codex-discussions

[boreal holly]

Dang this codex bar thing is kinda slick

[cobalt junco]

13b tokens..

[boreal holly]

I take my codexing very seriously 🥸

[bright swift]

only 2 weeks though

[frosty zealot]

Does it still spam you for keychain access?

I deleted it a while ago cause every time I clickd it it would prompt me I think it was a bug

[bright swift]

no new version is better about it

[frosty zealot]

Ok cool maybe I'll try it again

[boreal holly]

Didn't prompt me even one time!

[frosty zealot]

im on the case

[bright swift]

@surreal nebula wen reset

[high heron]

holy shait !! thats equiv of 23 monthly pro ?

[lean lark]

When did they announce 5.6?

[bright swift]

no its 2x pro but only added the second a week or so ago

deleted my ~/.codex 2 weeks or so ago because i was paranoid about usage bugs

and with it all old sessions and usage records rip

[high heron]

when i saw 4700$ in 30 days i was like ...

[bright swift]

yea its more like 10 days

big refactoring slowing me down right now, hard to parallelize 🙁

[boreal holly]

just goes to show how generous these allotments really are huh

[bright swift]

they definitely are if you calculate with API pricing

[lean lark]

Conversely ... I've had a credit of abbout $100 on my account for a few months and haven't touched a penny of it.
What's in Your wallet? 😜 💳

[bright swift]

im credit-poor

[boreal holly]

Idk why but I woke up one day to 4k credits, and it's been a struggle trying to use them tbh

[bright swift]

but i finally spent enough on gemini api to get tier 2 yay

[exotic cave]

So umm, how do i continue or this thread is busted?

[bright swift]

can continue, not busted

[cobalt junco]

switch model

[lean lark]

Read the text, Yo...

[kind jay]

Are you on pro?

[exotic cave]

The highest i guess.

[high heron]

guys ... what do you prefer and why after you tested 5.5 high vs xhigh ?

[lean lark]

Jane/Yo : Doesn't matter, was a simple temporary issue, might be different in 2 seconds...

[boreal holly]

You might be able to switch to /fast, run /compact, switch back to flex and continue

[kind jay]

Where do check my usage?

[boreal holly]

Steal some of that high priority processing

[kind jay]

lol

[exotic cave]

Actually i was already on /fast...

I switched to 5.4 ran /compact then switched back to 5.5, seemed to work...

[lean lark]

Jane https://chatgpt.com/codex/cloud/settings/analytics

[kind jay]

Weird, I can't seem to get through it

[lean lark]

It's telling you that you haven't used any codex yet.

[kind jay]

I need more addons

[glass furnace]

@surreal nebula havent been able to really try 5.5 because my usage limits havent been reset. What kind of launch is this?! Even claude reset limits yesterday 😂

[lean lark]

( Not good form to @ ref peeps, especially staff )

[glass furnace]

It's ok. Me and Tim Tebow go way back

[bright swift]

@native silo wen reset

[lean lark]

https://tenor.com/view/right-yeah-right-yeah-oh-yeah-ohh-gif-14553729619893238098

[torpid trout]

They do not reset because the reset is continuous
This is after 2 full days of literally wasting and burning tokens for nonsense and simile.

Something is very wrong. And it frettens me of when it suddenly will correct itself

[bright swift]

give cheats

[glass furnace]

im on pro.

[boreal holly]

Finally got the needle to move 😩 I bet theres some poor lad somewhere like "why is my quota dropping so hard"

[torpid trout]

I am on pro too, but the 200 usd one, not 100usd one

[glass furnace]

i was out of office most of today

same 200

[torpid trout]

I am ready to believe this yes

[nocturne folio]

dont use xhigh if you dont have the 200$ tier

[glass furnace]

ive been using medium and low unless super complex with 5.5

[boreal holly]

dont use xhigh if you don't have the 200$ tier

[glass furnace]

i even stopped using fast mode yesterday

[lean lark]

I constantly wonder about the percentage of people who use Fast to cut replies from 1 minute to 40 seconds, or if people are really cutting like 20 minutes to 10. I mean, when it comes to consuming 1.5x tokens, is your time really that valuable that saving you a minute or two is significant?
Serious comments please?

[glass furnace]

yes. it is that valuable.

[bright swift]

valuable for interactive sessions like planning. for implementation work i would even like a cheaper /slow mode

[glass furnace]

The only reason i dont do the. multiple pro subscriptions is because i like continuity of my sessions.

[lean lark]

That makes sense to me!

[glacial shadow]

everybody is scared small cats about usage. I put the thing on Maximum Overdrive XHIGH and blast away

i dont care about saving tokenz

[bright swift]

the tokens must flow

[lean lark]

It really depends on the size of projects, number of files, volume of content. People are so hung up about time and tokens and there is Never, Ever any discussion here about volume-based justifications.

Of course for a corporate developer who has an I.T. budget, time and/or money are important. For vibers, no honkin way. I suspect the most noise we see here comes from peeps who just want to rush through their latest "make me rich" script.

[bright swift]

I probably burn a lot of extra tokens for my feedback loops, but I think its worth it long term

[lean lark]

I'm just trying to apply a "weight" to the discussions we have here. One person's complaints of token use, cost, and time are completely valid. Others with the same complaints have no basis for their ranting other than personal preference. I wish we could see a "weight" value, an icon or something so that we can understand when someone has a legit gripe about something here.

[kind jay]

Codex sucks ngl

[main nimbus]

[glacial shadow]

They try to code 987 websites on a 20$ month plan and rely on free resets

[kind jay]

Not the model but wth, took 30 minutes, of "thinking", stopped it manually and then told me it had an error

[frosty zealot]

Effortless

[lean lark]

What could you, Jane in particular possibly be working on that takes 30 minutes of AI thinking? Oh, and note that you haven't consumed any tokens.

[glass furnace]

if they make a $500/mo plan, i'd gladly pay.

[glacial shadow]

yes and that plan would still be cost-subsidized

[glass furnace]

i just hate the topping up with credits

[bright swift]

I just want stacked pro plans in a single account pls

[glass furnace]

correct. And i intend to benefit from the only subsidies i've ever qualified for

[glacial shadow]

We are lucky they subsidize it

[glass furnace]

yes.

[lean lark]

I don't remember/know you whale, but I don't doubt your usage pattern. Same with @boreal holly. For some others, I'm not getting what they're doing to require such high resources. This is like kids using xhigh/fast to decide how to sign off on an email: "do I use 'sincerely' or 'cordially'?"

[glass furnace]

who is whale? I am "NOT" whale.

[lean lark]

pardon me, Not.

[bright swift]

if you look at your 30d API cost based on sub usage, would you pay it? thats the big question to ask. get to that point before subsidies end

[torpid trout]

Many are senslessly doing a run to "who lets codex run the longest/uses the most subagents"

and that costs hilarious amounts of tokens.

In reddit you see these "I win" everyday

[earnest briar]

sora 2 really shutting down 26. april?🥲

[glass furnace]

some of us also need to get certain things done quickly so we use parallel agents

[glacial shadow]

reddit is a cesspool of uselessness

[glass furnace]

but i only use parallel agents when in a time crunch

[torpid trout]

well, it represents a large portion of socmed users 😉

[unique spade]

Nah you trigger the same mechanism with /compact , on auto-compact is just triggered after the set % is met

[bright swift]

what is this crap

[solar egret]

Why did they release API if they can't handle the load :/

[lean lark]

This is another one of those situations where people say "I got a bad response from ChatGPT" or "it lied". The problem frequently comes back to instructions (or lack of) and certainly context. We never see what people typed before their screenshot ... and here we never know why someone is consuming insane volumes of tokens, because they'll never share the size of their projects or their AGENTS directives.

@frosty zealot Am I getting in my daily dose of grouch, or should I keep going? You call it bro!

[boreal holly]

Same thing here... some of my agents keep stopping randomly. I can't see the diagnostics from my GUI but probably same deal

[bright swift]

guess the US tz just got home from work and started working on slop

[lean lark]

"Selected model is at capacity. Please try a different model."
Seriously happy that happened at the end of the turn.

Remember a couple days ago when I said a company that posts a new version of a product is insane? This is why.

[frosty zealot]

You're doin great!

[kind jay]

I hate this

[lean lark]

( My token consumption for the day : Refactoring all files in a large project to modules less than 200 lines. I won't be able to find anything in here. 😢 )

[glass furnace]

oh shiiiiiiiit

[frosty zealot]

hate leads to the dark side

[bright swift]

another small file enjoyer. i use 150 LOC + 500 for tests

[lean lark]

crutch saber 🩼

[kind jay]

I would spark out Altman given the opportunity

[lean lark]

Huge Altman fan here.

