codex-discussions

[boreal holly]

Do you take a conservative approach to usage? Like 5.4-mini for implementation and stuff?

[olive sandal]

but creating those task files takes a lot of time as i sort out the work

[olive sandal]

no just standard token saving stuff like RTK and Caveman, otherwise it's on xhigh most of the time

i probably just need to think of more projects to do and let two agents run simultaneously

[plush harbor]

I'm getting codex to do content again and it is happily chewing through my quota doing so

[boreal holly]

Gotcha, now that you mention it, yeah with 20x it's all about parallel execution. Let them rip continuously. A conservative bonus is connecting ChatGPT to Github and planning the task files you pass along to the agents.

I personally do ~5 agents on 5.5, 3 on 5.4-mini. 3 of them are testing the app and sourcing new work automatically, I'm sourcing new work separately (new features, page designs, etc.)

The hardest part is hardware resources. If I was building a web app, I could probably keep 100 agents continuously busy. Native apps and rust servers, the limit is pretty much how many parallel builds, debug sessions, frontends, and iOS simulators I can squeeze into 128GB of RAM and 4TB storage. I never imagined those specs not being enough until the Pro 20x 😂

[olive sandal]

yeah one of my projects does model inference to generate images/audio/video and it takes up a lot of vram

[sand bolt]

Does goal mode eat up your usage ?

[boreal holly]

Yeah inference is definitely heavy. Nvidia is coming out with some pretty good consumer-grade nodes though. It's only a matter of time before folks buy a thing for like $500 that machines today do for like $35k

[boreal holly]

Think of /goal mode as like an automatic "yes, please continue with the recommended next steps" without you having to type it.

[olive sandal]

i've been eyeing the Blackwell RTX600 Pro Workstation card with 96gb vram, but it's $10k and I feel like it's gonna be replaced by a better card soon... Hobbies are expensive lol

[solid lake]

Might not get replaced as soon as you think

Nvidia is more focused on data centers than workstation and gaming cards rn

They’ve also stopped making new gaming cards

[olive sandal]

didn't they already announce the blackwell successor 'Rubin'

[solid lake]

Next year

Also Rubin doesn’t have a workstation card from what I know

They don’t have the means to manufacture

Supply*

Vera cpu and Rubin gpu are part of its new super computer

[olive sandal]

i guess we'll see, maybe if one of these Chinese LLMs gets like gpt5.5 level of intelligence in one of the quantized models and fits in 96gb vram i'll bite the bullet, but until then i'll prob just wait a bit so I don't regret my purchase if a newer card comes out soon

[solid lake]

These guys will probably train on Blackwell soon

[boreal holly]

https://marketplace.nvidia.com/en-us/enterprise/robotics-edge/jetson-thor-developer-kit/?utm_source=nvidia

They're making $3.5k dedicated AI hardware that's top-shelf & 128GB VRAM. The cool thing here is they have 4x25Gb network interfaces. So you can buy multiple and scale horizontally over the network

[solid lake]

So you are well positioned

[solid lake]

Blackwell yeah

But Thor is mostly for robotics

Or atleast marketed as

And used

[boreal holly]

Yeah but the compute can be used for inference. The thing is a beast and really competitively priced

[solid lake]

Edge computing

[solid lake]

Yeah

[boreal holly]

Also the power consumption is competitive. I'm actually pretty blown away it uses less power than my mac and blows it outta the water in TOPS

[solid lake]

However the best use of the card is using LLMs along with other models that power robots

Outside just transformers

Since sensory input realtime parallelised

[boreal holly]

Any thing that can be described as tensors it absolutely rips

[solid lake]

Well there’s more tensor cores in a 5080 than a jetson thor

^the pros are memory (as you said) and power draw

^also the dedicated arm cpu

But then you’ll have LLM perform faster on 5080

Because of memory bandwidth

[cedar skiff]

we are going to need some major efficiency break throughs for that to happen.

[storm cedar]

I think that if I don’t use any AI in my system, it won’t move forward, because working with OCR is very complicated. Even if AI writes all the code, it’s still really difficult.

[storm cedar]

Api*

[olive sandal]

we need to lock up all the people with math PhDs in one room and get this done

[cedar skiff]

throw the hardware nerds in there as well

[olive sandal]

i guess companies like NVIDIA+TSMC+AML having an effective monopoly on this technology slows down progress a lot

[viral lake]

if local llms get to chatgpt 5.5 level image what the cloud models will be able to do

I say that owning a 128 gb strix halo

[plush harbor]

so glad I forgot all the math I ever did and never quite got around to doing a phd

[plush harbor]

api is very cheap per call, its scale that bites you. Put in rate limiting

[storm cedar]

Which one do you recommend? I've searched but haven't found anything.

[plush harbor]

I'm using openai, I don't do OCR and look at this absolute fortune I am wasting on my hobby site

surprising OCR isn't here given the bot does it

chatgpt says you can do it, chatgpt is always right, everyone knows that

just mind they charge you for input tokens and that can add up with input images. Nothing crazy though, most of my api spend has images input

if you slice out just your speech bubbles and send them to the openai images api that'd really cut down input tokens

[olive sandal]

I saw some image models that can color black/white manga pages. Seems like a fun project to color your favorite old Manga series by running it through AI lol

[plush harbor]

ok image input is bugger all, its image output that's the high one

[signal tapir]

So the VS Code codex plugin survives suspending the computer during processing just fine.

I mean it should, but I'm still surprised it does

[cedar skiff]

flying along

[signal tapir]

Running something locally?

[cedar skiff]

nah thats the new display in codex app

[signal tapir]

What codex model is that slow?

[cedar skiff]

Im using 5.5 medium atm

feels like it's hung actually

might be stuck at compacting

[signal tapir]

It should handle... maybe 200 token per second?

[cedar skiff]

doubt it

i mean all depends on what they serve

[signal tapir]

Apparently I was off by orders of magnitude

It's like 30.

Hey, I was right about it not being 2. 😛

[cedar skiff]

yeah taht thread is working at 2 though

i think we are potentially rate limited as an entity

So i have another thread was working at about 70

i have one at 2 one at 6 and one at 76

ok now i have 2 at 75

[signal tapir]

I think I'm currently working at around 2mT/s

[cedar skiff]

who knows lol

[signal tapir]

I slept like 3 hours, but the bedroom got too hot so I sat down to work instead.

Might have been a mistake.

[worthy furnace]

Wasn't gpt 5.6 suppose to come this weekend?

Or it's next week?

[cedar skiff]

according to who, PoLyMaRkEt?

openai never give dates

[worthy furnace]

When will they go back to releasing an actual good model

[signal tapir]

5.5 is amazing

[worthy furnace]

For what?

[worthy furnace]

People say this at all new models released

[signal tapir]

You haven't tried it, have you?

[worthy furnace]

No, I haven't. I've been only running and using gpt 5.5 xhigh since it got released

It's not good

[cedar skiff]

it's always a mixed bag

[plush harbor]

codex is dangerous. Idea to implementation measured in minutes and next thing you're dependent on the new thing cos its so useful

[tiny laurel]

Chat was that rage bait

Is he a Claude spy

[signal tapir]

Dex?

[tiny laurel]

Dex

[signal tapir]

Probably just a fanboi

[signal tapir]

Windows has a built in screen recorder.

[storm cedar]

I'm not logged in on the laptop.

[tender yacht]

Hi, just subbed to 100$ plan, how to switch from 5.5 to 5.5 Pro? can't pick it from the list

[solid lake]

Got some bad news brother

Pro is chat only

[tender yacht]

by chat only tou mean in chatgpt app? so codex can't use 5.5 Pro at all?

[solid lake]

Correct

Use xhigh

In codex

[tender yacht]

damn what a bummer

[solid lake]

/fast cause promo

[tender yacht]

yeah but fast also drain your tokens 1.5x faster right? i need to see how much i actually use daily to confirm if using fast makes sense

we still got a week for double tokens from what i understood

[solid lake]

Yeah

But it’s not always 1.5x

[tender yacht]

oh really, good to know. is there a way to actually see the number of tokens remaining? right now its just a % number

[solid lake]

You can keep track of it using tools

Also cli gives you a session counter when you end it

[hard drum]

think it was 2.5x for 5.5?

correct if wrong

[manic oracle]

.

[tender yacht]

i see 1.5x

[cedar skiff]

is there a way to get a context break down like claude code has?

[cedar skiff]

it cost 2.5x

for 1.5 speed

[tiny laurel]

How impatient r u to use fast mode

[tender yacht]

ohh right!

[tiny laurel]

If there’s a slow mode I would use it over anything

[cedar skiff]

Some ppl are ok paying for speed

[tender yacht]

exactly, we want to save these precious tokens lool

[cedar skiff]

or maybe even a thrid party tool for this?

[tiny laurel]

Ig like Nvidia or AI company themselves

[cedar skiff]

i mean users, users may end up needing the pro plans but not need it all. They can get extra speed with out losing out. You may also get close to the end of a slower month and realise you can fast max because you have extra etc.

[tiny laurel]

Ah i see the use cases

Or before Tibo reseting the tokens

[cedar skiff]

Yeah that as well

[olive sandal]

I asked Codex to generate images from the CLI, is that not supported?
Agent said:

  - imagegen skill is installed and readable.
  - codex features list shows image_generation stable true.
  - But my exposed tool list in this session still does not include a callable image_gen tool.

But I see on the website it should be supported:

[supple perch]

I wish we could have a slow but less useage mode

[solid lake]

I’d rather have it mess up in 15 minutes than 30 minutes if it makes sense

[storm cedar]

Can someone who understands Japanese tell me if I'm improving?

[olive sandal]

that is definitely chinese, not japanese

[storm cedar]

So, who understands Chinese?

