codex-discussions

[undone patio]

too important

[sand bolt]

deleted 45 minutes of thinking progress cuz of this falase flag

[undone patio]

gg no re

[high girder]

join program, no flags.

[regal kraken]

man codex is in rough shape still on windows

just freezing

[frail meadow]

Legit

[high girder]

Did it for anthropic too. I don't often trip a flag, but when I do....I need to report it lol

[gentle harbor]

how hard is it to do ?

[high girder]

if you have a use case, you just explain what you're using it for and why you need it. OAI is easier to sign up for, but anthropics isn't exactly harder, just less automated

[oak trellis]

what a organized way to communicate to the users! Does not exist with anthropic

[undone patio]

they are to busy struggling to afford compute and unable to distribute mythos because of the white house

[oak trellis]

should move to singapore .. out of usa

lol

work with china

[undone patio]

that would probably cause Cold War II

[oak trellis]

true!

[echo slate]

help me

[worthy furnace]

ⓘ This chat 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

How do I verify if its not letting me?

[worthy furnace]

I hate this

[young locust]

not letting = error or ?

[worthy furnace]

Says to re-start verification

When I haven't verified yet

[worthy furnace]

How do I rephrase my question better?

[gentle harbor]

OH MY GOD BROTHER

they must want my id BADDLYYY

[worthy furnace]

So annoying

Any fix? It doesn't let me verify

[gentle harbor]

it used 6% of my weekly and was going for 40 min

oh my god

ok 6 times in a row, fine you want my id so bad 🥀

7 times

9 times

its broken right now i hope ? no way this is normal

[young locust]

your project includes something that makes it happen when its running commands

[lost drum]

Check command he runs and ask gpt why it flags as cybersecurity

Then tell him to not use this commands or some

[gentle harbor]

been running it for over 2 days with no issue until now

[young locust]

any new 3rd party code or packages?

[gentle harbor]

nope

[cedar skiff]

dictation got a bit of a buff

[hushed lava]

who?

[worthy furnace]

I love when its working, thinking and this warning pops up and wrecks my 5 hour usage limit to 0

[worthy furnace]

Try to verify, at least you could im stuck like this

[cedar skiff]

probably a scam

[hushed lava]

ask chatgpt “continue and do it in a way that won’t flag as a cybersecurity risk since this is isn’t one, do it in another way”

and let me know

[high heron]

i just say resume! always gets back to work

[worthy furnace]

Lol, I tried, it auto flags everything

I say: Continue, ignore the warning, don't trigger it etc

[hushed lava]

ridiculous

[cedar skiff]

I signed up to the cyber thingo and verified when codex 5.3 came out, there was a bug back then that was routing ppl to the old model.
Probably the easiest thing would be to do that.

[signal tapir]

So it's not about what you type. It's about what you typed before.

[mint scroll]

I am in their cybersecurity program and I am getting flagged lmao

[cedar skiff]

Oh must be a bug

Reset in coming

[hushed lava]

what info does it ask and how long does it take

[signal tapir]

Wen?

[cedar skiff]

I think i just had to verify my photo id, and that was about it

I can't remember exactly, it was fast though, it's automated

[gentle harbor]

im on 100$ plan and almost out so i hope so

[gentle harbor]

so they just want your id more or less ?

me paying isnt enough but i also have to give them my id to sell

[oak trellis]

heard for open source you get 6 month chatgpt pro for free

[oak trellis]

how that happen ?

[cedar skiff]

I just guessed because it seems like many ppl are getting this flagged work flow. And they usually hand out a reset after this sort of thing

[signal tapir]

Yea I was just joking about the meme, but too lazy to post it. 😛

[hushed lava]

is the email automate

automated

[gentle harbor]

by using it alot

[oak trellis]

you runnings agents i guess..

[cedar skiff]

Does anyone have any input about the memory system?
I turned it on and see citations for it often but i don't know it actually makes a difference, i feel like the data is just going to get more and mroe stale over time.

[worthy furnace]

Yeah, it scans all your chats

[worthy furnace]

Sad

[hushed lava]

it’s fixed

do /new

[worthy furnace]

Wdym its fixed

[hushed lava]

do /new or compact like the guy above me said

then mention the context and say continue exactly where you left

[oak trellis]

gpt 5.5 is my new friend

frend

[grim tiger]

hello, does anyone know how do i get my skills indexed on skills.sh?

[sage barn]

hello, is 5hr limit gone?

[shell coral]

is there a documentation about how goal feature works?

[oak trellis]

no

[inland sonnet]

u wish

[oak trellis]

feel like fanboy .. those who i make fun off

[sage barn]

-13% weekly usage in a request that took 1m 52s

[inland sonnet]

Relay on /status

[inland sonnet]

Depends on tokens?

[steady vigil]

│ Weekly limit: [███░░░░░░░░░░░░░░░░░] 15% left (resets 06:23 on 5 May) │

This is Pro (!) something has changed. I've never been able to use anything close to this

[sage barn]

no chance in hell 😂

[inland sonnet]

Damn usage lowerd?

Claude 2 incident???

[fleet geyser]

hopefully there's a codex update so the weekly usage resets 😂

[steady vigil]

just had one, no reset

[fleet geyser]

oh thats sad

this week ive used codex way more than i usually do, and now i have 4 days at 15% limits

[low cosmos]

guys if the weekly quota runs out but 5 hour doesnt can i still use it until 5 hour does?

[inland sonnet]

no

[low cosmos]

i hvae 0% on weekly rn and i can still do stuff

[inland sonnet]

odd make sure u arent paying 😂

[low cosmos]

i dont have anything linked lol

[inland sonnet]

then ur good i guess

[low cosmos]

lol how is this possible

[inland sonnet]

may they let you use the 5H limit even if weekly is out

idk

[low cosmos]

possibly

[inland sonnet]

prob one lasttime

just to be nice

[plush harbor]

I refactored my entire render pipeline and did a heap of other stuff recently and it hasn't made much of a dent. Shudder to think how fast that would go on claude

[inland sonnet]

Decent you are on pro?

[plush harbor]

plus

[inland sonnet]

Model and thinking level?

[plush harbor]

5.4 and whatever it defaults to, how do I even check

[inland sonnet]

meduim then

If u had used 5.5 it would have gone faster

xD

[plush harbor]

its made a few mistakes today. Mostly from not paying attention, like I wrote a damn spec and all and had to point it back at that more than once

all placement stuff. Explicitly told it to put new content under title, or under the entire thing and it kept putitng things at the end of the body content and breaking the layout. Grr. So a few re-dos

[inland sonnet]

rip

[cedar skiff]

Yeah, thats what it was for me, they use a third party id verification.

[plush harbor]

you know its been a good day when you successfully had a nice conversation with codex about cabbages and swimming and have nice working code to prove it

should have done a semantic cluster of codex banned words instead. Racoon, goblin, pigeon ....

[tiny fulcrum]

anyone else having problems when Codex App has to deal with images on Windows?
It constantly freezes and slows down

[lost drum]

Bro the 5h is just a limit that prevemts You from spendig Your whole weekly usage. It wont let u use that dsily

[oak trellis]

not using it because IDE blocks the port

2 days full coding only spend 20% ..

crazy

some cases i had 2 agents doing different tasks.. heavy content .. because designing backend logic..

in cc that would be already 90% of weekly limit use

[cedar skiff]

You can get tonnes of milage but getting 5.5 to orchestrate codex 5.3

[tiny fulcrum]

why 5.3? 5.5 low should do a better job

[jovial finch]

Hey guys can I get some advice please, I moved from claude because usage limits were bad, I got Plus plan (tho might upgrade go business with my friend) since we need it for work (well I do mainly)

I design backend/frontend websites and sometimes I need the ai to be creative and generate a non ai-template looking site as well as fix some logical issues when I come accross them.
In particular, lime making the login page more "reactive" and modern, or some visual hooks like moving stuff, etc... so enhacning the design basicaly.

What model is best for this and what thinking level, and the /skills are all activated by default, no? or I have to instruct in AGENTS.md to use them? How does it work?

[cedar skiff]

5.5 low is not as smart o.0

[torpid trout]

Stitch.
Gpt not particularly good at fe.
But if, then 5.5

[cedar skiff]

In fact 5.5 low is not as smart as 5.4 low afaik

[tiny fulcrum]

based on what?

[cedar skiff]

Some bench marks i saw, let me go see if i can find it.

[tiny fulcrum]

I mean I use this https://artificialanalysis.ai/evaluations/artificial-analysis-intelligence-index

and cost efficiency wise low > 5.3

[torpid trout]

Eric, what’s that banana doing to that onion.

[tiny fulcrum]

their "Score" for GPT models

basically 5.5 low should be beating 5.3 medium

[cedar skiff]

The bench was talking about 5.5 token efficiency being the reason low was a worse performer iirc.
It's easy for me to try low i have a work flow right here that i can try it with.
I'll try it and see.
Its pretty definitive because it works with 5.3 and if it doesnt work with low on the same task then...

[lost drum]

Wen 100

[cedar skiff]

real world work doesnt always line up with the bench marks

[tiny fulcrum]

I don't get it ^^

this is the cost breakdown, 5.5 low >>> 5.3 efficiency wise easily and should be quality wise too unless it's 5.3 xhigh

[cedar skiff]

right but again, real world doesnt always line up with the bench marks o.0.

Ill just try it and see.

