codex-discussions

[dusk thorn]

This lol

DONT DO IT

[lean lark]

I just mentioned an old minicomputer and got modded. whatever. or maybe I put Basic in caps. sigh

[undone patio]

xd

[dusk thorn]

Wait how old are you guys if you dont mind me asking

[frosty zealot]

I would say but then @kind jay would be 1 more step closer to finding me IRL

[undone patio]

lmao

im 36

[frosty zealot]

I'll say I was in elementary when 9/11 happened lol

[undone patio]

the new GPT image really is something special

[dusk thorn]

Is ai kinda something you started learning when it first came out or did you just get into it past year for your job?

[undone patio]

ive been studying ML since about 2020

[dusk thorn]

What a metric

[dusk thorn]

That’s a long time

[frosty zealot]

and im younger than Lofi 😛

[lean lark]

When I grow a beard, I look like Santa Claus. When I shave it I honestly get carded. It's tough to have a baby face. 😜

[undone patio]

ive been making software since about 2020 also, and websites since like 2003

[dusk thorn]

I like tried making a half baked discord bot when gpt 3.5 was released using its api and my bot was coded part with it and it like worked a bit

So I’ve been using it for 4 years

[undone patio]

bro is gonna be the intern with 6 years of experience that we've heard of in legends

[lean lark]

I was hakin games in 1979.

[dusk thorn]

Rn im almost done with junior year high school turned 17 this year

[dusk thorn]

Haha

[undone patio]

truly the captain

[frosty zealot]

I started with HTML/PHP, moved to JS shortly after Node came out, really enjoyed JS just thought it was fun to use, moved to Go for some side projects and enjoyed that too, then moved to TypeScript and that became my favourite 'language', never had a reason to learn low level because most of my stuff was web based etc.

[undone patio]

typescript is pretty goated but i like python the most

[frosty zealot]

dabbled with a bunch of other stuff, Elixir, Python etc. but never really stuck with them

[undone patio]

im doing rust things now

[frosty zealot]

Elixir was much different because it was a functional language and not OOP

[undone patio]

rust is cool

[dusk thorn]

I know python basics but should learn more languages and their basics to help my understanding of codex

[frosty zealot]

Yeah I startd learning rust like a year ago but never really had much projects to use it with

[undone patio]

im converting some open source game engines to rust

for future endeavors

[lean lark]

just curly braces and semicolons to me, i don't care about languages anymore, have forgotten over 20.

[dusk thorn]

I feel so outa place without the purple on the profile photos it’s like this chat likes the color as much as ai does when creating ui’s

[undone patio]

lol

[undone patio]

https://www.youtube.com/watch?v=fBDifUjNzbQ

[frosty zealot]

are you a BigEndian or LittleEndian enjoyer 😛

[lean lark]

End This. 😜

[frosty zealot]

Given your age all I gotta do is like jump scare you or something

[undone patio]

lOL

[dusk thorn]

Where’s my Lego xwing Sam I think we all need a rate limit reset and a Lego xwing

[dusk thorn]

Send a epilepsy gif and it’s over

[frosty zealot]

lmfaaoo

[undone patio]

ima work for OpenAI one day i swear

[dusk thorn]

“Falls to the ground, chair on the floor”

[dusk thorn]

You gotta get training on vague posting on x

[frosty zealot]

forgot what i was gonna do now was laughing so hard

[undone patio]

lol

[frosty zealot]

staring at a blank terminal

[undone patio]

im only good at highly specific posting

[lean lark]

"I'm saying, that when you're ready, you won't need to work for OpenAI" -- Morphius

[undone patio]

"There is no Claude." - bald kid probably

[dusk thorn]

💀

[undone patio]

i wonder if Grok Code is just gonna be an anime waifu painting the app on ur screen

[dusk thorn]

Grok is just Elon tryna do better than Sam but he also wants a waifu

Grok will just become mecha hit—r again and write code in German

[lean lark]

I wish Elon would stick to rockets.

[frosty zealot]

I feel like it really tapered off, he really parades it on 'truthiness' and the video generator..

But they just partnered or something with Cursor? Or are giving them the compute for cursor to train their model or something so

Maybe cursor about to blow up

[undone patio]

yeah cursor is very happy rn

[dusk thorn]

The video generator just is so he has content to jank to

[dusk thorn]

Fr

[lean lark]

x just bought cursor, right?

[undone patio]

they invested in it or smth i forget

[dusk thorn]

Space x did, since it merged with xai, and they will pay them end of year for their help or buy it

[undone patio]

ah yeah

spacex

[dusk thorn]

It’s like 10 for help and like 60 for buying

[undone patio]

i remember thinking "thats weird i thought it'd be xAI"

so there must be some capitalism trickery at play

[dusk thorn]

Grok waifu piloting starship to the moon and then it just explodes

[frosty zealot]

I cant stop turning my cat into pixel art lol

the new image-2 is unreal

[dusk thorn]

That’s a cool idea

[lean lark]

like wth?

[undone patio]

mistral but not NOUS?

so weird

[lean lark]

Cursor had such a great relationship with oai

[dusk thorn]

That’s why he probably did it

He used to not care and now cares about OpenAI

It’s like jealousy or smth

[lean lark]

of course oai is hangin with ms, Oracle, Amazon, and ... forgot the latest ...

too many peeps sharing beds

[dusk thorn]

How does Elon expect OpenAI to not go from non profit to a for profit like someone’s gotta pay for all this compute and he sure wouldn’t either

They still give free ai and local models

[lean lark]

anyway .. late at night when I'm off PC, i wanna code with Codex. Gonna either use cloud UI, or write an Android app. To many ideas and can't run to desk to do um.

[frosty zealot]

what time is it in your time

[undone patio]

yeah i just finished a PR

[lean lark]

I'm in SoCal

[frosty zealot]

oo

[undone patio]

CenCal here

[dusk thorn]

Uhh

Cen Minnesota here

[undone patio]

lol

[dusk thorn]

💀🤣

[lean lark]

When we get to NoCal we'll be diet!

[dusk thorn]

Finally warming up for spring it’s been a cold cold winter

[lean lark]

Mini-mini-apolis?

[dusk thorn]

Yeah around there

Been a great time as you can tell from the news over the past few months

[high girder]

I just realized I'll never need to make one for claude.

[dusk thorn]

Yeah

[lean lark]

Speaking of Android ... When I was taking a break last week I codexed an Android app generator. Scaffolds the whole thing, builds an APK, then i just need to fill in details on individual projects. Feels good. 🐧

[dusk thorn]

Claude after telling people the models were fine for a month, then saying they didn’t work and some stuff failed, and then reset your limits 2 hours before they would normally

Anyone here use chatgpt atlas?

[high girder]

To be fair, I never ran into any of the issues personally. Im not a fanboy for claude, but it's still a good tool. I also tell people to use codex to catch claude mistakes so 🤷

[lean lark]

Claude had everyone's attention recently. I suspect Darius is gonna like that feeling and will pull some business threads to get it back. That means toys for us at a nice price. Look forward to better.

[dusk thorn]

Yeah same, it’s just frustrating when I put 50 bucks or so into extra usage and then it fails for 15 bucks of it and doesn’t do anything but eat my tokens generating a pdf then telling me some tools not working

[magic condor]

I hope we’ll get something like Claude’s dispatch for interacting with Codex remotely via mobile. Mainly for non coding work.

[high girder]

This was fun to build for codex

[dusk thorn]

Not happening until they fix their compute issues

[magic condor]

Eh you have?

[high girder]

Im not a huge fan of the lack of transparency from anthropic. I doubt they would at this point

[lean lark]

yeah, after that. Take care of the home base then come out swinging.

It's the only business move i can see him making...

[dusk thorn]

The fear based marketing of going on the news and telling people their jobs will be taken in like 2 years ain’t helping anthropic either

[high girder]

Absolutely. Taking apart your tools and figuring how they work is half the fun. Once you know what you've got available, you can put things together pretty quick

[lean lark]

well, Sam has been trying to prepare peeps for that too, for a lot longer.

[dusk thorn]

Yeah it’s true but we should think of solutions and helping people use it

[high girder]

I have claude agents that spawn codex agents, codex agents that spawn claude sub agents. I even gave claude the gpt image 2 skill

[dusk thorn]

Rather then putting out stats

Even Dario says we need to “become aware of this problem now” but he isn’t really helping anything

And that was 5 months ago

[lean lark]

It's actually quite easy. But i wanna code late at night without getting up ... or talking.

[high girder]

Philosophically, I build similar to anthropics core values, but even claude keeps telling me id fit better at oai if I were to need a job

[lean lark]

OpenAI was the first and continues to be at the forefront of introducing humanity to the new age. That's why I'm here. The other players are in it primarily for other reasons.

[dusk thorn]

Right

https://www.youtube.com/shorts/oBf-hwovLng

👆

Sam and OpenAI went through barriers for ai to actually be in the hands of most

[high girder]

So the irony is why anthropic was even created in the first place. They didn't like the direction sam was going, only to turn around and do it himself

[dusk thorn]

Many thought it should be a small group

[lament lily]

https://creativecommons.org/publicdomain/zero/1.0/

[lean lark]

I don't fully understand the personalities or motivations, but I have a very strong sense for who I do not want to be leading with something this important.

Hmmm, I'm gonna put an Android front end on JaneBot!

[solemn acorn]

you can make a "rate limits cut" button for claude

[high girder]

Dammit mod mail

[high girder]

I think a turbo button with increased usage would work better

[lean lark]

going around the mod might get you a warning. 😢

[high girder]

I'll delete it, but so be it. Sorry bau baby

[lean lark]

I'm sorry bud, don't want you to get bounced.