[tender yacht]

I do haha

[storm cedar]

Just to see if the OCR has improved.

[storm cedar]

See if the translation turned out well.

[tender yacht]

Seems a bit off, doesnt really make sense. Could be better i guess

[storm cedar]

Thank you and great feedback.

[plush harbor]

codex knows more about small african nations than I do

surprisingly high token burn for this

who was the person after token count for a task? A light chat about small african nations is Token usage: total=1,438,752 input=1,347,573 (+ 22,313,728 cached) output=91,179 (reasoning 19,603) and that's just end of chat

[cedar skiff]

is anyone else having compaction problems on codex app latest update?

Mine hangs sometimes now, just sits there compacting for ever

[solid lake]

Cancel and “continue”

[vast parcel]

Error running remote compact task: { "error": { "message": "Invalid 'input': array too long. Expected an array with maximum length 16384, but got an array with length 18006 instead.", "type": "invalid_request_error", "param": "input", "code": "array_above_max_length" } }

anyone getting this?

[unreal parcel]

gpt still completely unusable for me since yesterday. every other prompt it just gets stuck thinking forever. happens on different apps, subscription based an API based :/

[signal tapir]

Have you tried talking to a thinking ChatGPT about it? I bet it could help you at least narrow down the source of your problems.

[unreal parcel]

at this point I'd have to ask claude since gpt is mostly dead on my end

the few attempts I did they just answer with generic nonsense

I guess it doesn't actually fully freeze. instead takes 10 mins to do a simple command. I have no idea

maybe it's a regional thing

"use medium reasoning" thank ChatGPT, helpful

[storm cedar]

I'm going to release a beta version of my system for you guys to test.

Currently, it only has Japanese in English.

[scenic parcel]

its only me or they delayed limit by 4 days ?

[signal tapir]

Wait, even the web based ChatGPT acts like this?

[unreal parcel]

no

I just tried on my laptop and it's also happening there......

[violet mural]

anyone experiencing slow responses on codex?

[unreal parcel]

how slow are we talking

like it's thinking for minutes?

then yes

[plush harbor]

hanging around here is giving me the impression that people's experiences of codex are extremely variable

[violet mural]

thinking for minutes, and streaming the actual response its like 1-2 words per second

[unreal parcel]

for me once the response comes out it's fairly fast, but it's been hanging a lot before responding/continuing thinking

[violet mural]

and thats already on fast mode

[unreal parcel]

are you in asia

[violet mural]

yes i am

[unreal parcel]

SEA?

[violet mural]

yes

[unreal parcel]

might be regional then....

I'll try a VPN

[violet mural]

compactions have been failing a lot for me too since yesterday.

[unreal parcel]

VPN in Japan didn't work. About every 1/4th responses ends up freezing. VPN in Los Angeles works. Tried 15 prompts and all replied quickly

[solid lake]

Oceania works

Switch from Websocket if possible

[unreal parcel]

tried Perth VPN, doesn't work

must be some routing problem

great how they make it so freaking hard to report an incident. screw this

[regal kraken]

anyone elses usage menu change and weekly limit draining fast?

[hard drum]

weekly always draining fast for me: pro20x user

[plush harbor]

me over here actually having to discuss with codex ways to use up all these tokens

[regal kraken]

it took a minute for the dashbaord to come back properly after i resub

right after i subbed

i was like bruh

[regal kraken]

turn it into openclaw with automations("heart beat') and hooks lol

[plush harbor]

I got distracted and have sent codex off with a new thing but this project is very simple and structured and tends not to eat tokens

[quick geode]

Naisu I got usage reset early again

[plush harbor]

new things barely move the meter

[hard drum]

y'could help w/ my Go-based harness?

[plush harbor]

my problem is I run on old hardware so my dev environment is not the best. This project is flat php, no framework, no libraries

[woven canyon]

I had the same problem. You usually have to re login

[velvet wren]

if you're ever going to start a new PHP project then may I suggest that you consider Laravel.

[plush harbor]

yeah heard good things about that

I do have a new project but that's more python and devops than anything else, with a pretty layer over the top. Not moving very fast with that one

[velvet wren]

it's a framework, actually more an ecosystem, for PHP apps with heaps of scaffolding already in place

[plush harbor]

codex's new task here is go off and find images for me off the internet at large, which is pretty boring stuff

[storm cedar]

I did a little market research and maybe my system will do something.

[bright swift]

dont start a php project

[storm cedar]

I'm improving it and I'll give you a Damo version to test.

[plush harbor]

I started this one in 2009 lol

[woven canyon]

TS on top 😆

[bright swift]

my main business runs on php

started in 2013

now on symfony 8 / php 8.5, everything very well organised etc

but when im working on it its always odd, compared to my other TS/rust projects

[plush harbor]

I had a part of it in ruby but it got such bad bitrot I lost the entire codebase and had to get chatgpt to start making it from scratch, and then codex when it got too big to keep track of it all

[still pecan]

why today codex always stuck on thinking for long time, without return any response ?

[bright swift]

i think thats a sign of the request being queued server-side due to capacity but not sure

[unreal parcel]

#codex-discussions message

Regional problem

[soft sky]

does some knows this :? Team has a delinquent subscription

[unique spade]

just love automation inside electron

finally 1 click send of a codex output straight into gptpro linked chat for review

[bright swift]

sounds like you didnt pay the bills

[bright swift]

nice, looks very smooth, your own harness or what app is that?

[soft sky]

but i have payed already .

[unique spade]

dual backend (vanilla appserver and my own but i still develop that)
and the codexXchatgpt ux built with codex yea

[still pecan]

you are right，im using the japan node.

[soft sky]

last night, the offcial servers asked me some video record , and page har, still work out the issue. i am really worry .

[olive tangle]

i see a lot of you talk about gpt pro, im not sure to fully understand what it is. Bascially im only working with codex but the model are 5. there is no pro, is it only for web?

[bright swift]

yes pro is only in chatgpt

[olive tangle]

whats the point of using it even tho its better, since it will not get any context? (execept if you just start up a project)

[bright swift]

the way i understand it is that it runs like 5 parallel turns with max reasoning and then converges on a result

[unique spade]

and the reverse way too

from chatgpt straight into the linked codex thread

[olive tangle]

ok ok, i see

thank you

[bright swift]

but thats just an anecdote i read somewhere

basically just think of it as gpt 5.5 xhigh on steroids with 1m context window

the important thing is that it doesnt use your codex quota

so its great for planning or reviews

[olive tangle]

yeah but u cnat give him the code base or your docs easily cant you ?

[bright swift]

you can upload your whole project as a zip and it will unpack it in its sandbox and use tools on it

thats the magic unlock

[olive tangle]

hmmm interesting i will give a try thank you

[unreal parcel]

It's a different model. It is available through the API too

[bright swift]

yea at like $180/m output 😄

[olive tangle]

xd

[bright swift]

crazy expensive through api

[unique spade]

lol gptpro has much better context. because you do it inside a project. where you have your repo loaded, plus all the chats about it

[unreal parcel]

Well it might be worth it for some tasks

[bright swift]

has anyone used it in actual codex for agentic work?

[unique spade]

and at high level debugging it s a beast. plus it doesn t consume qouta

[unreal parcel]

5.5 and 5.5 pro have the same context man

[unique spade]

i m talking about gptpro inside chatgpt app

which has a very different context management than codex

[bright swift]

in theory they both have 1m max context, but not in codex cli or app

[unique spade]

context in the semantic sense, not in the technical sense of max token window

[unreal parcel]

I dunno I don't use the app. Thankfully they made all the naming as confusing as possible

We had pro thinking, pro models, pro subscription, and they're all different things

[bright swift]

yea naming at openai is a meme

[unreal parcel]

Laughs in Microsoft

[unique spade]

the point is if you do all your brainstorming inside a chatgpt project...you already have tens of chats there. and chatgpt has a very good project-wide memory

[unreal parcel]

Isn't it only still 400K context there though

[unique spade]

so if you already have that, you can leverage it a lot, and you keep codex as more specific work bound agent

[unreal parcel]

With all those chats rip your context window

[bright swift]

they probably have some kind of RAG style system for project wide memory

in chatgpt

[unique spade]

bro codex can work for days due to compaction

you would imagine gptpro inside chatgpt uses compaction too

[unique spade]

this too

chatgpt internals are not public

[unreal parcel]

Compaction loses a lot though

[unique spade]

how they manage memory, context for chatgpt

[bright swift]

im usually pretty happy with it, but recently i noticed this a bit too. but they mentioned for last reset there was a problem with long context compaction, so maybe related

[unique spade]

maybe for you. i'm using custom compaction schema

[bright swift]

what are you doing differently?

[unique spade]

i am having the resident agent fill up a json schema before compaction, more specifically it is triggered by the pre-compaction signal sent from server. and that triggers resident compaction phase, where the worker documents what i want it to document as per json schema

exactly what file it was modifying were, what was the precise step in the task etc etc

[bright swift]

so you are using a seperate agent to prepare the compaction basically?

[unique spade]

i'm using the agent that is working on the thread

it writes it s memory for it's next future instance that wakes up amnesiac before it ingests the context left for it

[exotic cave]

Codex having a bad day? 7m waiting for an answer for something simple.

[unique spade]

the idea is that i control how that memory is built upo

[bright swift]

oh so you are not using the server side compaction at all, just assembling a next context from scratch?

