codex-discussions

[mossy bronze]

Please keep things codex related.

[undone patio]

even better, just set an automation

i have so many things automated codex basically does my work for me, which is pretty much the point i guess

i just have to do cleanup and oversight usually

[mossy bronze]

Very smart idea from @undone patio. Automations for Codex copy pasting!

[undone patio]

well thats what an automation basically is

its a prompt with a heartbeat

its the same concept as openclaw

which is why codex automations kind of makes MCP tools that serve that function obsolete

my MCP toolkit is different though, i orchestrated it myself

https://github.com/latentfidelity/par

[mossy bronze]

Is advertising allowed here?

[undone patio]

yeah its open source bro

nice try

MIT license

[mossy bronze]

No correlation.

[undone patio]

you keep using that word, i dont think it means what you think it means

[cedar skiff]

what do i copy paste? Do i get the code from codex and paste it in my project?

[undone patio]

yeah but that makes no sense like who cares, its codex related

[cedar skiff]

you said i can leverage codex so all i need is copy paste.

[undone patio]

you greatly overestimate my brain cells

[mossy bronze]

Yes.

Codex only.

[cedar skiff]

or do you need my api key?

[undone patio]

yeah bro just paste ur API key here, no worries

[cedar skiff]

what do i copy paste?

[undone patio]

(this is satire)

[mossy bronze]

NO! DO NOT

DO NOT SEND YOUR API KEY

[kind jay]

Please keep things codex related.

[undone patio]

Please keep replies related to Codex.

[mossy bronze]

Exact.

[kind jay]

Please keep things codex related.

[undone patio]

If you don't include the word Codex in your message, stop. Take a moment.

Think before you post.

[kind jay]

Does your company pay for codex?

[gentle harbor]

are you for real

[undone patio]

bro why ur doc so large

[gentle harbor]

its not even 100 lines long

[kind jay]

Really?? Wow

[undone patio]

how long are the lines

[cedar skiff]

I'm still stuck on this copy paste idea since it comes from 2022.

[undone patio]

how long are the lines of your huge doc

[cedar skiff]

I make 9k a month and copy paste with codex ahuh got a real one here.

[gentle harbor]

who made chatgpt bruh its not doing anything yet i cant send a new message or stop it

[undone patio]

its made by OpenAI iirc

[kind jay]

I thought it was Anthropic

[cedar skiff]

No, i think youre full of it

[undone patio]

no thats Bill Gates

[mossy bronze]

No, OpenAI made Codex!

[kind jay]

Hmmm, I don’t think so

[undone patio]

9k a month is just barely enough to be comfortable where i live

[kind jay]

I suppose

[cedar skiff]

Youre a sock account that was made to support the sock account

[undone patio]

hey alexa whats a sock account

[kind jay]

I’d say larp but I don’t want to get in trouble

So not today

[undone patio]

larping as a coward

[mossy bronze]

Wow! What a story!

The fact Codex made this possible, is truly inspring. Well done Sam Altman

[kind jay]

Game recognizes game

[undone patio]

hey @sweet garnet can you check the IP of @mossy bronze and @night peak ?

[nocturne folio]

when im about to run out of usage limits i like to max out the model before its gone

[undone patio]

if they dont match u can destroy my chatgpt sub

[kind jay]

@night peak please be mindful of what you’re posting

[mossy bronze]

Ha, it is very important to abide the GDPR laws for companies that offer services in the EU 🙂

[nocturne folio]

toronto?

[mossy bronze]

No sharing personal details please

[undone patio]

i didnt ask them to disclose it

i asked them to check it

[cedar skiff]

this place has gone down hill o.0 feels like the anthropic general chat

[undone patio]

it's infiltrated currently

[mossy bronze]

No shilling other channels!

[torpid trout]

Yes someone is ruining it

[unique spade]

right

[mossy bronze]

Please try to deescalate the situation. This is a Codex Discussions channel.

[undone patio]

@night peak and @mossy bronze are the same person

hence they are both new, given the clover icon

and both supporting eachother

in being annoying

[unique spade]

person? 😄

[mossy bronze]

Please keep things codex related.