I get both of them to write the same integration tests with the same pormpt

[whole estuary]

How to get access to 5.4-cyber? I am

[tiny fulcrum]

sure, personally 5.5 medium is sweet spot for me at work, Low sometimes but it just isn't as thorough

I use a pro plan so I run high all the time at home anyway

[sick axle]

So what's the deal with context window size? Is 256k the max for 5.5? Can 5.4 still do 1 million?

[cedar skiff]

im on pro plan as well. I can orchestrate 1 agent all night long with 5.3 while i sleep. But i cant do that with 5.5 med or high ill run out of usage. If low works better ill use that but it remains to be seen.

So 5.5 high does the orchestration and 5.3 does the work.

[lost drum]

I can do it on 5.5 xhigh

Depends what project you working on

[cedar skiff]

you will run out by the end of the week

[lost drum]

I done it last week and left 3% at the end

[cedar skiff]

you cant run continuous 5.5 on xhigh for 8 hours a night.

[lost drum]

• use it thtuhut a day

Pro plan is just the best fit if You want him to work foe whole week without "I need to check usage"

[cedar skiff]

I can burn 15% a day just working with 5.5 on high during waking hours

i have pro plan

[tiny fulcrum]

20x? 15% a day, seems hard if it is not fully scripted and fast mode

[sick axle]

How do you run anything that continues for 8 hours?

[lost drum]

Nah u just need HUGEEE goals and project to work with

[cedar skiff]

orchestration, a list of task that get delegated to subagents in a sequential manner.

5.4 would always stop during the night, but 5.5 keeps it going

[shell coral]

I think the more important thing is that whether 5.5 complete more task or not, there's a chance that because 5.5 is more intellligence and token efficient but with higher price/token it ends up completing more task and faster too on that same 5h window eventhough the limit is going to 0% faster

[lost drum]

For ex I made in 1 week of pro 20x whole ecom automated now I try to optimize it to make it work for any brand any service even launching yourself a brand

Crazy how much you can achiev this days just vibing with it

[cedar skiff]

it is more token efficient but the maths works out that it isnt cheaper because of it. it mitigates it a little bit. So it doesnt cost the extra 1 to 1 amount more, but lets not forget this model is more expensive that opus 4.7 per token.

[lost drum]

Had plus plan but 5h usage was spended in like max 4 messages so I switched to pro 20x after I spend like 3k credits

[nocturne folio]

my chatgpt plus plan is 2.85$ more than urs

[cedar skiff]

https://www.youtube.com/watch?v=c5QTCX4HlfQ this video has an explanation of pricing - i dont know how accurate it is, but i did watch it and it seems about right from what i feel when i work

[sick axle]

Does codex do this on its own or with a bit of prompt engineering or a third party tool?

[cedar skiff]

i write the orchestration just by yapping into the mic, i do spend a bit of time preparing skills to support the work that is needed, but that pays off immediately. This is always repetitive work, like audits -> fixes, or writing unit tests/integration tests. Things that can be trusted after skills have been refined.

So i'm not implementing a feature from scratch or something like that.

[tiny fulcrum]

define an orchestration skill, codex app can help do it

[cedar skiff]

ok i have 5.5 low working in a subagent implementing an integration test that codex 5.3 already successfully completed. It renamed the original, if it completes ill get 5.5 high to audit them against the requirements.

to be fair, i did have the skills refined by running loops with codex 5.5 high. So i could probably do that and get 5.5 low working fine.

[sick axle]

I feel like they didn't really answer my question

Is just giving skills what allows codex to automatically do orchestration and trigger sub agents

[cedar skiff]

i dont use anything third party

the skills are for performing the tasks eg: intergration tests the way i want them.

[tiny fulcrum]

talk to the codex app, let it explain how orchestration works and how skills work, then go on and create a skill for your specific orchestration workflow

[cedar skiff]

not running orchestration, you can just tell 5.5 to run orchestration.

[sick axle]

So the answer is it's built in, ok

[cedar skiff]

No, there is nothing built in to do it, you have to make it happen your self, there are third party skills etc to help. But its not a built in feature.

I just talk to codex and tell it what i want and to write a prompt for it

then i use the prompt in a new session.

takes about 10 minutes - to get the prompt i want

can take several hours to get the skill working the way i want

[shell coral]

looks like it's using GeneBench benchmark as assumption, if you are working on scientific data analysis then that video explanation should be relevant, but if you are on different kind of works it might differ

[sick axle]

A skill is just text that codex reads so I see that as built in? Third party would be like installing autogen or mempalace.

But I don't really speak to other prompt engineers so don't know the terms or cadence

[zealous pawn]

What is the maximum number of subagents I can run on the Codex?

[signal tapir]

A skill can include programs it can execute in its sandbox environment.

[cedar skiff]

I took built in as referring to feature that codex has built in, like some automations or skills it comes with. Your definition of built in seems like everything.
Third party to me is something someone else made that you use, like the other dev is the third party.

[cedar skiff]

this was the first run....

The gpt-5.5 low-reasoning worker saw the tracked deletion, restored the same test file, formatted/analyzed it, and ran it.

classic ahha

[torpid trout]

OpenAI should pay more attention to their cli docs

https://developers.openai.com/codex/cli/slash-commands seems to include only stable released commands... which makes it almost impossible to test out things without guessing

and well, stuff that is not there at all is still documented https://developers.openai.com/codex/cli/slash-commands#inspect-config-layers-with-debug-config
argh

[cedar skiff]

here is prompt i had 5.5 high use, i got this by telling another sessions what i wanted to do and telling it to write a prompt

[oak trellis]

5.5 medium is perfect for me

balanced ..

[lost drum]

They want to save usage on those who cant make their adjustments. Everything is done for a reason especially missing zibwabwe in cuda official drivers.

[cedar skiff]

it failed, but i could get it to work with a skill refinement loop so not really a fair test:

Comparison headline:


5.5-low added localization-based label lookup and extra “profile name prompt not shown” assertions.

5.5-low loosened one important style assertion: for the standard header path it passes expectWordOnly: false at line 81 (line 81), so the helper allows a premium icon if present at line 133 (line 133). The 5.3-high version asserted the constrained standard header was word-only/no-icon.

Android still did not reach the actual Connections interaction. It blocked at authenticated /home/women because the women-list subtree threw a production Fault before add_woman_button appeared.

Next opportunity i get ill give it a fair go, i have this working atm. So ill just leave it for now.

[tiny fulcrum]

sure thx, I'll experiment a bit with it myself

[ember venture]

Does anyone know what features should work in the codex app when using the api.

[vocal badger]

Hey there! I'm facing a issue in codex. does anyone know how to fix this? : So Browser Use seems to initialize correctly, but it cannot open any webpage. Basic tab creation works, but when I try to navigate to even https://example.com, it fails with:

Codex auth token is unavailable

So the issue seems to be that Browser Use is missing the required Codex authentication token for page navigation, even though the plugin itself loads.

[hard drum]

They have done lobotomising 5.5 again

5.5 has become a bit lazy for me

&& they seem to really be doing lots of fake implementations for some effing reason

[turbid axle]

/goal is amazing

[strong lily]

guys how to earn money with codex? As a teenager

i started using codex since february and mostly volunteered projects now i want to earn money to buy higher limits for ai

[vocal badger]

Maybe video editing with codex? Codex have a plugin called remotion and you could use it and maybe canva and like make a tool for video editing. if you somehow then calibrate the software carefully you might be able to like get quality results. you could maybe then upload youtube videos, video editing services maybe. but building it is kinda hard tho. remotion is mostly for motion graphics but if you can somehow calibrate it, it would be really cool when the edited video has motion grpahics.

[strong lily]

i have knowledge in web development so i can understand architecture and problems

but mostly i just give the problems to codex

to fix

[vocal badger]

hmm, you might be able to do it i guess

[strong lily]

right now i am building startup to participate in different competitions. It seems like english learning app on expo

[vocal badger]

but calibarating it is the hardest part 😅 u might spend your entire weeks quota for that

[radiant pecan]

what version you on, I updated but cant seem to find /goal

[strong lily]

so main difference is that users will use vr experience while practicing speaking

in phone

[turbid axle]

0.128 cli

[strong lily]

they can make form for vr glass by using hard paper to put their phone into

[hard drum]

So it took less than a week for gpt-5.5 from being pretty good to being pretty effing BAD

holy crap

[zealous pawn]

i dont have it bro

[radiant pecan]

same I dont understand why im on 0.128

[turbid axle]

only in cli rn

[radiant pecan]

https://klipy.com/gifs/who-is-bro-1--k01KQHVKNYJDKNSJ237NWVSVW3X

[turbid axle]

try updating again?

[livid relic]

its an experimental feature, you need to enable it in your config.toml

[turbid axle]

oh wait, you have to enable it first

[zealous pawn]

how

[turbid axle]

[features]
goals = true

also, bye bye tokens

[zealous pawn]

What is its main effect?

[radiant pecan]

If I understood correctly it is basically ralph loop now built in codex

[turbid axle]

it will keep going forever until it thinks the goal is achieved. amazing for large refactors for example.

im having it split a gui lib which was very much intertwined with my application into a standalone gui lib right now so I can reuse it. so far it looks good

[strong lily]

do you know other features that codex can do?

i really impressed by news that codex can edit video

[zealous pawn]

I thought it was a built-in feature of the Codex, and didn't need to be enabled.

