codex-discussions

[uneven kayak]

Oops, accidentally deleted the local server env file 😂

Good thing I saved a backup the other day

Man I was worried there for a second though lol

[lost drum]

how could you acidently do it lol

[uneven kayak]

Got 11 secrets in the server env file, maybe I should save them to Nordpass

[lost drum]

notepad

or send gmail to yourself XD

[uneven kayak]

I asked Codex to migrate my monorepo into separate repos and I forgot about .gitignore when I ran the cleanup job

[lost drum]

or write it on sign in minecraft

[uneven kayak]

Lol plaintext email, yeah that's real secure 😂

I didn't give Codex enough credit at first. I had everything consolidated into one repo to make it "easier" for Codex to contextualize everything. But then I learned that wasn't needed, Codex can work in multiple repos simultaneously. Seriously impressive.

[zenith linden]

I would just give them the project files. I wouldn’t charge a lot because I wouldn’t be managing it.

[lean lark]

someone mention Minecraft? 🤪

Random comments...

1. I have workspaces with multiple projects in their own repositories. Codex has no issues.
2. I have workspaces with multiple projects, all in one repo. No issues with Codex.
3. I have a ChatGPT Project dedicated to discussing game development. It rocks.

All of the pieces are there, we just need to keep things organized, specific, well-defined, manageable, and stay on top of the messes that we create.
Cuz when you don't understand your own project anymore, aint no amount of vibing is gonna result in a high-quality product.

[zenith linden]

Just curious, do you use ChatGPT to help you prompt codex?

[lean lark]

Hmm, I often plan things out with ChatGPT, token-free, long fruitful exchanges, and then take the plan to Codex for implementation. I also use ChatGPT for getting AGENTS directives correct, as much as possible, then fine tune with Codex in each project.

[zenith linden]

That’s pretty much what I do

Def worth the money each month

I also had codex automate some tasks for me

It’s awesome

[supple perch]

Yea I build skills for codex to do video editing and music mixing for me, its really nice

[zenith linden]

Which plan r u on

[supple perch]

teams

[ivory yacht]

What are the usage limits like for the free plan of codex?

[zenith linden]

Plus plan gets you a decent amount for low level coders

[ivory yacht]

This says that with the free plan you get to “Explore Codex capabilities on quick coding tasks.”

[zenith linden]

You may get a trial Idk. I wouldn’t count on more than an hour of use

[nocturne folio]

dude whats wrong with openai, why cant they fix the reconnecting issue

ughhh

i got this so much times

[cedar skiff]

maybe its your connection?

[vital bone]

is this normal to complete weekly limit 32% in 1.5 days with 5.5 low in 100$ sub? i think the limits are like 10x ryt

[cedar skiff]

i would be like -150%

I had to stop my automation becuase im almost out

[zenith linden]

Nah it does it for me too. For example If I automate a task that takes about 1 hour to complete, it does the reconnecting thing prolly 4 times throughout the hour. My WiFi speed is 2 gigs up and down, also everything in the house is hardwired

[cedar skiff]

Right but what about the actual conenction to the provider?

i have runs that go for upto 12 hours and i dont get it

[nocturne folio]

yea completely

i wish openai increased limits

[vital bone]

damnn i thought for atleast 3 days it wont cross 40% even with high reasoning as per twitter hype🤡

[nocturne folio]

i mean with 5.4 yea

the model is 2x more expensive

[solemn acorn]

I spent two days doing a bunch of ethernet troubleshooting and somehow only burned through 15% of my weekly usage

pretty great

[vital bone]

model?

[solemn acorn]

5.5 medium/high

the thing sips tokens

there were only maybe 3-5 compactions total which is wild considering all the stuff I was having it do

[vital bone]

nice, i used goals for complete one day the one goal ran for 5 hrs when i was sleeping and 15% used for it with 5.5 low (mostly reverse engineering tasks)

i've to figure out properply using goals when needed instead wasting tokens

[solemn acorn]

I will be a little sad come the end of may when the $100/mo plan gets bumped down from 4x regular usage limit

ah it's only double

that's not too bad

[vital bone]

we need 10x for 100$ , 20x for 200$, thats correct math
5x for 100$, 20x for 200$ is not fair

chatgpt web is goat, with pro thinking i did the planning with it, then codex didnt consumed that much for just execution/debbuging, but the automation i did ate most of my limits

[solemn acorn]

weirdly chatgpt web seems to run a lot better in firefox than my enterprise sub does in microsoft edge

granted I don't do nearly as much with the pro model on my personal account as I do at work

[vital bone]

wdym? u mean personal sub is good compare to enterprise?

[solemn acorn]

no I just think edge is bad

the enterprise subs are slower to get model upgrades, but to my knowledge they are running the same UI

[nocturne folio]

yea

[uneven kayak]

Man this is a bit of an issue. Codex was getting ready to merge a PR so it was following the PR merge flow and running a smoke test like usual, but the smoke test failed on Google auth so right as the agent started to diagnose the auth failure, the context was compressed. After compression, it lost track of where it was in the smoke test and merged the PR even though my agent rules explicitly forbid merging a PR with a failed smoke test. It wasn't really a big deal this time but it exposed a gap in the way agents work that I hadn't really considered before.

[trim rapids]

working around that pretty much requires custom tooling of some kind beyond just plaintext agent instructions, stuff like explicit checkpoints through your workflow that generate unique tokens that are required as input to get past gates on future stages

[vital bone]

i feel u😭 , before when it think it compacting it should do something to preserve key things, i dont know may be it can be done by changing compact prompt or some trigger that can save key things when it near compact

[uneven kayak]

Unique tokens, I like that idea

Well at least it understands why the merge should have been blocked. I adjusted the repo rules to explicitly inherit rules in another repo, I think my monorepo migration might have been a bit loose.

At any rate, it was a simple fix, turns out it wasn't a Google problem at all, it was a stale refresh token for the back end

I stopped using Playwright and browser tests and started running all client tests as full stack, using an Android emulator, so the agent can directly address real world conditions during development. That threw a huge wrench into the project architecture so I've been ironing it out all day.

[cedar skiff]

Block damaging git commands and force them to have explicit permissions.
This bad timing of compaction i think might be part of why the 5.4 xhigh scored worse in some bench marks, it might have built up more context with more thinking and investigation, then could compactsjust before or just after implementation starts or even before planning finishes.

[uneven kayak]

Yeah that's a sound decision, but I'm a risk taker! Lol jk, yeah I'll probably end up doing that soon.

Blocking damaging commands I mean

It seems like it usually does the compacting at proper times, like just before it starts editing files after it has a good plan to get rid of all the extra thinking nonsense that it had to do to arrive at the final plan

But this time it was like, right in the middle of a smoke test, very unusual

Might have had something to do with all my followups, it was a bit too branching, too many different things going on, I should have started a new session.

[vital bone]

is 5.4 high and 5.5 low usage is same? or better to stick with 5.4 high to avoid 2x cost for coding?

[cedar skiff]

5.3 is as good at coding and cheaper

it is not as good at anything else.

I use 5.5 for everything except implementing code

5.5 still does the same mistakes where it brings conversation history into the context of documentation - 5.4 did this as well. I some what fixed it with some developer instructions

[undone patio]

ive found that i code better if i put mayonnaise on my sandwich

i call it the moist method

[plush harbor]

me: gives codex a wireframe on paper
codex: makes a beige website with terrible marketing copy
me: sigh. Fine, off to style.css

[undone patio]

dry sandwich coder

[plush harbor]

yes, mayonnaise time

[undone patio]

the anthropic mind cannot comprehend

[plush harbor]

I got teal buttons not tan this time

[undone patio]

sounds like you used lined paper

[plush harbor]

its mildly amusing just how bad it is at this. But hey, I got my site outline at least

[cedar skiff]

if you are using web try taste skill or impeccable.

impeccable has a whole bunch of util in it

[plush harbor]

I'll do it myself like a stylesheet editing peasant

[boreal holly]

Mayo and cucumber sandwich, hold the bread please 🤤

[pale pond]

Which channel is to ask questions?

I wanna know if i upgrade from my current plus plan to the pro plan does the may 31 promotion apply or its just for new accounts

[cedar skiff]

There is a joke in there somewhere

[lean lark]

Please report that. There's probably a 👎 at the bottom of the thread and it opens a textbox for an explanation.

[lean lark]

This is an ongoing pain for you. Create a ChatGPT project that's hyper focused on CSS/SCSS and enhance it as you go.

Or a workspace with serious CSS Skills where you can take your tough problems.

[plush harbor]

oh I finished yesterday's website, and this morning's website, this is a third one

the first pass is horrid

I'm so far behind on what I'm meant to be doing its getting beyond a joke but at least I'm getting lots of stuff done on the computer

[lean lark]

It's almost 1am and I'm lying awake in your CSS pain. 🤦🏻

just fix it already. 🤪

[plush harbor]

I only just started this latest one. Been on the end of my to-do list for a while but codex speeds things up so much

[lean lark]

ok, i feel better now. gnite. 🤷🏼‍♀️

[vital bone]

got it, so if i do planning with 5.4/5.5 and 5.3 can execute that, but if i change the model then cache tokens will miss?

[trim rapids]

ideally, the plan you create with the better model contains enough context about the task to not have to actually re-read everything that was required to create the plan in the first place during execution but isn't the cache TTL only like an hour anyway?

[cedar skiff]

I just setup 5.5 as an orchestrator to get 5.3 high subagents to do the work. There is no real cache problems here tbh.

[vital bone]

interesting, will try that

[lucid mason]

My readme.md turned into an agents.md by content, but not by name. I have GPT clean that up now.

[gaunt perch]

guys

what is going on with this: ⓘ This chat was flagged for possible cybersecurity risk

i ask it to read documentation it it does this

and here we are now:

ⓘ This chat was flagged for possible cybersecurity risk
If this seems wrong, try rephrasing your request. To get authorized for security work, join the Trusted Access for Cyber program.
https://chatgpt.com/cyber

› HELLO? U THERE?

