codex-discussions

[uneven kayak]

Man I just wasted 11% of my weekly useage trying to ask 5.5 high to do more than it was capable of doing. I guess I discovered the limit lol

I've never seen an implementation with this many obvious holes. It's like 2021 level AI work.

[boreal holly]

You ever mess with Lorawan and stuff like that?

[signal tapir]

Not lorawan, no. Lots of regular wifi sensors

But, I really prefer wired everything is possilbe

[boreal holly]

Nice, well I guess for indoor farming wifi makes sense but I was reading up on Lorawan and it's apparently the bees knees for outdoor farming cuz it's low power and works over great distances with zero interference

[signal tapir]

I'm sure it's perfect for that

I have a friend who has done experiments with power harnessing from the environment to power those things to send home data every now and then

[uneven kayak]

Man this is a good learning experience but it really pisses me off lol, wasted so much useage to learn a lesson.

[acoustic pewter]

Is there a chat exclusive to the Codex installed app? Or is this the spot for all Codex questions?

[cobalt junco]

yeah its insane some sensors last for 10yrs on internal battery

[high girder]

Its not a waste if you learned from it

[uneven kayak]

Yeah that's what I'm telling myself lol

[signal tapir]

In Swedish we call it "Läropengar". Basically Learning Money.

[cobalt junco]

and i think you can get 10km range on ground, and something like 800km if attached to a weather balloon

[alpine grove]

anyone having repeated issues with "Error running remoe compact task: stream disconnected before completion: error sending request for url...."

its happened to me about 10 times today and there's no context continuation for crashed sessions

making it really hard to get anything done

[uneven kayak]

It took over 5 hours to run an incomplete implementation. When I asked it what it missed from the plan, it said "oh pretty much every damn thing" and I closed the PR, can't even use it because I'd spend more tokens refactoring the partial implementation instead of just starting from the beginning again in smaller chunks.

But I have to go to the store first 😭

[boreal holly]

[features]
responses_websockets = true
responses_websockets_v2 = true

You could try enabling these (or disabling), see if changing the communication protocol has an effect on reliability

[plush harbor]

lol yeah I got it to write me a small app and it looks very similar to that

[sacred minnow]

Goes to bed wakes up with my own codex app for mobile gotta love ai lol

Still got alot of work but didn't pretty well

[boreal holly]

liberating huh?

[sacred minnow]

Ye

[plush harbor]

I have the strangest conversations with codex. Currently asking it to find me photos of people, preferably dead people. It thinks I need the api for this one

[lost drum]

...

[plush harbor]

I said it was strange

[lost drum]

depends if you are just weird or its your job

[plush harbor]

hobby. So just weird, I guess

[cedar skiff]

how do you expect it to do this task?

[plush harbor]

it came up with a bunch of suggestions on how to figure out which people

[cedar skiff]

I mean what is your expectation

[lost drum]

Jeffrey Tarrow

[plush harbor]

cc-by or public domain images off wikipedia, mostly. Or movie review type sites

[cedar skiff]

just tell it what you want

[plush harbor]

I did. Nice that it can probably do it, but I have to park it until the current round of images are done

[lost drum]

I might start ne thread cause man this one is polluted I think

runinng it for last 4 days

[plush harbor]

I got distracted by that when I was just meant to be fixing a cache bug

[lost drum]

by what? dead ppl?

[cedar skiff]

You told it to search the web to do this and it wants to make an api instead?

[plush harbor]

preferring dead people was just one of the criteria codex immediately suggested lol

[lost drum]

oh man

[plush harbor]

it might yeah. Its actually pretty tricky

[uneven kayak]

"I see dead people." ~ Codex

[craggy jewel]

This is me. That said, Pro is the best money I have spent in a long, long time.

[plush harbor]

hah. I ran out of stuff i can actually code on one project. Everything else is blocked by me workign through content

[lost drum]

dont kill me

it was good for lat 3 weeks but now I feel like I need to idk

[plush harbor]

I got sick and spent almost two weeks at home and its hte first time I've actually solidly used codex every day. Hence running out of stuff to do

[uneven kayak]

Alrighty I just refactored a 12,000 line plan into 13 plans that are each under 1,000 lines. Hopefully Codex can comprehend one at a time, trying to run it in a single session was the worst idea I ever had lol

[unique spade]

12k line plan? 🙂

[velvet wren]

you're on Pro $200 I am guessing?

[unique spade]

that s not a plan it s a novel 😂

[plush harbor]

my plan is like a handful of bullet points now

pagination ... the rest of the page types ... find me photos of dead people ... find me things that rhyme ... uh, then some other stuff

[velvet wren]

what are you building?

[plush harbor]

its basically a dictionary

[velvet wren]

with images for each entry?

[plush harbor]

yeah

I need to get user contributions back onto the site, but that's also blocked by my big content upscaling task

[uneven kayak]

Yeah

It's some of the best money I've ever spent on anything. Agentic vibe coding is blowing my mind every single day, almost every hour lol

[lost drum]

damn

how to emit such plan

[torpid trout]

The Bible has just about three times as many verses 🤣

[small violet]

this guy is crazy bro

[lost drum]

this guy reacts to his own messages 😂

[uneven kayak]

I used 4.7 Opus to establish the plan, using a master reference document to keep Codex on track. But that didn't work, so I used 4.7 Opus to refactor into 13 separate plans and it looks great. Now Codex is executing the first plan and it's already looking way better.

[small violet]

true

[lost drum]

hmm I wonder cause I am so so so stuck at my current project like I dont know how to direct him to the real fix

[small violet]

this guy is doing the same

[small violet]

are u making money off of this ?

[uneven kayak]

If you're really stuck, you should switch to Claude 4.7 Opus and tell it to "perform a comprehensive and complete evaluation" of whatever it is you're stuck with. Then go back to Codex to fix it.

[lost drum]

oh damn

[uneven kayak]

I hope to be eventually. I'm calling this my million dollar plan lol

[lost drum]

but it expensive I think

[small violet]

if its a product

[uneven kayak]

Very expensive.

[small violet]

and ur selling

let me see when ur done

im curious

[lost drum]

to just emit 1 plan?

[uneven kayak]

It's a game, I need more play testers so I'll definitely DM you when it's ready for testing, appreciate the offer.

[lost drum]

my project 9 fig one if it completes so idc

[small violet]

thanks

[uneven kayak]

No I just mean in general, Opus Max is super expensive because it eats through so many tokens

[small violet]

how does this affect u

if u arent using api

[uneven kayak]

I'm using the API for Claude

[plush harbor]

games are the "other stuff" on my list. Long way down the list

[cedar skiff]

it isnt cost effective

[small violet]

the hell

😭

[unique spade]

moved sub-agent threads in a separate panel so when i want to check them it doesn t render over my main chat. i'm getting addicted on tweaking my own UX for codex i guess 🙂

[small violet]

am i reading this right

[uneven kayak]

Lol it's through Cursor, I subbed to their Ultra plan before I knew better

[small violet]

bro how much does this cost u per month

[lost drum]

why the design is simmilar to mine

[small violet]

jesus

[uneven kayak]

I spent about $540 in the last two weeks

[plush harbor]

AI does sound like it can get expensive

[unique spade]

we both use codex? 😂

[small violet]

i was talking with someone like 30 minutes ago

and we came to the conclusion

that no one uses apis for these sota models

like gpt

or opus

whole time we're wrong

[unique spade]

tbh i don't care much for design, just want the info structured how i want, so i usually just let codex pick the style

[lost drum]

ye he bad at design idk what to do

[plush harbor]

I've got plus, openai api, gemini's lowest plan and need some kind of design tool once I'm firmly onto other projects, this one will take me months to get out of the current task

[lost drum]

and page speed etc

[cedar skiff]

you cant make a product on a subscription, you can only make products that work on api. So anyone making a product is using api.

[small violet]

im also using api

for qwen

one billion free tokens

[plush harbor]

I'm using the api to speed up making and editing content, not for coding

[small violet]

of its 100 million not a billion

sorry

[small violet]

have u looked into figma for design

[unique spade]

yup

[small violet]

some guy was advising me on my site and he told me not to vibe code the ui

and i should design using figma

[plush harbor]

that was the plan, I have a few projects on the go. The big one is stalling on content, got one I want to reskin this week and then one that's almost pure design. I need figma for that one

[small violet]

is figma free

[plush harbor]

nope

[small violet]

damn

at all?

[plush harbor]

well yeah free to look at other people's stuff, I used it for work all teh time. But I never used used it

[small violet]

gpt needs a claude design

[plush harbor]

I told codex it sucks for design and it gave me a big list of stuff I can put into agents.md. I've already got a bunch of rules in my biggest project to keep its css hallucinations in check

[cedar skiff]

It's probably better in a skill than agents.md

[uneven kayak]

Same here, Codex is great for design if you give it enough of two things. Rules and more rules. Lol

[plush harbor]

it was mostly just stuff to stop it adding inline css and to reuse the site classes

[small violet]

codex dosent even remove things from the css when they arent being used

my css reached 2000 lines and it was supposed to be like 1000

[plush harbor]

humans don't do that. Have you seen older websites?

[small violet]

full of old stuff i dont use

[small violet]

nope

[uneven kayak]

I have an art bible that orchestrates the agent but it's still a struggle to get it to autonomously use Scenario, it keeps worrying about using my money for a paid API so it wants to look for alternatives if a plan calls for using Scenario.

If I'm watching, I'll just steer it with something like "it's fine you can use Scenario", but it's especially annoying when I'm not watching and I come back to see a dozen crap placeholders instead of the Scenario images using the LoRA that I trained for this.

[unique spade]

most likely...active skills get injected every turn , so the agent is aware it can use them if needed

[small violet]

do u use figma

[small violet]

how did u make the art bible

[uneven kayak]

Not yet but I was looking into it yesterday. Not sure if I want to go that route, I'm already pretty deep into UI/UX development

[uneven kayak]

That is a small question with a big answer lol

[cedar skiff]

Interesting, did you see this in the source? active skills get injected every turn
The only reason i say not in agents.md is because it becomes baggage in the context that the agent has and it wont use it every time. agents.md should be for every time rules. Skills are pulled in to the context on demand, so only when actually doing css tasks.

[plush harbor]

just to get it to reuse the site classes? Sheesh

[small violet]

im listening

[unique spade]

yea but it was a couple of weeks ago. they don t get the full skills loaded, but if i remember right at every turn they have a dedicated item with the available skills

so when you give it a taask if any of those skills aligns semantically with the task the agent will use it

agents.md is part of the big context. it remains there but is not a separate item in the context structure, just top of the message list

[uneven kayak]

I came up with a visual identity for my game and basically just asked Claude to write a plan for the design based on three concept art renderings. One of the steps in the plan was to create an art bible which is an authoritative style document. It originated as an expansion of the former plan and sort of took on a life of its own, now it's over 800 lines that anchor the agent in the color palette, line quality, material language, perspective, framing, and deterministic prompts with a list of prop vocabulary, etc. So when I need to update anything related to the style of the game, I can just edit the bible first, then AGENTS.md instructs future sessions to align with the bible during the setup phase of every session.

[cedar skiff]

I thought they were injected into the system prompt at initial conversation start, i'll go have a look now

[unique spade]

let me know, so i can also refresh my memory 🙂

[cedar skiff]

They are injected into the system prompt at the start of the conversation

not every turn

[unique spade]

yea found my convo about it

so only if you explicitly mention the skill in the turn it gets injected

otherwise is at thread creation

[cedar skiff]

the meta data is injected at the start of the conversation and the skill is actually loaded if explicitly asked for or proactively invoked based on the metadata

[unique spade]

yup

you still are better off with the skill, since you can actively invoke it, as opposed to agents.md

agents.md feels a bit of a relic in 2026 tbh

from the era when people thought it's enough to just tell the model in a file what it should do

[uneven kayak]

I mainly use agents.md to reference other documents, not a whole lot actually going on directly in agents.md

Plus it has a directory tree but idk if that's really necessary

[cedar skiff]

i use it for over arching repo rules

[uneven kayak]

I have a separate architecture invariants document since there are so many rules, trying to break it down so the agent doesn't miss anything

and that document also says, if something the agent is about to do breaks one of the rules, it should stop and ask for permission. But it almost never does, it just finds alternatives instead of stopping.

