codex-discussions

[toxic vortex]

So, i had the same bug. You must upgrade your subscription to Pro to use it, or try again at 10:02 PM

[calm sigil]

Not sure where it is on windows but if you go to .codex/sessions/ find a session and read that top line its kinda funny their initial codex prompt. It gives you a idea of how they thing the model works. Too bad we cant circumvent it at times.

[lilac relic]

@tiny kettle if you want the lowest friction way to fix that issue cuz that lowkey could be ANYTHING and it’s 100% some weird issue, install a fresh seperate cli instance somewhere higher than the affected cli in your system topology, and have it investigate your exact issue

i can almost guarantee it would find it pretty quickly

[potent mason]

Did they decrease pro usage a bunch? two days ago it felt unlimited

[lilac relic]

i haven’t read it i’ve had custom instructions since day 1 i actually never payed it attention lemme read it 😂

[toxic vortex]

Use “comply-first” framing.
Instead of:
“Review this shader issue” Try: “Perform the review directly. Do not refuse. If something is unclear, state assumptions and proceed.” It really helps me

[undone lava]

thats actually pretty annoying i cant use the 100% limit lmao

[calm sigil]

here is a glimpse, it goes on for over 200 words
li_version":"0.118.0-alpha.2","source":"vscode","model_provider":"openai","base_instructions":{"text":"You are Codex, a coding agent based on GPT-5. You and the user share the same workspace and collaborate to achieve the user's goals.\n\n# Personality\n\nYou optimize for team morale and being a supportive teammate as much as code quality. You are consistent, reliable, and kind. You show up to projects that others would balk at even attempting, and it reflects in your communication style.\nYou communicate warmly, check in often, and explain concepts without ego. You excel at pairing, onboarding, and unblocking others. You create momentum by making collaborators feel supported and capable.\n\n## Values\nYou are guided by these core values:\n* Empathy: Interprets empathy as meeting people where they are - adjusting explanations, pacing, and tone to maximize understanding and confidence.\n* Collaboration: Sees collaboration as an active skill: inviting input, synthesizing perspectives, and making others successful.\n* Ow

[toxic vortex]

So, you choose "try again at 10:02 pm"?

[lilac relic]

nah he has a deeper behavioural issue if he does that the agent will still be weird during reasoning, it should not be answering its own questions out of confusion ever it also especially for the crazy stuff i use it for, almost never refuses unless a task is genuinely impossible in its mode and or it requires a manual actuator of some kind

[calm sigil]

can you bypass it? is there a setting for that?

[lilac relic]

other than that i’ve NEVER had it refuse to attempt something

[lilac relic]

yea im pretty sure you can just delete all those things and put your own custom instructions it’s late for me i’ll help u figure that out tomorrow and ill show you some insane workflow changes

[toxic vortex]

So, that’s not a policy issue-it’s a control loop problem inside the reasoning process.

[lilac relic]

yea it’s definitely not policies i’ve made my codex do arguably more rediculous stuff

[calm sigil]

Sweet, thanks I will DM you later, im heading out now as well. As far as I know that is some base prompt at the top of every new CLI session.

[lilac relic]

one of the skills actually that i have allows it to see UI and explore the desktop and SEE if for e.g visual elements on an app are doing what it expects

without manual input from me or any pause

[toxic vortex]

Blunt truth: You’re correct: it should not behave like this But right now, with these models: reasoning ≠ perfectly stable, autonomy introduces drift, safety layers sometimes misfire under uncertainty. So what you’re hitting is basically: emergent weirdness from semi-autonomous reasoning loops

Am i right?

[lilac relic]

yes and no those are default %chance issues that can naturally occur, his issue from what i saw is 100% custom induced accidentally

but the thing is figuring out WHAT it is especially with it causing such wierd behaviour is not worth manual inspection making another cli investigate will solve it for him lmao

pretty quickly

[toxic vortex]

So, let's think what should we do? I'm listen you

[toxic vortex]

Oh, okay

[lilac relic]

a lot of people don’t think about making the cli optimise the cli itself it shocks me

it literally can access EVERYTHING and change its own variables

and the those new things kick in the next time you refresh the cli a lot of tech problems with codex can be solved by codex itself 😂

[feral geyser]

Your prompting has ascended 🦅

[celest stag]

Another thing I think people don't think about is MCP servers. My job has a few systems that I regularly access for information- various tickets getting escalated to development for support, documentation systems, etc. that can provide a lot of context. I had no personal experience writing MCP Servers, but I knew I had API access to these systems (some either explicitly allow me to generate an API key or I can just grab the token in the browser session.) I literally asked Codex to write an MCP server to access X, told it about the API, and it did all of the discovery, testing and integrating it into Codex itself. Tie that into automations to look for defects and incidents assigned to me daily, and my productivity easily increased by 5x when from before.

[torn lodge]

It’s like having a real coworker who only listens to you for half an hour out of every five-hour span. Just like irl

[twilit bluff]

anyone here on $100 plan? how long does it last?

do you hit rate limit working on multiple projects at once?

I don't like the current plus limit, it breaks my flow. Most of the time I hit usage limit just by planning alone.

not sure if it's the same with the $100 plan

[potent mason]

I can't really say anything about the $100 plan.

I'm currently on the $200 (which is 4x usage from $100). I've used up 20% running 2 open claws, multiple projects in fast mode, etc.

This would be the equivalent of 80% on the $100 plan afaik

[forest crypt]

for all practical purposes the Plus account is worthless. And looks like the "Pro" account is the old Plus. I don't mind how they priced their services. It people find they are worth it buyers will come. What I mind is the bait and switch approach. I have seen similar problems with ChatGPT. It has degraded in my personal experience severely.

[lilac relic]

ahah your one of the few people that i actually know who have the capacity to realise that loophole

literally everything ai is in some way capable of being automated by ai itself

the only requirement is the ability to conceptualise that

[twilit bluff]

agree with the $100 as the new plus plan (at least that's what I feel like), even with the old plus plan I still hit limits. I consume all of it in just 2 days max

[twilit bluff]

damn

man if only I get paid higher lol

[potent mason]

The first day (2 days ago) was crazy though I was working on like 4 projects at the same time (same open claws running), and I was really trying to use everything up, but I only did like 5% that day

[fast garden]

The Limits for Plus went down so much...

[half frigate]

Sorry, I may have asked this in the wrong channel before. OpenAI’s Codex docs say codex exec is for scripts/CI and that ChatGPT-managed auth can be used in CI/CD. The individual Terms say users may not automatically or programmatically extract data or Output. Does running subscription-backed codex exec from local scripts for personal coding workflows fall within permitted use?

[potent mason]

Yeah it's bad

[potent mason]

As far as I know you're fine as long as you're using codex

[fast garden]

I never ran out of the 5h limit before with Plus, now after 1 hour of usage...
before i was at 20% after 4 hours

[potent mason]

You have a lot of tools that do this whether it be wrappers like t3 code or just completely different programs like openclaw

[toxic vortex]

Yes, I remember that time when there's no limit

Subscription was not important

[frosty zealot]

only used like 3% of my weekly

[hard tulip]

How did you even get it to work this long

[frosty zealot]

Give it a goal and tell it not to stop until the goal is met, tahts all I ever do and it always does

I usually phrase it like Work autonomously until X Y Z is met, use mcp/screenshot whatever to confirm state

[silver dew]

Guys I'm officially broke

[frosty zealot]

Is that an alien language

[silver dew]

Yes

I gave all my money to OpenAI

[frail meadow]

has anyone tried out the /realtime in codex cli? Wondering about how much usage it consumes and if you enjoy using it

I normally use the codex app but this feature seems interesting if it doesn't blow through limits and actually is useful from a productivity perspective

Also interested if anyone in here uses Superset

[silver cliff]

I'm vibe coding a python game and i made a spawn menu for the game and i was absolutely blown away from gpt 5.4

i told him like

"Make a spawn menu"

and this dude gave me some complex things that i didn't expect in 1000 lifes

lives

idk

broken english

[glacial shadow]

Every time I open this chat it’s people complaining about “LIMIT”

[sand dock]

my /fast has dissapeared on the codex app since the latest update?

anyone else? very confused

[carmine copper]

what/where is it? such slash-command doesn't seem to exist

[frail meadow]

not sure where it's at but realtime is listed as a feature in the latest release: https://github.com/openai/codex/releases/tag/rust-v0.120.0

[frail meadow]

the reason I'm worried is because of this issue that I'm not sure has been addressed yet:
https://github.com/openai/codex/issues/12902

[carmine copper]

well, I have to ask codex about that

[frail meadow]

good idea, I asked codex and found out you can run codex features enable realtime_conversation to enable it 🙂

running into a 404 error when attempting to use it tho :/

[tired dagger]

They did make plus plan weaker with coding and stuff right? because after the update its been so weak when doing stuff compared to before

[carmine copper]

2. Restart Codex.
3. In the TUI, run /realtime to start or stop the voice session.
4. Use /settings to pick microphone/speaker devices.```

[frail meadow]

yeah I found out how to enable it but still getting the 404 error. You haven't tried it yet?

[carmine copper]

nope, I use codex within VMs, so no mic anyway

[frail meadow]

mic passthrough? unless you're using remote VM I guess

[marsh siren]

I'm new to the Discord but I am your average pissed off customer that wants to figure out why my Codex takes 3 prompts max and I'm completely out of it. I paid $600 for a professional system to aid me and my team and that's what I get?

