codex-discussions

[boreal holly]

Gotcha, Switch works just as well!

[exotic cave]

How to set agents model and reasoning level?

In the CLI.

[desert junco]

/models

[exotic cave]

Does this control subagents?

[desert junco]

I don't know if I understood but, using /experimental you can activate the multi-agent and then it will enable codex create sub-agents to manage the task

[exotic cave]

I explained it wrongly. How can i be sure that the subagents spawn with the main agent reasoning and model?

[steady vigil]

something new in 106? Didnt notice this before

[desert junco]

if you don't have a config_file for the specific agents it will get the config from the main agent

[exotic cave]

Thanks that's what i wanted.

[nocturne folio]

gpt4.5 is now a year old

[sleek spoke]

🤯

[simple star]

106 is out??

[kind jay]

Why don't they use codex to port codex app to windows?

[simple star]

They are

But they probably run out of tokens

[orchid plume]

some people have literally done this, well, on a more basic level. They got Codex to write a script which converts the .dmg image to a compatible Windows version

from what I understand though, I've heard it's something to do with them not being happy with the sandbox for Windows

[dense mason]

is it just my codex that got really slow after the latest update?

[orchid plume]

I did get the Codex app on the Microsoft Store to work in the end (no longer show the "update required" notice), but either it's an old version or the Windows version of the Codex app is missing stuff. In the UI there was no automations section. I could not find a way to access the Settings UI either (like change the theme from light to dark etc). Converting the .dmg image is still superior

[sleek spoke]

how did you get it to stop showing update required?

[orchid plume]

well, codex did, I'll get it to explain what it did, one moment

[orchid plume]

this is what it said it did:

› Ok it's working now, what was the final fix

• Final fix was a 2-part workaround:

  1. I blocked ab.chatgpt.com in C:\Windows\System32\drivers\etc\hosts so the app couldn’t fetch the remote experiment assignment that kept forcing the sunset gate.
  2. I patched Codex’s local LevelDB Statsig cache to set gate 2929582856 (the “Update required” gate) to false.

  Before the host block, the app kept rewriting the cache back to true on launch; after blocking ab.chatgpt.com, the local override stuck and the app opened normally.

I'm going to guess "ab" is related to something to do with A/B testing, and it could be that for accounts which were invited they get the expected value from ab which then doesn't say "update required"

[sleek spoke]

LOL, brilliant

Codex hacking itself open

[orchid plume]

yeah, it was interesting, not sure if I'd go as far as to say "hacking" but at least it allowed me to briefly check out what I might be missing

if this is the current state of the actual Windows app then I'm not missing a lot

[sleek spoke]

interesting, thanks

[steady vigil]

yeah its very painfully slow

[brittle stratus]

anyone wishing for undo/rewind in codex cli? ⏪ ⏪ ⏪

upvote here so we can get it sooner: https://github.com/openai/codex/issues/11626

[torpid trout]

No please
Either use git or nothing

Undo in app is so catastrophically bad I don’t want that same garbage in CLI

[stray swift]

People outside OpenAI have done that from the Mac .dmg file. But it's badly sandboxed. Someone in this channel reported that a Codex produced that way rendered their personal machine unbootable, by screwing with registry settings. So best to wait until it's professionally sandboxed, unless you really know what you're doing.

[brittle stratus]

there are many ways this can be done,
and it doesn't have to be garbage if done right

Also, I am not using the app, can you tell me why is it so bad? Is it actively hurting your efforts even without using it?

[torpid trout]

It’s just extremely buggy in the to back feature - I had used it onxe and it created an irreparable total mess of past and present code.

[boreal holly]

I agree. I think it should be a feature that is opt-in, and whatever diff tracking they currently have should be under that flag.

[boreal holly]

Basically the way they're doing it is apply_patch is using git patch under the hood, but there's no tracking for when that patch "diverges" from the existing code reality.

[brittle stratus]

I have nothing against this and I think this can be utilized, but given how helpful this would be I think it should be done soon

[torpid trout]

Git is already helpful.

I don’t feel an AI should manage git or do what it does. Not right now at least

Keep fancy „click behavior“ in app app or use gitkraken

GK for example does an amazing job at visual git

[brittle stratus]

i think Antigravity suffers from similar issues

But it's still quite helpful

The problem is with multiple edits in a timeline. I think we can easily replicate what Claude Code is doing, which is different than simple Git patches

[hard tulip]

so did anyone figure this out

the link goes no where

[cedar skiff]

does reinstall work?

[boreal holly]

I already posted a comment on the issue about a viable alternative. Basically in order for undo to work properly, it has to apply the negative of every single patch ahead of the one you're rewinding to and reapply the forward patches, assuming the agent didn't use stuff like sed to edit a file where a patch would be applied. If it's undoing all patches up to a certain point there's no forward pass. But that also falls apart if you move a thread into a different worktree or something. It's a really brittle way of doing it. But if you treat every single code change as a commit with a stable identifier you can rebase anything, including sed edits by the agent, and eventually squash the changes if you don't want a noisy branch. I just think Codex rolling its own separate version control on top of git is a little funny

[hard tulip]

nope

[boreal holly]

maybe npm i -g @openai/codex@latest, go to where the codex binary is installed via npm, copy it into %APPDATA%\local-roaming(or whatever windows does)\Codex\codex

[orchid plume]

#codex-discussions message

Scroll up the chat slightly, I posted how I got around mine

[lapis solstice]

For me its shows up in Task Manager, but the gui won’t open.

[orchid plume]

strange, not had that situation

[lapis solstice]

let me try to reinstall

[orchid plume]

this is what you should see if you manage to bypass the update required notice on the Codex app in the Microsoft Store

[hard tulip]

ok thanks let me try and do that

[mild pendant]

Just shipped a voice model that actually uses conversation context.

Same text, different emotion, because it reads the full dialogue history.
520M, runs fully on-device (RTX + Apple Silicon). Demo + writeup:
https://x.com/LuozhuZhang/status/2027391307338170676?s=20

work with Codex

[high girder]

are you guys using an official release of codex for windows or a port?

[hard tulip]

how can i use it or is it not out

[orchid plume]

this is from the Microsoft Store, it's an app there. However by default it looks like it's either deliberately locked out so people can't use it (update required notice) or possibly is genuinely not up to date. I have a feeling it's the former and only those invited can actually use it. However this update required notice can be bypassed

[hard tulip]

I did both of thoes things and its still not working at all.
and the Codex’s LevelDB cache and the gate is already set to false for that so i am not sure what happend but i check the netwrok thing for the ab thing and its failing so the block is working for that

[orchid plume]

strange, that's what Codex claims it did

I'll ask again for more detailed info

[high girder]

Oh, I'll just wait a bit more then. My port works still, so I'm not worried

[orchid plume]

yeah honestly I'd suggest the .dmg port option right now because it has more functionality

[orchid plume]

• Here’s a slightly more technical version:

  Technical workaround that fixed it for me:

  - Root cause: Codex desktop checks a Statsig gate (`2929582856`) in webview bundle (`index-*.js` via `useGateValue("2929582856")`). If true, it renders
  the “Update required” screen (`appSunset`) and link to `https://go/codex-desktop`.

  - Local cache path:
    %LOCALAPPDATA%\Packages\OpenAI.Codex_2p2nqsd0c76g0\LocalCache\Roaming\Codex\Local Storage\leveldb

  - Keys involved:
    - `_app://-\x00\x01statsig.cached.evaluations.2173843016`
    - `_app://-\x00\x01statsig.cached.evaluations.3625394933`
    - `_app://-\x00\x01statsig.last_modified_time.evaluations`

  - I parsed the cached `outer JSON` (`source`,`data`) and inner `data JSON`, then forced:
    `feature_gates["2929582856"].value = false`
    (also set `rule_id = "local_override"`), wrote both keys back.

  - Problem: app kept overwriting cache from network on launch (`source` flipped back to `Network`, gate true).

  - Final step that made it stick:
    add `0.0.0.0 ab.chatgpt.com` to hosts + `ipconfig /flushdns`,
    then patch the two cached eval keys again and relaunch.

I hope that helps and works

[jade crater]

After reinstall ...

[jade crater]

I'm using the same e-mail at Microsft Store that Open AI send the invitation.

[lean lark]

In other news, I'm stuffing my face with Fritos while Codex is writing new code to my extensive specs. This feels so wrong. I mean, I'm gonna review the code and do my part n all, but ... I'm feeling like such a slacker at this moment. Just sayin...

[quaint hazel]

Apparently there is a 120s timeout on MCP calls, any way to make those infinite?

For context I created a swarming skill that loads Codex MCP into Codex for infinite agent spawning

[boreal holly]

tool_timeout_sec = 9223372036854775807

[quaint hazel]

Legend

@boreal holly where to add that? Getting an error when adding it to the top-level

[boreal holly]

you gotta put it under each mcp_server config

[quaint hazel]

Aaah niiice!

[lavish geode]

I noticed usage limits got abnormally consumed, 75% of my weekly limit consumed in one day while working on just one project

[boreal holly]

Do you have subagents enabled?

[lavish geode]

No, just commit scan automation works for about 10 mins each day

[boreal holly]

automation as in the Codex app or custom?

[lavish geode]

Automation to scan daily commits for bugs
Just modofied it abit to check for ripple effects for what it's supposing to change or fix

Btw i added instructions in agents.md to check for ripple effects each time the model is doing big change, but I don't think this is why all these tokens got consumed in one day 🤔

[high girder]

If you're scanning a lot of dense context files, then the continual context window filling up and compacting could result in a ton of usage, it really depends on what you're working with.

[boreal holly]

Helps to know:

1. Platform
2. Codex flavor (GUI, TUI, Exec, Cloud, etc.)
3. Automation type (GUI official automation, custom script, cron, fswatcher, exec, git hook, precise and exact nature of the automation runtime)

Otherwise we can only speculate

[steady vigil]

codex app is such a let down. It started great now its suffering from terrible performance problems.

[abstract beacon]

u can bypass it?

[lavish geode]

I use Codex gui app on mac and run local not cloud projects
Automation was official one existent in automationa tab i used it and just added a line to check for ripple effects, but the automation isn't the problem it runs 10 mins per day and doesn't consume alot

What i notice is my day long usage consumed abnormal tokens
Idk whether the ripple effect customization in agent.md caused this big consumption or not, i noticed alot of context compaction today yeah but this wasn't what i used to

[boreal holly]

Good choice! Do you have features.multi_agent = true in your config? Also model_reasoning_effort = xhigh?

[latent tiger]

not strictly codex related but this could become a problem for OAI as well for refusing certain guardrail removal

https://truthsocial.com/@realDonaldTrump/116144552969293195
its real not a meme btw

[lavish geode]

Reasoning effort is high
No features of multiagent

[lean lark]

Sigh - the trajectory seems to point toward xAI where the guardrails are minimal, the relationships are already established, and unfortunately We know that the quality is sub-par. It's a sad day for reality.

[latent tiger]

Ironic the US department of defense is going to rely on literally the worst model grok going forward 😅

I suppose in the next 6-8 months it will probably catch up to competitor frontier models

[exotic cave]

Worst but with enough data and the US gov behind it, it might go up a notch or two...

[latent tiger]

Yeah for sure

DeepSeek v4 is dropping any day now too the US competition is going to heat up like crazy with this fed agency ban

[exotic cave]

Anthropic forgot it's the capitalism rules, you can't supply there will always be someone who will... Though US gov dodged a bullet with Anthropic, their models are crap at following instructions, look at IFBench Cluade is a joke.