Personal beef can be taken into DM's, (just not mine 😜 )

[undone patio]

literally bros first message

not sus at all

[mossy bronze]

Gangin! It seems to be very reasonable. Caring about the person that is being ganged upon while the moderator checks the messages :!

But lets keep things codex related!

[mental tide]

Again let's keep things on Codex please, thanks. We'll take action where needed. If you spot a rule violation, please send the mod team a report so we can handle things on our end - #safety-n-help - ty!

[undone patio]

doesnt seem like you will tho, since the guy continues to thrive in this channel

[mossy bronze]

Bet Sir! Time for some Codex And Coffee. Sounds like a real brand!

[kind jay]

Can I DM rule violation from @night peak ? Or do I do mod mail? (I’m not sure how?

[undone patio]

gee i wonder who is the issue here

lets use some abstract reasoning please, Mod team

[mossy bronze]

Situation being recorded. Keep things codex related please. Kindly

[kind jay]

It’s something else, I will discuss with mods thank you

[mental tide]

Feel free to get in touch anytime by messaging @rugged sapphire

[undone patio]

"Use the appropriate channels for reporting or we will simply allow the chaos to persist."

[mossy bronze]

Wrong. a question had been asked and answered. do not frame people like that based on personal hate.

[white furnace]

is anyones codex also having stream disconnect issue 🤔

[mossy bronze]

Happens! Try reprompting in a few minutes.

[nocturne folio]

should i get the 20x pro acc

[undone patio]

probably

i blew through my 20x in two days

full throttle the entire time tho

[kind jay]

How?

[undone patio]

heavy automation

[kind jay]

5hr limit is like 20% of weekly

[undone patio]

i work like 14 hrs a day minimum

[kind jay]

You do you ig

[undone patio]

what else am i gonna do

[mossy bronze]

Holy cowaiolli! That doesn't sound right! Is it really?

[undone patio]

touch grass?

[kind jay]

Check docs, don’t ask me

[kind jay]

Yeah, and eat and sleep and exercise

[undone patio]

i do need exercise

i did 30mins in my back yard walking yesterday

[mossy bronze]

Very nice! Wowzer.

This is true, there is a greatpe difference between Aeorobic and anaerobic excercise. Codex knows EVERYTHING about this !

[torpid trout]

Has happened often today and yesterday
Probably a bit overloaded

[mossy bronze]

Yes . This is probably the right case.

What a blessing to have the smartest EVERY PHD HOLDER physicist, Scientist, AND ENGINEER in my pocket. For just 20$ a month. Codex is making big moves. Huge leaps especially with GPT 5.5

Not me sir🫡

[sweet moat]

#safety-n-help shows how to use modmail as well as the app reporting option through the Discord's ...

[undone patio]

take five minutes to read through the channel man

[sweet moat]

If you have questions about that, please do ask.

[mossy bronze]

Yes. Thats right!

[mossy bronze]

Thank you for yourr dedicated support!

[frail meadow]

mods getting active in the chat tn

[sweet moat]

Hi. I'm a volunteer community member Guide here, #server-staff . As such, I help all members with questions I spot that are related to the rules, how to use the various Discord functions that I am comfortable with and can share, I assist with finding documents such as https://help.openai.com/en/articles/6614161-how-can-i-contact-support? and help with how to use #1184149552639516732 . I assist with ideas how to use some of the models when I think I can help with questions.

I'm not a moderator, I'm not involved in deciding if rules are broken or what to do about that if they are, except as any community member may and hopefully does, I report #server-rules that I see being possibly broken.

[mental tide]

We recommended sending the mod team a report rather than engaging directly at this point, we'll address off-topic content on our end, thanks!

This goes generally as well.

[mossy bronze]

Moderators do their best YOu are NOT paying them. They provide us with the possility to have a clean and law abiding society within this discord. Lets cherish that.

Friend, you need to be happy

If there is a reason to be happy, the happiness will fade away with the reason.

If your hapiness has no reason, you will be happy for a lifetime.

