codex-discussions

[plush harbor]

weird. I've had the api hooked in for a week now and everything is zero other than the money they are taking off me

[hard drum]

It's also a social coercivist, as i'd like to call.

[torpid trout]

wait.

That is new

They removed the usage menu?!

[plush harbor]

oh great so they broke something right when I implented the api

[torpid trout]

And now it is a dahboard ugliness? And yes, it is zero, and should not, clearly, one of my api keys is massive in usage and was reporting usage in the old screen for ... now years

[hard drum]

there's a reason you need something like this in memor/custom instructions:

Does not want unsolicited advice, guidance, reassurance, suggested next steps, coping strategies, interpretations framed as what they should do, or suggested wording/scripts in emotional, interpersonal, imagined, dream, memory, trauma, or hypothetical scenarios. Do not provide 'you could/should/might say/do/try,' 'the safest goal would be,' 'a better/more accurate thing would be,' or similar constructions unless the user explicitly asks for advice, guidance, action steps, or wording. Stick to direct answers and interpretation only.

general-purpose

it could be made better, of course

[native agate]

Anyone tried the new Deepseek free model? legendary

[hard drum]

it's basically the whole, "did i ask?"

[plush harbor]

oh well, guess its not just me then. I've just only had this for a short time, noob problems

[torpid trout]

what even is this new measuremnet dashboar dsupposed to be... a vibecoded insult?

[plush harbor]

lol I regret mentioning this now 🙂 Clearly I missed something good

[torpid trout]

It was horrible before, but at least one was able to see what is used where and when

[boreal holly]

That's funny! I think if you simply replace the system prompt, so rather than reading like "you are a helpful assistant who assists the user in every conceivable way" it's "your job is to pass the butter." It totally avoids those things

[hard drum]

this is for the web version

since the web version is a freak

[boreal holly]

Oh yeah, web version... I still have not tried it. Sounds frustrating!

[cedar skiff]

I asked 5.4 how to change the system prompt about a month ago and it told me i couldnt o.0 i guess i should have pressed it to search harder

[hard drum]

if i had feelings, yes, it would

It's just annoying in thought form

[torpid trout]

not utterly good but at least it was working, showing data
Now it is literally zeroed out across all years of usage, although the bars appear, it still says 0

Meh. Too long I let that account run unsupervised (I have api and codex on two different accounts)

[boreal holly]

Oh dude it has a huge problem with editing system prompts, even if it's for a totally non-codex agent lol. Also refuses to allow reasoning tokens to show up anywhere, even hidden behind developer/debug flags for diagnostic purposes only

[plush harbor]

the irony ... I only implemented API because images 2 is so horrible, and they probably broke that dashboard around the same time

[cedar skiff]

I ended up just inserting some stuff via developer_instructions

It's not ideal because it could have conflicts

[plush harbor]

is the two different accounts for redundancy? I've heard stories of people having their regular accounts locked

[boreal holly]

Yeah, if you look at a rollout log you can pull the Codex provided, fully formed system prompt out of the first message and refactor it to fit your needs. For example I told em "you don't need to use the apply_patch tool if there's a better way to edit something. Python works fine." And these suckas do incredible amounts of refactoring with that simple tweak

[hard drum]

Bro, you're my personal agent layer thingy infodump with all this stuff

Mine's been using python3 for a lot of edits, too, for some reason

also in middle of rebooting OABTW as OAL

[cedar skiff]

Thanks for the heads up, one question - how do i look at the rollout log?

[boreal holly]

It'll be one inside ~/.codex/sessions/**/*.jsonl, and the first line typically has the base instructions

[hard drum]

i should let them use python3 more for it

[cedar skiff]

Cheers 🙂

[boreal holly]

I think they put that "You must use apply_patch for edits" line in there because in Codex Desktop & VSCode, every apply_patch shows up in the UI with an undo button, so they want the agent to use it as much as possible. I see it as a hinderance. They're pretty clever with their python-fu!

[white rover]

hello codex gang 🫣

bit of a pickle im in, any easy way to get codex chats from one device over to another device, or letting codex that has the chat get evrything it needs, files, handover sheet etc, then downloading what it sends, and sending to the other device??

any help would be great gang 🙏

[plush harbor]

never tried but you might be able to lift and shift the session directory

[hard drum]

ask codex to make some sort of tool for EXACTLY this!

[white rover]

main codex is on ios and other device is windows 🫣

[hard drum]

ios mentioned!!!

🍎

[white rover]

🥵

[cedar skiff]

you mean macos?

[hard drum]

i raise you... so much

[white rover]

yes

[boreal holly]

Additionally to copying the ~/.codex/sessions folder, there's also a file ~/.codex/state_5.sqlite which has metadata, like name of the agents or what have you

[hard drum]

one is iOS game thing, other is android emu for reference

[plush harbor]

I'm over here on command line like its 2002

[hard drum]

i need my file browser && stuff.

cannot imagine using terminal agents unless i had "vscode but terminal" equivalent

[white rover]

can i send you dm bro?

[hard drum]

&& if you say neovim, i will personally force thigh-high socks on ya

[plush harbor]

no agents here, just codex. Old site, small codebase. I have cron jobs not agents lol

[hard drum]

Notepad++ when?

|| even better: nano

[plush harbor]

sublime text

[hard drum]

use nano like real people haha

[lost drum]

no podcast?

[hard drum]

Ew

🤮

[boreal holly]

A real chad uses sed and awk

[hard drum]

oh god

beautiful hole

[nocturne folio]

somethings off here

[tulip osprey]

FINALLY

it is about time they got a disocrd server

[boreal holly]

Silent Hill 2 memes be off the heezy

[tulip osprey]

https://tenor.com/view/judy-hopps-scuba-tiktok-dance-baby-boo-gif-10715200134141212483

ew no pic perms...

[hard drum]

those 150 discord notifs, yes

[nocturne folio]

its a bunch of pending requests

[hard drum]

how do people live with so many notifs? i'd shut stuff down if i had 150 notifs jfc

[nocturne folio]

idk i dont care about it

[plush harbor]

I have 10,000 unread emails but nothing anywhere else >.>

[lost drum]

what?

just dont go to ecom communities

[hard drum]

'ecom' ?

English, please?

[cedar skiff]

o.0

[lost drum]

e-commerce

[cedar skiff]

I bet your phone notifications feed is endless

[cyan wing]

so what do when Codex team members just fake click bait like this

i thought today was super app day

[lost drum]

no not at all I turned off every notif

[cedar skiff]

This seems like it's true, codex app has browser, computer use etc now, i just expect they will keep iterating on it until it can just vibe code anything

[strange leaf]

Does changing this option make any difference?

[cyan wing]

Still no mobile app...

[cedar skiff]

There still heaps of stuff to be built, including the models that run it.

[red mulch]

codex is tuff

but too less credits

tbh

[nocturne folio]

Maybe check my profile

[cyan wing]

dont check their profile

[glacial shadow]

Lol

[steady vigil]

looks like my weeky limit was moved forward without actually resetting it. (pro) │ Weekly limit: [█████████░░░░░░░░░░░] 43% left (resets 06:23 on 5 May)

this should be day 3 from the last reset

anyone else see this?

[frail meadow]

I thought they reset last night

[steady vigil]

yeah the timing would make sense

my % left not so much

that makes this literally a reduction in usable credits vs no reset for me

[frail meadow]

Where do you live?

[steady vigil]

canada

[frail meadow]

Hmm interesting, I feel like not everyone gets the same resets or at the same times

I'm in US and got it around midnight Central Time last night I think

[steady vigil]

I usually get it at the same time as others, we'll see I guess

[nocturne folio]

definitely check mines

[boreal holly]

Their techniques are becoming more elaborate 😩

[drowsy turtle]

made codex connect to claude using prive p[ythopn script private connecytion with user

[frail meadow]

What are you guys working on tonight?

[frail meadow]

What's the difference between this and using claude with the -p flag via codex?

https://code.claude.com/docs/en/headless

[lost drum]

someone make a scan of every message taht comes when seearching "codex" in here and make a complexed guide for new commerse here

and create a command like /guide

its out now bro

[inland sonnet]

is this ragebait?

[inland sonnet]

u do it.

[lost drum]

10 upvotes and I do it

[undone patio]

cleaning up a PR for an affiliate program landing page and stripe integration

[ebon skiff]

guys, please help, my Codex windows cant use 'Browser Use' Plugins

[oak trellis]

how is the limits they coming to an end tomorrow ?

[ebon skiff]

I’m trying to use the Codex in-app browser automation through the browser-use plugin / browser-client.mjs.

The in-app browser itself is open and working correctly, and I can manually navigate/login in the browser UI. However, Codex cannot attach to or control the browser through the Node REPL automation bridge.

The failure happens when running the Node REPL browser setup code:

const { setupAtlasRuntime } = await import(".../browser-client.mjs");
await setupAtlasRuntime({ globals: globalThis, backend: "iab" });
The returned error is:

Node runtime too old for node_repl (resolved C:\Program Files\nodejs\node.exe):
found v22.18.0, requires >= v22.22.0.
Install/update Node or set NODE_REPL_NODE_PATH to a newer runtime.
Important detail: the system also has a newer bundled Codex runtime available:

C:\Users\yusuf.cache\codex-runtimes\codex-primary-runtime\dependencies\node\bin\node.exe
v24.14.0
But the node_repl tool still resolves to:

C:\Program Files\nodejs\node.exe
v22.18.0
So the issue seems to be that the Node REPL used by Codex/browser-use is not picking up the newer bundled Node runtime, even though it exists locally.

