codex-discussions

[boreal holly]

So the quota && the deficit increased? I think your codex bar is drunk and needs to take a nap

[hard drum]

probably

but wait..

let me check codex web

[boreal holly]

Yeah cuz earlier it said your 5hr was like 40% and resets at the same time as the second pic, and the weekly went from ~60 to 71%

Codex bar be trippin

[hard drum]

i can't find codex usage place

just API page

[boreal holly]

I gotchoo https://chatgpt.com/codex/cloud/settings/analytics

[boreal holly]

Wow dude I would reboot the mac ASAP. Something wack is going on

[hard drum]

reboot as in... restart or reinstall?

[signal oak]

28% rem only for this week

[boreal holly]

I mean like you might have detached codex processes working autonomously eating up quota

[signal oak]

how you're gonna spend weekends 🙁

[hard drum]

OH CRAP

[hard drum]

sleeping && maybe looking at maternal 🌽 on pixiv LOL

[boreal holly]

I've had it before with my command-parser, easiest thing is reboot

For example codex exec --ephemeral "some prompt" is sneaky and does not produce a rollout log file, and you might have agents that spawned other agents like that or something, and if codex bar relies on rollout log files to do stats it will throw off the numbers

[lean lark]

hmmmm, my code would do that. 🙁 https://github.com/CaptainStarbuck/codex-usage

Taking requests for next round of enhancements...

[torpid trout]

theres a link in the very desktop UI yo uwhere screenshotting before 😉

[hard drum]

what else do you know? looks i need more stuff for OAL to add

[boreal holly]

You're probably good! As long as you don't use codex exec --ephemeral you will get accurate stats

[hard drum]

brb reboot

[lean lark]

but if agents are doing that on their own??

[hard drum]

Hook block

[torpid trout]

Codex VS cod extension update twice a day
noice

[lean lark]

And it seems the only updates to VSCode itself these days are in the GH Agents processing.

[boreal holly]

The only way an agent can spawn another agent like that is if they run in --yolo mode, or if they're in workspace-write and they spawn via codex exec --yolo --ephemeral actually they hit write sandbox in ~/.codex/sessions so no risk here. Or they can escape the sandbox with tmux new -s some-agent "codex exec --ephemeral" which makes tmux the process owner and escapes the sandbox. Or if they use docker

Anyways there's a number of ways they can do it, they tend not do do it without provacation. In borrowcheckbwt's case they're building an orchestration thing so the risk is higher than normal

[torpid trout]

You can also enable sub-subagents, no?

[boreal holly]

Yayyy!

[boreal holly]

Yeah I think it's "thread depth" limit > 1 allows recursive sub agents

[torpid trout]

this button still so smol

🤏

[hard drum]

https://raw.githubusercontent.com/openai/codex/refs/heads/main/codex-rs/core/config.schema.json find full stuff here

[torpid trout]

It got annoyed by your username

[lean lark]

It got annoyed by so much complaining

[torpid trout]

Like, duh, bro, I can do the same, but more annoying... let me print 500 recconnecting

[torpid trout]

well, that too probably lol

[hard drum]

    "experimental_thread_config_endpoint": {
      "description": "Experimental / do not use. When set, app-server fetches thread-scoped config from a remote service at this endpoint.",
      "type": "string"
    },
    "experimental_thread_store": {
      "allOf": [
        {
          "$ref": "#/definitions/ThreadStoreToml"
        }
      ],
      "description": "Experimental / do not use. Selects the thread store implementation."
    },
    "experimental_thread_store_endpoint": {
      "description": "Experimental / do not use. When set, app-server uses a remote thread store at this endpoint instead of the local filesystem/SQLite store.",
      "type": "string"
    },

these are the only thread-related things, rob

[boreal holly]

[agents]
max_threads = 1
max_depth = 1

[hard drum]

that's general

but yeah, could have written specifically those two

already got those two

[boreal holly]

Leave 9000000000000000000000000000000000000000000000000 alone!

[hard drum]

why have it at 1 though?

i mean the threads bit

why not like 2 4 6 based on sub?

plus/pro5/pro20 ?

or 1 3 5

[boreal holly]

I mean I don't use multi_agent so I set it to 1 as a failsafe. And you mentioned earlier there is a multi_agent_v2 that I did not disable so I'm kinda glad for the fail safe

[torpid trout]

why, your alt?

[boreal holly]

Negative, 9000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000 is an upstanding citizen

[hard drum]

i did disable it for mine now as well

[torpid trout]

Holy brew

==> New Casks
eez-studio: Visual tool for GUI development and T&M automation
factory: Native AI agent interface to build, manage, and ship software by Factory
open-webui: Desktop application for Open WebUI

I wont sleep tonight lol

[nocturne folio]

its happening again and again, my 30 queued messages are failing dude

[boreal holly]

Try

[features]
responses_websockets = true
responses_websockets_v2 = true

Seems like the HTTP SSE path might be bricked up

[boreal holly]

This skill is absolutely insane: https://github.com/robertmsale/.codex/blob/main/skills/design-worker/SKILL.md

Can finally do UI work without babysitting a designer agent

[hard drum]

mine has taste-skill upstreamed

i may add impeccable to the list

i used to have a custom ver of impeccable-esque on oabtw

[boreal holly]

Oh yeah, this skill has to be combined with another design skill like impeccable or taste-skill. On its own it's not very helpful. I combined it with impeccable.

But what the skill does is it creates a work pipeline for design tasks. The workers have to render screenshots of the real app, render reference designs with images v2, and the orchestrator approves redesigns based on specs. Basically the way I was doing it before is I would pass reference designs to a designer, have them work on it, and I'd have to constantly iterate on the designs.

[hard drum]

ooo

gonna upstream that thing

nice!

[lost drum]

oh man idk what to do, I delgated to codex app 10 huge convo blocks I had with previous threads every like maybe 4kloc or more and he compresses it to max 500loc (I am switching from VS ext) and he extracts eveyrhitn about how it went and stuff like that then I will let him compile the huge huge plean to polish to the max my project. the issue is I have no clue how will I delegate a plan that will be 5k loc to codex without him drifting off.

[hard drum]

5k loc...

[lost drum]

idk how much it will be at the end by I just assume it

[hard drum]

my proglang + its tools are like 120K now

[lost drum]

cause now I was sending him blobk by block convo and he already made summaries worth 15k loc

[hard drum]

still managable

why isn't your "AGI" being capable now?

[lost drum]

cause he broke

[hard drum]

&& i could attrib quarter of it to rust being verbose

[lost drum]

he drifted off to much after leaving him for 3 days

and now I decided to do the ugly part of scraping eveyr idea of mine and sturcturing it in to the plan

[hard drum]

@boreal holly

[lost drum]

• I need to scrape every peice of inof about possible enviiroments for that system so then it can choose which one is best for this project

Spend whole day sending messages to him from previous 3 threads

[plush harbor]

mine chewed through far more of my block of tasks yesterday than I thought it would so now I'm ahead and have to re-plan my day lol

[boreal holly]

Happy problems 😎

[plush harbor]

yeah codex actually told me to stop for the day

[lost drum]

wow

[boreal holly]

Oh thank goodness the agent caught that empty line and patched it 🤡

[hard drum]

setop now is even more beautiful!

also nice efficiency

[plush harbor]

if i keep burning through my personal to-do lists at this rate I'm going to run out of things to do and might need to get a job

[lost drum]

pierogi?

[boreal holly]

They see me greppin', they hatin'

[hard drum]

pierogi? where?

[plush harbor]

y'all got millons of tokens and I call it a good day when I get this "Token usage: total=1,368,911 input=1,268,786 (+ 22,299,776 cached) output=100,125 (reasoning 18,067)" and at this point I'm fairly convinced that a 5 hour window is 30M tokens

[hard drum]

i imagine that nigerian duck singing that

[lost drum]

I need to eat some now but too late, so tomorrow pierogi! hehe

[hard drum]

ahaaa

how come you know pierogi word?

[lost drum]

been in Poland and ate "ruskie", "miesne", "black berry", "lamb"

best food

but cant afford when purchasing codex sub

[hard drum]

pierogiiii

will go to poland somewhere summer hopefully

wanna meet my brother && all

[boreal holly]

i wanna meet Marcin

[lost drum]

30min and he in codex for 200h bro what ou building

[boreal holly]

omg dude I understand the 200h thing now! Yeah I have Zed open 24/7

[solid lake]

hell yeah

[boreal holly]

For some reason it says I have model_cache.json open but I don't lol

[stable obsidian]

Whats this do

Let codex design good?

[boreal holly]

It's the orchestration pattern for design so it can be automated. It still requires a design language skill, but it enables design tasks to be completely automated

[stable obsidian]

Thats cool

[cedar skiff]

What is the ballpark estimation on how many tokens we get on 20x?

[boreal holly]

~3 billion per week I think for 5.5

I mean 3 billion, with average input/cached/output ratio

[plush harbor]

me: right so is this enough to start working on this abomination
codex: So yes: the abomination is now specified well enough to code.

it better not use that as a variable name

[boreal holly]

You: Do not use abomination as a variable name
codex: Got it - uses static var not_abomination =

[plush harbor]

its a guild picker. It really, really, REALLY wanted to just use an autocomplete widget. I had to get chatgpt to craft it a small essay on why that is the worst idea known to mankind

[boreal holly]

You mean like a dropdown instead of a text field? How did you get Codex to want to not put a text field for an enumerated value???

[plush harbor]

i have 70,000 guilds and I swear half of them are either legion, have fun, or phoenix. I feel like codex has not met humans. Humans suck

[boreal holly]

I ask Codex to build forms quite often. Many parts of the forms have a small, fixed set of possibilities. Codex will make those text fields and provide error messages if the user types something that isn't one of the possible values 😂

[lean lark]

Common UX is that long lists don't belong in a dropdown/combo.

[plush harbor]

hah. I already had helper functions to turn exposed filter into fixed dropdowns and it actually found those

[plush harbor]

hence autocomplete, but that fails spectactularly when you have 10,000 items very similar

[lean lark]

Anyone who doesn't understand UX, just look at a page or book that explains differences between check, radio, list, dropdown, grid, and massive-honkin-selections-from-billions-of-records. 🙂

seriously though, how about breaking down the list into categories of some kind?

[plush harbor]

oh yeah hence the essay I got chatgpt to write for me on what this should do and why. I don't like CLI for long typing

[boreal holly]

Aw dude, for Codex it's like "feeble humans! Just read the whole list and type the option manually! No auto complete you peasant! Use your eyes!"

[lean lark]

Oh, I understand what you're looking for with guilds...

[plush harbor]

lol its put the reasons in the help text. This is hilarious

[lean lark]

yeah, because gamers tend to reuse the same keywords for everything in some games, you kinda gotta put your foot down and just make those ignore words. put a label around the box that says "don't enter 'legion', 'phoenix' or 'blah', these common keywords are ignored. Or put them into a little dropdown so peeps can see it.

Maybe I haven't read up enough here, sorry if I've missed context, I'll back out and flog myself if you wish.

