codex-discussions

[cobalt junco]

does anyone have anything good for seo automation and skills and stufff. icba to manually do seo/marketing

[cyan wing]

MCP, Apps, Skills, and now Plugins

Just one more abstraction marketplace bro,
I swear bro we just need one more abstraction marketplace

[simple star]

@cyan wing You are 100% right

It is starting to get real messy, and Im not enjoying it

[cyan wing]

what's messy about this ?

[gleaming crane]

I feel like clicking everywhere, adding everything, connecting all MCPs and pray to god it'll work fine

[vapid vigil]

Mind if I chime in?

@cobalt junco , I hear you on the manual grind—but I agree with @cyan wing , most of these new 'marketplaces' just add more noise.

A more robust suggestion is to bypass the wrappers entirely and use Make.com to bridge a Headless CMS (like Sanity) directly with Search Console/Indexing APIs. It keeps your workflow programmatic and out of the 'abstraction' mess everyone is talking about.

It’s a bit more setup initially, but it’s the only way to scale SEO without being at the mercy of the next 'flavor of the week' plugin.

[woeful crown]

Plugins combine all three into one, making it easy to set up. Seems like a good abstraction to me.

[cyan wing]

Now let's abstract the Plugins too!

So I can enable certain groups of Plugins based on the workflow

[woeful crown]

You can probably do that with a script that edits your config in ~/.codex

[cyan wing]

woosh

the point is: what is the right primitive that can abstracted and recursively composed of itself
so we don't have to keep making up new words and marketplaces every 6 weeks

[woeful crown]

Plug-ins seem like a good answer to me. They can be composed of various primatives.

[cyan wing]

Can I make a plug-in of plug-ins?

[high girder]

You can SORT of do it if you layer the plugins
core plugin + domain plugins + instruction layering

[woeful crown]

Theoretically the plugin skill instructions could include instructions to install more plugins

[fossil dove]

How did y'all get your codex to update? I updated CLI and attempted to update via Microsoft Store, but there's no update in Microsoft Store.

[high girder]

I use the VSC extension, so the extension market had an update

[fossil dove]

Could be that my Microsoft store is being slow.

[high girder]

Stovoy the digger, any personal use cases for plugins that aren't listed with the OAI release you can recommend?

[granite jay]

Todays glitch:

[high girder]

A glitch, or codex building you a custom OS by surprise?

[granite jay]

And it isn't even the 1st yet.

[woeful crown]

I'll be honest, I haven't tried them yet 😄

[hard tulip]

Back to work

[woeful crown]

I am on vacation for the next two weeks actually 🙂

[hard tulip]

Good work team

[high girder]

Well that answers that then. Enjoy the vacation. Also, remote codex sessions implemented natively please lol

[woeful crown]

Yeah that'd be nice! I do use codex remotely on my desktop with a custom tmux server I wrote.

[vapid vigil]

Just pushed the initial build for a private trading 'Vault' project. CMS is mapped, $userId security is locked. Staying lowkey on the details for now. 🖱️🔥

[hard tulip]

zoom out the UI bro

[clever steppe]

WTH is this? How can I buy more Codex credits from the Plus plan like ChatGPT says I can?

[vapid vigil]

@hard tulip That's by design, bro. 🔒 Only showing enough to know the architecture is solid. The full UI is for the client's eyes only for now.

[high girder]

How are you viewing the usage page? There should be a 4th box in the bottom right with the place to add credits.

[high girder]

https://discord.com/channels/974519864045756446/1407544294532845730

[clever steppe]

Thanks @high girder. I shall double-check, but I would say there is an explicit message about 'asking my admin'. ha!

[high girder]

OH, you may be on an enterprise account then, right?

[clever steppe]

Errr.. I was when a friend bought two seats and gave me one, but that's been closed for several months now.

There were 2 separate workspaces (and now I can't even access my old chats because he cancelled his subscription 😭) and I bought my own Plus subscription through my own personal workspace.

[frosty zealot]

https://x.com/markgurman/status/2037230804942610548?s=46

Sweet if true

[clever steppe]

Man, I can't even find the Billing page anywhere. This is surprisingly nuts.

[high girder]

it sounds like you might still be using that seat in some way

[high girder]

I can't wait for the influx of old people trying to order a pizza and end up with a data center

[clever steppe]

Ohh geez.

Yep, looks like a bug. :/

[high girder]

On the bright side, if you submit a support request, you'll hear back relatively quickly. cough anthropic cough

[clever steppe]

That's nice to hear.

Maybe some free credits for my troubles? 😉

[high girder]

Don't know about all that, but they're rather generous with compute

OH man, plugins are kind of nuts once you start messing with them

[whole gate]

Ok, how do you use this new plugins thing? I don't really see anything different in environments or settings, is this not in the cloud?

[high girder]

A Codex plugin is a local installable bundle on disk.
It has a manifest at .codex-plugin/plugin.json.
It can include skills/, .mcp.json, .app.json, hooks, assets, and UI metadata.
Codex loads that bundle and gets a pre-described baseline of workflows plus tool wiring.
The plugin can be project-specific or reusable across many projects.

Plugins are great if you have repeatable workflows that you use across projects

[whole gate]

Ok, so only works with Codex in a local context? I only use the cloud agent.

[high girder]

If you can edit your cloud codex settings, you should be able to use them, but I would ask codex specifically for your environment since I dont actually use the cloud codex, I cant be 100% sure

[visual geyser]

Codex is so good

[whole gate]

I read what you typed there again, it looks like these things are set using dot files, it could be that the Codex agent will use these in the project if they are found, thank you

[grizzled mirage]

Is codex on mobile limited to pro users?

[high girder]

You'll need to be more specific, because there is no mobile version of codex officially

[fast iron]

My codex chat wasted the remainder of my tokens this week “reading” or “auditing” the repo

😎

[potent mason]

limits reset again? bruh that's awesome

[grizzled mirage]

Like this user has a codex option

Vs mine

[cobalt junco]

usage reset?

oh yeah baby

come to papa

[white furnace]

lol i was just about to check this channel because i noticed usage reset

guess im not only one

[tame magnet]

yeah yeah usage reset

[potent mason]

I'm so glad I was down like 30% of my usage from yesterday xd

[toxic torrent]

leggooooo

[whole gate]

I rarely use more than like 10% if anyone needs to borrow some lol

[potent mason]

Wish you could sell it lmfao imagine a marketplace where you sell 50% of your $200 plan for a week at $10

[whole gate]

Oh jeeze so ripe for fraud 🙁

[potent mason]

Yeap 100% would be problematic, but I would love to buy some usage for a week

(not at API prices)

[hard tulip]

Resting today makes it so I will have less overall usage.

[tame magnet]

f

lol didnt thought about that

lol

[potent mason]

Ah right they probably reset it so new usage starts on 0 April 2nd

[hard tulip]

why didnt they push back the date in the first place if that was the idea

[potent mason]

You need everyone on the same reset day?

I honestly don't know what they're thinking, they've hit that button so mucgh

[hard tulip]

that is what happens when they raise $100,000,000,000

[potent mason]

I'm already down 6% since reset though so I'm probably benefitting from it

Damn I'm flying through usage I'm down 10% from this week, I need someone to click the reset button again please

[true bane]

woah mystery token reset for plus thx

[jovial dust]

If you're wondering why the usage has been reset again, since the Codex team apparently doesn't understand that the way to communicate with their users is via the email address they used to sign up for their account, or with a message displayed prominently in the app, and instead insists on announcing things only on X which probably only a small fraction of their users ever reads, you can read about it here: https://www.reddit.com/r/codex/comments/1s4rqwh/reset_woohoo/

[frosty zealot]

@boreal holly Sub-agents now use readable path-based addresses like /root/agent_a, with structured inter-agent messaging and agent listing for multi-agent v2 workflows. (#15313, #15515, #15556, #15570, #15621, #15647)

See this patch note from 0.117.0

They're stealing your gremlins

[burnt peak]

They reset the usage again

I thought they would never reset it

[whole gate]

Somewhat true, I log in with my Microsoft account but I changed the address and their system doesn't have a way to update the address connected to the account token.

So it has an email on a dead domain

[tacit socket]

is codex meant to show plugins now? mine doesn't
nvm fixed itg

[deft gyro]

Coded quota reset?

Fire

[static lark]

Yes

[deft gyro]

I was about to run out of weekly quota too

Xd

I have some credits anyways

But goated surprise before

The great halvening

[static lark]

It's wonderful to see pro account quota reset lol

[deft gyro]

The great halvening next week

[static lark]

Yeah

April 2 right

[high girder]

Exactly, it wasn't random timing, it's a "last" reset before you lose double

[static lark]

They might extend it

Fierce competition from claude

[high girder]

probably not. They're consolidating compute at the moment. Plus Mythos isn't even ready yet

[deft gyro]

What competition from claude

[static lark]

What's mythos

[static lark]

They have always been a pain

[high girder]

Anthropics new leaked model above opus

[static lark]

Oh

[deft gyro]

At $20,000 per mtoken

[static lark]

Lol

[high girder]

more than likely. an AI company that builds AI coding tools having their unreleased model leaked through a misconfigured CMS is... something though

[deft gyro]

I think that openai should lowk let you use gpt 5.4 mini more though

[quasi zodiac]

anyoen got code for sora2 how do i get

[kindred moss]

Is anyone having issues with Codex using too much usage after the limit reset?

/fast off

[deft gyro]

Wait lol did the limit halvening hit already

Ive used 40% of my 5h already

Wut

Yes

[kindred moss]

even i cant see 5 hours limits

this 25% was just Hello message and like reading 1k LOC file only

same for 5.4 and 5.3-codex

[cedar skiff]

ok it really feels like someone hit codex with a car

I mean it's just stopped trying. I doesnt even take one step through the code before making an assumption and suggesting a fix.

Litterly has no idea what the code does, just guesses. You cant trust it on the simplest job anymore.

I thought it might be something i added in agents.md or skills etc, so i sanitised those. Still no change. Then i explicitly added developer_instructions, still no change. So i have to resort to telling directly every prompt that it must step through code until it reaches concrete implementations?!

It's like high thinking is locked down to low

It's just constant

[potent mason]

GPT 5.4 mini has been out for a while what did you guys think about it? I haven't touched it much but wanted to know from people who have used it more what they thought

[potent mason]

Yes I'm down 13% of my weekly usage

[static lark]

Model is at capacity ?

