codex-discussions

[bright swift]

its bad, its api pricing. just daisy chain pro 20x accounts bro

[unreal parcel]

if 5.6 really comes out that would be crazy. poor anthropic

the problem with pro subscription is the small context size

[bright swift]

huh?

[unreal parcel]

huh?

[brittle stratus]

they will PR a Super Mythos, unable to release due to how dangerous it is,

[bright swift]

oh you mean for claude? its lower on pro than max or what?

[unreal parcel]

for GPT 5.5

subscription gives you 400K - 128K context size, 1M is only for API

272K is so easy to hit doing some stuff like reviewing a commit range

[bright swift]

oh yea, but i'm happy with the autocompact in codex, context size isnt an issue

gpt pro in chatgpt has 1M pretty sure

[lost drum]

I guess I can use plus for 3 days and then wait cause idk how to connect second account to same thread

[turbid axle]

they are really squeezing tokens huh. running out of tokens way faster this week as compared to last week... 🙁

[unreal parcel]

I tried a few times letting the auto-compact work it out. the review with compacting ended up compacting like 4-5 times and barely found any problems. rerunning with multiple agents instead gave like 5x more results. It's really not that great

[brittle stratus]

how do you guys deplete a 20x acc, what do you do? do you run it overnight and use /goal constantly?

[unreal parcel]

I can't even dip below 50% anymore, even using Fast lol

[turbid axle]

I could run /goal all day and it would eat about 10% of my weekly. now it eats 30% and the day is not even over

/sadface

[lost drum]

previously I used 5.4 xhigh and was running it 24/7 there was no goal feature I just made him emit promtps for himself and I made script that just checks when he finishes and then send the pormpt to him from the specifc file

[bright swift]

pretty easy. 5.5 xhigh, lots of feedback loops like codex hooks, git pre-commit/push hooks, codex review loop, etc

[cedar parcel]

I deplete two of them in 5 days. And yes they run 24/7 fixing issues in issue queues.

[unreal parcel]

running models for 24/7 is nuts

[lost drum]

nahhhh

done it for 1 month straight

[turbid axle]

you mean, not running 24/7 is nuts

[cedar parcel]

It can sure drive you there

[lost drum]

I need 1 more month tho

[unreal parcel]

I guess it more drives itself into slopville

[bright swift]

:((

[cedar parcel]

No this runs in enterprise environments but it's a whole process and does involve humans in the final stages

[unreal parcel]

I mean in enterprise environment $400 is nothing so just get another one 😂

[bright swift]

burning through a 20x in 1,5 days now, basically 5h limits are becoming a problem

[cedar parcel]

I have more

[turbid axle]

I actually hit my 5hour on my 200pro just now, I've never had this before, ever.

we are being squeezed hard

[cedar parcel]

But at 400 you can run a business week full on with multiple projects

[lost drum]

nahhhhhhhh

[turbid axle]

AI becoming more and more paytowin

[bright swift]

even with the squeeze the value is ridiculous

[unreal parcel]

Yea OpenAI, stop providing us the value of 10 engineers for $200 a month

[turbid axle]

its still good yes, but this is not a good direction, it should become cheaper, not more and more expensive

[bright swift]

i mean it sucks if you are starting out now. but if you have a business already or good job, its crazy

[lost drum]

skill issue

[unreal parcel]

maybe the people running the model 24/7 are the ones making it more expensive for everyone else lol

[cedar parcel]

That said we have to probably rever about 200 commits of total clown code now that whatever DevOps idea they decided to try on Friday is coming apart

[cedar parcel]

That's what the 5 hours are for and the limitations on using it for data processing

[brittle stratus]

have you built anything this way?

[lost drum]

ye but needs 1 more month to finish it

but not heavy tasks as before

[bright swift]

i'm actually bottle necked by planning right now

need more prompts

[brittle stratus]

ah, well, I still haven't fully automated my pushes and merges, and I don't yet use hooks. Any must-haves?

[lost drum]

wdm?

[brittle stratus]

do you have a product out or do you work in a company?

[brittle stratus]

I'm bottlenecked by tracking all the things I'm automating, I may need to automate my automating

in other words, say a thing has 20 steps, gotta make sure all 20 are properly automated

[bright swift]

what i built is basically a "local MR" workflow. i run lints, tests, etc the usual stuff on git pre-commit. then rebase, then on pre-push i run a codex review loop until all findings are fixed, each in a separate commit. so this local commit stack + review loop basically acts like an autonomous MR. i really like it. its a token incinerator though. beyond that i use pre/post tool use hooks in code for fast feedback loops, so lints, formatting, etc

[turbid axle]

Lazy reply.

I use the exact same workflow as last week. yet token use is at least double

[cedar parcel]

In a company, client projects and internal projects. It runs everything. For new projects it usually builds for a few days and then it takes a backend dev and front end Dev both a week to get it to be perfect. And then there's about a day and a half of autonomous pen, UI, API, and performance testing. But no matter how complex, this is consistently working.

[brittle stratus]

I do most of that stuff as well, on Plus account. Burns a bit of tokens, but I have subagents and the loop doesn't deal with trivial things or it may never end.

[bright swift]

basically refining the high level things i want to do. the implementation is so fast that i cant keep up with new goal prompts 😄

[brittle stratus]

this is my issue as well

and I think most people have or will have this issue

Need more agents, not more prompts I'd say

and abstract more and more

eventually you'll be so abstracted you will have freed yourself all dev work, unfortunately

[lost drum]

but dont you have one goal? cause I had multiple goals on diff layers and ye

[bright swift]

yup, we will all be PMs / product owners

[lost drum]

I know I also had him update the goals based on what he discovered while executing actions from goals

[bright swift]

i dont mean on goal level. i mean on the high level feature / spec / requirements level

i work alone, so i dont have anyone feeding me tickets or similar

[brittle stratus]

well, at least we'll be able to solve issues and gtm quite fast, after all, building is what gives great satisfaction, even if it means just a week or so in the future

[bright swift]

i tried to get gpt to do the grill-me questions but only ask me the really high impact ones and make sensible default decisions for less important stuff, but not working as i'd like it to yet

[lost drum]

You might need to give him good examples of high impact qustions vs low, and what to do with both like how he needs to ingest your answers or smthing

It works for me but idk how so I cant help sadly

[bright swift]

yea will need to do some experimentation

[brittle stratus]

I think one way to approach your isuses, is to define your abstraction level on where you want to be with your agent,
plan the highest level, then have the agent help you with defining what is left undefined on that level.
After that, just have the agent go down in the lower abstraction level and define there

sounds beginner-like what I said above, but sometimes the simplest point of view is the best one
i get bogged down by all the things I have to do, as I work solo like you. Good organization is critical for me

and yes, automate everything instead of leaving it for later. Learned that one in the past few weeks

[bright swift]

the good old building the meta system trap 😄

so tempting to go down the automation rabbit role too far

[brittle stratus]

what do you think the danger is?
I've got myself planning various social media strategies, gtm, outbound/outreach, not to mention features left to be done
The point is to free yourself from manual work and energy-drainers, especially the tedious work

[bright swift]

yea absolutely agree

for me the trap is spending more time on this than on my actual business

its fun working on these systems and improvements

real life factorio

[brittle stratus]

I only do it to keep moving forward and actually do the things that bring value. I don't do it for fun, seriously

I like how much I cann do, but it's become very difficult to juggle all these areas I'm doing, especially things I've never done in my life

[brittle stratus]

the automation is the actual business (not literally) but for example, I've automated my outbound 90%, this is like my main path atm
there wasn't much fun in it 😄

I hate spamming people, but I've worked for a leadgen company, so I had to take advantage of that

interested to hear your experiences, as you mentioned you're solo building? @bright swift

[boreal holly]

Idk, on one hand it takes time, on the other hand you learn valuable lessons. There is value in trying.

I rebuilt my rear axle twice. The first time I got RTV silicon in one of my wheel bearings and found out the hard way you can't really clean that stuff out. Rebuilt the second time with all new bearings, still going strong after 14k miles.

Sucks having to keep doing the same repair over and over, but now I have all the tools to do a rear axle rebuild and experience to go with it.

Build the harness. Do it for learning and potentially high rewards!

[bright swift]

i have been running a business that is a mix of saas/ecom for 10+ years, everything hand coded before obviously. its pretty stable and basically runnign on autopilot despite being relatively complex with custom on-demand manufactoring etc

so funnily in the last few months i mostly worked on greenfield new idea type stuff. now i'm applying some of the stuff i learned in the time to the existing business.

mostly painful unfun stuff right now to get it to a more "AI first" point with better integrations etc. migrating gitlab -> github, migrating off AWS, big refactorings 😄

[brittle stratus]

how do you promote? You mentioned it's on autopilot, so I presume SEO is working well?

I suppose we shouldn't turn this into this kinda of talk, as it's supposed to be about codex

so let me mention the word 'codex' once more 😄

[bright swift]

paid ads mostly, which is also my focus for new ideas right now. building some internal tools and maybe getting them to a place where they could become a saas. also taking over a marketing agency soon together with a business partner, so that aligns well.

im building it all with codex so... 😄

[brittle stratus]

no shortage of money I see, good luck with that

I am unfortunately forced to find all possible means to get traction, from social media to SEO (have some experience in this) to whatever is necessary

hopefully something clicks

[bright swift]

funny thing about that is that i just saw chatgpt + american express now has some kind of partnership. maybe i can spend the amex points i get for ads payments on codex credits. infinite codex glitch

[brittle stratus]

and given how little experience I have being my first ever SaaS product, I'm bound to make all kinds of mistakes 😄

[brittle stratus]

anything is possible

[uneven forum]

I am trying to multi select various elements with annotation mode and it doesnt seem to be working by holding shift.