[plush harbor]

so it actually made something quite good - its sorted by highest activity first to the dead guilds drop to the bottom of the list - but the help text basically says you humans suck

I'll give it better text when I stop laughing

[lean lark]

well, we do, stating the obvious is rather insulting.

[plush harbor]

you can't let the humans know though, that's just not done

[lean lark]

hmm, trvth

[boreal holly]

Yeah dude, this is a killer setup. Finally, no stopping at "good enough" 😩

[lean lark]

I have no idea what you're doing bro, TLDR with no context. But I trust it's absolutely Amazing! 🤯

[plush harbor]

I get the funniest issues with this site cos its a gaming site. I had one a while ago entirely about cleaning up emoji ridden pastes from discord

[lean lark]

which game? one of the Clash or Survival types?

[boreal holly]

Basically the most frustrating part about UI design is constantly iterating on something. If I have a reference design, the orchestrator makes sure it matches the reference and doesn't let them get away with half-finished designs

[plush harbor]

hero wars, which is objectively awful but this site pays for all my AI tooling for my less embarrassing projects

[lean lark]

Totally get that. Forge of Empires was gonna be another guess. We pick our guilty pleasures and sometimes need to suffer with others to support the habit.

I mean, hey, just look at my face, rite?

[plush harbor]

I've left it as default black-on-white text for months, given how easy this guild rewrite thing has been I've got loads of time to spend making the UX look poncy now

[hard drum]

lucky

i'm always at a loss but having something that could pay for my codex sub would be neat

just don't know what because i'm not money-oriented

[plush harbor]

I gotta make myself something that pays better than adsense at some point or I'm nver going to get ahead, I have no spare in my budget to save for stuff

[static maple]

Every time I update the desktop version of Codex, my configured Supabase MCP gets removed.
However, it still exists in the backup of the configuration file.
Why is it designed this way????
By contrast, Chrome-Devs has no such issue at all.

[lean lark]

OS?

[lost drum]

okay the plan is not that crazy 1500loc

[cedar skiff]

are you aggregating data?

[lost drum]

holy

I wish I could have such setup

maybe tomorrow

road to 1000h?

[plush harbor]

you're still planning your thing hey

[lost drum]

I spend whole day on trying to see on what stage is my project and ye I need to read that whole documentation that was done and see what I need to do

but it is promising

atleast😭

[plush harbor]

I've just burned the entire useable to-do list on my second project and now I'm blocked by staring at a piece of blank paper and holding a pencil. So I'm getting coffee

[lost drum]

wow

[plush harbor]

this thing is too good, I was expecting this to take the whole week and I only started yesterday afternoon

[lost drum]

damn

[plush harbor]

you seem to have almost the opposite problem to me

[boreal holly]

now the boys are cookin

[lost drum]

but yu have main one that orchestrates them?

[boreal holly]

Yeah, just need the one orchestrator to manage all the other ones

[hard drum]

what exactly is it that yours has?

I could harness that if you tell me exactly what that is

[boreal holly]

I made it so when an agent generates or views an image it shows up in the GUI. But more importantly, the orchestrator has reference designs, the agents produce goldens showing their work, and the orchestrator makes them continue until the UI matches the reference design. a very tedious task i no longer have to manage manually

[toxic torrent]

/fast does not feel /fast. anyone else feeling the same

[cedar skiff]

you can be sure you will feel the usage differences. 🤣

[lost drum]

I would love to ingesst that setup after I finish my real agent so he can then review the UI at the end stages would be insane I just need to know how to propperly do that and cheap cause I plan to stop the pro sub and in future even sell my service so would need to think how to resolve it tuth cheap api or some

[boreal holly]

sorry 😅

[cedar skiff]

You could just stop trying to make your assistant and just move forward with work.

[hard drum]

could i harness this for OAL?

[boreal holly]

absolutely!

[hard drum]

tell y'what

make a SKILL.md for other agents to read into && harness what the user might see fit from it

at this point why don't we work on OAL together && merge into one

seems like missed opportunity to put two minds together

[quick geode]

i feel so far behind when it comes to agents

[cedar skiff]

that feeling is never ending, it's like there is always something better just a few prompts away

[boreal holly]

I’m down to collab on something, but if your goal is to have something robust that people like and find easy to set up and use, probably dont wanna merge my configs 😅

[hard drum]

i don't merge a lot of them, but i meant "merge" as in come together && implement as a proper harness part of the layer itself

i got the grill, you got the raw meats

[boreal holly]

haha when you put it that way 👨‍🍳

[hard drum]

eeexaclty

i already got the big stuff going

just needs that little nudge of expertise

[boreal holly]

alright tomorrow I’ll see what i can contribute 🦀 sounds like fun!

[hard drum]

i'll add more settings to OAL to be able to actually toggle multi_agent multi_agent_v2, set the max_depth && max_threads

&& any specific other thread-related issue resolvances, if any

[hard drum]

🦀

OpenDex is OAL's attempt on Robdex

[boreal holly]

1 hour to do 10 designs 😭 5.5 is insane

[hard drum]

now to work on opendex rust binary

[plush harbor]

I'm spending today doing one

[boreal holly]

thats a much better name 😂

[hard drum]

ooooh indeed it is

[boreal holly]

I thought for sure I’d let em rip through the night and maybe be done early morning, but straight up 1 hour for all 10 designs fully implemented

[fringe shuttle]

I'm pretty sure the settings on the windows codex app are just broken. Nothing works. I change settings to fast and it still defaults to slow. I change from 'for coding' to 'for everyday work' to test it and nothing happens. I set Git settings to 'merge' and it still squashes instead.

[dull grove]

Chat when is desktop automation coming to windows 🥺

[hard drum]

lmao CodexBar scared me

showed my usage drop to 0% i thought i got reset

unfortunate...

[dusk thorn]

ngl, i wonder if anyone else agrees, i would rather have a 5.4 mini seperate rate limit instead of 5.3 codex spark

feels better overall and the speed isnt really that big of a thing

[plush harbor]

my problem is I'm not a designer, so I now have a bunch of interface elements, a sticky footer image, and two "inspiration" type images from gemini and its all rattling around my head. I need to turn this into css

[dusk thorn]

i know 5.3 spark is like a testing thing for future stuff, but i think it might be better with 5.4 mini as a second rate limit especially for openclaw + automations

[boreal holly]

at this rate, 5.6 is going to extract the design from your brain without even requiring a prompt

[plush harbor]

looking forward to that, its the brain <-> computer interface that slows things down

you can tell when I got sick and stopped leaving the house. And when I ran out of things to code and had to think about the next project. This thing is great

I suspect my entire backlog on project 2 is going to collapse in only 2 days though

[honest yacht]

What happened to the Anthropic channel?

[meager bobcat]

yeah.. if you gonna wait 29 minutes might as well wait 34 😄

[velvet wren]

this is the OpenAI server

[meager bobcat]

better way to take a break and think about it or move on to something else

[honest yacht]

I am aware

[velvet wren]

well there is no channel for Anthropic and never has been

[honest yacht]

Never mind, I was asking if something happened to their server

[nova solstice]

when asking codex to make something is it a better idea to give it the full feature at once or do i need to break it down into bite sized modules because it failed my very specific requirements when i gave it a literal step by step pseudocode of how something was supposed to work

[high girder]

there's a verification in the top of the welcome section incase you haven't done that yet

[high girder]

Try it both ways. Give it all of it and ask it to break it down into smaller actionable sections, or give it pieces and see which way gets you closer to what you wanted it to make in the first place

[nova solstice]

although tbf i dont think it resembles anything it would've seen in it's training data

[dusk thorn]

when claude just deleted my entire website html

thank god for recovery from git lmfao

yes because rebrand stuff seems like a perfect thing to clean

dude istg codex never does this stupid crap

[plush harbor]

I'm letting mine loose on a drupal site at the moment and its mildly amusing watching it get frustrated at it instead of me

[dusk thorn]

dude letting models install stuff and do tedious crap is so funny

i had codex setup openclaw for me connected to my subscription and it had a wild time

[plush harbor]

don't ask me what its doing right now, I told it we're making a dashboard with pretty charts and stuff and didn't say what goes in the charts. But I'm sure its having fun whatever its off doing

[dusk thorn]

it will probably make charts based on how much time you spend on codex and how much more you could

[plush harbor]

probably just pinch one of the existing charts or something. I have a lot

[hard drum]

mad deficit

[plush harbor]

we peasants on small plans have to stop when we run out of prepaid stuffs

[dusk thorn]

i got 23% left on $100 plan

and reset on the 11th

its only the 6th

smh

[hard drum]

i'm 20x

[dusk thorn]

i need to get that

codexbar?

[hard drum]

codexbar

[dusk thorn]

ill install

codex will lmao

[plush harbor]

I'm on plus ... but I'm going to be really scratching for things to code soon

[dusk thorn]

i went from being on plus not knowing codex even existed till early march, using it, then getting claude pro to go with it, then getting codex pro

it upsold me based on the fact i use it like everyday

[plush harbor]

yeah I only installed codex a few weeks ago and only been actually using it properly for like 2 weeks and I've got so much done its a problem lol

oh cool it actually did invent a dashboard

[dusk thorn]

@hard drum haha i use too much

[sonic dune]

Anyone did the math? Is the 10x on the $100 plan equal to a $200 plan?

[dusk thorn]

no

200 is 20x

wait like the amount of usage?

[sonic dune]

Ah ok so it is bringing it up to 10x

the $100 dollar plan isn't normally 10x?

[dusk thorn]

i think with promo its just half of 200 the 100 plan

[dusk thorn]

nah

normally its 5x

[sonic dune]

Wish they'd just give us raw token numbers, this is so opaque

[dusk thorn]

but since they released, its 10x until late may

[sonic dune]

Ok thanks, I'm thinking of swapping my 200 dollar a month claude plan to codex for a month to see how I like it, but if I could get the 200 dollar equiv for 100 dollar that would have been nice, thanks for helping to clear that up

[dusk thorn]

np

[oak trellis]

something changed .. it comsumed much faster .. 65% weekly left .. restart is 12th .. it was good until they reset the limit .. suppose to get the limit reset yesterday but the latest reset limit took literally now 5 days...

[oak trellis]

i have 12 too .. yes they did a global reset to know exactly that everyone is included in that .. something is up i can feel it

[sacred minnow]

think ive once had codex delete my stuff and it was my own fault i see all these post about ai deleting stuff and never have restricted access for the model im using and never have problems o.o

[tropic karma]

I'm seeing a lot more of this: "Error running remote compact task: You've hit your usage limit." lately. I think it used to follow through with tasks rather than erroring halfway through, which is pretty annoying. Anyone know whether this was changed?

[uneven kayak]

Would be nice if they would show that info in the Windows app

[hard drum]

That’s not even a codex official thing

It’s CodexBar

[uneven kayak]

Ah alrighty

What is a deficit? First I've heard about that

[oak trellis]

does 5.4 min low makes a difference vs 5.5 medium related to usage ?

[uneven kayak]

Yeah a huge difference. Like 5 times more useage probably.

