codex-discussions

[sleek spoke]

ah I see.....

[lean lark]

AGH!! Another #announcements !!

and this time the info isn't a 404. 😜

[wicked briar]

waiting for 5.4 codex

[boreal holly]

But yeah, the perfect "I want to go from 100% to 0% weekly as fast as possible" configs would be

model="gpt-5.4"
model_reasoning_effort="xhigh"
model_context_window=1000000
model_auto_compact_token_limit=900000
service_tier="fast"
[features]
multi_agent=true

That setup will allow codex to read a couple of files and then you get the sweet sweet 0% weekly

[exotic cave]

Don't think it's going to happen, and personally i am not into codex models anymore.

A coding agent should have general knowledge which can help it tackle various projects that mimic real life business logic not just how to code in x language.

[lean lark]

The codex models aren't "just that".

[wicked briar]

uh because 5.4 is very token hungry I used 25% usage in one day on pro plan

I think plus plan users are already done with weekly limits

[worthy tide]

Coding with 5.4 means I can chat about the product with the model. For coding games, it can help with making gameplay decisions. Codex is too uptight and focused only on code. I prefer this as these two areas are often interleaved

[lean lark]

Today, we’re releasing GPT‑5.4 in ChatGPT (as GPT‑5.4 Thinking), the API, and Codex.
-- https://openai.com/index/introducing-gpt-5-4/

[cobalt junco]

bro codex just coded me a gateway for iot devices lorawan hooked up w humidity sensors in the ground to monitor crops on my farm

what the hell did they train ts on

[exotic cave]

5.4 is not token hungry scale back to high, don't activate /fast and don't use multi agents.

[wicked briar]

pentagon

[cobalt junco]

makes sense

[worthy tide]

If they can make a general model that is as good at coding as a specialized model, then the general model is just better in every possible way. Seems they have achieved this from what I have experienced so far

[eternal quartz]

Applied for the oss thing, i really hope i will be chosen🥲

[cobalt junco]

facts

i like openai a lot

been testing their models since 2022

[exotic cave]

Yep that's what i am saying, a general model is much much better at understanding business logic and have the necessary general knowledge that can help it tackle various projects.

[weary stag]

Any chance codex app comes to Linux 😅

[exotic cave]

They did say something about it, but i am not sure 100%.

[earnest rock]

Anyone test out or use 5.4 yet?

[warm pilot]

it's my daily driver … day0-driver 😉

[exotic cave]

5.4 is fast and furious xD

If they continue with this trend i am afraid that by the end of the year we will look back at 5.4 and think it's primitive...

[analog ingot]

Excited about Codex OSS, while it builds up understanding of large codebases and their security surfaces over time, is there a plan for how that accumulated knowledge is governed? Not worried about today, more thinking about the long term. Codebases change slowly, but the insights Codex accumulates about their architecture and weaknesses compound. Curious how OpenAI thinks about the stewardship of that kind of knowledge.

[lean lark]

This is exactly what 5.4 is already.

GPT‑5.4 brings together the best of our recent advances in reasoning, coding, and agentic workflows into a single frontier model. It incorporates the industry-leading coding capabilities of GPT‑5.3‑Codex
-- https://openai.com/index/introducing-gpt-5-4/

[potent mason]

You should use medium it's basically the same as high according to benchmarks and much faster

[lean lark]

I was just reading through the 5.4 doc. It's usage is very nuanced, needs to be tuned for specific use-cases. Most people won't get that. Unfortunately even developers have lost their sense of nuance and detail and expect things to "just work" without configuration. As this tech evolves people need to go with the flow. Yesterday we needed a different model for code, today we don't. Yesterday context was limited. Today the limit has been raised but there's higher cost for using it (which is one of the reasons why it's not the default!).
Read the docs. Understand the tools.

[worthy tide]

That's what I'm saying, dude. Someone said they were waiting for 5.4 codex and I was explaining why it's not coming or even desirable ...

[noble quest]

Hello there! I just want to know what is your feedback about using Codex 5.3 vs 5.4? I'm currently using 5.4 and Codex CLI, and I like it so much! I didn't use the 5.3 codex, but the 5.2 codex was very bad with me! Any feedback from Codex 5.3 users?

[lean lark]

You could read up a few lines for comments already made, no?

[boreal holly]

I hear medium is just "high only if necessary" mode

[boreal holly]

Tbh it's not unreasonable to suggest spending >=50% of your time researching Codex and the rest actually using it.

It's one of the only tools I've ever used where productivity is disproportionately higher if you spend more time studying the tool than using it. It's like a wrench that builds the car if you stare at it hard enough

[meager sapphire]

Anybody using codex as a sysadmin tool to manage on prem hosts and cloud instances.

[exotic cave]

You want to be very very careful handing this things to AI currently. it can be done and successfully but be careful.

[boreal holly]

Yes, with approval_policy = "untrusted" for that agent it works pretty well

[meager sapphire]

Its phenomenal does vulnerability scan and can patch and all. No need for qualys 🙂

[viral jewel]

what is the point is recommending WSL but having codex deny access to WSL? why is this being so difficult

[lean lark]

You're just encountering a config-specific issue. It's not telling you that Codex can't use WSL. I have it right in front of me, as many of our colleagues here.
Make sure you are specifying the correct distro and that you have installed a distro that can use Codex. Ubuntu v24 is excellent.
For protection against disaster you really want to restrict Codex use to a "codex" user that does not have root privs. Set the default login to that user and if you need to login as root or otherwise, add the "-u" option on your wsl CLI.
I've written about this recently in this channel, look for my comments that include related keywords.

[viral jewel]

Question for you. I am currently testing the Codex App. Lets assume I successfully clone X repo in WSL ubuntu distro. And change agent env from windows native to WSL. Do I still pilot the agent from this app? Or is the piloting experience all done through WSL instead?

[lean lark]

You can operate on WSL content from the app and also from the CLI ... that's one of the benefits of the app, it adds a more centralized location for processing as well as allowing for multi-agent processing in the same environment.

Side note: If you want to clone the repo and work on it via the remote VSCode extension and the Codex extension, you may find that you need to clone from the integrated terminal. That ensures that the terminal is using the correct perms and GH auth. If also overcomes a potential issue where direct CLI access may not have the same perms as the IDE. Disregard if this doesn't make any sense to you yet.

(Disclaimer: I haven't installed the app yet. For that specific component I'm just going by what's been said in this channel and what I've read.)

[fast garden]

i got codex-5.3 spark as a plus subscriber

anyone else got that too?

[viral jewel]

thanks i'll test this out. i think when the Windows App was released, the sandbox introduced ACL perms that resulted in default no write settings to .git (maybe other directories as well, but i noticed them in .git for my repo). Do you know if this was by design? Prior to the windows app release, I was using the VScode extension without issue to read and write to .git.

[mint scroll]

Does someone know how to get codex security on my business subscription?

[frosty zealot]

I ust cant comprehend how GPT will implement code, but then the review will find bugs time and time again, yet if its aware enough to identify those bugs, why did it write the code that way in the first place

I added

        able.
    44 +
    45 +# Incremental Review Rule
    46 +
    47 +Implement in small, reviewable checkpoints instead
        of one large final diff.
    48 +Prefer slices that can be validated and reviewed in
        dependently before continuing.
    49 +During implementation, regularly run targeted `/rev
        iew` passes on the current checkpoint and fix issue
        s before stacking more changes on top.
    50 +When the feature or task is complete, run a final `
        /review` against the full diff and continue autonom
        ous review -> fix loops until no material issues re
        main.
    51 +Do not defer most refactoring, regression catching,
         or cleanup to one large end-of-task pass when smal
        ler in-flight review loops are practical.

To my AGENTS.md, see if it offers any sort of improvement.

[bitter parrot]

Is there a channel for Codex feedback?

[frosty zealot]

https://discord.com/channels/974519864045756446/1053565716500074526 https://discord.com/channels/974519864045756446/1070006915414900886

[boreal holly]

Was just gonna suggest that! Also it's not unusual to have short burst code reviews. It's better than 16k line change, 25 review comments, another 16k line changes, another 25 review comments. If you do it in small bursts it's good for the reviewer and the editor, and most of all the diffs per commit

[frosty zealot]

Dude I just had a 120 comment PR

of like @codex over and over

Im tired of it lol

[torn mist]

I'm seeing the same issue, maybe they are rolling it out slowly still

[main nimbus]

I think the same. I had a similar experience when they launched Codex in the cloud.

[mint scroll]

How long did it take?

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

Wait.. I just got access.

[main nimbus]

lmao same

it shows as aardvark in chatgpt

[mint scroll]

Yeah I saw.

That is the old name of it too

[main nimbus]

I love that instant swap when we are discussing it 😂

[mint scroll]

lmao.

Oh we got a quota.

I think that is per 5h?

[main nimbus]

The "Get more scans" button in the header allows you to request the paid version.

[mint scroll]

I don't got that?

[main nimbus]

I think this is the total repositories it can monitor.

[mint scroll]

It are scans?

Maybe if I just remove a repo?

I just have to figure it out ig.

[main nimbus]

yeah a bit confusing ngl. I can't launch multiple scans on the same repo for different timeframes, maybe at once which makes sense.

[mint scroll]

I just launched it on 3 projects.

[main nimbus]

Same repo

[mint scroll]

Different ones.

Wait.

It seems like you can update repo's without rescanning?

So I will look what happens when removing a repo?

[frosty zealot]

I am not going to make it till Wednesday with 33% left

Which means im gonna have to use poopus

[orchid plume]

not sure how many here are aware, but Tibo about an hour ago posted this on X (source: https://x.com/thsottiaux/status/2030017068540776891):
"We are investigating reports of higher usage drain than expected for Codex when WebSockets are enabled, the team is investigating and we will provide updates as we go"

interestingly I also have websockets enabled and it did feel like my usage was draining a bit faster than usual, but I thought that was probably just because of GPT-5.4

[feral arrow]

-_-

[lean lark]

I'm still using 5.3-codex, no token issues with any reasoning levels. Just sayin...

Arg, new 404 page from https://chatgpt.com/codex/security/findings
That's an impressively useless use of technology. 🙂

[boreal holly]

Gotcha! Add that to the list of "give me 0% weekly ASAP" configs. They're not in the official docs! But Here they are:

[features]
responses_websockets=true
responses_websockets_v2=true

They're false by default in codex-rs/core/src/features.rs:745 & 751 as of 0.110.0

[boreal holly]

It's so secure you can't even use it!

[ember venture]

Do you need both those settings for websocket support?

[boreal holly]

I'm not sure, but by default it uses the tried and true SSE transport

It's possible you need both enabled, or at least responses_websockets

Btw I am not recommending turning them on as they appear to be a reason one of the many experimental features or knobs causing people to run out of usage quickly

[main nimbus]

I recommend trying on high. it does appear to overthink problems.