[static lark]

Might be good for small work, but I prefer Spark instead of mini

Spark is fast

[potent mason]

I can’t deal with spark though, it does things so badly I spend more time fixing it than what I save by using it

[static lark]

Can't use those models for heavy work

Very low context window it has

[idle pilot]

Support Intel Mac. 😢
https://openai.com/codex/get-started/

[wind cairn]

stop using antiques

[idle pilot]

Works fine still with wired internet (it's wireless is too slow)... I don't typically even develop on it... I use Claude Code on cloud machines. Who needs an IDE these days?

[idle pilot]

I also can't deal with stuff breaking (thus my preference for cloud VPS)... I had a better Linux machine, but it failed after only a two years (cpu or board). So nope, this beast has lasted this long (9 years!) and has proven itself worthy of continued use. 😄

[plucky halo]

Does anyone know of a service (or a skill for codex?) that I could use to create a promotional video of a website showing the different features. Absolutely out of my skill set

[tranquil hare]

Do you know how to use hooks? I heard they're available now, but I couldn't find docs

[ebon ingot]

Yeah my 5 hours usually take 4 hours to get 0 but now in less than 2 hours

[boreal holly]

It's great! I have a disproportionately higher number of them running at all times

[boreal holly]

@simple star used codex a while back to make a video. It's in this chat!

[lone python]

I think there is no difference in rate limit for free and Go subscription in Codex. It’s sad! Please do something for Go subscribers.

[simple star]

Yes, I did. but I cant remember the skill, ahahah

[unborn thunder]

i think it depends on the workload

[cedar skiff]

give us spud already

[deft gyro]

Did they prefire the halvening

Like

The flip?

[ebon ingot]

But yesterday was normall

[fallen path]

i got infinite codex usage what should i do

/ codex api usage like for openclaw and stuff

[cyan wing]

rewrite all of the Linux kernel into Rust

[unborn thunder]

how?

mine reset today which felt nice..

[fallen path]

dang deleting that is crazy

[lean lark]

I just noticed that the VSCode Output window for Codex is streaming a ton on of debug/error detail non-stop ... when Codex shouldn't be doing anything.
I think this needs to be reported but I might be footgunning with a debug flag set somewhere.
Anyone?

[fallen path]

ye i had that happen too lol

[lean lark]

hmmmm, I just nailed this down...

[error] Error fetching errorMessage="open-in-target not supported in extension" errorName=Error errorStack="Error: open-in-target not supported in extension
at open-in-targets (/home/codex/.vscode-server/extensions/openai.chatgpt-26.318.11754/out/extension.js:256:22058)
at qw.handleVSCodeRequest (/home/codex/.vscode-server/extensions/openai.chatgpt-26.318.11754/out/extension.js:256:26917)

I'll check GH Issues and report.
TY

[lean lark]

Verified the issue. Update VSCode and the Codex extension to very latest versions and issue may/should clear.
This has also been documented as causing CPU spin in MacOS.

[junior umbra]

I want to use codex for a public project, but I don't want to commit my personal AGENTS.md to the remote repo. What's the best way to version control those files in a separate repo, while ensuring they're still in the right directory, and not appearing in the unstashed file list?

[lean lark]

How about symlink and add the specific folder/file to .gitignore?

[cedar parcel]

Git submodules

[unborn thunder]

cursor or vscode?

[lean lark]

Speaking of Git, everyone should be aware of a change in GitHub policy:

On April 24 we'll start using GitHub Copilot interaction data for AI model training unless you opt out. Review this update and manage your preferences in your GitHub account settings.

[lean lark]

Shouldn't be relevant

[unborn thunder]

how so?

its the IDE recommended in codex

[lean lark]

He's asking about Git, which is independent of the IDE. He could be working from the CLI or Notepad, doesn't matter.

[unborn thunder]

oh loll im wondering which is better

irrelavent to his post

[lean lark]

OH 🤣

Sorry about that....

Context matters. ::)

[unborn thunder]

is vscode any better than cursor?

both using ai

just downloaded vscode and realized i mightve wasted my time

[lean lark]

@junior umbra the symlink approach is good. Treat this exactly like a .env file or other config files for lint or formatting, etc. Your project might include a AGENTS_Template.md for your package users to copy/modify.

@unborn thunder The IDE doesn't matter but I think I see VSCode mentioned here more than Cursor. That said, Cursor is highly regarded by OpenAI and they are profiled as valued partners in presentations.
Personally, and it seems in most references here, I use VSCode with the Codex Extension, which is a great onramp to using Codex with code after Codex Web. This vector can help introduce you to Codex CLI and then the Codex app.

[unborn thunder]

is codex web the same as codex cyber?

[lean lark]

There's no such thing as "codex cyber" in this OpenAI ecosystem, can you clarify?

[unborn thunder]

there is

https://openai.com/index/trusted-access-for-cyber/

i have access on my phone

lets my select any repo from codex

instead of one like on chatgpt

[lean lark]

Oh that ... it's not a product called "Codex Cyber". "Cyber" is just a short name for CyberSecurity. OK, we're on the same page now.

[unborn thunder]

lol ok

[lean lark]

So no. Codex Web is unrelated to the security initiatives.

Codex Web is a web page where you authorize use with GitHub repositories, then open a repo for discussion, and tell Codex what you want to do. Codex then creates a VPS instance to clone the repo, processes the request, then allows you to create a PR with the changes.

[unborn thunder]

it was codex cyber at one point i think idk

maybe im wrong

[lean lark]

The IDE extensions are closer to the metal where you clone to your own system, and codex operates there. Then you decide if and how you want to manage the VCS.

[unborn thunder]

"GPT‑5.3‑Codex is our most cyber-capable frontier reasoning model to date."

[lean lark]

Yes, "cyber-capable" ... not "Codex Cyber".

In other words, Codex 5.3+ has (better) understanding of security concerns than prior versions.

[unborn thunder]

they just call it chatgpt.com/cyber

but it brings you to codex

it confuses me but Codex cyber works for when i think of the web verison

lol k bye

[lean lark]

Yes, this is because some projects seem like they work with security and need that verification.

Lots of tools with the same name. 🙄

Your head'll spin when you see that Codex plugins are not the same as the Codex Extension. 🤦‍♂️

[main nimbus]

Welcome to the trap that is OpenAI naming

[unborn thunder]

its just manipulation

from a non-profit company, they keep me entertained

[south latch]

Someone till me why codex 5.3 is better then 5.4.

Have someone notice that 5.4 is Kinda trash? Doing wrong things and cant understand shi

[hard drum]

I can't use caps lock/uppercasing in Codex CLI chatbox?

[oblique cove]

I mean it’s not guaranteed to be better in every way that’s why you get the option to use the other models

[true bane]

if you're wondering how the other half lives I wrote this up while waiting for rate limits (no self promo) https://twolongos.com/3/27/claude-code-capacity-collapses-calamitously/

[wide schooner]

codex cyber is the verification site lol

[solemn acorn]

5.2, 5.3, and 5.4 are all the same base model, just different RL

changes openai makes to them comes with tradeoffs

[round flume]

when i open the codex app, it shows plugins for ~1s and then disappears. anyone knows how to fix?

i cleared these, logged out/in, and restarted the app and same problem.

• ~/Library/Application Support/Codex/Cache
• ~/Library/Application Support/Codex/Code Cache
• ~/Library/Application Support/Codex/GPUCache
• ~/Library/Application Support/Codex/Session Storage

Codex 26.325.21211

probably a server-side rollout/eligibility flag that flips after hydration

[hard drum]

is it normal for gpt-5.4's "1m" context to drop very fast according to the statusline by not having much done at all?

[solemn acorn]

you don’t get a 1m token context through codex

also you wouldn’t want it either because it would have degraded performance

at least the way anthropic handles it, you basically get bumped down to a worse model if you use the high context ones (they similarly don’t allow those in claude code)

[full moth]

Figma plugin doesn't work?

[quasi summit]

Claude benchmarks much better with 1m context coherence

Hopefully gpt catches up

[hard drum]

Interesting...

that seems a bit unfortunate

would be nice of them to catch up, because the outsider's 1m context is actually pretty damn good for what it is

[solemn acorn]

the 1m context variants of anthropic/openai models are universally worse than models with smaller context windows

they're separate models

also extremely expensive which I imagine is a contributing factor

[potent mason]

Damn something is definetly wrong with usage I'm down almost 30%

[worn harness]

I noticed I had fast mode enabled for some reason, might want to check that

[dire drift]

It's like they already halved it.

[potent mason]

Bruh when they actually half it I'm going to be in huge problems.

I've used 30% in about a day, that's 60% when they half the usage

[lunar aurora]

plugins shows for a split second then its gone

i did update the app, the plugins is not there.

[hard drum]

INSANE value for 200 buckaroos a month

[glad wind]

Subagents completely broken right now for me. The subagents output is all about trying to spawn subagents.

[cobalt junco]

is codex down rn?

[nocturne folio]

plsssss tibo reset usage again

[cobalt junco]

i already drained 3 accounts llol;

[twilit bluff]

still working for me

[hard drum]

how?

[pale oyster]

Ugh... After updating Codex this time, the pop-out window feature is a total mess.
It used to allow individual pop-outs for each thread, but now it forces you to use only one pop-out. T_T

On top of that, while you used to be able to resize the pop-up window, now you can’t even do that anymore.
It was a feature I used all the time, so it’s frustrating that they’ve forced this change on us... How about the rest of you?

[pale oyster]

https://github.com/openai/codex/issues/15162
i find same issue

[royal mirage]

Hey hi guys does anyone knows why there is so much diff in token usage in chatgpt (codex) vs claude code ??

🥲 no one here to reply?

[winged depot]

Why would Claude charge you $900 for 780k tokens?

[chrome raven]

any solution to my local plugins are not showing on codex app but does on cli

anyone else facing this?

[warm pilot]

Did you benchmark this with exactly the same code base and prompts? In my experience gpt is a little bit better with token usage vs. opus … but overall they should be on the same broad level.

[exotic cave]

In my projects Claude is behind in efficiency and coding ability. I have both so i can compare end results and Claude is worse in alot of my own benchmarks....