[oak trellis]

oh wow .. should have know earlier haha

its pretty good too for most tasks

[uneven kayak]

Nice, I wish I could use a cheaper model 😭

[oak trellis]

why not ?

[uneven kayak]

I have to use 5.5 xHigh

[oak trellis]

oh wow

i always worked with 5.5 medium got done most ..

[uneven kayak]

Or maybe 5.4 xHigh would be fine too, but I use 5.5

[oak trellis]

yes even medium i do a lot of analytics stuff..

[uneven kayak]

I tried high and medium multiple times and it lets me down every time, xHigh does the job right the first time but medium and high miss all sorts of stuff

[oak trellis]

oooh ok

you have a 200 plan

[uneven kayak]

I had to throw out the entire shop implementation because 5.5 Medium missed so many things

[oak trellis]

i never used 5.5 xhigh lol

[uneven kayak]

Yeah

[oak trellis]

yes not made sense i actually lost 48 hours because it missed some thing and all my data was poison

[uneven kayak]

Damn that sucks, I only lost about 5 hours

[oak trellis]

yes but that was opus 4.7 :/

had to rebuild a lot and upstream so much data..

around 1-2 tb

[uneven kayak]

I even asked it, do you see what you missed? And it was like "oh yeah I didn't follow the plan at all! Oops my bad!" Lmao

But xHigh hasn't let me down yet, it followed the same plan Medium totally messed up and did a fantastic job

[unreal mist]

Same here

I think they are putting a limit on this

It's a real shame, it's was a great difference VS the way antrophic work

[oak trellis]

i might go to a 200 plan ...

[oak trellis]

have the same feeling .. time is up for us haha

[uneven kayak]

If you're working on complex tasks that touch a lot of different systems and have lots of references, xHigh is the only way to go imo

OpenAI gives you a lot of useage for that $200, it's totally worth it

[oak trellis]

will try that out ..iam on 64% and reset is 12th .. it went from super perfect after the last reset they did to consume way faster haha

[oak trellis]

yes will update

[uneven kayak]

Strange, lots of people are talking about increased useage but I'm not seeing that at all. Same as it was last week.

[oak trellis]

so i had around 95% when they reset the limits .. my limit reset was 6th .. then they reset it .. then somehow my usage next day dropped to 84% .. and now down to 64% ..

was so strange ..

[uneven kayak]

I'd run through a Plus subscription in like an hour and a half, it just depends on your task complexity

[oak trellis]

if you use high yes haha

[uneven kayak]

Yeah I don't really have a choice, I've tried high and medium and I always have to go back and finish what they missed. xHigh doesn't miss anything, and I don't have a problem waiting a while longer and spending more money if that means the job is done right the first time.

[hard drum]

It means my usage is faster than the limit could keep up with

[sacred minnow]

xhigh also overcomplicates stuff sometimes

[uneven kayak]

I'd rather it think through every possibility than miss something because it didn't think enough, over complication is fine most of the time

[sacred minnow]

yeah i usually tell it to do research and stuff first so it doesnt end up rewriting functions and commenting out code i hate -_-

cc was so bad for that

[uneven kayak]

xHigh loves to check for existing functions before it does anything, something I really appreciate. I have a lot of production gates and status checks and a complex PR flow and a lot of invariances, among other things, and xHigh catches all of that and doesn't skip things like high and medium do.

[sacred minnow]

i use it alot but switched to medium for a little and use xhigh for the planing when i do use plan

[uneven kayak]

I mean don't get me wrong, I want to save money. I tried medium for a while but it let me down so many times that I feel like xHigh is the only reasonable choice for my project

[plush harbor]

it really does. I just completely ran out of things to get codex to do. Again. So that's two projects now ... however this one needs a human on it for a bit

[sacred minnow]

dont think ive ever hit limits and i ushally run big subagent teams with split from spark agents xhigh and so on

[plush harbor]

I came very close to hitting limits yesterday

[uneven kayak]

Codex useage limits are such that I actually can use xHigh and not have any problems. So I might as well lol

[sacred minnow]

o i somehow hit 5hr yesterday sorry my bad

[uneven kayak]

Codex is SUCH a better value than Cursor.

But that's probably because Cursor is sort of a premium service and uses the API

[plush harbor]

limits are apparently not my issue, thinking of things I can palm off to codex is the actual limiter. It has put some css framework in this pass but it still looks bad

[uneven kayak]

What you need is a more complex project. Try making a mobile game lol

[sacred minnow]

lol i work on a game and i made a program for everything its so broken i love codex/ai

i just redid our live website for fun

[plush harbor]

ikr, none of my stuff is particularly complex. This one I could have got away with ignoring but it was blowing up cache/google's crawl and killing my server so this was a performance pass. And the retheme because it looks embarassingly bad

[sacred minnow]

imagen in the uis has been alot of fun

[plush harbor]

I got gemini to make me some ux icons, chatgpt to make a repeating background and codex has done all the scaffolding. I have to glue all this lot together

[sacred minnow]

i did a full ui remake for fun it turned out really nice it was just codex are we allowed to post a screen shot o.o

[uneven kayak]

When I run out of credits on Scenario, I'll switch to GPT Image 2 through $imagegen. Spent $115 on Scenario credits so I'm gonna use every last one before switching over lol

[sacred minnow]

this turned out nice it was just codex hopefully this is allowed or i can delete

cleaned up very nice i was really happy with it

[uneven kayak]

Nice

[sacred minnow]

i still remember the first days of gpt xD it taught me so much pyton and sql now its just dang

[uneven kayak]

Yeah the good old days of copying and pasting snippets, figuring out where they're supposed to go, so you pretty much have to know what you're doing. And then all the compilation errors lol

But now the agent just hands you code that works after testing it autonomously, it's blowing my mind every time I think about it. It's almost like we live in a sci-fi movie.

[plush harbor]

bit better than its usual default beige for sure

[oak trellis]

hmm the limit ration .. i need to validate that but right now 24 % of daily limit consumed 4% of weekly limits. Was messuring that in the morning on 100 USD plan for 5.5 medium

[sacred minnow]

ive been playing with the uis so much i still have not decided what one ive made i should make all my other programs follow lol

[plush harbor]

my biggest project is all blocks of bright pink, green, blue, purple. The one I'm doing now is stone and bronze. Variety and all that

[uneven kayak]

My UI design plan involved two stages, the first stage is complete with placeholders and CSS, now Codex is working on step 6 of the second stage, replacing absolutely every single UI element with a raster. The game is already looking ten times better.

Then I had the UI/UX functionality implementation in between those two stages, that took a couple days of just telling Codex "Very good, please proceed" lol

[plush harbor]

I had it do this by default. So now I get to bludgeon it into submission with css, cos that's not what I want

also its apparently really, really bad at marketing copy

[uneven kayak]

Looks decent enough for a simple graph

[uneven kayak]

Claude 4.6 Sonnet is great at creative writing

[plush harbor]

the graph is fine. Its the stuff around it that's all wrong. That's just a trimmed down version for a teaser

[plush harbor]

so is gemini, I've been using it for all my teasers

[uneven kayak]

Ah alrighty

I'm so much looking forward to posting a screenshot of my game, but not just yet 😆

I'm aiming for the highest possible quality

Basically the idea of, if the big apps can do it, so can I

[plush harbor]

I'm casually redoing the guts and theme of this site over 2 days and then will YOLO it at prod. Moved it onto staging now

[uneven kayak]

Lol nice

[plush harbor]

not sure how I feel about being the slow one around here, bottlenecking this release

[uneven kayak]

I posted a video the other day of a very early build because I was too excited to wait, but the final design is shaping up to be very different than the video showed #codex-show-and-tell message

[daring forge]

Sorry if this has been posted before, but I've given 5.5 a true run in the Cursor Codex extension, and it's the first time since using Codex for the last 8 months or so that I feel it's a downgrade within the IDE. Config.toml actively ignores context window command, and my AGENTS.md pointers and guardrails are too robust in many repos that I hit auto compaction in two or three turns, and then spend significant time having Codex reload the same invariants and enforcements instead of focusing on the actual task. 5.4 could do all of this even without the 1 mil context window. At a loss here.

[signal wraith]

Hi, I just upgraded to PRO because I ran out of credits, but Codex keeps telling me I don't have any credits. Has anyone else had this happen? What should I do?

[daring forge]

Hi, how are you interfacing with Codex? Extension?

[signal wraith]

Nop yet for me

[daring forge]

I had that happen when I was using my personal tokens, then bought my Pro sub. Had to log out and back in on the extension. Perhaps same thing if you're using it elsewhere?

[signal wraith]

i will try to connect and reconnect first it may be that

[daring forge]

How are you using it? Web/CLI/app?

[signal wraith]

Web-App : CLI may be one day 😉

[daring forge]

Have it tell you how to transition 😛 My recent post reads negatively, but I am overall impressed by Codex. But if you're on web, I bet a log out and back in would help. Then check usage page.

[signal wraith]

you was Right Bro

[daring forge]

Yay!

[signal wraith]

-> Full Deco , and Reco

Thanks you so much

[daring forge]

Happy I could help! Happy coding.

[signal wraith]

I just Coding a DemoScene Demo Maker 🙂 if you know about theses time

[daring forge]

Not familiar with how it's done currently, but sounds interesting for sure

[signal wraith]

Pm is ok for you ?

[daring forge]

Sure!

[oak trellis]

feels like the generous credits gone haha

[daring forge]

Feels like a case of "Codex everywhere" and no one can reasonably account for all the unintended negative effects when you do that kind of scope.

[red prism]

There are some attempts here and there. It runs well on NixOS with this https://github.com/mwmdev/codex-desktop-nixos

[zenith linden]

Im new to codex and coding. If anyone has some tips it would be appreciated

I work in tech and would be interested in integrating codex into my workflow and making money with it

[undone patio]

lol

[plush harbor]

careful. As a codex newbie, this thing is fun

[turbid axle]

codex might as well be called crack

[karmic gulch]

Anyone else goign back to 5.4 and even 5.3 because 5.5 is eating up the usage? I got the 200 plan, and its starting to act like the 200 plan on opus...

[turbid axle]

I use 5.5 medium on 200pro. works alright for me. but I have to manage myself from using too much /goal

[karmic gulch]

Guess I`ve gotten lazy, and setup my workers with to much automation that eats up usage. But there is a huge difference in usage from 5.4 to 5.5. Not sure 5.5 is worht the extra usage ..hm

[turbid axle]

yeh, I cut down on a lot of that. I use the oai memories system, but subagents and such I turned off among other things that eat tokens like mad

[karmic gulch]

My solution for a while was to use the 50 dollar plan on minimax, and setup up local mcp that codex can call on for the dirty work, but they can say what they want, minimax does not give nearly the same quality output as codex/claude, so I just killed that thought

[turbid axle]

thing I find with 5.5 is that it is just really good at tool use and following plans. so I can remove a lot of ensurance automation type stuff

[oak trellis]

its eating now limits muuuuch faster ..

🙁

[cedar skiff]

If a cheaper model can do the task it's 100% worth using the cheaper model.