[turbid axle]

yeh, its oai's special version of ralph loop is how they market it

[zealous pawn]

Sometimes it runs even when I run out of tokens, even though I'm not using Goal.

[radiant pecan]

I think this feature is huge ,now it hopefully will debug without stopping every 30 min

[turbid axle]

if you run out of tokens it will complete its current job still oai stated yes, this is pretty new I think though. it was not doing this a couple weeks ago

[zealous pawn]

hmm

[turbid axle]

it will yes. my refactor is running for 3.5 hours now

[zealous pawn]

how

[turbid axle]

also it neatly commits and pushes every task it completes neatly, then just keeps moving towards the goal. honestly, this is nuts

now all I need is infinite tokens

[signal tapir]

I'll also have some of that, please.

[zealous pawn]

good trick bro

[radiant pecan]

my only concern now is context rot, after one task codex cant even fix an easy bug, especially if you have a large codebase and interlinked libs

[turbid axle]

test code and testing tools are key

I had codex built an entire computer operator tool for app testing so I can do it all by itself

[worthy furnace]

What are the best experimental features?

Any hidden ones not in the list?

[turbid axle]

if you can make it deterministic, the ai can solve it

[worthy furnace]

Aside from the new /goal I guess

[worthy furnace]

Can the ai invoke /goal itself or you have to manually do it

[turbid axle]

you do it.

/goal improve the app until you deem it amazing incredible

[unique spade]

just clone the codex repo and start a thread inside it. and ask all these questions to codex. you'll get all the answers

[zealous pawn]

Does it run until it's finished?

[turbid axle]

/goal find a definitive answer to why we exist

[turbid axle]

until it thinks it is finished yes, whatever that means

[worthy furnace]

Finally, no stopping until the work is actually done

[signal tapir]

or the credits run out

[worthy furnace]

Most of the time that'll happen

[turbid axle]

so far it ate 10% of my weekly at 3.5 hours of work. but a lot of that time is it running tests too

nom nom nom

but hey. drop 50k in an account. give it a well specified goal. and you might just have a wacky magic app making machine

[zealous pawn]

Did you buy the Pro 5x or the Pro 20x?

[zealous pawn]

i think it can

Just tell it to add /goal before it.

[turbid axle]

its just 10x or 20x for codes stuff actually. I have 10x rn

[zealous pawn]

Is it possible for the main agent to create a subagent and then add /goal?

[turbid axle]

try it

/goal implement issue 354, making sure you conform in full to the rules as defined in devrules.md

context imo, aka AGENTS.md, never was the right place for instructions. it should be a /tmp working memory buffer only

[unique spade]

a lot is possible, not sure if what you re asking is possible yet in current vanilla implementation though

[unborn thunder]

Does the app have gpt 5.5 now I only see 5.4

[floral garnet]

I suppose that's what we get because 5.4 and 5.5 just drains your usage x.x

[turbid axle]

/goal cure my empty wallet

[potent mason]

Do we get /goal in the codex app or only CLI?

[torpid trout]

I told /goal that it has CWD at disposal and is free to create - the goal being, "generate 100USD daily as passive income, with minimal HITL requirements"
It is going since 56 minutes, not sure what the hell it is cooking up.

[signal tapir]

There's no way that can go wrong. 😛

[torpid trout]

Yeah right

[torpid trout]

Goal uses substantially more tokens than normal flow.

That, or today I’m on the paying end of the load balancer

[honest creek]

is this how much it usually consumes? (on low) its a lot, maybe medium or high is better? (more efficient or something)
(plus plan since 2 days ago around, not overly using it nonstop)

[rocky fog]

5.5 very high takes a lot

medium is a bit more like 5.4 very high on usage I would say
probably still a bit more

[cyan wing]

Just ran /goal in a loop overnight, ran for 8 hours
will report findings shortly 🫡

[honest creek]

I was using low the whole time for 5.5 I suppose, addon/plugin modifications with junior experience to say it bluntly

[cyan wing]

set it to 5.5 high, it used ~15% of my weekly Pro sub

[rocky fog]

yeah I am also rather settling on high than medium, seems a bit better

[honest creek]

you mean for mine or?

[wraith spoke]

should i use codex

[rocky fog]

no, I just mean 5.5 medium compred to high

[honest creek]

honestly idk how to manage it with not much coding exp on my side.

[rocky fog]

you just have to try and see what works best for you
if you see it making issues or just not good enough overall, try to go higher and see if thats better or not

also depends on what you are working on
so it might be different for everyone

[honest creek]

game mods, personal proff websites etc, to sum it up.

[rocky fog]

its more about complexity of that project

[turbid axle]

5hrs and counting, buddy still going

I have to leave soon, guess I'll tell it to take a break?

[honest creek]

well, for websites its mostly visual stuff, where to place what, keeping it mobile friendly, considering how everything interacts with everything else etc, last part is about the modding.

and now I had it help me make an wow addon for the UI.

that 50% of weekly got consumed by website adjustments and by the addon.

on low*

[potent mason]

We need more models in collaboration with Cerebras, please OAI

[worthy furnace]

For what?

[worthy furnace]

It will solve all your problems 😂

[cyan wing]

To improve performance, frame stability of my
ThreeJS (V8) <-> Rust native game engine runtime

[torpid trout]

It is doing an awesome work, but, the idea it had is totally (disallowed word because server rules are silly).

A practical kit with email templates, a call script, a follow-up tracker, and a 7-day workflow for turning awkward client silence into clear next steps.

[still fiber]

Did the update this week move Skills and Personalization out of the Codex windows client? Those options under the settings bar disappeared across all my environments, but persist in the web version

[torpid trout]

🤷‍♀️ as if someone would buy such things even for 1 usd, and it wants 39 for it lol

[wraith spoke]

i wish codex was more accessible

[torpid trout]

But, I have to say, it created a nice website, a full blown actual product, sales tracker and so on
Well, it is still going, but I will have to shoot it down soon, tokens are vanishing

[torpid trout]

You mean cheaper?

you cannot be serious lol

[wraith spoke]

no, i mean to people with no technical knowledge

[torpid trout]

ah. Well, you have your accessibility right there: ask it

Like, literally. Tell it you do not know, and that it has to hold your hand.

It will, happily, even tell you what buttons to click and so on.

[wraith spoke]

oh sweet

[torpid trout]

Its GAI my man

it can do about anything as long as it is not phisical action. Answering questions, guiding, explaining, yada yada,.. it excels at that.

I bet you could even ask it to create a video for you - but of course, that would be a tad expensive hehe

[rocky fog]

basically nowadays you cant say that you "dont know how" something
you can always just ask LLM and it will give you step by step for how to do/use anything 🤣

[torpid trout]

This was largely true with the advent of the internet and google in particular... just that now it is crazy fast and more concise too, so the excuse window narrowed massively

It does of course endanger human interaction, which is why we should also never reject to answer particularly precise questions 😄

[boreal holly]

Codex in VR now

Finally can play Beat Saber and develop software at the same time

[rocky fog]

yeah its funny, nowadays you can develop while gaming 😆

or also just remote to my pc from phone and do it anywhere (or cloud, but I dont github and its more usage)

[boreal holly]

Codex makes anything possible and I think that's cool

[hard drum]

I don't know what i'm doing wrong anymore because my own stuff keeps breaking && 5.5 stopped listening

[torpid trout]

but why

I saw that recently with another app, I think it was panic's terminal app. Could not really understand why someone would want that.

Is it like more easy to control the thing with that virtual stuff?

[boreal holly]

The truth is, Meta Quest is the only android device I have so I was setting it up for android app testing. What’s cool though is I can test the app and tell the agents about issues without like transferring screenshots and taking the headset off or whatever

So there is a tiny bit of reason to run codex in there, but definitely prefer desktop

[worthy furnace]

I forgot, it's still going!

[worthy furnace]

That's so cool

[undone patio]

jarvis, make no mistakes

[astral nest]

just realized i can see usage anytime in the codex app menu bar

ty codex dev team ❤️

[torpid trout]

So the codex thing sees what you see whole you have the goggle on?

My sole and only custom instruction in codex, that's it. nothing else, no skills, no plugins, no agents.
I might add one more about subagents but for now... thats it
Condensed and efficient

[ornate ledge]

I keep seeing this error while I try to give a /goal

■ Failed to read thread goal: thread/goal/get failed in TUI

[velvet surge]

So did I, it happened when pasting a large prompt.

[torpid trout]

can you put goal in a file, and prompt it that goal is to implement that file?

That worked for me with a huge prompt (goal)

[ornate ledge]

I see this even when I give a 2 line prompt or when I ask it to implement a md

[torpid trout]

Weird, it works on my end.

[velvet surge]

How are people enjoying this new goal feature though.

[torpid trout]

well its not bad, is it?
it is basically what it should do by defauilt, so the only I hve to complain about it is... it should not be necssary

[velvet surge]

Yeah, however I think that there is a significantly excessive burn on tokens it's insane.

[radiant pecan]

how do I set a goal in the cli?

[velvet surge]

/goal

[torpid trout]

yes, it is buttery

[torpid trout]

codex features enable goals
Then, use /goal in cli prompt thingy

[velvet surge]

That was from one goal for a simple addition took 11min

[torpid trout]

makes sense
Here it worked about 2hrs straight on my millionarie status, and consumed about 5%
(of weekly)

[torpid trout]

When in tui, type /goal {your goal here} and then hit enter
Ensure you have goals enabled, as I instructed above.

