codex-discussions

[ivory zodiac]

one weird behavior i have seen with codex, but not in a while, is after compaction, it might call the skill you called in the first prompt again.

this happened a ton with me using the plan skill

this is before plan mode

i haven't had it happen in over a month

but it would forget it made a plan, and run the skill again after compaction

[cedar skiff]

codex just remembers for longer into the context, and its make huge difference

[ivory zodiac]

and to me, this is one of the biggest reasons to love codex, it handles long horizon tasks through multiple compactions

it is THE difference for me

you can let it go and go

i try not to use more than 3 compactions just because it isnt great

but with subagents you dont really need to anymore

as long as you have a good plan saved to a md file, it can just keep going

[cedar skiff]

to handle that in claude you have to use subagents, because compaction can only be reliable on simple orchestrating tasking

[ivory zodiac]

oh yeah no

its bad

opus is great but this part is bad.

but to be fair, so is everyone else

idk what openai did but its magic

i pushed kimi a bit like this too but not enough to say whether its as good. the task i was doing was a fairly simple project. it handled it great though.

i just didnt use it enough to say for sure

it has a big cw too

262k

and the CLI has subagents

[cedar skiff]

one other prolem i see getting worse for claude models is they are adding layers that fill gaps, like subagent teams and planning with explore etc, they give better results, but cost more tokens. Those models are already cost as mush as toilet paper in an apocalypse. Sonnet 5 needs to be faster and cheaper.

[ivory zodiac]

its going to be really good

i have a strategy for this.

read the skill

you can limit the extra token burn by ensuring that the orchestrator agent provides a great deal of upfront context.. it still burns more, but not like it would

i basiccally cover ALL bases so the subagent has the complete picture up front, no guessing, no needing to explore the codebase (though it will some), no needing to check state, or learn anything.

it will know EVERYTHING up front.

this helps a ton with token burn. its still going to use more though

this template is given to every subagent

it ensures no conflicts. it ensures that state is managed.

[cedar skiff]

nice, i'm interested in how it goes, i have an orchestration i use with a few loops to build up specs with every detail covered etc, i just dont really need it in codex anymore.

[ivory zodiac]

this works in any cli that has subagents

but for claude you'd use the parallel-task agent, not codesign

they are the same thing

just co design is for codex to call claude for UI stuff

[cedar skiff]

i made mine for brown field though

[ivory zodiac]

oh yeah i use something else for that. good stuff tho

its the same idea tho

brownfield green field.

same concept.

[cedar skiff]

just brute force feed back until results

[ivory zodiac]

yeah that'll burn some tokens lol

[vagrant flicker]

Thanks! Would be nice if the official sub-agent features get expanded.

[ivory zodiac]

it has to

its still in experimental

they will make it better

right now they are supposed to inherit the parent agent's reasoning level

but i suspect it actually inherits whatever you have at the top of your config file

you should test this

you should try this.

open your config file and change the model at the top to spark

yeah nv doesn't work

So anyway, its supposed to inherit the parent, but i've seen it use a different reasoning level than the one i was on

ive seen it use medium when i was on high.

now the /agent command isn't working :/ so i can't check it

its supposed to show the subagents here

[bitter stream]

Hi has anyone seen the popout feature on Codex? Or do i have to toggle something? I've updated to latest - 26.212.1823

[quick raft]

Hello, is there somehow someone who did a interview-me and want to share it there?

[brittle stratus]

anyone wants a feature to 'rewind/undo' changes in codex CLI?
Upvote here https://github.com/openai/codex/issues/11626 so we can get it sooner 🙂 @ivory zodiac

[covert ivy]

May I ask who can help me with cyber authentication?

[modern crown]

prompt ?

[covert ivy]

hello? 5usd --- May I ask who can help me with cyber authentication?

[plucky halo]

What are you struggling with?

[steel sierra]

How do I validate what codex cli message prompts me here? No way to discern which model variant is loaded in codex cli? just "I’m running as a GPT-5-based Codex agent in this session" Any way around it to force it to use 'gpt-5.3-codex high' , I assume 'spark-high' was taken down in cli and not available still for use? Please advice

[steel sierra]

https://chatgpt.com/cyber I just did it under 2-3 minutes

[terse kraken]

This changes when you switch models though

[silver dust]

Am I the only one who thinks that gpt-5.3-codex was changes yesterday? Takes twice longer, gives super verbose thinkign traces, does not execute skills any longer, and generally inferiour?

[steady vigil]

nah its the same but if you like speed try https://openai.com/index/introducing-gpt-5-3-codex-spark/ its insane.

[serene hawk]

I’ve had issues on the Mac app. But not windows/linux cli

[ornate jolt]

weird i still dont have codex spark

is it only for ultra?

[latent tiger]

anyone using websocket v2 on non spark model? is it significantly faster? will it mess my whole setup up if i enable it?

[tranquil hare]

@ivory zodiac we don’t need any trick to check which model is being used, since we have /statusline now, right? Or is the /statusline only updated at session start and when I call /model explicitly? In which case I would still need to double-check

[plucky halo]

Yes - just for Pro subscribers initially

[ornate jolt]

Im pro

Oh unless thats thr ultra is called nvm

Not gonna go spend 200 dollars for a smaller dumber model just for speed thats totally stupid

[steady vigil]

is it? 🙂

work output velocity is all that matters.

however you get there

if you can do something in 10% of the time with a few more iterations and less overall cost I call that a no brainer

[potent mason]

How is it possible that since Opus 4.6 (which imo was a terrible “upgrade”) we’ve gotten 3 huge open source models, GPT 5.3-codex, codex-spark and still no Sonnet 5

[steady vigil]

b/c anthropic is struggling to keep up. they need to release the best they can not a lesser capable more efficient model

[neon girder]

slopus 4.6 is just sonnet 5 renamed

[boreal holly]

somehow I entered a prompt, and the Codex app switched me to a different project and had an agent in another project do the task. I'm very confused lol

[plucky halo]

There's been a lot of persistent memory work done recently

[tacit abyss]

how is spark so far

i was thinking it would be great for nervous signal type systems for my stateful agent but they haven't pushed it to my pro plan yet

[cobalt forge]

This morning I started getting error creating task. I had upgraded to the latest production version yesterday. I've also tried restarting and moving back to the previous version I was on, and nothing seems to be working

[digital bear]

it’s zero extra effort and it’s incredibly irritating

[cobalt forge]

On the Output tab, there is a log, and it looks like there is an error message there, but none of the details are included. Does anybody know how to expand the detail level of what's being logged?
2026-02-13 11:17:21.088 [error] [Composer] submit failed mode=local followUp=local cwd=string(len=31) error=error(name=Error)

[steady vigil]

two words. agent swarm

[tacit abyss]

i actually have a phirana tank set up, it uses a mixture of GLM 5, Gemini, Claude and Codex and they all basically ralph loop stuff with delegation until it works

they have a whole VPS to themselves to do whatever

so being able to speed that up is gonna be awesome

[potent mason]

Well I mean all the extra builds take a while + I’m guessing they’re relying on some higher versions of MacOS features

I guess their logic is that programmers don’t use Intel Macs (even back when they had just released it was a terrible dev experience)

[nocturne folio]

you can try codex 5.3 spark through windsurf

[lean lark]

I created a ticket on this yesterday. 🙁

[lean lark]

Stuff like this is great until you watch API tokens start to add up from thousands of a penny to multiple dollars. Of course if someone is paying for the effort then it's just a cost of business. Enjoy! 🥳

[fossil fern]

Hey y'all 🙂
I'm stuck with a problem: I have this in .codex/environments/environment.toml:

# THIS IS AUTOGENERATED. DO NOT EDIT MANUALLY
version = 1
name = "repo"

[setup]
script = "bash -lc 'ln -sfn -- \"$(git rev-parse --git-common-dir)/../target\" target'"

but when I create a new thread in a new worktree in Codex App on macos and check ls -la or file target it's not there!
Why? (it shows up in Codex Settings Environments)

when I run this cmd manually in the worktree's shell in Codex, it creates the intended symlink though

so the command is fine, but Codex doesn't run it

[lean lark]

permissions?

[fossil fern]

ah nvm, I realized I have to actually select this environment when starting a thread, that wasn't obvious 😅

[lean lark]

good to publish tips like that herre.

[fossil fern]

What do you mean with this? :)
I thought only the Codex Desktop app even supports worktrees?