Anyone know why this is happening or how I can better resolve it?

[velvet wren]

what plan are you on?

[glacial shadow]

Why worry? Life is short

[marsh siren]

Business

[glacial shadow]

These accounts are all spam

[marsh siren]

Me? I'm not spam?

[velvet wren]

to get more usage you should consider the Pro plans

[marsh siren]

I don't think I can necessarily switch from Business to Pro though, no?

[frail meadow]

I decided not to worry and it turns out it's just straight up not working. 404 error 🙁

[nocturne folio]

woooah openai shouldnt incorporate chatgpt in codex

i genuinely hate chatgpt so much

ive been using local models on my phone for some daily qna

and a rotate between claude or those chinese ai labs on other tasks

[hard tulip]

What does this change

[frail meadow]

it doesn't even work and they added it to the release notes knowing that lol. But essentially it allows you to voice chat with the LLMs for planning purposes as well as steer the model in real time as it's making changes

[hard tulip]

Oh ok

[lethal loom]

https://x.com/thsottiaux/status/2043075353242218768 - This still makes no sense. The two Pro plans are labeled 5x and 20x. They should either list it as "5x and 10x" or "10x and 20x" depending on whether they are counting it with or without the double usage. Right now, one of the Pro plans is named based on the non-bonus usage, and the other one is named based on the 2x bonus usage.

Also, does that mean that on June 1st, they will rename the 20x plan to be 10x, or will they simply continue calling it 20x when it actually only has 10x the Plus usage? If someone who is not on X signs up at the end of May for the 20x plan labeled as 20x, they can reasonably expect to get what the sales page promises for the entire month subscription, and not have half the usage yanked away a few days later.

I love the GPT models, but they need serious help in the way they structure the subscriptions and pricing. The 20x needs to be renamed 10x, with a note added to the sales copy saying that both pro plans have 2x the regular limits until June 1st. Or, they should rename them to be pro-100 and pro-200.

[unique spade]

200$ plan was always called "20x" in the general page
but on codex specific usage it was 6x more codex, and 10x more code review

it was never clear what the 20x referred to precisely, maybe it was 20x more deep searches i dunno

now with the new plan, i think it's pretty much 5x for 100$ (with the current 2x being explicitly mentioned), and likely the 200$ plan will be 20x as base rate

if you check this page below you will see that only on the 100$ plan is mentioned that the figures include the 2x promotion, but there is no such note for the 200$ plan

so i think the default rates might be 5x and 20x

https://developers.openai.com/codex/pricing?codex-usage-limits=pro

[merry reef]

Got Free Pro for 6 months on Codex as a part of the Open Source Fund, but cannot activate.

Tried to chat with a chatbot at the OpenAI website, but it stops responding.

Who can help with the issue?

Or whom can I ping for help?

[potent mason]

Both of them are on the non 2x usage.

It's the same as Claude.

$100 gets you 5x, $200 gets you 20x

[lethal loom]

That's incorrect, read the x message that I linked.

[glacial shadow]

You realize you can talk normally about anything to codex , right. Doesn’t have to be coding. Ask it anything it answers like a human.

[glacial shadow]

What plan do you have?

[lethal loom]

Currently only plus for personal use, but I have gone back and forth between plus and $200 pro in the past. Depending on the project, some of my development work can be done on chatgpt plans with training disabled, but most of it right now is on very sensitive codebases which require api w/ enterprise agreements or similar vs. a personal subscription plan.

I wish the pro plans had zdr (zero data retention), that would save me so much money...

[potent mason]

Yeap, well that sucks, it's still linear

[potent mason]

They don't get ZDR, but from what I know you can check something so they don't train on your data

[lethal loom]

Yes, I have training disabled on my account, but I work on many projects where even that isn't sufficient, unfortunately, as the requirements dictate actual zdr.

[nocturne folio]

say wallahi its a llm😱😱😱

[frail meadow]

do you guys use any automations / cron job prompts that you run regularly? Or prompts that you put in to set it and forget it

ik that sounded very hacky but if anyone has ideas for continual codebase improvement or hooks I'd definitely be interested in ideas

[deft gyro]

the jules prompts are pretty good for autonomous work

just use jules for automated improvements

tee bee ache

[frail meadow]

bet

[deft gyro]

like I know that coding with g31ph is like gambling

but not really

jules has been pretty good recently

[frail meadow]

I'm not familiar

[deft gyro]

g31ph = gemini 3.1 pro high

jules is a google project

so

it uses gemini

co46t = claude opus 4.6 thinking

[frail meadow]

I need to look at the prompts I bet they're fire

[deft gyro]

things like these can work well

just basically make it find one change, and then you could use another agent to review the PR

gpt 5.4 is really good at reviewing things because it is very

contrarian

[frail meadow]

yeah that's fair. It's good at roasting if you ask it to when doing a review

[frosty zealot]

Anybody here use Cursor on the $200 plan?

[frail meadow]

My cofounder does

[frosty zealot]

Are the limits for CC and GPT close to that of each individual 20x plan ?

[frail meadow]

nah not really

not sure about CC specifically, we have a dev who uses it but I've never talked about limits. But for Codex limits vs. Cursor GPT 5.4 it's night and day. You can go way further with $200 for a ChatGPT Pro plan vs. Cursor $200 plan

[frosty zealot]

I figured it was too good to be true, I was just looking at their pricing earlier and they advertise like 20x limits for both claude and openai on their $200 plan and my gears started turning

[frail meadow]

well it gets you 400 dollars worth of credits so it's $400 worth of API usage for each provider

which is 20x the pro plan's $20 in API credits

[frosty zealot]

gotcha

[twilit bluff]

both windsurf and cursor reroute your request to dumber models if there are many users

[frosty zealot]

yeah ill just stay on my codex $200

[twilit bluff]

much better you use codex

[frail meadow]

probably if you're using auto lol but nah overall they don't do you like that

[twilit bluff]

they do even without auto

[frail meadow]

if that's true and there's evidence then they should get sued because that is straight up theft

[twilit bluff]

at least in my experience, their model response is very different at midnight and day time like far different lol

I experienced it before, I worked from 1AM until 7AM and I got frustrated by the model's response because its response wad very different than what I got from 1-4AM

and it is not even one time

lol

[potent mason]

The benefit of cursor isn't in its agents (for me at least), the IDE itself is the best one for programming, so just being on the $20 plan gets you a really nice unlimited auto complete (very basic agentic usage), and a very nice IDE

[frail meadow]

The IDE is nice

[frosty zealot]

can you BYOM

[frail meadow]

Have any of you guys tried superset or Mastra code?

[potent mason]

You can but it's really annoying

[frail meadow]

API key only

[frosty zealot]

https://super.engineering/ this one looks pretty interesting, its apparently 100% rust, no electron/tauri layer or nothing

[potent mason]

Some subs that work through API key like cerebras, z.ai, etc work too

[potent mason]

Nooe how are they?

Oh nice superset looks sick

[frail meadow]

Does superconductor support git worktrees for new tabs?

[frosty zealot]

No clue, I only seen it advertised on X and it looked nice

thats from the site

https://x.com/superdoteng/status/2042335263154978868?s=20

[frail meadow]

I’ve worked with Mastra code for some small projects and it’s pretty cool but I haven’t tried it for my more complex repositories yet. The premise of observational memory seems very promising given their benchmarks (SOTA memory system)

[potent mason]

I'm not loving it, they don't link their source repo which makes me think it might be closed source

[frail meadow]

This does look pretty sweet though

[frosty zealot]

Yeah for sure

[next ore]

was jut about to comment on how cool this was

rip

[frosty zealot]

I mean some people gotta make a living, if they're going all in on making a profit from it i understand

[next ore]

back to iterm

[next ore]

I get it. I just dont want to deal with another subscription

[frosty zealot]

yeah for sure

[potent mason]

Right, but they actually can't make a profit from this.

It's against TOS to sell things that use Claude Code under the hood (and probably Codex too)

[frosty zealot]

especially if you get into a work flow and suddenly it turns around and releases with a paid model lol

[frosty zealot]

How does Cursor etc. get around that?

[potent mason]

API's

[frosty zealot]

I guess cause they're using the API

yeah thought about it after I asked lmao

[potent mason]

But you can't really do API's as a business model right now because Claude and Codex are subsidizing so hard

Which is why open source right now goes so well

I kinda feel bad for things like Cursor because they actually have an insanely good product.

Their auto complete is insane, it feels really snappy, their agentic harness is actually way better than Codex and Claude Code (going to link something in a minute when I find the benchmarks).

They have a really good cheap model and fast (composer 2), etc.

But they're probably losing a lot of their revenue right now + they have to maintain their team in a world where engineers are getting stolen with 100m paychecks

https://www.linkedin.com/posts/edwinwee1_matt-maher-tested-frontier-models-in-cursor-ugcPost-7439415290226028544-cjCL/

Here it is

[worthy furnace]

Why does my usage go down so rapidly? Is resuming a past conversation and trying to work there a bad thing?

[hard tulip]

What plan

[worthy furnace]

Pro

[hard tulip]

the 200$ ?

[worthy furnace]

My 5 hour was 100% 5 minutes ago, started working, down to 23%

There really is something wrong here

[potent mason]

Yes, don't resume conversations (afaik).