⚠ Your conversations have multiple flags for possible cybersecurity risk. Responses may take longer because extra safety checks are on. To get authorized for security work, join the Trusted Access for Cyber program: https://chatgpt.com/cyber

ⓘ This chat was flagged for possible cybersecurity risk
If this seems wrong, try rephrasing your request. To get authorized for security work, join the Trusted Access for Cyber program.
https://chatgpt.com/cyber

› whats 5+5

⚠ Your conversations have multiple flags for possible cybersecurity risk. Responses may take longer because extra safety checks are on. To get authorized for security work, join the Trusted Access for Cyber program: https://chatgpt.com/cyber

ⓘ This chat was flagged for possible cybersecurity risk
If this seems wrong, try rephrasing your request. To get authorized for security work, join the Trusted Access for Cyber program.
https://chatgpt.com/cyber

anyone else facing this? this has crossed into literally unusable territory

[uneven kayak]

Oh yeah I'll do that

[velvet wren]

you need to get verified for cyber trusted access

[gaunt perch]

I know but at the same time I am paranoid about privacy and stuff

😭

[velvet wren]

just do it, it's required

[gaunt perch]

if they wanted my info they'd have found it by now 🤣

[trim rapids]

yes, they already have it

they're making sure you have it.

[gaunt perch]

😂

[velvet wren]

just grab a photo ID and go through the steps, takes less than 10 minutes

[gaunt perch]

Don't really see any other choice at this point. It is either that or I am back in vim typing go by hand 😂

[trim rapids]

odds are actual privacy ceased to exist before you were even born anyway

[gaunt perch]

that's todays world man

[trim rapids]

indeed

[gaunt perch]

I can compile my kernel from scratch but I still use google services

I am just lying to myself

🫠

[rocky fog]

just on topic 😄
https://cyberinsider.com/mozilla-mullvad-proton-sign-letter-opposing-uk-age-verification/

[oak trellis]

yes 5.5 does a looot of mistakes

[rocky fog]

why is robert telling everyone that you MUST do ID check just because you triggered the cyber messsage once 😄
just do a new chat and try again
you just must have something triggering it in the chat or code/files

[cedar skiff]

I find it does a really nice job at high, it just cost too much i need to get more done

[oak trellis]

i use medium and have to do few rounds until it gets it right .. from my usage i think i have to go over task to fix wrong logic at least 4x ..

so i waste a lot bc of the quality

so you think 5.3. codex high is better ?!

[plush harbor]

... "back in my day" ....

[cedar skiff]

It is cheaper and the code is about the same

I mean do you do more complicated tasks now than you did when you used 5.3?

It just works like it always did

Im building a cli tool for a plugin atm i could do it with 5.3 but it just easier with 5.5, 5.5 is the smarter layer it knows how to read between the lines, understand this, that, it way better then 5.3 and 5.4

with 5.3 you really have to say what this is or it might just grab the wrong this and mess your junk right up.

But for a controlled work flow can't go wrong

[lucid mason]

TIL that revision-history files can mislead future prompts into thinking they contain current design choices. 🤔

[trim rapids]

saving history in plaintext on disk is the wrong direction in general

what you want is a db you can query relevant current information from

[lucid mason]

I am moving back and forth between Codex and ChatGPT, history in plaintext is the easiest way to move the information along.

[rocky fog]

codex app users
when you go to View and "Toggle file tree"
does that work and how does it look?

It just does nothing for me, on windows, project in wsl

mostly used vscode extension and I keep going back to it because of stuff like this in codex app 😄

[covert kayak]

Hi folks, got multiple bugs with codex chrome plugin today...

It seems that when I run chrome codex plugin in windows, every time I starts a new session the tool is lost, and I get " the callable tools available here don’t include Chrome" in response

[trim rapids]

it's like the context window needle in a haystack problem but the needle it can't find is the fact that the information no longer applies, the only way to prevent the problem you're encountering is to make the information unavailable to the agent without the agent explicitly retrieving it as historical data

if it's just there in plaintext it'll show up in grep output etc and then the agent will open the file and seek directly to the section that it thought was relevant to the task at hand and never even see the gigantic warnings you could leave literally all throughout the file about it potentially being out of date info

[gaunt perch]

Back in my day? Back in my day vim was a novelty! We were handwriting assembly in our notebooks!

With feathers and ink!

[lucid mason]

Which means to harden maintenance wording against that, make sure the file is only read when needed and even cleanup wording in the history file from time to time.

But it's still a culprit to look out for, that's for sure.

[covert kayak]

Anyone got this bug from CodeX today? It only happens if I start a new session

[simple star]

Living the dream

[cedar skiff]

openclaw with breaking changes and regressions every other update :/

[rocky fog]

gotta put everything through your own llm nowadays
then you can git sync and prompt "what did they screw up today?"
🤣

[uneven kayak]

Delivering packages in a thunderstorm, at least I can run Codex while I'm gone. If I remember to leave my computer on this time lol

Still have 8 parts of an 18 part plan to implement, then it's on to the tweaks and adjustments before starting the next plan

But I wouldn't even have to be delivering if I just arrived at the station not even 3 minutes later. They closed due to the weather, but I got my packages just before they closed so I have to go out and deliver in the rain

[inland sonnet]

just do it no crying 😄

[cedar skiff]

oh the power doesnt go out

waiting for one of these resets

got 8% left before i have to try and wrestle with glm or deepseek

the hype about deep seek is a bit uh

it feels two steps behind

[uneven kayak]

It's just really messed up. If I got there 3 minutes later, I would have been paid to go home. So why do they even close at all if they're going to still expect people to deliver? It's not like they cancelled my route when they closed the station, they sent me directly into the thunderstorm

[uneven kayak]

I still have 34% and only one day left lol

But I probably won't have time today to do much

[unique spade]

reset is in about 34 hours for me. and i m at 27% should be fine

[inland sonnet]

Enough time to yap here

[cedar skiff]

that is his digital avatar

[uneven kayak]

Sorry, didn't realize my messages were bothering you so much. Block me lmao

[inland sonnet]

Just giving you common sense

[uneven kayak]

Wow you're a real piece of work. I have common sense thank you very much.

Don't make assumptions

[uneven kayak]

Lol ok point taken

[inland sonnet]

Yappers blocking me 😂

Keep complaining you will stay broke

[uneven kayak]

Well, on second thought, I should have some time for Codex today. Just gotta finish up here first, should only take a couple hours

Gonna have to run two agents to use my useage limit lol

I can probably have one build the web store while the other takes care of a list of UI tweaks I've been procrastinating on

[lost drum]

Yooooo has anyone tried in codex app using android studio?

Cause if it works as gemini in it then I can work on my other project

[cedar skiff]

It can work on it via terminal

apparently there is a plugin for it

[sacred minnow]

i been having the same it says active then in chat and in plugins its not

[signal tapir]

Risky stuff to let AI handle

[tired ermine]

Im looking to create insta/tictok posts with my agent. Does anybody know of any skills that are posted that are comparable to elevelabs. Trying to avoid using there API. or other work arounds to create content

[main nimbus]

if you live in 2024.

AI doesn't magically solve every applicable problem to your requested solution. Like traditional development, it takes iteration across many disciplines.

[remote quiver]

Is there a way for me to use 0auth for chatgpt inside of my own app? (I'm not a coder... but I am vibing)

[jade plover]

This little dude ❤️

[final walrus]

I'd check with the ToS first before thinking about the implementation

[remote quiver]

yeah, that's fine, openclaw and hermes agent both have it, don't they? shouldn't be against their tos then

[vital bone]

started using 5.4 high from last 4 hrs for coding/planning without goal, only 2% weekly limit used and got somewhat better compare to 5.5 low

[grizzled badge]

Been having similar performance outcomes

I still stick to 5.4 medium and high, great results and my limits aren't threatened. The $20 plan (is it Plus?)

[uneven kayak]

UI for the game, yeah that's a bit of a gamble, but the web store is pretty straightforward.

I mean, it's a gamble when you let Codex decide UI design. Not so much when you're making very specific modifications like I'm doing. I set up the initial UI with Codex and now I'm going though and adjusting it

[wide hull]

I moved a folder that a Codex project was located in, how do I point it to the new location?

I'd rather do it with a method that doesnt lose all my chats if possible

[uneven kayak]

Move the folder back to where it was, then ask Codex to move it to the new location.

That's just one way of doing it

[wide hull]

I didnt know it could do that

[uneven kayak]

Sure, it can access your entire computer via command line

[strong hemlock]

is it just me or it's faster to reach quota nowadays

[uneven kayak]

I don't think it's just you, a lot of people have been saying that.

[wide hull]

will it update the Codex desktop app to know its in the new location?

[strong hemlock]

dang, now i'm stuck waiting for my weekly quota to reset

[uneven kayak]

If Codex is the one moving the directory, then yeah Codex will make sure everything is in order. Just tell it something like, "Please comprehensively migrate this directory to this new location, taking into account all appropriate Codex related associations with the directory" or something like that

[wide hull]

I'll give it a shot

[uneven kayak]

Lol I still have 30% left and less than 24 hours to use it 😭

[wide hull]

I've basicalyl been exclusively using Low intelligence to stretch it out and I'm still reaching it a lot faster.

[cyan gyro]

/goal

[strong hemlock]

i would use other models but my claude is 2 days from weekly reset and gemini outputs buggy code

[wide hull]

I dont have the cash to afford more than one coding agent lol

I pay for Codex and I have a lingering free-for-college-students on GitHub Copilot but that's gotten even worse tbh

[strong hemlock]

yeah, i only got 1 month of claude cause codex ran out so fast

[uneven kayak]

Oh I should mention, best to use xHigh for migrations so it doesn't miss anything. Low might move the directory just fine, but xHigh will move it and check for any references and update those too.

[wide hull]

oh

I already sent it with Low.. if it doesnt work I'll undo it

[uneven kayak]

I've been using xHigh exclusively and I still can't hit my useage limit lol

[wide hull]

or roll back

[wide hull]

lucky you 😭

[strong hemlock]

what plan are you on?