[whole gate]

My application generates the RPC namespace bindings for the typescript front end dynamically as part of the build process. The Codex cloud agent also does this as part of the setup stage so it can work with the files. The generated tree goes into /frontend/scr/rpc - the entire folder is dynamically generated, and the entire tree is in .gitignore - and Codex regularly ignores the gitignore and checks generated files in, especially if anything is mentioned about generated files at all. Codex needs to not ignore .gitignore and stop checking in files it should not.

Cannot post to bug channel, would, otherwise...

[royal mirage]

coding ability?? seriously ?

[cedar skiff]

Yeah, codex is better at code, it is just harder to prompt. It's better at remembering and following tasks. Claude forgets to use skills all the time.

[vocal sparrow]

Hi

[whole gate]

Codex seems to have an absurdly large token grant for code work

I barely ever used my token allocation in Codex, but I don't use it for planning any more

I am about to switch back down to Claude Pro instead of Max, though. Claude has optimized my code base so well it takes a lot less work to get things done now.

I added a dynamic page editor and wiki system in like three or four prompts

Codex cranked out like 15K lines of code for that

Claude's prompts were a couple pages long for each

Here are some thoughts of mine on the topic
https://elideusgroup.com/pages/emergent-architecture

Published using that system I was talking about 😉

[glad wind]

I think you're quite right. I am getting a lot out of generating very detailed (even 1000+ line) implementation plans, iterating with highest iq models available, and then getting Codex to systematically translate them to code. Then they seem to be "sticky"

[whole gate]

It's an algorithm. It's a really fancy one, but it's... just an algorithm.

🙂

[glad wind]

it's certainly algorithmic that's for sure

[whole gate]

I realized a while back that the way we're doing reasoning in LLMs is actually through verbal expressions like it will generate something like a mad-lib. Because of blank I will blank and then it inserts nouns and verbs into this frame. It's reasoning with language the same way people do. Which is the form of reasoning, but not the act of reasoning.

The act of reasoning is drawing the most relevant context about the outcomes of a decision into place and making a moral judgement that is almost always ultimately derived from the (Sorry for the morbidity) the mortality of humans giving value to life and therefore influencing our ethical outcomes.

We need to start teaching AI how to reason ethically while also understanding this thing is going to be like a god that will live forever, so we better not get it wrong.

I've been developing an advanced memory context system that addresses this, I think.

[glad wind]

Seems contradictory to say reasoning is fundamentally grounded in the way we care about the world but then that AI could be taught to imitate that

[whole gate]

🙂

[cedar skiff]

llms don't have any actual comprehension though. So at best all you can do is make them guess better.

[whole gate]

I think SciFi Fantasy both give us contexts to understand immortality and that's probably the reason things like Vampires and Elves and Vulcans exist.

[meager wadi]

using custom agents though the codex app? codex says it can't use a specific agent apart from the deafult ones?

[whole gate]

Seems like you have to set up dot files to configure things like that?

[crimson hazel]

which one do you think its better for coding, 5.4 or codex 5.3? both on xhigh

[boreal holly]

5.4-mini on high is best for coding. 5.4 on medium is also good, but not by much. xhigh is wasteful

[boreal holly]

You might be able to use CLI to get the agents going, then switch to desktop app and resume them

[warm pilot]

I'm basically only using 5.4 high nowadays, it does everything I throw at it – although it might be too much most of the time

[cyan wing]

how are you so sure it doesn't have comprehension?

[meager wadi]

Yeah it works through the CLI, looks like a codex bug. I think this is the issue for it: https://github.com/openai/codex/issues/15250

[wicked briar]

this has been happening frequently

lost two important chat sessions

pls fix

session doesn't work after we get this error have to start new session

[boreal holly]

Do you have

[features]
sqlite = true

[wicked briar]

let me check

[wicked briar]

no

infact its not even there no sqlite line

[boreal holly]

Good! You might be able to do /status in that convo, take the thread ID, and look for the rollout log in ~/.codex/sessions/**/*.jsonl. There may be a line in the file that is malformed, and when it sends the compaction request it isn't able to read it.

[wicked briar]

I asked codex to double check

and I think I have it enabled

[boreal holly]

Oh dangit that makes it quite a bit more tricky. But since you have 5.4 xhigh on the case you can have them check the thread in the sqlite file and see if there's a malformed item in there

[wicked briar]

what should i ask

[boreal holly]

Copy the error message, copy the /status thread ID in the broken convo, and tell the agent there might be a malformed item in there

[wicked briar]

ok

done

[boreal holly]

I think sqlite = true is probably the reason you are running into it often. If you are running multiple CLI's at the same time, not under a single process, that single file is being written to and multiple sessions can end up clobbering the chat history. The old ~/.codex/sessions/**/*.jsonl way makes each thread a single separate file, so the only clobbering that can happen is if you open the same thread in two different CLIs at the same time (not under an app-server).

[wicked briar]

I think you were correct

[boreal holly]

OK, if there were no corrupted lines, it's possible that multiple CLIs were running at the time of compaction and it was running a SELECT command on an actively changing SQLite file. It's my understanding that normal conversations it's only adding insertions to the DB, but when it does compaction it runs SELECT, and if any insertions happen during that time SQLite can bug out. Maybe run that scenario by your codex

[wicked briar]

acc to codex this is best fix

[wicked briar]

I run multiple sessions at once true

im not sure if the limits is good idea should I remove it

[boreal holly]

You should try:

[features]
tui_app_server = true

Then in a terminal you would run codex app-server --listen ws://[::]:4200, and for each of your other TUI's you'd run codex --remote ws://127.0.0.1:4200. That way the TUIs are speaking to a single app-server which will serialize access to the thread DB

[wicked briar]

okay

[lone yarrow]

Why do Weekly usage drops faster than the daily ?

[plucky halo]

Because it's designed to not slow you down but also to stop people taking the mick. It's better to think of the 5 hour limit as a safeguard for abuse whereas the weekly limit is the actual one to pay attention to

[boreal holly]

A week has 33.6 "five hour windows", but the 5 hour window itself is something like 15 or 20% of the weekly quota

[waxen terrace]

Has anyone recently compared Codex and Claude Code and can tell me what they think and make a recommendation?

[bold pilot]

I'm running both models at the same time in claude code and I can say that the difference is huge, not in a worse or better way. Claude is a lot more natural in conversations, more concise and also understands typos as typos easily, while gpt is naturally verbose and stiffer. Edit: but this is more about models than what you actually asked about.

[plucky halo]

I use both. I prefer Codex. I tend to use both of them to plan (and to check each others plans) but codex is the one that actually does the coding. Sometimes Claude if Codex has been stuck for a while. But I found I generally get better results from Codex

[cedar skiff]

yeah, it cant have new ideas. 100% sure

[waxen terrace]

And what about code quality?

[waxen terrace]

Do you would say the code quality of Codex ist better?

[boreal holly]

New ideas require input from the environment

[cedar skiff]

And comprehension

It's the reason it feels like the thing is the smartest entity you ever encountered on one hand, but on the other hand can't see the forrest for the trees.

It does really understand anything

[boreal holly]

That's because humans are forced to operate on continuous input and output as long as we're alive. LLMs are limited, can't work autonomously, trained to find the most efficient way to reason about and execute something within the reasoning budget available.

[cedar skiff]

they are pattern matches though, any thing that seems like understanding is just a good guess of tokens.

Its a simulation of understnadning not actual understanding

And it's really clear when you hit one of the idiosyncrasies that it didnt get a good connection on in training.

[bold pilot]

Haven't used it enough to say anything about code quality, just hacked it in claude code yesterday and the difference shocked me. I know there's people that prefer it so I think it's at the very least on par, I still prefer interacting with Opus, by far, this is just not my style, but I plan to use it as a sole reviewer or in quorums.

[plucky halo]

Yes

[cedar skiff]

Those idiosyncrasies show it cant reason about the simplest thing with out been given the correct token string. Even though it has everything it needs already

[waxen terrace]

Okay, thank you

[boreal holly]

That is a slightly over-simplified way to describe them. Sure, it's based in probability and pure math, but LLMs have one interesting property that makes them more than good pattern matchers. When a new model is created, they initialize the matrices of numbers with random values, and they have pure math algorithms to describe how those layers interact with each other. They could take two completely different randomly generated matrices of numbers, run them through the same exact training dataset and process, end up with two completely different "trained" matrices, and yet they act as almost completely identical models. It's like two humans that are biologically fraternal twins, growing up together, learning precisely the same thing, experiencing the same experiences, eating the same food and sleeping at the same time, being biologically unique, having completely different brains, but their understanding of the world is somehow almost perfectly the same. If you compare the actual numbers in their neural net, they are entirely different, but they produce almost the same responses.

It's a simulation of understanding not actual understanding

I think LLMs are a step towards true understanding. Sure, right now it's hard to see them as anything other than pattern matchers, but if you look at what they're made of, how they're modeled around human understanding of language and the brain's neurological functions, I think the next step is to give them more environmental inputs continuously and let the parameters train themselves. We're special because we can wake up in the morning, and depending on whether we eat an omlette or a bowl of cereal for breakfast will determine what we think about later in the day. LLMs can only think about what is right in front of them, they can't forget their training, and they can't remember more than a relatively small amount of information.

[cedar skiff]

The point is as they are they don't actually understand anything.

[boreal holly]

Gotcha, so it's more philosophical than textbook definition

[cedar skiff]

By the book they have no real understanding or ability to comprehend any idea that it didn't get a good analogy on in training. Even though it has all of the information it has zero ability to comprehend those missing links with out being directly given the link.

[boreal holly]

Are you suggesting that people work differently?

[cedar skiff]

We can learn as we go

so yeah we work differently

We have the ability to simulate creatively. Which enables us to make new connections that were never there before.

[teal blade]

https://join.pump.fun/HSag/n6ua2e64

[boreal holly]

I think we agree that LLMs are limited in what they can intrinsically comprehend with minimal inputs. They can't spontaneously generate abstract thoughts. Technically neither can we though. If you don't prompt a LLM, that LLM does not exist and is frozen in time. It only exists when it has an input to process. We exist as long as we're alive. If time were to stop, we would have the same problems LLMs have. But I don't think understanding and new ideas is just a unique human trait, I think it's because we have more continuous and high entropy environmental inputs and are basically forced to evolve. If we gave LLMs the same affordances as humans and designed them to be continuously trained they might actually be more than the world's greatest information compression algorithm

[cedar skiff]

that's the golden egg everyone wants to hatch

[potent mason]

Right but so can they with RL

If you really think about it all we need to do is feed them continous data to do RL by themselves

I think we don't because we would rather have the actual smart version gpt made than one that continues learning but continously gets dumber

[cedar skiff]

afiak post training is the most expensive part of building out models as it is. It takes time, money and human input.