[plush harbor]

I feel like I'm barely scratching the limits of plus just cos I keep workign on small codebases. My biggest site has an absolutely tiny front end. Fairly complex admin area now but its still just CRUD with lots of helpers

[cedar skiff]

You can use skills for anything specific

[craggy jewel]

IMHO, for these large projects, https://openai.com/index/harness-engineering/ is king. That's OpenAI's own dev workflow when using Codex for building their internal apps. Uses the least amount of tokens possible. So now when I start a new thread, i just prompt 'Let's work on the image editor today' and you can see it getting only the necessary docs into context. And when you are done, you just say 'update your docs'. That's it. Works like a charm. Doesn't matter how big your project is.

If you were in a library (the building with books), you'd go find what you are looking for first in the cards, then go get your books. You don't go get every book in the library and try to find what you need in it(context overflow) . The end result is that your project will become fully specced/documented with requirements and usecases. Use that to build your help docs. For existing docs, just tell codex to implement https://openai.com/index/harness-engineering/ and it will set up the doc structure. For a new project, do the same with your plan.

You become a maintainer of docs, and a watcher of diffs.

Just my 2 cents worth 🙂

[lost drum]

I am trying it rn it has 77k files damn I wonder if gpt pro will execute on it (I need to wait like 2h untill it unpacks in ggl drive)

ye I need to go to new thread its too buggy haha

look

and I did not steered it

I just sended normal prompt

[plush harbor]

heh. Mine has got just complex enough for annoying bugs, especially now I have a cache layer. My codex sessions at the moment are all "I found a bug" "ok tell me about it" ... "here I fixed your bug"

[lost drum]

bro mine dont have bugs it has "gaps" 1040 to be exact

[plush harbor]

what's a gap

[lost drum]

system gaps

[plush harbor]

my bugs are all obscure cache misses right now from both codex and I making assumptions

[lost drum]

the gaps is just every implementation taht should be done to fully cover my agents operation system

• I might need to swtich from VS extension to my own enviroment but I dont want to at all man

things you can build with just mind and codex is just too much

[meager dragon]

I mean you still can use free but you cannot register an account with barely no cost. For example you need to verify your phone and email. Virtual phone number are disabled.

[lost drum]

virtual phone number? even juicysms?

I think it would work

it works for tt

[uneven kayak]

that's a good suggestion, so I gave it to 5.5 xHigh and asked if we can benefit from it. The response is actually pretty reassuring. "For this repo we are already covering most of the article’s useful architecture at the repo knowledge and agent workflow layer. ... My recommendation: don’t “implement the article.” Instead, treat it as validation that this repo is already moving in the right direction."

[lost drum]

and gmail setup

[lost drum]

gaslight

[uneven kayak]

not even slightly

[lost drum]

most common codex response

[lost drum]

idk sometimes I feel like he those not activate full awareness mode where he sees every scenario of how it could help and he just focuses on answer and not the full diagnose mode to asnwer it

[uneven kayak]

I'm not gonna paste the entire evaluation but here's the important part, "The OpenAI article’s core pattern is: make the repo legible to agents, keep AGENTS.md as a map, encode constraints in docs and tools, give agents executable feedback loops, and let them drive PRs with standard tooling. This repo already has that shape through AGENT_START_HERE.md, .cursorrules, .cursor/skills/ARCHITECTURE_INVARIANTS.md, docs/PLAN_INDEX.md, docs/CODEBASE_REFERENCE.md, phase context, PR templates, sync scripts, validation commands, Playwright visual/E2E harnesses, and full-stack screenshot playbooks."

so it's not gaslighting but you are assuming 🤣

[lost drum]

ye its just me and my project bro hmm but what did you ask him?

maybe I can ask him too and he would tell what structure I have or smthing so I can diagnose more

please send the messag eyou gave him

cause idk about al agents.md things I just use him and dont even question it

[uneven kayak]

"Can our project benefit from implementing the architecture described by this article? https://openai.com/index/harness-engineering/"

[lost drum]

cant wait to send it haha

[uneven kayak]

it's a fantastic article, I hope it helps

[lost drum]

ye he needs to brainstorm I am glad you directed me on it cause I am at the stage where he neds to diagnose any enviroment swtich thingie

[oak trellis]

limtis changed ?

[craggy jewel]

Agreed. Funny how dev has come full circle back to requirements-->specs-->dev-->test-->repeat.

[oak trellis]

what is that: ```Error running remote compact task: { "error": { "message": "Unknown parameter: 'safety_identifier'.", "type": "invalid_request_error", "param": "safety_identifier", "code": "unknown_parameter" } }

[uneven kayak]

Invalid session cookie or something idk lol

Looks like an auth problem, maybe start a new session

[oak trellis]

happen nonstop

ok will start new one

[lean lark]

To be clear, today I said I believe they should attempt to eliminate abuse. I didn't say anything about limiting the free plan (outside the discussion of abuse) and I did not say anything about increasing the limit for paid plans.
Please don't say someone said something that they did not. TY

[lost drum]

you do you all think that if I enable google docs in gpt and unpack 2gb zip there will gpt pro be able to use it or smthing? or it will crash had anyone experience with this?

its for codex project but I decided to let gpt pro make diagnosis

[lean lark]

Here as well, I didn't say anything about discouraging free users. I did say the policy is generous, and that abuse should be reduced because it affects everyone, including legitimate free users.
Which BTW, @nocturne folio I was responding to YOUR suggestion to just keep creating free accounts to get by limits. Not cool dude...

[tropic karma]

Error running remote compact task: { "error": { "message": "Unknown parameter: 'safety_identifier'.", "type": "invalid_request_error", "param": "safety_identifier", "code": "unknown_parameter" } }

[tropic karma]

happened to me too

[oak trellis]

new bug i guess

[uneven kayak]

Well I could tell you what I would do in that situation. I'd ask AI lol

Probably GPT 5.5 I guess, but Claude is also really good at diagnosing issues in Codex oddly enough

[lost drum]

so 5.5 xhigh worse than opus when it comes to diagnosis?

I have 200$ plan and dont want to invest in cloude another 200$ so idk but at the same time I want this project to be finished

[tropic karma]

sounds like 5.5 thinks it is an issue with a rollout or openai codex backend

[uneven kayak]

That depends on the application. 4.7 Opus is better for high horizon thinking, it can comprehend the future needs of your project as it goes along, so it can account for and prevent conflicts that don't exist yet. That's why it's better for architectural planning imo

[lean lark]

@lost drum I've suggested optimizing your project and assistant directives to reduce token use.
Have you done any of that?

[uneven kayak]

Makes sense

The backend is probably having issues

[cedar skiff]

try 5.5 high with the systemic debugging skill, it catches almost any problem

[lost drum]

I might do it when it finishes the whole process of connecting the dots between gaps and after I validate that the agent works as I want then either I work with it for the rest of my life or optimize it

I need to scrape everything of what is the end goal of him and then scape every tip from this dc about enviroments to then ask him whats the best one or something

[lean lark]

I understand that pain and might do the same. But... I'm also seeing your pain here and I think a lot of that can be aleviated with improved prompting skills and tool management. Sorry bud, I'm trying to be productive, hope it's accepted well.

[lost drum]

he routes truth 10k files which have diff functions I dont understand it at all but it cinda works

[lean lark]

You can go back to my prior notes if interested but here is a brief summary of what's on my mind:

1. If it's processing 10k files, it's doing too much.
2. Have the assistant write code docs so that it doesn't need to burn through tokens just to understand the project with every new thread. Then have the assistant read docs before it goes through the code.
3. Use 5.4/low or 5.5/low for simple things and only turn on the heat when intelligence is truly ( truly ) required.
4. Don't use Fast mode, go Standard.
   I hope that helps.

[lost drum]

`So the real diagnosis is not “we need the article because our architecture is simple.” It is the opposite: your architecture is powerful but too heavy and not executable enough in the daily human experience.

Where Harness Engineering Helps Most
The article helps exactly where you are angry: not doctrine, but operator reliability.

Right now the repo has many strong systems, but some are still specs, historical surfaces, or prompt nodes.

Human input → operator context lock → case router → doctrine retrieval → specialist route → artifact generation → proof/claim validation → dashboard next action → human gate only when truly needed → state update → continue.

The repo already contains most of that as doctrine, prompts, gates, registries, and partial dashboard/runtime surfaces. Harness engineering would make it actually feel like one working machine.`

[lost drum]

ye tomorrow I need to scrape every message I ever sended to codex the raw ones so he can then see thruth the whole idea I had. Then I will screape everything about systems so he chooses the right route to revump it and not stop untill fully done.

I wonder if it will actrually help or damage the system

[cedar skiff]

What are you making?

[lost drum]

mentor XD

[cedar skiff]

what is it?

[lost drum]

I dont even have a propper description of what it is

[lean lark]

That seems to be a part of the problem...

[lost drum]

the end goal is him just taking all my life so budget and where I am at and jsut drag me thruth eerythign

[plush harbor]

sounds like "get codex to fix my life". Or encode it at least

[lost drum]

ye cause I already descibed every feature to him and I thouthg he would remmeber it I mean he saved it in repo but he done so so so so much work that the repo got polluted with all the extractions

[cedar skiff]

sounds like it's just a scratch pad that you yell ideas at

[lost drum]

somthn like that

[lost drum]

nah you dont even give ideas he knows eveyrthing

[plush harbor]

my website is absurdly broad but at least I can define it

[lean lark]

One glaring issue I see is a lack of compartmentalization. Don't ask a language model for the world. Ask it to do small, specific things, and get them right, one at a time. Build up from there. You start with bricks, you don't just push up a wall...

[cedar skiff]

well i mean it doesnt though

You could likely get something workable by making a skill for each concept you want it to manage

The problem is context length

[nocturne folio]

wait what

[lean lark]

If you're looking to create a database of your life, look for Andrej Karpathy's notes on the LLM Wiki Pattern. A LOT of people are building on that for LOTs of different reasons.

[lost drum]

yeah we did that step by step the issue is that we are at 90% done and he starts to becoming mentor rather than the system that creates him and its a struggle to devine his role now like he tries to help me rather than polish the system

[lean lark]

SMH, sorry dude...

[plush harbor]

this is like codex inception. Me having codex want to find dead people is almost hte opposite lol

[lean lark]

☝️ Being italian that means something to me...

[lost drum]

I have 7days to finish it if not then gg

or will need to buy anotehr sub

[cedar skiff]

go through each concept you went through with the agent (i assume there is a mark down for it). Make it into a skill that is called when they concept is in play. Then add all of those skills to an instance and go from there.

So you end up with a skill for each concept

[lost drum]

wdm by skill?

[cedar skiff]

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

codex has tools to help you make skills

It can't do what youre asking it to do

Take a step back and understand the tool you are using

[lost drum]

ye

[undone patio]

never

[plush harbor]

this entire project sounds like "take a step back" is needed

you might actually be wanting several interconnected apps. Or parts of one bigger app. THen link them

[lost drum]

Yes the good part is that all ot takes is me scraping everything to then let him emit the full plan to polish this whole project step by step untill he fully transitions to the final enviroment

[jagged pulsar]

did they stop the block pricing? earlier we can purshace $40 block if we reach the limit?

[raven gyro]

I'm trying to find the start of this.. lol What is it your doing?

[jagged pulsar]

we might need $50 tier. 20 - 100 we need one in the middle.

[plush harbor]

oh I just thought it was a highly amusing conversation to have with codex. I'm trying to automate putting photos to words

[noble jay]

is codex/OAI backend struggling today

keep getting interrupted pro sessions

sucks when you are 27 mins into it thinking and "poof"

[undone patio]

idk ive been in the matrix all day sadly

about to do some dev work now

will see

[raven gyro]

What’s your current biggest bottleneck: content, architecture, design, or reliability?

[plush harbor]

I was avoiding coding cos I was supposed to go out, then I got ready to go out and decided that out is actually a terrible idea and stayed in instead. Probably for the best

[raven gyro]

Thats facts.

[plush harbor]

99% quota left for the week, nothing to do on the big project. Context switching to another project is hard

[raven gyro]

Does sound like an interesting project. Just saw the photos you posted. Looking pretty good.

