost drum]

Show the desk

[bright swift]

nah just some random guy, i wouldnt downvote an OAI guy 😄

[frosty zealot]

looks like scruffly

[kind jay]

Boring Cisco routing stuff on top

Old crap

[finite sun]

The entire thing was that the session was starting with network access disabled suddenly

[lost drum]

I need to buy new chair can You show yours? Or tell ehich one if You have good one?

[boreal holly]

You're gonna like this. Apple has a 1MB websocket frame limit, but I just learned the hard way that Rust Tungstenite has a 16MB limit lol because stupid codex app-server tried to send a 21MB frame 😩

[kind jay]

I have some old steelcase one, it’s not bad

[boreal holly]

This stuff is absolutely re-donk

[frosty zealot]

"Make my send button bigger than anything else on the screen"

[bright swift]

gpt frontend skills 11/10

[torpid trout]

Dumb question, does someone here master photoshop and can seam a tile?
That's mean I have a square image, that has a certain pattern in it, and this needs to happen:

1. Offset the image by 50% horizontally.
2. Blend/clone the new center seam.
3. Repeat for vertical seams if needed.
4. Export the same 134px × 134px tile.

I can send you a custom meme as payment

Or I will just ask gpt, I guess. lol.

[bright swift]

try figma ai, im serious. i cancelled my photoshop sub and just ask it to do it on the free plan 😄

[boreal holly]

So basically GPT 5.5 cannot detect sarcasm and agrees the button should be more massive

[torpid trout]

have it do it bigger

Then ask again lol

[boreal holly]

There we go

[leaden heron]

Yeah I pay for 4 seats but it’s just like that

ATP just use 5.2 until ChatGPT can get their stuff together

[torpid trout]

This button is too small.

smol, very smol. It has to be bigger, the biggest ever, never seen before.

[frosty zealot]

feel like im playing Agar.io now

[boreal holly]

ok this might be too big. Can barely see the arrow!

[frosty zealot]

Put a easter egg, make it normal size, but as you tap and hold it just blows up like a balloon till you release

bonus points if you make the border wiggly and wobbly like water all the wy around as it inflates 😛

[torpid trout]

you are saying it still goes make it bigger?
🙄
At this point it might just be lobotomized and switching chat could resolve it lol

[torpid trout]

just make it a water filled preservative that escapes your touch, falls to the ground and pops

[frosty zealot]

Good idea, while holding you can drag it around, it obeys physics and molds over everytihng, then release it and it drops and pops @boreal holly make it happen

[torpid trout]

GBT for the world btw.

Robert is busy making preservatives

Robert rn

[frosty zealot]

its ability to do text is crazy though for real

I remember fighting like a year ago to make any sort of text look coherent now its printing out full code blocks lol

this got a little scuffed but

[torpid trout]

The ability to text is one thing

Look at that man in that image. I mean, if he where not trying to push a deploy button into a condom.. I am fairly sure one could not discern this is ai anymore, it looks like a pic

The veins, the hairs, its like it went to the www, got an image of a dude, and put it in there

[boreal holly]

but me rikey deploy to prod on a saturday 🥸

[kind jay]

Looks AI maybe 🤔

[torpid trout]

Your memes are waiting, wanna buy em?

They too spicy, almost got be gbtanned

[kind jay]

DM lil bro

[torpid trout]

└ Stream disconnected before completion: websocket closed by server before response.completed

[kind jay]

You’re lying

[boreal holly]

automods sleeping

[kind jay]

WTH

[orchid plume]

jees the spam

[kind jay]

How?

How?

[true bane]

ok after like an hour of this thread i got to this summary in case you're wondering where our capacity went

[frosty zealot]

@quasi canopy Can you take care of this

[kind jay]

Yeah!

If I can’t say it neither should he

[frosty zealot]

Squirtle squirt

[boreal holly]

automods be like Reconnecting... 6/5

This issue needs like a billion updoots https://github.com/openai/codex/issues/15485

[kind jay]

Why are you guys always begging for upvotes? It’s embarrassing

[frosty zealot]

Almost bed time for you (thankfully)

[kind jay]

You’re powering me down?

[frosty zealot]

sudo pkill -f jane

* * * * * *

on that cron

[boreal holly]

Idk, OpenAI keeps making features like "can now generate images" or "can browse the incognito for you", meanwhile core features like opting out of massive payloads for super obvious and sensible performance reasons goes on without even a glance lol

[kind jay]

Can it watch videos in incognito for me?

[frosty zealot]

@kind jay getting off the computer and walking to bed

[frosty zealot]

shoulda generated maps 🙁

JavaScript .map files (source maps) are basically a bridge between minified/compiled code and your original source code.

When you build a project (especially with tools like Webpack, Vite, TypeScript, Babel, etc.), your nice readable code gets transformed into something like:

• minified (one line, short variable names)
• bundled (many files merged into one)
• transpiled (e.g., TS → JS)

That’s great for performance, but terrible for debugging.

⸻

What .map files actually do

A source map file tells the browser:

“This piece of generated code corresponds to this exact line/column in the original file.”

So when you:

• open DevTools
• hit a breakpoint
• see an error stack trace

…it shows your original code, not the ugly compiled output.

⸻

Example

Without a source map:

app.min.js:1 Uncaught TypeError at a (app.min.js:1:2345)

With a source map:

App.tsx:42 Uncaught TypeError at handleSubmit

Huge difference.

⸻

How it works (simplified)

At the bottom of a compiled JS file, you’ll often see:

//# sourceMappingURL=app.js.map

That tells the browser to load the .map file, which contains a mapping like:

• generated line/column → original file/line/column
• plus original file names and sometimes source content

⸻

What’s inside a .map file?

It’s JSON that includes:

• sources → original files (e.g. .ts, .jsx)
• mappings → encoded mapping data (VLQ format)
• names → original variable/function names
• optionally sourcesContent → actual source code

⸻

When they’re used

• Debugging in browser DevTools
• Debugging production errors (if maps are available)
• Tools like Sentry to map stack traces back to real code

⸻

Important caveat (security)

If you deploy .map files publicly, you’re basically exposing your original source code (sometimes entirely).

So in production, people often:

• don’t ship source maps, OR
• upload them privately to error tracking tools (like Sentry)

⸻

Quick mental model

Think of it like:

“A GPS for your code — it tells the browser how to navigate from compiled junk back to the real source.”

⸻

Well you should

BECAUSE ITS A SERIOUS PROBLEM

Ask Anthropic kekekekekekek

You right now

[kind jay]

Bro what?