Question:
How can I make Codex / node_repl use the newer Node binary? Should I set NODE_REPL_NODE_PATH, update system Node, restart Codex, or configure this somewhere inside Codex/plugin settings?

[oak trellis]

now thinking buying codex or cc .. with all that new codex stuff

their limit reset tricks

[undone patio]

codex is SOTA

[oak trellis]

which one you prefer right now ..

[undone patio]

anthropic is behind and lacking compute

codex ez

i have all three

[oak trellis]

yes thought that too ..

[undone patio]

i have codex, claude, and google ultra

and i dont even use the other two

lol

[oak trellis]

loool

you do backend ?

[undone patio]

ML/Fullstack

[oak trellis]

yes same here .. backend right now and then frontend

[undone patio]

i like frontend more but its hard to not be fullstack these days

[oak trellis]

what about their limit you think they continue upcoming month ?

[undone patio]

idk wym tbh

[oak trellis]

they have 2x limits for a certain time

but i think with the reset mess they doing .. its no more haha

[gentle harbor]

why is it black ?

[strange leaf]

What? I'm asking which option I should select.

[cedar skiff]

what is that?

looks like adobe?

[strange leaf]

OpenAI Codex

[inland sonnet]

Its just a theme... do what you preffer

[strange leaf]

Does it interfere with anything? That was my question.

[inland sonnet]

No just other colors

[strange leaf]

Okay, thank you very much.

[inland sonnet]

[strange leaf]

[gentle harbor]

how does fast mode work ? does it just tell the model to hurry up or does it just give it more power ?

[solemn acorn]

it gives you priority over other users which gets you tokens faster, but has double the usage

that's a lot of how openai seems to be handling their pricing now--with a single model and pricing depending on how quickly you want an output

[gentle harbor]

oh ? thats it, then they should make a slow mode as well that saves tokens

[solemn acorn]

there is in the API, but it's very slow

like you're fine waiting a few hours slow

[gentle harbor]

well

in 2 days i used 76%

[solemn acorn]

are you on plus?

[gentle harbor]

on normal mode

[gentle harbor]

pro 5x

[solemn acorn]

uh

[gentle harbor]

or what ever the 100 usd one is

[solemn acorn]

how?

[gentle harbor]

no clue, 5.5 medium burns through rate limits

[solemn acorn]

I can say that's definitely not true on the $100/mo tier

[gentle harbor]

i thought 5.5 med would be about the same as 5.4 extra high but nope its still higher

[solemn acorn]

I've been using it pretty heavily over the past two days and I'm in no danger of running out of usage lol

[gentle harbor]

HOW

[solemn acorn]

idk it just doesn't use that much?

granted a lot of what I'm doing requires waiting for things to build which doesn't actively use tokens

[gentle harbor]

maybe its plan mode ? i do spam that alot, i use normal gpt 5.5 thinking to refine my prompts then i put it into codex plan mode and then send it, seems to give better results but kills my rates

[solemn acorn]

maybe, I use plan mode sometimes but usually only once or twice per thread

[gentle harbor]

i was using it for every message unless it was simple

[solemn acorn]

I think maybe you just need to start making separate chat threads for things

[solemn acorn]

yeah you really don't need to do that

granted I have opinions on things and already know how I want codex to do them

but a big part of effectively prompting is knowing how to communicate what you want

[gentle harbor]

Yea maybe thats on me, i also have a issue with the ui, i know gpt 5.5 is lacking behind on that is Claude the best for ui right now ?

[solemn acorn]

claude or gemini, though I would say no LLM is super great at UI design, they all have their telltale styles

kinda on you to bring the creativity in that regard

even if you bring them a template, they'll apply their telltale AI slop style if you ask them to use it in a design so if you do want something original, kinda just have to mock up the UX flow in figma yourself and give the model that

unless its a very opinionated UI library, that can help

[gentle harbor]

im very bad at ui so that wont be happening sigh, this is gpt 5.5 after a quick 5 min prompt

just as a test at how good it can do it

[solemn acorn]

doing effects with CSS is one thing, using those effects tastefully is something all LLMs struggle with rn

claude loves its purple gradients, gpt loves putting everything in poorly spaced squircles, gemini does wacky stuff which is cool but then has no ability to tweak it upon request

[gentle harbor]

huh, it cut me off when im at 1% not 0%

for only 100 more i might get 200 usd plan

when i upgrade that would apply right away right ?

[solemn acorn]

for what you're using it for, you could probably bump 5.5 down to low reasoning

[solemn acorn]

it would

[gentle harbor]

the 5 hour limit is so annoying im not sure why they dont let me burn through my full week usage at once

[solemn acorn]

probably because it costs them a lot more if you do

you're getting subsidized tokens, so there are limits on them to prevent you from burning through them all in the same way that paying full price would afford you

spreading out the usage is cheaper for openai and also makes it less likely that you use 100% of your weekly usage

[gentle harbor]

i wonder, how much less is 500 credits compared to a 20 usd plan ?

[solemn acorn]

the $20/mo sub is roughly 150-750 credits worth per 5 hours

probably closer to 150

I don't think there really is a way to compare the credits to the plan limits though since they're calculated differently

[gentle harbor]

sigh, time to wait a few hours until my 5 hour timer is up, should i wait until im at 0% weekly until i upgrade so i go back upto 100% weekly ?

[solemn acorn]

considering you're burning through tokens faster than even the $200/mo sub would get you in a week, I would maybe take a look at what you're doing that's using so many lol

it would get you 4x more tokens for the week, but that is a jaw-dropping amount of tokens that you might not need to be using

actually the $100/mo tier is temporarily doubled so it would only be a slightly more than 2x increase through may 31st and then only 2x from what you have right now after that

[gentle harbor]

im assuming its the over use of plan mode and the very long prompts when they dont need to be that long

[static maple]

Does updating the Codex version affect the MCP configuration of Supabase? Why do I keep having to re-add my Supabase MCP every time after an update?

[jovial finch]

They pushed an update after my complaint to them, now everything is working, simply update and you should be good.

[short linden]

$100 plan is cheap than buy extra credit.

It's token base usage anyway.

[prime ice]

When we will see codex available for Linux?

[tiny fulcrum]

roughly ~2000 credits per week, at least this is what it roughly showed for me when I was maxing it on business before we started buying additional credits

[oak trellis]

hmm is 5.4 more dumb now ?

very very bad

[low cosmos]

why does my codex randomly stop responding

like for no reason at all

mid request

[chrome raven]

Damn bro we have 5.5 , you don’t want the goblins to get you?

[oak trellis]

hmm 5.5 is truly better ? does it consumes more ?

[chrome raven]

Where have you been, 5.5 is truly better

[oak trellis]

ok ok .. i thought using 5.4 to save limits..

[split oracle]

it consumes far more tokens as long as i experienced in return of that performance

[oak trellis]

its worth it ?

[split oracle]

dunno exactly, but i guess it worths if what you are doing with it has risk in any form, not just project for fun.

[oak trellis]

yes very risky stuff

[split oracle]

i am building several apps with codex just for my personal use so i dont risk anything perhaps codex does a mistake. i can just pull the latest commit of my project

[split oracle]

then it will def worth it

[oak trellis]

thx so much for the feedback!

[oak trellis]

wow 5.5 is ultra fast

lol working entire day and only 4% of weekly used

please openAI do not change .. i know you going to bankrupt as your CFO said you cant pay the compute contracts anymore in near future.. just let me finish my project and let me print some coins ..

iam greatful for your service

[hallow iron]

Already paid one more time, but thnx

[vast crater]

Anyone else has their Codex App on Windows struggling hard with the path when used with WSL2 agent?

5.5 always tells me that node, bun, npx etc aren't on the path, it can't see the MCP... I have to default back to the CLI for a full experience

[unique spade]

Tell it to put them on the path

[wicked briar]

dude said same thing in anthropic server before the downfall

😭

[unique spade]

Started to use my own UX for codex. Maybe it's due to always loving to customize the software I use, but psychologically is so much better to be able to have all the features I want and need, where I want them

[nocturne folio]

its supposed to reconnect, idk why its just abrupt

[dawn seal]

so I realized power of automation

with this I can do agentic work so easily

create like, an agent that manages my website

and such

[oak trellis]

100%

[torpid trout]

API or codex server?

[unique spade]

Using appserver via codex cli exec as default

But separately I'm building my own direct connect path to oai server for gpt subscription

I ve experimented for last half of year with custom functionalities done via codex fork

But it's too much overhead to keep fork updated to codex releases, most of which have 0 impact on my own functionalities, but they still change a lot the core crates so I end up having to re-align the fork every release.

So for the harness, I'm just switching to custom implementation from scratch

[low cosmos]

no like the codex app itself stops responding

[unique spade]

Gptpro 5.5 is really fast

Asked it to prototype me a few design variants for my custom UX surface to also have a multiplexing view for multiple codex threads

It took it 2 minutes and they seem to be using the design I already use (it has access to the repo in that thread)

[oak trellis]

yes its ultra fast!