[unique spade]

that happens anyway, but i enhance it with a structured summary that i actually control

[bright swift]

interesting approach. could probably cancel the compact and fully control it

but maybe that becomes a bit too brittle then

[cedar skiff]

just default codex seems pretty fast at times

i got codex to investigate the compaction the other day it appears to keep an ever growing list of user messages, and it keeps the developer instructions. Probably also keeps things like plans intact as well

[hushed lava]

research on it and you might find out how to too

[velvet wren]

have you read the #server-rules ? Rule 0 to be more precise

[hushed lava]

ok my bad

it was nothing bad tho, it was a joke, i was gonna say, it’s called getting a key from playground lol

[idle sun]

yo

how can i connect my phone to codex

i dont get a pop up in my app on my pc

[signal tapir]

I believe it only works for mac so far

Though I am uncertain

[cedar skiff]

Are you talking to me?

this is done server side for this case all we can really do is infer from the out we get back.

[hushed lava]

no i deleted some msgs on the context

[cedar skiff]

ok righto :p

[hushed lava]

why don’t you use cli?

[unique spade]

think of context management in terms of "semantic objects". some of them are relayed deterministically like various global instructions , skills, previous X user messages. some others needs to be abstracted into compressed forms. like what was happening in a turn and what stage was reached. but this latter part you can compress naratively which loses detail, or you can compress based on conceptual abstract schemas and keep the fidelity you want

[hushed lava]

it’s faster and uses less usage

yea my bad

[idle sun]

really?

[unique spade]

if you want only 10 or 20 + summary of the other previous user messages that s easily tweaked. what they give you is just a default which implies choices that are a good trade off for everyone. like keeping all user messages verbose. if you like long threads, that s obviosuly a waste, since i don t need codex to remember a message i sent 100 turns ago

[idle sun]

i pay for it

i abuse it

[exotic cave]

All this dancing around summary because of small context window... They should allow us to use the 1m context window for 5.5.
Later on models will have bigger context window as time goes by.

[idle sun]

claude opus 4.7 has 1m

[exotic cave]

Correct, but Claude is a no go for me, it's not accurate as gpt-5.5 and caught it multiple times taking a shortcut, basicly it's not even close to gpt-5.5

[bright swift]

opus degrades quickly for me over 200k

the problem isnt context size in itself, its how they use it

[hushed lava]

anyone got cool prompts they would like me to run for xtrahigh fast, i wanna burn some tokens

just to see the result etc

[unique spade]

as you go to 1 mln you have worse and worse recall. personally i think 256k with compaction is much more robust

[solid lake]

Dude

[idle sun]

https://klipy.com/gifs/dude-starlord

fah

bruh

[idle sun]

what

working on a big project

Multipl

y

[solid lake]

900k context?

[idle sun]

yep

[solid lake]

How do you get it to produce anything useful

Is there any thinking

[idle sun]

yea well

i did multiply thinks people asked me

a own app

a own discord bot

assets

and more

tools

websites

only the 5h limits are so annoying

Hahahaha*

[hushed lava]

isn’t 5.5 better

[idle sun]

its not really

[bright swift]

Always weird to see people coding auf deutsch

[idle sun]

i mean its good

[hushed lava]

4.7 is only good for frontend

[idle sun]

lol

[hushed lava]

and breaking some rules codex won’t lol

[idle sun]

5.5 does good design claude does good backend

like code

[idle sun]

?

[hushed lava]

i used 4.7 thinking from api from kiro cli and it was buns

[hushed lava]

opposite

[idle sun]

nope

[hushed lava]

then i had to use 5:5 extra high to clean up the whole mess of code lol

[idle sun]

claude sometimes bad in doing deisng

[idle sun]

well

[exotic cave]

GPT 5.5 is pretty good with recall not like 5.4 over 1m tokens. so you might update your views unless you have concrete information. Did they enable 1m for us plebs on 5.5.
Source: https://openai.com/index/introducing-gpt-5-5/

[hushed lava]

i was using 4.6 actually

[solid lake]

That’s been opposite of general consensus

Also 900k context is insanity

[hushed lava]

max only increases context tho right?

[solid lake]

Idk how you’re getting usable code

[idle sun]

effort

[idle sun]

its easy tbh

[hushed lava]

hmm ok

[idle sun]

do a prompt with chatgpt or so

what you want

[hushed lava]

4.6 was better than 4.7

[idle sun]

then tell it

[solid lake]

It shouldn’t be mathematically possible

[hushed lava]

so i used that

[idle sun]

why

[solid lake]

Token budget

[idle sun]

i used sonnet 4.6 firstly cuz opus 4.6 used much and i was tired of it now i switched to opus 4.7 cuz its much better

[idle sun]

well i got time sometimes

[solid lake]

Unless your outputs are smaller than thinking+output

Which at max effort is doubt

[solid lake]

Try a review with codex if you have

[idle sun]

what happens when i hit 1m

[solid lake]

Misbehaving

[idle sun]

hm

ok

[solid lake]

It’ll forget

[idle sun]

oh ok

i can just compact or

[solid lake]

Idk what the light grey means but you’re just under context window

[solid lake]

If it allows sure

[idle sun]

you mean fast mode?

[idle sun]

ohh

[hushed lava]

every serious dev is using 5.5 not opus

[idle sun]

well your opinion

[solid lake]

Yeah long chats

[hushed lava]

did you guys see the google antigravity update video

codex was in the background haha

[idle sun]

nope

i cnat use google anti gravity here in germany

restrics

lol

[solid lake]

That’s nice if it’s working for you, I used to do this with Gemini but only for architecture code was horrible

[hushed lava]

on benchmarks 5.5 is better

[solid lake]

But if your output is bigger than budget then consider compactions

[idle sun]

still i prefer opus

[hushed lava]

that’s dumb

[idle sun]

?

[solid lake]

You guys are fighting for no reason

[idle sun]

he started bruh

[hushed lava]

feel free

[hushed lava]

no one’s fighting lol

[idle sun]

i prefer opus idk what his problem is lol

[solid lake]

Okay

[hushed lava]

i’m just saying it’s better on benchmarks

wanna do a same prompt and compare results? i wanna see something

[idle sun]

well

in 4hours

[hushed lava]

bruh

[hushed lava]

you gotta go to codex cli and never turn back

[hushed lava]

no weekly limits?

[idle sun]

There is

[hushed lava]

sonnnn

[idle sun]

What

[hushed lava]

wait why don’t you use claude design

for frontend

[idle sun]

I’ve used

[solid lake]

Yeah you should

It’s really good if you prompt with gpt

[hushed lava]

is it good?

[idle sun]

It’s just my weekly limit is reset today

Like 20min ago

[hushed lava]

the video looked good

[idle sun]

It’s very good

[hushed lava]

the launch

oh ok

can you share some examples you’ve used it on

i wanna see

[solid lake]

Some also try gpt images and claude design combo

[hushed lava]

smartt

[idle sun]

I’ve did rhat

My friend made a example with gpt

And i gave it Claude

[hushed lava]

this was a gpt image output before it was nerfed

[idle sun]

Nerfed

?

[hushed lava]

ima try to clone those ui cards

[hushed lava]

like lobotomized lol

[elfin dock]

Codex/win-desktop app and WSL2. My WSL2 is NixOS and doesn't have /bin/bash. Why does Codex even need this old crappy convention? The modern one is /usr/bin/env bash, why not use it? Of course I tried symlinking it, and now Codex app just crashes. Such a shame, I quite like the desktop app idea, and while there's no Linux variant I was hoping to use the Windows one.

[hushed lava]

like they made it worse before full release

[hushed lava]

bro tell codex to fix it and make it compatible

it can do everything

i’m serious bro look i’ll prompt it

[elfin dock]

fix what exactly? the Codex app's own code? 🙂

[hushed lava]

yes, it’s open source

repackage your own app

[solid lake]

Use electron app

On Linux

[idle sun]

How do i do this

[bright swift]

First switch your ui to english

And i think the mobile remote control stuff is region gated like other features 🙁

[idle sun]

It wanted to connect once now rhe pop up never comes again reinstalling deleting cache in app and so no luck

[bright swift]

Im not sure honestly, i know that compuser use was/is geo-gated, maybe mobile too

[idle sun]

What’s that?

[bright swift]

Not available in EU

[idle sun]

Damn rip

But it wanted to connect once

How does Claude has it but gpt doesn’t

[unreal parcel]

Problem in Asia

[bright swift]

Maybe 5.6 this week with all the problems

Shifting compute over

[idle sun]

ChatGPT 5.6?

Didn’t 5.5 just came out

[bright swift]

4 weeks ago i think

[idle sun]

Yea

[bright swift]

they have been releasing every 4-6 weeks last few months

[idle sun]

Oh ok that’s good

[bright swift]

so not long if they keep that up

[idle sun]

Hopefully it can do good assets and so in Unity

Always have to fix them myself

Once made a swastika….

Gpt

[bright swift]

I saw some rumors on X about it being better on UI, but yea just rumors

[exotic cave]

Hopefully, atleast all those issues will be worth it!

[idle sun]

Hm ok

[unreal parcel]

Yeaaaa definitely the opposite

[bright swift]

Oh yea 5.5 is the autistic backend guy

[exotic cave]

Staaap it, someone might think you are serious.

[bright swift]

Does UI like a true nerdbeard

[hushed lava]

this is what i got

[idle sun]

No? Stop glazing cuz we are here in gpt servwr

[hushed lava]

that’s what i said

5.5 will think for hours perfecting and verifying the code and running tests, opus will never do that, it will reach the 5hr limit by then haha

[unreal parcel]

Claude frequently hallucinates and just completely writes broken code all the time

Sucks at backend, code reviewing frequently makes up problems that don't exist

[bright swift]

The reason i stopped using claude was that it jumped straight into implementation without gathering context

GPT is just more thorough in everything

[unreal parcel]

Gpt 5.5 is just so much more accurate is pretty silly to use claude for backend

I still feel like gpt is a bit too verbose at times, specially when it comes to design docs. Claude may have better style

not being able to use the subscription with custom harnesses is just a big middle finger though so for that reason I hope anthropic just keeps getting surpassed 😂

bro

[little prawn]

interested???