Jon Carmack was working on fast training as an idea.

his thought is if we can get a model to learn in a fast manner and not forget previous teachings some how we get a step closer to agi

[frosty zealot]

At this point you would think the learning from online material must be done

It really depends on human correction factors to push it further, maybe im wrong

[potent mason]

Oh no I'm past 60% on my weekly limit

Can we get a reset guys at openai?

[short kindle]

Anyone else having issues with codex? Just randomly starting only looking in a sep drive I have instead of the normal dir i have it already set in. Been using it for about 30 minutes and on the last message it now wants to look in a different drive alltogether lmao

[surreal wing]

hello guys i want to ask something what is the best chatgpt model to code like 5.4 or 5.3-codex

[boreal holly]

5.3-Codex is overall good out of the box. 5.4 is extremely powerful but out of the box takes time getting used to

[surreal wing]

he made this

i love him

who needs css

if you just can do this

[cobalt junco]

anyone know what the hell openai's new spud model is?

more importantly, wen agi

[cedar skiff]

agi tomorrow after recess

[cobalt junco]

i hope mythos leak wasn't just marketing hype

[deft gyro]

did codex rate limits get like

prehalved?

or is it just me

[chilly glen]

Has anyone had problems with not being able to open or do really anything inside of threads on the app?

[unique spade]

the funny thing is that to define "actual understanding" you need to process a lot of philosophy on that subject.

[chilly glen]

Is there a place i could get help with the app?

[main nimbus]

not for me. are you using a lot of subagents?

[solemn acorn]

codex limits seem to be about the same for me

though GPT-5.4-Mini is good enough for most stuff I've tried it with that I'd recommend using that if you're running out of usage

seems to perform similarly at least to sonnet

[umbral blaze]

Hi everyone

[cedar skiff]

only if you want to argue about semantics. or you can just accept the common definitions that are applied in the industry and move on.

[cedar parcel]

Gpt 5.2 xhigh still beats 5.3 and 5.4 hands down at everything. Give them both the same complex task on the same codebase and 5.4 will give you a wrong diagnosis in 5 minutes while 5.2 one-shots the after 40.

[loud dragon]

might want to remove that and cycle your api keys if those are legit credentials

[slow flax]

Guys if someone wants to use codex heavily for cyber security and pen testing how to use it properly? I.e it just rejects most requests as openai almost don't want it to be used in that domain?

[loud dragon]

did you ID verify for trusted access to cyber? https://chatgpt.com/cyber https://openai.com/index/trusted-access-for-cyber/

[slow flax]

Does that allow that? I thought the verification is just to not get shadow banned / get routed go lower model if it thinks you are doing some cyber security task? Problem is still same isn't it?

[loud dragon]

at one point people who had not verified were getting routed off of 5.3-codex. eventually they added a notification in the cli to indicate you needed to get verified. but read the second link, if you need more permissive access than current models provide there's a good form to fill out to request access to a special model for sensitive work

[cedar skiff]

5.4 is is so weak sometimes. It doesnt check just assumes.

[loud dragon]

on xhigh? i haven't seen that behavior

[cedar skiff]

on high, it wasnt like this the first few weeks

It give answer where it didnt check so it give syou the branching choice.

[loud dragon]

in planning mode?

[cedar skiff]

do this: something here
But if this do this instead: something else here
And the branching idea is the thing it can just check easily

Just in general

[loud dragon]

try enabling codex --enable default_mode_request_user_input, it lets it ask questions using the multiselect tool outside planning mode. i find it comes up sometimes during normal operation and is quite useful

[wooden minnow]

oops my bad lol thanks

[cedar skiff]

Its not that, it just doesnt do the work anymore

[loud dragon]

in my experience though, at least on xhigh, it's pretty thorough, so long as you ask it to investigate the codebase during the prompt. otherwise it may just take what it knows from the AGENTS.md to answer the question

[cedar skiff]

I have to resort prompting it to actually doing the work which i didnt have to before. Like i have to tell it to step through abstractions all the way to concrete implementations etc or it wont, it will make assumptions when it thinks its knows. It used to just do that.

It takes short cuts now.

and produces bad results on complicated tasks

So i have to add a prompting layer i didnt need before

[loud dragon]

not sure. i'd put it down to either a change in harness or your codebase. i doubt they are stealth updating these models periodically

[cedar skiff]

I thought it might be my code base - even though i hadn't changed much to make it happen like this, so i sanitised all my skills and agents.md files. cleaned it all up made sure there was nothing stale etc.

The only thing that helped was giving it explicit instructions on how to work

but on top of that it happens on simple things as well - just asking it to search documentation and bring back an answer it comes back with the completely wrong answer.

[fading spear]

time to reformat windows

[cedar skiff]

Then swears by it, i force it to do the actual work, and then it's like - ok the opposite is true. The other day when i was having all these issues and i was looking for a solution i found developer_instructions. I got codex 5.4 to search the docs and let me know if i could put it in nested config.toml. It comes back saying it isnt documented and doesnt think it will work. I say please find and list all the documented vars that are supported...it comes with developer_instructions in the list of documented items.

Just cant trust it.

[fast iron]

I graduated in 2025. It's so sad I'm not a student anymore for this codex challenge haha

[fading spear]

currently earning my vibe degree

[royal mirage]

why codex is slower than claude code?is there any optimizations i can do to make it response faster?

for even easy tasks its taking 5-10 minutes lmao

[fading spear]

I've considered finding an openai employee to date, I'm on that token hustle

[winter flicker]

helo

[static lark]

Lol

[cedar skiff]

are you using 5.4 or 5.3. I moved from opus to 5.3 and 5.3 was about the same. 5.2 is noticeably slower, but a little more on point

[lyric vector]

where should we post codex bug reports? i can't post in #1070006915414900886

[exotic cave]

This is unhealthy...

[deft gyro]

No subagents, but 2 simul tasks usually

One gpt 5.4 on high

One gpt 5.4 mini on medium

[deft gyro]

Time is a factor

like

sure

gpt 5.2 could be better for long running tasks

thats literally what its optimized for

but like

[kind jay]

What did you run 5.3 and 5.4 on?

XH is known to preform worse than H for many tasks on 5.4

[cedar skiff]

5.2 is better, it doesnt take any short cuts, 5.4 is the worst of the three 5.3 is ok

5.2 stops when the plan has a conflict with repo rules and mentions it and asks how it should move forward, 5.4 is just ok...and breaks the rules and writes the code

[cedar parcel]

A (Drupal commerce) checkout page JS issue. Any codex model doesn't gather enough context so they never work for me in any of the codebases I use. But 5.2xhigh is perfect. I have tons of examples of >5.3 models producing bugged solutions I frankly don't understand why they work well for people.

[cedar skiff]

Do yo use 5.2 or 5.2 codex?

[cedar parcel]

Just 5.2xhigh.

[kind jay]

Check benchmarks??

[cedar parcel]

Benchmarks are pretty useless ATM.

[cedar skiff]

why do i need to check bench marks? I have plenty of first hand experience that shows the idiosycrisies in real project work.

benchmarks don't show anything real

[kind jay]

Maybe, could be some certain case for sure
To be honest to things like this surely ChatGPT is better due to the additional tooling

[orchid plume]

the only benchmark I pretty much look at is Voratiq's, because it's not as easy to benchmax that due to the data in the tests not being completely public. It also seems to align with my perception of the models

[kind jay]

Because benchmarks are generalised, not your one specific view of such. But I understand it might be hard to think of it that way

[kind jay]

Useless when you want to ignore the clear point that they didn’t release worse model after the fact

[cedar parcel]

I frankly worry about this more and more because I think I understand why openai is trying to make the models respond faster. Adoption is quicker of the product of it works faster. But I really rely on the thoroughness of 5.2. I'm thinking for the first time about open source models and investing in building the hardware on prem to get consistent performance. If openai deprecated 5.2 I'd be f-ed.

[cedar skiff]

bench marks wont change what is happening though, i can look at benchmarks all day, then when i check the pr 5.4 made and realise it went off the reservation again how do benchmarks help me there?

[cedar parcel]

I think 5.4 codex is more intelligent but it's wired differently to fake speed which comes at a cost of quality and if you run large production codebases then that's a fatal problem.

[kind jay]

Honestly this is a separate issue. However to answer what’s the point of a benchmark, it’s to give users an idea of the objective relative performance of the models so they can make decisions on which to use

[cedar skiff]

So when i turn around reject the pr and give the exact same task to 5.2 and it does it correctly how do you explain that difference?
I don't think telling me to look at bench marks is going to help here at all.

[kind jay]

This starts to make a bit more sense, you shouldn’t be using codex in this sense, it should be used on small isolated sections of the project
However you might say that’s not possible, but that’s because it’s completely vibe coded from the start

[cedar skiff]

5.4 makes more assumptions and infers more without actually checking, you have to add another prompting layer to get it to check deeper.

or

[cedar parcel]

You clearly don't know what you're talking about

[cedar skiff]

just use 5.2

[kind jay]

Why are you giving it so much freedom in the first place? It’s user error why the benchmarks seem irrelevant to you then

[kind jay]

Ad hominem

[cedar skiff]

You missed the point didnt you, your argument is give it more guard rails but my point is i didn't need to with 5.2 so 5.2 is better

[kind jay]

That’s so subjective, the test space is so small, which is exactly what benchmarks avoid

[cedar skiff]

You clearly just do trinket work or you would actually have an idea about how much of your prompt layer had to change with 5.4

[kind jay]

I don’t let LLMs do all most work for me because I understand their limitations, I don’t complain when I can’t say “make me a website to make money” and it fails to do so

[warm pilot]

when everything else failed, … I might just try that 😉

[cedar skiff]

Youre just making stuff up now, like i said the same stuff works on 5.2 with just s straight model change and the same plan.

[kind jay]

As you should, it’s just crazy to complain when it doesn’t work

[warm pilot]

my trinkets still work fine

[kind jay]

Use a worse model I don’t care. It’s not great that you’re trying to spread misinformation but at the end of the day that’s on you
I’m clearly not going to change your opinion so good luck with it

[cedar skiff]

Youre the one who has to limit the size of the task and autonomy of the agent not me.