[amber scaffold]

I know that Codex permits uses their OAuth auth for OpenClaw. Does this imply Codex permits using their auth for any third party application? I've made a computer use agent for local macOS, but naturally the API is expensive, so I would prefer that.

If you are expecting AGI you will be disappointed. It's an excellent tool that can really accelerate your work.

[main nimbus]

They absolutely permit it. I don't have the sources, but this is well covered. A twitter thread somewhere.

Theo (t3) has discussed this recently when comparing OpenAI and Anthropic.

[orchid plume]

Codex CLI is open source, which means you can easily re-use the authentication aspects of it as well in your own projects 😉

[amber scaffold]

That's really helpful. I'll let deep research find that thread for me

[amber scaffold]

I know it's technically possible, but I really don't want to risk my OpenAI account, since it's pretty ancient (pre ChatGPT)

[boreal holly]

OpenAI hired the guy that made OpenClaw because of how much they loved it, and announced practically official support for it

[main nimbus]

look up discussions over the phrase "app-server" for the auth discussions

[soft glade]

issues i'm experiencing:

1. apply_patch is always rejected (windows)
2. work trees do not show up on left hand side and so starting a thread as a work tree makes it difficult to find unless you're in the "chronological list" view instead of "by project" view
3. i can't select anything but 5.4 and high in a thread. the dropdown works but any option i select does not stick

[spiral gorge]

I’m at the edit step now. I’ll try the normal patch path first, but if the patch tool flakes again I’ll fall back to direct file writes so this doesn’t stall.

im using windows 11, powershell. is powershell HORRIBLE for codex app?

[boreal holly]

Restart computer and see if that fixes it

[spiral gorge]

"I’m switching to direct file writes now. The patch tool failed again, and the changes are substantial enough that waiting on it would just waste time."

[boreal holly]

powershell HORRIBLE

Affirmative on all counts

[spiral gorge]

then why did openai set it as default for windows codex app

[boreal holly]

There's a South Park episode about voting that describes it best Powershell ardy installed, wsl not so much

[main nimbus]

"windows" -> I see your problem

[boreal hawk]

Wich codex model is now the best?

[frosty zealot]

Do you guys think that the high usage right now is due to maybe no cache?

Im assuming being a new model it would have to build the token cache as it gets used

[main nimbus]

Codex model is 5.3-Codex. The best model to use IN Codex is likely 5.4.

[boreal hawk]

Aright thx.

[lean lark]

I think you're hitting exactly that issue that I documented where VSCode is running under different perms from the connection used from the Windows app. Honestly I don't talk about this much because I think some peeps are gonna call me an idiot for not expecting code . permissions to be different from wsl -d .., but it's one of those things that needs to be understood when we have multiple clients hitting the server with different creds, profiles, aliases, etc. Good Luck!

[lean lark]

It's so secure we can't even tell you about it!

[boreal hawk]

Guys do you know where can i try the new Codex Security becaus i cant find it

[frosty zealot]

Its for enterprise only I believe

[boreal hawk]

Ahh okay

[lean lark]

Now available: https://openai.com/index/codex-security-now-in-research-preview/

We plan to expand the program in the coming weeks so more maintainers have a direct path to better security, stronger review workflows, and support for the open-source work the ecosystem depends on.

Like other initiatives like this it seems we need this third party, OpenAI, to decide if a package that we maintain has sufficient merits to justify allowing the developer(s) to use the security software. I really dislike this common practice. I understand it but grudgingly disagree with how these companies apply it.

[smoky ermine]

I don't think that's how context cache works. It is usually stored per user for a short period of time(one session) AFAIK. And is actively reused within the same session. This is why your messages close to the context limit don't take eternity(a bit exaggering here) to answer.

[frosty zealot]

I thought there was a server side cache too, so the model didn’t have to recalculate things it’s already been asked

[lean lark]

You guys are providing the definition of "context" ... context is cache. It's pruned just like cache.

[junior knot]

https://voratiq.com/leaderboard/ not quite as good

[lean lark]

Um, except that with compression it gets summarized up-top.

[boreal holly]

Yep, it's 15 minutes TTL, and it's the KV Cache. It hangs on to tokens to it doesn't have to prefill inputs from 0 to n

[smoky ermine]

There is, but it's still limited to your machine. And it has to be the exact context for the cache to be reused.

Caching only works if two requests share the same prefix and land on the same machine

https://developers.openai.com/cookbook/examples/prompt_caching_201/

If they were to store cahce for all messages ever, there wouldn't be enough storage memory in all the world, as they store the actual latent space, or whatever it's called, that takes GBs of memory per cache.

[exotic cave]

As i said high it is.

[boreal holly]

Haha yep, a 1M token context window can be well over a terabyte of VRAM in most cases

Conversations are cheap, KV is very very expensive, almost (if not absolutely) exponential in comparison

[simple star]

5.4 is making stupid decisions all day long...

I have been working with it all day, and you cannot imagine how many times it has told you. "Because I made a mistake"

[smoky ermine]

Can you provide some examples?

[simple star]

They are very case specific

Like duplicating code to solve potential bugs, or changing API contracts

5.3 would never do this

[smoky ermine]

I see. For me it was pretty accurate.
As opposed to to 5.3 it seems to try to understand the project context better before making changes.

[winged kettle]

Code

[simple star]

It's being a total disaster for me, really

@winged kettle what?

[winged kettle]

What

[simple star]

blocked

In any case, I am not really having a good experience

[boreal holly]

I noticed 5.3-Codex is really good at following strict processes. 5.4 forgets to run static code analysis before commits, so I kinda see what you mean

[winged kettle]

Sobre o code convite sora

[simple star]

I might actually move back to 5.3-codex...

[boreal holly]

To me it writes much better code but idk, still discovering the right way to enforce strict processes with 5.4

[simple star]

I feel like it is working a lot against me

[winged kettle]

Sim sim

[smoky ermine]

I feel like 5.4 gives you opportunity to steer it towards the right direction more than 5.3 that would basically make all the changes and research silently and you wouldn't have any clue what it's doing/thinking.

[winged kettle]

Sora

2

[simple star]

also /fast is a trap. Dont use it. You will melt your quota in 3 days

After April, it is going to be a disaster

[orchid plume]

given there's still an ongoing quota consumption bug, apparently anyway, it's definitely not advised to use it 😄

it's still unclear to me as to what /fast does for Pro users, is it separate from the "priority processing" feature of the Pro plan

[simple star]

it fasts you 😛

No idea, I guess it drains your wallet faster 😄

[orchid plume]

😄

[simple star]

By the way, looking at the Codex App... do automations work with CLI?

[orchid plume]

I assume it uses the CLI to faciliate them yeah, well, the codex-app-server aspect of it

[simple star]

It uses the CLI to run Codex, for sure. But managing the automations, is App-only?

[simple star]

Breaking

[boreal hawk]

Do you think codex 5.4 can creat good design bc i need a ki that can make me good Pdf design

[simple star]

@boreal hawk Graphic design? no

[boreal hawk]

Is there a ai that can make Graphic Designe?

[simple star]

Yes, but this is not the place to discuss it

[boreal hawk]

Dms?

[vital light]

Codex Security looks cool @winged ore
Was just reading a user’s comment on another platform about this and folks were dragging the idea lol.

[boreal holly]

Gotcha. Well hey, at least there is a community elsewhere that you can talk about Claude. #ai-discussions or the Anthropic server

[simple star]

@vital light Is there a way to use it without the Codex App?

[violet anchor]

I think my biggest frustration with codex is the continual hand holding it demands. I can tell it 10 times in a row, to not stop until it's done something to 100% completion. Every single time it stops short of 100%, admits it stopped short, apologizes for not doing what I asked, then asks if I want it to keep working towards 100%.
Like yah, just like I told you to the past 10 freaking times heh

[simple star]

@violet anchor That is true

[lean lark]

Sounds like you guys need better AGENTS files and a better sandbox. When I see notes like these I tend to think "user error".

[north bison]

Sora isn't working for me right now, do you have any idea why?

[lean lark]

Wrong forum

[simple star]

How did I not think of that?! /s

[toxic torrent]

i'd be very happy to test the remote codex capabilities

and have plenty of use cases

[boreal terrace]

Can I use chatgpt subscription with openclaw model?

[tight chasm]

Does the gpt codex mac app has a checkpoint feature where i can undo/redo progress locally instead of utilising a git worktree

[unkempt falcon]

Does Codex Security Application work with VS Code extension?

[simple star]

@boreal terrace wrong channel

[boreal holly]

Oh yeah, quick life hack. If a Codex model like 5.3 or otherwise wrote all of your AGENTS.md and SKILL.md files, make 5.4 rewrite them with your guidance. This is the quickest realignment you can do today to help with weird 5.4 misbehaving

[rocky fog]

could use another limits restart soon 😂

[simple star]

People hoping for server bugs, so OAI resets limits lol

[velvet wren]

I wouldn't complain if that happened 🙂

[simple star]

calling all the hackerz 😄

I am seriously worried at how I will survive in April 😛

[cedar skiff]

you got one month to make fortune 500 start up

[simple star]

"Hey Codex, make me a fortune 5000 start-up"

[frosty zealot]

Something will have to give come April, I don't normally complain about usage, but something is eating, and this is simply just unsustainable for the price to usage windows come april

only thing keeping me afloat is the 2x

[high girder]

make no mistakes

[agile parcel]

My Codex got stuck thinking and cant stop

[simple star]

That happens to humans too

[amber scaffold]

Did you reply to me, and had your reply deleted because you mentioned Claude Code?

[toxic torrent]

5.3-codex > 5.4?

[cedar skiff]

no

[simple star]

@toxic torrent At times, I think yes

[cedar skiff]

5.4 has been better for me i havent looked back

[amber scaffold]

How are people so opinionated on this? If the 5.3 was silently switched with 5.4, I don’t think I’d notice

[simple star]

opinionated?

[amber scaffold]

“Opinionated describes someone who holds strong, often stubborn beliefs and frequently expresses them…”

[main nimbus]

Feels roughly the same to me. The improvements are getting harder to notice.

[simple star]

Im not sure what makes the opinions here shared stubborn or strong

as for "frequent"... well, the channel is called #codex-discussions . No sure what you expected to find here.

[glad wind]

If you look at difference in evals between 5.3-codex and 5.4 they're pretty similar on coding specific tests

[cedar skiff]

5.4 is better at documentation and planning ideas. I also notice my automated code reviews dont flag as many issues with its code. I use code rabbit and codex review.