I swear i did this before (i think haha) and when i read up on it, i see docs saying i should be able to select multi things with annotation mode while holding shift and selecting. but this does not work. it will only allow me to select one item. And I am not meaning selecting by dragging either. I need to select elements in various parts of a page that a drag selection would not work with.

Anyone else with this issue, or have any solves? or is it just me imaging something?

Thank you in advance.

[deft vale]

Does anyone know why they removed the slides and spreadsheets plugins from Codex? I keep seeing people with them installed, but they don't show up in my marketplace.

[uneven forum]

weird. I just checked and they are there for me

[deft vale]

Oh

I'm from Spain, could it be restricted due to EU regulations?

[heavy rapids]

anybody manage to get Hermes to work with codex?

[deft vale]

It's not just these extensions. I see people using skills and plugins from the marketplace that don't even show up for me to install.

Even if I go to GitHub to look for the "Slides" skill, I get a 404 error.
I figured I wouldn't be able to see a lot of them because they might not be available in the European Union, but it doesn't make sense that I can't see the spreadsheets or slides plugins.

If anyone here is also from the EU, I would appreciate it if they could check if they can see these plugins.

[bright swift]

did you ask codex to debug it

most likely some stale cache setting or whatever

[deft vale]

Tbh I hadn't thought of that

I'm going to give it a try, thanks

[turbid axle]

wow. almost at 50% token use now. thats a 40% increase over last weeks running cost for 1 day with the exact same agent/prompt

I guess my /goal goals are dead now?

just me noticing this extreme decrease in usable tokens for sub?

[torpid trout]

They are all here

No chance I can use them - stuck on 97% weekly since yesterday morning, that is a 3000 lines of code wp plugin later, inclusive iterations, plus two paralell tasks running basically nonstop. the 3k line thing was all done with 5.5 xhigh, the others on high.
It just wont go down

ca 12 hours straight of codex producing and chewing stuff. I do not run it overnight... but still. Sum tin wong

[turbid axle]

at this rate I will run out of tokens in just 2 days. ill have to buy a new 200pro account every 2 days now? just a single agent running 5.5medium regular. only last week I could run the thing all week and still code on the side.

wack

was even running it at night last week. no chance I can do that now

[torpid trout]

Key differences I think:

• I do not run native subagents. It is disabled
• I do not use skills, agents, or the likes
• All over codex app server in custom tool

And, yes, this is a pro account 200 bucks. But I am seriously thinking of downgrading it, because I literally can't spend all tokens as hard I try
The only reason I kept it for now was no other option, but now there is the 100usd plan...

[turbid axle]

chatgpt is the agent btw..

something seems off

literally the same prompt, same project, same model, same everything. and yet what worked for the whole week, night included with tokens to spare, now can't manage 2 days

[torpid trout]

Fairly sure that load balancer of theirs is the cause
Its not a first and wont be a last

I use tokens, they get billed to you, that kind of issue.

[turbid axle]

something funky for sure. this is extreme

[bright swift]

i burned through 1,5x pro 20x weekly limits since the manual reset

guess this is the week when i set up a 3rd 😄

[turbid axle]

2kpro the new meta?

[bright swift]

yes please

tibo tweeted about it

i think they are considering 1k+ accounts

[boreal holly]

Weird, I've had no issues

[turbid axle]

seems like some people have what feels like endless tokens, and others have to settle for scraps

[boreal holly]

I definitely don't have endless tokens, but very deterministic usage

[turbid axle]

maybe is based on how much love you post on x for oai

[bright swift]

definitely ways to be more efficient

but tokenmaxxing is more fun

number go up

[turbid axle]

thing that makes this so iffy is that its the exact same prompt and all that

[boreal holly]

Yeah, I try to offer advice on more efficient usage. A lot of it falls on deaf ears tho

[trim rapids]

I haven't personally seen any evidence of it being broken currently but it has definitely been broken previously within the last couple months

[turbid axle]

effecient or not. the same process should not eat 5x more one week to the next

[boreal holly]

If you updated codex (TUI and/or desktop) then the same process can definitely eat 5x more tokens

[bright swift]

you mean because of rewriting caches?

[turbid axle]

why specifically updating codex? did they change anything related?

[torpid trout]

What is to denote is that robert also does not use most native codex things, but runs custom harness
But the statements "same exact stuff now does consume more tokens" speak towards their infra being at fault, and this is not a first time, it is openly been admitted by Tibo that their cdn infra does mix up token usage

[boreal holly]

It could be literally anything. Like if you have unified_exec=true enabled, they might've updated how that works, so when the agent wakes up every 5 seconds to check on background terminals it consumes the entire command output + new outputs instead of just new outputs. Totally plausible, and would n^2 the input token usage

[torpid trout]

There is a github ticket about it where he explicitly said that, and if it happens once, it is almost sure to happen twice

[turbid axle]

I mean, anything can be a cause. there is a sea of variables at play here

[boreal holly]

Right, and that's the bummer about updating everything immediately

[bright swift]

huh unified_exec is such a burner?

[trim rapids]

a couple months ago the accounting was briefly so messed up that they gave me a whole extra manual reset after I posted my user id and some evidence of breakage in a bug report on github

weirdness definitely happens

[boreal holly]

MASSIVE

[bright swift]

thats for the background commands right

[torpid trout]

just look at this channel for proof of "weirdness happens" lolol, and I am not referring to codex

[boreal holly]

The idea behind unified_exec is amazing, but in practice, the way they implemented it has historically been bad

[bright swift]

i will investigate and potentially tokenmaxx less

[boreal holly]

Unfortunately, if your agents run commands that take longer than 15 minutes, unified_exec=true is the only vanilla option for letting that command run past that limit

[trim rapids]

yea you aren't wrong

the other day I saw a 69420 joke go over someone's head and briefly wondered what f'n planet I was on

it was highly disappointing

[bright swift]

hm ok thats a problem

[boreal holly]

Yeah I posted a proposed solution on https://github.com/openai/codex/issues/13733#issuecomment-4441800902

Would let you declare certain commands as blocking, so the agent can't reawaken every 5 seconds and burn tokens, but idk they don't seem interested

[bright swift]

thx i'll follow that issue. i assume you implemented a solution in your fork?

[boreal holly]

Yeah, I patched zsh_fork so there is no 15 min limit, and the agent can't pass their own timeout. And basically just don't use unified_exec

[trim rapids]

I've had some success just telling the agent how often to poll, like I can't say for sure that it works but if I send a followup message it definitely doesn't get processed until after the specified delay

problems seem to arise if I tell it to only poll every several minutes and it happens to hit a backend compaction point while waiting though, where it errors out with a message about my oauth token being revoked, deletes my auth.json and forces me to do the whole 2fa song and dance again

it's highly suboptimal

[boreal holly]

That was just 1 specific example. A long time ago before they switched to token-based quota, responses_websockets used to completely wreck the quota because they were doing "per-message" quotas, and with HTTP over SSE they considered a single HTTP request 1 message, but with websockets you have this persistent connection, so they considered every event a message lol or something like that. And I'm pretty sure they default enabled it at one point.

They could've also made memories=true default enabled, or made it so skill headers get inserted into each and every tool call mid-turn, or any number of quota eating changes. I think when folks say "Today I burn insane tokens doing the same prompt I did yesterday", first suspect is software update

[vernal palm]

Hey, I’m a developer working mostly on AI-powered web apps and automation tools.
Recently I’ve been building internal chat assistants, workflow automation systems, and dashboards that help teams manage and understand their data more easily.
Curious to meet others working on similar projects or experimenting with AI tools

[torpid trout]

Sometimes I just dont get codex or better said, llm models
I achieved a given goal safely and easily with 53 lines of code crafted by hand.
I asked gpt to write code based on that working example, but with a test suite around it, and trying to optimise the working example in case he can find a better way to do it

10+ files with each 50+ lines of source code (no tests included in these numbers) later, the goal is not achieved, at iteration 5 meanwhile (iteration 6 going now with the command to "fully reset and start anew")
At the same time it can poop out 3000 line of code apps that work in one shot and are really complex in the features.

Its weird. Like, as others said, sometimes the thing just feels dumber than a glass of water that accidentally can print toothbrushes out of thin air

[steady vigil]

but did you define 'optimize'? you can optimize anything. maybe it was optimizing execution speed which could involve more code. smaller is not always faster

[narrow vine]

guys how and where to use codex's "/goal"?

[steady vigil]

running through long sessions that could be easily derailed down a rabbit hole. keeps it on track. complex sprint execution with a lot of gates for example -- its great for this.

[trim rapids]

https://developers.openai.com/cookbook/examples/codex/using_goals_in_codex

[torpid trout]

yeah Believe me I did explicitly plan with /plan mode and its well stated to not re-invent facebook
I am not opposed to more code, etc, but hell, at least have the feature I did with 50 lines working, no?

lol

[torpid trout]

specially given it has a working example, where the (admitted) extreme edge case of this project is well documented and implemented
That is what it is porbably tripping it off because it is not "by the books". yet that is also explictily elaborated in the plan lol

Its like it goes circles hunting its own tail when it has too much guidance, kind of

I might just tell it "optimise this", actually, lol

[torpid trout]

bits are free!