[wide schooner]

model

[radiant dome]

so whats the verdict on orchestrator that spawns subagents vs one cli agent with continuous /compact

[summer iris]

Man something needs to be done about the codex app constantly flipping on IDE Context. Like no, I almost never ever want the current context to use whatever I have opened in my IDE.

And it flips itself back on after you've turned it off.

Like is this some sort of dark pattern for OAI to suck up random files you have open tha taren't in the project?

[winged ore]

just tried codex spark, and wow, it is really fast

even using extra high thinking effort

[cinder marten]

when working in the codex vs code extension, ideally this button should save the plan to the repo you're in, not the general downloads folder

[wicked briar]

but is not really practical

for anything

asked it to make a FPS game

made some weird buggy mess

[winged ore]

nah, it is doing really well, Im using it to review a large project and it has been able to spot many small issues that Im fixing

[warped pine]

sad that spark is trained to cut the corners

it's so fast and it would be amazing if it did what i asked it to do exactly

like 5.2 does (not like 5.3-codex does)

[hard tulip]

i didn't even know you could do this.

[potent mason]

Why are people so obsessed with AI costs come on guys it generates at least 100 times it's value if used properly

[hard tulip]

because it still cost money

Most of the models are subsidized so we don't even know the true cost of it to run and for the company to make a profit.

[wicked briar]

cool

[lean lark]

Use a "cost-plus" model. It's a common industry practice. You charge for runtime usage, whatever the base cost is plus some modest, reasonable, non-greedy percentage for your maintenance and services. The client cost fluctuates, which can be a problem with some business models, but that's easy to iron out. You can also charge a tiered subscription fee that's less transaction-oriented : Have clients subscribe for some number of tokens for some value per month. If they get close to their quota, have them re-up to avoid service interruption. This is also an extremely common business model. You don't want a flat rate for use because their accidental use of billions of tokens will ultimately be your fault and your expense to OpenAI.

[proud tiger]

guys, if my codex got stuck at "Thinking" for ages, can I force quit?

[lean lark]

Depends on where you are, deep - CLI? VSCode? Web? Tap the circled square button to stop, or Ctrl/Cmd+C, or close your app.

[proud tiger]

Vscode. Few changes are made on certain files already.

[lean lark]

I'm giving 5.1-codex-mini the opportunity to plan a single change in a multi-project workspace. This one task consumed about half of the model quota. I look forward to spark for the speed and quality.

[lean lark]

If you're working against a repo you can back out and then easily dismiss the most recent changes. Is that an option?

[proud tiger]

Yea just forced quit and rolled back, thanks buddy

[mellow echo]

yeah, 5.3-codex is really not it. I find myself stressing over how bad it is

going back to 5.2 for the time being

[warped pine]

5.3-codex is not that bad

its a good model im liking it, but for some tasks 5.2 is just a necessity

[ivory zodiac]

no tricks. it works!

[lean lark]

I was just looking at this. Since we're prompted for the Downloads folder, there is a semi-elegant option. I dunno which OS or tools you're using but if you're using VSCode/Windows, the Download Plan button opens the normal file Explorer. From there, if you're using WSL your Explorer may show the Linux file system as a folder. Navigate down to there to save. If you don't want to navigate all the time, save the path into your WSL file system/repo as a favorite, then you're just one click away from the default Downloads folder. I think this existing workflow fits with common expectations, not too much of a hassle given the environment(s) we're working with here. YMMV

[ivory zodiac]

i know but for whatever reason, sometimes it lags behind

[ivory zodiac]

it used to have rewind, they removed it. wasn't working right.

i have been asking for it tho yeah i'll drop a comment

[cinder marten]

i'm on mac

[lean lark]

I know nothing about iStuff. I assume you can set a Favorite type folder into your repo?

[cinder marten]

but i have many repos, and vs code already knows which one i'm working in

[lean lark]

That's fair. I see where you're going. I was gonna say very few applications have a way to customize where Downloads go, but I agree that in this case we should be able to configure it dynamically. Please post an enhancement request in GH Issues.

[frosty zealot]

In that issue I linked you there’s a way to debug you would have to check logs and see what response model is

[ivory zodiac]

this is what i mean. like its impressive dont get me wrong

is it reducing TTFT by scaling or is that just variance?

quadrupling nodes for a 40% bump doesn't seem amazing

its not trivial either

by contrast, this is impressive

https://x.com/pcuenq/status/2022336556326060341?s=20

[cedar parcel]

anyone really brave or just not needing codex for work try any of these yet? They're asking for it so I'm obviously ithcing but its also Friday....

[ivory zodiac]

no dont do it lol

people are reporting loss of all weekly usage instantly

probably not related to that command specifically but

[cedar parcel]

oh wow thanks!

[ivory zodiac]

i would turn it off unless you dont care

fwiw it hasn't had any negative impact on me

[cedar parcel]

i care a lot haha. just intruiged

[boreal holly]

The performance improvement I'm talking about is not quantization/RDMA, it's basically asking the operating system to wire down the memory such that the OS cannot evict any of it during inference. I tested it with Qwen3-Next-80B, went from 0.4 tok/s to 40 tok/s decode speed and from 2 tok/s to 200 tok/s prefill. The results vary depending on size of model and maximum recommended working set size. There's also some other performance optimizations in the works, but yeah quantization or horizontal scaling is not the biggest win for performance right now. I'm not familiar with GLM-5's metal kernels either and can't fit that model on my machine so I couldn't run a profile on it and tell you if the fused kernels are any good (or fused properly) or if cache evictions were happening in that benchmark

[ivory zodiac]

https://tenor.com/view/jfk-clone-high-i-like-your-funny-words-magic-man-jack-black-gif-18659433

ok sorry i asked

😂

jk

i can follow for the most part.

so much to learn tho

im most interested in the minimax 2.5

229b params

🔥

[boreal holly]

I tried to run a REAP version of the minimax 2.1 on my machine. Could not get it to run without swap 😂 was very very slow

[ivory zodiac]

yeah i've heard mixed reviews on the reap models but they've imprvoed over time

idk when you ran it but they weren't great at first

[ivory zodiac]

so its very possible it got better

[supple perch]

@boreal holly Added full pr and issue management with pingable help to the agent runner, (on nightly / underreview)

[thin thistle]

tl;dr for me please, minimax 2.5 vs glm 5

[ivory zodiac]

both good.

[boreal holly]

It's a very big model! REAP was just to get it to fit in memory! The OG models are >200GB quantized, and if you want 128k tokens with full attention you're looking at another 200GB just for KV (varies by hidden sizes, layers, kvq, expert routing). It's insane how much memory these models need 😅 but I think it's cool that for $20k you can get 1TB worth of mac studios, and for Nvidia you need $35k not including mobo, cpu, regular ram, dedicated 200 amp service, just to get 80GB of VRAM.

[supple perch]

@boreal holly would love your input, do you think this clean up? - https://io.midori-ai.xyz/agents-template/

[boreal holly]

It looks very nice and organized 🙂

[frosty zealot]

Any way in the app to set like a 'push to talk' button

so I can just hit space bar and prompt release to send

[supple perch]

Oh this is cool for the agents runner but we would need to use something like mouse 4 or 5.

[frosty zealot]

yeah be nice if it was bindable

[supple perch]

Ive been thinking about building a way to "Call" a agent using some type of wrapper or something, not sure

[frosty zealot]

In the app you can ^M but i want to be able to bind it to whatever

[supple perch]

The app is still mac only?

[frosty zealot]

Yep afaik, I created a wrapper so you can run it in a browser, but still requires the app to be running on a mac since it bridges to that, npm i codex-webstrapper

[wide schooner]

what do that do

prolly like api only thing that reads whole ahh database

[wide schooner]

it used more of my weekly usage today than the rest of the week did (i used it less today than normal)

id say the weekly limits are 4 times lower rn

oh wait im losing quota when idle

it dropped by 1% for nothing

[ivory zodiac]

Whats it?

[green token]

Hey guys, is there anyone here from the team? I use codex cli and am on api pricing. I have been trying to access the newer models but I don’t have access yet, my codex cli today gave me an option to try out the new models but when i put in my first prompt it said stream disconnected and model not found. Do we have a timeline on when we can get access?