[unique spade]

I might end up brainstorming with it instead thinking 😂

I could never use gptpro for brainstorming cause the time I had to wait for any answer was way too long. But things seem to have finally changed

[wicked briar]

YES

very fast

how is the usage on $100 plan

[unique spade]

For gptpro? Not sure is a difference. I never hit a limit inside chatgpt, but is true I don t send gptpro tons of requests.

But right now, my mental model

Is that the key difference between the 2 pro subs is mostly codex qouta

Otherwise for practical purposes (inside chatgpt) they seem quite equivalent

[oak trellis]

very very good

[wicked briar]

alr

[wicked briar]

okay

I am planning to downgrade

[cedar skiff]

on the x20 chatgpt is unlimited

[unique spade]

Switching back and forth is easy

I did the same once 100 came out

I said I check that first. And in case I hit some limitations I'll reassess. Not the case so far

[oak trellis]

i code all day long 5% now LOL

[wicked briar]

hmm

[oak trellis]

sam hear me out .. if you there do NOT change antyhing this next 2 month..

[unique spade]

The only difference they show in app between the 100 and 200

Is the 5x vs 20x generic more than plus

[wicked briar]

he probably has a codex agent reading the chat

[oak trellis]

100%

suck all the data and send to mossad

iam fine with it for now..

[wicked briar]

bro like anthropic also sucks data but doesn't really provide value thats why they are losing market

[oak trellis]

yes thats why iam saying not even mad there

anthropic sucks it and kick us when we down already ..

openai at least gives us gud service

better evil

hope elon lost in court too ... he the biggest leacher

[wicked briar]

yeah bro elon just mad he sold his shares

codex computer use is black magic

[oak trellis]

yes haha

[torpid trout]

So, I like to write articles from time to time, hobby investigative pieces
I used GPT a lot for research, rewriting, etc
So I had this brillinat idea, you know. Codex could create me a full newsroom, basically an app that replaces BBC's offices, and lets me write professional real investigative pieces

[oak trellis]

he would like to use openai to pump is spacex scam..

[torpid trout]

But, life is a b***

[oak trellis]

they all try to get into sp500 .. retirement funds..

if they get there .. auto funds coming in

[lost drum]

but we know how to trade hehehe

[oak trellis]

that is so crazy .. didnt know that .. as soon you hit sp500 .. its auto buy .. your stock

[oak trellis]

short first..

wait for bottom .. long

[lost drum]

brez went crazy today

[oak trellis]

saw it .. ya wow

[wicked briar]

💀

[lost drum]

predicted it (u know how)

[oak trellis]

i still need to read it didnt get there yet ..

was about to write you .. we can do soemthing nice with API endpoints those free ones

[lost drum]

mind gets blown

[lost drum]

already on it haha but the fre ones limit too much

[oak trellis]

there is a website what uses a lot .. tracking war and so on .. gud api aggregator ..

they track all infos .. related to war zones

[oak trellis]

clickhouse!

[lost drum]

we can later see and use it tohgether

now I try to improve it

[oak trellis]

will dm you tomorrow .. building a profiler since 3 years ..

[lost drum]

sure haha

[oak trellis]

cant explain but codex right now is insane .. long time ago i had that speed .. and accuracy .. wow

lucky i ditched cc

[oak trellis]

still so amazed 5.5 is so smooth .

cant believe today i used 6%

from total week

wth

[lost drum]

yo what is the best approach to create android app cause I previouslyt thouthg its a good idea to ask gpt 5.5 pro to pormpt gemini in android studio to create whole app the issue is gemini is not as great as codex when it somes to executing prompts so I now think its better to tell gpt pro to build himself whole app in html and then somehow make codex make it an app but idk about the gradle building and other stuff that codex cant do within his enviroment (its just app for me not cmmercial use)

[oak trellis]

its crazy 5.5 just always update the docs now ..

before i had to beg opus ..

i have early cc feelings ..

i know it will disapear soon

[cedar skiff]

If i was doing android only i would probably just use kotlin and compose. If you wanted it to be cross platform flutter would be my choice.

[lost drum]

thank you

[unique spade]

My UX philosophy i can now materialize with the help of gpt/codex

[lost drum]

sounds like team of 20 to work on it

[spring remnant]

Day 3 of telling codex desktop team to remove comment in annotation

[oak trellis]

5.5 is so FAST

[boreal holly]

Yep, same here! the switch from 0.116.0 to 0.125.0 was almost like switching from Windows to Linux (different in every conceivable way under the hood, massive migration effort)

They must have hundreds or thousands of gremlins cooking away at light speed

[lost drum]

ye best purchase

[oak trellis]

6% ..

wow

[oak trellis]

its so good unbelievable

like self reflecting questions ..

[signal tapir]

I just had the classic "I'll remove mention of the problem instead of handling it" bug. 😁

[torpid trout]

Just recently codex wisely told me that it fixed the bug instead of updating the documentation to justify the bug.

I was like... wowz. I appreciate it fam.

Did not know the other thing was an option really.

[signal tapir]

I could have used that today 😛

[torpid trout]

🤣

[signal tapir]

I've seen this applied to tests by tons of green coders

"All the tests pass now!"

[torpid trout]

Yeah well, that is like the essence of tests isnt it

🙄

[signal tapir]

I've even seen it in fully test driven projects.

Or.. I guess in projects that were supposed to be fully test driven. 😁

[boreal holly]

Yeah, whenever I find a bug of any kind I make the worker fix it, prove it with integration tests, then fire up the simulator and manually follow the repro in the test to confirm it's fixed. That way I know the fix is real, the test actually tests for the thing it's supposed to protect against, and the behavior is locked in.

[inland sonnet]

GPT 5.5 flags reading or fixing a captcha solver in my codebase as "This content was flagged for possible cybersecurity risk. If this seems wrong, try rephrasing your request. To get authorized for security work, join the Trusted Access for Cyber program: https://chatgpt.com/cyber"

Like bro..

[celest stag]

I got that message yesterday when I wanted it to delete a bunch of build pipeline runs and I was feeling too lazy to do it manually (I was doing exploratory testing and didn't want to clutter the build pipeline history). I just did the ID verification, whatever, still annoying though.

[rocky fog]

you can just tell codex to sync to upstream and resolve any conflicts
it will just rebase it on latest and apply your "patches"/commits on top (there are also different ways)

I also tend to mention to codex right when forking something that we want to be syncing to upstream later and resolving any conflicts in the best possible way and to try to keep our changes compatible with syncing to upstream and to note all that in agents.md

then you just tell it to sync to upstream and resolve any conflicts any time you want to update

I fork a lot of stuff that way and it works fine to sync and keep your own changes and include new

I guess unless you really change it a lot/fundamentally

[woven canyon]

Guys, can you use codex desktop app on the free plan?

[young locust]

yes with very low limits

[jovial finch]

ah wym, so you canceled sub? either way mine starting working today after an update in case you were curious - just decided to let you know

[oak trellis]

weekly me working hole day is 94% ..

wow

[celest stag]

I used 33% of my 7 day quota for my $200 plan on day 1 of the reset. I'm taking it easier and turned off Fast mode for now. 😢

[torpid trout]

we are buddies, i am also on 94%, weekly with daily 8hrs run

[oak trellis]

crazy feeling right ?

me coming from cc

yesterday 20 minutes in it ate 58% of the 5 hours and 10% of the week

that ended my sub ..

[torpid trout]

It will not always be like that.
As earlier said, I have a feeling their load balancers randomly mis-assign unsage to other accounts

Are you accidentally somewhere in the Southern American area?
It would be intersting to see where who is when they have high usage and low usage

I am in a Souther American area, so likely the load balancer sees me as either americas or latin america

[oak trellis]

In bali

[frozen shore]

can anyone help me there is no gpt 5.5

[oak trellis]

think so too .. i talk about anthropic CC.. i cancled that crap yesterday

[frozen shore]

its in the codex app but not IDE extension

[dark sun]

Does anyone know what the codex rate limits feel like in comparison to the claude code rate limits?

I've been hitting my limits with claude code a lot recently, and I'd like to know what people's opinions are without paying $200 right away

oh, i'm on the claude max 20x plan

i used my 5 hour in like an hour

then i used $50 usd of extra usage in 10 minutes

😂

[frozen shore]

how big is your project and mcp etc

[dark sun]

sigh

I daily Opus4.7 xhigh

the real question is: will i be able to finish this task without running out of the free api credits i got

[frozen shore]

if you want to extend the usage use the smart models for planning and the cheaper model for writing the code.