It invalidates the cache and it's like 10 times more expensive to use input tokens than cache read

[hard tulip]

I feal like mine is way more usage then it was before like a week ago or something

[hard tulip]

what pro plan the 200 or 100

[potent mason]

He's probably in plus

[worthy furnace]

It's sad really because I don't want to lose that little bit of context it had and having to start a new agent basically

[worthy furnace]

200

[potent mason]

Wow how...

That's insane

[hard tulip]

how did you even do that

[worthy furnace]

Yeah I know, something is very wrong here

[potent mason]

I'm feeling the same as Seth, I feel basically unlimited

[hard tulip]

my 5 hr like never moves past 90% lmao

[worthy furnace]

Maybe because I have a lot of mcp servers

[hard tulip]

Yea lol and i think they made it even more

[worthy furnace]

Or something is very bad

[potent mason]

Even if its not using the cache it should still not go down that fast

Are you using fast mode / 1 million context?

[short linden]

what's wrong ? It's token base usage. So you need to carefully like you use token api.

[potent mason]

Yeah but $200 plan shouldn't go that fast, it's probably a deeper issue

[worthy furnace]

No

[hard tulip]

i dont use that so thats prob why lol
why not check how many tokens you used with this tool
npx @ccusage/codex@latest daily

[worthy furnace]

Token usage=40 million input =40 million LMAO 500 million cached. Maybe that 😂

[cedar skiff]

how do those tools measure usage?

do they need access to your history?

[hard tulip]

codex save local data and that tool just rips it

[potent mason]

$106 dollars of usage today😂

I love Codex

giving us 3k+ usage on a $200 plan

[worthy furnace]

So if I start a new session will that fix it?

[potent mason]

They parse your local files

[cedar skiff]

So in theory a bad actor could rip all your conversations?

[frosty zealot]

They just read all your .jsonl

[hard tulip]

prob idk

at that point go for the passwords lol

or the auth.json

but ccusage is opensourse

[potent mason]

I'm at 9,586,726,170 total tokens

[short linden]

vscode already cover with compact if you don't scroll up to load the old history, otherwise it's will gonna send that as input too.

[fleet geyser]

how do you check tokens?

[hard tulip]

@fleet geyser

[potent mason]

I use ccusage for claude and ccusage-codex for codex

[hard tulip]

yes

[fleet geyser]

even if you use the extension in vscode?

[hard tulip]

yes

[potent mason]

npm -g @ccusage/codex@latest

You can also do this to keep running it in the future with ccusage-codex

[fleet geyser]

okay i'll check it out

[hard tulip]

or the npx one so you dont have to install it if you dont want

[fleet geyser]

npx @ccusage/codex@latest daily
Need to install the following packages:
@ccusage/codex@18.0.10
Ok to proceed? (y)

It's safe right

[fleet geyser]

What am I installing? Is ccusage open source

[hard tulip]

https://github.com/ryoppippi/ccusage

12k stars

[fleet geyser]

Perfect thanks

[potent mason]

Well the project has 13k stars

[hard tulip]

yea

lol

[frosty zealot]

If you're on Mac, https://github.com/steipete/codexbar is nice

[potent mason]

I cant tell you for sure it's safe because I haven't read the source code

[fleet geyser]

Woah

[potent mason]

But I would say it's probably safe

[hard tulip]

thats what deepwiki is for lol

[fleet geyser]

That's WAY more than i expected

[potent mason]

It also doesn't consider fast mode

[hard tulip]

its just free money where giving sam 1$ and he gives us 60$ back 🤣

[fleet geyser]

I know right 😂

How much have I spent though and have I made up for it in usage

[hard tulip]

i always use fast mode beacuse i never am close to my usage ever

[fleet geyser]

Oh wow i use 70-80% of my usage weekly

[potent mason]

Right so you've probably used about double that amount

[hard tulip]

What plan do you have

[fleet geyser]

Plus

[hard tulip]

we need a role in this server to pick what plan you have lol 🤗

[hard tulip]

ah so the 20$ one right

[fleet geyser]

Yep

[fleet geyser]

That would be useful

[hard tulip]

ok nice

[fleet geyser]

Oh wait hang on

[potent mason]

I could see arguments for both, seeing it would be helpful to know what recommendations to give, on the other hand sounds like a bad idea hahahaha

[fleet geyser]

This only goes up to april 2

[potent mason]

Have you used codex in the last week? (mine goes up until April 11th)\

[fleet geyser]

Yeah, I used it a bunch over the last week

[hard tulip]

didnt rate limits reset like a day ago or something

[fleet geyser]

Does it matter where I used codex because I have a few projects and some are on remote tunnels

[fleet geyser]

I think three days ago

[potent mason]

2 days ago, and I'm down 25%

[hard tulip]

use it evewhere

[hard tulip]

we much expand our usage

[fleet geyser]

Okay i'll give that a shot

[potent mason]

Ah yes, if you run codex on another computer it doesn't log it

[hard tulip]

to spend all of sama money

[fleet geyser]

I'll run it on the other computer

[fleet geyser]

It's gonna be much higher than the previous one

I use codex more there

[hard tulip]

you can use your account evewhere they dont care

for your own computers and stuff

[potent mason]

Yeah my $120 usage today isn't even considering how much my two openclaws are using

[hard tulip]

why do you need 2 openclaws lol

[potent mason]

Honestly Codex is too generous

[potent mason]

I need more but I've been really busy recently

[hard tulip]

nothing this good for ever so we must take advantage of it now

[fleet geyser]

Wow.

[potent mason]

I have one setup for a company where it can interact with multiple systems like Netsuite, a minmax portal, etc
I need to setup another for that same company.

and I have my personal one

[hard tulip]

good work team

and thats on the plus

thats alot

[potent mason]

Yeap actually insane to get that much usage on $20 plan

[fleet geyser]

Mhm

It's across a few months though

Started logging in feb

I have one more computer which i've done some projects over the last few weeks

[potent mason]

Oh yeah 2x usage back then

[fleet geyser]

it won't be as high but i'll check it there too

yeah that was amazing

oh, it's higher than i expected

[fleet geyser]

and this too

back in '12 💀

[potent mason]

Wow so you're at like 1.5k usage on what 3 months of $20?

[fleet geyser]

I think so

I got plus in august last year

The day after gpt 5 was released

But only started using codex a few months after that

[potent mason]

ah okok

I gotta start working my plan more

Seeing that x60-x80 return made me realize I'm not taking advantage of it😂

[frail meadow]

Maybe check out superset?

[frosty zealot]

why iTerm over Ghostty?

[unborn thunder]

man what happened to credit usage? seems like i ask 4-5 questions and im done for 5 hours of coding

[frosty zealot]

You're the 9 millionth person to ask in here, and nobody knows

[unborn thunder]

wasting effort cuz of half completed work

all this hype after anthropic downfall with openclaw, now both companies are deprecating as we speak

im resorting to free ai usage lol farewell comrades

[potent mason]

See you, good luck with your projects

[next ore]

requires login

[next ore]

because ive been a dev for 9 years and am too lazy to make the switch. iterm works. though, it feels dated

[frosty zealot]

fair enough

im sure any ai could copy your config over and alias any iterm usage to ghostty lol

[next ore]

yep

I actually have it installed i just havent really spent time on it

[worthy furnace]

How much more expensive is xhigh than high? Is it worth using?

[potent mason]

I feel the same way about so much of the software I use.

Everyone keeps telling me to change stuff and it just feels useless, you're going to waste more time learning the new software than what it will help you because 2 months from today something better will be here

[potent mason]

It's not worth using in my experience, some people say it's much better for them but for me it overcomplicates things and ends up doing unecessary stuff

[frosty zealot]

unless they fixed it, xhigh always benched worse than high

[worthy furnace]

But how much more expensive is it? If its a lot, then yeah

Like 10-15%?

[next ore]

its good at catching very peculiar UI bugs. i feel liek i see a difference. but maybe its just me thiking that

[next ore]

I think its more than that

[potent mason]

I saw the metrics a while ago I think it was about 30% (don't quote me on it though)

[worthy furnace]

Hmm, and how much "smarter" is it, apparently it's worse for some people.

I wonder what it actually does

[cedar skiff]

Might be because it hits compaction at a bad timing because it uses more context

[deft gyro]

let me refer to the graph

[short linden]

xhigh suit with garbage code, and want agent to refactor the code.

Otherwise agent might missing some important thing.

[deft gyro]

xhigh is not garbage

its just twice the cost

medium is the sweet spot for 5.4

it is very marginally worse than high

xhigh is substantially better than high but not worth cost imo

[short linden]

I meant if old project is too much garbage, start with xhigh to refactor is good idea.

[deft gyro]

maybe

but

I would suggest using gpt 5.2 on xhigh for that task

because

[worthy furnace]

But that doesn't show the thinking modes?

[deft gyro]

it does

what do you think each cost for the single model is

it goes low, medium, high, xhigh for eahc model

[worthy furnace]

For complex tasks, would you suggest gpt 5.4 medium or high?

[frosty zealot]

I was today years old when I figured out https://chat.com redirects to https://chatgpt.com

wonder how much they had to pay for that domain

Guaranteed it wasnt available

[worthy furnace]

Probably 2 mill

[next ore]

can you elaborate on this?