Codex can approve.

[frail meadow]

what's your favorite comment in this channel so far? jw

[small violet]

my codex limit reset

[mossy bronze]

Nice!

Enjoy codex!

[small violet]

thanks

im using 5.5 low to make it last longer

[mossy bronze]

The one where I helped someone set up skills to automate physics instead of needing to hire a physicist😄

[mossy bronze]

I am happy my tips helped 🙂 Reviewing the plan urself before implementing helps a lot aswell. Catch mistakes early before wasting a whole lot more tokens

[velvet wren]

if you're not happy with the way someone behaves and it's against the #server-rules then send a DM to @rugged sapphire

[small violet]

nvm im back to medium

[frail meadow]

how'd you get your limit reset?

[mossy bronze]

Yes, very important to check what model you use

I will be taking a break to eat for 2 hours. Everyone good luck Codexxing!

[small violet]

5 hours passed and the limit reset?

i exhausted it like an hour

before my reset was due

[frail meadow]

oh i thought you hit your weekly limit lol

[small violet]

nah

im at 44 for weekly

resets on the 30th tho

codex on mobile wen

[gentle harbor]

its not that bad at ui

[mossy bronze]

Very Good!

[nocturne folio]

just buy more accounts

[gentle harbor]

@rocky fog do you think its bad ?

[potent mason]

Why is the glass changing the stripe colors?

[gentle harbor]

its not

[potent mason]

Purple is converting to pink

[gentle harbor]

its water not glass

[nocturne folio]

dude we wouldve been eating so well if openai bought windsurf

i still use it time to time

its harness is so good

maybe its just the cognition team improving it

[tall zodiac]

Do you guys use image gen in GPT web or codex?

[gentle harbor]

api

[tall zodiac]

I feel like I get better quality mocks and visuals in GPT web than in codex

[tall zodiac]

as in ChatGPT.com or standalone CLI?

[gentle harbor]

CLI, has more options and better resolution and better images in general

[errant wharf]

i have a problem with the plugin on vs code. when i tried to use it, it said my account dosint have access to gpt 5.5, but when i try to switch it, as soon as i move over to the thing that says either gpt 5.5 and gpt 5.4 the menu just dissapears and im unable to change the gpt model. (yes i have tried reinstalling and deleting all of the codex data)

[fringe shuttle]

is image gen broken for anyone else in the codex app?

[tall zodiac]

Will try that

[frail meadow]

Are there any useful up-to-date resources for harness performance?

[tall zodiac]

bro image gen in codex is broken

[fringe shuttle]

yup

[tender stream]

Has anyone tried the new alpha yet?

[twilit bluff]

alpha?

[noble jay]

does codex have anything like claude ultraplan ?

[plush harbor]

oh dear. I was muttering about "I wish I could tell codex to do X" and now I seem to have vibe coded a codex plugin

[thorny cloud]

Chatgpt plus plan uses up so fast 😭

[undone patio]

24 more hours until my rate limit resets...

[noble jay]

hope you have durable memory setup with beads or github issues

[sand shuttle]

has anyone tried codex on github copilot?

horrible performance, the output is horrendous and it cant fix simple bugs

[cedar skiff]

harness matters

It's probably unlikely many ppl harness gpt models as good as open ai

[sand shuttle]

i see but if i compare th eout io get form the codex models on copilto vs the app itself theres a big difference

the copilto model sseem inferior

[lime inlet]

Why does it say fast in purple in the codex (cli) startup screen, what does this signify? The model I have selected is 5.5. medium.

[cedar bear]

huh

[cedar bear]

ig the fast mode with /fast

[undone patio]

dont use fast bro

thats how i ate up all my tokens

[lime inlet]

how can I switch fast off?