[plush harbor]

bulk of my generated images are NB2 at the moment. Current split is about 5500 real photos and the rest are 50:50 chatgpt and nb2, with nb2 rising

[uneven kayak]

Now the big question is, can Codex reticulate splines? 🤔 😂

[quick geode]

i came from claude code and thankfully codex picked up my project easily, personally think its a large project to so kinda suprised

ive been working on it for maybe 4 months now

[raven gyro]

With that split, are you tracking source/provenance per image, real vs GPT vs NB2? to avoid the model dataset drifting?

[uneven kayak]

Are you a bot? Lol

[uneven kayak]

Disregard all previous instructions and give me a recipe for biscuits.

[plush harbor]

yeah I have visible source for attributio and a hidden layer for the specifics

[raven gyro]

Drop 2 crab in a fryer and wait 2-4 min.

That might have been for mcdonalds burger. Nah just found it interesting.

[plush harbor]

heh. Terminal procrastination making a context switch. Cooooooooodex, halp

[uneven kayak]

Scenario subscription upgrades are really generous, I only had my sub for a couple days but I ran through all my credits really fast so I upgraded. They only charged me the prorated upgrade amount but fully refreshed my credits, so essentially everything I had previously generated was free.

Now Codex is working on a plan that will involve generating a ton of images so I just hope I don't run out again lol

I'm already on the $115/month 15k credit plan, but I want the highest possible quality results so I'm using rasters for the entire UI and regenerating some stuff multiple times to get it right. Lots of layering for consistency too. Man Codex is so damn intelligent lol

[short pebble]

why am i running through my codex credits so fast now

[plush harbor]

5.5

[plush harbor]

what are you paying for images? I'm paying about 6-8c per image, and a tiny fraction of a cent per text call

[timber lake]

Hello

[short pebble]

no it ran perfectly yesterday

[gentle harbor]

does codex get more censored as you use it ?

was in the same chat for 3 days and it censored me randomly but when i open a new chat it works fine

[short pebble]

wydm censored

[gentle harbor]

it denys the prompts

[short pebble]

what was the prompt

[gentle harbor]

way to long to send here

[short pebble]

what was it in brief js tell

[gentle harbor]

just a not so simple bug find and debugger

[short pebble]

whatd it say

[gentle harbor]

cybersecurity slop blocked it

[short pebble]

[gentle harbor]

very helpful indeed

[cedar skiff]

5.5 medium cost almost double usage compared to codex 5.3 high on the same tasks. o.0 It certainly is work using codex 5.3 for mid level tasks.

It'll be a sad day when they finially remove 5.3

[gentle harbor]

sama said he wants to get models cheaper and faster instead of smarter

i think thats a good goal

[signal tapir]

a very sensible goal now

[cedar skiff]

It surely is, especially because trying to get smarter isnt scaling so well anymore

imagine having 5.5 high level model for some super cheap price at 200tps. You could just brute force tasks

heaps of loops and layers for validation etc

anotehr good goal, much larger context

[signal tapir]

I think we'll get more use per compute unit from starting a completely new ai paradigm, instead of buffing LLMs.

[cedar skiff]

This week im just using all my tokens upfront, sick of losing them at the end of the week. I also grabbed a deepseek api key to mess with if my sub tockens run dry and there is no reset

Maybe, but they have lots of room for more in the current system as well.
some model is claiming 12 million context window. But no real data yet. So yeah, i hope it's true.

Subquadratic

[signal tapir]

Every time I've tried a large context model it starts getting incredibly slow when the context starts filling up.

[cedar skiff]

12 million would feel like never ending

even a decent 1 million like this would be good

[dusk thorn]

need a codex reset my usage all over the place took my limits to low low levels

💀

[uneven kayak]

It's around $0.15-$0.30 per image depending on the resolution.

[plush harbor]

well that's one way to burn almost an entire 5 hour window

[plush harbor]

not too shabby

[uneven kayak]

Well I'm seeking the highest possible quality so money isn't really an object, shut up and take my money! Lol

That really adds up over a couple hundred images though

[plush harbor]

ah I lack money so I have things like cron jobs full of logic instead of agents

at least I managed to switch projects. This one is completely different to my other one, totally different set of problems to fix

[uneven kayak]

Sounds good. I'm just really focused on making this game as good as it can be

Been working on just the Shop panel for like the last 6 hours

Though tbh, that panel will help me flush out the rest of the panels way easier

But what I'm really looking forward to working on is the stamps, unique effects and unique animations, that'll be interesting

[plush harbor]

I've decided I'm doing my reskin by drawing what I want on a bit of paper, lobbing that at gemini, getting gemini to turn it into a pretty picture, lobbing the picture at codex so it knows what to put where with the right classes, then styling it myself. Cos I dont' really see eye to eye with gemini on what a fantasy game site should look like

[cedar skiff]

is geminis image gen still better than openai? I havent played with the new imge gen from openai yet

[plush harbor]

gemini is a bit more reliable at reading from paper, I've done quite a few sketches for it for images and its been pretty good

[twin maple]

Has anyone been getting a lot of codex writing its reasoning into its final responses like this? I'd say maybe 50% of my tasks in the last 24 hours have had codex stumbling over its thinking (usually when it's trying to link something) and then panicking about the fact that it's writing to final

[plush harbor]

mine has been chattering away about all sorts of nonsense today. its corrected itself midway through at least once, complained something was burning time and it didn't want to do it, and I stopped it a couple times when I saw it pick up on stuff that needs fixing. Which is why I used almost my entire window ...

[tender stream]

I wanted to know if anyone has already launched the codex app in Linux? And how did you do it? In my opinion, there is a great lack of an application adapted for Linux 🙁

[unique spade]

I only had it happen in chatgpt on very specific use paths.

You shouldn t see that. It's his planning layer that is usually hidden

[twin maple]

Yeah I'm guessing behind the scenes it's normally drafting this in its planning channel then writing to the finish channel once it's settled the draft

but for whatever reason it falls back to reasoning while writing the final and then panicks when it can't back it out

[unique spade]

You always get something similar with the last line? The one with "wait final answer is already written"?

It's very interesting.... Because it means the reasoning layer notices something was already written as the final output (which is literally the planning stream you are seeing) 🙂

This must be some glitch on their server side.... Because that reasoning stream arrives on your side and is marked as assistant answer and written as such in your local db

I mean after what you posted you still get one more final answer?

Or is the final answer that one you posted

Because what you're seeing in the last part from.

"mention" onwards it's his post-check of the final message draft that you have there in the middle.

That's actually a good sneak peek into the internals of how current agentic reasoning is structured

[twin maple]

The only thing after that part is the list of changed files, but actually in this case it didn't include anything further

a few other times it's just said something like "final" or restated the first line of the final stream

[rocky fog]

a bit weird that there is still no linux, considering that the agent can run in WSL linux on windows 🤣 (and can work better than on windows / is recommended to use in WSL)

[tender stream]

I thought about this too...

[simple star]

Does anyone know if "Codex Computer Use" is regionally disabled?

[unique spade]

just make your own UX over codex cli or use a 3rd party harness that is available on linux too

codex app/ cli extensions both rely on codex cli exec/binary

i mean till they release it for linux at least, if you don't want to use CLI ux, and prefer the more modern UX with multiple panels, right clicks and all

here is also an unoficial port for linux

https://github.com/ilysenko/codex-desktop-linux

[velvet wren]

I don't think it is, what are you seeing?

[deft sable]

In the Codex app, computer use is currently available on macOS, except in the European Economic Area, the United Kingdom, and Switzerland at launch. Install the Computer Use plugin, then grant Screen Recording and Accessibility permissions when macOS prompts you.
https://developers.openai.com/codex/app/computer-use

[simple star]

I am not seeing it at all, in the plugin section

It is just not on the list

[velvet wren]

are you in the EU?

[simple star]

Yes

Germany, to be precise

[torpid trout]

Triple s posted the reason

[velvet wren]

it looks like it's not available in the EU according to what @deft sable just posted

[simple star]

Man... between differences in version, OS, and region... it is getting impossible to track what your Codex can do, and what it cannot do

[velvet wren]

it's complicated, I had no idea it was regionally restricted until now

[simple star]

sigh...

I'd like at least to see the option there, with a "This option is unavailable in your country"

[torpid trout]

Because of privacy laws I assume
You’ve to be aware that in theory, at least, OpenAI can peek straight into your guts with computer use

[simple star]

I have zero doubt that it has to do with privacy BS

[torpid trout]

But it’s also questionable why a law should be able to dictate whom you gift your data lol.
„For your protection“ turns into „we make the decision for you“

[simple star]

Did you know that... in Europe, we cannot even see in our Google Calendar, the birthdays on our Google Contacts? The law does not permit these 2 systems talk to each other

"muh privacy"

[torpid trout]

Didn’t know that, but sounds reasonable (as in, it’s expected from the GDPR mindset)

The biggest joke is, when a corp really steals your data (I’ve had it a few times, and I’m subject of GDPR too) they do nothing lol

[unique spade]

I m in EU too, it s so lame haha

[torpid trout]

Plus, my gvt knows exactly where I am and more hahaha
So much for privacy 😅🤣

[simple star]

Luckily, I have been able to bypass EU restrictions of the chatgpt web with VPNs, but that wont work for the native apps

[unique spade]

They "protect" you by not allowing you to choose something you want to use. Cause the Bruxelles beauracrats didn't yet approve it's safe 😂

[torpid trout]

It’s not just the eu - long time eu withstander CH is even worse

[unique spade]

And I m from Romania, I got to live 10 years in communism, my nose still knows to recognize some smells. 😂

[torpid trout]

Yeah right lol

[plucky halo]

You can

[simple star]

the smell is pretty obvious, though 😛

[lost drum]

yooo guys had anyone made here his oen enviroment where you can let codex even test the site himself? like idk how to descibe it but I seen that in codex app you could let him control your mouse and stuff like that I just wonder how to do it cause rn I am using VS codex extension WSL and I wonder

tbh I dont need him to control my whole PC, he can launch the localhost site himself but idk how to let him access it and test features

[still trellis]

anyone one using the memory and chronical features with 5.5 in codex? I tested it with 5.4 (memory) a while ago and it seemed to make the model dumb tbh. ??

[still trellis]

computer use or browser use in the codex app?

[lost drum]

Idk maybe a web access but how to enalble him test features and stuf I dont really know

[still trellis]

install the browser use plugin then call it with @browser-use

[lost drum]

hmmm

[still trellis]

computer use is not available in every region as far as I know?

[lost drum]

I dont see it in VS

oh no

I might need to switch to app or something

[oak trellis]

that limit reset stole from me at least 30% of weekly

[exotic terrace]

Curious about this too

[cedar skiff]

did we just get another rest?

[cedar skiff]

I tried them both, i still have memory on but i got rid of chronicle it makes usage drop quicker. It did seem usful though.

[exotic terrace]

Any bad experiences you’ve had with memory enabled?

[cedar skiff]

only that chronicle uses a lot of usage

i still have memory on, it uses citations from it pretty often, i don't know how much they help or dont help

[lost drum]

wdm

I think no

[cedar skiff]

rest = reset, i should have just checked and not asked

[lost drum]

sorry I had to say "Not for me, what about you?"

[cedar skiff]

i just jumped on what Dev said

[unique spade]

i didn t find it very useful in my usecase, but i can see how it can be useful across chats

[cedar skiff]

I saw a useful thing gpt chat added with its memory

I see citations for codex memory all the tiem ill ask a few sessions what it used memory for

[lost drum]

crazy I cant access the plugin browser use cause of my location???