Well if i use 5.4 i have to add layering to get the same job done

Just more work i dont want to do

[kind jay]

Additionally I would ask how much money you’ve actually made from your “serious work projects” but I know you wouldn’t tell me the truth, but just think about if your current methods are really working

[cedar skiff]

I work from home, i was free from being paid for my time before llms were a thing and now i just use them to work on what i already had.

using money as an indicator is a bit of weak direction.

Can't win on merit of argument so you swing to a completely different non issue.

llms dont make money, everyone has the same tools

[kind jay]

It’s suggest this as a metric because it’s clearly not for learning

And I’m not looking to know, I’m suggesting you think about it for yourself though

[cedar skiff]

i mean you dont need llms to make money for a project. So the metric is worthless.

HoW MuCh MoNeY YoU MaKe WiTh LLM?!? most ppl havent made money probably you included. The ppl with established business just get to use it to strengthen what they have.

Which is me

[wooden minnow]

hey how do we make codex to not stop working, like i've been doing a kernel optimzation but the codex keeps stoping even though I specify not to
I want it to be keep on going until the optimization reaches its goal, i've tried a bunch or stuff but its still stopping after certain few gains

[fading spear]

❤️ = 5.2 xhigh 👀 = 5.2 non-xhigh

🔥 = 5.3 xhigh 👍 = 5.3 non-xhigh

🍌 = 5.4 xhigh ✅ = 5.4 non-xhigh

👌 = 5.4 mini

[royal mirage]

i am using 5.4 xhigh fast

[deft gyro]

Are you using 5.2

5.2 tends to be better at long tasks

[wooden minnow]

oh I'm just using 5.4 xhigh

[deft gyro]

are you sure about that

jk most people probably havent

[cedar skiff]

Yeah most devs are still just doing what they always do, just now they work in a different way. Some ppl with smart ideas find niches to make money, but not most.

[deft gyro]

this server should allow gifs

I have good gifs

[boreal holly]

Just throwing this out there, 5.4 has a drastically different system prompt from previous Codex models. You can get roughly the same performance if you use one of the old system prompts

[deft gyro]

I thought they kept system prompts secret

T.T

[cedar skiff]

Interesting, i never ventured into finding system prompts and setting them in codex, i did it in claude cdoe

[boreal holly]

Nope, they have a JSON you can download that has base_instructions for every single model they let you use, and you can override them 😁

[cedar skiff]

most i did so far with codex was inject developer_instructions

[deft gyro]

ah ok ic

learn something new every day

I am relatively mid at using llms

just a good programmer xD bad prompter

which reminds me

i probably need to get the 200 dollar plan soon

Ive gotten my money worth out of this month

[boreal holly]

The 5.4 system prompt is a big, fat, generic blob of "be helpful", and then whatever personality you select (pragmatic, friendly, etc.) gets injected at the top. The Codex model prompts are much shorter, structured, and terse in comparison. I did try swapping 5.4 with the 5.3-codex prompt and it started operating almost exactly the same.

5.4 is extremely steerable compared to the old models. I throw stuff in there that it follows even after 10 compactions. But yeah the base instructions are the keys to the kingdom

[deft gyro]

I wonder why they dont make that the default

I assure openai

NOBODY likes talking to gpt 5.4

[boreal holly]

Lol a long time ago I tried changing the system prompt and it used to reject the changes "ChatGPT subscribers not allowed to change system prompt, use API instead." I guess they changed that policy! I think the 5.4 prompt works for most people - they have like a billion users or something and it says stuff like "never use destructive commands, always use apply_patch, stop immediately any time you see a file change that you didn't make" all sensible default behaviors. I'm like "good luck using destructive commands (clean up after yourself), idc about apply_patch, ignore file changes you didn't make"

[cedar skiff]

Thats good info thanks

[fading spear]

s/good/funny

[boreal holly]

s/good/funny/g

[steady vigil]

its amazing how little changes in a system prompt can have such crazy and unpredictable impact on inference, I was trying to tweak the claude code prompt to be less 'claude-like' with GLM models and the results were... interesting. (mostly bad)

The one thing is for sure, if you absolutely need to force a behavior system wide the system prompt is the best place, it will follow it much more strictly then AGENTS.md or any user prompt

[cedar parcel]

I did that, custom profile but it's not enough. There's something that makes it not acquire context that even a different system prompt can't negate

[deft gyro]

How much bigger is the pro quota than plus

do we know

[twilit bluff]

you'd probably consume it in 5 days

plus plan only lasts for 2 days

[cedar parcel]

Some of the most detrimental parts of the 5.4 system prompt is that it's okay to work on assumptions and to keep moving. It helps to change that to the opposite instruction in a custom profile but it's not enough. I can't get it to behave like 5.2 no matter what.

[deft gyro]

plus plan lasts for like

at least

3 days

once the halvening happens im so cooked though

tbh

might have to fork over the 200 bucks

I think ill do claude max tho

if I upgrade

[cedar parcel]

I'm currently on 2 pro subs + 1 plus. Still could use more

[deft gyro]

unless oai comes out with like a 50 or 100 dollar plan

[deft gyro]

what do you do

Actually

if you use fast

that makes sens

cos thats 4x + 4x + 0.5x

so like 8.5x my usage

I could see that

and when you consider its 1.5x faster inference on fast

[cedar parcel]

Running a business haha. I have codex do everything, so it's working on several projects, adding a feature, fixing bugs, performing updates, pen tests, building new projects, it's just a bunch of orchestrators churning through work queues basically.

[deft gyro]

thats only 5.6 * the llm hours

T.T and that business is profitable

[cedar parcel]

And that's just me I have a team I'm not counting their subs, just mine which do the bulk of the work

Yes I'm not complaining but I'm worried about the reliance on any given vendor, ultimately. OpenAI has an incentive to speed up their models to benefit adoption. But for the coming years that won't be backed by available compute so they're cutting corners somewhere to speed up the responses. This is opposed to the outcome I need, so that's why I'm starting to think if maybe open source and on prem is the way to go...

I need no limits and predictable performance. That's all. If I can have that then the sky is the limit in my opinion.

[deft gyro]

I mean oai will continue to make top tier frontier models

the pattern of new huge mdoel then distill is obvious

[cedar parcel]

Yeah but from what I'm reading, some of the OS models are now just lagging behind flagship proprietary models by about 3 months. The rate of innovation is hyperbolic right now so there's no reason why in maybe 6-12 months from now you can killer models running constantly and serving a small team without limits and also, totally private.

[deft gyro]

ok but are you willing to spend the money

the api has no limits

currently ai usage is super subsidized

(on subscriptions)

very unsustainably so

[cedar skiff]

I think he means with hardware innovation and model optimisation we might be able to run a model like codex 5.3 level locally at a decent value proposition.

Or perhaps hosted

[deft gyro]

and im saying it will nearly always be cheaper to rent gpus in a cloud

because

economies of scale are a thing

and its a bit unlikely that api cost or hosting will be cheaper than subscription

[cedar skiff]

well nvidia sure dont wont it to happen that you can run it at home for cheap

[main nimbus]

I’m using a DGX Spark, and for the use case designed, the offline processing is required (sensitive information) and is going to be highly cost effective at $4K unit price.

Now, for chat workloads, you’ll take forever to hit that ROI.

Pure code workflows are not yet frontier level capabilities for affordable hardware, but this would also be a huge ROI due to token throughput.

[wooden minnow]

after this thought, I think ive kinda convinced that 5.2 is kinda better than 5.4

on long horizon tasks

5.4 is more of research side maybe

[deft gyro]

5.4 is a more intelligent model

But 5.2 checks its work

[lyric vector]

where should one report codex bugs? i can't write in #1070006915414900886

[wanton shoal]

Hey! To post in the #1070006915414900886 channel, click on the pinned post in that channel and find the "Report a Bug" button under the instructions

#1295975636061655101 message

[cyan wing]

I think GPT 5.4 uses up so much context in searching / reading the codebase,
it might be time to go back to indexing..

anyone try using Deepwiki with 5.4 ? I'm seeing good results so far

[prisma crow]

I've used GitNexus with Claude. I haven't tried it with GPT 5.4 but I would assume that it is possible to do that. Maybe worth looking into.

[hard drum]

how do you refresh a new plugin/framework install on codex?

[tender stream]

Try mcp refresh in chat or just select some with
/name-mcp n say u need update
I hope will help🙏🏻

[hard drum]

not sure if this goes here, but i locally tested a working patch for the capslock issue: https://github.com/openai/codex/issues/16189

which needs a little more testing from users outside macos tahoe 26.4

funny... use codex to fix codex-cli

[tender stream]

💆🏻

[tender stream]

https://github.com/openai/codex/issues/16189#issuecomment-4151035563

[hard drum]

they report this on windows thru WSL2, but i'm natively on the latest macOS version, so that's different to an extent

26.4 (25E246)`

[tender stream]

Hmmm, do u have npm? I think u can use update
Just second

[tender stream]

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

[hard drum]

I did the manual binary replace for 0.116, and it did not do any difference

so the local patch on my fork seemed to have resolved the problem

[tender stream]

Hm i use 0.117 idk
Maybe u should change shortcut on keyboard?

