codex-discussions

[solid lake]

Or if it’s stalling just esc and type in continue

It’s not that tho, it’s you basically regurgitating the same thing you say in the other server

[alpine wasp]

Tried that, but hitting esc doesn't really fix it. It just completely freezes up

[quick thorn]

how delusional can you be

Claude servers crashing is a different issue than lacking intelligence

[solid lake]

What platform are you on

[solid lake]

And if you think claude doesn’t drop in intelligence you have grass to touch or are just new to this

[plush harbor]

you may wish to try gemini

[quick thorn]

you’re delusional bro

[solid lake]

Have fun in your bubble

Keep complaining

[quick thorn]

bro swear he works for codex customer service 🤣

[golden imp]

claude is def dumb in its own way too

you should take biggies advice

[quick thorn]

Yeah but less so

[golden imp]

dont be a loyal hoe, just switch each time a model does its thing for you

if you build an agnostic workflow its not a big deal

[quick thorn]

Yeah satvik is one of those brand slaves

Ready to die for codex

[golden imp]

3 months ago its was all 4.5/4.6 as the implemtation agent for me

[quick thorn]

🤣

[solid lake]

Please look at his messages there bro is just larp

[golden imp]

now 4.7 is only allowed to plan and be my hostile reviewer

and that may change again next month

[quick thorn]

I find codex to be the better reviewer

[golden imp]

they both find things the other doesnt

[quick thorn]

True

[solid lake]

?

Is this true chat

[quick thorn]

codex doesn’t love you bro

[golden imp]

really thats the trick to all of this, two models etc, cause you certainly cant trust any of them

[solid lake]

They give a lot of perks ngl

Prolly love us more than daddy Dario

[alpine wasp]

ubuntu

[quick thorn]

Yeah I’m down to one model rn that’s probably my biggest problem

[solid lake]

Are you using the electron app

Or cli

[quick thorn]

Claude + codex is the absolute peak

[solid lake]

If former update if latter clear sessions (this works for me)

[golden imp]

f only they could sync up and not both be rtarded at the same time, would be great

[solid lake]

You’re not gonna believe this…

[golden imp]

in the scheme of things though, still pretty awesome from 6 months ago

[quick thorn]

yeah + $400 a month gets a little excessive

[golden imp]

conspiracy time? 😉

[quick thorn]

Plus tax

[solid lake]

Nah you can sync them if you want

[quick thorn]

So really like 450

[solid lake]

Just use claude code through codex

[quick thorn]

Bro has the answer to everything 😭

https://tenor.com/view/kanye-west-you-aint-got-the-answers-no-answers-answers-gif-5300959

[plush harbor]

I'm locked to codex until I finish my project, earn some money and buy a new computer so I've got no basis for comparison here

[solid lake]

I do study it so I’d hope so

[quick thorn]

Using codex only is going to bring me to an early death

[golden imp]

oh i do, i ment the models syncing up i.e both being usbale and at full power at once

[solid lake]

Nah bro just has all the usage here btw

[solid lake]

Tough ask

You should ask the other guy he seems to love it

[unreal parcel]

Do you think they both drop in intelligence or just claude

[solid lake]

Both

Claude more so because yk, less compute

[unreal parcel]

And yet....
https://marginlab.ai/trackers/codex/

[golden imp]

new model = full power for a few weeks, gets hyped up, users subscribe/migrate, gradually reduce compute/quantatise the model over time, then its rinse and repeat between them

[solid lake]

I don’t get it, it proves my point

[golden imp]

for me the most excitement of a new model is not it can do new X, its i get two weeks of full processing power

[unreal parcel]

There's literally no evidence of this

[quick thorn]

No the real answer is, whenever their enterprise client like J.P. Morgan or whoever else want to do something they throttle all of us

That’s the real answer

[golden imp]

lol the evidence is in using it day in day out over the cycles, its obvious

[solid lake]

Don’t tell them what compute is

Or what gpus are

[unreal parcel]

Then you need new eyes

[solid lake]

Blind leading the blind here

[quick thorn]

so anthropic doesn’t have access to colossus?

Answer

Exactly

[cobalt fjord]

Yeah same here especially running via oauth for Hermes..3 days now.

[solid lake]

You think colossus has only 1 model?

[unreal parcel]

"I have a feeling" isn't evidence

[solid lake]

Answer

[quick thorn]

Idk

[solid lake]

Read computer architecture book and comment

[plush harbor]

but you aren't doing exactly the same thing every day

[solid lake]

Have a nice day

[alpine wasp]

what a joke, the page status say 100% operational

[quick thorn]

Lmao this guy definitely wants to work for OpenAI

Good luck buddy

[golden imp]

close enough, anyways i see with my own eyes, thats my opinion not going to argue over it

[solid lake]

Baseless larp

[quick thorn]

the delusional jester was comical

[unreal parcel]

It's a tired opinion from 2023

[golden imp]

k

[cobalt fjord]

Yeah..gonna try VPN now lol.

[unreal parcel]

You get amazed by new capabilities, then eventually it becomes the new normal, then it screws up (as it's always done) you think "it's getting dumb!"

[quick thorn]

@solid lake why didn’t you answer this fact

[solid lake]

Twitter post btw

Bro just copy pastes same message in both servers

Have fun

[quick thorn]

deflect

Avoid

[solid lake]

Oh bros just an LLM

Throws words around

Issok

[quick thorn]

yup I’m an llm that’s how delusional this jester is

[alpine wasp]

same result with vpn, it's completely stalled. i have to wait 5 minutes just for one response

[golden imp]

glad you're having a good time with it and its working well for you

[quick thorn]

you think an llm has the ability to call you a delusional clown

[solid lake]

no way he’s using jester in this context

Is bro that brain rotted

[quick thorn]

Nah I’m a llm

[solid lake]

Oh okay

[quick thorn]

🤣

[unreal parcel]

@robust lintel @winged ore this has been happening for over 3 days now, codex is just completely unusuable for a lot of people just frequently freezing. Why is it taking so long to even acknowledge the problem? Many people here reporting it, in twitter too, I sent a help request with a lot of info too. https://x.com/i/status/2058754716889489872

[quick thorn]

this chat gave me my laughs for the day I’m not even mad

bro said I’m a llm 😭 I wish

[solid lake]

You wish you were a stochastic parrot?

[quick thorn]

https://tenor.com/view/court-jester-gif-13754025198523329817

[solid lake]

Hope you find peace in life

[quick thorn]

as long as I only have codex there will never be peace here.

[mild island]

anyone tried /goal here?

[cobalt fjord]

Vpn seems to be helping so far lol

[shell ice]

Has anyone else having issues with goal?

[urban beacon]

can someone tell me how much usage has codex?

in plus plan?

[mild island]

yeah

used it in production for refactoring a codebase - no feature implementation or bug fixing

i don’t understand how they say its persistent when I can’t even close codex while the goal is active - and secondly, context management is not done well in my opinion

[weak patrol]

Enough at 5.5 on high/xtra for about one day max. 5.3 would last you the week depending on what you are doing

sadly i do not think despite you can buy a quantity of tokens is there a visible UI showing exact usage

which makes a really crap user experience for those trying to actually budget or optimize what they do

[plush harbor]

Token usage: total=921,155 input=835,996 (+ 29,337,088 cached) output=85,159 (reasoning 22,031) << is a 5 hour window a million tokens on plus or something?

[solid lake]

Promo

[plush harbor]

yes lol

[weak patrol]

is that fromthe cli you pulled the stats?

the ui just has generalized %s

[plush harbor]

yeah, the end of a session

[weak patrol]

where/what might you use for that?

[plush harbor]

idk, it just tells you that when you leave a session. its default behaviour

[umbral depot]

How do I access codex in my phone I remember there was a website no pc

[solid lake]

App

[plush harbor]

if you only have a phone you aren't going to get far with codex, I think it needs a desktop one to connect to

or just do everything cloud

[umbral depot]

So they don't allow phone anymore that's unfortunate

[unreal parcel]

still happening and VPN doesn't completely fix it

[plush harbor]

you might be thinking cloud via the website, or some other 3rd party site

[deep arrow]

anyone use cursor? if i get both a cursor and codex sub is it possible to use my codex sub in cursor like in opencode?

pls ping me in replies too

[dull imp]

能借用一个英国号码吗

用来验证 codex

收个验证码就行

[dusky panther]

Has anyone successfully connected Codex to GLM5.1? Please share your step.

[umbral depot]

How do I install on termux

[signal tapir]

I asked chatgpt and got a few suggestions, and it looks very similar to how one would do it on regular linux.