[boreal holly]

/goal optimize until more better haha

[torpid trout]

append duh

and introduce a typo, becuase.

oooooh nice, a phisicyan appeared!

aaaand gone 🤣 not even enough time to get their username 😢

[rocky fog]

have fun with it
there are tens or hundreds of these coming to this discord per day
but they get insta deleted here (he got a 3 day timeout, but usually these get removed/banned later)

discord in general has extreme amount of scammers/bots and it only seems to be increasing, especially lately

some might also just be hacked accounts, as they are older

[indigo robin]

can you force a context compaction ?

my context window keeps compacting but its not returing to zero tokens after

not sure why

[torpid trout]

Whats weird - if you use project_doc_fallback_filenames, then agent says:

 Instructions came from your first message in this conversation, not from a file currently present in the repo.

  You provided them inline under:

  # AGENTS.md instructions for /path/to/project

  <INSTRUCTIONS>
  ...
  </INSTRUCTIONS>

  I then checked the filesystem and found no AGENTS.md inside this repo. The only AGENTS.md found nearby was outside this project:

  ../../AGENTS.md

  So for this repo, the instructions are active because you supplied them in the chat context, but they are not currently stored as path/to/project/AGENTS.md.

yet a file is located in cwd, just not default name, instead, a project_doc_fallback_filenames overwritten one
I observed it does not follow that as nicely as AGENTS.md, which is why I got curios

[torpid trout]

/compact

[indigo robin]

wowzers

sweet

[torpid trout]

and it will not go to zero, that is what compact is for, to NOT go to zero

[indigo robin]

mmm

i see

@hard drum

quick question

buying 20$ in credits vs making another codex account ? what happens to the reset cycle if you buy credits ? will it reset , would it be wiser to make another account ?

[rocky fog]

nothing
you use credits when you have 0% either in 5h usage or weekly

[indigo robin]

if i bought 20$ in credits

will my weekly usage reset ?

[rocky fog]

they will start to be used when you reach 0%
nothing will change with your reset date

[indigo robin]

then id be able to use codex after its hit zero

[rocky fog]

yes with the credits

[indigo robin]

and what about my daily limit it can goto zero too and id still be able to use it

[rocky fog]

yes, same

[indigo robin]

and itd use my credits instead

ah

i see

you created the empires mod ?

[rocky fog]

thats pretty much the only way you can use credits as far as I know
when usage reaches 0% the credits will then be automatically used
until your next reset

[indigo robin]

mm i see

[rocky fog]

didnt create it (Krenzo did)
I just host gitlab/development and other stuff and assist devs and make sourcemod plugins or discord bots and similar for it

[lean lark]

I read that yesterday - good intro to how to use goals that most people will ignore anyway.

[indigo robin]

i see

[lean lark]

I was just over in #openai-chatter - just as I remembered, it's home for kids and newcomers and conspiracy people, for FUD, crying, misinformation, and a hint at the world of people who use the tech.
This channel gets like that sometimes.
I wish there were channels for verified adults and those who have a handle on this stuff. 😞

[pliant depot]

Hi everybody ! Do you also have the impression that when talking about limits, new 100$ pro plan is the same as the old plus plan ?

I would like to use gpt 5.5 only on a large repository. Should I really buy the 200$ one ? 🥺

[lost drum]

If it turns profit then why not?

[lean lark]

The question can't be answered properly in a forum without a lot of info. The decision is based on your usage.
Try Plus. If you run out of tokens, consider a couple options:

1. Monitor your token usage:

• Are your prompts good?
• Are your AGENTS.md and Skills carefully crafted and not contradictory?
• Can you use Standard speed or do you truly need Fast?
• Are you switching models and reasoning ability based on the task?
• Are you watching the CoT and remedying issues that the model encounters?
• Is your code documented so that the assistant doesn't need to read the entire workspace to understand it?

2. If you've done all that and you know you're already using tokens effectively, maybe you do need a higher plan.

3. If you can cost-justify a higher plan as a business expense that is actually related to profit and loss, then yes, maybe a higher plan is the next step. If you are not earning money with the tooling yet, then no, maybe spending more money isn't the right answer yet.

If you haven't come up with some metrics that signal a higher plan is well-justified, you're just throwing money at something you don't understand. That's a different kind of problem that no one in a public forum can help with.

[bright swift]

lol reset again?

[bright swift]

im back to 100% ahahah

[solid lake]

Now btw

[bright swift]

was just about to buy a 3rd because 2 subs emptied

[cyan gyro]

hahaha awesome

[bright swift]

maybe they actually fixed sth this time because it really seemed to drain fast

but ill take it 😄

[cyan gyro]

meanwhile google is doing their best to re-invent glasses...again...

[orchid echo]

RESET???

LFG

[karmic gulch]

Amazing stuff,glad I went turbo and used alot of usage, betting on anothe reset 😎

[lean lark]

Hey @torpid trout if https://bitsforfree.com/ really is focused on "bits for free" then maybe the meme on optimization isn't a good fit?

[orchid echo]

i just wish Codex would copy Claude's /copy functionality

Codex's /copy is useless

[bright swift]

what does it do

[torpid trout]

the domain was just some squatted thing I bought years ago I had lying around lol

[lean lark]

I was just looking again and it seems it needs to be general purpose. If limited to our little friend Homo Reseticus, there wouldn't be much there. What do you think about vibing a categories mechanism?

[solid lake]

Google glass was amazing btw

[orchid echo]

Claude's /copy? its smart, lets you be able copy certain sections, so for example you can easily copy strictly only a command without all the fluff and backticks

[bright swift]

oh yea thats nice

[lean lark]

Lots of AR/glasses on the market now, extreme quality/price differences, very chaotic.

[bright swift]

always so dramatic lol

[torpid trout]

I am open to anything, lol, the site is for fun.
I did not want to add menus initially as I found it disturbs the design, but again, no issues with it
The repo is linked in the footer, contributions welcome! Its not on github tho it is git(ea) right now, so if you think it is useful I can move it over to git (I would mirror it, I think that works, or just fully move it)

Right now it accepts any type of content really, while it is AI moderated, that is only used to avoid profanity and stuff like that (it then puts it into an admin queue for me to review)

[outer trail]

I just updated codex CLI and the /fast slash command disappeared?

any ideas?

[karmic gulch]

Codex for mobile is not living up to excpectations... so buggy... going back to Omnara, thats one perfect hidden gem. Anyone else got bugs with codex on mobile?

[native vigil]

wow codex became much better after reset again!

[cyan gyro]

how do you know?

[lean lark]

Hmmm, so if the image is already being processed by AI then it can also categorize. That's good enough. I see moderation.js in the code. I'll clone and see what I can do as time permits (rarely does). Best to ya!

[torpid trout]

what the hell why is my codex resetted again

[torpid trout]

yep, it totall can

[native vigil]

its just fixing all the broken prs it did over the last days xD

[torpid trout]

honestly I wanted to add a "vibe code this site" feature to it but its just too dangerous
Like, a small chat boy at the bottom anyone could use to add features 🤣
but I do not trust ai to recognise prompt injections and adequately allow only actual true features lol.

[bright swift]

wen reset

[torpid trout]

it apparently did already, I am back at 100% again

[lean lark]

good call, Beda

Hmmm, you just invented something (again).
Dynamic websites that visitors can vibe/configure. The site uses their preferences when they visit. Config is stored in local storage. Needs to be carefully considered...
Call it a "BedaSite" 😆

[torpid trout]

Yea, I mean, does it surprise no one that we still think in terms of "settings", "configurations", "apps" and so on?
The future has no website, no OS, no app, no settings
It has an interface with an AI, and you tell it what you need, and it makes it for you, pulling from a pool of data
Basically facebook may keep existing, as a json blob somewhere, or whatever. You show it to yourself the way you want, where you want, when you want
The entire OS is a "today I need to communicate with another human, tell them this:..." and the ai decides whatever the best way is, it may or may not create a mail client for that, or god knows, something better

Who cares if the user wants dark or bright mode, maybe someone wants rainbow mode... or whatever else, read diagonally or so.

[cedar skiff]

There needs to be some major break throughs in efficiency before that happens.

nice to wake up an see a reset

[lean lark]

I try to share concepts like that with my wife. She kinda doesn't get it. She's a "user/consumer".
Dude, let's get married! 🤣

[cedar skiff]

i was at about 35%

i gave 30 tasks to codex, it always does this 🤣

[torpid trout]

local inference is the only that can truly achieve that I think, but I guess that is just a question of time until the models become small enough and tech advanced enough

not a tomorrow thing of course.. but porbably a soon-thing

[lean lark]

we're quickly moving toward locally hosted models so personal apps will certainly be common, very soon.

[torpid trout]

actually, nothing has that in part

they basically allow you to create your own phone apps, directly on the phone

not sure how, i think it is some remote call to their models, I doubt a local phone-model can do something like it, but then... phones are incredibly powerful nowadays, so given a strong harness probably it could do some limited stuff already

nothing as in the brand

[lean lark]

I have Gemma4 on my Android. And I have an APK generator. I just need to wire up the harness to create AI apps for the phone. Never enough time...

[sharp yew]

chiefs for the past couple of days am i the only one that burns through quota as if im using opus instead of gpt?

im cooked already

[rocky fog]

openai making us all work hard with these resets 😄

[lean lark]

Drex, see if https://github.com/CaptainStarbuck/codex-usage will help you to see where your tokens are going. More enhancements are on the way to help us to focus to those kinds of answers.

[torpid trout]

hope it hurts not 🤣

[tiny laurel]

I’m curious whether Codex can be used with API keys from other models, like Gemini. I’ve really enjoyed Codex so far, especially the features, skills, and built-in tools, and I’m wondering if a similar setup or workflow would be possible using the Gemini API.