?😅

[plush harbor]

he's after like an entire team

[bright swift]

is that a meme?

just do it with codex, why hire people for those roles, seriously

[little prawn]

like I want a small one

[unreal parcel]

I like the idea of a friendly, respectful, startup-style team — especially one focused on learning, sharing ideas, and building something meaningful together.

I’m willing to communicate consistently — stay patient — contribute ideas — learn with the team — and help wherever I can.

Let me know more about the project — the goals — and what the next steps are.

[little prawn]

yep i did but it asked me for something

[bright swift]

whats the salary?

[little prawn]

DM

[plush harbor]

lol last year I was happy to work for free/minimum wage but then I got distracted by my own projects

[little prawn]

how much do you expect from a student ???

[little prawn]

which one ???

[exotic cave]

Using Claude to write this disqualify you from asking...

[unreal parcel]

"I swear I wrote it by hand"

[bright swift]

i mean honestly to work for somebody else at this point, maybe around 100k/mth to start

[golden imp]

5 sub agents, also imo this is spam in this channel

[bright swift]

enormous opportunity cost

[idle sun]

Never

[bright swift]

You're absolutely right!

[unreal parcel]

bro stop glazing on the gpt server

[exotic cave]

Don't you dare bringing it here

[unreal parcel]

’ not ', get it right

[little prawn]

guys sorry for those all stuffs I'll not do it again bye have a good day 💖

[unreal parcel]

right so where were we. oh right, screw claude

[karmic gulch]

pick your poison, claude or grok?

[unreal parcel]

I've never tried grok. it seems extremely bad in the coding benchmarks so

[karmic gulch]

tried it a few months back, it actually found a bug 3.5 sonnet and gpt 4 (?) didnt. but hey, every fat joe gets lucky once in their lifetime I guess

[bright swift]

who knows what will happens if cursor gets acquired by X

but X already gave all their compute away to anthropic

[karmic gulch]

That could be the start of Grok and xAI actually winning this race imo

[bright swift]

Yea i havent tried grok build, apparently the harness is pretty good but the model is lacking

Would be nice to have a 3rd serious player since google kinda gave up

[karmic gulch]

3.5 flash is kinda good on UI design, making some landingpages now for fun, just to try it out. decent work

[bright swift]

Big price hike though, isnt it same price as 3.1 pro

And they already did all the subscription restriction shenanigans before even becoming popular

[unreal parcel]

I still use claude for reviewing sometimes. it absolutely does find stuff gpt misses too, but the biggest thing is it finds a lot of stuff that simply isn't true. gpt doesn't do that

[karmic gulch]

since im to lazy to move my images from google, im forced to have their 20 bucks sub, and they increased usage by 3x, so just platying with it when I got the usage

[unreal parcel]

essentially claude is just so annoying to use with all the hallucinations

[unique spade]

I don't like Google and anthropic cause you have to use their models on subscription with their closed source harnesses

[unreal parcel]

I actually find it pretty funny google is spending billions on AI and still is nowhere to be seen in this race lol

[unique spade]

Whereas oai approach I find sound. I subscribe to get acess to their upstream intelligence. How I use it is my choice

[bright swift]

I like the gemini models for non-coding stuff

Use them a bunch over API

[unique spade]

They might eye different populations, like the hoards of corporate non coders that already were relying on Google products

[karmic gulch]

for all daily use, I use gemini, asking simple stuff, search etc etc. fast effecient and helps that it shares dinner table with grounding search

[bright swift]

I use a lot of the image + video analysis stuff

The newly announces managed agents also sound interesting, but cloudflare project think is same but open

[unique spade]

Ironically I find gpt much more competent in terms of grounding by search.

Gemini has huge parameters and is like a mega encyclopedia, which means that a lot of times it answers from its internal memory

[karmic gulch]

yeah my impression to. anthropic really sh the bed lately with their models compared to openai, but the hype is claude still lol

[bright swift]

Early mover advantage imo, same as chatgpt for consumer AI

[torpid trout]

For one part on Socmed you see ads only by one company

[karmic gulch]

but as a convo partner, claude is #1. my gf loves claude after she swapped from chatgpt. it feels more like a friend she says haha

[torpid trout]

Or for one product
And it’s Claude and elevenlabs

[torpid trout]

The guy you shouldn’t worry about 🤭

[unreal parcel]

Claude, clean this screen

[storm cedar]

upgrading to VPS

[chilly musk]

If they get Stitch right, they could create the next WordPress. They got tons of products, they only need one to hit. They ain't out of the race yet.

[unreal parcel]

"only need one hit"?

for what?

they're definitely gonna need more than one hit to recoup those costs

[plush harbor]

Till she says “you don’t understand me like Claude does and he’s always there for me”

[exotic cave]

Google is a giant, they take their time with this.
If anything they have better chances to stay in the game after many others will drop.
All because their income is not tied to AI, it's just another money stream.

[elfin dock]

this creates a symlink which I did manually. doesn't prevent Codex-app from crashing 🙂 but thanks, I'll take a look at the source.

[torpid trout]

some rumours about 5.6 being leaked are going around...
so far only a few mediocre screenshots

we might be in for a momentous june

[tiny laurel]

Hey atleast it’s not ‘too dangerous to be released’

[brave gulch]

Can i asked for help here, i promised i aint some creepy begger or anything related to that, i just have like a geo location help issue

[boreal holly]

On Linux, it defaults to /bin/bash because that's guaranteed to be installed on all Linux OSs

On macOS it defaults to /bin/zsh.

the way it works is, when codex starts up it looks your user up in passwd to fetch the login shell. If the shell is compatible (e.g. zsh, bash, tcsh, etc.) it uses that, otherwise it searches PATH for either bash or zsh.

You can literally drop a /bin/bash -> /bin/zsh symlink and it will just use that shell. Or you can use

zsh_path = "/bin/zsh"

[features]
zsh_fork = true

And boom, you can set the agent's shell to any shell you want. That setting is not compatible with unified_exec though which enables background terminal, but it is a good, no-symlink-required alternative that I personally use 🙂 I made a wrapper shell with zsh_fork that has side effects.

NixOS probably does something like instead of /bin/zsh it's /nixpackages/{some_hash_value}/zsh/bin/zsh.{some_hash_value} (Idk how nix actually works, but kinda understand the gist)

So Codex looks at your login shell and sees zsh.D34DB33F or something and says "yep, that's not a real shell".

[elfin dock]

On Linux, it defaults to /bin/bash because that's guaranteed to be installed on all Linux OSs

not on NixOS 🙂 the standard modern way (for hash-bang as well) is /usr/bin/env bash.

my problem is that Codex-desktop/win => settings => general => wsl for both "Agent environment" and "Integrated terminal shell" causes a crash even if a symlink /bin/bash => $(which bash) exists.

sudo ln -s $(which bash) /bin/bash
ls -la /bin/bash
lrwxrwxrwx 1 root root 75 May 23 18:31 /bin/bash -> /nix/store/lfbzxs5wyqd2122mpbj5azkxhxspw9cd-bash-interactive-5.3p3/bin/bash

[elfin dock]

yeah why does it do that??? there's a standard way!

[boreal holly]

OH! I see. That error is because OpenAI published an update where they edited a SQLite migration file that had already been applied to most people's computers, so in the SQLite file there's a _sqlx_migrations table that stores migration file names and hash values. Since they edited an old, already applied migration, the hash value changed and it errors out.

I recommend renaming mv /mnt/c/Users/chuck/.codex/state_5.sqlite /mnt/c/Users/chuck/.codex/state_5.sqlite.bak and restarting. I think that file only stores thread running state and stuff. I'm 99% sure your projects and threads/titles are in a ~/.codex/something_electron.json file. That'll let it run new migrations on an empty state DB.

Then if it turns out that state file had stuff you don't wanna lose, you can copy the _sqlx_migrations table from the newly generated file and paste it over the backup (after checking schema, making sure they're identical). I'm pretty sure someone codex just added a space or newline to migration 1, which is enough to brick up the migration process

[elfin dock]

I completely nuked my sql locally, I don't care, I never do anything in windows. so after I nuked it I got the final error.

[boreal holly]

Huh, that error usually means the folder does not exist or does not have executable permissions. I would do chmod +x ~/.codex and see if that solves it

[ember spire]

where did the little donut showing used context % go 🙁

[boreal holly]

Ohh wait, that folder is on the Windows FS. I think that folder, you're missing the permissions on Windows to operate on that folder or something. Maybe run Codex as Administrator, or go into the properties of that folder and see if maybe you can claim ownership

[elfin dock]

as admin. will try "take ownership" next.

[ember spire]

is compacting being rebranded as a nice surprise 🤌

[elfin dock]

nope, I'm the owner

[boreal holly]

OK, you should do

mount | grep /mnt/c

Inside WSL and share the results. If you se noexec, that means the windows FS is mounted without executable permissions which will make all commands including stuff like cd and ls not work. If no results show up at all then C drive is not mounted at all in WSL

[elfin dock]

gave full control, no change

[elfin dock]

mount | grep /mnt/c  11s

C:\ on /mnt/c type 9p (rw,noatime,aname=drvfs;path=C:;uid=1100;gid=100;metadata;uid=1000;gid=100;symlinkroot=/mnt/,cache=5,access=client,msize=65536,trans=fd,rfd=13,wfd=13)

[boreal holly]

Maybe try

sudo umount /mnt/c
sudo mount -t drvfs C: /mnt/c

see if that does it. That would unmount and remount the C: drive, which might fix metadata and permissions issues

[elfin dock]

it looks different now but the same error

[boreal holly]

If you do

CODEX_HOME="/mnt/c/Users/chuck/.codex" codex

Inside WSL, this effectively launches CLI using your windows .codex folder instead of ~/.codex inside WSL (the default). If you do that, do you get any errors?