[uneven kayak]

Sounds good, it will probably work just fine but it's best to be a little cautious about AI moving entire directories

I use Codex for everything these days, becoming a little too dependant on it lol

[uneven kayak]

Pro $200

[strong hemlock]

OpenAI really needs to figure out a way to add server capacity stat

[strong hemlock]

I wish i could afford to pay this per month

[uneven kayak]

I'm downgrading to the $100 plan next month, don't really need so much useage

[wide hull]

im on the $20/mo plus plan

[strong hemlock]

I'm also on the 20/month

[wide hull]

ya it moved the folder but didnt update the codex files lol

[uneven kayak]

It is really nice being able to use xHigh without even thinking about hitting the limit, but I'd use it even if I had to pay extra for it lol

[wide hull]

ill try on xhigh

hopefully this doesnt burn the rest of my weekly

[uneven kayak]

Yeah, xHigh is the solution. Just be sure to mention "comprehensive", that's a key word. "Comprehensively modify all appropriate Codex associations with the directory"

Or you could just do it manually if you don't want to burn your useage. You could ask the web based ChatGPT how to do it and that's free

Well, free enough for your purposes anyway lol

These days, I sort of jump to the agent for everything before even thinking about doing things manually. Just being able to ask the agent to do stuff for me is intoxicating lol

[wide hull]

Yeah that’s come back to bite me lately

12k line script 💀

[cyan gyro]

same here, think that's the lifehack you kind of learn along the way. To try to not re-invent the wheel yourself.

[uneven kayak]

Damn 12k lines is a lot for a script

[wide hull]

And I can barely understand it

It’s for my game, it’s for the main mechanic

But I def need to split it into modules

[uneven kayak]

Oh alrighty, so not just a script then

[wide hull]

The games main mechanic is a “game engine editor” and the script handles how objects can be edited

[uneven kayak]

Interesting

[wide hull]

It’s all handled in one script but I should be splitting it into modular scripts for each editor tool

[cyan gyro]

throw the zip file in the web app and let it make you a proposal

[wide hull]

That’s honestly what I’ve been doing, giving 5.5 extended thinking the relevant files and having it devise Codex plans to help me break up the file. It has the Godot documentation and everything

[uneven kayak]

Well here's a good rabbit hole for you, ask 5.5 xHigh to "Evaluate the entire structure of my project and make comprehensive changes to improve the architecture and make it machine-readable for Codex agents." Then watch it work for three and a half hours lol

[wide hull]

Codex usage doesn’t affect regular ChatGPT usage right?

[wide hull]

I don’t think my Plus plan can afford that luxury lol. Maybe regular GPT can help tho

I hope one day I can get strong enough hardware to run local agents.

So I don’t have to be bound by usage limits like this

[uneven kayak]

That is really cool, I'm surprised I've never seen a puzzle game with that sort of approach, seems like an obviously good idea

Plus I like the Animal Crossing style vocals lol

[wide hull]

It’s primarily inspired by Deltarune and Mario Maker

[uneven kayak]

Ah, two games I've never played lol

[wide hull]

RPG and Mario game with a level editor

[uneven kayak]

Omg I say lol too much... Lol

[uneven kayak]

Belated answer but yeah that's right, Codex useage is separate

[wide hull]

as much as Codex has helped accelerate the coding side so I can focus on the other aspects of the development, i do want to try and ween myself off it so I dont become reliant on it

[uneven kayak]

Well your game is definitely coming along nicely

[wide hull]

thanks :D

[uneven kayak]

Oh man don't say things like that or Sam might show up with shackles 😂

[wide hull]

har har lol

I think I've gotten a better handle on the engine lately, but I dont want my default coding method for this to be just have codex do it all for me lol

[uneven kayak]

Tbh, I wouldn't have been able to make the game I'm making in the way I'm making it, if it hadn't been for agentic vibe coding. I can understand complex constructs and orchestrate interwoven projects but I'm not that great at the actual coding. So a lot of the more advanced functionality like the entire server/client idea would be beyond my ability. So if I want to make the game the way it's being made now, there's no way I could go back to not using an agent. I'd be stalled out worse than a Cessna in a tornado.

[wide hull]

I'm graduating in Computer Science so I know the code, it's just that I'm better at thinking at a high-level for the operation and Codex can type faster than me

usually I prompt it with comprehensive, detailed descriptions of what I would like to implement, and then let it do its thing

[uneven kayak]

I have a degree in CS too, and the CompTIA trifecta, but I suck at coding 😂

[wide hull]

Nice

I got CompTIA Security+ but thats it lol

[uneven kayak]

That's a good one to have, more useful than A+ for sure

[wide hull]

Theres also the fact that I wanted to spend my energy on school assignments first as well

and the game was more for-fun anyway

so I was more comfortable using codex for the game project than my school stuff

[uneven kayak]

Makes sense

[wide hull]

Plus im getting a proper SWE job soon and I def don't want to let my skills slip for that lol

[worn forge]

Hi, just wondering.. I started to use codex like 14 days ago, but Im getting a lot of recconecting messages like past 4-5 days... is there a way to eliminate them?

[uneven kayak]

Very nice, but your Codex experience will still help a lot in SWE these days. Of course you also have to actually know what you're doing, but agents are making the lives of software engineers so much easier. Sam Altman recently said in an interview that SWE will be 100% agentic in the next few years

The way it's been progressing, I don't doubt it.

[uneven kayak]

Does that happen while the session is running? Or only when you send a followup?

[worn forge]

while session is running

[uneven kayak]

Interesting. High demand I guess. Nothing to be too concerned about except it's a waste of time, but it doesn't harm the session.

[worn forge]

ye but it take more tokens/credits or w/e it is from my rate limits

[wide hull]

i mean ya there are some upsides but I dont want to have my coding ability be rate limited 🤣

[uneven kayak]

It uses prompt caching so those sorts of connection issues don't actually use very many tokens

[worn forge]

oh ok, thanks

[uneven kayak]

Lmao, pay enough money and you'll never have another rate limit again 😆

[wide hull]

I barely can afford Plus bro 😭

[uneven kayak]

Dude I can't afford Pro. I'm about to have to start pawning things to pay for AI subscriptions lmao

But it's like a very high priority for me

[wide hull]

and thats exactly why I want to avoid becoming reliant on these things

[uneven kayak]

I've been wanting to make this game for years, but it always seemed beyond my reach because I never wanted to put in the time and effort to actually learn how to make it. Until now.

Using an agent is like hiring an employee who knows way more than I do lol

As much as I love to yap, I should probably get back to working on the game, I don't have a lot of time today

I'll talk to you guys in a bit

[wide hull]

ya

same here

cya

[inland sonnet]

wow that dude was still yapping

and then complaining about time

dude will never make it

trust me

[meager dragon]

Well I tried use GPT 5.5 on web instead codex after my usage is gone, and I find out the Web version is like worse than GPT 4o programming skills

I told him add debug on the code and he added it and said I have to place a config into .env and he didn't even read the .env file in the code

normally fixing a bug with CodeX only take 1 round chat, and now took me over 40 rounds still not fix it

[lost drum]

why would you do that? dont speedrun cause thats how you loose. I mean dont gamble like this and just plan it out proepperly.

[uneven kayak]

Lol if you could only see my repo, you'd probably say "why so much planning?" 😂

Well I just started the agent with the first trial contract. I set up an agent contract system last night and I just gave it a long list of specific UI related changes so this should be interesting

23 items on the list and some of them are pretty extensive, but the idea of the agent contract is to keep it on track and record the progress so hopefully it will keep the session going until the end without more problems caused by compression

Beautiful, it's already been blocked twice and fully stopped itself. That's a great test, I'm glad it's actually stopping when it should.

[lost drum]

I wired soemthign similar yesterday now I try to make universal agent that identifies your project from scratch and works on it selfly haha but ye I think I am too broke for now to afford next months sub so I will save up, and just use plus plan and in meanwhile develop some skills and freeze my project mabe even better ideas will come

I wanted to speedrun it was so stressed but then I touhg "hey I donte really need to finish it now"

[uneven kayak]

Yeah makes sense, sometimes it helps to take a step back

[lost drum]

not even step back but slow down

[uneven kayak]

But right now I'm full steam ahead lol. I know exactly what I want, the trouble is getting from here to there, but it's coming along nicely.

[signal tapir]

I've only got 2 more days until reset. 😛

[uneven kayak]

Still making great progress, been thinking about this game for a long time

[signal tapir]

enforced vacation. 😉

[uneven kayak]

I have about 10 hours, gonna get the most of it, doing game client and website updates at the same time

Both on xHigh lol

[signal tapir]

I've been installing a new version of Fedora, and plan to study as well. Slow down, my butt. I just need another direction. 😛

Hmm. I wonder how good Codex would be at fixing configuration issues in an OS installation.

[uneven kayak]

If it can be done via command line, Codex can probably do it. It can just search the internet for anything it doesn't already know, or you can give it reference websites to pull information from

[signal tapir]

I've been using ChatGPT all day, and it could not find a solution for me.

[uneven kayak]

Well the nice thing about Codex is, it can actually try stuff out and check error logs

[signal tapir]

I would need to write some sort of script that started it up again after a reboot.

[uneven kayak]

One word. Autonomousosity. 😂

[signal tapir]

And also, connect a KVM to let it hard reboot when things get stuck

I mean I wouldn't let an AI loose in the root of my system, but theoretically

[uneven kayak]

Or you could tell it to stop when a reboot is needed so you can perform the reboot manually

[signal tapir]

Rebooting is the most annoyuing part. 😛

[uneven kayak]

Lol so true

Man this UI task contract is going so damn well so far, really loving the changes it's made already

I mean, the UI changes as well as the workflow changes, love em both lol

[lean lark]

Did a recent update to Codex delete .codex/AGENTS.md, maybe in favor of Skills?

Um, and my config.toml file was replaced too.
I guess I'm not reading the changelogs. 🙁

[signal tapir]

Not here

[lean lark]

Oh, nvm, I was looking at the .codex folder in Windows instead of .codex in Linux. 🤦‍♂️

[signal tapir]

They're both still in my project

lol

[deft gyro]

did rate limits get screwed for weekly recently

[velvet wren]

no, why?

[deft gyro]

nothing

just seem to have ran out earlier than expected

guess it has been a heavier week

[velvet wren]

are you on Plus?

[deft gyro]

ye

[signal tapir]

That happens to me every single week.