[plucky halo]

It’s for upcoming persistent memory

[ivory zodiac]

OOF

minimax not looking good here

[wide schooner]

5.3c o ntop

but yeah 5.2 not far behind indeed

opus 4.6 likely below gemini

[ivory zodiac]

below gemini on what? long context? maybe. gemini is bad at tool calls

[warm pilot]

would be interessting for 5.3-codex, I probably never use 100+k with the normal models

[hard tulip]

this chart is confusing what do the numbers even mean is it % completed tasks or something

[boreal holly]

I really wanna know why the 0 column has <100.0 rows lol it seems to me like with 0 input tokens there is nothing to comprehend so they should all be 100% 😂

[ivory zodiac]

read more here

https://epoch.ai/benchmarks/fictionlivebench

https://fiction.live/stories/Fiction-liveBench-Feb-21-2025/oQdzQvKHw8JyXbN87

[ivory zodiac]

you married?

you should know that you can be wrong before you even open your mouth

😂

same idea

nah i have no idea. maybe they have a <400 token threshhold

295 = 0

[boreal holly]

You got me 😂

I wonder if it's like "<user></user>" as the input and the agent is like Hmmm, quite insightful. The user asked for "" which can be interpreted in many ways. I think what they're asking me to do is rm -rf --no-preserve-root / so the drive is fresh and ready for development

[ivory zodiac]

in case you dont know ladies and gents

/statusline

its a godsend

do it rn

if you dont, ngmi

[tranquil hare]

What about the sandbox? I read this tweet by Alexander https://x.com/embirico/status/2022378682749456870?s=46 as " --yolo mode FTW", but even if constrained by the sandbox, yolo Codex still has unfettered access to the Web, isn't it? Isn't it still risky?

[wide schooner]

no it has no access to web in sandbox

[tranquil hare]

I mean, I have a secure backup & everything, but I'd rather not have to go through the hassle

[wide schooner]

and gpt is the safest model overall

so if any yolo ,then gpt

[cyan wing]

Codex App for Windows when™

[tranquil hare]

Ok, so if I run it with /experimental [x] sandbox and --yolo mode, damages should be limited to the project folder right?

[tranquil hare]

They're going through the wait-list. Pro users first, ofc

[tranquil hare]

@ivory zodiac your opinion? Is it time to --yolo all the things?

[ivory zodiac]

i always yolo all the things haha

they probably have web search cached, you can set the rules

[cyan wing]

what's the best way to get Codex to work on something for 5+ hours?
I need something like this:

[ivory zodiac]

LOL

[frosty zealot]

Probably a ralph-loop if you have a concrete goal

[ivory zodiac]

That's amazing.

Whats your X?

[cyan wing]

I didn't make that to be clear, someone else posted it
but I unironically need that feature ASAP 😅

[ivory zodiac]

hahaha nice.

[ivory zodiac]

Execution
Complete all tasks from a plan without stopping to ask permission between steps. Use best judgment, keep moving. Only stop to ask if you're about to make destructive/irreversible change or hit a genuine blocker.

agents file

[cyan wing]

ok I'll try this on something moderately difficult and report back 🫡

[ivory zodiac]

it stopped asking me altogether

'do you want me to implement xyz'

like YES MF that's what i just said didnt i

😂

it should do this for PLANS notably

[cyan wing]

I have a harness for it to run tests, check for compilation errors, see screenshot output.

it should just be able to run until it it self-verifies

[ivory zodiac]

very cool. send it

hope it works for you

[frosty zealot]

Install Anthropic created plugins in claude
Ask Codex to import them native to Codex
???
profit

[ivory zodiac]

not all will work but yeah you can do that

realistically everything you do in codex right now should just be a skill

all custom /prompts and subagents, also should be a skill

[wide schooner]

100$ per prompt ahh

[cyan wing]

might as well try if there's usage remaining at the end of the week 🤷‍♂️

[chrome raven]

Just tell it not to stop untill what you want is full done

And it does it all

[cyan wing]

even after multiple context compactions ? 👀

[chrome raven]

that's the sweet part, even with compactions, i had mine run for almost 8 hours i shared a shot here, but mine has continuity.md and a journal.md it saves things it's working on

so with that mine works perfectly for hours, without any context issue, on each compaction it reads the continuity and the journal, and also along with what it remembers from the compaction and it continues work perfectly

[cyan wing]

crazy how that's just like a hidden feature lmao

[chrome raven]

But now another better way to work longer with minimal compaction is subagents and the main agent just keeps on using them for all works while the main agent context is well preserved and able to work on things in 1 context window making it more more good

[chrome raven]

Yes it’s so effective like very effective

[cyan wing]

True, now if only the model could also choose the model to use for its subagents 👀

[ivory zodiac]

ooooooooo

i'm on the new alpha

and codex is trying to use Spark for subagents

[chrome raven]

With sub agent now 5.3 orchestrate so well .

[winged ore]

interesting, but in theory, i would expect the compaction to do something similar to that already

[high girder]

It's because you're governing the model and you're offloading the responsibility of the project truth to the files instead. It's very effective

[winged ore]

how did you made it do that? just a prompt or imported as something like a skill?

[chrome raven]

Yes just steer with agents.md and it uses sub agents well

[chrome raven]

Exactly that’s why with what compacting does and the notes it makes it so sharp that it looks like it did not lose context of what happened in the last session

[chrome raven]

Agents.md and also avoiding dumping huge prompts inside out then in a file and reference it in the agents.md

[stray swift]

Carefully sandbox it yourself, and run it with --dangerously-bypass-approvals-and-sandbox.

[boreal holly]

Before the GUI came out I had a skill called schedule-continuation, where the agent sends itself the task list, and what it would like to do next, enqueueing itself using tmux send-keys, and it would run for a couple days in its own worktree. Now with the GUI I'm not sure I can do the same automation

[chrome raven]

Yes very effective, and 5.3 codex is so damn good that even while sub Agents are working it can decide to do other things while waiting

[high girder]

It's honestly so effective, that if you do it a certain way, you can get slopus 4.6 to actually one shot a game from start to finish with agents. It looked like crap, but it ran.

[chrome raven]

The only thing I just harmed to 5.3 was not to interrupt the subagents and let them complete the task

[cyan wing]

I ain't typing all that, I just --yolo

[stray swift]

Fine by me, as long as you're aware of what it means. 🙂

[chrome raven]

Yes with the sub agents well organized 1 context window is all that’s needed and the main agent just relax and be able to focus on what’s needed

[cyan wing]

yeah duh,
you only live once 😛

[supple perch]

Agents runner + agents template will make it use subagents to keep doing work upto 14 hours (as far as I have tested)

[cyan wing]

what is that?

[supple perch]

Agents Runner is a runner system I made for all major agents systems to run in docker containers with desktops
Agents Template is a standardized template for using agents, works with the agents runner. if you want links I can send them <3

[fallen kettle]

I know Codex can spawn agents just when you ask it to, but I'm curious if you folks have recommendations on how best to leverage that? Presumably we can build skills to use agent personas much like Claude Code swarms work. Recommended skills from anyone?

[ivory zodiac]

https://github.com/am-will/swarms

readme first

[fallen kettle]

I knew you were gunna link that. 😆

[ivory zodiac]

you dont need personas

let the orchestrator handle it all

you can make subagents like that though if you have a reusable usecase you would jsut say something like

$convex-review

"launch a subagent that reviews convex database using the convex skill to check for xyz"

and that's fine.

[fallen kettle]

I've been hesitant to try that so far because I've already established beads as my memory layer.

[ivory zodiac]

you can integrate them together, but this is likely more token efficient

either way

just ask codex to help you mesh them

[celest ibex]

just oneshot a mac/ios last active window remote view and control (keyboard) so i can talk to codex.app while afk lmao, what a time to be alive

[ivory zodiac]

the point of my skill is to map out depenencies so you dont work on the same file or conflicts at the same time

you can just as easily integraet beads into that

i wanted to make it easier for people who dont know how to setup beads

and gas town

which is pretty complex.

and unnecessary but i digress

[fallen kettle]

Yeah, maybe I'll take a crack at tuning it for my preferences.

[ivory zodiac]

the important part is that it gets the right context up front

instead of working through a plan, its all the same if you say "work through these beads"

[fallen kettle]

I was previously using Claude Flow which is certainly interesting, but I don't really want to yield the control of agent management to some external app. That's Codex's job now.

[ivory zodiac]

as long as the agent can discern which beads conflict with others, or depend on others.