[signal tapir]

I don't see why it wouldn't work.

Scam much?

[barren lotus]

Nope, what u want in proof.

[cedar skiff]

it's cheaper to buy a $200 sub

[digital saddle]

Hi, is this channel dead?

If not, what's the difference in quality for medium vs high temp?

Is the play to run medium by default and crank it up to high when medium can't cut it?

Or just run high always? xhigh seems overkill ngl.

[olive tangle]

how can u know if its overkill or not ?
Im turning always on 5.5 xhigh but i actually have no clue which one to chose

[tiny fulcrum]

depends what your plan is and how you are using it

[digital saddle]

Usage limit drain.

[tiny fulcrum]

xhigh takes longer, sometimes medium does the same in shorter time, reasoning only makes more sense if the problem is complex enough where medium reasoning doesn't get the job done

[digital saddle]

Coding mainly. You could say 50% new development like API design along with frontend design and the other 50% debugging and fixing filed bugs.

Also, yeah, I use the chrome extension heavily for verifications.

Or automation.

Is it normal for subagents to show up here even after they are done?

[golden imp]

no this channel is just endlessly compacting atm

[cedar skiff]

low does a good job looking at code for investigations

so does 5.4 mini high

[plush harbor]

I didn't realise how hard it nags you to switch to 5.4 mini until I used almost a full 5 hour

[unique spade]

first it's not temp, it's reasoning effort.
if you want a simple rule of thumb, think of the higher the effort the more inference passess the model makes

a higher effort has better chance to grasp more depth, spot more edges etc

[solid lake]

Xhigh my beloved

Will miss these limits

[unique spade]

are you on pro 100?

[solid lake]

Yes

Prolly will get the big plan but haven’t decided

[unique spade]

same. probably will move to 200

[solid lake]

Waiting for models to drop

[unique spade]

since they decided to keep 200 at 20x , it s economically rational lol

[solid lake]

Yeah looks like

[unique spade]

since it will be 2x times more cheaper than 100

[digital saddle]

Right. Reasoning. So what's the sweet spot? Medium or High?

I use Opus 4.7 at xhigh always.

[solid lake]

Whatever produces the best code for you

This depends very heavily on your codebase ngl

[unique spade]

if you use xhigh always, you already know the answer for codex too

[solid lake]

Yea

[digital saddle]

But Codex seem to be running into rate limits?

[solid lake]

Are you on plus

Or pro

[digital saddle]

Yeah, I don't have the time to run the same task with different reasoning levels to compare the output honestly.

Pro 5x.

[solid lake]

I’ve been on Xhigh /fast all month

[digital saddle]

Do you use subagents like I do?

[unique spade]

then pay someone to work for you then lol

how could you anyone on this discord with 0 info of what you actually doing know what effort is best for you

[solid lake]

lol if only there was a tool we already pay to work

[digital saddle]

I'm already paying Anthrophic. I could get Claude to run multiple tests to verify.

[solid lake]

Jokes btw

[digital saddle]

False. I have shared context above #codex-discussions message and folks are more than welcome to ask for further context.

[unique spade]

models are very relative to workflow. they re not like videocards having objective framerates in various game resolutions

[solid lake]

If you’ve used xhigh opus all this time, xhigh 5.5 should be your choice. Limits are higher if you have equivalent plan

[unique spade]

i m speaking on this convo here and now

[digital saddle]

Well, you should jump into a convo after reading context.

[solid lake]

Also milk it till promo ends in the next 5 days

[unique spade]

again, xhigh will work best with ambigous or underspecified intent

[digital saddle]

I'll probably try /fast mode. It does appear to be pretty slow.

[solid lake]

Yes

[solid lake]

/fast is bearable

[digital saddle]

This is with High. Not even xhigh.

[solid lake]

Yep

[unique spade]

yea maybe go do that, if that's already working fine for you

[crimson zodiac]

did they lower the limit again?

[bright swift]

doesnt end for pro 20x

i hope performance gets better when all the plus and pro-lite plans are reduced

[solid lake]

https://x.com/thsottiaux/status/2043075353242218768?

[bright swift]

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

[unique spade]

plus is not on promotion anymore since 3 months ago

[solid lake]

Eh

[bright swift]

o rly

[unique spade]

only pro 100 currently on 2x promo

[solid lake]

No way i got that much usage without promo

Bet

[unique spade]

? you re on plus?

[solid lake]

No i have 2 plus accounts too

Cancelled 1

Might cancel the other and get pro 200

[bright swift]

do it

[solid lake]

But instead of that I’ll prolly keep the 200 and one plus

[bright swift]

do 2x pro 200

[solid lake]

Idk let me first max out the first one

[coral remnant]

200 plan is decent. I got about 2.1 billion tokens on /fast over 2 days before I ran out. 600m into my second plan with no /fast usage, 14 hours in and at 82% usage

Depends on your usage of course

[bright swift]

i dont use fast, prefer to have some things going in parallel instead for same total usage

[quick thorn]

what did you accomplish with 2 billion tokens

[solid lake]

That’s it?

2 billion?

[coral remnant]

Finished my iOS emulator, debugged problems with a certain binary not properly decoding encoded scalar, built a private dashboard, some MCP servers etc

[coral remnant]

Yeah that's about what you get when you use /fast

[quick thorn]

Noice

[solid lake]

Interesting

[bright swift]

fast is 1,5x speed at 2,5x usage right?

[coral remnant]

I ran it from the minute it reset to the minute it ran out on /fast. I may have been running 2 chats at one point - lots of long running goals, 2 days of near 24/7 usage and 2b tokens of output. By that metric you can expect ~5b normally, but obviously that is outputting slower so by the time you've used 5 billion tokens you're likely to be at / near the reset

[coral remnant]

Correct. It's not worth it imo unless you have a deadline which I did

[bright swift]

i dream of running /fast with API key and being able to afford it 😄

[solid lake]

Need to check usage now

Brb

Yeah sounds about right

[coral remnant]

API pricing is rediculous. I think I've done ~9b in the past 2 weeks, which would have been ~$6.8k iirc?

[solid lake]

More

[bright swift]

Yea I think you get sth around 10k out of a 200 plan, maybe even more

[coral remnant]

It's heavily heavily subsidised, especially with the reset schedule Tibo's got us on

but I'm going to get blue balls when they stop

[solid lake]

X20 was also given the extra usage

Ig now on a rolling basis

I remember it was super confusing to decipher how much they gave tho

[coral remnant]

I honestly don't pay attention to it too much. I use what I need to, if I need more and it's worth it, I'll buy more.
I only use Codex for projects where the only important part is the output being correct. If I care about the internals, I need to put a bit more care into it.
That said, I'm pretty happy with how a couple /nuclear-code-reviews turned out, got from 2 15k line scripts to a lovely balanced directory with max 600 lines per file

[bright swift]

Thats the skill from cursor i believe?

[coral remnant]

Yeah I just ripped it and installed it in cursor to try it out

[bright swift]

Nice, i have it bookmarked still have to test it

[solid lake]

Is this the skill from Twitter

[coral remnant]

I've got a couple others, like I downloaded the Jira confluence guide and made an agile skill, I have a rust best practices one I use in Plan Mode

[solid lake]

Like from yesterday or a couple days ago now idr

[coral remnant]

I've been testing out codex-lb, and honestly, not amazed. It's buggy and not reliable, and the docs are a mess.
I was thinking about contributing, but there are PR's that are approved and not merged, I don't really want to get involved with a bunch of merge conflicts and versioning issues

[verbal zenith]

is it just me or each tool call takes ages?

[bright swift]

i worked on this concept i called chaos review, inspired by netflix chaos monkey, basically pick a random file, then read the high level docs, and review its part in the system, if the tests are good etc

[solid lake]

Too late buddy sama is approaching your location

Tibo is already in your room

[coral remnant]

I asked in here and I was informed 2 accounts is fine

[coral remnant]

good i can ask him to add some features :))

tibo give me the reset button

[coral remnant]

Seems fine on my end

[bright swift]

i talked to codex about it and it suggest a bunch of options to implement this better

incl hotswapping during active turns

needs a codex fork though

[coral remnant]

Oh there are infinite ways - I presume the project was started by a non-developer.