[velvet wren]

yeah me too, it's hard

[deft gyro]

never happened to me till now

man

gpt 5.5 mini when

maybe ill fork over the 100 bucks

[velvet wren]

I am thinking of resubscribing to the $200 pro plan

[deft gyro]

ugh

and codex desktop aur build is broken rn

wait

it was updated today

how is it out of date

[inland sonnet]

Codex is becoming claude at this point

with this cyber secure bullcrap

Setting up a random username API is considerd possible cybersecurity risk

And thats why you never should sub to one provider

id rather use a chinese one then then this bs

Both take my data anyway 😂

if its usa or china who cares

[signal tapir]

I'll stick with the foremost product, which is OpenAI

[lost drum]

wait can I work in codex on trading project or no let me know?

[inland sonnet]

If you use 5.4, there no cyber risk

:p

[lost drum]

wdm cyber risk?

I use 5.5 now

and idk if to stop him cause I jsut started

[inland sonnet]

How long u dont get any issues it should be good

what do you mean else?

why wouldnt it be allowed

[lost drum]

I started it 5min ago

I mean will tehy ban my whole codex acc or just this tread?

[inland sonnet]

You won't get banned, how long you dont try to jailbreak or bypass safety filter

dw lol

[lost drum]

oh I see

I dont plan to bypass no clue what the safety filter is tho but ye and why its related to trading

[inland sonnet]

if u request something its not happy with it will tell you and change behavior or not do it at all.

[lost drum]

he seems to not care

he just follows so I guess my project is ot breaking rules

it just spams "not live trade, just signals"

Sorry, I was just scared I will loose my main project to some side hustle

[inland sonnet]

Nah haha ur complete fine

[lost drum]

thank You

[plush harbor]

seems per-project for me. My site I spend the most time on doesn't move the token meter much. But I ran out of things to do on it

[inland sonnet]

Also know u always can ChatGPT itself if it would allow that or not. 🙂

[plush harbor]

then I ran out of things to do on a second project, finished a third project I only just thought of (rebuild of an old archived site) and now onto a new project. Still got quota. See how long it lasts I guess

[lost drum]

I rather ask ChillPepe than a machine

[inland sonnet]

I'm AI

fooled

jk

[bronze obsidian]

So yeah

Am I doing something wrong ? Seems like my context window is progressively getting more likely to hang for several seconds, too.

[torpid trout]

That’s your clue.
Yes you’ll get permanently banned from using any OpenAI service if they decide you’re automating trading.

Be careful. There’s a user who was whining a while ago exactly about this - allegedly created a „just a signal bot“ and got suspended. For good, no appeal.

Their tos have wording about this, so be sure to read it first

[lost drum]

oh so even if its just a signal they can ban?

oh man

then itdk what to do

I think I hsould delete it archive it

[torpid trout]

The point is, automating

[lost drum]

but it is automated

[torpid trout]

read their tos or have gpt read it
It’s as usual and of course vague af

[lost drum]

I mean it just sends on discord alerts like entry here and there

I mean it also auto takes trades but only in virtual acc that is created within scirpt so no official platfrom integrated at all but ye the risk is still here

[torpid trout]

That’s too late lol.
It’s not about the work done it’s about the inferred answers

Main question for me is … when do they check. Immediately at inference? Later on? In which case as you say if I delete a chat or dont improve llm they can’t?
🤷‍♂️

What I do know is, it’s not worth losing access to Oai for the build of a useless and not-working treading automation

You’ll lose money any way

Maybe you could get around all that with „purely academic“ reasoning

[lost drum]

ye..

[torpid trout]

As much I’d effing love to meddle with it lol

What you can do is make you help train your own model to trade

That’s not using gpt to automate trading so it’s fine

What you can’t do is use gpt to interpret these signals and give you or do trades based on it

I think it’s a worthy distinction to make

[lost drum]

I mean they should atleadt give warning or some

[torpid trout]

… warning == TOS

[inland sonnet]

It allows openclaw etc

why wouldnt this be allowed?

They also added /goal

the problem is clearly not that it keeps running.

[plush harbor]

talking automated trading bots here

[zenith linden]

My connection is fine I promise you lol. Now OpenAI servers I can’t vouch for. Sometimes it does long tasks fine and sometimes it disconnects

[solid lake]

Try switching from Websocket

Or to

If you’re on one or the other

[cedar skiff]

tried cancelling and then buying sub again to get a reset. doesnt work

[plush harbor]

that smells like desperation

[bronze obsidian]

Roger, will try, thanks

[inland sonnet]

cause its not instant cancel.

😂

nice try

[inland sonnet]

api req fail beter then mid stream fail.

[mortal phoenix]

codex wont let me save my mcp server

i really doint knowwhy

[cedar skiff]

what do you mean?

[solid lake]

Correct

[cedar skiff]

My subscription ended yesterday, i cancelled, leaving me with no access since the month had ended. My weekly rollover is tomorrow. So i purchased again, but they maintain the weekly data record.

Cancellation was finalised and subscription was ended.

It stayed at 8% weekly.

[mortal phoenix]

not trying to beg but does anyone know the issue or is there a support section

[inland sonnet]

They dont instantly cancel, even if your next month start and you can't get billed

[cedar skiff]

I just said the cancellation was finalised

[inland sonnet]

so you went to free?

[cedar skiff]

Yes

[inland sonnet]

Then they keep tracking it seems and sub doenst matter to bad i guess.

You beter did try a new acc

[cedar skiff]

How many ways does some one have to explain the same thing to you 🙄

[solid lake]

The /status has been buggy for free I tried too it gave me 97% for the longest time till it updated to 2%

[inland sonnet]

I understand you but they keep track no matter sub i just said

[cedar skiff]

Yes thats what i said in the first place. Glad you caught up with the idea.

[plush harbor]

well if they keep phone numbers on record for a month they probably keep usage data for a month too, which makes sense or everyone would be tryng this

[solid lake]

I mean if you’re paying it shouldn’t be that bad no, only if you abuse the free plus plans

[plush harbor]

chatgpt is picking on my commit messages 🙁

gonna scare my users when the site goes from black text on white background with brown buttons to charcoal/red/gold with metallic gold icons and buttons, without warning

[lost prairie]

i got the same problem

have u found any fix yet

been tryna look for so long

[west dirge]

going to try this and report back

with the $20 plan

[idle cypress]

what ??

[plush harbor]

I have a $5 patreon tier, so anything I do on my site that's enough to convince people to give me $5 a month I'm ok with

[solid lake]

Noticing that extra high is the only level where it consistently refers to my agents md and skills everytime I compact context

For long chats

[solid lake]

You shoud queue the message instead

For after tool call

[idle cypress]

no??

waht does that have to do with anything

[solid lake]

You’re interrupting the tool call

[idle cypress]

the agent just finished in those states

[idle cypress]

im not

[idle cypress]

you can tell because it has the copy / thumbs up / thumbs down icon

[idle cypress]

the agent itself has no idea if it was actually interrupted or stopped itself prematurely - theres nothing in the prompt

[solid lake]

After you asked if it’s done?

Or did it stop before you asked it

[idle cypress]

it stopped before i asked it

you can also see the improperly formatted diff

9idk

idk

🤷‍♂️

[solid lake]

I don’t use much of the codex app tbh so idk if this is app bug

Oh

“If only the codex app existed for Linux officially”

[remote rapids]

Guys is uploading files down for anyone else

[zenith linden]

How’d it go?

[verbal zenith]

Sup

Is there anyway i can trigger my 5h period from my phone without any laptop access?

[west dirge]

nothing crazy yet. it first tried to go on a paid survey website and earn money from there, but i told it to move to something else since it didn't want to make submissions/answer questions without my input

doesnt even want to progress past login screens

will probably give it a throwaway email after my usage resets in 10 min

currently it's trying to make a storyboard template pack and sell it

but cant progress without login

[solid lake]

• type /status
• 12 may
• look at date
• 11 may
• repeat till date says 12 may

[frail meadow]

Do you guys still use deep research?

[river bramble]

4% for 18h ... 💀

[blissful basin]

longest running goal yet, and it is still going

[solid lake]

Yeah

[last meadow]

umm , chat does anyone know why

i have 2500$ from yc and my card after being added as a payment method 5$ were deducted but i am still not bumped up to tier 1

anyone here that can help in

[velvet wren]

this is off topic for here

[stoic needle]

Why am I paying money for something consistently broken. Lately Codex chats especially in VS Code Editor will just refuse to load no matter the t/s steps taken ie. uninstall reload reinstall do a rain dance make a sacrifice..nothing works. It's just so buns and not worth the monthly price anymore. Get mid project and damn thing wont load to even gather chat history to keep working on a thread or flow of work 😐

[last meadow]

which channel do i go to rn ??] also mb

[velvet wren]

#off-topic

[plush harbor]

seems the thing I got codex to do knocked my time for importing one Thing down from 48 seconds to 6 seconds, gotta be happy with that

[silver dew]

Can anyone explain how /goal works?

Is it another markdown trick

I want to know whether I can trust it and how to use it correctly and for what tasks

[chrome raven]

https://developers.openai.com/codex/use-cases/follow-goals

[silver dew]

So OpenAI is listening to lazy vibe coders good

/goal “Build a billion dollar business make no mistakes”

[chrome raven]

it's kinda what things should be in the first place, with this and the right spec and details, full products can be built

[silver dew]

It’s sad that some people still don’t know how good Codex is and use Claude Code

I used to be very mad at Claude all the time

[oak trellis]

@cedar skiff correct codex 5.3 is much better .. using high its same as 5.5 medium but uses much less crdits

tokens

[silver dew]

5.5 can reason about tasks unrelated to coding tho and can make less mistakes when choosing tools and recommending stuff

[oak trellis]

yes beside code i use 5.5

but pure coding 5.3 codex high works better for me as 5.5 medium somewhow

so many mistakes it dies

does

[silver dew]

What are you building?

[plush harbor]

no shortage of claude ambassadors coming into this discord to tell us how good claude is

[warped minnow]

Does it must be on github or hugginface is also possible for codex oos?

[silver dew]

Claude is good for small projects

I wouldn’t trust Claude in building in a serious repo tho, not even a mobile app

I like Sonnet 4.6 for its speed, price and temperature tho

[craggy jewel]

Can you imagine: What Codex will be like in a year. Two years? I mean for my dev work, 5.5 High could be the last model ever released and I would be fine with that. Probably have local llm equivalents after 2 years.