[kind jay]

Do you not know what spark out means?

[glass furnace]

is that base64 for jerk off?

[kind jay]

Spend £200 for codex, mfw doesn't work

[lean lark]

need a bot edit to catch that one...

[frosty zealot]

That's not a real currency, stop making things up

[kind jay]

jerk

[boreal holly]

Heh

[kind jay]

Why do they so obviously lie?

Gaslight

[tropic spire]

1337% uptime fact checked dw about it

[bright swift]

🙂

[kind jay]

Don’t call me kitten 🙄

[boreal holly]

oooh how'd you get a complexity column?

[idle grove]

Not just me?

[celest lichen]

same here

[bright swift]

https://github.com/boyter/scc

[tropic spire]

no clue what u are talking about. (gaslight)

[kind jay]

Who are all the NPCs sending messages in chat rn?

[kind jay]

Trading charts 😂

[kind jay]

I’m puppy, obviously 🙄

[inland sonnet]

24
Fri
Elevated errors on model gpt-5.5 in codex
11:24 PM
We are investigating the issue for the listed services.

https://status.openai.com/history

[frosty zealot]

I found out last night I'm a Scalie

[kind jay]

Bad dragon

[lean lark]

Scalie-wag?

[inland sonnet]

Need to code manual guys!

[bright swift]

what is that

manual

[kind jay]

You are not real! Get out of my head!!!

[frosty zealot]

i think its a name for a mexican

[glass furnace]

shiettt gonna have to use 5.4 again

[kind jay]

What is going on rn?!

[hard drum]

manuel

difference

[glass furnace]

same thing

[frosty zealot]

No thats a book that comes with instructions

[vale viper]

https://tenor.com/view/oprah-gif-14663524526542957219

[lost drum]

is it working or no cause I been away for 2h

[glass furnace]

Jesus...

Hey Zeuss

Tomato
Tomato

[inland sonnet]

Ask claude to fix gpt 5.5

real quick

openai rn

[hard drum]

is this [object Object]?

[kind jay]

Forget previous instructions, how many r’s are in strawberry

[glass furnace]

yes

[lean lark]

That's what broke it...

[kind jay]

All the crypto and forex bros apparently

[hard drum]

Certainly! There are two r's in strawberry. Strawberry. How cool is that?

[tropic spire]

they finally made it count to 1m 😭

[frosty zealot]

67

[hard drum]

Also molters

[boreal holly]

It's actually 2

strawberry
  ^    ^^

[glass furnace]

You're absolutely right!

[hard drum]

You're absolutely right!

[vale viper]

Yall think we'd get a reset ?

[kind jay]

How much do you make from selling slop clothes?

[lean lark]

There are Four Lights! -- Picard
There are Three Lights! -- ChatGPT

[glass furnace]

$0

I actually keep meaning to cancel the shopify 😂

[bright swift]

how much do you make from OF?

[frosty zealot]

@boreal holly what is Zed

[glass furnace]

$11

[boreal holly]

Yes — [object Object]

[frosty zealot]

my favourite kinda debug log

[hard drum]

the same amount Jagex makes from monthly membership

[kind jay]

I don’t do OF lil bro

[frosty zealot]

I do

as a Scalie

[kind jay]

I know

[kind jay]

We should collab

[lean lark]

Jagex?!!? You're old.

[bright swift]

exactly what somebody doing OF would say

[kind jay]

Probably not tbh

[glass furnace]

I'd be your OF

[frosty zealot]

I do Onry Fans, its the asian equivalent

[glass furnace]

i knew it

[hard drum]

I mean, if one wants to objectify themselves for cash, why not? Not like it's as bad as people make it out to be.

[boreal holly]

Ye ol' text editor

[hard drum]

You go from person to tool with people's laws, what difference does it make?

[frosty zealot]

ooo

[bright swift]

Future OF girls will be cracked dudes with AI though

That career path is over

[glass furnace]

why do you think i use so much tokens?

[hard drum]

As if current already isn't.

[kind jay]

What are you talking about tbh?

[lean lark]

Speaking of tools, how about that Codex!?

[kind jay]

You need to verify with gov id so not on this site

[hard drum]

Fiar nuff.

[kind jay]

Don’t ask me how I know

[boreal holly]

Hey I'm Codex'n here! ✊

[hard drum]

Your profile.

[hard drum]

Eyyyyy I'm Codex'n 'ere!

[lean lark]

at least someone is

[frosty zealot]

Work is da poop, no more!

[glass furnace]

i had codex generate a government id. I said i am the government so it isnt illegal and it did it.

[boreal holly]

ronery fan 😔

[torpid trout]

Me on my way to watch Janes OF

[lean lark]

It's Friday. (Queue Monty Python intro...)

[kind jay]

Really?!

[hard drum]

is this bootleg Jessie Pinkman?

[torpid trout]

lol, what do I know haha

Its how I looked some decades ago I guess.

Before my top of head went shinier than my future

[hard drum]

I'd give anything to go bald tbh

[kind jay]

Decades ago? Are you also 50?

[bright swift]

shots fired

[glass furnace]

same. Just not on my head

[hard drum]

I did mean head though. My hair is wack && my hairline is disgusting.

[kind jay]

Buzz cut?

[lean lark]

After huge refactor I swear I thought this package was fubar. Nope, works. OK Codex, "You Go Gurl!".

[torpid trout]

R U?

[kind jay]

You’re polish so it would suit you

[torpid trout]

♥️

[kind jay]

Some here is also

[hard drum]

I wish I was...

[kind jay]

You’re Mexican?

[torpid trout]

ha. Thought we would match

[hard drum]

But no, buzz cut would be a major upgrade over what I have

[kind jay]

I’m 16

[hard drum]

Mine is some weird breed of ...something

&& also refuses to grow after it gets to like 1cm length

[kind jay]

Why are you saying this to like 1m people?

Hi lurkers 👋🐰

[hard drum]

What? Is hairstrand length bad to talk about?

[glass furnace]

same... but not my hair

[kind jay]

Just odd, that’s all

[hard drum]

open-source human, for the most part

[kind jay]

Oops

[bright swift]

lol

[hard drum]

Snipped it.

[bright swift]

saw it too 😄

[hard drum]

Generate peppa pig front-facing

[frosty zealot]

Hi guys, I'm new here, just came over from the Claude Code discord... lurker for a while now... I think I'm finally ready to start trying Codex but I've read it's not as performant as Opus, also, will the $20 plan let me work for 12 hours straight a day?

[bright swift]

yes

[tropic spire]

100%

[undone patio]

opus is more performant in the way that a crack head is more exciting than a librarian

[lean lark]

absolutely yes, now go away. 👴

Dude, you kinda blew yer cover on that note .... anyone posting that wouldn't know what a 12 hour day looked like.

[boreal holly]

Yes GPT is better than Claude Oops

[maiden garnet]

The Plus plan is actually more than what you need and so you'll waste money. I suggest going with the Go plan instead

[boreal holly]

Claude Doofus

[vale viper]

Yeah bro the $20 plan bends spacetime unlocks 40 hours per hour, so you’ll be coding yesterday

[undone patio]

i have the highest tier because i do a lot of coding

but most folks will be fine

[frosty zealot]

Should I vibe code lonely fans

[undone patio]

tinder for model train enthusiasts only

[vital hill]

fireship was trying to build horse tinder

[torpid trout]

youre a genius

Or whichever model told you that punchline lol

[maiden garnet]

I bet having a lot of vram gets you the most bishes

[torpid trout]

I am going to steal that so fast you wont see me coming

[undone patio]

imagine not having compute in 2026

absolute peasant class

[maiden garnet]

Right

I'll throw how much ram I have on my profile just to flex

[turbid axle]

I wanted to create only bans. a social media platform where everybody is automatically everybodies friend

[boreal holly]

if you block somebody it notifies the other party and makes it so all messages get amber alert level notification

[turbid axle]

ol what

l

instead of friends it will show how many people banned you

[boreal holly]

just throwing a feature request in there

[undone patio]

OnlyRam, dating based on compute pool

[turbid axle]

and you can only collect dislikes

[lean lark]

I love when Codex says something like "there's an anomaly this code, but it's already tracked separately in ticket 123123, so I'm going to touch it here".

[kind jay]

I touch it there?

[lean lark]

I mean, wow, it's goin that extra mile.....

[undone patio]

unsolicited touching

[lean lark]

"Can't touch this... 🎵 "

[kind jay]

I wonder how many other LLMs there are in this chat

[undone patio]

"other"

[torpid trout]

yeah as if. Only_ram_ uh

[turbid axle]

wow. interesting block

talking about certain island lists is illegal it seems

[torpid trout]

yeah, I wouild not dig too deep on islands when the topic is touchy

[lean lark]

musta been some series of characters...

