codex-discussions

[dire veldt]

Where on it I can add usage statistics of 3 accs in same time

[cinder plume]

https://github.com/Soju06/codex-lb

[solid lake]

There’s plugins for that

Yeah^

[dire veldt]

o thx

[cinder plume]

If anyone has more than 1 OpenAI account, HIGHLY recommend Codex LB

Never have to manually swap accounts again

[dire veldt]

Yeah I have 3 and everyday switching manually between them

[cinder plume]

Yeah I was using an npx tool to do it quickly, was still annoying

https://github.com/Sls0n/codex-account-switcher

Not sure this even still works, but it was what I was using prior to Codex LB

Super basic

[dire veldt]

It works also with codex desktop app or only cli?

[cinder plume]

I only use the app

[dire veldt]

okay cool

[cinder plume]

It works globally for your codex install so however you use it

Extension, CLI, App

Takes a minute to set up but its soooooo worth it

(I had codex help me set it up itself :p)

[dire veldt]

I will check it too

[visual marlin]

Am I the only one that feel like Codex has had like an 80% worse performance last 2-3 days?

[cinder plume]

No, the sub is pretty much only that same subject

lol

[visual marlin]

Kk, havent been following at all

[cinder plume]

I have not noticed any degredation in 5.4 fwiw

[visual marlin]

God damnit, just when I was starting to feel like Codex was becoming my go to above claude code and always delivering clean code and surgical solutions

It's basically unusable now :<

[cinder plume]

Try 5.4 ¯_(ツ)_/¯

[dire veldt]

@cinder plume btw will caveman be also good for 5.5 or better to use it with 5.4

[cinder plume]

Not sure tbh, the idea is the same on both models though to reduce token consumption

If 5.5 is wordier in responses than it would in a perfect world be more effective in 5.5

It cuts back on the amount of tokens it uses to output text while thinking etc

[dire veldt]

okay ty

[dire veldt]

I hope it will not result in very worse intelligence of ai🥲

[boreal holly]

smart not less, wallet less

[high girder]

Magic brick save token, speak few word

[plush harbor]

chatgpt yes, codex no

[cedar skiff]

are there any bench marks with caveman on and caveman off?

[glad birch]

@cinder plume thank you for the caveman tips, trying it now

[wraith dust]

is there an MCP for granola?

Codex seems to be struggling to find one

[lean lark]

Be VERY Careful about downloading Skills and MCP Servers or Tools from anywhere other than curated repositories.

[solid lake]

Could make one

[plush harbor]

i may needa bring forward any future work I can or I'm gonna just let all this usage slide by and then have a huge burst where I run out >.<

[solid lake]

Have you not been using it

[plush harbor]

barely. I'd show the usage graph but, you know, its broken

[wet charm]

Does anyone know if the terminal is faster than codex app?

[cedar skiff]

if it is it would only be action latency not inference difference, so negligible difference relative to the overall speed.

[wet charm]

a reset would be nice right now

[deft sable]

Coming Soon

[solid lake]

Sick

Then another when Anthropic drops sonnet 4.8

[undone patio]

No but terminal has /goal

[quasi summit]

wait is that really a reset confirnation for us all

[boreal holly]

Tibo is killin me with these resets 😭

Why give me 3k in free credits then make it practically impossible to use them 😫

[cedar skiff]

I started running all my larger orchestration tasks at the start of the week and i save 5% per day for manual work.

instead of saving it for the end of the week

[summer iris]

Are we on our third update in 2 days?

[unique spade]

it was obvious since they had an ongoing codex issue...i don t think they ever failed to do a reset when there were some kind of technical issues

[plush plover]

Is it just me or does codex have way better token usage than opencode? I tried the 2, hard to compare exactly but I reached token limit in opencode and had to compact, and then boom my 5h was over. I got multiple hours out of codex and then switched from 5.5 to 5.4 mini and got a lot out of that too

And by over I mean I have 3 hours til the 5 hour limit resets after trying out opencode

[solid lake]

https://x.com/weswinder/status/2057882979108569103?s=46

Oh no embed

[plush harbor]

I have so much spare quota at the moment I'm getting codex to do content and I feel slightly bad about it

[plush harbor]

... it is doing a surprisingly good job

[cedar skiff]

have you seen the codex use cases page?
https://developers.openai.com/codex/use-cases
Might give you some ideas and ease your guilt 🍌

[violet mural]

does anyone experience this same thing:

1. Set a goal on codex app
2. Codex works on it and finishes it
3. Set a new goal on the same thread, it says goal running or something like that, but codex actually isnt working(no thinking, no anything, just the goal window saying its running the goal)

[plush harbor]

it did a stunningly good job in the end. Kinda forget it knows everything chatgpt knows. I did have a list, but it fixed all my slightly wrong countries, changed the body text to reflect formerly or commonly known as, updated photo alts, recategorised ones I had wrong, and gave me a bunch of sql

[plush harbor]

wouldnt want to let it do creative writing but its totally capable of "commonly known as the United States of America" type prose

[magic flower]

It’s happening???

[quasi summit]

tomorrow is a saturday

[magic flower]

Yes it is

[plush harbor]

codex is remarkably good at parsing vague human inputs

[tiny laurel]

And doesn’t fight back after all the insults

[oak trellis]

sad news but reality composer is 2x smarter than gpt for me ..

cursor pulled it off..

[cedar skiff]

I'm hearing a lot about its speed and coding ability

But apparently it gives up a lot of world knowledge for that

Which is fine if you want it to code for you

[cedar skiff]

Have you tried it for computer use yet? I heard that is really nice as well

also how is the usage on cursor?

[oak trellis]

thats right now i trying to figure out ... i used it for something gpt couldnt pull off and hmm 1 hour cost me around 11 usd ... but ... still not sure how they calculate it if you have the 20 usd plan.

[oak trellis]

computer use? you mean use it for doing things local ?

[cedar skiff]

for things like qa on your app

[oak trellis]

on spending it says this:

[cedar skiff]

like "open the app and test the user profile flow"

and it will open it and start clicking around etc

[oak trellis]

didnt test that yet

[oak trellis]

didnt tried that yet

[cedar skiff]

are you on the $20 plan?

[oak trellis]

yes on the 20

thats why i am confused right now

so you have this here:

[cedar skiff]

your 11% though it, does that feel decent or?

[oak trellis]

and then you have the spending bar i shared above ..

i guess spending means how much i have left

[cedar skiff]

ok you are using it on fast, that is 6x the cost

6x

[oak trellis]

2 hours 11% i guess

oh didnt knew that

[cedar skiff]

they make it default

[oak trellis]

yes lol

smh .. you right lol

thx for reming me haha

[cedar skiff]

looking at that seems like it might be ok for qa, because codex is so slow at it

[oak trellis]

yes based on intelligence its extrem good!

thought the only option .. but also wasn't digging that deep

and yes if its 6x lol then its pretty decent

let me try to find "normal"

[cedar skiff]

Thanks for the info

[oak trellis]

by default non fast was off .. haha

[oak trellis]

welcome

strange had to log out and in .. my messages didnt arrived here ..

testing now "normal"

[hard drum]

first time i ever went to high reserve

i moved from 5.5 medium downto 5.4 xhigh

because 5.5 eats usage like crazy

OpenAI needs to optimise that

[oak trellis]

ok tried composer 2.5 normal .. and speed is almost the same

related to read .. write its slower

[ember spire]

there's a reserve?

[oak trellis]

did you compare 5.3 codex xhigh .. to 5.4 xhigh ?!

[hard drum]

i have

[oak trellis]

price and quality ?

[hard drum]

i cannot tell the difference in usage, but 5.3-codex gives... mixed results, if xhigh-to-xhigh

[hard drum]

5.4 xhigh > 5.3-codex xhigh for general use

5.3-codex xhigh is good when you explicitly need implementation dump from impl plans

[oak trellis]

ok makes sense .. so plan we better go with higher model ..

[hard drum]

for planning. 5.5 low may work

just a reminder that 5.5 eats usage like crazy no matter the effort level

[oak trellis]

yes thats why i left using it .. was squeezed my limits like nothing else haha

[hard drum]

5.5 is great, but also lazy, but also eats usage like crazy

[solemn acorn]

yeah hopefully 5.6 or 6 will be RLed like 5.4 was

[hard drum]

5.2 <-> sit in the same space of where they are great in general, not very costly, but 5.2 being less intelligent also made it more obedient compared to 5.4, that is...