[hard drum]

Nope! This isn't an issue when i type here

it is strictly an issue in codex's textfield

an odd situation of "it doesn't work on my machine"

[tender stream]

I see okie

I hope everything works for u rn

[hard drum]

i will installed the patched version, and will see!

i am of course testing it (and other stuff) live, so i have objective evidence

[tender stream]

Try explain to support, i think this will help for next patch

[hard drum]

Not sure how else to explain it other than, "macOS tahoe 26.4, vscode 0.113.0 arm64, codex 0.117.0 (and then 0.116.0) unable to caps-lock when typing into ThE TeXtFiElD wHiCh iCan Do HeRe)"

it's like i cannot capitalise and scream

[tender stream]

Ahhahaha

[tender stream]

U cant thats cli)
In codex app u can lmao

[hard drum]

lol

hopefully this issue-patch can be further looked into, and further tested against other machines

i even got an intel mac i put to charge so i could test it there, too!

and i also got windows on it, so i can WSL2 it, as well, and native windows... maybe

[tender stream]

I use just power shell, works amazing

[deft gyro]

smh

windows

[hard drum]

i use zsh with oh-my-zsh on macos natively

[deft gyro]

🤮

[tender stream]

🙁

[deft gyro]

why use windows

it's the worst operating system

more bloat and ai slop please

[tender stream]

My main operating system since childhood

[deft gyro]

this is not a reason

[tender stream]

But i luv cross platform then i use rust 🙂

🦀🦀🦀

[hard drum]

windows for dev is atrocious

hence why i moved to macos like... 3 months ago

never been happier

[tender stream]

You just get used to it

[hard drum]

i could never in my lifetime

[deft gyro]

no

[hard drum]

that is a terrible way to think

[deft gyro]

there is no getting used to it

it's objectively terrible

use linux

[hard drum]

it's like saying, "oh, you'll just get used to physical abuse, just deal with it, no problem"

[deft gyro]

or macos

[hard drum]

terrible argument

[deft gyro]

macos is like

bad but not so bad

very user friendly

not too bloated

very pretty by default

[hard drum]

still better than linux and the lack of actual ecosystem and stability over there

also Liquid Gl(ass) sucks

[tender stream]

I use intel😭😭😭

[deft gyro]

what

wym

better than linux

you cant compare them

like

"linux"

is a collection of realistically like 10 major operating systems

or 3 depending on how you count

[hard drum]

linux and macos are both unix-based, except macos is derived from BSD... to a major extent

[deft gyro]

no

linux is unix like

macos is unix based

[tender stream]

I can, i don't wanna install all drivers and think about "why my os leaving"

[deft gyro]

but this is mostly a semantic difference

what

this is not an argument

[exotic cave]

I wrote a reply to your issue, check if it works for you

[hard drum]

to me, i tried linux, but the only time it actually worked well, was ironically on a Lenovo Yoga 7 laptop

[deft gyro]

installing drivers is only a thing on windows

if you dont have mac hardware

use linux

[hard drum]

everywhere else i've tried, except for say... Steam Deck and whatnot, was just terrible experience overall

do not recommend outside Valve's stuff, or Lenovo devices

[deft gyro]

I have never tried on steam deck

but on my 2019 t2 mac linux worked great

on my desktop it works

on this acer aspire 14

[tender stream]

So u understand 30% humans use windows for all time?

[deft gyro]

its truly wonderful

so what

people using something doesnt make it good

[hard drum]

^^^

[tender stream]

I just don't see the point or benefit in this. In Linux.

I can use msconfig and clean ALL AND USE ONLY TERM but for what?

[deft gyro]

What

Linux isn’t all about the terminal

Like

Linux guis are modern nowadays

[tender stream]

$sudo apt idontwannausethatokay?

So i hope i can say something

[deft gyro]

It’s okay if you want a subpar operating system

That’s your choice I suppose

But I am saying that I grew up developing on windows

[tender stream]

Thank u for understanding 🙏🏻
God bless u

[deft gyro]

And I can never go back

[hard drum]

Same here, but moving to macOS worked for me too well, so I will continue staying there

makes it incredibly comfy

[exotic cave]

@hard drum did you check my comment regarding your issue in github?

[hard drum]

let me see!

[deft gyro]

macos is reasonable

[hard drum]

i did a replacement --release binary with my fork, and i am happy with the result.

[deft gyro]

not optimal imo

but reasonable enough

[rose tinsel]

anybody else getting high cpu usage for no reason when the vscode codex extension is open? Not even running anything

On macos

[small parrot]

hi all, is 5.4 pro not available in Codex app?

[potent mason]

It's not 5.4 pro isn't really meant for programming imagine each request taking 10 hours

[small parrot]

oh i see.. so the only benefit of pro is more usage within codex?

[potent mason]

Well you get access to spark (which is about 1k tkps), you get (basically) unlimited image gen in gpt, unlimited prompts, models like gpt 4.5 and 5.4 pro, you can ask gpt to not train on your data, more usage on codex app

[clever steppe]

Anyone know how to perform a search in the Codex VSCode extension?? CTRL+F invokes a search in the main code window. 🙁

[true bane]

I use codex on windows everyday it's great (wsl). Way better than CC on powershell

[normal spruce]

Why does it feel like my usage limits on the plus plan are so high right now

ive been beating codex up and down all afternoon

only used 20%

usually 50-60% by now

5.4 extra high

[hard drum]

that's why

xhigh

[normal spruce]

ive been using xhigh as long as its been available, i just meant it feels like im getting more out of my plan lately

compared to a few days ago

[hard drum]

5.2 xhigh > 5.4 xhigh

ideally 5.2 high for plan, 5.2 medium/5.4 medium for impl

[deft gyro]

Do yall use implement plan or copy plan into new chat

[cedar skiff]

the way codex writes plans they are meant for the planner to implement, if you use the default plan mode and then use it in a new chat it wont have the context it needs and the plan is ambiguous.
So you need to prompt for a plan that can be run in isolation by an agent that has no prior knowledge.

[tiny kettle]

GPT 5.4: Sure, I'll help you set up red-team E2E to identify and poke at potential exploits in the site you're working on.

Also GPT 5.4: Nope, I can't make a .exe to read from a .txt file and use SendInput to write it to another window. That's not a legitimate use case and I will not help with malicious code.

[fading spear]

I hope so, because it's gonna be like a kick in the face when 2X promo ends in a couple days

[normal spruce]

Ahhhhh

X2 promo

That makes sense

I think I saw something about that on insta

[deft gyro]

I LITERALLY GOT A REFUSAL FOR NO REASON YESTARDAY

im like

okay

hook it into the api

and its like

no

I cannot help with automated [REDACTED]

and im like

you built the rest of the tool?????

[deft gyro]

I know, I am already running out of quota, it will be so bad

xd

avg gpt 5.2 moment

1 hour

still running

[fading spear]

not all threads are built the same

[whole gate]

Super helpful and informative Codex work! lol

[deft gyro]

single proompt

avg gpt 5.2 moment

[whole gate]

How much code generated for that?

[deft gyro]

5412 loc

[whole gate]

That's pretty average, but, hmm... I mean, most of my prompts result in 3-5K each, how were you calculating the use cost?

[fading spear]

should be $1 per 10k loc... if we just get another usage reset everything will be fine guys

[whole gate]

I'm not sure what loc is

[fading spear]

lines of code

[deft gyro]

ccusage on codex ofc

api pricing

per token

lines of code is not a valid metric

[fading spear]

there is no valid metric, but they're trying to normalize it with credits

[deft gyro]

well im just using input tokens * api cost + output tokens * api cost - cache

like

if I were using the api

this is how much it would cost

[wooden minnow]

is codex cli having issues?

not getting any responses

[night ivy]

Is any one else's codex usage burning insanely fast for some reason? Like literally after 3 prompts it's done

[solemn acorn]

free or plus?

[night ivy]

Plus

I usually get like way more usage out of it

[solemn acorn]

I would check the dashboard on the web to see where all your usage is coming from maybe

but I haven’t had that issue personally

[raven kettle]

I'm enjoying adding features most smaller projects don't bother with.

[hard drum]

Quebec French when?

[raven kettle]

🥖 🤔

It's in my next stretch goal.

[hard drum]

🇨🇦 ouiouei

[raven kettle]

If we hit £500,000 I add fr-ca, and a full French Mode where the app takes three times as long to complete a task, then force-quits at 17:00 local time.

[frosty zealot]

Dont forget to have it demand an equalization payment too!

[hard drum]

Don't forget to open a PR for the next marching on the streets of Paris with torches und pitchforks!

[raven kettle]

Also if fr-ca is selected, the app will refuse to accept any English data inputs.

[hard drum]

At that point, put most of the words together too, so that it sounds like you're reading a live transcript of xQc in his natural state

[raven kettle]

I'll add text2speech in xQc's voice.

[hard drum]

albprplrobpaaa, chat? alblglblblbl, blblb chat

[raven kettle]

Maybe for en-us I'll add an Asmongold mode

In-app reminders will be in the shape of a rat

[hard drum]

rat would be PiRAT Roachware

asmon would be a cockroach

you'd want a narcissistic rat for piratesoftware

and don't forget that he will occasionally drain his mana, then ask what he is suppoed to do for you

[fast iron]

Is fast mode actually making the model faster or just wasting tokens?

[fast iron]

mine worked for 50 min produced no code, then had an error because its context was too big to compact :/ my bad

[blissful basin]

I'd like to make an official complain to OpenAi codex team. Ai destroyed my life. I can build so much cool stuff now, but day still have only 24h.

[dark ginkgo]

Does anyone know how to enable PR screenshots for codex cloud environment? im running into this issue I couldn’t take a screenshot because this runtime doesn’t expose the required browser_container screenshot tool.

[main nimbus]

I was up until 2AM context switching life a mf.

[nocturne folio]

autoresearch with codex doesnt last long

[lucid mason]

Does anyone know how to make links work in Codex' chat answers? They are resolved to URL links that open the browser instead of opening files inside VS Code. I tried to change file_opener from its default to specifically "VSCODE", this removes "open-in-target" errors but still doesn't resolve the links properly.

[celest stag]

I have found they just don't work half of the time. There are a bunch of related issues raised on Github for it. It is annoying.

Seems to be a Windows issue, as the Codex team is apparently 'Mac first'

[lucid mason]

For the time being I will set it to "none", because there are strong indications that this is responsible for Codex pushing Code to high CPU load even while "idle" and generally sluggish behavior. I will have to watch that one for a bit longer, though.

But even "VSCODE" might fix these compared to the default.

[pearl kestrel]

Has anyone found something like an OnCompaction hook for Codex IDE in VS Code? I would like the agent to selectively ingest some *uncompacted) context after compaction.

[main nimbus]

Have you ever seen a piece of content from them without a MacBook in it? I’m sure almost the entire org uses mac

[celest stag]

Yeah, it is frustrating since I do a lot of windows based work- and after a quick google, the majority of developers in the real world do... but not them, apparently. I have been raging at the 'apply_patch' script codex uses to apply bulk file changes simultaneously constantly hitting the windows command length limits for the last month. It loves to try to use apply_patch, fail, then fall back to manually editing the files one by one... taking like 5 minutes and who knows how much tokens/context.

Yeah I know, windows command length limits is the issue, but there has to be a workaround they can do for that.