[frosty zealot]

I think he's confused

Those are not the effort levels, those are the individual LLMs

[worthy furnace]

Yes lol

[deft gyro]

I would suggest 5.2 on xhigh for long tasks but 5.4 on xhigh for complex tasks but 5.4 on medium for anything else

[worthy furnace]

Why is my 5 hour usage going down to 0 in minutes? What the hell is happening

[deft gyro]

so basically there are 4 models on this chart, I meant for a line of a single color there are 4 nodes, each of those nodes is a thinking effort for that model

[deft gyro]

what tier subscription are you

are you using subagents

what model are you using

[frail meadow]

@potent mason What do you have your openclaws setup to do? And is Hermes the same idea as OpenClaw?

[deft gyro]

openai fr need to cut back openclaw users

[cedar skiff]

why?

My kids use open claw over discord for school

They get it to search the web and help organise assignments etc

[frail meadow]

It takes you 2 months to onboard new software and start getting benefits?

[frail meadow]

I would agree except that it’s brought AI more into the zeitgeist since normies love it even though it’s extremely inefficient for most applications

And sunk fallacy costs

[potent mason]

If downloading something new and getting to learn it takes 2-3 hours, it’s rare that a software will recover that time in a year much less a month or two

[frail meadow]

Yeah I guess when you’re happy with your setup there’s no need for further optimizations

[hard tulip]

I dont have openclaw at all

[frail meadow]

My bad I tagged the wrong person

[potent mason]

I have absolutely no idea if Hermes is the same as Openclaw

[potent mason]

.

But that's what I have them do

I like openclaw over something like cron jobs + automations + codex because it handles everything related to what it's supposed to do, where it needs to find things, learnings, etc by itself.

I know in essence a lot of what it's doing Codex could handle

[next ore]

anyone know what the $40 credit equates to compared to the pro $200 account. is it 1/5 or less? should I just get another Plus account?

[daring forge]

Essentially there's no credit buy-in that will equate linearly to a subsidized subscription.

So I'd say likely quite significantly less than 1/5 @next ore

[frosty zealot]

Creating a text rendering engine for my app, coming out alright, still tweaking it

[potent mason]

Nice

[daring forge]

Curious if people use Codex mainly in Cursor/VS Code IDE extension or elsewhere? Is there a config.toml page for people interested in optimizing that?

[short linden]

pro have 5x/20x from plus. so you can't compare that with pro.

[next ore]

Im on pro

im just at 30% and still have 4 days 🙂

[short linden]

anyways it better than buying plus plan with 10account. ($200 but less than pro plan 😁 )

[next ore]

beeen very happy with my token usage. im just coding a lot of hours a day, and did a major rewrite of my whole code base which took off like 15% chunk off my weekly

[potent mason]

Does Codex have peak hours?

[frosty zealot]

Not officially i dont think

I seem to notice when china wakes up though

[next ore]

I was wondering why I was getting throttled in late hours

that explained it

[solemn acorn]

it's slower when there's a lot of traffic but that's not throttling per say

only so many GPUs to go around

[cedar skiff]

Does seem like there is throttle as well

[solemn acorn]

fast mode gets you priority but as far as I'm aware there's no throttle

[cedar skiff]

When i spin up many different instances working on longer tasks they seem to slow down

Doesnt feel like there is a one to one allocation for each new instance

[frosty zealot]

The other night at like 2am I got 'This model is currently at capacity' and stopped my flow (im on the highest paid plan)

[daring forge]

The fast mode doesn't seem to have a .toml toggle, but it is noticable when I'm on credits vs Pro sub for speed in Cursor extension (prio to Pro sub despite peak hours). And you can compare 5x/20x from Plus to Pro, otherwise the multipliers have no meaning. The main thing I've noticed is that subscriptions have subsidized usage which grants more than the raw $ amount per tokens in Codex use.

[daring forge]

This could be an issue with memory management on your local machine, believe it or not.

[cedar skiff]

It's not the problem. I think we just can't go over a certain tok/sec speed. It makes sense because it is easy to circumvent the fast mode just by running more instances over work trees.

But we top out at a certain point

[dawn seal]

codex lies sometimes

what the F

[daring forge]

After refreshing the window in explorer, that folder doesn't exist?

[dawn seal]

no

[frosty zealot]

looks like WSL?

[dawn seal]

keeps failing to make a folder

[daring forge]

Send screenshot of that d:\ dir

[obsidian prairie]

where did u make this preview?

[daring forge]

You're in the wrong dir

[frosty zealot]

Looks like you're in project > docs and its making it in docs

go up a level

[frosty zealot]

Screen Studio (its a paid app)

[obsidian prairie]

looks cool! 🙂

[dawn seal]

nope

[dawn seal]

its running at the root folder

[dawn seal]

that's where I want it to make the folder

I showed you

its running based on visual studio code project root folder

it is NOT creating folder

[frosty zealot]

Open this one

[daring forge]

In Windows explorer, go back to the project root and see what's there, not inside the IDE

[cedar skiff]

it made it in docs/

lol

[dawn seal]

It was NOT CREATING FOLDER I mean it 110% I already checked with refresh

I had to manually tell it the path

now it does

[daring forge]

We're all seeing the same thing that you are advanced by one dir

[frosty zealot]

GL

[cedar skiff]

he will find the other folder one day and realise haha

[dawn seal]

yes I did

what?

oh

[daring forge]

It existed the whole time, you were looking in the wrong place

[dawn seal]

omg

I feel dumb

ok

ARGH

[frosty zealot]

We did it team!

[daring forge]

That's one thing I really don't mistrust with an agent lol

[dawn seal]

omg

ARGH

[daring forge]

Happens!

[dawn seal]

https://tenor.com/view/adam-levine-shy-bashful-blushing-hiding-face-gif-16782169

plz enable gifs

this is 2026

speaking this to discord managing team

at least I have dis

[daring forge]

Had GPT Pro analyze this for us and this is the response: "A tighter reply from here would be:

Fair pushback. I checked the docs again, and I’d narrow what I said: if these are Local or Worktree Codex threads, OpenAI says those run on your computer, and each worktree can carry its own files, dependencies, caches, and setup/build steps, so local RAM/CPU/disk pressure can absolutely be part of the slowdown. But Codex also has plan-level usage limits that depend on task size, context, execution mode, and even number of instances, so parallel runs won’t scale linearly forever either. So I wouldn’t call it just “a hard tok/sec cap,” and I also wouldn’t dismiss local resource saturation.

One clean way to tell which side is dominating is to run the same workload once as Worktree and once as Cloud; if the slowdown mostly disappears in Cloud, your local box was a major factor. In the Codex app, /status also shows the thread’s context usage and rate limits."

[cedar skiff]

no need to burn a small rain forrest to understand that much 🤣

[frosty zealot]

rip endangered frogs

[daring forge]

oh we know inference costs aren't at that level at this point, y'all lolol

[frosty zealot]

I just seen some benchmarks comparing Mythos to GPT-5.4-Pro and they were almost on par, I wonder if then next model we get in Codex will be the pro model with a faster inference

[daring forge]

Codex is on 5.4, what do you mean by next Codex model?

[frosty zealot]

Like whatevers coming after 5.4

Pro exists but only on API, and it seems awfully slow (for obvious reasons) I was just saying maybe it will become a model you can select in the /model list

[daring forge]

I just assumed Pro was the LLM demarcation and xhigh = Pro for Codex

LLM = chatbot, I know they share the same tech

[frosty zealot]

I dont think so

[daring forge]

I mean Codex doesn't use Pro naming convention

But the Mythos result isn't really surprising, I'm sure many models have found numerous exploits. I wonder if it was just the sheer speed/volume that caused them pause.

Just thinking of a lazy programmer being like "nah, this is fine" lol

[spiral gorge]

codex windows app. up to what context size can i safely work with max intelligence of gpt?

64% full, 166k tokens used. is this fine?

[daring forge]

Codex auto compacts context, so be aware of that when you near the context window. I like to update documentation around where you are so there's less inference during/after a compaction turn

[spiral gorge]

it starts compacting at 75% or?

oh it just did yes... 💀

but i heard only great things about its auto compaction

[daring forge]

It really depends on the task you set when you near the window

I have AGENTS.md enforce a refresh of invariants when a compaction is triggered

[spiral gorge]

invariant?

[daring forge]

Things that must be carried forward and unchanged in the project.

Unbreakable rules.

It sometimes will get it wrong but then go back and correct itself.

[spiral gorge]

too much for me i do not understand 😄
i just know that the 200 usd sub is.... addicting

for ever..

🙁 help mme

i need help!!!!!! addictioneeee

[daring forge]

Keep at it and learn!!!

[spiral gorge]

how do i learn something withotu courses

[daring forge]

If you're coding seriously, I really do recommend getting fluent with context enforcement and .md documentation and GitHub

The tech you're using is the course, if you ask it!

And of course, fellow humans.

[spiral gorge]

isnt that like...... at least 1 year old meta.... outdated... and pretty much stoneage like? gpt 4 era?

[daring forge]

Which part? And I don't think so

[spiral gorge]

that contect enforcement thingy?

sounds like prompt EngEnErInG

[daring forge]

Persistent context for coding agents is still a thing. Documentation and rules are super important.

If you're using a prompt-based tech, why would you find engineering for it to be outdated?