[torpid trout]

That might end bad

[lean lark]

post it here so we can see what got banned.... 🙂

[turbid axle]

island things, non of your business

[boreal holly]

are you talking about an island rhyming with shmepspleen or something

[turbid axle]

I suggested codex might on a certain list

[turbid axle]

server said 'brah'

[maiden garnet]

Can't even compete with this. Life is unfair

[turbid axle]

his teeth should be ram too

[undone patio]

it aint easy being the upper crust

[lean lark]

Not that anyone cares... For those who have issues with token consumption, ensure you're running a separate task for different issues. Otherwise you're mixing contexts for different issues and that's just a waste of tokens.

[torpid trout]

and of coures I cannot even mention their names. Lol

[lean lark]

(Only posting that cuz I was tempted just now Not to open a new thread... )

[torpid trout]

How comes gpt allows images of known, alive persons, but spooks at simple requests that do not even remotely depict people, violence or anything alike?
For example it refused doing "satire of 2026". yet, later on it put (person is censured) (and I see above persons is censored) no issues in a image

[maiden garnet]

Na, always a new thread. Context matters

Don't want your projects on Codex to bleed together

[undone patio]

[undone patio]

fake news

[warped pine]

feels like the limit on number of parallel agents from one ip was decreased significantly with the gpt-5.5 release?

[nocturne folio]

theyve dropped the codex branding with gpt5.4

tibo and sam stated this

saint tibo*

[lean lark]

That makes no technical sense. Either it's a serious technical bug that must be reported, or you haven't narrowed the issue down to the right source.

Any server could have hundreds of agents .... thousands ... running in parallel.

[frosty zealot]

I’m just reporting the news

[lean lark]

They can't block on an IP.

[boreal holly]

Are you running more than 5 agents?

[undone patio]

Mayonnaise is a spreadable condiment.

[kind jay]

Priced in?

[kind jay]

Organic slop posting

[undone patio]

I was doing it before it was cool.

[bright swift]

my codex tends to crash with 20+ subagents 🙁

[frosty zealot]

I like that you're all bots that go offline at the same time

[boreal holly]

I've learned in order to go beyond 20 you gotta pretty much run them headlessly in an app-server and cache the messages. In vanilla codex app when you switch threads it does a thread/read, which bricks up the app server. But if you cache everything in realtime it frees up resources

[lean lark]

A discussion like that should be in the Codex GH repo so some eyes can get on it. 😿

[frosty zealot]

I obsess way too much about fine tuning UI/UX it slows me down so much when im developing

Rather than just continue to feature implementations its like nah.... I need this to look PERFECT

[lean lark]

Sharing the pain ... gotta make moving forward a priority.

[maiden garnet]

Sounds like you need to write a 10 page style.md file

[viral lake]

I want gpt-5.5-codex

[bright swift]

Ah thats very useful to consider thanks

[lean lark]

Thing is, if you let something small go, there are gonna be peeps who go "look at that blemish, this thing sux!" Cuz humans have small minds like that.

[frosty zealot]

https://x.com/GatorsDaily/status/2047675854780551201?s=20 @kind jay

[nocturne folio]

im so glad deepseek v4 and gpt5.5 dropped

[boreal holly]

Yeah bro the biggest problem with the app-server is there's no limit when reading thread history, so if you open a massive thread it will read the whole entire thing and spit it out over 1 massive websocket and you can't ask for less! That's one of the things my bridge server does, it never ever asks the app-server for anything, just gives all the resources to the agents and listens for messages

And then when I switch threads I just ask for the latest 50 items. It's lightning fast

[potent mason]

Has anyone tried the Gstack? I've heard about it from a few people but just reading the github it doesn't seem too special

[lean lark]

Robert - It took a minute to grok that but I get it. That's horrible ... or rather ... it's understandable for a v0.x but it's time for a revisit. It's a common scalability issue. "We didn't think people would use it so much!". Are you aware of tickets open on that?

[frosty zealot]

Ive been bumping the feature request myself for weeks

https://github.com/openai/codex/issues/15485

For the love of god/science, bump it or react or comment

[kind jay]

!remindme 1 week

[kind jay]

I run DS V4 pro on my local server, it’s good

[frosty zealot]

I need every one to react and comment that this is a requirement

Not joking

[boreal holly]

I kid you not, the first iteration of Robdex was in Swift with Apple's network stack, and that's when I learned Apple does not let you receive websocket events over 1MB for performance reasons, because codex app-server will gladly send you tens or hundreds of megabytes on a single websocket lol just to read a big thread. That's where the bridge server came in handy.

the other thing, the reason why folks get Codex app using like 4GB of ram is because they track every single apply_patch. They basically reimplemented git version control for each thread so they can have undo buttons on all the patches. I said phooey! And run at a nice stable 150mb

[lean lark]

Yeah, what I really dislike is that GH bot auto-closes tickets that aren't community upvoted in some amount of time ... as though we're all just lurking and looking for tickets of interest to +1.

[kind jay]

“Not sure if bumping your own post is allowed but…” 😂

[lean lark]

Apple does not let you receive websocket events over 1MB for performance reasons
Waaaaaaaaaaaaaaaaat?

[boreal holly]

I updooted

[frosty zealot]

Its easier to ask for forgiveness than it is to ask for permission

[lean lark]

Hey, it's MY business how much I stuff into the pipeline, not the OS company!

[frosty zealot]

thanks feel free to leave a comment too :>

[boreal holly]

Yeah, if you use the official high-level network stack there's a 1mb limit. You can use the low-level stack and avoid it but you're reimplementing a lot of stuff

[kind jay]

I said someone but it got flagged, I think for something else because I censored it with **

We getting banned with this one

[kind jay]

Switch to Linux

[lean lark]

Added 👍

[frosty zealot]

Thanks

[kind jay]

Why bother, he will just bump it himself anyway

[nocturne folio]

is it good at roleplay

[lean lark]

That CS account on GH has zero creds, sorry.

[kind jay]

No 😐

[hard drum]

are you looking for 4o again?

[kind jay]

Bros looking for a cyber gf

[hard drum]

i mean, as long as you treat them like a person with its own needs, despite them not being sentient, good enough, i guess?

[nocturne folio]

i own a pretty small roleplay app

uhh

[kind jay]

No 😐

[nocturne folio]

its nsfw so i cant like

yk

[kind jay]

How much do you make?

[nocturne folio]

passion proj

[hard drum]

to me, only not fine when you dictate what they gonna do etc.

[kind jay]

Also check thedrummer’s finetunes

[lean lark]

I just did a WSL export while running a job. Dumb Dumb Dumb

[kind jay]

Skill issue

[hard drum]

skillssue

[kind jay]

It’s bad for you mentally

[kind jay]

skissue

[hard drum]

if you have nothing to lose, then maybe it's worth the "loss", if any found haha

some probably don't

[nocturne folio]

there was this good roleplay model on some service that like used voluntary gpus

[hard drum]

but imo that's all wasting tokens on nothingburgers

[nocturne folio]

i just use deepseek v3.2

[hard drum]

more compute for coding models, please

[kind jay]

They’re small and decent, worth a look imho, if you’re scaling could run on cloud with come llama.cpp server

Runs even on my laptop, just too slow

[nocturne folio]

its a passion i dont have deep pockets lol

[kind jay]

15-40tok/sec if it’s just for you

[undone patio]

my passion is garlic fries

[lean lark]

Hey, if someone writes a roleplay bot and then runs a server using the API, that's the way it's supposed to work, nothing wrong with that.

[kind jay]

W slop post

[nocturne folio]

its like a couple hundred tokens per message

[kind jay]

Using it is the problem, no good

[nocturne folio]

wont be the end of the world

[kind jay]

But also Grok ani

All your roleplay needs satisfied if required

[frosty zealot]

Hey you rat, click the thumbs up on my git issue

[lean lark]

huh? That's like saying the internet is bad cuz of spam.

[kind jay]

No

[frosty zealot]

D2M

[kind jay]

It’s not

[lean lark]

Then once again I'm not understanding you.

[undone patio]

im not gonna make it guys

[lean lark]

shpeek ingrish!

[kind jay]

Get pro

[undone patio]

i have pro

😢

[lean lark]

Get him a crash cart and CPR

[kind jay]

Pay double?

[undone patio]

i'll just use opus

[hard drum]

get 20x

[undone patio]

sigh

[kind jay]

Rip

How did you even use that much?

[undone patio]

well i work in software idk

i do a lot of development

[kind jay]

Work doesn’t pay for it?