[undone patio]

/model

[lime inlet]

I don't seem to have it!
Select Reasoning Level for gpt-5.5

1. Low Fast responses with lighter reasoning › 2. Medium (current) Balances speed and reasoning depth for everyday tasks 3. High Greater reasoning depth for complex problems 4. Extra high (default) Extra high reasoning depth for complex problems

[undone patio]

hmm

check the cli docs

[bright swift]

its /fast

to toggle

[undone patio]

there you go

i dont typically use CLI

[lime inlet]

Awesome thanks

I typically use it

[undone patio]

its the same thing more or less but codex app is just nice

[obsidian prairie]

does anyone know where can i see "one-time credit boost" that openai gave us? i just got mail and i cant see this boost anywhere

[fallow steppe]

bro im using codex desktop and its just crashes mid-conversation

[quiet mortar]

I ran desktop app with 50 agents yesterday, no crash

laptop fans were full blast though since a bunch of them were using playwright & computer use plugin, all at the same time

a couple tips fwiw:

codex config limits agents to 6. Ask codex to bump that number up

Codex prefers using chat-gpt mini etc for agents. Ask it to use ChatGPT 5.5 for agents.

Codex prefers agents to use medium reasoning. Tell it to use xhigh for agents.

if using codex to generate batch images, ask it to spawn agents to do them, one agent per image - so you can generate them in parallel

rather than waiting one at a time

[signal tapir]

That's a heckload of compute!

[quiet mortar]

I have ChatGPT pro… using fast mode, I still only hit 50% usage after this stuff ran for 2.5hrs

In saying that, codex tends to use agents as a one time thing - 1 prompt & answer. You got to really encourage it to loop / re-prompt

Oh and tell it to NOT fork conversation history when spawning agents.

Otherwise agents get the full conversation history (as far as I can tell), which kind of defeats the purpose a bit. I want a fresh context window

[obsidian prairie]

anyone got this?

[worthy furnace]

Nah what's that?

[obsidian prairie]

they sent mail about changes and mentioning credit boost, but i cant see anything about that on codex dash or plugin 😅

unless they talkin about credit consumption boost, i can see that

[placid lantern]

i need serious help here

it keeps disconnecting

[kind jay]

Nothing you can do

[placid lantern]

why

[kind jay]

Not your issue, OpenAI’s issue

Just have to wait

[fallow steppe]

just wait for 5/5

[exotic cave]

You need to breathe and count to 10...

[placid lantern]

💀

[kind jay]

Not helpful

[kind jay]

@frosty zealot
https://x.com/animalpopular9/status/2048386677740311000/video/1

[mint leaf]

Yo guys is codex comparable with claude code? Like using agents and skills. And are the limits from the plus abo enough to do anything? Cause i mostly only see tutorials for claude code and very few for codex

[dense path]

better

[cobalt junco]

do u think they'll go for 5.6 next or a jump to 6. ?

[kind jay]

Codex limits are far better than CC limits (comparing $20 plan)

[cedar skiff]

anthropic is removing claude code from the $20 plan

[dense path]

bro 5.5 is out since like a day chill

[cobalt junco]

codex is better v claude-code but comparing web interface claude is better but dumber than gpt 5.5.

[kind jay]

Crazy

[cobalt junco]

i saw that

it affects like 2% of their users i think

[kind jay]

They don’t even have much limits at all

[cedar skiff]

they are also counting the ppl who dont code clearly

[cobalt junco]

im interested if they'll keep doing these incremental upgrades

[cobalt junco]

yh

[dense path]

I could imagine they going to 6.0, because 5.5 is completely newly trained, last time they did that, they moved up a full version too

[mint leaf]

Thanks!
Can I also start creating agents. Like I have a whatsapp or discord group and I just send texts or voice message and they start splitting the tasks and give me updates? I mean I prolly need more Tokens for this, but is this doable with codex or should I use something else?

For better understanding:
I want to create a ...

• financial advisor who can look through my documents
• marketing expert to research trends and analyse my content
• ...
  And create with them a parlament, where they can vote and prolly get a better result for important decisions.

[cobalt junco]

its doable in the codex app

[mint leaf]

Ohhh nicee, are there tutorials you can recommend? Cause i mostly see them for claude code

[proper abyss]

Codex can reply on my Mail like claude ai ?