Any orchestration i get it to design i always make sure it designs it for 5.3

[short pebble]

claude users got 2x usage

[cedar skiff]

the 5 hour window got extended to 2x not weekly

so they can burn usage faster lol

[short pebble]

💔

[cedar skiff]

it'll take a little time for them to get the infra setup and working for them

then they might see some usage boosts

[plush harbor]

musk probably just rented them some server space out of spite because he doesn't like altman. Gotta love personal vendettas

[cedar skiff]

probably something to do with it

[plush harbor]

its taking me longer to fiddle with this css than it took codex to do the entire rest of this body of work

[short pebble]

unless you have thousands of dollars to drop on claude credits i dont see being able to sustainably use it

even on the 200$ plann

[cedar skiff]

Codex has much more usage currently even with the large cost jumps from 5.3 -> 5.5

codex also gives image gen and chat unmetered

[uneven kayak]

Nice, just merged my 200th PR for this project. 😄

[brave kettle]

can someone confirmed that Codex CLI is not (yet) supported MCP's prompts (https://modelcontextprotocol.io/specification/2025-06-18/server/prompts).

my MCP's prompts are works on another agent but not on Codex CLI or i missing something?

[uneven kayak]

omg codex can make fonts but it's like pulling teeth getting it to understand certain things, "what the hell even are you doing, just stop with the nonsense and make a letter A that actually looks like the letter A" was in my last prompt lmao

despite the minor frustration however, it's coming along nicely.

[oak trellis]

i need weekly 2x .. the 5hour window lol there iam on 40% usage max

[oak trellis]

is it still available ?

[wicked briar]

they always do opposite of what userbase wants

[oak trellis]

always something ..yes

the 5hour window is just marketing has 0 impact ..

[plush harbor]

this is why I am doing this css myself

[uneven kayak]

I just want a font that I can use in production and not have to pay $1,000 for a commercial license, and I don't want to use a system font because I want it to match the style of the game, so I figured I'd try making a custom font for the game.

It's actually almost done, been working on it for a couple hours and I'm really liking how it's turning out

I'll post a screenshot in #codex-show-and-tell soon

[cedar skiff]

Yeah, it's deeper in the menu, i just tell the agent to use a subagent with that model

[oak trellis]

oooh and that saves you a lot ?

5.3 ?

[cedar skiff]

Yeah its considerably cheaper, 5.5 is more token efficient but that isnt even close enough to make the cost difference

[tiny fulcrum]

not really a fair comparison, 5.3 high is closer to low than it is to medium

[cedar skiff]

just doesnt make a difference, 5.3 is cheaper for mid complexity tasks and it handles them fine

[tiny fulcrum]

there is quite a difference between low and medium

[cedar skiff]

its not even close, low cant do the tasks 5.3 high can.

[tiny fulcrum]

probalby will cut the costs in half

at least

it depends on the task

and I also doubt that benchmark, would have to see for myself where 5.3 high finishes a task and 5.5 low doesn't, while also using significantly fewer tokens, seems unlikely

[cedar skiff]

I had plenty, it just doesnt cut it. I tried many times with different iterations doing repetitive tasks. 5.5 is more expensive in every case. and some times 5.5 low doesnt finish it uses a blocker as an escape hatch.

[tiny fulcrum]

yes 5.5 is more expensive it's pretty clear; but can't really say 5.3 outperforms 5.5 low when I have to use 5.3 medium or high, usually ends up the same or more

[cedar skiff]

fine you do you, i have many repetitive tasks to try this on and the results are always the same.
All i have to do is change the configuration in the orchestration and see the difference.

that above is xhigh vs xhigh

[tiny fulcrum]

sure, it is task dependant

I usually do c# stuff which is not front end heavy, so it might be entirely different to what you are doing

[cedar skiff]

I think you are just kidding your self, you just think it is what you think with no real actual idea.

[oak trellis]

ok so sub agent ok i use them now

[cedar skiff]

I just cant see how you could possibly come to your conclusion in any other fashion

[oak trellis]

thx sir like always i have 54% weekly left

until 12 :/

[tiny fulcrum]

not really, I'm actually looking at the analytics and break down the token usage, especially since this is new for us and people who actually make decisions need some idea and evidence about budget

[cedar skiff]

ahuh

[turbid axle]

main tasks 5.5 is better at I find is when it needs to use tools, or for multistep jobs

[tiny fulcrum]

I'm not doing some youtuber stuff, or watch some videos, I look at the analytics and even set the guidelines for our software team, so I have some idea since it is part of my job

[turbid axle]

I had it write my own computer operator type app some time ago. and 5.5 is really good at using it, pretty magical to see an ai build and test native apps with a tool it wrote by itself 😄

[cedar skiff]

5.5 high is really good at orchestration

it just never stops

[turbid axle]

really the only wall I see is tokens and time. I need more of both

you can now build anything, really is starting to feel super close

[cedar skiff]

1.8 points difference in swe bench o.0

xhigh vs xhigh - 5.5 is much a better generalist, it tool calls better, it understands semantics better, it does better orchestration, it's so much better at longer running tasks it's not even close. But it didnt jump much in coding. Trying to say 5.5 low == 5.3 high in coding is just objectively wrong. at xhigh it's only 1.8 points higher than 5.3 in swe-pro. It's much cheaper to have 5.5 plan and orchestrate 5.3 implementers than having 5.5 do both.

[tiny fulcrum]

I never said 5.3 high = 5.5 low, I said it gets the job done at equal token/price efficiency or better depending on the input / reasoning / output

on that note the models aren't getting that much better anyway, this is stricly reasoning thing, where 5.3 high might use a lot more reasoning tokens to get a job done, which will take longer and could cost more

it really depends on the task

[turbid axle]

where 5.6

[chrome raven]

wen reset

/goal has dried my limit out 😄

[turbid axle]

yeh careful with that beast

[languid kernel]

with that 2x on pro 200 dollar

i cant hit my weekly limt

its really crazy how much usage you get

[chrome raven]

i hit mine too easy

probably you are no fast?

[languid kernel]

no fast indeed

[chrome raven]

yes it's crazy useful

[languid kernel]

i just have multiple goals running every day to check for things in the codebase

[chrome raven]

yes if you enable fast it will drain faster but also work faster

[languid kernel]

security. improvements

etc

i just parralize so 2-3 projects

so when 1 is done i can look at an other

[languid kernel]

yeah fair i just dont want to get use to fast mode

when the 2x si gone

ahhaahha

[chrome raven]

yes that's what im facing now, i cant get off the fast mode

[languid kernel]

hahahahaa

[turbid axle]

do you also drive a lambo? jeez

[cedar skiff]

Youre just guessing and i'm telling it isnt cheaper and it's dumber. It would have to be 53.3% more output-token efficient. to break even. But it can't even get the task done. You also said its not a fair comparison becuase 5.3 is closer to low?! not really a fair comparison, 5.3 high is closer to low than it is to medium You just make it up as you go.

[languid kernel]

i wish

ahahah

but also reasoning on medium helps a lot

[meager wing]

Can someone confirm gpt-5.5 pro is available in Codex MacOS app and CLI For ChatGPT Pro $100 plan?

[orchid plume]

it is not, it's only available in ChatGPT

however with the Codex app you can use Browser Use to indirectly use the Pro model via the ChatGPT website, login on Codex's internal browser to ChatGPT and make sure it's set to Pro model

[boreal holly]

If you need help getting 5.5 to be more efficient lemme know. I think it’s legitimately better in every conceivable way to all previous and competing models, and costs less overall to use. I read through your discussion and the fact that 5.3 out performs 5.5 on your setup is pretty shocking, but likely also fixable.

[woven canyon]

What UI is that?

[cyan gyro]

Always interested to read personal experience/tips

[woven canyon]

How 💀

[boreal holly]

My entire ~/.codex is version controlled, so today when I get back at my desk I'll try to compile down all the commits since 5.4 released til now and see if I can slap together a list of highest impact changes.

[lost drum]

yo idk if its a thing but in VS codex ext I enabled in config goals but I dont see the command /goal when I type it But I just slamed it anyway and it works I mean Idk if it works but he those not stop I mean if he stops and emits summary he just goes back in to executing somehow

and I now dont know if thats how goal feature works or my thread is glitching

[prisma pelican]

Has anybody else had issues with Codex Desktop in-app browser and computer use on Windows. Everything else in Codex Desktop works fine, but these 2 just don't.

[slim anchor]

gm

[dim jacinth]

feel so good , after updates , use codex app5.5 xhigh in my scaffold can run a long task,and it almost keep working 100hrs now , really excited to see the progress,won’t be disrupted easily .🫡

[torpid trout]

Can someone shed light on this?
https://www.reddit.com/r/codex/comments/1t66opm/this_may_have_went_unnoticed/

If that is true, that will be hurting very much
I can almost not believe it because the two have nothing to do with each other, they do not share any feature, at all

Often not even model or effort.

So how could using chatgpt count towards a codex limit.

Asking gpt on the subject it said:

If you’re just using the normal ChatGPT app (regular chats, GPTs, voice, uploads, etc.), that usage does not consume the separate Codex 5-hour / weekly quotas.

Those Codex quotas are for actual Codex surfaces/features, such as:

* Codex CLI
* Codex app
* IDE integrations
* OAuth-based Codex tooling
* Codex agent sessions

The shared limits language applies across those Codex surfaces, not ordinary ChatGPT chatting. OpenAI explicitly phrases it as “everywhere you use Codex.”

So the reddit post seems nonsense/misinterpretation/fearmongering

[uneven kayak]

great I just wanted to change the picture showing my font and now I gotta wait for the slow mode cooldown lol

[uneven kayak]

well, I don't see anything abnormal about this. They probably used Codex Cloud in the browser and misinterpreted it.

[turbid axle]

I would actually prefer it all just be one token count personally. I use codex way more than chat. would nice if I can eat those tokens for codex as well, much will go unused im sure atm

just gimme a huge pile of token to nibble from

ill be happy

[uneven kayak]

Alrighty I made a custom font for my game 😄 I'm pretty happy with how it turned out but I might make some minor adjustments, plus I need to get it in front of some people who speak languages other than English for testing. #codex-show-and-tell message

[torpid trout]

yeah got it. Buffled as of what they expected then lol

[turbid axle]

image model 2 is excellent for memes 😄

[torpid trout]

Free tokens or what?

[turbid axle]

I wish

[uneven kayak]

lol I wish

[torpid trout]

I asked it once to do that, then I discovered it copied a 500USD font (realised that after I searched for similar fonts) As I asked how, it said "it was wrong" - too late buddy. Font was made. Lol
But it took a heavy amount of pressing, and some manual rework after - not the perfect result all in all.

[uneven kayak]

damn that sucks lol

but that's not possible with the way I made Bureaucracy, it started off trying to copy other fonts and applying graphics on top but then I switched gears and ended up using extremely specific examples and prompting and a lot of tweaking, well over 100 iterations going back and forth about every letter and every tiny detail

I had like 4 follow-ups about the number 3 alone lol

and the infill took like three hours to get right

as far as agents go, that's a long time lol

[uneven kayak]

burned through 16% of my weekly useage working on that font

but I also had another thing going on with another agent so idk, probably closer to 12%

so not terrible but that is Pro

[cyan gyro]

it's quiet on the codex front...