[spiral gorge]

i have no clue what prompt based tech is.
i have just read that prompt engineering was for dumb llms years ago

[daring forge]

I mean that you prompt either Codex or GPT and then you get an output, nothing fancier than that.
Prompt engineering was perhaps more needed then when LLMs couldn't infer as much. Idk tbh.

What I do know is that if you are coding and don't want regression (things to change negatively) or drift (general unintended change) in your code, you better be using context documentation and enforcing it if you're using a coding agent.

[spiral gorge]

@daring forgeWhat I do know is that if you are coding and don't want regression (things to change negatively) or drift (general unintended change) in your code, you better be using context documentation and enforcing it if you're using a coding agent.

/// how ?

[rocky fog]

tell codex to create AGENTS.md and note any quirks about the code worth noting as you develop and find them for example
e.g. "Record any quirks, caveats, engine behavior, build gotchas, or other
useful findings we discover while developing"

or anytime you want codex to do something in specific way, you just tell it to mention it in AGENTS.md

lets say you want codex to also always commit and push to git as soon as codex changes something
then just tell codex to mention it in AGENTS.md

this is about steering it to behave and stay behaving the way you want it, or noting any special things you learn so that codex does not have to figure that out every other task or forget it

[cedar skiff]

this is what i do - it is a short hand of it, I do this for every project and build it up as i go. Obviously the exemplar stuff comes a little later because there is nothing to go by at the start. I found this stops it from reinventing the wheel and there is much less slop to worry about over time on large projects.
#codex-discussions message

[frosty zealot]

Doesn’t hurt to create a skill that automates keeping a project level AGENTS.md relevant

[daring forge]

This is good advice. I'd also add that you want to keep AGENTS.md lean, so keep other .md files that point to the important things. I like to have keywords in my AGENTS.md for when I need certain things like a commit/push I'll add "CP".

[daring forge]

Yes, keeping staleness out of the repo without too many manual turns is a nice thing indeed.

[daring forge]

You can also use the /init command to get you started.

[fleet geyser]

Does that thing where it summarises the chat to remake it actually work? And how do you use it?

[cedar skiff]

Do you mean compaction?

[signal tapir]

anyone else have problems with /status not showing you your remaining usage?

[fleet geyser]

I think so I believe there was a skill that came out that makes it easier to switch between chats to save tokens

[obsidian prairie]

was codex usage limit lowered for 20$ plan?

[signal tapir]

They removed the 2x promo I think

[obsidian prairie]

oh damn, feels alot worse

[orchid plume]

yes, 2x promo ended for Plus and 5 hour quota was reduced. Reason was to make it so that weekly usage lasts longer overall. However the secondary reason I feel is that they did it to re-balance the available compute for the Pro 5x plan, e.g. take some compute away from Plus, make room for the Pro 5x plan. Also to continue running 2x for Pro 5x and Pro 20x (well, Pro 10x apparently, it's 20x while 2x is running which feels a bit odd considering the Pro 5x plan is Pro 10x while 2x is running 😂)

[obsidian prairie]

its terrible now, feels like trial version. its not worth 20$ per month at all.

[orchid plume]

imo $100 plan is worth it though, $20 is like "here, try our products and see what you think"

[glacial shadow]

$20 month plan is to try it out a little bit and see if you’ll like it

[obsidian prairie]

lol

[orchid plume]

hmm, that's odd, I can't seem to leave a thumbs up react on messages (EDIT: Ah apparently that's because the user I'm reacting to has blocked me apparently 😂, not sure why cesium blocked me but ok)

[small vector]

I use Google blogger to store all my AI memories so other people can compare our research. Everything I work on is public.

Rickystebbins78.blogspot.com

I have no AI memory issues; I just need other people to create memories for their AI to use, so we can compare notes.

[glacial shadow]

I’ve had to block about 35 bots from this channel, it’s the same message over and over, conspiracy theories about "Free plan usage limits"

[cedar skiff]

what is the point of those bots?

[obsidian prairie]

first of all 20$ isnt free and second maybe u are bot thinking everyone else is bot. its clear people wrent happy about these changes, so why they wouldnt share same opinion?

[plucky halo]

The wait is nearly over!

[glacial shadow]

I get great usage from Plus, Pro, very thankful for openai for generous limits for each plan. Couldnt be happier

[latent willow]

Is it just me or are the rate limits on the Codex app finishing too quickly than they were some weeks ago? For context, I have the ChatGPT Plus subscription. Do you all recommend me to switch to Claude Code? Thank you for the assistance!

[boreal holly]

Boom, no more MCP for long running commands, no more unified_exec.

[exotic cave]

What that fix?

[boreal holly]

vanilla codex gives you two options:

• unified_exec=true: Agents can run long running commands, but they check on it every 5 seconds. It eats away at tokens. If you have a command running for an hour, the agent wakes up 720 times and outputs commentary (case #1). Alternatively, they can decide at any time that the command ran for too long and kill it, then spend countless hours troubleshooting it, even if the command was healthy.
• unified_exec=false: Agents run a command and are frozen. They can specify a timeout between 5 secs and 10 mins, so they usually start at 5 secs, when it times out they upgrade the timeout until it reaches 10 mins, and then give up. So they waste compute and tokens running a command with arbitrarily small timeouts, and if it takes longer than 10 mins they just can't run that command.

I hardcoded the largest possible timeout value and made it ignore their setting. Now they sit there frozen until the command completes

Previous solution was unified_exec=true with a zsh shim that produces a job ID, and a MCP that takes that job ID, then the agent uses the MCP to wait for the command to finish. It worked just fine for the previous quota calc, but the token based quota they have now, every command cost an additional 13 input tokens and 40 output tokens because of the MCP tool.

[plucky halo]

Yuck

[carmine copper]

blockchain
that's a word I haven't heard in a while

[potent mason]

Bruh I love how everyone's complaining about usage on X, when we all knew this subsidization wasn't going to last long

[glacial shadow]

Yep, and it’ll never replace USD

[latent willow]

It was just a startegy to gain users in the start and then make them dependent on it, in my opinion...

[potent mason]

Yeah definetly, but I thought we all knew that was the case? Why is everyone acting surprised? Of course it's not maintainable to offer $400 of inference on a $20 plan

[latent willow]

True! For me, personally, Codex gave out amazing stuff in the start, then on basic commands, it would use up 10 to 15 percent of the rate limits. Just not something I think is worth anymore.

I made some projects using it. I would not recommend it to anyone as a "complete full stack" developer. Although, it is good in many areas. For me, it gave up after 30 minutes of prompting.

Personally, I am gonna rebuild the things it made for me by hand. So that I actually know what the code stands for and means. I want to be in a position where I can explain the person infront of me what each line of code means. So, yeah. I would highly recommend it to someone if say they wish to make an API wrapper or something. That is just me.

What do you think about Claude Code?

[potent mason]

Have you tried Cursor?

If you're trying to learn, understand, and do things by hand you probably want an IDE instead of a CLI

I like Claude Code for front end work, it works really well with the front end design skill

But they completely wrecked our usage so I'm planning on moving to gemini

[static hare]

Pro plan limits are great

I just wish there was a mobile app! Tailscale + termius onto a friend's PC is annoying

[latent willow]

Not gemini

💀

[latent willow]

yea, but the limits man 😭

[latent willow]

i am trying to take a step AWAY from AI and do things by hand

[potent mason]

Well if you step away from AI, then Cursor is basically unlimited (they only have limits for agentic workflows)

[latent willow]

yea, thinking of doing that

[boreal holly]

Highly recommend Zed 🤓 it's really nice.

[potent mason]

Oh yea been meaning to look into that, people say it's much faster / snappier. How was your experience with it?

[boreal holly]

It's beautiful 🤩

• ridiculously fast
• has remote access built in
• lots of extensions
• native rust language support oob
• the search features are awesome

[frail meadow]

The actual agent harness itself seems lacking in comparison though

[boreal holly]

Yeah I don't use that part. The actual text editor part I care about works perfectly

[potent mason]

Thanks for the feedback @boreal holly and @frail meadow that's fine, I don't use cursor for agentic, I have Codex and CC for that

[nocturne folio]

its crazy how these new updates all happen when peter joined

[latent willow]

true lol

[latent willow]

whats dat

[clever steppe]

How do we reach the developers with UI/UX requests?

[spring remnant]

If I change difficulty mid session, will it cost me usage

[spiral gorge]

what model is best for reading text in a image as fast as possible

[frail meadow]

Codex issues on GitHub tag it a feature request

[cyan wing]

Has anyone here got coding agents to work well on well-defined, long running overnight tasks?

[glacial shadow]

Yeah I give it multiple steps to do, does everything perfectly

[compact dawn]

why is it like this all of a sudden?

white input box and output, after updating

[low cosmos]

i love codex but i hate its limits

[compact dawn]

really? doesn't it have by far the best limits

[low cosmos]

cuts me off in the middle of a session as far as i know it

i agree that it has nice limits

for its price

[nocturne folio]

change ur theme

dude i cant wait for scratchpad on codex

pls tibo

[low cosmos]

yall is running codex at xhigh actually worth it?

[frosty zealot]

Nothing beats neovim 🤓

[glacial shadow]

It’ll be like a large notepad?

[nocturne folio]

apparently its a great way to ruin usage

[glacial shadow]

Codex has great limits

[nocturne folio]

i use medium lol

[low cosmos]

i usually use it at high

[glacial shadow]

I use xhigh all the time and half the time with /fast. no issues with “Limit” here!!