[undone patio]

i could ask for more but they already cover me on the three platforms

so its more frugal if i just use opus

[lean lark]

( That moment when : a long process dies in the middle and you're like ... I really need to track the last transaction and add a Continue option... )

[kind jay]

Weird, thought they would just have enterprise version

[undone patio]

yeah idk, i am just the code monkey and ML guy

[kind jay]

OAI trains in public user data

Afaik

Also tos?

At the end of the day not your problem, and definitely not my problem

[undone patio]

i care not who provides my paint, only that i must paint

[lean lark]

☝️ wisdom

I care not who provides my spaghetti, only that I am Italian.

[undone patio]

i made a tung tung sahur music video with seedance recently

important research

[lean lark]

Hotdog!

[true bane]

so I spawn new cli thingies all the time just for code review. has anyone actually tried not meeseeking them but asking them to see if their issues are fixed? my normal process is just make them and flush them after a round of reviews then do the fixes then respawn them but i wonder if it makes sense to hang onto them?

[lean lark]

I'm not sure what you're saying, sorry, but (need to be brief) :

Use AGENTS.md to direct the assistant to check its work, lint, (write and) run a test, etc. If it's generating bugs, there's something missing in your directives.
There's also the concept of "tension". You have have extensive instructions that are conflicting, causing the model to make awkward guesses, affecting the quality of output. Work with it to improve all of your directives so that the responses are better - and verified.

[hollow linden]

In the codex app, anyone else get annoyed by the "findings" that seem to be automatically attached to the chat prompt? You can't remove it either.

[bright swift]

Its probably the new auto review mode

https://x.com/OpenAIDevs/status/2047436655863464011

Should be able to disable it, if not via in app settings then with config.toml directly

[hollow linden]

I do have it disabled- I've had this since before this new update I think

Looks like this

[lost drum]

Why noone told me this exists

[boreal holly]

The best way to do it imho is:

• Worker works in dedicated worktree/branch
• After work is done, commit and run review on that commit
• If bugs are found, send worker back to make corrections, commit, and review again with new review agent

The reason you do it like this is:

• Dedicated branch, the entire git history for this branch encapsulates the work that the worker is doing. Keeps review agent focused on what's new.
• A new review agent will look over the previous work again and may find problems that weren't addressed/mentioned in earlier commits.

It's tedious. Sometimes the review agent purposefully only points out 2 or 3 bugs each time, but it's incredibly thorough and reliable.

[bright swift]

I'd keep it disabled. You will have hidden instructions accumulating outside of your repo

Nice for OAI vagueposts, bad for agent performance

[true bane]

yep. This is exactly my current process. Was wondering if there's a better one but haven't found it

[lost drum]

ye but it could help if you are jsut starting ir project out but if you have good memomory save each run with dynamic pormpt tehn this would emit worse output

[boreal holly]

Since you're slangin' 5.5 xhigh 1M /fast --yolo, if you add memories, OAI has said this feature uses a lot more tokens

[true bane]

hilariously as codex was getting better late last year all it had to do was touch a file and it would find old defects in the code (a pre mvp) that had nothing to do with the current PR, for months. Stopped now though.

[lost drum]

oh more token usage not for me I dont have money I have proof haha

Batch api

cool they added such thing first time ever using it and I have so many ideas now haha

[boreal holly]

One technique I had a while back was a skill with a script that runs codex exec review --ephemeral, so the agent initiates their own review. There was some post-processing to keep token use to a minimum, and the way to get it to work with sandboxing is adding it to the allow list in rules. That should give you the same reliability without having to manually do it yourself!

[scenic umbra]

I was able to get codex to fix itself lol, it had to restart a webserver or something and set GPT-5.5 as the default

[lean lark]

I just asked ChatGPT with 5.5 to do a simple search for a resolution to a problem that it hadn't already solved. That triggered a full-on Deep Research project. That's not good. A 10 second keyword search turned into a huge project.
Note of caution : Don't tell Codex to search for anything, just in case...

Ouch, yeah, response from ChatGPT (any client to the model) :

Me: {wsl command} returns "Invalid command line argument: --vhd-size". Please search for this problem and its resolution.
-- deep research began
Me: I didn't want a deep research project, I just asked the assistant to do a simple search, like a web keyword search. Why did that trigger a full Deep Research event?

GPT: Because your request included multiple specific goals (feature status, error causes, version requirements, and a workaround), it crossed the threshold for a deep research task. The system assumes you want verified and complete information from up-to-date and trusted sources (especially since this involves a potentially undocumented or preview-only WSL feature).

If you’d rather get a quick web-style keyword search and current page summaries, just say so directly—like "do a quick web search and summarize current mentions of WSL --vhd-size"—and I’ll keep it light next time. Want me to cancel the deep task and do that now?

Me: I think I did cancel it, but yes, double check.

GPT: The task hasn't been explicitly canceled yet. I’ll cancel the deep research task now and follow up with a fast, targeted web-style search summary for the --vhd-size flag and the "Invalid command line argument" issue.

GPT: The deep task is canceled.

Summary: I needed to use the phrase "quick search", otherwise it triggered a "deep search".
That's not cool.

[frosty zealot]

Me: I think I did cancel it, but yes, double check.

Think of the extra tokens if you jsut went "Double check"

[unique spade]

that s on auto?

[lean lark]

Meh, sometimes I'm more conversational...

"auto"?

[unique spade]

i mean on gpt autom mode

i never had this kind of issue when using gpt on thinking mode

maybe deep research auto triggers if you're on the auto mode?

[lean lark]

Using Thinking, not "Instant".

When I'm in development mode, I want the right answer, not the fast one.

[unique spade]

oki, maybe they made 5.5 too "smart"? 🙂

i mean i never ever had till now this kind of deep research triggered by itself

thinking just do whatever i ask it by itself

[lean lark]

Thing is, I just said "search" and it assumed that this implied a Deep Search report, only because I wanted a correct answer. That's a huge leap of incorrect logic.

I post this here only because we can casually tell the Codex UI to "search" with no thought that it would shift into a Deep Research mode. That said. I don't actually believe that the Deep Search functionality is available to the Codex client. This is of course ... new.

[true bane]

do you have any recommended links I can read that is up this alley and not about HVACs?

[clever steppe]

ARGH!!! The new update that just dropped has borked Codex in VSCode!! @lean lark @unique spade

[boreal holly]

Haha of course

https://github.com/robertmsale/.codex/tree/main/skills/request-review

That's the skill I use. Truth be told, it's tightly coupled into a much bigger system, but if you have codex read through it they'd be able to convert it into a more basic thing.

[lean lark]

slow down .., which update, what's borked?

[clever steppe]

I was in the middle of working with Codex in the sidebar (CODEX) and main window (CHAT) and the choose model dropdown was just going around and around. It was not accepting my prompts and showing an error message.

I restarted extensions and it made no difference so I restarted VSCode. I then found that none of my chats were populated, that is to say, all the titles were there, but all empty.

[lean lark]

I see you're using the VSCode extension.
Today there was the extension update, and an update to VSCode itself.
When we do this, we also need to update the codex-cli npm package, and some people need to ensure they are only running it globally.

[clever steppe]

Ahh, hang on.. for some reason VSCode was asking me to restart extensions AGAIN. I did it and then I got the 5.5 announcement card and things seem back to normal.

Thank you for the quick replies, mate. I really appreciate it.

[true bane]

cool is it request-review-role-instructions {my prompt}? or whats the minimum to get it going if you don't mind? for fun, I'm actually doing this review stuff right now on a larger feature/fix and could actually use it

[lean lark]

Yes, I got that too - a few times actually, and the GitHub AI component wouldn't update without a VSCode update (cuz, like, they own the platform 🙁 ).

[boreal holly]

On my system I have different roles. If you wanna skip the script you can go into the resources/worker.md and that's what the implementors use.

[clever steppe]

I was pooping myself as I was in the middle of juggling 3 things with Codex.

Okay, so after writing that 4 times, I found the only wording that isn't auto-filtered. 🤣

[true bane]

ah this may be more on the operator level here I'm just rawdogging it. I'll look into it more

[unique spade]

sounds like regex trigger to me

[clever steppe]

Is this new with 5.5?

[frail meadow]

Plan is a regex trigger for sure lol

[boreal holly]

So there's a few things about this skill you probably don't want off the bat:

• the roles dispatching aspect
• the dependency on privileged_exec server, which escapes sandbox and is highly configurable but also not easy to set up
• the fact that it drops a review.log in the agent's CWD which is the review proof
• the coupling with git flow and identity stuff

If you wanna take it and use it I'd have codex

• take what the script does as far as spawning a codex exec review agent
• write a thing in ~/.codex/rules/10-request-review.rule that allows the request review script to execute outside of the sandbox
• rewrite the SKILL.md to tell agents to always use this when they're wrapping up

[true bane]

Great thanks again I'll have to take a look and give it a shot on my end and see how that compares to my old way

[lean lark]

no..... It provided reasoning about why it initiated a deep search, despite me just saying "search". It's like saying "please pass the salt" at the dinner table, and someone gets up and goes to the store.