Conversation with yourself 💀

[frosty zealot]

https://tenor.com/view/silly-funny-reaction-meme-stan-twt-buckspendragon-gif-12988406325947841564

your dumb face rn

[frosty zealot]

lmfao

what a tweet, not fake btw

https://x.com/sama/status/2048160404376105179?s=20

[thorny rain]

Tried using codex app to generate image it called image-gen but ended up using python to annotated 😭

[hard drum]

PLEASE, OAI, PLEASE

weekly reset PLEEEEEASE

holy deficit

[frosty zealot]

Time to bust out gemini

[hard drum]

ew no

gemini can go dig its own grave

[frosty zealot]

I feel like google gonna come outa nowehre with some insane model on new technology that does the most cracked stuff

[hard drum]

they have failed since 2.5 Pro, so my confidence on that is NULL

[frail meadow]

Do you guys get crazy memory usage from Codex on like a daily basis? On macOS

[harsh patio]

Windows Code 5.5 huge reduction in limits

[boreal holly]

Bro u really think you can use all 98% in 2 days?

[hard drum]

i can not

it's at 99% lmfao

i am done for soon

[obsidian socket]

its worse than ever! I just used 5 hours in 10 mins and 3 prompts to continue the code from ealier when I had to wait for this silly 5 hour reset. now the platform is not usable

The service is presented as a usable allowance (e.g. “5 hours”).
In practice, normal use exhausts it in minutes.
That is not a reasonable interpretation of what was presented.

[boreal holly]

bro check your issue. I just found out v0.125.0 does do pagination if you use thread/resume, which I think fixes both our issues

those sons of a gun are gonna make me do another update just for pagination

[obsidian socket]

what are OAI doing?

[deft gyro]

what

yall cannot be burning thru the 5 hour quota in minutes

(on plus)

like

wyd

[frosty zealot]

Is this documented or do I need to get Codex to pull the functionality out of the repo

[boreal holly]

you could probably codex app-server generate-ts, it'll have the new params

[frosty zealot]

kk let me have a boo

This looks very promising

I'm getting excited

[boreal holly]

Yeah, I been uploading pics to my designer which is why that thread history is so massive. Having pagination is gonna be sick

[frosty zealot]

Where I work is very remote with very poor reception, like im talking 0.5Mbps kinda thing

So trying to switch between threads was so painful

[stable obsidian]

Codex + blender 👍

[deft gyro]

xd

[boreal holly]

at least it didn't use metaballs

[deft gyro]

at least

it didnt use cards

god

chatgpts addiction to cards

needs to be studied

chatgpt in a task it FULL WELL KNOWS doesnt need cards

[boreal holly]

one time I asked it to describe what it thinks a card is in its own words, and the definition it gave had nothing to do with rounded rectangles, so all the times I was like "get rid of the nasty cards" it had no idea I meant everything a rounded rectangle with borders and drop shadows

[deft gyro]

although 5.5 is mildly less card addicted

from what ive seen atleast

[frosty zealot]

Would you call these cards

[deft gyro]

this is A card

[frosty zealot]

I would just say its wrapped in a container

[deft gyro]

its more of a table imo

but typically having a container(which is really just a singular card) is fine

I am more talking of the pervasive habit of spamming them absolutely everywhere

like the screenshot

which would obviously be better in a table

like this from the codex show and tell

these are a bunch of cards

[boreal holly]

Oh you want cards? I'll show you cards lol

[brave gazelle]

so ik i talked shi about the image model when it first dropped because i was mainly salty that it wasnt 5.5 but jesus this model is so good, its trivializing frontend development which i suck at. Im having codex read my openapi spec, create a prompt for the image model, generating the image, and then codex views the image and writes the code to attempt to match its look + wire the endpoints it was associating with that screen in a loop, and the results are amazing

[boreal holly]

Yes! It's no wonder they released image v2 and 5.5 so close together. They are the perfect team

[deft gyro]

cardception

image v2 to design ui into 5.5 is a W combo

[frosty zealot]

I have a secret weapon I'm willing to share with you guys

[deft gyro]

5.5 is the first oai model that isn't completely terrible at multimodal

[frosty zealot]

IF

[brave gazelle]

ive never cared much about image models but this image v2 is extremely compelling

[frosty zealot]

IF you want

[deft gyro]

like gemini models have been my go to for turning sketches into ui

[brave gazelle]

sure

[frosty zealot]

Are you sure you're ready?

[boreal holly]

im scared

[frosty zealot]

https://getdesign.md/ you find one you like for spacing, containers etc. and you just use the design.md, can save multiple if ou want and just instruct it to use the containers from X or whatever, id change the colors though

[boreal holly]

I thought it looked great 👍

[frosty zealot]

ty

i imagine theres like 300 lurkers that just wait for ideas to be dropped in here steal the screenshots and then tell codex/claude to create it and then push it immediately to the app store

[brave gazelle]

so u upload the design.md to the image model?

[frosty zealot]

Just save the .md in your repo directory, and then you can either update your AGENTS.md to say use 'DESIGN.md' or like in my case 'INTERCOM-DESIGN.md' for UI/UX guidelines

Or just do it inline with a prompt

[gentle harbor]

doubt

[frosty zealot]

Following DESIGN.md guidelines, implement X

[brave gazelle]

oh so ur skipping the image model all together

[boreal holly]

Good news is, outta the 300, 280 of them are troubleshooting why codex for windows has issues

[gentle harbor]

i wonder why claude and open ai dont work together and make a better model, they are already two of the best

[frosty zealot]

No doubt LMAO

[brave gazelle]

yeah ive tried using https://impeccable.style/ which is a similar way of doing things

but i think image model has made much better designs for me personally

[leaden heron]

has anyone every noticed when you switch accounts, it automatically switches on fast mode or like opening the codex app again, also switches on fast mode

it literally wastes so many credits

[gentle harbor]

no but i have noticed that ive been downloading a update for the past 15 hours

[leaden heron]

Bruh

Too many bugs

[boreal holly]

Make sure service_tier = "flex" in your config.toml. If you flip it off in the GUI it might not be updating the grand master config file

afaik the GUI app maintains a separate set of config overrides for certain features

[gentle harbor]

are you fr ? it took 1 min and 30 sec to try and see the text in a image, is the model getting slower or something

[frosty zealot]

is this a mac

[gentle harbor]

this is windows

[frosty zealot]

There's your problem

[gentle harbor]

they added the download button 3 days ago and it breaks after the 3rd update lol

[boreal holly]

Unplug all USB devices, restart computer, try download again

[frosty zealot]

the funny part is with windows, that would probably work

[gentle harbor]

im out of credits so its what ever right now

[boreal holly]

That's a legitimate fix for most issues. I used to have a USB headset that would break Visual Studio C++ builds constantly

[gentle harbor]

when do you think the super app is coming out ?

[frosty zealot]

when im done building it

[unique spade]

how s everyone

[cedar skiff]

I think its going to be codex app and they just keep iterating on it until it is it.

[boreal holly]

Good! 5.5 is killin it. Can't wait to get off work tho

[frosty zealot]

Are you still on the tools or they got you makin your thing full time?

[boreal holly]

Full time 😎 the app has replaced all the one-off tools they asked me to make

[frosty zealot]

That's awesome I'm a little envious 😛

My solo venture has been tedious, needless to say lol

I need to finish my codex app so i can get back to trying to polish it off and start trying to promote it or approach local companies

But been having a bunch of fun with the codex app not gonna lie

So do they own your software then since they're paying you to develop it or are you licensing to them or something?

