codex-discussions

[orchid plume]

well, not unchecked, but likely more on the side of vibe coding

[cedar skiff]

i'd think every line is reviewed by a person

[cobalt junco]

u gotta be stupid to fork the leaked claude src code

[nocturne folio]

openai please say the magic words, please keep 2x permanent

or maybe 4x

anthropic increases usage based on model drops

its about to be

[exotic cave]

Wait till it's gone i think it's up till the 2nd.

[hard drum]

Who lobotomised the GPT models? Suddenyl they're all being lazy arseholes

[potent mason]

Oh good they got some more funding they should phaze us out of 2x slowly

[cedar skiff]

they folding code reviews into main usage and dropping back double usage, next will be dropping spark off the usage model etc etc

[hard tulip]

$122B rased give us codex 10X usage for the next 100 years 😭

[cedar skiff]

hopefully we see something good out of turboquant

[cyan wing]

Is there no way to set the subagent model in the config.toml?!

I'm trying to use gpt-5.4-mini for subagents..

[cedar skiff]

did you ask 5.4 about it? because it's probably making up the answer

[worthy furnace]

Are you kidding me? I had full 100% 5h limit and weekly limit 10 minutes ago, and all my 5 hour limit went to zero and my weekly to 70%

How is that even possible?

Yeah so annoying

Do we get a usage refund now or have to wait?

[cyan wing]

yes i asked it 🙃

[cobalt junco]

use orch systems

if you facing usage drainage, use orch systems like ruflo for context management and offloading light tasks

has a higher success rate

and cuts usage by like 30-50% in my cases

[lucid mason]

Starting a new Codex in VS Code chat, asking it to read all relevant docs. Now I can resume work on the code. 😄

I read the required docs, benchmark-store files, and maintained PowerPacer.* scripts, then verified the live package state against the code.
Somewhere around 500k the behavior starts to become more erratic, though, and around 700k is becomes mostly useless. So the 1M context window is never fully usable anyway.

[cedar skiff]

codex always tries to steel man keeping stuff and adding complexity

[tender stream]

Thx 🙏🏻

[cyan wing]

Codex wants to support the legacy API for the downstream users
who don't exist 🙃

[tender stream]

Lmao

[cedar skiff]

not even that just edge cases that dont matter

I worked out to tell it to steel man the removal

[lucid mason]

I wouldn't mind compacting or older "memory" getting lost. But I repeatedly reach a point where even the most current chat input isn't answered sensible anymore. Like GPT not looking at a file I just uploaded and then stating old numbers instead of new ones. Or it outright stating that I didn't upload anything until I question that answer.

It really becomes unable to even handle new chat input at one point. I also had multiple instances where it repeated an older answer to my new chat input, ignoring what I just wrote. So it doesn't lose old data, it loses new data right away instead.

[cedar skiff]

is there a 1 million context window?

[uncut ridge]

hello

[lucid mason]

I just posted the corresponding screenshot. So at least theoretically, yes.

[uncut ridge]

hello

[lucid mason]

You set it via config.toml.

[uncut ridge]

hello

i have a question

important

for

me

[cedar skiff]

Oh right i had no idea, i guess they limit it to the context they think performs well

[uncut ridge]

hello

[rain kettle]

4 small-medium tasks done with the use of subagents (yesterday) vs 25 big tasks done without subagents (today). Lesson learned: subagents should be used only when the task really requires their help

[feral geyser]

Straight from the model's mouth!! What do you all think? I looked into some of the risk with connecting third party plugins -- the bit about sharing recent "context and intent" should raise caution for any businesses or entrepreneurs because the ambiguity of what constitutes these terms 'context and intent' is extremely flexible and thus as a means of consumer right and protection, it may fall short here. What do you all think?

Imagine you are using Codex to roll out your new FinTech related tool. Oops, Codex just shared all your business strategy and plans in an attempt to "help Stripe perform", all because you were recently discussing these subjects and they got categorized as "context and intent".

Curious to hear everyone's thoughts on this -- including the mods!

[nocturne folio]

gpt 5.4 is like both so good and so bad, it keeps hallucinating frontend requests, but will go on and create ground breaking discoveries in the backend

[cedar skiff]

probably not going to open a pdf from a random discord chat o.0

let me ask chatgpt if i should do it and then put the answer in a pdf and paste it in here

[royal mirage]

chatgpt down?

[boreal holly]

I got these suckas so dialed in they dont even listen to me anymore 😂

[royal mirage]

anyone else? can check please

[cedar skiff]

classic

If bro can miss-understand bro will miss-understand

It's like an internal directive

[nocturne folio]

i hope nothing possesses openai to increase the price of their next models

dude 2$ and 10$ was enough

uggghhh and the crappy cost increase for going over 200k tokens

[cedar skiff]

i wonder if the have a ten trillion param model in the works like anthropic

[boreal holly]

This is pretty awesome. 5.4 knows if a message is not prefixed with [End of Turn], [Approval Request] or [Some-Agent-Name] it's a user message, but they've been running for so long they just act like user messages are the devil 😈 set in their goal of building the perfect software. They will not rest until the whole thing is built and perfected

Gotta have the authority to talk to these digital folks

[wintry sonnet]

damn, plus plan is definitly not enaugh for developer task

the 7 week limits can be used up in just only one days

[bright geode]

anyone know for the claude code codex plugin how you would run on fast mode? I think it doesn't have the option, I guess this is something that can be turned on with ~/.codex/config ?

[cyan wing]

excuse me sir, is that Codex on an iOS mobile application?! 👀 🔥

[cyan wing]

and does it sync with your desktop chat sessions
so you can continue them when you're on the go ?!

[cyan wing]

shut_up_and_take_my_money.gif

[boreal holly]

i mean i like chillin in the couch while the robots are killin it

[cyan wing]

idk why Codex team is not making mobile handoff natively part of Codex already

[boreal holly]

a bridge server that caches messages and brokers communication makes it all possible

[cyan wing]

ya i made one of those, but posting to the app store is soo annoying 😩

and i made it with Expo so I have to deal with all the development server / expo app bs

[boreal holly]

the only concession i made is I do not use Apple’s network stack. they impose an arbitrary 1MB websocket limit that I discovered the hard way.

Codex sends huge websockets

[cyan wing]

ya prolly want the bridge server to only show latest N messages in a thread
and don't need to send all the response output of a command / tool call
bad for battery life + network usage

[boreal holly]

it shows 50 messages, but optionally can read the entire chat history

[cyan wing]

And does it support $skills?
plan mode, model switching, and reasoning level?

[boreal holly]

it supports all but plan mode. Skills are a must, certain roles have model and reasoning. most of em 5.4 medium and 5.4-mini high. but plan mode is for me

[tender stream]

I installed fedora....... I did this all night...

[deft gyro]

cool

[cedar skiff]

once i got a mac i never looked at windows again

[deft gyro]

good 4 u

[deft gyro]

skills?

I dont use too many tbh

just like

web design

[cedar skiff]

i kept a gaming pc for a while, but then i replaced gaming with claude code and codex o.0

[cedar skiff]

sold the gaming pc before the ram hit a high

is it all local data or do you run a server?

[deft gyro]

robdex orchestrator skill

I forgot u have like 20 million agents

I have prompt optimizer skill

web design skill

and the default codex skills

idk

all the github skills

imagegen

[cedar skiff]

most of my skills are topical

[deft gyro]

I feel you dont need that many

I should probably remove some of them

I dont need imagegen

just context bloat

[boreal holly]

thats their communication system, so they can coordinate with each other. The name is misleading

[deft gyro]

ah ic

do you just have like

a markdown file

that they write to

or is it something special

[boreal holly]

The orchestrator is the memory system. Workers do their work and get archived

[deft gyro]

(i have never done agent orchestration b4)

seems like it would be a mess to me

to have multiple llms touching the same part of a project

[cedar skiff]

codex has openai.yaml you can set like this so the front matter is excluded from the context and they are only manually invokable.

policy:
  allow_implicit_invocation: false

[deft gyro]

yeah I know

but sometimes I dont ask for it to use X skill

and I just want it to be able to call it on its own

[cedar skiff]

right but you said about context bloat

[deft gyro]

yes

like everything

its a balance

oh wait is that for individual skills

I might have to do that

im a lil slow

[cedar skiff]

yeah but this wont be in the context so you can have your cake and eat it too

yeah ahha

[celest patrol]

anyone knows how to fix this. im upgrading from plus to pro but i can't get it approved.

the support is useless as well. and i know it's not to do with my card (i tried 2 cards)

[deft gyro]

are you sure about that

[cedar skiff]

get a third card

[deft gyro]

vro

i swear

automod

no fun

[boreal holly]

The orchestrator keeps them from cross cutting edits. The workers have their slice and thats their entire life. QA pilots the apps in a simulator and report bugs, the orchestrator spawns agents in response to the bugs and aggressively reviews their work

I just speak with the orchestrator

[deft gyro]

ah ic makes some sense

[celest patrol]

ill try it 😂

[deft gyro]

but how much does it cost tho

[boreal holly]

its dialed in for efficiency

[cedar skiff]

Do you just install the app via xcode?

[deft gyro]

I need to upgrade to pro soooo bad

holy

I ran out of weekly yesterday

[cedar skiff]

get a cheap model for the easy work

glm or minimax

[deft gyro]

I use my gemini sub for the easy work

and the really hard work

cos AG has co46

[boreal holly]

Yeah i have some entitlements and capabilities that Apple does not like on the app store so i go directly from Xcode

[deft gyro]

ong i hate apple

apple is such an opp

but I still use iphone

im just complaining really

the app store restrictions SUCK so bad

[boreal holly]

it grows on you lol I love Linux but I also like a machine that has good battery life and stuff so mac is a good alternative

[deft gyro]

idk

my laptop gets 12h with light work

like

normal dev work

[whole gate]

If I put a config.toml that configures access to my app's MCP server will the Codex Cloud agent use it?