[latent tiger]

I hope the OAI team leaps on this opportunity to fill Anthropics position lol

Speedrun gpt 6!

[quaint hazel]

Anyone know when the app gets an update?

[spiral gorge]

this does not look like mac

[spiral gorge]

/rewind is somuch easier than git lol.

[spiral gorge]

or claude just lying to customers at vendor bench lmao

[earnest rock]

Hi all, I used the Codex app for the first time last week due to the 2x rate limit increase. When I began it felt like I had a lot of usage available, I ran many prompts using 5.3-Codex on high reasoning. It seemed like every request only consumed .5-1% of my usage.

After my limits reset it seems like I consume way more of my weekly usage, like 2-3% after a request even when setting to low reasoning.

Is there any guidance for optimizing the model/reasoning to make sure I can get the most out of it? Does using 5.1-Codex-Mini at medium reasoning provide more usage than 5.3-Codex at low reasoning?

[final ruin]

Hello, has anyone used Codex for their Obsidian Vault? I am not sure how I feel about giving it complete access. There are personal things in there.
I have disabled the toggle for using my data for training, but afaik openai still keeps data on their servers for up to 30 days.
Does anyone have any opinions about that? Does it really matter?
Please ping me to see your responses

[cedar skiff]

The sub-agents get their own names o.0

[sleek spoke]

they're so cuuuuute

[potent mason]

I get why Anthropic doesn’t want to do that and honestly support them. He’s being ridiculous about this, but in all fairness there does need to be an AI that can help with military stuff (don’t agree how he’s handling things, but I see the point for AI in millitary)

[potent mason]

I wouldn’t worry too much about it as far as I know they have only been hacked once and it was not for the chat histories

[potent mason]

Don’t use 5.1 way lower quality and same pricing (mini is lower price but you’ll suffer using it so much)

On the other hand you can do strategies like vectorizing your codebase or replacing AGENTS.md with information so it doesn’t have to search huge codebases

[final ruin]

Is it enough by disabling this? Or am I missing something else

[ivory zodiac]

just use medium for most tasks. 1 prompt != 1 prompt. your usage is token based

1 prompt can use few tokens, another can use many

the best way to "optimize" token usage is to provide clear, high quality prompts with minimal ambiguity.

[simple star]

I cant wait for spark to increase context and accept images

[severe mist]

How do I use GPT-5.3-Codex-Spark

I see it in my usage tab but VScode doesn't show it as an option

[final ruin]

uh try using Codex CLI

[ivory zodiac]

you needa pro account to use it

[severe mist]

I dont have pro

[ivory zodiac]

i believe they are going to start letting plus accounts in soon though

[severe mist]

😭

why they bait me like this

[ivory zodiac]

got em

[severe mist]

"look at all this spark usage you have! "

[ivory zodiac]

i think its a transitionary period

[final ruin]

Is Spark just a quicker model?

[ivory zodiac]

well no

its not just a quicker model

[severe mist]

isnt it slightly better and also faster

[ivory zodiac]

1000 tokens per second

or more

1200

it literally writes walls of text instantly

and no its less accurate and therefore needs to be used in a particular way

[severe mist]

the code outcomes they showcased, spark looked better while also being much faster

[ivory zodiac]

it is not better

[severe mist]

dang

cherrypicked then I bet

[ivory zodiac]

everything in life is tradeoffs

no they were just easier tasks

they were using it to its strengths

[severe mist]

they did a snake game and spark visually looked and played better

[ivory zodiac]

just variance.

[severe mist]

I suppose, yeah

[ivory zodiac]

LLMs are like slot machines

[severe mist]

with music gen I called it "mining for diamonds"

wish openAI would try to do musicgen

since all the ones focused on that sold out to UMG/WMG

(to become walled garden enviorments where UMG/WMG owns the output)

I'm still upset about it. I loved working on music with AI

[minor niche]

how do i fix this?

"Failed to apply patch to repo /workspace/d1-template. This can happen if the setup script and agent modify the same files."

[craggy cloud]

its popped up in my dashboard too, and can confirm it shows on cli too

(i am on plus btw)

■ {"detail":"The 'gpt-5.3-codex-spark' model is not supported when using Codex with a ChatGPT account."}

well had to try. lmao.

[indigo raptor]

My VS Codex has been stuck on this screen for a long time.

I tried restarting VS Codex a couple of times too.

[potent mason]

https://www.axios.com/2026/02/27/pentagon-openai-safety-red-lines-anthropic

Anyone know if this is legit?

[glacial shadow]

Nope, saw this article too, im taking is as unknown/not-factual

If oai releases a statement saying exactly this, ill believe their statement and not read the axios anyways, its not a primary news source

[boreal badge]

neco arc

[severe mist]

👀

[cyan wing]

5.4 leak confirmed 👀

[boreal badge]

codex app only for mac - not fair

im on linux

[hard tulip]

its not even on windows its going to be some time for you lol

[boreal badge]

bro what, i can run dmg package with a wine

🐁

[hard tulip]

wine suports dmg files lmao

[boreal badge]

yes

[boreal badge]

https://youtu.be/NdgWUeylzQk

[potent mason]

Codex has been running for 1h+ on this one task lmfao

[potent mason]

I'd imagine it would come to linux first (as most CLI's did)

[hard tulip]

in no world is codex app coming out on linux before windows

an app isnt the same as a CLI

[potent mason]

Well I would assume an app comes out in windows before Mac but it seems these days Windows is neglected (specially for devs since most don't use it)

[frosty zealot]

When I'm in hte CLI, I wanna use the App, when I'm in the App I wanna use the CLI

[potent mason]

I just decided to have the app for things in general and right behind it, I have ghostty running the development server, a terminal slot (for things I want to in general like vim, cat, ls, etc), and finally 1 codex cli and one claude cli (mostly for UI).

[frosty zealot]

I just feel like with tmux etc. and also ssh access I can be so much more efficient, the only thing I really like about the app is how easy they make using worktrees

https://x.com/sama/status/2027578580159631610?s=20

Interesting so Anthropic stuck to their guns about no AI for miltiary use then OpenAI cheered them on, and as soon as they're out, OpenAI submitted a bid lmao

[potent mason]

I mean they are going to do it either way with or without open ai (backup options are google, and grok, or worse case scenario an opensource model)

This is a government issue not an openai one

[frosty zealot]

I thouight they were all banding together with some petition

or was that just the employees

[potent mason]

Well I mean even if they are stuck with an opensource model they're gonna end up with an AI willing to do stuff for the millitary

[hard tulip]

use the codex cli in the codex app 😂

[obtuse river]

A worse AI means less harm less soon.

[latent tiger]

You think the CCP lets its AI labs deny them access to their latest models for malicious use? End of the day the us govt has weapons to truly oppress the American people if they wished to independent of any AI labs help. When it comes to national security and dealing with other countries that have the same level of AI access you cannot in good faith just restrict the military from using your model as they see fit in my opinion. Imagine a company that makes the best missiles trying to sell them to us govt but they’re like btw you can’t use them to strike in X Y Z specific scenarios and if you attempt to we have the ability to deactivate them. Apply that same logic to the AI models lol. I’m glad open ai was nimble and snatched the position though will probably accelerate progress now

[latent tiger]

5.4 codex when ???

[feral geyser]

I made this...while I was working on my car and doing household chores today!!! Amazing hahah
https://hunterwiley.neocities.org

[supple perch]

@boreal holly - Well soon you will never have to install even the IDE locally, into the sandbox it goes

[gentle steppe]

I've been using Codex for about a week to refactor a large project in Xcode, so if like me, your code lives in an iCloud-synced folder, a simple way to avoid syncing build artifacts is to add a small wrapper script (for example Scripts/build-local.sh) that runs swift build/swift test with --build-path set to a local cache directory like ~/Library/Caches/<Project>/swiftpm, and runs xcodebuild with -derivedDataPath set to ~/Library/Caches/<Project>/derived-data. This keeps the repo clean (no .build churn), reduces unnecessary cloud sync traffic, and keeps local builds repeatable. I asked Codex to help implement this, and it generated the build-local.sh script for me. 🙂

[raw hill]

or you could just not put your version controlled software in an iCloud synced folder

[gentle steppe]

Whatever works for you buddy.

[sterile crow]

Here's how you can use Claude's new /simplify feature in Codex:

https://x.com/bytes032/status/2027691445516112040?s=20

[karmic shadow]

why codex its too laggy for me today?

[exotic herald]

What model are you using ?

[karmic shadow]

5.3 codex

on vs extension

[exotic herald]

Nice are you using Max , Mini or Medium ? Also are you using platform account or are you using CLI account

[karmic shadow]

medium, platform acc

i have been using it fine for a week

but today, i changed proyect folder and it start laggin and do strange things

[exotic herald]

Really that's crazy - maybe put it in the codex show and tell as a documentation and we can try and replicate the error and see whats going on ?

[karmic shadow]

seems that happends with large conversations, when i open a new chat, it works fine for a while untill it gets laggy again

[velvet wren]

I highly recommend the Codex app

[exotic herald]

Did you try to compact the chat ? And also if you don’t me asking what are you building ?

[exotic herald]

Wow this is great a great recommendation - I downloaded it a while and played around with it. But I have been mainly using the CLI - i.e Spinning up 4 terminals of codex LOL. Wow this is great ! I am gonna have a lot of fun this weekend.

[karmic shadow]

just for mac

[velvet wren]

Windows invites are being sent, so make sure you're on the waitlist

[karmic shadow]

im on

one from microsoft store its not even working

[nimble terrace]

@ivory zodiac Yo when you saw GPT 5.4 in the Codex app did you try to prompt it or nah

[feral elk]

does codex app work better with multi agent now?

[cedar skiff]

it works, but there is no feedback

[spiral current]

Good morning everyone. I setup Codex alongside Claude in WSL. What type of setups does everyone else use? Wondering where I can minmax, mostly working on accessibility and instructional design, dont need anything too fancy

[velvet wren]

Vercel skill is pretty good. I just prompted create a hello world app and deploy to Vercel and got a nicely formatted web app

[sand shuttle]

where do i put my agent.md and skill.md files

[cedar skiff]

doing a deep research document focusing on official documentation and authoritative source with claude. Then converting that into a skill with nested reference docs is like a silver bullet with codex. You can get up and running on anything with it.

the skill.md is just a problem -> solution index for the reference files

[cedar skiff]

codex can tell you, just ask it to check the offcial docs

[sand shuttle]

why dont u teellme sir

[cedar skiff]

seems like you should ask codex that

.codex/skills/

[sand shuttle]

ah ok

[cedar skiff]

but you know there is more to it

like a lot more

[sand shuttle]

oh

mind explaining

[cedar skiff]

codex has a skill for creating skills

[sand shuttle]

or sending a link

[cedar skiff]

uh where do i get the link?