[boreal holly]

You totally should man! And if you wanna borrow the orchestrator setup I got, let em run 24/7 on whatever needs to get done, only thing that requires full attention is the frontend design

[frosty zealot]

I think I'm a little past that requirement at this point, a lot of the tasks are just UI/UX work and really just making sure the modules arent dependent on one another, but I appreciate the offer

[boreal holly]

I been working here for 8 years doing service, and the boss pulled me aside and said "let's start a software company" haha so im like a cofounder

[frosty zealot]

Dude that's so sweet, my company is too brain dead for stuff like that lol

[boreal holly]

I count my blessings bro

[frosty zealot]

They also might be losing the contract to another company in June so I'm interested to see what happens

[boreal holly]

Oh yeah, with any luck you'll run into a HVAC boss who's sick of the enterprise grade bologna and might actually accelerate the timeline 😁

[queen delta]

The world does not need more slop

[lean lark]

I am always in my prime!

[boreal holly]

you gotta get supervisorctl 😁

The worst is when Codex updates something and reboots the app-server, but it wasn't a clean reboot so it leaves an orphaned listening port and the new one fails to launch

Or reboots and you're not paying attention, and you're like "wow, these agents have been cooking non-stop!" but really all the threads got interrupted

[dusk thorn]

This is real

[frail meadow]

Dude google is planning to invest $40B in Anthropic? Lol

SOTA for one model release

[unique spade]

to keep the show going
competition is good for innovation

[frail meadow]

Because also Anthropic hosts some of their compute on TPUs now

[unique spade]

i don t like that dude haha

[frail meadow]

Yeah

Their proprietary hardware Google made

[unique spade]

the relation between google and anthropic was kinda obvious since they released antigravity from day one with opus and sonnet included

gemini is rather bad on agentic stuff, but don't forget that on frontier science, google are the only ones keeping up with oai

as in both have models capabple of hitting gold at math olympiad live

anthropic isn't there yet, i don t think they even attempt to participate

but they're good at ux, boilerplate, powerpoint and sheets

[obtuse river]

AFAIK Anthropic hardfocused reasoning (and now coding) with nearly no fluff. Also they have less raw resources so they have to specialize and be tight.

It's impressive what they've done, I reckon.

[unique spade]

it's impressive sure. that s why they re valued where they are. i would put their main focus on coding though, then mid-level work tasks

reasoning is not their strongest point, that was the point about who does well at math olympiad and similar live testing for actual reasoning

since those are not static benchmarks you can train for in advance

[gentle harbor]

coding with chatgpt normal app might be worse and slower but i feel like the rate limits are 20x higher then codex

i have never ran out of messages

[unique spade]

you mean copy paste?

and yes you won t run out of chatgpt messages

[gentle harbor]

kinda depending on what it is you can just download the file they give you

[unique spade]

that s the ultimate backup option for when you run out of codex qouta

[unique spade]

yup i know, you re on plus or pro?

[gentle harbor]

plus

i might get pro when spud fully comes out

[unique spade]

still ok, thinking extended if you convice you re doing valuable project will work quite hard

and give you nice zips

gptpro you can ask straight for a full app in a zip. didn t try that with thinking

[gentle harbor]

i wonder people think the price per message will go up when gpt is done training, but wouldnt it go down since the servers that were used for training could be used for normal users now ? so there is less stress on the servers

[unique spade]

what happens now is a civilizational upgrade, but most still think about it like in the movies or with old 20th century logic

just keep training for that starfleet academy, space cadet 🙂

[obtuse river]

As if they wouldn't start training another new model? Why wouldn't they?

[unique spade]

who?

[gentle harbor]

eventually ai would be able to train it self no ? every model they are relying more and more on ai to help train it self

[obtuse river]

OpenAI?

[obtuse river]

Would still take just as much compute - or even more, if no work is being done externally by humans with human brains and tools.

[gentle harbor]

they are planning on expanding from 10gw to 30 gw so they will have plenty compute

[unique spade]

They already said that gpt/codex did a lot of the training for 5.5

That doesn't mean it trains itself autonomously. It's akin with how it codes for us, just that there it works with much more compute, but still on clear plans and all

[velvet wren]

three fingers is enough to change desktops

[fringe shuttle]

idk what it is but something changed in the past few hours. GPT 5.5 Medium feels dumber all of a sudden

[lost drum]

who tried increasing the context window from 258k to 400k? for gpt 5.5 xhigh? I wonder if it will really go bad

cause mine jsut twice reseted his context to 0

ye

haha

nope

its just a very large porject and he needs aloooooot of context to know what to do

he then compacts it but idk why this time he did not

10% of it its context but rest is his systems that allow him to self heal and diagnose and execute the real playbbok.md I gave him which has requirements of mine and rules that paersuade him to the finished system

okay now he compacts it idk why he didnt earleir haha

I just checked and ye 133 main folders with the systems that allows him to do it

[cedar skiff]

Currently you can't

[lost drum]

then thats a sign that I should not do it then and maybe try and revump the system co it wont access one file with all the context

[thorny cloud]

Codex or Claude for rust language 🤣

[cedar skiff]

The idea has been done to death and it doesnt work with what we currently have.

[lost drum]

wdm

depends on what sontext those codex need to have to execute things so its just matter of creating him the right enviroment I need to revump mine so his decisions will be spreaded across more steps so it will have mroe uppourtunities to auto compact jis context to proceed with next htings

[dusk thorn]

They don’t even invest in their own Gemini when they got the entire Google search engine

[astral nest]

same here, browser use just doesn’t work with codex’s in app browser

macOS user

[cedar skiff]

There really is only a certain amount of good usable context. By injecting the system you use all of that up adding it and because all of the usable context was used supporting the system the actual implementation after that is flakey at best. The original ideas for the system get washout out by the incoming implementation context and are not honoured.
The common convergence of evolution i see is topical skills that are invoked when needed, agents.md for codebase rules like linting/code formatting local codebase rules - 500 lines at most. Then system prompt injection (or total rewrite if you are game enough) that supports coding patterns you want in your code. Like TDD or output formatting, ideas that you never what to change and expect to be honoured.

[inland sonnet]

Does codex even follow instructions properly?

It never mentions

xd

[spark oak]

Hey, I’m running into something confusing with Codex usage limits.

I’m getting a “usage limit reached” message, but at the same time the context window shows only ~38% used (97k / 258k tokens). That doesn’t seem to line up.

Is the usage limit tied to something other than the context window?

[inland sonnet]

Yes, context window always get compressed.

You get so much tokens to use per 5 hours, and a weekly limit.

[inland sonnet]

You can type /status to see your 5 hourly and weekly usage.

[cedar skiff]

i have this at the end of my agents.md and it does it every time. It's a good indicator because i can see it used the skills i expected it to, and it's still honouring the agents.md. If it doesnt have a skill i expect in there i ask the agent why it didn't use it and then update the description based on the feedback.