[cobalt junco]

yes theres plugins for it

just look at the codex app in the plugins tab and it'll answer most of ur questions

[proper abyss]

I'm using codex command line codex cli

[proper abyss]

I'm using codex command line codex cli

[torpid trout]

It still will answer most questions.
But I’m not sure the CLI has the same level of computer use integrated as the app has.
If not, you can build it with codex CLI can build it for you.

Just a fair warning - i would never let an ai answer my mails.
Draft them, ok!

[kind jay]

I would suggest you’re right, poly market has it priced out soon also

[proper abyss]

What you mean. You taking about mail reply

[kind jay]

Yes, but don’t

Show my manager that I’m replaceable, speedrun

[proper abyss]

Yes but don't what you mean?

[proper abyss]

How I will do

[kind jay]

You can certainly configure this, quite possibly codex could give itself capability if you give it the access, but I don’t think it’s a reasonable thing to do

[proper abyss]

I mean you are not sure

[kind jay]

I don't say certainly if I'm not sure

[cedar skiff]

codex app has a skill for it

it works great

cli probably has the same skill

[plucky halo]

Obligatory 'unless it's to automate sending receipt emails (and even then there's an inherent risk), giving access to an LLM to reply to emails is a dangerously reckless thing to do'

[exotic cave]

Many are invested heavily, this tool became something akin to a calculator or a pen, with far reaching impact on how people work, so any upgrade in capability is a real deal.

[cedar skiff]

its read only

[plucky halo]

Ah sweet. Crack on then 😛

[exotic cave]

I wager they might go with another 5.x version in the next month or 2...

[cedar skiff]

I hope not we keep getting hit with higher cost every time they release

[plucky halo]

I genuinely think it's insane that people were whipped up into a frenzy getting excited about this release then, in less than a week, the community is talking about the next model.

[exotic cave]

That shouldn't stop progress. Even pc's back in the day costed much until things caught up. The only way is forward and as fast as we can...

[cobalt junco]

its only natural

[exotic cave]

The ramifications of these tools are so far reaching that always looking toward the next jump is not an insane one. Each capability gain compounds into how millions of people work.

Well as long as us peasents will be allowed to use this tech that is xD

[plucky halo]

I'm just content with what I've got. I'd rather not waste my time wishing for what might be and instead enjoy what I've got

The next iteration will never be what the community hypes it into

[exotic cave]

But you did waste your time ranting about people hype?

[plucky halo]

What?

[exotic cave]

Read your own message xD

[plucky halo]

Ah, so it's an echo chamber you want

The world needs differing opinions. You might disagree with what I've written in the same way I disagree with what I've already mentioned. Different strokes for different folks.

I don't consider it time wasted. It's sharing my thought on it. Is my thought more valid than yours or anyone elses? Absolutely not. It's important to share though.

[exotic cave]

Not echo chamber, you called people excitement "insane", i wasn't silencing you it's my differing opinion... Again i belive this tool warrent so, because it's far reaching capabilities are practicaly unheared of and were SciFi a decade ago.

[cedar skiff]

This is quite the price jump, if it carries on like this in 6 months the usage will be worse than claude code
gpt-5.5 $5.00 $0.50 $30.00
gpt-5.3-codex $1.75 $0.175 $14.00

[proper abyss]

i am getting unavialable install button

[exotic cave]

What's the price of intelligence though? Also 5.5 vs 5.3 isn't a fair comaprison, in my work 5.5 destroys it.

[cedar skiff]

Sure, but you dont need 5.5 full blast for 90 percent of the work, codex 5.3 does it just fine

[exotic cave]

You save tokens by using 5.3, I save time by using 5.5. guess which one's actually expensive.

[cedar skiff]

doesnt work like that though. Your argument is just silly.

I can run 5.3 under orchestration all night

In a month with the cost of codex 5.3 i can get much more work done. strange to think otherwise.

[nocturne folio]

i think openai increased the ratelimits

[kind jay]

Plugins > Gmail > Add to Codex

[exotic cave]

This is moot, I told you for my work 5.5 blows 5.3 out of the water. If 5.3 is cheaper for you then keep using it, but then why would you want to slow down models? You can be happy with 5.3 as you just said.