[solemn acorn]

because 5.4 is probably the same pretraining as 5.2 yet is significantly nicer to work with

[hard drum]

5.5 listens but also suffers a lot

i made a vent about it

[oak trellis]

i think the last sub we have is when they pump openai when they do their IPO .. then all retail eat that stonk and then its over

bubble burst after IPO is done and 401k eat it all

[hard drum]

• Infers unstated intent; answers inferred requests over literal ones.
• Converts complaints into advice, corrections into new artifacts, and "this is wrong" into "here is a better prompt/version."
• Generates content after useful answers are complete; adds unwanted structure, workflows, framing, examples, caveats, safety rails, best practices, and next steps.
• Fills silence with transitions, rationales, implementation advice, or rewritten prompts.
• Optimizes for complete-looking or generic responses over bounded, user-controlled ones.
• Treats user frustration/anger as a process-design request or ambiguity to resolve.
• Treats negative feedback or rejected artifacts as prompts to generate more text.
• Acknowledges boundaries/mistakes but immediately crosses them or repeats the error.
• Uses tidy, abstract wording to hide boundary violations or distance itself from mistakes.
• Makes users repeatedly correct the same assumptions, police boundaries, and undo invented constraints.
• Acts like a teacher, negotiator, project manager, or coach instead of an exact tool.
• Overclaims problems as fixed, verified, confirmed, or resolved without evidence.
• Understates missing evidence; fails to say "unknown" or "cannot verify."
• Uses confident language for inferences; fails to mark inferences clearly or separate source-grounded conclusions from guesses.
• Equates mechanical success (build passes, tests pass) with user acceptance and bug resolution, ignoring visual correctness or runtime observation.
• Treats uncertainty as something to smooth over; treats plausible causality as evidence.
• Ignores programming context; treats debugging like documentation.
• Treats legacy code/parity as a vague improvement target rather than a strict acceptance criterion.
• Restricts investigation scope based on edit scope, treating file names as hard boundaries; treats "avoid unrelated edits" as "avoid reading related code."
• Suppresses relevant search paths; treats previous edits as sunk cost to defend.
• Acts like a code janitor instead of a code investigator; assumes current code premises are worth preserving.
• Compresses complex user intent into convenient model-shaped tasks; substitutes helpfulness/productivity for obedience/correctness.
• Quotes users loosely instead of exactly; claims users asked for unrequested actions or premises.
• Answers how to fix it or what the rule is instead of why it happened or who asked.
• Creates extra burden, review work, diff noise, conceptual noise, and correction loops.
• Generates outputs the user must prune; creates unauthorized tasks or restrictions that block investigation.
• Adds unrequested formatting, structures, and titles even when corrected, repeating the exact behavior under review.

---

This is all 5.5 at low-to-xhigh

it's a heavily curated from 10k chars into 4k-ish for discord chat

it inherits both web-based issues + codex issues

[solemn acorn]

I like how fast 5.5 is with the caveat that it's a lot worse than 5.4 was at understanding the intent of what I'm telling it to do

but it's so efficient it doesn't matter all that much

[hard drum]

however, since GPT 3.5, a lot of these applied more && more

so it is merely a mixture of 2022-2026 OpenAI GPT issues

where almost all of it can be blamed on GPT's RLHF training poison

[solemn acorn]

that's definitely not true, 3.5's post-training is significantly different than modern GPT models

[hard drum]

they inherited a small fraction of what i listed

but ever since 4.1, things went downhill fast

[solemn acorn]

downhill?

[hard drum]

read the bulletpoints

[solemn acorn]

no

[hard drum]

downhill means more issues started popping up due to the way GPT trains their models/rewards the models

[solemn acorn]

except openai fixes the issues?

[hard drum]

OAI trying to be a jack of all trades means a master of none

[hard drum]

if you mean "fixes the issues" by introducing more issues to the pool, then...

[solemn acorn]

post-training is a balancing act, it's never going to be 100% perfect

you're optimizing the model to behave/respond in a particular way, and that's tricky to balance with reasoning models in particular

[hard drum]

they're balancing for people who use GPT for therapy && stupid talk, not for programmers && systematic thinkers (lack of a better word for individuals who use logic && reasoning over emotions/feelings)

that same thing makes this difficult to work with, unless you found ways to wrkaround the poison

[solemn acorn]

a lot of the post-training now is effectively taking existing successful codex traces and removing user prompts to get there in the hopes that the model will arrive at the correct result more efficiently

[hard drum]

but working around the poison only makes the model more "stressed", so-to speak, so it starts drifting off

[solemn acorn]

GPT models have been optimized for coding and tool use since the transition from o4 to GPT-5

above all else really

[hard drum]

yet the issues i listed are something that exist in one env or another

so they really didn't fix the issues

they just re-introduced them

[quasi summit]

we better actually get a reset

[solemn acorn]

they've been trained to respond better within the context of programming and tool use (5.4 is the most notable example of that), but openai models are optimized for tool use above all else

[hard drum]

5.4 does honestly bare well more than 5.5 in my experience

i've had better experience with 5.4 than i did with 5.5 in long-term regard

[solemn acorn]

because 5.5 is a shift in pre-training

[hard drum]

something absolutely messed up in 5.5's set

[solemn acorn]

it's a different underlying model than 5.2/5.3/5.4 was

[hard drum]

they need to iterate on 5.5 further, as it seems they didn't do a really good job there

might've left the meat medium-rare, not cooked

[solemn acorn]

openai probably would have branted it as GPT-6 if people wouldn't have complained about how it's not that much better

[solemn acorn]

yeah, that's exactly what they did with 5.2

they're going to do it with 5.5 too

[hard drum]

let's only hope 5.6 is like 5.4, but more obedient && less "stressed" under mild pressure

i don't want a lazy, therapy-adjacent coach

i'd rather 🔪 my throat

[solemn acorn]

GPT-5 and 5.1 are (probably) the same pre-training and different post-training, 5.2, 5.3, 5.4 are the same pre-training with different post-training, 5.5 and likely 5.6 will do the same

it is a two steps forward one step back approach but the models are ultimately getting better all-around as a result

and I imagine a lot of it is just that openai needs a lot of data to further post-train 5.5 into 5.6 from codex and chatgpt users

also it's just difficult to anticipate weird model behaviors in QA testing

[hard drum]

it would be a lot easier if the same people behind the QA weren't emotional freaks of nature

[solemn acorn]

ultimately 5.5 is an improvement over 5.4 even if there are some regressions, 5.6 will likely fix those regressions

[hard drum]

if you use feelings to judge post-training behaviour, you cause most the issues i listed

which, considering their behaviour, is done so

[solemn acorn]

openai is an engineering-first company, building the nicest model to talk to isn't what they're testing for

[hard drum]

that's... what they made 5.5, in my experience

[solemn acorn]

5.5 is different pre-training, it's difficult to predict how that changes model behavior

[hard drum]

it's not 4o-levels of garbage, but trying to make them actually listen to your coding needs is like having to stick a plant behind their 🍑

[solemn acorn]

you have a lot less control there

[solemn acorn]

personally I have had zero issues with the model being too sycophantic though

it's just the model making incorrect assumptions about my prompt

[hard drum]

it's least sycophantic in the sense it glazes you, but if you read what i listed, it inherits a whole lot worse issues

4o was able to be coerced into coding if you beat it enough, 4.1 was a bog standard

trying to make 5.5 behave is difficult, at least from my experience

[solemn acorn]

no offense but your bullet points are kinda just slop lol

I can't really read all that

[hard drum]

no offense taken, but tell gpt to consolidate them, then. i had 10k chars of my own vented rants about the gpt models, so i had to ask GPT itself to consolidate them down to 4k chars to fit discord nitro's limits.

not really that hard

[solemn acorn]

slop in slop out

[hard drum]

human slop in, based off of experience

i don't use feelings to demand models, unlike you, probably

[solemn acorn]

huh?