[plush harbor]

the stats page has a graph of how much you talk to it and my graph is like ... 0 before I got sick and a near linear trend up from there

[craggy jewel]

Like, what is the killer feature it is missing? (speed, longer context window etc)

[velvet wren]

I am really not missing anything with Codex as it is now, for me it's pretty good already

[craggy jewel]

hopefully evolves to more usage, lower price

[velvet wren]

yeah, I wouldn't mind more usage

[plush harbor]

probably be more usage higher price tbh

[craggy jewel]

I'm thinking a local llm appliance with 1TB gpu mem is in our futures, as common as a printer.

[velvet wren]

my current local LLM appliance is a MacBook Pro, M2 Max with 96GB RAM

[plush harbor]

mine are all pretty small, but one is alarmingly well aligned to what openai does

[uneven kayak]

I'm trying to come up with something but every time I think of something, I'm like "no, it can already do that..." Lol

[craggy jewel]

mmm 96GB would be nice...my 16GB RTX 5060 just teases me what could be

[plush harbor]

I'll get myself a shiny new computer when I actually have more spare income. Gotta get the AI to build me the next big thing first. 1. AI 2. Stuff. 3. Profit. I hear it works

[craggy jewel]

hopefully all this AI hardware research drives some cool new hardware in several years. Going to be a bit of a consumer hardware desert til then i think.

[velvet wren]

Apple just discontinued the Mac Studio with 512GB RAM, now the beefiest Mac you can get only has 256GB RAM

[plush harbor]

probably ram prices

[craggy jewel]

wandering off-topic here, but it just occurred to me that we could now build a very realistic Data from Star Trek, and have it do all the Data stuff. Heck, GPT would understand humor better 🙂

[neat hound]

I can't preview md files with the formatting anymore, i see them raw, did i touch something or did that come with an update ?

[uneven kayak]

Man, adding my custom font to the game really makes such a huge difference.

[plush harbor]

oh it finished the font eh

[still trellis]

codex app unusually slow at the moment:
Codex 26.506.31421 is currently the newest version available.

[uneven kayak]

Yeah, I posted an example of it here #codex-show-and-tell message

The visual aesthetic perfectly matches the style of the game

[plush harbor]

hmm I'm aparently mostly using open sans, inter, roboto slab and comfortaa. The last one was a gemini pick though

your font reminds me of Brazil, the movie

[uneven kayak]

Well if you decide to make a custom font, let me save you some trouble. Tell Codex "do not take shortcuts by overlaying SVG graphics on top of any preexisting font, instead, hand draw each letter entirely from scratch". If I had done that when I started the font, I would have saved myself like an hour of frustration.

[oak trellis]

what ?? ```Invalid prompt: your prompt was flagged as potentially violating our usage policy. Please try again with a different prompt: https://platform.openai.com/docs/guides/reasoning#advice-on-prompting

[uneven kayak]

Well I'm curious, what was the prompt? Lol

[oak trellis]

Investigate why L0 logs show Jupiter 403 at startup; provide root cause and fix.

i try to debug 403

[uneven kayak]

Huh weird

[oak trellis]

yes

[uneven kayak]

I say try it again in a new session

[oak trellis]

first time i saw that

[velvet wren]

have you registered for cyber?

[oak trellis]

no

[velvet wren]

that's why

[uneven kayak]

If that's the case, it normally shows a message about being blocked due to cyber security

[oak trellis]

i mean we should fix 403 errors without cyber haha

[oak trellis]

ah ok

Searching the web for redis HSET key without expiry memory persistence hash key count
redis HSET key without expiry memory persistence hash key count```

[velvet wren]

I agree with you, I didn't make the rules 🙃

[oak trellis]

ok its back working

got almost panic attack.. without codex no life anymore in the codespace. forgot how to code

💀

5.3 codex i used 2% in not sure like in 3 hours lol .. its really cheap i have the feeling

[uneven kayak]

Yeah it's definitely slow right now, wonder why

[cedar skiff]

I'm on 4% it's getting hairy

[uneven kayak]

I have a little more time than I thought, got until 5:30 PM for this week

[cedar skiff]

Mine resets at 8am tomorrow, its 6pm atm

[uneven kayak]

Ah, you'll use up that last 4% way before then

I still have 29% left, but I have enough stuff to keep it running pretty much all day

[plush harbor]

I have 45% left and a day and a half. I may have to hound the machine into being useful

[velvet wren]

I have 31% left until May 13

[plush harbor]

I'm gonna end up with a public facing openai button and a bigger api bill if I give in to this urge

[turbid axle]

I have 21% left till tomorrow

[north tendon]

11% codex 20x tomorrow 🙂

[still trellis]

mine is bigger than yours

[cedar skiff]

I still have automated tasks to run, had to stop so i could carry on with the manual stop

[turbid axle]

100pro was not enough, 200pro is challenging me to push more, must eat more tokens

[cedar skiff]

doing gods work for the rest of us i see

[turbid axle]

alright george... alright...

I had ai write a detailed spec doc, and just just have /goal cycle through this thing forever more. I guess im using AI now to train my application until it's weights align the spec doc?? 😄

just 24/7, every now and then I test some stuff. and things just 'improve'

pretty crazy

[still trellis]

yeah i do the same but i dont touch my machine on weekends or after 5... discipline lads and lasses.

[turbid axle]

yeh I dont look at it, sometimes I get curious so I test things, or guide it a bit to fix things more directionally, but overall, its totally on its own

[uneven kayak]

Oh wow, adding the archive role is like a walk through the past lol

[still trellis]

anyone went down from 200 to 100 pro ? I might need to consider it but im greedy

[turbid axle]

I did, but went back to 200

100 was totally fine, but then /goal released

[still trellis]

exactly and when i tried symphony it ate my 200 pro like a fat kid loves cake

[turbid axle]

buddy is coding away for like 80hrs+, I need the tokens 😄

[uneven kayak]

I don't really understand /goal, does it do anything that Codex didn't already do before? I use agent task contracts to make sure it works until it's supposed to stop

[still trellis]

I hear you, but i find goal does more testing and refinement so often the outcome is more aligned?

[uneven kayak]

Yeah makes sense

[turbid axle]

its oai's version of this https://github.com/snarktank/ralph

[uneven kayak]

Oh, does /goal run fresh instances? That could be very helpful

[turbid axle]

not sure how it handles context tbh, it just stays in the same thread

its mostly just easy than some custom setup. just /goal some goal, and done

you can pause, and you can still steer it. etc

[uneven kayak]

Ah I was just asking because Ralph uses fresh instances, that would be really cool, could save tokens by not contextualizing irrelevant things

[turbid axle]

I understand, not sure how it handles context. I think it keeps the same context and relies of compaction.

context pruning is something I am still hoping for. the ai should just clear resolved patterns from its working memory imo

API

[uneven kayak]

The API doesn't have bonus useage, you pay for what you use. So I'd just say hello to ChatGPT and welcome to the family lol

[turbid axle]

this. oai is slowly aligning them all it seems though. would not be surprised if that just merge it all into a single thing, stop with the split usages etc.

[velvet wren]

no the API is the most expensive way to do things

[uneven kayak]

There is a project recently posted in #codex-show-and-tell that claims to prune context to save tokens, but I'm a little iffy about it because I don't want to lose something that is actually needed by accident. So I'm just gonna keep using more tokens 😭

[turbid axle]

codex is an app, its free. what you want is model access. API is pay for usage, its clean and predictable, but more expensive.
most of us use just chatgpt accounts. these give you an unknown sized bag of tokens. oai can adjust this based on whatever they like essentially. effectively its much cheaper than API. but pricing and usage can vary quite a bit

[velvet wren]

no, don't do Codex on a Business plan

if Plus is not enough for you then subscribe to the $100 or $200 Pro plan

[turbid axle]

perhaps you can just talk to oai support. most likely you just want 100pro or 200pro accounts. business is really for groups, and gives you business type features. if you dont care about those, just use normal accounts/plus/pro, etc.

[uneven kayak]

The $100 plan is still the best choice right now until the end of the month

Best value in agentic model access overall, across any platform

Though I have to admit, Replit is a hell of a lot faster than Codex lol

[turbid axle]

hm interesting. but ill wait for oai to add it 🙂

[uneven kayak]

Me too

[turbid axle]

im hoping for continuous learning this year. AGI 2027?

[cedar skiff]

I'm hoping for a few more years to keep milking what we have

if we get some sort of agi those $$ opportunities in our domain go away

gotta get that cheddar

[turbid axle]

well, sure, but also, just like most startups these days. im pretty confident that oai will steamroll the financial system too.
I'm pretty sold on the idea that everything we are doing now, its pretty much pointless other than a temporary way to give ourself meaning, 'something to do'

that big pile of money about to go up in flames im afraid

[ivory cairn]

There isn't just pay for it and u will get the bonus usage which exceeds the amount u pay for via API anyways

[turbid axle]

if you want to build wealth, build/own compute and power. that might give you a chance, if for some reason these models don't manage to make it all dirt cheap and abundant

but hey, I like cheddar

[cedar skiff]

You can provide value via a service in other domains by using the skill you have in this one. If you are building software to support software then i agree, but filling a niche can make money.

[turbid axle]

I don't disagree you can make money, at all.
I just think money, as is, will become utterly pointless in the near future

[cedar skiff]

uh doubht the ppl who control it will agree. resources have to change hands some how.

[turbid axle]

hard assets is where its at, hard resources, power, and compute.

[cedar skiff]

Yo here ill give you 500 trillion tokens for that oil tanker filled with oil isnt going to cut it

[turbid axle]

sure, but look at countries with extreme inflation, that probably models how money will be viewed as a nice visual

[silver dew]

Not meaning but purpose
And no, I disagree that it’s pointless. Money can buy resources even if you’re right

[turbid axle]

we will find out

[silver dew]

Electricity cost and compute will come down to zero, nobody needs compute from a datacenter on earth when you have racks powered by the sun in space

That’s what is cooking rn basically and it’s so obviously right

[turbid axle]

a huge thing money currently encodes is effort. we are about to remove a huge chunk here.

scarcity will remain part of it. that is true

at least in terms of physical reality

[silver dew]

How does it make a difference?