[unique spade]

in theory yes, but if you want to keep stuff clean long term you want to check what needs updating. codex by default is great at patching, but if you keep do it like this after 10 releases your custom will be over the place in many core files

[dark sun]

that's already what i do

[unique spade]

or just use gptpro for high level design grounded in the code you already have which you upload as a zip

and then use codex to add or align the prototype to 100% aligned code

[unique spade]

i only use opus in antigravity, but i think the overall consensus is that codex quotas are more generous.

granted right now we re still in 2x promo period, plus almost every week you get some reset from Tibo

you can start with the 100$ since it's literally half of the 200 plan insofar codex usage.

plus (20) plan is trickier to compare because they made some tweaks to the 5hour limit there and not sure how it compares with the 5hour limit on the pro plans

[boreal holly]

Lol this was really interesting. The worker basically suggested "archive me and spawn a new worker so worktree creation passes" and the orchestrator did it

[torpid trout]

Got one now lol
"Hey codex its not working"
"ok let me see ah user used the app wrong so let me adjust the code so user mistake is now no mistake anymore"
🔫
ctrl+c you have to tell me when I am wrong! enter lol

[unique spade]

lol between 125 and 128

1015 files changes 🙂

[main nimbus]

A million times better

But keep in mind that’s 2x and essentially what a regular $200 plan feels like

[boreal holly]

Flippin crazy man 😩 I just wish when they released new models they wouldn't be like "you must update harness to use new model"

[rocky fog]

I dont neccesarily mean that it will be only doing changes in separate files which patch the main code
(although you could also do that)

it just puts the commits over the top (rebase)
and GPT calls them "your commit patches" 😄
(alternatively you can also have your commits in between, proper timeline, but thats much more messy)

I still change/update everything and it merges/fixes/resolves conflicts during the sync I ask

or of course you can also tell it later to put it all together better in another pass

to "try to keep it compatible with syncing to upstream" does not exactly mean that it will only be making separate patches on top in the files

so far didnt have much issues with that and it works great (time to just fork literally everything open source and customize/fix/improve it lol)

but the whole point is that you can just have codex handle that, or also ask how to best set that way of working up

[torpid trout]

This is the first time Codex created me something I have no freaking idea how to use 🤣
I literally need to ask it "guide me through the app because it is my first time using it"

Holy hell. This could probably kill newspapers in one go.
It allows you to run a whole editorial enterprise in one tab.

[signal tapir]

Happens to me when I have it build dev tools

[torpid trout]

Inclusive factchecks, research, analysts, yada yada yada

[signal tapir]

Ease of use and integration with other tools is important for that purpose

[boreal holly]

You gotta be like "pilot this web browser. Show me how to do this using the browser" haha and when it struggles go like "see? The answer should be obvious"

[torpid trout]

Yeah.. I am going to pitch that like so:
"Hey, this app is so powerful, I cannot tell you how it works"

[mossy nest]

Use computer plugin removed after update codex what should I do?

[boreal holly]

Windows or mac?

[mossy nest]

Mac mini m4

[torpid trout]

srsly tho

[mossy nest]

I also reinstalled but same issue

[torpid trout]

This is too much for me, I do not know where to start, I will go back to asking GPT to write me the article lolol\

[boreal holly]

Hmmm, check Privacy & Security settings, make sure codex has screen recording and accessibility enabled. See if there's a new feature flag in the configs or app settings

[boreal holly]

How the fact 🤣

[torpid trout]

yeah, how the f was taken

🤣

[mossy nest]

I was using use-computer it was working by these settings but it happened after updating

[boreal holly]

Gotcha, sometimes after an app updates, macOS makes you re-enable the security settings. That's my first suspect.

Either that, or part of their update was relocating how to enable computer use. Like the button exists somewhere new now

[boreal holly]

You know what, try enabling

[features]
apps = true
plugins = true

I have them disabled, and for me it says computer use is unavailable.

[unique spade]

Yea sure I m doing my stuff with codex lol

[mossy nest]

Use computer and codex still in privacy settings enabled but after updated it shows plugin not available

[mossy nest]

Where to update these lines

[boreal holly]

~/.codex/config.toml

[hard drum]

does disabling apps also disable the annoying codex_appsMCP?

[mossy nest]

Thanks let try this

[boreal holly]

Afaik, I have apps = false and my agents do not have that MCP

[lean lark]

@torpid trout Those are some seriously impressive screens. You may recall my ongoing mantra here about using AGENTS.md to direct the assistant to document code functions, and to generate and maintain user documentation for all features. That would include the README, features.md, detailed docs for each primary feature/screen, and developer documentation to explain how things work. This is not only good for us to understand what's going on, and for users - the assistant gets to read the relevant documention in each thread, significantly reducing the time to process changes. Otherwise the assistant needs to learn the app with every new conversation - that's wasteful of time and tokens, and significantly increases opportunities for errors.

[torpid trout]

that is baked into my agents, yes
yet even so... I mean, you know, I am not a freaking journalist lol

Hence my earlier meme

#codex-discussions message

Bsaically for once I bite off more than I can chew lol

Now I want the same thing for Developer biz lol
At least there I will understand the domain-specific terms haha

[lean lark]

The other day I had that same moment : I don't understand what it just did for me. Sometimes I don't understand the challenges it notes when it finds something wrong and I need it to explain ... in my own project.
I suspect those of us who care to discuss the projects with AI will all have these moments from time to time and the weirdness will become the norm.

[small violet]

wheres 5.5-cyber at

[lean lark]

Isn't cyber a separate product?

[signal tapir]

I think you can request access to it

[young locust]

yeah its separate with id check

[torpid trout]

What is that?

Me wants, i dont know what I will get but me wants.

[signal tapir]

I believe it's a cybersecurity specific model

Correct me if I'm wrong

[torpid trout]

aah

[lean lark]

https://openai.com/index/trusted-access-for-cyber/

[torpid trout]

That this here:
... yeah you faster. There was also a discord channel, but cannot find it now

before it was called something else, not cyber, but security or so

[lean lark]

https://openai.com/index/strengthening-cyber-resilience/

https://openai.com/index/introducing-aardvark/

Introducing Aardvark: OpenAI’s agentic security researcher
Now in private beta: an AI agent that thinks like a security researcher and scales to meet the demands of modern software.
March 6, 2026 Update: Aardvark is now Codex Security, and is available as a research preview.

[torpid trout]

lol

what the heck

[lean lark]

[ those flashes happen when people post BC stuff and other ads or off-topic images ]

[small violet]

https://x.com/sama/status/2049712078836170843?s=20

[torpid trout]

I am sure we had a discord channel for that?
cannot find it 😄

Also I did never complete ID on my account... but I can access that thing in the cloud, or at least was able to, under the PRO sub

Is that over?

[torpid trout]

I am appreciating the fact that it deems my memes as on topic.

[lean lark]

https://developers.openai.com/codex/security

Codex Security works with connected GitHub repositories through Codex Web. OpenAI manages access. If you need access or a repository isn’t visible, contact your OpenAI account team and confirm the repository is available through your Codex Web workspace.

I'm surprised but pleased that your cat meme survived filters. 🙂

[torpid trout]

Its because codex creates them lolol

[torpid trout]

So this is distinct to aardvark?

[lean lark]

Ardvark was renamed to Code Security.

[slender plinth]

guys what does feeling codexy mean

im new to this

lol

i got plus sub for like 1 week

[signal tapir]

maybe that someone was feeling like using codex?

[lean lark]

Where did you see this new term?

[slender plinth]

x

https://x.com/thsottiaux/status/2049880535577854338

[torpid trout]

Release incoming

[potent mason]

He usually says it when there’s a big codex update on the way

[lean lark]

Hmmmm, that could indicate a feature drop.
More likely it just means Tibo is getting more use out of the platform today.
We get into a groove using these tools...

[torpid trout]

Dang, so it was not model release, it was codex release 🙁
5.6 still pending then grin

[boreal holly]

It's probably gonna be some big feature like "upload the picture of the back of your head, Codex will generate a picture of your face with 90% accuracy"

[lean lark]

#announcements !

[slender plinth]

is he gonna reset the limits

[torpid trout]

I will try the back of something else immediately, if this drops

[slender plinth]

thats. not smth for codex lol

people say goblins or smth

or usage rest

[torpid trout]

Daily dose

[slender plinth]

LMFAOOOO

[torpid trout]

off, off, off, off?
ReallY?

😛

[slender plinth]

im 14 okayy 😭

dont insult me pls

[lean lark]

Just please don't ask about resets. 🙏

[torpid trout]

reset wen

[slender plinth]

oh why

[slender plinth]

reset wen

[torpid trout]

A bit young for discord aint you mate

[slender plinth]

nope (:

[lean lark]

🤦‍♂️

[slender plinth]

im allowed

[torpid trout]

no one planning to fam

[signal tapir]

we all hope for them, but nobody can affect when it happens

[slender plinth]

well the reset wen photo ):

[torpid trout]

hows that an offense lol

[signal tapir]

It's posted here every day

[torpid trout]

Did you drop your humour somewhere?

Sorry deeply if said meme offensed you, I will duly report it to the machinery that created it

[slender plinth]

ty

[boreal holly]

I know what you mean by that 😂

[lean lark]

As a buck-toothed brainless troll, I was feeling a bit insulted here by that meme too...

[torpid trout]

Yeah, I had a feeling

[hard drum]

people use watches on their right hand?

[lean lark]

lefties, yeah

[hard drum]

i thought it was "if you RH, then watch on LH, && in reverse"

[hard drum]

oh, so i didn't think wrong?

[lean lark]

you thought "right"

[hard drum]

i thought some people in my country telling me i'm wearing my watch wrong was true

but am right-handed, so left-hand watch

[lean lark]

u do u

[boreal holly]

Adds weight for more impact during fisticuffs

[hard drum]

i do mostly what people tell me to do

[signal tapir]

I just do what ai models tell me. 😉

Actually, if I did, I would be in so good shape. Mentally and physically.

[lean lark]

That's dangerous - Like a politician who believes when ChatGPT tells them how to write new policy to defend against AI 🙄

[signal tapir]

hehe, yea.

But I do get some good advice most of the time. Especially when it comes to health and cooking.