[opaque ginkgo]

兄弟，去 DeepSeek

[cedar skiff]

So i base my opinion on my anecdotal experience

[rose mantle]

https://x.com/gapdeepry/status/2030108833100558589?s=46

for anyone still on Intel Mac… slight workaround to use codex app

[simple star]

Does anyone know how to access Codex Security?

[cedar skiff]

Get enterprise

[simple star]

It is also available to Pro

But I have no idea where to find it

[cedar skiff]

oh nice, let me get onto that as well o.0

[simple star]

I have yet to find it (I cant believe OAI releases this stuff, and makes it so hard for people to find it)

[main nimbus]

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

[simple star]

@main nimbus Thanks

[main nimbus]

It's called Aarvark if you go to chatgpt.com

[cedar skiff]

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

[simple star]

Sounds like a norse god

[cedar skiff]

You have to use codex cloud

[honest yacht]

what is GPT-5.3-Codex-Spark

nvm

"Today, we’re releasing a research preview of GPT‑5.3‑Codex‑Spark, a smaller version of GPT‑5.3‑Codex"

I stopped reading at smaller

[main nimbus]

Codex Security made me feel better about my vibe coded mess of a chat app. 1 medium issue which isn't an issue, it's a feature. I started the app in December of 2024 as an experiment to see if AI could truly develop an entire app without me doing anything other than directing it. It, by all means, should be a complete disaster of a codebase.

[honest yacht]

I want a model that I have to wait a day while it churns

[simple star]

@main nimbus Try /review and see if you still feel good 😄

[main nimbus]

I mean vulnerability wise, don't make me go digging for the skeletons.

[honest yacht]

which one is better 5.4 or 5.3 codex?

[cedar skiff]

5.4 imo

[simple star]

well, /review will also look for vulnerabilitites

it is all it find me, as a matter of fact

@honest yacht Divided opinions

[main nimbus]

There's a benchmark that shows model's ranking by how many HUMAN hours of work the model can complete. It's getting quite high, despite being 20+ min runs, not hours.

[honest yacht]

I'm guessing 5.4 is a better model and 5.3 codex is better trained on the codex interface and thus more agentic?

[simple star]

"more agentic" sounds more and more like fluff, lately

[cedar skiff]

I used to use gpt 5.2 for documentation, building skills, web searches etc. and 5.3 codex for code implementation and planning. Now i can just use 5.4 for everything.

[reef oasis]

Can’t wait for codex mobile

[main nimbus]

Typically but OpenAI is anything but consistent. GPT-5.4 includes better tool calls and GPT-5.3-Codex's improvements.

No idea if we'll see a GPT-5.4-Codex. If we do, it won't be anytime soon.

[honest yacht]

perhaps, but it still means something, better tool usage, designed to hand-off tasks to other agents in pursuit of a goal that supercedes the agent instructions

[cedar skiff]

5.4 is much better at tool use

it has tool search tool now

[main nimbus]

when o3 introduced CoT tool calls, that was eye opening for me.

[cedar skiff]

Did you have set this up or does it just do it?

[simple star]

It just does

[cedar skiff]

Is it the same security audit the cloud version does?

[simple star]

Probably no

[tall zodiac]

@tacit heath https://x.com/banteg/status/2030061859110293898?s=46&t=9l1KwyEX9uwBxqfL3uX3zA

[main nimbus]

The security side creates a detailed threat model of your application, then scans the changes introduced (likely against the threat model).

[simple star]

It infuriates me to no end, when Codex ignores my questions, and just changes stuff

[tall zodiac]

Give an example

[main nimbus]

I imagine there are questions like "Why is this happening?" and it finds and fixes a bug.

[honest yacht]

codex needs something like claude codes new /loop skill to reissue skills/commands in intervals within a session

[main nimbus]

oh lol completely off

[simple star]

1. He asks me something totally unrelated to what we were talking aobut
2. "Hey, what the... why are you asking me about this suddenly?"
3. Here's the plan

[cedar skiff]

I dont think it does?

codex just uses the skills as needed

It's much more proficient with skills than claude

[cobalt junco]

i just cooked 20% of my weekly limits with one feature request

lol

[simple star]

@cobalt junco medium effort?

[honest yacht]

if you mean claude, it does have /loop

[cedar skiff]

I mean i dont think codex needs it

[honest yacht]

ah ok. sometimes it comes in handy though

[cedar skiff]

it doesnt have the same alzheimers claude does

[cobalt junco]

xhigh

[simple star]

Yeah, take it down to medium, if you want to make it to next week

[cedar skiff]

what does the /loop do i havent used claude for a bit?

[simple star]

I also used to xhigh all the things

[honest yacht]

something like this:
/loop 2m check if the database migration in tmux session "migrate" has finished, and report any errors from the log

[cedar skiff]

so it checks every 2 minutes?

[honest yacht]

within a session, yes

[cedar skiff]

ahh ok

I dont think you need it for skills, you can trust codex/5.4 will use them when it needs them.

[honest yacht]

I prefer claude probably because I am more familiar but I respect codex. I'm trying to force myself to use all of the CLIs. even gemini

[cedar skiff]

try codex app, its really nice.
I am going to give anti grav a run sometime in the next week

I'm on mac though - windows only just got codex app and it's a little buggy from what i can tell

[honest yacht]

I really tried to like antigravity. many like it but I just didnt. I think Ive just become partial to CLI over IDE

[cedar skiff]

since i started using codex app i never looked back, I still use ide for a few things.

when i touch the cli i cringe

[honest yacht]

I feel like we are at a place where the models are differentiated enough that learning all of them is a good move. I like claude's task harness, codex for reviews and complex tasks, gemini for UI and dashboards

[cedar skiff]

Yeah maybe, for the time and effort i think it might also be worth filling gaps in one model

Like for ui, just adding a skill for codex instead of configuring gemini

[raw hill]

sfunny i felt like we were in a place when it was worth using different models, and now they are all converging

[honest yacht]

yeah agreed. gemini cli is really behind. the gemini app actually has a nicer UI builder

[honest yacht]

When I run a plan or code review through codex and claude they still seem to complement each other. gemini seems redundant if anything, although it can create an amazing dashboard

[raw hill]

yeah maybe. the alternative is just asking the same llm for counterpoints, alternate options, devils advocate etc

[cobalt junco]

bro

im not even joking

what the hell

codex is too good

[honest yacht]

1m context is nice sometimes. notebookllm is great to cheaply create a 300 source RAG chat

[cobalt junco]

never going back to claude.

[cobalt junco]

i do this sometimes, codex writes, opus reviews

[honest yacht]

trying to reconcile your last two statements

[cobalt junco]

i meant claude code

i use opus in windsurf if i need a quick review

[honest yacht]

what was your issue with claude code? I see enough complaints to give them legitimacy even if I dont experience them

[cobalt junco]

tbh i have no issue with claude code, but gpt 5.4 xhigh writes impeccable code and requires less loops. the good thing about claude it writes fast and you can reloop, but with gpt 5.4 sometimes its just one prompt and its good enough

[cedar skiff]

claude requires more effort to get the same results, you can't trust it to do the task as expected with out more prompt layers.

[cobalt junco]

^^^^

[cedar skiff]

It also has a terrible memory, like it wont use skills proactively after the first few turns

[cobalt junco]

That's true as well, context rot % is higher in claude

[granite lava]

Anyone have feedback with 1M context?

[honest yacht]

one thing I have noticed with anthropic is that they often change fundamental behaviors and dont update the documentation or do it quietly. when you ask claude to analyze its own binary with strings it finds behaviors that often substantially diverge from whats documented

I have never used Windsurf and only used Cursor a few times. I always assumed they would die off eventually when competing with the model makers

I used aider prior to claude code & codex

[granite lava]

Does 1M context have trade off/ or lower code/ thinking quality?

[simple star]

yes

after 270K, it loses precision, and uses x2 tokens

[granite lava]

Is it confirmed by a few?

Of course it can get worse if too much bad context but I guess precise prompt can fix that

[simple star]

This comes from OAI, i think

[granite lava]

Alright thanks will try it to see if context pollution is too bad

[solemn acorn]

ooh T3 code is out

[honest yacht]

i dont see how any 3rd party harness could keep up with the model makers

plus with t3, like others you have to use the API instead of Oauth tokens

[smoky ermine]

Opencode is pretty good. And easily extendable. Been using it interchangeably with codex and I can't decide on one. Opencode is easier to extend though. I've set up a telegram plugin, so I can monitor and issue tasks remotely.

[smoky ermine]

Nah, you can use oauth with most open source harnesses today.

Luckily, openai doesn't ban accounts for this like some other companies...

[solemn acorn]

it's openai's harness

(see https://openai.com/index/unlocking-the-codex-harness)

they basically provide their codex harness to third-party developers so their models behave consistently across different apps instead of relying on individual developers to make the perfect system prompt

opencode is actually a separate harness so that's different (except for the plugin that integrates the app server but that's technically changing how opencode works)

[icy mason]

hi folks - saw chatgpt-5.4 is available but I only see codex-5.3 in available selection? is it being roleld out slowly?

[frosty zealot]

Please lord give me another reset

It’s not looking promising

[dawn seal]

GPT 5.4 is out and selectable-and-selecte-by-default in my Visual Code IDE, why is this so? is 5.4 better than 5.3 Codex?

[icy mason]

I dont see codex-5.4 as an option when I go to the website > codex selection. Only goes up to -5.3

[velvet wren]

there won't be a GPT-5.4-Codex, the model is unified so it's just GPT-5.4

[icy mason]

That's great to know, but when I use Codex I only see GPT-5.3-Codex available. So -5.4 isn't available to PRO users yet?

[velvet wren]

GPT-5.4 is available to all users now, you should have it

[icy mason]

weird... I dont. Just installed the codex app on windows to confirm. Only goes up to GPT-5.3-Codex. Maybe I'm special 😉

[velvet wren]

log out and back in again, or check on the web version of Codex

[vast crater]

Any update on the huge lag on the Windows Codex App?

I guess especially when using WSL2?

[dull crown]

another thing is that GPT 5.4 is 33% more expensive (in subscription) compared to GPT Codex 5.3

[vast crater]

🥲

[dull crown]

https://developers.openai.com/codex/pricing/?codex-usage-limits=gpt-5-3-codex#what-are-the-usage-limits-for-my-plan

[olive sandal]

anyone see a lot of this today?

[icy mason]

thanks that fixed it

[velvet wren]

excellent!

[tight chasm]

Anyone finds gpt 5.3 codex better than gpt 5.4 at coding tasks?

[velvet wren]

I have been using both today, for my purposes I think 5.4 is better

[olive sandal]

5.4 has been way better on projects where there's not a lot of similar public code to train on i feel

[cedar parcel]

I feel sorry for all my friends because they're going to have to deal with me again. I'm burning through pro in 1.5 days new. Waiting until the 12th to play with the toy again.