[turbid axle]

maybe think of it as a language, where an entire class of words and meanings is simply deleted because it no longer has purpose. then consider what would happen to all the dictionaries.

[cedar skiff]

Yeah ill stick to making money for the time being anyway

[turbid axle]

im sure money, or some kind of 'exchange' token will remain. but I wonder about the shape of it

[silver dew]

People may stop believing in the system you say?

[velvet wren]

this chat probably belongs in #ai-discussions now

[turbid axle]

yes I stop

back to /goal talk

[cedar skiff]

still on 4% it goes down slow without agents running in the background

[turbid axle]

im at 19%, try harding to get rid of it today, without just randomly spawning xhigh fast mode bs

if only I could store them, and then distrubute them in exchange for things somehow

[craggy jewel]

That, and you need to trust that proxy with your data.

[uneven kayak]

Bingo, that was my other concern

I'm at work while Codex is continuing with the strict, specific UI adjustment. The task contract seems to really be helping, it actually seems to know what it's doing lol

It's on hour 4 so far

[craggy jewel]

Did you ever figure out your hosting re: AWS being expensive? how about AWS LightSail, good free tier.

[uneven kayak]

I'm using a small EC2 instance, looks like the total cost of the EC2, ElastiCache, RDS, load balancing and storage comes out to about $5/day

[craggy jewel]

running linux?

[uneven kayak]

But that doesn't scale with useage, I can get thousands of players before that cost will increase

[uneven kayak]

Yes

[craggy jewel]

for testing, $10/month linux on lightsail. there's a migration path i think from lightsail to ec2 if you need.

still would have storage costs though, but cheaper than ec2 until there is a large load.

[velvet wren]

I moved from AWS to Vultr, haven't looked back

[uneven kayak]

Good idea

[uneven kayak]

I'll look into that as well

[plush harbor]

is the desktop codex different to the cli one? I have not had it do anything this phenomenally stupid before

[velvet wren]

without context I can't discern what is going on there

[uneven kayak]

Lol it's only as smart as the instructions it has access to, you can get pretty strict in the agent rules and skills

[plush harbor]

it made a new edit field and decided to put it inside a deliberately hidden section of infrequently used fields

like the ONLY hidden section

[torpid trout]

The models aren’t different. You just got a bad inference artifact. Close that chat and restart.

Happens.

[plush harbor]

just weird as all the other new fields I've had codex add its stuck them in sensible places, in their own little groups. This form is getting big so its sectioned up

[uneven kayak]

So add a new architecture invariant, something like "Edit fields should be unhidden unless explicitly instructed to hide them."

[plush harbor]

I'm just happily wasting tokens here cos I have half my quota and one day

[torpid trout]

That’s the way

[plush harbor]

I now have a button that says "juicy" when I press it and presumably costs me money

fabulous use of tokens if I do say so myself

[blissful basin]

uhh just tested credits how it looks compared to subscription, and damn, credits are being drained insanely fast 😄

[turbid axle]

Yeah I too found that strict spec docs to guide it really helps it stay on track.

[blissful basin]

loaded up 250 credits 5 mins ago, now at 160 😛

[plush harbor]

I think my juicy button is cheaper. Not showing up yet though

[lucid mason]

How large (as in appr. line) is such a contract that makes it work for 4h+?

[uneven kayak]

The agent contract for the UI work is about 500 lines

I can send you my agent task contract scaffolding later if you're interested

Just not at home right now

[lucid mason]

Thanks, I'd like that to learn about these things. Like "is it very specific or is it just a lot of areas to touch". And even more so how it gets Codex/GPT to do this in one big commit instead of it constantly trying to do smaller transitional steps. Yesterday I reminded it that for personal projects I don't have deadlines to follow and don't have to ship working versions to customers/end-users. Kind of "just work towards the end-result, don't give me hybrid versions of old and new all the time".

And also "I can wait, just finish the work (on the plan)."

[uneven kayak]

That's all covered by the scaffolding, I think you'll really like it

[lucid mason]

Thanks, it's appreciated. I often try to start long tasks shortly before my 5h or weekly limit runs out to make it finish said long task on tokens over the limit. You have to make sure not to send any follow up prompts then, though, because that will make it stop right at the limit, leaving half-finished edits behind.

[uneven kayak]

Actually I'll go one step further and create a modular repo that can be used to add agent task contracts to any agentic project

[lucid mason]

Does this work because it uses sub-agents or would the main agent/chat also adhere to it and do a properly large chunk of work in one go?

[uneven kayak]

So a running session doesn't get stopped by the useage limit? Nice

[lucid mason]

Yes, a running session doesn't get stopped, but if you ever (often accidentally) post a follow up then it really stops right in the middle of the task. That can be quite a problem.

[uneven kayak]

It doesn't explicitly use sub agents. The contract has to be checked at the end of a session and it will keep the agent working if the acceptance criteria aren't met

[plush harbor]

I am unsure if I should let chatgpt name my juicy button

now I need to look at what codex actually called it

[lucid mason]

And for ChatGPT you get some similar problem. When you reach the limit of a chat (as in too long) and the last chat output includes a download file link then that link is always broken and cannot be recovered. You have to start over the last prompt in a new chat then to get a new file (created from scratch/last implementation revision).

[uneven kayak]

Sauce intensifier is the clear winner imo 😂

[plush harbor]

psht. Just audio with help text Hear pronunciation (AI audio) which is far less interesting. Also RIP my API balance

[lucid mason]

I didn't look into contracts yet. I had ChatGPT write me one for itself and it then proposed an "even sharper" one for Codex. But these are less than a page long, not hundreds of lines.

Your complaint is valid, though. I have often defaulted to incremental transitional maintenance steps because that is the safer general behavior: reduce blast radius, keep each patch reviewable, avoid inventing missing constraints. But for your projects, that is often the wrong bias. You usually want the model to resolve uncertainty locally and deliver the largest coherent finished step it can.

[plush harbor]

some api is absurdly cheap. My three test presses of my new button? 0.001c

[lucid mason]

@uneven kayak I had GPT write a "very thorough" contract and it ended up being 409 lines long. Let's see how ChatGPT and Codex handle that.

[uneven kayak]

I just asked Codex to create a new agent task contract framework repo for me, it'll be ready for my review when I get home in a couple hours. I specifically asked it to make the repo installable into any agentic project using Codex, Claude Code, Cursor or Replit, and to ask the user a series of questions to tailor the framework to their specific needs.

When it's done, I'll post it in #codex-show-and-tell

[vital reef]

does it account for every loophole and or condition?

[lucid mason]

Likely not, especially since it is a first try and doesn't have the tailor made questions Andrew just wrote about. I never looked into contracts before (and am only in my second week of GPT and LLM based coding in general).

[vital reef]

hmm ive no idea only used chat apps and some agent work in ide's

[lucid mason]

My weekly Codex limit ran out last night, so I will have to wait for it to renew to test in Codex, but I included a contract into my package now at have GPT improve it based on Andrew's last post. Will try in ChatGPT first until I get back my weekly.

[chrome raven]

i am surprised no resets this last week, damn

[lucid mason]

For my project structure it seems better not to ask tailoring questions, because most stuff is already documented/written down:

the better version is not a generic “agent task contract framework” that asks many tailoring questions. Your projects already have strong owner docs, module maps, revision rules, and maintenance policy. Adding a questionnaire layer would likely reintroduce the exact bloat and confirmation-loop behavior you are trying to prevent.

[rocky fog]

careful what you wish for
too many resets and then "whoops we need to lower usages overall" 😄

[chrome raven]

that's true too 😄

[uneven kayak]

I'll make a note that the framework installation process should read existing documentation and ask only relevant questions (or potentially none at all).

[lucid mason]

Since I have to go back and forth between (too limited) Codex and ChatGPT anyway, I make it a habit to use up my weekly limit about half-way in. Got me some nice resets out of the last few weeks. 😉

[signal tapir]

I expect a reset tomorrow, a couple of hours before my scheduled one. Just to mess with me. 😛

[cedar skiff]

or a reset 2 hours after your reset time pushing your reset time forwards

[signal tapir]

yea, had that a few times

[rocky fog]

yeah thats what they have been doing lately
then you might have a goal running and then "Whoops they already reset"

[unique spade]

having a lot of fun testing my reasoning harness on ProgramBench tasks

doing a re-run with 5.4-mini agents only right now, since vanilla 5.4-mini has 48.1% on that task

claude models sure spend a lot of money on this tasks compared to gpt models lol

[west dirge]

update: $0 made, it is scared out of its mind to do literally anything

surveys don't work, and it doesn't want to log into any website without my approval

[lost drum]

W for actually trying

[west dirge]

thanks lol wanted to see for myself

it's currently trying to file a vercel vunerability report on hackerone

i have to approve of it to do anything

[signal tapir]

GPT-5.5 is incredibly capable. Can probably do everything Claude can do, except possibly frontend design.

[velvet wren]

Codex with 5.5 is the best there is right now

[plucky trout]

What appeals to me about OpenAI is the fact that I can use my Codex subscription anywhere Anthropic prohibits the use of harnesses other than Claude-Code (I’d like to use OpenCode or Kilo). And for the frontend, it doesn’t matter because I’m a developer, so I like to keep control over what the AI does

[unique spade]

haha, boosted gpt 5.4-mini from 48.1 to 93%

[signal tapir]

Surely both AIs can answer your questions?

[velvet wren]

then you will be more than happy with Codex and 5.5

[lost drum]

agi?

[lost drum]

I want to test mine how to do it

hehe

I mean nahhh idc he is not designed for such tasks

but I could iterrate it to be designed for such thingie

[unique spade]

https://programbench.com/

this is the benchmark. codex can figure out how to run it from your side

[signal tapir]

I'm so glad there are people out there working not just on AI but how to utilize it properly

[lucid mason]

You could start with a Plus subscription to try it out and switch to the more expensive Pro if it convinces you.