Or will only the local CLI use it?

It would be exceptionally useful for Codex in the cloud to be able to access my MCP server that exposes database schema details and RPC reflection metadata.

Currently, I'm using Claude for planning because it is easy to configure this in the Claude desktop client

[oak trellis]

back here .. cc got me rekt ..

[whole gate]

Useless Dev hey bud haha

[oak trellis]

long time haha

[whole gate]

RIght?

[oak trellis]

yes horrible

i used /resume .. 4k tokens used 13% of my session

unbelievable .. its like i have a 20 usd sub ..

one member said its like 100->20 .. 200->100

and codex fixing so much code issues i had with opus too

i hope openai releases a 100 usd sub ..

[boreal holly]

Codex some else’s computer Cloud is only able to use tools inside the repository

[deft gyro]

I mean codex sub is about to get nuked

so like

the halvening

istg

gpt 5.4 is TRYING to ragebait me

why does it not

read the damn files

[whole gate]

Ok, so if you tell it specifically to connect to an MCP server in Codex Cloud it will try. I think I just need to correct my token auth flow to handle it.

This won't work in this particular scenario, but I ran a few more additional prompts and it definitely will use a configured MCP server in the Codex Cloud. I got it a step closer, but I might need to adjust my auth flow to use a true bearer token, currently my auth flow is set up to require interactive access token from Microsoft OAuth provider.

[whole gate]

In effect, it's basically saying if I turn on DCR it will just work.

[chrome raven]

happy reset

[orchid plume]

the codex team are amazing, they didn't have to reset but they did so for two reasons 👍

[eternal quartz]

wdym fraudulent

Times of generous codex are over and now we increasingly getting thowards claude treatment?

[orchid plume]

well we don't know what will happen on or just after April 2nd. There's changes to the Pro plan in the pipeline, just hidden right now. Pro 5x (new $100/mo plan) and Pro 20x (replaces current Pro). If Plus's quota goes back to what it originally was and they release the new Pro plans then Pro 20x should effectively be like the bonus 2x rate wasn't turned off

[tawny island]

can you s t f. u?

[tawny island]

the amount of b. itching in "llm" community is insane.

[eternal quartz]

Wow thats not a nice thing to say

@normal ore

[tawny island]

exactly the kind of behaviour i expected of someone like you

[eternal quartz]

Can someone please ban this toxic guy

[tawny island]

they havent rugpulled anyone you get insane usage forf 200$ month and if you cant make more than 200$ from the usage you get then thats on you

earn money with it and buy more subs to make more money. its not a charity

codex team have not done a single thing sketchy

[orchid plume]

also evading potential automod filters (such as bad language potentially) could land you in hot water, see rule 10 #server-rules

[tawny island]

yes i did and the mods here are not anthropic mods. they agree, dont you think they get sick and tired of everyone being so insanely ungrateful and entitled?

[fast iron]

I love it! Thank you codex, now I can keep building!'

[eager forum]

I felt rate limit consumed faster then before this week. But happy reset!

[supple perch]

@boreal holly turns out gpt 5.4 has a 92k rolling window context, so its best to compress at 80k if you have that setup

[worn harness]

holy another reset

[turbid axle]

is nano any good? as in, usuable as a daily driver?

[worn harness]

depends what you want from it

[wide forge]

Nice! Just noticed

[main wraith]

getting these since yesterday, what does it mean?

[celest stag]

I find you can ask codex to diagnose issues with MCP servers and tell you more information. Hell, I have had it build brand new MCP servers for systems where I knew I had API access just by asking it to. It should be able to explain it.

[worn harness]

hmm, usage bonus is over? burn seems 2x faster

[orchid plume]

seems fine here at the moment

[worn harness]

you have codex cli?

[orchid plume]

using a different harness currently

[stable obsidian]

Anyone else finds 5.3-codex way more reliable than 5.4

Especially when it comes to frontend

[quaint sigil]

how is it for front end been struggling with 5.4

[stable obsidian]

It just doesnt break the frontend, usually when I task it to work on some frontend task the frontend is never broken, just exactly what I asked it to do without any flaws

On 5.4 I always get broken, hallucinated or non-functioning components

[worn harness]

5.4 for my projects at least is more reliable on frontend (if we ignore design)

you're doing JS?

[azure cape]

I'm using CODEX CLI with model 5.4 high in FAST mode in windows WSL, and I'm getting a lot of "fail to apply patch". Almost all file patch execution attempts are failing and the CODEX is going to go back to Python to try to edit files. Anyone else getting the same error?

[boreal holly]

Nano is really good at these things (in terms of cost):

• Classification: Can read large bodies of text and classify it quickly and fairly accurately (e.g. is this command destructive or safe? rm -rf / "destructive").
• Structured Outputs: Pairs with classification - force the model to respond with an enumerated and finite set of values
• Routing: Can determine complexity of a request and upgrade conversation to a smarter model

I've used it for these things with pretty good results. But for actual coding tasks, it very quickly degrades in performance/accuracy. In the app I'm building, I use nano until the user asks for something with medium or high complexity and then it upgrades the conversation. It also detects whether the user is asking for something relevant (e.g. business or just messing around) and downgrades the conversation if it detects chit-chat. For this it works extremely well 🙂 but for coding I think 5.4-mini is legitimately powerful.

[deft gyro]

5.4 mini is only useful for good prompters though

I find 5.4 mini really tries to get away without reading the codebase

and it always fails when it tries not to read the codebase

much more than even 5.4 or 5.3 codex

5.2 my god always reads 10 trillion files before CONSIDERING to start

not saying youre wrong of course

5.4 mini can code

im just saying for the average user

who isn't all that great at prompting

its not the best choice

I think if you really want something done right and you dont know how to prompt the order is

5.2 > 5.3 codex > 5.4 > 5.4 mini

maybe switch 5.4 and 5.3 codex depending on what youre doing

[boreal holly]

I don't make 5.4 mini navigate the whole codebase. 5.4 pieces together the work and passes it down to 5.4-mini with enough detail to carry out implementation. If you tell 5.4-mini where to look beforehand it does a fine job. Sounds wasteful to make 5.4 navigate, but a majority of the quota gets burned up on output tokens and tool call outputs as input. Navigation is relatively cheap, and keeping the smart model as a coherent navigator while the junior model carries out guided implementation keeps the 5.4 agent from degrading as fast

[main nimbus]

robert, you and another mentioned an orchestrator setup. is this just done through agent files in codex?

[boreal holly]

Yeah you can do it with multi_agent=true and setting up agent profiles. But then the agents that get spawned are "sub agents" owned by the parent agent. I have mine set up as "peer agents", normal threads, and give them a tool to pass messages to each other

[vast crater]

Is there any fix or solution to avoid having a very laggy Windows Codex App please?

[odd fog]

I am getting burnt pretty hard by https://github.com/openai/codex/issues/13762 on Windows Codex using WSL integration. The worktree integration is nice but a bit silly it clones to a windows mount...

[vast crater]

Same for me, no matter how I tried to fix it, it still goes through Windows to create worktrees

I tried modifying the environment.toml in .codex-app/environments in WSL2

but no luck on that

[odd fog]

my current fix is i created a symlink between /mnt/c/..../.codex/worktrees onto the wsl filesystem

[surreal wing]

2 prompts

[vast crater]

Also I don't think git folders on Windows is good

[odd fog]

the agent is fine with it, but the UI gets buggy as it can't see the folder

[vast crater]

so I might avoid that, but didn't research further than that

[odd fog]

Yeah it screws with permissions and bitflags on the file

[vast crater]

Yeah

[odd fog]

it's very frustrating

[vast crater]

🥲

[odd fog]

and can't even fix it afaik as it's set app side

[vast crater]

do you have a laggy app as well? @odd fog

[odd fog]

sometimes

[deft gyro]

I have a laggy codex app

no clue why

[vast crater]

Mine is stuttering/lagging periodically, the cursor is laggy af

I don't understand where it's coming from, no huge RAM usage, GPU or CPU as well

[boreal holly]

You could use Mutagen

[odd fog]

I don't think that would help? It's codex app that is trying very hard to write to the windows directories

and you just lose unix-file information when writing to the windows filesystem

[boreal holly]

Right, and with Mutagen you can restore the file permissions, users, and groups. Also if you turn core.filemode=false then git will ignore the "windows drive -> all files 777" permission stuff

[odd fog]

core filemode could be a good shout

But the path of the worktree is being set to the /mnt/c so unless i redo the mount entirely that would be quite hard to inject mutagen in?

[boreal holly]

I mean like, on Windows you set it up so that C drive worktree folder Mutagen syncs everything except for .git back into WSL, and you can set directories to 755 and files to 644. That way the agents are writing to files inside the WSL container, and mutagen pulls those changes back into the c drive worktree. That keeps the agents doing their apply_patch edits and stuff on real unix files inside WSL with proper permissions. And with code.filemode=false git wont care about file permission changes

The only thing idk about is if Codex app sets their CWD to the worktree on the windows drive. That might be the only blocker with that setup (sandbox would scope to windows mount)

[odd fog]

Yeah codex app is very stubborn on setting cwd

[north jolt]

Is anyone using Cursor CLI through Codex? I’m curious if I should set up the Cursor CLI as some kind of subagent in the Codex config or make a skill

[true bane]

as usual, thanks to whoever or however my weekly tokens on plus were reset to 100% unexpectedly early. 👍

[cyan wing]

Does anyone know when the 2X rate limits end exactly?
Like what time of day on April 2nd?