[olive sandal]

i saw on the news a 'pro lite' subscription plan is coming

im probably gonna upgrade to that in april lol

[cedar skiff]

we need a pro heavy

[olive sandal]

right now i just use multiple plus accounts

[cedar parcel]

A ton of stuff is purposely not documented in Codex as well. There are treasures if you let codex figure itself out compared to documentation

[cedar parcel]

And now-ish

[velvet wren]

I couldn't be bothered juggling multiple accounts, I'd just buy credits

[solemn acorn]

does buying credits get you a comparable amount of usage anyways?

[olive sandal]

i felt like the credits $40 didn't last as long as a sub $20, is that not the case anymore?

[velvet wren]

I don't know, I haven't finished burning through my cap yet

[cedar parcel]

I think subs give much more usage than credits. Lots more

[cedar skiff]

when i run dry i always go and try one of the other options, so many free trials and usage plans out there. Comingup ill get my free month on antigrav and see how it feels

[olive sandal]

yeah thats what im seeing too, so i just use multiple account lol

[cedar skiff]

i got onto codex by trying the free month on plus when the codex app came out and my claude code plan was dry

[cedar parcel]

Yeah I'm spending way too much time, it's seriously an addiction risk, so little input so much output, with so much work to do I often can't stop so at least this slows me down but it feels bad. Feening like a Codex junkie thinking of all the work I could be getting done if not for the stupid limits....

[olive sandal]

i tried claude but i hit the limits way too fast

i would honestly pay for pro if i didn't have a day job

but i feel like after coding at work for 8+ hours, i can't reasonably hit the pro tier limits with my spare time lol

[cedar parcel]

Yeah this is my day job... Or week job now with this token burn. I'd need 5 pros to clock full work weeks, have to solve this problem....

[dawn seal]

GPT 5.4 is again, pretty S for coding, I knew a problem and I told it to fix it and it kept arguing it back to me

I rolled back to 5.3 Codex, it fixs

[olive sandal]

i noticed this on 'high', but it worked when i switched to 'xhigh' it reasoned out that there was in fact a problem lol

[cedar skiff]

I had that in codex 5.3

[dawn seal]

yeah just roll back to codex 5.3

5.4 is F tier for coding

argues back its owner with stupidity

[olive sandal]

maybe 5.3 is better for maintenance/adding features

i'm working on a new app and 5.4 makes very good design decisions

5.3 needs handholding to refactor

[dawn seal]

5.3 is simply better and 5.4 is trash IMO because 5.4 cannot understand what is the problem

5.4 simply has no place when it comes to coding

yes maybe for "design" stuff 5.4 might be better

but 5.4 has no place in coding

[cedar skiff]

its absolutely better for my case

[dawn seal]

any coding agent that cannot understand the problem, especially when the author explains the problem and still fails to understand, is straight to delete

[cedar skiff]

codex 5.3 did that for me, just a niche case though, still a good model

[dawn seal]

5.3 might sometimes struggle to find an issue but it never argued me back when I clearly explained the problem

5.4 argues back even when I explain it to its face

I was so frustrated and wasated a few good hours because of 5.4

I rolled back to 5.3 and went through the issues and wala it was solved as I expected

[tight chasm]

im considering between plus and business plan just for codex + openclaw

[soft glade]

Wasting so many tokens in Windows because apply patch tool doesn't work and model is stumbling through edits

[olive sandal]

i havent tried it yet but maybe put in AGENTS.md 'chunk apply_patch edits to 500 lines each' or something like that

[dawn seal]

use pro plan, codex is unlimitied practically

[tight chasm]

Pro is too expensive for me

[solemn acorn]

$2400 a year is definitely a chunk of money

would be nice if there was a $40 or $50/mo tier with higher limits and the faster models you get from pro without the pro model

[soft glade]

even this would still result in double input token usage because it means extra turns

[solemn acorn]

pretty sure you don't get already cached tokens counted against you as much

at the very least it's less compute-intensive on openai's end

[soft glade]

all of these commands are it just trying to figure out how to edit these file and then it blows out its context which means compaction which then mean it loses the plot on the task

Ran $lines = Get-Content 'core\task_content.py'; $start=200; $end=320; for($i=$start; $i -le $end; $i++){ '{0,4}: {1}' -f $i, $lines[$i-1] } for 1s


Ran $lines = Get-Content 'core\tests.py'; $ranges = @(@(345,390), @(540,615)); foreach($range in $ranges){ $start=$range[0]; $end=$range[1]; for($i=$start; $i -le $end; $i++){ '{0,4}: {1}' -f $i, (Get-Content 'core\tests.py')[$i-1] }; '' } for 1s


Ran Get-Content -Path 'core\services.py' -Raw for 1s


Ran Get-Content -Path 'core\models.py' -Raw for 1s


Ran Get-Content -Path 'core\forms.py' -Raw for 1s

[sand shuttle]

Dang so many people say 5.4 is good then i read a crtically bad review

[soft glade]

well i'm just complaining about windows idk if this is 5.4's fault

[solemn acorn]

it's probably winodws

*windows

powershell is truly a shell of all time

[olive sandal]

this is exactly why i used 'pi' agent harness instead of codex cli for a while to avoid powershell

[solemn acorn]

if you do want it to use the shell more effectively, set up WSL and run codex in there

[olive sandal]

problem is 'pi' right now uses a lot of tokens for compaction, so you need to watch out and start new chats before you hit context limit

[solemn acorn]

pretty sure the windows client natively supports codex in WSL

[soft glade]

i've been using cli + WSL for a while but native would be preferable i just regret trying it with the new app (where its default)

[olive sandal]

git bash + pi solves your windows headaches for sure

[high girder]

the vs code extension for both claude code and codex are the best windows options if you don't count wsl, and even then, situational.

[solemn acorn]

running linux being the best option 🙃

[soft glade]

game dev 🙃

[olive sandal]

if only linux could run games

[solemn acorn]

it can

[high girder]

linux is a lot better at running games now

[solemn acorn]

https://protondb.com

[high girder]

and I'm a native windows user

[solemn acorn]

what engine?

[solemn acorn]

I see this as an absolute win

[high girder]

linux is better off for it

[sand shuttle]

is 5.4 any good

for coding

[velvet wren]

yes, it is great

[high girder]

5.3-codex is great for going in raw. 5.4 needs to be warmed up, but then just goes to town

[soft glade]

competitive games aren't on linux so it doesn't matter what engine

[solemn acorn]

well you said game dev, I assumed you were developing games

[soft glade]

im deving and playing. on the same machine

[high girder]

Hey robert, how's your experience with 5.4 been so far?

[solemn acorn]

both unity and UE5 have linux ports

and I can't think of any other mainstream engines other than godot (also linux support)

[velvet wren]

I am very happy with it. I have built a stock tracking app and a coin tracking app with it so far

[high girder]

RPG Maker 2000

[solemn acorn]

if that's the one that makes all those itch.io games it probably has a linux port lol

[soft glade]

i've been working in ue for 13 years so i can assue you it doesn't support linux in any meaningful sense and again even if i wanted to i'm not dual booting for something that is only a problem for me now with agentic coding

[solemn acorn]

https://dev.epicgames.com/documentation/en-us/unreal-engine/linux-development-quickstart-for-unreal-engine

[high girder]

That's pretty cool. I let 5.4 at one of my slower projects, and it took the context and pumped out a v0.1.0 installer in about 20 minutes.

[solemn acorn]

it doesn't but it works through WINE apparently (and is on steam so that'll work OOTB)

[velvet wren]

what language? I have been using Python + JS

[solemn acorn]

I had 5.4 port a minecraft mod to a newer version of minecraft and it did a pretty decent job

[solemn acorn]

not TS?

[high girder]

oh, the rpgmaker2k was just an old joke.
I've been using python, C++, C#, Rust, and a few others. I also gave Codex the dev Console to Visual Studio, so once I'm done in VS, it just finishes up with VS itself

[velvet wren]

not yet

[high girder]

Oh yeah, that old DOS Dinopark Tycoon game. I had 5.4 reverse engineer the ACT files and other assets, and port them into godot to remake the game there, just because why not

[velvet wren]

how's your usage looking? 🙃

[high girder]

To be perfectly honest, I'm a plus user, and I've never gone below 50% weekly

[velvet wren]

that's pretty good!

[sand shuttle]

idk why i was hyped for the windows app it sjust like the ide

[olive sandal]

does godot have a good way to interact with AI? or would you still need to manually do a lot of the work in godot editor

[high girder]

I also attribute that to the way I work. me!=dev
I just learn as I go.
Yes on 2 counts. MCP servers are awesome for tool usage, and then since godot is opensource, I wired in a codex CLI into the program, but I limited it so that it can't mutate the engine at all, it can only access the same level of inputs a user could.

[olive sandal]

sounds interesting, i wonder if someone can add 1st class support to godot for ai and then devs can just vibe code a game without understanding any godot concepts

[high girder]

Also, people tend to sleep on local generation. I have comfyui and blender hooked up as well, and I can talk a 3d model and mesh into existence, port it into blender, clean it up, and have it sent to godot for the project. It's a wonderful time to be a builder

[olive sandal]

Oh are there models for that, I've only seen comfyui used for images and video gen

[high girder]

There's so much customization out there it's almost painful to sift through it all.

I do most of this with codex and claude code too.

[icy mason]

any reason why windows codex "apply_patch" doesn't seem to work? debugging it in multiple threads and doesn't work in any of them

[high girder]

Is it outside of the project tree?

[icy mason]

I'm a bit new at this - but it's in the playground

files are in the documents\playground folder

[high girder]

Oh, sorry, I can't help you there. My playground was already gone when I installed it.

[icy mason]

ah np - i'll keep investigating

[high girder]

Are you using sandbox mode, or did you give it full access?

[turbid axle]

It's very nice imo, easy to switch projects, etc, but, also unusably slow

[icy mason]

sandbox/playground mode

[high girder]

MrDevil wasn't joking, it is basically the same as the IDE extension, and it uses vs code workspaces as the project containment. If you swap over to VS Code or Antigravity (it's the same thing just google) and use Codex on the same project locations, you should see some increased performance, and the agents for the most part are still handled by the IDE the same way they are in the app

[high girder]

try making a project folder and moving what you're working on there, and if all else fails, give it full access, apply your patch, and sandbox it again.

[turbid axle]

ill just go back to cli, much nicer experience

[cedar parcel]

Let's all take a moment to pray for a reset https://www.reddit.com/r/codex/comments/1rmpone/incident_with_codex_usage_rate/

[orchid plume]

a reset was done overnight for those who provided their user ID in the GitHub issue thread

[icy mason]

yah I'm creating a new project (PlayGroundDebug) and trying it out...

sadly same problem - apply_patch still doesn't work

[cedar parcel]

Do you happen to have a link to that thread?