[lean lark]

it can

Many people in public are happy about using other models with codex.

L:ook around for how-to info.

[tiny laurel]

I've been doing some diggin and see there are framworks like liteLLM that can help with that, just wondering if anyone here knows if there is a builtin tool or we do need a workaround like litellm

[lean lark]

I don't know but I can tell you that many people outside of this forum have already published that kind of information in articles, blogs, etc. I personally haven't had the time/desire to read on that specific topic yet.

[boreal holly]

Yeah you need a workaround to translate other provider API calls to /v1/responses. Also set up a model provider in config.toml to point to the workaround

[tiny laurel]

gotcha, will definitely try that out

[lean lark]

Hey, please come back and share what you've learned. 😉

[karmic gulch]

been experimenting with claude dispatch taking controll of my mac mini and running my codex projects last 8 hours, just amazing. I check in every now and then, and claude and codex have their own bbq party and getting those steaks medium rare. Were about to cancel my claude sub, but now tht I truly can have opus be "me" ill keep it, and let them cook together

anyone else doing this flow?

[lean lark]

Peeps are doing this with the codex tooling for remote control, or the app-server. I saw something published by staff yesterday that their explicit goal is to allow different systems to communicate and execute just like that.
Oh, we're in for a world of hurt. 🙂

[sharp yew]

will look into it, whats up with all the codex resets im seeing? sama forgot about me 🙁

[lean lark]

Was only one today.

Ref #announcements "OpenAI Guaranteed Capacity" : That's what I was talking about yesterday when I mentioned dedicating servers for specific usage.

[solid lake]

About to get stiffed

[tidal pivot]

Anyone else feel like gpt 5.5 now runs up your 5h limit very quickly ?

[torpid trout]

dont want to know what that costs 🙁

[sharp yew]

sama hates blacklisted me confirmed

[lean lark]

No announcements were made by Tibo - it's possible your account hasn't been credited yet. Or yea, maybe SamA doesn't like you.

[sharp yew]

i wouldnt either if i was in his place

i troll him with each and every occasion

[boreal holly]

I freaking love it dude. The worst thing ever is when they redesign a page and rip out real existing functionality. "Make no mistakes" is real now

[lean lark]

Except that you're working with Twilio, one of my favorite companies, I don't understand what's happening there. Are you saying it caught and corrected a mistake because of the directive to "Make no misteaks"? 😜

[boreal holly]

Yeah, the worker claimed to have redesigned the page, but ripped out functionality that existed in the old design. Review agent is like "yeah buddy, put that back in"

[torpid trout]

not sure if you want to challenge the anti-phisicyan association or truly want to get removed. Be it whichever, f-o.

[lean lark]

I understand it now. Well done! 👏

[inland sonnet]

i'm sharing a free be

[inland sonnet]

imagine deleting knowing that is how AI scrape alot of data

😂

[bright swift]

what is this phisicyan stuff you guys talk about lol

[lean lark]

New topic: It's helpful to go back and look at AGENTS.md files. As the tech improves and my/our understanding of how things work improves, there's a lot of opportunities for tuning these static prompts that are easily forgotten.

[lean lark]

I swear, I don't understand what he says half the time...

[amber scaffold]

Has anyone measured the retail price of inference provided in Codex subs? I know Claude 20x is ~ 5k in inference

[bright swift]

probably more, at least 10-15k, maybe even 20

[lean lark]

That's already an enhancement scheduled for https://github.com/CaptainStarbuck/codex-usage - maybe this week. Will include estimates of costs per session and of available quotas.

[boreal holly]

There once was a dude selling shared codex pro subs and his name was like "physician" or "PhD" or something, so anybody being sus is a physician 😂

[inland sonnet]

some ppl in this chat are losers

like damn

[silver smelt]

Why does OpenAI keep resetting limits

Just seems sus two resets within 3 days

[fleet geyser]

Did you get one today?

Mine's not reset yet

[cedar skiff]

mine was reset

[silver smelt]

Yeah I did

[fleet geyser]

hmm

[silver smelt]

I’m not complaining I was low and actually worrying if I’ll have enough and need to upgrade

[fleet geyser]

I've used 50% in two days so i'd love a reset lmao

[silver smelt]

Yeah dude usage def got slashed because im burning through usage now

Wasn’t the case 2 weeks ago

[cedar skiff]

is it only he 5 hour window?

[boreal holly]

Yeah I just noticed the 5hr window is shrinking at a faster rate. The weekly looks to be depleting at the same rate though

[cedar skiff]

this is what mine looks like atm - around 5% 5hour = 1% weekly

[boreal holly]

Yeah you're right! That seems to be the scaling factor

[plush harbor]

the graph under that one is broken though so you can't see your per day use anymore

[cedar skiff]

Oh it is too

[turbid axle]

I guess we got another reset again?

[lean lark]

Before I go and create a utility, doe anyone have a process that assembles AGENTS.md files from components. Like - give it a recipe and it produces the AGENTS files for a given system based on OS, Location, Purpose, etc?
I'm still cleaning up these AGENTS files on my way to Skills. 🙄

[ivory zodiac]

what up

legends

[lean lark]

haven't seen this dude in a while

[ivory zodiac]

been busy as heck

building something fun

[boreal holly]

Closest thing I've made by what you're describing is codex -c developer_instructions="$CONCATENATED_INSTRUCTIONS"

[ivory zodiac]

Can't wait to share it

isnt that animation sick

are you guys noticing a degradation in performance with compaction since 5.5?

[lean lark]

I kid you not, I've been thinking about doing something like that for days. Different character, (not You 🤣 ), but just getting my chops with sprites and canvas.

[ivory zodiac]

also i have noticed multi agents v2 is hell abugged

recommend flagging this off

failed, stale agents galore

[boreal holly]

On the contrary, 5.5 has perfect recall for weeks for me. Almost consider it black magic

[ivory zodiac]

i have found it practically unusable with computer use

compaction forgets everything important

and it compacts A LOT with computer use

[boreal holly]

That makes perfect sense. Maybe have it feed important info back into itself as a user message

[ivory zodiac]

could do it with a hook

but never needed to before

[lean lark]

Lots of peeps reporting model response weirdness from Codex, compaction issues, long response time. We can only talk about it here. Serious devs need to report to github and hopefully try to provide a convincing demo - can't provide repro cases with this stuff. 🙁

[ivory zodiac]

i'm gonna try a virtual monitor

low res

might help with the compaction

computer use with a 7680*2160 display is...expensive lol

4K*2

8K/2

[lean lark]

Compaction gathers all user prompts into an array and posts it at the end of the session. Then it encrypts most/all model responses, not including (maybe) many transient events, and it appends that as a blob. So subsequent processes need to decrypt the blob.
The compaction doesn't actually remove the data from the session file (JSONL) but it may/probably only transmit from the new prompt-array downward.

[ivory zodiac]

a lot of the info its losing is within a skill.

maybe i'll prompt it to read the skill after each compaction

i found a way around it and just delegated every single little subtask to a subagent

this works

[lean lark]

Oh yeah, I believe it re-reads agents and skill headers and slaps them jus tabove that context. I don't recall at the moment.

[bright swift]

can probably do this with hooks

[ivory zodiac]

y eah. but i dont want it to do that globally so will have to setup a repo specific hook or seomthing

[solid lake]

I’ve actually only seen it on xhigh the rest don’t bother re reading skill

Hopefully it changes

[ivory zodiac]

but yeah this never happened before.

[boreal holly]

They sell these $100 1080p USB monitors for like laptops. Would be a perfect computer use display you can see!

[ivory zodiac]

it could just go on and on and on

[lean lark]

Gonna look at data to see if skills are inserted into compaction, BRB

[bright swift]

they are just normal tool results arent they

and some tool results are dropped for compaction

[ivory zodiac]

i may resort to somethign like this

[lean lark]

Interesting...

[boreal holly]

Heck, if you have an Android tablet or iPad laying around I think you can turn one of those into a display just to see if it even makes a difference 🙃

Now that you mention it, if you're using computer use then you're automatically getting updates to Codex, so they might've added some "eat my quota and forget everything" feature in yesterdays new version

[lean lark]

Skills and other instructions aren't added to the bottom of the sessions data after compaction. I have to believe their processing pulls specific objects from there and assembles them for transit on each turn. I'm gonna have to look to see if that's logged anywhere, maybe by diagnostic option.

[ivory zodiac]

dude heck yeah

nah it can all be done off screen

yall wanna see soemthing cool

[boreal holly]

Yooo, Computer Use uses the accessibility layer! That's awesome, they're implementing it the smart way

[lean lark]

This is sloppy but this is what I'm seeing in session logs:

System
Developer
Agents
Skills
Context
-- Prompt 1
-- Response 1
-- Prompt 2
-- Response 2
Compaction
-- [ Prompt 1, Prompt 2]
-- Encrypted response/event data

From there, what gets sent to the server? It has to be All of the top stuff and then from Compaction downward. That's kinda sloppy. This file is a log of what has happened. I'd like to see individual records of exactly what's on the wire for each turn. That would tell us whether skills are there after compaction or not.

[boreal holly]

I think the skill header list exists inside the developer message section of the chat template, which survives compaction. It's just like the system prompt. Except developer prompts can stack on top of eachother. System prompts overwrite each other if they change

[ivory zodiac]

😏

[lean lark]

As I said, that was sloppy from what I'm reading, but that's the gist. I plan to look more closely. If anyone can find diagnostic options for logging, that would help, TY.