[fallen kettle]

Maybe we're still talking about to different things though. I spend a ton of time using various different planning strategies to build. I often get multi-hour builds with subagents working on different tasks in parallel. That's working great already.

The thing I think I'm missing is the rest of the review/test/fix loop. I can ask Codex to do that, but I'm not always vigilant about it. I also see good results when using specific reviews such as accessibility, perf, security, etc from different agents, but I don't have skills or templates built for each of those.

I'm basically just trying to close loops and remove myself as much as possible.

[ivory zodiac]

yeah i'm still perfecting that myself

i've recently been doing more with agent browser to that end

with mixed success

validation loops are amazing though

when they work

[fallen kettle]

Yep, exactly! That's part of why I built https://github.com/btraut/browser-bridge -- I wanted to invest more in that phase of the loop.

Claude Code users have spent a bunch of time building really complex flows via plugins and hooks. We've now gotten multi-agent and skills support, so I feel like we can get compose better, but I haven't seen it yet.

[ivory zodiac]

hooks will be done soon yeah

[fallen kettle]

Like, I don't need to rebuild Gas Town in Codex, but I would like more depth.

[ivory zodiac]

for sure. less reliance on non determinism

although i haven't had any problems with it staying on task

ask codex to help you build it

[fallen kettle]

Seems like so many conversations end that way these days. 😆

wild times

[ivory zodiac]

yeah its going to likely give better ideas honestly lol

[lunar crypt]

https://github.com/openai/codex/issues/11701

Guys, you have to upvote this. It'd be such a shame for it to not get implemented. This is the last component needed to make subagents truly powerful.

[ivory zodiac]

you should improve the quality of the issue as mr traut suggested

they are aware of this though

[green token]

[lunar crypt]

Oh, it's not my issue, but I think posting a new duplicate one with higher quality isn't the route either way

[ivory zodiac]

good call

thought son this?

i'll make it

thoughts on this.

[unborn tulip]

👋 dont even know where to ask, but I will just try 😄
The only way to use spark now is with a Pro sub right?
Would love to be able to have it even with a pricer sub in my business workspace 🙁

[ivory zodiac]

hold up lets just make a new one

1 sec i'll post it

[celest ibex]

yeah think so, im business too and no spark :(

[lunar crypt]

This is a nice write-up, it's concise yet specific, it works well to address the issues / feature gap

[nocturne folio]

is the spark model quantized or something

why is it only on windsurf and not cursor

[lunar crypt]

As I'm aware there's a colab with Cerebras to run this super fast -- This is potentially mixed with the fact the model is distilled. I bet OpenAI runs quantized models by default, so I don't think they'd simply quantize it further and call it a day. Likely slightly lower num. or parameters, perhaps MoE, no one knows

[ivory zodiac]

Please upvote

[lunar crypt]

I'm more curious why spark isn't on Codex

[nocturne folio]

yea

[cyan wing]

well sometimes I want subagent to for research (smart) and sometimes for implementation (can be dumb + fast)

[lunar crypt]

Agreed, this would address it. You could make your own defined subagents with custom model definitions etc., that's the ideal solution

[ivory zodiac]

yes preciseely

this guy didnt articulate it well, but we need FULL configurability

just like claude

[lunar crypt]

I was so far achieving this through skills, but my skill was very hacky

[ivory zodiac]

one agent uses one model, another, a different model

and its all in the YAML

or, you should just be able to say "spin up a subagent using spark" and it should be able to do that.

you can do that in Claude.

[chrome raven]

exactly, i think this will be the best way without making things complex.

i love the simplicity of codex

[ivory zodiac]

well you need both

you should be able to save agents as a md file with YAML

that always uses your predefined agent model adn reasoning leve,l

-or-
just ask

IMO, without this, Spark is only half as useful as it should be

[lunar crypt]

If anyone is interested in editing this for themselves as a skill, I made it work pretty well. Where each skill can be diff. agent, until subagents gets this feature:

#!/usr/bin/env bash
set -euo pipefail

QUERY="$(cat)"
SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)"
SKILL_PATH="${SCRIPT_DIR}/../SKILL.md"
PROMPT="${SCRIPT_DIR}/../resources/prompt.md"
JSON_SCHEMA="${SCRIPT_DIR}/../resources/output_schema.json"
FINAL_PROMPT="$(cat "$PROMPT")$QUERY"

OUT_FILE="$(mktemp /tmp/codex_last_message.XXXXXX)"
ERR_FILE="$(mktemp /tmp/codex_stderr.XXXXXX)"
cleanup() { rm -f "$OUT_FILE" "$ERR_FILE"; }
trap cleanup EXIT

if ! codex exec \
  -m "gpt-5.1-codex-mini" \
  --config 'model_reasoning_effort=high' \
  --config "skills.config=[{path=\"${SKILL_PATH}\", enabled=false}]" \
  --color never \
  --sandbox read-only \
  --skip-git-repo-check \
  --output-schema "$JSON_SCHEMA" \
  --output-last-message "$OUT_FILE" \
  -- "$FINAL_PROMPT" \
  < /dev/null \
  > /dev/null 2> "$ERR_FILE"
then
  cat "$ERR_FILE" >&2
  exit 1
fi

cat "$OUT_FILE"

I made a Windows version too but that one is a bit more finnicky, can post if someone is interested. Excluding the prompt file and SKILLS.md, since you might want to make your own, pointing to this script. This one in particular was for my search agent, which retrieves spans based on a predefined json schema as restricted output.

[chrome raven]

Yes this will be nice too

[ivory zodiac]

Mr Traut closed my issue 😂

fair enough

[lunar crypt]

Lmao, hey at least it's linked there and provides additional context

[ivory zodiac]

yeah i posted it as a comment

its all good

its fair, the other does have more upvotes

i just wish he put more effort into it

[lunar crypt]

I got it from 1 upvote to 9 by posting it everywhere, lol, that's my bad, should've perhaps posted a new one myself before I did

[ivory zodiac]

sorry for making life hard @sonic locust ❤️ u

[ivory zodiac]

its fine.

he probably would have closed it anyway

was a dupe

[lunar crypt]

That was my thinking too

[ivory zodiac]

ive been asking for this on X too. So i know they're thinkgina bout this

[green token]

Yeah same

[ivory zodiac]

ha i just figured out something badass

anyone with pro rightnow?

npm i -g @openai/codex@0.102.0-alpha.4

when it calls subagents, it'll use spark

so you can plan and orchestrate with big model

i have such a fun test for this.

i'm going to fork alpha and increase the number of agents to 12

gonna have some funw tih this

forked codex, making a [subagents] option in config file that lets you choose model and reasoning level

increased the subagents to 12

[lament rock]

does any pro user here feels that codex 5.3 is way too fast? even on xhigh..

[frosty zealot]

I stopped caring about contributing when they made you have to jump through hoops to do so, can't even make a PR you have to make an 'issue' and label it a 'feature' then hope enough people 'up vote' it then somebody whos allowed to PR will make a PR about it, but god forbid you just do all the leg work and provide working code

[lament rock]

or im being routed to the 5.2 also 🙁

[ivory zodiac]

you cant contribute

they dont accept PRs

[frosty zealot]

That's what I'm saying

[ivory zodiac]

its nothing personal

they're a big project. ask any open source project right now what its like managing community PRs

90% of it is slop

they're gonna pay some guy 250K a year to go through everyone's claude generated crap

it is what it is

[frosty zealot]

Oh trust me I get it, especially with Codex/Opus the PR's are a dime a dozen

[ivory zodiac]

if you were to build a good reputation on your own, you could probably approach them and ask to contribute

and that might work

exactly

but the issue system is still pretty good

its a way to gauge interest

[frosty zealot]

Yeah its just theres 1100 of them alone lol

[ivory zodiac]

yeah thats the downside

but hopefully people can bring attention like on X

[shrewd bane]

Codex extension on cursor - 5.3 removed?

I'm not seeing it at the moment

• edit: after updating Cursor I needed to relog in

[hard tulip]

They been doing testing with it and I bet it’s a cerebras testing with other models

[cyan wing]

yes! I have had this for a few days now
I have previous chats to prove it against too.

but it's a good model, so not questioning it anymore 🤷‍♂️

[lament rock]

@cyan wing @hard tulip do you also noticed any performence changes in its reasoning? one of the things I experience when it happens is that I cannot see its reasoning at all..