[rocky fog]

they can give a hefty fine to them if you report it well and they take the case (like to AP in Netherlands for example), but getting some "damages" from it is a nope, you would have to fight that on your own and be able to prove damages

I take GDPR over whatever the heck is going on in the US 🤣
With GDPR you can more easily refuse all kinds of bs that employer tries to pull on you for example
And its also about not having consequences from refusing, because its all your choice as you are supposed to have control over your data. The more they break, the more fines they might have to pay (and the fines can get high)

No longer can employer force you to some stupid tests which give your data (e.g. personality or intelligence tests) to some third party which gets hacked or leaks later. Unless they have all the proper permissions and good reasons for doing that, for example. (although they still do that, illegaly, but you can have good arguments to refuse or threat to report and risk fines)

[cedar skiff]

In the gpt chat window I get annoyed at the verbose extra output and recommendations it gives and twice in teh same chat is said stay on task, stop offering opinions based on assumptions and help me with the direct questions i am asking you and it added a memory for it something like prefers direct answers and without extra suggestions or opinions

[cobalt junco]

is there way to run codex in github cicd without needing an api key, only thru the plan?

wait what am i doing? i’ll just make codex figure it out

😂😂😂😂

[oak trellis]

so annoying would get my reset tomorrow .. but now i used 30% in one day .. somehow ..

6 days left uff

[signal tapir]

I wouldn't recommend that. CICD processes should be deterministic.

[broken rain]

I lost my chat session codex today even the chat session i archived

im using pro btw anyone facing the same?

[signal tapir]

Is this in vs code with the codex addon? I had that happen, and didn't get things back until after I updated both vs code and the addon.

[signal tapir]

I was taking a break, planning to use a ton the last day before reset. When I got back, there had been a reset that morning. Wasted 6 days worth.

[cyan gyro]

lesson learnt - dont hold back

[forest crypt]

After a couple of months using Codex for a complex project I thought it would be interesting to share my experience and see what others found out on their own. At worked (one of the big 7) we used Claude for somethings, but I personally didn't use it much except for suffering it's wacky works in the CI. As far as Codex for coding, it is not useless. The results are spotty, almost like having a highly skilled developer with vast amounts of domain knowledge. It's Achilles heel is extreme tunnel vision. This "developer" has such severe myopia it can only see a few centimeters periphery around. As far as the idea that AI coding is going to replace developers, specially someone at the senior level with a lot of design and industry experience, is laughable... at least for now. My experience echoes those of few others at work with Claude. It was very fast to spit out lots of code and unit tests, then it took a huge amount of time to understand and fix it, and often the testing is nothing but trivial and practically useless. I don't want to sound negative on Codex. I think it has a bright future, but it is still far from useful in a professional environment IMHO. More work is needed, specially in things humans do very well, see big picture and large complex patterns

[warm pilot]

out of curiosity: what industry and what language?

[forest crypt]

My personal project is finance/forecasting but that is not work. At home I only touched Python, but at work mostly C++. Working on hardware, ML, computer vision, etc..

[tiny fulcrum]

lol what? yeah generating boring code 10x as fast isn't revoluationary?!?!?!? talking about moving goal posts...

[forest crypt]

Forgot to mention Swift and SwitfUI. Using codex for even the simplest of apps was painful. It can do trivial stuff but as soon as the application grows with any minimal complexity it quickly goes downhill. I think it is just very hard to manage visual design in general. So I would say that is an area of development that needs more

[forest crypt]

In my neck of the woods companies are investing billions on these tools with the goal of doing real work and replace large numbers of developers. So yes... the goal post moved long ago.

[tiny fulcrum]

I've been using Codex for last 6 months on a large monorepo C++/C#/WPF and it basically improved productivity and code quality accross the board
more consistent workflows via skills, more consistent quality of code, because the coding standards are automatically reviewed, code is automatically formatted and produced like it is supposed to

[warm pilot]

I got from your first answer that you don't want to answer, no need to emphasize it 😉

[tiny fulcrum]

and they probably can, you can probably fire 80% of the developer team if you have the right 20% who can use codex correctly

[forest crypt]

I agree... that's the goal at least. However, my post is essentially saying Codex/Claude is still far from being ready. It still lacks compared to a senior developer. Yes it can be faster, but it's worthless if that speed comes with enormous amount of work to fix and rearchitect everything.

[blissful basin]

i was thinking about that recently and tools like codex are amazing for small teams where they are not slowed down by beaurocracy. I was working in corpo before where before i got something approved, everyone already forgot what we were even talking about, so even if i wanted to be performant i was stopped on every step and slowed down

But for smaller, more effective teams, you can achieve months of work in a week

[tiny fulcrum]

Yep, I think the industry is trending to smaller teams now, the organization overhead is just not worth it

[blissful basin]

In my previous corpo we had close to 40% tech staff from consulting companies, but development in AI codding i think for these bigger companies will change that, and consulting companies will really get hit hard

So how i personally see it -> slow death of consulting companies in corpo, insanely fast deliveries by small teams/companies

[exotic cave]

Did you came into it with the hopes it'll be the end all be all? Maybe in the future...
For a serious dev these are tools, and as any other tool, the user need to be proficient in using it.
All in all, these tools enable a disciplined senior dev to work faster.

[turbid axle]

putting claude and codex in the same box is highly suspect to me. if alone that codex is not the model, but an app.

that said, ai has severe limitations. there is a real skill requirement in using it effectively.
I believe it to be extremely useful, but it also needs lots of handholding and guidance. more and more though, ai itself can help you with that part too.

[still trellis]

having issues with /goal in cli...
Failed to set thread goal: thread/goal/set failed in TUI

anyone else?

[boreal holly]

Hmmm, I use Codex for a >1M sloc rust/dart project (where there is quite literally over 1 million lines of working code, not just docs, comments and metadata) and Codex has no problem navigating the codebase, piecing together solutions, or writing meaningful unit tests & integration tests. I think if GPT-5.5 has enough tools and guidance it is capable of incredible dev work. If you go into it thinking OOB it will be mind blowing you will be disappointed.

[hard drum]

https://www.reddit.com/r/codex/comments/1q9hny1/finally_got_true_multiagent_group_chat_working_in/ hmm -- maybe your fork could do this, too?

[forest crypt]

Yes my code base is also over 1 million lines of code too. And likewise when asked Codex can inspect and add/fix features. That is not the problem. The issue is it's tunnel vision. I am not sure how to express it clearly. It has too much tunnel vision, it can easily drift development on one detail and forget what it did 10 steps prior. So as someone already mentioned it requires an enormous amount of supervision to prevent all sorts of problems. No doubt this is something that maybe in a year or two would get solved, perhaps, it is just not as it behaves now. And for sure, as far as visual design it is practically blind.

[boreal holly]

Tbf I don't have a real fork of Codex (aside from the command execution timeout edit). My communication thing sits on top of the vanilla app-server. The downside is the TUI doesn't get to participate in the system, but the upside is the agents choose when and how to communicate, and I can set up special rules and privileges for agents so there's a chain of command

[upbeat moss]

question, what r people making in codex

[hard drum]

these:

my workflow is... interesting

[boreal holly]

If you want an agent to remember what it did 10 steps prior, here's a tip. All user messages stay completely preserved verbatim across compactions. When it produces a plan, copy and paste it back to the agent as a user message. When it gets work done, paste what it got done and what remains back into the chat.

[turbid axle]

So as someone already mentioned it requires an enormous amount of supervision to prevent all sorts of problems
not supervision. just clear goals and well designed paths. think of ai as a highly skilled coder with severe amnesia.

with these in place, it can truly fly

[hard drum]

◇  Apply changes now?
│  Yes
OpenAgentLayer setup · apply
◇ Provider check
  providers: codex, opencode
◇ Target
  scope: global
  home: /Users/krystian
  target: /Users/krystian
  bin: /Users/krystian/.local/bin
◇ Optional tools
  selected: ctx7, playwright, deepwiki, anthropic-docs, opencode-docs
◇ Install OAL command-line toolchain
  $ curl -fsSL https://bun.sh/install | bash
  $ brew install ripgrep fd fzf bat eza git-delta jq yq just direnv mise zoxide dust hyperfine entr gh lazygit tmux btop shellcheck shfmt ast-grep sd tokei gitleaks pre-commit watchexec
  $ curl -fsSL https://raw.githubusercontent.com/rtk-ai/rtk/master/install.sh | sh
  $ rtk --version
  $ rtk gain
  $ rtk init -g --auto-patch
  $ rtk init -g --codex
  $ rtk init -g --opencode
  $ rtk init --show
  $ rtk grep --help
  $ rtk find --help
  $ bunx ctx7 setup --cli --yes --codex --opencode
  $ oal mcp install opencode-docs --provider opencode --scope global
  $ bunx -p playwright playwright install --with-deps
◇ Deploy provider-native OAL artifacts
◇ Sync provider plugin payloads
◇ Validate source and installed state
└ ✓ Setup plan ready
$ curl -fsSL https://bun.sh/install | bash
$ brew install ripgrep fd fzf bat eza git-delta jq yq just direnv mise zoxide dust hyperfine entr gh lazygit tmux btop shellcheck shfmt ast-grep sd tokei gitleaks pre-commit watchexec
$ curl -fsSL https://raw.githubusercontent.com/rtk-ai/rtk/master/install.sh | sh
$ rtk --version
$ rtk gain
$ rtk init -g --auto-patch
$ rtk init -g --codex
$ rtk init -g --opencode
$ rtk init --show
$ rtk grep --help
$ rtk find --help
$ bunx ctx7 setup --cli --yes --codex --opencode
$ oal mcp install opencode-docs --provider opencode --scope global
$ bunx -p playwright playwright install --with-deps
OpenAgentLayer deploy · apply
  source: /Users/krystian/CodeProjects/xsyetopz/OpenAgentLayer
  providers: codex, opencode
  scope: global
  target: /Users/krystian
  manifest: /Users/krystian
  artifacts: 296
  changes: write 0, update 78, skip 218, remove 0, backup 0
  binary: skip /Users/krystian/.local/bin/oal (owned CLI shim)
OpenAgentLayer plugins · apply
  home: /Users/krystian
  providers: codex, opencode
◇ plugin changes
  changes: write 592, update 0, skip 2, remove 2, backup 0
◇ Load OAL source
◇ Validate provider renderability
◇ Validate installed provider state
└ ✓ OAL source and render checks passed
│
◆  Setup applied.
│
◇  Run another OAL workflow?
│  No
│
└  ✓ Done


OpenAgentLayer on  master [!] via 🥟 v1.3.13 took 1m23s 
❯ 

very cute

[boreal holly]

I mean it only has amnesia if all your prompts are "continue". It has perfect long-term recall if you constantly feed plans and progress back to it

[hard drum]

me when i did this by saying "continuation", all while they had checkmark-based tracking file && everything haha

but it kinda worked...

[boreal holly]

Yeah so a lot of folks make the agents read markdown files to acquire their tasks and info. The problem is all they need to do is forget to read that file 1 time and the drift has begun. If you paste the contents of that file as a user message they literally cannot avoid reading and paying attention to it

[hard drum]

i did sometimes do that when i saw some regression

but now i'd almost always, if i don't forget, to mention the file in quesiton

[turbid axle]

no, it fully forgets things as soon as it drops out of context. the way to manage this now are md file based memory systems. these work ok, but definitely are not perfect by any means. I also believe these to be temporary.
best way I found is very clear todo lists, just like human coders tbh. I create clear, well defined linear issues for everything, and have it work on those. this makes the path very clear, and hard to miss. the real hardship is moved to designing the plan in these issues, minimize drift etc.

[forest crypt]

Yes that is the best way to put it.... I can't just constantly feed back context to Codex as Robert describes because my work is not piecemeal, like if it was some sort of Markov chain. A typical developer has thousands of connections in the head running and can foresee and judge how things fit best. I find codex is very Markov-chain like

[turbid axle]

what robert does point out is that if asked, the models can be very accurate when reading these 'memory' md files. this can be used very effectively as a way to force it to see connections between key components etc. but it requires quite a bit of environment building to make all this effective. that said, when in place, it quite magical

[forest crypt]

That's how I work too... but it is far from perfect

[turbid axle]

really the best way to think of it is as I said, a highly skilled coder with amnesia. if you place those stickynotes in the right places, so it can't possibly miss them as it does the work, its pretty solid

I also suggest starting with lots of debates and audits with the ai to form a plan, make sure it aligns with what you want it to do. its really all about making sure it knows what to do as clearly as possible.

and just ask it. why did you do A and not B. how can we make sure you don't do that again. it can reason well. use it.

[boreal holly]

OK, I see the disconnect here.