[vital hill]

when does the extra usage end?

[uneven kayak]

May 31st

[vital hill]

word

[short pebble]

EXTRA USAGE?

[vital hill]

yeah i was using the desktop app for like 9 hours straight and didn't even hit 0% usage

on the plus plan

[boreal holly]

Hmmm, if that's the case, 1 run of GPT 5.5 Pro on ChatGPT would eat like 10% weekly in one shot

[short pebble]

huh

same

im at plus

no unl usage for me

[vital hill]

probably because i've been using codex 5.3 medium

[short pebble]

whys ur differnt looking

[vital hill]

codexbar

[wicked briar]

@cedar skiff hey

[short pebble]

tyty

[short pebble]

so its only for 5.3 medium?

i just spam 5.5 high or ultra

[vital hill]

the extra usage should be for all models

[short pebble]

when did the extra usage start

[vital hill]

not sure. maybe sometime eawrlier this week

[short pebble]

why do i not have it then???

[vital hill]

well you are using 5.5 high or ultra

[short pebble]

right now high

nvm it was on extra high

DOES that matter tho

[solid lake]

Wait is there an ultra

[boreal holly]

Yeah, the difference between high and xhigh is double the quota usage for less than 1% higher accuracy

[solid lake]

Does anyone remember ultrathink in claude

Is there anything like that in codex?

[vital hill]

yes cuz 5.5 with max reasoning will waste your usage % quick

[cyan gyro]

I dont have factual evidence but to me 5.5 xhigh appears to be smarter than 5.5 high though, maybe not in coding itself, but at least in understanding and thinking along.

[short pebble]

xhigh tends to overthink alot i only use xhigh for q&a's

creates its own variables

[wicked briar]

xhigh is best

[short pebble]

.

[torpid trout]

I honestly have not liked 5.5 xhigh. high seems the golden standard. High is always better than too high, or not high at all

[obsidian lotus]

AI has looong way to go https://programbench.com/ 🤣

[short pebble]

give me opus 4.7 unlimited creds and ill rule the world ong

[vital hill]

codex better

[short pebble]

5.5 or opus 4.7 creds and i wont sleep ever againn

[torpid trout]

I am very nuttily excited about this
Codex MCP driven project management from a to z, inclusive but not limited to working off gh tasks, deterministic orchestrator, ai driven Project Manager, agents and skills, project knowledge, repo awareness, and yes, little chatbot that helps you set up the whole app and improve the app's code if you wish so

The real-life test begins now.

[lost drum]

ye I switched to high after a week and I dont see any diff

[rocky fog]

I was on xhigh first with 5.5
took a ton of usage and took a while until it did stuff, perhaps it was also overcomplicating some things

on medium, usage barely moves, but I think it was making more mistakes or missing things

settled on high and that one seems fine in all ways

[boreal holly]

I love that Codex is practically self-improving.

[torpid trout]

Yes, well, I did not yet try to actually have it edit its own code in that chatbot lol, but, in theory at least it should not be an issue until reload (which is when the new code would be loaded into memory)

Anyway, that chat bot is just a nice to have addition, after all the "do not create another chat bot" i wanted to let codex have some fun lol, and it did a quite nice job actually, it seems to love chatbots 🤣

It was particularly helpful for understanding the app (which supposedly "I" created)

I think I will make it a standard in any app built with codex from now on lol, along with the obligate documentation-no-one-ever-reads

In this specific case it can also actually create agents or so, from scratch, based on my "blunder"

[steady vigil]

anyone using codex security?I did an initial commit for a large mono repo... it found one medium issue, and now is back to waiting for commits.

[wicked briar]

how do you use it

[steady vigil]

its in the cloud

[wicked briar]

hey greg

long time no see

damn used to talk back in the day in claude server

[steady vigil]

yeah I have not much to complain about so 😛

[wicked briar]

haha

[steady vigil]

codex 5.5 working well

[solid lake]

Claude refugees

[wicked briar]

indeed

[steady vigil]

I'm enjoying Pi also. For me its my second fav after codex. much nicer then opencode

[wicked briar]

hmm

[honest creek]

aaaany minute now...

[hard drum]

WYSI

[woven canyon]

Reset when 🥺

[torpid trout]

I have a new problem tho

[lean lark]

Dude, you SO have your finger on the pulse. 😍 🤟 💌 🙇‍♂️

[torpid trout]

I am not sure if you have my wife's hidden sarcastic humour or if this is genuine brotherly love lol
I will take it as the latter 🤣

[lean lark]

latter, truly

I dunno about brotherly yet, but I sure as hell appreciate your spot-on insight on this one topic. We'll see what happens over time. 🥀 😆

[torpid trout]

BTW gpt has decided to baptise the character in the meme as Ponzi Homunculus

[gusty pagoda]

how much better is the codex standalone app as opposed to the extension on vscode? i'm currently using the extension and i feel like it's just fine, but i'm on ubuntu and there isn't a standalone codex app for ubuntu just yet. i can use my macbook but it might be a hassle to change it being my main workstation

[lean lark]

With a more indepth analysis of the challenge, ChatGPT returns the following options:

Quota Homunculus

It hits the character cleanly: a small, malformed, impulse-driven creature whose entire worldview is "reset wen?"

A few stronger Latin-flavored options:

Homunculus Quotarius
"Quota little-man." Best pseudo-Latin form. Sounds academic, dumb, and insulting without being too obscure.

Resetus Goblinus
Not real Latin, but instantly funny. The "reset goblin" who appears whenever quotas run low.

Tokenis Incinerator
"Burner of tokens." Fake Latin but readable. Good for the second image.

Homo Reseticus
A mock species name. "The reset-seeking man." Very memeable.

Homo Quotavorus
"Quota-eating man." This is probably the best coined taxonomy-style insult. "Vorare" means "to devour", so "quotavorus" reads as "quota-devouring."

Homunculus Resetensis
"The homunculus of the reset." Sounds like something found in a developer dungeon.

Codex Resetivore
Plain-English/Latin hybrid. More understandable than pure Latin. Means "one who feeds on resets."

---

I don't think we can get Homo Reseticus through the filter all the time, so far I'm inclined toward Homunculus Resetensis or Homunculus Quotarius.
Any anthropologists here?

Sorry for dropping that after your good question, @gusty pagoda .

[boreal holly]

Turn on image generation 🤪 it burns tokens but is also somewhat useful

[lean lark]

@gusty pagoda Are you using the Codex CLI at all yet? That might be a stepping stone to the app. You can also use ChatGPT to operate on your repositories. That's extremely cool. I've also recently published a very simple wrapper around Codex that creates and modifies projects and does any other CLI operation for you. It's a braindead OpenClaw. 🙂

[boreal holly]

Homo Reseticus absolutely slaps

It's the Idiocracy-flavored evolution of man & machine

[lean lark]

Unfortunately the word Homo can be perceived by anyone (who resembles the character) as being slanderous. 🙄 Given that it's a legit Latin word and capitalized as-Latin, I'm inclined to go with it, and damn the sensitivities of idiots who can't distinguish millennia-old language from contemporary slang.

[boreal holly]

The benefits far outweigh the consequences

[torpid trout]

Why would I get Session not found for thread_id: 019e036e-7be5-7653-b623-263898e27375 on a codex mcp codex-reply passing a thread id?
That thread id is the id codex just gave me upon a codex (first message)

[lost drum]

disgusting piece of creatures

but its beautifull

[lean lark]

I believe @torpid trout is the appropriate person to present that meme created by @boreal holly to Tibo, a gift from #codex-discussions.

[simple star]

wth

[lean lark]

If you don't understand the subtlety ... don't worry about it.

[torpid trout]

This is dangerously important to me, anyone ever used the codex mcp server?

[lean lark]

not here, sorry, watching to get an education. What are you doing that requires the UUID?

[torpid trout]

supposedly you can use codex mcp like codex exec or codex itself, and you have two tools, codex and codex-reply, wherease the latter supposedly takes a thread id so you can resume a thread started with codex

That, at least, is what gpt cooked up

Yet, it constantly fails on that "thread not found" error, leading me to believe it is all hallucinated spaghetti, which would destroy a fairly advanced thing it made for me lol

[lean lark]

researching...

[torpid trout]

codex-reply uses persistent thread IDs to maintain context across multi-turn conversations, enabling session continuity even after restarts. These threads, typically stored in ~/.codex/sessions, allow users to resume, fork, or archive conversations using specific IDs.
From some random online doc, since openai doc on the subject is a bit measle

So I see no reason this woudl not work... I guess I will have to do a minimal sample to see what's going on here.

[lean lark]

Are you using the original "structuredContent.threadId" ? I dunno if there is a different threadId prop.

[torpid trout]

Are you using
I will ask codex what he is using, lol

I just confirmed that at least the /.codex/sessions are there, so... it should really be resumable.

[lean lark]

codex-reply only knows in-memory sessions, not persisted disk sessions.

[lost drum]

I can give you prompt

maybe it will help

[lean lark]

That's a bit too vague @lost drum ...

[torpid trout]

huh?
So each time you do mcp-server its wiped?
But the chats are in the session folder!

Did you ask codex for that?
It told me something similar in a previous step, then stepped back from that

[lean lark]

This is the pattern I'm seeing. If the session is persisted and then resumed there might be a different thread. Looking...

const first = await mcp.callTool("codex", {
  prompt,
  cwd,
});

const threadId = first.structuredContent?.threadId;

await mcp.callTool("codex-reply", {
  threadId,
  prompt: "Continue from the previous turn..."
});

For durability across restarts, codex-reply may be the wrong interface. Use Codex App Server thread/resume or the Codex SDK resume-thread flow instead.

[torpid trout]

yeah but codex app server is an entirely different thing 🙁
diu!
Typical "trust the thing" pitfall I guess.

await mcp.callTool("codex-reply", {
  threadId,
  prompt: "Continue from the previous turn..."
});```

this to me however looks just exactly as it should
passes a thread id... so...

[lean lark]

Is your process stopping after the first enquiry?

[torpid trout]

I guess so, no - more precise, yes
Because it should not matter. These threads are not ephemeral, they are on my disk

I am having mr. codex try to keep the process open instead to see if that works, but it would still cause issues since my process can pause (in the UI) and I would not want a codex MCP to just run there idle on a machine. well, maybe that is not so scary as it sounds after all

I started off from this https://developers.openai.com/cookbook/examples/codex/codex_mcp_agents_sdk/building_consistent_workflows_codex_cli_agents_sdk#orchestrating-multi-agent-workflows but that is like just the seed of a massive forest meanwhile, but you get the gist

[lean lark]

nooo, In general terms, and remember I don't have experience with these specific libs yet... Thread IDs are ephemeral. While they are UUIDs/GUIDs, their use is generally not persisted across sessions or processes after a server has been restarted. Persistence would only be guaranteed if we're sure that the entire session is persisted under that UUID key. I kinda doubt that's happening.

That said, Codex App Server is intended to support durable persisted sessions. MCP codex/codex-reply do not seem to be designed for that use-case.