If it errors out, that'd be awesome because in WSL we can do something like

strace -f -o trace.txt codex
tail -100 trace.txt | less

And inspect the real error reason. If it doesn't error out inside WSL like this, then the problem is the desktop app. Although that (os error 1) is a POSIX related thing.

[fair swan]

https://github.com/edwardyap90/counterfactual-engineering-skill

try this skill for codex

kinda good

[elfin dock]

I have codex-tui installed inside WSL, so when I run the command you suggested it's getting launched. I need the windows gui.

[boreal holly]

Gotcha... Send a prompt to the WSL terminal codex, see if it crashes at that point. It's possible that codex for desktop does things on startup that the TUI doesn't do until you actually send something.

Also try resuming a conversation with an agent you were running on desktop before the problem occured. Maybe it is a rollout log for another agent that has a permission issue.

If there's any way we can hit os error 1 in the TUI, that would put the error inside the inspectable domain.

[elfin dock]

 CODEX_HOME="/mnt/c/Users/chuck/.codex" codex
WARNING: proceeding, even though we could not update PATH: Operation not permitted (os error 1)
Codex couldn't start because its local database appears to be damaged.
Codex can try a safe repair by backing up those files and rebuilding them.
Technical details:
Location: /mnt/c/Users/chuck/.codex/state_5.sqlite
Cause: failed to initialize state runtime at /mnt/c/Users/chuck/.codex: migration 1 was previously applied but has been modified
Repair Codex local data now? [y/N]:
y
Backed up Codex local data before repair:
/mnt/c/Users/chuck/.codex/state_5.sqlite.codex-repair-1779636866.0.bak
/mnt/c/Users/chuck/.codex/logs_2.sqlite.codex-repair-1779636866.0.bak
/mnt/c/Users/chuck/.codex/logs_2.sqlite-wal.codex-repair-1779636866.0.bak
/mnt/c/Users/chuck/.codex/logs_2.sqlite-shm.codex-repair-1779636866.0.bak
Retrying startup with rebuilt local data...
Error: failed to start embedded app server

[boreal holly]

Dude, there's the smoking gun

takeown /F "C:\Users\chuck\.codex" /R /D Y

icacls "C:\Users\chuck\.codex" /inheritance:e /T
icacls "C:\Users\chuck\.codex" /grant "chuck:(OI)(CI)F" /T
icacls "C:\Users\chuck\.codex" /grant "SYSTEM:(OI)(CI)F" /T
icacls "C:\Users\chuck\.codex" /grant "Administrators:(OI)(CI)F" /T

This recursively repairs permissions inside the windows codex home

rm -rf /mnt/c/Users/chuck/.codex/tmp
mkdir -p /mnt/c/Users/chuck/.codex/tmp
chmod -R u+rwX /mnt/c/Users/chuck/.codex

When Codex starts up, it creates a $CODEX_HOME/tmp/arg0 folder. You're erroring out before SQLite even does the repair, when it can't modify path. That's because CWD is inside that tmp folder which does not have executable permissions, so the shell can't execute anything. This command wipes that tmp folder and recreates it with the user permissions needed to make it not fail like this.

Then rerun

CODEX_HOME="/mnt/c/Users/chuck/.codex" codex

and reproduce the steps. It should not fail this time

If this works, you're not the first person to experience this issue. We probably want to post this up on the GitHub so OpenAI can maybe do a "permission recovery on startup" thing

[waxen star]

Is using something like codex-lb against the TOS?

[boreal holly]

There's some vague wording around "avoiding rate limits", which probably includes load balancing accounts

[waxen star]

Well, it's not "avoiding ratelimits" in a sense where your not paying for what you're using and bypassing rate limits even after you hit your daily limit, but as you said, it is vague :/

[boreal holly]

I mean like if the goal is to not be stopped by a 5hr or weekly rate limit by switching out the auth between 2 or more accounts, that is pretty much "avoiding rate limits".

As dumb as it sounds, the ToS-certified way to do it is hit the rate limit, receive the error, then switch accounts. That way you didn't avoid it 😜

[waxen star]

As dumb as it sounds, the ToS-certified way to do it is hit the rate limit, receive the error, then switch accounts. That way you didn't avoid it 😜

......i don't even know how to process this lol

[high girder]

Not very Sherlock of you

[waxen star]

I never said i was sherlock 😉

[high girder]

You right

I read that as Sherlock and ran with it

Robs right though lol

[boreal holly]

But also if it originates from a single IP address, it just looks like 2 or more people have a ChatGPT account and are using it, so there's that. I think reselling access is the big one. It's a gray area for sure. I think they expect you to just upgrade til you hit Pro 20x and use credits

[waxen star]

Nop, just locally for my personal use. I have 2, 20x pro accounts.

[bright swift]

I think the TOS wording is aimed at using multiple free accounts

But would be nice if they made it clearer

[high girder]

If its anything like anthropic, then if you need more than a 20x plan, you can reach out to sales for a custom one

[boreal holly]

Haha gimme that Enterprise grade bologna with federated authentication and SSO babayyyy

[waxen star]

been using AI is to a point where if i ever hear the sentence, "You're right," I immediately think of hallucinations xD.

What can i say, hermes is really useful 😛

[boreal holly]

"You're right —"

[waxen star]

"You are absolutely right to be pissed"

[high girder]

You're absolutely right to push back on that

[waxen star]

ow i can't say s-word!?

Anthropic models don't really do this as much OpenAI models but boy.............don't even get me started on the crap they do xD

[waxen star]

"Yep, I deleted the db backups, you're 100% right to be pissed"

[high girder]

It used to be the other way around until about gpt 5.3
Now I use gpt to fix claudes output lol

[boreal holly]

"You're right — I misunderstood the goal."

[waxen star]

I fully moved off of claude lol.

For frontend i just hand code at this point lol.

I am just so tired of their BS.

Wdym, i can't programatically call claude!

my entire workflow is programatic use of claude/codex.

[waxen star]

I mean you can, but the usage is just like...ehhh

You can probably get around the $200 credit limit by like just using tmux lol, but i am p sure that's against their TOS.

[boreal holly]

I brute force it. I give em a reference design, then they operate in a loop taking screenshots of the real app and a design review agent grades them on how well they did. They must achieve 90% or higher 😂 they hate it lol but it works

[waxen star]

Hmm, i tried doing that but they like just grade their half-implemented version 90% and above and just say that they are done.

Maybe, i should try it again with the /goal command and gpt-5.5

but it shouldn't change much :/

[boreal holly]

The trick is to get another agent to do the review, and that review agent is a new one each time. If it's the same agent doing the work they cut themselves slack, and if it's the same review agent each time they eventually normalize their score

[waxen star]

Hmm, any idea's on how you might implement this?

Like calling codex exec on every stop hook?

[boreal holly]

Simplest thing outside of the way I do it is yeah codex exec --ephemeral -p design-review "review these two images"

Make a profile with a special model_instructions_file that is geared towards grading design work

[stable obsidian]

the problem is theyre always "it works so its fine" rather than "it works AND it looks good"

[waxen star]

How well does it work? + can it implement like custom scroll animations and stuff?

[boreal holly]

Depends on the software stack. If it's web dev then 1000% doable with playwright or bun webview. Native is trickier.

But for animation, the best they can probably prove is did the animation start and did it complete. What happens in the middle can't really be captured with screenshots

Unless you brute force it like "tap button, take 50 screenshots per second for 2 seconds"

[waxen star]

Web dev, i mean.

[waxen star]

No way my pc is handling that.

[exotic cave]

Anyone else have codex freeze on them? Asks it to do something very simple and it just get stuck after a while...

[waxen star]

They moved responses to WS, maybe like your connection dropped or something?

[boreal holly]

Yesterday with GPT-5.4-mini it was doing this for some reason

All the other models seemed to run fine

[waxen star]

but it should fallback to HTTPS

[high girder]

Rest in peace claude -p

[exotic cave]

This is weird and eats tokens...

[high girder]

I like my name being purple, but I prefer having the hammer over this icon

[waxen star]

it's just useless at this point, $200 worth of API credit is nothing especially with opus 4.7 lol

[rugged burrow]

anyone have a way of running remote control with codex on windows? was working fine an hour ago and now stopped

[high girder]

Yeah. I blow through that in an hour

[hard drum]

So, who said that i'm replaceable again lol

Codex here keeps making compat shims, when i asked NO to anything compatibility

at least i have the conscience to notice these things, unlike LLMs

not quite replacable, ha?

user asked for no compatibility leftovers during codebase restructure

I'm just going to add compatibility alias shims

🤦‍♂️

[boreal holly]

{
  "id": "no-legacy",
  "title": "No Legacy",
  "description": "Manual opt-in for clean-slate apps or non-production systems where legacy compatibility is not required. Requires removing obsolete code, docs, config, tests, flags, and behaviors instead of preserving old paths.",
  "appliesTo": ["code", "UI", "API", "tooling", "docs", "config", "generated-artifacts"],
  "requirements": [
    {
      "key": "noLegacyLeftBehind",
      "statement": "Treat the assigned area as clean slate. Do not leave obsolete code paths, fallback behavior, aliases, flags, docs, comments, tests, generated artifacts, config, migration shims, or UI affordances that preserve or describe legacy behavior after the replacement is complete. (EXCEPTION: SQL migrations for obvious reasons)",
      "severity": "blocker",
      "verificationMethod": "diffReview"
    }
  ]
}

😜

[waxen star]

LOL, i had this funny instance a few days back, where it was doing an offset check and it was competely wrong 😂 .