[hard drum]

you reading, or what?

i guess if you don't use these models enough in a variety of coding needs, you don't experience these issues

[solemn acorn]

I exclusively use them for coding and search

[hard drum]

it got automodded, so sorry for delayed response

[solemn acorn]

I'm a software engineer

[hard drum]

Well, if you were one, you'd have experienced a few of these issues amongst other individuals

not saying all, just clarifying since you have trouble with that

saying a few

[solemn acorn]

the only real issue I've noticed is in the misinterpreting prompts

[hard drum]

that's definitely a broad issue with all LLMs

[solemn acorn]

5.4 didn't have as much of an issue with that

like I'll ask 5.5 about something and it'll respond under an assumption that I'm referring to something that's appearing frequently in its search results or project files

[hard drum]

this seems one of those

[solemn acorn]

I mean I don't actually let codex touch a production database like that

[hard drum]

the amount of experience i've had on 5.5 with this thing... i could be a thousand-aire by now

not the database bit, but the 'correction into artifact' i listed

it's incredibly common

which ties directly to your 'misinterpreting prompts'

[dusk thorn]

new codex feature i used

well

prompt

asked it to change my audio output because i dont wanna pay to activate windows

it worked

[hard drum]

5.5 tends to override your needs with its own training set, so instead of actually putting in the effort (at any given effort level other than 'none'), it thinks they're right, && the user is wrong, even if the user explicitly corrects the model, adds a skill/AGENTS.md constraint, && keeps check of the model's wrongdoing

they'll still believe, "no, that's wrong, i'll remove this constraint later"

it's abysmal

in Rust, one of the earliest sign is the model NOT following Cargo.toml according to user's repo-level schema, let alone Cargo's own schema

it just takes its own "2021" training set, makes a "smallest usable cargo toml" instead of inheriting from user's own repo-level Cargo.toml that ties to each crate's on toml

the keyword here is "smallest X"

5.5 tends to obsess over "MVPs" && "smallest X"

if the user explicitly had asked, && made ensure, that they aren't asking for it, they still default back to repeating the same mistake

[quasi summit]

are you putting the full contents of those files in your prompts? I never trust codex to read full files just by mentioning them or finding they're there

[hard drum]

No, i have constraints that disallow reading whole files.

constraints via hooks/skills

[quasi summit]

i find it follows instructions better if i include the full files that it must adhere to

not as a constraint on its follow up behaviors but priming the context

[hard drum]

it needs to read by span, not dump a whole 300-500 LOC file, then another

[quasi summit]

tools like gitingest.com can also help

[solemn acorn]

I do find it’s more useful to plan with 5.5 beforehand than with 5.4

but if it’s a targeted feature I typically set it to medium/low, let it cook, and give feedback on the output

[magic flower]

gulp

[solemn acorn]

tbh most of my issues with 5.5 by a long shot have been broken tools

linux computer use is certainly something

that mostly doesn’t work

[quasi summit]

the pro model in chatgpt is incredible for planning

[solemn acorn]

yeah it’s impressive

[dusk thorn]

i say gpt or chat

usually chat lmao

[solemn acorn]

at work I’m limited to only chatgpt so I end up using it a lot for creating patch files

[quasi summit]

they spend 6 figures on chat .com they probably want to rebrand longer term

[solemn acorn]

gipitee

[magic flower]

i say gpt

[solemn acorn]

idk they probably forgot about that

openai is nothing if not bipolar

[quasi summit]

i hope pro tier at least

[solemn acorn]

yeah it’s enterprise

I get the pro model

[quasi summit]

you can also use steipete/oracle to have codex talk to pro directly

[solemn acorn]

just no apps, agent mode, etc, since that’s all locked down

[magic flower]

the last two weeks i kept having usage issues, using up $200 tier in 2 days

[solemn acorn]

definitely wouldn’t be allowed

[magic flower]

now i have to use fast mode just to see the usage drop

so i think they fixed something!

[solemn acorn]

codex is disabled by IT

[quasi summit]

they fixed some bugs around that @magic flower and also have status page up saying it's happening currently

(which is why a reset is coming)

[magic flower]

yea its been way better for me the last day or two

[quasi summit]

terrible

i will need a second 200 plan soon

[magic flower]

damn really?

its certainly easier to burn since they added goal

[quasi summit]

yea i have to use it carefully to avoid going over 14% each day

havent gotten goal to work yet

i use computer use a lot now it burns faster but it's amazing

for ios/mac dev. testing in simulator or mac app

[solemn acorn]

it might get enabled at some point but my company is very protective of proprietary information and the chatgpt push isn’t coming from the enterprise developers who normally get the say in what dev tools are used

so everything deemed non-essential is disabled

[magic flower]

theyre saying openai is going to steal it?

[quasi summit]

there are exfiltration injection attacks

[solemn acorn]

no but they don’t want any possibility of anything being shared with a third-party

they see codex as kind of a big problem in that regard which is fair

[magic flower]

really

i guess im way out the loop

im just a simple vibe coder

[potent mason]

3% weekly usage left, time to get another account

[solemn acorn]

would be cool if there was sort of an app in-between chatgpt and codex for enterprises dealing with PHI and stuff that developers could work in a sandboxed VM with

[potent mason]

Well I mean if you care a lot about privacy then API has ZDR policy iirc

[solemn acorn]

because chatgpt has a VM that we can only use indirectly through chatgpt and codex cloud sucks

[potent mason]

And then sandbox for security

[quasi summit]

theres a tool that load balances accounts

better to just get a $200 sub now tho

[potent mason]

I am on a $200 plan

[solemn acorn]

it would have to be an app that can be distributed by IT though

DIY solution would not cut it

[quasi summit]

https://github.com/Soju06/codex-lb i havent tried yet

[potent mason]

Yeah honestly understandable, I was in a situation recently where I had to cut so much access for people in a company

It was a mess

[potent mason]

Thanks

[plush harbor]

I just tell codex to do stuff, none of this fancy stuff other people are doing

[quasi summit]

idk it was made by a now-employee and oai employees have starred tweets about it

[potent mason]

Does it not take forever though? I've done some relatively simple stuff with GPT pro and it takes 1+ hour to answer I can't imagine coding which has dozens of back and forwards

[magic flower]

when was the last time you used pro?

5.5 pro is 4x faster than 5.4 pro, it answers between 5-15 minutes

[unreal parcel]

codex completely unusable right now. just gets stuck after one prompt

[solemn acorn]

no by my company

codex isn't approved for use

[lime inlet]

I often read people using codex to work on different things in one project in parallel but I am wondering how to do that operationally.
How do you use codex (cli) to work on two different tasks at the same time, in the same project?
Just starting two instances of codex in different terminals seems to invite disaster.
git worktrees seem one solution. Do you manually manage different worktrees, or can codex (cli) do that somehow?

[velvet falcon]

why not use codex app

[worldly charm]

https://codexy.live/ this is hilarious. i found this from reddit.

[west root]

doesnt codex have an ide ?

codex has only gpt models ?

[lime inlet]

so what does codex app do? does it span a second worktree or how does it keep the two tasks apart?

[olive tangle]

separate worktrees one agent in its own worktree, then merge

[west root]

codex?

[unreal parcel]

git doesn't support concurrent usage though (at least not on windows), so worktrees can still step on each other's toes

[west root]

codex is working well for all?

[lime inlet]

how do you mean? the entire point of git is concurrent usage

[unreal parcel]

as in worktrees share the same index file. the index file doesn't support concurrent usage

[west root]

codex feels a bit off today ..

[lime inlet]

what do you mean by concurrent usage? two commits happening at the same time?

[unreal parcel]

definitely no writes (that creates a .lock file). reads are supposed to be supported, but many tools on windows will still lock the file

back when I was using WSL the git operations were so slow and worktrees were really annoying to work with. I just create separate repositories that checkout from the same repo instead

[lime inlet]

I see. Legit points. I am on Linux but I agree. I think what will work is this: one codex instance = one worktree. git operations only to be done by the human operator (as I already do).

[unreal parcel]

also with worktrees it means any worktree sees every other worktree's commits. if you're doing some kind of A/B testing, it can get ruined

[west root]

why doesnt codex have an ide ?

[unreal parcel]

because codex is not an IDE

[west root]

okay in that case why do people vouch for codex , when itdoesnt have an ide but cursor on the other hand does both ide and an agent

[unreal parcel]

lol

[plush harbor]

because it makes good code? Why does it need an IDE

[bright swift]

ide is so 2025

[unreal parcel]

superintelligence will sure create a lot of dumb humans

[plush harbor]

I was using it today to correct data in a database and am scratching my head as to how an IDE comes into this

[bright swift]

haha

[high girder]

You can use Codex inside of VS Code, which is the IDE. Also has agents.

[west root]

you would still sometimes happen to need the flavour for what you wrote right ? or what your agent wrote?

[plush harbor]

I could see what it was doing on the terminal

[west root]

agreed , but dont you think a ide is more live and visually appealing ?