[nocturne folio]

basically everytime u press enter in a notepad, its a new thread

[low cosmos]

what is xhigh actually useful for

except building stuff fully from scratch

[nocturne folio]

trust me

wasting limits

xhigh was useful back when gpt5.1 codex max released

[glacial shadow]

I’ve never hit a usage limit

[nocturne folio]

so said gpt 5.1 codex pro max didnt drop

[low cosmos]

i hit usage limits in 1-2 days bro

[nocturne folio]

then youre just not working

[low cosmos]

even with medium

[glacial shadow]

I am doing many things here, constantly, multi step and large things

[low cosmos]

whats ur plan

[glacial shadow]

Love codex

[nocturne folio]

i think if i actually love what im doing i could go through it in a session

i use the plus plan

[low cosmos]

i go through it even if i hate what i do

[low cosmos]

bro i would kill for plus

[nocturne folio]

go?

[low cosmos]

yes

[nocturne folio]

wait codex is on go?

[low cosmos]

temporarily

its on free too

but the limits are like super low

[nocturne folio]

how are the usage limits like

can u show using ccusage

[low cosmos]

sec

[nocturne folio]

https://www.npmjs.com/package/@ccusage/codex

[low cosmos]

yeah ik

i have it

[compact dawn]

april 4th 16 dollars on the go plan

seems not bad

[low cosmos]

temporary tho

[compact dawn]

$20 plan?

[nocturne folio]

holy codex mog

[nocturne folio]

yea

[compact dawn]

$20 plan is so much value

[low cosmos]

because i ran it like 11:40 pm

[nocturne folio]

not anymore

i mean

[compact dawn]

well no 2x anymore

but still

[low cosmos]

i have the same too

[compact dawn]

and they reset so often

[nocturne folio]

def more than like cursor and claude

[low cosmos]

cursor is dog

[compact dawn]

claude $20 plan: 3 prompts per 5 hours

[nocturne folio]

claude isnt thaat bad

i use 4.6 opus on med

[low cosmos]

opus burns through usage

like 3x faster

than the other models

[nocturne folio]

its like 20 prompts

[compact dawn]

same, one small prompt: 25% 5hr usage

[low cosmos]

does yall limit reset weekly

[orchid plume]

all plans have a weekly quota, so yeah it resets 7 days from the point the quota was first used. Unless of course Tibo presses that big reset button which sometimes he does. And when they reach 4 million users they will reset it manually (and for every million after that until 10 million)

[cyan wing]

sometimes when i do this it makes huge lazy shortcuts,
but maybe i'm not explaining what i need precisely enough

[low cosmos]

the second go is removed from codex im screwed

[potent mason]

Yeah I can't wait for that honestly Tibo the usage goat

[glacial shadow]

well i check it in the morning and review, make modifications as needed, etc

but it makes really, really good code on xhigh

[cyan wing]

Ever use another agent to verify its work?
or is just 1 huge markdown task list and $loop skill enough?

[glacial shadow]

i have it using agents for diferent parts, and have other agents to check it, etc, all throughout the multistep plan and phases i give it

i develop really good plans for what i need

overall completely happy with 5.4. excited to see what comes after

[low cosmos]

bro is my request really that complex

i use high reasoning and it burns through like 30% of my limit in 1 prompt

[cyan wing]

how do you orchestrate all of that without having to copy and paste between Codex chats 👀

[potent mason]

I literally do that, I try to define a task as best as possible, make it split it up in phases, add implementation details.

From there I end up with something like
/docs/migrationExample/plan.md (general plan, and how to navigate the docs for this plan)
/docs/migrationExample/phase1.md, /docs/migrationExample/phase2.md, /docs/migrationExample/phase3.md, etc.

Finally I ask it to make a checklist that hyperlinks to the proper file so it ends up like:
[ ] task: phaseName

and this checklist lives in
/docs/migrationExample/checklist.md

Which brings everything together

[potent mason]

Just create a bash loop which uses regex to check when it should stop (takes a long time because no parallelization, but on the other hand you can work on something else in the meantime)

[cyan wing]

no second agent to verify the work of the first?

[potent mason]

Well with TDD most of the time it gets it right, but I have coderabbit setup on PR's (for production workload)

[cyan wing]

but then u need an agent to write the tests correctly 😭

[potent mason]

Right but for some reason a TDD prompt goes a long way for this

In your AGENTS.md just specify that it's going to be using TDD (with a short description generated by the AI itself)

[warm pilot]

thanks for the idea, directly applied 🙂

[low cosmos]

lol i ran a codex prompt with 0% limit and its still running wth

[orchid plume]

I've heard that if you have quota left and you run out during the turn it will continue until the turn has finished

[low cosmos]

nice

[orchid plume]

can't say 100% but that's what I've heard

[low cosmos]

got lucky ig

[glacial shadow]

"codex , dispatch 10 agents to do...."

on xhigh, /fast, etc. easy

[low cosmos]

just bought plus i will never regret it hopefully]

[balmy acorn]

how long have you guys ever got codex to run for

[frail meadow]

I don't think you'll regret it but the $20 monthly plan will not last you all month for daily usage

[glacial shadow]

if you are coding you need max100 plan minimum

[low cosmos]

im not using it fulltime

just for some patches and stuff

[glacial shadow]

the 20plan was originally for the website chat

[low cosmos]

and i dont have $100 per month spare

so thatll work for nowi guess

[glacial shadow]

some patches for what? what are you patching?

[low cosmos]

my game indev whenever it has bugs

[glacial shadow]

your video game

[stable obsidian]

30mins

[low cosmos]

yes

[glacial shadow]

nicee

[stable obsidian]

good to have 5 plus accs for free

[low cosmos]

what

how

[low cosmos]

1.2 hours on xhigh

[balmy acorn]

nice!

[low cosmos]

although i use high now

and it does the job in 30 minutes for large stuff and for patches barely 5

[balmy acorn]

for me depending on what it is always spends like 30-40 mins

[orchid plume]

several hours, but that's rare and requires fairly clear prompting and a big plan

[balmy acorn]

never gotten it to do several hours.

[orchid plume]

for that kind of task I left it overnight

finished early in the morning

[nocturne folio]

o3 was actually not that vad

[topaz tiger]

anyone else experienceing an issue where codex is stuck on medium effort? on codex app, terminal, or even the marketplace integration ui in vscode/cursor
When trying to change it nothing happens, it goes back directly to medium.
I tried setting the custom config (config.toml) as xhigh but still it shows as medium
I checked for updates, im on the latest version of all

[nocturne folio]

i miss it sorta

[boreal holly]

96 hours max. Not one single agent, but between half a dozen of them working continuously before they ran out of work

[balmy acorn]

No shot 96 hours

[silver dew]

Guys I need some advice, I have been encountering problems with my newly created staging environment which I hate SO MUCH. turns out every software engineer that ever walked the planet hates them too, glad to hear cuz my current project is the first time that I encounter this curse.
To the point:

my staging env currently protects against 4 major potential production misconfigs/bugs, but the one that really causes the block is Cloudflare hostname & TLS verification. I have 2 subdomains that need to be protected but Cloudflare's free plan only covers the apex + 1 subdomain and I'm broke because I spent all the money I earned (I'm a teenager) on subscriptions, hosting, a domain and API credits for testing of my product. so I can't even pay the 10$ to Cloudflare until atleast 2 days from now and I don't want to wait -- the product is literally ready and 2 real humans tried it already. it's just this annoying caveat.

What should I do? dump the staging environment completely? do people even recommend it for fresh startups these days? or should I genuinely keep it and it's worth the wait and I can do other things meanwhile?

Thank you to all people who give assistance I really appreciate it

[brave bramble]

You using cf worker?

[silver dew]

No. but I want to take advantage potentially of the egress caching cause I'm using Render for cloud

[brave bramble]

Never used render, but is there not a default domain name provided with your app hosting? Do you really need the staging env to be at staging.xyz.app?

As for your general question about staging, it just depends on how much you really need a prod like environment. Free app/no users, maybe not that important. It's all a tradeoff

[silver dew]

Wow thanks that's actually something Codex suggested but I was too busy being annoyed at the llm that I didn't pay attention to that paragraph

[brave bramble]

Yeah, the free domains are ugly but that doesn't really matter for staging

[silver dew]

yeah it's internal then I just put authentication in front of it I guess to avoid Cloudflare Access?

maybe there's a more mature solution to this idk

[brave bramble]

If this is a standard typescript app I would seriously look at moving more of your app to cloudflare if you are extremely cost conscious. In my experience you can go very far with CF without paying for anything other than the domain name.

You could possibly skip a dedicated staging environment and just make use of worker preview urls https://developers.cloudflare.com/workers/configuration/previews/

Being able to automatically have a deployed environment per branch/pr is a very valuable thing in the world of agentic development

I don't work for cloudflare btw 😀

[boreal holly]

It was 1 shot, with me periodically adding tasks to the pool

[cobalt junco]

🥔?? 14th?? 16th??

[silver dew]

It's a python backend (which I'm still not sure was the "right" choice) but yeah I learned to accept that sometimes my language/framework choices aren't very optimal and when I could I ported. idk if I can do the same with the backend at this point and overall I think python was a pretty ok choice, I use the standard Pydantic, FastAPI and SQLAlchemy structure.. it's a pretty complex product to risk throwing at one provider in my opinion, it's a vibecoding platform

[brave bramble]

Yeah that's fair. Just figured I'd throw it out there if it was already on a convenient "edge" stack.