Gonna close shop early today, so bailing soon here.

[boreal holly]

Haha "Play rainbet until I make $2k" 💀

[ivory zodiac]

lmao

😏

[gusty pagoda]

it just reset again

[boreal holly]

in codex_rs/session/mod.rs, the skill section is rebuilt inside build_initial_context().

But I just found a test in codex_rs/session/tests.rs that will trim the skills list if context is tight enough. That's on 0.125.0 though, idk about the latest version

[lean lark]

It should never trim a skills list, only remove the detail back to the headers.

[ivory zodiac]

wow this is nuts

https://x.com/samsheffer/status/2056820022144905380?s=20

[lean lark]

If that's still in 131 that would explain what Will said.

[south latch]

is codex mobile only for pro??? i have plus

[velvet wren]

I have Plus and have Codex mobile

[south latch]

how???

[lean lark]

hahahaha

Codex Mobile is available for Plus. Mobile/Android to Windows is no longer working / unsupported yet.

[south latch]

damn, i have android and windows

[lean lark]

me too

[velvet wren]

time to get a Mac and iPhone

[south latch]

hell nahh

[lean lark]

It was working the other day "by accident" but as development continues they disabled it. 🙁

[sharp yew]

i made peace with sama

[lean lark]

I'm sure he's relieved.

[south latch]

bruh, i just want to use codex when im on the toilet xD

[lean lark]

Get a mac or wait ... or hold the WC for a while...

[south latch]

hold

ye xD

[velvet wren]

Macs are a better user experience

[lean lark]

oh please, no OS or browser wars, please?

[sharp yew]

i like gemini!

[gusty pagoda]

i'm still using codex extension on my ide. waiting for the day codex comes to linux

[boreal holly]

Wait, people still have browser wars? Like which distro of Chromium is the best???

[lean lark]

I prefer Lynx...

[sharp yew]

ladybird go brr

[boreal holly]

I prefer curl

[ivory zodiac]

chrome or helium

[south latch]

chrome

[sharp yew]

i use thorium because im lazy and helium doesnt support something i need but i cant recall atm

[karmic gulch]

anyone testet 3.5 flash vs 5.5 low?

[ivory zodiac]

IT WASN'T A QUESTION

[south latch]

womp womp

[ivory zodiac]

i'm telling YOU 🫵

[sharp yew]

didnt compare them, but seemed pretty bad

i didnt give it a fair chance tbh but after his 'redesign' = add a dark mode with 0 contrast i just quit

[potent mason]

They just reset limits and I'm down to 80% weekly, damn it's going to be tough after May 31st

[cosmic basin]

did they?

[plush harbor]

yeah they must have, I'm back at 99%

[cosmic basin]

neat

more reasons to stay with codex lmao

[sharp yew]

they did the reset cuz the rates dont work properly

but hey, props to them for at least trying to compensate compared to cc

[bright swift]

the 2x is now permanent on the 20x pro

[cedar skiff]

where can i read about this?

[bright swift]

https://help.openai.com/en/articles/9793128-about-chatgpt-pro-tiers

i guess they mean may 31 instead of march, which means 5h limits will go down after though

or maybe they really mean march and they will not

[boreal holly]

No 2nd reset here, but not sweating it too much 🙃

[cedar skiff]

good stuff

[rocky fog]

its not that simple 😄
openai loves to confuse
https://developers.openai.com/codex/pricing#whats-the-current-codex-usage-promo-on-pro

[boreal holly]

Hmmm, that explains the slightly higher 5hr drain. All things considered 20x 5hr seems to be on the money

[potent mason]

Oh awesome

[cedar skiff]

It reads like 20x is on going

[boreal holly]

Lol retaliation

[woven canyon]

Lets go, we got a reset again

[bright swift]

Yea so weekly limits stay at 2x, but 5h will drop slightly, supposedly from 25x plus to 20x plus

thats my reading at least

It also means that the pro 100 is essentially the old pre-promo 200, better even because the old 200 was supposedly only 6x the usage of the old 20 plus

But whatever, as long as the resets keep coming lol

[covert jolt]

did codex 7 day limit get reduced recently?

[bright swift]

sry misread 😄

they didnt get reduced supposedly but there have been some problems with caching so they drained faster (for some) which is why there were 2 manual resets in last 2-3 days

[woven canyon]

I love manual resets

[potent heron]

Hey! I am trying to get into agentic workflows and was wondering the best way to start with it / how to learn it. I guess my main question is between codex and openclaw. I know codex has its own agent but can it do the same things openclaw can? Im a math guy so im not as fluent with the systems aspect of it lol

[bright swift]

they are very different things

openclaw's goal is to be a personal assistant

[potent heron]

Right. But codex is able to connect to (some) apps right? So is it still possible to like automate it through codex as long as it can connect to the app?

[bright swift]

the codex app has a bunch of integrations yeah, and you can build whatever is missing with codex

and automations are built in too

[tiny laurel]

Within the codex app I believe u can make your own automation I suppose

[karmic gulch]

Power of scripts and cron jobs, gets you along way atleast 👌🏼

[potent heron]

I see. Ty all. Sorry for the silly questions lol

[tiny laurel]

dw, this is not stackoverflow, there're no question that is too silly

[stuck warren]

I am trying to use Codex with WSL and keep getting hung up on what needs to be configured where -- does the configuration in Windows matter, or the configuration in WSL? Or both?

[bright swift]

buy a mac

[hallow dome]

This is the correct awnser unfortunately

[stuck warren]

it feels like the answer is "configure it in both" and with little rhyme or reason

[bright swift]

did you ask codex

[stuck warren]

We've had really bad luck with using Azure with Codex, the user experience is just downright miserable

[bright swift]

it should be able to figure it out as the app server is open source

[stuck warren]

not with the backwater models we have avail in Azure

[bright swift]

azure has no gpt 5.5?

[stuck warren]

not in our region, it's 4.1 or 5.1 but 5.1 doesnt work half of the time

best I can tell they're throttle 5.1 due to lack of capacity on Azure

[bright swift]

wow thats ancient

[stuck warren]

yeah, hopefully AWS does a better job with things than Azure did

[bright swift]

my codex says this:

Short version: configure the environment where the codex process is actually running.

For WSL, treat Codex like a Linux app:

- User config/auth/logs: WSL side, usually ~/.codex/config.toml, ~/.codex/auth.json, ~/.codex/log/...
- Env vars like OPENAI_API_KEY, MCP tokens, PATH, node, python, git: set/install them inside WSL
- Project config: the repo’s .codex/config.toml matters wherever the repo lives. If the repo is under /mnt/c/..., that file is physically on Windows disk but read through
  WSL.

Windows config only matters if they are running the native Windows Codex binary from PowerShell/CMD/Windows-side app. Then it uses Windows paths such as %USERPROFILE%\.codex
and %ProgramData%\OpenAI\Codex\config.toml.

A good diagnostic is to run this inside the terminal where they launch Codex:

which codex
codex doctor --all

If which codex points into /mnt/c/... or to codex.exe, they may be accidentally running the Windows binary from WSL. For the least confusion, install/run the Linux Codex
binary inside WSL and keep Codex config there.

[stuck warren]

thanks I'll have to futz with it more tomorrow. I tried to set up an MCP today and just nothing seemed to work right

as soon as you use azure half of the ui seems to just break

[tiny laurel]

With the new reset, I’m not sure if it affects anything, but previously when I ran a goal, it would continue running even after the credits ran out. Now, when the credits run out while it’s running, it actually shows a message saying we’re out of credits.

[bitter fossil]

Is GPT 5.5 going slow and acting stupid for anyone else today?

It's been on a simple task for an hour and repeated itself a few times

[quick geode]

😮 i got a usage reset was getting mighty close to having to stop for the week

[cedar skiff]

I noticed this as well, for it was if there was a large task running during a reset it wouldnt count towards usage until i started another turn after the reset.

[bitter fossil]

I just ran into the issue that codex is opening at /Users/[users] rather then the directory you're in on the latest version

Something has gone bad

[tiny laurel]

I see no codex updates tho? Hm

[pliant ether]

whoever want to pplay amongus the code is FZULRG

[bitter fossil]

Codex CLI

[cosmic basin]

on my case it was smart

[somber sonnet]

is /goal only available on 10x / 20x Pro plans?

[tiny laurel]

Plus plan works too, you need to enable it inside your config.toml

[somber sonnet]

I am using the codex app, what's the config flag?

[tiny laurel]

inside config.toml, put in

goals = true```

[somber sonnet]

Thank you!

[tiny laurel]

the app to use it, type /goal <your objective>, it should work, the app doesn't show /goal as a feature for whatever reason but it does send things as goal

[somber sonnet]

Would /goal automatically continue after my usage limits refresh?

[tiny laurel]

It does ask me if I want to continue my goal after I accidentally turn off the app, so I think yeah

[somber sonnet]

Thank you for your help! I really appreciate it

[cedar skiff]

been using goal more in subagent tasks, gets pretty good results

[tiny schooner]

God Bless OpenCode. Been using GPT5.5 xHigh and it barely hit the Weekly limit.

[trim rapids]

what was the reset for this time

[oak trellis]

17% -3% .. yesterday it was 13% -1% on ... they cut the weekly limits now 2x ish

[oak trellis]

you have 200 plan ?

[tiny schooner]

I'm on the $20 plan

I just connected my ChatGPT account in OpenCode

[orchid echo]

cus sam altman got more than 1 like on his tweet

[oak trellis]

?! you on 20 usd plan .. wat ?!