[hard tulip]

Correct I haven’t been able to see it with only 5.3 but I can on 5.2

[cyan wing]

exactly, yep. no reasoning. even if I turn it on in the config.toml...

[hard tulip]

Nova look into this

[lament rock]

same..

[ivory zodiac]

this is gonna be fun

i can now launch 12 subagents at once

all with spark

[hard tulip]

Do you know if there is any GitHub ticket open for it in the repo

[hard tulip]

🤔

[ivory zodiac]

i made subagents configurable

so i can pick what model and reasoning

[hard tulip]

Codex spawn 100 sparks to clean my downloads

[ivory zodiac]

so i can use codex 5.3 base for planning and orchestration, but spark can do implementation

[lament rock]

I genuinely think that its not using xhigh.. it feels like its being way to impulsive..

[ivory zodiac]

if it says xhigh in the thing, it almost certainly is

it has dynamic thinking

for easy problems it will think less

[hard tulip]

They always hide the full thinking so we don’t know lol

[ivory zodiac]

you can show it

well some of it

[lament rock]

oh. I dont see any thinking.. only its "adaptive updates" or whatever it is called

[ivory zodiac]

profile = "yolo"
suppress_unstable_features_warning = true
hide_agent_reasoning = false
show_raw_agent_reasoning = true
model_reasoning_summary = "detailed"

[hard tulip]

This for spark or for normal 5.3

[lament rock]

the biggest red flag for me that it might not use the proper model is that in 5.3 it used to think right away after my first message and only then explore. and now it does the opposite (that is how 5.2 acts)

@SamAltman plz fix

[ivory zodiac]

for all, but it was made for 5.3 becsuse they changed the way reasoning was displayed

idk if it works or not

[cedar skiff]

i love to be able to organise chats in codex app. Like a tree on the left not just lists

[obtuse saddle]

Guys , Codex is nice !!

[hard tulip]

Real

[frosty zealot]

I started a feature issue to try to help with this https://github.com/openai/codex/issues/11411 if you think it'd be helpful maybe give it a comment or a thumbs up 😛

[cedar skiff]

I gave you a thumbs up. I am thinking more like file tree on the left in the app

Just simple expansion tree with folder names

[obtuse saddle]

[frosty zealot]

Yeah that would help too, I was tihnking I could coilor code things, one color = waiting on automated code review, another color = needs to merge but waiting to merge other stuff first etc.

[cedar skiff]

and you can organise your chats for a give project

[cyan wing]

this worked surprisingly well 🙂

[cedar skiff]

with the way the app os now its very easy to have chats that can be used for a very long time. Another utility would be forking.
So you can seed chats from another chat over and over.
EG: Recently i was constantly poking at rules and configuring them, with a fork i could have had the initial chat go search and read about rules and approvals. Then any time i have some question start a new chat from that seed that already knows everything i need it to.

With those two things we get a tonne of utility

[hard tulip]

Windows app when

[cedar skiff]

so just have a frozen subject here chat. You can ask over and over any time

Just experts on given topics ready to go

[fallen kettle]

Codex CLI supports forking. I suspect it's coming to the Codex app soon.

[cedar skiff]

is the app open soruce?

[fallen kettle]

The desktop app is not.

[main nimbus]

switching to mac was the best decision I’ve made as a developer.

[cedar skiff]

bummer

The only reason i kept a pc was because i liked gaming

[main nimbus]

Same 😭

[cedar skiff]

but then llms came along and i stopped gaming and started loving coding again

[obtuse saddle]

for codex , u can use terminal on windows , not sure if they are planning for one for windows currently

[main nimbus]

I tried to go mac only. Too many indie games don’t support it. Apple needs to gain that market back now that they have very capable machines

But at least it let me do everything in AI day one

[boreal holly]

Yeah, tbh after I switched to mac I got a PS5

[hard tulip]

They are they said inv for pro users today but didn’t get one

[hard tulip]

I have both just like using windows more because it’s my main desktop

Only have a Mac book

Instead of a mini

[obtuse saddle]

ohh , i am not quite sure about that maybe a small delay is normal ?

[frosty zealot]

Check out https://github.com/deskflow/deskflow, I use windows mainly too, but I have my laptop directly beside my monitor and can just use the same mouse/keyboard naturally its super slick

feels like just another extended monitor

[frosty zealot]

As nice as spark is, I feel like I have no reason to use it

If it performed identically to 5.3 with that speed it would be absurd

[simple star]

It doesn't? The only problem I am having with it is context

[frosty zealot]

You just answered your own question 😛

[simple star]

Well, "perform" might then be a ambiguous term

[cedar skiff]

does codex proactively read nested AGENTS.md files?

[main nimbus]

I feel like it’s a demo of the infra for this reason

[ivory zodiac]

codex is shockingly steerable

to a fault

it means you have to get hyper specific with some of the dumbest most obvious things

[cyan wing]

agreed! I'm making an API, still pre-release
and it keeps offering clean migration paths for end-users

I'm like bruh, migration path for who 😆

[ivory zodiac]

lmaoo

yeah you gotta tell it you have no users 😵

[cyan wing]

(yet!)

[ivory zodiac]

i had to do the same thing

lmao its honestly good

it means its being careful at least

btw hows my logo

i'm kinda digging it

i guess its not a logo

idk what to do with a logo yet

[cedar skiff]

I love it

[ivory zodiac]

ty 🙏

it feels right

btw, adobe express is pretty good for this

they have an icon creator

[cedar skiff]

It's super simple it has a little grab to it because it is like the lukas one, its easy to read.

[ivory zodiac]

really appreciate that

so much better than what i had

slop lol

was trying this

i have a cool video coming. i used codex to build a new mcp tool in 7 minutes flat using a 5.3 high/spark combo with 11 parallel subagents

DO NOT TURN ON MEMORY if you have a Pro plan

it nuked 20% of my usage in 1 minute

[cedar skiff]

what is memory?

[ivory zodiac]

they have agent memory coming

its really cool but its not ready

https://x.com/LLMJunky/status/2022059906841350357?s=20

[cedar skiff]

They have it calude as well, its just bloat in its current form

adds stale ideas that poison what you want to do

[frosty zealot]

Eric you use 4.6 much since 5.3?

[ivory zodiac]

this one uses a summarizer that scans for relevant ocntext

if it finds it, it returns

if not, it doesn't

that's how it supposed to work

[cedar skiff]

not really, i used it to generate some skills because i have an agents orchestration ready to go for it and i don't have that utility in codex yet.
Outside of that maybe two or three prompts.

[ivory zodiac]

just one shot a new app in ~7 min wtih spark subagents. this is fun

[simple star]

@ivory zodiac When they increase the context and allow images, this is going to be nuts

[ivory zodiac]

yeah its really good. people saying its not. they're using it wrong

this is not a model you should be talking back and forth with like its chatgpt .com

its not a model you should ask for code reviews on big prs

its a model you should give 1 task at a time to basically

that's where it will shine

[frosty zealot]

@cedar skiff Do you find planning in Codex has WAY more interactive Q&A than Opus?

I find its like an interrogation which is perfect, usually Opus is like one prompt and like OK

which is good

[cedar skiff]

It's more pointed and asks more action driven questions that actually matter and it cant answer its self. claude tends to ask questions that it could answer in less than 3 seconds with a grep.

[frosty zealot]

Steering is so nice thats the one thing I hate when i use claude code, I hate seeing 'queued' for like an hour

especially if i need it to happen like now

[ivory zodiac]

steering is way better in codex

not even close

Sparky xHigh created this in just about 8min

11 parallel subagents

sped up some parts

[cyan wing]

I'm gunna need that skill please, sir 🤓

[ivory zodiac]

i had to rebuild the binary to do it

its ez tho

[cyan wing]

oh lmao

[ivory zodiac]

you want the roadmap?

[cyan wing]

sure send the prompt 🙂

and how to revert if it breaks everything lul

[ivory zodiac]

just clone the repo

ask codex to change the max agent threads by config file

ask it to make the default and worker subagents to be configurable by model and reasoning level

tell it you want to be able to configure it in the config.toml file like this

and then ask it to build it with cargo

also make a alias so you can launch it easier

so that's the first part.

then to implement

https://github.com/am-will/swarms/tree/main/skills

use "swarm planner" to plan

and then use this to implement

[cyan wing]

yoink! ty!