If that does not work, it has to be OS or else, because it works here and on other instances (making it almost impossible to be a broader codex bug, but perhaps some compatibility issue)

[vale patio]

yo i have a question am i the only one not being able to connect his phone number when trying to use codex app ?

[radiant pecan]

is there a char limit for /goal?

I want to use a long prompt but it wont let me

[torpid trout]

BTW, I just had to, this was a calling, forced and unavoidable, i could feel it in the air, my destiny is and stays....

[torpid trout]

I have added it a goal that had 1200+ lines in an MD file
So, no.

[honest creek]

cost-wise (Tokens) is 5.4xhigh practically half the pricing of 5.5 xhigh for about a few % in coding intel gains? i.e. right now 5.4x high is more worth it?

[exotic cave]

So much for mythos scare mongering... GPT 5.5 has better cyber capabilities. Anthropic are full of themselves.
Source: https://www.aisi.gov.uk/blog/our-evaluation-of-openais-gpt-5-5-cyber-capabilities

[torpid trout]

No, 5.5 is more worth it

5.4 is an absolute trash anyway
You either go 5.3, or 5.5
Given the intelligence of 5.5 you will consume a fraction of the tokens - because it is more efficient and more intelligent - compared to any other model.

No need to go x-high on 5.5
Just medium or high is way far enough.

[torpid trout]

Different marketing strategies

OAI is like apple for the masses, Antrhopic is like blackberry but only if we like you

[exotic cave]

This is not the case, most of the tokens are spent on Input, output is not what eats token for me. Couple it with small context window and multiple compactions and Input becomes the issue and what eats tokens like no tomorrow. Still 5.5 is superb and i use it as much as i can!

[torpid trout]

I am not sure I follow.

Of course that is the case, 5.5 uses about 1/2 of the tokens for the same task as its predecessor, it is literally measurable
I am not stating where that shaving happens - just that it happens.

[boreal holly]

Yeah depends on the ratio of tool call input and resulting output. also if you have unified exec enabled, all the polling it does

[honest creek]

what is the sweet-spot for most mid-level coding tasks, file code checking etc?

[exotic cave]

Question is, do you eat through tokens and your plan is not enough until next reset? If yes then use 5.4/5.3, if not keep rocking 5.5 because it's really really good.

[honest creek]

3 days on plus on 5.5 low standard if that answers the question.

to be honest I just started giving the context to claude and it makes a brief to give to 5.5 so maybe that will improve things.

[exotic cave]

I am not familiar with anything lower than high/medium.

[exotic cave]

unified exec? care to elaborate?

[torpid trout]

Simply use 5.5 on standard

Default.

Whatever it’s called

[honest creek]

its low on standard yes.

[torpid trout]

Cheap code is expensive
It’s very simple
The single codex token on 5.5 might be expensive - but it has quality

[torpid trout]

Is it thou? If so, change it medium.
Low for coding seems a bit underdone

Im running all the things on medium or high
As so far… exponentially less waste and massively better outcome

Mind you, I don’t even run any skills or such

Zero. Nada. Just a tiny agents file

[honest creek]

do you think low wastes tokens?

[torpid trout]

All these addons are polluting the context. Thousands of nice looking sentences - when it’s all already baked into it…

[torpid trout]

Can’t say, I never used it, if anything it probably just doesn’t reason enough?

KISS

Don’t overthink it

Just use and be done.
Turning too many knobs has sent so many folks down a rabbit hole they never came back from

Did she now.
If I’m not wrong … are we entering phisicyan times again.
That website is for sale

So I call bull

I think the physician is you

[honest creek]

okay I hear ya..

[honest creek]

thats what I am trying to avoid.

[torpid trout]

She never scammed me or other serious users here afaik.
The only source saying so is you and a newbie who was in here ystrd .
Neither has shown proof to me, and there’s about 50 Jane’s in here

Now, if it was, then so be it. She still didn’t show it to me.
And the url you just shared was fake. So, that’s where I stand and where I’ll stand until proven otherwise

Also, this discussion is over.

Did you now.

I’ve factual proof about a certain inverted situation buddy

So stop. Now.

[slender plinth]

SO CUTE

[maiden garnet]

BonziBuddy is about to be reborn.

[honest creek]

phew that was fast wasnt it.

[inland sonnet]

nobody will beat bonzibuddy

[maiden garnet]

Haha

[inland sonnet]

GPT 5.5 will prevent to destroy your pc so it cant mimic real bonzi!

[maiden garnet]

True. haha

He tried, though. Poor Bonzi

You can have codex make your own, apparently, but the /hatch command doesn't work for me, just /pet

[torpid trout]

He was in here way too long, to be honest
I wish there was a rule in this server that this channel can only be joined after X amount of time being proven member of the server, with real interactions showing real intent.

Frekkin scammers, everywhere.

[dark sun]

does anyone have experience using claude and codex on the same repos? is there anything interesting about codex that I should know before i symlink my .claude directory and pray to god that i can press go and everything works

I'm about to toss $200 at OAI

[honest creek]

I got like 2 dms from just being here today for the first time from people who were here for a long time.

they will like rejoin with an alt btw.

[torpid trout]

Did they also scam you?
Because that former one was not on here since long

What I mean is, prove your intents in other channels first. For months, not days.

[honest creek]

I mean , I have a degree, basic stuff like that wouldnt do much, basic social engineering... cmon

I hate to code and I have AI do it, doesnt mean I dont know...

but then again, it was obvious, like.. "dm me to know how" obviously cmon...

then deleted that of course.

[torpid trout]

Of cours it is obvious, but, A) there is people who fall for it and more importantly B) do I really need to get verbally attacked by an a 🕳️ telling me OTHERS scammed, when he/she is the real scammer here?
That is undefensible, it makes me rage, and when I rage I produce memes, and that is not good either

Seriously thou, on the other hand, it is a welcome daily source of fun too lol

[honest creek]

anyway, I get help from the AI to make a free addon for world of warcraft since I found no other that covers my needs, find a problem, seek help by AI to get a solution that is sufficient, share.

[torpid trout]

BTW when another user comes saying someone was banned in future, an easy way to check if that is true is search the sidebar for that username,
If the messages are still here, it means they where not banned - just left
Liek Jane - their messages are still here, but you cannot mention them - because they left
More a note to myself than anything else but might be useful to spot fakers

[undone patio]

this guy was banned

[torpid trout]

Jane was not, the one saying so was. That is correct. And just.

Justice was served.

[undone patio]

no i meant you

Jane was found to be propagating the fake codex website however.

I have evidences.

[velvet surge]

I'll be honest messing around with that goal function some more and it really is what it should have always been you are very correct. Really powerful though from my testing.

[undone patio]

She was banned from anthropic discord for it

[torpid trout]

Really now. Then tell me why her messages are still here, while banned users arent here - not even the messages
And why NO ONE else ever reported something like that, unless... banned users?

[undone patio]

i didnt say she was banned here

i just said she is guilty of spreading the fake website

and trying to earn a commission from it

[torpid trout]

I believe that only once I see it

I am friends with Jane, like not in real life, but here in Discord, and she was in this channel for a long time, before all the phisicyans popped up

[undone patio]

yeah well

im not interested in convincing you, the screenshots are already in the hands of anthropic. if i notice it happening here i'll report her here as well

[torpid trout]

I dont know what anthropic is. A species of human?

[boreal holly]

You could've just said she uses OpenClaw and they would've banned her for that

[undone patio]

cant tell if trolling or just clueless

[undone patio]

so you think spreading fake codex websites is cool, i guess

[boreal holly]

No, I'm just saying Anthropic will ban people for any reason and no reason. Sorta like the "red scare" but AI edition

[torpid trout]

He most absolutely did not state that

[rocky fog]

when banning on discord, you have a choice to remove x hours/days of users messages

you should just check whether the user still has this shared discord with you in their profile

thats when the server/guild bans it
when discord itself does, that might be different

[boreal holly]

Here, you gotta be a real nut jub to get banned here

[torpid trout]

If you just leave the server, you also have no server in common anymore.

[rocky fog]

yeah true

[boreal holly]

Jane leaves and comes back. She's done it like twice now, literally to get the newcomer badge back lol

[torpid trout]

Considering the amount of useless stuff I shared here so far, I have to hard agree
I even clashed with mods at times.

What I did never do is sell crap, and neither did Jane, (or Cantcrit, for whats worth, who also seems gone)

Kind of sad because it added some character to the channel

[rocky fog]

thats how that default discord ban menu looks

but there are also many bots that handle that

not sure if they usually remove messages here, might depend I guess

[boreal holly]

Cancrit is grinding don't worry. Idk what Jane is up to

[torpid trout]

Interesting!
Yes the last ban here def removed their messages too.
Not sure how far back, tho

[lean lark]

Gonna change name of 'JaneBot' to something more generic. 🙁

[torpid trout]

Lol, stabbing it right in the feels this one does

[lean lark]

Meh, it was just intended to be an inside branding anyway. If that account is gone then it doesn't make sense to keep the branding.

[remote yew]

Stupid question. If I’m on codex, what’s the difference between 5.5 and 5.3 codex models? Like ik 5.5 is newer ofc but is 5.3 codex better for coding than 5.5? Assuming im using both at highest level what’s the difference regarding token usage, speed and general intelligence?