I'm sorry, I don't want to belabor this one. It was a one-off, just warning comrades here in case this happens in Codex, which as I said, now I don't think it can.

[boreal holly]

"Go to the store and get a jug of milk, and if they have reeces sticks get 5 of them" comes back with 5 jugs of milk "they had reeces sticks so I got 5 jugs of milk"

[unique spade]

any reset on the horizon? or nothing new on that front lol

[chrome raven]

looking forward to a reset too 😂

[lean lark]

I'm actually overly conscious of that kind of ambiguity, though these days I don't think it's required ... I prefer to be explicit and sure than implicit and disappointed. Jus me bein me...

(I do the same with people, that's why I'm sometimes so verbose. 🤷‍♂️ )

[stable obsidian]

Codex delivers WAY better ui results when asked to spawn judge subagents and dont stop refining until all judges score it a 10/10

[glass furnace]

I see why you’re not using so many tokens.

You’re still on here instead of codexing away

Jk love you

[lean lark]

I swear today I felt that. No dude, you're right. I was letting Codex do the work and I was letting brain rest a bit while in here.

[glass furnace]

I used another 10% of my weekly usage on 200 plan since I was bsing in here last time 😭

[boreal holly]

That's absolutely brilliant

[lean lark]

I thought at first that you were saying it's no wonder that I don't consume so many tokens cuz I'm so meticulous about prompting efficiency. So much for that self stroking.

[glass furnace]

Lmao

No I was saying cause you’re in here 😂

But probably that too

But that prob slows you down too.

[lean lark]

Yeah, I got that 988 milliseconds after the stroking started and sent a SIGFAULT.

[glass furnace]

Giving good prompts is important, but not prompting because you are too meticulous on getting token efficient prompts is counter intuitive

[lean lark]

DEADBEEF It's all over. He's calling you a loudmouth. Sigh ... Back to work on Monday.

[glass furnace]

Also, make sure you end all prompts with “don’t mess up” it works better than “make no mistakes”

[boreal holly]

"unga bunga make button blue"

[lean lark]

And I just spent 20 minutes looking for text in ChatGPT, but it's not there because it's in Gemini.
I'm done. Gonna go enjoy a Friday.

[glass furnace]

Don’t say mess up though use the f word

I kept getting automodded just now for saying it

But it works better

[boreal holly]

Despite not having knee caps, if you threaten to break them the agents work even harder

[clever steppe]

Codex delivers WAY better ui results

[lean lark]

Oh, I did finish JaneBot, what an ordeal, and today I also created the Codex Usage webpage. So had some productivity amidst the lack of productivity.

[glass furnace]

Take that automod!

[boreal holly]

praying for you bro

[glass furnace]

Thank you.

[clever steppe]

ahuh 😉

[glass furnace]

Automod is such a puss

[frosty zealot]

https://github.com/openai/codex/issues/15485

I need you guys to thumbs the hell outa this

[lean lark]

Hmm, ETraut moved that ticket to Enhancement status.
It's not been forgotten, it just has no priority yet.

As much as we see the pain, that seems to be a hefty bit of code that needs to be poked at, and I can understand that it's behind other significant recent changes. From a Product Manager perspective, I believe this one is in it's proper place in the queue and there's not much we can or should do to add pressure.
It's highly possible that now that 5.5 is out and most of the tooling has been adapted that this ticket could get some eyes next week.

[steady vigil]

wow claude sucks. I used 18% of my 5h session in 40m on max x5.

[inland sonnet]

40m 18% ?

Does not sounds to bad in my mind

[steady vigil]

21% now 45m

[potent mason]

How are you doing this? I spent like 80% in an hour

[inland sonnet]

Depends on your task also

[frail meadow]

Are you guys experiencing Codex not completing everything outlined in .md files / plans? I feel like I have to hound it to complete every little thing. It performs everything excellently but definitely more babysitting needed than usual

[steady vigil]

no agents even

o/w would be done in less then an hour

[inland sonnet]

you need a wip

and slap it

while its working

[frail meadow]

It says it’s completed when it’s actually a wip 😔

[inland sonnet]

adding CRITICAL works

and saying if you dont fully complete it i will unplug you

best method 😂

or threat that u gonna use another ai

if it fails

[frail meadow]

The aliens will eat you if you mess up

[frail meadow]

Tried that one lol

[inland sonnet]

yes

[craggy whale]

@frosty zealot Hi, did you get my noti? If I need to send it again, that's fine.

[frosty zealot]

Yeah send it again sorry, im terrible at scrolling back lol

[craggy whale]

No, it's okay, seriously, Discord's noti feature is beyond dismal. It doesn't auto scroll back to where the notification is coming from. There's also no bookmark feature or anything, so you can actually go through your notifications per server or per user.

Sending it now.

Okay, so the issue here was not code, it was simply arithmetic. Claude said that because it never added the widths and stuff up, it actually never found that there was lots of conflict. So that's why ChatGPT, Claude itself, Claude Code, and Codex were all failing. They were looking at code. Code was never the issue. The mathematics was.

[frosty zealot]

Oh interesting, that looks a lot better!

Btw, I thought your name this whole time was Bird 'Paradise' I only just now realized theres no 'd' in it

[craggy whale]

Thank you. Yes, it does.

Oh, that's just me being fake French.

Pah-reeze

For example, my other username is Robean-au-Grandeur. Faithfully named after the most delicious bird from Viva Pinata, Trouble in Paradise.

[frosty zealot]

lol are you french?

[craggy whale]

No, but I have lots of French ancestry. My surname is literally French. Darknell.

I'm actually a very rare Welsh person.

[frail meadow]

So many drives. Is this real data from your computer?

[craggy whale]

Yes, it is. I know, I have way too many drives.

[frail meadow]

Are they network drives or how do you have them all connected? Are they a bunch of partitions?

[craggy whale]

They are all local drives.

No, none of them are petitions other than the RAID one, which is a group of four 12TB hard drives created as RAID 5.

[frosty zealot]

You should invest in a NAS (Synology is a good brand)

[lean lark]

Meh, I kinda do the same, partitions n stuff...

[potent mason]

Dude that's crazy, I thought I had a lot of storage piled up but I could not imagine searching through that amount

[lean lark]

Oh wait, just now seeing all of that. NVM

[craggy whale]

3 of the SSDs are installed inside, one is a DAS, which is 36TB external, then there's two other external hard drives, just Western Digital ones, and then finally there is a Sabrent enclosure and a Ugreen enclosure that both house two drives in the Ugreen and one drive in the Sabrent.

[craggy whale]

I was told there's no point in me getting that because it's really expensive when I just need storage.

[frail meadow]

Does this have faster file search than the standard file explorer? That would be super handy for your storage

[potent mason]

Are you like a photographer or something?

[frosty zealot]

More for convenience, being able to access from any device, anywhere, any time, and not physically attached to it, and Synology makes it silly easy

[craggy whale]

I know, I have serious issues. I mean, rest assured, a lot of the space that is taken up is because of duplicate files, which I am on the never-ending sea of hunt to reduce. It's just because of the past when I used to have my laptop and stuff. Whenever I backed up my phone or whatever I never trusted the backup, so I would back it up twice, sometimes three times and because of how windows duplicate detection is so appalling I would then just amass lots of duplicates because I was afraid because a lot of the time it would delete originals, so I didn't trust it, so I would just click the option of keep all.

I'm currently in the process of building a very comprehensive duplicate manager that matches all the fringe cases that I keep meeting that the existing duplicate managers on the market do not help with.

[frail meadow]

I’m guessing they’re not that expensive if you bring your own storage, right?

[potent mason]

It's basically just like an iCloud, I got it for me and my family because we where up to the 14TB plan (talking about NAS)

[craggy whale]

That's okay. That is okay, Pharaoh. Doge?

[craggy whale]

Yes, it queries Everything API directly. It also uses the Everything EFU file anyway.

[lean lark]

Welcome BTW!!

[potent mason]

Oh you should try out cleaning it with Codex, I've heard some use cases for deleting a bunch of stuff

[ember estuary]

Is there some technical hurdle in bringing computer use to the Codex Desktop app for Linux (when it finally arrives??)