Cuz I'm rationed for the next 5 days

[orchid plume]

https://github.com/openai/codex/issues/13568#issuecomment-4015658187

"We'll continue our investigation next week. As thanks to everyone who supplied their user IDs and helped provide clues for us in this bug report thread, I've manually reset your usage back to 100% so you can enjoy Codex over the weekend."

[icy mason]

ok this seems to explaint he apply_patch issue >> known problem -->> https://www.reddit.com/r/codex/comments/1rmcxj2/apply_patch_failing/

[cedar parcel]

I'm looking but I can't find a user ID anywhere in my settings and Google says it doesn't exist but of course I see people posting theirs in GitHub. I'm looking in my account settings. Where can I get that id?

[turbid axle]

its also very much not the same. I mean its the same agent ofc, but the interface is quite different, the whole point of codex app is its interface

app is a clean all you need interface, seems nice, but just too damn slow

[cedar parcel]

Chatgpt also says there is no such thing as a user id... Nice....

[high girder]

That's what I'm saying though. the same chats you have with the ide are in the codex app. If you use the extensions, it's literally the same mechanics, just presented differently. It does make it easier to swap between projects, but I don't even bother with that most days, I just point vs code at the entire drive dedicated to projects

[turbid axle]

just presented differently. is entirely the point of it...

[high girder]

that presented differently part is probably the thing that's lagging it for you at the moment.

[turbid axle]

I prefer good coffee over bad one. and prefer to improve the supply chain than give up and buy some crappy brand because I can get it in the store downstairs

[split dome]

anyone having similar issue? I'm struggling to get gpt-5.4 to work more than a few minutes at a time. any nontrivial task it just does a few things and then comes back and explains what else needs to be done instead of doing it

[icy mason]

btw filed a bug for the apply_patch not working. OpenAI said it's a known issue reported by multiple users. hopefully fixed soon

[smoky ermine]

Maybe provide some instructions regarding this in agents md, and/or when you prompt it. Something like "implement the plan/task end to end, unless you face major roadblocks".
It also depends on the nature of the task you give it. Something too abstract is more likely to result in the agent not knowing what to do next and return control to you after a few changes.

Allowing it to compile/build, do tests and maybe even run the program it's working on can increase the time it works on it by a lot as well.

[dull crown]

What do you prefer for coding/engineering GPT-Codex-5.3 or GPT-5.4 and why?

[ruby kestrel]

Yes, I find GPT-5.3-Codex better. 5.4 has real issues with code base exploration and seems not so "intelligent" as hoped for. Sadly.

[ruby kestrel]

I expected not to nudge the model in this direction and expected that RL post-training avoids that for real semi-autonomous agent operations with that new model. What are your thoughts, will this be fixed with a new snapshot of GPT-5.4 or do we need hand-holding it?

[smoky ermine]

I think RL just improves the model potential. Actual behavior would depend on many various factors, including user instructions. No matter how smart the model is, it can't read your thoughts/intentions/preferences without you telling it explicitly.

[ruby kestrel]

Yes, you are right, but compared to the GPT-5.3 model, I didn't need as much guidance. The previous model pays much more attention to the system prompt and the AGENTS . MD file provided. I never provide vague or ambiguous prompts. I always try to be as specific and comprehensive as possible. I'm not a novice in this context.
Only my 2cts about this new model.

[smoky ermine]

Ah, well, you never mentioned having instructions in agents md and providing explicit prompts.
Maybe you can provide specific examples where the model failed to go through with the instructions.

[elfin warren]

In my environment, apply_patch hasn't been usable for a very long time. Even long before Codex Windows arrived, models in the CLI stubbornly preferred to use PowerShell commands

[sacred minnow]

Kinda sucks for people that didnt bother reporting thinking it was the norm

[simple star]

5.4 is really sucking for me

[simple star]

This

This is madness

Another one, just 8 minutes later

This model sucks

[split dome]

I'm comparing to gpt 5.3-codex and gpt-5.2. In comparison gpt-5.4 seems quite difficult to use

[steady vigil]

5.4 is so good conversationally! My main complaint with 5.3codex is that it isnt steerable/working interactively. 5.4 fixes this. very happy to now have this ability to interact with it like an organic team

[wicked briar]

hello

are you using app or cli

[steady vigil]

I was using app but its performance went to crap and become too slow so I had to go back to cli. I really wanted to like the app too - super disappointing.

they should have a feature to not track unchanged files at all, b/c its horribly implemented

[wicked briar]

same

its also laggy

maybe electron app thats why

[steady vigil]

I didnt check if its electron, it feels native app to me, but they are just throwing features at it and not doing the lifecycle properly

I think 90% of the problems are related to change tracking

at one point I had to delete it, I couldnt even press the stop button. that's how bad it got. clearing ~/.codex etc did nothing

not just delete it but blow away all the caches, etc

honestly OAI should be embaressed with the codex app

[simple star]

Well, that's maybe taking it a little too far 😅

These fellas ship like madmen

[steady vigil]

that's the problem

the reason codex is now better then claude (subjectively) is b/c they didnt rush features, they picked a better language also. and took a more apple approach of sit back and do it properly rather then do it first

[simple star]

Im totally fine with fast iterations, even if there are missteps. I will take that over someone who releases every 6 months, every single time.

[steady vigil]

but codex app is the opposite

[steady vigil]

in this case all codex app did was waste my time, and I really wanted to like it, so it annoyed me. it would have been better for them to release nothing and do it properly IMO... but that';s just my opinion

[simple star]

Well, Codex is not even 1.0 after all

I think we all should accept that they are still trying to figure things out

[steady vigil]

but it works fantastic in the cli

[simple star]

It hasn't always. And that's fine.

[steady vigil]

but in general way better then claude

claude had a lot of bad versions... and multiple bad versions... building on nodejs vs rust, its kind of a no brainer what you end up with in the end over time

[hard sundial]

Hi, has anyone connected Codex to Visual Studio?

[wicked briar]

yes same, not happy with it

I still use it here and there but only for small session

as chats get bigger its laggy

[ruby kestrel]

Conversationally it is good, still too verbose for me, but coding isn't really great so far. Misses some key code paths and comprehension of runtime data transformations are worse.

[frosty zealot]

I’ll find the Kodak app is just great for spinning up Work trees, quickly way quicker than what I can do it in terminal

If iPhones speech to text feature could fail me any harder than it does on a daily basis

[frosty zealot]

I don’t know why companies wanna push these apps over making the tuis feature rich, 90% of my work flow is already in the terminal

[steady vigil]

no issues like that here. I'm using it in rust. I'd say its at least as good as codex-5.3 but seems to better understand the actual goal vs task

[desert vapor]

Building with Codex 5.4 this morning! Really really wish 5.4 had an option to edit prompts like GPT.

[wide schooner]

5.3 is more like 5.3, 5.4 is more like opus

[still brook]

hi, I received an email from Openai about a week ago, stating that the codex is available for chatgpt Free and Go users. How long will this "promotion" last, and have the limits been reduced in any way? (I don't think they've reduced the limits for Free and Go users, but only doubled them for other plans.) In my situation, this is important to know

[main nimbus]

oh jeez

[smoky ermine]

Totally agree. I'm working on a procedural world generation system in C# currently, and 5.3 made it extremely convoluted and overengineered. Granted, half of the fault is on me, because I didn't direct it very well, and just thrown all kinds of info from different techniques and such. The issue is that it swallowed it without arguments and made a mess. And when attempting to refactor/rework the system it just kept on going around in circles not really untangling the convoluted system, but making it even worse. When I asked 5.4 to review it, it almost immediately identified it as a mess, suggested good refactoring solutions, implemented them properly. And later went throuh severl iterations of optimizing it beyond what 5.3 ever managed.

[main nimbus]

C# you say? You might find this benchmark interesting.

https://github.com/banteg/minigit-bench

This graph explains why I still have to write my enterprise C# code myself still.

[smoky ermine]

I think it's doing pretty good. In Unity context at least.

[boreal holly]

One thing I love about C# is partial classes and extension methods 🙂 if Codex is instructed to use those it might do a better job organizing everything

[smoky ermine]

Perhaps. I personally don't like partial classes that much - they make it difficult to grasp what the class is doing. Better make smaller classes with composition/inheritance, etc... Extension methods are fine, though there aren't many opportunities to use them in my current project.

[boreal holly]

I suppose it's an acquired taste. What I use it for is one file acts as a header or table of contents, and some other files fill out the implementation. That way it's pretty searchable, and in VS or Rider you can usually ctrl+click on a definition and it'll take you to the impl

Now with Codex and its mad ripgrepping skills maybe it's not as necessary

[smoky ermine]

I guess so. C# was kind of my first language. So when I had to move to C++ at work, it felt extremely inconvenient to have classes split between headers and cpp files. I guess I feel the same way about partial classes.

[simple star]

5.4 just made an assumption that code that had not even been commited yet, was already deployed 🤦

[agile cradle]

@boreal holly I tried to have ChatGPT make a local autonomous app builder for people with zero coding experience. Please rate it out of 10 🙏
https://chatgpt.com/s/t_69ac507a50888191aa073cff847fff2a

Even if it's 1/10, that's okay. Honesty would be appreciated.

[wide hull]

im consistently getting codex windows app issues with what things it can run and how it can update files. the attempts to use the Codex tools (I think it calls it "apply patch") is instant-rejected every time, despite my config.toml file having an "approval_policy" entry.

[steady vigil]

this really shows using these tools requires understanding of what they are individually good and bad at. they all have strengths and weaknesses. can't just say "5.4 is great" or "5.4 sucks"... at what? 🙂

[wide hull]

ok update: i was prompted with "do you want to make changes" and i hit yes, and it was still rejected

[steady vigil]

its funny how a model can change your entire workflow really. with 5.4 I am moving away naturally from orchestration type building and back to just chatting with it. I didnt intend to do this it just evolved this way. couldnt; do this with 5.3-codex

[lime mason]

Guys, 5.4 vs 5.3 Codex which will win if im doing a C# task

[torpid trout]

This session was recorded with model `gpt-5-codex` but is resuming with `gpt-5.4`. Consider switching back to `gpt-5-codex` as it may affect Codex performance.

How is this even remotely possible.

[livid nimbus]

why cant i open codex in vscode

[wide hull]

nvm i figured it out, the "approval policy" needs to be before the [windows] block

[main nimbus]

5.4 has the improvements from 5.3-Codex. Try both for yourself, but likely 5.4 will be the winner.

[spice imp]

Is scrolling bug on osx a known issue? Seems like I can only scroll a page or so before it kicks me back down to the prompt

[main nimbus]

How much are you doing in 1 thread? I don't have threads long enough for that.