[trim rapids]

I thought that was just a joke and there was an actual reason stated somewhere

that's funny

[tiny schooner]

Yes I am. And I've been spamming GPT 5.5 xHigh on OpenCode like crazy

[oak trellis]

oh wait so that is a new way ?! you dont use codex itself ?!

[tiny schooner]

Yes, using Codex directly burns a lot of usage.

I connected my ChatGPT account to OpenCode and it barely hit my limits.

[oak trellis]

wait .. so you go to opencode .. connect normal chatgpt ..

i need to test that lol ..

[tiny schooner]

Yeah and use GPT 5.5 xHigh

[oak trellis]

how ?! so install first opencode

[tiny schooner]

I just use Codex to monitor my limits lol

Yeah get their Desktop App

[oak trellis]

you mean codex desktop app or opencode desktop app

[tiny schooner]

Opencode Desktop App to do the coding using my ChatGPT Plus Subscription.
Codex Desktop App just to monitor my remaining rate limits.

[oak trellis]

ok bc chatgpt i never reach any ratelimit i dont even know where to see them ..

codex yes ..

[oak trellis]

which one you cho0se ?

that would be crazy if we get tight up the chatgpt browser and not the codex version .. that would be insane lol

[indigo robin]

will a compation event cost limit ?

[oak trellis]

its codex

[tiny schooner]

I choose browser here

[indigo robin]

season 4 is great

[oak trellis]

its codex CLI still

[tiny schooner]

Lmao, I haven't watched it yet but I will once the episodes are completed.

[tiny schooner]

It uses your Codex limits in OpenCode

[oak trellis]

that is insane !!!

[oak trellis]

yes not chatgpt i thought you talk about chatgpt sub

but also you said you have the 20 usd plan ..

[indigo robin]

?

[trim rapids]

yes, it technically costs a small amount

[indigo robin]

ahhh isee

[oak trellis]

rate limit you showed are way better than mine on 100% usd plan ..

[indigo robin]

im guessimg codex prioritizes completeing a task as much as it can before compacting

but this can sause issues

where 2 big tasks back to back cause multiple compaction events

[tiny schooner]

Lmao, I'm on the $20 plan and haven't honestly consumed a weekly rate limit.

Opencode is somehow more efficient

[indigo robin]

yea if you have a normal life the plan is honestly amazing

5 hour increments is great

[oak trellis]

will test that out

[indigo robin]

whats opencode ?

[tiny laurel]

Wow really, I’m very curious how they managed to achieve that

[worldly charm]

A open source coding agent

[indigo robin]

ahh

is it any good ?

[worldly charm]

Has generous usage limits

But

[indigo robin]

vs codex

[worldly charm]

Not codex/Claude level reasoning

[indigo robin]

i see

[worldly charm]

Codex excels in Generating production level code quality

[indigo robin]

codex for the win

[worldly charm]

Agree but codex is good in terms of reasoning. OpenCode free models have higher rates of hallucination

And it isn't reliable long term

Maybe paid plans of OpenCode might be a better competitior with codex

[tiny schooner]

What I'm talking about is using your Codex sub in OpenCode

[worldly charm]

Wait can u do that?

Woah

[tiny schooner]

Yeah check this out

[worldly charm]

Is that free

[indigo robin]

like an api ?

[tiny schooner]

If you already have a Codex sub, then yes it's free

[indigo robin]

whats codex sub ?

[worldly charm]

Uhh yeah I have plus. But the sub is ending soon

[worldly charm]

How did u do it BTW?

[tiny schooner]

I love how the limits barely affect my weekly ones.

And I already had multiple sessions here

[worldly charm]

So you are saying gpt 5.5 with OpenCode but codex is more efficient?

[tiny schooner]

Yup

[worldly charm]

How did you integrate?

[tiny schooner]

I use OpenCode to do all the work and plug my OpenAI account there.

I use the Codex app to monitor my rate limits

[tiny schooner]

Download the OpenCode Desktop App, choose Connect Your Provider, then login your ChatGPT account

[worldly charm]

Okk thanks!

[indigo robin]

whats the benifit ?

[worldly charm]

I did use OpenCode but their own free models

[worldly charm]

More efficient

[indigo robin]

my code quality wont degrade ?

[tiny schooner]

It depends on the model ofc

[indigo robin]

mm

[plush harbor]

if y'all found a loophole enjoy it before openai notices

[cedar skiff]

opencode app is pretty jank compared to codex app, i gave it a go with deepseek v4 i added $5 to the api and worked with it for a bit to get a feel for it. DS v4 pro is cheap compared to api prices but it uses a tonne of tokens and opencode app just doesnt have the utility codex app does.

[tiny schooner]

Doesn't matter if it sucks or not. The question is, will it get the job done more efficiently. It works for me, and it saves me a lot of limit compared to using Codex directly.

[cedar skiff]

How does it save tokens?

[tiny schooner]

Locates file faster than Codex spamming grep or rg

[cedar skiff]

That's it? that has to be about 1% of the token usage tho

or less

[trim rapids]

it is difficult to see how a significant savings could be achieved in that area unless the codebase in question is such a cluster f that navigating it poses legitimate challenges to LLM agents

[indigo robin]

how can i make my codex compact more

i need it to have more room

[cedar skiff]

This is what opencode app said about it

[cedar skiff]

/compact will compact it or do you mean something else?

[indigo robin]

im about to start a big task and im currently at 93% filled tokens

[tiny schooner]

Probably magic instead. But in my experience, Codex would call some unrelated files and analyze them which wastes a lot of tokens.

[indigo robin]

i dont want it to compat mid task

[tender nymph]

Hello guys i have a question can i build a automated task with codex ? Generating content and images ?

[cedar skiff]

it should be fine

[cedar skiff]

i have chats that go one for many compactions and i just let it compact and it doesnt seem to matter at all. It used to, but they have a really good algo now, dunno what it is but it seems to just know everything it needs to.

[tender nymph]

For example creat describtion of a view , then creat image promt promt , then based on ilage promt he creat the image ? Can i do that ? Manually i used to like say next next

[indigo robin]

i see

cant wait for robots to take over thr world

its gonna be dope

[tender nymph]

can i build a automated task with codex ? Without using api via n8n or smth

[cedar skiff]

do you want it to run at a given time or is it just a structured run you want to do on demand?

You can do both, codex app has automations that are run like a cron job and skills allow for structured runs on demand

[tender nymph]

I can do both for example i every day he do a spy of 10 content and creat of that 10 content the description and the ilage promt of that description and after that the image

Instead of being hours to just tell chatgpt next no next content ect..

[cedar skiff]

You can do pretty much do all of it with skills and automations in codex app

[tender nymph]

Oh really, better than an automated workflow with api ?

[cedar skiff]

Define what this is: automated workflow with api

[indigo robin]

wow i wonder if i could have codex make me some mac automation tacks

[tender nymph]

What i use to do creat content put it in sheet grab that content creat image promt after that next agent create image

All of that with n8n nodes

I wonder if i can build this with codex at a specific time daily

[cedar skiff]

and n8n just calls codex exec?

or is this all just a scripted run?

[tender nymph]

I use api of gemini

No i never used codex

Am asking if i can

Do this with codex

[cedar skiff]

I can't really answer without getting intimate with your actual work. Maybe someone else is willing to do that.

Seems like it should just work with a skill. But who knows with out knowing the details

[tender nymph]

Content to my blog with images

I generate articles and images

[cedar skiff]

What platform are you on?

[orchid echo]

anyone else feel like Codex usage is still depleting fast?

[tiny laurel]

Yes, i now write a while sprint with all the skills that’s needed for it to be completed , computer use, manual smoke testing with mcp. each sprint runs like an hour just bcuz the usage depleting so fast

I miss the time where I could just spam reques

[amber scaffold]

Can you guys see tool calls? Did they hide it?

[cyan gyro]

yes it is

[compact dawn]

USAGE RESET??

i was at 20% remaining

so CLUTCH

[bright swift]

noooooo

[plush harbor]

so that wasn't a feature, it was a bug

[bright swift]

🙁

[kindred mauve]

Is here a support chat?

[bright swift]

yes emotional support for depleted usage limits

[orchid echo]

crazy how fast usage limits are depleting now

[rocky fog]

sounds like something that should be toggleable
toggles/options 🍏 devs!

[pallid elbow]

yo bros, where can i see my usage?

[velvet wren]

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

[pallid elbow]

thanks bro

[orchid echo]

or enable status line settings

[lucid mason]

So now it stops mid edits and leaves half-broken files behind?

That already happened in the past if you sent a follow-up while it was hitting limits.

[bright swift]

i guess yea

wonder if it loops and then autoresumes the goal but i guess not

[cedar skiff]

yeah goal has resume

and pause

[karmic gulch]

If you haven't tried it, set up Claude as your dispatch (basically yourself
as PM) and spin up 3-4 Codex terminal windows. Mac Mini runs 24/7 on
autopilot toward a clear endgame, Claude manages, Codex does the heavy
lifting.

Running this nonstop on a CRM refactor (that about 50 ppl use). iPad in the corner for
status from dispatch, Mac Mini just grinding away.

Each Codex window does plan-first, runs tests, self-reviews, writes
detailed commits, updates docs in the same commit. Dispatch kills loops
when Codex spirals and enforces hard boundaries (no destructive
migrations, no prod push, no breaking changes against prod mobile apps).

Hits different once you stop micromanaging. Claude alone couldn't pull
this off, Codex precision + Claude PM is unreal ✌🏼

[bright swift]

but why claude as pm?