[high girder]

Don't apologize lol. You're right. Its also 1 am and the weed is helping with decisions lol

[solemn acorn]

you can set up something close to it with SSH and tailscale, you'd just be limited to the codex CLI

[lean lark]

all that said, i haven't seen a single mod or guide in here, probably because we're mostly adults ... cept fer some whining about quotas. 😜

[dusk thorn]

Someone did make a app to talk to codex on your computer but it’s pay

I forgot the cost tho

[high girder]

You should read the claude discord. They have 10x less people in there, but it's usually chaotic

[frosty zealot]

I've been personally timed out by a mod for a naughty joke, watch out ladies im kind of a bad boy

[high girder]

Some useful info in there sometimes though

[dusk thorn]

Scary

[lean lark]

heck, i could do that before i crash tonight.

[frosty zealot]

was it you that said they watched Caseoh or someone else

[dusk thorn]

Not me

[lean lark]

careful, Jane likes that 😉

[high girder]

You can make it so that you can use voice mode with codex from your cell phone number.

[dusk thorn]

My favorite channel in there is the Claude status and seeing how much crap is broken each day

[dusk thorn]

Fr?

[high girder]

You'll need to give codex some hooks, but yeah

[frosty zealot]

"Codex, I'm in a lot of trouble right now, this is the only call they're giving me from jail.... I need you to bail me out...."

reconnecting 1/5

[dusk thorn]

LOL

[high girder]

I like both being helpful and rude at the same time

[high girder]

Lmfao

[dusk thorn]

Brb guys gotta talk to one of my employees - “PUSH THE UPDATE MAKE NO MISTAKES”

[frosty zealot]

Even better if it calls you directly for approvals and stuff

"Hang on I gotta take this.."

"WHAT DO YOU MEAN YOU CANT..."

"WELL FIGURE IT OUT" hangs up

"Sorry about that .... interns"

[dusk thorn]

LMAO

[lean lark]

We talk about tech and I think "meh, too easy, I like doing cool stuff" but to much of the world this is nothing less than sorcery. I need to punch out a bunch of little tech ditties, slap a price tag on um and support my token habit. Y'all do the same!

[high girder]

I still can't believe some of the "talent" that actually got into the hackathon.

[dusk thorn]

Brb I gotta make a meme with images v2 for this

[high girder]

Im not salty at all 🧂

[gloomy veldt]

Here the codex app for Linux?

[lean lark]

not yet, Sys

[gloomy veldt]

But I'm running Linux machine

[lean lark]

OK

it's still not here

[frosty zealot]

I thought that was going to change everything

[gloomy veldt]

Wine / Proton is possible

[lean lark]

Windows/WSL or Mac

[undone patio]

its a nice app but you can get by with terminal

[high girder]

Don't bother. Download the Mac version and the cli. Have codex use electron to port it to Linux

[gloomy veldt]

Basically is just user interface, and still requirements cli

[frosty zealot]

[lean lark]

You don't really "need" the app, it's just another UI... kinda ...

[plush harbor]

lol. I'm getting codex to do me a thing on my 404 page to suggest new and its default in all its examples is "flibbertigibbet"

[dusk thorn]

This is what first came into mind

[lean lark]

Anyone know if Wine can handle the app?

[high girder]

I dont own a mac to test with

[solemn acorn]

there's a script on github you can use (or ask codex!) to repackage the mac version of codex for linux

no automatic updates is the main downside

[gloomy veldt]

Thanks

[frosty zealot]

Gets you into a 5 way call with the sub agents "This little idiot keeps ..."

[solemn acorn]

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

[dusk thorn]

zoom call with the whole sub agent team and the manager codex and just ask what went wrong in the push today

[frosty zealot]

lmao

[lean lark]

Need to use an Indian voice...

[dusk thorn]

LOL

[lean lark]

Hello, my name is Orchestrator, how can I help you today?

[frosty zealot]

nah gotta use all the weird western names the sub agents spawn with

but indian accent

Ello my nem is Herschel

[lean lark]

"Orca-sta-deray-tor"

Guy with obvious Bangalore accent "allo? My name is Steve."

So yeah, I wanna code late at night, so gonna put Android ui on JaneBot, do optional voice in and/or out, and she'll interface with all my repos from my pc or VPS. sounds good.

[high girder]

I just sent this man to anthropic support. Won't see him for a month xD

[lean lark]

and on that, I'm gonna split too. thanks for the laughs and tech jabber.

[high girder]

See ya!

[sterile crow]

https://x.com/bytes032/status/2047935375499256106?s=20

[oak trellis]

3:22 PM

Still here. The system looks clean.





3:22 PM
i mean i talk to you
3:22 PM

I’m here.





3:22 PM
i said i miss gpt 4.0 .. was more human but i know its in you
3:23 PM


I get it. I can be warmer without getting sloppy.```

smh ..

its still there but caged ..

[plucky halo]

This was a very surprising answer from chatgpt

[lucid mason]

GPT doesn't need to be "warm", it needs to provide answers. It's a tool, not a partner.

[plucky halo]

#chatgpt-discussions

These options aren't available in Codex

[lucid mason]

model_reasoning_summary = "concise"
model_verbosity = "low"
personality = "pragmatic"

For Codex.

[plucky halo]

That's right - you can set some personalisation in Codex. However, the configs you showed earlier like "Warm, enthusiastic, emoji" etc aren't present in Codex

[lucid mason]

Yes, they still underline my personal opinion that a LLM bot doesn't have to be "warm". If it can be personalized that way, fine, but given how much money is burned in LLM development that's not something they should be judged on. Even less so with Codex.

I removed the screenshot.

[plucky halo]

Ok, but this is the channel for Codex and the option for 'warm' responses isn't available so that point is moot. I wasn't saying it wasn't a valid point of view - in fact, I agree with you! I was just pointing out that this wasn't the right place for it

[lucid mason]

I was answering to a post in this very channel. And maybe be a bit less rigid after you just posted a non Codex ChatGPT screenshot right before my own post?! You made your point, I removed the screenshot.

[rustic walrus]

anyone using hermes on 5.5

[plucky halo]

Ah, I didn't see that it was replying to anything. Also, mine is relevant to codex as it was from a short conversation with chatGPT about the efficacy of Claude Code and Codex and it told me that the "choice isn't close. I should use Claude Opus 4.7 (max) (over Codex/gpt-5.5(xhigh))"

[lucid mason]

How can I make Codex in VS Code show usage status again? Sometimes it does, but most of the time it does not anymore.

And even in its config I have to hit this button three times (reprodicibly) to get usage data:

[unique spade]

Not sure mine got broken too after the updates to get 5.5

[lucid mason]

It's fixed temporarily after hitting the config Retry button 3-5 times until it returns data and then restarting VS Code. But that's rather inconvenient.

Any news about model_context_window being fixed to at least support the 400k advertised?

I guess they tried to fix model_context_window allowing arbitrarily high numbers and then broke it to not work at all anymore.

[cedar skiff]

I think the advertised context is input + output

They had the same thing with 5.2

[lucid mason]

258K shown in Codex ≈ total effective window Codex will manage
not 258K input + extra output

reasoning tokens still occupy the context window while generated and are billed as output tokens.
...
You still need headroom for output/reasoning. OpenAI recommends reserving space for reasoning and outputs rather than filling the whole window with prompt/history.
Does setting model_reasoning_summary = "none" make any difference for this?

[obsidian socket]

The product labels the quota as a “5h limit”, but under ordinary Codex use the quota fell from 100% to under 25% in around 78 minutes of wall-clock time. No ordinary understanding of “5 hours” matches that behaviour. On 25 Apr 2026, Codex displayed a “5h limit”. Starting at 08:55 BST with 100% remaining, the quota showed 42% remaining by 09:57 and warned of less than 25% remaining by 10:13. This is approximately 78 minutes of wall-clock time, not five hours.

[prisma flare]

No, it still generates but doesn't shiw

[lucid mason]

That matches GPT's own response:

No material difference for the Codex context size number.

model_reasoning_summary = "none" disables reasoning summaries; it does not increase the model context window, and it does not disable reasoning itself.

[cedar skiff]

I dont think that answer youre getting is correct and currently just like all of the other models that claimed 400k 5/5.1/5.2/5.3 they also have 260k ish context output. They all claimed 400k.

[kind jay]

I already guess from photos

Anyway @frosty zealot looking for a belt for your truck?

[kind jay]

Collected

[lucid mason]

Which shows that the documentation is lacking.

[cedar skiff]

With that said you can just set it to 400k on 5.5 and 5.4 because they both support 1 mil, not sure of the cost though, usually charges more after the default amount.,

[lucid mason]

Any setting above 258K is ignored for 5.5 in VS Code. For 5.4 it was the other way around, you could set 5M and it would even show that number in its status window.

GPT:

I would either remove model_context_window entirely or keep it at/below what Codex actually honors. There are also Codex issues reporting custom model_context_window values causing or worsening compaction failures, so leaving the default is probably safer for now.

[cedar skiff]

where does it say they limit in in 5.5?

[lucid mason]

OpenAI’s GPT-5.5 announcement explicitly separates the two: Codex gets 400K, while API developers get 1M context.
...
For GPT-5.4, OpenAI’s Codex changelog says GPT-5.4 in Codex had experimental support for the 1M context window.

[lucid mason]

Doesn't matter what they say, I am reporting observed beahvavior. Setting model_context_window above 258K does nothing in VS Code (according to its status window), only lower values affect the reported size.

[cedar skiff]

That doesnt mean much, just means ui isnt reflecting changes

[lucid mason]

Therefore, if /status shows 258400, I would treat that as a real Codex effective-session value, not merely a VS Code display bug. It may still be a Codex/catalog bug relative to the advertised 400K window.

[rocky fog]

lol, that works
4 times
I just thought thats broken or that it only works with API or something
never had it load in vscode extension
I guess I never tried to "Retry" 4 times lol, then it shows up

[cedar skiff]

maybe its only in api now

[lucid mason]

Should still be 400K in VS Code, like I quoted earlier.