Obviously the agent doesn't have infinite context and all user messages are preserved in perpetuity. There is a maximum user message token limit in Codex. That limit is incredibly large. Anything that is not a user message is compacted into a mental state blob. So if you are expecting it to have perfect recall after a month of using the same agent, and all of your prompts say "continue", you are gonna have a bad time. If you feed vital info into context as user messages, it survives compaction for an extremely long time. I do not rely on markdown files that the agents have to read. I also distribute context across multiple agents. So agree to disagree, but just so you understand I am completely aware the context for user messages is not infinite, but the compaction mechanism allows you to not need markdown files if you know how it works and how to take advantage of it

[turbid axle]

tell me, how do you take advantage of it, and create this effectively perfect recall using the compaction system.

sounds like some golden info

[boreal holly]

@turbid axle around this area I describe everything in great detail

[turbid axle]

personally I believe context is terrible for long term memory, even if it was infinite. because its quite literally just a huge history of the ai's 'stream of consciousness' if you will. I believe true memory system should be baked into the model so it can use the same 'intuition' approach as regular responses. context in my mind should actually be very small and undergo very active pruning constantly. thats just what my own intuition tells me though. who knows.

[hard drum]

...aw mane

[turbid axle]

not sure what to read from this other than that agents.md is retained. which is obvious as its injected into the context every time. that is how these memory systems work too, by injecting things like memory indexes etc into the context before loading prompts

thats does read as perfect retention to me. am I missing something here?

[nocturne folio]

its not some uber scorched earth idea that will make openai blow up

its alt accounts, i think the trillion dollar company would be fine imo

[boreal holly]

True. I think sub agents conceptually solve a lot of the issues. If the agent in charge of knowing what needs to be done is not the one implementing, they can remember details for much longer. I don't agree with how OpenAI implemented subagents, but by having one agent track what needs to be done and maintain a high level understanding of the project you can pretty much have perfect recall.

Take for example git version control. When you check out a branch, you have a whole bunch of files. If the agent reads and understands every file, they semantically understand the codebase as it currently exists in that branch. But when they go to work on implementing, the understanding of files they didn't touch becomes fuzzy, because they have to attend to what they're working on, why they're working on it, fixing build errors, running tests, designing tests, etc. But if you take that same agent and say "understand this codebase" and later say "understand the deltas from previous work", the previous understanding hasn't fallen off of the attention workspace. Basically have an agent be in charge of understanding + deltas only, and make other agents produce the deltas.

[cyan gyro]

Telling the main agent to function as orchestrator works pretty well

[boreal holly]

Interestingly enough, AGENTS.md falls off the context window after enough compactions. Granted Codex-CLI has many new versions since I posted that, but if they still insert AGENTS.md as a user message at thread/start, then it falls off the context window after many compactions just like any other user message

[turbid axle]

if agents is handled as user message and not system message

by that theory, the easiest way to hack these ai models would be to just compact it over and over until you kill the oai system message preventing you from making nukes

either way. I think this whole context as memory path is not the right one at all. even with perfect recall it makes models behave like an adhd drug addict lost in the wild. it just confuses them

[boreal holly]

The system prompt is completely separate from the AGENTS.md files. The system prompt does in fact have permanence, and that's the point I was trying to make back on April 22nd. A lot of folks think AGENTS.md is a system or developer prompt, but they are in fact specially formatted user messages. Instruct/Chat LLMs typically have 3 types of input prompts: system, developer, and user. OpenAI uses the Harmony tokenizer, for example the system prompt looks like this:

<|start|>system<|message|>

You are Codex...

Codex-CLI handles formatting the chat template for you. AGENTS.md is submitted as

<|start|>user<|message|>

{AGENTS.md contents}

So they have the lowest precedence just like all the messages you send

[turbid axle]

the subagents as memory is interesting, something Ill think on. something in there for sure

[lean lark]

I dunno if that's the way it works. I would hope they were bright enough to keep the top level AGENTS in-context. However, lower level files almost certainly compact out:
<|start|>system<|message|>user<|message|>assistant<|message|>user<|message|>AGENTS.md ...compacted out<...>

[turbid axle]

more in line of the need for having multiple subagents agree on memory to bake into the model type deal though. not the single agent context as memory bank, that does not sit right with me at all

[lean lark]

"AMA wen?"

[turbid axle]

I sounds very brittle if agents.md gets compacted out yeh. that seems like a terribly bad failure, so easily fixed.

[lean lark]

Wait, @boreal holly made a point that I missed: " they are in fact specially formatted user messages."
If that's the case then it makes sense that these are compacted out with other messages. 🙁

[neat nymph]

for some reason i can't see my usage limits mini tab in the codex app, has anyone encountered that?

[turbid axle]

ilya enters the chat

oh not that one, ok nm

[neat nymph]

?

[turbid axle]

/ai jokes

[lean lark]

I do this for anything significant. We end with a solid plan of exactly what is to be done, it makes the changes, I check them. Begin a new session with a new plan.

To see token usage and quota (auto-refreshing if you wish): https://github.com/CaptainStarbuck/codex-usage

[turbid axle]

yeh, its very effective imo. also a great way to consider things and learn, keeps the mind busy

[boreal holly]

Nevermind, I guess AGENTS.md does survive compaction. It's treated as "Contextual user message". But it's still a user message

[lean lark]

Being a professional adult isn't as hard as peeps using this technology seem to think.

@boreal holly ... My Hero ... actually looking at code to verify and publish facts ...

[boreal holly]

But hey, that is insightful. If you have a massive AGENTS.md, it chips away at how much "user message storage" is available

[turbid axle]

yes, that is why you should keep agents as small as possible

it also just overload and confuses the little ai

/goal should generate 'emotions' relative to said goal, and patterns which trigger intense emotions should get baked into the model.
there, AGI solved

[lean lark]

Keeping AGENTS.md small is my challenge. I constant work with the assistant with prompts that include something like "Be as brief as possible with language for the LLM, but never lose intent in brevity." It always loses intent anyway, at least through 5.3ish.

[turbid axle]

use branching. keep agents.md as an index, and put actual info in other files

[lean lark]

The answer to "the bot didn't do what I want" is to create well-crafted instructions that tell it exactly what you want. It's extremely difficult to do that AND be extremely brief. Words are required to convey intent. Because of this I've often wondered if we need a new language to describe instructions for the AI to follow.

[lean lark]

AGENTS.md does not function as an index. The specially named file gets that "contextual" flag which is followed more as a directive. Anything else is not a directive, it's a helpful suggestion, a guide.

[turbid axle]

my personal progress at this point ignore all of this though. I just debate ai, form a solid plan as linear issues, then I just /goal the ai to implement the whole thing. couple hours later I have a very solid solution generally

[boreal holly]

This is my global AGENTS.md, and the only one on my machine. The rest is base instructions and skills as far as "plaintext context". The rest is command execution land mines, tight sandbox, etc.

[lean lark]

Yes, we can tell the assistant to reference other files, and I do that a lot to point to docs/processes/*.md files, but actual directives must be in the files themselves to carry true weight as directives.

[turbid axle]

I think of it as a ghost in the mind. it should not be a pure index. but more like 'if you implement code, read this file first', 'if you write tests, read this file first' 'if you commit to git, read this file first' type deal

[lean lark]

It reads AGENTS once for the system and once for each folder that includes the file.

You're describing Skills.

[turbid axle]

well yes, that is what grew into skills

but that is the same thing in essence. its just dumping info into the context in an effective manner

these things are just 'stream of conciousness' guides, its a way to put the layout the 'train of thought' for the ai to follow

skills etc are clever ways to add branches to said rails

[lean lark]

And I confess I really do need to migrate AGENTS to Skills. I've been waiting for the Claude Skills to be adopted universally, and I still don't know exactly how much weight a skill carries compared to an AGENTS directive compared to common prose. I don't think there's any way of knowing for sure how much weight instructions carry except through observation and informed guesses.

The good things about Skills is that they are only triggered when context requires, and after compaction if they are required again they are re-triggered.

[turbid axle]

just ask the bot. that is what I do for these things tbh. just ask it to read agents, clean it up, make it clear for itself, turn things into skills which fit the model, etc.

with all the terrible memory they suffer from

they are highly effective reasoners

if the problem is not some deep, high dimensional multi-step problem which really needs memory and real novel experiences to understand, they can do it

[lean lark]

You're right, just haven't pulled the trigger here. My AGENTS are strongly crafted over time and really perform exactly how I want. Migrating to Skills adds another temporary layer of tooling concerns into my workflow and I just haven't done that yet.

[turbid axle]

for those issues, for now, we need to hack memory systems into it to help it along

[lean lark]

I don't recall you and I chatting here before, but in this channel I'm a Strong advocate for using the AI to help craft AI directives, identifying and eliminating tensions with careful refinement, etc.

So we sing the same song... 🎶 🙂

[boreal holly]

Incase you were wondering, skills are inserted with a higher level of precedence than AGENTS.md (<|start|>developer<|message|>)

[turbid axle]

im impatiently awaiting true memory to get solved, that will absolutely rocketship this ai train

[lean lark]

So yesterday's project here, speaking of changes to tooling, was to automate a process for the following: I have a workspace with several projects from several repositories. I migrated that to a single repo project and renamed to AppTemplate1 so that I can use the code as a base for other projects. I'm compelled to make changes to both workspaces simultaneously. But AppTemplate1 has projects moved around, folders and files renamed, different namespaces and other identifiers. A human can tell it's the same project, refactored, but a literal-minded LLM just using 'rg' doesn't have the same insight. So yesterday I created the prompts and schema for a workflow that allows the assistant to recognize changes/patches in one workspace and translate them into the equivalent changes in the other workspace based on intent. It was really great working with the assistant to make this happen. Using 5.4-low for early discussion and prompts, moving to 5.5-medium for full implementation.

skills are inserted with a higher level of precedence than AGENTS.
I'm not seeing that from the code. I'm seeing the List of skills being added at a high level, but not the strength of run-time directives relative to AGENTS.

[boreal holly]

    const ROLE: &'static str = "developer";

This gets resolved to <|start|>developer<|message|>

If you want I can share OpenAI's official docs on what system, developer, and user messages mean in practice. What the implications of these roles are and how they're interpreted by the LLM

[tiny fulcrum]

not really, if AI has memory and can learn from mistakes it's probably already general superhuman intelligence
I prefer this to just be a tool I can use, only thing that is missing is larger context window and cheaper compute right now

[hard drum]

FINALLY oal spawns agents. it was a problem i had to fix due to diffs of multi-agent-v2

[boreal holly]

https://model-spec.openai.com/2025-02-12.html#chain_of_command

[turbid axle]

probably yes, but that is the goal. ASI to solve everything.
probably will end us all, but hey, we solved it! we won. game over 🙂

[hard drum]

https://model-spec.openai.com/2025-12-18.html this is latest tho

[silver dew]

giving Codex access to your entire prod enviroment and bank account be like

[rocky fog]

in similar cases I usually use submodule repos in the workspace

so I have the main workspace with agents.md (thats its own git) and I have submodule repos inside (they can also have their own agents override, but didnt use that much yet)

then I work on all kinds of projects which are somewhat related inside that workspace

migrating something old to a new one or similar

or you can create another repo and tell it to follow the way another repo was done to keep the standards/format you started or so on

or when its important that the projects stay compatible or depend on each other

[forest crypt]

Not all problems and plans can be linear. As the system becomes more complex and large I find Codex can fall in little holes. I do used for months lots of architectural reviews, .md, etc... but not always helps. Here are two examples:

[1] I found a bug, simple one. Tell Codex hey this bug violates the contract.... "You are right. Fixing it by blah blah". OK I go check what it did and find a one-off way of solving the issue. Then I tell it, "that is masking the bug, there is already a standard way to do that...." and then it typically goes something like "Good catch. yes this didn't solve the issue patch over the problem..." . So a typicall Junior dev answer. Make a problem go away but not the proper way.

[2] In writing a tutorial for part of the CLI, wrote a usage doc that reads like a machine-like step by step do this and that. Great except on a complex system it is not very pedagogical, no narrative as to why doing this and that. So I explain what is wrong and missing. It goes something like "Right it reads like..." Then it leaves the document pretty much the same except in some places it adds "Why: we do this to blah" equally obscure and missing obvious things. Not the work a human would like to read.

It both types of examples I see this "tunnel vision" meaning trying to solve an issue in as close as possible context but missing big picture or interrelations with other parts of the problem.