[cobalt junco]

was there usage reset?

[true pendant]

Hi guys, I found a shell-init inconsistency between Codex and Claude on zsh that breaks command aliases/functions set in ~/.zshrc.

On my machine, both report /usr/bin/zsh for echo $0, but the shell flags differ:

Codex: echo $- → 569X
Claude: echo $- → 569JNRXghkl

This causes aliases to be broken, in Codex:

type gh
# gh is /usr/bin/gh

In Claude:

type gh
# gh is an alias for op plugin run -- gh

My gh alias is defined in ~/.zshrc via 1Password shell plugin setup. Claude picks it up, but Codex does not.

I'm on linux and installed codex via npm.

[cobalt junco]

interesting

[royal garnet]

guys do you prefer playwright mcp or playwright cli?

[drifting granite]

cli, easy.

Any CLI > MCP

[cyan wing]

anyone seeing this error a lot on the new TUI version ?
on Linux

[manic thunder]

Hi

[stable obsidian]

What is that

[cyan wing]

ooh why is that?

[nimble terrace]

i remember hearing few weeks back that gpt 5.3 codex spark would release to plus users

i don't think that ever happened

[royal garnet]

have you tried it for playwright?

[drifting granite]

So an MCP injects context for how to use it into every conversation. Let's say you are using playwright for one task, then you need to switch to another where you aren't going to be using playwright. Codex still recieves instructions for how to use playwright, even if it isn't going to use it. With a CLI you can create a Skill for how to use it, then the skill only gets calls when you say "Do this with playwright", then Codex knows how to use it and how you want it used.

The skill also injects some context, but it is much much much lighter than any MCP.

[worn harness]

Try github cli vs mcp, and watch your tokens get evaporated

[tawny island]

anthropic limits the performance of claude code depending on which plan you are on I just found out from the src code

[sharp gazelle]

how will everyone cope with the sudoku of 2x tomorrow

should we form a self-help group and talk it through

[past prairie]

Hey guys, do you think we'll really stop not having the x2 anymore tomorrow ? I really feel the usage will go down really quickly

[past prairie]

it feels ok, it's to not use too much tokens with a little plans

[sharp gazelle]

I think they will indeed be removed, but immediately a $100 plan pops up with 5x. I just hope they don‘t forget about the poor business plan users that do not qualify for enterprise but can‘t have individual monthly based subs due to company policy (only tenant / seat billing on a yearly cycle).

[true pendant]

do codex shell tool calls do not inherit the user's real terminal auth/session context?

[tawny island]

yes but only if its transparent that lesser plan levels actaully performe worse, mybe it is transparent and they say that idk

[plush harbor]

I usually use codex for small tweaks, I got it to write an entire application last night ... if I kill this session and restart it does it just forget our old conversation or what? I'm really not a fan of living in the terminal like this

[true pendant]

you can resume the conversation via codex resume

and it should also appear in the conversation history in app/extension

[cedar skiff]

It's not a surprise they arent being transparent. They never have been.

[plush harbor]

I want to give it an image of how I want the page laid out and apparently you do that with codex -i

I can't use the app, this is all cli

[true pendant]

you can paste images in the cli

[plush harbor]

... how?

[true pendant]

ctrl+v

[plush harbor]

yes but how, like what do you paste? Path to a file?

[true pendant]

you paste the entire image

[plush harbor]

but its all ascii

[true pendant]

what is ascii?

the cli?

[plush harbor]

codex. The whole thing is just ascii

[cedar skiff]

did you try?

[wise heron]

yes path to a file will work too.. but just copy/paste the image works

[plush harbor]

like I said, I'm really not a fan of being stuck in the terminal

[true pendant]

dont worry, it will appear as [Image #1] but codex will realize its an image

[plush harbor]

ok that is really freakin weird lol. I wish I could use the app 🙁

[true pendant]

you can try ide extension if you really dont want to use the terminal but cant use the app

[plush harbor]

I just thought I'd randomly vibe code up an entire application. As you do

[wise heron]

on next updates will be also available in the app as in claude where they have in app all.. chat / cowork / code

[vagrant lily]

do we know what time today the x2 rate limits end?

[cedar skiff]

Youre going to build your own codex ui?

[plush harbor]

hah no, just a small application. It did a pretty good job of what it did so far, if you like somewhat teal buttons

[cedar skiff]

depends what the buttons do 😁

[plush harbor]

like all of them, its used teal for save/add buttons and the highlight colour on menus

failed to find what I wanted online, thought screw this its 2026 I'll get codex to do all of it

[safe holly]

Can I run codex on a remote SSH target? I want to deploy to a VPS but it kinda sucks losing codexs ability to work on the code.

[hard drum]

can you disable hook messages?

[cobalt junco]

is 4bln tokens in 30 days healthy usage or am i doomed

[scenic remnant]

Is Codex still available for Free tier?

[sweet heart]

codex keeps creating an empty .codex file idk why

[cedar skiff]

you mean .codex/ folder?

[exotic herald]

Is anyone experiencing hitting their rate limits faster with codex ?

[urban gazelle]

is codex acting strange for anyone else

[nocturne folio]

peter steinberger gets that much in a few minutes dude

you need to pump more slop

[cyan wing]

ya we need another month of 2X rate limits
cuz this is ridiculous!

[burnt peak]

I feel like the usage is draining after this morning reset

[plush harbor]

2x rate limit ended today

[nocturne folio]

is codex down

[nocturne folio]

tomorrow night

[exotic herald]

Yo I thought I was crazy like both gpt and api accounts moving crazy

[nocturne folio]

not todday

[plush harbor]

I thought it was the 2nd, who knows what timezone. Its afternoon april 2 for me

I'm just over here testing writing an entire application without touching the code myself. Going good so far

[nocturne folio]

openai is american

so it wouldnt be like

yk

[plush harbor]

yes I've noticed, chatgpt always has the day wrong. Wouldn't kill them to add localisation

chatgpt got the day wrong YET AGAIN today so that reminded me to tell codex to put configurable timezone in this app I'm making it write. Or I'm gonna be disappointed when I put it on a server that's running at UTC 0

[exotic herald]

I’m in London it’s 2nd April and I’m hitting rate limits

[cedar skiff]

kimi devs have a new way to manage attention, looks like it could increase context stability. Lots of big steps forward coming out lately

[vital hill]

how long does the go plan $8/month last in codex?

[cyan wing]

anyone else experiencing apply patch tool errors in new Codex version?
I'm on Linux

[slim badge]

any unity devs using codex? how's your experience so far?

[karmic grotto]

I'm using codex $20 (USD) plan, and it offers basically as much as the competition's $200 plan

[cedar skiff]

not even close tbh

[vital hill]

depends on the project

[cedar skiff]

tokens are tokens

[oak trellis]

I like your new profile pic

[cedar skiff]

Ive always been bananaman :p

[oak trellis]

By the way codex still lit ? Vs cc ? Used yesterday without any limits …cc hitting non stop

[oak trellis]

Not the lego versión

[cedar skiff]

codex is still pretty good, they both have their problems

[oak trellis]

Agree! Using both right now. Cc for init , codex for clean up and continuation

Cc opus has always missing stuff .. do not follow instructions etc

[karmic grotto]

yeah, cc tries to skimp and skirt the work a lot more

[hard tulip]

[slim badge]

im trying it out lol. ive worked on other projects with codex and they were awesome. wanted to explore unity too

[sand shuttle]

did they reset limits?

[slim badge]

it resets itself every 5 hours and every week

[cedar skiff]

yeah the model is incredibly smart but has a real attention problem, the leaks prove it. Every single turn claude code injects the CLAUDE.md.

I wonder if they do the same thing with gpt models

[solemn acorn]

you can check but I don't believe so

codex is (intentionally 🙃 ) FOSS at least

[solemn acorn]

In a normal ongoing Codex session, it is left at the top of the conversation history rather than freshly re-injected on every new user turn.
(chatgpt given the codex repo, also the official docs corroborate that)

[vital hill]

i love how vague i can be in a prompt and the codex desktop app jsut gets it

[solemn acorn]

it is re-injected during compaction but that's because the entire context is rebuilt which is fair

[cedar skiff]

gpt models must be much better at attention

i mean they call skills really well

[sweet heart]

No, an empty file

[cedar skiff]

where is the file?

[sweet heart]

in my project root directory

[royal mirage]

anyone know which plugin we have to use inside claude code so it can run codex/chatgpt inside it?

[plush harbor]

Interesting. I just got codex to write me an entire app. Token usage: total=897,183 input=823,967 (+ 12,276,096 cached) output=73,216 (reasoning 6,704)

[royal mirage]

chatgopt 5.4?

[plush harbor]

yeah

did the whole thing in one session so that was what i got when I closed codex

[wintry crypt]

Anyone notice how codex superbowl stuff has had no update

[barren hinge]

Have you experienced malicious command from bitdefender if you use codex ? or R0? Should we treat those as a necessity or can it be better ? Someone expect more browser_container (e.g. dart) in codex for screenshots ?

[lusty nimbus]

are the 2x usage limits gone now?

[barren hinge]

It was mixed in 1

[barren hinge]

Task and review are in the same count

[lusty nimbus]

meaning they both count towards the same limits now? What about the 2x rate limits? Are those still in place?

[cedar skiff]

code review still showing up for me

[barren hinge]

Yes, 2x is replaced by that, If you use code review you most likely won't experience it

[cedar skiff]

Cna you explain this Task and review are in the same count?

I think i misunderstand what is going on

[barren hinge]

If the review do the work you don't do before then you can keep it like that

[cedar skiff]

uh i can't understand this o.0

code review bar is for github review

I think they are folding that into normal usage right?