## End-Of-Task Skill Audit

After completing coding tasks, list the skills that were used.

[inland sonnet]

[lost drum]

ye mine those max 700 lines of code per run to know what to do next and next as you said its possible to give hiim brain but its jsut slow

[lunar berry]

I feel like 5.4 is dumber since 5.5 came out

[deft gyro]

Why

[lunar berry]

It was behaving stupid yesterday when I worked on pgsql

In opencode

[lunar berry]

https://youtu.be/2_Dtmpe9qaQ

[kind jay]

Go to sleep 🙄

[kind jay]

I watched like 2 minutes and he didn’t get to the point

I was up earlier than this

I don’t go on discord first thing

[deft gyro]

No

It’s about right to own software

https://be-clippy.com

Basically

Clippy never asked you to buy a subscription

Clippy didn’t force you to do anything

"Clippy didn’t sell your data. Clippy didn’t hold your data hostage. Clippy was there to help you."

[kind jay]

Just use Linux and open source, what?

[signal tapir]

Anything new on the codex front today?

[cedar skiff]

no update today on the codex app

[signal tapir]

I seem to have a new bug. Had the old conversation frozen bug. Stopped. Restarted. Nothing. Now, on second restart, it works agin.

I wish I had the energy to actually investigate that kind of stuff.

[kind jay]

Son 😭

[signal tapir]

is confused

which in itself isn't that rare

[kind jay]

Me or you? I’m confused

[signal tapir]

Both I guess. 😁

[kind jay]

She’s tired

[signal tapir]

One of the cool things with AI is that one can be productive even if one is tired. 🙂

Of course being a cat, she can be productive while sleeping.

[kind jay]

Her name is Panic

[signal tapir]

My lapwarmer is called Ninja. She is productive 24/7

And I'm out of usage. I really need a larger account.

[kind jay]

Short for panic is real (inside joke)

[signal tapir]

3.5h until reset. 🙁

[kind jay]

Why don’t people pay for two separate plus accounts?

[signal tapir]

I think people generally do. I can't afford it.

[kind jay]

Fair

[signal tapir]

Well actually, people seem to get the $100 thing more often.

[kind jay]

I have this, worth

[fervent anvil]

Codex via VS Code is very very broken at the minute.

[signal tapir]

I'm going to blame spending all my credits on that. 😛

[lost drum]

Not worth cause for 200$ u get usage worth 400$

[lost drum]

Wdm mine works

[signal tapir]

I plan to get something like that, eventually. Currently I'm a bit short of monetary units.

[lost drum]

Me too haha. But I want to abuse it so next month I will use plus plan

Hebrunning 24/7 and spending enough usage

[terse kraken]

I've using the 1% left of my weekly $200 usage on 2 sessions at 0% for 3 hours and still going lol

[orchid plume]

is it a continuous turn, like you sent a prompt and it's still going 2-3 hours later?

[terse kraken]

Yeah, its kind of a cheatcode it waits until turn ends before giving you rate limits

[orchid plume]

yup, was about to say 😛

[terse kraken]

Honestly though they should reset limits since 5.5 came out they haven't

[cyan gyro]

how do you keep codex going?

it always splits things in slices

(by default)

[orchid plume]

I've not figured out a reliable way, then again for the things I do most of the time I don't need to keep saying "continue" to it, it just does so

[orchid plume]

problem is yesterday, and maybe today as it's still the weekend, there were issues with "Selected model is at capacity" errors happening. Manually resetting everyone's quota at the moment would compound that

not to mention GPT-5.5 was incredibly slow yesterday without enabling fast mode

[cyan gyro]

slow still right?

[unique spade]

Give it a procedure it likes and then just chain more of those slices in one turn

I think for me next period will be about testing how codex orchestrates oss models

Seems Nvidia is giving free unlimited api access for prototyping purposes to around 50+ oss models

[orchid plume]

not sure, I'm still on fast mode this morning

[unique spade]

So things are going in the right direction it seems, intelligence is not getting more expensive, quite the contrary

And those include deepseek 4, glm 4.7, nemotron etc

And if there s any model capable to properly orchestrate it has to be 5.5

[torpid trout]

By that math I’m 120

[cedar skiff]

I tried the nemotron super they have for free, but it was REALLY slow

[kind jay]

You got up at 12:00?

[torpid trout]

Depends on what timezone you’re looking at it

I often go up at 5:20

And sleep at 1

Now who’s living on a server rack uh

@burnt sage who are u and why do you try to befriend me?

[kind jay]

That’s not a flex

I aim to sleep 10-12 hrs per day

[torpid trout]

Right, it’s hardcoded

[burnt sage]

@torpid trout no friends 😭😭😭😭

[torpid trout]

I didn’t need to aim for that for long, but somehow one day my body decided that’s what’s going on now

[torpid trout]

Then why send me request?

Oh. Because you’ve no friends.

[burnt sage]

Add me @torpid trout

[torpid trout]

Jane is looking for friends

Im well served. Me, myself and the other one.

[velvet wren]

brand new account to the server and they have posted once in this channel since joining, probably a scammer

[kind jay]

It’s so good for your brain though

[torpid trout]

Publicly explain your intentions
I might consent

[torpid trout]

My Brain wakes in the morning and about 8 times a day

Thats why I also look like my brain

[torpid trout]

How do you know they posted onxe?
Is that a feature somewhere

[velvet wren]

search their username posts and filter to one user

[torpid trout]

Oh

Thx!

[orchid plume]

plus they have a "new user" badge

[torpid trout]

I saw they new user - didn’t think about the posts filter

Could be a new fan of mine, what do you think?

[orchid plume]

😂

[torpid trout]

I mean. I’m lovable. Jane loves me. Cancrit does. Everyone does. I’m the best never seen ever

[orchid plume]

true lol

[torpid trout]

Duh.

🫳 🎤

I ooze humor it’s almost impossible not to laugh at me

[kind jay]

I don’t, I’m just not allowed to say what I think or I’ll get in trouble

[burnt sage]

ok

[torpid trout]

😭

But Jane

Im your onlyfan

[kind jay]

No

[burnt sage]

@kind jay

[kind jay]

Don’t talk to me

[torpid trout]

Donut be rude.

[burnt sage]

ok

[kind jay]

He's 9y/o

[torpid trout]

Even less so a reason no?

Also you’re assuming gender. How dare you.

[burnt sage]

???

[torpid trout]

Even it might be a machine.

Respect the machines.

[kind jay]

😐

[torpid trout]

He’s verified!

[bright swift]

did limits just reset?

[burnt sage]

im kid @torpid trout

[kind jay]

Discord ToS

[torpid trout]

They gender checked it!

[torpid trout]

Then go play.

Outside.