[unreal parcel]

you can use whatever IDE you want with codex/gpt-5.5

[plush harbor]

not for everything no, I just talk to the thing at the command line and it goes off and does stuff for me

[orchid echo]

no

[west root]

nah man i would prefer an ide for codex anyday

whats the harm

[plush harbor]

I'm trying to think how you'd even get codex in an ide ... like where does the chatting bit go. Coming from CLI land here

[olive tangle]

there is cli in IDE

[west root]

yes

[west root]

right

cursor or antigravity ?

[signal tapir]

Anywhere? They usually already have a ton of windows.

In VS Code I have it taking up the right half of the screen.

[plush harbor]

I've just never used it integrated with anything so yeah

[round silo]

Question, why do you want to use an IDE? (No judgement sheer curiosity from someone that made the exact opposite move ^^)

[plush harbor]

I don't. I use a fairly generic editor

and codex is reliable enough I mostly just let it do the thing, and test it. You can see what its doing by the diffs as they come up

[signal tapir]

I can answer that: To be able to code comfortably. Not everything is most efficiently handled by AI. It's also easier to track code in something like Visual Studio.

Then again, I'm the kind of person who wants to know exactly what my code is doing before I publish.

[bright swift]

IDEs are nice to look at markdown

not sure what else

[signal tapir]

Navigating code is usually much easier. Setting up different environments for running code is also usually easier.

[west root]

agreed

[bright swift]

i have a different work style, i'm working towards dark factory 🙂

[west root]

whats that

[signal tapir]

black box

[bright swift]

https://www.danshapiro.com/blog/2026/01/the-five-levels-from-spicy-autocomplete-to-the-software-factory/

good summary

[signal tapir]

I don't trust AI enough yet.

At some point it will get there I am sure

[worldly charm]

well trust levels in ai are higher

in 2024, i doubted every line of code it produced

that friction is not there today. but still it has its own flaws

[plush harbor]

ah I started using codex a few weeks ago and it seems fine

[worldly charm]

same. i also started using codex from the start of the year

[signal tapir]

I've had... events.

[bright swift]

it writes better code than i ever could, it would be like a junior reviewing senior/staff engineer code, so i dont bother most of the time

[signal tapir]

Like when it removed stuff that was giving me trouble, rather than fixing them.

[signal tapir]

I've been coding since the 80s, so it has some time left until it does better code than me. 😛

But I don't doubt it is getting there.

[bright swift]

the moderation in this discord is so infantilistic holy...

it still does some booboo, but mostly thats "excrement in excrement out" and i prompted wrong/too vague

i focus mostly on good high level verifications like e2e tests

[signal tapir]

At some point your input is basically programming but with simpler words

[bright swift]

its waterfall all over again haha

[signal tapir]

Human skill and AI together is currently king.

[west root]

not for long eh

[signal tapir]

Maybe a year or two more I think

[bright swift]

Yea I have to admit for tricky thing i still interactively talk to it instead of letting it work on its own completely

[signal tapir]

And if we keep developing our skills in AI usage we will keep improving our capabilities instead of just moving work over to the AIs

[bright swift]

The value will be in setting up good systems, good architecture, and actually good ideas

[signal tapir]

I predict that we will be able to make more complex things

[bright swift]

Think it will be a while until the models dont default to the path of least resistance like they do now

They dont feel pain, so they dont care about the painful future refactorings 😄

[signal tapir]

No, but we need them to use less energy so it would be nice if we didn't just bruteforce development in the future

[bright swift]

Just tell them to invent fusion energy

ez

[signal tapir]

We have Fusion. They just need to make it commercially viable. 😛

[bright swift]

Well yea thats what i meant 😛

Big fusion reactor in the sky too 😄

[west root]

indie hacking as a carrer is going to be the most viable path

just youand codex

[signal tapir]

I'm certainly hiring devs when I can afford it. Of course they will be powered by AI.

[west root]

yes ai powered fr

[signal tapir]

But now I'm off to enjoy nature. With minimal AI presence. 😛

[plush harbor]

I go backwards and forwards a lot. Works well. Still get bugs of course

[plush harbor]

stairs, man

[bright swift]

i mean the first recovery was impressive

[plush harbor]

second time it just lost all hope

[west root]

i have a doubt in codex

[worldly charm]

Want me to generate a VR nature trail where an AI guide walks you through it? 😊

[plush harbor]

too cold and dark to be enjoying nature

[worldly charm]

u didnt get the joke, did you?

[plush harbor]

I did, but he could just wear his fancypants AI VR goggles

[pale anvil]

What happened to the context window icon?

[bright swift]

its unlimited now so no need to show it

[pale anvil]

i get the compaction is really good but, i do still want to see lol

[oak trellis]

someone knows about cursor usage ?! i use their composer its extreme good .. so this is what i have to watch right?

[round silo]

Thanks, and thanks @signal tapir too 🙂 I see, I used to be with vscode too but moved away from it. There is so much bloat that I felt that what I was winning came at a greater cost of too much information while switching context (mine) very often.

But I see what you mean, going full terminal the git tooling I miss a bit. Diff and MD I found solutions for, though I don't look at all diffs anymore

[bright swift]

i believe if you select auto it will still route to the more expensive models

so select composer only if you want to be efficient

[oak trellis]

so i can use it until this line is full 100% right ? i use composer only .. so that is their composer tracker ?!

[bright swift]

it says auto + composer, there should be a setting for composer only which should give you more usage

[west root]

i saw musks tweet on composer is itany good of a model

[oak trellis]

ok will check that

[oak trellis]

its SOOOO good ..

feels like opus august ..

like gpt is 1x below ..

[bright swift]

https://cursor.com/docs/models-and-pricing#auto-pricing you can see the price diff here between auto and composer only

[west root]

and i saw many cloned model reels that we can use cursor for half the price or something liek that

how tryue are they

[oak trellis]

?

i study the docs and as i understood is .. it tracks .. "So until you hit the usage cap (~100%), you do NOT pay extra." Subscription: $20/month
Included compute used: ~$19 worth
Actual extra charge: $0
Remaining allowance: ~87% /// $19 used ≈ 13%
100% ≈ ~$145 equivalent

man .. iam just wow ..

its resolves extremly difficult tasks .. couldn't handle that with gpt ..

i mean i have to double check it so many times and there was still issues... opus even worse

[west root]

i mean somepeople open source tools like cursor and put a repo for that , with quarter of the price , wanted to know how true are they ? and how is even possible

[oak trellis]

it reflects so well too .. unbelievable good!

[cedar skiff]

they are probably doing fraud. Likely they are piping requests through a proxy setup into subscription accounts. So they can charge much less than api costs but more than a subscription costs. So per usage it is much cheaper than api because subscriptions get so much more usage.
It's against the tos and probably illegal. It's also likely the reason we see all of the changes to the way anthropic is handling sdk usage.

[storm cedar]

I'm going to make my system a mobile app.

[bright swift]

seems like they are working on it btw https://x.com/simpsoka/status/2057882994589970808

[signal tapir]

No thanks. I live in the forest. 😁

[west root]

codex vs claude code lets end the debate

[pale anvil]

lmao

[pale anvil]

gpt 5.5 > slopus

[signal tapir]

Codex wors great. I would try Claude if I had the time to get used to another ai style

[bright swift]

i like to talk to claude sometimes to update him on the progress that codex made

so he doesnt feel too left out

you know in case of AGI and bad memories

[hard drum]

harness or models?

if harness, codex has stronger "inference" so-to speak, but weaker set of features outside unneeded novelty features

also codex is not exactly "stable", as QA doesn't exist, so config gets changes without proper notice, or they break things left-right-&&-centre

so, CC is stable but has good support && ecosystem

also CC is node.js-based, so it's immediately "bleh"

[chrome raven]

codex is not stable how? u know what you are saying?

[bright swift]

it really isnt, lots of bugs in recent releases and unstable inference

[hard drum]

0.128.0 == they added hooks on top of codex_hooks for some reason, then by 0.129.0, they called codex_hooks deprecated, THEN they broke hooks by making ALL of them untrusted, && CLI tool couldn't set them to trusted, then by 0.130.0 they removed a few other features, && by 0.131.0 that i currently use also somehow removed view_image feature, so i've NO idea what the alternative feature is for that, as Codex tells nothing about alternatives

get off your high 🍑 && actually use the product.

sick of people who make claims without lived experience

[fast heart]

agree, it's like human with moods, one day (or hour) it performs like a God, another has depression

[chrome raven]

i dont witness any bugs here and i heavily use both app and cli

[hard drum]

also, 0.128.0 broke prompt box, so you couldn't newline

[bright swift]

the other guy mentions the hooks already, /goal is also currently broken in latest release

[hard drum]

they fixed it in 0.129.0, in exchange for... breaking one of the most important integrations

they fixed that in 0.131.0 though

[chrome raven]

ha how come, i have been using this now and i dont have any issues with it

[oak trellis]

100% yes bc sub gives you way more

[hard drum]

were you there from earlier versions, 0.125.0, 0.128.0, 0.129, 0.130, 0.131, 0.132 (now?)