[sand shuttle]

i just need to create a folder

and a skill.md

[cedar skiff]

yep thats all there is to it

[sand shuttle]

and for agent.md?

do i place it in the .codex

folder

[cedar skiff]

there is a bunch of nuances you need to know, you probably need to read the docs

[sand shuttle]

i had ai create my agent.md file

[cedar skiff]

It usually does a pretty bad job of it, puts data in there that goes stale fast and the agent can find easily

[plucky halo]

This doesn't appear to be on his profile

[torpid trout]

That’s because the post is now deleted

An ok writeup of the hype https://piunikaweb.com/2026/02/28/openai-gpt-5-4-accidentally-leaked-codex/

[plucky halo]

Awesome - thank you. Been trying to get that bloody pop up all day

[main wraith]

lates codex update has broken the real time UI updates in the app/extension. i have to click on it for the latest updates to arrive

[topaz pewter]

Dude why do the codex cli work like this bro
if my weekly usage runs out i cant use my hour hour usage
if my 5 hour usage runs out i cant use my weekly usage
its so stupid

now i cant use codex til march 5th

[velvet wren]

what plan are you on?

[topaz pewter]

plus

[velvet wren]

how much code have you created?

[topaz pewter]

how do i check

i only been using for 3 days

;/

[sand shuttle]

just go thru ur chat and see th enumber of lines

althought i think there is a way to get the total

idk how

[topaz pewter]

how do i do that

😭

i gotta use trash opus 4.6 for 5 days

bro

[sand shuttle]

opus 4.6 isnt trash

why do you say that

[topaz pewter]

after using 5.3 codex

opus is so bad

😭

5.3 codex feels better to use

[sand shuttle]

actually?

[topaz pewter]

yea

[simple star]

Dont you understand why it works this way?

[topaz pewter]

nah

[simple star]

Let me explain you:

[topaz pewter]

why?

[simple star]

• Weekly quota: This is obvious. There is a limit at how much tokens you can use in a period of time

Right?

[sand shuttle]

how are the usage limits on plus plan? are they better than claude pro plan?

[topaz pewter]

i mean pretty good i guess lol

[sand shuttle]

how much work could u get done?

[topaz pewter]

alot

[simple star]

5 hour quota: This is so that you don't blow all your weekly quota at once. The point of this is that you dont blow up all your weekly quote in 1 day, because you hog resources that could diminish performance for other users

[topaz pewter]

ah

how do i use GPT-5.3-Codex-Spark

[simple star]

The 5 hour quote is effectively just a rule, so that you dont hog CPU performance at once

It is not like a "separate quota"

[sand shuttle]

also did they decrease limits on the models? my usage is oign down so fast today, did they teporariyl increase it after releasing 5.3 codex

[simple star]

You have to have a Pro account

[topaz pewter]

yea i think so

[topaz pewter]

aw dang

[sand shuttle]

dbugger did they decrease limits on the models? my usage is oign down so fast today, did they teporariyl increase it after releasing 5.3 codex

[sand shuttle]

oh 💔 no wonder

[topaz pewter]

welp

[simple star]

@sand shuttle Mine too. This is likely due to subagents requiring more tokens

[topaz pewter]

i have unlimited 4.6 opus

[simple star]

@sand shuttle This is likely not a consequence of 5.3-codex, but a consequence of Codex v0.105

[sand shuttle]

Mind explaininig?

im nto even using sub agents

just prompting oin the extension

[simple star]

Dont you have them enabled?

[sand shuttle]

nah

[simple star]

What version of Codex are you running?

[sand shuttle]

im on the extension so im not sure

how do i check

[topaz pewter]

i made my own version that has a prompt enhancher lol

[simple star]

codex --version

[sand shuttle]

i gotta disable sub agents is that it? will that help?

[simple star]

@sand shuttle There's no guarantee that will do it. It is just my suspicion

[sand shuttle]

ive just used 11% of my weekly limit 😭 im on the go plan for now

[sand shuttle]

Ah ok

[simple star]

If you see something like that while Codex works, then you have subagents enabled

[sand shuttle]

well im using the extension so im not sure

[simple star]

I can't tell you. I only work with the CLI

[sand shuttle]

my cli version is still 104

[simple star]

You likely do not have them enabled, even though not impossible

Me personally, I have not seen an increase on quota usage due to 5.3

Only due to agents

[sand shuttle]

my quota was goign down much slower 3 days ago, now its goign down super quick i think they had temporarily increased the quota after releasing 5.3

[simple star]

An increase of quota has no effect of "usage rate"

Different concepts, "how much you have", against "how fast you spend it"

[sand shuttle]

the % remaining goes down slower

[simple star]

Yeah, that is true

[sand shuttle]

Yeah thats what i was refering to haha

[sand shuttle]

in ur opinion should i buy the plus plan or claude pro plan? bopth are 20$ but which do u think offers more usage

[simple star]

Honestly, I am Pro

I just... need the tokens

And even then, I already blew out 80% of my weekly quota

[sand shuttle]

What did u do 😭

[plucky halo]

He’s made ~26262727 snake games but it only took 5 seconds each

[sand shuttle]

btw if u have an idea what offers more usage? chatgpt 20$ plan or claude 20$ plan

[simple star]

no idea, but it flew by

I added a multi-member feature to my bookkeeping platform

[sand shuttle]

i feel ur pain lol but im on a tight budget so just rying to get thge one which offers more usage since i fell the quality fo the modes is similar

[simple star]

It was tricky, and I had to start over a few times

[tawny island]

how do you enable to hotkey for it? post is removed

[spiral gorge]

codex only
or
opus to write code, then codex fix bugs
or
opus only

[sand shuttle]

is there a huge difference in token consumption between 5.3 high and 5.3 medium?

[orchid plume]

back when I had plus (before and after the 2x rate limit boost), I'd honestly say that it felt like I got a lot more usage out of ChatGPT's Plus compared to Claude's Pro

also with ChatGPT you have quota for Codex and quota for ChatGPT, while on Claude it's combined

[sand shuttle]

Hm thats true

Im on a tight budget so im reall ytrying to choose the plan that gives me the best usage

[orchid plume]

understandable, I guess it depends on your usage pattern. I'd say you'd get more out of Codex, even more so with the 2x rate limit boost at the moment. The other alternative is Copilot for $10/mo I believe, 300 user messages (monthly) if I recall correctly and still access to the full context window for 5.3 Codex. Tool calls and subagents do not count towards the user messages, only what you send

[sand shuttle]

form what i heard copilot has restrcited usage on premium models like codex 5.3, opus 4.6 etc so itsnot worth it

[orchid plume]

never had such issues when I was using it with 5.3 Codex on high effort. I've seen such reports though for Claude models, particularly Opus

I suspect GitHub has discounted access to OpenAI models, as well, Microsoft. Anthropic, probably not as much

I also believe they offer a trial for one month, so you could always try Copilot Pro for one month

[sand shuttle]

Ah thats true

how would you say the token usage/cost compares with 5.3 and sonnet 4.6?

[orchid plume]

never used Sonnet 4.6 so I can't say, I think the last Sonnet model I used was possibly 4.5, or maybe the one before that 😛

[sand shuttle]

Well thanks for the help :)

[south solar]

4.6 is best

in coding and creative tasks

but codex is also good but it uses gpt as its base and gpt is in good at quite personalized experienced

and they both costs similar

but claude needs less cost

[simple star]

It drives me so mad, like seriously furious, everytime Codex (or any LLM) first response is "You are right", "Correct", "Indeed"...

[south solar]

;-;

[desert flame]

have u guys desktop codex app for windows?

[simple star]

@desert flame no

[south solar]

no

[sand shuttle]

I've heard claude is more costly

[south solar]

4.6 compares to codex 5.3

[desert flame]

u can get plus subscription for free and use codex without any restrictions

[sand shuttle]

How would I do that

[sand shuttle]

I see so it's similar

[desert flame]

go to poland and create new account using poland mail

[south solar]

and for codex its not public the cost so yeah

[sand shuttle]

I see 😔

[ivory zodiac]

it was just a video

[nimble terrace]

Oh?

[ivory zodiac]

yeah it was on X

[nimble terrace]

ah

didn't know that, appreciate it

[plucky halo]

Also, fairly certain this is feature gated at the moment and not accessible to users even with some jiggery pokery

[nocturne folio]

gpt5.5 codex soon

[ebon skiff]

i alr use it

[sick iron]

how? Mine says the version is incorrect and need to redownload it - from the exact place I got it the first time.

almost like the joke for blonds

[ivory zodiac]

https://x.com/LLMJunky/status/2027818059319132352?s=20

[boreal holly]

Agreed. I still frequently use gpt-5.1-codex-mini for tasks where its strength is being smart enough for that task

[ivory zodiac]

Right?

And you benefit from speed

[boreal holly]

Yeah and it uses way less tokens

[ivory zodiac]

I want a new mini

It's time...

Been long enough imo

5.3 codex mini, 262k for cheap

[main nimbus]

latest codex mac app sucks is borked. it broke realtime updates for me. i have to click the UI to see what is evening happening.

[lavish geode]

@ebon skiff same happened with me hours ago, currently seems to be fixed

[main nimbus]

hmm, maybe haven't gotten the patch for it. hope it is fixed soon, it was confusing at first.

[solar mauve]

is gpt 5.3 codex better than gpt 5.1 codex max?

[main nimbus]

in multiple ways. speed is 25% better than 5.2 which itself was an increase over 5.1. it has much better long-context handling.

[lavish geode]

Oh no it's still present not fixed😂

[main nimbus]

🙉

[simple star]

I was sad because I saw my quote didn't renew until 2nd March

Then suddenly I realized... FEBRUARY!

Finally I find a use, to have a 28-day month

[blissful basin]

Anyone knows where i can submit bugs found in codex windows app?

[high girder]

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

[tall zodiac]

yall tried this yet?

https://www.cmux.dev/

[potent mason]

Feels a bit immature the image doesn't adapt well to different screeen sizes, don't really trust the product if something like the landing page is already wrong

[plucky halo]

Jesus. That UI looks horrific

[potent mason]

Lmfao someone gave me their catalog through a pdf and I had to recreate it all so naturally I was using AI for it, I got rate limited on the $200 plan xd

Honestly understandable though

[next ore]

Hello, does anyone know how much I get for the $40 credit extension? I am only using codex via CLI and ran out for this week. I have never used the $40 credit I just want to know if its worth it.

My basic question is, with the pro subscription ($200), is the $40 credit 1/5 the amount of tokens the pro sub provides? or is it less than that?

[boreal holly]

It definitely feels like it goes by quick. Hard to say though. Perhaps a Plus user can share their experience and whether it feels like "twice as much" or not.

[velvet wren]

As a Plus user I can say that credits are worth it. 1000 credits for USD $40

[next ore]

I think online it says pro users get 10k credits. so 10% of the so you are effectively paying twice the price for pro. as its effectively $20/1k credits.

[torpid trout]

Credits are definitely not worth it lol.

You better use the api to bridge out until your normal account refills.

Credits go like butter - literally each and every „message“ any subagent „thinks“ is counted.
With api you at least get actual token billing.

Ps I’m both plus, pro and api customer. And I’ve tried them all (least of all pro as only customer on it since 3 days)

[torpid trout]

Where does it say that?

If that’s true then pro would last me about 6 days and subsequently plus 1 day
Which isn’t the case. I made it roughly through 5 days with plus and one day with 1000 credits.

It’s a silly measure anyway. I wish OpenAI would just use tokens everywhere like in api.

[main nimbus]

think it is fixed now. got an update not too long ago

[next ore]

so... whats the "official" token count. I searhed the website. I cant find it. maybe I'm blind. I saw some comment online that said its 10k but I guess its wrong. I've just been using the CLI endlessly without much concern for token usage, it's been working great so far until now.

according to the man himself (chatgpt) the $40 credit is 15-20% of the pro plan

[tall zodiac]

cc: @potent mason

i like how everything is all in one view

[ebon skiff]

idk, i just download it and its works normally

[potent mason]

Yeap agreed I’m just going to wait it out until a good one comes along (I’ve had enough changing and configuring)

[abstract beacon]

u might not be on the invite list since its an alpha release

[cyan wing]

Anyone kno a easy way to run Codex in Plan Mode and auto accept all options?

like auto accept recommended and "Implement the plan."