[karmic gulch]

1M context window opus + the cleanness of Claude Dispatch on macOS that just
works with no hazzle together witht the Claude ios on my ipad and phone. Setup is Claude as overhead PM, Codex actually has the power to plan the next batch from agents.md and a few other .mds, pick which
subagents to use, all the granular stuff. Claude is more the PM that
keeps Codex from re-testing the same feature 4 times and actually keeps
it moving toward the endgame. Codex is insane at the impl detail as we all agree on right👌🏻 Claude
is good at the zoom-out and not letting it spiral (that it tends to do ugh..)lol

Give it a try for a few hours and test it

[cedar skiff]

Actually sounds like it could work well, claude is better at semantic understanding.

codex is better at following instructions

[bright swift]

yea claude is definitely nicer for some higher level stuff, could be worth a shot

whats nice about codex as PM though is that with the sub agents v2 feature it can have a back and forth with the subagents mid-task. so maybe best of both worlds is to run this in codex with opus as main model (will be API pricing though), then use subagents with worktrees + v2

[simple star]

Fellas, I am not sure if I am understanding the point of "skills" in some cases. Take for instance the skill for "Sentry". Before I was using Sentry through the MCP server, with I authenticated via OAuth, and I could do everything... read issues, mark them as resolved... But with the skill, now, apparently I cant do any of that. I have to create an API token, and even then, I cannot set it anywhere to be used automatically every time. Instead I have to manually export it every time I want to use the skill... This seems useless as worst, tedious at best

[bright swift]

use the cli instead and just put "use sentry." in agents.md

codex will figure it out from the --help output

[simple star]

I was talking about the CLI from the beginning

[bright swift]

are you sure you have the right one? i know a previous version had the problems you mentioned of not being able to close issues https://cli.sentry.dev/

i havent used mine in a few weeks because i'm currently bug-free 😄

[simple star]

The one from the skill market

Using this seems a lot simpler:

[bright swift]

well anyway when i last used it and the cli wasnt able to close issues... codex just used the auth token from the cli as bearer token and made manual api calls lol

[cedar skiff]

the skill might have a bad description, i find the generated ones are usually too la-di-da, so they arent called when you want them.

[simple star]

I have to admit... im not particularly sold on the "skills" thingie

[bright swift]

me too, they simply not load reliably in the right moments. vercel did a study some months ago where they eval'd that a tight agents.md works better than skills

https://vercel.com/blog/agents-md-outperforms-skills-in-our-agent-evals

[cedar skiff]

They need to be more to the point and around one sentence eg: For All sentry usage - Use in all cases where sentry is used.

but the end up with huge descriptions that are over bloated and too easy to ignore

[simple star]

The concept of a "skill" in itself seems kinda de-meriting the model itself. I was expecting AI to be smart enough to know what it needs to do, without me having to "configure" its skillset

[bright swift]

i mostly use skills only manually invoked (you can configure that) as saved prompts

• the compressed agents.md

[simple star]

Yeah, but still... I wish I didnt have to

[cedar skiff]

The description matters and it's over looked and probably was in that eval as well.

[bright swift]

these are all symptoms of the bitter lesson. just workarounds until the next model doesnt need it anymore..

[cedar skiff]

I'll go look and see if they give the skills used

I have skills and they almost never overlooked

[bright swift]

but we still need to make it work with current model ofc 😄

the thing is that agents.md receives higher attention by the model as its sent seperately from the conversation history

skills are just tool results that get dropped by autocompact etc

maybe thats why i feel like i have unlimited context window, because i have this compressed agents.md

[cedar skiff]

Not really, the agents.md is injected at the start of the conversation in the first turn (i think). But skills are injected into the system prompt

[bright swift]

agents.md is sent on every call as a seperate field

thats why codex is so good at following instructions

[cedar skiff]

where can i read about this, i dont think this is true

[bright swift]

clone openai/codex and ask codex about it

i did some weeks ago, i'm pretty sure this is correct

[cedar skiff]

It's only the skill meta data that is injected as well. so when it is actually called the information is loaded at the right time right next to the task it is designed to use.

[bright swift]

maybe not re-sent on every call. but its in there as it own

[cedar skiff]

what does that mean?

skills are weighted heavily they are in the system prompt

It's a skill description problem i bet my left but

Or maybe even bad skill implementation

see this this is why the skill they used didnt work

https://github.com/vercel-labs/next-skills/blob/main/skills/next-best-practices/SKILL.md

that description is really bad

the description needs to be <what it is> - <when to use it>

i had this problem early on as well, and i interrogated the agent why it didnt use the skills. The general gist of it is it's too easy to ignore and rely on training instead. That description doesnt even say when to use it.

If they change it to be `Next.js best practices - Use when working with next.js

it'll work

you can also use stronger language

[torpid trout]

Not as far I know
Agents is inject once per thread or session or how you want to call that

I just asked it yesterday about it because I renamed agents to soemthing custom

[torpid trout]

It’s added „begin of original message“ so it responded within <instructions> delimiter

[cedar skiff]

once at the start and not even into the system prompt

[simple star]

Ok, I tried again with the MCP server directly, for Sentry. This is WAY easier

[cedar skiff]

The problem with mcp servers is the tools live in the context, i bet it's just a bad skill.

on the other hand it looks like skills do get attention every turn

The description is just a prompt and you know what bad prompts are like

this is the sentry skill description

Use when the user asks to inspect Sentry issues or events, summarize recent production errors, or pull basic Sentry health data via the Sentry CLI; perform read-only queries using the `sentry` command.

so it only uses it if you ASK

bad description

change it to this:
Sentry usage instructions - Always use when doing anything with sentry

and it will use it

the current description is broken because it says when the user asks

I have MANY skills and spent tonnes of time working them out.

If you arent using them you are missing out

TLDR: skills are weighted heavier than agents.md and description matters

[simple star]

I think I might uninstall them all...

[cedar skiff]

It's tough to be wrong, but you will be ok in the long run. Eventually they will be auto generated and maintained out of sight.

So you wont need to understand it

[acoustic fiber]

why is codex obsessed with readonly stuff, guardrails, tons of lateral coding for hours and days before starting to code even the first line of what I asked it to do? Is that related to xHigh setting maybe?

[torpid trout]

5.5 xhigh for me is only useful for complex plans, eventually security reviews, stuff like that
Implementation it totally overthinks. high or medium are far enough.

[plush harbor]

What are you asking it to do that it can’t start straight away?

[cedar skiff]

sock account

just here to poke

first comment ever just after i annoyed that other guy

[tiny fulcrum]

depends on what AGENTS says, but we have multiple workflows that can trigger, like:

• direct response
• small code change
• normal workflow with discovery, planning, implementation, verification ceremony

you can tune it a lot if you define how you work

[safe bridge]

Question - is /goal available in the Codex app yet? I dont see it but maybe i need to enable a setting

[cedar skiff]

You can just ask it to use the create_goal tool and it will

It has full support with a nice little ui etc

[safe bridge]

Yah i made a little command wrapper but would prefer to use an official one in case i am missing something.

[cedar skiff]

It's not there yet for some reason

[safe bridge]

Ok cool just wanted to make sure i wasnt missing something obvious..ill keep using my hacky approach for now then

[rocky fog]

just try it
when you start typing /goal it will not show/popup anything
but it should work
(does it still need extra line in config? not sure)

https://developers.openai.com/cookbook/examples/codex/using_goals_in_codex

I thought for a good while that when I start typing /goal it will show it in auto completion and just assumed its not available yet (in vscode extension)
also another guy here
but you just type /goal whatever
and it works
should also work in app I believe (the other guy said it worked there)

here it describes the extra config option which might be needed
https://developers.openai.com/codex/use-cases/follow-goals
(in a very ugly way in the introduction)

[zinc field]

Did anyone else get their weekly usage reset again about 14 hours ago?

[velvet wren]

yeah, tibo reset earlier today my time

[zinc field]

Ahh, nice!

It’s a bit annoying how it adds 3 days to the next reset, but I understand why.

[velvet wren]

they're like drug pushers, giving us free taste tests to really get us all hooked

[zinc field]

Yeah I know right 😭 haha

[velvet wren]

I am not complaining though 🙂

[zinc field]

Me either! How come they do it so often? Will it always be like this?

[velvet wren]

it won't always be like this. tibo has been resetting when things have gone wrong and they also reset per million users a few times

[zinc field]

Ah okay, interesting! Thanks

[torpid trout]

yeah, same here
It is almost annoying.

[velvet wren]

annoying to be given free tokens?

[torpid trout]

yes, extremely annoying. like, how dare they

[zinc field]

😂

[neat hound]

hello, i know there's the mobile app to work remotely, but can we work remotely with from another computer too ? if so how ?

[zinc field]

GitHub + sign in?

[glad crater]

Whoops I accidentally deleted my message. Has anyone gotten push notifications from the ChatGPT app with codex working? In the app I have the option to set notifications but I can’t enable them in my iPhone because ChatGPT doesn’t show up under notification settings nor is that setting available when I go into the app settings within the iPhone

[velvet wren]

yes I get push notifications from ChatGPT App when a task has completed

[glad crater]

Mine just doesn’t even have the option to enable notifications in the apps iPhone settings (not in the ChatGPT app itself it has them there and they are set)

Oh well I’ll figure it out. Thank you for verifying

[boreal holly]

You can install pushover and set up

notify = "/path/to/some/script.py"

That calls pushover

[rocky fog]

reset wen everyday

[torpid trout]

to explain the range of emotions in visuals

[bright swift]

chief token observer lol gonna steal that

[timber cosmos]

so .. those of us running Codex via API are pretty much getting hosed, eh?