[torpid trout]

between 5.3 and 5.4, you can definitely say 5.3 is better for code (referring tot he -codex version)
between any and 5.5, I cannot see a difference
You do not need to use xhigh for 5.5, medium is enough.
It will be about the same as 5.3 at xhigh

(with a lower or equal token consumption allover)

[remote yew]

Thanks!

[shrewd flint]

they were shilling some kind of illegal service in people's dms

[shrewd flint]

just ignore it bro..

[lean lark]

Hey, does anyone use Codex here?

[low cosmos]

ive managed to burn thru my weekly plus quota in 2 days

[torpid trout]

Laughable

The person before who went all out badmouthing the member is now banned... because that person did exactly what they accused others of. LOL

[lean lark]

We should have a survey here, for when people announce that they've blown their quota:

• Which model do you use primarily
• Low, Medium, or High
• Standard or Fast
• Do you choose different models based on the task
• Do you use the same model for documentation as for code
• Do you ever use Plan mode
• Do you write code or just let the model do everything
• Do you look at generated code
• Are functions documented
• Are modules documented
• Do you have a system-level AGENTS.md
• Do you have a workspace level AGENTS.md
• Do you have any AGENTS.md files in sub-folders
• Are you using Skills
• Do you use different threads (sessions/conversations) for different tasks, or do you use one thread for many efforts

[undone patio]

i know

i have screenshots

[hard drum]

Nintendo of AI

[lean lark]

"Are you tired of off-topic discussions? Ask your doctor if Codex Focus is right for you."

[inland sonnet]

just prompt better

[shrewd flint]

wow

[lean lark]

(I was talking about this channel)

[shrewd flint]

why is that guy defending the scammers? is he one of them aswell?

[undone patio]

probably

[torpid trout]

2x blocked, since mods wont do it.

[digital thunder]

<:book_icon:1363314738255364126> Rule 4: No spamming.

-# Refrain from spamming or flooding through actions such as mass pings, all-caps, or repetitive posts, as they disrupt the server.

<:book_icon:1363314738255364126> Rule 3: Stay on topic.

-# Be mindful of what other users in a channel might find helpful or interesting when posting. Stay on topic in order to keep conversations focused and productive.

-# Consider posting in #off-topic or an appropriate channel.

[plush harbor]

I've been home sick for a week (normally I'm out most of the day doing stuff AI can't help with) and blown through pretty much every backlogged code task I've wanted to do on this website for, oh, years and not hit limits. I've run out of stuff I want to code now 🙁

I have no idea if I'm using this thing right. Whatever "right" is. I just tell it stuff like "ok, on a word edit page, for a cluster parent (id = related word id) add a button next to both word save buttons "rebuild cluster" that does a cache rebuild on all the words that share that cluster parent" and it tootles off and puts buttons on my cabbage clusters like a good bot

[orchid plume]

can't believe how much time I'm spending on customising the Pet in Codex 😂

[boreal holly]

"Codexetron" side effects may include severe productivity, sleepless nights, peripheral neuropathy, death, sneezing, and building an orchestration system

[turbid axle]

7hrs and counting, refactor still going smoothly

/goal goals

this is a game changer

[plush harbor]

hah I refactored most of my site 2 days ago. I think I have a smaller codebase, it didn't take that long

[turbid axle]

its a pretty big job, its splitting of pretty deeply integrated gui code into a standalone gui library from my app.

its redesigning all the widgets to stay functionally the same, but implemented in a generic way. im just impressed after impressed with this atm

[torpid trout]

Can you share more about how you do the /goal?

Like, do you use skills, agent files, additional Md files like plan, design, and so on?
Or just one concise short /goal {do your job}?

Im interested as I’ve a major OOP PHP codebase with about 5 years worth of tech debt I want to migrate to a container oriented code … and it needs major careful feature parity checks from old to new

SERIOSULY

GET OUT OF HERE.
I’m not even allowed to use the f word in here but we have to accept like more spammers a day than there are meaningful messages in here!?

[undone patio]

@winged ore i am available for interviews if you need a discord engineer (you do)

[torpid trout]

Yes, that’s your best candidate.
Then we will all be in a very safe place.

There’s no need to elevate people to roles they can abuse.
There’s a need for the existing mods to do their work when users are reported

Not say „well look“ and then nothing happens.
Just my thoughts.

[undone patio]

well im assuming the actual mods are just volunteers

[shrewd flint]

whats a discord engineer

just a normal software engineer?

[undone patio]

yeah, wherein the software is usually the bot serving as moderation/oversight/logging etc

and organizing the server itself

[torpid trout]

It’s very simple to block people who spam by having a proper onboarding process - discord has tons of features for that.
But seems they’re only interested in how many members they have, not what members they have

[cedar skiff]

subtle rage baiters will always slip through unless you have overly tight moderation.

It's that nature of the beast

[winged ore]

oh boy.. if it where that simple...

[high girder]

i revisited the desktop app again with 5.5. this is almost too easy now

[plush harbor]

when I grow up I want my own discord bot for my own channel. Fortunately its 2026 and there are ways to make that happen without actually doing any work

[frail meadow]

What’s too easy? Dev?

[frail meadow]

Aren’t there discord bot frameworks too though so you can still customize your bot however you want using code?

[plush harbor]

idk I've never written a discord bot before. I still want one

[frail meadow]

Have you googled it before?

[boreal holly]

Idk bro if they moderated harder we wouldnt have such dank meme fuel

[plush harbor]

naw, I just keep puttting it off because its low importance but I'm starting to run out of things to code, codex is speeding up my backlog destruction too fast

[boreal holly]

I take it the great refactoring is going smooth?

[plush harbor]

done for the most part, need more content for the second pass. That'll take me weeks

[frail meadow]

What content do you need?

[plush harbor]

about 3000 images

[high girder]

Planning, building, even reverse engineering is going smoother

13m and 38s from start to finish on a little web game based off of the chao adventure mini game for the VMU. Looks like a 1st pass, but not bad

[tired ermine]

anybody having issues with Chronicle eating up CPU and slowing down system? at times it completly freeezes up on me. I had to disable it its gotten so bad

[nocturne folio]

dude gpt image is so peak dude

its like almost there but i already love it

[ornate jolt]

what is pretooluse and posttooluse?

is that new?

[turbid crane]

The "Codex app" should not translate "Slash command". I can't use English to enter slash command now!

[torpid trout]

Pre-tool use and post-Tool use
Hooks for before abd after a tool is used

I don’t think codex has them, it’s a Claude thing afaik, but I could be wrong

Yeah I’m wrong- apparently codex has it too

https://github.com/openai/codex/issues/14754

[gentle harbor]

some times i feel like this isnt the real sama

[cyan wing]

Uhm.. what is 5.4 xhigh saying here 🤨

[unborn thunder]

Is heavy use mode just the same as chronic use?

😆

[unborn thunder]

His true colors are coming out

[tawny void]

does pets work on the PC codex ? Cant seem to get mine to work

[undone patio]

did u type /pet ?

[tawny void]

yeah just keep going in circle

[undone patio]

show me

[tawny void]

"I found the issue: Codex was still running, so it overwrote the state file after I wrote custom:doomguy."

/pet is not a recognized Codex chat command here.

I started the Doomguy fixer again. Fully quit Codex; it will relaunch with custom:doomguy selected and the pet overlay state set to open.

[undone patio]

did you update the app?

make sure you have the pet and hatch skills installed

[tawny void]

yep updated to current

just checked

[undone patio]

idk fam ur cooked

[tawny void]

installed

looks disabled

does it work on Plus accounts?

[undone patio]

idk man i wouldnt survive on a plus account

im on the 20x sub

[tawny void]

we supposed to have the option in settings to turn off and on ? i dont have it

Codex pets are optional animated companions for the app. In Settings, go to Appearance and choose Pets to select a built-in pet

[tawny void]

this is all i got

[undone patio]

scroll down to the bottom

[tawny void]

thats the bottom

dam burned a ton of tokens trying to get ot to work

[undone patio]

then u dont have it idk

[tawny void]

keeps looking me since it trys to set the overlay starte on and restart turns it off

[simple star]

Codex team has been giving kinda of a middle finger to Linux users recently... when the hell will we get an app?

[undone patio]

when you buy a macbook

[oak trellis]

His mentor said he is a serial liar

[gentle harbor]

his mentor ? who is that

[oak trellis]

Some guy from y combinator

Lemme check

[inland sonnet]

Your whoile live is a lie!

[gentle harbor]

much easier to make a app for macos and update it then both linux and windows, the windows app is also behind the mac

just use the cli

or buy a mac

[oak trellis]

Paul graham

[inland sonnet]

Mentor or dictator?

[oak trellis]

Oy vey https://www.instagram.com/reel/DWymH2qD_XM/?igsh=MWszZHVvMmsxMmJxaA==

New Yorker article

Is ChatGPT good in design now ?

[undone patio]

thats what i said and they emoted on me

literally if you work in tech and you dont own a mac, everyone secretly thinks you're weird

just letting you know

[balmy olive]

define design

[gentle harbor]

i got it to make this, id say its bad at making new ui but alright at copying stuff

[balmy olive]

well it's not a good judge of aesthetics and doesn't have taste, so it can't create the UI on it's own, that's why every vibe coder design looks the same, but if you're a designer and obviously have a taste and good judgment, you will create great things with codex, basically don't use for aesthetic pursposes, but use it as a fast executer

[simple star]

yeahhh.. no

[gentle harbor]

???

