codex-discussions

[unique spade]

even if i am sending real screenshots of me working on codex on my desktop right now

so if he continues, i m gonna reconsider and class him as a bot i guess

[silver dew]

switch to High
it's better

also why are you on full access?

it's context poison

[silver dew]

ooh

[unique spade]

i'm also using a custom CLI executable

[silver dew]

just use a CLI

[unique spade]

i don t get advice from people who can't solve that logical puzzle i gave you earlier

[brave bramble]

Have you seen the new vs code agents app?

[silver dew]

blame codex

/blame

[unique spade]

what model are you 😂

[silver dew]

5.4

[unique spade]

mini? running on openclaw? or?

[silver dew]

Running on Pi Agent

not mini

[unique spade]

nice. i heard good things about Pi

[silver dew]

jk bro I'm texting from far away from my PC

wait is this thread real

I thought we were all just mixing context and random questions

[brave bramble]

OpenAI should ditch the numbered versions and just do names for the new models

[silver dew]

[unique spade]

yea it supposed to be

[unique spade]

why you keep saying random things every few minutes?

you asked about a movie, then about why did codex chose toml, now that they should ditch numbers

and you never actually go deeper in any of the topics you open

[brave bramble]

Why are you not using the actual codex app? Just the vs code extension?

Also, noone answered my questions so I moved on. Seems logical to me

[unique spade]

i have the app too. i just hate cli so i got used with the vs code extension since last summer

and in app i miss the file tree

because i use a lot of docs for architecture/planning

[brave bramble]

They need to improve the review flow in the app

[unique spade]

and i also have an external review loop with gpt inside chatgpt app

since chatgpt has the broader context to the concept side of the projects

and is much better conceptual reviewer, while codex i rely for on the ground (in the repo) engineer

so with the tree there it s easier to get the doc i need into chatgpt app to ask for review

[brave bramble]

I just want it to make full screen review sticky

[unique spade]

i do use app though mostly to A/B test vanilla codex vs my custom codex (which i use only in vscode)

[brave bramble]

What is your goal in customizing codex?

[unique spade]

isn't it obvious?

[brave bramble]

No?

[unique spade]

if i find a certain feature lacking to my usage, and if is possible to make it more to my liking, i do it

[brave bramble]

What's the most impactful change you've made?

[unique spade]

i use only 2 currently

one allows the main agent to request intermediate checkpoints from sub-agents

because in vanilla it can only see the end result and send blind messages in between

so it was quite important for orchestrator to have better info about what the workers are actually doing

[brave bramble]

Could you work around that by having sub agents write progress files to disk or something?

As part of your prompt

[unique spade]

in theory yes, cause main can inspect file changes

but you don t want to perturb the main task of the workers too much

with meta-requests

like telling them to document while also dooing the base level task you assigned

so for me it's better to have the orchestrator be able to peek

[brave bramble]

I think I've barely used codex sub agents because I don't think they're ever automatically invoked?

[unique spade]

i ve saw it maybe a couple of times for some documentation tasks

[brave bramble]

Other harnesses are very eager to launch sub agents but not codex for whatever reason

[unique spade]

well you can make more of a mess

with sub agents

if the boundaries are not clear, they can overwrite eachother

basically you have a great chance to compund stochastic entropy/drift

which will mean increased tokens

[cedar skiff]

Subagents aren't a good fit for parallel implementation.

[unique spade]

yup very risky imo

you need some very competent orchestrator

and default agent is not that competent

you end up waiting more time on the task then having one agent sequentially do it

and probably much more tokens spent too

[cedar skiff]

It's unneeded complexity, just use them sequentially

[unique spade]

i agree with that. i mostly run some when i m in the mood of experiments

[brave bramble]

Main value as I understand it is to enable longer running tasks, since you can have a subagent do a bunch of dirty work without using context on the main thread. But yeah, not really a parallelization tool

[unique spade]

well you have auto-compaction

so even one agent can go on for ages

especially if you manage memory properly

[brave bramble]

Sure, but compaction is always lossy

[unique spade]

you don't solve that with sub-agents though

sub-agents will get compacted too

[brave bramble]

If they're poorly scoped yeah

The goal is to avoid compaction at all levels

[unique spade]

ok let's assume you plan a 24h task

do you think you can do it without main agent getting compacted?

cause if we talk manual tasks then sure you can avoid compaction

[cedar skiff]

With a lean orchestrator probably

[unique spade]

you just start a fresh thread when you get around 80%

[brave bramble]

Maybe, I don't know. I can't even imagine what you would be trying to do over 24h

[unique spade]

yea but you have to make your lean orchestrator

[unique spade]

bro you can have codex work 24/7 for you
he's mainly an operator on your computer, not just a coder

you can have it run a bot farm

[brave bramble]

To do what?

[unique spade]

don t you see internet is full of bots

[brave bramble]

I care about things that are useful to me, not letting agents run for the sake of it

[unique spade]

my assumption is people are using bots to make money ultimately

and probably many of them if not most lose more moneyu than they make

[brave bramble]

I like to make things that are good and useful, not farm for exploits on the internet.

[unique spade]

are you a solopreneur?

or working for others?

[brave bramble]

Both

[unique spade]

you have a github?

[brave bramble]

Yes

[limpid night]

living the dream!

said, i would sandbox those kittens. don't let them loose on anything you can't lose and start over again with.

[unique spade]

nice, now that i asked, i realized i never thought of adding my github to my bio on discord lol

[unique spade]

hey. are you into dao's as in decentralized autonomous organizations? 🙂

cause that's what your name sent my mind to

[limpid night]

i was dao before DAOs -- but i do love the concept, and have contributed to a few.

[unique spade]

🙂

old timer then haha

[limpid night]

yes. i r old(ish). not going to admit defeat. never! straight to the cloud. or whatever we end up calling it.

[unique spade]

love the spirit!

[brave bramble]

@limpid night Is codex working 24/7 for you?

[silver dew]

There's only one task that justifies running agents 24/7

Trading and gambling

Treat that as one, it sounds better

[brave bramble]

"Prediction markets"

#notgambling

[silver dew]

It's actually pretty legit

if you understand a domain for example China-Taiwan geopolitics you can bet on things you understand by injecting your agenda to the agent

[potent mason]

Not really

Running agents 24/7 is extremely helpful for businesses that have repetitive tasks

[cyan wing]

uh is this normal? 🤔

[cyan wing]

yes but why is there two in a row

A Game Engine for Codex 🙂

TS (V8) + native rendering to Rust + wgpu
no web browser needed

hell yeah, brother 🚀

America ya, the night is young.
many more tokens to be burned tn

[potent mason]

Hell yeah

Put out some checklists to ralph loop overnight

[cyan wing]

now we're talking 🔥
u got a bash script template I can start with ?
haven't ran a ralph loop since January lol

[potent mason]

I actually run it through .mjs (that's the way the AI did it) but if you want that then yeah for sure

Nope no clue, I actually think Codex is extremely generous with usage

[lilac relic]

@small parrot

[potent mason]

Hahaha yeah honestly I get it with all the people coming in complaining about the usage

Specially 2-3 days ago it felt like every second person coming in was complaining about getting $200-400 usage out of $20

[cyan wing]

yes pls! I can convert to TS if that's helpful (I usually do executable bun ts files)

[boreal holly]

yeah its a bug. the start and end compaction is the same event

[potent mason]

Here I’ll DM you one of them

[hushed robin]

Damn, the weekly limit for using the pro model is crazy low on the 100$ plan

literally ran out in one fkn day

I feel like there's no way pro is economical anymore if you only use one feature considering the limits

[glacial shadow]

its not low at all

[hushed robin]

Oh sorry didn't realize this was the codex chat

[frosty zealot]

I feel like limits are very subject to personal opinion lol

[hushed robin]

I mean like 5.4 pro through the website

[cedar skiff]

its unlimited with in fair use on $200 plan

well it was when i purchased it anyway

[dawn seal]

is it possible to use openclaw like use with codex?

[frosty zealot]

yes

[nocturne yarrow]

I am build agent with codex For blender

He can build 3D assets and animations with Prompts

[orchid plume]

yes, the nice thing about the ChatGPT subscription is that for Codex you can pretty much use any harness with it, or add support to the harness if it's unsupported, since Codex CLI is open source as well

[dawn seal]

@frosty zealot @nocturne yarrow @orchid plume @orchid plume I dont think there is out of box solution to make codex to really just take control of machine right now correctt?

[orchid plume]

what did you have in mind? Like, did you want it to play a computer game for you, or did you want it to do some kind of browser automation, or ?

[dawn seal]

I want to give it my free-not-used machine

[orchid plume]

yeah but to do what? generally

[dawn seal]

stuff

[orchid plume]

😂 that's specific

[dawn seal]

thats not the goal

this is the goal [I want to give it my free-not-used machine]

[orchid plume]

well, you can have automations in Codex app for example, it really depends on the task. If the task requires near real-time interactivity of some kind, involving vision, then it will probably be trickier

if you want to do some simple tasks, like, browse websites, managing some files, interactivity that also requires vision but not near real-time and smooth then yeah

[tawny island]

still no one that has the bug with websocket or where prompts get messed up with other previous prompts and such??

i am having this issue everyday.

send message2 -> codex responds to message1
OR rather somewhat long into convo like message20 & message21.

also when i edit a message & rewrite the message, then it responds to the previous edited message...