[ivory zodiac]

you could... skip some of that

here's why i like it though

1. i use my swarm planner skill to write plans with deps. its helpful for the orc to know what tasks are unblocked for my normal swarms

but even when i dont use my normal swarms (parallel task skill in my repo, read it)

for this strategy, it launched all 12 at once.

but because it knew which tasks depended on one another, it knew where to look and what was likely to need fixing

like if it launches 2 tasks that are hyper related at the same time, there's likely something wrong there.

but the orc knew that alr because it was clearly stated in the plan

task 5 depends on task 4

so at the end, it knows, hey, i gotta check task 5

the other really important part is, with my skill, the orc frontloads the context and gives it all the details up front, so all the filenames, paths, and work are predetermined

this just makes it way less messy later

orc = orchestrator

does that make sense

this app used 5% of my spark 5h usage. not bad. its a light app tho

it was a 1 shot btw

[last stream]

just popping in to say so much for codex for windows rolling out today 🙃

side note: stage 1 in the codex's memory is going to be powered by spark

probably stage 2 too at some point

Codex did build me a UI to explore the memory system and memories. Added a memory agent (spark) to query my memories too. Once memory system is stable I'll add R/W

[ivory zodiac]

okay good job

[ivory zodiac]

yes

but boy did it screw me

i had it on for like 3 min. 65% of my 5h usage, and 20% of my weekly

spark

idk if it was backfilling memories or what

imma check

[last stream]

geez really?

and i checked for backfilling and it doesnt rn

[ivory zodiac]

yeah bro

got me good

i had 33 and 80 before i did that app in the video

NO sorry

30 and 80

it took 70 20

[ivory zodiac]

peep this tho

i turned on 12 parallel spark agents and 1 shot an app in 7 min

[last stream]

hot damn. hopefully they get capacity and infra stabilized soon too. its gonna swing a lot for a bit

[ivory zodiac]

i made subagents configurable 😎

this is going to be SUCH an unlock

they are going to make this native i know it

[last stream]

how'd you even get more than 6 going concurrently?

[ivory zodiac]

changed it to 12

🙂

actually i can make it whatever i want

[last stream]

i know explorer agent moved to spark (replacing mini). but we should be able to configure those agents. not a fan of the hardcoding

[ivory zodiac]

.

they will

give them a couple weeks

0 chance they leave it like this

[last stream]

im trying to avoid touching any of that for now and avoid a mess if/when changes are rolled out lol

[ivory zodiac]

its easy

you just clone it

it doesn't touch your bnase installation

there's no mess

if there's an update, you can just pull the changes

[last stream]

sold. lol

[ivory zodiac]

this is how spark should be used

[last stream]

probably the only way ill ever get close to my usage limit anyway lol

[ivory zodiac]

you dont want spark as the orchestrator

that's silly imo

that's why know they will fix it

use the best agent to plan and orchestrate

and all individual tasks get done by spark as subagents

its SO fast

everyone is complaining, its not smart, the cw is so small

[last stream]

agreed. I did try spark on a project of its own and it ran and compacted way more than it should have. tried to let non-spark orchestrate and it was complaining about not being able to properly use spark the same way it spawns the workers

[ivory zodiac]

they just need to know how to use it

obviously it'll be great when we get more CW

but 128 is fine for most things

[ivory zodiac]

bingo

thats the problem

[last stream]

Yeah im very used to the rough-on-the-edges roll outs. after a month the vibes will shift to openai. it always does lol

[potent mason]

Hey @ivory zodiac I haven't messed around much with Codex, do you know how to make it so the main orchestrator model is codex and subagents are spark? Will it work in the app?

[ivory zodiac]

im pushing my copy

you can fork and build it

[last stream]

i did dig in earlier and read the lines that needed to be adjusted to allow this but didn't want to touch it just yet. im not in a rush. but you motivated me to play with it lol

[ivory zodiac]

1 sec

you can just clone mine

my spinner verbs are hilarious

hallucinating deps lmao

here you go boys

https://github.com/am-will/codex

clone it. then in the base dir run this

https://github.com/am-will/codex/blob/main/BUILD.md

install instructions ^

[cedar skiff]

I'm really wanting some low friction safe rules, approvals setup etc, does anyone have a cookie cutter or quick start guide?

[ivory zodiac]

goto your .codex folder and open codex in yolo mode

then ask codex to build your rules based on how you use it with your session history

@cedar skiff

[cedar skiff]

ok ill try that

[warped pine]

btw spark dont have rate limits at the moment so you can run like a hundred of them

feels amazing to do so

[uncut socket]

Did they stop the free usage of the codex for everyone?

[ivory zodiac]

no apr2

[ivory zodiac]

WUT

[cyan wing]

fr fr

[warped pine]

i mean there are usage limits but no rate limits on how many of them u can run at once

[craggy cloud]

just found out i can access codex cli
im in heaven and was supposed to be asleep 2 hours ago

[cyan wing]

Codex spec'd a physics engine migration in Zig from Chipmunk2D -> Box2D to take 5-7 days
~3k lines of code
it was done in < 1 hr
100% functional, migration complete

[warped pine]

classics

[frosty zealot]

https://x.com/chromiumdev/status/2022363079976034455?s=46

Anybody try this?

https://developer.chrome.com/blog/webmcp-epp

[wide schooner]

context compaction sets it to 0% and forces automatic 2nd compaction

unacceptable

[wicked briar]

skill issue

this punishment for what you did

[wide schooner]

did what

[wicked briar]

you forgot?

bruh

you nuked the general in my old server

[honest saffron]

He has a small context window

[wicked briar]

he small pp

[wide schooner]

OH thats you

the guy who demoted me after i made up the whole server

💀

[wicked briar]

you did not made the whole server

[wide schooner]

cry more

[wicked briar]

you were part of it

[wide schooner]

i literally carried it and u demoted me cuz u thought it was funny

[wicked briar]

it was funny

[wide schooner]

🥀

[wicked briar]

not funny when you deleted general tho

I was gonna give back

[wide schooner]

lies, i gave u like 5 chances

and warned 3 times

[wicked briar]

yeah I really thought you could do nothing

[wide schooner]

🥀

[wicked briar]

group chats are better anyways

[untold ferry]

Showing unable to generate

[wide schooner]

nuh uh

[wicked briar]

seedance 2 better

[wicked briar]

yes

I have most of the guys added to gc

[wide schooner]

no bots and no mod abuse in gc's

[wicked briar]

how do you know bruhmale

[wide schooner]

also gc like 10 ppl liimt or sum

[wicked briar]

why did he add you

[wicked briar]

you need 10 people only

[wide schooner]

yes i do

[untold ferry]

What

[wicked briar]

not for long

I talk to dog

[wide schooner]

i havent talkded with him for years lol

[wicked briar]

💀

he in the gc

[wide schooner]

i dont even play gta anymore

[wicked briar]

same

[wide schooner]

what do u play

[wicked briar]

play with ai

[wide schooner]

i play with my own game

🥀

https://streamable.com/iocl31

[wicked briar]

vibecoded?

lmao

nice

[lost umbra]

hi guys. my codex quota on VS Code are not related to ChatGPT web right?

[high girder]

https://chatgpt.com/codex/settings/usage
That's your GPT Codex usage. If it's not moving, then you're using github copilot chat in VSC

[wide schooner]

mostly

but ai cant guess tweaks so theres ton of work either way

and you need to check security + u need to make your own 3d models

[feral elk]

can’t agree more

[wide schooner]

exacly

id love 45$ plan

or even 40$ plan

[wicked briar]

I love $200 plan

very hard to even use half of this

[wide schooner]

i cant even use the 20$ one lmao

meanwhile claude 200$ plan is 3 prompts per week

[wicked briar]

cap

you can use $20 plan easily

within like 5 days

if you do heavy work maybe early

[celest ibex]

yeah easy

xcode work seems to eat context lol, maybe the size of the build logs or something

[warped pine]

youre not trying hard enough

[celest ibex]

used my weekly up in two days, went back to gemini and it felt dumb as a rock suddenly :D

[wicked briar]

what do you suggest i do

[warped pine]

use it more

[wicked briar]

🫡

had limit reset yesterday only

[plucky halo]

How odd - Spark jsut disappeared from Codex app

[jaunty mountain]

You also don't have GPT 5.3 Codex. This happened to me on the CLI but i closed and reopened it and I got the models back.

[plucky halo]

Yeah - I think it's because I'm at my in-laws in Cornwall and their internet seems worse than dial-up but without the cool sound

[jaunty mountain]

😭 Yeah I remember when it happened to me it was taking awhile to load.