It's just that I already know it all. But AI don't care. AI halp anuwai!

[mossy nest]

still same issue .. Computer use plugin unavailable

[boreal holly]

You should run

/Applications/Codex.app/Contents/MacOS/Codex

In the terminal, which will launch the GUI, but you may see errors being logged in the terminal. Also if you open the Console.app (macOS built-in logging tool), you may be able to find official logged errors from Codex GUI pointing to the real issue

[exotic cave]

in 0.128.0 js_repl has been dropped, good or bad? and does code_mode works the same?

[boreal holly]

Here you go bro, I enabled apps & plugins, and I had to manually install it in the plugin marketplace. See if there's an add button

[lean lark]

If js_repl is only used internally as a tool then I don't think it matters. If the JS is surfaced, or if we have the ability to inject a dependency into the call, then I'd consider it to be a concern.

[boreal holly]

I'm curious to know as well what this feature even is and why it's better than opening a terminal

[lean lark]

I believe js_repl is only used for internal calcs and determinant processing.

[exotic cave]

It gave a persistent node backed js runtime with top level await, local imports, package imports, helper access, and reusable bindings

the model can run js logic to call multiple tools, transform output, loop, filter etc... while staying inside one tool flow.

[boreal holly]

I think I prefer giving the agent a jupyter kernel for stuff like that

[halcyon ingot]

Hello, I have a question, please. Does the ChatGPT Go subscription have better limits for Codex than the Free subscription?

[lean lark]

It seems to me that it is very Jupyter-like but only for internal use by Codex. Since Codex can do this with Python, I'm thinking they're deprecating the internal dependency.

[lean lark]

Interesting question.

[chrome raven]

this /goal how can we use it?

[lean lark]

I don't think so, @halcyon ingot The pricing page does not show a detail item for Go between Free and Plus : https://chatgpt.com/plans/go/

[halcyon ingot]

Because I'm looking for an affordable subscription to do some development work in my free time.

[boreal holly]

I mean with jupyter they can replay the steps they took to do something and you can see the entire sequence + state of each variable & add markdown cells. I'll probably keep that repl disabled

[lean lark]

Many of us do all of our professional development with Plus, with no requirement for additional credits.

[cyan wing]

anyone else not seeing /goal command?

[rain gorge]

same issue

[torpid trout]

Js repl is sort of a browser tsting thing

That’s gone now!?

Guess I didn’t update yet since it still used it here just a minute ago

[torpid trout]

Probably one of them released but not released features

[lean lark]

well, it's just an internal tool - again, if it's surfaced for our usage then I'd like to see docs and understand how it was intended to be used. I can think of some applications, but don't actually "need" them right now.

[boreal holly]

Probably have to press esc 4 times to bring up the goal menu 😂

[torpid trout]

The model uses it to test what it created
Instead of headless chrome

At least that’s what I’ve understand so far looking at it when it used it

[lean lark]

That makes sense... again, for determinant content.

So if it generates JS for us it should be able to build/transpile that code just to verify it. If it can't do that I'd be annoyed. That said, if the Code functionality in 5.5 is now so good that it doesn't need to build code to verify it, I'm kinda OK with that too.

[boreal holly]

I wonder how it works with sandbox enabled haha

[torpid trout]

it asks a million times for permissoon

unless you have the guided permission on

I gues they left again

They did so before, due to Phisicyans

[boreal holly]

I'm pretty sure codex with unified_exec=true already has repl. They just launch node, then pass

cat << 'EOF'
let hw = "Hello world"
print(hw)
EOF

to stdin, and it stays alive even after they finish their turn. They just keep passing stuff into stdin on that process and it keeps repl'ing their commands. No need for the extra feature, already obeys the sandbox

[lean lark]

I'm looking through the repo for notes about js_repl. It's all very interesting but lack of available time forces me to avoid such pleasures.
https://github.com/search?q=repo%253Aopenai%252Fcodex+js_repl

[small violet]

who remembers the "Physics PHD" guy

that was his name]

[torpid trout]

Yeah, that one

[small violet]

i doubt he has a phd

[torpid trout]

Phisicyans are scammers

[small violet]

lol

all of them are just trying to sell their product

[lost drum]

lawyer too

all in one

[boreal holly]

Oh yeah? The physics phd guy was selling something?

I just remember him enforcing the "be noice" rules

[small violet]

codex for cheap

let me show u

[lost drum]

haha

this guy too

[idle cypress]

new codex update might be the best one yet

[lost drum]

which one

am I missing osmething

[idle cypress]

its clean af

[small violet]

bro is behind

[lost drum]

ye I jsut use the vs codex extension so ye I might be

[orchid plume]

goals feature was added to Codex as well, for long running tasks, like a ralph loop

[small violet]

im jk i also dont know what the update is

[small violet]

dosent everyone do this

[lean lark]

Actually the Physics PhD guy seemed to note channel decorum only after I did. I think he was mimicing or being facetious.
He disappeared after insisting on some detail at OpenAI that I verified was completely incorrect. As in HHGTTG, he disappeared in a puff of logic.

[lost drum]

Buy from me

[small violet]

i think he left or got banned

[lean lark]

time for 🥪 🥣 🥡 ☢️

[small violet]

yum

[lost drum]

he just got payd to advertise that service

[small violet]

whats the last thing

[small violet]

can i ask u a question

[lost drum]

yes

[small violet]

if u had to pick between these 2 names for your software dev startup would u pick 88tools.net or crasule.com

and if u had 88tools.net would u rebrand to crasule.com

[small violet]

whos paying him to do that

[lost drum]

depends what you selling

[small violet]

nah just answer

with that context

[lost drum]

i have whole playbbok about positioning your brand

[small violet]

a book?

[lost drum]

ye it has like 33 chapters

[boreal holly]

Probably realized it was a dead end and went back to pressuring old ladies into buying lots of gift cards

[small violet]

and youve read everything

[small violet]

he's stupid for trying

i thought 5.5 cyber was going to be generally available

bro thinks he's mythos

[cyan wing]

Hey guys you have to do codex --enable goals
to activate it 🤣

[lost drum]

ye I made it

[small violet]

is it free

[small violet]

codex cli

[lost drum]

its just mine its not for sale

[small violet]

oh u made it for urself

why would u make a book for urself

😭

ggs guys

[lost drum]

#off-topic

[nocturne folio]

i havent seen codex cli in so long

[nocturne folio]

88tools seem chinese

the name atleast

[small violet]

go to offtopic

[nocturne folio]

ok

[rain gorge]

probably using some sort of hacked account

to power the service

[lean lark]

I need to schedule, dedicate evening time just to keep up with Codex features. 🤦‍♂️

[chrome raven]

nicee

[unique spade]

yup lol

[chrome raven]

how do you get access?

[lost drum]

is it better or what?

[chrome raven]

Better for cyber security

[south compass]

I miss the commit button and diff view from previous update..

now I have no idea what it did without clicking in the UI and im stuggling to find the commit button

[lean lark]

@teal cargo @left pecan Any chance to get regular updates in #codex-updates for changes to Codex NPM, Apps, Web, or Extension? There's no unified "Codex" place to go for the latest info - info is all over the place or simply non-existent. Thanks.

From the latest Codex Extension update:
"Introducing" Fast mode : The feature that's been in the product since v0.0 that most people should actually disable to avoid having their token quota gobbled up.

[small violet]

lol i used imagen for this

[lean lark]

I dunno when this happened but it looks like after I updated to 11.13.0 the permissions for Codex changed from Full Access to Default. Did I miss some notes in here about that?

[grim perch]

i dont think full access was ever default but the introduction of auto-review mightve reset it back to default? if thats what you are wondering

[lean lark]

I'm not saying full access Was default. I have a good sandbox so I have it set for current use - at some point something changed it back to Default (limited perms).
Whoops.... Looking at config.toml, the one project I just opened is not listed with a trust_level. That must be the issue. 🙄

Hmm, no, that's not it. Oy. Will look later.

[plucky orchid]

Anyone know if GPT 5.5 is available on AWS Bedrock yet ? On OpenAI website they said that yes but on AWS there is nothing at all.

[rocky fog]

💀

[nocturne folio]

if openai acquires github i think they would fix the uptime

[solemn acorn]

oh that's cool, the codex desktop linux project has computer user support (for people who do have access which I don't yet :c)

[jolly lily]

Honestly unlike Opus 4.7, GPT 5.5 seems to spend such a little amount of the model limit

[torpid trout]

where do you get this from

Jane was not banned. Nor did she sell anything (not on her anyway)

[boreal holly]

Jane leaves when the "newcomer" badge disappears and eventually comes crawling back and pretends to be a newcomer

[torpid trout]

That is a smart move, allows for more tailored ragebaiting.
I might need to adpat that

Leave every day and rejoin and drop an offeffeffeffensive memememme

[boreal holly]

uwu 👉👈 is opus better than gtp?

And then somebody comes in and says "use gemini"

[cedar skiff]

it's ust better over all, it has it's own issues, but they are much more reasonable than claude models

[neon girder]

slopus

[jolly lily]

Yeah, as always, hopefully OpenAI sustains the large limits in long term

[torpid trout]

If they are any smart, they scoop up all users, lock them in, then slam the prices

[cedar skiff]

they are squeezing it as well, they are just a few steps better than anthropic, i mean 5.5 cost more than opus per million.