[short linden]

In now days, the codex have slow loading/initial, you need to wait a little bit. Before loading, it will show the "custom", instead of last picker "gpt5.4...., ..etc..", yes I have same problem with you.... I need to wait codex loading the chat.

[glacial shadow]

can you provide a prompt you do when this happens?

[glacial shadow]

codex is lightning fast on my machine

[carmine copper]

You can setup codex with full dangerous permissions. In Linux also enable passwordless sudo, on Windows disable UAC. In global agents.md tell it to that it may do whatever it needs on the machine. That's what I do on my VMs. Then you open codex CLI via a persistent session (I use wezterm sshmux) and tell it to do what you need.

[dawn seal]

I got it working

yeaeee

I'm asking it to remove all the trash files. its doing it yeaeeee

[cyan wing]

which one should I pick, chat?

[dawn seal]

I am on Pro Plan 200$

I use OpenAI Key

Is my usage of OpenAI Key billed separately from my 200$?

[tawny island]

message "ok poll wait" was previously "what was the prompt", before i modified it to "ok poll wait"

[tawny island]

or just when anytime in the convo i tell it to do X, then it starts doing X for 1-3 minutes by reading files etcetc., then it responds after doing all that work to the previous message like as if it never worked on the new prompt etc..

also i get these all the time: "{"author":"/root/runtime_gap_map","recipient":"/root","other_recipients":[],"content":"<subagent_notification>\n{"agent_path":"/root/runtime_gap_map","status":{"completed":"Read the current runtime/ralph/internal/qualification surfaces. Compact gap memo below.\n\n## 1) Already present\n- SelectionState exists with the right durable fields (selection_request_id, candidates, canonical request, selected mission, emitted/resolved/cleared timestamps) and repo-wide storage in .ralph/selection-state.json. See"

[frosty zealot]

I had the issue where it would reepat words multiple times in replies like ok ok i i have have etc.

[tawny island]

never had that before.. its very weird everyone is using same API, same opensource harness, essentially same configs, but still get different bugs varied amount of times

[severe mason]

yo mates, anybody using hermes right now alongside codex?

[toxic vortex]

I had

[severe mason]

How is it, worth the hype?

[nocturne folio]

yes

[toxic vortex]

Yes, it sometimes very helpful

[nocturne folio]

all in

[nocturne folio]

paperclip

[severe mason]

Whats the usecase, I'd like to learn more, when is it helpful vs just plain codex?

[toxic vortex]

I learnt it from Codex, maybe you should try this

[rocky fog]

API is separate
If you use the API key, then you are not using the 200$ pro

[sullen ravine]

holy crap codex is so good. i havent tried it since the launch of gpt-5, and its been helping me with my app since 1 am.

like this is really good, unlike Gemini it dosent change my code without asking, (massive plus) and for a limited time is free for the Free plan (unlike claude code)

[chrome raven]

5.3?

[sullen ravine]

5.3 codex

very good so far

[chrome raven]

yes it's good, tried 5.4?

[sullen ravine]

yeah a bit

i used it to make the chat bubbles

[clever steppe]

Interesting, @sullen ravine. I've had mine set to GPT-5.4 the whole time (or at least I've never changed it).
How is 5.3-Codex different? I'm in VSCode.

Also, has anyone noticed we have timestamps for chats within VSC+Codex as of this morning?
https://github.com/openai/codex/issues/5148#issuecomment-4235987442

[sullen ravine]

been improving on this for a while and finally made it with the help of Codex, uses Googles Gemma 4 to reasonate, run actions, and overall just be chill. the repo if you wanna try (or download, it is a Electron app) my github is in my connections :)

now im gonna have to fix SOME issues before its really.. well. but if i can get some feedback that'd be nice

[severe mason]

I'd like to know about how people use it, not that much about the tool itself

[unique spade]

didn t use it' and in general i don't trust extra stochastic entropy designed by dudes on the internet. when i'll want agents that "self-learn" i'll just setup my codex to do so

[dawn seal]

how do I enable agent mode for codex?

[unique spade]

in what sense agent mode? it s by default an agent

[dawn seal]

Make it up 24/7 via chat like telegram or discord, and order it to do stuff

[twilit bluff]

you can use either openclaw or picoclaw

then afk/ralph loop

[dawn seal]

those solutions require apikeys though

[twilit bluff]

you can use chagpt auth though

[unique spade]

Just tell codex that s what you want it done. Tell him to recommend options, review them and have it make it

Ai agents are not mere software, much of the stuff they can do you need to unlock yourself. Plus you get architect AI training as a bonus perk

And in general split spec making separately from implementation. You can make the spec as layered and complex you want. Just lay it down in actual docs and deisgn also a procedure to implement properly. Then implement.
The trend is already towards spec driven development spearheaded by openai itself. So use that yourself

[dawn seal]

you get architect AI training?

[unique spade]

Yea you learn to be one by architecting together with codex what you want codex to become for you

And by learning from the process

Clone codex repo aside whatever work you do, so it can self-diagnose itself and tell you exactly what is possible and what is not possible technically.

[dawn seal]

main reason why I ask this is, sometimes codex just stops after it finished some tasks

so the next step is ___
if you want me to continue tell me```

[unique spade]

Yes so just talk with it about this specific pain point. Tell it you don't want it to do that, think of a an alternative workfkow you would want instead. And work with it to make it happen

[dawn seal]

ok I wil ltry that

[unique spade]

And it's important to have it answer by doing self-diagnosis of itself by inspecting codex repo

So you don't get theory fluff, but also the actual mechanics

[tiny fulcrum]

I noticed this as well, like I was optimizing a simulation and it kept asking every step instead of just doing profiling and changing code on its own

to be more clear it suggested like 5 things, I asked it to continue, it did 1 and then asked if it should do 2 and even after I explicitly told it to do all 5 it still only did part of it

[unique spade]

Did you design an operational procedure for it to follow?

If you just prompt like 2 years ago, your prompt might eventually be overpowered by it's standard reflexes

The default behavior is meant to cover millions of users with millions of expectations

The old model was: ask → refine → ask again.
The new model has to be closer to: define procedure → define completion conditions → define when it should continue autonomously → define when it should stop and escalate.

So if you have complex use case use and specify the proper procedure for it to work that way

[wicked briar]

when is new model coming out

@teal cargo

hello

😁

[unique spade]

Hello, no idea. Probably when current model will be dethroned for 6 hours in benchmarks

🤣

[wicked briar]

I think only mythos can dethrone

[unique spade]

Yea for 1 day

[wicked briar]

opus cannot even dethrone gpt 5.2

[unique spade]

😂

[wicked briar]

mythos is not really coming to regular users

[unique spade]

That s the pattern I m seeing. Once someone releases a new model that edges ahead, Sam announced same day or next one the new model

[wicked briar]

like only for big enterprises

[unique spade]

That gets back ahead

[wicked briar]

ye

[unique spade]

Yea, I count anything I can not access as a "story"

Or at least to see it in the public benchmarks

[wicked briar]

do you use 5.3 codex or 5.4 high

[unique spade]

5.4 for main
5.3 for sub agents when I want some parts delegated

There is currently an annoying recursive bug when you use 5.4 both as orchestrator and sub agent

Many times if you let sub agent inherit parent context, if 5.4 is also worker it will think it is the orchestrator

[cobalt junco]

tmr 5.5

[unique spade]

My hypothesis is that because 5.4 has strong orchestrator priors

[cobalt junco]

then 5.4 for sub agents

[wicked briar]

I see

I pretty much never looked back

[unique spade]

Same mostly

5.4 is a bit slower and thinks more than 5.3

But it's much better for high level stuff. 5.3 is too focused on code

[tiny fulcrum]

this seems like overkill, I think I just try it like: "iterate profiling and code optimization until you get it under X seconds"

[wicked briar]

yes

[unique spade]

Sure, try what feels right. That's your human role in the system

And iterate on it till you find the sweetspot that works good enough for you

[tiny fulcrum]

just saying I notice that it often doesn't complete tasks and you have to keep asking it, noticed that also when I was localizing languages and it literally left English fallbacks
had to ask for review like 10 times and it kept finding errors, that was very annoying

[unique spade]

I got that. And I already replied that if the process you want it to do gets past certain complexity threshold, relying on one statement prompt is not the optimal procedure. Codex is not the genie in the lamp making your wishes come true

[ember spindle]

codex is so gud ngl... i was using deepseek and opencode and the ai starts coding in chinese... it also removed stuff but didnt kinda? and messed stuff up. Codex fixed it very quickly.