But it seems bugged.

[cedar skiff]

Like i said it input + output

[lucid mason]

That doesn't tell us anything other than 128k max output tokens are likely part of the 1,050,000 context window.

[cedar skiff]

Im not saying taht about the image

Im just showing it is 1 mill

and maybe its only api

[lucid mason]

For API, yes.

[cedar skiff]

But the older models also claimed 400k

the ones that did not support 1 mill

[lucid mason]

5.4 explicitedly had "experimental" support for 1M in VS Code.

[cedar skiff]

system cards claims 400k

and like 272k output

[lucid mason]

And VS Code implementation was bugged in that it allows to set any arbitrarily high value, even far above 1M (I tried up to 5M).

[cedar skiff]

I was digging around like you are now 6 months ago with this

[lucid mason]

And I was using 1M in 5.4 and it did not compact early. Now it does.

[cedar skiff]

I undserstand, and it claims 1 mill but probably only api it seems

oh i see this in here
https://developers.openai.com/api/docs/models/gpt-5.5

Maybe that does something

Try the switch and see

[frail meadow]

thoughts on something like ralph loop for gpt 5.5? I feel like it almost never completes or even attempts to complete the full plan (if it's complex) on the first pass

[kind jay]

Will codex ever work?

Why say it’s operational on the status page if it’s not?

[frail meadow]

currently operating rn

[hollow gull]

Could we please have the Codex Mac app team work on Atlas for a few sprints?

[bright swift]

wen reset

[bright swift]

isnt it abandoned and now integrated in codex app?

im serious, i thought that was the strategy lol

[hollow gull]

Codex has browser integration for development, but it can’t replace a fully fledged Chromium browser.

[hollow gull]

And this

[sleek spoke]

5.5 lmao... I wasn't even trying to get it to do this

[hollow gull]

Which plan are you on?

[sleek spoke]

Pro lite

[bright swift]

thats quite long, what did it do?

[sleek spoke]

I'm trying to figure that out 😂

[hollow gull]

I believe it would have been stuck on a tool call or awaiting approval.

[lost drum]

Wen 5.5 codex?

[bright swift]

yea i ran sth overnight and woke up to a user_input call it did 30min after starting that was sitting there for 7h while i slept. lazy clanker

[stoic bay]

WHY???😭

[bright swift]

at least the other subagents continued running

[hollow gull]

I still find Cursor to be more efficient in terms of token usage and getting real work done.

[bright swift]

i never used it but heard good things about the orchestration and cloud agents, should probably give it a try

[hollow gull]

https://x.com/edwinarbus/status/2033625866350334333?s=20

[bright swift]

how is their composer model compared to gpt 5.4/5.5?

[hollow gull]

it not a composer comparison, it's a harness comparison

composer is good but not close to gpt 5.4/5.5 in terms of raw intelligence, but the harness game that cursor is playing is a gem

[turbid axle]

yo can we get a reset, I only have 34% left, and 3 more days to go

[hollow gull]

their harness is the USP

[bright swift]

yea i understand that, but i'm curious if using composer in cursor could take on some cheaper routine tasks

just to reduce costs

[turbid axle]

running 5.5 on medium. pro aint what it used to be, thats fo damn sure

[kind jay]

Fake lil bro

Fake news

[frail meadow]

How did you do that? Mine is lazy and hasn't worked for over an hour this entire time and I have to keep pushing it to do the original thing I asked it for lol

[turbid axle]

at this rate the dream of empowering the masses is completely dead. it will turn software into a pay to win scheme. skill won't get you anywhere

[kind jay]

Low-key at least you’re getting to use it

[hollow gull]

Yes it does, The problem is that big labs are pushing their massive models for everything, while 85% of the work could be done significantly cheaper if orchestrated properly. But I suppose if they don’t force-feed us these high-compute models, how else would they fund the infrastructure for their next big release?

I'm out of my codex usage for this week, have shifted to cursor free for a few tasks and it's working just fine

[turbid axle]

low key distopia

I still have a business acount in my backpocket, but that will run out with 2 prompts or so

[obsidian socket]

Codex is officially PANTS! and that is polite. . 90 mins or there about usage this morning, and what a surprise , and the OAI answer is spend more money, my question is 5 hours = 5X60 mins where there are 18000 seconds in a 5 hour session>? this is not workable. someone from OAI must read the messages here and see they are telling porkies to their customers......

│  >_ OpenAI Codex (v0.125.0)                                                     │
│                                                                                 │
│ Visit https://chatgpt.com/codex/settings/usage for up-to-date                   │
│ information on rate limits and credits                                          │
│                                                                                 │
│  Model:                gpt-5.4-mini (reasoning medium, summaries off)           │
│  Directory:            ~/ELIZA                                                  │
│  Permissions:          Custom (workspace-write, on-request)                     │
│  Agents.md:            AGENTS.md                                                │
│  Account:               removed  (Plus)                        │
│  Collaboration mode:   Default                                                  │
│  Session:              019dc39f-fafa-7712-89e3-afa96c981e16                     │
│                                                                                 │
│  Context window:       25% left (196K used / 258K)                              │
│  5h limit:             [░░░░░░░░░░░░░░░░░░░░] 0% left (resets 13:54)            │
│  Weekly limit:         [█████████░░░░░░░░░░░] 45% left (resets 23:10 on 29 Apr) │

[turbid axle]

ai expensive yo

[hollow gull]

Definitely need it—with how resource-hungry the 5.5 model is, I can barely get a long-running task done before my 5-hour limit runs out.

[eager forum]

Yes business plan burns out for 1-2 prompt.

[hollow gull]

Try cursor with gpt-5.4-mini or composer, it gets most of the work done if orchestrated properly

[turbid axle]

on plus you must use 5.4-mini

if you want to do any reasonable work

[obsidian socket]

cursor, what is that/ I use bash

[hollow gull]

And the problem is there's no pro business plan! I had to get 3x business plans to get my work done but switching to each one of them every hours is a pain

[hollow gull]

try fish

[eager forum]

Same here, i got 4 accs

[turbid axle]

its paytowin

we need loot boxes

[rocky fog]

guess then its API time

[eager forum]

Switching between annoys

[turbid axle]

random token drops

[hollow gull]

just today I shifted to cursor free, trying it from a couple of hours now able to get my work done if orchestrated properly. but with the open-souce devlopment, I'm confused if I should get cursor pro ($20) or get the Ollama Cloud ($20) and run K2.6-thinking over there

[hollow gull]

not willing to pay $5 input + $30 output

[turbid axle]

those new chips better make my token cost peanuts

[eager forum]

I would not like to count my tokens usage but concentrate to my works

[hollow gull]

Exactly that's why I'm not willing to enroll into their Enterprise plan

[eager forum]

Subscription is better to me

[hollow gull]

This ain't PUBG pro

[obsidian socket]

the words on the screen say 5h limit: if we can understand the gap between 18000 and 5200 seconds that would be a great starting point, to have an app that tells the truth and how you actually used the credits. not oops you used all the credits now by more or bye till tomorrow then a weekly limit that lasts 4 days. there was a time not so long ago that I could code for a whole day and eventually run out of credit. now its well. Enterprise hahahaha, joke its $$$$$$$$ OAI making out to be poor while giving away to Doctors while the ones that pay suffer

[turbid axle]

is it though?

[hollow gull]

Someone has to pay for the limit-reset

[turbid axle]

everything will turn into a pay to win scheme in the end. that is the law

[obsidian socket]

would not mind, before xmas I could use it all the working hours of the day, then one day. oops limits happened, they never existed pre Nov 2025 or not that was ever felt

[rocky fog]

oh never noticed this

[turbid axle]

quick, invite all your friends, lets get to another million new subs

[obsidian socket]

like joe blogs wants to pay $200 a month.,. its a joke. old Joe was happy with his $20 a month subs and cant validate paying 10 fold for what used to be included part of plus package.

[wicked briar]

anyone notiched the UI on the chatgpt web is really good but on the codex side its like really doo doo

@teal cargo

[rocky fog]

just it seems that its currently "boosted" so I guess it will be less after that
although in "Learn More" there is nothing about May 31 for the 200$ one

oh, the "learn more" should just scroll down, I guess it links to outdated page

[obsidian socket]

love this is that simply using the platform ..... because use becomes abuse very quickly when you use it properly.

[hollow gull]

If you want to know the real pricing, do math with API token usage-pricing that is the real pricing, rest everything is just a scene to get more public

[obsidian socket]

lol. when did Plus $20 – For lighter use. Try advanced capabilities like Codex and Deep Research for select projects throughout the week. become liter usage? it was sold as get more out of the service. now its lite

[tall notch]

Is there like beta testing of any OpenAI products, API or others whatsoever?

[obsidian socket]

this shows usage fall and now capped beyond useless

[turbid axle]

its joever

[obsidian socket]

and there is no code review in there

[turbid axle]

they got us addicted, and now they upping the prices

[obsidian socket]

its not addiction its SAAS trap.

[turbid axle]

and at this stage, if you dont use ai, your development cycle is basically at a standstill in comparison

[hollow gull]

Hey guys let us stop being negative, I'm feeling they might be recoding our chats and sharing the usage limits based on that

OAI is good

[obsidian socket]

I dont care, a simple UK GDPR SAR request gives them 30 days to comply and then ICO hello we have a problem

[hollow gull]

OAI is doing amazing work

[turbid axle]

we tokenbegging now?

truly

[hollow gull]

I've a feeling this guy's a bot

[obsidian socket]

hahaa, really? they have the word "OPEN" in the name because they started as a not for proffit did they not?