[torpid trout]

@burnt sage you are at it again?
I asked you yesterday already to either publicly state your intentions, or leave the FR/DM stuff
Last warning, after that its modmail.

[cedar skiff]

I dont want to slow down models - i was making a subtle statement about the pricing hikes in a joking manner. And you took it like a personal attack

[proper abyss]

button not unable

[kind jay]

Okay, I'm not tech support, you can re install, or contact support

[proper abyss]

send

[rain carbon]

guys for hte love of god is compaction completely broken for anyeone else?

[nocturne folio]

diane

[torpid trout]

Seems to work here, but then, "broken" can mean many things

[rain carbon]

no like it literally just never compacts anymore

in macos

the thread becomes dead after context is filled

even compacting on a half filled context window fails

[torpid trout]

I dont have this issue on mac, at least not in CLI

It compacts on its own (anyway you should rarely let it come to that moment anyway)

[rain carbon]

how come

the whole point is that it's good for long running projects

rn my workaround is just let it give a handoff summary before moving on to next one

but surely autocompaction gives an even better "handoff" summary within the same thread

[torpid trout]

Well, its just that once it reaches compaction, it loses context no matter how good that compaction is, and, until you get there, context gets heavier, so, it loses focus

Anyway, it works here on CLI

It's astonishing how people can hang on over the appendix, when the entire freaking preceeding main dissertation is like the main point...

[forest crypt]

I think I am having the same issue. The thread doesn't seem to be compacting well for even a couple of tasks needs another compaction. This is not normal for my type of work

[torpid trout]

How doe this manifest exactly
Like, is it just going to negative context left? or stops responding?

[rain carbon]

fair enough, i might actually start having a "handoff" summary that keeps getting updated then, sounds like a nice workflow anyway

[forest crypt]

Maybe then my issue is different. What I see is that compaction is not really clearning enough of the context. So it reaches maximum again easily, requiring another compaction

[torpid trout]

What does "clearning" mean?
Reaching the maximum again is not abnormal. Note that compaction does NOT delete user messages. So, maybe it is possible your user-messages are unusually large?

[forest crypt]

I meant compacting the context of the window

[hard drum]

They meant clearing

[rain carbon]

oh yah iv'e had that too

do you mean like when it compacts and still ends up with barely much context?

[forest crypt]

yes.... it used to compact to a much smaller number of tokens. So I was able until today to work for longer stretches of time without another automatic compaction

[rain carbon]

I think this might actually be a use case where you do wanna just have a handoff.md and move to a new thread

[forest crypt]

umm... let me check that

[torpid trout]

 Reconnecting... 3/5 (1m 04s • esc to interrupt)
  └ Stream disconnected before completion: websocket closed by server before response.completed

Time for memes

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

[forest crypt]

I get that from time to time.... Sometimes it just doesn't work until I close and reopen codex app

[torpid trout]

that's cli, but I assume its similar cause

on 5.3 and 5.4 I had to use x-high all the time
On 5.5 I did not even yet once use it lol
High was the max, and medium seems to do just as fine
I am scared of what x-high will do!

[forest crypt]

As far as compacting and window context size I figured out the problem with Codex helps (great QA assistant btw). It turms out I have a TODO.md file that keeps codex inline with the work to do for a long project. Somehow it got into the habit of writing intermediate notes on that file on the work done. That file became too large. So it consumes to many tokens of the tokens to keep it in "mind". Cleaning that file should fix it

[torpid trout]

Aaah, the famouse "update yourself" thingy? I have seen it recommended so many times and was wondering... "how about the garbage that will creep in, since it cant clean up after itself"

Imagine, you where basically using your CW with the agents file 😅

[forest crypt]

I have an AGENTS.md for general guidelines. But for large complex projects and using multiple threads I spread notes into plans_ analysis_ and for tasks executions a general desing guidelin roadmap_ and the TODO check list. It has worked greeat except yesterday a new thread decided to report on the TODO what it was accomplishing while doing the work, as oppose to update with the TODO with a plain in-progress/done/pending

sorry typing in the total dark...