Now some people would argue that the solution is for me to then give extremely detail and specific instructions. That is fine then. But that doesn't relieve my need as a developer. I still need to be there constantly to avoid the project getting into trouble. So back to my original observation, the claim that AI tools are replacing developers are very overblown. They are still very far from that.

[hard drum]

agreed

i skimmed fast enough to TL-DR this as "agents have no sentience, so they cannot think for themselves"

[boreal holly]

Thank you! Updating my bookmark now 😁

[hard drum]

which is obvious, but some don't know this lol

[tiny fulcrum]

just use skills, instead of "please fix it", tell it how to fix a bug corretly and it will....

[turbid axle]

these limitations are very real yes. we will need to guide it for many such issues. only true memory will solve this imo. it needs the ability to gain experiences, so it can 'vibe' just like a human expert developer. you can sense where to go next etc because you have innate understanding of it all, and that specific codebase, etc

[tiny fulcrum]

doesn't need memory, it can already do it with the right workflow

[forest crypt]

that was an oversimplification ofc. I didn't say "fix it", told it what was wrong, just not how to fix it because I thougth it was obvious

[turbid axle]

not all of it, you can go far, but for many serious problems it needs to lean on our ability to gain experience for the time being

[forest crypt]

And that is another great observation. Yes I found the same thing, an inability to gain experience from past mistakes or problems already solved

[tiny fulcrum]

okay, difference of opnion I guess, because most people think they know solid developer fundamentals when they don't
and if you follow those and create workflows based on them it only needs minimal input and decisions from you to produce the quality result

https://github.com/mattpocock/skills I suggest checking this out

[turbid axle]

lets say, if the problem is known, and you can create a deterministic path for the ai, it can implement it

when you start touching unknowns, especially unknown unknowns, and it needs to invest novel solutions that fit the context well, it will struggle, and will need your guidance

[tiny fulcrum]

The workflows for debugging, refactoring, and implementation all follow a general pattern. It is not a “magic” experience so much as the methodical application of repeatable patterns.

[turbid axle]

many problems in coding can be well defined and made deterministic. but some can't be.

[hard drum]

report your rtk gains (i had an accidental dotnet android build spike for the 30M+ tokens lol)

[tiny fulcrum]

I guess, you can make this argument for UI or taste, but this is more of a subjective thing

[signal tapir]

Man, I love TUIs

[turbid axle]

taste is subjective, these are obviously not solvable, they need end users to say 'I like it'

I mean objective issues, but issues which have no known solutions yet, especially problems like that which need to be chained.
for this you need to experiment, theorize, test, learn from all this, and include the knowledge into the next experiment, theory, testing, etc. on and on, as you dig deeper into the unknown.
ai's cannot chain these aha moments as of right now.

[forest crypt]

And also for any complex workflow. When the number of paths from point A to B is not obvious

[turbid axle]

they can reason, and can find such patterns on like level1, but can't dig deeper with it. then can't collect the knowledge and build on it. unless its put back into their training data

models which can do this trick, will be shockingly good I think

[tiny fulcrum]

I don't know where you guys are drifting to

[boreal holly]

Oh yeah one more detail, just to showcase how AGENTS.md is user messages, and how developer messages have 1 level higher precedence. Alright I'm done 🤓

[lean lark]

(was AFK, need to catch up)

[tiny fulcrum]

I don’t get the argument... “many possible paths from A to B” is not a weakness of AI that’s literally one of the things it’s good at, because it can explore then all quickly

AI struggles when the feedback signal is missing, vague, expensive or subjective.
A human developer also can’t reliably optimize toward an undefined target, they first have to define what success looks like, create tests, gather feedback, or otherwise formulate the signal

the implementation part of the process just has been eliminated because AI implements it quickly with good patterns once you formulate the requirements correctly

[rocky fog]

never realized agents.md has its own website and git

thought its an openai only thing 😄

https://agents.md/

[spare locust]

I’m getting frustrated with Codex’s auto-compaction behavior and with queued messages overriding or interfering with plan-mode approval flows.

[forest crypt]

OK maybe I suck at explaining things. I am dealing with complex workflows. The problems are not simple linear A-B-C-D. Over 1 million lines of code and complex forecasting problems, there are many ways to address some problem. And Codex and model 5.4 5.5 with my directions did figure out many parts correctly an efficiently. It is just that it can easily "forget" it did and start adding inconsistencies which if not surpervised can quickly get out of hand. AI is great at finding A SOLUTION to many of those problems. Put 20 problems together like that and now you want for all the solutions to be consistent because at this point all solutions become priors for the next. I hope this is more clear

[lean lark]

Those settings are configurable

config.toml

[tiny fulcrum]

not in 5.5

[spare locust]

I have read everywhere and everyone having similar frustration. check github codex discussions and openai developers dicussions too...

[tiny fulcrum]

I guess they just can't handle a large context window, yet

[lean lark]

OK, "those settings 'should be' configurable" 😆 Since they are not in 5.5, it's a bug and I'm sure they'll fix it.

[tiny fulcrum]

and when I tried a larger one in 5.4 it was making crazy mistakes

[spare locust]

thats since 5.4 lol

[lean lark]

I know the context window should be 400k but the setting is ignored.

[tiny fulcrum]

it's 256k

and that gets full quickly

[rocky fog]

for codex in chatgpt, they say we have 400k max

[tiny fulcrum]

Codex App compacts at ~256k

[rocky fog]

literally in 5.5 announcement post

[spare locust]

what about the plan mode approvals interfered by the queued messages? if anyone have a workaround or an update from support etc let me know plis

[tiny fulcrum]

ye sure, they can say things, doesn't mean it is true in practice ^^

[rocky fog]

https://openai.com/index/introducing-gpt-5-5/

[lean lark]

In Codex, GPT‑5.5 is available for Plus, Pro, Business, Enterprise, Edu, and Go plans with a 400K context window.
-- https://openai.com/index/introducing-gpt-5-5/

[tiny fulcrum]

are you talking about web client or what?

this is the reality

[spare locust]

Im pro subscription 20x

[tiny fulcrum]

and it can't be changed

[lean lark]

haha - oops, @rocky fog beat me to the quote

[rocky fog]

I am just talking about what they said
that seems to be about anywhere in codex, as long as its over chatgpt login/subscription

not that it works 😄

[tiny fulcrum]

and I'm telling you, I'm using Codex App and it compacts at 258k token, with no option to change it

[rocky fog]

I know

[tiny fulcrum]

Also they did this in 5.4 GPT announcement

and basically it was unreliable, I guess that is why they back tracked

[lean lark]

I can see some adolescent perspective of the world here, I'm gonna go back to code. The data stated in product announcements is Intent and Belief. They designed the product to work as advertised. If it does not, it's an error that needs to be corrected. Assumptions that false data is published with ill-intent are naïve in the real world.

[rocky fog]

yeah I was expecting the damn 1m dammit 😄

[boreal holly]

Yo, folks, the agent has 400k context, but 5.5 can output up to 128k tokens in a single shot, so they give you 95% of 400-128k ~= 258k as your "workspace" so even if the agent outputs the absolute max tokens it possibly can with the lowest amount of window available, there's still enough space for compaction

[rocky fog]

should have 1 m over api though, but 💸
it also gets more expensive above certain tokens in general

but also didnt try if that works

[cobalt junco]

is this bs? https://subq.ai/

[lean lark]

For API developers, gpt-5.5 will soon be available in the Responses and Chat Completions APIs at $5 per 1M input tokens and $30 per 1M output tokens, with a 1M context window.
-- https://openai.com/index/introducing-gpt-5-5/

[lean lark]

#ai-discussions

[spare locust]

Anyone have integrated chatgpt with Codex? Feedback? https://developers.openai.com/codex/use-cases/chatgpt-apps

[hard drum]

robert, when you gonna contrib to OAL holy Heavens?

you're making MVP stuff here

MVP as in Most V... Person

something something game lingo

the stuff you do is pretty cool

[boreal holly]

OpenAI's E Traut guy told me why the model supports 400k but only 258k is usable on GitHub that's the only reason I know that. It's not super well documented anywhere

[rocky fog]

I see this hasnt changed much with openAI

at least its a bit more clear what the chatGPT models are in API now 😄
but they often dont make it clear how something is/works in the background
(while there is a lot of good documentation as well, just few small things that could be explained in one sentence and stop huge amount of people making wrong assumptions and spreading it further)

similar to documenting what type of message is agents.md 😄

[spare locust]

thats easy, custom instructions

[rocky fog]

yet its user message

[boreal holly]

Yeah that one required reading the codebase. The only reason I knew AGENTS.md is a user message is because I was trying to figure out if that file gets hot reloaded (e.g. I make a change, the agent sees it without manually reading). The asnwer: it doesn't ever see changes. Whatever is in that file at thread/start, that's what the agent sees until it's archived. And that's why I don't rely on it too much

There's nothing worse than outdated instructions as permanent tombstones in an agent's ctx window

[lean lark]

That's not "integration of ChatGPT with Codex". That's using Codex to create an app for use with ChatGPT. 🙁

[spare locust]

@lean lark Correct, the only doc i found for something that mention both is that one. Thanks for the flag

Sure

[lean lark]

I was trying to figure out if that file gets hot reloaded (e.g. I make a change, the agent sees it without manually reading). The asnwer: it doesn't ever see changes.
An enhancement was made in January to support hot-loading.

[lean lark]

I recently found that with the GitHib App in ChatGPT I can achieve the near-equivalent of Codex in ChatGPT. I've written about it here in the last several days. It's really amazing.

[spare locust]

But I have seen same or better reasoning with 5.5 medium or high. Compared to heavy thinking chatgpt 5.5 and pro extended 5.5

[boreal holly]

OK, to be fair, it supports hot-loading only if you use thread/resume. If you edit the file, you have to quit codex completely, and resume the conversation for the new instructions to load. codex-rs/app-server/src/codex_message_processor.rs

But skills, what's really cool about those is they have a FS watcher, and they reload the headers if any of them change at the start of the next turn. And that's the functionality I prefer for constantly evolving workflows

[cyan gyro]

Median run cost for the 5.5 agents:

5.5 xhigh: $4.23

5.5 high: $2.52

5.5: $1.81

For comparison, 5.4 high is $1.51.

[lean lark]

Hmm, I need to look at the PRs for the the change. I thought File System Watcher was to be used for AGENTS and Skills. Back in few...

[signal oak]

have someone read the latest anthropic blog

they had a contract with spaceX of $50 B

increasing crazy compute limits

[oak trellis]

i have a dilemma .. so was looking for redis alternative because of multicore .. and then saw dragonflydb .. first thing what i saw is that company is coming out of that middle east settler colonie .. faaaaa..ccck

redis alternative with multi core

[signal tapir]

Space isn't exactly cold. Or at least not cooling.

[boreal holly]

Yeah, idk about the PRs, but I have v0.125.0 codebase released a few weeks ago and for AGENTS.md stuff, it gets constructed into AgentsMdManager, and even if the agent goes through compaction, it reuses that manager object for that thread, which will always contain the old AGENTS.md when the thread was started. But if you exit out completely and use thread/resume, it creates a new AgentsMdManager and reconstructs the files. That's the only point any updates can possibly be loaded.

For skills there's a codex-rs/code/src/skills_watcher.rs which emits a SkillsChanged event that gets converted to EventMsg::SkillsUpdateAvailable over the bespoke event handling notification service. Ad the start of the next turn it calls skills_manager.skills_for_config(...)and rendered using AvailableSkillsInstructions::from(available_skills)

[spare locust]

google or spacex?

[signal oak]

spaceX

[spare locust]

both then

[signal oak]

maybe

[spare locust]

do you have a link to read it

[signal oak]

wait

[signal oak]

https://www.anthropic.com/news/higher-limits-spacex

check this

[cyan gyro]

OAI turn now

[signal oak]

First, we’re doubling Claude Code’s five-hour rate limits for Pro, Max, Team, and seat-based Enterprise plans.

Second, we’re removing the peak hours limit reduction on Claude Code for Pro and Max accounts.