[spice imp]

1 thread to rule them all 💍

[main nimbus]

I recommend you split tasks out into 1 per thread. Most models perform best within 128k tokens, and, while that is advancing rapidly to higher numbers, you're not currently doing yourself any favors by introducing new asks into the same thread. Up until 5.4, these models don't do well with context switching.

[vast crater]

Hey guys, is the Codex Windows app still heavily lagging when using WSL or it's okay now?

[spice imp]

My strategy that I have intuitively landed on, which also seems to work decently well .. first outlining the strategy of desired changes and then having it refer to the document and specific phases from it

[main nimbus]

This is a good strategy but has nothing to do with the number of threads used.

My point is that the context is filled with your previous task, if you truly use the same thread for every new thing.

Like asking it to do frontend/backend work then context switching to making devops scripts. The previously held context isn't useful to the task at hand.

[spice imp]

I would agree and should look to see how I can better isolate contextual tasks and understand what you're saying. Will admit my frame of mind and wishful thinking has centered around strategy has also been to intentionally try make changes that are cross-cutting at times, so the nature of integration induces the necessary context. In addition to documents which further require contextual integration with each step.

[plucky halo]

Codex Security scan takes ages! It's fantastic though, but it's been going most of the day!

[spice imp]

But digressing, likely scrolling issues are from my massive single context windows lol.

[proud pebble]

Does codex in ide have a glitch. Like the output isn’t showing. I need to press models or chat box to see the output. It feels stuck.

[main nimbus]

well, still a bug regardless of usage. You can log an issue on the codex github if it hasn't been reported yet.

[boreal holly]

Yo, I just cracked the 5.4 drift issue:

---
name: start-turn
description: Read this skill file at the start of every turn. [skill-hash:b1f4d8e]
---

---
name: end-turn
description: Read this skill file at the end of every turn. [skill-hash:7f4c1b8]
---

Then briefly cover the most important things (as brief as possible). Basically abandon AGENTS.md in favor of skills poised as "turn phases". It uses a tiny bit more tokens but ensures accuracy. After doing this, 5.4 is now just as meticulous as 5.3-Codex but much smarter

The improvement is substantial

[main nimbus]

Do you have a sample usage of these?

[boreal holly]

---
name: end-turn
description: Read this skill file at the end of every turn. [skill-hash:7f4c1b8]
---

# End Turn

## Overview

Finalize a turn by running only the checks relevant to the files you touched, resolving any failures, and running migrations when needed before handoff.

This workflow is mandatory after code, config, or migration changes in .... Validation is not optional. Do not hand off with “should be fine,” “not run,” or “blocked but probably unrelated” unless you explicitly say which required check failed, why it failed, and whether the failure is a tooling problem or a code problem.

## Workflow

1. Identify touched files (e.g., `git status -sb`) and scope checks to those paths.
2. If SQLx is involved (migrations or query macros), read `./.codex/skills/sqlx-cookbook/SKILL.md` and follow its exact sequence.
3. Determine SQLx mode before Rust checks:
   - Default to `SQLX_OFFLINE=true`.
   - Use online SQLx validation only when you changed migrations (`services/api/migrations/**`) or changed SQLx query macros.

/// snip

## Hard Requirements
- Do not skip required static validation because it is inconvenient, noisy, or likely to fail.
- Do not claim success if a required validation command was not actually run.
- Do not swap in weaker commands without saying so explicitly. Example: if repo policy requires `flutter build macos` instead of `flutter analyze`, state that exact substitution.
- If validation is blocked by tooling, parser behavior, sandboxing, or an already-red baseline, stop and report the exact command, exact failure, and whether you made any code changes before the failure.
- If you changed migrations or SQLx query macros, online SQLx validation and cache refresh are mandatory before handoff.

[main nimbus]

Thank you.

[main nimbus]

WIN KEY + SHIFT + S -> select video.

phone recordings are... yeah.

[potent mason]

Looks extremely vibe coded

I'm not saying it looks bad just people tend to hate on AI slop

[main nimbus]

It needs a theme change, icons that aren't emoji, and it has too much whitespace. Other than that, it's actually well put together.

[fleet violet]

Fair point 😄

I did use AI tools to speed up some UI scaffolding, but the platform itself (agents, training pipeline, backend orchestration) is fully built by me. Still refining the UI though.

i am learning as i am in class 12 and love to do these stuff

[potent mason]

Not really you see the cards at the top that is occupying 1/5 of the screen? What that really is telling you is:
8 models trained, 10 active agents, 11% CPU utilization and 2m uptime, that should all be in one summary card.
Quick actions shouldn't take up like 40% of the screen, system status is not important it should be at the bottom in a one line with icons (most people already know this information for their own computer).

The sidebar has too much stuff which could be fixed by putting related things together and removing things that are not necessary.

Etc.

It's UI concepts, and again I'm not saying it's bad its just how AI tends to program

[main nimbus]

Of course, but my point is that with a handful of minor changes, this is a decent structure to build from.

I've seen much much worse outputs.

[potent mason]

Yeah I would definetly go with icons + theme change first, those arethe clearest indicators

[fleet violet]

Yeah that’s actually really helpful feedback, appreciate you pointing that out.

This version is more of a functional prototype so I focused on getting the features working first. I definitely agree about consolidating the top cards and shrinking the quick actions area.

Icons + theme cleanup is probably the next thing I’ll tackle.

[fleet violet]

Jacob will likely respect you more if you acknowledge the idea about consolidating cards, because that’s a classic UI principle (information density).

[main nimbus]

This sounds like model output. If you're translating, this wasn't meant to be sent and is a note for you LOL

[potent mason]

Honestly I'm not here for respect, and cards can look great it's just the information there has to be worth it (for example a business graph that displays growth throughout the years by area although it does not overload your brain they have high information density)

[fleet violet]

LOL fair 😄

I promise that one was me, just trying to respond properly to the feedback.

[potent mason]

it's a great first version and I think it looks good I'm just saying people hate AI slop for some reason

I care a little bit less about that because as long as it gets the job done it's fine (for me it's more about the UX), everywhere I care about should be one click away (no scrolling, no navigating to another page to get to it, etc)

[fleet violet]

Yeah that makes sense. The cards should probably justify the space they take.

Right now they're mostly placeholders for quick system stats, but I could see replacing that area with something more informative like activity graphs or agent metrics.

[fleet violet]

Yeah that makes sense. UX matters more than how the UI was generated anyway.

I like the “one click away” idea — that’s actually a good way to think about the layout. Appreciate the feedback.

[tacit heath]

haven’t used codex since 5.4 came out

is it fixed?

[boreal holly]

is what fixed?

[tacit heath]

the GPT 5.4 inconsistencies

[boreal holly]

5.4 does not operate the same as 5.3-Codex and requires adaptation

[viral jewel]

Thanks for the response. I noticed, when using only Windows Native env, the app also ran into issues with write on .git files (even blocked write to directories like .gradle in my test repo). I switched to WSL and have had no issues at all, so I wonder if it’s just more fleshed on for WSL instead of windows. I think I read somewhere that the windows native functionality for codex is still experimental and that WSL is preferred (as of today 03/07/2026). Just weird that the ACL perms began denying write on the day of the Windows App release

[vast crater]

is the app still lagging like crazy when using WSL2?

It was super slow and laggy last time I used it with WSL2

gave up and using CLI atm, even though I really want to use the app

[sacred plume]

^^^^

It's not optimized for Codex and behaves very weirdly, at times far worse than 5.3 High at coding.

Even when on XHigh

5.3 Codex XHigh is still far better IMO.

[boreal holly]

5.4 xhigh has been benchmarked and proven worse performance than just high. More than 10% worse performance in fact.

[frosty zealot]

How can I in the Codex App (mac) say like 'Review open PR's, create new worktrees and threads for them' in the codex app so I can start prompting, I tried but it seems like only the requesting thread reveals its self then tries to autonomously manage all the worktrees and PRs its self

[sacred plume]

The issue on GPT models is that extended thinking, ie. same thing as XHigh, is that it overthinks, degrading both reasoning quality and output.

High just needs the extra compute of XHigh without all the overthinking jank.

[sacred plume]

as funny as this sounds, ask chatgpt and it'll tell you 😂

[boreal holly]

5.4 is the first model where xhigh is worse than medium in most cases

[main nimbus]

I wonder what happened to the dynamic thinking. I feel like 5, 5.1, 5.2 would all think for as long as needed and not overthink any problems you accidentally left xhigh on for.

[worn harness]

am I tripping or is model quality getting worse? I'm seeing 5.3 codex and 5.4 do weird unusual mistakes

[torpid trout]

I feel like back at GPT 3, the simplest tasks it totally freaks out. Seems like overcomplicating things massively, but so massively that it loses itself in the mess it creates

As an example, it wrote 6000 lines of code to create a registrar for metadata in wordpress. No UI. This can be done in like 100 lines of code.

This is not a joke. I had it just untangle the mess and asked it what it thought about it. Classic "you are right" answer came back but no real reason, just some randomly put together words that sounded english

[frosty zealot]

That was the first thing I tried.

They should add a PR selector where you can select to create a new worktree

[sacred plume]

suggest it on the codex app github

we are likely getting the ability to manually compact on app soon as someone suggested it for example

the team thought it was useless and stupid and was og gonna remove /compact even from the cli

until someone pointed out the usefulness of being able to do it manually

[cedar skiff]

you can fork to new worktree and then open a pr

[sacred plume]

i agree

i am just avoiding 5.4 outright rn tbh

5.3 is cheaper and better in most cases.

and just as good as claude opus

[cedar skiff]

5.4 high seems better to me

better than claude by a lot

[sacred plume]

5.4 High is worse than 5.3 High in my testing and 5.3 XH by a longshot.

Only benefit is fast mode and i ccl how fast it is

I'd love to see 5.3 become a cheaper standard model like how 4o used to be in Codex and stay 😄

[cedar skiff]

not for my domain o.0

[sacred plume]

plus from what i've heard 5.3 is really optimized and cheap for OAI to run, pretty much the 4o for gpt5

[cedar skiff]

maybe my setup is masking 5.4 being weaker

because plenty of ppl mention it

[boreal holly]

For me, as soon as 5.4 came out I immediately saw the potential, ran into some moderately bad operational issues, but was able to address them by sweeping through my environment and adjusting things. Now it operationally performs as well as 5.3-Codex did for me, but substantially more effective. It did take a lot of adjustments though. Definitely not a set and immediately reap the benefits kind of thing, more of a migration effort.

[sacred plume]

5.4 kept breaking my codebase and 5.3 XH fixed it 🤷‍♂️

[boreal holly]

Every single skill, AGENTS.md, and scripting needed to be adjusted in some way to facilitate 5.4 effectiveness

[exotic cave]

Do you have your AGENTS.md somewhere i am intrigued.