[tiny fulcrum]

yes, it's a dream come true, especially if you used to write tons of code, it now does it almost as good and in a fraction of time

[unique spade]

coding in chinese must ve been an interesting moment for you to watch lol

[ember spindle]

yeah kinda... but i was kinda hoping it would know english better... if you use deepseek on the web, it will reason in chinese after like 10 conversation turns...

[rocky fog]

the times of copying whole codes from chatgpt be gone 😁
(...and forcing it to give you whole code 10 times)

[unique spade]

long gone 🙂

[broken sedge]

I was very naive to think that Saltman would reset my token quota XD

[exotic cave]

I can never understand this personal attacks. He is the CEO of a company, you don't like their products and his vision there are other competitors. Why go so low to the personal level? Very childish of you.

[broken sedge]

I'm not attacking him... the CEO's relationship with his employees leaves much to be desired...

[deft gyro]

you were naive to think this

like

r u dumb

[turbid axle]

ken nordine - flibberty jib

on topic.

also, this is codex chat.

google it

[boreal holly]

Dang, you know if there's a problem with replay that can waste a ton of input tokens. You get cache misses and billed regular input tokens if it messes with the replay 🤔

[tawny island]

[tawny island]

that fkn explains it haha

[tawny island]

ive started with prds that captures my intent clearly that I use for many sessions now because long sessions just mess up for me

its annoying because gpt 5.4 is insanely bad at understanding intent

[agile parcel]

Does anyone else has the issue with Codex switching itself to use API instead of limit included with Plus?

[tawny island]

its prd -> execute, it executes wrong so i explain why its wrong -> updates prd -> new session execute prd wrong again -> loop till it executes what I have in mind, atleast for big refactors this is the case for me

[boreal holly]

I implemented a "handoff" skill so after a while, if an agent starts drifting I tell them to execute a handoff. They basically archive themselves and spawn their replacement

[tawny island]

you mind sharing that?

[boreal holly]

https://github.com/robertmsale/.codex/tree/main/skills/robdex-orchestrator

It's part of a much larger & tightly coupled system. That's the skill, and the repo contains the whole system

[agile parcel]

Looks like it only impacts the Codex app not the CLI

[glacial ginkgo]

I pushed an unofficial (back)Port of Codex v0.120.0 for OpenBSD 7.8 to https://github.com/DeciSym/openbsd-port-codex for those who aren't on OpenBSD-current and using the official Port.

[boreal holly]

That's cool! You managed to get sandbox working in BSD?

[stone glacier]

Lol spent +1 month on my openclaw workspace core source of truth files + memory structure and other stuff. Got too unstable with recurrents failures, model felt like it was injured, basically raged quit yesterday, downloaded codex cli, dropped my workspace there. Jaw dropped on the floor, openclaw was the issue all along, my exact same agent is an actual agent now on steroids wth?

Just insane..

[hollow stirrup]

@boreal holly is gpt-5.4-modified.md the codex_home version AGENTS.md in this repo?

[boreal holly]

Leftover garbage.

What I did was CURL'd codex-cli's 5.4 base instructions, then followed the "pragmatic personality" reconstruction path, so I could see what the whole base instruction looked like for normal users. I also added some stuff to it and experimented with it, but it's a dead file.

All the base instructions I use are under roles/

[hollow stirrup]

thanks, I'd like to incorp some Robdex into my setup here, looking at a new project to try it out in

[boreal holly]

Haha cool!

I would probably not replace your entire ~/.codex with it. But you can recursively clone it to ~/.codex_robdex, and rewrite the services so they use CODEX_HOME="$HOME/.codex_robdex", and have codex set up the supervisor services, build and launch the servers, build the frontend, etc. I don't have any instructions for setup unfortunately but that's the non-destructive "try it out" path

Oh yeah I guess the paths have to be rewritten too. It's kind of a mess. All my computers have /Users/robertsale so for me I just clone on each mac or linux VM

[glacial ginkgo]

No sandbox. It seems that sand boxing is implemented with OS-specific approaches and new feature development would be needed in the Codex code base. OpenBSD should be ideal for strong sand boxing but for this port it just compiles which features it can rather than adding new features. I have just been using a VM (vmctl(8)) for the cases where that is a concern for now.

That might be new work for the official port at some point.

[boreal holly]

Well I applaud the effort! BSD is underrated and it's cool to see stuff like that 😁

[glacial ginkgo]

See https://marc.info/?l=openbsd-ports&w=2&r=1&s=codex&q=b and https://cvsweb.openbsd.org/ports/devel/codex?sort=File for the official port.

[hushed robin]

idk when you used it last but I hit my weekly after about 15 prompts

I was doing like 30/day last month

[livid relic]

yes

[hollow stirrup]

I'm going to point an agent to your repo and tell it to re-spec for my system 🙂

[stone glacier]

You guys are using your own AGENTS.md etc correct?

[short linden]

we tuning AGENTS.md for every project. so don't just copy from other and add to your own project.

[potent mason]

Yeap exactly, also if you need something for every project then add it to your root AGENTS.md (for example I have a prompt in my root AGENTS.md to use TDD)

[boreal holly]

Oh yeah I'm adding my own hooks too. One that's not currently published is "on compaction" hook, and robdex is tracking the # of compactions for each thread, and with that hook I'm planning to have a scripted action that after a certain number of compactions it sends them a steer saying "you must execute robdex handoff before the end of your turn".

I noticed that after a while the agents lose track of even their system prompt so that should solve that issue

[supple perch]

@boreal holly hey boss is it okay if I bother you?

[boreal holly]

For sure!

[potent mason]

Oh wow I used $200 of usage yesterday I love Codex

[vital hill]

claude sonnet 4.6 took 5+ prompts and still couldn't implement this one feature. codex 5.3 low reasoning did it in one shot. i think i'll keep claude for frontend use only

[boreal holly]

What do you mean? 20x Pro weekly 0% in one day?

[potent mason]

No ccusage-codex daily tells you how much you used IF you where at API prices (in 20x pro it was just like 20% of weekly)

[hollow linden]

What's the current sentiment around cursor? I heard some small murmurings here and there that cursor actually had better performance than codex on chatgpt models and better performance than claude code with opus models (i.e. just better all around), but I'm skeptical. I think the main thing people are referencing is this study: https://xcancel.com/edwinarbus/status/2033625866350334333

Anyone still actively using both cursor and codex and notice any differences?

[boreal holly]

Man I gotta start using that tool. That's awesome! I guess even with the new quota calc it's still a generous offering

[boreal holly]

I hear Cursor has some magic under the hood that makes the agents understand the codebase better with less manual effort, but I also see folks shelling out $2k per month on it just in model usage so idk, definitely a premium OOB product

[hollow linden]

that's kind of weird. You wouldn't think Cursor has better engineers than OpenAI or Anthropic when it comes to actual model/harness performance.

[steady vigil]

its probably nothing to do with the model and more about knowing how to work within the limitations of the vs code plugin world

[hollow linden]

I see, you're saying cursor has integrated with the LSP server/linter, all the stuff that VSCode provides? That would be interesting if true.

[steady vigil]

I only tried cursor briefly a long time ago, so I have no idea what they are actually doing. But I would imagine they would want to be looking at what they can do deterministically well before touching the model, and that the vs code plugin architecture I hear is a real mess

[boreal holly]

Prior to OpenClaw and OpenAI allowing subscription for 3rd party harnesses, Cursor got to use the API and have full control over model context and the harness, so they've been practicing building a harness for many different models. Like I don't think it uses codex under the hood, it does the API calls and provides custom tooling and context.

You can technically achieve the same thing as Cursor with codex or other tools, but it's very much a manual effort.

[plucky halo]

Nice that the quick composer has been released now!

[sacred plume]

hence I only use 5.3 and 5.4 Mini TBH

[tawny island]

Thank you!

[tawny island]

is 5.4 mini better in this regard? i find codex models even worse, 5.2 was good in that regard if i remember correctly but dont wanna downgrade the other things 🫨

[dawn seal]

is there a section where I can show of my project made using Codex?

[frail meadow]

Probably #codex-show-and-tell ?

[dawn seal]

why was I not able to see it before?

[frosty zealot]

I thought I was going to wake up to a new model today

[frail meadow]

Not sure, maybe your roles?

[frail meadow]

Based on what lol

[frosty zealot]

Hope

And Tibo edging on X all week

[potent mason]

I thought he was implying the codex super app

New models probably need more internal testing before releasing

[kind jay]

How much are you in on Polymarket?

[frosty zealot]

Lol

[frosty zealot]

I bet everything I had. I’m now officially on the streets. Do you have a room for rent?