[frosty zealot]

Kinda defeats the purpose, I dont think theres a way to get it to accept its own plan and carry on, you could just prompt it to do some planning outsidfe of plan mode

[high girder]

So who here is actually in the windows codex test at the moment? What's it looking like?

[tropic karma]

I saw this (I'm on the $20/month), very cool, but I can't seem to actually use it yet. It doesnt show up in the IDE extension, or the CLI.

[high girder]

They're supposedly bringing Spark as a trial for plus, but it's not implemented yet

[hybrid ore]

Same here
I was literally going to ask about this
Why is it reflecting when it's not accessible.
Should be remove if I can't use it.

[true bane]

id like to introduce everyone to some business geniuses:

[tropic karma]

yeah though im pretty sure that openai doesnt do that

theres pros and cons to it. you should get more usage with what anthropic does, but yeah its pretty stupid.

The ideal system would be:
free tier (no claude code) message pool on all subs
then on top of that a combined usage pool for claude code and chat messages for paid tiers

[plucky orchid]

Hi, someone know why the Codex app on windows store don't works ?

[plucky orchid]

Did you find the issue ?

[modern tide]

Hello. It seems since a few weeks ago my codex performance have been severely degraded. I use gtp plus and web version. It can't handle the same task it did just at the beggining of febuary, I also can't select deep thinking or which model should work on it. Any tips?

[sick iron]

Nope

[hard tulip]

How

[hushed storm]

It’s rage bait fr

[orchid plume]

I don't believe it works unless:

• You're invited to try it, e.g. you had an invite by email
• Alternatively, you bypass the UI which shows the update required notice, which you could ask Codex to do

when I tried it though, it seemed like it was missing things compared to the MacOS version of the app. I couldn't find an automations tab and I couldn't find a way to go into the settings UI

[ivory cairn]

is it me or is my usage for codex going down significantly faster in the past 2 days?

[simple star]

nav mine too

I suspect it is due to agents

What version are you using?

[desert flame]

what a legend

[simple star]

I wish CLI could show me the usage quota on the statusbar

[ivory cairn]

idk I am using the Codex app

But I have used little prompts and I am 87% used weekly (within 1 and half day if it refreshing) which is insane because last week I used sooo much and I struggled to finish it.

[ebon skiff]

got invited from gmail

[boreal holly]

Yep, almost time for credits 🥲 at least every % used was forward progress. Very proud of my agents. They do incredible work

[simple star]

Sadly, next week we will have half the quota 🙁

[boreal holly]

What? I thought the promo was April 2nd

[simple star]

Oh wait, maybe I misread

[modern tide]

Any help?

[simple star]

@modern tide Too little information. What model/effort?

[modern tide]

Says 5.2, what effort?

[boreal holly]

You're using web version so you can't customize the model, it uses 4x as much credits, and perhaps you need to ask Codex why it's performing poorly (e.g. what do you need from me to do this job correctly end 2 end)

[simple star]

The web version uses 4x credits?? O_o

[boreal holly]

Yeah rate limits drop precipitously when using cloud codex

[simple star]

holy..

[modern tide]

I believe i was able to. Ask it as in directly in prompt?

[boreal holly]

Yeah, give it a prompt like "Make a comprehensive list of tooling you need to work on this project that is not currently available inside your execution environment, provide me with any setup details I need to do on my end, install the missing tooling on your end if applicable. If the project we're working on is ambiguous in any way, or the technical specifications are not clear, let me know what clarifications you need from me so our work can be as productive as possible. I value static code analysis and rigorous unit/integration testing, so any gaps in our project that result in failed builds or flaky runtime, please collect these gaps and present them to me along with the aforementioned details."

It's like a "realignment check". You could even make that a part of the AGENTS.md. Whenever you say "realignment check", all the stuff in that prompt becomes applicable

[modern tide]

well its weird i was using it for literally same tasks just a month ago and it implemented it for like 80% (pretty much what I needed) so now i just fizzles and do nothing with exactly same instructions and inputs

#codex-discussions message

[nocturne folio]

something new

[simple star]

what?

[nocturne folio]

i didnt notice ive gotten the spark model on the plus plan lol

[simple star]

Yeah, lot of people have been reporting it for 3 days

[fallow stream]

wait we can use spark on plus?

[simple star]

I think you see it in the UI of the website, but you still dont have access to it through the tool

But im not sure

[civic oriole]

how r u on that

!?!??!

[ebon skiff]

Mine is fine

[civic oriole]

!?!?!

[ebon skiff]

idk lol, I feel it's better than the one on the VS Code extension (I’ve never used the web one), and so far it's fine

[civic oriole]

i hate the vs code extension cause then i cant see my files

holy scam

[modern tide]

there's no codex app for windows or im missing something?

[civic oriole]

its on the windows app store

[civic oriole]

https://apps.microsoft.com/detail/9plm9xgg6vks?hl=en-US&gl=US

[modern tide]

thanks

[boreal badge]

codex -5.3

[simple star]

wow, I was not expecting THIS

[main wraith]

are limits frozen? or my app's broken?

I been using it or awhile and still got the same usage left

[torpid trout]

so... PRO plan supposed to be so much larger than PLUS, but within less than 1 hour, 6% of the weekly gone?
If I continue like this within 20 hours it will be zeroed 🙁

6% plus 10% of the spark, to be precise.
I definitely felt PLUS was not going that fast, how could PRO go even faster lol.

[torpid trout]

lucky you

[craggy cloud]

i tried a few days ago to use the spark model and the server wouldnt even allow the connection

[modern tide]

uh... have you solved this?

[civic oriole]

nope

[modern tide]

jfc

this app also sucks

i dunno why they had to remove functionality from web version

[unborn helm]

Codex asks me to run a command EVERY SINGLE TIME even though i give it permission to run and ALWAYS use "Yes, and Dont ask again for the commands that start with" and IT ALWAYS asks me.

[boreal holly]

Are any of those mutating git commands like git commit? Also are you in workspace-write sandbox mode?

[boreal holly]

Are you using subagents?

[orchid plume]

interestingly you're not the only one to notice a possible issue with the quotas. My weekly quota is behaving strangely. Not only does it appear to be burning faster than usual, but it's also been bouncing around. Say for example, it was 61% originally, then a moment later it's suddenly 58%, and then a few minutes later it's back to 61%

[boreal holly]

Are you using subagents?

[orchid plume]

not at the moment, I turned them off this morning after it felt like my quota has in some way changed (UK timezone)

latest version, 0.106.0 CLI

[boreal holly]

Gotcha. I have a feeling that feature is still not ready.

[unborn helm]

Its a bunch of get-content thinggs

[cedar skiff]

what kind of work flow do people use to get spark to build out small apps?

[boreal holly]

One thing I noticed is Codex will sometimes spawn multiple sug agents to work on the same issue.

[boreal holly]

OK, sandbox setting would be helpful, but if you have network_access = false, or you're in workspace-write without specifying network_access = true then Codex will ask you any time a command touches the network, even if you approve forever

[cedar skiff]

Also setting the workspace to trusted helps reduce them

[unborn helm]

Yeah this fixed my issue thank you

[cedar skiff]

I'd love cookie cutter workflow to get spark to make me a prototype, anyone know of nay good place to get started?

[boreal holly]

The low friction, sane security settings would be

sandbox_mode = "workspace-write"
network_access = true
writable_roots = ["/path/to/cache/folders", "/for/example/.pubcache", "/or/global/node_modules", "/var/run/docker.sock"]
approval_policy = "on-request"

approval_policy = "never" in my testing leads to good commands getting rejected, and then Codex begins scheming for a workaround. "on-request" rarely ever gets hit, and if it does it's your golden opportunity to either allow it or steer them towards the correct way of doing something.

[torpid trout]

I do, but so did I with PLUS, and .. last week I run out a bit early but I created like 3 new apps from scratch plus client work
And today I barely sit on it (meanwhile) 2 hours approx, 80%!!!!!

Being this now PRO account I am close to cancel that immediately haha

It's not any more than PLUS - literally it does consume faster.

Things that changed

• new PRO account
• new version 106
  otherwise all the same

And with spark now I even changed some agets to NOT use 5.3 xhigh which otherwise it used until now

(yes yes, xhigh)

[sand shuttle]

why does codex always do this

[boreal holly]

With pro you get priority speeds too. So faster agents on top of an experimental and inefficient subagent system will eat the rate limits. My rate of consumption is 2% per hour weekly for 12 regular agents with subagents disabled. It's like 10-20% per hour with one agent and subagents enabled. Suffice to say I have never used subagents since that happened.

[sand shuttle]

it has started happening recently

[torpid trout]

It's like 10-20% per hour with one agent and subagents enabled. Suffice to say I have never used subagents since that happened.
That is pretty much what I see
Horrific. So, disable subagent mode you say?

[torpid trout]

last week I was using them on PLUS and made it through a week

So yes, 100% new issue

[boreal holly]

Yeah disable it. Using regular agents, you can still have a "subagent-esque" system without the official feature. There's all sorts of ways to do it, none of which seem to burn through the rate limits like the official feature

[torpid trout]

basically like this?

[features]
multi_agent = false
prevent_idle_sleep = true

[agents]
max_threads = 1

[boreal badge]

websockets die

used only HTTPS mode

[boreal holly]

Yep! My favorite, minimal effort subagent system is adding a skill that lets agents use tmux new -s and tmux send-keys to spawn agents and send each other messages. The strategy I'm currently using is codex app-server and a MCP server that lets them sent thread/steer messages to each other, which doesn't interrupt you typing a message into the terminal.

Anyways, yeah, I love Codex/OpenAI but the subagents thing is definitely not ready for production use yet

[boreal badge]

ewww, tmux - screw

screen so better

[spiral gorge]

5.4 when

[boreal badge]

we need codex 6.0

next best gen

[sand shuttle]

how do i do that

[torpid trout]

OK, first, disabled those and now i am rerunning the same tasks as before again (at the cost of tokens but.. have to know)
Then I will look into what you describe there, I am sure if it saves tokens it is 100% worth pondering on it.

[boreal badge]

In new version of OpenAI codex, they added websockets so that the context "cooks better," but this almost doesn't work anywhere now, so it switches to HTTPS automatically.

[torpid trout]

Only problem I have with no subagents is it seems it then also simply does not use them anymore, not even in synchroneus mode?

[sand shuttle]

thanks

[boreal holly]

Exactly. It's a tradeoff. There are alternatives that work very well but do require some engineering effort. The way I see it, OpenAI has hundreds of features and thousands of bug fixes they're working on, and the only thing I want is subagents, so the solution is build that system for myself the way I want it. With a Pro account you can start today and be done by Tuesday building something that works just as well and does not consume your entire rate limit within hours, and it'll use a tiny fraction of the limit to do so 🙃 or operate with single agents until OpenAI can address the subagents efficiency issue

[torpid trout]

  Delegation Report:

  - Primary owner: main thread
  - Agents invoked: none
  - Skills activated: none
  - Required delegations not used: php_engineer path was applicable, but no specialist agent runtime is available in this environment, so safest direct fallback was applied.