[coral remnant]

Honestly as long as you keep the websocket connection stable & open I don't see a reason why you couldn't direct individual requests to different accounts

The codex client is blisfully unaware of what happends behind the abstraction of that websocket, as long as it replies in the required format

(to my knowledge, I haven't delved deep on these things)

[solid lake]

Websocket hasn’t been stable enough for that ngl

[bright swift]

well you have to be mindful of the caching still, so dont switch constantly. but codex already has a built-in reload auth function, its just not publicly exposed

[solid lake]

Http fallback still better rn

[coral remnant]

This is true. I presume it'd actually be less effective to do mid-turn swapping. Honestly all I care about is a convenient way to

1. Align my reset periods
2. Swap from Account X to Account Y when the ratelimit is maxed out

[bright swift]

yea and without the manual logout/login, thats the annoying part

[coral remnant]

Login is and always has been awfully buggy for me on OAI products

To the extent I've considered cancelling my subscription out of frustration many times

[bright swift]

works fine for me, just multiple accounts with 2FA etc its annoying

[solid lake]

Threatening gpt with this works too btw

[coral remnant]

I've looked into the auth flows / requests and I still don't understand it. It's that preauth cookie error bug that's been reported 100 times but never gets a clear fix / response

I'm currently still logged out of the ChatGPT app on my Mac. To make matters worse, every time I tried to log in to one place it would force log me out of everything else. I just gave up. I use ChatGPT on the web or mobile, and Codex as the app as it seems to work

Could be worse though. A\ took 3 months to respond to a support ticket I opened about a new hidden limit they applied on everyones plans (200$ included).
I cancelled my subscription days after reporting it. Told them I switched to Codex 🤣

[bright swift]

i use cli only because the app always seems to pull in additional stuff like tool descriptions i dont need, since they share the app server + central config apparently. easiest solution to the auth stuff seems to be using the CODEX_HOME env var

probably also for my cli/app issue

[nova turret]

https://x.com/drummatick/status/2058429773370773599?s=20

[bright swift]

like CODEX_HOME="$HOME/.codex-1" codex

[nova turret]

How???

😭

[coral remnant]

I'm using experimental_bearer_token because managing env vars is a pain for apps on MacOS

[bright swift]

pretty sure thats a fake screenshot, for me codex switches to "1d 0h" after "24h".

[coral remnant]

I've had longer then this without a completion - and a clearly defined goal lol

[coral remnant]

Nope. At least not in the MacOS app.

[bright swift]

hmm ok. longest goal for me was 4d so far

[cedar skiff]

what took 4 days?

[bright swift]

it was a big refactoring, there was a bunch of wall time for test running, compiling, etc

[coral remnant]

May be wrong on this

[bright swift]

so not 4 days of inference

[coral remnant]

I had to optimise one of my programs because each turn was taking 3h+ due to runs being so slow

Made me realise that iteration speed is really important to accomplish things in decent timeframes using LLM's

[nova turret]

Please help

Jow

How

[coral remnant]

It's often better to stop and optimise performance then to have it keep drilling on

[nova turret]

Is it some secret prompt engineering

[nova turret]

My codex doesn't work for more than 40 mins

[coral remnant]

It's just a goal that's defined clearly yet difficult enough to achieve. It will always keep going until it resolves the goal

[bright swift]

I had this stupid idea for "paired coverage" where each source file needs to have a sibling test file and only coverage from that one counts. That slowed down everything to a crawl haha

[coral remnant]

I was working on top of Unicorn and writing a lot of features that instrumented at the instruction level with runs that execute over 20 million instructions.
I had to do quite a bit of optimisation to get runs down from 40min+ each to 12 seconds lol

I only realised because I came back to Codex after leaving it on overnight to only see 3 new messages

Waste of 12 hours, but you learn to take these things into consideration

[bright swift]

i recently migrated my main project from gitlab to github, incl a lot of CI with auto deploy to prod etc

codex wrote typescript tests for terraform, yaml, cloudformation, you name it

[quick thorn]

/goal doesn’t work for me codex refuses it says “can’t do goal”

[bright swift]

absurd token burn

[coral remnant]

I've never been a Gitlab fan to be honest. I tend to stay away from hosted version control now though as a lot of my projects are quite sensitive. Just local vc and onsite / offsite backups.

[coral remnant]

hahahhahhaah

I mean atleast you know your infra wont fail

that's gotta be worth something

[nova turret]

Umm super duper 4 digit frontend thisend thatend developers please help

[solid lake]

Tests in golang are annoying

[bright swift]

yea i removed the obviously stupid ones ofc 😄 but it seems pretty solid now, before i had this handwritten 1k LOC deploy.sh haha

[coral remnant]

It does take instructions very very literally now though. It's good because it reduces instruction following issues, but it can lead to some really annoying situations

[coral remnant]

My deploy.sh is normally just a docker build, tag and push. Hook portainer / kameleon up to the registry and have it auto pull and deploy. Saves me writing a bunch of IaC stuff

[solid lake]

It’s also burning through usage rn apparently

/goal.

[bright swift]

yea i had sth in my goal about dont break prod, so this probably caused this

[solid lake]

Stuck on user approval

[coral remnant]

meh I had this and 2 other chats I've been using and my usage seems normal

[solid lake]

Seems like a growing trend rn

Could’ve been patched but timeline wise doesn’t look like you’re far from the reports

[coral remnant]

If it's prompting for permissions for silly stuff, you may have to add
approvals_reviewer = "auto_review"
to config.toml

I had that issue yesterday

[solid lake]

Nono this is for the /goal burn

[bright swift]

and of course it added a smoke test lol

[solid lake]

Apparently the permissions are causing infinite loops

[coral remnant]

I've seen it doing that recently, then I saw that meme on X lol. Seems like a 5.5 thing though, can't remember it ever writing smoke tests before

Both of the times it's written them for me it's been extremely helpful though. One was to ensure a unicorn python oracle output matched that of the rust rewrite, and another was to validate a login flow on an old deprecated microservice I was rewriting

[bright swift]

yea i mean smoke testing prod after deploy is reasonable

[cedar skiff]

i have some similar works flows took a few weeks to complete some of them, i ran orchestration and stopped it periodically to check and change lanes etc.

[coral remnant]

I have yet to reach prod with this lol

[white bane]

we need to validate the validation, and to validate the validation validating the validation and have a typeclass for the validation for the telescope validation and ask forty wrappers if the 30 lines of code has permission to exist

[cedar skiff]

the limiting factor with goal is its 4k chars

[coral remnant]

I honestly feel that's not a limitation. If your goal is too big, it's likely too big to be a goal

[bright swift]

/goal goal.md

you can even edit it while its working

[coral remnant]

A goal isn't meant to be an entire Epic, it's meant to be more of a Story from what I've inferred (sorry for the Agile PTSD)

[solid lake]

Smoke tests being synthetic ? I find codex ignoring test data and go make its own synthetic cases

[solid lake]

EW

KANBAN

[cedar skiff]

The create goal tool limits it to 4k afaik

Ill try that though

[coral remnant]

Those tasks do not need a bigger prompt. They need a persistent objective.
From the Goals doc

[cedar skiff]

Right but if you want the goal to be a work flow?

[coral remnant]

How many story points to create our own in house version of Codex?

[cedar skiff]

so it has some structure and doesnt just do it how ever it likes?

I just use orchestration usually

[solid lake]

5k

[urban beacon]

guys codex or cursor?

[coral remnant]

The whole point of a goal is that it will loop until a specific measurable outcome has been reached. They're not meant to run workflows, they're meant to keep Codex locked onto a specific problem without veering off track and being told to "Continue" every 10 minutes

[solid lake]

Actually excessive maybe 2-3

Depending on milestones ofc

[coral remnant]

They're two very different products. One is an assistance tool built into an IDE, one is a standalone product meant to write code autonomously. Depends what you're looking for

[urban beacon]

write code on java

i cant find a solution on this really if i pay for cursor they gonna reset the limit every one month which is soooo muchh

[solid lake]

Do you want to watch it be written

[cedar skiff]

You dont have to do that anyway "Continue" every 10 minutes Where do you get this definition from?
It's not very useful if you can't give it structure with in the task. Large code bases need to follow architecture and you need work flows to bring those into line.

[solid lake]

Or do you want to spend time chatting with agent

[urban beacon]

watch it to be written

[solid lake]

Ide then

Codex in vscode or cursor

Idk about cursor limits codex definitely more rn

Unless composer

Which is just a kimi

Which you could get cheaper than cursor

I think

[cedar skiff]

Passing files and using skills overcomes the 4k limitation, but i already use orchestration like i said

[urban beacon]

so u suggest intellij?