[plucky trout]

I use it for work, and I'll hit its limits too quickly without really getting a sense of its capabilities

[plush harbor]

You can do quite a bit with plus. I have 40% left and a day to use it lol

[spring remnant]

I'm having same reconnecting 1/5 problem

[rocky dragon]

i have the same issue

[spring remnant]

Is all your file on ssd or main drive

[spring remnant]

My friend told me it's because my file on ssd, codex can't find it

He didn't experience such problem with file on his main drive

[rocky dragon]

all my files are on my pc

[spring remnant]

Is it put inside codex project

Folder

[cedar skiff]

It's better than claude at coding, it has been since 5.2, 5.2 and 5.3 are hard to prompt though because they dont get semantics as well as claude it seems like they are worse. 5.5 gets it though. Fun fact 5.5 is about the same in swe pro bench as 5.3.

I swapped at 5.3. The random dumb patches that claude has just got too much, there is no way you can trust it.

codex is rock solid

[plucky trout]

let's use GPT with opencode now, bye claude

[spring remnant]

How to fix reconnecting issue, I tried everything I can, still the same issue

But my other friend is not experiencing this issue

Seem like it's a 50 50 problem

I have exhaust all the method, re install, and power off my Mac, still not fixed

[lucid mason]

I had a lot of these using ChatGPT yesterday and it also started to happen again this morning. But now it works fine. My speculation is that certain server (clusters) are more problematic than others. Maybe starting a new chat helps, because it could switch you to another server (cluster).

[uneven kayak]

Alrighty, the agent task contract framework is nearly ready, just performing some final checks and tests

[lucid mason]

Great. I will present it to my chats for GPT to work it into my two projects. Thanks to memory GPT knows both projects well and can discuss/decide how well it fits into my workflow and project structure (like earlier mentioning that we already have authoritative instruction files for decisions).

Thanks for introducing me to contracts, that should help with improving efficiency of the workflow.

[uneven kayak]

No problem, it's my opinion that any project beyond a certain level of complexity should use agent contracts

Making some adjustments now, not too much longer

[wintry dune]

what kind of level is 10b tokens in, 700-800m tokens out in 1 month for a solo dev? running single instance only

[uneven kayak]

10 billion in, most of those are going to be cache. 1 billion out, that's hard to believe

[wintry dune]

yeah i checked like 70-80% cached for in

[uneven kayak]

That would cost like $1 billion lol. Edit: I'm wrong, it would cost $30,000

[wintry dune]

oh sorry it was below 50m out

[uneven kayak]

That's not bad but it is a lot for one person

I guess it depends on what you're using it for

@lean lark Did you put a trojan in your Codex useage repo? I strongly doubt it, this is probably a false positive, but Windows Defender found ClickFix.EE!MTB in your repo when I ran the report.

[oak trellis]

wf again .. ```Invalid prompt: your prompt was flagged as potentially violating our usage policy. Please try again with a different prompt: https://platform.openai.com/docs/guides/reasoning#advice-on-prompting

Most likely trigger in this thread: aggressive/abusive phrasing in the prompt text. When that happens, the API rejects the prompt before execution.```

low iq parser from openai

[vital hill]

please don't yell at your future ai overlord

[celest stag]

Does anyone have an issue with conversations in the Windows desktop codex app seemingly vanishing after they were created? Its not the first time. Not talking about archiving.

I just had Codex fork a conversation into a new worktree, gave it a plan to work on, and it started doing its thing. I clicked into another conversation and now I cannot find it anywhere. I didn't even close codex. Its not archived, I looked- and its only been like 10 minutes anyway. I was able to find a missing one the other day by right clicking the tray icon, but its not there this time.

[sand shuttle]

got models on github copilot are so bad 😭

[proper cosmos]

How do I add chrome to codex?
Is it randomly not available in the UK

[vital hill]

use a vpn

EU's privacy laws probably prevent it from working fo you guys

[proper cosmos]

so is it not in the UK then?

fair enough

[vital hill]

US only, for now at least?

[spring remnant]

Did you just connect to cloudflare mcp?

This issue started after I connected cloudflare mcp

I'm still trying narrow down the problem

[rocky dragon]

Yes, I recently installed the Cloudflare plugin.

[uneven kayak]

alrighty @lucid mason here you go #codex-show-and-tell message

[lucid mason]

Thanks a lot! 👍

[uneven kayak]

no problem, happy to help if I can. Feel free to DM me if you have any questions.

[lucid mason]

Oh, and you did quite a bit more than just share a file there. Nice work.

[uneven kayak]

oh yeah I turned it into a complete modular framework

pushing an update right now but it's inconsequential

[wraith spoke]

my codex has been stuck like this for 2 hours

[still trellis]

had the same issue earlier this morning, restarted my mac then left it on and it working again

[nova swan]

is the Remotion plugin still under development?

Because when I use it in Codex. It bugs out.

[opaque cargo]

how can i talk more to codex without giving more moneys

[lucid mason]

If you are interested I can post you GPT's assessment of how to incorporate it into my workflows. It's a lot of text, so maybe better via DM than filling the chat?!

[rigid igloo]

best way of learning codex? utilizing its full potential?

[lucid mason]

@uneven kayak TLDR GPT answer:

So: inspiration yes; installation no. The external framework is solving a real problem, but your projects already have the harder part. The best next step is a small native amendment to the existing plan contract, not a second framework.

[uneven kayak]

well you jumped the gun on me lol, obviously you don't need task contracts if you already have them. But you can still use my framework to augment your contracts.

[lucid mason]

I only have a single agent execution contract that I had GPT build after learning about them from you. 😉

But I already worked with a lot of maintenance files to handle things.

[uneven kayak]

oh wait a minute, that wording sounds singular. Did you start a new chat session outside the context of your existing contract task session? The major difference here is that my framework is going to apply to all future sessions, not just the one you already have a contract for.

well it says "the existing plan contract" but it also says "projects" with an s. so idk lol

[lucid mason]

Before maintenance changes, read:

• README.md
• docs/agent-maintenance.md
  - docs/agent-execution-contract.md
• docs/active-context.md
• docs/module-map.md
• active revision headers across maintained .ahk files

[uneven kayak]

ah nice

[lucid mason]

That's in my agents.md

Same for the other project I am working on.

I will now update both with the "inspirations" taken from your framework. 😉

[uneven kayak]

then yeah, you went ahead and did it yourself but I hope my framework is still useful to someone else. Or tell the agent "Determine if there is anything useful in the framework that we can incorporate into our existing plan contract."

[uneven kayak]

sounds good

[lucid mason]

Or rather have GPT do that. It checked your framework and then takes in what is useful for my two projects.

Changed:

Added native substantial-task contract fields to .agent/PLANS.md.
Kept contracts as routing/verification anchors, not approval gates.
Updated execution/maintenance/bootstrap/module-map references.
Added concise Task contract sections to:
docs/plans/keyboard-powertoys-import.md
docs/plans/ux-finish-and-hardening.md
Updated docs/revision-history.md.

[uneven kayak]

ah very nice, looks good

[lucid mason]

Here is what GPT avoided for my use-case:

Do not add:

docs/agent_tasks/INDEX.md
JSON task contract files for every substantial change
a second validator unless you actually want CI/task-contract checks
PR-template requirements for a repo that is not PR-driven
multiple agent adapter docs unless you actively use those agents
mandatory approval gates before implementation

Those would duplicate your current agent-maintenance / agent-execution-contract / .agent/PLANS.md split. SysToys already says shared agent behavior belongs in a small portable maintenance layer and project behavior belongs in README, module-map, active-context, and narrow owner docs.

[uneven kayak]

yeah those are all personal choices pretty much

looks like you got what you needed and didn't get what you didn't need, gotta love AI lol

[lucid mason]

Still a very useful addition, because the whole contracts thing might introduce a higher authorative structure than what my old maintenance files did on their own. I willingly keep burning tokens on making sure GPT/Codex keeps referencing them and adhering to them, but many things still slip through that should not.

[uneven kayak]

you'll be running those 5 hour sessions now, no problem 😂

[lucid mason]

But in general I already found that this kind of structure helps a lot against longer chats starting to rot. Many reminders keep the memory fresh (although Codex 5.4 with large context window still turned from senile to lobotomized above 500k context size).

[uneven kayak]

comprehension is helped a lot by having an actual document to refer to, not just session memory

I've been having pretty good results so far

[lucid mason]

Plans already help a lot with even making ChatGPT implement features/changes in bigger chunks. And if nothing else they allow to just post something along the lines of "Continue with the next largest coherent phase of the plan."

[uneven kayak]

yeah for sure, some of my plans are thousands of lines so they have proper sections and a table of contents with the actual implementation phases or steps almost at the very bottom of the plan, that way the agent is encouraged to read through all of the information in the plan before it reaches the phase. Then I tell it something like, "Phases 1 through 4 have been implemented. Comprehensively execute Phase 5 and don't touch anything related to Phase 6 or beyond."

[lucid mason]

If I hadn't build up this maintenance infrastructure over the last few weeks already then starting with your framework would likely have improved and speedup the whole experience a lot for me. It just came 1.5 months late. 😛

[uneven kayak]

well the reason I made it into a repo the way I did is so hopefully it can help other people who are just getting started

[uneven kayak]

oh I meant to reply to this

Open up a chat session and say to it, I want to use Codex for <whatever your reason is>. Formulate a series of questions to ask me which I will answer so that you can get an idea of what my goals are and what my current level of knowledge is. Then after I answer your questions, use the answers to formulate a comprehensive and complete guide to help me learn more about Codex to achieve my goals.

[prisma flare]

When computer and browser use in EU 🙁

[opaque cargo]

?

[velvet wren]

what plan are you on?