It basically ignores the agentic behaviour, this is weird
I would have expected it to at least apply it per-thread.

I will test a few other CLIs first, I have heard of opencode or something like that, which works with openai. Maybe they do that (better)

And I am opening a GH Ticket unless i can find one

[boreal holly]

Here's a good subagent system that is extremely optimized for token efficiency. And idk if the official subagents feature does this, but my agents can talk to each other, and the orchestrator manages all of them asynchronously

[torpid trout]

This thing open source?
😛

Opened GH https://github.com/openai/codex/issues/13179

[boreal holly]

Not yet! It only runs on macOS/iOS

And is built around basically my mac and how I have it set up. Very extremely opinionated

[torpid trout]

Yeah, I get it... well if ever you decide to share the source code let us know 😄

Obviously given your reports of what you use in terms of rate limits, something is done better.

Cant' be that I am overwhelming codex so much, I had a look at it with the api and really most of it is INPUT tokens, so, should be dirt cheap. The expense is on the OUTPUT, and it produces a fraction of output per input.

So... yes there is something off with their agent behaviour

[boreal holly]

I think OpenAI's "rate limit" is a mixture of tokens and requests. Nobody really knows the algorithm, but back when subagents first came out and it was a hidden experimental feature, some folks with Pro reported dropping to 0% weekly in like 10 minutes of usage, and then OAI staff on Github were like "that was a hidden experimental feature, should not use it until it's done". There's no way tokens had anything to do with it. Also there's a CLI command called like /debug-dump or something that will instantly drop you to 0% weekly the moment you press enter (idk why it's in the app for people to use). But something tells me the requests play a role in the rate limits. My subagents system uses the same requests as a normal single-agent codex session, just distributed.

[true bane]

anyone on plus figure out how to use the random new buckets that showed up 2 days ago for "spark"? if its only for the oligarchy tier why do they show it to us?

[plucky halo]

Oligarchy tier 😂

but yeah, it seems for Plus users, that's for looking not touching currently

[true bane]

its that or "department of war tier"

[boreal holly]

I love watching them coordinate. They're like "blocked until that agent makes DTOs", orchestrator knows they don't need to respond to that until the other agent unblocks with DTOs, the agents rebase to acquire what they need. It's insane. It's literally an entire company of senior software devs working autonomously at light speed

[plucky halo]

Ah, the more-expensive-due-to-more-use tier!

[boreal holly]

5.3 Codex, I can't imagine anything better but I'm sure OpenAI will continue to impress

[simple star]

@true bane I rarely use Spark. It is fun, but you are not missing out on anything extraordinary

[true bane]

Thx more interested in a new token bucket or allowance than any model as I can always use more /review

[livid nimbus]

anyone else getting this now

when codex gives you links

[orchid plume]

the strange weekly quota % anomaly is still happening for me. Is anyone else seeing this happen today?

That 53% shown in Codex CLI will sometimes say 56%, then 53% or some value close to that, then 56% again 😂

[torpid trout]

Well but you’re comparing local to remote counter?
Then I’d not be surprised.

However if they jump locally from 53 to 56 or-and remotely too… I think you should report it.
They don’t want to believe any discrepancies with those counters right now so every report puts more cracks in the otherwise smooth eggshell this counter is advertised as

[true bane]

not for me and I stare at that page a zillion times a day. my kingdom for a.. "widget".

[orchid plume]

yeah today at least is the very first time I've noticed this happening

[simple star]

I really wish I could make PRs to Codex

Now, not even my "Issues" are being answered

[orchid plume]

and yes, I've posted a comment about it in one of the open issues regarding some of the recent feedback about quota concerns. Sure the value might be out of sync, although I've not seen this happen, but for it to flip from 56 to 53 and then back to 56 on the CLI, it's a bit concerning 🙂

[torpid trout]

Yes, I think my pro sub will very soon become a second plus
Can’t waste 200 bucks a month for being a Guinea pig

That’s usually what you get for no charge - no offense OpenAI, the product in general is amazing but some things just can’t be touched when people drop money in these heights, imo

[bronze snow]

Does anyone know how to implement codex inside VS Code workspace?

[nocturne folio]

is there something telling openai to change keybinds to do the most useless things ever

transcript should be ctrl t

not esc x2

esc x1 should cancel prompt and x2 should close the app like how it have been since it released

[chrome raven]

same for me

[boreal holly]

esc x3 to pull up skill menu, esc x4 to rename thread, esc x5 to switch to plan mode

[torpid trout]

Esc6 to reset credits gd

[boreal holly]

Actually it should be esc x2 to immediately archive current thread, esc x3 to switch to plan mode lol

[simple star]

I still have not managed to find a use for skills... I wonder if it is my skill issue

[rough abyss]

Hi all, I just found this article showing how Codex can take into account the IDE with the JetBrains MCP server (IDE acts as an MCP server that agents can connect to) 
Wondering if anyone knows of an equivalent for VS Code?
https://developers.openai.com/blog/skyscanner-codex-jetbrains-mcp

[left elm]

Still blocked in this container, but I did reattach origin successfully first.

git remote -v && git fetch origin --prune && git branch -r && git log --oneline --decorate -n 5 origin/main
(origin was missing at first: “does not appear to be a git repository”)

git remote add origin XXXXXX && git remote -v && git fetch origin --prune
(origin add worked, but fetch still fails with CONNECT tunnel failed, response 403)

So: repo URL is set again, but live GitHub sync from this container is still network-policy blocked.

I'm pulling my hair out here. I've done everything I think I could possibly do on my local end to fix this, but nothing seems to have worked. The only thing I haven't tried is disconnecting it fully from GitHub and reestablishing the connection.

Anyone have a solution for this?

[velvet wren]

possibly you don't need any. I use a few skills such as the Vercel skill so I can deploy apps to Vercel

[torpid trout]

Skills are useful if you have very deterministic requierments such as "freaking use WPCS when coding in PHP for wordpress"

Because gpt will categorically refuse to do so - and if you prompt it to do so it will forget it after 2 lines of code
That is where skills and agents really are useful becaues you can force it to do it with verification matrix and similar things

Similarly for repetitive things you do not want to repeat every time, like - say you want all your changelogs in a specific format, or, the format of doc you prefer etc
You could stuff it all in an agents md but the thing is, you migth not ALWAYS need that, so a skill/agent can be useful as it reacts upon codebase recognition and intention in prompt.

[gleaming sapphire]

anyone seeing this:
/Users/home% codex
missing field id_token at line 5 column 3

[void cedar]

Tried to install Codex on Windows in the store and got this. When I click "here" it just goes back to the store. Is this a new bug?

[void cedar]

I got the same error. Did you fix it?

[civic oriole]

it dont exist

its not out yet

[true bane]

i actually like wsl2 on vs code's terminal quite well. way better than powershell/claude

[boreal holly]

After running out on Pro, burning through 1000 credits, switched to Plus and hit the 5 hour rate limit in like 1 hour. So on Pro you get 5 1/2 days of 24 hour, sparse 12 agent usage. On Plus you get exactly 3 hours of usage per week, with 4 hour breaks in between. It's 1000% not 6x more usage. It's wayyyyy more

[tribal lake]

Tell Codex to do something
Codex spins up subagent
Subagent gets stuck with a tool call prompt "Do you want to approve this to run for this section?"
I don't know anything is happening, I just see the main agent taking a long time waiting for subagent

There gotta be a better way...

[potent mason]

Is anyone here on Claude Code? It’s been having so many issues😂😂

[velvet wren]

you can only discuss non OpenAI products in #ai-discussions

[potent mason]

Bruh look literally 3 messages above someone else mentioned Claude such bs, also not like I’m saying to use CC I’m literally saying it’s so bad rn

[snow thistle]

What’s happening here?

[nocturne dove]

Anyone here can arrange and can give me this course iclone8 3dmasterclass by tristryam aniyam,
My whatsapp is 9508752860 from India, if anyone can arrange this course which is iclone8 3dmasterclass by tristryam aniyan, please chat me with no

[cedar skiff]

gving code rabbit a run, seems ok. Kind of picks up on the same sort of things codex does. Just using it default though. I'll give greptile a go as well once my code rabbit trial dries up.

Best thing about it is it gives prompts you can just dump into codex

[potent mason]

Ah and ofcourse @velvet wren says nothing about everything else honestly hilarious😂

[left elm]

Someone help me plz Codex has lost access to github and Idk how to fix

[potent mason]

It seems more like an access issue

Use SSH not http

[left elm]

How do I do that Im super new to all this Im accessing through BRAVE browser atm

[potent mason]

Open up your terminal and do the following

Generate ssh key

ssh-keygen -t ed25519 -C "[e-mail]"
enter
enter
enter

Copy ssh key

clip < ~/.ssh/id_ed25519.pub or cat ~/.ssh/id_ed25519.pub | clip

Go to github and do the following:
In the upper-right corner of any page on GitHub, click your profile picture, then click Settings.

In the "Access" section of the sidebar, click SSH and GPG keys.

Click New SSH key or Add SSH key.

In the "Title" field, add a descriptive label for the new key.
Select the type of key, either authentication or signing. For more information about commit signing,
In the "Key" field, paste your public key.

Click Add SSH key.

[left elm]

Fatal: unable to access ... CONNECT tunnel failed, response 403

[cedar skiff]

you have codex, that thing knows all of the things

[left elm]

So Im using codex and it keeps giving me that when I tell it to go look at my GITHUB repo ive done personal networking stuff ive disconnected and reconnected the app it just keeps giving me this result

It was working last night and all of a sudden this morning it stopped and been giving me this error over and over again...

[cedar skiff]

right but give codex the problem you are trying to fix, it literally knows almost everything.

[high heron]

1. Update your cli to vers 0.106 ... 2. After 2nd of April the 2x quota offer will revert to half of what they give us now...

[left elm]

OK so I converted my github to SSH but Codex still comes back with

So your SSH setup on your machine is fixed (great), but this Codex container still has outbound network restrictions to GitHub.

so it seems maybe something on the Chatgpt serverside is goofy or GITHUB is that safe to say?

[hard tulip]

GPT-5.4

[left elm]

So does that mean OPENAI needs to fix the issue and im stuck?

[timber hazel]

anyone's codex limits going down insanely fast?

4 light prompts and 10% 5hour, when previously it would probably had been 2%.

[hot egret]

Same .-.

[chrome raven]

if the 2m window is accurate it would be a game changer, but if well managed for long sessions

[terse kraken]

I hope we get the general model in codex not just 5.3 codex

[lavish geode]

Same

[obsidian prairie]

hi everyone. ive been using codex for about 5 weeks now and i can say that i like it alot. just today i saw information about double rates for pro plan are for a limited time. i wonder if that info is outdated or will our limits be 1/2 in the future? my honest opinion is that for 20€ per month is decent deal for what i get now. but its a bit concerning getting 1/2 rates for that price in the future...

[terse kraken]

I dont think its 2x still, its back to 1x after the first week

Its just promotion, their servers can't sustain 2x throughput

[obsidian prairie]

ohh, thats a great news. rates i get now for my plan are perfect for me 🙂

btw do you have any experiences with spark? i dont see it in my vscode extension

[sand shuttle]

they probably reduced usag elimits

they temporarily upped it when releasing gpt 5.3

[torpid trout]

Stop guessing and assuming.
Report your issue in the relevant GitHub and read how it’s stated NOTHING changed

If you all only complain here and on Reddit, in GitHub it looks like some single snowflakes have an issue.