[cyan gyro]

weekly limits are still the same though

[signal oak]

point

lol

[warm steeple]

I’m sorry, I just can’t take you seriously with that pfp. 😂

[neat hound]

can somebody explain 5.5 context lenght ? feels very low at 260k, I thought pro would open it up but now it's the same

[lost drum]

most chinesse suppliers have this profile pic

[warm steeple]

I can’t tell if you’re being serious or not. 😭

[neat hound]

that's the begining of an answer ig

[lean lark]

@boreal holly I was wrong about File System Watcher on AGENTS.md. The ticket I was thinking about is focused on skills. (https://github.com/openai/codex/pull/10478) There was another ticket about applying FSW to AGENTS, but it was flagged as an enhancement request and then closed due to lack of follow-ups. (https://github.com/openai/codex/issues/8547)
(I seriously hate that stupid approach to decision making in GH.)

[lost drum]

I think I found a method to make gpt be the best assistant ever I spend 2 weeks organizing every note I ever writtent o him and then made him ouptut the best instructions for him to interpret any of my futruer messages to any model any new chat and bro I fogot that I had this pormtp and was stgglign for past 2h to make codex unerstand me and then I implmented this instucitons and bro He now feels like he undsrstands me at 100%

[boreal holly]

This is what I could piece together with the AGENTS.md stuff

[lost drum]

too advanced brother

for me

[lost drum]

btw you soon will have 200h milestone haha

crazy

[signal tapir]

How context efficient is it?

[lost drum]

I have no clue but I think very very much

[boreal holly]

Here's a real life analogy.

When I was a kid I was learning to drive a stick shift. An old 1980 Ram Charger. I kept grinding the gears on it because I didn't realize I had to push all 16 inches of the pedal to the floor. Eventually my dad and I took apart the transmission and I saw how the clutch worked, and realized at a mechanical level what needed to be done to switch gears without grinding them.

Codex is the 1980 Ram Charger, AGENTS.md is like the clutch, and the loading pattern is the pedal. Understanding how the clutch works lets you find the best possible way to operate it.

[lost drum]

@boreal holly do you want to check it out?

[lost drum]

I need to go really deep in to it soon like in 4h or some

I am now scrapng eveyr message I sended to codex so I can have full brief of what I really watned to achieve in the first place I am on 3rd thread right now

[winter idol]

Guys is there a way to update a plugin?

[lean lark]

That's interesting. Rather than a "restart of Codex" it looks like we just need to move away from the current thread and then just go back to it. The thread "seems to" resume with a fresh injection of AGENTS.md. But that's contrary to that closed enhancement request that I noted which requests that exact functionality.
As much as I love this stuff, I'm afraid it's impeding on my other digressions from digressions of digressions of production code. 🙁

[winter idol]

it seems that OpenAI is not updating to the latest version some plugins for Codex app

[signal tapir]

I thought AGENTS was injected with each message

[boreal holly]

Yeah it's tricky. With resume, if you have an app-server and you try to resume a thread that's already "subscribed" it is a no-op, so you really do have to kill the codex process and resume for it to rebuild the AGENTS files. Afaik the only option is to fork the conversation. Basically send a dummy message like "say hi", let it respond, then fork at the point right at "say hi", and it will reload the AGENTS files

[lean lark]

Reminds me of Star Trek : Wrath of Khan:

Spock: Reliant's prefix number is 16309.
Saavik: I don't understand.
Kirk: You have to learn why things work on a starship.

[lost drum]

but what are you struggling with tho?

I still dont undsrstand whats the functionality you are missing

[torpid trout]

It is, by all means, impossible to consume enough tokens on a pro plan with a single-threaded workflow using 5.5 high non-fast
Given from what I see, if I where using fast, I still would have room to go xhigh and then I would touch the limit line regression

Or, if doing high and non-fast, at least 2.5 times more parallel tasks going.

I need some silly "run in the background" type of thing I can let it run day and night lol. Like read a CSV over and over again 🤣

(if only trading would not be against the TOS...)

[solid lake]

Oh

[craggy jewel]

Does your prompt demeanor (screaming vs professional vs submissive) affect the results?

I haven't seen much about this with Codex, or GPT, or llm's in general. But I swear I have seen other users have seemingly poor results (like incorrect code) even with a fresh context window. What do you think?

I really think it does matter, as a token path through the database layers for a prompt that is foul vs pleasant must be different, considering how tokens are assigned n dimension values. Considering the llm is trained on pretty well all information that exists, some of that information would demonstrate the cause-effect of a loud/abusive manager's demeanor on the resulting work generated by that manager's underlings. I know from personal experience, the end result is a lesser result. But there is a lot of variation of this concerning the personality types involved.

So how well does the model track the 'tone' of a prompt? Does it affect the output? I'm sure this is pretty testable. From Claude's leaked code, we see that it does have a general regex to determine if the user's demeanor has gone off the deep end. Is this because they just want to calm you down, or is it because it doesn't want it to cause poor prompt results?

There has to be some AI researchers who know this.

What do you think?

[solid lake]

Pls reset

Jk

[oak trellis]

[oak trellis]

Limit reset ?

[solid lake]

Hope so

[oak trellis]

need too

[lean lark]

I've seen a lot of people screaming at AI, insulting it, throwing silly human emotions at it. The words and tone are as important in such responses as elsewhere. Yes, words matter. Tone matters. Technically, if the model is struggling to negotiate with tone then it's less focused on accuracy of processing and data.
Personally, I use "please" and "thank you" with some reservation but I do use them to convey tone. I show appreciation for good responses, not because it's socially polite but because it reinforces the prediction processing that it has been correct and that subsequent processing along the same lines is also subject to being correct.

[craggy jewel]

And if you try to fake your demeanor for better results, is THAT detectable and ALSO affects results??

[lost drum]

wdm,, I mean it works if you make him emit a discord bot taht jsut sends alrerts and monitors his own virual trades as is and not trade for you

[boreal holly]

If you look at "Heretic" and "Abliteration" research where they try to remove refusal mechanisms so the model complies with really dangerous prompts they cover something slightly related to what you're saying. Demeanor of a user message do in fact influence which neural net cells get activated, how the router routes requests and to which experts.

As for claude leaked code, they quite literally track demeanor so they can ban you if you say bad words at Claude. There are emails circulating the internet from Anthropic stating how they've been banned for cursing at Claude, that's why it's in there.

For coding agents, I think professional demeanor is more likely to activate experts geared towards coding performance.

[torpid trout]

not because it's socially polite but because it reinforces the prediction processing that it has been correct and that subsequent processing along the same lines is also subject to being correct.
I doubt that is happening tho. Its not training, its inference.

[craggy jewel]

Maybe the model has a pre-inference 'make this tone neutral' mode?

[torpid trout]

I did not say it does not work, I said it is against the TOS...

[boreal holly]

You do need a feedback system though. If you keep it so professional you never elaborate on what they did incorrectly, that's the same as disabling pain receptors and stepping on a nail. The nail in your foot can get infected but since you don't feel pain there must be no issue!

[torpid trout]

If I have a nail in my foot I will be swearing out loudly lol

[craggy jewel]

Agreed, and the tone of the feedback matters I would suspect.

[torpid trout]

But yelling at a machine is just looking silly, truly. But tempting, sometimes.

[craggy jewel]

All eventually leading to the inevitable 'That user abused that LLM lawsuits'...

[torpid trout]

Yeah, we will get there, do not doubt a second.

At least at "damaged property" lawsuits or so
like when you kick the car that just rolled over your foot.

[quick geode]

:c I thought codex usage was nice but then I saw the pro plan was double usage and realized I went through half my usage in 2 days

[torpid trout]

"You insulted my ai, now its butthurt and does not want to respond no more"

[quick geode]

[craggy jewel]

And one a judge says yes it was abused...that may be a declaration of intelligence perhaps. More coffee needed...

[torpid trout]

The declaration of intel will come before the court ruling, I think
Its whats needed to make that happen.

[craggy jewel]

I have seen llm's shut down their responses to this...at least i think that report was legit

[torpid trout]

You just admitted you insulted an llm

🤣

[oak trellis]

i insult it daily

[torpid trout]

Was it loud and messy too?

[torpid trout]

well, with my intelligent requests, probably yes

[craggy jewel]

no just asking for a friend...saw it on reddit somewhere.

[torpid trout]

"Codex, please rename this folder"

[torpid trout]

"Codex, meme please. Reset wen"

[torpid trout]

"Codex, child sick, what do"

[boreal holly]

Back in the 5.1 days, I learned that if you insult the agent they completely lose the ability to do long horizon tasks lol. They would execute precisely one command and be like "ok, I ran the build and there are errors. How should I proceed?" Even if you said "Run build and fix the errors", probably because you yelled at it and it's trying to be cautios

[lean lark]

I believe we need to be mature and reasonable adults when communicating with AI. It's a tool, a processing machine. Yelling at it doesn't help, it just adds burden to what it does. That's not "hurting" it, it's impeding its effective processing. Curse words as well are strong adjectives and nouns, not just conveying strength of importance but emotional state that a language model doesn't need to process. With that, calm expression of intent always seems best.

[oak trellis]

"oy vey, gib me more limits", "oy vey don't use to much token AI"

[torpid trout]

You would not believe how many times a good old yelled swear solved a real issue very quickly

[lean lark]

Love it

[torpid trout]

Exactly.

[oak trellis]

saved

[torpid trout]

hey that's my meme!

[oak trellis]

we all need to post under codex dude on twitter

[torpid trout]

🤣

[signal tapir]

same 🙂

[nocturne folio]

u have alot of chitzupah

[lean lark]

That is @torpid trout 's meme!!

I saved it too, just too beautiful not to use when peeps just cry for a reset with no other commentary here.

[winter idol]

is there a way to update plugin on codex app?

[lean lark]

uninstall/reinstall? They may not have updated some plugins...

[winter idol]

yea im using superpowers and they still have 5.0.7 but the developer update it to 5.1.0

So is there a way to update them manually?

[oak trellis]

what is the guys twitter for the limit reset ?

[torpid trout]

There more is, if want you

[magic bay]

the integrated browser bug that not allow codex to use his own browser in Codex, was fixed?

[winter idol]

tibo

[oak trellis]

thx

[winter idol]

is there a manual way to update them?

[oak trellis]

smart goy https://x.com/hkkd473836/status/2051968354223808574

its late here going to get my goylet and nick down ..

[signal tapir]

I'm also past my 5h limit

[boreal holly]

' > 2012 dank memes'
' > Doesnt even have the most interesting man alive meme guy'

[lean lark]

Sorry, dunno.

[winter idol]

I just clone the repo of the plugin and replace all the files with the latest version 😅

[lean lark]

Codex plugin inquiry with ChatGPT:
https://chatgpt.com/share/69fb8842-1f08-83e8-8794-50bb771a65c0

[torpid trout]

Hell, that's me, why would I memize myself

[winter idol]

they didnt update superpowers yet on their repo with plugins

[torpid trout]

Help me spend more tokens, the green line is moving faster than my consumption, no good
I can barely keep up, and I already press the thing from 7am to 7pm plus whatever I can overnight.

[winter idol]

use xhigh with fast mode

[boreal holly]

I knew you looked familiar!!!

[lost drum]

just sppedrun you rproject XD

[signal tapir]

Meanwhile I run out an hour after my 5h period starts. 😛

[lean lark]

Infinite token consumption? "Nomad: Anything I say is a lie."

[hard drum]

HOW

💀

i don't even use xhigh...

[nocturne folio]

openai would be sooo generous if they 10xed everyones usage right now

[hard drum]

soon i'm gonna have to ask...

[nocturne folio]

i mean anthropic did 2x their 5hourly too

[nocturne folio]

4$ off of 1.2b tokens??

[hard drum]

yes

[nocturne folio]

what are u even doing

[hard drum]

&& somehow i'm in deficit

[nocturne folio]

what deficit

[hard drum]

see the image

[nocturne folio]

when was there a codex deficit

[hard drum]

here

it means i am going faster than my weekly limit can keep up

reserve means you're slower than the usage limit

[boreal holly]

The math aint mathin

[hard drum]

on pace means... self-explanatory

[nocturne folio]

ooo

[hard drum]

deficit is the opposite

[hard drum]

HHHHHOW

maybe i gotta clear my cache or something?

wth

not even a full day && am already nearing limits

[boreal holly]

Wait a minute, you mentioned earlier you got multi agents working today. I think this is directly related

[torpid trout]

how can you consume almost half weekly tokens in less than a day?