[static lark]

How to use this code review in codex app ?

[static lark]

It can only be used inside browser codex ?

[cedar skiff]

It's for github, you setup github with codex and then it reviews prs. But i think they are folding that usage into the normal limit soon anyway.

[tacit socket]

man i really hope $100 plans come out after 2x finishes

$200 is too much and i don't need that much =/

[cedar skiff]

I think that is kind of the idea of the pricing gap

[tacit socket]

there were some leaks that showed $100 plan floating around. so i think it's coming, just a matter of when

[cedar skiff]

ahh ok, interesting.

i wonder if the plus plan will take a hit then

[tacit socket]

sh*tification is always the path forward

[cedar skiff]

it moves fast though, there is lots of progress in performance and efficiency

[lusty nimbus]

is anyone else getting "Selected model is at capacity. Please try a different model."

[orchid plume]

Yes i hope after 2x ends. Pro 5x for $100, pro 20x for $200 last i checked. Depends if plus goes back to original limits, if it does then 20x will feel like 2x didn't get turned off

[tacit socket]

Yeah I hope so. Otherwise it'll be another month of 2 accounts and swapping between them. Rather the $100 5x for laziness sake lol

[worthy furnace]

My usage limit was full 5 minutes ago, did 1 task and its all gone and my weekly is down to 40%

Are you serious? Lmao

Who is going to fix that?

@robust lintel

@winged ore

This is just theft at this point

[hard drum]

i would very much appreicate a 100-buck plan tbh

[muted tinsel]

Codex VS code extension KO ?

[winged depot]

Are they really going down the Claude route, where they make their affordable plans significantly worse, in order to push people towards the more expensive option? That surely is one way to encourage account sharing.

[orchid plume]

only way we'll know for sure is what happens after 2x ends and when the new plans release, but the hidden slider and such on the plans page is Pro 5x and Pro 20x

my hope is that Plus will go back to near enough or exactly at its original limits for the quota, in which case Pro 5x essentially nearly becomes what Pro $200 was at 1x bonus rate, while Pro 20x which replaces the original Pro $200 (at 6x~) will feel like 2x wasn't turned off

[past prairie]

real, i don't know either, what I'm sure is even with 20X plan the quality of the model often changes without any transparency (for claude, not codex)

[cedar skiff]

bring on spud

lets see how it goes

[past prairie]

I hope

[orchid plume]

if the monthly model cycle is continuing, we might see spud today, if not, next week I guess 😛

they're probably also waiting for Anthropic though

[wide forum]

здарова

[fallow steppe]

uhh guys, codex is down?

[cedar skiff]

its working for me

[waxen terrace]

Is Claude Code still better than Codex?

[orchid plume]

for features, well, it has more hook coverage, perhaps a slightly nicer looking TUI, an experimental teams feature, probably missed a few things which would be nice to have in Codex CLI at some point as I don't really use CC. Codex CLI however has what's needed pretty much, just improved hook coverage would be nice to see soon. Codex CLI's code is also far more polished compared to Claude Code's code (the leaked code that is)

[worn harness]

claude code is better if you hate your money or if you need heavy frontend design

[cedar skiff]

Codex is better at code. Claude forgets too soon, so you have to spend a lot of time and effort enforcing the following of instructions. Codex doesnt forget like that.

[sharp gazelle]

is 2x gone yet

[worn harness]

hope so, because usage burn at least feels 2x faster,

[cedar skiff]

code review is still showing in the dashboard

[waxen terrace]

What about the usage limit?

[cedar skiff]

usage on the top plan is similar for code alone, but you get unlimited image gen, unlimited gpt chat, more model choices, codex app, gpt-pro access, weekly spark usage above normal usage.

Claude is better at research, documentation, and writing imo. but you have to pay for it in your token allocation

[waxen terrace]

But is the coding limit much higher than Claude?

[cedar skiff]

If you were to us claude for only coding and codex for only coding i think it feels about the same amount. BUT thats just my exp, nothing definitive. codex also has a github code review usage. They did say they were folding that into normal usage though.

They havent done it yet

[wooden minnow]

I feel like codex limits are consuming fassst

is it the 2x over?

I think codex should keep 2x limit as defaults

[cedar parcel]

I think they should replace the 2x by 4x then we're closer to where we have to be again

I wonder what would happen if they made a $999 unlimited tier

[worn harness]

any idea what "other" here is? I have only codex cli

or their new version is logging stuff wrong

[orchid plume]

Codex's current usage limits are far superior compared to Claude's current usage limits, well, even on x1 rates I imagine it's still going to be superior 😂

one other nice thing is that Codex and ChatGPT have separate quotas

hopefully we'll have a nice surprise today, not a nasty shock 😛

[orchid plume]

they'd probably end up chasing fradulent accounts if it was truly unlimited

[hard drum]

Codex doesnt forget like that.
GPT 5.4 wuld like a word with you

[cedar parcel]

Yeah people could infinitely resell the capacity it would create a black market for tokens 😬

[waxen terrace]

Is it also that bad if you use Claude with auto dream and plugins like everything on GitHub?

[hard drum]

Only add a rust library for this existing TS library
agent decides to not follow this, and actually add something extre that was never asked for

gpt 5.2 doesn't do that

gpt 5.4 is an arsehole

[sleek spoke]

😍

[plucky halo]

Disappointing on the pro plan

[sleek spoke]

what is that message?

[plucky halo]

Telling me gpt-5.4 is at capacity in Codex

[worn harness]

area? Never seen that message in EU

[plucky halo]

UK

[sleek spoke]

it can be 'at capacity'? 😮

[fervent anvil]

Codex 5.4 is unusable with anything like a large codebase. It goes one step forward, four steps back. It is so frustrating. It breaks the last thing it fixed when fixing something else. Its like it has amnesia.

[worn harness]

what thinking are you using, do you have memories enabled?

[orchid plume]

I haven't seen that message yet but my guess is that many are making the most of the bonus rate which ends today

[plucky halo]

Ignore my last message - I misread yours

[past prairie]

got this error yesterday i'm in eu

[past prairie]

you prefer which model for large codebase ?

[vital hill]

maybe you fed it too much context

[sleek spoke]

do you reckon they'll launch the $100 plan today to coincide with that? it seems plausible since it sounds like the new model is coming next week

[orchid plume]

I'm hoping so, otherwise this week could be a bit of a dull week 😛

[waxen terrace]

New model?

[sleek spoke]

"Spud"

tl;dr = Greg has said in an interview it's a new pre-train and it's coming, Tibo tweeted a hint that something big is coming next week that will increase Codex user growth

[waxen terrace]

What do you think about Claude Code and Codex?

[sleek spoke]

I've never used Claude Code, so don't know

[waxen terrace]

Are you happy with Codex?

[sleek spoke]

sure

[vital hill]

this is hilarious, it's running live too https://iscodexgoodatfrontendyet.com

[worn harness]

5.4 will never make decent frontend design, I think I have wasted 20 days trying to make it but it's just impossible

[vital hill]

codex 5.3 was pretty good at recreating an html site to a custom wordpress theme

but coming up with a design from scratch? forget about it

[worn harness]

like yeah if you copy an existing design it will do it, but if you want a unique design it's just a dead end
just an army of lists and cards

if I see this color I reflexively gag now

[quaint sigil]

whats your guys main coding model been switching been 5.4 high and xhigh and cant seem to justify the thinking between high and xhigh

also seen some people still on 5.3 codex

[plucky halo]

I very, very rarely use xhigh

[worn harness]

high 90% time, if stuck then try codex 5.3 or 5.4 xhigh

[quaint sigil]

yeah been seeing that pattern hm okok

[plucky halo]

Find I get better results instead of using xhigh when it's stuck, handing off to another model to look at

[vital hill]

i stay with low until a bug loop

[quaint sigil]

main pattern i see is xhigh seems to overthink and trip on itself

[plucky halo]

nested and nested and nested

[worn harness]

and that

[orchid plume]

have you tried using Impeccable or Taste skills? high reasoning effort

[worn harness]

yup, very minimal impact

from 5 to 6 as best

[orchid plume]

that's unfortunate

[drifting granite]

I've got a theory we are getting 5.5 today for Codex. The last releases have been on the 5th of the month, the 5th is a Sunday, and today the 2x limit ends.

[potent mason]

I don't think we are getting a new model this month, we just got 5.4 and both labs are working on something "big" which should take a couple more weeks at least for RL

[worn harness]

i think they are waiting to match claude release

[drifting granite]

I'm pretty sure they mentioned doing monthly releases for models.

[past prairie]

probably next week based on what tibo said

[cyan wing]

source?

[sleek spoke]

this

[orchid plume]

yeah while it would be nice if the model arrived today, I don't think it will. Hoping instead that the new Pro plans will be released at the very least, curious how 5x and 20x will actually be

[past prairie]

https://x.com/thsottiaux/status/2039550028444991515

[potent mason]

They're going to change pro plans?

[orchid plume]

yeah, there's a hidden slider on the plans page, on Pro specifically, which says 5x and 20x. 5x will be $100, 20x will be $200. There's also hidden mentions of these in the javascript too

[toxic torrent]

is pro plan already 20x?

[ivory zodiac]

they nerfed 5hr limits?

[orchid plume]

Pro $200/mo is currently about 6x~

[orchid plume]

so, there's two possibilities:
a) If they return Plus back to the original quota limits then Pro 20x will feel like the 2x bonus is still ongoing
b) If they reduce Plus's quota limits compared to what they originally were on x1 then Pro 20x won't necessarily feel like the bonus 2x is ongoing

so we need to wait and see what happens

[toxic torrent]

voice input has been failing/timing out a lot today

unsure if client or server issue with whisper

[nocturne folio]

gpt 5.5 will release tomorrow and 2x will stay

well i hope

[sand shuttle]

has codex performance dropped recently?