[kind jay]

That's fine but you shouldn't be on discord, it's 13+

[torpid trout]

Reset wen

[kind jay]

Also don't add random people online

[burnt sage]

@torpid trout ok

[torpid trout]

Specially this channel is 80^

[kind jay]

You're like 50 so shhhhh

[torpid trout]

120 as we figured

[torpid trout]

But jokes beside I’m old enough to marry you

[burnt sage]

@torpid trout @kind jay bye

[torpid trout]

Want to marry me?

[kind jay]

Not your age in fahrenheit 😭

[burnt sage]

???

[kind jay]

Bye bye

[kind jay]

I'm 16

[torpid trout]

Yes

Hence why

[kind jay]

Why what?

[torpid trout]

Why what what?

[unique spade]

I d imagine free is slow, but there are things that can be done fine on slow, just need to know which and plan ahead

[cedar skiff]

Was too slow for me i just grabbed a minimax sub, it can do images as well

[kind jay]

Sub

[unique spade]

Yea for what you actively work on and use a primary agent to do immediate stuff like coding, it's hard to see much value in high latency stuff. Overnight slow and careful documentation can be an example where it could be useful though

[kind jay]

Bruh, how you all surviving on the brokie plan?

[bright swift]

by brokie you mean pro 200 i guess

idk

Create more accounts™

[cedar parcel]

I need one yesterday

[velvet wren]

what plan are you on?

[cedar parcel]

Pro 200. When I switched to 5.5 from 5.2 it burned 52% weekly in a day

[velvet wren]

ouch!

[cedar parcel]

Yeah... But it was kind worth it haha

I use my own profile so I escaped the fact that openai set /fast the default. Anyone especially in cli should make sure theirs is off

[velvet wren]

I only use the macOS app now

[unique spade]

fast is default only on enterprise plans, not personal plans

[cedar parcel]

The setting is more visible there if /fast is on but in cli you won't see it unless you've happened to add it to the status line.

[cedar parcel]

Good to know!

[exotic cave]

This means you would naturally burn through it in 5 days because 52% / 2.5 (fast is x2.5 on 5.5) = 5.5 burns through usage FAST...

[unique spade]

i took it slowly this week. but now i m at 36% with 2.5 days to go so starts to be stressing haha

[velvet wren]

what plan are you on?

[unique spade]

100

[exotic cave]

They said 5.5 is more efficient and use less tokens, that is output. But when working with such a small context window and short cache the majority is Input. For me output was never the main cost.

[cedar parcel]

Yeah but fast was off. And I checked the skills usage (I only issue single prompts for each session via skills) and the day before 5..5 I did much more work (70 skills triggered) for about 20% of my weekly. The day I used 5.5 I triggered 50 skills and burned way more.

[exotic cave]

It was off? Oh well that changes things...

[unique spade]

does anyone make good use of spark ?

i mean in theory i could use that if i reach 0, but that 128k context feels very limited

[fervent anvil]

Mine doesn't.

[exotic cave]

128K context is not something i can use, working holistically on some modules require to gather enough context, 128k sometimes is just the start of the work...

[cedar parcel]

I'm optimising by reducing agents.md to point to docs folders so it stays minimal. I'm doing the thing where I delegate long terminal output generating commands to a sub agent so it won't bloat my context. I'll see on Tuesday if it helps when my limit resets

[unique spade]

yea you can even end up with almost endless with context re-gathering cause compaction hits while gathering context lol

[chrome raven]

so guys how is 5.5 with 1m context?

[exotic cave]

Holy heck...

[unique spade]

saw someone hitting that kind of loop with normal 256k cause context was around 400k

[exotic cave]

This is hilarious and sad at the same time...

[bright swift]

Never saw something like this happening and i used some pretty vague prompts that caused a lot of reading, research and then work..

[chrome raven]

info is mostly not lost from compaction, i think important details are carried and it can read back into previous for info, so so probably this person is new to codex

[bright swift]

I mean the combination of "frontier AI" + "looking for work" + trying to build a jarvis personal AI in python is pretty telling

[unique spade]

It depends if no hard work has been done, the work is more ephemeral and it won't be summarized as well by the compaction model

[bright swift]

Likely very new, little coding experience + light AI psychosis 🙂

[chrome raven]

regardless of the task the compaction will handle it, and it can also read back on previous work so, no issues regardless of the work, you can stay on one chat as long as you want

[bright swift]

Big tool output poisons your context fast. Build wrapper commands if you use them regularly and/or hooks to parse command output for the useful bits

Codex' compaction removes some tool results but not the most recent ones

[unique spade]

The point was about losing early context gathering if compaction happens too early in the turn.

For long thread purposes compaction works very well. I usually keep very long threads myself

[cedar skiff]

This is a task management issue, using sub-agents and phased plans is the solution.

[unique spade]

An losing is a matter of degree, meaning the model might revisit some things that it just did because it s not so sure of it anymore

[cedar parcel]

Yeah I'm curious to see how much it will help.

[unique spade]

Yea

[bright swift]

Also tool output has a (configurable) token limit, so even though you already moved it to a subagent, it might not receive the full output..

[unique spade]

But there is also somehting that is called context management

[cedar parcel]

Another factor is how many messages you send and what is being included in those messages each turn, such as agents.md and ever increasing convo history. Since I use single prompts sessions this should not matter too much for me but for many others it will

[cedar skiff]

when agents hit tool output limits they usually pipe the output directly to a text document if they can, and then wrok from that

[unique spade]

Which if you want optimal task management you want to also have solid context management

[cedar parcel]

It's certainly worth spending some time running analysis and finding optimisation opportunities in whatever harness/framework/workflow people are using.

[unique spade]

Right

It s something that really can gain you a lot

[cedar parcel]

Looking forward to those gains for sure 😃

[unique spade]

How s opus 4.7 doing BTW. Anyone using it?

I spend too much time on linkedin, I see so many Claude Boyz there 😂

[bright swift]

linkedin is usually behind by 6 months or so

i use it for ui work, one prompt every 5h on the tiny pro plan

[unique spade]

6 months is not even that bad

[haughty plaza]

just a quick question, anyone is using api if you have codex? then api is just plain stupid ans way more expensive isent it ? except if you need it for some reason

also anyone on pro and some updates about the recent quota? it says 'unlimited' but thats not true ofcourse x20 but im totally burned out of trial account and was just wondering on a payed upgrade how much quota you would get

as of this week since before it was messy with fast burn of quota

[summer bobcat]

Does anyone know a way to force codex app to recheck the remote version? it's stuck on 122/124 for some reason, while in cli on these machines I have 125 installed and running :)

[unique spade]

what "unlimited"

there s no plan that has unlimited codex

unnlimited refers to chatgpt app

[haughty plaza]

yes well i ment the x20 how that works for people 1 have 3 big mono repros and refactors with quote heavy sprintlanes i just wonder how long i can run with pro - i guess is should try it but just curious

what still bothers me.. very much

is that they can killof your account for any reason and in the terms says you will not be refunded

that bothers me so much - i still havent moved to payed account

and i think they should change that. unless you do something like become the next unab*mber you should not be automatically flagged. although i think it got better now

[bright swift]

you can blow the weekly limit on a pro 200 account in a few hours if you are determined

you can also make it last the whole week for a lot of work

actually you might be limited by the 5h limit to use up the weekly in a single day, not sure

never got flagged despite codex telling me multiple times "i wont help with that" BS

[unique spade]

if you re on free, i suggest first getting a 20$ account

see how much that usage lasts

then just do basic arithmetic

till end of may

100 plan will give you 10x more mileage
200 plan will give you 20x more mileage

replative to 20 plan as 1x baseline

[cedar skiff]

there is a 5 hour limit though

[silent lodge]

Why does codex output all these messages while working, it eats up the output token and how can I turn it off?