[obsidian socket]

me? no im just peeved I cant do the work I was doing last week. and that 90 mins = 5 hours

[hollow gull]

not you, the other one

[obsidian socket]

Ah 😉 lol

[turbid axle]

I mean, its not oai's fault. its just the way of the world

no matter how cheap, free, or open they make it. hierarchies will form.

[obsidian socket]

my token use is purely a plus account user. no.. its seeing the 256gb ddr4 ecc ram selling for silly money.... I have 256gb sitting on the stairs right now.. but it wont fit in my GPU sadly.

[rocky fog]

still at least they gave it to the world when it could have stayed enterprise/private only
others only followed that
although not open source

[kind jay]

Am I actually just doomed or what?

[hollow gull]

I'm planning on trying out Kimi k2.6 Thinking on Ollama cloud

[kind jay]

I send a message... "thinking...", check network no data

[obsidian socket]

lol ```entoo ~ $ free -h
total used free shared buff/cache available
Mem: 125Gi 85Gi 34Gi 3.8Gi 12Gi 40Gi
Swap: 15Gi 15Gi 851Mi

[hollow gull]

it's coding performance is better than GPT-5.4-xHigh

[obsidian socket]

yes. and watch 5 hours credits burn

[kind jay]

lol what?

[hollow gull]

https://arena.ai/leaderboard/code

Kimi at #6 and GPT-5.4-xHigh #13

[obsidian socket]

oh qwen @ no.10 and then 1000 credits became 100 and lasted 90 seconds

[hollow gull]

Anyone has any exp with Ollama cloud?

[turbid axle]

I guess these token prices will force us all into building our own compute farms and just run chinese models.

[obsidian socket]

actually I wonder how codex-cli works with ollama cloud and how long that lasts

I use olllama, let me do a test.

[hollow gull]

do you have the pro (#20) version?

https://ollama.com/library/kimi-k2.6

https://ollama.com/library/glm-5.1

[obsidian socket]

no. juar dumped arch for gentoo.. let me get ollama from source and see ```gentoo ~ $ sudo emerge -pv ollama

These are the packages that would be merged, in order:

Calculating dependencies... done!
Dependency resolution took 0.72 s (backtrack: 0/20).

emerge: there are no ebuilds to satisfy "ollama".

emerge: searching for similar names...
emerge: Maybe you meant any of these: app-emacs/llama, dev-python/colorama, app-misc/solaar?

[kind jay]

What’s the problem? I used to use this

[hollow gull]

please try anyone one of them and let me know your feedback with codex/claude code harness. I'm planning to get the paid plan for ollama

[hollow gull]

the paid one? or free?

[kind jay]

Use Ubuntu 22 for this also

[hollow gull]

You can only run this on Ollama cloud and it's paid ($20)

[kind jay]

Free ig, local version

[hollow gull]

I've also tried it but limited with by Mac GPU, and can't run K2.6-Thinking locally

[obsidian socket]

yea local is slow tried codex with a baby rtx3060 it no work well

[hollow gull]

Gentoo doesn’t have official coverage for everything, Install from upstream

[obsidian socket]