[solid lake]

Haven’t used actually

[coral remnant]

I'm not entirely sure what you refer to when you say "definition". If you don't have a goal set, when Codex finishes a turn, it will wait for user input. When you do have a goal set, it will automatically resume itself after a turn ends until it decides to mark the goal as complete.
I cannot really advise you on whats best in your specific situation as I do not have knowledge of the codebase that you are working in, or the "workflow" that you speak of.
What I can tell you is how the Codex team intended for goals to be used, how they've instructed end-users to use them in order to derive the best outcomes from them, and what my personal experience has shown to be the best way to utilise them.

[solid lake]

Vscode or vim rn for me

[bright swift]

i tell gpt pro to read gpt 5.5 prompt guidance and goals guide from openai docs and then write a goal doc for me, works well

[coral remnant]

Robust, long standing, heavy startup, menus that go on for kilometers.
If you like old software with 1000 options, it's for you. If you prefer a .json - probably not

[bright swift]

https://developers.openai.com/cookbook/examples/codex/using_goals_in_codex

[coral remnant]

Good language support though. They customise editors for each language

[solid lake]

I see

Idm the extra customisation

But i also don’t want more than what i get from vs

Or need

[coral remnant]

For Java - VSCode, Zed, Vim etc are a bit of a pain to use because you have to execute all the build commands yourself, manage the extensive files that java build systems require etc.
IntelliJ, Eclipse etc manage Maven / other build systems for you, so you just have to hit run. It's a lot easier to work with

[solid lake]

Yeah i mostly use go and cpp

Some python

Vs is okay with these (although the language server takes a while to refresh sometimes annoying)

[coral remnant]

I'm a massive Go fan if I'm manually writing code, but I find agents write really good rust, and rust is quite a robust language - the compiler is very strict and can prevent a lot of footguns

[nova turret]

Umm super duper developers

[bright swift]

also strict clippy rules etc

[nova turret]

Was this /goal thing there when gpt 5.4 launched

??

[cedar skiff]

You definition is this: The whole point of a goal is that it will loop until a specific measurable outcome has been reached. They're not meant to run workflows, they're meant to keep Codex locked onto a specific problem without veering off track and being told to "Continue" every 10 minutes

Where do you get this from? You don't need to use goal to keep a loop going. You can just use orchestration and have been able to for a long while.
In orchestration i use structured prompts for worker sub-agents and debug sub-agents. Then there is the actual orchestration instructions for the main agent. This can easily be 4k chars.

You're not actually explaining anything to me i don't understand.
I just think orchestration already solves the problem in a better way.

[solid lake]

Yeah i have a rust rewrite for one of my project in a roadmap but wont do it just yet

[bright swift]

no its relatively new, few weeks

after 5.5 launch i believe

[solid lake]

Yeah

Just look for Ralph loop mentions on Twitter

You’ll see exact timeline

[solid lake]

Subagents and goals are a risky mix from what i hear

Not stable enough

[nova turret]

Ok so you people weren't able to make gpt run more than 3 hours before the launch of /goal

[bright swift]

who says that?

[coral remnant]

I mean I don't want to be pedantic, but that's a statement, not a definition.
Orchestration is meant to solve a different problem then goals are. I've been saying from the beginning that goals are not built for large problems. They're not built to run workflows.

If orchestration solves the problem in a better way, why are you so persistent on trying to use goals for it?

[solid lake]

I haven’t even used it 💀

I should

[coral remnant]

I used omo and its horrific subagents. I'm glad I ditched it, but it was nice when not much else existed

[urban beacon]

the problem with vs is he is only talking english

[nova turret]

In codex you cannot spawn more than 10 subagents right

And the subagents aren't like gpt 5.5 pro level

[solid lake]

Smaller

4

[bright swift]

you can spawn unlimited, just change the config value, you can even nest them

[solid lake]

Or 5 I think

[nova turret]

Where

It says multi agent true

In config

[coral remnant]

err not sure. I had 20 odd going in OC a while back when I was trying to emulate a company structure. I had an agent per employee lol. Was a cool experiment

[bright swift]

[agents]
max_threads = 10
max_depth = 1

[nova turret]

In cc you're able to run 70+

[nova turret]

Can you give some other stuff from your config

That I can add in mine

[solid lake]

God i want to read a codebase that was made using this

[bright swift]

model = "gpt-5.5"
model_reasoning_effort = "xhigh"

web_search = "live"
personality = "none"

tool_output_token_limit = 25000
# Leave room for native compaction near the 272–273k context window.
# Formula: 273000 - (tool_output_token_limit + 15000)
# With tool_output_token_limit=25000 ⇒ 273000 - (25000 + 15000) = 233000
model_auto_compact_token_limit = 233000

hide_agent_reasoning = false
show_raw_agent_reasoning = true
model_reasoning_summary = "detailed"

#suppress_unstable_features_warning = true

[agents]
max_threads = 10
max_depth = 1

[features]
goals = true
fast_mode = false # fast mode consume usage at 2x rate
apps = false # buggy, burns tokens
default_mode_request_user_input = false # interrupts /goal runs that are supposed to be autonomous

[notice]
hide_full_access_warning = true
fast_default_opt_out = true # opt out of the default fast mode added in 0.124.0

[nova turret]

No js repl?

[bright swift]

whats that?

[coral remnant]

I'll send you a project I worked on. I gave up on it but I had 20 subagents running for a good couple of days. It's a funny repo to look at. You had agents making PR's and other agents reviewing them

[solid lake]

Sure

Max i spawn is 5

That too only in audits

[bright swift]

i tried 20 or so and it crashed

[solid lake]

Ram

[bright swift]

like it was still running, but it didnt render anything anymore

[bright swift]

64gb ram

[solid lake]

Try as much as ram allows

[solid lake]

I’m on same use a bit less

[hard drum]

wish i had 64gb ram

or even 48gb

stuck at measly 32gb

[solid lake]

Lots of memory hungry processes

By codex

Even cc

[hard drum]

maybe when M4 Pro macbooks get cheap at some point, scoring one for 1.5k here in EU would be neat

[bright swift]

20 subagents running rust compiles in parallel kinda blows up any cpu/ram you have

[hard drum]

y'could tax your whole machine just running cargo tests alone

i did froze my machine twice doing that

[coral remnant]

offload builds to a load balancer with remote compile farm

then you can run all the agents you want

[hard drum]

something is so wrng w/ Cargo build/test on workspaces

[nova turret]

[features]
unified_exec = true
shell_snapshot = true
ghost_commit = true
apply_patch_freeform = true
skills = true
undo = true
memories = true
image_detail_original = true
responses_websockets_v2 = true
runtime_metrics = true
js_repl = true
js_repl_tools_only = true
codex_git_commit = true

[bright swift]

i switched from subagents to one codex in its own remote dev vps

and just spin those up as i need

[coral remnant]

not a clue what you're building / testing to warrant this lol

[hard drum]

compiler

not really a compielr but a whole thing

tools, lsp, vm, etc

[coral remnant]

lol this was a forseeable issue

[hard drum]

it's like 120K of absolute 🍑 ery that I haven't touched in a week

it has a lot of reductions due

[cedar skiff]

Right so we agree about the limitation then?

[coral remnant]

It's not a limitation

[cedar skiff]

I'd also like to know what you base your statement on

[coral remnant]

https://developers.openai.com/cookbook/examples/codex/using_goals_in_codex

[neat hound]

anybody knows how to prevent this on windows, going mental

[bright swift]

buy a mac

[cedar skiff]

You dont have to have a goal set for this, you can simply have completion criteria. You don't need to create_goal for it at all.

and the link calls out constraints, boundaries, and an iteration policy which is a workflow

[coral remnant]

Yes, I'm aware that workflows exist

[cedar skiff]

well your link contradicts your statement

Either way, it limits chars at 4k.

which is a limitation

[coral remnant]

You were asking why you couldn't use over 4000 characters for a goal. I told you goals weren't meant to be big. You continue to talk about workflows for some reason. I tell you I don't have enough knowledge of your set-up to tell you exactly what you should be doing. You proceed to tell me that a goal isn't the only way to keep a thread going, and that orchestration solves your problem better then goals can. I ask you why you are so persistent on trying to use goals for something you claim orchestration does better. You don't answer the question

It's not a limitation because goals aren't meant to be big!!!

[cedar skiff]

I never asked that

why you couldn't use over 4000 characters for a goal

[coral remnant]

the limiting factor with goal is its 4k chars

A goal isn't meant to be an entire Epic, it's meant to be more of a Story from what I've inferred (sorry for the Agile PTSD)

Right but if you want the goal to be a work flow?