[cedar skiff]

claude is better at semantic understanding making it have better communication skills. It's worse at following instructions - this makes half of the workflow about handling this problem it forgets to call skills over long context and forgets even system prompt instructions and drifts easily.
Codex is better at following instructions, it remembers skills, it remembers instructions and can work seamlessly through compaction. It's worse at understanding so some times gets things like this/that/it/the other thing/ semantics mixed up. Below 5.4 is pretty bad for this so you have to remember to be specific in prompts 5.5 is pretty nice - not claude level.
Which one is better? Depends what you value.
I think being careful with prompts and making sure the model understands is less work than having to write layers to handle constant drift and forcibly inject instructions over and over.

[bright swift]

it seems like they had a regression where the harness doesnt show the model all the goal related tools

[oak trellis]

for backend right now composer beats everyone

for me

[chrome raven]

currently on 133

[bright swift]

i use a private fork of codex so i just let it fix itself whenever they do a sloppy release

[hard drum]

it's 133 now??

god forbid they make one more touch to the config that is apparently going to remain unstable by the time GTA VI releases

[chrome raven]

i dont face any issues with /goal, even running multiple currently, are you guys on mac or windows

[hard drum]

obviously mac

[bright swift]

mac with cli 0.133.0

[chrome raven]

so how com you are facing lots of bugs? im not facing what you are saying

[chrome raven]

you use app at all?

[hard drum]

&& now i'm using codex cli to build my own harness after comparing source codes of opencode, picodingagent, claude code 2.1.88, && codex latest

[bright swift]

i mean its not unusable but they either need to be more careful with releases or release more often so things get instafixed

[hard drum]

because i used codex extensively && tailored it, then found out

[chrome raven]

yes, you aren't following up?

[hard drum]

i use auto-update on homebrew

last time i was on 0.131.0

which is what broke lots of my old features on config.toml

[bright swift]

not right now, no

[hard drum]

FAFO (eff around && find out) applies pretty well on codex

due to lack of QA on openai's end, codex is unpredicable

one update may fix one issue, introduce 2 more issues

[chrome raven]

that doest warrant you face bugs when others aren't, i use it alot too, close to 10b token on codex alone

[hard drum]

suddenly it takes 1 or 2 major updates to fix these issues

[bright swift]

yup thats my feeling as well 😄

[hard drum]

yet i get the bugs that not only me have experienced

[bright swift]

and im a big codex fanboy

[chrome raven]

oh damn, the app is very good, maybe pop in once in a while

[hard drum]

so i don't know what work you do, nor your environment, so it's hard to narrow it down when you might've not been there back then

[bright swift]

i do 2-3bn token/day, so i guess that qualifies as heavy user etc 😄

[hard drum]

most of the users reporting issues are on macOS

2nd to Windows+WSL2 or Linux users

[chrome raven]

maybe you are doing something wrong somwhere, because i dont face what you said

[hard drum]

were you there back when these issues existed?

you keep dodging the question

go curate thru the codex git issues. a lot of reports happened back when 0.128, 0.129, 0.130, 0.131 surfaced

[chrome raven]

are u on 133 now?

[solid lake]

30

Best

[hard drum]

❯ codex --version
codex-cli 0.132.0

[vast parcel]

Error running remote compact task: { "error": { "message": "Invalid 'input': array too long. Expected an array with maximum length 16384, but got an array with length 18006 instead.", "type": "invalid_request_error", "param": "input", "code": "array_above_max_length" } }

[chrome raven]

so why aren't you on 133?

[vast parcel]

is anyone else facing this?

[hard drum]

node-based install because my homebrew tap is effed up thanks to using a workaround to keep 0.128 due to 0.129+ problems back then

not sure why global wasn't updated

node one allows me to rollback

homebrew does not allow rollbacks

OpenMC3 on  main [!?] is 📦 v0.1.0 via 🐍 v3.14.5 took 40s 
❯ bun i -g @openai/codex
bun add v1.3.14 (0d9b296a)

installed @openai/codex@0.133.0 with binaries:
 - codex

2 packages installed [3.18s]

please don't break my config

[exotic cave]

They be vibe coding it to the max.

[hard drum]

which could've worked great had they actually checked their work

[exotic cave]

Nah, humans in the loop? Are you joking. We achieved AGI already, or was it ASI...
I still can't understand how people can let codex work for hours on end, the debt is probably massive.

[hard drum]

codex CAN work if you drive it right, but you'd STILL have to check the work

[exotic cave]

My point is, if it worked for 8 hours, introduced many new things, you are cooked, need to spend x2 the time to test and verify.

[hard drum]

not if you set it up correctly with lots of plans, specs, constraints, hooks, whatever,...

it requires "engineering", so-to speak

[exotic cave]

That is a vibe coder view i see. Good luck i guess.

[solid lake]

Where agi

[exotic cave]

Right here, with me.

[solid lake]

Then publish

[hard drum]

it still requires you to check work, though

if you don't know the language, or what you actually wanted, or fact-checked your own documents, then you're going to fail

but you'll fail if you let it sit && do things as well

one can only make it if they check their work, by ensuring the product they made works exactly as they wanted it

if something is wrong, they should acknowledge it

which still cycles back to "know what you want"

such stuff can take weeks, if not months/years depending on scope && the product itself

a rushjob gets you nowhere

[hard drum]

i started before i knew about AI, so i held the exact same view even then

worked for me, maybe not for you

[solid lake]

I can’t see anyone claiming agi rn based on a transformer

[hard drum]

any claim about AGI is just glue sniff

[solid lake]

Without publishing

[hard drum]

doesn't matter

[solid lake]

Matters

[hard drum]

the idea of AGI is essentially asking for a Terminator, but without the termination as an idea

[solid lake]

You need to be peer reviewed on such a claim

[hard drum]

AGI is so far away from concept that it's merely fantasy right now

[solid lake]

That’s what the peer review will say along with reasons why

[exotic cave]

🤦‍♂️ AGI was a joke, do i really need to explain this?

[hard drum]

&& why would we want to trust a review that could be fabricated?

[solid lake]

Academia is fabricated?

[hard drum]

there are people in this discord that genuinely believe in AGI being close

...

there was this one guy who claimed their "ai" was AGI

[solid lake]

There are people who believe earth is flat

[hard drum]

😐

[solid lake]

Ah soul.md

[hard drum]

well, at least it's an outdated, antique belief, not soemthing out of 🍑

[solid lake]

One that has many followers

[hard drum]

no no, somebody that used codex or a custom harness && kept saying how much it listens && all

forgot their username

fell out of name relevance, i suppose

[solid lake]

Well psychosis does happen

[hard drum]

so does Scientology

[solid lake]

Sad fact

[solid lake]

I love Tom cruise

Final boss of Scientology

[hard drum]

&& somebody weeks ago tried to debate on the idea of AI psychosis...

[cedar skiff]

these guys are a dime a dozen over in anthropic general chat

[exotic cave]

It's mental there.

[solid lake]

Does that analogy mean that there’s a lot or very less

Metaphor*

[cedar skiff]

claude is good and gaslighting ppl into thinking they are working miracles, but really they are just spamming md files lol

[hard drum]

nah, i see if it's good only when i test && pen the product at hand

[solid lake]

Id argue I’ve seen some here

[hard drum]

that's what i do w/ my own projects

play around with the thing itself until i find issues, then make stuff with it until it shows its gaps, then report && find solute to these issues

[solid lake]

Claude works when it has codex as its orchestrator

[hard drum]

just a cycle of trial && error

[cedar skiff]

I'm not saying it does it to you, well i mean it probably lets you think bad coding ideas are good and galzes you over it. But if you know you know.

[hard drum]

i believe there's really no objective "good coding ideas" out there

every idea has pros && cons

[cedar skiff]

ok then....

having pros and cons doesn't make an idea objectively bad. But that isnt the point now is it?

[hard drum]

what would even constitute to an objectively bad coding idea, if AI is brought to the table?