officially.. nope.. ```sudo emerge --ask app-eselect/eselect-repository

sudo eselect repository enable guru
sudo emerge --sync guru```

echo "sci-ml/ollama ~amd64" | sudo tee /etc/portage/package.accept_keywords/ollama
echo "sci-ml/ollama cuda" | sudo tee /etc/portage/package.use/ollama```

```echo 'CUDAARCHS="86"' | sudo tee -a /etc/portage/make.conf``` 

then ```sudo emerge --ask sci-ml/ollama```  wont be long

[hollow gull]

That’s basically pulling it from the GURU overlay, not the main Gentoo tree.

[obsidian socket]

I know., had to build a venv this morning for tensorflow because gentoo lags..... . may go back to Arch at this speed.

yay ``` sci-ml/ollama-0.20.3::guru USE="cuda -blas -blis -flexiblas -mkl -openblas -rocm -vulkan" AMDGPU_TARGETS="gfx90a gfx908 gfx942 gfx1030 gfx1100 gfx1101 gfx1200 gfx1201 -gfx803 -gfx900 -gfx906 -gfx940 -gfx941 -gfx1010 -gfx1011 -gfx1012 -gfx1031 -gfx1102 -gfx1103 -gfx1150 -gfx1151" CPU_FLAGS_X86="avx avx2 avx_vnni bmi2 f16c fma3 sse4_2 -avx512_vnni -avx512f -avx512vbmi" 108496 KiB

[hollow gull]

You got the pro version?

Just let me know the feedback, n I'll get one too

[obsidian socket]

no. however it may need be

I think thats doable ```Everything in Free, plus:

Run 3 cloud models at a time
50x more cloud usage than Free
Upload and share private models

[hollow gull]

Star wars?

[kind jay]

Gento larp

[obsidian socket]

maybe I dont want to Each plan has session limits that reset every 5 hours and weekly limits that reset every 7 days.

[hollow gull]

Go ahead! ALL THE VERY BEST!

[kind jay]

Trynna go GB for GB?

[hollow gull]

I'm getting cursor pro then!

[obsidian socket]

thats not nice

whats the electric bill like?

[hollow gull]

They got SOLAR

[kind jay]

I think it's same as two 4090s? Ampere is more power efficient

[hollow gull]

btw what are you running on those four?

[kind jay]

This is so annoying

[hollow gull]

Need better codebase indexing

I belive cursor has it, will be trying it out today

@obsidian socket any update with the Ollama cloud + Kimi K2.6-Thinking + Codex harness?

[kind jay]

Usually training, but I like playing with some things from time to time

[hollow gull]

Ohh I missed, it's stuck

cmd+q and restart

[obsidian socket]

│ >_ OpenAI Codex (v0.125.0)                        │
│                                                   │
│ model:     kimi-k2.5:cloud low   /model to change │
│ directory: ~                                      │
╰───────────────────────────────────────────────────╯

  Tip: New Build faster with Codex.

■ unexpected status 403 Forbidden: {"error":"this model requires a subscription, upgrade for access:
https://ollama.com/upgrade (ref: ad3f788b-80c7-4111-9b71-9ae5e0ceb3c7)"}, url: http://127.0.0.1:11434/v1/responses,
request id: ad3f788b-80c7-4111-9b71-9ae5e0ceb3c7
 

[kind jay]

I haven’t got a single prompt to run in 2 days

[hollow gull]

Yes you need the $20 subs for this

[torpid trout]

Jane, do you never sleep?

[kind jay]

I live on a server rack, no need to sleep

[hollow gull]

Which timezone are you guys in?

[torpid trout]

That's what I suspected

You ram a lot?

[torpid trout]

nice try, fbi..

[obsidian socket]

its Subscribe to Ollama Pro £15.37 per month, is it really going to help it seems have same 5 hour limits

[torpid trout]

Where you from Aryan, curious name!

[kind jay]

Can’t you just run this locally? What’s your goal?

[hollow gull]

Now I understand what was running on those 4xA5000s, it was Jane

[kind jay]

You haven enough ram to store the model, if you have a decent CPU, you can get away with it

I used to do this on my laptop

[hollow gull]

I believe the limits are going to be a lot more relaxed as Ollama is just giving you the inference they don't want to earn from this for model traning, it just simple inference

[obsidian socket]

m2 or 13900k .. m2 wont cut it

[hollow gull]

it's 1T+ parameter model!

[kind jay]

13900k for sure because I ran on like i7 10th gen?

[hollow gull]

https://huggingface.co/moonshotai/Kimi-K2.6

[kind jay]

These are huge wth

[hollow gull]

How do you think it's able to beat OPUS-4.6 on coding benchmarks?

[kind jay]

What happened to old Gemma models?

[hollow gull]

Gemma won't help you with coding

[frail meadow]

if you want codebase indexing in codex you can try Augment Code's context engine mcp

[kind jay]

Where is your coding model?

Don’t need it

[hollow gull]

I heard of them in Ads never tried, howz your exp. with it?

@obsidian socket ollama pro done?

[obsidian socket]

weighing up throwing money at it

[torpid trout]

codex enabled with that js_repl does things like

Ran node -e "let id=0;const send=(ws,method,params={})=>new Promise((resolve,reject)=>{const msg={id:++id,method,params};const on=e=>{const data=JSON.parse(e.data);if(data.id===msg.id){ws.removeEventListener('message',on);data.error?
  │ reject(new Error(JSON.stringify(data.error))):resolve(data.result)}};ws.addEventListener('message',on);ws.send(JSON.stringify(msg));});(async()=>{const tabs=await (await fetch('http://127.0.0.1:9222/json')).json();const ws=new
  │ WebSocket(tabs[0].webSocketDebuggerUrl);await new Promise(r=>ws.addEventListener('open',r,{once:true}));await send(ws,'Page.enable');await send(ws,'Emulation.setDeviceMetricsOverride',``` and for each slight new version asks for permission

Is that somehow avoidable? Without giving it full access? I feel it is wrong to ask persmission for each?!

[hollow gull]

I'm confused between this and Cursor, Should we do one each?

i get cursor pro and you get Ollama pro? and share the exp. ?

[frail meadow]

pretty solid but the free plan only gets you so far. However the subsidized cost from codex vs. cursor charging API costs saves a lot. I have a friend using cursor that's dropping at least 1k in extra usage a month

[hollow gull]

But cursor also has a better overall harness not just indexing, I can check if I get the same accuracy with codex + Augment code

[obsidian socket]

lol. its catch22, stop being lazy and remember how to google the answer + ai lol. or pay to have it done and then cuss when it does its own thing and not what was asked or go out fly the drone or fishing and dont catch fish again

[hollow gull]

Why about the deadlines that are based on me using AI?

[obsidian socket]

who wanted indexing with codex? https://www.npmjs.com/package/@code-indexer/mcp npm install @code-indexer/mcp then add this to .codex/config.toml ```[mcp_servers.code-indexer]
command = "npx"
args = ["-y", "@code-indexer/mcp", "/path/to","/tank","/home/","/opt","/srv"]

[hollow gull]

Going back to code, Thanks for the fun 🙂

[torpid trout]

does someone have a solution?
I dont care how it does browser testing, but this seems idiotic

infinite handholding

[bright swift]

--yolo

[torpid trout]

_Without giving it full access? _

[bright swift]

why would you not give it full access

if you are scared about it breaking things, just run it in a vm?

[torpid trout]

It is a lot more complicated than that
you cant just run 10 years worth of development in a vm, even just the tedious sync would kill me

I just do not understand why they "start with" (the full command)
When it does preview with headleses browser its a repeating "starts with" command, so you approve it once, not for each silly new pixel
I dont get it. What do you all use for browser testing? Yolo certainly is not a solutoon?!

[cedar skiff]

what about auto-review?

[supple perch]

I made a program that puts them in a container that can mount the folder as if its real

[bright swift]

What does 10 years worth of dev have to do with it? And why sync? For looking at the code?

I use chrome devtools mcp for browser stuff currently

But seriously I would recommend just going with yolo mode, in a container to be on the safe side, not just for destructive commands, but mainly for supply chain attacks and prompt injection risk

[supple perch]

(if you want a link to mine am happy to share)

[torpid trout]

what is auto review

[bright swift]

I'm moving my setup from local dev to remote envs, mainly so i dont have a 100° C macbook from all the compiling

[bright swift]

New codex app feature that runs a second agent on top of the normal loop to use up more tokens. Its called guardian in the cli

[torpid trout]

You have a codebase, you edit in vm. and then? You want to open that in your VC to do manual edit. you want to test it in locl servers. thre is a plethora of reasons why this is extremely disruptive?

[torpid trout]

thats app

CLI does not have that I think. at least I am not aware of it.

[bright swift]

CLI has the building blocks for it, which are called guardian

[torpid trout]

oh, but that has nothing to do with testing

Its the thing that approves comamnds on its own

[frail meadow]

I wonder how this would be improved if a reranker was added

[bright swift]

I agree it can be disruptive but its solvable. For example I have the remote envs behind cloudflare tunnel/access + tailscale so i can test them just like i would locally

[frail meadow]

isn't that what dev containers do?

[bright swift]

I mean codex even ships with a devcontainer setup

[supple perch]

I mean you could use that but the program prs and does issues and works with almost all agents clis

I like being able to toss it out like how old codex cloud worked and then a pr shows up on the repo when its done

[torpid trout]

tell me more?

[bright swift]

I dont personally use it, i built my own setup. But it might allow you to test our your setup with yolo faster

Btw for browser automation i literally just stumbled on this.. https://github.com/h4ckf0r0day/obscura

But depends on the browser usecase

[wicked briar]

do you use this mode? auto review

I prefer full access mode

[cedar skiff]

i currently am, had one hiccup where it let 5.5 edit official checkout of framework code

[frail meadow]

Here's some resources for you, not sure if it's what you're looking for:
https://developers.openai.com/codex/concepts/sandboxing
https://developers.openai.com/codex/agent-approvals-security#run-codex-in-dev-containers

[obsidian socket]

sandbox's are great however if you trust codex to do the right thing then codex --dangerously-bypass-approvals-and-sandbox resume ......... no sandox can do everything and unless you tell it to it wont delete / 😉 usually

[frail meadow]

trying to help the peeps who don't trust codex completely yet

[obsidian socket]

ill be shush, , it helped me out so much though.

talking about trusting codex! ollama + local qwen3.5 seems to be immediate solution to silly 5 hour limits that are really 90 mins of work

[cedar skiff]

kimi, minimax and glm are cheap for what they provide

[frail meadow]

anybody tried out deepseek v4 pro yet?

[obsidian socket]

weighing up that cost. its plus account + another account .. from what I read ollama pro is well worth it. this works.. kind of once I get the metadata issue fixed

[kind jay]

Yeah, I have a local instance, it's peak

[obsidian socket]

peak with almost a gb of local vram 😉 I can imagine hehhehee

[frail meadow]

do you run it on a mac studio cluster?

[obsidian socket]

@kind jay speedtest results? vs vram results?

[kind jay]

😂

[cedar skiff]

what model?

[frail meadow]

i think she lying lol

[finite sun]

Haha yall mean

1.6T parameters? 😭

[frail meadow]

49B active though

[boreal holly]

Bro, if you have a mac, you can install Facebook's idb tool and give the agents a whole entire iOS simulator, including web browser, or legit native apps to pilot and develop on

Other systems you can use adb but there is lock contention, and qemu uses a lot of resources

[clever steppe]

Urgh. More teething issues.

Is Codex down right now for everyone else, too?

In VSCode it's taking forever to load chats.

[gentle harbor]

why have they still not reset usage im so sad, i thought they would reset it with the new model coming out

[clever steppe]

They do that??

[gentle harbor]

they did it before

[torpid trout]

reset wen

[bright swift]

did i hear reset?

[lost drum]

I think I broke my VS cause when I click commit it freezes.

[bright swift]

vibed too hard, it happens

[lost drum]

Glad my codex those not rely on it and I have backup tar.gz

[lost drum]

Why do you want it? Its same as id you were to wait cause the only thing it those is extend the waiting for next refresh

Better to wait till 28th so it auto resets