[bright swift]

goals with big detailed goal files work fine for me

[cedar skiff]

Right i called it a limitation i didnt ask WHY?

[coral remnant]

It's not a limitation

It's purposeful. The feature is not meant to be used how you want to use it

[cedar skiff]

back up a second, you said i asked why its limited. I didnt.

the premise of your argument is wrong

It is literally limited to 4k

the create_goal tool limits it to 4k chars

which you probably didn't even know until i said it

[bright swift]

what are you even fighting over

[cedar skiff]

I actually went out of my way to inspect the source code

[cedar skiff]

Youre right im off

[coral remnant]

Yes because goals aren't meant to be long. You are not meant to use goals in place of workflows. You asked why you couldn't run goals as workflows.
You can't because there is a 4000 line limitation. To prevent you from trying to use the goal tool in a way it wasn't intended. Read the document I linked you about the architecture of goals, who owns a goal, what state a goal carries.
You're not meant to use a goal as a workflow. I said this from the beginning

You didn't directly ask me why you couldn't use over 4000 characters for a goal, but I'm trying to explain to you why that limitation exists, and you can't seem to grasp the concept

A good Goal is more than a larger prompt. It is a compact contract for how Codex should work, what counts as success, and what should happen if success is not yet reachable.

Goals are implemented as persisted thread state, not as global memory and not as project-level instructions. That design choice is important: the objective belongs to the thread where the relevant context lives

At the architecture layer, a Goal is a durable, thread-scoped state. It records the objective, lifecycle, budget, and progress accounting that Codex needs to evaluate the thread over time.
Codex treats that state as a contract between the user, the model, and the thread.

This is the architecture to keep in mind: a Goal is a thread-scoped completion contract. It combines durable objective state, lifecycle controls, continuation policy, budget accounting, and evidence-based completion. The point is not to make Codex loop forever; it is to let the objective persist until the evidence says the work is done.

Do not use a Goal when the finish line is vague. Make this better gives Codex no reliable completion condition. Refactor this code is also weak unless you define the expected end state, tests, and constraints.

Goals are strongest when the task has three properties: a durable objective, an evidence-based finish line, and a path that may require several turns of investigation.

[bright swift]

you dont need to inspect the code, it tells you right in the UI

like i said this works fine for me, i havent seen a difference really. i use it with well-scoped instructions though

[cedar skiff]

how do you get this?

is that in the cli?

[cedar skiff]

So we agree on the limitation.

[bright swift]

when you type /goal and then paste something >4000 chars in the cli

[coral remnant]

A car not being able to fly is not a limitation. What you want to do is outside the bounds of what goal is meant to do.

We do not have to continue further with this discussion. I have tried to relay everything in as clear a manner as I can, citing the official docs, but you don't seem to want to listen. Good luck with your goalflows

[cedar skiff]

Oh i see its if you give it too much, never actually got that far haha. After i inspected the source i was like ahh ok ill stick to the orchestration because of the 4k limitation. I have used it a few times for some smaller stuff though

[cedar skiff]

Your analogy is just silly, the create goal tool actively limits to 4k. unlike a flying car. If the car could fly and they limited it so it couldnt that might the same. But i think the idea is missed on you. It was my mistake for trying to talk to you.

I wont make that mistake again

[coral remnant]

It wasn't the best analogy, no. You want to use a tool that was designed for purpose X to do purpose Y. The developers knew that this tool would not perform purpose Y well, and therefore tried to stop people from doing purpose Y, so they would not complain about its poor performance.
You seem to be upset that I am telling you that Goals are architected specifically for well defined tasks with a visible end. A workflow is not a goal.
Not being able to use a goal to run a workflow, is not a limitation of the goal tool.

[cedar skiff]

I dont want to do that, i never said that. You just make stuff up as you go. You want to use a tool that was designed for purpose X to do purpose Y
I just simply said it's limited to 4k so i perfer orchestration.

[urban beacon]

ffsx

so where should i pay for

[coral remnant]

You can call it a limitation if you'd like to, but that is not the point I'm trying to make. The point I have emphasised in every single message to you is that the "limitation" is the point. It's on purpose. It's to prevent you from doing exactly what you're trying to do.

[lost drum]

pls reset

[signal tapir]

No

I need to use up my tokens first

[cedar skiff]

Youre clearly taking it out of context, i am implying that orchestration is the answer.

like i said several times

[coral remnant]

You say:

I just think orchestration already solves the problem in a better way.

I say:

If orchestration solves the problem in a better way, why are you so persistent on trying to use goals for it?

You avoid my question entirely, and say

Right so we agree about the limitation then?

[cedar skiff]

If orchestration solves the problem in a better way, why are you so persistent on trying to use goals for it?
This is a loaded question

i dont do this!

i said many times i dont but your are just like talking to a wall

[urban beacon]

so cursor or codex? which one should i pay for

[cedar skiff]

You just make it up as you go, you misunderstand everything and make it a truth in your mind.

[bright swift]

just stack codex pro 200 plans

[coral remnant]

the limiting factor with goal is its 4k chars

I honestly feel that's not a limitation. If your goal is too big, it's likely too big to be a goal

Right but if you want the goal to be a work flow?

I have stated many times

You're not meant to use a goal as a workflow.

[signal tapir]

Stack?! How many can you need?

[urban beacon]

what do u mean stack

[cedar skiff]

again the implication is that orchestration is the answer

[coral remnant]

Okay, then we agree on the architectural answer: workflow-style behavior belongs in orchestration, not in Goals.

My point was specifically about the 4k cap being framed as the limiting factor. If the thing you want to express needs workflow-sized instructions, then it has probably exceeded the shape of a Goal. At that point, the answer is not “make Goals larger,” it is “use the orchestration layer.”

So yes, orchestration is the answer. That is also why I do not think the Goal character limit is the core problem here.

[bright swift]

just add more once the first runs out after a day or so

[cedar skiff]

No i think a small workflow can work in a goal.

it just needs a completion criteria.

[cedar skiff]

It is a limitation by the very implementation limited to 4k chars

I never actually met that limitation

i only ever used it for smaller tasks

[coral remnant]

Yeah, a small workflow can work inside a Goal for sure, as long as it still meets the criteria for being a Goal.

The issue is that Goals and workflows are architected very differently. The Goals page describes Goals as thread-scoped completion contracts with evidence-based finish lines. The subagents/workflow page describes orchestration as spawning agents, routing work, waiting for results, and consolidating outputs.

Those are inherently different abstractions. Something can overlap both categories in simple cases, but that does not make them interchangeable in the general case.

That is the point I was trying to re-emphasize: a Goal can contain workflow-like steps, but a Goal is not a replacement for a workflow.

So if you take a workflow and try to put it into a Goal without adapting it to the Goal shape (compact objective, clear completion criteria, verification surface, constraints, and blocked stop condition) then the 4k cap is not really the core problem. The problem is that the thing has not been shaped into a Goal.

Goals are meant to be compact. Workflows are usually procedural, reusable, and potentially open-ended. So while a small workflow can fit inside a Goal, that does not mean it should. It only belongs there if it has been reduced into a bounded, compact, evidence-checkable objective.

[lost drum]

so we have 5 days untill they remove 2x on pro damn

[cedar skiff]

You're just yapping like you have some magical information to teach.

[bright swift]

on pro 200 its permanent

[urban beacon]

sweetheart u mean to put more money in there?

[lost drum]

Really?

[bright swift]

yes feed the slot machine while its subsidized

[urban beacon]

how much should i feed it

how much dollars sweetheart

[lost drum]

Then am happy haha

[bright swift]

i think 1x pro 200 / day is a reasonable beginner goal

[urban beacon]

200 a day? im not rich sweetheart

thats too much

[umbral depot]

How do I start

[signal tapir]

By asking chatgpt

[bright swift]

just tell codex to make you more monies

[urban beacon]

im out of limit ask him for me please

[bright swift]

busy making me monies

sry

[umbral depot]

I did what it said but I keep getting error

[hard drum]

codex != chatgpt

ask the web gpt

codex limits don't apply there afaik

[umbral depot]

What does that have to do with how to install

[signal tapir]

Ask it how to get better at using it for tasks like this.

[umbral depot]

I mean I keep getting error installing it

[signal tapir]

Ask it how to get better at using it for tasks like this.

Say something like "I end up needing to ask humans for help on things like this. How do I get better at using you to get to the bottom of problems like the one I am having now?"