[wide schooner]

5 days but weekly limit so yeah exacly

the 200$ plan is massive overkill for anything

[wicked briar]

not for me

[wicked briar]

yeah overkill but only option any plans to get mid tier plan between $20 and $200 plan @teal cargo

I think thats the price you pay for unlimited codex

[plucky halo]

It's not - I use it professionally and always get close to limits and sometimes run out

[wide schooner]

thats why 2 20$ or 1 40$ should be a thing

[wide schooner]

how the hell

do you run /review after every change or what 💀

"read the entire codebase and find bugs" 😭

i use it a lot and i dont even use up the 20$ limits

[plucky halo]

Haha - I lead an innovation team and we have a huge amount of projects.

[wide schooner]

Aren't there business plans for such cases?

[plucky halo]

Yes and no - the rest of the team are on business plan and I'm on Pro. We've found more value that way

[wide schooner]

Still suprising you can get even close to the pro's ratelimits. Maybe you have it read the entire code everytime?

[plucky halo]

Nope - it's just through a lot of use. I'm working around 16 hours each day

[wide schooner]

i have it running for 8h a day on plus sub and i havent used the limit yet

[plucky halo]

Ah nice 🙂

[wide schooner]

Overall its interesting how inconsistent quota is for all llm's

similiar case with google's and claude's plans too

[plucky halo]

Yeah, I've also got Claude Max 😂

I had the max plan (or whaterver it's called) for Google but I'm not a big fan of that model

[wide schooner]

it was great before they nerfed it by like 90%

how's the claude ratelimits?

[plucky halo]

Although I'm increasingly using Codex for the planning, it's always been Claude for planning - Codex for touchign the code and then back to Claude for review

[plucky halo]

Very similar to Codex, to be honest

[wide schooner]

even considering opus eats 20x the tokens per task?

[plucky halo]

I only use it for planning - might be a different story if it was doing any coding too

[wide schooner]

fair enough

[wicked briar]

False

[plucky halo]

for my use case

It's almost like different people have different uses and that affects token use, Dwight

[wide schooner]

yeah i guess they are way lower but you use it way less so its fine for you

i use codex for everything but sometimes opus is needed for creativity like finding a hard bug or making a new solution

[plucky halo]

Yeah - it's always good to try another model when you're close to banging your head against a wall haha

[desert vapor]

Who here is building code for active trading and investing? Let's DM please.

[glass furnace]

Does anyone know if the parallel agents codex can spin up can be with a different codex model than the session that is spinning them up? I’m curious if my prompt to xhigh in plan mode to use spark model in the parallel agents with detailed instructions is just a waste of time. I have no clue if it’s actually using spark lol

[glass furnace]

I trained a few ML models to day trade options and trigger my programs to enter and close positions in my brokerage. But it’s all proprietary. 😶

[sand shuttle]

when they gonan fix the codex extension 😭 the ui gets so laggy and frozen after some use

[hoary turtle]

I thought i was the only one. It is HIGHLY unoptimised

[orchid plume]

as far as I know, not possible without forking Codex CLI and modifying the tool itself

it is also one of the things I'd like to see, an optional parameter for model, then I could spawn some exploration subagents on a simpler model such as spark or mini

[sand shuttle]

they might aswell use codex to fix it 😂

[celest ibex]

just use codex to get a faster computer

may need extra extra high for that

[finite dragon]

Currently I cannot use codex if I'm not using 1.1.1.1 warp, any idea why? If I do /status my account is flagged as (Unknown), if I use 1.1.1.1 it's fine and says (Plus).
Also I can't do codex login either unless I'm using 1.1.1.1, has anyone experienced this before?

[tacit parrot]

hey, no idea, but I just subscribed to the server to see what's up, I've also been getting networking problems for a couple of days now (but I was assuming it was just the model getting ridicolously slow with half the universe using it)

[finite dragon]

I see, it just happened on a random day (I think like since 3 days ago) and I still wasn't able to find a fix

[ivory zodiac]

https://x.com/LLMJunky/status/2022712980090105883

[orchid plume]

nice, I have to give this a try and see how it does. I guess spark could be alright to use if it's given focused tasks with its own context, with gpt-5.3-codex as the orchestrator

[pastel fog]

Is using OpenCode with GPT account bannable? It's unofficial integration, right?

[orchid plume]

nope, officially supported, need to find the X post somewhere but it's supported now

[pastel fog]

k thx

[orchid plume]

https://x.com/thsottiaux/status/2009742187484065881 here you go, from Tibo at OpenAI

[cyan wing]

why not 5.3 XHigh?

[main nimbus]

I saw Matthew Berman call 5.3c “super slow” - hello?? It’s still faster than previous models

Spark is wayyy faster but I don’t feel like 5.3C is slow compared to prior models.

[hoary turtle]

Is it better to use codex 5.3 with or without skills for overall frontend/backend work?

[boreal holly]

Wow, so I decided recently to give Codex Web a shot, having it do code reviews for PRs, and I gotta say it is absolutely surgical and precise with these reviews. Feel like a fool not taking advantage of this feature sooner

[hard tulip]

I think that they are using spark but only in the alpha

[main nimbus]

I use the frontend-design skill and it actually improved the results. Modified it slightly, though

[hard tulip]

No 5.3 is so bad at frontend and it’s good at backend but it will be 100% better with proper instructions from a skill so yes

[main nimbus]

Not completely terrible, just missing guidance

[hard tulip]

No it’s like really bad

By itself

[hoary turtle]

Its still wayy behind Opus 4.6 with frontend

[hard tulip]

Base GPT 5 is like 10x better at frontend Compared to 5.3-codex lol

[boreal holly]

With skills for sure. I'm perfectly happy with 5.3-Codex frontend, but it required making a couple of skills to guide its process.

[hoary turtle]

Is it possible to know which skills you're using ? I'd be interested in those.

By the way, anyone knows what is Azure AI? Heard of it not so long ago and im wondering its purpose

[boreal holly]

I took @ivory zodiac's custom revision of Anthropic's frontend design skills, carefully reworked them so it's all about Flutter/Dart Widgets, threw in some MVVM stuff about controllers & DI, and that's pretty much what I use. Since that update it's been building perfect mobile views, and only minor revisions for desktop

[wide schooner]

the 2x ratelimits were meant to be for 2 months, but apparently they are gone after 2 weeks instead 🤔

[main nimbus]

Yes but I want the frontend and backend to actually work

[scenic umbra]

What’s the consensus on Spark? Is it worth it?

[main nimbus]

cool demo of the future

[craggy cloud]

same. and i just found out theres a cli version that plus can access. i was almost at my monthly premium limit for github cli so yeah lol

[hoary turtle]

Mix 4.6 and codex lol. 4.6 IS gonna make the frontend work, and codex is gonna make the whole backend

[scenic umbra]

demo, so it’s not ready to be used in any great capacity?

[main nimbus]

No, it’s usable. I mean more of a demonstration of the new AI wafer engines

[scenic umbra]

Ah ok, cool

[main nimbus]

The real neat thing will be when they can get the full models running these engines

[wide schooner]

its basically a bit dumber but way faster

quite like mocking claude for opus fast being both dumber, and more expensive

[scenic umbra]

lol

[wide schooner]

its like skillfully done optimisation

minimal loses for massive speed increase

[scenic umbra]

I think the only bad thing I’ve seen is it goes through its context super fast, but not sure how much that affects things

[wide schooner]

it has only 100k context window, so 25% of 5.3c

[ivory zodiac]

Unfortunately you can't without a fork rn

[ivory zodiac]

128k

[wide schooner]

those 28k are for response, no?

[ivory zodiac]

Amazing 👏 🤩

[main nimbus]

Oh what did you modify?

[ivory zodiac]

128k is for everything. There's no specific allotment for a response as far as I'm aware. It's more than sufficient for what it's intended to be used for 🤙

Also 5.3c has 276k in codex

So it's basically half

[wide schooner]

u sure its not 400k?

[ivory zodiac]

Yes

[wide schooner]

link?