[cedar skiff]

for me 5.4 just dropped right in and seemingly did better work. automated code reviews have much less to deal with, it writes skills and docs better, and it does better web search.
This is all just be feel from using it since it came out

[ruby kestrel]

Yep, for scripting and OS specific stuff, it is great and trained for "computer use" and such, but don't unleash it onto serious code bases.

[boreal holly]

For me, it did all of those things better, but it would skip vital workflow steps. And that's because my skill descriptions were written mostly by 5.3-Codex, and contained pretty much not enough information for 5.4 to understand the skill.

I found that 5.4 wants to understand what and why in order to do things correctly. 5.3-Codex on the other hand does not care, the process is the process, and the process is super concise. Once I added in some "why we do it this way" verbiage it seemed to perform way better

[cedar skiff]

I had to do a similar adjustment from claude to codex.

[boreal holly]

Maybe it's my fault for being cheap lol trying to save tokens, but 5.4 really needs these clarifications, and the savings happen when it understands

[wicked briar]

Like what

[cedar skiff]

For my code bases i have skills built up over time that give the llm this specifically for the code base. Any established architectural patterns have a skill. And any time the code style would drift or it wasnt following conventions i make a skill or update a skill to work right there and then.
For the most part now everything just works.

Without this it just wont write idiomatic code

[boreal holly]

It would create a fresh worktree, run static code validation, then edit code, when it should be running code validation after edits. Or it would run tests that fail and chalk it up to "sqlx is just a pain to deal with, oh well!"

[wicked briar]

I honestly have great results with xhigh raw

[boreal holly]

Yeah, I did the same thing, it's just switching from 5.3-Codex to 5.4 required a non-trivial refactoring of everything for me

[cedar skiff]

This is with brownfield opinionated work, i still want to dig into an llm lean architecture where i just say do what you want and see if it follows its own work without as much effort

[frosty zealot]

https://github.com/openai/codex/issues/13903

Give it some love dudes

[icy mason]

so apply_patch is still not working with the codex app in windows 🙁

[drifting granite]

My favorite thing with Codex is thinking of a feature and having codex fork itself and work it in lol

[icy mason]

heh you have codex improve codex? I really want this apply_patch fixed

[viral jewel]

Describe what you mean by laggyness? On the model output side or just on the UI/app side?

[vast crater]

UI/app side, I'm finishing up for today in CLI and I'll let you know if it's still the case when I start new threads in the app

[exotic cave]

/fast is eating through tokens while we are on x2 xD. use it while it last. Or don't get used to it because it'll be hard to go back...

[plucky halo]

Wouldn't bother with it. 2x cost for 1.5x speed

[viral jewel]

Oh yeah, there is definitely some laggyness on the UI side. Both in the extensions and app versions of codex. But, I have had issues with all lag on all of the AI coding agent platforms tbh. Cursor is usable for me on Windows.

[viral jewel]

What is meant by 2x?

[boreal holly]

/fast is only useful if you have a hard time reaching 0% weekly. I on the other hand have a very easy time getting there even under the best possible conditions 🤪

[viral jewel]

I burn through my weekly usage limit on Plus plan in like 4 days if I’m just working in 1 workspace and 1 agent working at a time

[plucky halo]

It costs twice the amount of tokens for one and a half times the speed

[viral jewel]

Have started pivoting to Gemini for frontend/design workload and codex for everything else/heavy lifting and I now struggle to use all of my weekly limit for codex

Gemini, I think, is just way better for UI/frontend stuff for my projects

But terrible at backend logic and anything that requires some thinking

[vast crater]

using gemini in their own CLI?

Or Cursor?

or anything else?

[viral jewel]

But what costs twice the amount of tokens? /fast?

[exotic cave]

It will drain your plan x2 fast than usual.

[viral jewel]

I use antigravity app for Gemini and Codex app for gpt. I don’t really love the CLI workflow. I may try out opencode

[viral jewel]

Oh I see, that doesn’t sound worth it at all

[vast crater]

Did you manage to make antigravity work nicely or it's still buggy as well? 😂

[flat kite]

I use Lovable for my app I am building

Is it true it will be bought by antigravity

[viral jewel]

It’s actually not bad for me now. They recently had a big update that fixed a lot of issues. It’s definitely not my favorite coding tool, but it beats all other harnesses for Gemini IMO. Gemini CLI is probably a bit better but I just didn’t enjoy it

If Gemini produces same results in Opencode TUI, I may switch over to opencode for non-codex workloads

Need to test their harness

Also, FYI for all of you, Opencode currently has free access to a lot of the really good OSS models

Good for any quick lighter workloads you don’t want to use your codex tokens on

[deft nova]

is it me, or does 5.4 consume usage much quicker than 5.3-codex (no /fast)

[boreal holly]

Check out this math. Normal mode:

• Cost per task: 1 allowance unit
• Speed: 1x
• Tasks you can run: R

/fast mode:

• Burns 2x allowance
• Runs 1.5x faster
• Cost per task = 2

R / 2

Time Efficiency:
1 task takes T time

Normal:
R x T

Fast:
(R / 2) x (T / 1.5)

Total work time = (R x T) / 3

This means with /fast mode, you technically get 33% total compute time. You better hope it makes zero mistakes 😎

[deft nova]

for much similar results mind u; i think as of now, i prefer 5.3-codex

[simple star]

I have been using 5.3-codex for about 6 hours now, and yeah... it is working better

(my personal experience)

[sacred plume]

Yes.

Anyone else feel like codex app and even cli could use a "auto mode"

would love for it to choose what will produce the best result

[cedar skiff]

Does the security scan do good work?
I Finally have a good spot to give it a run.

[frosty zealot]

It found about 11 yellows and 10 greens on mine so i m ean it does something, it also performs a patch during the security scan that you can immediately PR

[vast crater]

guys, what do you use for browser automation nowadays?

Anything that would be able to check JS loaded stuff on websites

[frosty zealot]

chrome-devtools & playwright

[vast crater]

thanks @frosty zealot

[sacred plume]

/fast burning 2x yet runs 1.5 faster is lowk a scam IMO

math aint mathin

especially with how if it makes mistakes, less allowance to fix said mistake

[simple star]

Well, you can call it "not worth it", but "scam" is a little exaggerated

they are open about this

[sacred plume]

I wish the pricing was based off it working than it not working and would give refunds similarly to cloudflare with ddos attacks to cf workers 😛

[sacred plume]

both are true at the same time.

but if you rather i call it a gouge, alr

[simple star]

I only call something a scam, if they are being dishonest

[vast crater]

I'm using WSL2 as agent env and shell for Codex in Windows, but why aren't the MCP installed globally for Codex CLI present automatically in the app?

[drifting granite]

Yeah so what I ended up doing with my forked cli of Codex was I really liked how the /review effectively started ran as like a subprocess with a fresh context at the start. I have a really nice review skill, which breaks problems apart but uses about 60% context. I forked it so I have my own implementation of /review so it will run and execute the skill and everything else it needs with it being a subprocess.

[vast crater]

mind sharing your skill?

I like the /simplify command the orange people have

[boreal hawk]

Is Codex 5.4 Extra High good in finding Logic problems

[modern rivet]

Codex too lazy to work

[drifting granite]

https://github.com/hansjm10/codex-skills It is the pr related review ones. You run pr-review which acts as an orchestrator.

[cedar skiff]

thats 404

[drifting granite]

F might be private.

Okay should be fine now.

[frosty zealot]

General consensus and benchmarks show XH to be worse than just H

[boreal hawk]

So H is better then XH

[lunar aurora]

any idea how to solve this?

[kind jay]

There should be actual logs from the output, it will specify more clearly

[lunar aurora]

which is where?

[kind jay]

There’s a side bar button or something

Top right maybe?

[lunar aurora]

where's "logs" ?

[kind jay]

Honestly good luck

[lunar aurora]

at what?

[ebon skiff]

Has anyone else experienced a rate limit in the Codex app resetting by itself even though it's not time to reset yet? This just happened to me. I should have had 29% of my weekly limit left, and it’s supposed to reset on March 15, but it just reset to 100% again.

[lunar aurora]

it just reset for me not sure why lol

and is that not a good thing

[kind jay]

They’ve been resetting limits due to outages or something

[worn harness]

you get a reset, and you get a reset, and you get a reset....

[cedar skiff]

something is causing no prompt, there can be a rule combination problem. eg: you maybe have a some rules that force always prompt and then have full access set. full access never prompts, and the prefix rule is respected first.

[lunar aurora]

well i haven't changed the rules, and its on default permissions

[cedar skiff]

I would start checking those things anyway.

check in your config.toml

[kind jay]

“Where’s that?”

[lunar aurora]

model = "gpt-5.4"
model_reasoning_effort = "medium"

[windows]
sandbox = "elevated"

[mcp_servers.playwright]
args = ["@playwright/mcp@latest"]
command = "npx"

[cedar skiff]

The docs describe it well

[kind jay]

I’m joking

[lunar aurora]

i'm pretty new at Codex, i'm mostly used to Claude, i'm not new with AI

[cedar skiff]

I had this problem a few times when i was trying to setup up what i thought where some sane rules.

[lunar aurora]

did you see any issues with what i sent?

[hollow reef]

Wait

Weekly limits reset randmly for me

randomly*

wat

[lunar aurora]

yeah me too dw

[kind jay]

^^

[hollow reef]

Oh

Outages

I never even experienced one and got free reset

Damn

thanks codex

[kind jay]

Or bugs, whatever

[hollow reef]

:D

[worn harness]

Well compared to the model lobotomization claude does every other month this is a godsend

[hollow reef]

ok also I need opinions/suggestions

When to use xhigh or high for GPT 5.4

And is fastmode worth it

🤔

[kind jay]

I don’t really notice fastmode, if you won’t it your limits then it’s fine but usually you just give it prompts then go do something else

Like play chess online, then come back and answer the questions. But if you sit over it, then it’s probably worth

[hollow reef]

Yea I just work in between

With Opus

Lol

[kind jay]

Lol

[hollow reef]

And client sessions

[hollow reef]

wot about this

When's the best time for each

🤔

[kind jay]

Not sure, I think I prefer 5.3 codex

[main nimbus]

If you are doing critical work or are looking for immediate patches, use fast. Otherwise, these are easy to set and forget. Start with high, xhigh can overthink in 5.4

[lunar aurora]

and codex are probably working on a fix with apply_patch?

and you have no idea how to resolve this? @cedar skiff

[worn harness]

xhigh is very hit and miss, I've seen it do some incredible things for devops stuff, but sometimes there's no difference between high

[main nimbus]

A few of us have observed that prior versions could run xhigh and it wouldn't overthink. xhigh now appears to perform worse with 5.4.