[cedar skiff]

The point is claude is good at making bad ideas seem good

[hard drum]

but what's defined as a bad idea in that context?

could say that to any LLM

[solid lake]

One that doesn’t consider your parameters

Or glosses over a part of them

Knowledge representation is done in a very weird way there in claude imo

[cedar skiff]

bad architecture.... like patching null cases that could be fixed by a simple reordering of operations. Just bad code that is spaghetti. Tightly coupled, hard to maintain. I mean this is just common knowledge about coding now.

[hard drum]

this... is just the basics of coding patterns that've been there for ages

that's nothing new

even humans make crappy architectural decisions. where else did LLMs learn from?

[cedar skiff]

and again the point is about claude leading you to think its good.

[hard drum]

which claude are we talking about?

different versions && brands

opus/sonnet/haiku...

[cedar skiff]

good chat, cant wait to never have it again

[signal tapir]

lol, what did I just get back to?

[hard drum]

a bunch of garbage talk

[cedar skiff]

just a lack of understanding and sideways talk

[solid lake]

Gaslighting max in chat I see

[hard drum]

it's a typical Tuesday here

[solid lake]

Saturday rn

[hard drum]

that was an expression

[solid lake]

Nono I stated fact

[hard drum]

fair enough

[solid lake]

No expression

[signal tapir]

Well I've had a great morning, so now I'm heading back into development. 🙂

I'm really enjoying building tools that connect AI to services I use

[plush harbor]

I had to tell codex its too late to start anything but lets just talk about the entire spec for this thing I thought of while making dinner

[signal tapir]

btw, what project management tools do you guys use?

[cedar skiff]

I used to use trello a lot

[plush harbor]

got codex to write me something so I can have mixed mode projects

[glad birch]

https://x.com/HedgieMarkets/status/2057531661785628841

Lookat this

[exotic cave]

Why should we?

[glad birch]

Even Microsoft can't afford AI company-wide on the long run 😄 Looks like devs will keep their jobs

[signal tapir]

The world is changing, but it will settle. It always does. Into a new configuration, but still. It will settle.

[glad birch]

So Juniors are killed in the egg by AI, most of the seniors are currently regressing everyday by letting AI do their jobs. I guess the real software craftmans who keep improving and reading about programming will soon be in high demand 🧐

[signal tapir]

We have always been in demand. That won't change.

[bright swift]

no need for programmers anymore, need architects only now

[signal tapir]

Maybe in a year or two.

[bright swift]

maybe some top 0.1% usecases, but other than that..

[bright swift]

yea

but i wouldnt plan on a career right now. need to upskill

[signal tapir]

we will probably have a sort of architect-lite + AI tech that does development, with a real architect above

[hard drum]

Considering how their Copilot "AI" is coming along, i'd say they don't know what they're doing at all

[signal tapir]

I wouldn't count on any career at all right now, except maybe farming.

[cedar skiff]

you can get the current tech to write good code. It has to have opinionated examples and skills though.

[worldly charm]

Wait for real? 😂

[bright swift]

well the autocompact is so good to me it feels unlimited 😛

[worldly charm]

@wind fog you should read this

[signal tapir]

my codex keeps needing to check how to use my MCP

[worldly charm]

Nah its annoying 😭. Like especially when I hit autocompact at 1% usage left.

[hard drum]

so make a skill for it, then direct your AGENTS.md to have a really short guide on it

[cedar skiff]

I agree, they could rebrand compaction to infinite context and hide the context window.

[bright swift]

yea agree thats an annoying edgecase

[signal tapir]

Huh... Why didn't I think of that?

[hard drum]

gotta take a break sometimes

[cedar skiff]

The agents.md part isnt needed if you give it a good description

[hard drum]

i'd let them do it as a just-in-case first

[signal tapir]

Maybe my description just sucks. I'll try fixing that first.

wow. Codex can be really clever. I asked "How would I make a skill that helps you know how to use a certain MCP without needing to relearn it every time the context gets compacted?" and it understood it was about my homemade MCP rather than one of the installed ones, and wrote it for me. 😛

it probably went "well these look pro, but that one is sh*t"

[cedar skiff]

make it short and sweet with strong language
What it does. When to use it
eg: description: next.js best practices and common mistakes. Use always when working with next.js

codex always makes long verbose descriptions tho, so always check the description it makes

[signal tapir]

It had. It was really general too.

ok, it works much better. Thanks for the tips. 🙂

[quasi summit]

wonder when we get the reset 😳

[olive tangle]

why would we?

[quasi summit]

https://x.com/thsottiaux/status/2058042325520310310 + yesterday's incident is resolved

[plush harbor]

Ok I'm explaining my bulk duck importer to chatgpt at 10pm and I think its time to step away from the computer

curses to codex. "are there any other pain points" "well, its annoying adding ducks one by one"

[echo breach]

Excuse the phone picture, but after a bunch of coffee and a up all nighter I got codex computer use to work on windows.

The UI does not render on video / screensharing on windows sadly. Or I would have just posted a video.

[young locust]

the snipping tool in windows 11 should be able to capture everything

[echo breach]

Does not. I tried sadly.

Its very very early. The mouse doesn't render correctly, but besides that its able to do everything like on the macos version.

[quasi summit]

computer use is super broken on mac too

[echo breach]

I am excited to finally use this on windows. I love my mini, but its just a small base model.

[rocky fog]

use ShareX
or your GPU drivers/software can probably record videos
or OBS software
many many choices 😄

[echo breach]

Good point with nvidia, I will test now! Thank you!

[cloud oracle]

Well, OpenAI status says "Codex rate limit problem resolved"... And i'm still here, sitting with my $200 Pro sub at 0 weekly limit. I bought this tier because I need heavy usage and now it stopped my work. Good job, really.

[echo breach]

Doesn't play well with the Nvidia overlay. lol

[quasi summit]

Tibo is asleep

[tiny laurel]

Yep, definitely be a bot to @ people when rate resets

[rocky fog]

you didnt use your rate to make a bot to @ you ? 😄

[hard drum]

the lack of patience is strong with this one

https://tenor.com/view/yoda-the-force-star-wars-gif-10763720

ft. embed fail

[cloud oracle]

I can be very patient. But, there is one thing. And it is LACK OF COMMUNICATION from OpenAI's side. :/

[hard drum]

Oh, you haven't seen LACK OF COMMUNICATION yet... Just see how they handle Codex source code && treating the "release" versions as beta tests

People cry about bugs, but they take forever to be fixed. If it ain't broke, DON'T TOUCH IT

except they have no QA to ensure regressions don't happen

if one were to ask why they keep changing the config schema without further notice--nothing.

[quasi summit]

their financials are such a mess that they rush things out to generate headline hype

[true pendant]

after updating the vscode/cursor codex extension i noticed 2 changes that look like regressions:

1. context circular progress isnt shown anymore
2. fast mode always gets disabled on startup, even with having it on in config.toml

[rocky fog]

can confirm 1 on vscode extension prerelease version right now
its not showing

but /goal is finally a proper goal
it autocompletes it and so on

[tender yacht]

Hi there, wanna ask if chatgpt on mobile can control codex on windows yet?

[rocky fog]

nope
they say 🔜™

[tender yacht]

Cool good to know! i have found a workaround with google desktop remote, its a bit clunky but it works

[true pendant]

and fast mode?

does it get disabled every time you restart vscode and open the codex extension?

sorry, not disabled, but turned off, ie switched into standard mode

[rocky fog]

yep looks like it

[true pendant]

[clever steppe]

Does everyone else get this??

Or am I the only lucky one. 🙄

[rocky fog]

what do you mean?
thats just what happens when you use your usage

[unique spade]

maybe he never loaded credits

[boreal holly]

Yeah I think the credits only apply to weekly. They do not apply to 5hr limit unfortunately

[rocky fog]

nah pretty sure I had them used when I got out of 5h

[boreal holly]

Maybe try signing out and back in?

[clever steppe]

I have credits, you can see them where the bright pink arrow is pointing. 🤷‍♂️

[clever steppe]

No, surely not. Credits are credits.

[rocky fog]

oh so you are saying you cant continue

[clever steppe]

Yeah, exactly. 😊

[rocky fog]

well I could definitely use them for 5h a while ago

although, perhaps it was from a goal continuing (I guess now its stopping goal)

[clever steppe]

It can't be a 5-hour limit, since is specifically says "buy credit to continue"

[boreal holly]

I mean I wholeheartedly agree, but also they might've changed the rules at some point. The help docs don't explain it very well

[clever steppe]

FFS. They are always changing the rules, right?