[simple star]

No way I am going to buy a new computer, and use a whole different OS that I do not even like, just to use Codex

[indigo marlin]

Anyone else seeing in this error when trying to access the codex computer-use?
Name" : "SkyComputerUseClient",
"exception" : {"codes":"0x0000000000000000, 0x0000000000000000","rawCodes":[0,0],"type":"EXC_CRASH","signal":"SIGKILL (Code Signature Invalid)"},

[gentle harbor]

well then use cli and dont complain lol

[simple star]

Use the CLI

I am

Dont complain

I will

Complaining is how people get the things they want

[gentle harbor]

linux and windows are harder to develop for then macos even if you complain it will never be updated as fast as macos

[simple star]

Harder to develop? Where does that come from?

Leaving that aside... what is up with that /goal command? It is not appearing in my CLI

Dont tell me it is a Codex-app only feature

[gentle harbor]

you have access to gpt as well it couldnt be easier to search up no offense

[simple star]

I have no idea what this sentence means

[inland sonnet]

you have access to gpt as well it couldnt be easier to search up no offense

[simple star]

Thanks for the repeat

[inland sonnet]

And if you want to complain use the right channels.

Stupid

[simple star]

Oh look. A chance to test the block command. Thank you!

Much better

It seems to work

[inland sonnet]

what a moro

n

[simple star]

Yeah, it works. Thanks

[hushed lava]

are any trials going on?

[balmy olive]

Testing moderation ? 🫠

[undone patio]

do yall like my pet

[plush harbor]

Truth

[simple star]

I cant my /goal to run more than 3 minutes 😄

[shell coral]

anyone tested whether /goal is auto continue the task after usage limit reset if it was failing because limit is reached ?

[turbid axle]

bruh

[plush harbor]

codex is a pain if you're used to doing front end. "put it in wiht the title" puts it in the utility bar "no put it in with the title" puts it in the body "no put it in the title area" *finally puts in the title area - would you like me to change the css too?" NO

[plucky halo]

Use the inbuilt browser and the commenting feature - will cut down this back and forth dramatically

[plush harbor]

what inbuilt browser

[plucky halo]

In the codex app, if you use it

[plush harbor]

I cant run the app

[plucky halo]

Linux, I'm guessing?

[plush harbor]

mac

[plucky halo]

Ah ok - intel-based then? I do wonder why it's gated to silicon macs since it's just an electron app

[plush harbor]

psht, all that for a tiny little bit of subtext

[plush harbor]

lazy devs

[plucky halo]

M is for Main.
See also: Yolo, let's push straight to main

[plush harbor]

I have to restrain myself whenever I look at G is for Goblin

still, refactoring an entire routing/rendering/caching pipeline and adding that little clustering feature would have taken me a lot longer without codex

[undone patio]

how force push looks at me after 5324773 merge conflicts

[cinder plume]

What is the codex pet format for custom pets?

[simple star]

We CLI guys are going to be delighted of all the pet talk

[undone patio]

enjoy your terminal, nerd

[simple star]

wasnt "nerd" supposed to be cool in 2026?

[undone patio]

always has been

[simple star]

In that case, thank you

[cinder plume]

lol okay, the sprite generator for pets is seriously mint

[turbid axle]

link?

[fickle rune]

anyway to make codex make better game graphics? like im making a game and its just not designing good. this is how it looks right now ive seen people online get much better results with codex.

[velvet wren]

I don't know the answer but I suspect using gpt-image-1.5 or gpt-image-2.0 would be a part of it

[livid relic]

ask it to use gpt image

$image-gen

[fickle rune]

What's that?

[livid relic]

its a skill that you should have my default

to tell codex how to use image gen

[cedar bear]

Is there a way to change the pet?

[zealous pawn]

I think the goal wasn't very good atfer limit quota

[worldly charm]

best custom codex pet

[worldly charm]

good qs

For now, going with Null Signal

[chrome raven]

same 😄

[vale patio]

did they fix the login phone number issue?

[chrome raven]

https://codex-pet-share.pages.dev

[boreal crown]

why do you have pets on codex windows and i dont?

did they blacklist me from feature roll outs cuz i've been using the pets feature for like half a month via flags

[zealous pawn]

how to have codex pet

[livid relic]

in your settings

you can also ask $hatch-pet to create you a new pet via image generation

[livid relic]

ask codex

[young locust]

update codex and then activate /pet

[simple star]

It's not been a week, and I am already sick of the pets thing

[boreal crown]

i got an update today but didn't get it

[young locust]

is your codex lang english? the command is localized

[boreal crown]

yuh

[unreal root]

I'm hosting LMStudio on my windows host, and trying to access the codex CLI on the WSL client. How am I able to access it?

The server is online

[dense depot]

After updating my Codex app is showing "GitHub CLI unavailable".

I have github cli installed and it was working before the update. Does anyone know how to get Codex to find the installed CLI?

[zealous pawn]

i dont have codex pet

[green wren]

launching the browser crash Codex Version 26.429.30905 (2345) on MacOS 26.4

but not in all cases

[unreal root]

Update; set up a reverse proxy and now i'm getting this LM studio is not found.

[boreal holly]

Don't use --oss, you wanna configure it as a regular model provider. I ran into something similar. I think the --oss flag was abandoned.

[boreal holly]

Make sure the model provider name is not exactly lmstudio. I went with lmstudio-local. If you name it plainly lmstudio it follows the --oss code path and doesn't work (it ignores whatever host:port you set in --oss mode)

[torpid trout]

@boreal holly - are you using codex hooks?
I mean the ones you can config via toml + hooks.json

looks interesting, honestly
You have start turn, end turn, user input, etc...

[boreal holly]

I am using them, simply as a reliable source of "is running" hooks. I wrote my own hooks on top of it for stuff like "startingCompaction" because 5.4 would after like 10 compactions start drifting, but 5.5 doesn't have that problem so I don't use most of the hooks anymore

Oh yeah I have one for when an agent spawns, it creates the worktree, rewrites their CWD to the worktree, and appends a thing in the initial prompt. And when an agent gets archived it reaps the container stack and if the worktree merged it cleans up that worktree/branch. Other than that not much

[wanton quartz]

https://github.com/adamdroberts/agent-skills

Two agent skills:

1. Adds code changes transparency and ensure that the changes match what you asked for, if any unplanned changes get added, it will state what they are.

2. Adds Wiki style docs, architecture, API, methods, LLMs.txt, and agent skills.

Install with:

codex plugin marketplace add adamdroberts/agent-skills

Open Codex then navigate to plugins and enable 🙂

[torpid trout]

OK... I am at the crossroads of three options:

• just wait until codex cli actually implements orchestration, because rn, its a joke still
• try and use hooks and again go back to using skills and agents, from which I came away, but... its annoying having only an agents.md because you have to be careful what rules you pass, as it is global
• build my freaking own UI, basically like you did... which means lots of work and I actually need to work (for money), not work (for fun) for once lol

[boreal holly]

I mean if you wanna try my orchestration thing you can! I get it dude. Back in August 2025 I was like "I'm gonna build my own thing". Went through so many iterations. First I tried "task-based" orchestration where workers get precisely one task, then "system-based" orchestration where there's a worker for each system of a given project, and finally landed "role-based" orchestration where you have task-based, but you have additional roles that source the tasks (QA, designer). Yesterday I rebuilt the rear axle on my truck, fixed a coolant leak, replaced the air filter, and the agents worked tirelessly on the project (that's why I was pretty much not here all day).

Could probably fork it, take the bits you like, discard the rest. Have spark set it up for ya

Anyways yeah it's a lot of work. It feels like you're gonna be distracted from real work. But it totally pays off

[torpid trout]

You’ve the entire thing on git meanwhile?
I’ve to check it out again, last time I looked is a while ago lol

I mean… I’ve had it buold crazy stuff in no time, but for some reason when it comes to buold stuff for itself codex seems to struggle

Even just deciding agent vs skill, it always messes up the concepts

I guess it’s a challenge of enhancing yourself lol

I’ve a feeling that their system instructions are part of the issue too

Tingling to fork codex and rip them out

[boreal holly]

I gotcha. The way I set it up is like this

• System Instructions: narrow the scope of their work down to something they cannot drift, and these instructions are totally permanent. Survives compaction indefinitely.
• AGENTS.md: Tell em not to run broad find or rg commands in ~/ and ~/Library because sandbox issues
• Skills: Define "sanctioned" ways to run privileged commands. All the agents run in workspace-write, so if they wanna use mutating git commands there are skill scripts designed to safely execute non-destructive commands without sandbox issues. There's also simulator piloting skills that tells em how to provision and pilot their simulator. Think of it as using try {} catch {} as control flow for agents. Let the sandbox guide them to the proper way of doing things. Running in --yolo is basically running without try {} catch {}.

And that's pretty much it!

[boreal holly]

codex -c model_instructions_file="/path/to/new/system/prompt.md"

[lost drum]

hmm I need to learn as much as you cause I see I am way behind

and I see that it would improve my system

[quick geode]

🙁

[signal tapir]

I must code the most boring stuff. I never get that warning.

[quick geode]

trying to security test api xd

[hard drum]

same

[floral garnet]

did 5.3 codex get dumbed down, it searched online for so much stuff now, i worked with the same stuff before and it used to figure things out really fast, now it just spends 5-10 mins searching trying to figure out problems

[fringe shuttle]

I don't see the /goal command. Is it only available for macs?