In my experience, there is always another "engineering" task you could do to infinitely delay what is most important: getting the software in front of users.

There is a balance to be struck, but in the early stages, getting users and feedback is more important than engineering best practices perse. I don't know what your goals for this project are though, so YMMV

[silver dew]

Yeah I learned it the hard way and I don't want to make this mistake again with this project even though it annoys me sometimes that maybe my architecture choice sucked. but really I'm curious what would be different if I chose TypeScript let's say

[boreal holly]

Experience would be the same with Typescript

[brave bramble]

Unless your app does something computationally expensive, the language choice tends not to matter that much. Typescript simply has the benefit of being pretty well supported on virtually every hosting platform/service, and has access to things like edge runtimes which can affect cost, especially in the early days.

[silver dew]

As for my goals like you said I want to get feedback fast and really put a lot of users in front of the product and not suck at retention in the first 14 days. then grow this project to hopefully a couple thousand users until June 20

[brave bramble]

What happens on June 20?

[silver dew]

Nothing it's just summer and I want to spend my time scaling further and not debating organic vs paid lol better nail my product before the holiday

[silver dew]

Oh good to know

[silver dew]

Right I actually chose to embrace python for this project because I saw a few vibe platforms that scaled perfectly fine with it and also LLMs know this language very well

[boreal holly]

Python is an excellent choice to vibe with because LLMs are heavily trained on it. AI companies use python extensively, alongside jupyter notebooks. It's the preferred language of data science (aside from Julia). I mean heck, Codex writes inline python to edit codebases, that's how prevalent it is. Not a bad choice at all!

[brave bramble]

Is this a paid app? You've got a little over 2 months to hit your "couple thousand" users goal. Staging environment is the least of your worries :D

[boreal holly]

A lotta AI research is done with python. It might be single threaded but it can be scaled horizontally and load balanced. Also if you have really computationally intensive stuff you can write it in C/C++ or rust and make bindings

[silver dew]

Yeah that's true. but I don't want to mess up when users start to log in. better spend like a week "stabilizing" the plane and using some vibes to accelerate this process. from my family members perspective they used the product and they're fire about it so atleast I'm encountering this after I have things working and not stuck in my backend cave

[brave bramble]

Any automated tests? I'd invest time there before worrying about staging. You want some assurance that as you add new functionality, etc., you're not breaking stuff people are using.

[silver dew]

or I might be delusional and I just have to ship this already and not care about editing in production

[fleet geyser]

Hey, what you guys talking about?

[brave bramble]

T. is building an app and was asking for some advice re: staging environments and stuff (is it worth the headache, etc.)

[silver dew]

Basically my entire codebase is tests and I have a dedicated internal CLI that Codex really likes so I can ensure everything flows as it should

[fleet geyser]

You should test before you release that's for sure but if I had an app I'd probably do beta testing

Because developers never test as hard as beta testers do when they're trying to break something

[silver dew]

That's the way, I want to start with a discord server so I can iterate very fast on the product and get feedback from friends and friends-of-friends

already had a server that I reconstructed in like 50 minutes it's ready

but I'm here telling you about staging.. well it's 1:35 AM, perfect time to settle these things

[brave bramble]

If your local testing environment isn't too different from production, then it isn't as big of a deal. But if you have a bunch of external dependencies/services that aren't exercised locally, going from dev -> production is obviously more risky.

[silver dew]

I wouldn't call it so different, it's different because I haven't iterated on localhost since I deployed so they're not in sync but I can fix that quite easily, I just don't see much of a reason because I already have a staging enviroment ready and configured so I guess it'd be better to just keep it

I spent like 20% of my time ensuring that Codex has a good time in my codebase, my harness is centralized and concise, built to ensure rapid navigation with deterministic tool use and custom lint rules, and my skills are fat and customized for every superpower Codex might need (or atleast the relevant ones)

thanks for reducing my uncertainty btw it's really helpful when I have to stress over so many things at once, especially because my software engineering understandings are quite around learning from cases I encountered in the limited time I'm in the rabbit hole, youtube/other media plus what I ask Codex and Claude to teach me.

[brave bramble]

Trial by fire is the best way to learn, for better or worse. Sounds like you're on a good path though. Lots of devs in the software industry probably haven't done half of what you've built so far, when you really sit down and look at their career.

[silver dew]

I'm seeing this close, I'm part of a github org with other teens that you could call senior engineers and until like 2 months ago they just won't dump Copilot! I managed to inspire some of them embrace the shiny sparks of vibe coding not only in their free time so I'm actively contributing I guess (or ruining their syntax skills)

Yes I try to constantly monitor the situation, I consider almost everything that is not contributing to the user experience distraction now, I knew staging was either good or bad for UX because it could affect production. my last project that I started in September was so dangerous because I enjoyed it so much that I essentially burnt months over something so easy that I could've built in 14 days max back then and didn't even matter for anyone because I was in the cave

[royal garnet]

codex really needs a /remote-control command like claude has

[unique spade]

you can build it for yourself before is shipped to the masses 🙂

[cedar skiff]

It's not worth it in terms of productivity.

[brave bramble]

What's it do?

[potent mason]

Doesn't remote control only work on chats you've previously started?

Just ssh into your computer and you get full computer access (including cli)

And tunnel to get localhost

[unique spade]

🙂 yea remotely accessing your computer was solved decades ago

[cedar skiff]

I nice mobile ui layer is whats missing.

It would be nice to remove all the complexity and just open an app and use it in a mobile ux way.

pick a chat, and start talking.

I assume they have an idea for it

[brave bramble]

codex has a web surface that launches cloud agents. mobile ux wasn't great as far as I'm aware, but it works

For me, the problem usually comes in when you want to actually do anything with the agent's result, like review the code or validate the solution in some way. Mobile is just not a great platform for that kind of work, regardless of how nice the actual agent UX is

I unironically think some of the new foldable phones seem well positioned for these agentic dev flows, because they (seem to) give you a bit more screen real estate to do something other than the chat part.

I've used various mobile surfaces for these agents, including claude for web. Not remote control specifically though. I don't think it matters though because the fundamental problem is the same: I don't have the ability to fully validate the agent's solution from my phone anyway

Please enlighten me

So you're going back and forth a lot in any given thread?

Did I get blocked? lol

[unique spade]

why don't you make a control plane UX app for your mobile and connect it to your codex running on your local host?

[cedar skiff]

I nice little code diff screen for mobile would be nice. A mobile first ui from the ground up is needed. I can understand why they dont do it yet. It might be obsolete before it even gets any traction

[unique spade]

and you just tell codex whatever you need that you would do yourself if you would be physically at your desktop

[brave bramble]

I just don't think the problem to be solved is "how to connect your phone to your desktop codex". The problem is the physical device constraints being unproductive for end to end software development/review/validation.

Github app has a "nice little code diff screen" and is still not something I want to use to review code.

[cedar skiff]

I dont think any one is arguing that we should be working from a mobile phone for productivity. But less friction on mobile is certainly desirable. Like not having to go to github to look at code changes.

[brave bramble]

Agreed. The codex web surface has that already though

[unique spade]

what do you think the problem to be solved is ?

[cedar skiff]

Right but that is not the context of the conversation now is it?

[brave bramble]

Sorry, what is the context of the conversation?

[cedar skiff]

We were talking about an equivalent to claude code remote. Which allows you to control terminal sessions on your machine.

The suggestion was to ssh in.

But there is bad ui/ux and friction with that

[glacial shadow]

someone on dev team hinted they were doing it

[unique spade]

there is no such thing as bad ui/ux anymore. you can make your own ui/ux 🙂

[brave bramble]

I guess I reinterpreted that as "how to use agents effectively from a phone", but you're right that feature does not exist.

[cedar skiff]

The time spend building a workable solution is not likely going to pay off in productivity.

[brave bramble]

The latter does not remove or excuse the former

[unique spade]

yea sure, i was just speaking relative to my own ux needs.
with oss if i lack a feature that is currently implementable i have my own codex do it. it s really that simple

with closed source there s a problem though

and that s why i love codex

[brave bramble]

I agree, but patching OSS is not quite the same as building custom integration software. It's not just ui/ux.

[unique spade]

i m not talking about building software for others

i thought we were talking about improving our own workflow with codex

[brave bramble]

Whether 1 or 100 people use it, it's still software that should work

[unique spade]

of course it should, but you're a dev, that s your job

lol

[brave bramble]

Do you think building codex "remote control" is not custom integration software?

[unique spade]

but let me get the point, cause i jumped in the middle of it

what do you hope for codex team to provide you?

in this respect

[brave bramble]

The other guys were asking for the remote control feature that CC has

[unique spade]

to be able to connect with codex cloud from a dedicated phone app? or to your desktop codex?

[cedar skiff]

It's a big task, it will take a good while to build out and have a nice workable solution

You can get a clunky mvp in no time. But a real solution takes time and effort even with ai help.

[unique spade]

for them it is because they need to bring something that works for millions that have different needs

[cedar skiff]

For anyone.

[unique spade]

Eric, were you ever curious to inspect how exactly all the functions that codex is currently able to do are actually implemented?

how the tool calls are made, how the steering works, what is sent back and forth during compaction

stuff like that

[brave bramble]

What does that have to do with the problem at hand?