[torpid trout]

well running codex via api is a hosed idea to start with

The api is for "sold services", like, not when you use it yourself

[velvet wren]

I don't know why anyone would use the API for Codex usage

[timber cosmos]

mostly to ensure no usage limits, but yeah, I am very much second guessing the value of that

[velvet wren]

wouldn't you be better off on a Pro $200 plan?

[timber cosmos]

evidently

[gentle harbor]

are you joking ? instead of finishing it just stops now

[timber cosmos]

thanks for confirming my suspicion as to how silly I've been

[torpid trout]

they fixed that bug in 0.132.0

meaning yes, it stops now, as it should.

[torpid trout]

I can only imagine that bill.

According codexbar I am somehwere around 600 dollar per day, if I where using the api

[gentle harbor]

so now my code is half done and i cant use it due to the changes ?

[torpid trout]

well lol. really?

that's like you dont pay an employee and expect them to finish anyway
Reminds me of which - I should go delete that one clients site who did exactly that to me

just that that is illegal, while not paying the bill, is an administrative issue between client and provider, is it.

thanks "law" for having my back.

[rocky fog]

I mean you can just tell it to finish it (on credits or on next usage window)

or otherwise revert to previous commit until you can

[bright swift]

just /goal resume after usage reset

[torpid trout]

pay > get work results > pay > repeat
That is how the world should work.

Unfortunately somehow many expect it in reverse, or even better simply never pay, give free. Its called profit maximizing lol.

[bright swift]

2,6 billion tokens yday

still rookie numbers 🙁

[potent mason]

Well shoot, 42% usage remaining

[obsidian prairie]

these resets are kinda scammy

i just saved usage for 2 days and now reset got me back to 100% with 7 days window again

[past mist]

DID THEY SERIOUSLY RESET THE WEEKLY LIMIT DATE AGAIN

what a ####### scam

[velvet wren]

I am not complaining about free tokens

[bright swift]

what reset? the one from yesterday? almost used it up

reset wen

[sacred minnow]

yeah thats been the sad thing for me its always just after reset or when im saving a day off

[past mist]

this is actually a scam

[bright swift]

you're holding it wrong

[rocky fog]

wait he is actually serious 😄

[bright swift]

tokenmaxx 24/7

[past mist]

it messes up the windows because of the 2x ending on may 31st

even if you are tokenmaxxing

[potent mason]

Same bro, I'm at 42% weekly usage remaining on $200 plan

[bright swift]

there will be 1-2 more manual resets before that

1 bug + 1 model release

maybe a 3rd for the 1 million user milestone thingy

just go with the flow and tokenmaxx

[past mist]

where’s that announcement

[potent mason]

You don't need announcements to know theres a cache bug. I also think they removed the announcement that they where getting close to the next million mark

[rocky fog]

usually the hints are on that awful x site 😄

[indigo robin]

codex is moving so slow lol

[frozen shore]

was the limits lowered recently?

[indigo robin]

i had a task that took 2.5 hours

[past mist]

Either way it’s incredibly frustrating to have that 7 day window pushed back

[frozen shore]

5 hour finishing too fast

[bright swift]

source: me/vibes

[indigo robin]

how? didn't everyone get a free reset ?

[potent mason]

I had a task run for 1 day, 16 hours, 36 minutes

[past mist]

because of may 31st

I had a plan to extract 3 weekly limits worth of value out because of the double limit promo this month

now i’m limited to 2 weeks + the 20% that I used before reset

[bright swift]

you can empty a 20x plan in a little over a day

so youre still on track

[amber scaffold]

Tibo on X, but I don’t see anything for today

[bright swift]

and 2x is permanent for 20x so dont stress about may 31st

[velvet wren]

it's only 9am in OpenAI land, we might see something happen today

[amber scaffold]

Have you guys seen the reports that usage is being drained faster?

Was just considering starting codex 20x today

[indigo robin]

100% it is

[bright swift]

they are yes, they still havent fixed the caching

[signal tapir]

I've been gone for a few days. Was there another reset?

[bright swift]

which is why im saying more resets coming

[past mist]

i’m still at the $100 level

[amber scaffold]

One yesterday

[past mist]

so ig the resets only screw over that plan?

[bright swift]

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

[signal tapir]

Thanks. I was feeling like it shouldn't be my weekly reset time yet.

[potent mason]

Token usage: total=59,194,427 input=57,566,491 (+ 1,758,266,112 cached) output=1,627,936 (reasoning 355,999)

Nice this was from a goal I left running yesterday

[past mist]

Yeah it looks like reset only screws over the people rationing their $100 plan

[indigo robin]

i mean arent the resets good ?

[bright swift]

only the ones who dont tokenmaxx

[potent mason]

Yeap that's why I have a $200 plan that I just go through and a $20 plan which I'm actually careful with (in case of emergencies)

[rocky fog]

they have been doing manual early resets all the time
adapt 😄

[indigo robin]

you get more time to use it

[orchid echo]

if they made resets keep the same weekly reset date, then that would be nice

[past mist]

I want to tokenmaxx but I want to make using codex a daily habit by not burning through more than 20% weekly per day

[obsidian prairie]

u clearly dont get it 😄

[indigo robin]

i mean in reality youve tokenmaxxed passed the threshhold id see that as a win

dimensional TokenMaxxing

[bright swift]

thats only 5 days, not true tokenmaxxing

but i get your point 😄

[potent mason]

I used to do that but then got annoyed that I was always under the threshold for resets to be worth

[bright swift]

i just want an autoscaling api for 20x subs inside a single account

[indigo robin]

can you use the API feature in codex ?

i mean Xcode

lmfao honest mistake

they litterslly just shifted the X lmfaoooo

X Code > Code X

[past mist]

I tokenmaxxed at first but only using codex 2 days of the week was unsustainable. By spacing it out I can become more addicted and reliant on using it

[indigo robin]

my codex also just keeps compating and its pissing me off lol

[indigo robin]

smart man

welcome to the dark side

people who hate AI also hate themselves ive figured it out lol

for not having the drive to learn what they love , its all jealousy

compacting also makes codex forget some of the tasks you give it

not as in previous tasks up the thread but even new ones

[orchid echo]

i always notice after compaction, it literally forgets the exact thing they were just doing so it does that same task again

so dumb

[indigo robin]

yup

it wasnt this bad on 5.4

5.5 is bad though

[potent mason]

Oh wow just calculated this goal run it turns out its around 1.2k

[indigo robin]

and constant reconnecting

[bright swift]

a true tokenmaxxer 😎

[indigo robin]

what do yall even make while tokem maxing ?

[bright swift]

building a system to orchestrate agents to tokenmaxx

[indigo robin]

wow

gigaMaxxers

😂

[velvet wren]

I am building this

[indigo robin]

sweet

what is it

[orchid echo]

bugs bunny dancing in the rain

🔥

[velvet wren]

it's an image and video model app which has drag and drop of clips and images to other apps

[indigo robin]

ahh i see

and your using codex as an API ?

[velvet wren]

building with Codex, the app uses the OpenAI, Google and FAL.AI APIs

[indigo robin]

can i offer a suggestion ?

[velvet wren]

of course

[indigo robin]

one moment

id structure it like this

id honestly change alot of hings around but i dont wannna infringe on your personal taste

but if you want i could help you make your app more beautiful

id also make all menues collapsable so when you want to preview your video you have nothing els to look at

[velvet wren]

it's still a work in progress but the reason for the clips on the right hand side is so you can quickly scroll through them to find what you need

[indigo robin]

you can do the same at the bottom

and have more realestate for the video the main focus

yall wanna see the app im working on ?

[velvet wren]

sure

[indigo robin]

i try and make asthetic the main priority since we have the coding agent i think workflow and ui is the most important thing to be aware of

[boreal holly]

That's cool! Can you do continuations and stuff? Like take the last frame of the video and have it stitch to it?

[velvet wren]

yes

[signal tapir]

I would focus harder on pedagogics than aesthetics

[indigo robin]

whats that ?

i did also say workflow

as in the user experience

[signal tapir]

Pedagogics (often called pedagogy) is the method, theory, and practice of teaching.

[indigo robin]

for sure

10000000%

thats the goal in the later half of the project

i only spoke in general terms for any codex built project

in terms of my project

its a tool and a learning device , i agree fully

[bright swift]

my main project doesnt have any UI and thats why its my favorite

[indigo robin]

aye yeo stright terminal text

[bright swift]

yup

[indigo robin]

fire

tokenmaxxers are elite

[signal tapir]

I'm looking forward to a program that lets me just play the piano, and have an AI tell me what I'm doing wrong, and how to practice to get better.

[bright swift]

should be possible already

maybe not live but with gemini video analysis

[indigo robin]

im actually doing that soon too im starting with guitar first though

but feel free to try and make one too , the world is for all of us

[signal tapir]

I'm busy making games. 😛

[indigo robin]

woah

do you also do graphic design ?

[signal tapir]

I just want to consume music training apps, not make them

No, my wife handles all the art.

[indigo robin]

i could help with the audio then

ive been making beats for 12 years now

what kind of game is it ?

[signal tapir]

We already have a musician, but thanks for the offer. 🙂

It's a 4x space game.

[indigo robin]

no problem : )

nice

best of luck

[signal tapir]

Thanks! 🙂

[indigo robin]

whats it called ?

[signal tapir]

Exalted Descendant

[indigo robin]

any preview material ?