[hard drum]

No, you have to go to ~/.codex/config.toml, then find top-level [features] && add goals = true

[fringe shuttle]

wth

[hard drum]

yyyyeah

weird how you cannot just toggle it on normally

[torpid trout]

Does someone know why there are https://github.com/openai/codex/blob/main/codex-rs/core/gpt_5_codex_prompt.md prompt files in core, but also https://github.com/openai/codex/blob/c5e2921e1dbd43fc1324d3daed206212b0bccd1c/codex-rs/models-manager/models.json
The prompts are differing.

[wraith spoke]

tried getting codex to do pixel art in aseprite and it looks awful

[floral garnet]

5.3 codex definitely has been nerfed, gave 5.5 medium the problem, solved it in 2 minutes....and here I was hoping to use codex 5.3 for as long as possible

[boreal holly]

Yes, because they migrated away from putting the prompts in markdown files. In fact, the prompts in models.json are not even what gets loaded at runtime if you use a ChatGPT subscription, those are if you use the API and it pattern matches the model.

With subscription it fetches the base instructions from the backend API, takes your personality setting, and interpolates the personality into the top of the base instructions. That's also why folks can suddenly be like "GPT was doing good yesterday, now it does weird stuff like search the internet for half an hour". They can update the base instructions in the backend and you just receive them regardless of the version installed

[torpid trout]

OK
so we factually cannot know what it actually loads into instructions (even if we overwrite with model_instructions_file, because in theory at least, they could always re-inject something)

Sidenote, just found Cards are kept at 8px border radius or less unless the existing design system requires otherwise. in the 5.5 instructions lol

[fringe shuttle]

Looks like they made this little diff accordian more narrow after the latest update? why..

it's useless now

[boreal holly]

No, if you overwrite with model_instructions_file that is the least surprising path to know what instructions they have. When that is not set you get whatever the server provides. And the way to confirm is when you fire up a thread, look for the ~/.codex/sessions/**/*.jsonl, and the first line will contain the base instructions they received. If you've been using the defaults provided by OpenAI you'll see that base instructions evolve over time. When you override with model_instructions_file that sucker never changes

[potent mason]

Did the commit and push button disappear?

[boreal holly]

I have a question about the images 2.0 in Codex, for folks generating images, does anyone know if it uses the quota to do that? I know doing it in ChatGPT does not use the quota, never tried it in Codex, would consider it if I knew the impact

[unreal root]

Gotcha, weird now it's giving me this error

[boreal holly]

You should launch with codex -m "google/gemma-4-e2b" -c model_provider="lmstudio-local" instead of codex --oss

[unreal root]

Mm okay i'll check that out

That seems to work, when I run /model it only shows the GPT models ~ is there a way to show my loaded models there?

[rocky fog]

https://developers.openai.com/codex/pricing#how-does-image-generation-count-toward-usage-limits

[boreal holly]

You gotta configure your own model catalog

[unreal root]

Gotcha, thanks man!

[high girder]

1 image took 1% of the 5 hour limit on the plus plan

[wicked briar]

hi

[boreal holly]

Oh ok, that's not as much as I thought!

[wicked briar]

do you guys use high or xhigh

my usage is dropping really quickly

so im planning to switch to high

[high girder]

I use both. xhigh for xodex, the other for me

[wicked briar]

but honestly im going with xhigh cuz why not

I want it to get done correctly in one go

[rocky fog]

might also overthink it and/or waste context
but whatever works for you 😛

[hard drum]

nice savings

somehow it's running on the first iteration of oabtw v3, not the newly-made OAL

my weekly went from what... 62%-67% overnight...?

nvm 68% now

only 9% in deficit from 4%

[cyan wing]

@boreal holly do you have your mobile swarm setup connected to a Mac + computer use?

[boreal holly]

I been using it for two weeks with all supported commands. The savings are not too shabby!

[hard drum]

i am amazed because it's RUNNING ON THE OLD OABTW framework i had

the initial v3.0.0

like, i used to have 48-50% back then

[boreal holly]

Yes they're connected to my Mac. No computer use enabled

[hard drum]

idk what happened

&& of subsequent v3.x.x versions broke it so mcuh that my efficiency went down to 12-20%

so, slightly modified v3.0.0 somehow helped me... somehow

to create OAL v1 of course

but it's also fiercly grinding on osu-droid

[cyan wing]

why not 🙃

[hard drum]

if you're an EU user, you're not getting it

best you can do is force computer-use MCP

i tried...

you have to go thru some manual scharade just to get the mcp working

[boreal holly]

they each get their own iOS simulator and docker compose stack

[cyan wing]

can they navigate chrome in there?

[hard drum]

Browser Use, i guess?

[boreal holly]

They navigate with safari. Simulator doesn't have an app store

But they don't really use the browser, they just pilot the native apps I'm building

[cyan wing]

idk there's just something magical about seeing that little cursor do stuff in the window

[hard drum]

i wish i could try Computer Use, but effing EU, man...

[boreal holly]

ChatGPT atlas is pretty cool! Use it all the time to find parts for my truck

[hard drum]

sometimes EU does good, like making Apple use USB-C, then there's... welll... this

[cyan wing]

agreed! I just need programmatic / agents to have access to it

like I wanna fire off an agent doing something in my chrome browser
via my discord agent

[hard drum]

@boreal holly wanna give OAL a try on Codex?

[honest creek]

https://tenor.com/view/sweating-intense-sweating-nervous-anxious-waiting-gif-15207841

[hard drum]

i just told them to make a patch change for codex, the model instruction file thing

[boreal holly]

Gotcha, you should try idb and iOS simulators. Yes, it uses Safari, but it's not that bad. Maybe you can find a Chromium for iOS and push it in there. Or use adb on an android emulator

[hard drum]

the one you said about

[boreal holly]

Have a link? 🔗

[cyan wing]

what's the benefit over the Computer Use skill in Chrome?
security / sandboxing?

[boreal holly]

I can have many different agents operating independently in the simulators without stealing focus from each other or me. So less about security and more about concurrency

It does however use a lot of resources. Like 1GB RAM per agent, so typically 6-8GB constantly unavailable for other tasks

[cyan wing]

so like those automated social media phone farms 😆

but all virtual, and for actual productive work / verification of code ??

[hard drum]

i can't send here, but on my profile

[boreal holly]

Haha yes, exactly! Because unit tests & integration tests can only take you so far. At some point you gotta actually pilot the thing. Also when they write tests they typically do not cover all the edge cases so I make em find the edge cases via piloting and turn them into real fixes backed by tests

[glacial shadow]

Good morning, sorry I’m late, what are we scheduled to complain about today? Lmk

[cyan wing]

still no Codex mobile app...

[merry sluice]

my codex is always reconnecting and thinking... only complaint today

[worldly charm]

I don't know. Maybe your Codex is outdated?

[worldly charm]

Oh 🤔

[boreal crown]

tell me ur version number?

[worldly charm]

It appears in settings too. Much convenient I would say

[worldly charm]

I am on my phone right now. Not opening my laptop now since I am about to take a flight. But you can check online for the latest version.

[cinder plume]

This is awesome!

[torpid trout]

In codex it uses expontentially more tokens (they say twice to trice)
You should use the app for images (or the api, but since you have pro, use the app)

[small violet]

how do i activate pets

[boreal holly]

Gotcha, yeah I been using the app with massive, unrealistic levels of success. It's only a small amount of extra work to generate, drag & drop into codex anyway

[plucky halo]

/pets

install the hatch-pets skill to create them

[torpid trout]

It's interesting, I accidentally went xhigh and .. ."why is codex so dumb"
turned it back down to high and it was good again

[signal tapir]

That's.... strange

[small violet]

is this only available in the codex app

[plucky halo]

Yeah

[small violet]

or can i use it in the vs code extension

[cinder plume]

Does anyone else's pet have some bad pink fringing?

[torpid trout]

yeah, sorta, but also, not so much
I think it thinks too much, kind of reverting the effect of thinking

Why is everyone talking about pets suddenly
Is that a new insider joke y'all excluded me from

[plucky halo]

New feature in codex app

[exotic cave]

Are they going to allow 1m context window for 5.5?

[signal tapir]

I've stopped wanting huge contexts. It just eats credits without being worth it.

[torpid trout]

Its allowed since 5.4

just not set as such since mostly not what you want.

[exotic cave]

That's a poor view on it. It's used mainly for big refactor reviews etc... not used for day to day.

[signal tapir]

That makes sense

[exotic cave]

Yeah for 5.4, but 5.4 MRCRV2 is horrible for long context.

[torpid trout]

since 5.4
5.5 too

what is mrcrv2?

[exotic cave]

5.5 is much better at recall/accuracy over long context window.

source: https://openai.com/index/introducing-gpt-5-5/

That 1m context can make a difference in big reviews and more when needed. right now the compaction is so lossy it's annoying.

[torpid trout]

Understood... to answer your question, it has a 1m window if you enable it.

Is it useful? That is up to you to decide
In my experience it cant even keep a 256k window well managed unless you compact or restart often. But it might depends on the task

[exotic cave]

When enabled after the first prompt 5.5 turns back to 258K

hopefully they will allow us to set it at some point.