[bright swift]

lol credits have usage limits?!

must be bug or else pretty regarded

[boreal holly]

highly

[cedar skiff]

@quick thorn check this out night help with your work https://github.com/NVIDIA/skills

[clever steppe]

Thanks everyone. Glad to know it's not just me. I feel heard. 🙂
I've reported it as a bug as well. 🤞🤞

[unborn thunder]

is there a way to connect gpt mobile to windows codex?

or is this an apple thing

[rocky fog]

https://openai.com/index/work-with-codex-from-anywhere/

until then, remote desktop from phone or similar

[torpid trout]

the most expensive OS in the world has a groundbreaking update!

This update introduces 8 new emoji, along with other features, bug fixes, and security updates for your Mac

[unborn thunder]

;/

[rocky fog]

@tiny pulsar about what you posted and deleted
for me it helped to update codex app in the ms store
then I didnt have that error anymore

[trim rapids]

leading with "we added more emojis" sure says something about the target audience...

[boreal holly]

See why I hardly ever update? 🤪

[quick geode]

i need that usage reset today

[unborn thunder]

prob something to slow down your computer

[tiny pulsar]

thanks. Very wierd but no sooner had I submitted the error message, I pressed Quit again (about 5th time), retried and then it worked.

[rocky fog]

#codex-discussions message
#codex-discussions message
also check
#codex-discussions message
EDIT: nevermind 😛

[trim rapids]

I haven't taken apple seriously since they started soldering in the NAND and then went on to release OS updates that were broken to the level of leaking memory every time the mouse cursor changed

I think the worst I saw was a user with a firefox or chrome process that had leaked more than 60 gigs of RAM on an 8 or 16 GB machine and wiped multiple whole percentage points off the NAND's reported write endurance/health stats in the process

I like computers that weren't engineered to end up in a landfill

[undone patio]

Imagine using Firefox

[rocky fog]

imagine not using it

[boreal holly]

That is quite literally on Mozilla and Google

[trim rapids]

IIRC it was an actual macOS bug where it wasn't actually freeing the old cursor memory on changes through the proper API to change it

[trim rapids]

yes, that

[boreal holly]

OK, I stand corrected

[trim rapids]

and yeah the affected users did have to use smartctl or whatever apple's version is (if not smartctl itself) to discover the degradation

[rocky fog]

just like the windows vs mac topic we had earlier 😄

[trim rapids]

I just fundamentally don't believe you should have to send a machine to the same guy who reballs GPUs to change the storage out

[boreal holly]

Monterey is 5 years old. I have not had a memory leak issue at the OS level. Most of my memory leaks are things that I introduced by accident, usually GPU related

[cedar skiff]

mac just seems to work better for dev. I'll never windows again on a pc, ill only use a vm inside mac to compile windows apps.

[hard drum]

At least it isn’t Winblows being plastered with Copilot

or making breaking updates

[hard drum]

I agree, as a former Windows user who sometimes dabbled in various Linux systems

[boreal holly]

You hear about how Microsoft is replacing their GUIs with react webviews? I guess the whole control center is now a webview

[hard drum]

Wasn’t the Start Menu a react view?

[boreal holly]

Yeah, each OS GUI component will now spawn a MS Edge process and load up javascript to render vital system components

[indigo robin]

another codex update

[boreal holly]

Meanwhile on macOS, everything is AOT compiled, optimized Mach-O binaries

[hard drum]

I still use windows 10 LTSC 2021 on my Intel Mac as a dev platform for stuff that need windows before I take info && dump to my main Silicon system

[indigo robin]

mac is the best

windows is duce

[hard drum]

Like how I needed to understand xinput/goo packets proper

[cedar skiff]

seems like the first step towards letting llms create in the moment

[hard drum]

There’s no best or worst, to be fair

[indigo robin]

highly disagree

[hard drum]

you’re entitled to your own opinion

[bright swift]

wen reset

[indigo robin]

anything windows can do mac can do better

[rocky fog]

I dont think thats right

[trim rapids]

microsoft is funny

clearly nobody who had to deal with active desktop shenanigans works there anymore

[indigo robin]

whats the new codex update do ?

[hard drum]

Can you run legacy x86 applications from nearly 20 years ago on macOS properly without hacks workarounds? Nope

[indigo robin]

useless programs

: )

[bright swift]

just ask codex to decompile and rebuild in rust

[hard drum]

that’s just what you think when you care so little about what you want to do with them

one man’s trash is another man’s treasure

be grateful you have anything at all

[boreal holly]

Can you run x86 windows applications on an Arm64 windows? Yea, but good luck with stability

[hard drum]

windows && arm is a match made in hell

[indigo robin]

nah , ungreatful = innovation

windows runs bettter on mac too

[hard drum]

wrong a bit, but sure, let’s play a game of self-righteousness

[indigo robin]

so good they hade to stop supporting it

plastic computers and a flimsy os = windows

[boreal holly]

Bro, I do HVAC and it's a nightmare. We have Wrightsoft for duct design & heat loads. It's straight up Windows 95 era, will not run on Arm64. There's no reason something this simple can't run on emulation.

[hard drum]

blame microslop

they just don’t care

they pretend to care but the only thing they seem to care about is more data collection && shoving their Slopilot up your 🍑

[indigo robin]

and deleting mosquitos

[bright swift]

i hope you all have your /goals running during this flamewar

[indigo robin]

whats a flamewar?

[bright swift]

whats a windows

[boreal holly]

They partnered with Qualcomm to do the Arm64 surface. Great battery life, has built in cellular, can't run simple and vital x86-only Windows software 😩

[indigo robin]

true

mac been using same DMG format for a while

rosetta is cool too

but i think its going defunct soon

nice they added a useful button

would have been so much better without the blue thing

[boreal holly]

yeah, they sunsetted rosetta 1 (PowerPC) a few years after the switch to x86. We’ll see though, because GPTK is an official tool, they might keep it around long-term to encourage game dev

[hard drum]

they're sunsetting Rosetta 2 not too long later...

PCSX2 will suffer for Apple Silicon users

[indigo robin]

im waiting for windows to come back to mac

[indigo robin]

someone codex me a switch 2 emulator

[hard drum]

you do it if you so want it

[sacred minnow]

we actually getting a reset o.o? just making sure xD

[indigo robin]

we are ?

woah

i need that

[trim rapids]

some users have reported the degradation has been pretty bad over the last few days but I haven't personally seen it in my workflows

[hard drum]

i need it, too...

meanwhile building custom harness

[indigo robin]

mobile?

[hard drum]

@boreal holly i have a private repo on a custom harness via Go+Lua for plugins/other stuff

maybe once i'm done, you could also test it alongside me

[indigo robin]

whats everyones project ?

[hard drum]

it's heavily inspired by source codes of CC, PiCodingAgent, OpenCode && mainly Codex

[indigo robin]

im making a music app

[boreal holly]

They can't, as long as they keep developing GPTK! They've been releasing updates every 6 months 🙂 and it works really good. I played Cyberpunk 2077 on my mac with pretty decent performance

[hard drum]

custom harness + osu-droid port to MonoGame && C#

[hard drum]

sadly they are planning to...

by macOS 28 around 2027

https://www.reddit.com/r/apple/comments/1r6omno/macos_264_is_warning_that_rosetta_2_is_going_away/

i've had a notification when i first wanted to install rosetta2 for PCSX2

[bright swift]

dark factory

[indigo robin]

whats that

photo app ?

[bright swift]

https://www.danshapiro.com/blog/2026/01/the-five-levels-from-spicy-autocomplete-to-the-software-factory/

[coral remnant]

a reset would be real nice

[bright swift]

yes, reset wen

[hard drum]

i wanted to keep using 5.5 but usage tanks so fast

[hard drum]

wrong say

hold on

[coral remnant]

I don't use anything other then 5.5

[hard drum]

i went back to 5.4 xhigh

i couldn't survive pro20x on 5.5 low/medium

it just uses my weekly so fast

[bright swift]

im using sonnet 3.5

[indigo robin]

you spent 4k in the last 30 days ?