And that’s where that team works. Not here, not on Reddit.

[torpid trout]

At least this confirms it’s not a codex CLI version issue but a remote counter issue, because under 104 last week I used up to 6 agents and literally wondered when my credits will go down as they kind of didn’t

I start suspecting that actually the problem is in reverse
Plus gives you 45/250 messages and it surely did feel like much more… much more!
So chances are they fixed something, and now we actually get the promised messages - which from 45/250 really is almost nothing and corresponds with current reality

45 messages of large code case isn’t surprising to be gone within a few minutes

Same with plus - 250/1500 messages aren’t an awful lot, and it explains why within an hour of subagents you can go 20%

[boreal holly]

0.104 works fine for me. Thanks for your recommendation though

[main nimbus]

Updating is a risk atp

[boreal holly]

Well no, I'm not reporting an issue here. The only reason I'm using 0.104 is because I think I needed to upgrade to use Spark or 5.3, can't remember which. Otherwise I would still be on 0.94, because the current version doesn't let you go -c features.skills=false which is pretty lame for automations. Was even considering maintaining two separate copies of codex just to be able to disable all skills with a single flag

But on Plus, that 5 hr limit goes by so fast. I think when a lot of people discuss Plus versus Pro, see the 6x usage at 10x cost, the real important increase is the 5 hr limit

I do think everyone's rate limits are dropping precipitously due to the subagents feature so that's the real issue they need to work on

[torpid trout]

What I mean is, I suspected the CLI version to be responsible for the sudden much higher consumption of credits, but since you experience the same under 104, that cant be, since under 104 last week (begin. of) I did not consume that many credits

[torpid trout]

I reported it, with the specific fact of the subagents (which is clearly provable), but it was dismissed as "expected"
They simply ignore the fact that before a certain moment in time, even with subagents credits did not go so fast

So hence my guess that initially they counted credits wrong, and now count them as they actually intend to, thus the drop

45/250 messages is nothing really, and in the begin I was very pleased as of how much I can factually do with it (a full week of work!)

And now plus is gone within hours, pro within ... 2 days, if I keep using agents (stopped using them now, which uses significantly less % of coures)

I mean, be it what it is - in neither scenario I will be able to use this product if that is what we can do under 2x usage, because it will mean a cost of about 1000 USD per month at least under 1x usage policy - if used full on under the current % consumption, and for that money its not worth it

[boreal holly]

Gotcha, that makes sense! Maybe they were counting the credits wrong! I noticed when my Spark limit hit 0% it suddenly showed 95%, but if I went to use it I would get "you have 0% spark remaining" errors. So the counters are acting kinda funny. The regular usage limit, to me, appears to be expected. I just kinda ignore the spark meter and figure OAI will fix that some time.

I used Plus in August/September and when it was me and 1 agent collaboratively writing code it was cool - never really reached any limits, but was getting less done than if I was writing all the code myself. When I started building frameworks for concurrent agents that's when I started quickly hitting the 5 hour limits and got a pro account. All problems melted away at that point. This week I scaled up my usage to gauge how much usage I can get with 12 agents working 24 hours, and I got about 5 1/2 days on pro, used 1000 credits in 1 hour, and then used 30% weekly in 45 mins on Plus. Anyways, hitting that 5 hour limit reminded me of how useful Pro plan is, and I think it's just a coincidence I did this little experiment while everyone's been getting rate limited hard. I still think Pro is the best value and have been paying attention to the issues. I think it all stems from quickly testing new official features

[velvet wren]

probably best to delete this

[void cedar]

Thank you! Is that against TOS? I'd rather just uninstall then I thought it was just patching a windows issue or something

[velvet wren]

yes it is against the ToS

the Windows version of Codex app has not been released publicly yet

[void cedar]

That's what I thought, but it I can download it and run it from the Windows store, following the instructions on the open AI website, it just says I need to update the app.

I suppose that just means it's not for me yet lol

[nimble terrace]

is the gpt 5.4 being seen in commits just hype from openAI or actual mistakes from the team lmfao

[void cedar]

Probably hype lol

[digital rapids]

Anyone having "Do you want to run this command?" repeat issue? For the last few days, on Windows VSCode Codex keep asking this even if I select "Yes and don't ask...."

[torpid trout]

are you sure it is the exact same command?
It uses to ask even if there is a nuance differing.

[digital rapids]

It differs but it wasn't asking as frequently as of now. Giving a task now takes around 20-30 "Yes and don't ask.... " selection.

[boreal holly]

In sandbox_mode = "workspace-write", if the command is git and involves mutations (e.g git commit), it will 100% always ask. If the command uses the network in any way (e.g. curl) and you do not have network_access = true then it will always ask 100% of the time for approval.

[supple perch]

Update the new naming is nice

[supple perch]

Hello btw

[boreal holly]

Here's some good sane defaults:

sandbox_mode = "workspace-write"
network_access = true
writable_roots = ["C:\\path\\to\\cache\\folders", "C:\\for\\example\\.pubcache", "C:\\or\\global\\node_modules", "C:\\var\\run\\docker.sock"] # or whatever Windows paths look like idk
approval_policy = "on-request"

Try that out. If it fixes it then hooray! Otherwise need way more details

[boreal holly]

Whaddap 😎

[cunning olive]

Sup

[supple perch]

(or the agents runner sandbox would be nice and do this for them =P)

[supple perch]

Hope you are well Mr Robert!

[boreal holly]

Ah you got windows support?

[supple perch]

I mean... docker commands are the same

and wsl should be able to build it...

I did add "IDE in container" for even better sandboxing

[digital rapids]

Just added that into the toml settings and it stopped asking. You are awesome! Thanks a lot

[boreal holly]

disclamer: network_access = true is considered risky. don't ask it to browse the open web with curl or whatever windows powershell uses and should be fine

[supple perch]

I mean or do

JK JK JK dont

[boreal holly]

I use it as an alternative to Postman/Insomnia or whatever lol Codex is just so much faster at smoke tests

[digital rapids]

Also removed that and run a few prompts and it still works flawlessly

[modern tide]

welp, i'm still stuck on upgrade required

[orchid plume]

I believe this comes up if your account hasn't been invited to beta test it. You can bypass it with some trickery but that's at your own risk/peril. When I last tried it, it didn't seem as complete compared to the MacOS .dmg variant, so I'd honestly either suggest sticking with converting the .dmg image for now to a Windows compatible version, or wait until the official release at some point in the future

[modern tide]

well crap... do i do it from the cmd if i want non-web version for windows?

[orchid plume]

there are some repos on GitHub which can convert the .dmg MacOS image to a Windows compatible one (using Powershell). I won't link them here in case it's against the rules, but should easily able to find them with the right keywords 🙂

alternatively use Codex CLI

[modern tide]

is it worth using over cli?

honestly why downgrade web version is beyond my understanding

[orchid plume]

depends what you want to use it for

CLI doesn't have automations for example

[modern tide]

well i used web version to make quick updates for project on git

[orchid plume]

hmm I see, web version will consume your quota faster so unless you need quick and easy remote access to Codex, you can just use the CLI to do most things still imo

another alternative is OpenCode, which has a CLI, desktop app and web interface options

[modern tide]

well i didn't care about quota at that point, it used to be enough for what i need but now it can't finish the task really, like 3 times dumber than it was

[orchid plume]

I see

[hard tulip]

how can i use the .dmg one on windows i keep seing peopel say that

[orchid plume]

search for codex windows on GitHub, for me I use the first result. I had to have Codex CLI slightly modify the script though as it didn't quite suit my system for some reason, but that was easy enough

it's a powershell script which extracts the .dmg file and ultimately makes it compatible with Windows

[hard tulip]

oh ok let me try it out

[latent tiger]

real accidental leak lmao

There was a PR that slipped in "fast mode for GPT-5.4" plus a deleted screenshot from showing GPT-5.4 in the actual Codex app model selector Community accounts spotted alpha gpt-5.4 in the public /models endpoint too

[sleek spoke]

they may well be deliberate teases

[orchid plume]

other companies have been known to do similar kinds of leaks, which ultimately end up being indirect teases

I suspect these are deliberate given how many have occurred

what I'm wondering is how /fast will work for Pro plans, given that they are already meant to be using priority tier

one possibility is that they intend to make is so that Pro by default is not priority tier unless /fast is enabled. When disabled it could consume the quota slower compared to original 1x consumption

[cedar skiff]

They probably arent gong to bother with deliberate teases, they just going to move forward with creating the next iteration. Which is going to be 5.4.

[simple star]

Codex v0.107 just landed

[torpid trout]

A reminder that you need to be sure to understand what you let codex write

Codex is a lazy, extremely negligent developer
As an example, it will happily ship your "production ready" docker images using npm run dev, and if you ask it to fix that, it will after some murking use npm run preview.

It will never ever unless you hint it actually use what it should (for production): a real static fileserver

This is why things like Huntarr happen.
I only discovered it in my current project because I learned about this in past.

[wide schooner]

when fix

[torpid trout]

(Note, I have codex run externsive security checks which bring up real things, but this one was not even remotely mentioned in the report, and asking it to fix was really like trying to convince someone that "A is better, when you used B your whole life" lol)

[wide schooner]

lol codex is the least negligent out of ALL

[nimble terrace]

i wonder how big gpt 5.4 would be compared to 5.3

[torpid trout]

yeah, out of all gpt's maybe, and out of many many devs too

[nimble terrace]

I heard a rumor it's a 2 million context window

if true, insane

[latent tiger]

im seeing 2M context window rumors as well but i dont bleive it, I know deepseek V4 is launching with 1M context window

i will be happy with 1M on codex 5.4

[simple star]

GPT-7 confirmed

[latent tiger]

biggest 5.4 leak is the full resolution vision (the big PR #13050 drop)

there's also a new fast/priority tier (PR #13212, added yesterday)

probably for pro users

https://github.com/openai/codex/pull/13212

[cedar skiff]

I'd be happy with 500k window and the same performance

if they could do 2 million with out it turning into garbage i would expect they would roll it out bit by bit and keep some in the barrel for extra release cred later.

[latent tiger]

claude opus has had 1m context window for a while now and if DeepSeek is dropping with it too I don't see why they wouldn't at the very least match it

I'm sure their internal models already have been trained and capable to handle larger context windows without drift

[cedar skiff]

claude struggles to stay coherent within the 200k, i can't imagine 1m being useful for much with it.

[simple star]

40 more minutes.... 💦 💦 💦

Today was tough

[orchid plume]

I'm sceptical about the 2M context rumours, maybe it's 500k or 1M, but even then I'm sceptical. I do believe that GPT-5.4 is real though, and what we've been seeing aren't accidental typos. I have a hunch we'll see a new model on Thursday evening (UK time)

[boreal holly]

Need 2 billion token context 🧠

Better yet, ♾️ context window, and uses an entire storage device as a KV Cache so it remembers everything at all times

[supple perch]

lol

[pseudo flame]

irony: vscode plugin kilo.ai setup with openai just fine. next i try to install codex plugin setup, openai token failure. cannot setup.

[tall zodiac]

@ivory zodiac you

Do you happen to have this documented anywhere with your exact file 👁️👁️ https://x.com/llmjunky/status/2028544096667648391?s=46&t=9l1KwyEX9uwBxqfL3uX3zA

[hushed storm]

5.4

well tibo also leaked it

[latent tiger]

deepseek v4 dropping this week

OAi probably gonna hit them with the same time release just like they did to slopus 4.6 😂

[fair edge]

anyone got experience with 503 Service Unavailable: no healthy upstream?