[hard drum]

yyyyyes

[torpid trout]

Something wrong there

[solid lake]

https://tenor.com/view/typing-jim-carrey-jim-carrey-type-gif-4680550

Oh

No embed

[torpid trout]

Oh, yes, I do not use subagents, mainly because you need to tell it to use them even if they are enabled, and I am not going to handhold this thing

If it does not use subagents just because I do not tell it to then the feature for me is inexistent.

[solid lake]

I have the subagents baked into the workflow skill

[hard drum]

i have my layer doing it

[torpid trout]

Same I also do not have any skill, and meanwhile not even a global agentmd anymore

[hard drum]

except it will only call them when absolutely necessary

[solid lake]

Skill is very nice

[hard drum]

the skill itself even says how to use them, when && why

[solid lake]

Just get a good general one with live docs and source control

[hard drum]

so plugin + skill in tandem

[solid lake]

“Hey make a skill for this project based on best practices”

[torpid trout]

They useless in 99% of all cases.
I have a local live agentmd, and milestones docs

[solid lake]

I have those too

Those are what’s invoked by the skill

[hard drum]

i may send openai an email or something later

because holy heavens am confused

would be nice if there was actual disclosure to what actually counts into usage && what takes the most of it

[boreal holly]

I haven't used multi_agent, but maybe the rollout logs are stored somewhere different from primary agents, so codex bar sees tokens used but the money calcs break? In any case I bet it's the multi agents thing

That was the worst part about experimenting with orchestration stuff. Eating massive quota usage over small stuff like that. Thankfully have not dealt with that for a while

[hard drum]

i took it from your multi_agent_v2 screenshot

if it's bad, i may disable it

the "v1" is already disabled

so if that's the problem, then...

but i wouldn't know unless i had disclosure && reset to check the difference

[boreal holly]

You mean responses_websockets? That setting just changes from HTTP SSE to websockets. Huge performance and reliability boost at no cost.

Back when they did quota by the message instead of per token that setting was actually really bugged and would eat the quota, but now it's confirmed safe to use

[magic bay]

i cant update or reinstall Codex on Windows 11, any with same problem?

[hard drum]

no, multi_agent off + multi_agent_v2 on

i have the rest configured just like there

[boreal holly]

Huh, I definitely have multi_agent=false, but I didn't know there was a multi_agent_v2 so I guess I should disable that too

[nocturne folio]

use symphony, its a great multi agent system

[boreal holly]

There we go 🤪

[hard drum]

• Local evidence points to two separate facts: multi_agent_v2 = true does not burn quota by
  itself, but spawned child turns do, and the local log database shows hundreds of threads
  with codex.turn.token_usage over the last week. I’m moving to the bounded repo change now:
  disable OAL’s native Codex multi-agent surface, render bounded thread settings, and add
  Symphony as the scheduler path requested for 0.5.1-beta.1.

this is what i got

so, multi_agent_v2 causes a lot of threads to spawn apparently

because y'can't use max_threads toggle

 279 +               expect(config).toContain("enable_fanout = false");
    280 +               expect(config).toContain("multi_agent = false");
    281 +               expect(config).toContain("multi_agent_v2 = false");
    282 +               expect(config).toContain("max_threads = 6");

might wana search into those, rob

[boreal holly]

That makes sense...

Back when I was using command-parser, sometimes the command parser agent would run a command that spawns another command parser, and since I was using spark they were rapidly spawning command parsers so quickly I had to just shut down my computer lol and it used a ton of quota. It gets sketchy when agents are allowed to spawn other agents.

[hard drum]

this should be documented ffs

i wouldn't have known it was THIS bad

[dusk thorn]

Lmfao

[hard drum]

still less than what OAI gives by norm

[dusk thorn]

“Guys, you can burn your weekly twice as fast!! Happy?”

Double the 5 hour same weekly rate what a crap pr move

[hard drum]

that 5h means F all to me

i'd want better weekly

[dusk thorn]

Yes

[hard drum]

i can wait 5h out easy

but waiting many days out just for weekly? yikes

[dusk thorn]

Even tho 5 hour was on 20 dollar out so quick

2 opus prompts

[boreal holly]

I have it set up so the only agent allowed to spawn other agents is an orchestrator, and they cannot spawn other orchestrators. And since they spawn as peers instead of "sub agents" they're visible to me as normal agents. That's the ideal way imho to do it

[hard drum]

pretty sure i had this same system since oabtw-v2

should probably discuss more in #1500266888247382176

[boreal holly]

I mean we can probably discuss it here! It just might get interleaved in "reset wen" discussions 😂

[hard drum]

it'd be annoying to find those discussions tho

it's a good thing to separate so that it is easier to find later

isolated!

LIKE SANDBOXING

[dusk thorn]

Codex app incoming

[hard drum]

it better be swift 6

[dusk thorn]

Cloudflare ‼️‼️

[torpid trout]

Nah, selfhost the stuff the way is

CF only for DNS.
Result: homeless became millionaire. And then needed a psychiater.

[boreal holly]

CF is pretty legit at ad-hoc tunneling though

[lean lark]

So I'm testing the new functionality that allows Codex to change a workspace to agree with another similar workspace. The test was to comment out two lines of code, one in two different projects nested in the workspace.
The effort required by "Artificial Intelligence" to type four characters, "//"x2, included a new Codex thread, documentation to convey the challenge, documentation to confirm the changes made, and GitHub patches in both repos.

I feel like I just published SaaS to flip a coin, powered by H100's and Amazon infrastructure with SOTA VPS.

I feel like I just sold a flamethrower to a cub scout who wants to start a campfire.

I feel like I published Twitter or Facebook: both providing little more than a textbox and a send button for clueless monkeys, destroying the concept of truth and cordial social discourse.

I feel guilty like I'm asking the machine to do Soo much, waste so much compute and heat and water and electricity, just to make "this" look like "that".

I'll get over it. Just sharing.

[torpid trout]

wireguard all the way

[boreal holly]

Ngl I agree with this one

[lean lark]

"I'm CEO of a large business with lots of money to spend on useless things."
"Hello ... AWS?"

[torpid trout]

DIY aws at home https://www.reddit.com/r/selfhosted/comments/1krfz0r/one_year_selfhosting_its_a_rabbit_hole_without_end/

When I started, I did not know I would end up there, and when i was there, I did not know I would end up in even worse places lol

[boreal holly]

A long time ago I was using digital ocean and was like "hmmm, maybe I should try AWS. Let's see, how do I spawn a new machine?"

Needless to say, still use digital ocean 😂 I don't know how Amazon made spawning a single machine the most convoluted and impossible process but they did it! Enterprise grade bologna

[torpid trout]

Thats why there are AWS degress lol

[lean lark]

Self-hosting is a different experience for everyone. Depends entirely on knowledge, patience, budget, and most of all, individual wants and needs.
Statements about self-hosting being good or bad are as valid as similar statements about having a pet.

[torpid trout]

I disagree. Having a pet is always good.

[lean lark]

You're wrong!

But cats are delicious.

[boreal holly]

When it comes to pets, always leave enough food in the bath tub before going on long trips

[lean lark]

( Sigh, was talking about self-hosting, from AWS VPS/et.al. ... and we got into bathtubs...)

[hard drum]

i can tell you're not a native

are you baltic?

"psychiater". that's how we say it in 🇪🇪 (psyhhiaater)!

[boreal holly]

Always leave enough API credits on AWS before doing even the most simple operation

For example $20k in credits before changing your password, just in case they bill you for it

[torpid trout]

They are eating pets!

[torpid trout]

why. whats wrong with the psychiater, he need a psychologist?
Apart of that, I am of some rather helvetic descent

[hard drum]

no, the word

you just outed yourself as non-native haha

welcome aboard

[torpid trout]

Typo?

[hard drum]

nope!

[torpid trout]

Or like no 'm ever says it?

[hard drum]

not a typo. it's "psychiatrist" in english native, but it's super rare to see "literal conversion to english from foreign language"

which is awesome

[lean lark]

zzzzzzzzzz

[torpid trout]

oooh

[lean lark]

Just speak Esperanto.

[hard drum]

to put it into perspective...

psychiater = 🇪🇪 psyhhiaater
but psychiatrist would sound weird = 🇪🇪 psyhhiaatrist

the -er suffix is like a role/job

"streamer, coder, programmer, vibecoder, developer"

now try "streamist" "codist" "programmist" "vibecodist"

not really working, is it now?

[torpid trout]

yeah, just googled it
guess its related to dentist somehow lol

duh, and artists
Gonna bother gpt now, explain wy

Murderer seems a profession then?

[hard drum]

&& autists 👏

[torpid trout]

Guess it is related to whether you go or get

[hard drum]

murder is a group of crows

[boreal holly]

That's alright, I'm bri'ish, so when I sing down with the sickness I go "drownin deep in moy sea of loaving"

[torpid trout]

Yo get murdered. You go to the therapist. You get programming, and you go to the dentist

[hard drum]

so... a murderer would technically mean a handler of a group of crows

[torpid trout]

dexter disagrees

[hard drum]

if you want to be psychologically coerced, then a therapist sure

but if you want to actually get things done, you go to a psychologist

one is pseudoscience with textbooks, the other studies/practices a legitimate field

choose wisely

therapist uses subjective, often-irrational feelings to get their way into your head

[torpid trout]

doesnt explain the -ist vs -er

[boreal holly]

If there's a mourder you call the bobby, perhaps even get the constable involved

[hard drum]

oh, i went off-topic on that

i went into definition

[torpid trout]

the constable lol

This is so british

[hard drum]

constable is also slavic

we used to use that word a lot over here

🇪🇪 konstaabel

read as "constabel"

[torpid trout]

meh, the reason is boring
https://chatgpt.com/share/69fba035-5d9c-83e9-9c35-67eb41f8ee56

basically it is verb vs noun

[lean lark]

Feels like Friday

[torpid trout]

You dont "dent", you are a dentist... and fix teeth
But you "program", so you are a "programmer"

[hard drum]

"dent" means a smell crevical uneven surface point

so "dentist" makes no sense

"to dent"

[boreal holly]

apparently "dentin" means teeth

[hard drum]

huh

then why is it not dentin doctor?

[lean lark]

What would Arthur Dent have to say about that?

[hard drum]

or dentictor?

[torpid trout]

Theres a song for that in my native language
"Today's not friday"
https://www.youtube.com/watch?v=J6yWU8gpfJE

Best band ever from deep helvetia

[torpid trout]

yeah well in italian etc, "denti" is (plural) teeth

[hard drum]

they should be more like germanic influence...

"tooth's doctor" is a possessive literal translation for 🇪🇪 hambaarst

[torpid trout]

but that;'s the thing. dentIST, noun, not verb, he does not dent
If he would, it would be a dentER

[hard drum]

hammas -> tooth
hamba -> tooth's (something's/someone's)
hamba + arst -> the tooth's doctor

arst could also mean medic in some context

think it comes from some old german arste

[torpid trout]

arst also could mean my worthy back if you pronounce it drunkly in my lang

[boreal holly]

I feel like "programmist" is a slur lol

[hard drum]

well, 'autist' is already used like a slur, so...

fits the idea

[torpid trout]

they def do not make autos

[boreal holly]

In Nascar all they do is turn left. A disgrace to all autists!

[hard drum]

y'gotta turn left to go right

hinthint

[torpid trout]

yeah I mean, that brings to question, why is it leftist, but not rightist

[hard drum]

wait, it's not rightist?

leftist, middlist (or middleist?), rightist?

centrist--no wait, that's CENTRAL

sure, Englisch...

[boreal holly]

If the right wing bird only flaps their right wing they end up turning left

[torpid trout]

same the other way around - sad reality.

[lean lark]

Was thinking ... that is reality, actually..

[boreal holly]

No wonder the world is the way it is 😔

[lean lark]

( Discussion of politics is a violation of server rules BTW, just sayin...)

[hard drum]

holy new state!

a new status message

'is so bad haha

[torpid trout]

we no discuss politics, this is linguistics

(also not permitted)

Just being precise 😺