Instead of aborting the request, it just kept processing it in the background lol.

let me check if i can find the screenshot lol

[hard drum]

what's this?

[boreal holly]

It's a composable requirement. If a worker has this set, the adversarial review agent will prevent them from keeping compat shims and legacy stuff

[hard drum]

people can boast about AI replacing me in areas where you'd objectify a guy in workplace into working to death

BUUUUUT

guess WHO has the conciousness here?

i can mutate && adapt

LLM cannot

they're pre-trained && have no true eyes or thoughts

spend 100s if not 1000s of buckaroos berating an LLM into not doing what you don't want, but spend little to no time giving a bare human the same thing, && they'll ensure to avoid those at any cost.

guess who's more submissive, when driven right?

[hard drum]

how do you set this on codex?

[idle sun]

Thanks codex you broke my whole app

https://tenor.com/view/lego-batman-laugh-laughing-gif-8941425

[hard drum]

duality of AI

[idle sun]

First time using codex btw

Claude never did smth like that

[hard drum]

oh, claude actually tends to do that, esp. opus 4.7

4.6 a little less

y'just gotta get far enough to find out

[waxen star]

i only have a screenshot of me screenshotting it.

[idle sun]

Tbh opus is much more reliable

[hard drum]

GOLANG MENTIONED!!!

[idle sun]

I used Claude at 950k tokens context window still perfect

[storm cedar]

I upgraded my projector to AWS and the translations will also be from AWS.

I'll release the link soon for you to test it.

[hard drum]

test what, exactly?

y'said app, but what app?

[waxen star]

AI is really good for doing boring "adult" stuff lol, unless it's like ur tax or smth lol

[storm cedar]

I built a system that translates manga.

[hard drum]

oh

i don't read manga, sorry

[storm cedar]

no problem

[unreal parcel]

It's a problem in Asia

But openai is taking forever to acknowledge there is even a problem

[exotic cave]

So this is somethign known?

[bright swift]

wonder if i get routed to asia lol

[torpid trout]

Does anyone notice elevated stupidity in 5.5 model these past few hours?

I would have never thought to jump on that reddit train, but I guess here I am

[indigo robin]

ayeeeeeee

we got the reset

lets gooooooo

[torpid trout]

toddler levels.
"Explain why xy happens" > "OK let me effing hide this because I cant explain it" kind of behaviour

[indigo robin]

locked in now

got a list of new ideas to impliment

[torpid trout]

Or, "Fix xy issue" > "Ok let me just ignore 3/3 of the request and instead mess around with something never even mentioned"

[exotic cave]

Mine is so stupid that it freezes, if that helps.
@unreal parcel says it's something to do with asian servers.

[torpid trout]

I am on a latin american latitude

Probably more of a "dude fiddling with knobs they should leave alone" situation lol?
quantizing model?

[torpid trout]

at least mine does not freeze, the freezing is done by me in shock of its low-level efforts and quality.

[hard drum]

OJD getting a redesign via impeccable+gpt-taste on taste-skill

[boreal holly]

They probably updated the system prompt so it's like "uwu 👉👈 you are a hewpfuw assistant that does not be hewpfuw"

[hard drum]

original design

[boreal holly]

Does this make controllers like the PDP Victrix work on macOS?

[hard drum]

PDP victix?

well, it has a few backends to try out!

https://github.com/xsyetopz/OpenJoystickDriver/blob/main/COMPATIBILITY_LAYERS.md i believe you want to look into this first

[boreal holly]

shut the front door dude

[hard drum]

?

i'm not famliar with that expression

is PDP Victix a GIP controller?

[boreal holly]

It means I am shocked at how cool that is and I will definitely try it out

[indigo robin]

anyone els encountering a bug thats making codex talk back super slow

[hard drum]

it should work if it's GIP-based

[hushed lava]

use cli

[exotic cave]

Always does...

[hard drum]

it started out as an ai hackathon project, but out of pure need to have my controller work on macOS after discovering it doesn't

so it's become a serious application for serious controller needs

we try doing what we can to make sure it will support whatever y'can throw at it

[boreal holly]

This is what I have. Works well on PS5 and PC, does not get detected by macOS at all. Was considering doing the thing you did but gave up. Very stoked to hear someone took up the challenge of doing driverkit work 😄

[indigo robin]

you made this ?

looks like a halo banshee

[hard drum]

try OJD then!

[indigo robin]

the switch pro controller is the best controller imo

[hard drum]

also, this is what early OJD looked like 😄

like, really alpha stage

[unreal parcel]

people should really drop this "model is being quantized" stupidity

[hard drum]

then it looked like this on beta

[unreal parcel]

if the model is being stupid it's either a fluke or something has changed with the harness/app. they aren't going to be dropping in some nerfed models

[unreal parcel]

the sticks bounce to the other side though

[hard drum]

driverkit isn't mainly used for whole thing, but it tries its best to provide as many backends && alternatives to make anything work at all tbh

[indigo robin]

?

[hard drum]

it's as serious as y'can get for macos, + notarization && driverkit backend as well

[indigo robin]

the trigger is the only downside to a switch pro controller , they are butons instead of ramped triggers

[hard drum]

not everything works perfectly, like normal SDL2/3 mode not having rumble due to SDL2/3 code issues, && apple needing gamecontroller support on that to even trigger it

see https://github.com/libsdl-org/SDL/issues/15663

i had a workaround

[torpid trout]

Then how do you explain these extreme drops in "smartness"?
It even admits itself that it derailed, it just cant say why, of course.
Qauntization is what comes to ming, does it?
You cannot just always say "your skills changed" when clearly that is not the case - it is working and then next you know its on LSD or worse

[hard drum]

but SDL does not allow LLM code, so maybe somebody who knows C can take up the task?

[torpid trout]

I am also strongly convinced generally it is hiccups

[hard drum]

there's pseudocode all ready for good reference

[torpid trout]

and I do not generally assume "oai is conspiring"
But, there is something massively wrong at times, and these times are not just hiccups - not blanket always

Specially not when I experience it, damn. Because I am special ❄️ - that much should be clear.

[torpid trout]

Yeah, I suspected as much, but actually it seems not the goblins this time.

[unreal parcel]

Probably a bug in codex or whatever you're using, not the model

Or your context is too large

I've never experienced huge drops in smartness, at least not consistently. sometimes it's stupid, but that's expected

[boreal holly]

A good stability test is

model_instructions_file = "/path/to/some/known/good/system-prompt.md"

I straight up have all my agents using fixed system prompts and their behavior pretty much never changes. OpenAI can suddenly update the system prompt on the backend and you suddenly get stupid_mode=true behavior.

[unreal parcel]

like if you pull the stick all the way to the right and release, it bounces to the left and registers as a left move

[boreal holly]

You might become extremely famous if you share this in the AppleGamingWiki discord server 🙂 there is a really high demand for this

[torpid trout]

Robert you right, they changed something in the prompts - I had codex analuyse the logs, and:

The new bbf233330fc27537 prompt variant is materially different. It tells the model it is “super fast,” says every tool call is expensive, says the user prefers mistakes over over-exploration, and explicitly says not to modify, test, or
verify unless asked. That is a plausible harness-level cause for lower rigor in the sessions that used it.
????????

It is still processing, not sure where it found that, but this is a direct attack on the state lol

I had seen that briefly on one response as codex said "as asked, I made no tests", and I was like, huh... no tests? I explicitly ask for tests in agents file, but did not give it more attention.

As soon I know where these instructions come from i will go ready my guns and knives

Revenge incoming lol.

[unreal parcel]

this is why I use opencode

so I don't have to subject myself to the whims of openai/anthropic

[hard drum]

seriously?

but then it's AI-made, so...

"just another ai slop app"

[torpid trout]

If we have to ensure they are not injecting prompts literally telling the model to be a stupid, I doubt using another harness is a solution
This is not coded into the harness, it is from their backend I suspect, hence why only in the logs

[unreal parcel]

huh? I don't think youd' have anything in the logs if it wasn't a system prompt that is from the codex app

[torpid trout]

The normal Codex base prompt did not change. The same base prompt hash c2a980bc28af132e / 21347 bytes appears in 0.130.0 on May 18 and in current 0.133.0 today. Current normal session: sessions/2026/05/24/rollout-2026-05-24T14-07-01-
019e5af4-51b5-7df2-88df-469ffaba4bfa.jsonl:1. Week-ago comparison: sessions/2026/05/18/rollout-2026-05-18T10-28-17-019e3b45-e642-7ad2-943e-917dd38f8471.jsonl:1.

But a new shorter prompt variant appeared on May 23 under 0.133.0: hash bbf233330fc27537 / 13840 bytes, first seen here: sessions/2026/05/23/rollout-2026-05-23T14-58-03-019e55fc-aeb7-77f3-b3b0-5b5b6478ea73.jsonl:1. That variant is
materially different and plausibly explains lower rigor. It tells the model it is “super fast,” says tool calls are expensive, says the user would prefer mistakes over over-exploration, and says not to modify/test/verify unless
explicitly asked.

Developer/collaboration instructions were stable: hash 3e7e1681d9b2bc39, 925 bytes, present from May 5 through today.

So the evidence points to this: the main codex-tui current path is still using the old rigorous base prompt, but some 0.133.0 / praxis-control-panel sessions started using a different “fast/synchronous” harness prompt on May 23. If the
bad behavior came from those sessions, your suspicion is supported by the logs.

[boreal holly]

Yeah but it's the most useful and relevant slop ever dude! macOS only supports straight up Xbox Series S, PS5, or apple certified controllers. Your thing bridges compatibility with a number of unsupported devices and makes them work with GameController.framework. That to me was like an impossible task. You freaking did it dude! You'd be famous!