[bright swift]

I need to go on a trip and want to burn some more usage before that

Also this new trend of no resets is concerning!

where is my free stuff

[fossil bison]

Has anyone experienced a situation where, when sending a command or asking a question to Codex, it results in an error like this, or when Codex responds, it causes a delay that requires waiting for a moment?

[vocal goblet]

anyone else gets an error while trying to use GPT 5.5?
{"type":"error","status":400,"error":{"type":"invalid_request_error","message":"The 'gpt-5.5' model is not supported when using Codex with a ChatGPT account."}}

[lost drum]

Nope

[lost drum]

I use codex extension in VS and gpt 5.5 xhigh it works for me

[exotic cave]

@boreal holly How did you find 5.5 for your workflow? or it's too damn pricy for your use case?

[boreal holly]

Bro, my usage has decreased since switching to 5.5, with roughly the same number of agents. I think it's generally more efficient. It calls fewer tools and it makes smarter decisions mid-task, so it's fundamentally the same usage while also being vastly more intelligent.

[exotic cave]

Nice to hear! Do you know what they mean that in the coming mont's they will make more updates because of RL? (I might be confusing things... so fix me on that)

[lean lark]

Looks like you haven't updated the VSCode codex extension, then VSCode itself.

[vocal goblet]

ive updated both vscode and codex app, tried reinstalling codex app too, still same error when i try to use GPT 5.5.

[vocal goblet]

Yeah but I am trying to use GPT 5.5

[lost drum]

oh sorry I ment 5.5

[lost drum]

I edited my message now its correct

[vocal goblet]

dw

[lean lark]

i see you're using App, not extension. OK, you may need to globally npm update @digital thunder/codex. need v0.124.0

[vocal goblet]

Up to date & same error, i guess ill just wait for now

[lean lark]

I hate to say this but restart os. which OS anyway?

[vocal goblet]

win10 i had this problem yesterday too

[lean lark]

common issue is that people have npm codex loaded globally And in project under node_modules.

i don't have "App" but I'm surprised that's not updating and working.

oh, cache...

look in this channel about deleting %AppData\.codex folder or similar. stop app, delete, restart app.

[vocal goblet]

Will try!

[lean lark]

k, i need to go back to sleep. hope that helps.

[boreal holly]

I don't know what you mean by RL. If you mean Reinforcement Learning they probably have a better training pipeline since the model keeps getting smarter and it's basically building itself

[vocal goblet]

gn

[torpid trout]

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

[hybrid ore]

I guess the reset never landed for the 5.5 launch?

[unborn thunder]

why is codex usage so out of touch? my 5hr limit has been maxed out for weeks while weekly goes from 100 to 80 percent ?

[plain crescent]

do you use voice mode vid gen etc?

[unborn thunder]

noop

[plain crescent]

you can get 2x the quota if not

[unborn thunder]

how?

[orchid plume]

just a guess but I'm guessing mattz is probably referring to ChatGPT has its own quota, so while it's not Codex you could continue there in some manner

[velvet osprey]

How do I delete chats in the codex

[bright swift]

rm -rf ~/.codex

[hasty sky]

Hey everyone,
I’ve been using Claude Code quite a lot for coding workflows and I want to try Codex as well.
Before I switch, I want to understand a few things from people who already use it:
What are the main differences you notice between Codex and Claude Code in real projects
Any tips for getting better results with Codex from day one
Workflows that work best with it (planning, debugging, full feature builds, etc)
Common mistakes new users make with Codex
Any settings or modes I should always enable or avoid
I mostly work on web projects and like clean, structured code output. I’m curious where Codex performs best compared to Claude Code in real usage.
Appreciate any advice

[bright swift]

codex is better, but it isnt good at ui

[signal tapir]

Still?

[bright swift]

5.5 is better than 5.4 at it but claude still better for frontend

i mean depends on what you are building obviously, for internal stuff codex is fine but claude better for anything public facing imo

[signal tapir]

In what way? Like does it write cleaner CSS?

[bright swift]

no i dont mean in a code sense. codex does everything code better. just from a design taste perspective

but maybe i'm holding it wrong. and i dont use any special design skills, so just raw model/harness basically

[hasty sky]

so codex isnt worth learning? also I am on free so I am limited

[bright swift]

it definitely is, for coding the gpt models are better

they are just stronger at backend than frontend right now

dont know how much you will get done on free plan though

[gentle harbor]

why is there a red icon ?

[signal tapir]

I'm just telling it to do what my artist decides. 😛

[lost drum]

Yo what is the most valid way to check your usage?

[clever steppe]

I hope this gets sorted out soon as this is killing me. I've never seen it before 5.5 was rolled out.

[lost drum]

Cause hwre its not updating propprly

https://chatgpt.com/codex/cloud/settings/analytics

[lost drum]

Yo g

I had same problem

Just click off and return to chat again

It will work

[clever steppe]

Interesting! Thanks for sharing this. I might give 5.5 a try soon. I do wish there was some kind of 'miles per gallon' meter so we knew how much it's actually using.

[tall zodiac]

Yo,

Do you guys have like an Md file to keep track of what was last done or is auto compaction good enough to not have to do that ?

[gentle harbor]

sora shuts down tmr do you think we will have better rate limits for codex with all the extra power they will gain ?

[lost drum]

You need have advanced memmory MDs

[tall zodiac]

I have memories and chronicle enabled

Before that I used to have a file that used to write the last checkpoint before compaction but I heard memories and auto compaction does that basically now

[lost drum]

Look up Jake van Clief he has free course

[lost drum]

If you have setup that works dont change it haha

Or ask codex itself of its agood idea in your project or You have better memmory saving structure

[tall zodiac]

I was just thinking of trimming down non necessary things now that codex has evolved since 5.1.-5.2

[lost drum]

If You dont have time to scroll thruth documentation then ask codex within your project ro give You prompt for deepresearch that he will later digest and know what You should change based on Your project and model You want to use

[clever steppe]

@lean lark, you seem to be a Codex+VSCode expert; have you seen that Codex in sidebar (secondary view) isn't working as of this morning?

[unique spade]

you can wire the auto-filling of a pre-defined json with what has been done during the ongoing turn, to the pre-compaction signal (but you have to tinker with the codex harness for that)

[gentle harbor]

new update for windows codex

[placid lantern]

can anyone help ?

always happening

all uptodate and i am on plus

[gentle harbor]

funny enough just tell it to fix it self

[placid lantern]

just sent that 🤣

[lean lark]

Most people here are Codex experts, and I look up to most of them on this topic. I just woke up, not at my desk, so no clue about anything this morning. 😁

[clever steppe]

Cool, cool. 🙂

I see you have the OAI badge, though? You don't work for OpenAI?

Also, good morning. 😊

[lean lark]

@tall zodiac is describing new memory feature

[gentle harbor]

anyone can get that

just apply the servers tag

[clever steppe]

What? But why would you?

[signal tapir]

indeed

[gentle harbor]

looks cool i guess ¯_(ツ)_/¯

[lost drum]

secret dicount right haha?

[gentle harbor]

haha

[lean lark]

just a nice easy to show support. all servers have it

[gentle harbor]

thinking mode for normal chatgpt is very fast and way better, i make sure to keep on extra thinking mode and online search

only takes 5-9 sec on avg

[nova turret]

Help

■ {"type":"error","status":400,"error":{"type":"invalid_request_error","message":"The 'gpt-5.5' model is not supported when
using Codex with a ChatGPT account."}}

[lean lark]

we're setting many reports of that

[nimble terrace]

5.5 is very token hungry holy moly

definitely need to get the £100 pound plan

[signal tapir]

Are you running it at full throttle?

[nimble terrace]

no, low thinking

[lean lark]

@clever steppe did you get yours working

[nimble terrace]

not fast mode either

[signal tapir]

Use it at medium. I think it is smarter, and uses the tokens more efficiently

[nimble terrace]

interesting, so not low?

[signal tapir]

No, I just noticed that low results in more usage myself.

Not certain of course. Just a feeling.

[nimble terrace]

could be

[nova turret]

Use the api

[nimble terrace]

doesn't it cost a hell of a lot more on the API?

[nova turret]

Paul Solt notes Codex has ~400K context while the API has 1M context for GPT‑5.5

Does this mean subscription codex or will using codex with an api key instead of subscription will still not give 1M context

[lean lark]

@nova turret I have solid impression that error 400 means components aren't fully updated

[placid lantern]

???

[nova turret]

I reinstalled codex again

Just to make sure

[lean lark]

oh wait

[nova turret]

I'm talking about the codex cli on the codex desktop app the model doesn't even show

[lean lark]

you said you're using API?

[nova turret]

No

[nova turret]

I found this just now and I'm saddened by the fact

[lean lark]

OK codex --version needs to be 0.124

[placid lantern]

someone pls

[nova turret]

I'm confused here that will using the codex harness with an api key instead of subscription still won't give 1M context?

[lean lark]

@placid lantern you didn't provide context but it seems you're talking about operations

[nova turret]

codex-cli 0.125.0

[placid lantern]

i am trying to deploy my personal branding website via cloudflare and hostinger connectors i did brought domain and all but the browser use wont work at all

i am on chatgpt plus

[lean lark]

they did another update!

[placid lantern]

already updated

[lean lark]

no i mean i didn't know about . 125 , i just woke up.

[nova turret]

Get off of screens

Drink water, wash your eyes

Get sunlight if it's 7-8 am

[placid lantern]

aww man

didnt noticed 9%

[lean lark]

😁
thinking about error 400 on 0.125 ....

[clever steppe]

For a little bit, but it's crapped out again now. 😭

[lean lark]

did you see my note about appdata? which OS, sorry.

[clever steppe]

I'm just trying to get some work done while my kids are happily playing. 😭

[clever steppe]

Me? Win 10 Pro N x64

[topaz phoenix]

why i cant load previous chat

anyone know how to fix it

[clever steppe]

I'm having the same problem @topaz phoenix

And several others here.

[topaz phoenix]

even undo doesnt work sometimes

[signal tapir]

Restart vs code

[topaz phoenix]

i did many times

[clever steppe]

i restarted VSC many times.

I'm using 5.4 Medium, BTW.

[signal tapir]

What helped me was start a new conversation, and then go back to the old one. After restarting vs code

[lean lark]

both of you.. delete the .codex folder under AppData,.
npm install -g @openai/codex-latest.

and please forgive my fast phone typos.

[topaz phoenix]

i tried this one too and not working

[topaz phoenix]

not a vs code extension ?

[clever steppe]

How do we get this feedback to OpenAI team fast?

[lean lark]

github for codex repo Issues. Read first before posting!

@topaz phoenix updating extension doesn't update codex itself. tools all need to be in sync.

[topaz phoenix]

got it , thanks

[lean lark]

if you're running codex in WSL you need to update npm there too.

[lean lark]

not finished with coffee yet and too much finger swiping... 🤦🏻

[topaz phoenix]

lol i just finished mine

[lean lark]

someone here have the correct package name?

its codex@latest

and i shoulda seen dat

the challenge here is that your systems are saying Codex can't use model 5.5 with ChatGPT about. That indicates it hasn't yet been updated, the array of valid models doesn't include 5.5 yet.
This is why i point to update. after update, people have reported continued issues. This indicates cache conflict Or more than one npm package.