[hard tulip]

how do you know

[drifting granite]

vibes

[steady vigil]

that's more of an architecture and planning problem

[cyan wing]

how do you recommend people experiencing this issue can resolve it ?

[sleek spoke]

Well they're making one pricing change today at least... https://openai.com/index/codex-flexible-pricing-for-teams/

[steady vigil]

I dont know the specific problem they are encountering, but if you arent getting good results on a large code base it probably means the architecture is poorly designed. I'm using it with an ~2m loc project that is as easy to work with as a 20k loc project. Why? rust actor model. No state problems, no context bloat issues, its like working on individual small projects

[cyan wing]

do you find that Codex uses like 30-40% context to even understand the current implementation?

[steady vigil]

my point is it shouldnt with good architecture b/c there are almost never any global changes

you are dealing with one actor + potentially one hop away actors that's it

have you tried a graph / semantic view/mcp?

that can help with messy architecture to try to simplify/clean it up you just build the nodes and lines of what talks to what and then you can optimize

[cyan wing]

i have a Deepwiki index of my codebase, which has worked well actually
i think they use semantic search

[cyan wing]

is there a good public example of this i can see for reference?

[steady vigil]

well it depends on the language used

[cyan wing]

im using Rust ya

[ornate musk]

a good skill/agents.md fixes this

[steady vigil]

lemme find it, sec
try this:
https://github.com/jonmmease/jons-mcp-rust-analyzer

@cyan wing I'll release the actor model I'm working in a few weeks I'll let you know. I highly recommend using it, it will change your life 🙂

I buid everything using it now

[cyan wing]

Sickk thank you, much needed!
I have codex splitting out my ~300k loc repo into separate crates rn for better discovery / decoupling

[steady vigil]

yeah there will jsut be one time refactoring/migration pain but will be well worth it. then you basically have a set of re-usable actors that are like self contained apps with their own state. your error rate will drop to near zero, no locking issues, etc, its literally magic

[true bane]

yo i thought this** (INCREDIBLY USEFUL AND BETTER THAN YOUR COMPETITORS)** code review bucket with 53% was going away 2 days ago? any one know when its really leaving? shockingly internet searches not helping

[cyan gyro]

wow tokens are being eaten like crazy

[karmic grotto]

Does anyone else call GPT Geppetto?

[cyan wing]

better model accuracy, lower context use, and faster code changes?!
sign me up!

[manic thunder]

HI, nice to meet you

[orchid plume]

Still hoping for pro 5x and 20x today but I have a feeling next week earliest now

[sleek spoke]

it is hard to imagine that they will not acknowledge the end of the 2 months of 2x rate limits today in any way at all

[orchid plume]

Rohan on X at least said something about it being a war room about ChatGPT/Codex pricing changes or something, glimpsed a post earlier at least, so maybe it's still happening today

[sleek spoke]

@ivory zodiac seems to agree

[cyan wing]

I agree. kinda concerning actually

[sleek spoke]

oh wow, I assume this is the post you meant, that seems to pretty much confirm it

very interesting, how old are the current price tiers now? since the launch of o1-pro? (excepting the Go tier)

[orchid plume]

not sure on age but certainly quite some time, and yes that's the one

[cyan wing]

bug..

[sharp gazelle]

the new gpt business plans are not what we wanted… we wanted a more expensive seat with higher codex limjts

[frigid sigil]

I am having trouble getting .codex/rules/default.rules to do anything useful. I have massaged it in about 8 different ways, and I still have to approve every attempt to read anything. I wrote a helper function
"C:\Users\usx28880.codex\scripts\read-file-slice.ps1"
and put variations of
prefix_rule(
pattern = [
"&",
"C:\Program Files\PowerShell\7\pwsh.exe",
"-File",
"C:\Users\usx28880\.codex\scripts\read-file-slice.ps1",
"-Path",
],
decision = "allow",
justification = "Allow the standard read-only helper for pulling a subset of a file.",
match = [
"& 'C:\Program Files\PowerShell\7\pwsh.exe' -File 'C:\Users\usx28880\.codex\scripts\read-file-slice.ps1' -Path 'src\StoreApp\StoreAppBaseContent\ViewModel\Managers\InteractivePrintPreviewSession.cs' -Skip 1 -First 950",
"& "C:\Program Files\PowerShell\7\pwsh.exe" -File "C:\Users\usx28880\.codex\scripts\read-file-slice.ps1" -Path "src\StoreApp\StoreAppBaseContent\ViewModel\Managers\InteractivePrintPreviewSession.cs" -Skip 1 -First 950",
"& 'C:\Program Files\PowerShell\7\pwsh.exe' -File 'C:\Users\usx28880\.codex\scripts\read-file-slice.ps1' -Path 'C:\Users\usx28880\.codex\rules\default.rules' -Skip 0 -First 260",
"& "C:\Program Files\PowerShell\7\pwsh.exe" -File "C:\Users\usx28880\.codex\scripts\read-file-slice.ps1" -Path "C:\Users\usx28880\.codex\rules\default.rules" -Skip 0 -First 260",
],
)
(with the "&", without the "&", with single quotes, double quotes, etc.)
but still it asks me for approval every time it needs to use my script to read part of a file.

[meager wadi]

the 2x limits were so good 😭

[haughty plaza]

Must be an idiot but codex is totally unusable this cant be for real

No serious work can be done like 20 min some refactors -> wait 5 hourw

I understand it can become a bit more expensive but this is redicelous let them at lewst hive me credits till im
Internet millionaire then i gladly fork over 2500 a month

[worn harness]

there's no way you get only 20 minutes out of plus sub

[haughty plaza]

No im on a plus account using codex

[worn harness]

disable fast mode, dont use xhigh

[haughty plaza]

Guess thats the catch then they want me to upgrade to pro. Yup no fastmode and using medium

So they got me how much do u do on a pro subscription and anyone lnows if its tax deductable? Like in europe or it it like google ireland no tax deduction?🙏

[worn harness]

codex features list | grep true
collaboration_modes              removed            true
enable_request_compression       stable             true
memories                         under development  true
multi_agent                      stable             true
personality                      stable             true
shell_snapshot                   stable             true
shell_tool                       stable             true
skill_mcp_dependency_install     stable             true
sqlite                           removed            true
steer                            removed            true
tool_call_mcp_elicitation        stable             true
tool_suggest                     stable             true
tui_app_server                   removed            true
unified_exec                     stable             true

Check what you have, here's mine

[haughty plaza]

Ah good one im not behond desk but def no fast mode inteied that once no benefit to me

[worn harness]

it was enabled by default on some update

[haughty plaza]

Ok ill double check thank u. How much would u get out of it with peo its fair usage policy right?

[sleek spoke]

it's confusing but I believe that the feature flag just enables the /fast command in the TUI, it doesn't actually enable fast mode

[deft gyro]

lol

if ur running out thats a skill issue

[haughty plaza]

Whatnwcares me they can shut your account down no reason. No refund

[deft gyro]

of the 5 hour at least

[worn harness]

my token usage reduced instnatly after I disabled it

[deft gyro]

youre not using enough 5.4 mini

[sleek spoke]

interesting

[deft gyro]

also avoid 5.4

5.3 codex is mildly cheaper

[haughty plaza]

I once did claude subscription last year claude totallu scked gemini same dont thrust them anymore codex is only option for me

[whole gate]

I use both Claude and Codex together, Opus 4.6 does my planning, Codex in the Cloud does the actual code implementation.

[deft gyro]

I use ag and codex

[whole gate]

Codex is very good at writing code, Claude is a bit better at planning becaues of larger contex, at least that's the feeling I get.

[haughty plaza]

Claude is a lying ai agent outputting just crapp and never follows rules thrust gone maybe its better now

[deft gyro]

to be fair I have like 200 dollars of unused codex credits

[whole gate]

@haughty plaza This is a fair observation, even in planning I constantly find it violating rules and suggesting anti-patterns. I just do extensive reviews to make sure these kinds of things are caught. I spent at least an hour on most prompts and results reviews.

[haughty plaza]

What does fair usage means for u guys? How would that compare to a pro subscription of course i use the 5 accounts even though not allowrd

[whole gate]

I upgraded to Max and have never hit any limits for the last two months I've been using it.

[haughty plaza]

Yes its total headache it lief so many times to me i just wont soend a dollar on claude but anyone free to use it

[haughty plaza]

Claude max?

[whole gate]

I actually just spent several days writing new LLM guidance, which has helped a lot. (Yes, Claude Max)

Codex in the Cloud on the $20 plan I rarely use more than 15% of my tokens per week

[haughty plaza]

well im only interested in codex pro but.. like said there policy if they deel anythijng wrong they block account gone 230 dollars nonway to contact them decently that is just why i used pro at least then the hit is only 140 ush

[deft gyro]

sooo youre not using codex

[haughty plaza]

I ran 20 tabs working on 20 different repros tools researxh 5 saas business etc and that was just doable with 3x plus but now dunno how to ever finish mynstuff

[whole gate]

Literally writes almost 100% of my code lol

[haughty plaza]

Manually codijg is nonoption anymore

[whole gate]

Yeah, I'm very single threaded.

Because of how unreliable the coding agents are still, this has only improved my workflows slightly.

I barely ever write code these days, occassional fix or quick edit faster than going through a prompt, but to be fair, I usually spen 90% of my time as a developer planning, not writing code.

So this improved my workflows a bit as it cut that 10% writing code out of the equasion.