[fleet geyser]

But it's also more efficient right?

[cedar skiff]

not too sure, the usage limits are better one way or an other

[boreal holly]

Each token is higher quality

[fleet geyser]

Definitely

[lost drum]

5.5 xhigh worse gpt 2.0 worse than 3 days ago

only me?

[boreal holly]

Yes

[lost drum]

oh

[boreal holly]

Use medium

[lost drum]

wdm

why would medium better

[boreal holly]

More decisive

[torpid trout]

Medium rare

[lost drum]

he just seems to reason less than before now it feels like 5.4

[torpid trout]

I am suing a mix of medium and high on 5.5, I really so far not once used xhigh with it, completely forgot it exists.
Let me try.

hello codex create me a meme...

[boreal holly]

I removed xhigh so I don't accidentally select it, that's how useless it is

[nocturne folio]

lol

[lost drum]

wdm why

I dont get it at all

[boreal holly]

I'm not researching an entirely new branch of physics, or solving the mysteries of the universe, so xhigh is worthless

[torpid trout]

Did you ever talk to one of those guys who always overthinks it?

So, thats what xhigh can do to code.

You ask it to create a quick edit in docker compose file and it comes back with an assertation as of why this is not a new chatbot app

🤷‍♀️

[lost drum]

I feel like talking with not thinking model at all when using xhigh like idk he seems to not go deep enough when I tell him too

[torpid trout]

You sure you use 5.5
Because i can agree with you there on 5.3 model
5.4 was useless throughout the entire range, no matter what

[lost drum]

I mean maybe I was on stages in my project where it was more to finish and now teh polishing steps are blicking himeach run cause he assumes eveyrthing is validated and works

but I remember taht he used to create better plans 3 days ago with same context

[boreal holly]

I mean idk what you're using codex for but if it's software engineering then rest assured, it thought very deeply about everything you prompted it, even if all it did was change the button from blue to green

[lost drum]

idk that system so complexed that I wonder how the hell it works

[chrome raven]

seems /goal is pretty nice

[lost drum]

I dont have it in vs codex extension and he already has my goals I think its jus matter of me testing the system with him I just cant inget the fact that its now testing it out haha

[chrome raven]

i think only avliable through cli for now

[small violet]

5.5 low is nice

it dosent complicate things

[boreal holly]

I still gotta muster up the courage to do low. I hear it's a lot faster!

[glacial shadow]

Were you finding xhigh to trip over itself on simple tasks and take too long

[gentle harbor]

why is it black ?

is that a glitch

i love the black way more then grey

[boreal holly]

windows or mac?

[gentle harbor]

windows

[boreal holly]

Unplug all USB devices, restart computer

[boreal holly]

As a matter of fact, I found it performed the same as medium, but took way longer on each task and was not decisive.

[gentle harbor]

wha ? why would i unplug usb devices

[boreal holly]

Hardware interference. It's a real diagnostic step

I mean if your keyboard and mouse are USB then try different keyboard and mouse, or just unplug everything but those two devices

[glacial shadow]

Are you saying med is more decisive

[boreal holly]

If you have USB headphones, bad drivers or a bad device can cause the graphics stack to do weird stuff.

[boreal holly]

Yes, medium is more decisive. Especially during times of troubleshooting, it will decide to run tools to gather more information, whereas xhigh would think deeply about what tools to run

[glacial shadow]

Yes listen to him, he knows his stuff, unplug and restart

[glacial shadow]

Interesting

[lean lark]

( he's a bot, don't listen to him )

[glacial shadow]

Hopefully we see a flowing state of reasoning effort in future

[gentle harbor]

welp

[boreal holly]

There's another knob I haven't tried: "verbosity". But it seems to do well with medium reasoning and default verbosity in every situation I throw at it

[lean lark]

click Retry.

[gentle harbor]

just deleted my message and started talking about random things

[lean lark]

Goblins ... and ChatGPT ... not related to this channel.

☝️ only applies to tiers above plus

[glacial shadow]

Ya turn that up when you’re trying to learn something

[unique spade]

got the issue twice in a gptpro thread on the follow up prompt

that thread was odd from the beginning cause gptpro took 90 min oin the 1st task, even though in a different thread it did it s usual below 10 mins on a task that is not very different

[gentle harbor]

im guessing servers are being abused heavily right now

[unique spade]

yea at least chatgpt

codex seems fine

[lost drum]

I agree

my pro one is thinking now for 5min and telling he cant access zip

[jolly lily]

LOL

[unique spade]

:))) yea you're right

[jolly lily]

I honestly never thought the plus plan would be more than just "worth" it, hopefully they don't smack down the limits in the future

[shrewd flint]

what do you guys use codex for? i wonder if it can be useful for someone who isn't a developer / founder

[shrewd flint]

isn't there a temporary 2x limits haha

[jolly lily]

I myself use it to enhance my engineering/software progress as programming being my main field

[jolly lily]

I think it would've been worth it even if the double limit event weren't active

[boreal holly]

GPT Pro does not have the entire set of tools that regular GPT has. You should have GPT operate on the zip, and switch to GPT Pro when you need it to think about the contents

[jolly lily]

The limits are honestly very generous

[shrewd flint]

you're lucky

haha

[jolly lily]

Why tho?

[unique spade]

you can use it to automate anything you want

[lost drum]

how would I make gpt operate thuth files while make gpt pro think about them in normal chatgpt.com

[shrewd flint]

I'm doing a bachelor's in actuarial science and interning at a company where using AI isn't allowed, no idea how to make use of my codex

[unique spade]

no bro, gptpro has no issue with zips.

i give it full repo and ask it to add proto-modules. and it's all fine

what it lacks is stuff like rust etc

[shrewd flint]

CS students must be feeling like they're in heaven rn.. @jolly lily

[unique spade]

do you have rational cognitive workflows that also include in some manners interacting with digital data?

cause if yes you can delegate a lot of them|

from doing research on topics for you, readin/replying emails according to your prefereces, operate various apps on your computer

[jolly lily]

Um honestly it's sometimes the quite opposite, only using automated workflows actually weakens the ability of solving problems

[small violet]

what kind of company is this bro

[shrewd flint]

oh hell no this guy's here aswell

[unique spade]

one that will go bankrupt soon 🤣

[boreal holly]

Interesting. It's been a while since I tried it but maybe it's gotten better

[shrewd flint]

i think it depends on the way you use it, i feel like AI is so much better in coding/cs than in as

[small violet]

no ai rule is still stupid

[shrewd flint]

it makes sense tbh, we work with sensitive data and one mistake can cost us a lot

[unique spade]

that s all work from the repo zip

[shrewd flint]

i mean i'm an intern so i have less responsibilities haha

[shrewd flint]

not really

research yes but no ai allowed according to my uni

[boreal holly]

Oh I see so you're at capstone

[unique spade]

use it as DJ then have it play youtube for you

since you are not allowed to do advanced cognitive stuff with it 😂

[lost drum]

nahhhhhhh pro model bonkers

Investigating potential cause of ClientError It seems I might need to create the files in chunks, though the command length looks fine. The ClientError could be due to CPU overload, but I’ll wait to see if the processes timeout after 10 seconds. I’ll give it another 15 seconds to see if they die naturally.

[shrewd flint]

what does this mean

[shrewd flint]

haha, claude does this for me already

[boreal holly]

The end of the undergraduate degree program, so can't use AI

[shrewd flint]

no I'm a 1st year

it's just not allowed at all

[small violet]

intern at 1st year

how

[shrewd flint]

for any student

yeah

[small violet]

😭

[boreal holly]

Huh, I did my internship at the end of my bachelors

[shrewd flint]

I have a lot of issues w my attendance %

[boreal holly]

AI didn't exist yet (GPT 3.5 was not all that handy lol)

[shrewd flint]

due to my internship, but I wasn't going to my classes anyway so

[small violet]

is it a paid internship

[shrewd flint]

now I have something to blame it on haha

yeah

[small violet]

damn

career maxxin

[shrewd flint]

why would I do an unpaid internship

that should literally be illegal

the people I work with get paid a lot more than me though, about 10x as much

8x actually

[boreal holly]

I did... for 4 months. At a company that was not going to hire any of their interns lol

[shrewd flint]

oh god

why would u

[small violet]

alot of ppl do this

[unique spade]

maybe it's not allowed to use in the papers/homework you submit

they can't disallow you to DYOR with it

university doesn't own you. they can only impose limitations on certain artefacts you are required to produce

[boreal holly]

Experience 😕

[small violet]

because its a requirement for grad and most companies offer unpaid

lol

paid internship in 1st yr is crazy

i hope i get a nice internship

[boreal holly]

What really sucked was doing 4 months internship while finishing degree working as a HVAC service tech full time. Still don't know how I made it through all that 🥸

[shrewd flint]

honestly we dont learn that much during lectures etc, it's so annoying to be forced to sit there and learn nothing

[shrewd flint]

wow

if only codex could attend my classes/lectures for me

[unique spade]

no one "forces" you. you enrolled to uni out of your own will

[shrewd flint]

well it's mandatory, even though I could get good grades without them

[unique spade]

it s ok all of us who went through university went through that lol

yea :)))

[small violet]

imagine being in uni pre gpt

i dont think id survive

[shrewd flint]

maybe i just have adhd

[unique spade]

we all have, switching before various codex threads all day long is a sure way to get it haha

[small violet]

who are u to decide that

lol

[cedar skiff]

I have a degree with honours never needed to do anything like that. Not deserved?