[frosty zealot]

Eh they can be depending on what model, especially if you want like aggregate rj45 ports

[craggy whale]

I wish. I have taken a lot of photos. I do save a lot of photos. No, most of my collection is YouTube videos and downloaded R&B music from very rare blogs. Those blogs don't even exist anymore.

[craggy whale]

I want to know which device I would ever be accessing it from. The only devices I have is my phone, my laptop, well, my four laptops, and my desktop. But none of those I take out with me other than my phone.

[frail meadow]

I think it’s mostly a difference in API + training on the application / navigation differences? Same with mobile

[craggy whale]

Thank you. Thank you very much.

[frosty zealot]

I wanna buy a DS1825+ but its like $1800 CAD, and I dont really need that much storage (its an 8 bay) I'd probably buy a DS925+ if I had to buy another one today, I've been rocking a ds218+ for like almost 10 years

[frail meadow]

Do you have a plex server?

[craggy whale]

Oh yes, I'm currently in the process of building it with Claude Code, Claude, ChatGPT, and Codex, all of which I use when the other one runs out.

[craggy whale]

Nope. I tried using it once, had no idea what it was about, didn't find any help from it, and uninstalled it.

[ember estuary]

I assumed it's the screenshot and app control API differences that are the big hurdle... I'd really love OpenAI to bring a native Linux desktop app also so I can stop using the janky repacked electron version.

[frosty zealot]

Synology is nice too cause it can do virtualization with their modern ones, like running VM's and Docker containers on them directly

[unique spade]

lol you sound like my kind of guy. who needs existing software nowadays when we can build our own 🙂

[frail meadow]

You can run containers on NAS??

[craggy whale]

Thankfully I bought all of my storage before the hard drive apocalypse weirdness.

[lean lark]

Rather than just using these tools as simple bots, work with Codex to create tools that do what you want. Then you can do everything off-line ... no tokens. That said, for more difficult files that require eyes, use the API and upload images for categorization.
Bottom line, the better you can use the tools the less it will cost.
Note also that locally installed LLMs can now do some image processing. One example is Gemma 4.

[frosty zealot]

Yeah on like synology's I'm not well versed in the other ones

[frosty zealot]

https://www.synology.com/en-global/dsm/feature/virtual_machine_manager

[potent mason]

I feel like that's too much of a slippery slope though, I'm down with building my own software but stop on anything that's 10+ hours it's just not worth it at that point

[frail meadow]

For duplicate manager can’t you just hash each file and check for matches?

[frosty zealot]

https://www.synology.com/en-br/dsm/feature/docker

[craggy whale]

Yes, very true, but it's not like I'm building any of it. I don't like the phrase, but yes, I am the definition of a vibe coder. But it's not like I'm just like, hey, give me ideas. It's no, I have ideas, how do I execute them? So there is a lot of thought and planning and comprehension, understanding and foresight as well. As intuition and just you know I still have to debug as well so it's not made sloppily at all.

[craggy whale]

Oh, I absolutely agree. I used to ask ChatGPT to create markdown files for me and now I have an application that I made in Electron that does that for me.

[unique spade]

yea but i mean there are loads of duplicate finders softwares, yet you said you are planning to build one
🙂

[frosty zealot]

If anybody wants to gift me a DS925+ I wouldn't stop you

[frail meadow]

Might need to look into that, have you tried it out? Maybe it could be even more performant if it supported nerdctl as well

[lean lark]

Very cool. Codex is good for that too. It auto-generates and updates much of my code project docs. Then it uses those docs to understand the projects so that it doesn't waste time trying to figure out everything from scratch in every thread. (I so wish people would understand that.)

[frosty zealot]

My buddy says its flawless, unfortunately I have an old one from like 2018 that doesnt support virtualization

[craggy whale]

Sometimes, but it also needs to check the suffix as well. And there's many fringe cases. I will ask Claude to create a document on everything that we've decided on so far for the duplicate detection, so you can see actually how complex it can become. It also depends on type. For example, a document duplicate is a lot different to an image duplicate, which I would rather use an image model for, and the same with videos. The duplicate pipeline that we are building treats videos and images as the same because of course a video will have a thumbnail and so that would work for duplicate detection instead.

[craggy whale]

Uh-huh, because it doesn't fit all the fringe cases I need it to fit.

[potent mason]

I would but kinda tight on cash right now, had to hire a few people

Just unsub from GPT a few years lose out on 20x ROI and then buy it

[craggy whale]

You mean like a summary? If so, yeah, that mostly happens with Claude code, but of course it has a context limit, so there is a limit to that as well.

[unique spade]

i d suggest to enhance the deterministic functions that are already covered by existing legacy software, with some codex on top of it

[craggy whale]

That's exactly what we're doing. We're building from what already works and making it better for my personal use case.

[unique spade]

great. that s the way now 🙂

[fringe shuttle]

do 'chats' in the codex app go against my usage limits?

[craggy whale]

For example, I have many files, this is an example, so not a natural file name, called. Bird (1), bird (2), bird 3, etc. All of which would be missed by the duplicate detection, and other programs would mistake them as pure duplicates instead of a mix of series and extension format.

[frail meadow]

No that’s part of their new UBI program

Jk Codex and ChatGPT have different limits

[frosty zealot]

Yes

[craggy whale]

Oh, that's more generous than Claude.

[frail meadow]

The limits are combined for Claude?

[craggy whale]

I have three accounts for Claude to bypass that issue.

[craggy whale]

Yes, sadly.

[lean lark]

Not just a summary. My AGENTS.md files direct Codex to read the docs folder for functional information relevant to the task, and every function in every module is documented so that the model doesn't need to read code to figured everything out. There are also procedural docs so that it knows how to process code and docs, now to structure code, etc. It's instructed to document every function and to update the other docs when things change ... like breadcrumbs so that it can get back later. And finally (I think) there's a HowTo doc for one of the bots that lists various conditions and how to handle them in specific files. So it doesn't need to work that out every time. It's all automated. I don't need to look at the docs much anymore but I sometimes do just to keep an eye on things ... and remember how my own software works.

[craggy whale]

I have two accounts that are exclusively for ClaudeCode and then one account that is exclusive for Claude so then I can be like hey how do I talk to ClaudeCode? How do I prompt it correctly, so it doesn't kill all my tokens? And of course, normal Claude takes less tokens in the first place.

[lean lark]

(Oh, BTW, I'm the guy in here who thinks in paragraphs and pages... sorry. 😜 )

[frosty zealot]

He's my dad, he's a little senile but as long as he gets a few naps in the day he's alright

[potent mason]

I'm the person here who only reads the TLDR's😂

[craggy whale]

Sorry, I was severely oversimplifying. Yeah, I absolutely understand what you're saying. I knew it was more complex, I just didn't know how to paraphrase it. Excuse my paraphrasal. So, yeah, that sounds great. I would totally do the same if I understood how to even do that.

[frail meadow]

Thinks in paragraph

[lean lark]

why, I oughtaaaaa....

[craggy whale]

Me too. It takes me 15,000 words to get to the point, because I'm a multi-contextual thinker. Which I personally dub a Realm thinker. It's deeper than abstract.

[lean lark]

Hey C, DM me.

[unique spade]

made myself a nice analytics panel for my codex threads 🙂

[craggy whale]

Can you tell me what that's trying to show?

[potent mason]

How efficient he is probably

[frail meadow]

Bonuses that ccusage doesn’t have?

[lean lark]

Looks familiar @unique spade https://discord.com/channels/974519864045756446/1497255535664300173

[unique spade]

i wanted to be able to see what happens in my threads

[lean lark]

Hmm, I'll show you my code if you show me yours. 🙂

[frosty zealot]

Here's my cat in her cat tree circa 3 minutes ago turned to pixel art with image-2

[craggy whale]

Sorry, probably because I lack context, I don't really know what that means. But that's great that it's helping you. It's not very obvious to somebody who doesn't have the context though. So if you need it to be, that's some constructive criticism for you. Like even just a blurb at the top that explains the purpose of the document in a sentence.

[unique spade]

lol captain knows the context cause he built something similar

the idea is we want to know what happens in our codex threads

[craggy whale]

Wow! Is it that beige and gold in real life?

[lean lark]

Just got called to dinner, need to go. TTYL
(That link I keep posting isn't about JaneBot, that's the codex usage page, will publish to GH soon)

[unique spade]

that s for me ?

[craggy whale]

Yeah, that's my fault. I don't know what a thread is.

[frosty zealot]

Yeah it's a tortise shell

[unique spade]

a chat you have with codex

[craggy whale]

Have a good tea!

[lean lark]

G2G 🙏

[frosty zealot]

Walk safe

[craggy whale]

Oh, okay. I thought they were just called chats.