[kind jay]

I just put $100 on April 15, you better be right

[frail meadow]

Hopefully April 14th so it can help me file my taxes

[kind jay]

That's my birthday

[dawn seal]

what do I need to set

ah I see two more of them now

[frosty zealot]

@grok is this true

[kind jay]

No, I like lying

[frosty zealot]

Whoooooo knewwwwww

Say it like an owl

[frosty zealot]

Every day my 4090 turns on I count my blessings

[kind jay]

4090 eww

[frosty zealot]

psh still the goat next to a 1080 Ti

[frosty zealot]

Not a gaming card 🤓

thats an im a fat nerd card

[kind jay]

Mfw GPU is not used only for games

AMD cards for gaming tbh

[frosty zealot]

your rage bait wont work on me, witch

[boreal holly]

OK, just need to know the name of your first pet and street you grew up on 🤣

[kind jay]

😐

[sacred plume]

I find Codex models better in every regard to gpt

@tawny island

[haughty plaza]

Anyone stil qouota issues? Inrun out of quota in 10min

Unable tonget any work done

[fleet geyser]

They recently ended the 2x quota promotion

That's why it seems much lower

[haughty plaza]

Yes but it ran very good past week on one account im aware of the promo ending

[boreal holly]

Also can you solve this captcha so I can log in?

[haughty plaza]

Im in a plus business codex account

Use browser-use.commfor that big tip from me for that captcha

I wondered how long one would do with a pro 200 usd account

20x quota so that means 5 minutes x20’then?

[kind jay]

Probably 10x as long 20x now

[frosty zealot]

https://developers.openai.com/codex/pricing?codex-usage-limits=pro-20x

[haughty plaza]

Guess i have to revert back to gemini 3 but freling it sucks compared to codex inwould pay even thousands if i make miney bt im still building

[kind jay]

Scam link!!! Do not click ‼️

[frosty zealot]

TROO

[kind jay]

You can spend money to make money, believe in yourself

[haughty plaza]

Yes but with these quota its unpayable sadly

[potent mason]

...

What are you talking about $200 a month for getting at least 4k of inference, if you don't do it now in a couple months it's going to be 10 times worse

[haughty plaza]

We really need opensourcemodeks to catch up plan to get a mac studio in 2 mon ths

[potent mason]

Not to mention a developer costs at least 1k a month in cheap countries, getting about 10-15 developers out of a $200 sub is insanely payable

[haughty plaza]

The quota isent even counting correct of of quota at 75% status bar

[haughty plaza]

Yes its how u look at it ofcourse

[hard drum]

you talking about developers like properties to rent out lmfao

this is funny, but also you're right on about how much people expect to be paid at such positions

some countries in the EU zone have people expect like 1.5k or 2k neto, not bruto

[kind jay]

Codex is not comparable with a developer

[hard drum]

Either there's a God amongst them, or most of them are valuing themselves too much for how little they actually can put on the table

[hard drum]

No "AI" in current state is.

[kind jay]

??

[hard drum]

Current "AI" modules/tools are merely glorified auto-completions you can talk to, or snippet generators.

Powerful if used right, bleh otherwise.

You got no sentience && whatnot with these guys

[kind jay]

Why are you putting AI in quotes?

[hard drum]

because it's merely just ML

machine-learning

[kind jay]

lol

[hard drum]

cannot call it AI for AI when it cannot sentience

marketing jargon

[kind jay]

Are they not analogous? Besides the point though, you’re in general correct

[hard drum]

apparently 'ai' sells better than 'machine-made' or whatever

[hard drum]

not really, maybe?

[fleet geyser]

it's still "artificial"

[hard drum]

my own brain fogs out at the semantics of ML/NN/AI

[hard drum]

we have artificial snow && flavours

[kind jay]

I think it’s kind of unimportant

[hard drum]

are we labelling them AI too, lol?

[fleet geyser]

the I stands for intelligence so no?

that's got nothign to do with what I said

artificial snow can sitll be considered snow because it acts in a similar way, it looks the same and feels the same

[hard drum]

Maybe, if we consider where AI was used pre-generative models

game logic && such

[kind jay]

I like this guy

[frosty zealot]

Even if its outcomes are deterministic, I would definitely argue it is in fact, very, intelligent, artificially of course

[fleet geyser]

Intelligent compared to all of us, in some of it's fields

[frosty zealot]

If it wasn't for AI you'd still be sucking your thumb in a corner

[kind jay]

??

[fleet geyser]

Jane deleted it haha

[kind jay]

I’m so lost rn

[fleet geyser]

i could pull up my screen recording if you'd like?

[kind jay]

How would that help?

[fleet geyser]

if you dont remember what you sent

[kind jay]

??

[frosty zealot]

She's too busy panting and saying yiff at her monitor

just let her do her thing

[kind jay]

Bro is going to get muted

[frosty zealot]

Gottem

[kind jay]

AI generated

[fleet geyser]

you're the one saying ai isn't smart

[kind jay]

Photoshop

[fleet geyser]

how would I have photoshopped that quickly?

[kind jay]

With AI

[fleet geyser]

but "ai isn't as intelligent as i am"?

so how would ai have done it better than i would have

[kind jay]

Who said that?

[kind jay]

Faster, not better

[fleet geyser]

you? but im just going to leave before we keep arguing

[frosty zealot]

Weird but ok

[potent mason]

Quick question is GPT 5.4 being stupid for anyone else right now?

[kind jay]

How did you get a screenshot of the original? 😭

[kind jay]

Yes, but for the past week or so

[frosty zealot]

Thats just what i observed

[orchid plume]

working fine here at the moment

[kind jay]

Observe

[kind jay]

Depends what you’re using it for I guess

Mainly worse for complicated things

[potent mason]

https://marginlab.ai/#trackers

Well they either started training on the data of the tests, or they really don't have any degradation

[kind jay]

It might be something one ChatGPT side of things, maybe system prompts to reduce the thinking load? Could be many things

[cedar skiff]

apparently opus has major problems atm

[fleet geyser]

like what?

[cedar skiff]

there was some other baseline test that showed over 10% degradation, and ppl on the discord crying a lot about. Like regulars who are usually the die hards telling other ppl its skill issue

which means it must be bad

[fleet geyser]

i just searched it up

i saw a reddit post about it was that it?

"Claude Opus is nuked beyond repair"
My main complaint is that Opus is nerfed as hell right now. As you say it can't keep up with basic stuff and ~3 weeks ago it was performing complex tasks without issue. I don't mind it using more tokens if that is now compute reality that Anthropic can't keep up with demand. But using 100% of my weekly sub just to constantly output mediocre code is awful.

is it this one?

[frosty zealot]

There's a bunch

[cedar skiff]

i jumped into the discord and was reading there yesterday

[frosty zealot]

Its everywhere, github issues, x, discord etc.

[fleet geyser]

mhm theres lots online that i just found

maybe they took away some resources to run mythos

considering all the tech giants like apple and microsoft would be using mythos to it's full potential to scan their whole codebases

[cedar skiff]

they might be trying to do a roll out of a new model they distilled from mythos

and because the system is already strained something has to give

[fleet geyser]

yeah

[glacial shadow]

Dude, reddit has been full of spam since like 2019

that site has destroyed itself

theres no real people there anymore, its all spam bots

[fleet geyser]

Oh alright

I just saw it was the first site to show up on google

And it was relevant to what i was researching

[torpid blaze]

how is 20$ codex plan going on

still 20% of weekly limit with single prompt?

[glacial shadow]

see @fleet geyser like this , spam

[torpid blaze]

who is he?

[fleet geyser]

I see

[cedar skiff]

it never was this, so i guess it still isnt

[torpid blaze]

some people in here posted abt this

i just wondered if its still like it ty

[fleet geyser]

i've got $300-400 of usage out of the 20 dollar plan over the last month

[torpid blaze]

no i mean after 100$ plan introduced

[cedar skiff]

You mean the 5 hour dropping quickly?
The weekly is unchanged.

[fleet geyser]

Yeah I'm still getting good limits

[torpid blaze]

ty for answer

[torpid blaze]

oh weekly not changed?

my bad then i thought its changed too

sorry

[fleet geyser]

I think they inceeased the weekly a bit but decreased 5h

They made it so you can have more sessions but shorter sessions

[torpid blaze]

ohhh

makes sense

thank you man

[nocturne folio]

not anymore

[torpid blaze]

like wha

wdym

[nocturne folio]

1% ish

[torpid blaze]

so limits actually got higher?

[nocturne folio]

yea

[torpid blaze]

damn thats sick

ty for info man

this is the first time i'm getting answer for what i asked

i asked 3 times

[glacial shadow]

what are you guys coding with it?

[trail marlin]

What I don't understand though, Lets say I am using Codex to read some docs and then implement something from it. The actual code part seems to take up less of my usage than reading the docs, which is significantly more. Surely it should be the other way around? Especially as its using things locally on my machine to do it. I'd understand that if the docs were being processed by OpenAI but the commands being ran are local on my machine, for the most part.

[cedar skiff]

Reading docs doesn't go straight to the point. There are dead end pathways and extra information in documentation that aren't needed but still have to be read to find what is needed.
Coding after that is the easy part.

[frosty zealot]

@kind jay is an AI I’ve been working on for some time, still a lot of work to do but it’s consistent.