Yes?

[small violet]

i dont care

😭

[shrewd flint]

😭

this guy

[unique spade]

lol

[cedar skiff]

That's Mr Wimp to you

[small violet]

@bronze sable are u physicist

i changed it

[boreal holly]

Shared subscription kind

[small violet]

codex maxxin kind

[shrewd flint]

the scam one?

[cedar skiff]

phd in fine arts?

[shrewd flint]

a person called jane invited me with their referral link 😭

[boreal holly]

We get em a lot

[cedar skiff]

or maybe general humanities

[shrewd flint]

they had their whole dedidcated scam referral link it was so funny

[unique spade]

is that time of the day when this chat is mostly useless banter?

[cedar skiff]

It only takes one

[shrewd flint]

what are u guys talking about

no

[cedar skiff]

5.5 so good at working out how to get stuff going for skills that can be used by weaker models

[unique spade]

once you do your own UX/harness, vanilla feels so lame

[small violet]

phd in bafoonery

[cedar skiff]

i have a converstation with 5.5 i have been working on manual app testing and it's been through 5+ compactions and just doesnt even matter. It still just works

[unique spade]

for example codex desktop instead of telling me when was each thread last used, it tells me when it was created

which is very useless info in finding the threads in each project i am actually active

[small violet]

i literally dont care

[cedar skiff]

what do define as a UX/harness?

[boreal holly]

sounds like you're stalking angel yang. sus

[unique spade]

UX is just the interface, you can combine it as i'm sure you know with appserver via cli exec for backend

or you can also implement your own backend harness connecting directly to oai server upstream

[shrewd flint]

why are you lying?

[cedar skiff]

ahh right you mean making your own app layer?
I'm really happy with codex app.

[boreal holly]

Agreed 💪 OpenAI really cooked when they made the app-server. Straight from the heavens

[unique spade]

i like it too as an app

but it s still good experience to craft your own

and sorry, i thought you were Robert haha, he has his own ux

[boreal holly]

Tbh mine does not look nice - design is not my strong suit - but the performance and memory usage is what made me do it. Also removing features I don't want and putting in stuff I do

[cedar skiff]

I spent a lot of time working on this sort of thing early on with claude code. I had so many ideas of what i wanted and how i could make it better, i spent weeks working on things that would improve my work flow etc. But turns out i wasn't getting my real work done.
This is a common path i see, if your fine spending the time great go ahead and do it. I think it happens with devs because this is our domain of expertise, so it's where we see opportunity. I reevaluated my ideas and focused on what is making me money. Which are all outside of the software improvement domain.

[boreal holly]

Oh dude, I did the same thing. Since August 2025, I had like 2 other frontends and harnesses that ended up consuming all my time and providing no value (except I guess a learning experience). The latest iteration, I have not touched it in weeks. It just works! Finally happy with it

[cedar skiff]

It feels like a trap where the perfect solution is just a few more prompts away every time

[unique spade]

the small usability wins are what i enjoy

for example working on 2 or 3 projects at the same time

i'd rather be able to see those chats side by side in one workplace

instead of having to keep hopping from thread to thread like in 2010

[boreal holly]

I just have 1 chat view, but my bridge caches up to 50 messages so they load instantly when I switch threads 😁

[small violet]

and im idle now

[unique spade]

or this tiny tweak. when i have gptpro do reviews which i copy paste to codex, in current surfaces i end up with huge replies of mine that i keep have to scroll

so i just made my own messages to keep open only 10 lines with the option to expand if i ever need it

also loading by default only last 10 turns

with the option to load more if i ever needed it

for me all these, win time , but sure they are tweaked on my own habits

[boreal holly]

Oh yeah, you know the app-server supports structured outputs? I was gonna set it up so if an agent takes screenshots, the agents add em to an images array of absolute paths and the frontend shows em

[unique spade]

with only last 10 messages displayed by default, my chats load instantly though, even if they have hundreds of messages

[torpid trout]

Is it time for a meme?
Or are y’all going to be nice to each other?

I leave this chat for 5 minutes and all hell breaks lose

[boreal holly]

I think meme time is on the horizon

[unique spade]

any of you actively using claude too?

how it "the other one" doing lately

[torpid trout]

You read my mind.
I realized my latest meme wasn’t as funny as could be, and was thinking of doing a physician scamming lol.
Guess you beat me

But the joke landed late

I get it… I’ll scale back on the memes 😅🤣

[oak trellis]

I hope we don’t have a anthropic wave coming in and make things worse lol 😂

[boreal holly]

Well bro, if you were here when stuff was poppin off it would've been 🔥

[boreal holly]

It's only a matter of time 😬

[unique spade]

oh those are annoying af

[mental urchin]

is there any claude/codex skill/plugin for cf

[boreal holly]

I remember when the night time robert would tell em to take their A/ discussions to #ai-discussions

[unique spade]

what s cf?

[mental urchin]

cloudflare the company that protects like 20% of the internet

[unique spade]

it s good you didn t just say C then :)))

2 letters in the alphabet standing as acronym can mean a lot of stuff

[frail meadow]

I think they have their own skills, at the very least codex knows how to use the CLI really well and investigate the docs

[mental urchin]

lol yeah ik but cf is widely regarded to as cloudflare

[mental urchin]

hmm i gotta find those quality skills

[boreal holly]

Codex knows precisely how to use cloudflare CLI without even looking it up on the internet or reading the help. Might even be built into the training

[frail meadow]

You ever been on skills.sh ?

[unique spade]

i d imagine is built in too

[frail meadow]

Isn’t there a cloudflare codex plugin? I’m not on my laptop atm

[unique spade]

https://developers.cloudflare.com/agent-setup/codex/

lol

[frail meadow]

Literally first result on Google lol:

https://developers.cloudflare.com/agent-setup/codex/

I just saw it too haha

[unique spade]

yea i was on my pc so when you said you aren't i jsut copy pasted cloudfare codex plugin to help u haha

[mental urchin]

damn no

wait i might be missing smth

am i slow lemme check that out

holy there are like 90k skills but damn its not open source. could've forked it and deloyed it on my own domain

thx trippyprism

[dark sun]

I can use my entire opus 5 hour limit on the 20x plan in about 15 minutes

[shrewd flint]

how?

not saying you're lying but that's almost unbelievable

[dark sun]

I started working at 5:20, when I sent the first message

It's 5:25

This is about half-tilt, if I go full-tilt I can use about 20% in 5 minutes, sometimes up to 30%

This is Claude btw

[cedar skiff]

Seems to be a common story now days

[unique spade]

that's pretty much unusable for any serious work... max 20x should at least give you the basic of being able to have your own pace

[high girder]

Max 5x best value. Same with the gpt $100 plan. Both get you a lot more than what either $200 option would

[unique spade]

oai is orders of magnitudes better in this respect

[maiden escarp]

Is it preferred to use Claude Code and Codex together to build a tool or an app?

Or can use either one to build

[unique spade]

well for gpt, 200 is pretty much 2x100 since you get double codex usage

[high girder]

You can use either, but it comes down to preference. I use both

[unique spade]

currently my weekly workflow aligns quite well with the 100 plan so i am using that

[high girder]

Yeah, I meant instead of paying 200 for one plan, buy the 100 of each and get the best of both

[unique spade]

ah ok. yea that's fair. assuming you do find value in both similarly

i m sparsingly using opus for review work via it's availability via google sub in antigravity

but my experiences in actual implementing in code what i asked opus were rather bad, so i am not using it for anything serious

[high girder]

I understand why opus 4.7 gets the crap that it does, but that's also a workflow issue for the most part. It's the change in model behavior that people need to adjust to.
Ideally, opus is excellent at holding and sticking to a well drafted plan, codex is a little monster when you let it execute the claude plan. It just goes to town and cleans up claudes mess

[shrewd flint]

what's ths physics phd lore I want to knoww

[cedar skiff]

This is a common thought in the claude code community. That we see much less of in the codex community. The idea is you can solve it with a better work flow, maybe it needs more detail or another layer, some extra guard rails etc. Lots of the conversation in the claude code channel is about how to wrestle opus to do what you want it to do.
Moving from claude to gpt for coding is not intuitive. The first thin that happens is you notice a mistake codex makes that claude doesnt - some idiosyncrasy that needs to be handled. This leads to a miatake in thinking that because claude doesnt make that mistake fit's better. But claude requires much more work to get results, the difference is people are already used to it's quirks.

codex and gpt are just better at coding. It used to be that it was really only if you know what you your doing with coding it was better, but with 5.5 they really closed that particular gap. 5.5 just gets it better now

[high girder]

I agree, but I also think the use case comes into account there. I don't use either of them for just code. In fact, at code, I find codex is usually the better choice. Claude is much better at systems though. When you're working across multiple systems and keeping things managed, claude usually handles that much better than codex would. They both have things they're better at than the other.

[frail meadow]

Found another alt account

[cedar skiff]

5.5 closed the gap on things like computer setup, things like setting up open claw with profiles and what not on a headless server. that stuff used to be in claudes space more. Claude still holds the crown for research, documentation and communication style work.