[haughty plaza]

I dont even rename folders or move files a prompt is faster but well the main question is if codex will atill be worth it now. I wonder if there is some docs on pro versus plus usage

[deft gyro]

if youre on the 20 dollar plan and only use 15% of your weekly tokens AND YOU ARE USING CODEX CLOUD you are not using codex often

[haughty plaza]

Need to find some intell

I dont use cloud to much hassle i run locally only

[indigo sandal]

Is 2x done already

or today's last day?

[deft gyro]

done already

I only use codex local

its the first day of the wek

[whole gate]

Possibly, I generally run around 5-10 prompts a day, but given that it takes an hour to write a prompt and I don't code 24/7. If you're curious: https://github.com/elide-us/elideus-group/tree/test current stability branch progress. Project is not at all small, but this is also about 18 months of work.

[deft gyro]

and I have already used 5% of weekly quota

[haughty plaza]

Yes and its terrible feels like 20x

[desert junco]

guys I don't know why, last week I was using codex for everthing and I didn't remember if I reached the limit, but this week I run 3 commands and 80% of 5h limite over in 20minutes, something happen?

[haughty plaza]

Must be the x2

[deft gyro]

half doesnt quite feel like half

[desert junco]

I verified it before, and it is off

[steady vigil]

its really not, Just make a planning skill and you can make whatever kind of planning behavior you want. This is where gpt 5.4 is best. If you need inspiration check oh-my-codex deep interview

[deft gyro]

No, the quotas got halved today

[haughty plaza]

• “Codex now uses x2 more” = the model (whatever “codex” refers to in your setup — probably the underlying code model or agent backend) now consumes twice as much of the limiting resource per minute of runtime.
• That limiting resource is almost always the thing your subscription/quota is measured in (tokens, compute credits, API calls, etc.).
• Your total budget for that resource hasn’t changed.
• So if the agent previously burned through the budget in 30 minutes, it will now burn through it twice as fast → 30 ÷ 2 = 15 minutes max runtime.

[whole gate]

I have been using Codex for planning for the first 16 months of this project, only switched to Claude about 6 weeks ago, the results I am getting there are far better than the Codex plans. I did A-B comparisons. I might try it again, those tests were on 5.3.

[haughty plaza]

To
Me
Status also feels off and way more then 2x

[desert junco]

oh okay, I understand now, the quotas were 2x limite, and now it is the normaly, lol

[deft gyro]

Yes

[desert junco]

bro, the normal is not too good, I need more money for pay the 200$ plan

[haughty plaza]

Yes but claude code scks totally unreliable lying manipulating ai

[steady vigil]

yeah I mean I could see why you would say that, especially just comparing /plan vs cc plan mode, but definitely try your own skill. I felt the same way for a long time also

[whole gate]

I will note that Codex does repeatedly introduce regressions into my RPC namespace due to a unique AST crawling quirk. Claude details it very clearly and explicitly and Codex still falls back to "correcting it" because it really hates making references and variables not the same term 🙂

[haughty plaza]

Guess we need to buy a mac studio m5 at 5k and run opensource models

Shame they will never be as good

[whole gate]

Actually not even using /plan mode. I just use Claude Chat, I tell it to read the documentation for the project and tell it what I'm trying to do. I did experiment with CC/plan yesterday, it was faster, I'm not sure it was better.

[haughty plaza]

I wonder what advencements there are in opensource models havent looked into it laterly im on a macbook m1 def not able to run anything decent

[haughty plaza]

I guess only same repro instructions and compare output but indont touch claude anymore

[whole gate]

It's this, I basically had to beat Claude into submission to understand why this variable needs to be specifically 'module' but that it doesn't match the reference name. The app state has basically a registry of modules that are dynamically loaded on startup and registers them all by module name. Since that entire system is dynamic, there's no static code for the agent to undrstand why this variable mismatch is there.

I don't like AI touch my code directly, except for Codex. That is actually the only allowed pathway to prod.

[haughty plaza]

Nobody soijg multiple saas startups af the same time with codex i must be insane running 20 tabs

Well rhat dream is over now

I wait a few days fhen see if pro would be an option be in the mercy of opensi account closure anythine they want and nonrefunds discussions possible

[cedar skiff]

You can solve this with a topical skill

The same goes for the claude quirk, but it's not as relaiable as codex with skills so it will forget sometimes

[whole gate]

For sure, I basically just wrote a document that details this repeatedly and tell the agent to read it all. Actually this detail is also in the AGENTS.md where relevant as well, which was odd that Codex missed it. I'm doing a consistency pass in the namespace now, looks like there are several pattern deviations, which explains why it might fail. If it catches one instance of a more typical pattern it will lean into that.

[cedar skiff]

I noticed this as well it will lean into that. I was wondering why it kept defaulting to some function wrapper pattern instead of a class - it started doing it out of the blue. When i asked it it showed me the nearest pattern in the code base, which was also wrong. So letting that slip through can have it spread over time.

I was thinking i would wake up today to a nice shiny new spud

[whole gate]

I'm running a new A-B test on Claude vs Codex now. The prompt is too long to share, but it's a consistentcy and pattern refactor, about 140 lines long

So far Codex produced a plan, Claude is still thinking on it, but Claude is also spending time reading the MCP server for schema reflection.

[ebon ingot]

Codex 5 hours usage is insane it runs out in half of an hour😭

It didnt even finish half of the prompt 😂

[cobalt junco]

damn 2x usage is noticably gone.

[whole gate]

I might... consider switching back to Codex. I had Claude review the plan that Codex made and it did find an issue that Codex ignored, but it was a design issue on me. Otherwise Claude and Codex were in perfect alignment, but Codex did it about 3x faster... that's... appealing.

That being said, that was Claude Chat vs Codex Cloud

[worn harness]

highly suggest using cli, I've tried 5.4 over API via cursor and others and the quality is different compared to codex cli

[whole gate]

I don't allow AI to operate locally, in general, beyond just reading reference data.

I'm not sure I'd get any particular advantage out of that, my entire code build pipeline is set up in the Codex Cloud environment so it's working in a more realistic setting than my local system.

[worn harness]

I mean unless you are shipping live doesnt matter if you break the bindings every now and then, codex usually corrects pretty well unless you ship multiple bad versions

5.4 very rarely makes regressions, only then when I usually leave something open to interpretation

[whole gate]

This is all CICD live to test server

I do feature releases once feature/stability backlog sets are complete.

I literally just had Codex go through and correct 25 regressions it made earlier today lol

There are multiple notes in AGENTS and other documents calling out the pattern it broke 😉

[worn harness]

idk, would have to see code, hard to guess blind

[whole gate]

Oh, it's not a problem

Just observations, going quite well

This is just an A-B test betweel Claude and Codex on a pattern conformance task

Codex has been faster and similarly accurate, Claude noticed some details that violated the architecture that Codex missed.

They are all issues that could be fixed in subsequent reviews and tasks and don't break anything, so pretty minor.

I'm giving Codex the win on this round.

The main issue is that the objective here is that I get rid of this static RPC code that I do AST crawling on and instead generate the RPC typescript files from data in the database. The architectural violations come from logic in the service dispatcher. Which is what this is fixing, or trying to.

I've got the crawl populating the tables, making some additional enhancements to logging on the build generators to call out specifically found violations while crawling.

[jagged pulsar]

why cant codex rename its terminal?

[white furnace]

is the x2 rate for codex gone?

[nocturne folio]

im using 5.3 codex and it just feels better

[nocturne folio]

not yet

let them rip

[vital hill]

i am on the plus plan and don't see codex 5.4

where is it?

[tacit socket]

damn i really feel that 2x drop off. my 5h limit taking a beating ha

[potent mason]

Oh wow I just realized how much usage I'm consuming. I'm guessing it's the same for everyone? (because of the 2x thing)

[karmic grotto]

I just ran out of my weekly limits on Plus... until Apr 8 ;`(

never ran into session limits at all

[lean lark]

OK, I give up. VSCode Codex Extension returns this on prompts:

Your access token could not be refreshed because your refresh token was already used. Please log out and sign in again.

There is no "log out" or "sign in" feature after the extension is installed.
What's the magic key that unlocks the great door to the Holy Codex? 🙂

[deft sable]

Looks like we'll be seeing reset again very soon lol.

[tacit socket]

I just setup codex-lb to balance my Plus account and Teams account. Works well 😮

[lean lark]

There are known issues on GH related to the access token situation, but I'm not seeing anything yet that helps me. Output log for Codex shows:

[error] [CodexMcpConnection] cli: message="... WARN codex_app_server::message_processor: failed to force-refresh directory apps after experimental feature enablement: Request failed with status 401 Unauthorized: ... Provided authentication token is expired. Please try signing in again. ... code: token_expired
[error] [CodexMcpConnection] cli: message="... ERROR codex_login::auth::manager: Failed to refresh token: 401 Unauthorized

I'm not using MCP in this environment yet. 🙁
I don't know where it stores Codex access tokens or how it tries to refresh them.

I can go neanderthal on this and delete the extension and OpenAI folders to clear cache, but I'd rather not.

I'll continue to look at GH issues. 😢

Hmmm - I opened the CLI and it wanted to do an update. I did that and restarted. Then used the "/logout" command. Then restarted. On restart it needed a new login, launched the browser, did the login. Restarted CLI again and it worked. Then VSCode extension worked.
Solution to : "Your access token could not be refreshed because your refresh token was already used. Please log out and sign in again." 401 Unauthorized

[cedar parcel]

Gpt 5.2 xhigh, I never understood how 5.3 and 5.4 can work for anyone, really.