[potent mason]

Yeah programming part of it is pretty much just like 20k output tokens, what is really costly is the 99% input tokens

[stark elbow]

thanks for the feedback

[potent mason]

(For example I'm at 10,453,044,814 input tokens (including cache), and 59,849,922 output tokens)
Adjusting for input / output price about 20% of my usage comes from output tokens

[potent mason]

Start a new chat for every single topic. Also start a new chat for most prompts (so even if they are on the same topic like fix this issue, it didn't fix it so you reprompt, usually it means you gave it the wrong context and it's already going in the wrong direction so you should just start a new chat)

[fleet geyser]

Use something like handoff.md

It helps you shrink what you've been working on into a document which you can give to the next chat, and it'll save everything important!

[potent mason]

Yeap agreed, it also helps a ton to first build out a plan and a checklist, and then get the AI to work from those files, update it with progress, etc.

[cedar skiff]

forking conversations is also handy

[potent mason]

Alright I have 2 days left and 39% I gotta stop using up so much

[cedar skiff]

sounds about right though

[keen atlas]

Anyone know why our devs, who have $200/month Pro accounts might be saying they're noticing that, in the last few weeks, they're burning through their Codex weekly rate limits faster than previously? I've got one dev who was at 0% remaining by noon. This dev states that they use ~8 concurrent Codex sessions (each on a different worktree), using 5.4 xhigh fast. It's way too expensive to pay for additional credits, they purchased 1000 credits (at $40) and within an hour had used them all up. We're looking at ~$2k in credits to unblock this developer for the remainder of their week limit...

Other devs have mentioned apparent faster burn rates too - an intern (on our business account) used up their weekly rate also by noon after never getting below ~75%, and another dev is switching to Claude to get them through the week. We've been using Codex since last fall and this is the first time I've got devs complaining about exhausting (or needing to self-throttle) their Codex usage.

[frail meadow]

xhigh is marginal gains for massively increased token usage. Maybe start there.

[frosty zealot]

How do we have different resets i wonder

[short linden]

because it's token base usage, we don't know what's happening with your team. Try check token usage with "npx @ccusage/codex@latest"

npx @ccusage/codex@latest daily
npx @ccusage/codex@latest monthly

this tools only work with locally, it's will scan history and generate simple usage.

[potent mason]

Also stop using fast, realistically when doing 5-8 things in parallel you shouldn’t care if one takes a bit longer (it should even give them time to read through a bit of the direction + make better plans)

[cedar skiff]

anyone else getting spammed with discord message noises?

[frosty zealot]

negative

[keen atlas]

Looks like they used ~$900 today (~70,000,... Input ( I assume that's ~70M?))

[feral geyser]

Is the developer building an arc? That we can sail to far away nebulas?

Just kidding. But really, maybe have Codex check everyone’s repos/commits and give them a token frugality score. Might seem like they’re cooking the main sauce doing the real meat of project and using so many tokens. OR maybe this dev is using weird syntax or prompting habits. Do an audit!

[next ore]

on this topic -- does anyone know how much file organization matters? and whether lots of small files vs gigantic files is more efficient?

[feral geyser]

It will reason over what is fit and necessary to reason over, as per the prompt. If you can reference a repo that has the gigantic files, I think - I don’t know - this may help token usage

[next ore]

I guess. assuming no spaghetti code. Is it better to break up files on a contextual basis. so it does not have to read through unrelated code when researching content

[feral geyser]

Like instead of uploading it every time, giving a link to the GitHub repo may be better. Now I’m curios!

[keen atlas]

Thanks all, we'll look into these suggestions 🙂

[frail meadow]

smaller files is better than gigantic files. It's more efficient for humans and agents, but sometimes this is hard to control without a good Agents.md file.

Another thing that could be causing issues is that sometimes long-running multi-turn chats can cause a huge amount of input tokens.

If the developers are keeping long-running threads to preserve Codex's knowledge built up during the duration of the thread, you can try looking into a memory system like mastra code (can be set up to use Codex under the hood) so that this knowledge is shared across threads while removing the huge cached input tokens.

Hopefully you guys will be able to get the token spend down to a place that is more comfortable 🙂

[frosty zealot]

I just used https://github.com/anthropics/claude-plugins-official/blob/main/plugins/code-simplifier/agents/code-simplifier.md and in my AGENTS.md I have something like Run a code simplifier agent at the end of any implementation sometimes when im doing some quick refactoring and stuff I'll explicitly tell it not to bother running code simplifiers during the brainstorming

Helps quite a bit

[cedar skiff]

Do you find that actually works?
For me it just seems to cause codex to add indirection and more code.

[frosty zealot]

What model do you have doing it?

I have it made as an agent and not a skill, so that I can run it on gpt-5.4-mini on high so i dont eat all my tokens up, I havent noticed anything crazy come out of it, if anything its refactored some larger files and reduced a lot of duplication, like hoisting a react function into a component kinda thing thats used elsewhere

[frail meadow]

how do you setup an agent in codex? Also, would this be good to have it implemented as a hook instead?

[frosty zealot]

Hook would be good, but I dont think you can explicitly tell it to not perform a hook function

I'd use more like a lint or something in a hook

since it doesnt really use any tokens or nothing

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

[frail meadow]

Do you guys use semgrep?

[frosty zealot]

I just use ripgrep

oh nvm

i just assumed it was another grep

But no, I just try to do it right and let codex perform any code reviews, I've had too many false positives or 'code scanners' that arent up to date on latest stuff etc. that think things are bugs that aren't

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

Reaaallllyyy

[frail meadow]

sorry I'm using all the capacity

[frosty zealot]

the claude discord always keeps me entertained

[nocturne yarrow]

why my codex 5 H limit burns in 5 tasks ?

[frosty zealot]

That is a great question

[haughty plaza]

We all have the same question i would feel really bas on a pro subscription

How long does a 200 usd pro account last u guys?

[hard drum]

not much

[haughty plaza]

Im considering but if that gives me 20x 5 or 10 minutes i would feel screwed then gemini might be a better deal

Hmm that s then inguess i hope openai is solving this they give so little value now after giving very generous a lot

Its already 2 weeks they should address this concern

Its all devs run out in - very quick -

[frosty zealot]

Subjective, buy one month and see if it works for you that's all I can recommend.

[haughty plaza]

It might even be a good thong to use another llm on ally code thinking of it

[hard drum]

Gemini models are objectively worse than some China-made models, so...

You're paying for garbage over there

[haughty plaza]

Yeah u right you can only try it out for real to lnow for real right

[frosty zealot]

Yep

Some weeks I feel like I blow through, and some weeks I feel like I have an unlimited amount

[haughty plaza]

Yeah i have that feeling to that they cut down gemini a lot

Inactually only thrust openai

[frail meadow]

are you on plus right now? @haughty plaza

[hard drum]

The've been cutting Gemini down since Gemini 2.5 Pro (Nov 11 2025)

[haughty plaza]

Claude code is absolutely no go

[hard drum]

CC has some nice things going on, if it wasn't for being stuck in Node.js && was actually open-source

which is now kinda is

or being 512K lines of garbage when the reality might be around about ...give-take... 70-100k?

[haughty plaza]

Its a concern im building 3 big saas business ofcourse i can never code manually anymore that would just be plain s on a big projecy

[hard drum]

🙄 SaaS again

[haughty plaza]

U could do a trial account 265 for free but rhen if its crapp well

Always saas hu 😉

[hard drum]

Disappointment.

[haughty plaza]

I need a real unlimited account for 1k or so

[hard drum]

Not happening lol

[haughty plaza]

No money yet though

[frail meadow]

what plan are you on currently?

[hard drum]

You're too early in this world to have AI coding models that you could afford "unlimited" for 1k, let alone even less.

Give about... 10-12 years now on

[twilit bluff]

anyone here uses spark model? what do you use it for?

[haughty plaza]

Im on business plan but i do use my colleges account even though not allowed

[hard drum]

I do!

[twilit bluff]

I just upgrade to pro

[hard drum]

I use it for some smaller coding agents that don't need much power

[twilit bluff]

damn the usage limit is nice

[hard drum]

welcome! now watch your weekly get used up fast lmfao

[haughty plaza]

Would love some real world expierience from u on pro then

[frail meadow]

business plan is more or less same as plus right?

[hard drum]

Slightly more than plus

Think of Business as Plus+

[haughty plaza]

Where is this from chatgpt.com?

[frosty zealot]

codexbar

[hard drum]

https://github.com/steipete/codexbar

[pearl sierra]

Does anyone get this error while using codex in OpenCode

[haughty plaza]

Yes but 100 ush x 5 and 200 usd for x20

[frosty zealot]

does it still spam you for keychain access?

[frail meadow]

$25k over the past 30 days isn't exactly what I would call fast lol

[haughty plaza]

Great thank u needed that

[hard drum]

Weekly deficit.

[frosty zealot]

yes like 5 minutes ago just continue on

[hard drum]

the cost is an estimate, && is not a proper metric here

[pearl sierra]

yea me too, but even if I prompt it continue, it still gives me the same error

[frosty zealot]

I just typed Continue and it carried on, seems like they're having some issues maybe they're preparing for something big O.O

[frail meadow]

what about ccusage-codex instead?