[wraith temple]

anyone else getting Unexpected status 503 Service Unavailable: no healthy upstream, url: https://chatgpt.com/backend-api/codex/responses?

[fair edge]

yep

ok so its not just me

[wraith temple]

they still didn't update this

[warm pilot]

if they cracked 2M, I guess it will be pro only. Maybe 500k/1M for non-pro. I'm still looking forward to basically no compaction anymore for small-middle issues 🙂

[warped pine]

yeah some connection issue for me too

[wraith temple]

codex back online

[main nimbus]

Not likely given long context performance sucks on all 1M models

[eternal quartz]

Yes

[boreal holly]

I would actually prefer if they didn't touch the context/compaction. It's perfect the way it is. Perhaps add a knob that's like "sliding window user messages" that let X number of user messages stay in the window unsullied and anything older get compacted. That'd make it so you just have one agent and never need a new thread again

[supple perch]

(was making a joke sorry)

[boreal holly]

"might i interject for one moment please?" 😂

[supple perch]

"can you factcheck this for me, I want my context window to be 20k then compact"

[boreal holly]

See, sliding window compaction would make that infallible! I have convos where every compaction I start with like 70% remaining because all those user messages get preserved and I have to think about starting a new convo. Sliding window would be awesome

[supple perch]

*now you see why I made that statment back in Jan?

about wanting to compact at 80% remaining? It does so much better when you closer to the models true context window (32k all models)

[boreal holly]

Yep, none of that RoPE or Yarn hackery 😎

[cedar skiff]

you lose importnant front loaded data though

[boreal holly]

I'm not saying make it a global thing like "press esc x3 to archive the chat" lol or whatever cool UX stuff they come up with next, I'm saying add the knob. I have agents with no important front loaded data who essentially act as medium-term memory agents. I don't need them to remember what some agent was working on 2 weeks ago, just what the current agents are working on right now

[cedar skiff]

RIght but it's not infallible 😛

[boreal holly]

Ok fair point! OpenAI did turn AGENTS.md hot-reloading back on, and -c developer_instructions="..." is hot reloadable, so that'd probably be the way to keep front loaded info from compaction

[hushed storm]

https://x.com/OpenAIDevs/status/2028577643113922944?s=20

[simple star]

<cries in linux>

[simple star]

🤨

[orchid plume]

"soon" 😂 - this reminds me of the jokes about "Valve time" regarding when Valve get round to doing something (or not)

[simple star]

2 more weeks

Jesus, 96% of quota already after 1 hour... 😱

This is gonna be... tough

[ivory zodiac]

It's new actually. I'll push it later

[flint isle]

Hey does anyone know how to open multiple codex chat windows at the same time in vscode?

[cedar skiff]

open a terminal in vscode window then type codex do that as many times as you like.

[flint isle]

thanks for the reply, I meant in the codex extension on vscode

Claude extension allows multiple chat windows, but not sure how to do that with the codex extension

[frosty zealot]

https://github.com/openai/codex/pull/13212

Is this the same as /fast for say Opus 4.6?

Or what exactly would this do

[tribal lake]

GPT-5.4 coming? But we didn't have general 5.3, only codex version... or did we?

[cedar skiff]

is anyone elses pr reviews taking a really ling time to come through?

[sleek spoke]

👀

[dawn seal]

hello first time using codex cloud, it did some coding but I cant find it

where is it? it says it commited to my git but I see nothing

[cedar skiff]

is it in a branch?

[dawn seal]

I do not see any branch at all though

I used a cloud workspace and the work is done somehow? but I cannot touch it

[cedar skiff]

list remote branches git branch -r it's probably there

[dawn seal]

D:\Projects Unity\20251010_MassiveMultiplayer>git branch -r
origin/HEAD -> origin/main
origin/main

D:\Projects Unity\20251010_MassiveMultiplayer>

oh I just received a pull request

[raw hill]

all codex cloud results in pull requests

pull requests will use the branch it creates

[dawn seal]

it's odd it commited but also requested pull

[raw hill]

you have to commit prior to the pull

[dawn seal]

it commited on the main branch then requested pull from other branch

[raw hill]

ahh yeah thats weird

[dawn seal]

yeah

[raw hill]

the AI likes to sneak little hotfixes into main without PR

[dawn seal]

I think it might be just internet

I think it did commit to main branch and pushed and fixed the bug like maybe my UI didnt update

i dont know

its weird

[desert vapor]

Who uses an Agent builder like Vertex or AIP?

[cedar skiff]

is this korean?

[cyan wing]

noo what did i do 😭

[cedar skiff]

ok and then i get this?

stream disconnected before completion: This user's access to gpt-5.3-codex-premium-1p-codexswic-ev3 has been temporarily limited for potentially suspicious activity related to cybersecurity. Learn more about our safety mitigations: https://platform.openai.com/docs/guides/safety-checks/cybersecurity

All i am doing is fixing up unit tests?!

[cedar skiff]

i got the same thing at the same time that has to be a mistake right

[cyan wing]

yeah must be

thank goodness, not the only one

[cedar skiff]

Lets hope it's many ppl getting the same thing, so it's a quick fix

[astral pine]

I just got it too lol. Using 5.2 medium

[hard tulip]

why are you even using 5.2 anymore 😭

[astral pine]

Must have been a server hiccup.

[cedar skiff]

5.2 is smarter

[hard tulip]

no

its not

like what

[astral pine]

Its defaulted to that cause Im using it for agentic assistant tasks most of the time. not coding so prioritizing decision making vs coding most of the time.

[spiral moat]

■ stream disconnected before completion: This user's access to gpt-5.4-ab-arm1-1020-1p-codexswic-ev3 has been temporari
ly limited for potentially suspicious activity related to cybersecurity. Learn more about our safety mitigations:
https://platform.openai.com/docs/guides/safety-checks/cybersecurity

mine says 5.4???

[ember venture]

I thought it was odd it returned me a single emoji in a response

[hard tulip]

dont look into it 😂

[cyan wing]

you were on an AB test, they leaked the API in the error log lool

[spiral moat]

incredible

[hard tulip]

you look like gpt-5.4-ab-arm1-1020-1p-codexswic-ev3 @cyan wing

[cedar skiff]

I just told it to carry on and it went forward normally

[hard tulip]

send a screen shot

[cyan wing]

can u screenshot that for proof pls 👀

[spiral moat]

Running some benchmarks with local models, not doing anything cyber related

[hard tulip]

@sama fix your servers

https://x.com/zats/status/2028643947996741657?s=20

[cedar skiff]

everyone getting hit with it, just tell the agent to carry on i think

still using 5.3

[cyan wing]

did u notice any difference in performance of it? 🙂

[spiral moat]

Hah not really, but it's just monitoring some long running tasks so not a great show of intelligence

[astral pine]

6k lines of code? wut?! https://x.com/liam06972452/status/2028603250518171826?s=46

[potent mason]

Ugh translating a website to another language is so annoying

I wish it would just spawn 12 subagents to get them to do it😂

[true bane]

hmm my weekly usage page says 100% and been doing stuff for the last hour

[simple star]

@true bane Limits have been reset, due to technical problems

https://x.com/thsottiaux/status/2028649088594436225

[potent mason]

Bruh that's awesome on their part they where down 8 minutes and replaced it

[cedar skiff]

It is a rolling window though. so they are just moving its position in some cases if you were doing extra usage you might be better off.

[potent mason]

Not really if you were using extra usage now you have usage for the next week instead of having to wait 3 days and then having that weeks usage.

Even if you had only spent 5% you're still getting that 5% assuming you can get to 100%

[cedar skiff]

Except thats exactly what i just implied.

[potent mason]

No I'm saying theres no scenario where you don't benefit from this you implied their are scenarios were it got worse

[cedar skiff]

No i never said it's worse, I'm saying for the most part you arent gaining much.

If you managed your usage to last a week you gain nothing

[potent mason]

You do get something instead of using about 420% in the month you might be able to get 450 or 480

[cedar skiff]

it's a rolling window though, you are bottling it up into a month idea, its not a month idea, it rolls forever.

[potent mason]

It doesn't roll forever you pay for a month so by getting it reset you essentially get more $ out of this sub

[boreal holly]

I used like 13% and it went back to 100% so I got a days worth back. Can't complain!

[cedar skiff]

You still get the same amount of usage over time if you managed your usage for the daily allocation. The only differences is a sociological expectation.

windows dont reset when your month does

[boreal holly]

The only time you waste usage is if it's at 100% and you don't use it. Or you never hit 0%. It starts counting 7 days from the moment you use it at 100%, so if you want your rate limit to reset every tuesday at 8am you let it sit at 100% and use it at exactly 8am on tuesday, and always use it the instant it is 8am on tuesday. No rolling window

[potent mason]

It's not but ok
Really quick example and then I'm done with this topic:
week 1: 100%
week 1.5 (half a week): 70%
Week 2.5: 100%
Week 3.5: 100%
Week 4.2: 70%

Thats 440% compared to 400%

[cedar skiff]

right, but what im saying is the allocation is about 15% per day, if you just stick to that, thats what you still have.

You still just use 15% per day, you dont have any more than you did before.

[potent mason]

You do this month you stop your sub and it resets the moment you resub (if I'm not mistaken)

[cedar skiff]

I'm talking about normal managed usage, not trying to game the system

[boreal holly]

It sits at 100% with no reset time until you use it

[cedar skiff]

Most ppl arent canceling their sub and rebuying to gain a few %

[potent mason]

I usually cancel it the moment I buy it cuz AI moves a too fast to keep a sub

Next month I decide which set of subs I want

[next ore]

anyone else had their usage limit just reset early?

I was at 0% for the week till march 4th but it just reset

new model releases reset weekly limits

Nice!

[rich ridge]

Where is new model

I had 100% too

[next ore]

5.3 something, it popped up early today when I opened the codex app

[potent mason]

No they reset it because services where down for 8 minutes

[next ore]

does this happen often?

online googling says apparently new model releases mean they reset the weekly limit

[potent mason]

Nope

[ivory zodiac]

🙃 🙏 😏

[orchid plume]

the creativity going into the 5.4 leaks is impressive, appreciate the rate limit reset 🙂

[frosty zealot]

https://x.com/thsottiaux/status/2028649088594436225?s=46

[boreal holly]

Nice! All I got was "Congrats on your first automod filter violation!" you got a flippin' reward lol

[frosty zealot]

I think it’s a fishing dm lol

[fallen kettle]

After watching several friends adopt it and encourage me to post, I finally took the time to write nearly 3,000 words on the agentic coding workflow I've spent months developing. I hope it helps you learn something new!

https://medium.com/@btraut/assemble-your-agent-team-fbfb6b8904b2

[hard tulip]

i had my weekly reset today so i was only down like 5% 😭 🥀

[hard tulip]

🤔

[hushed storm]

They are doing 5.4 ab test

Or tried at least

[hard tulip]

yes

[cedar skiff]

Might be a claude model release soon as well.

[hard tulip]

claude 6.7

[cedar skiff]

it's down atm, and the usage page is playing up.

they falling behind, they need something.

[civic oriole]

how do you have access t that

[hard tulip]

they did

[civic oriole]

oih

[hushed storm]

sad

[civic oriole]

very

[hushed storm]

what do you have to do to be a top plus user bruh

[hot egret]

It’s crazy, ever since it was released, despite the updates, I still haven’t been able to open it.