[upper kraken]

.

[lean lark]

If you want to "resume" a session with this MCP process, after it's been terminated, you may need to look at that session data, extract the transcript, and pass that all back in to a new process, followed by a new prompt ... just like we do manually for other stuff like this.

[torpid trout]

yeah that aint happening, I cannot pollute the context window any more
If this is really the case I will have to switch to the codex exec or codex app server. I guess with usage of mr "i am sure"-ai it should not be that hard to do the switch

Still waiting for the current response, I will know soon I guess

[lean lark]

yeah, and don't trust me on any of this, I'm trying to assimilate the info as I go.
https://community.openai.com/ would probably be a much better place for this nuanced use of the current libs.

[torpid trout]

I am banned there, many moons ago when openai tried to charge users twice on their credit card (and I was only one of a high number of affected users), they tried to mob that under a nice rug of banning everyone who spoke about it in public
You can still find most of the threads, but the folks that really insisted are gone-d 🙂

[lean lark]

Ugh. That Discourse server is a good platform but it's mis-used by users and incorrectly used by their mass of admins - I feel the pain.

[torpid trout]

Anyway, here we go

The probe reproduced the real issue: codex-reply works in the same codex mcp-server process, but a fresh MCP process returns Session not found for the same threadId.
That means the current committed code cannot truly resume PM threads after closing the MCP server; I’m going to implement a PM MCP session manager rather than just
surfacing the failure.
(I did not tell it to do that, and I do not like that it does it... but I guess I have me my reason as of "why")

[torpid trout]

The admins there are weird at best, and (censored) at worst
I guess the face is more public so they try to keep stuff away that might shape them in a less-than-optimal light

[lean lark]

Yay ... kinda ... I found a GH item for exactly what you're doing. It was just closed last week for not having any +1s. (I hate the way they do that) But you can +1 it and maybe present a business case for it. It seems this may be possible with a fork+patch ... but few people are inclined to do that.
https://github.com/openai/codex/issues/12596

[torpid trout]

I am going to have give up stable MCP tooling and conventions with experimental messy propiertary app server integration
🙄

[lean lark]

Well, I dunno about experimental, but OpenAI does come up with a new "best" SDK every year. That's unfortunately a sign of the times. I think we're compelled to bite the bullet, modularize the hell out of what we do (good practice anyway), get specific modules to do specific things with whatever lib/tech is required, and just move on.

Summary: Maybe use a separate module just to support this specific MCP use-case?

[torpid trout]

well it all comes down to keeping the MCP session alive - which is another way of saying "if for whatever reason your mcp restarts, you lose all your thread histories"
I guess I could have codex build its own rehydration logic like that user in the GH did.

I am weighing the pros and cons of using an MCP vs the app server now. I lean towards hacking MCP as much as possible because at least it is a standard, the app server is literally a daily mood of codex operated in oai cantines

[lean lark]

I see MCP purely for function calls, for tools, for events. Why might we want to keep a MCP session persistent across sessions to continue a thread? Is there a common reference MCP server or tool that can be cited as an example?

[torpid trout]

their very workbook makes zero sense if threads are ephemeral unless it was purely "well, we can do that, but its useless" example

Who wants a orchestrator that cannot remember even its own previous run?

[runic quarry]

hi guys, how is your experience with codex?

[torpid trout]

but you are right, MCP usually is not used for that. I merely chose it because it is a standard.
They also would have the SDK, but codex said that is overkill for this

However, it is already moving the codebase to app server now.

[runic quarry]

i have the $100 plan for about 2 weeks and a half and it's amazing with gpt 5.5

[lean lark]

Most of us here love it, amazed at what we can do with it.

[runic quarry]

yap, i mean i started using codex since end of february and i was blown away that i can actually start expressing my ideas without worrying about implementation

and without having to always copy and paste errors it would produce, it checks and fixes everything by itself

and now with gpt 5.5 is insane

i m curios where we will be in 2 years

[torpid trout]

under a bridge asking for money, where else
We are actively developing our all doom lol

[runic quarry]

that s a good point :))

[torpid trout]

||sarcasm is not everyone's strength||

[runic quarry]

claude is insane with how expensive is and not so much time until codex becomes expensive as well
because right now they are burning money

it s unsustainable

especially with the need to scale the compute power

[lean lark]

@torpid trout Without more understanding of your use case and why you'd want to keep a MCP session alive for a while, I don't know if it's worth pursuing that approach. If a MCP tool exposes that much functionality where long-term persistence is required, I'd doubt the engineering of the feature-set. That is, the MCP tooling is subject to scrutiny if it's not designed as being a non-persistent resource. Maybe the answer is less about questioning the use of the Codex App Server (which I totally understand and havent coded into yet for the same reasons), but about questioning that one (?) server/tool that seems to have this very unusual requirement for persistence. Oh, and that make me wonder why the MCP server/tool itself wouldn't support its own threadId if it does require context persistence. 🤔

[torpid trout]

I am not creating an MCP tool actually, so the using MCP codex is an unusual choice - as said the only reason being is its a standard, like an API, and if it changes it can only change so much

Its all dumped already - codex boy is ripping through the codebase and migrating everything to python SDK wrapping app server (which is experimental, but hell, yolo)

[torpid trout]

This is what's it for

[lean lark]

ahhh

OK, looks like you're on your way. 🙂

[chilly mango]

Why do conversations get so laggy when they get long? Its so frustrating.

[torpid trout]

app?

[chilly mango]

Windows app

[lean lark]

General note: It seems the advanced users here are starting to use Codex for more of these "just do it for me" projects. I don't see that as "vibing" so much as lending more trust to the tooling as it has become earned. I continue to bend toward this mode as well, V.e.r.y. slowly. I know enough Not to trust this stuff, but I also can't deny the amazingly high quality that's been demonstrated in such a short period of time. I'm starting to design and allocate new projects where I'm not babysitting every function. I can't get myself to admit I'm vibing yet.

[lean lark]

That might be related to the connector being used, the streaming connector versus the non-persistent HTTP connector.

[chilly mango]

I'm not familiar with these terms.

[lean lark]

Open the Discord Search form here and search for "in: codex-discussions from: robertmsale websocket".

[torpid trout]

I am not vibing, I am abstracting. And honestly.. That is what enabled me to program 15 years ago. Not the understanding of PHP (I still dont understand how someone could ever code in PHP, yet I do) - but the ability to comprehend language, and abstract "I wants" into "a > b > c"

[torpid trout]

Its the whole electorn mess. Laggy not equal to slow in response, I think the websocket is related to communication with teheir server, not local lagginess due to crappy code driving the app and leaking memory on long chats

not here.
(pay attention to the dot at the end of that phrase)

[lean lark]

But aren't you/we giving more responsibility to the tools as "I trust it to do this thing that I want" ... It's an extended form of Declarative coding vs Imperitive, where we trust that tags in code will work so that we don't need to hardcode nitty-gritty syntax-driven functionality.

pronk - we don't know you, we will not refer you. We discuss Codex here, not employment.

[cyan gyro]

Reset! 🙂

[lean lark]

Anyone send that meme to Tibo yet? 🙄

[warped minnow]

When you request for codex ambassador do you receive a confirmation that they received your request

I am not sure but i think my email was wrong and wondering if there is anyway to check (i dont want to send again to not spam)

[lean lark]

( Why would someone who just joined the server feel qualified to be an ambassador? )

[gentle harbor]

are you fr

[lean lark]

Just join the Cyber program - your code tripped a simple trigger. It's easy and they don't bother you.

[boreal holly]

The agent has to provide the same CWD that the thread or MCP server started with for it to detect the thread.

So if you look for the rollout log with that UUID it might say the CWD in the first message, and the agent using the MCP has to provide that CWD. If the rollout log file doesn't exist then definitely a separate issue

[vital hill]

oo desktop update

[cyan gyro]

what changed?

[vital hill]

codex desktop app can now use chrome via codex chrome extension

[lean lark]

And a Codex extension update ... never a clue what changed. 😆

[chilly mango]

Sorry, I posted in the wrong discussion. Apoligies. I meant to be in community help.

[vital hill]

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

[lean lark]

TY

[vital hill]

thought they'd add a plugin for playwright but chrome works better

[lean lark]

Playwright is such a sophisticated developer tool, better to use it from code, IMO.

[torpid trout]

yea, I was just being noisy, of course we are vibing. probably with more understanding than actual vibe coders, but sure, this is not anymore the same as "I write code"

[vital hill]

yeah automation stuff. i just want the damn model to see where it messed up

[tulip harbor]

for some reason the chrome extension does not connect to chrome with CDP enabled

[vital hill]

might wanna report it

[lean lark]

Hmm, looking at the documentation I'm now wondering if they actually are using Playwright, simply not using the headless flag. I do this in my projects to open Chrome on a page when manual auth/challenge is required for content. The processing then continues operating on the pages.

[lost drum]

I am bruned out

I updated codex app and my chat got removed

or idk where it is

[vital hill]

i used playwright mcp in windsurf IDE, it uses chrome and i can see it scroll and open other links

been wondering how i can do the same for codex desktop app but then this update came along

[torpid trout]

My token reserve keeps growin faster than my consumption
I am already hunting the thing down and uphill lol
I might enable fast just for fun tomorrow.

Btw this CodexBar tool is a real GOAT

[vital hill]

set your reasoning to xhigh

[jolly lily]

Are the 5x and 20x plans actually worth it if used for producing projects/work?

[vital hill]

this thing is amazing

[torpid trout]

Did you just read my comment above lol

[jolly lily]

Thanks for the reaction but it doesn't help honestly

[torpid trout]

yeah. Well, I do not have five providers, but it is still amazing

[torpid trout]

You will not be able to use it up if using it with even just a little coordination. Does that help?

[vital hill]

lol that announcement is late

[jolly lily]

The limits you mean?

[torpid trout]

I am literally stamping out 3 projects a day and running it over client tasks at the same time and it just cant be killed

[torpid trout]

both, that and the actual quality of the model.

[jolly lily]

Are you using the 5x or the 20x subscription specifically if I may ask?

And do you think it has to do something with the current double limit event that it's difficult to hit the limits?

[coarse shuttle]

yoo guys, i was tryna verify on https://chatgpt.com/cyber but when i scan my id its always denied, does anyone have the same issue?

[boreal holly]

20x and plus here. 20x is amazing, wouldn't trade it for anything. Use plus for non-coding

[jolly lily]

Have you ever gotten close to filling the limits?

I assume you can't fill it up anyway

[lean lark]

When you say 5x and 20x, to which subs are you referring? Plus? Business? Pro?

[boreal holly]

Since August 2025, there have been times where I ran out, yeah, but it was mostly the way I was using it. Have enough experience now to not get below 60% weekly (right now with double quota)

[torpid trout]

200 usd plan is not doubled anymore.

[jolly lily]

Oh, the pro subscription is what I'm talking about

[torpid trout]

No proble, just enable subagents and it will be not enough.