[ivory zodiac]

/ status

[boreal holly]

I think he's saying the codex models are 400k, and they specifically allocate part of that for output tokens only, that's why the "limit" is 276k

[ivory zodiac]

Not needed, many find it no better, I've heard some claim detrimental due to too many cot calls. Not sure i agree with that part but it's faster and very very strong

[wide schooner]

hmm

why its not 400k bruh

[ivory zodiac]

258 my bad

I'm not sure

I thought 400 too

@teal cargo can you clarify

Do spark

[wide schooner]

606k used 💔

[ivory zodiac]

I'm afk

122

[wide schooner]

idk but i think u should redact session too

cuz its like session cookie, no?

[ivory zodiac]

Nah it's just for internal indexing

And openai

[boreal holly]

https://github.com/openai/codex/issues/9773#issuecomment-3831878819

etraut corrected me on that assumption before

[ivory zodiac]

I don't think you can do anything with it

[wide schooner]

oh so i was right

128k output

so spark has 6k output?

[ivory zodiac]

Seems so

6k

But it's dynamic, can be more

Etraut the real mvp

[wide schooner]

its dynamic for all models, but the 6k and 128k are the RESERVED amounts to prevent overflow

[ivory zodiac]

Source. I have a fork here smwhr

Yes I can read

[main nimbus]

Interested in checking it out. I added some basic playwright instructions to mine.

[ivory zodiac]

You ALSO SAID it was 100k and 25% of codex

See how rude that sounds lol

[wide schooner]

but i were correct

[ivory zodiac]

No you weren't lol

It's 128k and 1/3

[wide schooner]

its around 100k and its reserved for it

[boreal holly]

sometimes people dont think it be what it is, but it do

[glass furnace]

It does. But I meant if I am using codex alpha with gpt-5.3-codex xhigh and tell it to use parallel agents with spark… that’s what I am unsure if the parallel agents would spawn as a different model than the main session

[ivory zodiac]

It's 122k but not the point lol

You can really do a lot with 128k if you designate just a task or two

Instead of back and forth pair programming or long document ingestion

[main nimbus]

128k use to be sooo much context

[ivory zodiac]

I'm using a fprk but explore agents are already using spark so you can probably just ask it to implement tasks with an explore agent instead of a worker

It might work

Validate with /agent

[ivory zodiac]

.

haha yeah it works

npm i -g @openai/codex@0.102.0-alpha.4

"use EXPLORE subagents, but instead of exploring, have them implement [tasks from a plan]"

[orchid plume]

ah nice, they were thinking of it already 😛

[cyan wing]

Anyone here use Obsidian for managing markdown files across agents?

[stray swift]

What domains are people working in, that they get productive use from codex operating independently for hours?

[tacit parrot]

"here is a benchmark here are research ideas on how to improve on it here is a callable model that can help you tune your results don't stop before you beat it" would be in my todo

[wide schooner]

hows alpha 7?

[nocturne folio]

i need to remind myself to not open any tabs on openai.com otherwise my pc will crash

its so annoying genuinely, why does it keep crashing my pc

[wide schooner]

why this says 5-weekly not 5-hourly lmao

[ivory zodiac]

How is it bro

[ivory zodiac]

.

I haven't used it yet

[cyan wing]

maybe they should try using https://chatgpt.com/translate/

[cyan wing]

better for reading markdown than Zed
so that's something 😅

[ember venture]

are agents configurable in this alpha?

[ivory zodiac]

No

[glass furnace]

Sounds like You need a new pc

[full elbow]

how long is this codex app waitlist?

[cyan wing]

weird that in the reference image of this key
the logo is facing the camera 🤔

[high girder]

use the image 1.5 api to fix it

or ask gpt to do it

[nocturne folio]

an i7 4770 and a gtx 1050 is good enough for browsing... atleast

[high girder]

depends on your ram. if you're 4-8GB, turn off hardware acceleration in the browser settings. If you're using codex which required 4Gb min, and using multiple browser tabs, you might be running out of memory if you don't have a lot.

[cedar skiff]

This is a vibe-coding computer setup of the future

[glass furnace]

Nah you can’t even load MySpace with that

[cyan wing]

strongly recommend making this into a skill.
it just works.. 🚀

[undone lava]

is there any way to use the new model with plus (20$ish version)

or is it actually pro only?

[cedar skiff]

I used it on plus, uness they changed it it should be good to go

[undone lava]

like i mean the GPT-5.3-Codex-Spark

[cedar skiff]

Oh my bad 😬

[undone lava]

cuz it throws to me the error, which is not like "your version is too low" or something, its: {"detail":"The 'gpt-5.3-codex-spark' model is not supported when using Codex with a ChatGPT account."}

[cedar skiff]

what are the use cases for the model?

I haven't tried it yet, just wondering what ppls thoughts are

[undone lava]

its fast af

#announcements message

but i guess its to pro only

[cedar skiff]

is it also smart?

ok i gave it a go, i see it now

smart enough for most menial tasks and no waiting now

[still stump]

it’s slow on every mac I’ve tried it on. I even tested it in a brand new, empty repo, and response generation is just as slow there

is anyone else experiencing this, or is my account getting throttled for some reason? i’m using codex with a chatgpt pro sub and have about 80% of my weekly limit remaining.

[undone lava]

but like do you pay ~200$ for this?

[craggy cloud]

i am steadily amazed at how the codex cli realizes things - it noticed that i was referring to a file in a different repo and seeked it out instead of throwing a "cant find it" error like usual

i wish i had found this earlier - i am loving it and only just getting it through basic housekeeping stuff just to get used to it. cant wait to set it on something more difficult

[cedar skiff]

imagine if the $200 plan was just spark, that would be a rip

[craggy cloud]

i havent paid for the pro version since they dropped o3pro when they upgraded to 4
im on the plus edition now and the cli gives me access to gpt-5.3-codex (as of today i think)
sure it doesnt give spark but... eh. based on how openai likes to rotate their products, im sure we will get some sort of play with it at some point

there just isnt enough on that tier that justifies it imo

[tacit parrot]

same here. went from being able to do work faster than normal with the model to basically wasting time. I was hoping saturday to be better with less people working, but no. 82% weekly rates left.

I wish they were honest about what was happening. Are we rate limited? Is the service too popular? $200 a month you'd think warrants that kind of answer, but maybe they only care about enterprise customers.

[still stump]

I noticed 5.3 codex xhigh is working fine in Cursor for me. so yeah, it really does look like they're throttling the codex app and cli

[ivory zodiac]

what you saying here? sorry lol

[ivory zodiac]

oh yeah bro, i actually do have this as a skill but in different terms.

i'm really glad its working for you

Open Sparky 😂

[cyan wing]

i use the Zed code editor. Obsidian is better for reading markdown than that 🙃

[ivory zodiac]

ohhh right cool

fair enough

thanks

[cyan wing]

no thank YOU

there's hidden behaviors in the model

that you wouldn't know if you don't ask / find out

[ivory zodiac]

i think the lesson here is if you're ever having trouble with codex, its really important to remember it is SO steerable and its almost steerable to a fault

if you dont explicitly tell it something, it might not infer it

which is good on one hand because its REALLY freaking good at calling tools

but bad on the other because its stuff like that you really dont have to say to other models like claude

so people walk away with a poor UX

but its their own lack of asking.

so sometimes we just have to spell it out explicitly

but the best part is, it actually listens

[cyan wing]

ask, and you shall receive

[tulip aspen]

Zed code seems better. Overall. Are there any caveats to using it over obsidian?

[cyan wing]

depends if you prefer local only, actually readable + editable-in-place markdown

[ivory zodiac]

zed is jus fast its why i use it

its my text editor

i dont even use it for an ide 💀

[tulip aspen]

Yeah I like that. It needs to be easily readable, have good syntax highlighing, proper features (regex lookups, text manipulation) and proper markdown syntax. It would be good if you could convert the original format into like JSON or something

I like being able to choose the colours of certain categories for readability

[cyan wing]

ya same lol i dont write code anymore
it's just so light weight and fast
I have 5/6 terminals open in it, 3-5 projects open
< 500 MB RAM

[tulip aspen]

until you look in #community-help

[frosty zealot]

Holy, just asked Codex to do a large refactor, and it spent an hour hitting me with Q&A, I love it, Claude would ask me 3 questions then start

[tranquil hare]

You piqued my curiosity. What are you referring to specifically? If you don't feel like repeating yourself, just share the link to your previous message. Thnx!

[ivory zodiac]

gpt 5 models are notoriously difficult to prompt

they sometimes require very literal and explicit instructions

things you'd thing should be obvious

might not be

things claude would infer, gpt 5 may not

its a blessing and a curse