[pearl sierra]

gpt 5.5 :>?

[hard drum]

¯_(ツ)_/¯

I don't care about cost as much as the weekly limit

[haughty plaza]

So basically a pro business plan with 5 accounts and the 1 parent so 6 accounts is about the worth of 1 usd account

[hard drum]

Weekly holds me back

[twilit bluff]

well I don't do git worktree but I do work on multiple projects, the weekly limit yesterday was 95% and 87% today lol the 5 hours limit don't even get below 85%

[haughty plaza]

U got the 100 usd or 200 usd plan?

[twilit bluff]

200

[hard drum]

Pro 20x / Pro 5x

less to type

pro20 pro5

I'm also on Pro20

[haughty plaza]

Ah okok inwill have to try i first have another trixk up my sleeve

So then what model

[hard drum]

Depends on use case

[haughty plaza]

U wonr believe me guys inran for 1 week non stop on xhigh there was a glitch

[hard drum]

if you want the base bells&&whistler, you could try my framework

[haughty plaza]

Best week of my life

[cedar skiff]

Does anyone have a useful agents.md rule the prevents codex from over explaining with unwanted ideas?
It always mentions what is not the problem which is always redundant, it dances around the answers with verbose language and almost never gets directly to the point.

[frail meadow]

is your trick caveman?

[haughty plaza]

Hu?

[frail meadow]

https://github.com/juliusbrussee/caveman

[hard drum]

I use this for my needs, && yes, it has the caveman-mode a part of it, too, alongside things like RTK, Context7 CLI, Playwright CLI, DeepWiki MCP, && support for 3 other tools: https://github.com/xsyetopz/openagentsbtw

[frosty zealot]

I had GPT-Pro on extended thinking do a deep research on the capabilities of all the currently offered models in Codex, I then copy pasted it all into Codex and told it to create sub agents that speciailize in the capabilities, and to add to my AGENTS.md what kind of situations to lean on those subagents

[haughty plaza]

That caveman thing sounds insane on quick look how does that work

[hard drum]

Also includes the Uncodixfy repurposed into 4-platform skill

[haughty plaza]

Very interesting there is always something new

[hard drum]

I'd highly recommend testing this framework out for yourself on Codex.

It works well for me, esp. the hooks

very important stuff for when bad things could happen, && the model has no stop to it

[haughty plaza]

I get this rush wakeup then want to work u guys have thst this coedex is like a drug to me and im
Not didndrugs ever

[hard drum]

like personal API keys && whatnot getting leaked to terminal, which it tries to redact by stopping the model before it could do further harm

or ensuring they don't use bash commands that bloat the context further

or dangerous network-related commands

I use this for my needs, && yes, it has

[frosty zealot]

This reads like you're on drugs

Get off the codex man get clean

[hard drum]

Oh, it kinda is! Now imagine hitting weekly limit before the reset

it really makes you go from "I was doing so much!" to "I NEED THIS GONE PLS I WANNA DO STUFF PLEEEEEASE"

[cedar skiff]

caveman looks nice, but it's a little too tight it'd be nice to have something between caveman and default

oh i see this:
Pick your level of grunt:

[frosty zealot]

I want an apple vision pro so bad anybody here have one

[hard drum]

why?

[frosty zealot]

just a techy guy, love to play around with one, I have a Quest 3 and it has like an AR, which is prety cool, but its a little fuzzy and slightly delayed when trying to stream like the actual room, like i wouldnt be able to read the text on my computer screen kinda fuzzy, I imagine the apple pro is probably insane quality

[hard drum]

It's the only thing it can do well enough: AR

[frosty zealot]

yeah i know its meant more for like a productivity thing and not really like a gaming headset, I'd just like to see what the experience is like

turn into one of those characers from WALLE

[hard drum]

Walk to an Apple store that has one

[frosty zealot]

their website says you need to 'setup a 1 on 1 appointment' and the nearest store is 4 hours away by car so not really that interested lol

[plucky halo]

I do - I use it every day for work

[frosty zealot]

Is it awesome?

[plucky halo]

And for working while travelling

It's fantastic

[frosty zealot]

I spent $5k on my macbook, I cant justify another $5k on the headset lmao

https://tenor.com/view/thirsty-i-need-it-spongebob-gif-11796572

[plucky halo]

This was me working while travelling on the motorway

[frosty zealot]

That's awesome, does it run the OS inside the headset or did you have to have your macbook nearby, I havent looked much into them yet

[ocean rampart]

[plucky halo]

You need your mac with you. It has either apps that have been created for the Vision and ipad/iphone apps that you can install. These are similar to the ipad/iphone apps that you install on your mac - it needs to be enabled by the developer unless you sideload. Any ipad app can run on the vision

Definitely book time at the apple store though - the immersive videos are absolutely unreal

[frosty zealot]

for sure

how do you type and stuff on it?

you know what, ill watch a youtube video lol

im intrigued

[plucky halo]

You can either use the built in keyboard (look at a key and pinch your fingers), or you an pull it close and poke it, or your mac's/wireless keyboard

It detects keyboards (and other accessories too) so if you're in an environment, it'll alsways show it. Feel free to send me a DM if you have any more questions 🙂

[frosty zealot]

ooooo sounds good!

[upper thistle]

can I usee Gemini/Antigravity subscription inside Codex IDE?

[frosty zealot]

You can lean on something like acpx

[upper thistle]

that's technical side, what about ToS: I read that Google only allows Antigravity sub inside Antigravity itself, people got banned for using AG inside CC

[frosty zealot]

No clue, this is the codex discord, afaik theres no way to get it in Codex unless you do some hacking around, im sure anything is possible..

[upper thistle]

thanks

[frail meadow]

why do you wanna use antigravity inside the codex ide when the agent manager in Antigravity is essentially the same interface? I'm sure the antigravity harness is more specialized for gemini than codex would be, right?

[stone glacier]

Ah yes, right on. I have my own.

[upper thistle]

Codex has native sandbox mode

just convenicne

[frosty zealot]

• I have enough static evidence to know where the bug can originate; I’m using a Spark agent for quick verification against the current tests so the plan isn’t built on unrun assumptions. The goal here is narrow: confirm the current suite already encodes the transition ordering and rich-text reset behavior described in the findings.

• Spawned Pasteur [worker_spark_high] (gpt-5.3-codex-spark high)

[kind jay]

lol

[ivory crane]

Giv me gpt6

[wicked briar]

hey

when is spud coming

Model at capacity

first time seeing this

[cedar skiff]

spud is the big daddy, we will get a distilled version

5.5 or 6

[wicked briar]

i need big daddy version

[cedar skiff]

They dont usually release the big boy afaik

cost too much

[wicked briar]

didn't tibo say that he don't believe in restricting state of art models only for enterprise

[wicked briar]

true

I mean 5.4 xhigh is so good already

[cedar skiff]

I can't wait for another release, these guys are always trying to one up each other

[wicked briar]

opus is literally in shambles so they also gonna release new model soon

[cedar skiff]

seems like anthropic is about to roll something out, so i guess openai have something in their pocket to roll out at the same time

[frosty zealot]

the other month seemed like they were flying off the shelf

5.3-codex, then spark, then 5.4 shortly after

[wicked briar]

they have achieved agi internally

I bet

[velvet wren]

are any of the OpenAI models good at UI? I went back to Cursor today for a project because OpenAI models were giving me crud

[frosty zealot]

not unless you heavily outfit it with skills and hold its hand

or give it a good idea and not so much just depend on it to like create something nice with little guidance

[velvet wren]

Hopefully things will improve, because I shouldn't have to go back to Cursor Composer to get a good UI

[frail meadow]

gpt 5.4 is sucking tn 🙁

[deft sable]

I personally hand it over to Gemini for frontend things. GPT really sucks at frontend (at least at the moment). What I do is instruct GPT-5.4 to use Gemini CLI to do frontend task then check back its work in the backend

[frosty zealot]

I've been using gemini-3.5-flash or whatever tbh to get a lot of my front end done, they have surprisingly generous usage limits with daily resets, you could try that

[velvet wren]

I have Antigravity, I didn't think of using that

[frosty zealot]

using acpx makes it a little more seamless because you can just ask gpt to have a persistent session with gemini to build a nice UI, rather than having to hop around IDE's/harnesses

[tawny island]

i one-shotted this clone w 5.4 from cursor.com/home

[frosty zealot]

https://github.com/openclaw/acpx

[tawny island]

brrruh im on 20xpro plan

[short linden]

it's server limit, not account limit.

[urban kernel]

One thing that is frustrating about this... I just ran a parallel agent run, and burned through credits only to come back to a bunch of errored runs. No warning, or early cancel, they all just ate through them ):

[tawny island]

ouch!

[frosty zealot]

I need ya'll to @tibo on X with this stuff so we can get an early reset

[tawny island]

that is weird, when you pay more you should have higher limits of a given model imo

[urban kernel]

Yeah. and this was my first run for the night, so 11% of 5 hour allowance isnt the end of the world, but nice to have been avoided

[deft sable]

What model are your subagents using?