But sub agents are usually extremely useless the way they are implemented in codex

[boreal holly]

Yeah, there are settings and workflows that will precipitously drop your quota

[jolly lily]

Well, there practically isn't a subscription with higher limits (if we despite the pay-as-you-go model) so I assume it should somehow be enough to be able to produce something

[boreal holly]

Yeah with 20x you can do an insane amount of forward progress with it. Or you can turn on xhigh /fast + subagents and get hardly anything done as fast as possible

[jolly lily]

May I also ask if you're mostly using it for your own projects?

[boreal holly]

I use it for personal and professional software development

[jolly lily]

Like to produce projects/tools to make profit off them or just for your daily work overall?

Thank you!

[silver dew]

Codex 100$ plan is so refreshing
been actively using Codex for 24h total since and it doesn't run out!

[torpid trout]

profit.

Think 7k phase 1 projects done in the time and cost of 2k

But, codex does not magically bring you 7k projects. That's your job.

[jolly lily]

Thanks!

I really appreciate it

The last question though, would you rather have Claude 20x or Codex 20x any day of the week?

[boreal holly]

Codex 20x

[lean lark]

Suggestion: Try not to approach Codex from the perspective of limits. That's like shopping for cars that have 200MPH on the speedometer. Rather, become familiar with the tools, prompts, AGENTS.md and Skills files, and other options for getting the most of the tech. When you do this, you won't be concerned about limits, you'll be focused on the quality of the work that you're doing.

[jolly lily]

I also appreciate this!

[lean lark]

A focus on limits is just a focus on tokens, not what you do with them. It's an assumption that there will be a ton of waste. If the expectation of waste is reduced, then limits become largely irrelevant.

[jolly lily]

But how can I find out what I'm doing wrong with the prompts overall?

[lean lark]

Ask the bot. 🙂

[boreal holly]

To be fair, it's hard to tell which AI plan will be right for you until you try it. A lotta folks swear by Claude, Copilot, Cursor, Codex (lol they all start with C). I just know my life has changed since Codex came out

[abstract beacon]

does anyone know why the chrome extension doesnt work for me it says connected on the app but disconnected on chrome

[jolly lily]

I use a context file all the time, be as specific as possible with the prompts while maintaining understandable descriptions

[lean lark]

Seriously, ask ChatGPT or Codex to look at your directives in AGENTS.md and Skills files, to help optimize them. Explain your intent and ask it to help refine the text to match your intent. It will do so amazingly well.

Same with prompts. Tell Codex what you want to do, with the instruction not to begin coding but to discuss it with your first. Ask if the approach is correct or if the feature is valid, etc.

Another way I use to significantly reduce token consumption is to ensure Codex has good instructions (1) to document all features in a docs folder, and (2) to rely on the docs folder for information about the workspace. A ton of time/tokens is consumed when the assistant needs to learn the entire project with every new thread. ‼️ Document the code and features and the model won't need to hunt to figure things out.

[jolly lily]

I have two files on the same purposes called CODEBASE.md and AGENTS.md that I always tell it to keep updated throughout the progress, do they actually help?

It prevents the agent from scanning the whole project every single time over easy tasks

[boreal holly]

Set up an "eval" project that you use specifically for tuning agent behaviors. What I do is I have a throwaway project (clone some repo with a lot of code) where I want to test out context management strategies. The idea is this:

1. Craft a highly detailed input prompt with zero ambiguity & set up skills and agents files
2. Run eval (submit prompt), grade the results. If it does not meet expectations, tune the skills and agents files and try again.
3. Once it produces the expected results, add ambiguity to the input prompt. Remove details that would guarantee success, run eval. If it does not meet expectations, tune skills and agents files, and try again.
4. Keep doing this until the input prompt is ~50% ambiguous but the results are 100% what you expect.

This way you have a really strong baseline of instructions where you're confident the agent can handle ambiguity.

[torpid trout]

start with plus. use, if not enough, upgrade to 5x, and so on.

[lean lark]

That's a great start but a single file can't fully describe a large project.
Also, this is more subtle: AGENTS.md files are regarded as model directives, while other files are just informational. So don't put too much in the one CODEBASE.md file. Get the model to refer to files it needs when it needs it. Development of Skills will help with that (I have yet to migrate to skills myself but I know it's a worthwhile effort).

[torpid trout]

Dont even use skills or plugins or agents.
Plain simple.
Have one - one agents.md per repo where you tell it what it has to do in general (like, always do atomic commits, write doc, do/do not introduce backwards compatibility and so on)
have it write milestones. Then use /goal to implement milestone end to end do not stop until done and when done start all over again and spot gaps and deviations and fix them

Seriously. Less is better, specially to evaluate a product

[jolly lily]

I actually want to ask a sincere question, because I'm a little bit lost. I'm currently 17 and I've worked in the software development industry since I was like 13, and coding was actually the main skill back then. Now that we're in 2026 where we use AI-powered agents for the coding sections, what have my skills evolved into, like what even actually makes me worth something in the industry? I've worked with several game development studios most of the time and made successful projects which actually made good a amount of profit but I'm really curious, why would somebody even hire me today?

[torpid trout]

If you can code, you can code. codex cant code, it infers

[jolly lily]

Sorry if my long question is actually difficult to understand, I'm not a native speaker of English

[torpid trout]

And as such, the job/skill is not at risk, it just got a new, masively powreful tool

[lean lark]

Codex is a hammer. Be a carpenter.

Otherwise you're just a monkey swinging a stick.

[torpid trout]

I am that one

[tall notch]

It seems that I can't use the new way of using chrome in codex on Windows, do you have same issue?
Codex
26.506.20924

[jolly lily]

Should I also be happy that it is now easier for me to produce bigger projects on my own and in a shorter amount of time?

[lean lark]

hell yeah

[jolly lily]

Thank you, like I really appreciate the whole conversation, I'm actually now sure that perspective is what matters the most in the industry

[lean lark]

We don't see anyone here worrying about their jobs. We're all kinda giddy about having the ability to do what we've always done, but more quality and quality in less time. It's a shock for many of us.

[cyan wing]

how do i remove remote repos from the Codex app?!
there's no option to remove it....

[boreal holly]

That's actually the greatest benefit AI has to offer. Empowering the individual to accomplish in the same amount of time what massive enterprises can accomplish

[lean lark]

stick around here if you'd like. Good people, good info, good memes.

[boreal holly]

Quit app, find metadata in ~/.codex, delete the remote repos from metadata

[jolly lily]

Honestly yes lol, it's definitely the most professional environment I've seen on Discord lately while maintaining a fun aesthetic

[lean lark]

agreed

[velvet wren]

there are a lot of talented people on this server, I learn something new several times a week

[nocturne folio]

dudeeeee chrome? why doesnt everyone just use zen browserrrr

[lean lark]

and then there is 900000000000000000000000000

[nocturne folio]

the entire world would be a little better if we all use zen tbh

[nocturne folio]

missed a few zeros

[boreal holly]

Faith in humanity almost restored

[jovial prawn]

Where's the LINUX version?

[nocturne folio]

its actually 90 nonillion

[lean lark]

how nice

may we be more familiar and just call you "9e+30"?

[orchid plume]

anyone managed to install the Chrome plugin on Codex? I updated my Codex (Windows) but it's not listing such a plugin

[tall notch]

Does anyone knows if the new GPT-Realtime-2 can be used for basic speech to text?

[tall notch]

EU?

[boreal holly]

Yes!!!

[orchid plume]

ah yes UK, let me guess, another gated feature?

[tall notch]

Seems EU usually gets releases day later 😢

[boreal holly]

OpenAI probably has a plugins repo where you can clone the plugin and kinda force install it

[lean lark]

Twitter seems to be exploding with enthusiasm over today's Codex and Realtime updates.

[orchid plume]

true, time for me to investigate 😄

[chrome raven]

vpn unlocks them, but after a resestart most of the features are gone

[tall zodiac]

When do you guys use the review feature in codex?

Before or after pushing

[boreal holly]

The context window for the new realtime is 4x bigger which is absolutely insane

[lean lark]

never?

[nocturne folio]

yea

[boreal holly]

Before merge

[tall zodiac]

💀 so u never check the code?

[nocturne folio]

https://tenor.com/view/facebook-like-justin-timberlake-gif-19273316

whhhh

[boreal holly]

embed fail

[tall zodiac]

So not after every commit is pushed up?

[boreal holly]

I do a "logic/security" review after each commit, so the review agent isn't reading all changes in the entire branch and iterates on only logical and security errors at the commit level. Then before merge, an adversarial completion review.

[stiff leaf]

hello, can codex do tasks in my computer directly that require GUI ? not just commands ? (Windows)

[lost drum]

my chat is back🥹

[astral nest]

❤️ heartwarming:

/goal finally comes to the Codex app 🥳

[boreal holly]

Oh dang, let us know in a few months when it finishes that goal!

[stiff leaf]

300 app ?! what kind of apps ? web apps or what ?

[plush trench]

is /goal even useful?

[lean lark]

My approach:

1. Plan out what I'm about to do, maybe discuss with Codex or ChatGPT.
2. Give Codex the task.
3. Look at what's been produced. Take responsibility. Understand.
4. Adjust manually and/or with Codex.
5. Commit - manually, intently.
6. Push

I'm not using Codex to review its own work.
I'm doing my own reviews, testing, evaluation of what's been done.
I'm being the human in the loop, every loop.
Maybe I just can't let go of the keyboard. Maybe I'm just getting old.

[chrome raven]

damn useful

[lean lark]

Peeps in public are talking about Codex mobile. Today is Thursday but I don't think we're gonna see any more goodies today, and IMO we won't see mobile for a while...

[boreal holly]

No harm in that! If you scroll back up to August/September I was doing the same thing. Not so much anymore.

[chrome raven]

how are you using it?

[cyan wing]

/goal "Get the mobile build working on iOS with haptics, touch, screen rotation, and core APIs.
@tawny moss use as needed to verify.
Don't stop until ti's done."

[chrome raven]

nice even with no function poping when typping /goal?

[chrome raven]

thanks seenn

[lean lark]

I'm definitely months behind you in my evolution of use with this product.

[tall notch]

Yes!!!

[cyan wing]

i did it from the TUI

[chrome raven]

nice, i enter /goal and it worked

[boreal holly]

@lean lark I kid you not, wrote my own code alongside Codex, reviewed every single line. The product has evolved a lot since then. I trust em a lot more now

[unique spade]

how s everyone doing?

[orchid plume]

at least Codex app is much faster and /goal is now fully integrated into the app. Just need to wait for the new Chrome feature to eventually be supported in the EU region, either that or use a VPN I guess

[unique spade]

what s the new chrome feature?

[vital hill]

it's playwright but with extra steps

[chrome raven]

yes vpn , u just need vpn again to enable if you restrt the app

[lean lark]

This morning I added high priority items to my task list to scope out backburner projects (which have been there for way too long) for Codex, collaboration or vibing, with minimal effort on my part ... just to see how the other side is living. So some of those new little businesses that might never see the light of day just got pushed to the front of the line.

[unique spade]

it s about the chrome plugin they released?

missed checking news a couple of days

[lean lark]

yeah, big news is about Chrome.