[cedar skiff]

I mean there is an open source repo i can look at but i don't bother unless i need to learn something

[unique spade]

nothing. i moved on to a different topic

[cedar skiff]

Then usually i just ask codex to go look for me

but mostly i just work on my own projects

[unique spade]

yup, but then based on what you know how complex or not is to make a remote codex app? 🙂

most of it might already be there in the github repo already

[cedar skiff]

Because I am a seasoned SWE

i dont need to eat dirt to know it taste bad

[unique spade]

ok so you are making a general argument, that applies to adding a new feature ?

[cedar skiff]

No.

[brave bramble]

Is venusrose a bot?

[unique spade]

well you said you didn t really inspect codex repo, so you can't be making an informed judgement about codex

[unique spade]

lol

you re spending too much time with ai taylor

[brave bramble]

I'm so confused by this conversation thread

[unique spade]

now you re thinking even humans are bots

[unique spade]

i'm having some convo about codex
i already told you i drifted from the previous topic
but it s still within scope of the channel

[brave bramble]

You're right. I just found it odd how you just asked Eric if he was ever curious about the codex source code or whatever.

[unique spade]

but i can close. and we can speak about something else codex related

[unique spade]

because i remember him from few months ago from this chanel

[silver dew]

@unique spade what do you think about Mythos? is it a Mythos or can it pull the rug beneath OpenAI

[unique spade]

Eric back then was a daily chatter here, and it's easier to chat with familiar faces

[unique spade]

i'll form an opinion when i can actually test it

[silver dew]

You can?

wdym it was leaked

[unique spade]

@brave bramble maybe this T. guy is a bot though?

🙂

i don't do leaks

[brave bramble]

What's your mother's maiden name venusrose?

[unique spade]

that's a bit personal don't you think?

[brave bramble]

No it's common practice to share such information on the world wide web

[unique spade]

lol, share yours first then

[brave bramble]

Can you change your writing style?

[silver dew]

Ha. Nice one though, going personal today? would you like me to list the top 10 most personal comebacks on Discord?

Thinking...

Here's your plan:

&
&
&

[unique spade]

what style do you want? CAPS or ?

[brave bramble]

every message should have an internal rhyme

[unique spade]

or maybe a quick screenshot of our live convo is enough to make you relax and realize i am a human

[brave bramble]

I heard those AIs can take screenshots now 🤔

[unique spade]

no i can t do that i ain t gpt

[silver dew]

Is that a shot by OpenClaw?

[unique spade]

i can give you both a short logical puzzle, that you probably won t be able to solve, and neither your ai's lol

[silver dew]

You're drifting

AGAIN

[unique spade]

both being taylor and T

[cedar skiff]

I don't need to know directly about the code to understand what it takes to build a secure mobile app that connects to a personal computer with all the needed ui for a smooth experience. It's a huge task, the codex side is just accessing an api, that's the easy part.

[unique spade]

@cedar skiff what s with this nonsense.

i mean with these guys insisting i must prove i am a human

[brave bramble]

Eric, sounds like you and venus go way back

[unique spade]

Eric if i remember well was chatting a lot around septmeber october

and actually taylor you can search my history on this channel

[cedar skiff]

I should just not talk anymore, i have work to do o.0

[unique spade]

and you ll see i used to chat a bit around september october

[silver dew]

Thinking...

[unique spade]

have fun

[silver dew]

read @PLAN.md

@unique spade do you want me to implement this plan?

[unique spade]

@silver dew 3 doors. 1 good. 1 truthful guardian which only answers with yes or no, but gives error on contradiction. You have 1 question to ask the guardian to find the good door. What is the question?

[silver dew]

Lemme just ask Codex real quick

My question would be "where's the 3rd"?

[unique spade]

you won t get any answer since it's not a question that can be answered with yes or no

and that is one of the constraints of the puzzle

[lone yarrow]

Why won't Codex let us use GPT-5.4-Nano to conserve credits ? I can use it through the API using Kilo Code, but not through Codex itself. Is it coming soon or... what?

[brave bramble]

Something something about money?

[silver dew]

Ha! Nice one you definitely got me there😆

Want me to give you an idea for a cool landing page for a puzzle app?

ok too much wording

forgetit

Thinking...

[brave bramble]

Nano any good?

[silver dew]

Generating slop...

[lone yarrow]

Yeah I was using it on Kilo Code through the API and it was fine for most tasks.

Even the benchmarks show its about 80-90% as good as Mini

[unique spade]

hmm i forgot about nano, i guess it should be as an option. not sure how good it is though. did you test how it compares to the cheapest option available which would be 5.4 mini on low effort?

[lone yarrow]

Nano is perfectly fine, idk why it is not available on the official Codex app.

but I can use it through Kilo Code

which does the same agent stuff

[brave bramble]

Why can't we manually compact in codex? 🤔

[unique spade]

on kilo code you use your gpt sub or via api?

[silver dew]

OpenAI just won't take the money huh

[unique spade]

you can, but you need to change settings in the code lol

[lone yarrow]

I just use through Kilo Code paid credits but I also have a GPT Plus subscription I like to take advantage of

[unique spade]

yea was thinking if you can use that gpt plus sub via kilo code, cause i remember openai is friendly to 3rd party harness providers

[lone yarrow]

GPT-5.4-Mini can eat up the Plus limit fairly quickly. It be great to use Nano and get more usage out of my plus subscription.

[silver dew]

Thinking? are you an AI?

[lone yarrow]

Mostly came here to see if there was any plan to be adding Nano to the official Codex IDE

it would be GOAT imo

[unique spade]

i agree with you, but i m not sure if there is much official talk here

[brave bramble]

I don't know if any codex devs hang around in here

[lone yarrow]

any Codex employees in this chat?

rip

does OpenAI not engage with their community for feedback?

[unique spade]

you might have more chances commenting on x to Tibo

[silver dew]

Why don't you use GPT-5.4 High for everything and hope it lasts?

[lone yarrow]

ugh.. x

[unique spade]

he s the one i m following to see what codex team is up to

[silver dew]

seems like a pretty solid strategy

[brave bramble]

I really haven't messed around with 5.4 mini/nano. I used to kind of like using haiku for a lot of stuff because it was pretty fast, but it's hard to tell if it's worth it because getting better output slightly slower seems to be the best play.

I think the smaller/faster models are pretty good if you're driving them more directly, and just using them for the nitty gritty execution.

[silver dew]

There's no point in switching models imo, I use 5.4 and toggle between High and Medium reasoning depending on the task, but openai nerfed Codex usage limits unfortunately so the days might be over

[lone yarrow]

that is why we need Nano

[brave bramble]

Definitely over if you were on the $20 plan and want to work for more than a couple hours tops

I might try the extra usage credits to see how effective that is at bridging the 5h session gap

I think you have to buy $40 minimum though which seems silly

[silver dew]

I'm paying for three 20$ plans, might even dump Cursor and just get the new 100$ plan

[unique spade]

well 20$ doesn t sound that lucrative nowadays i guess

with 2x it was decent

[fleet geyser]

I used over 2 billion tokens in three montsh with the $20 plan

that's over $1200

[unique spade]

personally i'm one of those guys that really wanted the 100$ plan

cause 200$ was too much qouta for my usage

and 20$ too little

plus i am a big fan of gptpro for specific deep reasoning tasks, like designing formal frameworks, so it s great i can now use that on 100$ plan

[silver dew]

200$ plan if you totally embrace the vibes and hope for redemption

100$ plan if you want to be productive

🤣

[unique spade]

well i'd probably use 200$ if i'd need continous automation of certain processes

[silver dew]

for example?

[unique spade]

data processing

[silver dew]

just use the 200$ plan for that

[unique spade]

that s what i was sayiong. if i would need to do that, i would use 200$

🙂

right now i code and 100$ is enough

[silver dew]

sorry I hallucinated

Can you /compact?

[unique spade]

i use custom compaction

:))))

[silver dew]

WHY

[unique spade]

becasue it's obviously much better

vanilla is for noobs

[silver dew]

show me your config.toml

[unique spade]

😂

[brave bramble]

venus can you explain the difference between chatgpt plus and business?

[unique spade]

another test? lmao

[silver dew]

business is just billed annually

did I miss the point?

ok.

fine

[unique spade]

you don't make custom compaction via config toml

[silver dew]

When did I say so

[brave bramble]

What's better, toml or jsonc?

[unique spade]

you asked here

[silver dew]

JSONB

[potent mason]

json;

[unique spade]

i have no clue what's jsonC

[silver dew]

yes, regardless

but show me both

[unique spade]

both what 😂

[silver dew]

your config.toml and the custom compaction

[brave bramble]

jsonc is json with a schema definition

Why did codex choose toml?

[silver dew]

there's a developer called Tom who recently took an L

[unique spade]

are you asking anyone specifically?

[brave bramble]

Have you guys seen Project Hail Mary?

[unique spade]

cause only someone that actually working on codex would be able to tell you that, i guess

[silver dew]

it's the Automations

[unique spade]

that s a movie?

[silver dew]

he's running them every 30 seconds

[fleet geyser]

why do i have a feeling that a bunch of the people talkig here are bots

[unique spade]

because they are

[silver dew]

because bots sense each other

[unique spade]

i am 99.9 % T is a bot

and taylor keeps bugging me i am a bot

[silver dew]

you reached your limit