[torpid trout]

Can I ask a question that is not directly related to codex, but to code - as in how you interpret this image

• absolute nonsense
• it does convey the idea but does not have enough "code/tech" in the image
• it totally nails the idea

[torpid trout]

I am not sure that is expected. Its a toml setting, so... it should just apply.

[exotic cave]

yeah added this:
model_context_window = 1000000
model_auto_compact_token_limit = 800000

no luck with 5.5, 5.4 does accept this settings.

[rocky fog]

400000 might be max with chatgpt account
for more you might need API

https://openai.com/index/introducing-gpt-5-5/

so probably
model_context_window = 400000
model_auto_compact_token_limit = 370000

I never tried (yet)

[plucky halo]

I think it’s poor positioning. Customers don’t necessarily care if you’re ai assisted or ‘hand crafted’.

They want results. And they want results increasingly quickly.

That said, it looks real nice!

It gets the idea across though

[torpid trout]

I am positioning exactly where you are not buying stock diamond but "hand-dug gold crafted into a hand-forged ring" 🙂
However, your feedback confirms it gets that idea across, i guess
I was/am a bit afraid there is too much "not-programming" in the scene - it looks super nice but kind of weird for a programming domain 🤣
(I am originally boatbuilder, so when I told my story to gpt it went to that image lol)

[signal tapir]

I've done programming in the field of shipping.

That was maybe 12 years ago. Everything was ancient.

I upgraded a system from using punch cards to an actual SQL database

[plucky halo]

Oh nice! My father in law is a boat builder!

I get that but realistically, by the time someone is on your website they know what you’re offering. I think it looks great - it comes across as artisan programming rather than mass produced

[torpid trout]

Yes, you are right about that, by the time they reach that image, they already did read the main punchline, which is along the friendlier version of "you want a 200usd quick fix? There is the 🚪"
This is just a subpage (methodology) where I briefly tell them how I work (and kind of hint that a mm mistake in a boat makes the boat sink, so I port that same pedanticity to coding)

Your FIL still working on the field?

Its rare nowdays. That has been stockified time and time ago already 🙁

This was my last big one I had the honor to work on https://www.holzboote.ch/rum-runner.php, basically the nicest project the firm ever took where I worked

Well, off topic, so showing myself the 🚪

[signal tapir]

eek

I really really need another 20 accounts.

[torpid trout]

Its time!
Hopefully the hommage to the meme gods helps!

[small violet]

ts did not happen to bro 😭

[undone patio]

its his normal face wym

[small violet]

im talking about the accusations

[undone patio]

i know lol

[boreal holly]

who dat

[small violet]

search lorna hajdini

[boreal holly]

Dang that was quite the rabbit hole

[undone patio]

i bet you dont have these

cannons

[small violet]

he got this from wattpad

😭

pets in codex vs code extension wen

[velvet wren]

once I have hatched a pet how do I use it?

[undone patio]

/pet

you have to set it as the active pet in appearance tab of settings

[velvet wren]

thanks

got it 🙂

[karmic gulch]

After I updated latest on macos, my usage left dropped alot and it says I got pro lite, when I got the 200 dollar sub. anyone got same issue?

[deft gyro]

anybody getting stream disconnected before completion: failed to lookup address information: Try again?

[little hill]

does the codex review agent on gh use regular subscription usage bucket? i remember it used to be a distinct usage bar

[quick geode]

i see people using agents alot and youtube creators recommend them it seems. I cant find a use case for myself (i usually guide the ai/llm to what i want to create, a bit more involved than ai).

[warm pilot]

It's still a second tab, but it doesn't seem to be separated on the analytics page anymore. The json just has "code_review_rate_limit: null".

[little hill]

Ya it’s fried - I guess it just pulls from our pure codex usage

[supple perch]

hard caps at 77k

after that starts flipping facts

[hard drum]

think of yourself as the design lead

you tell your other workers what you want, why you want, how you want, && when you want

those 'agents' are your team in the company, essentially

it's almost like a pyramid-scheme-type corporate hierarchy

you're at the top--the human operator

then you got the orchestrator agent, && from there it goes into tree-sitter

does this help?

[torpid trout]

This is an amzing UI 😅
Makes me remember minesweeper back in the days when monitors where as big as trashcans lol

[hard drum]

God loves CRTs

[boreal holly]

You should see the new MS Paint! Now includes ads and subscription, >256 colors, tools deeply nested inside menus, requires microsoft acct and to be signed in

[hard drum]

LCDs are still inferior in everything but size && weight

don't we have all this modern technology to make thinner CRTs, || at least a spiritual successor to CRTs?

i mean, we've shrunken LCDs to mere millimetres in thickness

surely CRTs could harness newer technology advancements

[boreal holly]

Night vision goggles are CRT. It turns out the best way to do night vision is via particle acceleration, so they have actually advanced the CRT technology quite a bit for it

[sick axle]

can we get tokens refunded if it wastes tokens for hours

[torpid trout]

Seeing how good gpt image is, I am more and more thinking of just building websites with images now, adding positioned hotspots you can click, literally like back in the day animated acrobat sites where made

[boreal holly]

💀

GPT image is absolutely hilarious

[torpid trout]

I mean, the time it takes even AI to implement the HTML/CSS etc based on a design, and then thinking about responsivity.. in an image, you literally just have that, it works on god knows what outdated machine, no issue, it works in a 100 years still probably

Make it a webp, possibly sliice it so low internet does not choke, add interactive hotspots with some bimbo js script... and you can then literally do the "move that button a bit to the right" without having to explain to client "yeah but then we need to do hard coded positioning, which is suboptima;"
lol

the only unfortunate is, image edit with gpt is not good
it has to be a one shot currently, but that probably will be resolved soon

[hard drum]

THEN WHERE ARE OUR THIN CRT MONITORS && TVs??

if you can squeeze that tech in there, what's stopping LCDs from being seconded?

[turbid axle]

bruh

[torpid trout]

how many tokens did that use?!

This must be close to a 100% weekly gone?

[sick axle]

Codex not glitching for anyone else?

[exotic cave]

not here.

[boreal holly]

Try WSL

[sick axle]

Well even chat GPT was it glitching out on me yesterday?

I think it's on their end

I was just getting weird encrypted numbers and messages

But it's wasting my usage on codex

[boreal holly]

You're getting 400 Bad Request. First suspect, you're running with CMD instead of PowerShell. The message contains a bunch of unicode grapheme clusters \u{} presumably because CMD can't render those so it looks like garbage. Windows is janky. Switch to PowerShell or WSL

[sick axle]

I just resumed that session by a powershell and it's still giving me the same 404

Well maybe I'm just the first to get it. I'm hoping it's just going to be fixed on their end soon cuz I cannot proceed

[boreal holly]

It's hard to know exactly what the problem is because most of the output is not visible. It looks like tons and tons of multi-lingual text, then there's a

with length 6198 instead."
},
"status": 400
}

You very likely need to fork from a previous point in the conversation to recover, and if the agent read a file that contains non utf-8 text that's probably what caused the issue to happen

Like if it read a binary file or utf-16 encoded file, or something that isn't representable in utf-8, OpenAI is trying to parse the payload and finding a discrepancy in the reported length and is therefore rejecting it

[undone patio]

have you tried not using windows

[sick axle]

I do both

All three actually

[undone patio]

nice

i use mac and linux

prefer mac tho

using linux for my MCP server

since i opted for a cheapy minipc instead of a mac mini

this one

[turbid axle]

about 40%

[hidden zealot]

bro is this normal that tokens are dropping so much?

[sick axle]

Are you on fast?

[hidden zealot]

no

but three messages and down from 100% to 40% is weird

and I am using 5.5-medium

[undone patio]

what plan

[hidden zealot]

plus

[lost drum]

[hidden zealot]

now it is around 20

plus acts like free lmao

[undone patio]

[hidden zealot]

free-trial buddy

[lost drum]

weekly?

[undone patio]

i mean u cant expect much on plus

[hidden zealot]

no, the 5h limit

[lost drum]

tehn its normal on plus

[hidden zealot]

ehh i mean it was not like this before

like a week ago lol

[lost drum]

I had same thing on plus plan

then I went on 20x and no problem

[undone patio]

maybe they changed the tokenomics

[lost drum]

no, even before 5.5 it was like this

[undone patio]

20x is the only way tbh

if u make software for a living

[lost drum]

plus is good when you already created your whole project, plus will just maintain it but pro will build it

[undone patio]

theres always a new project to be made

[hidden zealot]

no money 😔

economy issues

[undone patio]

if ur a student i think they do discounts and sponsorships not sure

[lost drum]

eat bread with water and buy pro

[undone patio]

do you have local compute?

run kimi

or get google

i think they have good limits

[torpid trout]

Thats like the HODL fam, buy btc but sleep under the bridge lol
trust me it'll turn... trust me it'll make me money 😄

[inland sonnet]

if u dont wanna go broke u not gonna make it!

risk it or lose

[torpid trout]

Yeah, I can agree with that

[hard drum]

HODL? Hardware ??? Descriptor Language

if VHDL stands for VHSIC Hardware Description Language, what does HODL stand for?

[torpid trout]

You might be seriously asking, so I am going to seriously answer

Hold On for Dear Life

It comes from BTC trading, when btc does what it does best, many people panic and sell, instead of HODL

[hard drum]

this- oh

i didn't see LOL

[torpid trout]

HODLers are those who make money in BTC but also never really make the money since they... always HODL

🤣

[torpid trout]

TIL