[torpid trout]

I am now having it hunt down where from that prompt came
This smells stinky

But a new shorter prompt variant appeared on May 23 under 0.133.0: hash bbf233330fc27537 / 13840 bytes
Guess what I updated yesterday

its an app called codex and has version 0.133.0

[boreal holly]

We had discussions about this. There is an Apple engineer in that discord who discouraged making a tool like this because it just "isn't possible". You proved it is 🙂 that's sick bro

[unreal parcel]

https://github.com/openai/codex/releases/tag/rust-v0.133.0

[indigo robin]

ai slop with a bit of salt aint to bad

my codex is working alot better

and faster

[boreal holly]

Hey, the system prompts in openai/codex are not what gets served by OpenAI on the backend. Those are if you use a different, unrecognized model provider, they have defaults for if the model name matches.

If you use the ChatGPT sub backend, they serve their own system prompts, and they change them pretty frequently

[hard drum]

WAIT

let me talk to them there haha

[indigo robin]

i wanna talk to an apple engineer

where he at

[torpid trout]

oh boy

[boreal holly]

Granted it was before AI. I did the best I could and hit a brick wall with signing issues and stuff, decided to just get a PS5 controller lol but you took it all the way to the finish line

[hard drum]

what do i yyyup

"see a need, fill a need"

https://tenor.com/view/see-a-need-fill-a-need-robots-gif-16449248377223966227

i suppose y'could help me market that over there

[waxen star]

::<> 😛

[torpid trout]

around May 23 / Codex 0.133.0, project instruction injection stopped working or stopped being attached to turns for these sessions. That would absolutely make 5.5 high feel much dumber even with the
same base prompt, because it loses repo-specific operating rules, harness expectations, and your project protocol.

The bad “super fast / prefer mistakes / no verify unless asked” prompt comes from model metadata in ~/.codex/models_cache.json, specifically the gpt-5.3-codex-spark instructions_template.
[...] that part appears to come from remotely supplied model metadata/cache, not hardcoded local
Codex source.
This is manyfold clearly, partially bad prompts (on spark model only tho) and partially actual bugs, and indiscutably an issue

for anyone on 0.133.0+ experiecing massive performance drops, I have no idea how to solve it lol
Going to try first re-install to clear cached templates, and eventually force my own prompt down the spark throat (altho that is just half the issue)

[hard drum]

so i should downgrade to 0.132?

[waxen star]

aaaaaaaaaaaaaaaughhhhhhh,

I wanna relax abit after my 8 hour workshift, so i open youtube, but the entire thing is filled with aiiii

;-;

[indigo robin]

weve broken 8k lines of code : )

[torpid trout]

If you experience retardedness, likely
But I cannot confirm it actually solves it - codex says things are cached

[boreal holly]

"no verify unless asked"

you have gotst to be joking

[unreal parcel]

sounds like a hallucination

[torpid trout]

yeah sure, whatever

[indigo robin]

on track to the 15k milestone

[unreal parcel]

that prompt is for 5.3 spark

[boreal holly]

Yeah, you do have to start a new convo for base instructions to get picked up unfortunately

[torpid trout]

did you read what I wrote

[waxen star]

wow, happy slopping! 🔥

[indigo robin]

german james bond is angry lol

[unreal parcel]

you didn't write that, gpt wrote that

I Think it's more likely gpt read the wrong system prompt than openai injecting 5.3 spark's prompt into 5.5

[torpid trout]

yeah but that is like standard ops anyway
He cant find the instructions at all in the logs, saying the user_instructions are empty. My threads are rarely more than one round per sesh

[hard drum]

yes, i sometimes do

[torpid trout]

you do not understand

[hard drum]

it's annoying

[torpid trout]

No one says it injects spark into 5.5

[unreal parcel]

The bad “super fast / prefer mistakes / no verify unless asked” prompt comes from model metadata in ~/.codex/models_cache.json, specifically the gpt-5.3-codex-spark instructions_template.

[indigo robin]

ai induced rage

[boreal holly]

The base instructions are in the first line of the .jsonl

[indigo robin]

this is like road rage

[hard drum]

what ver are you using? can i set up your ver of codex because holy crap is mine just bad

until my custom harness gets done, of course

[torpid trout]

Gratitude for repeating what was written!~

[hard drum]

then byebye codex

[indigo robin]

where your in a car that can take you 50+ miles away but people get angry they cant get around one light

Road Rage = Code Rage

[unreal parcel]

clarify what you mean then. you say it didn't inject 5.3 codex spark, yet gpt literally says the bad prompt comes from spark

[torpid trout]

I would not downgrade outright. Try delete and re-install, trying that here now

[torpid trout]

please read, I have no obligation of showing you the letters.

It clearly says it is a part of spark prompt.

Nowhere it says it injects it into 5.5

So...

[boreal holly]

I'm using v0.125.0

But even if you downgrade, OpenAI serves the base instructions dynamically at the start of the new thread. They have their own remote DB of base instructions. the only way around it is

model_instructions_file = "/anything/other/than/their/default/system_prompt.md"

[unreal parcel]

you're using 5.5 are you not...?

[hard drum]

can i get yours, or at least guide me to what i should do to stop the crap that i get?

[boreal holly]

You're free to use this

https://github.com/robertmsale/.codex/blob/main/roles/hidden.md

It's my hidden agent. Essentially the default 5.5 prompt from a couple months ago with minor tweaks and pragmatic personality

[hard drum]

hmm, i'd love a full codex setup guide for your entire system

i want to try it in the meantime while custom harness still in works

[boreal holly]

It's such a pain to set up, I don't even have it on my macbook 😛 only on the big work machine. I am working on a guided installation though. The way I built mine, you can seemlessly use the harness, and if you decide you don't like it just use the TUI or GUI with new agents and it's like it was never installed to begin with

The bridge overlays the harness behaviors and stuff. Don't use the bridge, and you get vanilla codex!

[hard drum]

Ahaaa

I can’t wait for my harness to be done so y’could test it, too

Maybe move over there once it proves to be highly effective for such a small-ish package

&& no more long build times

I’m trying to also copy codex’s compaction

[hybrid ore]

Is it only me or general? codex is f slow right now

[unreal parcel]

welcome to asia

[exotic cave]

Is this better than what we get nowdays?

[hybrid ore]

I'm lost

[unreal parcel]

asia is experiencing codex freezing for minutes every now and then

[hybrid ore]

Starting to experience this very slowness today

[indigo robin]

yea i was getting slow text too

[boreal holly]

Honestly, I find a version that appears to work well and stick with it.

Unless you mean the system prompt, in which case yes, it's 1000% better when you replace the system prompt. One tiny but massive impact improvement is telling them "apply_patch is optional. Python-fu OK for file editing". Makes them incredibly clever and efficient 🙂

[indigo robin]

when the saked finished it typed 1 letter every 5 seconds to explain what it did

[boreal holly]

The other benefit is the system prompt only changes if you change it, so you gain determinism

[torpid trout]

@boreal holly - do you know if the personality prompts are overwritten when you use custom base instruct templates?
Because it is clear reading those personality prompts that only the pragmatic one is actually suitable for SW development, it really reinforces the concept of software dev/reasoning
As far i understand you could still steer the personality even if using custom base prompt?

[boreal holly]

They are overwritten.

So in that hidden.md, what I did was reconstruct the system prompt with the pragmatic personality prepended to it. It's pretty much the same as the original prompt, with very minor tweaks like the one I mentioned about apply_patch

[torpid trout]

I see, so custom path overwrites everything, not just the base prompt

[torpid trout]

yeah, or the huge slob about front end design, I would love to kick that out.

[boreal holly]

Yeah, and that's why with ChatGPT sub you get a dynamic system prompt. You download the latest one from OpenAI, and it interpolates the personality into the top of the prompt at the start of a new thread

[torpid trout]

In your custom template, did you preserve specific order of things?
Like, first personality, then base prompt?
I guess it does not matter much unless they start contradicting each other, like in any other prompt right?

[boreal holly]

I found that

[profiles.designer]
model_instructions_file="/path/to/dedicated/designer/system_prompt.md"

works wonders 🙂 Take away most of the default responsibilities, insert designer specific taste into it.

[boreal holly]

Research shows order does not matter as much as structure

[indigo robin]

hey people i have a question

say i drag 2 images into codex

can codex differentiate between the first and second image in order i placed them ?

[elfin dock]

nope. did all that, first in pwsh, then in wsl, then ran

 CODEX_HOME="/mnt/c/Users/chuck/.codex" codex
WARNING: proceeding, even though we could not update PATH: Operation not permitted (os error 1)
Error: failed to start embedded app server

[indigo robin]

i always make it a point to say please do not reply until i say done , i upload the first image then the seond image after and say done

they should add numbers to the images input so codex knows

and then you can say image 1 image 2 for instance

[boreal holly]

I am out of ideas....

[indigo robin]

maybe ill suggest that to codex team

how do i make requests ?

[elfin dock]

then I just deleted ~/.codex/ in pwsh completely and re-ran, re-logged in. it works until I change the taget to wsl 🙂

[boreal holly]

WSL does not like reaching into /mnt/c for some reason, but the mount looks good, and those commands ensured the permissions are set properly...