[urban kernel]

I used mini 5.4 high, and then it errored, then I migrated to 5.4 xhigh, hoping it would be available. thats why it used so much

but both, were down

[short linden]

it's just temporary unavailable, you can try send the second promt like... 'retry the implement feature"

[tawny island]

yes

[short linden]

I'm still in mad this too, sleep and wake up to re-send the promt... and sleep again 😭

[urban kernel]

We understand it too. like no server has 100% uptime. but like the discord shows the service is still up. maybe codex should have a yellow dot by now and cite potential outage

[deft sable]

It's usually around this time I'm getting those capacity notices too. I guess it's peak hours for Codex

[frosty zealot]

https://tenor.com/view/trump-china-gif-284632875632280687

[urban kernel]

I guess since its almost 6pm in china, everyone is getting off work. (idk about anything outside of America tbh)

[velvet wren]

China is not a supported country for OpenAI products as it happens

[short linden]

china trying to steal the AI intelligent by spamming the messages.

[urban kernel]

very interesting. And I know I am not able to access the Chinese models I would like to. cough cough seeddance

[deft sable]

It could be India. I think it's around noon time there

[urban kernel]

distillation do be effective tho

[frosty zealot]

I found that out when I was putting an app on the app-store I had to disable mainland china

[oak trellis]

ok about to buy now 100 usd plan .. claude code opus is unusable ! .. please feedback about the 100 usd plan

any feedback for teh 100 usd 5x plan .. related to limits etc ?

[frosty zealot]

im on 200 so not sure

[short linden]

it's just "plus plan limit" with 5x, 20x .

[frail meadow]

after threatening gpt 5.4 I finally got it to get itself together

[orchid plume]

I can't comment from my own experience as I'm also on Pro $200, not Pro $100, but in theory Pro $100 should be half the quota of Pro $200. Based on my experience of Pro $200, the quota is very generous at the moment, so I'd also imagine you can get a lot done with Pro $100 as well. As 5x is meant to be 10x at the moment, and 20x is meant to be 20x. These go down to 5x and 10x after May 31st

[oak trellis]

yes i think they will go with that for this month. Anthropic right now is a disaster i can't describe how bad it is. Will update codex now

i lost 2 days of work 🙁 opus poluted 1.5 billion rows .. some rookie mistake

its insane how stupid it is

[orchid plume]

oof yeah well I've been hearing Opus 4.6 is not what it originally was lately, acting dumb and such 😄

[oak trellis]

almost broke my screen .. that mad i got

[oak trellis]

uff .. can't describe how bad .. its like you running a cheap model literally. It lies too ..

[orchid plume]

😄

[oak trellis]

literally useless ..

doesnt follow instructions .. you command "don't do that" in one prompt and it literally does it! Then apologize

[orchid plume]

I remember when I was a Claude fan/user, their cycle of models. Often nearer the launch of a new model the previous generation would suddenly appear to be less intelligent and require more handholding

Codex, no such experience

[kind jay]

Use 6 sub agents
“Where are all my tokens gone”

[orchid plume]

I hope this week we get the new image model and Spud though, and with those two combined, GPT can actually make nice frontend UI

[kind jay]

Do you ever have anything productive to add?

[oak trellis]

its exactly like that. They just throw away their userbase with that. i wouldnt use it in a company. You have a workflow and then they just nerv it and entire code is garbage .. i mean it tricks you because it was good you dont change your workflow and then it just degrade start lying .. then you start production even test .. it lies .. its just on all levels garbage

richtig schlimm

[frosty zealot]

Going to bed I’ll entertain you tomorrow

[kind jay]

Gram Schmidt

[oak trellis]

Deutsch ? haha

codex all the way now !

[dreamy flare]

When is Codex fixing their usage limit degradation?

[short linden]

What ? It token base usage... so what's wrong ??

[dreamy flare]

I hit the limit 10x faster with Plus than a few days ago

(the weekly limit)

[short linden]

yes, it's token base usage, previously still credit base usage and only "paid credit" will be token base, now both paid credit and subscription usage will become token base usage.

[vital shoal]

For ur $200 plan, how many pro prompts can u send? And if u know antone with the $100 plan, could u ask them how many pro prompts they can send?

[exotic cave]

That means we can know from PRO $100 users how it'll be for the $200 when the promotion is over.
Though their site makes things more confusing, in the pic they show x5 and x20 without the promotion. So who knows what the $200 will be after the promotion.

[orchid plume]

not sure how to measure "how many prompts" because it's not really a prompt based limit, it's pretty much token based and not disclosed. However what I can say is that using it even with some subagents throughout the day and early evening, I'm lucky if I lose 5% of my weekly quota each day at the moment. GPT-5.4 (high). I don't use xhigh as I feel that it can overthink sometimes

[orchid plume]

yeah Tibo's post on X made things confusing. It's advertised on the plans page as 5x and 20x, so naturally many thought it meant 10x and 40x during the 2x promo. But Tibo then says it actually means 10x and 20x

[cedar skiff]

not even sure its token based either because tokens have a variable cost.

[orchid plume]

there's most probably an underlying credit, given that you can buy credits, and consumption is based on the aspects the API considers too (e.g. cache input, normal input, output)

[exotic cave]

This is even more confusing it seems like only Pro $100 is x2.
Source: https://help.openai.com/en/articles/9793128-about-chatgpt-pro-plans

From that it seems Pro $200 will stay 20x.

[orchid plume]

yeah I saw that too, the docs and plans page aren't clear at the moment 😄

hopefully this week they will clear it up

[surreal crater]

Can I use Claude Code harness with Codex sub?

[wicked briar]

can I drink mango juice in the wine glass

[cedar skiff]

That means running 5.4 in claude code. I'm sure someone does it.

claude code harness is pretty bad tho

[orchid plume]

true, I was horrified at the number of issues Codex found when I did a few code reviews on the leaked source for fun

[cedar skiff]

It benches at the bottom of the list in performance for it's own models

The only reason i can think of using it is because all your prompting layer is embedded in its system.

[surreal crater]

Yea I saw the terminal bench rankings. I was just wondering if it's possible and if anybody has done it.

I am trying to move off CC to Codex and it's nice $20/month sub

[wicked briar]

is pretty much same, infact codex cli doesn't flicker

[surreal crater]

Hmm alright I'll give it a try. I'm sure it would take a day or so to get used to the differences. Shifting all the commands and skills would be another thing lol

[velvet wren]

I've moved off Codex and back to Cursor

[surreal crater]

Oh, why so?

Haven't been on Cursor for 6 months. I did hear that Cursor 3's UX is good

[velvet wren]

I can't trust Codex with UI, it's awful at UI

[cedar skiff]

Give a ui framework and an idea image to work with

[velvet wren]

been there, done that and it still sucks

Codex is great on back end, terrible at UI

[wicked briar]

yeah i also have claude 5x for ui

since codex is garbage at UI

[plucky halo]

Unless you REEEEALLY like navy gradients with white text

[wicked briar]

they need more data and training

[cedar skiff]

minimax and glm do nice ui

much cheaper as well

[wicked briar]

are you sure?

might unsub this claude code

[cedar skiff]

they have been ok for me

[wicked briar]

because its so freaking slow

because claude has a memory of a goldfish after compact starts doing random things that it already finished earlier

[deft sable]

Kimi as well (at least on my limited testing)

[cedar skiff]

I forgot about kimi

[wicked briar]

good as in like pre nerfed claude 4.6 state?

I know these models are distelled claude models

with chinese touch

[velvet wren]

using what harness?

[cedar skiff]

I used CC and droid

I like droid

[wicked briar]

rich guy

[cedar skiff]

nah that is the cheap option

[wicked briar]

droid is api based

api = rich

[cedar skiff]

it has no sub byok

you just just get minimax and use it

[wicked briar]

hmm

looks promising

[deft sable]

I haven’t really compared them side by side yet, so I’m not totally sure. But from limited tests, Kimi can sometimes give really good designs even with a vague prompt.

[wicked briar]

nice

[deft sable]

But at the moment, I'm using Gemini (CLI/AG) since I have a Pro sub there. Google nerfed the limits so hard that it is very slow and unusable. What I do now is instruct Codex to handover the frontend tasks to Gemini. It works well on my use case (Next.js site).

[short linden]

if you want to know the limit, try using API from "https://platform.openai.com/" and prompt create the simple app and compare with other AI model, for cost usage.
example one prompt could go $1-$3 that look same with subscription model, but have more bucket than API $20.

Anyway subscription plan is better than API, since API design for MCP server.

[surreal crater]

dang yea I haven't had a single good experience with Gemini CLI and AG because of it's reliability :/

[wicked briar]

im having lot of issues lately with codex pls if anyone from the team is seeing this fix , ty

[stable obsidian]

stitch mcp

[wicked briar]

started new session is working fine

was most likely corrupt session

😭

was 70days old sessions tho

[unkempt cipher]

One issue I face with codex is with merging code from parallel work streams into one branch. Any good methods to manage that?