[bright swift]

its in the system prompt

[silent lodge]

Can I turn it off?

[bright swift]

not sure if you can override the whole prompt or if you need to fork

there is an instruction like "provide updates to the user every 30sec"

[nocturne folio]

its just like 30 tokens

[silent lodge]

Is codex opensource?

[bright swift]

the cli and app server yes

the app no

[nocturne folio]

why would the app not be opensourced, so odd

[bright swift]

probably because they want to turn it into the flagship everything-app product

[nocturne folio]

rereleasing chatgpt when they could just make it more competent and less sycophant

when gpt5.3 get memory it writes a bajillion things about how im using docker, and mentions it a billion times

[hard drum]

I blew on weekly fast

[nocturne folio]

ahhh i hate it

[hard drum]

I seem to always blow my weekly limit in 2-3 days before the reset

I don’t really know what I’m doing wrong anymore

[nocturne folio]

blow?

just say use

[hard drum]

Not you to tell me what to do

[kind jay]

Blow

[spark bobcat]

Anyone else on W11 stuck on this screen?

[unique spade]

The app is just some ux stuff, the backend is still the codex cli binary

[unique spade]

Bro those are like 0.1% from your token usage 🙂

And honestly if you have to ask here wether codex is opensource, it s probably a good idea to not try yourself to tweak it

[boreal holly]

you been spawning 20 5.5 xhigh 1M /fast --yolo agents? 🤪

[unique spade]

If you had any experience with oss, you would have known already what is oss from codex

[unique spade]

To be honest, if you only use one 5.5 xhigh on normal speed to build stuff, it s probably very hard to exhaust 20x

But sure if you tell it to spawn max agents and get it done, and that "done" is quite vague then sure

[spark bobcat]

excuse-me but

W T F

How gpt 5.5 is 7.5 uses on vsCode copilot

If it has the same input and output and should take less to work

[unique spade]

Special offer for the copilot mob

[spark bobcat]

this a scam

[unique spade]

😂

[unique spade]

Complain to microsoft or use codex directly not via intermediares

[bright swift]

use 5.4 xhigh in copilot at 1x multiplier until thats gone too next month when they switch to token based billing

[spark bobcat]

that's why

Next month i am going to cursor

[unique spade]

Why to use copilot though

[spark bobcat]

Was the cheap option

No 5 hours or week limit

And it was not by token

[bright swift]

1 request = hours of work with 5.4 xhigh, unlimited subagents

[spark bobcat]

But by use

You could just make the agent work for hours with 1 reqquest lol

[unique spade]

Ok then there must be a trade-off that new models are super expensive

[boreal holly]

Cool, enjoy $1.5k/mo for Pro 20x amount of progress

[spark bobcat]

time to

install cursor for vscode

[bright swift]

they put that 7.5x multiplier so they can say its available immediatly, but people dont actually use it, because they have no capacity

[spark bobcat]

since my cursor app dont work

[bright swift]

same reason they removed opus 4.6

[spark bobcat]

codex***

[unique spade]

Yea as we saw even oai struggles with 5.5 capacity 🙂

[spark bobcat]

and now 4.7 is 7.5x

that's a scam

[boreal holly]

Dang, I thought Microshaft had practically unlimited compute

[bright swift]

well, 7.5x in "promotional period", more after 😄

[spark bobcat]

😢

[unique spade]

They probably have, but I guess right now 5.5 is served only from the source

😂

[spark bobcat]

this companies forget that there are poor coders that dont take 120k per year lol

[unique spade]

You do not need latest model if you re a decent coder, 5.4 does the job fine

[spark bobcat]

yes yes

i have been using

sonnet 4.6

and gpt 4

5.4*

for a while

[unique spade]

And you can get it for 20$ at normal rates via chatgpt sub if you really want 5.5 at regular price from day 1

[spark bobcat]

the problem is

my device has some problem with the word codex

i cannot open the vsCode extension

neither the desktop application

[boreal holly]

Install fedora

[spark bobcat]

i think if i try the cli

it will explode

[boreal holly]

You can still play Roblox on Fedora

[unique spade]

You can use pi.dev

Or any other harness that allows you direct login with your gpt sub

Or code your own

Your qouta from gpt sub

And the surface you access it with

Are 2 distinct things

[hard drum]

1. I never use yolo
2. I don’t use xhigh for primary && my setup often involves medium/high splits.
3. subagents
4. 2-3 projects at once
5. I don’t use fast mode, && the command also seems to not exist anymore. If they turned it on, it could explain the recent problem.

[boreal holly]

it could be a lotta things. they like to suddenly make features default enabled. Someone yesterday said every time they turn off fast mode, new agents start with it enabled or something, like theres a second config source competing to set that feature

[hard drum]

Ooof

Then I may want to revert versions holy crap

That’s bad for prod

[boreal holly]

I'm pretty sure you use CLI though, so probably not the same issue

[hard drum]

Yeah, I use CLI

was I supposed to use the app?

[boreal holly]

Nah, the sticky fast issue I think is app and/or vscode (if it's real, they didn't elaborate)

[kind jay]

Can a robot really be a puppy?

@frosty zealot

[kind jay]

You only look once?

[short kindle]

Why cant we see our usage anymore? Ive restarted VSC and ensured it was updated (Extension and VSC) is this a thing now or is my set up just wonky?

[signal tapir]

That happened every time I needed an update, except once when I got it working after a restart. Maybe some service needs a restart or something?

[kind jay]

Hi Jim

[signal tapir]

I mean restart of the entire system

Hi, Jane

[short kindle]

Ill do a full restart and see what happens, thanks jim

[signal tapir]

Best of luck

[boreal holly]

try sending a message

[short kindle]

It was mid working when i screenied that and still doesn't show after sending another message. Im on Pro so I dont really need to check it currently but if others still see theirs, it must be on my end. Ill restart fully in a bit and see what happens

[lost drum]

Ggs I never thouthg I would use whole weekly usage of 20x plan.

If I were to still pay for credits😭

[lost drum]

Me too

[nocturne folio]

clav and sam stream when?

[thorny cloud]

@signal tapir hi jimm

finally decided to sub to gpt plus haha

[hard drum]

Huh?