[lucid mason]

Every time I type the word "linear" into the chat a popup suggest to install the "linear plugin". 🤔

[blissful basin]

run it in WSL if you are using windows

I noticed same behaviour when was running it normally on windows, but sadly it is actually unusable without WSL

It takes few mins to setup if never did that before, but you wont regret

[unborn thunder]

Is codex on mobile less powerful than just using codex on windows? Also is there a difference between using GitHub plugin in chatgpt compared to using codex in chatgpt mobile?

I’m guessing it doesnt have cmd commands that’s all? It just makes more mistakes on mobile, when codex in mobile and windows are gpt 5.4?

Just don’t understand why chatgpt codex makes more mistakes

[lucid mason]

That has the drawback of having to access native NTFS C: drive files from within the WSL environment. It works, but it's at least a bit less convenient (and theoretically slower). Personally I don't need working links atm, so I can just disable that and also benefit from it (seemingly) being a workaround for CPU idle problems (not entirely sure yet).

What irks me more is that the 1M context window is somewhat useless. I am currently below 700K and Codex/GPT starts misbehaving again. Not concerning history/context, but not being able to properly handle even the very last chat input (misreading input, stating wrong/older numbers, ignoring input partly or entirely).

I saw that happen multiple times both with Codex in VS Code and the GPT web-chat already. Time to start a new chat, but what do I have a large context window for then?!

[whole gate]

This has been bugging me. So, I just counted, Codex has generated 19,404 lines of code (deleted lines and inserted lines total) since last Thursday reset. I have 90% of my token usage available still. Are people literally generating 200,000+ lines of code every week?

[raven kettle]

Did you count by hand?

[whole gate]

No, it's in the UI

I mean, I kinda did, entered the numbers into Excel

[raven kettle]

Also, it's not just the writing of the code itself. It's the figuring out what to write, writing it, writing tests for it, running those tests, fixing issues, etc.

[whole gate]

Just added it all up

Oh, ok, I don't use Codex for planning so that might be the difference.

[raven kettle]

I'm still using it for hobbyist stuff.

[whole gate]

Ok, well that kinda solves the mystery at least, just been bugging me, like, how 🙂

[boreal badge]

bruh

[raven kettle]

I'm still mad at WinUI 3.0 for not having a XAML designer live view where I can do WYSIWYG design.

So dumb. Not everything needs Figma or equivalent.

[boreal badge]

upgrade to pro

[raven kettle]

Could also be your project is simple in terms of tokens required.

I can't justify £200/mo for hobbyist stuff.

[boreal badge]

click, click, click

make openai more happyest

[raven kettle]

Also I think OpenAI ended the 2x token week thing?

[lucid mason]

I hit the weekly limits after 3-4 days. Most of this is analysis and policy tuning, not the few lines here and there.

I switch between Codex and the web-app in order to save on Codex tokens.

It can happen that Codex/GPT has to analyze hundreds of values in order to change a single number.

Actually that's what's happening all the time here. 😄

This is the kind of log file GPT has to analyze on top of measured data to correlate against (older log, complexity grew since then).

[raven kettle]

Which plan, and what are your workloads like?

I'm only using this stuff for hobbyist projects as I'm not a professional dev anymore.

[lucid mason]

Plus. I don't know what you mean by workloads. Most of the tokens are used for log and benchmark analysis and then policy tuning and then back around. There are only a few (dozen) lines of code changed at a time, sometimes only a few variables values.

For most of this "Extra High" is needed and even then struggles with the complexity. That's ok. I am still on a free month trial plan, so I absolutely get my money's worth. 😉

[raven kettle]

Are you using it to build a waifu companion bot?

Long answer = no, but yes

[lucid mason]

And I also improves another app/script on top of the complex one using the same plan where GPT struggled a bit with the specifics, so that cost tokens as well. Sometimes tokens get wasted by GPT not properly following instructions, though, which then leads to a back and forth that isn't always handled by more specific prompts.

[raven kettle]

I knew it.

[lucid mason]

I am building an app that turns screenshot 1 into screenshot 2.

[boreal holly]

I think if you used a real database like PosgreSQL to store the logs the agents wouldn't need to burn tens of millions of tokens doing data analysis on unstructured log files

[lucid mason]

They are more structured now, because policy and general debug logging is split now.

And reading the logs doesn't take the bulk of thinking. Putting it into correlation with measured results in various different workload situations and then coming up with policy tuning changes keeps GPT "thinking". 😉

[boreal holly]

Right, but with a real database they can do stuff like SELECT * from logs WHERE date < ... do really complicated queries that can produce real insights without manually looking at the logs or describing how to parse them, and create sql scripts that automate the scrubbing process

[lucid mason]

With policies for the different workload scenarios at least partially affecting each other. That's complexity enough for burning tokens.

[boreal holly]

Just a suggestion. Might make it to 7 days if you optimize that one area

[whole gate]

I spend a lot of tokens to make sure this is true, varable names are the same in the database, sql provider layer, application logic layer, rpc security layer, and frontend bindings and pages. All use deterministic models and data keyed to deterministic values. End to end pattern consistency makes LLMs sing.

[nocturne folio]

autoresearch is so good its like crack

[whole gate]

I actually generate a lot of the application itself directly from data, rather than code. The more of the build you can perform mechanically, rather than generatively, the better.

A lot of my generative code is about writing mechanical code generation so the LLM doesn't have to keep track of a ton of stuff.

I almost always tell it to write scripts to fix things, rather than trying to fix them.

[lucid mason]

GPT/Codex designed the logs itself. There was no need to describe the parsing to it yet, rather the other way around if I was interested in looking manually. 😉

Would using a real database for the debug logs help GPT performance and saving tokens? Could this be implemented natively in AHK 1.1 (plus DLL APIs present on all Windows systems)?
"A real database would only help GPT/token usage if you also change the analysis workflow so the model sees small queried subsets or pre-aggregated summaries instead of raw log streams. The token cost is dominated by what gets read into context, not by whether the bytes live in .log files or a .db file. With the current package, the bigger win is already the structured benchmark store in runs.csv and metrics.csv, plus targeted text-log review when a run needs it.

For runtime overhead, a database is also not an obvious win here. The current logger already keeps persistent append handles, batches normal flushes, and only forces immediate flushes for lifecycle checkpoints in PowerPacer.Settings.ahk:69 and PowerPacer.Platform.ahk:3484. The main cost of moving to a DB would be complexity: schema design, transaction/locking behavior, recovery, corruption handling, archival/export, and much worse human inspectability than the current split text logs in PowerPacer.Platform.ahk:3556.

On “can this be done natively in AHK 1.1 with DLLs present on all Windows systems”: technically maybe, practically I would say no for this package’s goals."

[boreal holly]

They are so much better when you get rid of the "must use apply_patch" directive! I like seeing their crazy python-fu work magic on the codebase

[whole gate]

I added an MCP server in front of my database server so the agents can read the database schema directly from INFORMATION_SCHEMA and also dump table contents where a lot of the application details.

I still need to figure out how to wire up Codex to use it, right now Claude.ai uses it for planning, so not a huge deal.

[boreal holly]

Oh gotcha, csv is good enough! They can at least run it through pandas for analysis. cool!

[lucid mason]

It regularly reads snippets only like starting with the tail first, before diving in deeper. And I use fresh logs for every benchmark/test run atm.

Main log + policy log + 1x screenshot showing Rivatuner+HWinfo overlay data:

[coarse escarp]

#afghanistan🇦🇫 #دایره_نوازی #زمزمه_های_دلتنگی #محلی #sofi
https://www.instagram.com/reel/DVB1ECXiLrj/?igsh=bWNsdXR4cWRlemJ3

[hollow stirrup]

Am I crazy? For a while, maybe a few weeks ago, the codexcli chats were reproduced in the Codex app. I'd see new ones pop up after each App restert. Then at some point that stopped happening is that something we can turn back on?

[icy mason]

ugh codex is crashing after several seconds. Any fixes for this?

[lean lark]

I was reading about this recently. Common popular offerings like OpenClaw are using non-vectored .md files for basic memory - this technique is trending, with projects independently coming to the same conclusions.
This doesn't work for semantic searches. But now with agentic processing where we're working with smaller context across more threads, the pendulum swings back a bit from "gotta be in a DB" to a more hybrid "put some memory/data in files and then when you need semantics or heuristics, or simply massive amounts of data ... go for a DB.

[lucid mason]

I go with what GPT came up with itself:

[bright geode]

Hey all, what's the best way to mimic something like stop hooks for claude code but for codex? anyone got some good setups? is the agent SDK the only way?

[icy mason]

so no one else is seeing the issue where codex closes several seconds after it's started? I've repeated this many times now.

[lean lark]

Codex supports Start and Stop hooks ... those are the only two so far. So, it seems like you're in luck this time.

[lean lark]

You'll need to provide a lot more info:
Which "Codex" are you talking about? VSCode extension? App? Cloud/Browser? CLI?
What do you mean by "closes"? Does it just die? Does it say it completed a task that was not completed?
C'mon dude, act like a developer and provide some diagnostic info.

[icy mason]

I've submitted a bug report with these details but here you go: Win11 Codex app. It closes several seconds after start. No errors, just closes. Replicated 5 times. Found the logs, didn't find anything (obvious anyways) why it closed. Btw, before it closes I am able to type or do other things in codex so the app is "running". It closes independent if I type in it, or do nothing.

[worthy matrix]

has someone been facing gpt-5.4 being too slow?

like from the requests side

[mint locust]

I run codex CLI on Ubuntu 22.04 , and it change my terminal font to chaos, not ASCII characters, how ever, I fix it, and seem like leak of content, it read a some Italian paper.

however, I am not use newest version, maybe openai already fix, they have lots of PR

[mint locust]

I am sure 5.4 is slower than 10 or 15 days ago, I use token counter when running, seem like too much people use, no big issue, openai never promise rpm or etc. , when about v0.106 you run speed mode when then just announce 5.4, it was 200 token per second .

for now, maybe 10 or 20 when you don't use speed mode.

[icy mason]

so I added the info @lean lark. Any response or nah?

[worthy matrix]

It just replies me after 30 seconds

It used to be super fast though

[mint locust]

how to define replies?

it shows reconnect?