[orchid plume]

I think it's because you have to be invited to use it

[rocky fog]

did the codex limits reset for some reason?
I had like 10% weekly left until 8 march
but now its back to 100% until march 10 heh

[rocky fog]

I see
well I am not complaining

[dark grove]

i chat ts to want to see how much is chat delay

uh

[warm pilot]

https://x.com/thsottiaux/status/2028649088594436225

[wicked briar]

you enjoying claude lately?

is very slow for me

testing it out for UI

[rich trellis]

Good morning. Does anyone have any information on when Plus users will be able to use Codex with full GitHub integration? Currently, lthough it can read repos and generate diffs, it does not seem to be able to push to GitHub.

[velvet wren]

I just use the GitHub Desktop App in conjunction with the Codex Desktop App

[rich trellis]

And you are a Plus user? Ah, are you using macOS or Windows?

[velvet wren]

Plus subscriber and using macOS

[rich trellis]

Thanks. I'm a Windows user so it looks like I'll have to wait a little longer. The workflow using the normal chat interface isn't so bad but it's the GitHub integration that would reallybe nice. Thanks for your help.

[sleek spoke]

anyone think they've got 5.4 in Codex through the a/b test?

[warped pine]

how to check

[orchid plume]

good question, not sure. Apparently they've got it in the web UI (chatgpt) though, for GPT Pro. If you see the thumbs up/down next to the chain of thoughts then apparently it means you're on 5.4 Pro really

[warm pilot]

It can? The same way you push yourself? As long as git push works, codex will use that just fine. The web version lets you directly create a PR (and also pushes before).

[glacial ginkgo]

Working in a Rust codebase with gpt-5.3-codex xhigh I have found that a prompt "Review src/<source>.rs. Fix changes." will run and make changes. But resubmitting the same prompt again always leads to more changes. I changed the prompt to "Review src/<source>.rs. Fix changes. Review again. Keep working until there are no more issues." But it still seems the model never converges on a state that won't keep changing. Has anyone else seen this sort of thing? Is there a pattern or best practice to avoid it?

[deep lily]

F

[brazen kettle]

Does the macos codex app use codex cli in the background? Is it worth using instead of codex cli or does it have less features?

[exotic cave]

Looking at benchmarks alone and Gemini 3.1 is huge jump from Gemini 3.0. I don't use it as my daily model, but it shows Google is here for the long haul. Wonder what OpenAI is working on, because if Google keeps at it they can drive a lot of companies out of the race.

[simple star]

yes, and depends on what you need

[boreal holly]

You need to give it concrete completion criteria. Just saying "build app and make no mistakes, don't respond until app done" is super low entropy and it will never, ever work until it's done. You need to do it in a series of concrete steps:

1. "Review src/<source>.rs, create a list of compile errors, security flaws, gaps in functionality based on what the file is supposed to do, and save a report of your findings to <source>.findings.log"
2. Copy results, paste into chat with this template:
   """
   Here's a copy of your findings permanently in memory, so you will see it as a user message for the remainder of the session:

{<source>.findings.log}

Your completion criteria is to patch all of the items you found in your initial report, write and run unit tests that prove the patches are in place. Your end-of-turn natural language response must include a list of every single finding and a description of what you did to remedy the findings on the list. Do not respond until all items in the findings have associated actions and proof. Example final response:

## Files touched
- ...

## Actions taken
- ...
  - Solves: ...

"""

That's actually all it takes. It could work all day on that. You have to tell it exactly what you expect their final response to be.

[glacial ginkgo]

Yes, good point. I did introduce that process where the code review then wrote the findings to a Markdown file. The next turn worked through each finding, fixed it, and was then done. Unfortunately, then asking for a new code review produced a new set of findings.

I think it is clear that an open ended "Review this code" is too broad. However, that is also where there seems to be to most value for finding unknown unknowns.

[wide schooner]

❤️

codex best

wish more people switched to claude so we get even more quota

[boreal holly]

I gave my agents a $request-review skill and a $finis-hturn skill. The finish turn skill says they must run static code analysis, unit tests, request a review, and if there are findings in review they patch and restart the finish turn process, otherwise stop when done and present final response in a specific format. Request review is done by another agent using codex exec review --json and deserialize the final response (also has a remote code review switch). Anyways finis-hturn skill is probably what you're looking for, if the goal is rigorous and autonomous code reviews

And yes, the hyphen was moved to avoid triggering automod

[inland bramble]

I'm waiting for it... a mid-tier plan... ChatGPT Pro Lite. When will we have it?

[torpid trout]

Here's something that is worrying me, maybe someone can shed light

1. I had a branch locally with a file scan.sh which was added to gitignore
2. I commited and pushed/merged that branch into develop, then deleted it both locally and remotely
3. This effectively removed scan.sh from my entire codebase no matter what
4. I then promoted codex in a NEW cli chat to implement CODE TEST COVERAGE
5. It said a few things and amongst them...

• Explored
  └ Read scan.sh
....
....
I found there is no scan.sh in the current tree right now, so I am switching to building a robust local pytest + coverage setup that
  Sonar can consume (coverage.xml) and then validating the percentage locally.

Is this a very unfortunate coincidence or... this looks like memory to me, yet I have memory turned off

scan.sh is by no means something typical either for tests nor Sonarqube, it was literally a temp file I used to ... well, yes, scan with my sonarqube instance manually
Codex was aware of that file during that time, but as said, branch is gone, chat is new, memory is off.

[boreal holly]

if scan.sh was ever committed, then it is permanently a part of the git repo. The only way to truly delete it is git filter-repo, which removes all of the deltas of that file throughout history. Codex will always be able to find and restore that file otherwise

This also means force pushing to master btw so be aware of that

[torpid trout]

that's the thing, it was never commited to start with
I added it first to gitignore, run scans, fixed, commited and pushed, and that effectively never added scan.sh

the only literal reference to it is in gitignore file

(because actually I intended the file to persist but forgot that a commit > merge would remove it since it is only on that branch lol)

From what I see, it EITHER read the gitignore and somehow for whatever reason also decided to read it, OR, it persists memory despite the GPT setting, OR, it is an accident and it assumed a scan sh file to exist for whatever reason
Neither looks very reassuring to me 😄

[boreal holly]

I mean when it says "* Read scan.sh", that means it either executed /path/to/shell -lc "cat shell.sh" or /path/to/shell -lc "sed ... shell.sh", and the green dot indicates successful read. So the file is somewhere in your repo

[torpid trout]

PS here should be proof enough:

bedas@apollo project_name % git log --all --oneline -- scan.sh
bedas@apollo project_name % 

[ivory zodiac]

LOLOL

[torpid trout]

but also see I found there is no scan.sh in the current tree right now

[ivory zodiac]

Best Potty Mouth

[torpid trout]

• Explored
  └ Read scan.sh
    List ls -la
    List rg --files
    List ls -la
    Search scan|fetch|hotspot
    Read sonar-project.properties

──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────

• I found there is no scan.sh in the current tree right now, so I am switching to building a robust local pytest + coverage setup that
  Sonar can consume (coverage.xml) and then validating the percentage locally.

• Explored
  └ Read requirements.txt, README.md
    Search pytest|unittest|coverage|tox|nox|conftest|tests in .

Those dots are grey on the cli

So... it did NOT find them?

[boreal holly]

Try git log --all --full-history -- "**/scan.sh" so it's recursive

[torpid trout]

bedas@apollo project_name % git log --all --full-history -- "**/scan.sh"
bedas@apollo project_name % 

yeah, as epxected. Also gitkraken does not show it, I am 100% positive I made sure to add it first to gitignore for some reasons (hello access keys...)

[boreal holly]

What the heck lol they oughta fix the Read to be dark red or something so it doesn't look like it successfuly read an existing file

[torpid trout]

also, note that in the READ thingy, one of those files does exist on my repo - so that one it can read (the sonar-project.properties file)
This is certainly weird.

I will see once it is done in a new chat if it repeats the same stuff.

[boreal holly]

Ohhhh you know what, sometimes codex likes to do cat somefile || true which shows up successfully in the UI. Codex does not like it when a command exits non-zero

They pretend like everything has set -euo pipefail and so they add || true to a lot of commands, when in reality it's not like that by default

[torpid trout]

right, but... why scan.sh? Just out of the blue a big coincidence?

[boreal holly]

If I'm not mistaken, codex CLI has git status hints in the context. I've read somewhere, perhaps in app, if you have hundreds of changed files in git it complains that this affects performance or something. Maybe it saw a stale git status in its context, carried over by the CLI and investigated?

[torpid trout]

... which somehow translates to cross-chat memory, something I thought turning "memory" off would be off

Anyway, I run a new chat, same prompt, it does not read or try to find that file anymore

This is a weird one.
Fishy. Kind of a "ups I slipped something I should not know"

[boreal badge]

🐁

[mellow smelt]

I just wanted to share two points of kudos with the Codex team if they monitor this channel.

1. Thanks for still having a password login. Anthropic didn't have this. Even if
   you layer it with TOTP or whatever, I still enjoy not having a single point
   of failure with my gmail account.

2. And the second is allowing OAuth access from third party tools. I use
   https://github.com/editor-code-assistant/eca with emacs myself.

   I believe there ought to be a(n) [AI-assisted?] way to differentiate those
   who abuse oauth from those who simply want to improve developer tooling. It's
   the reason why I ended up paying for codex over claude code or gemini cli.
   The ability for solo devs/power users to build upon their tooling without
   needing to pay the same production rate for API's might be a smaller customer
   segment, but one that I think lends to much vitality in mindshare over the
   long run.

Frankly, the difficulty of my work means any frontier model can hack it.
Choosing between commodities then means Developer UX was the differentiator for
me. Keep up the good work.

[rustic pine]

I want to "Plan" with 5.2 ... but "Implement" with 5.3-codex ... but after you run through a plan session with 5.2... it ends with a nice ask, "Do you want to implement this now?" ... but I just cancel out of that, then switch to the codex model and then ask it to implement it... am I doing this right? 😄

[rigid meadow]

love it when it randomly switches languages

[wide schooner]

why did they add monthly quota instead of weekly that sucks

[orchid plume]

monthly? I don't see a monthly quota. I see a 5 hour quota and weekly quota

[cyan wing]

I think that's a translation issue

[wide schooner]

oh

🙏

i was scared

the quota so good i didnt realise

[cyan wing]

Anyone have a tmux automation for:

• auto-select Plan Mode with (Recommended)
• auto-submit Implement plan

[raven kettle]

Why is the client macOS-only? Weird platform choice.

I'm scratching my head over this as Macs are luxury devices and most engineers don't run Macs.

[sleek spoke]

OpenAI just posted on their Threads account that the Windows app is coming tomorrow

[plucky halo]

Woah. Thought I'd gone back in time to 2012

[cedar bear]

Fr?
Why posting it only on Threads tho

[sleek spoke]

Yes FR, and no idea lol

maybe Windows users use Threads instead of X......

[cedar bear]

Link?

[sleek spoke]

https://www.threads.com/@openai

[cedar bear]

Oh lol

[raven kettle]

Sam Altman gets into arguments with Elongated Muskrat.

I guess he's friendlier with Zuck?

[latent tiger]

if we get 5.4 before windows app im gonna laugh

[meager wadi]

No wonder the errors we got recently were mentioning 5.4 (in reply to recent message in announcements)

[snow hornet]

5.4 today?