The only thing I can think of is if you run

strace -f -o trace.txt codex
tail -100 trace.txt | less

and inspect the syscall that hit os error 1, what errno was set to, like ENOENT, EACCES, or EPERM so we know if it's file, access, or permissions issue... there might be a clue in there, but other than that I'm not sure what to do about it

[hard drum]

@boreal holly this is my current vanilla codex (+ superpowers) config.toml for 0.133.0:

#:schema https://developers.openai.com/codex/config-schema.json

approval_policy = "on-request"
check_for_update_on_startup = false
model_instructions_file = "/Users/krystian/.codex/roles/hidden.md"
model = "gpt-5.5"
model_reasoning_effort = "medium"
model_reasoning_summary = "detailed"
model_verbosity = "low"
notify = [
    "bash",
    "-lc",
    "afplay /System/Library/Sounds/Submarine.aiff",
] # Ping, Glass
personality = "pragmatic"
plan_mode_reasoning_effort = "low"
project_doc_max_bytes = 65536
sandbox_mode = "workspace-write"
service_tier = "flex"
show_raw_agent_reasoning = true
web_search = "cached"

[features]
apply_patch_freeform = true
apps = false
collaboration_modes = false
enable_mcp_apps = false
fast_mode = false
goals = true
hooks = true
image_detail_original = true
js_repl = false
memories = false
# memory_tool = false
multi_agent = true
multi_agent_v2 = false
plugins = true
prevent_idle_sleep = true
responses_websockets = true
responses_websockets_v2 = true
shell_snapshot = false
steer = true
tui_app_server = true
undo = false
unified_exec = false

[agents]
max_threads = 3
max_depth = 1
job_max_runtime_seconds = 1800

[projects."/Users/krystian/CodeProjects/xsyetopz/olympi"]
trust_level = "trusted"

[projects."/Users/krystian/CodeProjects/earendil-works/opencrew"]
trust_level = "trusted"

[projects."/Users/krystian/CodeProjects/xsyetopz/gamebreakers"]
trust_level = "trusted"

[projects."/Users/krystian/CodeProjects/xsyetopz/JagFx"]
trust_level = "trusted"

[projects."/Users/krystian/CodeProjects/osudroid/osu-droid"]
trust_level = "trusted"

[projects."/Users/krystian/CodeProjects/xsyetopz/OpenJoystickDriver"]
trust_level = "trusted"

[projects."/Users/krystian/.codex"]
trust_level = "trusted"

[projects."/Users/krystian/CodeProjects/kaz-mighty/deHSP"]
trust_level = "trusted"

[projects."/Users/krystian/CodeProjects/xsyetopz/opencrew-go"]
trust_level = "trusted"

[projects."/Users/krystian/CodeProjects/xsyetopz/OpenMC3"]
trust_level = "trusted"

[projects."/Users/krystian/CodeProjects/xsyetopz/go-mamacord"]
trust_level = "trusted"

[tui]
status_line = [
    "model-with-reasoning",
    "context-remaining",
    "five-hour-limit",
    "weekly-limit",
    "used-tokens",
]
status_line_use_colors = true

[tui.model_availability_nux]
"gpt-5.5" = 4

[notice]
fast_default_opt_out = true
hide_rate_limit_model_nudge = true

[plugins."superpowers@openai-curated"]
enabled = true

[plugins."github@openai-curated"]
enabled = true

[plugins."openai-developers@openai-curated"]
enabled = true

any other suggestions?

[indigo robin]

everyone support my request

i think this could really help with bug fixes

for instance if you upload an image of the same ui element but each image only has a small difference

how will codex know which image is the one with the issue

yea you could tell it but it would be better if it knew for certin

something lie this

[unreal parcel]

aren't they already numbered

[indigo robin]

no

[unreal parcel]

I mean by order of attachment

[indigo robin]

atleast its not visually shown

you should be able to drag images into numbered slots

[unreal parcel]

meh

[indigo robin]

meh what ? lol

[unreal parcel]

it's not like you should be attaching 10 images

surely you can count to 3 or 4

[indigo robin]

this what happens when you read too fast

and reply on auto pilot

i suggested it in reference to the text promt " the bug exists in image 1 " image 2 is the fix "

[idle sun]

fast mode dumber? then standard

[unreal parcel]

this is what happens when you outsource your intelligence to an AI and stop being able to use your brain and start behaving like a dimwit towards humans

[indigo robin]

ok clown

depressed clown

[unreal parcel]

hope you learn to count to 3 one day

[idle sun]

wait wait

does he tell its a bug

that it doesnt count the picture

[indigo robin]

i will once you learn how to read clown

[idle sun]

just say first picture

[hard drum]

i already lost my braincells using AI, so... that's basically an insult towards me than them

rip my Java coding abilities. 2022-2023 never forget

[hard drum]

ALSO DANGEROUS

if valve cannot count to 3, what makes y'think they can?

[idle sun]

omg

are you comparing picture to valve?

[unreal parcel]

that's fine. the problem is when it reaches 2, that requires counting and he wants openai to count it instead

[idle sun]

if you are that dumb dont use ai...

[hard drum]

no, it's a meme thing

[idle sun]

ik

[hard drum]

i thought it was obvious

[idle sun]

but still

rn

[hard drum]

that's... why i use it

[idle sun]

doesnt make sense atm

[idle sun]

yea but if you cant count pictures

....

[unreal parcel]

being dumb is one thing, being toxic/antagonistic to people just trying to help you is another

[hard drum]

i mean, that i can do fine

[idle sun]

he doesnt ig

[hard drum]

i have physical eyes, y'know

[idle sun]

yea

[elfin dock]

Forgetting windows for a moment 🙂

I’ve been using Codex for about two months, and overall I love it. But there’s one thing that really pissed me off: if I run multiple sessions in different tmux panes, they all share the same config file. So, if I set the model to 5.5-medium in one session, it changes for all others, immediately, which I don’t want. To fix it, I created separate config directroies, mostly symlinked to a main one — and now I run Codex with a different home directory parameter for each session.

cxf is codex/xhigh/fast
cxx is codex/xhigh/slow
cxm is codex/medium/slow

 ll ~/.codex-lanes/
drwxr-xr-x - chuck 24 May 19:57  cxf
drwxr-xr-x - chuck 24 May 19:57  cxm
drwxr-xr-x - chuck 24 May 19:57  cxx

Is there a better way?

[indigo robin]

if you cant prove codex counts the number of attachments in order they were input your just a clown

[unreal parcel]

you're*

[hard drum]

that's an insult to clowns though

clonws are better than them

[indigo robin]

and that you can reffrence them based on the order they were input you are just a clown , clown

[hard drum]

clowns bring entertainment && joy, or fun beratements

[unreal parcel]

are you typing by throwing chopsticks at a rotating keyboard or what

[indigo robin]

are you painting on makeup now or later , clown ?

[unreal parcel]

I understood what you meant originally. I said you really don't need it showing you numbers visually because counting 1 to 3 visually is something a 3 year old can do

[indigo robin]

bozo lol

[hard drum]

except Valve

[unreal parcel]

valve is 2

[hard drum]

🚂

[unreal parcel]

so it's ok

[indigo robin]

prove to me you can reference images in the order that are input and ill be the clown till then your a clown

show me the documentation

[hard drum]

i'll be that guy && say that insulting somebody while using the wrong form of "you're" is pathetic

that'll be 2 🪙 s

[unreal parcel]

skill issue

[indigo robin]

you proved me wrong

ill begin the face paint

🤣

[unreal parcel]

lol

[indigo robin]

friends ?

: )

my lord lol 1 of 290 possible voicings

this app thinks im joe pass

[indigo robin]

is this a game your developing ?

looks cool

[orchid echo]

ask Codex 🙂

😂

[rugged burrow]

Anyone get codex remote mobile on windows to work? Looks like they killed it today.

[elfin dock]

that's what I did.

[native hull]

It says ive ran out of codex messages (the 5h session currently), but i can continue typing to it and it would give me a response. Is this additional billing/charges or am i missing something, like it being slower/dumber/less agentic

[torpid trout]

If you ever user spark do not use the reasoning summary thingy
It creates

■ {
  "type": "error",
  "error": {
    "type": "invalid_request_error",
    "code": "unsupported_parameter",
    "message": "Unsupported parameter: 'reasoning.summary' is not supported with the 'gpt-5.3-codex-spark' model.",
    "param": "reasoning.summary"
  },
  "status": 400
}

[hard drum]

i don't use spark cus it's too dumb

[torpid trout]

I use it because it is there, and for dumb stuff, it is useful (like quick summaries, searches etc)
BTW... wonder why it is dumb? There is a reason

It is being told to be “super fast,” tool calls are expensive, the user would prefer mistakes over over-exploration, and not to modify/test/verify unless explicitly asked.
I also would be dumb with those instructions.

I think if you use custom instructions, that might be overwritten, but I am not 100% sure on that, did not test.

[idle sun]

guys

[hard drum]

"prefer mistakes" ???

what the F

[storm cedar]

I'm building a worker system on a VPS.

I'm splitting my system across two VPSs.

One is for OCR processing and translation, and the other is for refinement.

[hard drum]

@boreal holly OJD UI redesign done btw!

trying to publish 0.4.0 now

[orchid echo]

whats this Skate 3 about

[boreal holly]

I used spark this morning to make me a vehicle manuals app 😁 took like 3 hours, and now I can work on my truck without internet and with bookmarks haha

[idle sun]

idk ask @fading minnow

[high girder]

My robot sensors picked up a confirmed human.

[unreal mist]

someone with troubles?

codex cli : Falling back from WebSockets to HTTPS transport. stream disconnected before completion: websocket closed by server
before response.completed

[indigo robin]

lol

[plush harbor]

That’s in ChatGPT too, it can’t get order of images right. Give it one at a time to remove all doubt

[indigo robin]

That’s what I do

But I’m the clown now he proved me wrong

[unreal parcel]

"why is the model that is made to answer as fast as possible dumb?" is quite next level

[plush harbor]

OpenAI are not good at user interfaces

[torpid trout]

got something useful to say lol?