[bright swift]

i just invented "decision-sourced specification", wonder how that will play out haha

[umbral depot]

It still don't help me

[bright swift]

you need to have agency yourself to make the most out of AI

[umbral depot]

What do you mean I don't really understand English

[bright swift]

"In psychology and philosophy, agency is a person's capacity to act independently, make their own choices, and take control of their life. It is the psychological feeling of being the active author of your own actions rather than just a passive victim of circumstances."

[unreal parcel]

is codex still broken or do I have to still use a VPN to use what i"m playing $200 a month for

[bright swift]

i think still semi-broken, usage seems to go down fast

but i dont have connection problems

[unreal parcel]

I mean the freezing

[umbral depot]

I shouldn't ask for help?

[unreal parcel]

I have a problem where usage isn't going down 😂

[bright swift]

you need to learn how to figure things out yourself, even with a free chatgpt account you have the most powerful tools in existence to do that at your fingertips

doesnt mean dont ever ask for help

but dont ask for things that gpt can answer for you in a few seconds

[signal tapir]

I would say don't ask for things that needs multiple steps back and forth when gpt can help you step by step in a better way than humans can.

[orchid echo]

give me some

[gusty basalt]

lol Codex being able to reflect on images generated through the code it made is a really awesome feature

so much fun seeing it figure out what is going wrong based on the generated output XD

[olive tangle]

hello, anyone know when gpt pro reset? im ont $100 plan

[velvet wren]

go here -> https://chatgpt.com/codex/cloud/settings/analytics#usage

[olive tangle]

thank you !

but its all about codex, not gpt pro on chatgpt.com no ? or i missed something ?

[velvet wren]

if you're talking about plain ChatGPT then the plan resets each month

[unique spade]

you re out of gptpro messages?

curious how much it takes to get there. i never managed that lol

[olive tangle]

yeah, just received a message saying i hit the limit try again later

[unique spade]

yea doesn t seem to be documented clearly

[bright swift]

the pricing and usage limits are all such a mess

[olive tangle]

it says now limited until 29

[unique spade]

maybe it's on weekly basis then

[digital saddle]

@solid lake Fast mode no good.

[stable obsidian]

when planning frontend with chatgpt extended pro does it make codex give better results? or is it still bs regardless? did anyone try this?

[bright swift]

you can have gpt pro create an image of the UI and use that as reference for codex implementation

[stable obsidian]

ill give it a try

[unreal parcel]

6 minutes to answer "you're right"

[pallid elbow]

sup bros, i know its only been a day since i asked last, but just wanna check, not expecting a "yes" anyway, but, has the token eatage thing been sorted in codex?

[unreal parcel]

rename the model to gpt-snail-mode

[digital saddle]

It was verifying and double checking to give you a proper answer.

Appreciate the thoroughness or get out.

[spice kindle]

is codex really helpful in building something big?

[civic viper]

Are you using Medium or High reasoning when coding with 5.5?

[velvet wren]

I am using Medium with excellent results

[velvet wren]

yes it is

[spice kindle]

extra high

[spice kindle]

well what have you built so far , asking because i love to hear what ppl have made with ai

[civic viper]

Do you have Plus or Pro?

[velvet wren]

I have built a royalty free music streaming platform with an API and coded an iOS app to use that API. I have also built a desktop macOS app for creating images and videos using the larger image and video models which allows drag and drop of content to and from the app. Right now I am working on a game.

[unreal parcel]

😂

[spice kindle]

Go

[spice kindle]

what kind of game and are u using more ais because i heard codex is not really good at building the frontend

[spice kindle]

for best results shud i use high or medium?

[velvet wren]

a card game. I am only using Codex models

[spice kindle]

so which language and also u building it fully using codex or also coding urself

because i actually make things fully using codex

as i have no knowledge about coding

[velvet wren]

Swift is my main language

[spice kindle]

oh alr

[proper cosmos]

Is chatgpt like dying or something...
I regularly am using chatgpt while codex is working but I keep getting messages like this:

and codex keeps timing out here and there

[bright swift]

i think they still have a lot of infra problems over at openai

[velvet wren]

yeah, Codex is very slow right now

[proper cosmos]

do they not announce anything

[velvet wren]

there's been posts by Tibo and others on X

[unreal parcel]

it started 3 days ago and it's just slowly be getting worse and worse

it's really atrocious right now

[bright swift]

https://status.openai.com

[rocky fog]

like this?

[velvet wren]

yep and all the comment replies screaming about ratelimits and slowness

[rocky fog]

he didnt acknowledge the slowness though
or I dont see it

[bright swift]

probably busy building more pets into the app or other stupid features

[velvet wren]

nobody from OpenAI has acknowledged the slowness

[vital hill]

codex seems fine to me

[lethal pike]

dear gods above in openai heaven, let the 30th come sooner for reset 🙏

[velvet wren]

pets is really useful for me with multiple workspaces on my Mac, I can keep an eye on what Codex is doing even if I am outside it's workspace

[unreal parcel]

this is insane

[unreal parcel]

where are you located so I can VPN to a place that isn't completely degraded

[vital hill]

california

[velvet wren]

Maldon, Victoria, Australia

[unreal parcel]

damn I already tried LAX server

[vital hill]

northern cali

[unreal parcel]

it was working fine yesterday and today it's just super slow

[unreal parcel]

probably still just connecting to LAX

you can run this to check (colo=)

curl.exe -s https://chatgpt.com/cdn-cgi/trace

[rocky fog]

pretty weird that they take on every smaller issue and reset after, but now there is tons of people mentioning this slowness for few days now and they are just not reacting

[vital hill]

i use codex 5.3 mostly, haven't used 5.5 in months

[unreal parcel]

an acknowledgement would go a long way, I guess they're playing games with competition and want to wait to release gpt 5.6 or whatveer

well nobody has used 5.5 for months given it released last month

[trim rapids]

it's not slow for me currently, connected to ORD

[unreal parcel]

I'd rather just switch over to claude than use 5.3

what is ORD

[trim rapids]

not even sure where ORD is; I live in northern california but my ISP's routing is completely f'd

[lethal pike]

chicago 😛

[unreal parcel]

oh you mean colo=

[bright swift]

eu works reasonably ok for me, at least no timeouts

[trim rapids]

is ORD chicago? that would make sense since my f'd routing actually routes everything through chicago

[bright swift]

yes. https://github.com/Netrvin/cloudflare-colo-list/blob/main/locations.json

[trim rapids]

like half a decade ago there was this huge wildfire here that burned up all their network infrastructure and when they got it back up and running everything was routing through chicago even though it's like 2000 miles away

they never fixed it no matter how much I complained

I ended up just buying a subscription to exitlag because they have some servers I route correctly to

[unreal parcel]

#codex-discussions message

[bright swift]

i mean maybe its your isp to cloudflare routing. try cloudflare warp maybe

[velvet wren]

me right now

[trim rapids]

it's my ISP's routing to almost everything, the packets stay in their internal network until chicago

most VPN services are laggy as hell for me too because the packets still go through chicago before they get to the VPN endpoint

[bright swift]

starlink

just bypass legacy infra

[trim rapids]

what's really messed up is it's gigabit internet and when I use exitlag I ping like 10 to everything

moderate downside to exitlag is that their client only allows specific whitelisted programs (mostly games) so people can't use it for bittorrent etc and tank service quality for other users

...but when my ping only actually matters to me in a browser and in games, it's fine 😄

[vital hill]

this is new 👀

[proper cosmos]

How'd you get this, I wanna have a side chick in codex doing my cleaning

[vital hill]

idk, i just woke up

you just need to wake up and it'll be there

[vital hill]

side panel too

[velvet wren]

I just bought credits

[hard drum]

i refuse to

creds get drained too fast probably

[velvet wren]

it's not that bad really

[hard drum]

i'm not made of money

[boreal holly]

It depends on the plan. I mean the credits drain at the same rate on all plans, but the amount of quota per dollar on Plus and Pro 5x is roughly equivalent to credits. For Pro 20x, it legitimately is less value to use credits

[bright swift]

credits = API pricing

[rocky fog]

https://developers.openai.com/codex/pricing?codex-usage-limits=pro-20x&codex-credit-costs=plus-pro-enterprise-edu-legacy#how-do-credits-work

[boreal holly]

Proven mathematically, credits == API Pricing

But in terms of value, API Pricing is slightly more valuable because you get access to more models, more modalities, no expiration on API wallet, etc.