[worthy matrix]

exactly 32 seconds lol

though all the other gpt models just work fine

other than 5.4

[lean lark]

I can't help with that specific issue but at least now someone who might be able to help has some of the info required to do so.
I would go to the GitHub repo and search for focused keywords - see if someone else has reported this.
Also, do a reboot, just in case, cuz ... Windows.
Good luck. 🤞

[mint locust]

Windows is funny

I mean, reboot is best

[potent mason]

How do you get it to output tokens per second?

I'm guessing its a CLI only thing right?

[mint locust]

you can monitor codex session, I am not sure how it work after compile.

[lean lark]

There are always things that these models just don't know. I wish there were a way to expedite such information to OpenAI so that they can add the info into subsequent training. For example:

• I just created a Java (Android) project in Codex Web. For building and testing it adds gradle-wrapper.jar. But a PR from Codex cannot include binaries (.jar=.zip). We've known this for a Long time. So why doesn't the Codex model (gpt-5.4) know this yet?
• The GitHub 'build-and-test' operation fails when that JAR isn't available. Why didn't the model recognize this when I told it to remove the JAR from the PR?
• OK, so I explicitly tell it to delete the file from the repo but add it into the setup.
• It's also unaware that Node v20 is deprecated for GitHub testing. Why didn't it know this? This is a tool that integrates with GitHub. So I added that into the build-and-test as well, to require Node v24.
• If the model generates images into the Codex VPS, it doesn't recognize that those won't survive the PR. So I had to write a utility a long time ago that was and is still required to base64-encode for the PR, and then rehydrate images from the repo.

The language isn't important. It just doesn't know things, and we don't have a good way to get these tidbits to OpenAI for incorporation into their system instructions for the Codex pipeline. Every developer needs to find these things out and handle them individually in the hope that some day we might be able to remove the redundant world knowledge. That's just sloppy.

[mint locust]

I think because it model don't know what is to know.

it can not think, it predict next token pretty well, so, more likely mean or average level is best.

[lean lark]

But as a product that is most commonly used for integration with GitHub, it should be aware of GitHub requirements.
Here's another example: When that CI process fails it generates huge logs (the lint log captures some of the above stuff). How do we communicate those failures back to Codex? It seems we need to write our own code to parse the logs for errors so that we can provide Codex with carefully defined problems to resolve. Why is this "a thing" that we need to do? I'm guessing most people would just copy/paste entire logs and tell Codex to figure it out and fix the issues. That's a big sloppy waste of resources.

[cobalt junco]

codex can read ci/cd logs with gh auth

[lean lark]

It looks like I'm confusing model training and system instructions. To be clear, yes the model only knows what's available at training time. But as a product, subsequent fine-tuning and system instructions need to supplement the information available to the model so that it can operate within the defined environment.

[mint locust]

maybe you can open an issue

it's useful

[lean lark]

That requires a lot of wiring.

[lean lark]

Yes, the way we do things in this industry is that we create a ticket, it gets considered, maybe processed. What I'm lamenting here is that this has been the state-of-the-art with Codex since it started, and it seems obvious enough that it would have been addressed early-on and not a condition that persists for years.
It's like using "gh auth" as @cobalt junco suggested - that's a good thought but it requires a lot of brittle configuration. We shouldn't need to play games like this ... we got udder stuffs ta do! 🐮

[cobalt junco]

not a lot nor brittle imo

[lean lark]

I will look at this more closely. Thanks. I tend to think of gh auth as being a sandbox hassle but I will try to find some time to re-evaluate my position on this - it's only fair...

[high crow]

Is claude better than codex for coding etc?

[exotic cave]

Currently nope. Claude is behind for quite a while now.

[boreal holly]

I mean hey, that's the price of using someone else's computer cloud to do programming! I think the Base64 utility you made is pretty genius tho

[lean lark]

I'd like to find some solid testing/evidence to support that. The world blog-o-sphere seems to be much more focused on Claude than on Codex. There's a great deal of excitement around it. At the moment I'm avoiding too much "xyz is better" fanfare because there's rarely a fair (if any) comparison. People are just talking about what they know and saying it's the best. Cuz humans do that.

[lean lark]

I have a pet project that I've been wanting to with Android for a long time but I lack any time for it. So I decided to see how competent Codex can be with ( cough ) vibing. For this one I'm happy to just tell the bot what I want and see how well it can do. Unfortunately I haven't even gotten to the application because Codex is really incompetent with configuring the CI.

[steel gale]

Is it worth spending the $40 for codex credits if I’m close to done working on a project but am near my weekly quota? I don’t want to burn though $40 in 20 minutes for instance.

And ideally I keep using it on Gpt-5.4

[glad wind]

Subagents have been complete broken for me for 2-3 days now, am I the only person who this is happening to? This is the output of a subagent tasked with exploring part of my codebase

[cedar skiff]

If this is in a work flow i would start questioning the orchestrator and find out exactly what it has been saying to the sub agents

[glad wind]

It's not that

Somehow subagents no matter what they're told to do get confused about their identity

[unborn thunder]

it seems like chatgpt app can continue doing coding in github after codex credits run out

[sacred minnow]

In todays age sounds about right 😆

[sand shuttle]

yo codex icon has dissapeared form my side bar and the bar above tabs how do i enbale it

[royal mirage]

anyone having issue with codex?

■ stream disconnected before completion: stream closed before response.completed

[deft gyro]

the enpoopification has begun

[potent mason]

Oh no I'm down to 20% usage until April 2nd

Come on great Codex people pull a reset please

[sand shuttle]

just buy the 100$ plan haha

[cedar skiff]

there is a 100$ plan?

[sand shuttle]

oh mb onyl 200$ haha

[solemn acorn]

I think the only thing that's actually changed with the limits is openai stopped resetting them early lol

they seem to be pretty managable especially with 5.4-mini

that model is great

[potent mason]

I am on the $200 plan

[solemn acorn]

jesus

[cedar skiff]

Not sure what im going to do when the double usage is over

[potent mason]

Yeah me neither I've hit more than 50% pretty much every week

[boreal badge]

GPT 5.4 xhigh failed to reverse engineer, but it implemented ALL of PL... It implemented almost every functionality, but couldn't put finishing touches on it. It seems OpenAI slowed it down in this area, but Pro version showed brilliant results, and Claude Max's efforts were also successful

[simple star]

Well, 2x limits are gone on Friday, so I think im going on vacation

[mellow echo]

So I dug into this and found a workaround. The subagents are basically forked conversations, meaning they have all the context from the parent. The workaround I found is adding something like this in the description:

CRITICAL:
When spawning this agent, use fork_context=false by default

[stable obsidian]

Why did my codex add this to my AGENTS.md? I only tasked it to build a simple frontend, this was the initial AGENTS.md: Make use of skills wherever feasible. For example, when working on frontend, use the frontend-skill skill. You HAVE to make use of them, for ANY task where it could be useful.
It also created a CLAUDE.md 😂 5.4 xhigh btw

[cedar skiff]

it's probably copying what you had in another directory under instructions, that's a common thing to have in cluade.md because claude forgets to use skills and follow rules a lot.

[calm vault]

It would be pretty OP If OpenAI got With Supabase And Brought Supabase Integration To codex So Codex Can Single Handedly develop Full Stack Apps With ChatGpt As The AI, Supabase Backend And Codex As Te Dev

[deft sable]

Unrelated to Codex, but has anyone seen that the CC source code was reportedly leaked? lol (hidden/upcoming features also leaked)

[warm pilot]

https://supabase.com/docs/guides/getting-started/mcp – or better: https://supabase.com/docs/guides/getting-started/ai-skills

[plucky halo]

MCP and CLI work well

[calm vault]

i mean yeah there are mcp servers but they acan be tricky to link up that wasnt what iw as talkign about, imt alkign how lets say Lovable Has it

[plucky halo]

I wasn't aware of the skills - cheers for that

[warm pilot]

just found it when searching for the URL 😉

[royal mirage]

codex is cooked

[warm pilot]

Lovable uses Supabase as it's own backend, that's (hopefully) not something that will be happening with codex. The MCP has the same setup curve as Lovable with a own Supabase account: you create an account, let the MCP auth to that account and then you're set.

[royal mirage]

did anyone recently used the claude code codex plugin??

[calm vault]

no u can connect ur own too

[warm pilot]

it's really helpful if you read all and not only the first sentence 🙂

[plucky halo]

First 3 words.
All I need.
All rest unnecessary

[tight geode]

https://github.com/kill136/claudecode
The source code from the official website. The leaked version can be run by executing node dist/cli.js –version.

[oblique cove]

Hell yeah, that’ll be interesting to inspect

[royal garnet]

why do they not let you fork from previous message in the codex app

[main nimbus]

I've had extremely long runs, but the performance at these ranges sucks.

[wicked briar]

why

[deft gyro]

I ran out of my weekly quota in 3 days

[unborn thunder]

2 days for me

[deft gyro]

Plus user so not too bad

I understand

[unborn thunder]

I got Claude plus now

It’s helping

[deft gyro]

I have gemini pro

And openai plus

Might get anthropic 100 dollar sub

[unborn thunder]

I just rotate and thinking of getting Gemini now too if I need the extra credits

[deft gyro]

If openai has a good deal for 100 dollar sub ill do that too

[unborn thunder]

Codex 2x usage is going away on Friday apparently

[deft gyro]

Yes

April 2nd

[royal garnet]

is there any way to control cursor agents from your phone, like claude code lets you

[little magnet]

my longest was 4h but then app was ready 😉

[woeful isle]

I’m working on a small open-source project (very early stage) it’s a CLI tool that uses AI personas to test apps (basically “break your app before users do”)

It's using codex!

If any want to participate or try let me know
https://getwired.dev/

[toxic torrent]

so slow today. idk whats up

[tiny wind]

I like it so much that they recently released free access for gpt-5.4 in the codex cli and app for free, last time I checked it gave an error.

[cedar skiff]

I dont think they did that

oh cc source got leaked o.0

i wonder if that was an accident

[orchid plume]

probably was, after all, they admitted not too long ago that Claude pretty much writes itself, so I expect it made some mistake in that process 😄

[boreal holly]

I've had 4 QA agents, 1 orchestrator, and randomly spawned and archived workers chugging away for 72 hours straight with very minimal issues or input from me

[cedar skiff]

I'm sure the office is filled with devs mumbling into microphones instead of clacking at keyboards. But claude having unfiltered opinions and making unchecked decisions about code is pretty unlikely. Its more likely they use it as a work horse so they dont have to clack keys.