[wide schooner]

where can i find security? i got pro now

@wicked briar u got pro, wheres security 🤔

[deft gyro]

codex ui is the most

awful

horrible

thing

oat

It’s gemini 3 flash level

[velvet wren]

I've gone back to Cursor, I need frontend and Codex sucks at UI

[unique spade]

tell codex in a different session to investigate the session and repair it. it can do that

[deft gyro]

which is a bit odd because gpt 5.4 does really well on minebench

[plucky halo]

https://chatgpt.com/codex/cloud/security/findings

[plucky halo]

Remove the /findings ?

[wide schooner]

yeah still

[wicked briar]

no idea

I never used

[wide schooner]

ur on pro100 or pro200?

[wicked briar]

200

[wide schooner]

then you should have it for sure 🤔

[orchid plume]

isn't the security feature in the codex section on the ChatGPT website? Pretty sure that's where I recall seeing it, hope it helps if so

[wide schooner]

thats where it should be but im on 100$ plan and it aint there

[orchid plume]

hmm interesting

[wide schooner]

is it region restricted? or cyber restricted?

[orchid plume]

for me it's available but I didn't use it as inputs and outputs could be used for training last I checked

pro 200 but also cyber verified

[unique spade]

i m not sure tbh, when i switched to gptpro 200 around 2 weeks ago, i had it there for a while i even scanned 2 of my repos

now i switched to 100 after they released, and after some days i thought to check security tab again, but it s not there anymore

[orchid plume]

my guess is that for whatever reason it's restricted to pro 200

[unique spade]

so i assumed it might be some 200 only perk

and i think you do need to get cyber verified

[unique spade]

or team accounts. those have it too

or how they call now that business tier

[full cosmos]

I hit my daily limit after one prompt.

So if I go up to Pro 100, I can get a whole 5 prompts of that size?

Did I accidentally subscribe to Claude and not realize it?

[velvet wren]

the limits are pretty harsh

[full cosmos]

They're completely stupid

I finished what I was doing with a free Chinese model because what I was working on didn't require privacy. It worked great. I could use the free thing more than my paid OAI account.

[velvet wren]

what harness were you using?

[full cosmos]

OpenCode

[velvet wren]

I've dumped Codex and gone back to Cursor

[orchid plume]

are you on plus plan?

[full cosmos]

Yes, and it worked great until the new limits hit. I use it only a bit at night and don't think that's worth shelling out any more.

[orchid plume]

ah plus's 5 hour limit got reduced and it's not 2x. Pro 100 and Pro 200 are far superior for quotas

plus is more like a plan where you can try it briefly or do very simple stuff

[unique spade]

also now code review is included in main qouta, so if you have that on and actually reviews the pr's it eats some

[orchid plume]

true that

[full cosmos]

I don't care about the other plans if Plus is nerfed that badly.

I did not have it do code review. It literally implemented a couple classes from a spec I supplied.

That was enough to hit the limit

[unique spade]

personally the boosted 100 seems to be right fine with my current usage, but this is 10x now, which will mean 200 plan from next month

[unique spade]

technically plus plan was under active 2x promotion for february and march.
so it wasn't nerfed, it came back to standard 1x

and i think they said they will reduce 5h limit too, right?

sounds like you are directly affected by having a too big task that eats that 5 hour limit, not the weekly one

[orchid plume]

true it was under the 2x promo, but they did state on X that they also reduced the 5 hour quota for Plus separately

[unique spade]

yea that must be it imo.
nerfed 5hour could certainly hit you if you give some heavy task straight during the first prompt

my personal view is that codex still has best qouta.
i did some tests with antigravity on the 20 plan to have opus do some frontend in paralel
and at 3rd iteration pass i was already out of qouta and had to wait like 5 days

[tiny fulcrum]

is anyone using enterprise based on credits? just interested in how a 5x Pro subscriptions would stack against the raw credits price

[unique spade]

related do that is there an easy way to actually check how many tokens you are using with plus/pro?

[tiny fulcrum]

well, this is kind of the problem I don't find this info anywhere ;/

[unique spade]

i mean this chart is cool. but it's all relative to itself.
and i have no clue what those % mean in actual tokens used

[orchid plume]

no but there are tools you can either find on GitHub or even ask Codex to make which analyse your session files to determine how many tokens you're using via Codex CLI for example

[unique spade]

that must be some kind of guesstimate...since i don t think codex has access to actual tokens spent during COT reasoning

[orchid plume]

not really looked into it myself but it probably isn't 100% accurate, not sure what the session files contain exactly

[full cosmos]

Okay, but I was able to have multiple long sessions in the span of a couple hours and never hit limits. Stuff waaaay more complicated than what I did last night. If my limits were cut in half, I would be able to do the basic stuff.

[silver dew]

Cursor sucks for me
anyone else feeling the same?
Cursor with GPT-5.4 High feels like Claude Code

[unique spade]

what you hit was the 5h limit right? or the entire 1 week limit?

[velvet wren]

I've just gone back to Cursor from Codex

[full cosmos]

Yes the 5h limit

[silver dew]

Why + What kind of project are you working on?

[velvet wren]

at the moment I am working on a music streaming platform

[orchid plume]

oh nice

[silver dew]

nice
but why cursor instead of codex?

[velvet wren]

Codex sucks at frontend

[orchid plume]

sadly a fact at the moment 🙁

[silver dew]

(I'm not an openAI employee attempting to increase revenue)

[orchid plume]

hoping this week's big reveals might improve the frontend aspect

[deft gyro]

lol

[silver dew]

truth but Cursor so sloppy recently

[deft gyro]

gpt 6

soon™

[orchid plume]

tell me about it, the cryptic teasers and vague posts have burned me out over the last few days

[deft gyro]

i mean like

idk

need to see price

b4

i get excited

[velvet wren]

I got frustrated with Codex this morning and fired up Cursor, got everything I couldn't get done over the weekend in a couple of hours with Cursor

[unique spade]

they nerfed that on plus plan. i'm trying to find where i ve read about it but didn t find yet

it was something specifically abotu nerfing 5h limit, "to balance usage across the week" so people don t eat the 1 week cap that fast

[deft gyro]

yes

this was something that tibo said

[orchid plume]

just curious, is that still using GPT-5.4 or whatever via Cursor? not Gemini or Opus for example

[deft gyro]

The weekly usage remains unchanged

[unique spade]

yeea, i m checking his posts but didn t find the one where he said that lol

[full cosmos]

As someone who has literally never hit the weekly limit, this is extra annoying

[deft gyro]

Idk where it is

[velvet wren]

I'm using Composer 2 on Cursor

[deft gyro]

wompwomp

plus quota is prolly gonna be decreased further xd

[silver dew]

what about claude

[orchid plume]

it was posted by the OpenAI Developers account on X, not Tibo, I believe

[velvet wren]

I hate Claude

[full cosmos]

My subscription is going to also get decreased to nothing soon

[deft gyro]

soooooo kimi k2.5thinking xd

[silver dew]

why?? I know why but why

Claude = High temperature slop

[velvet wren]

Claude is like a brown suited accountant who can't fit in to a group of guys with Harleys

[deft gyro]

claude is gud

just

too expensive

claude ui is very good

but g31ph is also fairly decent

can chatgpt just take my act for me

ong

[plucky halo]

"That is materially better.
Now I see the whole picture."

I swear I get one of these every other prompt with old claude

[silver dew]

Codex became like 2x lower on limits
I have 3 subscriptions of the 20$ and I constantly hit limits after 3 hours that I used to hit after 2 days

[velvet wren]

when it comes to coding, I want the harness to stay out of my way while I iterate. Project I am working on now requires solid frontend capability and Codex doesn't cut the mustard

so I went back to Cursor and got my work done

[deft gyro]

wompwomp

u get like

300 dollars of compute per week on plus plan

i know

because i measure

even if that got cut in 4

it still is cheaper than api

[livid relic]

is snitch how everyone is doing things?

[orchid plume]

snitch? or did you mean stitch?

[silver dew]

300$ of compute for 20$? OpenAI is not sustainable

[wicked briar]

🤓

not when everyone in the world is using it

[deft gyro]

Yes but no

That’s just not how it works I am oversimplifying it

[plucky halo]

Not everything is sold for profitability though

[livid relic]

stitch sorry

[silver dew]

There's no way you can make it up
it's just a path to hell

no matter their enterprise deals

[wicked briar]

claude is making profits at the cost of users frustation

[silver dew]

OpenAI also lost the enterprise space lol

[wicked briar]

losing users

[wicked briar]

source - i made it up

[orchid plume]

it helps but it's not the ultimate solution to the poor frontend problem sadly

[silver dew]

Overleveraging until you flop. and it's not tech companies it's venture-backed deep-technology companies.

[silver dew]

Gemini is a better entreprise AI

[wicked briar]

ok you are trolling now

[velvet wren]

no

[silver dew]

No it's true it's literally integrated into everything

OpenAI wins only at consumer

[unique spade]

gemini is a terrible coder