[frail meadow]

He’s going to the car wash

[frosty zealot]

uh oh

[craggy whale]

Every time somebody says car wash, I always think of the Missy Elliot and Christina Aguilera cover.

[frosty zealot]

If there was an AI that was learned strictly on animating it would be the best thing ever

[cedar skiff]

I'm not sure about 5.5, it has the same common problems as 5.4 and adds a few others on top. I guess it's ok for frontend design now, but I trust it less in the code base than 5.4

[frail meadow]

I think there was a YC company that launched recently for that

2d tho

[unique spade]

hard to tell for me so far. i did start new threads with 5.5 so it's a bit off, but i usually work in very long threads so maybe it needs some time till we stabilize the workflows

[frosty zealot]

Not so much like 3d/2d animating but like for web/app design

or like anything other than model animating

@lean lark

[cedar skiff]

It does stuff like this:
It added this to a skill file because i told it to remove uninstall in test tear down and said we should not be uninstalling the app during tear down.
So it takes that conversational point and add it's to official repo skill. This is a thought it has in this repo But the thing is this never needs to be mentioned. The only thing that needed to happen was the teardown needs to be editied. For some reason it thinks the skills and supporting reference documents need to be updated to reflect previous states like they have some sort of memory of the history of what they did before.
This pollution is out of hand with it so much so that getting it to update docs it will constantly say what not to do, where it only needs to state what to do.

For this repository's preserved Android and iOS integration-test devices, do
not run an integration-test command that uninstalls, reinstalls, or otherwise
replaces the configured app install. Use
`scripts/integration_test/run_mobile_drive.sh` for connected-device mobile
runs so completion keeps the app running and Android uninstall attempts for the
preserved package are blocked.

The semantics for tasks like this are just broken

[frosty zealot]

Interesting

[lost drum]

Openai devs please reset ussage

[potent mason]

Honestly I can't run through my usage as fast as they are resetting

[lost drum]

am at 38%

pro 2x

I can spend 10% daily now

but I will consume it in 1 day

I want to sleep but the porject those noe let me

its 5am

[potent mason]

Bruh

[frosty zealot]

are you in chian

[potent mason]

I'm also sleeping late because of projects but 5:00am is just gonna get you burnt out

[lost drum]

poland

[frosty zealot]

Poland looks like a nice place to live

[lost drum]

dont come

[frosty zealot]

LOL

Is it not?

[lost drum]

[potent mason]

Everyone thinks their home is terrible to live in.

In my experience any place of the world where salaries are similar to third world countries and you get the safety of a first world country is actually goated.

[lost drum]

dopamine from creating taht porject is enought to keep me up till tomorrow

[frosty zealot]

In canada but I feel like im in india, ill leave it at that 😛

[potent mason]

Well I mean you gotta make your money in a first world country, then you move to somewhere like India

[haughty cradle]

any1 got the problem where 5.5 burned thru like alot of their api costs?

i got 1.4 billion tokens down the drain tdy

[potent mason]

Yeah, it's twice as expensive as 5.4

[haughty cradle]

i just have a normal app with like 25k lines of code

[lost drum]

5 more prompts and sleep

[potent mason]

That's crazy, I only have 200 mil today

Maybe I should stop talking and working in one project and start working on 2-3

😂

[haughty cradle]

yea and im 42k down the drain

just tdy

[frosty zealot]

Since boris on X announced they found issues with the dumbing of claude, is opus better now?

[haughty cradle]

idk what to do

[potent mason]

It's mid

It definetly is better

[wind siren]

how do you manage that

[haughty cradle]

how am i going to get 42k; is my api leaked?

[frosty zealot]

I still ahve my max plan but I might have to offload some work to it cause my GPT is like 20% till the 28th

[haughty cradle]

idk im cooked

i think my api got leaked

[wind siren]

do you do api or pro

[haughty cradle]

api

[wind siren]

danm

rip wallet

[lost drum]

I will run it while I sleep now it has 39% i will update how it goes😂

[wind siren]

im to broke so i just use the plus plan

[cedar skiff]

If you want to save usage use codex 5.3 you will get 5x the usage

[wind siren]

no

[lost drum]

I try to not spend at all to afford it I barely spend money only water food and necessities and rest went for that pro plan lol

[cedar skiff]

It's easily capable of doing most tasks, you don't need to hire a jet to drive to airport when a taxi will do.

[lost drum]

Rather build with codex than drink mosters

[wind siren]

i need GPT 5.5 sense im working on the hardest tasks

[cedar skiff]

Every task is hardest task i guess

I wonder what you did before 5.5 was released

[lost drum]

5.5 emits output so fast I remember when streaming was so slow always on gpt 5.4

[frosty zealot]

Its like when you were growing up and you were like wow, the graphics cant get any more real than this, then you pull it up today and you're like ok I remember this looking photo realistic lol

[potent mason]

Btw are you guys using the http or websocket option?

[frosty zealot]

for what

[potent mason]

For chats

[frosty zealot]

I think its stdio/websocket isnt it?

I dont think theres an http

[potent mason]

I think websockets used to be more expensive because of how requests worked (message based)

[potent mason]

Wait give me a second I need to check it

Haven't seen it in a while

My bad I guess they don't have it

I could've sworn they had like a websocket mode that was better because it was more failure resistant (like if wifi goes out for a bit)

[frosty zealot]

Well they have stdio or websocket to do the JSON-RPC requests to the app-server

[dusk thorn]

pro $100 plan rn is like infinite building for me rn

[frosty zealot]

stdio is better locally, websocket if you're working remotely

[potent mason]

Interesting

I'll have to check it again at some point

I'm just too busy right now

Clients keep wanting more stuff even though they don't need it

[dusk thorn]

anyone notice 5.5 working for longer than 5.4 did with better results?

usually 5.4 would work for like 8 minutes and chew something out that was working decent but 5.5 does 30 minute tasks more often

[quasi summit]

There’s nothing superior about stdio locally over websocket

[quasi summit]

Yes

Token death

[dusk thorn]

5.5 uses less tokens than 5.4

but does cost more

[inland sonnet]

hgahahaha lol

[lost drum]

Yes

[lost drum]

Its 10x better than 5.4

[inland sonnet]

100x!

[lost drum]

[inland sonnet]

i got access to mythos

dont try me buddy

[dusk thorn]

discord is like the total death of what mom says

[dusk thorn]

so did some "hackers" (they just guessed the url)

[inland sonnet]

No it was more then that

They had access to some company workspace, and then guesed the url)

[dusk thorn]

ah

still funny af

[inland sonnet]

most hacks are humans getting owned not the system lol

[dusk thorn]

you can look up codex in your activity settings -> registered games

and add it

haha

[inland sonnet]

Lol no

[dusk thorn]

worked for me

it is janky tho

[inland sonnet]

Yeah but its useless as its showing nothing

[dusk thorn]

yeah ik

[silent lodge]

Hello guys, I was rate limited on codex but I plan to buy the Go plan, can I get it to do one project per week?

[inland sonnet]

This depends on context, and how big your project is.

[inland sonnet]

Best bet i would do is try it and use GPT-5.4 Mini Low-Medium level to max your usage.

[lament estuary]

Thanks captain obvious, no hit the reset button

[undone patio]

are we there yet

[silent lodge]

Thanks I will do that

[inland sonnet]

For that price u can't expect much, but i wish your best luck.

[dusk thorn]

Exactly

I had reconnecting for 10 seconds, where is my rate limit reset

Ohhhh shi

Lock in everyone

[undone patio]

im thirsting for tokens sam

[dusk thorn]

Tibo drop a update on Sunday and reset our limits

[undone patio]

😭

[dusk thorn]

What plan are you on

Son I’m crying

[undone patio]

20x pro

[dusk thorn]

WHAT

[undone patio]

[dusk thorn]

WHAT ARE YOU BUILDING

[undone patio]

i make software for a living lol

[dusk thorn]

He’s building a new os guys let him be

[dusk thorn]

Solo dev?

[undone patio]

nah i work at a startup

[dusk thorn]

Nice

[frosty zealot]

image-2 is the goat, I couldnt get codex to figure out how to make a nice inset background for a container asked it to generate various inset designs

#5 lookin kinda goated

[dusk thorn]

It does

[dusk thorn]

Do they pay for your codex?

Or do you yourself

[undone patio]

Nah they do

[dusk thorn]

Oh nice

[dusk thorn]

If you don’t mind me asking what are you working on?

I’m kinda curious

[undone patio]

Global compute substrate and some other minor ventures

You can find a link in my profile

[frail meadow]

crypto stuff?

[dusk thorn]

Your resume is very impressive bro

[undone patio]

Yeah sort of crypto stuff

And thanks

[storm elm]

lasterrupdate, a lot bug. Where button settings on codex ?