[bright swift]

you also get better routing with API

if you can afford it, just fast mode 5.5 xhigh API yolo it

[spice kindle]

well is it normal to build things fully with codex?

cuz am a vibe coder

[bright swift]

you can build everything with codex

[velvet wren]

you can build a lot with Codex, it helps to know the language you are coding in because Codex can make mistakes

[bright swift]

the more you understand, the better your results will be

[spice kindle]

its just that vibe coders are hated idk why

[bright swift]

who cares about what future unemployed people think

vibe coding means you dont care about the output at all

thats what i mean. let codex do the work but also understand what its doing

[spice kindle]

but honestly codex is really good at fixing stuff if it makes any mistakes it fixes itself

[bright swift]

it can teach you stuff while it does it

[lethal pike]

like my dad? 🙂

[spice kindle]

well i do try to understand and find errors and give it to codex to fix it

and i do care about the output

its just i have decided to build stuff w ai because i cant seem to get the concept of certain languages in my head

[boreal holly]

Drop

## Sensei mode

Explain what you're doing and why so the user can learn SWE

In ~/.codex/AGENTS.md and your agent will teach you stuff

[proper cosmos]

Thats the spirit "future unemployed people" thank you

what is SWE

[boreal holly]

Software Engineering

[proper cosmos]

oh

I failed that IQ test

[boreal holly]

All good! Saves a few tokens, and codex understands what it means 😅

[vital hill]

all good man, you're all for the vibes

[spice kindle]

how do i put this into codex ?

like where exactly

[proper cosmos]

I guess at the top of whatever you're doing

[bright swift]

if you truely want to save tokens you can write your prompts as formal logic hehe

[boreal holly]

If you're on macOS or Linux, run

cat << EOF
## Sensei mode

Explain what you're doing and why so the user can learn SWE
EOF >> ~/.codex/AGENTS.md

If you're on Windows, try configuring a MCP server, which will open config.toml, and in the same folder as that file create a file called AGENTS.md and put the directive in there

[velvet wren]

finally!

[spice kindle]

okay

[vital hill]

if you're using the codex desktop app, you can put it in settings

[orchid echo]

in your AGENTS.md file in ur Codex directory

[spice kindle]

done

[sacred minnow]

oof

[velvet wren]

that is just insane 🙃

[sacred minnow]

i love goal lol

[vital hill]

is /goal the new yolo mode

[boreal holly]

Now when you fire up a new thread, ask it to do "Sensei mode" 🙂 and if it knows exactly what you're talking about then the instructions loaded successfully

[sacred minnow]

sandbox o.o and yolo

[sacred minnow]

oof xD dang

[vital hill]

yeah sandboxed for sure

[spice kindle]

okay then thank you

[proper cosmos]

have you cured herpes now

[unreal parcel]

it's not that unusual with cached tokens and many turns

100 turns of a 180K context and you reach 18M tokens

[trim rapids]

that count actually doesn't include cache

[tawny island]

anyone car to share how you have codex sessions be able to use your apikeys? without it being mad about having printed it in chat and that you need to rotate your keys? tried 1 password cli but it asks for my fingerprint in every action.

looking into fixing that but anyone that found a perfect solution?

[velvet wren]

Codex is frozen for me right now

[hard drum]

can i be considered a vibe coder if i actually do care how the code looks, or where i want optimisations paths/profiling, structure && whatnot?

[torpid trout]

you mean use the api key to connect codex with openai, or api keys in general?
For the latter, you should use .env and add that to gitignores which generally is respected by codex (but no guarantee), for the first, I think you can add it to config.toml?

[hard drum]

not just "never look, never check output"

[bright swift]

direnv

[trim rapids]

imo that's the difference between "vibe coding" and agentic engineering

vibe coders never care how the code looks, because they don't know what any of it means anyway

[hard drum]

i can't understand majority of the complex code i have on first or two glances (i used to!)

[torpid trout]

As if anyone is actually reading the ultimately thousands of LoC's codex produces in any given project 😉
You would be spending more time reading than there are hours in a day

[hard drum]

but i still can dig out any regressions that AI may have missed on testing, though

i care that my stuff works as it should

&& that the codebase isn't a mental overload for others (but may be for some anyway)

lints, analyzers, all that stuff

i need to ensure my stuff is actually coherent

not sure if that fits into vibecoding

i honestly don't really know what vibecoding even means anymore

[trim rapids]

real vibe coders are like that guy at the jiffy lube who takes the impact wrench and goes dugga dugga to your oil pan's drain plug

[unreal parcel]

agentic engineering, said no one ever

[hard drum]

because according to a ragebait tweet of human art claimed as "AI art" showed how people jump to judgement && cold input the moment something is labelled AI/vibe

so why would anyone make up a term for other things?

[torpid trout]

At least these times are over despite vibe coders

// remember why
$this?->get()?->loader()?->contract?->handler()?->get()?->the_thing(
    a()?->b()->c(fn($x) => $x->d()?->e()),
    new Factory(new Manager(new Provider())),
    [...$stuff, ...$moreStuff],
    fn(&$e) => tap($e)->mutate()->persist(),
    &$e,
)?->finally()?->unless(false)?->dispatch();

[hard drum]

they put all in one broad category anyway

[hard drum]

what the hekc is this?

[torpid trout]

basically what produces mental overload lol

[hard drum]

is this a custom js template or what?

looks like js

[torpid trout]

Seen often in so called senior developer projects. Specially the // remember why part

[exotic cave]

My eyes!!! Damn you should have put a warning on it.

[bright swift]

thats php

as evidenced by the dollar sign

[hard drum]

what

i cannot even read this

[tawny island]

the latter. thanks! but thats what I have been doing but dont like it due to security and agents sometimes printing or reading it then they get not very smart because they get context polluted by thinking about it so much...

doing my first project i really need to think about security so trying to up it . :/

[tawny island]

tnx! same with that though

[torpid trout]

yeah, but "you have to understand your code" right?
and then //remember why
My take is, even a vibecoder produces often better results than all the so called studied señor developers who are often very afraid of AI
(heck, not saying a senior dev cant code, just that it is often not that far away from hallucinated ai slop)

more like "copied it from stack overflow and now //cannot remember why)

[hard drum]

ai was trained on human slop, after all

but humans have a fragile ego

so of course they'd attack whatever they "feel"

stupid

[torpid trout]

you could do a .env outside CWD but you have NO guarantee it wont be read

[tawny island]

i'll give this a go probably, tnx!

[hard drum]

i think the age of Agentic coding really starts to show egotistic/narcissistic people out the zoo

separates people who just want to do stuff from actual lunatics/delulus

[torpid trout]

generally, I suggest just use a temp dev api key for what you consider "compromisable" that you rotate often'

[torpid trout]

me likes delulu, that word has meaning.

[tawny island]

yes!

[trim rapids]

what if you consumed the .env file through a setuid binary owned by an additional unprivileged user and restricted the permissions on the .env so only that user could read it

then codex could run that binary and that binary could read the .env but codex itself couldn't read it

maybe

[torpid trout]

mac and permissions...

[bright swift]

my codex has root access through aws cli with all production secrets and has never read any of them, it always builds this weird shell scripts to avoid it, so idk what you are doing 🙂

[torpid trout]

this might work on linux tho

[torpid trout]

how do you know it never read any of them?

Its not that it would tell you lol

[bright swift]

hooks

pre tool use specifically

[trim rapids]

it might work on a mac too, the filesystem has to at least support the setuid flag for stuff like sudo to work

[torpid trout]

"hey sorry I just accidentally read your xyz file"
hooks wont really help that do they?
Even if your hook registers every move it does

You still dont know exactly where it went?
Or you have hooks impeding read?

[tawny island]

for me it overfocuses on that it has read and "printed it in chat", just making the model stupid

[bright swift]

yes blocks reads etc

[tawny island]

oh

[torpid trout]

you mean specific file reads, I assume

such as said env

otherwise you would cripple codex to the level of absolute incapacity

[tawny island]

isnt it that they accidently print it in cli through weird commands? or can you block that too?

[bright swift]

i have a .envrc that loads the necessary secrets from aws secrets manager and exports them as env vars. codex doesnt need to do anything

[torpid trout]

well it still needs to read files lol

otherwise how is it going to make changes for you or answer a question you have about x in n file

[bright swift]

by blocking only access to .env files for example

but i moved on from that, i just have everything in aws secrets manager now, different accounts for prod/dev

i could make it safer by removing the prod profile for aws cli i guess but so far it hasnt deleted my prod db so..

yolo

[karmic gulch]

Been a gpt #1 on most things lately, but today I found a usecase Claude hands down did 10x better. Configuring my hermes every now and then, (alot last days) and Claudes insight into the system, ideas, how to, what not to, and everything vs GPT 5.5 xhigh is a mile between. First in several weeks I found gpt to be the lesser. But oh well, had to be something one of htese days lol. anyone with same exp?