[orchid plume]

yeah will look into subscribing to one, thanks

[lean lark]

just roll your own, subscribing or money for a VPN? sheesh....

[boreal holly]

@chrome raven mentioned if you sign in without the VPN it wipes all the plugins. Perhaps if you VPN to download them, go into plugin metadata and rename the plugin it'll stay permanently 🤡

[chrome raven]

oh there is a way to make them stay permanet?

[unique spade]

and it isn t available yet in EU?
don t see any chrom extension in my codex lol

[chrome raven]

use windsurf vpn

[orchid plume]

smart idea, probably would work lol

[boreal holly]

I mean once you download the plugin, there's nothing preventing you from copying the plugin, and going into the metadata and renaming it!

[orchid plume]

sadly not, it's yet another gated feature. I assume it's gated due to similar reasons that Computer Use isn't available

[chrome raven]

nicee, i will check this out

[meager wadi]

Has anyone had issues with chrome not showing up in plugins? surely it isn't region restricted

[boreal holly]

The only bummer is to receive updates you gotta VPN again and download the update

[unique spade]

it's gated due to our beloved eu haha
i think for every such feature they need to do some paperwork and get eu approval to make it available

[orchid plume]

true words there, "our beloved eu" 😄

[meager wadi]

nvm: https://x.com/OpenAI/status/2052480804971028879

do any of you have it? / are willing to share it. I got the extension already

[vital hill]

USA baby

[small violet]

hey

did they use codex for canvas

[hollow current]

Codex extension in vscode has broken ui after recent update

[small violet]

its been broken. u cant see usage limits

[meager wadi]

huh?

[boreal holly]

Added a new design review script 😁 the workers provide the reference design image and the in-app screenshot, and the reviewer does a graded visual comparison. The ralph loop of UI design

[small violet]

canvas got hacked lol

[meager wadi]

yeah I have seen, what does codex have to do with it though

[small violet]

vibe coding

[meager wadi]

skill issue

[small violet]

can i see the outcome

[smoky parrot]

Has anyone gotten deactivated for sora use that violates TOS despite Sora not being used since shut down? I got 3 in a row and they shut down my whole OpenAI account today out of nowhere.

[lean lark]

not related to Codex.

[smoky parrot]

Oh sorry I thought I clicked on a diff sub sorry

[stable obsidian]

Wheres the difference in me prompting with /goal instead of not prompting with /goal

[boreal holly]

Before, and 66% complete (still working on it)

[lost drum]

before is where

I think on left

[boreal holly]

The workers still like to put stupid developer commentary in the UI but it's coming together quickly now

[lost drum]

wait but what you are building hmmm

[plush harbor]

I'm still doing my own css like a peasant. But I now have inconsistency with classes cross pages and need to get codex onto that

[lost drum]

peasant XD

[boreal holly]

SCSS FTW 🙃

[plush harbor]

I did give it a mockup from gemini as a vague indicator of layout

[plush harbor]

probably shuold, this site was meant to be a small thing not a horrid high traffic beast that's making me actually pay attention to hosting costs, caching, seo, not breaking it when people are looking at it so I need a proper deployment workflow, someone hassled me for a changelog etc

its seriously cross AI in this pass. Icons by gemini, icons cleaned up by chatgpt, icon markup by codex, actually putting the icons in by me

[lost drum]

medieval

[plush harbor]

and they don't line up cos they're all different sizes and that bugs me

[plush harbor]

its a game site so yes

[lost drum]

use canva then

crop it a little

[plush harbor]

eeeh. Css. Chuck it in columns

[cedar skiff]

5.5 xhigh is only 1.8 points higher than 5.3-Codex xhigh in SWE-Bench Pro. That is not a big code implementation gap.
They didn’t even put the SWE-Bench Pro result on the main graph. They just slipped it into the test stats, which I think is pretty telling. If the coding gains were impressive, they would have highlighted them. They didn’t, because the gain is small.My claim here is not that 5.5 is bad. I use 5.5 for almost everything else. My claim is that for medium complexity code implementation tasks, 5.3 high is cheaper in orchestration and competes at basically the same level.
For bulk iteration, writing integration tests, and unit tests, 5.3 costs me about half as much and gets the job done. The price difference can’t just be ignored with a claim of token efficiency. 5.5 would need to be massively more token efficient to make up for the pricing gap, and for this kind of work it just isn’t. I tried every angle several times and there is nothing that can change the true outcome. 5.5 with less than xhigh is shown to be weaker than expected. https://www.reddit.com/r/codex/comments/1t5ipjd/gpt55_xhigh_is_the_strongest_coding_agent_weve/

So yes, 5.3 high for medium complexity code tasks is cheaper and performs easily to 5.5 that it makes more sense to use it. There is no real way around it.

The public stats are there to be discovered and my in house experience matches them.

https://llm-stats.com/benchmarks/swe-bench-pro
https://openai.com/index/introducing-gpt-5-5
https://openai.com/index/introducing-gpt-5-3-codex
Anecdotal - but relevant:
https://www.reddit.com/r/codex/comments/1szb4bs/models_usage_comparison_table
https://www.reddit.com/r/codex/comments/1szzt2u/is_anyone_else_moving_back_to_gpt_53_codex_gpt_54/
http://reddit.com/r/codex/comments/1sz1d2g/i_think_the_53codex_version_still_performs_better/
I'll die on this hill.

[plush harbor]

will get codex onto that after coffee, got these icons on 3 different page types

[lost drum]

or tilt the box so it will look like users screen is off

[plush harbor]

its apparently going to take me 3 days to totally reskin and do my backlog for this site. Where was AI in my life when I started this site

[small violet]

this is beauiful

slop eliminated

[plush harbor]

I did the flag maker with chatgpt in web and that was painful. One function, one file, but painful. Got distracted cleaning up my thing that makes the flags

[small violet]

looks nice

have u used ai to make logos

[plush harbor]

logo for this site was made by a human but I redid it when I did these icons, so it matched

[lean lark]

Can you be more specific about what you did that was painful? There might be a better way.

[plush harbor]

backdrop is a bit buggier than drupal so stuff that works with drupal api doesn't always behave in backdrop, so you have to route around it with creative markup instead

had I known this site was going to get this much traffic I would not have used a framework like drupal

[vital hill]

oo desktop app update

[lean lark]

What was painful about it though?

[plush harbor]

just hte number of attempts to get things to even work at all

that one feature took far longer than its taken me to do this entire pass, it was a good couple days work

[lean lark]

I mean, was the content the difficult part or was working with ChatGPT difficult?

[plush harbor]

was getting backdrop and the pixel pushing code to work. I'm hitting backdrop issues with codex too

codex couldn't get the batch api to work so now I have a button I have to press 100 times. Bleh

"lets just make a site with backdrop, this should be fun" ... yeah no

[lean lark]

I'm understanding that backdrop is the issue, not well recognized by chatgpt/codex. It wasn't the chatgpt/codex tool mechanics that were the difficult part.

[plush harbor]

chatgpt likes to get enthusiastic that it can still do it. At least codex gives up earlier

[boreal holly]

SWE-Bench Pro is an OK benchmark for knowing if the model can write code, no doubt. There are other benchmarks such as Tau2 which you would think has nothing to do with coding, but it determines a model's ability to plan and execute a long series of tasks in as few steps as possible and stopping once complete. 5.5 absolutely smashes that test at all reasoning efforts. There are metrics not directly related to code, but related to the process of doing things correctly and efficiently, and 5.5 is beyond SOTA. It sets the bar impossibly high in areas that are just as important and often overlooked.

With that being said, the OOB experience for Codex + 5.5 is pretty bad. 5.3-Codex definitely appears to win in cost with OOB setup. You make back the cost by removing or streamlining responsibilities, distributing context across multiple agents, etc. Granted, if you optimize like this, 5.3-Codex benefits as well, but if you set the bar to "Optimize 5.5 so it's as efficient as OOB 5.3-Codex" you get better results

[lean lark]

I understand. Personally I dropped Drupal for Wordpress about 15 years ago, have never regretted it. YMMV

[plush harbor]

backdrop is more like drupal 7 so it was very fast to get a site up and working, but its the details that misbehave

I wonder if I can integrate this with patreon

probably not worth it, I only have 13 patrons

[boreal holly]

I know 😭 I used to have a single designer agent and have to painstakingly iterate on UI. Now these suckas can just steamroll the UI and I can do minor touchups when they're done

[meager bobcat]

hah I thought it was happening to just me

so silly

[plush harbor]

all my small sites are jekyll or php that's barely a notch above static (template includes). I'm running on old hardware so I cant' get a lot of modern tooling to even work, so I just roll with older stuff until I get rich or something

[small violet]

we're waiting for the script 😁 😭

im thinking of getting gpt pro

but i havent made any money yet

ill get pro when i make money

[plush harbor]

I make money, it psys for openai plus, gemini plus, my hosting, and all my other vaguely tech related subs including my internet. This is not much money though

[small violet]

is pro 200$ or 100

[plush harbor]

yes 😛

[small violet]

the hell does yes mean

😭

[plush harbor]

there's two pro plans lol

[small violet]

oh im saying gpt

[boreal holly]

already posted up https://github.com/robertmsale/.codex/blob/main/skills/design-worker/scripts/design-review

Not usable OOB. It relies on a carefully engineered sandbox thing

[plush harbor]

yeah there's a 100 and 200 plan

[random cloak]

has anyone attempted to clone palantir yet?

[lean lark]

Consider modifying your business models, from "low ticket, high volume" (which is often "low ticket, low volume") to "high ticket, low volume". That shifts the income to come from bigger clients. Then you can use the cash for your other passions.

[small violet]

what does psys mean

[small violet]

i told codex to make my site look like palantir

[plush harbor]

means I just got up and I can't type straight yet. Pays

[small violet]

i did not modification after

[plush harbor]

adsense + patreon is nice hobby money but you can't live off it

[small violet]

can it comfortably pay for 200 pro?

[plush harbor]

no.

[small violet]

damn

i think ill just rock the plus then

[random cloak]

i want to clone a minimal mvp to see if i can find criminals in california and cash in on gavins 50k rewards lol

i don't even know what palantir does tbh haha

[small violet]

go through their site

reddit

and ask ai

then make the mvp

of what u envision

[plush harbor]

oh right you can't use the T word here

adsense is not a great business model anyway, its just small change. Goes up and down with [insert mad politician here] mood unfortunately, very tied to global economy

[small violet]

the t WORD

[boreal holly]

Tword?

[small violet]

read the other statement

u can connect

[plush harbor]

automod.

[boreal holly]

Twitter

I'm confused. It's OK, don't mind me lol

[small violet]

bro

u cant be this dumb

[lean lark]

Generally speaking, I kinda SMH when peeps talk about going Pro. Do something solid with Plus. If you can't do with solid, Plus ain't gonna help much - it'll give you more rope and you'll be in the same place minus some cash. People seem to think more tokens will buy success. That's a serious fallacy of logic.

[small violet]

brain for code only

who has the power over the global economy

[plush harbor]

I get by with plus. I'll get onto a project that isn't powered by vibes and user love later