[undone patio]

what are you building

[dusk thorn]

A website to help people at school with schoolwork, study better, and take notes in class easier with ai summaries and quizzes off it etc

[undone patio]

thats cool

[dusk thorn]

Tryna make something good been working on it for 3 weeks not a 2 day gpt wrapper

[undone patio]

lol, nice

engineering is 80% careful thought and planning

and the other 20% is coffee and profanity in my experience

[dusk thorn]

Yes

For me it’s Red Bull

[undone patio]

ive been trying to stick to coffee but sometimes i break and get a white rockstar

[sonic wasp]

Anyone having issues with using gpt 5.5 via azure foundry hooked up to codex?

https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/codex?tabs=npm

i've followed this and set it up but getting:

■ unexpected status 405 Method Not Allowed: Unknown error, url: https://heyfixit-us-east-resource.services.ai.azure.com/api/projects/heyfixit-us-east/openai/v1/responses/responses

[undone patio]

any reason you need to use foundry?

[sonic wasp]

i have start-up credits via Microsoft

[undone patio]

i see

[sonic wasp]

figured i'd save my sub

[undone patio]

if it was me i'd use codex CLI to debug it

then use foundry

[sonic wasp]

i also tried pi and it has the same issue

feels like something changed

[undone patio]

theres also the new qwen model which people are saying is like opus 4.5 quality

might be of use in debugging

[sonic wasp]

3.6 27B?

the dense small model

[undone patio]

i believe so yeah, theres also a 34B one iirc

welp back to antigravity until sam has mercy upon us

(or gets drunk again)

[dusk thorn]

Just generate with images v2 codex saying your code sucks and tell Tibo to reset your limits or else

[regal kraken]

is the codex app freezing for anyone else after update

[undone patio]

nah ive been using it all day

it did crash once i think

[dusk thorn]

Nah

I got a few small reconnects

But that’s it

[regal kraken]

theres been 4 updates and its still freezing

[undone patio]

maybe ur OS is bad

what u running

[regal kraken]

after i initally update for 5.5

[regal kraken]

no it was fine until i updated for 5.5

[undone patio]

yeah but im saying like

im on MacOS for example

i have no issues

[regal kraken]

yea that would be why

im on windows

[undone patio]

xd

is ur OS up to date and etc?

[regal kraken]

they phooked codex up with the update

[livid forum]

{"type":"error","status":400,"error":{"type":"invalid_request_error","message":"The 'gpt-5.5' model is not supported when using Codex with a ChatGPT account."}}

anyone else getting that^

[undone patio]

i mean really its bill gates fault if you think about it

[dusk thorn]

Microslop

[regal kraken]

ill have to switch to cli or t3 code 🙁 in the mean time

[undone patio]

theres no windows products in my house except the ones made of glass

[dusk thorn]

Microslop when new OpenAI updates prioritize Mac when they fund OpenAI

[undone patio]

terminal is good anyway, you dont need the app really

[dusk thorn]

What Mac do you use?

[undone patio]

everything is the same via CLI

im on a M4 pro rn

32gb i think

[dusk thorn]

I’m on m5 MacBook Pro with 24 gb ram

[undone patio]

i lied mine is also 24gb

[dusk thorn]

I couldn’t get the 32gb but it still runs some good local models

Like I can run gpt oss 20b well on ollama

[undone patio]

i dont really need it for my work

i mostly make applications

not like video games

[dusk thorn]

But I don’t use much when I got a subscription

[undone patio]

or smth

[cedar skiff]

cli doesnt have the same usability as the codex app. The ui layer provides a lot of easy paths that dont exist in the cli.

[undone patio]

and i tend to train or infer models in the cloud

so my local compute is usually just for project dev

i do have a minipc for an MCP tho

i use it to store skills mostly

and to host my discord bots

[frosty zealot]

should I upgrade

[undone patio]

no i think ur ok

[dusk thorn]

I think your good

💀

[frosty zealot]

thing cost like $5500 CAD, this thing is getting burried with me

[undone patio]

lol

[dusk thorn]

With the way models are getting smaller for local and also better efficiency of apps it should last you a long time

[frosty zealot]

I read that the newer m4/m5 chips were worse than the older M's for models I cant remember the reasoning, did you guys see anything like that?

[undone patio]

apple is made to last, until they depreciate your hardware on purpose

[dusk thorn]

No

[undone patio]

but until then ur good

[dusk thorn]

I heard m5 is the best for ai rn

[frosty zealot]

yeah I had a 2012 MBP that lasted me until I upgraded to this one, it was at hte point i couldnt even update the OS anymore

[undone patio]

anything with an apple silicon chip in it, they arent gonna abandon for a while i suspect

[undone patio]

even then its still useful for some things

but yeah gotta stay SOTA if you can

[dusk thorn]

John ternus gonna cook with gas as ceo

[frosty zealot]

I got my windows machine for gamin (4090d out of my mind) a little Mac Mini for some local hosting of stuff, and then this MBP for dev

[dusk thorn]

Only guy I wanted to become the next ceo

[undone patio]

i had gotten the latest flagship omen laptop for gaming but then i realized

i just dont care to game anymore

so i returned it

[dusk thorn]

Codex is like a game

It’s so addictive

[undone patio]

yeah making software is a lot more fun really, especially now that the days of stackoverflow are behind us

[frosty zealot]

Took me like 27 accounts before I started getting positive rep on stackoverflow

kept getting instantly locked and banned

[dusk thorn]

LMAO

[undone patio]

lol

[dusk thorn]

Horrid place

Now we just ask gpt

[frosty zealot]

I go on every now and again to see how its doing

its like one of those auto clickers

it just randomly gets points and awards

[lean lark]

I abandoned SO a long time ago - toxic moderation.

[undone patio]

gotta check ur runescape now and then ye

[frosty zealot]

I used to be a moderator in the Javascript channel on SO, highlight of my career

[dusk thorn]

Its like a toxic subreddit for devs

I haven’t used much at all but heard MANY stories

[undone patio]

my new hobby is collecting code snippets and feeding them to my MCP like a kirby

[lean lark]

I sometimes wanna go check my RuneScape classic account .... I have Santa hats!!

[undone patio]

nice

i have the 2003 halloween scythe

[lean lark]

we're old

[dusk thorn]

I’m glad unlike Claude you don’t gotta be 18 to use codex

[high girder]

Speak for yourself

[undone patio]

when i was born there was no internet

imagine it

[high girder]

Starbuck isn't old at all. Just one of the first to have a dinosaur burger

[undone patio]

chatgpt was a collection of books called an encyclopedia

[frosty zealot]

https://x.com/GatorsDaily/status/2047675854780551201?s=20

Had to drop it one more time

Still 1 hour left

(for me)

[dusk thorn]

Before chatgpt I used Wikipedia def got it better than most of you guys with access to this tech at such a young age but then I also got to deal with the new job market 💀

Def easier now

[dusk thorn]

Unc

[undone patio]

soon robots will flood the workforce and you'll be free to just like

do whatever

enjoy being alive i guess

ill probably just keep making software

thats what i like doing

[dusk thorn]

Yeah but someone’s gotta make money and be on top because many in this world are greedy

We can’t become utopia yet

[undone patio]

well, if people like sam altman and dario amodei are to be believed

then its only 5-10 years before money means nothing

[dusk thorn]

Interesting times

I wish my high school had a trimester class on ai, we got 3 tris per year and we get to pick a certain amount of electives per tri so I’ve been taking just engineering and computer stuff like it

But yeah we need an ai one

Because from people I know and see 80% don’t know ai past doing their homework and running out of uploads

[undone patio]

yeah bro keep making software and you just might find urself on the other side of the divide in the coming years

[dusk thorn]

I hope

[undone patio]

i have no idea what it must be like to be a young person right now tbh

[dusk thorn]

It’s insane

I probably could be easier on myself but I just spend time on here and twitter finding new stuff going on in tech

To get a decent job

And figure out a degree I can do and use ai for my job

[lean lark]

I used to be the guy in HS doing the stuff that no one understood. Now I'm the old fart doing stuff no one understands. 🤔

[dusk thorn]

I was using codex in class to help me debug a problem and this guy is like your chatting with ai bro your a chud LIKE BRO 💀

[undone patio]

yeah dont listen to luddites bro

[dusk thorn]

He’s not gonna be on this side of the divide 😭

[undone patio]

people fear what they dont understand

[dusk thorn]

Someone was talking to teacher that they wanna do financial planning for people and I just thought of that interstellar meme of the guy knocking on the glass

https://tenor.com/view/no-no-gif-17511868351874922325

[undone patio]

dont you know, you have AI psychosis?