[torpid trout]

wait until you see the ultimate guantanamo radio station chart winner:

((((($this->get->loader->contract->handler ??= new HandlerFactory())
    ->get($type ?? throw new Exception()))
    ->the_thing(
        a(b(c(d(e(f(g(h()))))))),
        match(true){default=>fn()=>yield from []},
        static fn &(&$e)=>$e,
        ...$args
    ))?->maybe()?->idk()))

(admittedly tho I think only gpt can come up with this type of nightmare when specifically asked to exxagarate lol)

But yeah, if you daddle a bit in PHP you will see lots of very similar looking code, logic and ... many persons implementing humans extending animals finally throwing babies or similar lol
I do love that language, but its messy nature invites for abuse.

[bitter fossil]

Codex going super slow for anyone else?

[torpid trout]

Many say so here

not on my end, yet

[torpid trout]

well, that is a complex question

[bitter fossil]

I questioned, due to it was "working" for 6+ minutes and seemed to be doing absolutely nothing lol

[torpid trout]

you cannot expect a straight forward answer
It might be seeking for what really is going on

Conspiracies and fact checks and all with that what comes

[bitter fossil]

Seems for the last hour Codex is having some issues

Guessing we will get a status change?

[torpid trout]

there are some users in here claiming the same issues since several days so you are apparently not alone, and no, there is no estimate nor confirmation of any actual bug or fix

I see some here saying with VPN it is solved

[boreal holly]

Good news, since you're using the TUI, downgrade to a previous version

[bitter fossil]

Any recommended version?

[boreal holly]

I would just do codex --version and install the version right before that 🙂 idk what features you use. If you mostly do vanilla codex with minimal features enabled you can do 0.125.0 which is the first one containing /goal. I use that version and have no problems.

[wooden minnow]

how many of you guys think that codex has been nerfed?

[vale ore]

i have to wait 5 days to reset, so idk anything about nerfing

[wooden minnow]

bro we are on API and it can't even handle normal FSDP training with vLLMs

[bitter fossil]

0.129.0 seems to be stable

[boreal holly]

A few folks really like that version! I have not tried it yet, but I hear that's the last version that supported hooks in a good way

[trim rapids]

what's wrong with the support for hooks in the current version?

they work for me

[boreal holly]

Have seen reports that hooks require approval or something

[bitter fossil]

Yeah, I believe it's the stable version for using Codex Mobile. I upgraded to the latest version last night, and today noticed a few bugs, then the issue with constant "thinking...", where it almost felt frozen

[bitter fossil]

Yeah, I kept getting denied permissions as well under the latest version, editing a file, viewing a file, etc all were failing permissions a bit ago

[trim rapids]

yeah, they go in the config file with a hash listed for the trusted version

for my project I just made the installer update the config file

[boreal holly]

Gotcha, I don't use hooks, just remember folks saying on startup it's like "these hooks must be reviewed before enabled" or something. But that installer thing sounds like the de facto solution!

[quasi summit]

codex seems soooo slow the last week or so...

on the $200 tier

[torpid trout]

where the hell are all projects gone on platform.openai.com?

There used to be a selector for projects, to add api keys, and so on

[plush plover]

How slow are you talking? I often wonder why some people get slow and others dont. Makes me think this isn't exactly region specific but more account based, maybe internal throttling

[quasi summit]

hard to give you a measure. just seems every task takes 2-3x as long now. wish I had a good way to measure it

[torpid trout]

you mean it depends on how slow somenone talks?

[quasi summit]

they onboarded a load of new users the last couple weeks

so maybe they are throttling

also the gui app is so laggy on my m1 max mac that it might be skewing my perception too

like 3-5 second lag on typing, switching sessions. insane

[plush plover]

I cant remember if it was claude code or codex but I read somewhere someone deleted a state file locally or some cache and it increased their speed by reportably 100x, so it could be something like that?

[quasi summit]

interesting

[ember flame]

if anyone is in highschool and in the bay area, and wants to join a hackathon dm, or if u wanna sponsor the hackathon also dm me

[signal tapir]

That's mighty vague in an international place like this

"The bay"

[quasi summit]

california

[plush plover]

insert [that-post-is-too-american-for-me-to-understand] meme

[vital hill]

thizzzz

[boreal holly]

San Francisco, Silicon Valley, the bay area 🙂 but the bay area is more than san fran. I think it includes oakland and surrounding cities

[signal tapir]

So USA? 😛

[torpid trout]

start by making sure you are not using memories

and if using the app, I never was able to use it for more than maybe 6 total chats, after that it goes brr and you need to start deleting chats otherwise it just becomes unusable, so test cli first to exclude app issues

[dense pier]

Yo if anyone need a website that gives 3k free credits for all claude models/gpt models/gemini models , if u use my link because i bought pro on it , dm me
IT ONLY HAS 3 REFFERALS BTW , ILL GIVE FOR FREE

[little swan]

Why does it say 'app unavailable' when I try to install the Stripe plugin on my Codex desktop app ?

[wanton iris]

Uh

I just installed the codex desktop app on my windows laptop and signed in and everything and i see no support for mobile anywhere

I tried opening it on the app to get it to prompt, tried restarting both. Checking for updates and everything.
I cant figure it out

[dense pier]

https://www.anything.com/invite/cp6vh7jd

Yo if anyone need a website that gives 3k free credits for all claude models/gpt models/gemini models , if u use my link because i bought pro on it , dm me
IT ONLY HAS 3 REFFERALS BTW , ILL GIVE FOR FREE

https://www.anything.com/invite/cp6vh7jd

[wanton iris]

Bloqd

You're too into that

Clearly a scam

[signal tapir]

Yup

[boreal holly]

Wowoweewa very nice deal

[wanton iris]

Robert do you use codex on mobile?

I seen you talkin a bit. Im getting frustrated lol

[boreal holly]

I use Temu Codex for mobile

What's going on with yours?

[boreal holly]

Oh yeah @lean lark posted the workaround a while back. Something about downgrading and running codex remote-something in a WSL terminal, and it lets you connect with Windows

Unless they updated it and added official support, but I hear the current solution is a workaround

[wanton iris]

Ah okay so they dont officially support windows then?

[lean lark]

That's no longer working.

[wanton iris]

Dang

[lean lark]

So Mobile > Windows is not supported yet and the workaround no longer works. We just need to be patient..

[wanton iris]

Oki

[signal tapir]

Should be trivial to write your own bridge, right?

[plush plover]

Its stuff like this which really makes me think if I made my platform it could be useful for others too

[lean lark]

Why bother when they might publish the official version today?

[signal tapir]

For fun only

[lean lark]

( Having said that ... I have a bridge in front of me that I wrote over the weekend. HAHA )

[boreal holly]

Bridge is what I did, but you can also just do codex app-server --listen ws://0.0.0.0:4200, then write a frontend that connects directly to the app-server over websockets

[quasi summit]

Thx I’m not using them

[quasi summit]

Oh no I have hundreds lol

[lean lark]

This isn't so much a bridge as a poor-man's OpenClaw, where I can use Codex from anywhere. It' does mobile too but that's just one client UI from my perspective.

It's like app-server but I'm using the SDK.

I think @torpid trout was doing something like this too.

[boreal holly]

I like the bridge method because you can customize how Codex works without changing the original code. I have a "onCompact" hook that tracks when an agent compacts, and after like 10 compactions it tells the agent "Time to execute a warm handoff", they write up a handoff initial prompt, archive themselves, and spawn a replacement worker to carry on. The bridge sequences that operation, and prevents agents from spoofing their identity and stuff

[lean lark]

"Time to execute a warm handoff"
Sweeeeet

As always, I think I'm about 6 months behind you in my evolution here... 🙂

[stable obsidian]

can they pls train 5.6 for better frontend capability 😭 how can claude deliver better frontend in just 30s that codex couldnt do in 8h

[boreal holly]

Have you tried giving it reference design images and ability to screenshot the real app?

[lean lark]

Some people/bots just have no sense UI aesthetic.

[stable obsidian]

I did yes, yet it still always manages to make it look template and slop

[signal tapir]

I find it pretty easy to get it to do good UI. All I have to do is let my artist set it up first. 😉

[lean lark]

See if you can use Claude (Sonnet) as an agent and tell Codex to ask for aesthetic guidance. 😉

[leaden hornet]

if i unistall codex and reinstall, will all my projects be gone?

[velvet wren]

on what platform?