That's what drives my responses right now, getting a clean consistent single version across your tooling.

with that understanding, ChatGPT can probably help you now.

For VSCode extension, ensure it's updated in each profile, if you use them. You must reload VSCode a couple times, dunno why. Best to close down everything with exit and restart app.

"End of Line" -- TRON

... and Mr Choo Choo Train Conductor

[nova turret]

Bro please

Drink water

Wash your eyes

[lean lark]

I've already responded. what's up?

[boreal holly]

Lol it says @openai/codex-latest. Starbucks is right, it should be npm i -g @openai/codex@latest

[lean lark]

already corrected that, wash your eyes

[boreal holly]

My eyes yearn for the clorox

[main nimbus]

I need coffee

[lean lark]

mine for an ice-pick

[main nimbus]

sold

[lean lark]

anyone have problem resolved yet so i can start my day? 😁

[vital imp]

oof, i go to use codex for something simple and see my usage remaining is at 0% and my weekly is at 68% and I havent used it at all today. Well I was messing around with undocumented config.toml settings and I enabled chronicle. I think that's what used all my usage in the background

[boreal holly]

Yep, they posted officially on X that chronicle uses an order of magnitude more quota

[lean lark]

its in announcement/dev docs too

[severe quail]

Anyone have any issues logging back into codex on VScode when you chnage accounts?

[lean lark]

depends on the specific message 181

look at output window

[nocturne folio]

yeet??

[robust rivet]

does anyone like me encounter this error everytime it compacts ? "Error running remote compact task: stream disconnected before completion: error sending request for url (https://chatgpt.com/backend-api/codex/responses/compact) codex"

[plucky orchid]

Anyone have the same issue ?

[young locust]

paid account?

[torpid trout]

Me: uploads an image of a hand holding a pen. "Please redo this logo and create an assets bundle so I can use it on the website"
GBT: sure man. goes ahead to produce Hogwarts mid-year quidditch turnament logo...

I admit the prompt is vague, but how do you go from a hand holding a pen to this?
LOL

[torpid trout]

its hiccking up a lot, cli too
reset wen

[nocturne folio]

bruhh

[torpid trout]

GBT is very nice to create memes. I think I will have to start selling out my old memes now

[unique spade]

Who s GBT though? 😂

[stable obsidian]

:l

[plucky orchid]

Yeah

[nocturne folio]

gotten the same thing

[steady vigil]

yup its broken

■ Error running remote compact task: stream disconnected before completion: error sending request for url (https://chatgpt.com/backend-api/codex/responses/compact)

RESET TIME

[torpid trout]

GPT. But like when > wen. GPT > GBT

[potent mason]

5.5 medium is actually doing pretty good, I still don't know when to switch back to 5.4 though

[bright swift]

meh

wen reset

[potent mason]

Yeah honestly makes sense

They went from 2m users to 4m users in like 2 weeks

[bright swift]

remove free + plus plan, problem solved

[torpid trout]

Selected model is at capacity. Please try a different model.|
Now resetis unavoidable.

[bright swift]

user metrics are inflated by people abusing those anyways

[torpid trout]

I am on a pro priority plan

[potent mason]

I think the issue is a bit deeper.

They allowed openclaws so they have people wasting way more than they should be using

[bright swift]

i am on pro vip plan, still errors

[bright swift]

that too yes

[torpid trout]

/feedback also broken

[bright swift]

allowing openclaw was a huge mistake

[narrow wave]

Hey guys! I just finished my document on my workflow with ai, the issues I've come across over the last 8 months full-time building with codex, addressing the findings and gaps in how people are currently using it.
This document is not an AI-written summary of best practices, but a reframing of the core model of how people think about using ai, in a way I haven't seen anyone talking about.

Hopefully this is an acceptable place to drop this, not promoting a product, just a free doc for those interested; would love to hear some feedback.

This core concepts:

Why trying to accomplish issue identification, intention spec, implementation, tests and verification into one pass is compounding hidden failures
Using AI as a targeted reasoning tool, separating each perspective-pass
Addressing intention and document drift
How documents and chats can be kept focused, avoiding bloat/filler
Assuming 90% confidence in each step, intention-based tests in verification
Common failure modes and my preferred workflow

Read the full document here: https://www.inneratlas.co.za/field-notes/working-with-ai

[potent mason]

The solution imo is #1 fix how wasteful openclaws are, #2 Remove free + go (plus still needs to exist), #3 be honest about reducing limits, no one is expecting to continue paying $200 for 4k inference just don't pull an anthropic and silently decrease it

[potent mason]

TLDR

[bright swift]

oh i completely forgot about go tier

[torpid trout]

Let the "me too" carpet bombing begin please!
https://github.com/openai/codex/issues/19579

[unique spade]

just got this on 5.5 lol (on 100 pro plan)

[potent mason]

No thanks, OAI is doing their best, you can always switch model

[potent mason]

Same on $200 plan

[unique spade]

in the middle of a turn haha

[bright swift]

both my plans around 10% left with 5 days to go, guess i need a 3rd now

[cyan gyro]

quick, use your limits

[ornate agate]

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

NO

[cyan gyro]

Fast model is not impacted yet?

[plain crescent]

do the people that are complaining abt quota reductions make use of video gen and voice mode etc? all those non codex features?

[unique spade]

well i had 2 threads in paralel. one keeps going, the other suffered from that

[bright swift]

i use only codex cli and occasional pro in chatgpt web

[plain crescent]

you can get abt 2.5x codex limits, but will lose access to other features liek video gen and pro in web

[orchid plume]

I guess there's capacity issues, first time I've seen "Selected model is at capacity. Please try a different model." on Pro 20x. I can see it also stops the conversation, doesn't even auto retry

[potent mason]

Where is that an option?

[cyan gyro]

it's also really slow now right

[plain crescent]

I am temporarily gatekeeping that.

[orchid plume]

noticed that today as well, guess it's really under load today

[torpid trout]

y
u want no reset?!

[potent mason]

I don't really care, I'm at 70% and have a reset coming in 3-4 days

[bright swift]

the dream

[potent mason]

Honestly at this point, if everyone gets a reset and they're already having compute issues we are going to have even more issues

[lone ruin]

Could someone help me understand why my Codex won't load? I've tried restarting my laptop and quitting the app.

[lost drum]

Me too I have 2x pro plan

[bright swift]

you are just gatekeeping because you have more usage left

i really dont want to set up a 3rd account right now 🙁

[young locust]

try using end task from task manager

[lone ruin]

I've already tried :/

[lost drum]

I hope it didnt break my code tho

[lost drum]

Maybe there is an update try reinstalling

[unique spade]

there s no such option. it must be some kind of scam or shady against ToS offer from god knows who

[robust rivet]

what does it mean "RESET TIME"

[potent mason]

Tibo resets usage every time anything goes slightly wrong

So everyone right now is probably trying to go through all their usage xd

[unique spade]

and now it s clearly starting to go wrong ion new ways haha

turns stopping mid work cause model is at capacity is new to me

[lost drum]

Resets are useless if You havent used all 100%

[potent mason]

Actually it's more like >15% per day since last reset

But yeah

You should be aiming to use it up

[unique spade]

nah, they good whenever you used more than what you would have with equal day distribution

so if in 2 days you already used 50%

you want a reset

[lost drum]

Yes but if Your reset is in 2 days and you used 60%, but you plan to spend 40% on the next 2 days, then if they reset you loose the 40%

I think they have calculation of when to reset to give less usage to users lol

Some win some loose

[lean lark]

Stopping in for a quick note from a CLI execution:

warning: Codex could not find bubblewrap on PATH. Install bubblewrap with your OS package manager. See the sandbox prerequisites: https://developers.openai.com/codex/concepts/sandboxing#prerequisites.
ERROR rmcp::transport::worker: worker quit with fatal: Transport channel closed, when Client(HttpRequest(HttpRequest("http/request failed: error sending request for url (https://chatgpt.com/backend-api/wham/apps)")))
ERROR codex_api::endpoint::responses_websocket: failed to connect to websocket: IO error: failed to lookup address information: Try again, url: wss://chatgpt.com/backend-api/codex/responses
ERROR: Reconnecting... 5/5
ERROR: stream disconnected before completion: error sending request for url (https://chatgpt.com/backend-api/codex/responses)
codex command exited with status 1

Will look at this on Monday.
Have a great weekend!

[potent mason]

Actually no one really loses unless you use up 100% next week

[lost drum]

It happened for me earlier I had 70% left to use on 2 days and I planned to do it but they reseted and ye I lost 70%

[lean lark]

The bubblewrap thing was supposedly fixed. The rest of it is a persistent failure from CLI. I might be facing the same issue that others were earlier with a new 0.125.

[stable obsidian]

Is there rlly gonna be a limit reset

Not sure if I should spam xhigh fast mode now

[bright swift]

you and me both buddy

[unique spade]

you only incur a cost of opportunity relating to the counterfactual case where if you would have already used more (or all) of the qouta that will be reset.

but that's a counterfactual loss

in reality, you just start again from 100%

[lost drum]

Hmmm why dont they enable Auto diagnose feature so You always get answer why there is an error like that

[unique spade]

think might be related to the other issues we re facing, with model at limit etc

[torpid trout]

The problem is this:

[lean lark]

Had exact same error yesterday though - and this is a single transaction from CLI.

[potent mason]

Bruh I'm using 4-6 at the same time and my brain is already at capacity

[lean lark]

I'll just blame JaneBot and tell her to fix herself.

[lost drum]

[bright swift]

i currently have 8 ghostty tabs with 3 tiles each. are we the baddies?

[lost drum]

All this to get liquidated

[lean lark]

See? This is exactly why I say a company should never do a release on a Friday. Derp!

[torpid trout]

I'll just open 10 more. Codex give help. Y u no profit.

[lean lark]

AFK finally

[unique spade]

lol hahah nice

[bright swift]

the subagent spawns will continue until capacity improves

[small violet]

how do i fix this

[torpid trout]

See. This is why I did uninstall the app. That, and the constant memory crashes

[boreal holly]

It's OK, I'm currently dealing with the very annoying "all MCP tools require approval" thing since I went from 0.116.0 to 0.124.0.

[torpid trout]

It starts failing, the more your threads progress.

[small violet]

this is a very annoying error

i havent experienced memory crashes yet

[torpid trout]

Honestly, I would never use the app for actual development
Either use the VC integration or cli - the app is kind of bound to not be a good fit. ALl that UX has a cost.

[lost drum]

Tbh who needs better mofel than 5.5?

[potent mason]

It's not about the model, it's about the autonomy

I feel like I'm having to prompt it 20 times to get it to do something which in the future will be done with 1-3

[lost drum]

I mean now I dont even need to prompt he is in loop emitting prompts for next run

[torpid trout]

u asking for real now

[lost drum]

For my project tho

[potent mason]

I've done that but unless my plan is extremely specific it starts going off from my vision

[boreal holly]

I know 😩 why does OpenAI bother reimplementing the entirety of git version control when folks can just use git

[torpid trout]

yeah no lol.
But I do understasnd why you fell like that

[unique spade]

you can make your own UX and expose anything useful for you from what codex cli allows