[uneven kayak]

He's on free, you can see "Get Plus" in the top left corner of the screenshot

[velvet wren]

ah ok, I didn't notice that

[spring remnant]

I was not able to fix reconnecting issue

Did oai say anything about it

[uneven kayak]

$20 goes a long way for a Discord bot if that's what you're doing @opaque cargo

So Plus might be enough

[stray tide]

use gemini gives 1mil context free

i switch between codex and gemini a lot

[velvet wren]

Plus has allowed me to build a fully functional macOS app for video AI and an iOS music streaming app, a Discord bot would be childs play and hardly eat into your usage @opaque cargo

[velvet wren]

we discuss non OpenAI products in #ai-discussions

[stray tide]

then the correct answer to the question he asked is simply wait for your context to reset lol

[opaque cargo]

On go

[velvet wren]

then to get more usage subscribe to Plus

or wait for your usage to reset

[uneven kayak]

oh huh I didn't even know there was a Go plan

[opaque cargo]

Ok brother so i will shift to plus plan 🙌 thnx

[uneven kayak]

I'm living that Pro life 😎 lol

[opaque cargo]

Yeah fs

[blissful basin]

first time i had to use some extra credits with pro, and ffs it is not worth it at all

Literally 10usd worth of credits is gone is 5mins

[uneven kayak]

oh yeah that's not worth it at all, using the API or extra usage credits after having all that usage for only $100 or $200 makes you really appreciate the subscription model.

[lean lark]

ABSOLUTELY False Positive. I'll look to see what's triggering that. Thanks!

[uneven kayak]

sent you a DM with the info Defender has, not sure how useful it will be

[unique spade]

Petition Ursula 😂

[prisma flare]

I will call her 🤣🤣

[unique spade]

I m in EU too when you find her number I ll call her too 😂

At least we have codex, and ai in general... Yet

[ember spire]

do 20x claude code users generally need the 20x plan on Codex when we migrate?

[velvet wren]

try the $100 plan then upgrade if it's not good enough

Codex is very generous, even the Plus plan goes a long way

[solid lake]

Im just scared that after the promo ends the limits won’t feel the same

[velvet wren]

what plan are you on?

[solid lake]

Pro

Pro lite

[velvet wren]

you'll be right

[solid lake]

But I ran out of usage on the promo limits

2 days ago 😭

Might just upgrade atp but not until I try the non promo limits ofc

[young locust]

how/what are you working with when exhausting limits so fast?

[solid lake]

Mostly cuda kernels

Some os management stuff but that’s very less compared to the kernel stuff

[young locust]

your tasks go through lots of stuff each time?

[solid lake]

Project specific I keep it to 3 docs at the most, agents md the skill and my living spec

And yes I’d spend maybe 2-3 hrs finalising plans before any actual coding through experiments and log driven dev

[runic rain]

Eventually, but there's not exactly a strong incentive from the few providers to do so. Infamously, nVidia's CEO indicated "the fact that everything is scarce is fantastic for us"

[torpid trout]

we have ourselves a dreamer 😆

If there's something that now generates profit that could be done for free, you will either first be bled out for it, and then sold the product that does it at a recurring subscription, or, you will never see the product reach any factory nor be deployed, ever.
Not unless something very inherent and deeply engraved into human nature changes.

That's like healthcare you know 😉
It should be free, it could be free, yet most of us pay for the most simple things outrageous amounts of money
(Hell, if tonnages of solar panels sent to space and regularly maintaned should be free... then health care should be like giving you money lol)

[solid lake]

It’s not just solar panels innit

[torpid trout]

right - compute also should be free 🤣

[solid lake]

Lots of materials science research has been carried out on ISS on this

[torpid trout]

yeah, nothing of which is nor ever will be free lol

[solid lake]

Not unless you get asml competitors and solve that bottleneck imo

[torpid trout]

they solved the bottleneck of about anything you pay nowadays long long time ago.

yet you and me still pay for it all

lets say water

[solid lake]

They didn’t

[solid lake]

I don’t see any asml alternatives popping up

On that scale

It was actually addressed by tsmc not long ago too

[torpid trout]

I do not think you understand what capitalism, or, better, human nature is in that respect

[solid lake]

I think you should look at labour and capital markets as a whole here

[torpid trout]

lol, ok.

[solid lake]

Unless you have things to add other than snide comments 🤷

[torpid trout]

Name me one product that is free nowadays. Just one.

You can be lucky the air you breathe is free, and that is about all that is free really.
Nothing else, that is there naturally, has tons of competitions, alternativees, be it tech or natural, invented or not, scaled or not, is free.

[solid lake]

I’ve seen people make whole websites from free ai

I don’t see your point

[torpid trout]

I do think you do confuse things at a large scale.

A gift is not a broad service provided for free by someone to anyone. Mostly, you will see those rare occasions spread here and there, and even rarer where there is no second thought behind it

[solid lake]

Boss this argument falls when you add open source to the mix

[torpid trout]

Open source that is mostly sponsored by large companies?

Like WordPress, powering 50+% of the web? (and secretly being owned by "angel" investors)

You mean that kind of open source? Or the sideproject of someone that has 400 stars on github and 30 active installs?
Or those that have thousands of installs and are not maintained, like the *arr stack?

good luck with that same approach in solar panels in space.

Anyway, I do not want to stomp on your dreams, really!

[solid lake]

More like Linux type of open source

[torpid trout]

ok, then you mean sponsored by huge companies

[solid lake]

Linux was a hobby project btw

No sponsor until later when people adopted it

There were volunteers worldwide btw

No external funding

It all comes out of need

[torpid trout]

which is why it did not yet fall out of the sky onto all of our heads, or, better, why it even managed to go to the sky - because folks with money sent their employees maintaining it for the biggest part
Imgine true OSS in space - each solar panel sent up there by the vibe coder of yesterday, who today has other prioroities

Things may come out of need. They stay because of profit.

Someone always profits. Otherwise stuff either goes away or becomes stale.

[solid lake]

??

No?

[craggy jewel]

Methinks open source and all it's crazy half/abandoned open source dependencies is on the way out. Mythos et al can find and exploit open source faster than we have a hope of fixing them. All my new projects top-line requirement is: No third-party libs. Back to the proprietary timeline we go.

[runic rain]

Can attest to the use of third party libraries being much more scrutinized in corporate environments, especially if they serve in any capacity as a contractor. Lots of SBoM requirements and security audits lately.

[torpid trout]

That's another issue, which i do not fully agree with (there are pros and cons to oss when it comes to safety)
Fact however is that unless someone makes a buck there's no bang.
That's literally how the world works even in the most non-capitalist societies left on this world
If its not money its something else. Scoping to the original topic of free compute and energy, we will first meet aliens before that becomes a reality

[craggy jewel]

Like i needed in a Winui c# project Dx12 compute shader functionality. Could have used some open source, or proprietary third party libs. But with Codex 5.4 High it wrote that exactly my specs and testcases and works like a charm No more wondering...do i need to update that thirdparty nuget package? Does it have any transitives that need updating? Are any of those vulnerable? No. Third-party libs are redundant when you can just make your own.

[solid lake]

I think you’re arguing against a much stronger claim than the one I’m making.

“Compute costs trend downward and access expands over time” is already historically true.

That’s very different from saying thermodynamics disappear and everything becomes literally free forever. (Which probably was the original post you replied to)

[torpid trout]

I do not dispute cost came down (1gb of storage costed something like 50k not too long ago)
I also do not dispute that through obsoletion of tech, invetion of new tech, and so on, costs will come down more, you may even get stuff on the dumpster in future that now we pay for
However, yes, original post said... something like Electricity cost and compute will come down to zero

Which is nonsense

To put it bluntly.

[craggy jewel]

And the real tricky thing with third-party libs is if they have established trust over time, and hey get hacked upstream in a package provider (npm, nuget etc) or even if the originating biz is sold with a valid code-signing cert to a bad actor, well then...not good.

[solid lake]

Yeah, There’s a difference between zero marginal cost tendencies, abundance economics, commoditisation and literal zero cost physics.

The first three already happen all the time in technology markets.
The last one is sci fi imo

[torpid trout]

It is - because even if you take a slave and make him think - you need to feed him, that means, in other words, no input no output
Its thermodynamics, no fuel no burn
So, free could only be achieved at the cost of someone or something else

(as in, we do it already, to be honest - we actually pump free oil from the soil. we are not paying for that oil, not yet anyway, in a sense that the true producer of said oil is being robbed, you could say if you stretch it a bit)

[solid lake]

The cost is less oil in a few generations

[torpid trout]

but that is just a contract we cannot see the end of, in our magnimous miopy

[torpid trout]

or none. and a ton of other issues as invoice lol

[solid lake]

Day 2 of asking for codex Linux app

[bright edge]

How to use open router free models or open code free model's in codex desktop app?

[spark bobcat]

It's possible to make a workspace context with codex? Like we do on VsCode, where we are in a workspace and the agent can work in multiple repositories? I need the agent to know the context of two different repos

[solid lake]

Enable memory

[ember spire]

mmk I think Claude Code is dead to me... nice one Codex! 🙌

[spark bobcat]

I really dont think it will help

Since i would need him to code in others repos too

[solid lake]

Who’s him

[spark bobcat]

Like sometimes i am working on the API, WEB and Game, need the same feature

[spark bobcat]

Mario

[solid lake]

New model

Hmm

[pliant wren]

Why do ai tools have dated conversion - like messaging app?

I did a conversation for a problem trying to fix something - ran out of limit at that time - and now I know the date when we discussed this but can’t easily find that conversation in codex (or any other ai tools)

What’s the reason for not date time stamping the conversation with the agents?

[radiant pecan]

is there a way to see how much tokens i used up the whole month in total?

[velvet wren]

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

[simple star]

6... more... hours...

[solid lake]

this doesnt give token wise right

[velvet wren]

no, it just shows you relative usage

[solid lake]

im confused what does this mean, the right bar is avg requests i make?

and left the amount of different chats?

[cedar skiff]

turns are conversation back and forth, threads are conversations

[solid lake]

its more back and forth than just me

yeah

makes sense

[karmic gulch]

You’re out of Codex messages

those parallel agentic runs a few hour before reset so that you leave nothing unused is stressful. Hate to leave % open before resets 🤣

Even tho its not easy to use it all on 200 plan

[simple star]

This week I will turn off /fast 😛

[torpid trout]

I am very sure I will not ask again what is prompt, and what is response in the log, from now on, thanks codex!
What I might ask tho is to add it just one more time, because obtuse, you know.

[plucky halo]

Daybreak! Aaaah! Fighter of the nightbreak!

[unique spade]

managed to get down to 3% with 36 mins to go till my reset . finally a week i optimized usage and time

[vital hill]

i used about 5% then it reset before my eyes

[plush harbor]

mine resets in almost exactly 24 hours

[radiant pecan]

please just work...