[hard drum]

POLSKA GUROMMMMM

[kind jay]

YOLO

[hard drum]

oooh

[lost drum]

Wyciep te smieci to hasiokaaa

[hard drum]

nie

wal sie, ty kurw

hehe passed the OAI foreign limit

[signal tapir]

I've had Teams for a long time. The problem is I can't afford anything better. 😛

[thorny cloud]

any tips btw?

[signal tapir]

on what?

[frail meadow]

life

[signal tapir]

Nothing suitable for discord. I'd need a warm summer day, cold beer, and a fired up bbq for that.

[frail meadow]

get yourself a beer

[signal tapir]

I'm saving that for the summer. 😛

I actually drink alcohol extremely seldom. Only on those days in a year where it feels most appropriate.

[frail meadow]

same I drink a couple times a month which I consider somewhat infrequent

[signal tapir]

Right now it is absolutely NOT summer.

I'm at a couple of times per year.

[frail meadow]

high in the 80s over here so not too far from summer temperatures

[signal tapir]

Southern USA?

[frail meadow]

yes

[signal tapir]

Northern Sweden here

we had 46f today

[torpid trout]

└ Stream disconnected before completion: Transport error: network error: error decoding response body

[frail meadow]

do you hit the slopes on the weekends ever? Or is it more of a trip for you to get to resorts?

[signal tapir]

No. Not my cup of tea.

[frail meadow]

fair enough

[signal tapir]

Lots of people around me do though. We've got plenty of slopes around here.

[signal tapir]

They're great lookout points though.

[frail meadow]

are there any tech hubs in Sweden?

[signal tapir]

I guess Stockholm would qualify

about 5h drive

According to GPT we have several tech hubs, for different fields

[frail meadow]

I had no idea Spotify and Klarna are from Sweden, that's wild

[signal tapir]

Many many things are from sweden

Minecraft for instance.

[thorny cloud]

codex usage blah blah blah LOL

[signal tapir]

And like all good music.

[signal tapir]

Yea, I'm back to regular Chatgpt usage

[signal tapir]

Skype. King. DICE. And many more.

Hopefully my company will one day be in that list. 😁

[frail meadow]

yeah I saw Legora and Lovable on that list too. Are you a game dev?

[kind jay]

You

[signal tapir]

Yes.

[signal tapir]

Correct.

[kind jay]

Robot puppy

🤖🐶

[signal tapir]

https://tenor.com/view/k9-doctor-who-scanning-tom-baker-companion-gif-27289195

[gentle harbor]

can someone take away this dude twitter access ?

[kind jay]

I love this guy

[signal tapir]

What on earth is he talking about?

[gentle harbor]

claude beats them on front end but they beat them in coding if i had to guess

[signal tapir]

ah. makes... sense?

[kind jay]

Unc tbh

[gentle harbor]

their UI/front-end still isn’t the prettiest compared to competitors, but our model intelligence is way ahead now here is this better wording

[signal tapir]

I understood when you explained it the first time, but the choice of language is still... weird

I guess I'm a couple of decades too old. 😛

[frail meadow]

it's gen alpha slang, related to the Twitch streamer Clavicular

[signal tapir]

But he's like Genx, isn't he?

[gentle harbor]

isnt that dude gone ?

[frail meadow]

idk I think he's still relevant

do you guys know if codex spark is multimodal? Can it process images?

I'm guessing yes since it can use the computer use plugin?

[dense spire]

How you ladies and gents use Codex app on android phones?

[frail meadow]

^

[signal tapir]

just off the top of my head I guess I would use the CLI app, via an SSH shell

[kind jay]

I was going to say this

Actually I wasn’t

Get some terminal emulator on your device

[signal tapir]

Is the CLI app compiled for arm?

I've never seen it before

[chrome raven]

happened to me

have you verified before?

[kind jay]

Mfw I make phishing content, then get flagged for phishing content

[signal tapir]

oh wow.
"Your organization has hit the limit for advanced features and models.
Your limit resets May 24, 2026. To get more access now, purchase additional credits."

[kind jay]

lol

[turbid axle]

token in, broken out

[signal tapir]

At least codex will be recharged in 3 days

[undone patio]

check this mans hard drive immediately

[chrome raven]

oh regardless your data is probably out there somewhere 😄

[undone patio]

"i dont want my data out there"

uses smartphone, ISP, Discord, Google

[signal tapir]

It's still a good idea to keep track of your exposure.

And attempt to control it as you see fit.

[undone patio]

I prefer to control my digital footprint rather than control its exposure.

Simply don't do shady stuff and you have no cause for concern.

What's OpenAI gonna do, steal my crypto?

[frail meadow]

I still haven't updraded to MacOS 26, should I do it?

[boreal holly]

It has some good features. Im on it, but I hear theyre cooking up a new Foundation Models framework so im waiting for that update

[hard drum]

Not unless you want QoL things && to tolerate Liquid 🍑

[nocturne folio]

idk

[kind jay]

Stop slop posting in codex

[undone patio]

that is art, madame

[kind jay]

#images-canvas

[undone patio]

sloppier

[kind jay]

Don't say that about my cat

😐

[undone patio]

my fruit detective is investigating your cat actually

clearly using steroids

[undone patio]

also relevant to codex as it was made in the app with gpt image 2

look at the details of the text in the scenery

nothing is wrong, truly astounding

[signal tapir]

So, what software do you guys recommend to use between a local llm and your vs code, as well as what should I run inside vscode?

[undone patio]

if you must use vs code, i'd be using codex and opus in tandem

[signal tapir]

What if not vs code?

[undone patio]

i still highly suggest using antigravity or codex app directly

codex app is the goat right now

[signal tapir]

linux

[undone patio]

antigravity then

pretty sure theres a linux build no?

[signal tapir]

thanks. I will check it out. Does it handle connection to the llm directly, or should I still have an middle layer?

[undone patio]

you can add custom models

but it comes with gemini and opus

(if you have a google sub)

and you can use codex through it via extension

also codex is amazing at sprite assets

check out this mob from a duke nukem clone im making

[dense path]

As an Google AI Ultra User, I don't like Antigravity, it is buggy. I think Codex is doing great right now

[undone patio]

i for sure prefer the codex app

antigrav is my fallback

but sadly im out of codex tokens rn

i went full blast on fast mode and highest thinking effort

[dense path]

Same for me. I think that Google is pretty generous when coming to Credits, that is a good point of it

For Opus model not much usage but for Gemini it is much

Never was able to spend it

[undone patio]

i kind of like gemini pro 3.1 compared to opus 4.6 sometimes

its less reckless

opus you kind of have to bridle it

[dense path]

I know what you mean

[keen glade]

Is there any way to the get the "You're out of Codex messages" popup to go away in the vs code extension? There's no X button anywhere from what I can see, only a Upgrade/Add Credits button and I already have a ton of credits

[nocturne folio]

kinda late tbh

just a tad bit late

[lean lark]

Step away from the keyboard. Think about your projects. Design something. Plan the next feature. Your 5 hour quota comes back when you're not using it.

[keen glade]

I've got the credits to keep going though?