[potent mason]

I liked 5.3 codex high, 5.4 is kinda weird works in some very niche scenarios

[boreal holly]

I guess I must be doing something wrong because 5.4 works pretty much flawlessly

I'm like the one and only person maybe on the planet who experiences nothing but excellent performance with 5.4

[cedar parcel]

If you give both the same task on a large codebase it should be that 5.4 consistently gives the wrong answer and 5.2 consistently gives the right answer. I can do that experiment all day long

[lean lark]

Yeah, with just a couple exceptions, I've been continually impressed with 5.4.
Robert pointed to a page that describes nuances of how to work with 5.4. See OpenAI docs.

[junior fog]

5.4 is highly effective, but its significant token usage causes the weekly limit to be reached too quickly

[lean lark]

Turn off Fast processing and you'll reduce your token usage significantly.

[cedar parcel]

It just doesn't gather enough context so it can't produce positive results that way I don't know how it could ever work.

[lean lark]

sigh - I wrote a paper the other day about this very topic. I need to clean it up a bit and publish somewhere. It got too long with too much detail, which is kinda required for this deep topic. I'll see what I can do. Watch this space. 😉

[sacred minnow]

na bro i love 5.4

[lean lark]

Oh right, and 5.4-mini is also a great way to save tokens.

[boreal holly]

You have to give it a personality. Not necessarily in how it responds to you, but how it's supposed to think. For example I give mine "adversarial, pragmatic software engineer". If it has no personality then it will basically be a thoughtless task executor. Also give it an output contract like "files touched and why, validation performed" so it knows before it responds to you, it has to look at what files it touched, and it has to perform some kind of validation. No output contract, it will resort to doing the bare minimum to finish as quick as possible

[sacred minnow]

i forgot rate limits changed and ate 10% of my pro plan xD

[boreal holly]

5.4 output contract is a huge win. Tell it how it's supposed to respond every time, and make the output contract in such a way that they must review their work and think about what they did

[lean lark]

We've discussed this a number of times : Document in AGENTS.md how to do what you want it to do. Get it to create comments for all functions and to update procedure docs, functionality specs, README's, and CHANGELOGs. When you give it a new task it will rely on existing documentation rather than having to relearn the entire project with every new task.

[cedar parcel]

I do all that, and it even went out to rewrite the task description so it could meet the acceptance criteria without having to do heavy coding work. If I handhold it I can make it produce some kind of results but still I have to fight it and check it. 5.2 just does the work and does it perfectly 8/10 times and the 2 cases it doesn't it's usually due to human error or it's just something needing more human insight and guidance (front end).

[sacred minnow]

Actually got some really nice results with front end 👌

[vital hill]

Thanks

[sacred minnow]

Its funny reading reddit I swear half the problems most have i have not experienced

[vital hill]

Thought codex 5.4 was out already

[potent mason]

Nah they're not releasing any more codex models (iirc)

[cedar skiff]

seems like they have picked path of general model

over all they have so many options but not enough resources (probably) so sora got the chopping block, codex got the chopping block. And they are building spud

spud probably going to come with a app (eventually) that brings everything together for everyone. Like a mega app

[leaden heron]

is anyone elses rate limit also bugged? literally 3 prompts is my entire 5 hr limit

[cyan wing]

i swear Codex is giving me nightmares about the words authoring and cutover
they show up everywhere for some reason
IYKYK

[next ore]

2x limit is extended

i coded 14 hours today to take full advantage of it before it expired

at least it got me working lol

[next ore]

you must be doing some insane prompts

[oak trellis]

codex is MUCH better than opus right now

did a test running both .. and oy vey ... cc messed up created me errors then fixed its errors .. and had still errors. codex super smart detecting it surg..fixing it

[next ore]

I totally agree

[cedar skiff]

I find codex thinks i shouldnt change code some times, it tends to make things more complicated that it needs to be, some times introduces abstractions for no reason etc. Then when i say get rid of it , it always comes back with it's doing real work but it isnt o.0. So i worked out i can to say steel man removing it. And it will do its best to justify the simplification rather than doing its best to be diligent and not remove something important.

Because it always finds a justification to keep almost every abstraction it made.

giving it the steel man task most times it comes back with this should be flattened or this is mostly ceremony

and it gets the right perspective straight away

It also highlights the actual real work it is doing (if any) not the flakey ideas it has when you just ask it normally

[solemn acorn]

could also just remove the dead code yourself

[cedar skiff]

whats the point of having a coding agent then?

[leaden heron]

nah its not even about the prompt or the 2x limits, if it was about the prompt it shouldn't be burning through 30% usage just looking at a couple files and reasoning, then another 10% to write the code. when it was 2x limits my average prompt only used 5%

on my life im not tweaking bro

[cyan wing]

nah you're not
they need to fix it

[plucky halo]

Second time in two days! 😭

[fallow steppe]

I’ve been using Claude Code lately, but since they’ve significantly tightened the limits, it seems to run out quickly. Some Claude Code users have started switching to Codex—I switched to Codex myself and felt a huge sense of relief.

[spring snow]

Good morning, what does this message mean ? "Selected model is at capacity. Please try a different model."

[spring snow]

having the same issue

[cedar skiff]

lets just hope its because they are rolling out a new model

[spring snow]

gpt 5.4-codex ?

[cedar skiff]

spud

they hinted at it being the next model

codex models are probaly a thing of the past

[cyan wing]

source? 👀

[neat sinew]

They're probably gonna introduce a new tier, Codex Max or something

[cedar skiff]

https://youtu.be/J6vYvk7R190?t=1378

[oak trellis]

ok couldnt send that but reached my weekly limit ..

question ... i reached my weekly limit within 2 days is that normal ?

i reached the weekly limit before my 5h limit

[cyan wing]

I thought it was next week cuz this 👀

[oak trellis]

it says 5h limit 93% weekly limit is 100% ..

confused

[neat sinew]

Yes, the double limit count is gone

So you're burning through your limit fast now that you're wired in

[oak trellis]

makes sense .. haha if i do biz 2 seats its worth it ?

[neat sinew]

You mean buy 2 Plus?

[oak trellis]

yes

2 seats .. Business plan

54 usd is ok based on the quality ..

[neat sinew]

If one can resume context from others

[oak trellis]

yes you can

[slow verge]

Hello everyone—are you able to use Sora 2 right now? For the past couple of days, my account has been showing “out of video gens.” when i gen video.

[oak trellis]

hmm now confused limits different i read biz has no more sub but is token based .. so better 2 subs plus ..

[young locust]

free users can no longer use it

[slow verge]

So do we need to buy the Plus or Pro plan, then?

[oak trellis]

question too .. would like to have like 50 usd plan ..

so the 5h limit eats up the total weekly limit ..

[leaden heron]

2 business subscriptions

its worth it but its like 2 plus subscriptions doe

same thing

[oak trellis]

so 2 seats are better ?

what you using ?

[leaden heron]

yes but you have to switch emails everytime for codex

[leaden heron]

3 seats in business

[oak trellis]

and you code all day never hit limits total?

i so 1 seat = 1 plus

[leaden heron]

well today i hit limits but i think its cuz of the rate limit bug

[oak trellis]

oh ok

[leaden heron]

ehh i think its a little more but im not sure

since i have used plus before and it was terrible

you also get access to pro which is pretty cool

[oak trellis]

ok makes sense. So in general you feel 1 biz seat is worth more then 1 sub plus

[leaden heron]

yes 100%

[oak trellis]

perfect will do that now

[leaden heron]

especially after they lowered prices

[oak trellis]

the 5.4 model is perfect for me right now (backend)

thx a lot for helping me out 🙂

last question if i update now the plus will be gone i guess ..

[leaden heron]

yeah youll just have to cancel your plus

[slow verge]

With the Plus plan, how many video generations can I make per day?

[plucky halo]

#sora-discussions

[oak trellis]

now i have that issue .. localhost refused to connect .. try to login and it gives me that

[wicked briar]

0

[wicked briar]

yo

I had some issue yesterday codex said token was refreshed please relogin

but it was after 2 months

[oak trellis]

ok so i just used different way login into the CLI .. but the extension is different hmm

problem iam logged into cli but not extension haha

[wicked briar]

had to turn off fast mode after 2x limits are over

[wicked briar]

that's weird

[oak trellis]

yes

[wicked briar]

maybe reboot

ide

[oak trellis]

man .. i should knew that too haha was bragging in cc about the speed yesterday ..

the code quality is same with fast off?

[oak trellis]

lemme try

[wicked briar]

yeah

I mean claude is still the useful to me due to front end task

codex just make slop

fr

[oak trellis]

yes frontend better .. backend GO codex 5.4 much better .. like almost made for it haha

[wicked briar]

yeah

thats why i have both

and will keep both

[oak trellis]

same same

2x codex 1x 100 max

[wicked briar]

waiting for new codex plan

$100 one

I wish codex team put some focus on front end

it just overengineer's simple stuff

its like factually doing what im saying but that UI is not appealing at all

[oak trellis]

yes agree!

[oak trellis]

the fast mode .. if set off in cli should adopt the same on extension right ?

[wicked briar]

yes

you should see the fast toggle in UI as well

[oak trellis]

yes found it .. was already on standard

[orchid plume]

yeah I was hoping it would be yesterday but we only got the codex-only pay as you go seat for business/enterprise it seems. I'm not sure if we'll see it today. At least in my country it's a public holiday today, not sure about the USA or the state that OpenAI's HQ is in though

Pro $200/mo expires on the 6th here hence why I was hoping it would show up 😄

[velvet marlin]

Hey all, any RN devs here?