[high girder]

I need to test more with 5.5 for the things I usually have claude handle then. 5.4 didn't cut it, but I'll need to try that soon

[cedar skiff]

5.5 digs more and finds ways to get it done

[frail meadow]

What kinda stuff?

[high girder]

I usually use a claude parent session for multi agent work. spawns in claude, codex, and other local models as needed, but if GPT 5.5 can handle that, then I can try something else with claude

it's one of my favorite things about opus 4.7, is the ability to follow directions and keep them for the duration of the session

[frail meadow]

One thing I will say is I’ve had much more success with long time horizon tasks using gpt-5.5 on xhigh mode as opposed to medium or high

[cedar skiff]

5.5 is the best orchestrator of all the gpt models. I go to bed and run automated orchestration sometimes, and i always got up and it would have stopped and asked me something(5.4) 5.5 was still going, it worked for 12 hours straight in orchestration and it only stopped because codex app had a memory leak that start caused apps to pause on my system.

[high girder]

that's one of the things Claude was much better at. I guess I'm about to go test it out

[cedar skiff]

This is the screenshot i took when i got up, it went for another 4 hours 0.o
#codex-discussions message

[glacial shadow]

What was he saying, what were people arguing about towards him?

[cedar skiff]

I read your message history @bronze sable it's all just spam

[high girder]

dude. It was a racist and chaotic mess. Leave it over there, keep this place clear. He also got nuked from the claude discord

[cedar skiff]

Wont take long for him to get the boot here either

[quaint hazel]

Has anyone found a way to make codex write less tests?

Bro writes 20 LoC on the feature, then 600 LoC on tests.

Currently running an autoresearch loop with the sole goal of reducing test LoC and so far it has removed 78.8% of the test code it wrote 🥲

[cedar skiff]

Do you still want tests?

[quaint hazel]

Yes, but just regular tests that a human would write, not 200 test cases

It's down 78.8% without losing actual test coverage. It's insane.

[cedar skiff]

I have a guard rail skill that is used anytime unit test are written that tempers it. has stuff like this in it:

## Core Rules

- Assert observable behavior: outcomes, state transitions, emitted effects, persisted changes, and externally visible interactions.
- Do not assert internal triggers, private mechanics, or incidental call sequences unless that interaction is itself the public contract.
- Keep assertions resilient to harmless refactors.
- Keep test-writing work focused on proving behavior; do not change production implementation solely to make a newly added or corrected test pass.
- If a correctly scoped test exposes an implementation bug, keep the bug evidence visible; do not weaken the assertion or rewrite the test around incorrect behavior.
- Keep test files concise and comment-free; encode intent with clear test, group, and matrix names.
- Keep touched test files analyzer-clean.
- Prefer the shared outcome assertions (`expectOutcomeSuccess`, `expectOutcomeFailure`) over bespoke outcome destructuring.

## Test Layer Selection (Cheapest Proof First)

- Prove behavior at the lowest-cost layer that can validate the contract:
  - pure logic: unit test
  - provider/notifier behavior: provider or notifier test
  - UI composition/semantics: widget test
- Escalate to a heavier layer only when a lighter layer cannot prove the contract.
- Avoid duplicating the same contract across multiple layers unless the additional layer protects a distinct risk.
- If the required proof crosses app/runtime boundaries and needs integration or emulator coverage, stop using this reference and switch to the integration-testing guidance for that stack instead of stretching unit/widget/provider rules to fit.

[cedar skiff]

There is a whole to more that i built over time. but the goal was the same as yours

I also have some stuff about using test harnessing so test are a little dry because it reinvents the wheel a lot.

Cuts down on the tokens and time taken to write and fix them.

[quaint hazel]

Thanks, I'll have a go at fixing it further thru prompting 🙂

[cedar skiff]

To make the skill i got a session to write the tests for the same code over and over. It was get a subagent to write the tests, audit for my requirements, update the skill, remove the tests, repeat.

One thing you have to be careful of is that it doesnt start getting explicit in the skill updates. It might start making rules in the skill that are specfic to one test.
So you have to prompt to be generalised fixes that will work in all cases

If you get one explicit instruction, then it will take that as a green light to do that all the time.

[oak trellis]

5.5 is so good

[quaint hazel]

had it go through the changes and propose a plan to make it better

Gave me quite a few promising tips that it's now implementing

[glacial shadow]

Tell it to write less tests

[oak trellis]

thank you openai for the great codex .. its sooooo good .. math .. speed

uff

just something the weekly limit something off but its ok .

[boreal holly]

<=5.4 models, I would archive them for being totally broken after a while. This 5.5 orchestrator has been cooking non-stop for days without so much as a “whoopsie”. Zero drift, infinite compactions.

[oak trellis]

cooking like nobody else .. i start to get annoying because its like a drug lol.. knowing they will do something later .. pump and dump emotions hahahaha

[boreal holly]

Yeah bro I just keep adding tasks. Itll have like 6 agents concurrently working on 6 completely different things, adversarially review their work, tell em to fix stuff, keep em busy. Even if a lot of steers come in all at once it never leaves business unfinished

[oak trellis]

beautiful .. what you using the codex app ?? iam still on their marketplace app in cursor IDE ..

got twitter API for 100x cheaper .. works flawless

[boreal holly]

custom build 🤪

[oak trellis]

ah damn nice! so you couldnt find certain features you needed ? oh its on the phone ?

yes was thinking .. is there a phone app for codex -> my server

[quaint hazel]

Robex was right there

[boreal holly]

Yeah, I just like messing around. If you built it yourself you know precisely how it works and it makes ot easier to use

[high girder]

well hot damn. I got 5.5 running behind my system for a bit to see how it goes.

[oak trellis]

doing the same .. was building own claude code market place app when i didnt like the first cl last year ..

[boreal holly]

With 5.4, If I had like half a dozen agents running, and 3 of em message the orchestrator at the same time, after a while it would handle one of the 3 and let the other 2 go idle. 5.5 literally does not lose track of anything

It could get completely dog piled with messages and still stay focused

[high girder]

That's a problem I would love not to have as often. We'll see how it goes. My longest Claude session was about 10 days, and I had some errors, but not a ton. I'll give 5.5 the same time frame

[icy hinge]

Did codex usage just got significantly trimmed?

[cedar skiff]

what task takes 10 days?

[boreal holly]

probably many tasks, one agent!

[high girder]

it wasn't a specific task, it was my agent

[cedar skiff]

Well yeah, but for example i can say write these integration tests which is the goal, then there is like 120 task. or what ever

But you get what i am asking right?

[nocturne folio]

oh my gooodddddd 1500$!!!! whereeee

[high girder]

I get you. So the parent session I used for the agent itself, even when not actually doing anything, has it going. It's linked to my entire stack. So it depends on what I'm working on. I have local image gen and modeling workflows, physics work, and a bunch of other projects in between. The parent session itself ran for 10 days. Sub agents take care of individual tasks. The main thing the parent session handles is governance and continuity

The main reason I left it for 10 days is I forgot to turn it off and let it run on one of the other machines

it made for a good test though

[cedar skiff]

No i understand how orchestration works, i just want to know what you did for the 10 day session.

[deep totem]

codex 0.128 has new feature /goal, but I don't see it in cli, how to use it?

[cedar skiff]

with subagents i can see 10 days being 100kloc

[boreal holly]

I was gonna say I spend at least 2 weeks at a time on external integration work. Just now with 5.5 its one agent. Been doing QBO sync for 2 weeks straight. Rigorously testing it

[high girder]

Oh, literally bits of everything. I'm what you would call neurodivergent, so I do lots of things either at once, or some days nothing. I've been working on building out. Also, I don't actually use either claude or codex for code work or scanning PRs or any of that, unless I'm scanning something I'm going to put out. I'm also pretty toasted right now, if I'm being perfectly honest, so hopefully I'm making enough sense

[cedar skiff]

Oh, i miss took what you were saying as an automated orchestration session that went for 10 days.

I was like, if this is real i gotta get in on that

[high girder]

Ohhh. I'm not against automation. I just only let things automate once it's been manually done by me, and the acceptance output is continually up to standard. even then, evertyhing is logged and inspected.

[oak trellis]

/goal make 10k within 10 days

[torpid trout]

Enable it via toml

[deep totem]

got it, thanks

[oak trellis]

just a reminder: please openai .. do not change anything within 2 month .. if you wanna do the model better that is ok .. but do not downgrade the limits even if you cant afford the compute contract. Just make a new oracle deal .. send the stuff to mossad.. everyone does it anyway.

[oak trellis]

if they keep it stable for 2 month i buy openai stocks

sentimental right now i would dump anthropic and buy openai stocks

[cedar skiff]

i didnt notice any degradation since i joined up in 5.3

Rcok solid the whole time

[oak trellis]

wow

i have paranoia .. hanged to much around cc and opus .. they did us the rollocoaster .

[cedar skiff]

It's just better i wish i had have moved to the light side sooner

[pearl zinc]

Why am I constantly getting? :

This chat was flagged for possible cybersecurity risk
If this seems wrong, try rephrasing your request or submit /feedback. To get authorized for security work, join the Trusted Access for Cyber program.

Frustrated.

[cedar skiff]

Have you verified?

I know you probably don't need to.

But it makes life easier

[cedar skiff]

I thought you got banned

[pearl zinc]

HAHA no thanks. Back to opus...

[cedar skiff]

slopus

[undone patio]

cant ban me