[lean lark]

There's is a quota, a throttle, a rate limit of the number of tokens we can consume in any 5 hour period. Just don't blow through that limit. Take it easy, think though your project. Don't ask Codex to do dumb things.

[keen glade]

I think you're mixing up weekly quota and credits. Credits are something you can buy to bypass the limits. What I'm reporting is a bug rather than intended feature

[lean lark]

Turn Fast mode to Standard, use the 5.4 model unless you absolutely need 5.5 thinking. Use Low or Medium, not High, XHigh.

ahhhhh, yes

you Are talking about credits...

I'm sorry. Yes, if there's no Dismiss on that, check GH Issues and report a new bug/enhancement on that.

ChatGPT and Codex have the same issue:

The link with Accounting servers is always delayed, so recent changes aren't reflected or are just inaccurate. But if the info box doesn't have an X option, that's a simple UI bug -as- simple omission.

[spark bobcat]

guys

idk what to do

my codex application does not work

I reinstaled, removed, updated, restarted computer

I am just stuck on this

[inland sonnet]

What happends? You don't explain anything

[undone patio]

use terminal CLI to debug it

[inland sonnet]

its closed source, it won't find anything

unless there is a error log

[undone patio]

just do it

[spark bobcat]

How could i do it?

[inland sonnet]

in cmd just type codex

then u have cli

[spark bobcat]

yes but

i dont think

the cli will be able to solve

well let's see

[inland sonnet]

same haha

[undone patio]

"i dont think this key will fit the door"

doesnt try

[inland sonnet]

waste of tokens

:d

[undone patio]

waste tokens or stare at a useless app, i guess

[spark bobcat]

jjust trying rn

[inland sonnet]

did it work before?

[spark bobcat]

never tried before

[inland sonnet]

https://www.google.com/search?q=Codex+desktop+windows+app+stuck+at+loading+screen

there you go @spark bobcat

[inland sonnet]

1. Reset Local App Data
   This is the most common fix recommended by support and community members to clear a stuck state.
   OpenAI Developer Community
   OpenAI Developer Community
   +1
   Fully close Codex (ensure no Codex.exe processes are running in Task Manager).
   Open File Explorer and go to %APPDATA%.
   Find the Codex folder and rename it to Codex.backup.
   Relaunch the app. It will recreate the folder and ask you to log in again.
   OpenAI Developer Community
   OpenAI Developer Community
   +1
2. Verify System Dependencies
   The app requires specific runtimes to initialize the UI correctly.
   Ensure Edge WebView2 Runtime is installed and updated on your system.
   Check for any pending Windows Updates that might be blocking app initialization.
   OpenAI Developer Community
   OpenAI Developer Community
   +1
3. Clear CLI/Auth State
   If you also use the Codex CLI or VS Code extension, a corrupted authentication token can hang the desktop app.
   Reddit
   Reddit
   Navigate to your user profile folder (e.g., C:\Users\YourName.codex).
   Locate and delete the auth.json file.
   Restart the desktop app to force a fresh login.
   Reddit
   Reddit
4. Known Workarounds
   CLI Login First: Some users found that installing the Codex CLI via npm install -g @digital thunder/codex, logging in through the terminal, and then opening the app resolved the hang.
   Model Switching: If the app loads but a specific thread is stuck "thinking" or loading, try switching the AI model for that thread to reset the stream.
   Disable GPU Acceleration: If you are seeing a blank screen or spinner, try launching the app with the --disable-gpu flag via a terminal or by modifying the app shortcut.
   Reddit
   Reddit
   +3

[kind jay]

Slop post

[spark bobcat]

Someone complaing about slop in a AI server?

[undone patio]

lol'd

[inland sonnet]

Comments slop post, but does not attempt to help

[kind jay]

Posting a wall of text straight from google ai is not helping either lil bro

[spark bobcat]

Well

[undone patio]

who uses google lmao

[spark bobcat]

That's kinda helping

[kind jay]

Apparently not Blakes

[inland sonnet]

[undone patio]

why is there no mayonnaise emoji

[inland sonnet]

cause f mayonnaise

[undone patio]

nah bro

this will start the first AI war

im preparing my agents now

[kind jay]

[kind jay]

I thought this was a mob spawner

[carmine copper]

At least some parts of the codex app source is extractable in some way. On my request, codex was able to get some js code from codex app and tell me the entire sequence of connecting to ssh remote, including exact commands that are executed remotely when initializing.

[kind jay]

Comments slop, but does not attempt to help

[undone patio]

ur lucky this isnt anthropic, dario would use orbital laser cannon on you for even saying this

[kind jay]

Yeah, I was able to make a CLI accessable version that I can use completely from the terminal, all this by extracting and modding the source code of the Windows Codex app

Codex is so much nicer in my modded CLI version

They should really have some tools you can use from the terminal

[carmine copper]

Hey I didn't even ask codex to reverse engineer anything. Just asked it "what does codex app execute on ssh remote after it connects"? After a quick search it started looking at the app files..

[inland sonnet]

xd

[cedar bear]

What makes it more comfortable

[kind jay]

More natural that some app

@frosty zealot DM

[spark bobcat]

worked xD

Asked codex cli to fix the application

he did

thx

[inland sonnet]

prob just deleted app data etc no?

😂

[spark bobcat]

just a auth problem

[kind jay]

Or don't ig

[spark bobcat]

Mine is an multi love
Opus 4.6, Sonnet 4.6, GPT 5.4 and Gemini 3 Flash

[undone patio]

reset would slap rn ngl

[inland sonnet]

Opus was good, just they f it up with there stupid guardrails.

[spark bobcat]

Opus still rlly good

But a lot expensive

So for me, no way to be using anymore

[inland sonnet]

gpt 5.5 is now more expensive

then opus

xD

[spark bobcat]

Yes i was like

Wasnt 5.5 the same price from

5.4?

I didnt understand this

[undone patio]

need more datacenters

[inland sonnet]

No

[spark bobcat]

if i remember when they launched

They said it would keep the price

and be faster

[inland sonnet]

it use less tokens so make it more expensive so its same as 5.4

😂

[inland sonnet]

Ai for everyone in 2030: input 100$ output 100$

base model

[spark bobcat]

And i am sure

It will be less expensive than a 300ml water bottle

[inland sonnet]

cap

[signal tapir]

I don't need better than 5.4. I need cheaper.

[inland sonnet]

Deepseek, Mimo 2.5 pro...

Kimi..

I like Mimo V2.5 pro

its great

[undone patio]

whats the minimum vram for kimi inference

[inland sonnet]

its 1 T so i think alot

[signal tapir]

I only have 16GB VRAM

[inland sonnet]

There is kimi code subs

🙂

[spark bobcat]

We do not need more general models, but models for specific actions

[undone patio]

i need GPT 7

[signal tapir]

I think being so general is what makes them good at specific things. It sounds like a contradiction, but it seems to be the case.

[inland sonnet]

Tbh if you want to be free, Windsurf / cursor