[wicked briar]

yeah waiting

I actually got my pro renewed for this month but next month I'll downgrade

[orchid plume]

just checked California's public holidays, today isn't a public holiday so maybe we see it today (optimistic perhaps 😂)

[wicked briar]

let's see

bro casually saying v2 fr

I wonder cli or app

[orchid plume]

that's a good question, or he might mean the web version 😉

Codex app is version 20 something, CLI is 0 point something, so maybe he means website

[wicked briar]

hmm, btw is codex app improved? earlier i used it was taking too much ram and lagging when i had bigger sessions

is it worth using codex app or should I continue using cli

[orchid plume]

not sure, I stopped using it as I got tired of the lag when switching between windows briefly, or threads

[wicked briar]

same

CLI is the goat

[orchid plume]

true

[oak trellis]

hmm how the limits are calcualted so if iam on 5.4 high (not fast -> standard) ..weekly limits hits faster than 5 hours limits .. dont think so right ?

[orchid plume]

not sure if it's the same these days, probably is, but in the past it used to be that if I managed to somehow consume the 5 hour quota, I'd lose about 20% or 25% of my weekly

[oak trellis]

that would make sense ..

[orchid plume]

only time I ever consumed my 5 hour quota on Pro was due to a bug I introduced when experimenting with subagents on Codex CLI before they were officially implemented

[oak trellis]

i mean should be at least 2x 5 hour daily and 5 days ... ?!

ok eric back ..

haha

[orchid plume]

yeah, I'd say 4 or 5 times

[cedar skiff]

I just aim for 15% a day

[wicked briar]

aim for

💀

[orchid plume]

yeah that's a good target, so you have quota for the week

[oak trellis]

how many hours and which model which level

[wicked briar]

what plan you on?/

[oak trellis]

20 usd now

[orchid plume]

oof, limits on that plan will be noticeably lower 😛

[wicked briar]

I was gonna ask eric

[oak trellis]

you have 200 ?

[wicked briar]

forgot to tag him

[oak trellis]

plus -> biz (hella said is better) .. i just wanna consum from my girlfriend hers quick then thought biz

[orchid plume]

yeah Pro $200

[cedar skiff]

I'm on pro i use codex 5.3 xhigh, 5.2 xhigh, 5.4 high. i use it anywhere from 4-12 hours a day, also have minimax in droid/opencode and i have openrouter on both of those as well.

[wicked briar]

damn

bro is loaded

[oak trellis]

ok yes pro of course another levle haha

[wicked briar]

i only have codex pro and cc max 5x

[cedar skiff]

i did have cc max5 as well but i didnt use it enough so i dropped to the minimax and openrouter

[oak trellis]

cheap plus wanna migrate to biz .. and 5x claude

[oak trellis]

oh ok

[orchid plume]

I just have ChatGPT Pro and Qwen 3.5 27b locally at the moment

[wicked briar]

openrouter is api based usage

afaik

[cedar skiff]

yeah, but it has a bunch of free models you can use i put in 10 credits and get 1k req a day on free models

so atm i use qwen 3.6

It can do decent work

[wicked briar]

hmm

qwen is good

[cedar skiff]

i mean when i just have a simple change that covers a few files it works good

[oak trellis]

@cedar skiff is there any noticable difference between medium and high on 5.4 ?

depends of what i do of course ..

[cedar skiff]

little bit yeah, it just feels like it short cuts a bit on medium. It doesnt talk back as much ahah

like on high and xhigh it'll think it knows better and argue

I just give the old, i don't need your opinion just do the work!

[oak trellis]

true lol ..

i like that it reflects itself while doing stuff .. better than opus for what i do

[cedar skiff]

I think 5.4 xhigh can be a bit worse because sometimes for a big task it can get almost all the way to compaction without having done much out side of research... then it compacts and loses all the research. So high seems like the sweet spot

[wicked briar]

are you using fast

with high

[sleek spoke]

damn, I never thought of that

[oak trellis]

oh wow yes so 5h limits on 93% and weekly is on 97%

its consumes fast now lol

[clever steppe]

@barren sluice We absolutely need the ability to rename a Codex chat in VS Code. I have tried starting a chat with "Title this chat FOO", but it never works.

Even a simple /rename function would do the trick for now if it's too hard to work it into a context menu.

[sour burrow]

We need a toggle to switch back to the old 'concentrated sessions' way of consuming rate limit. I can barely get a single decently sized task down now without getting throttled to waiting another 5 hours.

[past prairie]

Guys do you prefer gpt 5.3 codex high or gpt 5.4 ?

[ember spindle]

dude... codex is SOOO OP bruh... also i like GPT 5.4 xhigh...

dude... rlly? especailly in the openai discord?

[kindred meteor]

My bad

i thought it was the JSS server

[wicked briar]

what server?

[high hawk]

Codex is very powerful and very good tools for developers.i tred this tool for 1 week and it is found to be very useful for my coding projects

[sand shuttle]

hwo exactly od iuse code review?

on the codex website?

[wicked briar]

you can use codex app or just /review in cli

[sand shuttle]

can i have ti review and maek changes to local files

[wicked briar]

yes

[sand shuttle]

i used /review in cli

but its consumping my normal usage

not code review

i gave it some instruction after /review review my local files..xyz instructions

[winged depot]

The code review balance will only be used, if you trigger the review on the GitHub website or something.

[sand shuttle]

ah ok

makes sense

[sour burrow]

Triggering code review via the proper channel does not make local changes FYI. It suggests repo changes in the cloud (GH).

Also you can use Codex Security for free right now on business accounts and that doesn't count towards your usage either - I highly recommend it. You only get 5 repo slots but worth it.

[obtuse slate]

Try rename in the codex app. May also rename in vscode

[clever steppe]

Unfortunately the chats don't even show up in Codex.

[obtuse slate]

What about when you start them in codex app? And then switch to vscode

[clever steppe]

I have not yet tested this, but thanks for the suggestion. I'll install Codex.

Has anyone got both Codex & VS Code installed?

[woeful crown]

Yes, Codex CLI and the vscode extension should share chats. They both write to ~/.codex/sessions afaik

[tiny kettle]

Anyone use the Codex Security thing?

Am I missing something on how to trigger scans? It should see when I commit and rescan, right? I made commits yesterday it's still not picked up on.

or am I just being impatient and it's like a once a week thing

[clever steppe]

Unfortunately nothing visible on Codex. :/ thanks though.

[safe holly]

How does that work? Any more info? What kind of deployment are we talking here?

[sour burrow]

Ours took awhile to scan the initial few thousand commits on a couple of projects and then new commits took almost 2 days to start filtering in.. but then it happened much more smoothly since. Doesn't really help, just saying we had something of a similar experience.

It's not once a week though, we get more frequent scans than that.

Also make sure you didn't inadvertently pause the scan. I created more than 5 initially not realizing the limit and it paused the first ones I set up.

[tiny kettle]

I've got I think 6 commits and ~45k lines between documentation and code in this repo. It's not that big. I didn't pause it.

I've only got the one scan set up

[ornate jolt]

realllly want them to come out with a 100 dollar plan so i can finally ditch claude its legit the only thing stopping me

[nocturne folio]

oh 2x isnt permanent? okay......

[plucky halo]

Now it's 2x/2x

[wide barn]

I'm on Codex Web, and Codex says it couldn't capture the screen because the browser_container is missing. How do I fix this?

[deft gyro]

wait what does this mean

[deft gyro]

depends on the task, but 5.4 mini is my primary model with 5.4 for planning and 5.2 for advanced tasks

5.4 mini is simply comically cheap and can typically do a single task

5.4 can usually take regular human speech to 1 to 5 tasks

5.2 can create something from scratch

you can see 5.2 = 5.3 codex per Mtoken

in reality 5.2 is slightly more expensive because it plans more, but if you are using it, it is for that very capability

all in all 5.4 mini is simply too cheap to ignore

[cyan wing]

Y no 5.3

[vital hill]

codex 5.3 only in the codex desktop app

[deft gyro]

5.3 codex simply doesn’t fill a niche I need anymore

It’s mildly cheaper than 5.4

But i could just use the far cheaper 5.4 mini for most of those tasks

Allowing me to allocate more compute budget to those high complexity tasks

Which I would rather use the more intelligent 5.4 or the more thorough 5.2

[vital hill]

i wish they'd release the ssh feature for the codex desktop app

[undone bobcat]

😭

When I hit the limit and i'm mid-task, like now, will it pick up where it left off when my rate limit resets?

[frigid sigil]

FWIW, I do. And VS Code and Claude on another machine.

[frosty zealot]

Which feature you referring to?

[vital hill]

waiting for the codex desktop app to be able to work on remote servers

[frosty zealot]

Oh gotcha

[vital hill]

the desktop app is fairly new so hopefully they'll release that feature sometime soon

[past prairie]

why only in the desktop app ? I use 100% in terminal

[vital hill]

because i can quickly switch to different projects via sidebar

instead of opening hella terminals

[leaden heron]

and cuz last month there was 2x on the codex app

[worn harness]

aw

where's my 100 plan

[past prairie]

ah ok but nothing different for the quality I got stressed

[vital hill]

cli and desktop app should be the same quality/performance. not too sure about the codex extension though

for vscode and its forks

[twilit bluff]

it will continue until the task is complete but if it does context compaction midway it will stop haha, and yeah it is able to pickup where it left

[cedar parcel]

i dont know how but this happened....

[cyan wing]

ya that happened to me too, i think that's when the subagents are requesting feedback
but the UI is bugged and won't let you provide it

[cedar parcel]

now see this spooky action at a distance.. check the weekly usage stat. What's going on....?

5h jumps around too

the token burn bug in live action?

[glacial shadow]

Is this confirmed to be in the works?

[cyan wing]

Well that’s concerning

[cedar parcel]

yeah to say the least lol

[glacial shadow]

what is the problem here?

[leaden heron]

yeah that 5hr usage and weekly is wild

[cedar parcel]

also, look at various terminals showing different stats for 5h and weekly

[leaden heron]

that depends on if you used codex recently in that terminal

its not live