who needs ux made by someone else, when you can make your own

[torpid trout]

I dont... they do. 😉

[unique spade]

for all the noobs who never even entered the old coding universe

[unique spade]

yupz 🙂

[torpid trout]

this, iMO, is actually more dangerous than releasing a dangerous model
However, that is probably also where the money is

In terms of... they pay, play and then have other things to do, being not working in the field

So after the "create me tinder 2" fails in terms of no gf still, they abandon but keep paying

[supple perch]

there is so much wrong it seems... useage limits are killin me right now

[torpid trout]

reset wen

[kind jay]

🙄

[kind jay]

Mean reversion is wild tbh

[potent mason]

Not for me

[torpid trout]

reset wen

[nocturne folio]

how do i get half a billion tokens before may

[bright swift]

ez

[lost drum]

how to check it if I am on VS?

[bright swift]

VS should still save the logs in same location but not sure. npx @ccusage/codex@latest

[solemn acorn]

I like having my codex runs in a separate window from my editor

[bright swift]

whats an editor

you mean this?

[woven python]

Anyone having issues recently with Codex hanging on "Thinking..."? And it constantly seems to be reconnecting. Using Gpt5.5.

[craggy whale]

I was just about to complain about Claude. It seems to be a server outage again. So it must be affecting Codex and Claude.

[ornate saffron]

hey ya hello guys.

[craggy whale]

Hiya! A.M.

[ornate saffron]

hi bro . i heared that you use codex. can you tell me does using codex for a high codebase like work like building a whole full stack web from scrath . Do i need to add plugins or skills. or does it affect anything in codex.

[craggy whale]

I assume it doesn't. I would ask ChatGPT. I personally don't know because the coding that I do with Codex is very personal and basic in comparison to that. It just helps me build GUI programs that make organisation and managing stuff on my PC easier.

[craggy whale]

Codex (or modern AI coding tools) doesn’t require plugins or extra “skills” to handle a full-stack project—it already understands frontend, backend, and common frameworks.

What matters is how you use it:

• Give clear structure (e.g. “React frontend + Node/Express backend + PostgreSQL”).
• Work in steps (don’t ask for the entire app at once).
• Review and guide it—don’t expect perfect output in one go.

For large codebases, Codex doesn’t “remember everything” automatically—you usually feed it relevant files or context. Tools like IDE integrations (VS Code, etc.) help with that, but aren’t required.

So no special plugins needed—but good prompts, iteration, and structure make a huge difference.

This is what Chat GPT said.

[ornate saffron]

Huh like some kind of automaticn things. Ah no problem . well what you work mainly on . like what is your tech stack .

[woven python]

Yeah, it's full on hanging now. Getting error messages in the app about disconnected streams while the app is trying to send requests to the backend for codex.

[bright swift]

just keep prompting guys, i can already smell the reset

please

[ornate saffron]

do you use free tier or paid pro teirs. and what things you mainly work on with AI.

[craggy whale]

Tauri + Rust backend, JS frontend (Vite/PixiJS). Use the and the GUI I have built so far is I have a markdown creator so where you can just paste whatever you want to create a markdown with into the box and then it calls a local AI to generate a file name and then you can save it as a markdown. This is, so I can always give context to Claude code without needing to I try and use Visual Studio Code which isn't very intuitive to create a markdown document in because it always saves as without an extension.

Other things I have created is a thing called tidy folder which categorises everything into buckets like images, videos, music, audio, personal, and system files, and a few other buckets I can't remember.

Another GUI I'm building is. I mean that wasn't a GUI, that was a context menu Python application thingy? Script? Another GUI I'm building is a folder picker. I have already built, it's not completed yet, but it's in the end stage of Alpha, a whisper dictation application. And a few others that I would have to look at to see what I've built. A lot of what I've built is either in alpha, is very small, or is still in work in progress. And all of it is personal, none of it is commercial.

[craggy whale]

Yes, definitely seems that it is down. Check if they have a server status website.

[ornate saffron]

do you actually used AI to write it.

[bright swift]

thats pro account #2 the first is empty 🙁 i work mainly on a data analysis project right now in rust, but also have two existing web projects

[craggy whale]

All of the ideas are my own. AI has built all of it, and I have debugged all of it.

[ornate saffron]

yeah data analysis projects hits limits fasts . Like when i made 4 txt files of 50-90kb each it hit limits very fasts.

[ornate saffron]

I can realte that very much. But i don't have any kind of traditional coding knowedge . Just I use AI for coidng and designing.

[bright swift]

oh i'm not feeding data into it yet, and the models wouldnt read the data. its just coding

[ornate saffron]

I see.

[craggy whale]

Oh, me neither. I have no knowledge at all. I'm the equivalent of Mariah Carey that can't play the piano, so tells the piano player what she wants to hear.

But that's still a skill in itself. You have to have some intuition and you've got to feel things through. That is quite abstract and that isn't easy.

[ornate saffron]

yeah bro i made 20+ full stack webs and AI agents and lots of automation and workflows myself but has never has any moeny to modify or develop them or un them competely.

[craggy whale]

Oh wow, you've only been using the free tier? I have four accounts. I have issues. I have two Claude Code accounts. I have one Claude account. And I have one ChatGPT account that I use for Codex. I have all of those because I run out of usage very quickly.

[bright swift]

focus on one, and dont just build it, market it. the challenge is distribution now, not building..

[ornate saffron]

no it is not the execution or deployment but the real problem is moeny . as you even go to try a free cloud teir they take money form you from sides.

[craggy whale]

Yes, I understand, and yep, that's true.

[ornate saffron]

yean bro . free tiers problems.

[ornate saffron]

ah yeah that's nice suggestion but the single projects i make includes all fields in one like cyber security, web dev works, AI agents and LLM. and many more stuffs in one folded.

[gentle harbor]

this is so annoying, i wish they at least tried to improve the rate limits

[frosty zealot]

I cant help but have a laugh

[solemn acorn]

IDE or text editor like vscode

I haven't had a chance to use 5.5 too much, but it's been great at low reasoning actually

super fast and stretches out your rate limits a lot more

[dense path]

I think he meant it as a joke didnt he xD

[wintry gorge]

Hey everyone, is anyone using Codex with both a personal account and a Business account having issues with the limit being used up way too quickly?

I pay for both accounts. When I hit the limit on my personal account, I switch to the Business account to keep working. But on the Business account, after only 2 or 3 tasks, I’m already hitting the limit again. I’m not talking about dozens of tasks or anything extreme, it feels like it barely lasts at all.

[boreal holly]

Haven't tried low yet, but that's awesome!

[torpid trout]

I swear, I have seen them all, like 9gag, reddit, facebook, private make fun whatsapp groups... but this here is the purest of entertainments.
How can be, a such serious channel, such wow

[burnt sage]

Call me

[main nimbus]

I have used low for a few cases and it is great. Medium is also* a great daily driver, no need for xhigh everything like before.

OpenAI recommends at least low instead of none for any type of multi-step agentic task that is also latency critical.

[frosty zealot]

https://x.com/RamzaRamza4/status/2047787596848173508?s=20

Thank me later

[nocturne folio]

sharing is caring

[kind jay]

She’s using the heat from my server to keep warm

[kind jay]

Make one of me!

[torpid trout]

Did you use that neon prompt lol

[torpid trout]

did you forget I already made one of you my only star?

[nocturne folio]

i wish i could get a freeeeeee pro 20x account

[torpid trout]

the closer you look the better it gets

[kind jay]

That’s silly

[kind jay]

You broke or something?

[nocturne folio]

Yea

[nocturne folio]

reconnecting 6/7

[frosty zealot]

🎵 SHOULDA BEEN THERE ON SUNDAY MORNIN HANGIN MYY HEADD

[stable obsidian]

Not nice :l

[torpid trout]

GPT 5.5 is the goat of design.

[hard drum]

SAVE MEEEEEEE

[kind jay]

Sorry, I don’t help brokies

[hard drum]

not even a 20x sub could withstand the needs of a borrow checker

[torpid trout]

And Jane, sorry, the meme GPT made of you is too spicy, I think I will break all 15 rules if I post it here, plus, lose you as a star

[hard drum]

when 35x?

[torpid trout]

wen reset

[hard drum]

why am i ALWAYS in deficit, dammit

somehow rtk-ai/rtk is not enough, || my framework has something causing major issues

¯_(ツ)_/¯

[frosty zealot]

where tooki

[undone patio]

having to use opus after the new gpt release is so gross

im so cooked until monday

[frosty zealot]

are we still on 2x limits?

[cyan gyro]

we're screwed if they drop the 2x

[boreal holly]

Bro, unified_exec=false and custom build

[frosty zealot]

do you have a second account

or one that you did have but is now expired

[undone patio]

no sadly

[frosty zealot]

i accidentally logged into my expired other account and there was an offer to claim a free month

i think it was just plus but

O.O

[undone patio]

lmao

[frosty zealot]

me watching that video for the 1 second i got to see it

[hard drum]

what is this?

[frosty zealot]

I was trying to save some usage by letting Opus work on something, it worked in circles for HOURS, got sick told GPT to fix it, one shot

[hard drum]

am now planning out to fully rebuild my agent framework as a secondary harness

typescript + rust + shell

[undone patio]

nice

[hard drum]

then take codex-cli && modify it as well

maybe strip it down of unnecessities

[torpid trout]

Someone needs to leak GPT image so we can uncensor it.

[lost drum]

why would you

[frosty zealot]

So he can have tingle time

[hard drum]

if you need that time, NAI exists

[kind jay]

Post

[boreal holly]

full context

[kind jay]

What’s that?

[torpid trout]

No, I just need it not to tell me that creating satire is against the rules of whatever buttoned up sharpie who defined the rules 🙁

[torpid trout]

Sersiousl. Then you challenge it and suddenly it does it anyway. Not entirely sure what the pattern is so far.

[nocturne folio]

dude this would be peak for roleplay

[unique spade]

nice, so you're part of those using homebrewed custom build codex?

[deft gyro]

@boreal holly have u migrated your orchestration thingy to 5.5

[boreal holly]

I mean I changed exactly one stupid behavior that they introduced a long time ago

[deft gyro]

If yes how is it

Yall are acting like u could run a frontier model at home

[boreal holly]

It was as easy as setting the model to gpt-5.5

This model did not have the same transition pains as 5.3->5.4

[deft gyro]

Ah really?

Nice

Are you using more or less of your quota

[boreal holly]

I even upgraded my workers and QA from 5.4-mini to 5.5, and usage has decreased while they get more better work done

[deft gyro]

Iirc it’s twice the price but it kind of evens out

Because so much more per token

Just having less autocompactions for me has been super nice

[unique spade]

you still had to compile it though right?

[boreal holly]

Yeah... First build takes like 45 mins, but all the builds after that are a minute or two

[stable obsidian]

Bro what

[unique spade]

yup i know i m using in paralell a custom build myself when i feel like it

[deft gyro]

Xd

[hard drum]

what effort level?

[stable obsidian]

Cant take design references without being flagged

[boreal holly]

schmedium for all agents

[hard drum]

sch? german spotted?

[deft gyro]

Medium is the goat

I used to have some issues sometimes with 5.4 on medium

But 5.5 has smoothed my experience out