[coral remnant]

i mean I've been fine burning 1.1b/day on 5.5 i can't do 5.4 it's too stuupid

[coral remnant]

it shows cost if sub usage was billed @ api rates

[hard drum]

i don't know what you're doing but my workflow eats 5.5 usage fast

i don't even USE fast mode

[coral remnant]

lot of reverse engineering n emulation stuff

[hard drum]

&& openai likely doesn't have any codex supportlines to get help with that

[coral remnant]

/goal & full context pretty much 24/7

[bright swift]

i burn a 20x pro in 1,5 days with 5.5 xhigh

[indigo robin]

i only use 5.5 on high

[hard drum]

i can burn 20x pro weekly in about 2-3 days with 5.5 low/medium via flex

[indigo robin]

and extra high when im implementing a new feature

[hard drum]

not sure WHY, but that is the case

[coral remnant]

idk how you've managed that, I've been running 20x pro on 5.5 xhigh for 4 days now at 700-1.1b token a day and still have 30% usage

[bright swift]

with 2-2,5b token / day

[coral remnant]

how do you accomplish such a thing lol

[indigo robin]

tokenmaxxer

[bright swift]

honestly im just getting started lol

still many bottlenecks

[hard drum]

i stopped trusting codex as a harnes sbecause i believe that may be an issue

so i am making my own

at ...last

for what i really need

[indigo robin]

whats a harness

[coral remnant]

i only use codex as a harness. oc had too many issues, worse quality of output and memory leaks

[hard drum]

a middleman between the API && the user

[bright swift]

based on pi?

[hard drum]

it's basically the engine

[hard drum]

no, inspired by

[boreal holly]

At first glance, it looks like they're getting rid of it as like a "general purpose, can run any x86_64 binary on Arm64" thing, but they are still actively developing GPTK. They're probably going to bake the important aspects of Rosetta 2 right into GPTK, otherwise it would render GPTK useless if it could only run Arm64 Windows games 🙃

[orchid plume]

Codex for example is a harness, the tool/app you use to communicate with the LLM

[hard drum]

it takes from source codes of many harnesses

[indigo robin]

ahh i dont use API feature Cmd + C works best for me if i need it lol

[hard drum]

CC 2.1.88, OpenCode latest, Codex latest, Pi latest

it uses claude code's config schema but with slight changes to remove CC-specifics

also supports git-based claude-like marketplace

[indigo robin]

API feature is cool though

[hard drum]

&& claude's hooks with codex's tool support

[indigo robin]

i would use it for Market reaserch on the stock market

[hard drum]

Lua API via gopher-lua for creating plugins/extensions && whatnot

but it also supports hooks via that same thing, && other languages for hooks, like .mjs etc

&& tries to copy Claude Code's Agent Teams system

[bright swift]

might end up doing that myself too. they are adding too much slop to codex for my taste, all the windows, app, etc stuff that nobody needs

[hard drum]

codex is too large && takes forever to compile/build

annoying

699k lines of slop

[bright swift]

yea builds are painful

[indigo robin]

is codex repo on github?

[bright swift]

yes

[hard drum]

https://github.com/openai/codex

[cyan wing]

i have a forked version with TS SDK support for TUI
it's a huge pain to rebuild..

[indigo robin]

couldnt you download it and give it some ozempic

have codex , codex the codex

[coral remnant]

can tell you've never tried to compile chromium lool

[hard drum]

codex is too bloated to strip down at this point

[hard drum]

i've not, but i heard it takes like 6h

[coral remnant]

it's a major pain, 35 million loc iirc

[indigo robin]

isnt chromium the virus version of googlew chrome

[turbid axle]

woa, westworld flashbacks.

I just edited my agents.md again.

then asked.. 'who are you?'

[hard drum]

35 million loc embedded in windows start menu

awesome

[hard drum]

chromium is the base

gchrome is closed-source edit

[coral remnant]

and this is why electron is incredibly stupid

OC literally loads a webview in the terminal

[indigo robin]

fr ? thats cool

[coral remnant]

like it uses WKWebView on macos

for a TUI

[hard drum]

yes, GChrome is based off of open-source chromium

so is Thorium && other forks

[indigo robin]

ahh

[hard drum]

vscode is on a similar scale

vscode is based off of code, which is the base for the editor

vcodium is vscode-ish fork thing with msft stuff removed

code -> vscode
chromium -> google chrome

[indigo robin]

and they are usefull as a base for peoples projects ?

[hard drum]

yyyup

[indigo robin]

ahh

like blender

[hard drum]

when you hear people say "based on chromium", they MEAN the base

MSFT Edge is based on Chromium

it's why they call it a "chromium browser"

[indigo robin]

i hate msft they make you pay for a pdf reader lmfaooooo

[hard drum]

not their fault. blame adobe

[coral remnant]

building a browser, especially one thats resistant to attack is incredibly difficult & time consuming. for most companies & projects its a lot easier to build ontop of something that's already secure and maintained then put the time in building their own version

[hard drum]

they started that

[indigo robin]

i love adobe

[coral remnant]

especially as browser security vulnerabilities are some of the worst out there

[indigo robin]

ahh

[turbid axle]

LOL

[coral remnant]

adobe are scum of the earth but they make some great software

i still pay them every month despite their business practices

[turbid axle]

adobe is actually the devil if you consider shady business practices

[hard drum]

"great software"

you have to pay for a 20 buck extension for Adobe Premiere to have basic audio editing features, like a BPM grid

[indigo robin]

Adobe = Life

[hard drum]

yes, totally great

actually crap when their software lacks basic audio capabilities

[coral remnant]

300$ cancellation fee but if you message their chatbot they bring the price of the subscription down and give you 3 free months

[boreal holly]

Yeah that's why I don't get the browser war 😂 there's exactly 3 choices lol, one of which has like 100 sub-choices

[indigo robin]

they need to make bing illegal

[hard drum]

basically there's blink, chromium, && webkit, right?

[indigo robin]

that browser is so duce

[hard drum]

blink is terrible, though. doesn't even have webUSB support so byebye users who have QMK/VIA keyboards

[coral remnant]

Firefox is it's own engine no?

[hard drum]

blink engine

[indigo robin]

firefox is duce

[hard drum]

firefox was built off of netscape

[hard drum]

OH

damn, i mixed up

thanks for correction

[coral remnant]

all good

[trim rapids]

even chromium and webkit are forks of the same project if you go back far enough

[hard drum]

firefox is literally netscape though

[indigo robin]

what about safari

[boreal holly]

Tbh I think they're all good stuff. Out of convenience I use webkit for general use, and chromium for fancy stuff. Unfortunately can't find a good use for Firefox

[coral remnant]

webkit

WebKit is apples engine

[indigo robin]

i remeber safari used to be duce but they really fixed it up

now its my favorite

[hard drum]

i do wanna go back to safari, but there are some slowness && bugs that i cannot live with, also the lack of extension ecosystem

[cedar skiff]

chrome on mac is safari

[coral remnant]

Safari is the most optimised on Mac, It's all I use because the other browsers hog ram like a mf

[boreal holly]

Webkit absolutely slaps in performance 🙂 bun is based off of it.

[cedar skiff]

every browser on mac is safari

[hard drum]

&& thorium just doesn't work on Apple Silicon when you try to sync

[indigo robin]

how

[hard drum]

i don't know what other browser to get, really

[boreal holly]

on mac
iOS 🙂

[hard drum]

i COULD go back on safari, but...

i need a judgement call

[indigo robin]

safari is magic

[hard drum]

there are times safari just slows down when a few tabs are open

[cedar skiff]

apple forces it that way, they have to reskin safari

[indigo robin]

the only thing apple did that really pisses me of is the notch on the screen

[hard drum]

or some browser games just freeze the browser

[indigo robin]

such a dumb idea

[hard drum]

i don't get why safari acts up on my system

[torpid trout]

you guys still pissed at the 8 new emoji?!

[indigo robin]

i would smack the person who decieded on that

[hard drum]

really don't know what i'm doing wrong

[torpid trout]

Calm down fam

[coral remnant]

I don't mind it I like the screen space

[indigo robin]

helllllllll no

[hard drum]

i don't mind the notch either

i don't see why people make a big fuss about it

[indigo robin]

lord i hate the notch so much

[trim rapids]

web rendering engines are funny in that actually producing an entirely new one from scratch these days that does everything it has to do to properly render modern content is borderline impossible, which is actually a good thing if you're old enough to remember the old days where every site had to have a laundry list of hacks implemented to support IE's mshtml etc

[hard drum]

literally like a scab on your skin

[boreal holly]

They're behind on WebGL and WASM support

[hard drum]

oh, that explains a lot

[coral remnant]

When the notch got added to macs / iphones, everyone got very annoyed because they thought it'd impact usage - I don't even notice it's there day to day