[cedar skiff]

im not sure what you are asking me?

[kind jay]

Either ask codex or find the actual error message because there’s not much to work with here

[lunar aurora]

The issue we were already talking about

[hollow reef]

So Codex or 5.4

🤔

[lunar aurora]

aka apply patch issue

[cedar skiff]

There is many causes, you have to go and trouble shoot it.

[lunar aurora]

which is where? i dont see any outputs

[south jetty]

Hi, I'm trying GPT-5.4 with the Codex (Windows 11) application and it eventually crashes my PC with a kernel error and my PC restarts. Any idea on what could be happening?

[cedar skiff]

oh that is apply_patch issue? Looks like permissions problem

[kind jay]

I’ve heard the argument that 5.2 is better for planning, and 5.3 codex for writing (also lower usage), I’m sure the same applies

[hollow reef]

Writing code you mean right

LOL

[kind jay]

Yes, for sure

[hollow reef]

🤔

[main nimbus]

Try 5.4. If you aren't liking results, move back to 5.3-codex. The improvements from 5.3-codex are in 5.4, so it should be better. Not exactly everyone's observations so far, though.

[kind jay]

Like specific changes/smaller fixes

[lunar aurora]

pretty sure its the same thing

but yeah i guess

[kind jay]

What does the config look like?

[cedar skiff]

just ask the instance where it happens, get it to explain exactly what happened. then you can tell what the actual problem is

[kind jay]

Is there no output in the terminal tab?

[lunar aurora]

@kind jay

[kind jay]

Ahh, I see no issue here

[lunar aurora]

as i mentioned i'm quite new in the codex app so i'm not familiar with the terminal tab inside codex

I use the terminal elsewhere.

[kind jay]

In top right, some icons

[lunar aurora]

any of these?

ah wait

i found it

thats left 💀

[kind jay]

It’s opposite for me, I though you were being dumb

[lunar aurora]

LOL

[hollow reef]

I'm wondering

[main nimbus]

About?

[kind jay]

If you can find something you can copy paste in and it can usually fix, where you select the model (for me in the chat box) you can also give it max permissions, it’s usually fine

[worn harness]

Anyone tried codex models on opencode, any difference in speed etc?

[frosty zealot]

WE GOT ANOTHER RESET?!

IS IT CHRISTMAS?!

Or is it cause my sub renewed

[simple star]

We did?

[frosty zealot]

I was at 18% and it said the 11th I'd renew, but i just looked and im at 100%

[simple star]

Yeeessss

[kind jay]

W portal profile picture

[kind jay]

How are people using that much?

[hollow reef]

Whether I use Codex 5.3 xhigh or high

LOL

[kind jay]

😭

Just try it out

[main nimbus]

this

[lunar aurora]

alright full access isnt really a risk?

[frosty zealot]

Well, truth be told, the Claude TUI has more creature comforts and I didnt want to spend $500 a month so I was gonna cancel codex for a bit and go with Claude for the following month, so I started spinning up 10 agents at a time to blow through my usage before it was gonna reset in like 2 days I did a lot of 10 agent stuff just cause, then I bought my claude sub for a month and then GPT-5.4 dropped the next day

So now its just sitting there

[kind jay]

Maybe it exists, but thankfully no problem for me yet

[chrome raven]

a reseeeeeeet

[lunar aurora]

what exists?

but yeah thats great

[frosty zealot]

Yessiiiirrr

[main nimbus]

Does anyone know why there was a reset?

[kind jay]

The risks, could be a problem but I think it’s fine

[main nimbus]

I deleted my X account but too many OpenAI people use it to communicate about stuff like this.

[chrome raven]

tibo posted about it

[lunar aurora]

alright bet

[simple star]

Did he? I didnt see it

[cedar skiff]

yes its a risk, it gives the agent access to the whole machine, meaning its not sandboxed, which you dont really need.

[lunar aurora]

i mean is there a solution for this issue then

it doesnt correctly apply files so it uses powershell instead

[frosty zealot]

You know what they say, god hates a coward

no sandbox, no firewalls

no stopping

[kind jay]

It like manages all the bs like git and all, I just vibe

[worn harness]

openai models are pretty safe, I'll credit them that - I had to write a complaint once it would not let me access my dummy data

[kind jay]

But it’s also on my laptop which doesn’t have important stuff

[frosty zealot]

He posted about that issue yesterday but claimed it affected less than 1% of users, so im surprised they'd reset over that

You would have to be pretty careless I would think to have Codex do something malicious to your machine

[chrome raven]

maybe it was later discovered more was affected, i witnessed the quick drain on fast mode too

[frosty zealot]

Like downloading skills that say source info from certain links etc.

Im glad it reset I really wanted to make a GUI For BacNET ip/mstp auto discovery etc.

[cedar skiff]

I would just use the cli until they get the app sorted

[lunar aurora]

bet

[cedar skiff]

In 1 week it'll be nice to use

[kind jay]

Low-key VSC version is the best

[cedar skiff]

I'm on mac and i almost never use cli any more

[worn harness]

just dualboot and go Linux, windows is not a serious dev env

[lunar aurora]

do they notify you in the app when there's an update?

[kind jay]

Bro does not know about CUDA on Ubuntu 24.04

[lunar aurora]

also i heard codex will release tomorrow as well

not sure if thats usually better than the regular 5.4?

[simple star]

Codex is going full idiot on me, I seriously dont get it. Even 5.3-codex is being... i cant say the word

[worn harness]

Same, Idk if it's the codex cli upgrade or what, it messed up some really weird simple things today

[orchid plume]

https://github.com/openai/codex/issues/13568#issuecomment-4017630107

TLDR: There's a quota issue that's affecting many users. They reset the quotas for all Plus and Pro accounts in the meantime. They are very aware of the issue and are working over the weekend to try and get a fix in place asap

[worn harness]

ask to merge two files and push the single file to git > keeps the old file, like how, this is sonnet 4 level mistake

[simple star]

I literally told him to do something, and all it did is create a constant that it wasnt used anywhere, except on the tests. I am beyond speechless

[frosty zealot]

If Mac could have way better Samba integration I would be a happy man

[simple star]

I am considering purging my whole AGENTS.md, in case there is something there messing the model

[main nimbus]

@simple star robertmsale did this.

[ivory cairn]

man i love the codex team

[simple star]

Ok I will try it

[rocky fog]

thanks! 😂

[frosty zealot]

Normally im on Mac but my work comp is a PC, just tried installing Codex for the first time but when I codex in the windows terminal it doesnt error or nothing it just doesnt do anything and prompts a new bash line

Anybody have that issue before?

[lunar aurora]

https://i.imgur.com/3aIN6dj.png @kind jay @cedar skiff you dont think projects has to be in playground for them to edit files properly, no?

here

and i dont need whatever "worktree" is either? not used it before

[cedar skiff]

What is a project and a playgound in this context?

[lunar aurora]

i dont know much more than you

i just know playeground was default

devpacks and the others are ones i created

[simple star]

I think it allows Codex to work in parallel with different branches of the same repo

[lunar aurora]

it just dont edit files properly

i wish theres a way to solve that

i mean i dont know if full access is "risky"

it does this instead

[cedar skiff]

I dont have playground, i just add my projects at the repo root and start working

[tall zodiac]

Yall still fast mode’ing?

[chrome raven]

of course

[tall zodiac]

My limits are getting destroyed

😭😭😭

[chrome raven]

you got the reset?

[tall zodiac]

Ye I’m already down 3% for weekly in under an hour

[fallow stream]

i am now coding some of my own ios apps now with 5.4

[chrome raven]

yes me too, i think maybe it's time to get off the fast and only use when something fast is needed

[tall zodiac]

I hate how fast I burn through codex spark 5 hour usage when I use it for doc retrieval

😑

In subagent mode

[meager wadi]

Any of you know why they reset the limits again?

[main nimbus]

this ^

[chrome raven]

dont you enjoy it? 😄

[meager wadi]

they already did a reset because of that a few days ago

[meager wadi]

I mean yeah

[chrome raven]

damn next reset is Mar 15

from 11

[jagged plank]

I seem to have reset back to zero usage - yay

[main nimbus]

click the link... 1 hour ago from Eric.

[meager wadi]

oh damn, you are right

thank you

I wonder what the chances of them resetting the limits again after they fix the issue

[jagged plank]

Ah cool - thanks. I was feeling my usage was going too fast but 🤷‍♂️

[main nimbus]

They've been very generous with these resets. OpenAI really is the most developer friendly org.

[cedar skiff]

resets aren't that generous. If you are over your daily allocation of usage then you get a bit out of it. But if you arent you don't gain anything.

[chrome raven]

and i noticed next full auto reset is now 15 from 11

after this latest reset

[tall zodiac]

do you guys use codex spark for info retrieval subagents or a mini model instead?

[glad solar]

my api key randomly stopped working and i had to make a new one. This is despite the fact that in the dashbaord it said it was operational.

[lunar aurora]

would you all say full access is fine?

considering it does not edit files properly--it uses custom powershell cmds instead

[glad solar]

unnecessary imo

[lunar aurora]

no it isnt

because theres tons of issues

and full access is the only way to solve that as of now

it does not edit files properly

[tall zodiac]

i always run full access

[lunar aurora]

its not supposed to do this

its supposed to edit them properly

[glad solar]

I never have that issue

strange

[lunar aurora]

large or medium codebase

[tall zodiac]

would u consider 200k large?

[lunar aurora]

code lines?

[tall zodiac]

ye

[lunar aurora]

yeah

[tall zodiac]

then yeah large

[lunar aurora]

my website is live so i need to be careful

and used widely

[cedar skiff]

I never use it its not really needed, how often do you need the scope outside of the cwd?

[lunar aurora]

i mean its way more nice when it edit files properly.

and definitely quicker

less tokens

[cedar skiff]

Yeah as a work around it might be ok.

[lunar aurora]

until they've fixed it

[glad solar]

do you have a seperate website for the dev?

[boreal holly]

What? You don't want the agent reading code files???

[lunar aurora]

a dev site?

[lunar aurora]

what

[glad solar]

yeah like a dev version. dev.yourwebsite.com

[lunar aurora]

its not supposed to do it that way

this is how it's supposed to do it but it gets rejected.

[glad solar]

no fainxc is saying how it edits files using commands

[lunar aurora]

i got a beta yeah but i dont really need it

[boreal holly]

How do you expect Codex to read code files? That's the command that for reading files

[lunar aurora]

🤦‍♂️

you know damn well what i'm referring to

[glad solar]

no the issue is when codex tried to edit the files using commands

[lunar aurora]

yes

would full access genuinely solve it tho

and im pretty sure its quicker

less tokens