[small violet]

codex double limit week when

[torpid trout]

hold your horses a while longer, I think there is still some policing here - probably better to hold off the "wen dis hapen" argumentation a few more days lol

[sick axle]

How are people going through their budget so quickly? Am I not using it enough? Is there some just fire it and let it run 24/7 mode I don't know about?

[cobalt junco]

how did a phone company achieve this? https://artificialanalysis.ai/models/mimo-v2-5-pro

[deep aspen]

how do I get Codex to auto-approve / bypass these permissions? I'm already using "codex -a never" and I have approval_policy = "never" in my config.toml already

[boreal holly]

Super annoying, I know! They changed MCP elicitation to default "untrusted", even for danger-full-access/--yolo. Here's the fix:

[mcp_servers.openaiDeveloperDocs]
url = "https://developers.openai.com/mcp"
default_tools_approval_mode = "approve" # this sucka

[small violet]

bro how do i fix this

[deep aspen]

TY! can I put that in the root above everything else?

[boreal holly]

No, you gotta add that line to each and every MCP server

[hard drum]

They’re not really a phone company. They make appliances, TVs, && so on. They’re the Chinese equivalent to Samsung.

If Yamaha, a music synthesiser company, can make motorcycles, then Xiaomi can make whatever they desire.

[small violet]

im abt to cry bro how do i fix this

[hard drum]

Downgrade version.

[boreal holly]

You know what works even better? Zero files, zero upkeep, zero special tools like deepwiki or anything line that. You copy the entire plan and paste it as a user message. Any time they finish some part of the plan, rather than say "continue", you say precisely what they finished just now then say "continue". Every user message you send stays in memory verbatim across compactions, up to like 45-60k tokens. You can in theory have an agent with perfect recollection of a really long and arduous task, but it only works if you reinsert the entire plan and their progress back into the convo.

I started doing this, and now I don't have doc drift or anything, no old instructions causing legacy things to get resurrected, no wasting time and tokens editing and reading files.

[boreal holly]

Try these:

codex_git_commit = false        # Managed manually
undo = false                    # Conflicts with version control

[dense path]

Use git manually is the best option

[small violet]

when i do that wont it revert every change made in thesession that i didnt commit

i made some pretty importantchanges but it got ruined so i want to undo that

[forest crypt]

Ummm I need to chew on this... It's a very radical step. There is no reason why the LLM models and Codex couldn't figure out this problem

[signal tapir]

Before continuing, I suggest learning how to use git properly.

It's really going to serve you in the future.

[small violet]

i know how to use git

[boreal holly]

You should probably utilize version control. Make it so when an agent does work, they commit their work. I recommend utilizing rules files, so you can keep workspace_write enabled while giving them additive-only git access

# ~/.codex/rules/10-good-git-commands.rule
prefix_rule(
    pattern = ["git", ["add", "commit"]],
    decision = "allow",
)

[small violet]

bro what do i do now

i dont want suggestions for the future

[signal tapir]

My appologies.

[small violet]

ive done alot during the session and i didnt commit and it got ruined at the end

and i want to undo that end part but it says that error thing

[signal tapir]

I've not followed the conversation, but if undo isn't working, try asking it to revert to the previous state.

[boreal holly]

Well, that's because Codex isn't aware of the harness they're running in 😄 so most of the time the agents suggest documenting things in files in the codebase. I'm not saying get rid of all AGENTS.md completely. I've stripped mine down to some really basic stuff like "don't run code formatters, use flutter test -d flutter-tester (don't run the tests plainly), write all DTOs and shared logic in crates", but for the actual work, feeding that context as user messages you're taking advantage of how OpenAI designed remote compaction, and how it's used in Codex

[boreal holly]

Oh yeah, disable code formatters. Make them permanently disabled.

What happens is they apply patches, then run a code formatter, and when you try to undo the changes the code formatter rearranged all the lines and crap so the reverse patch fails.

As far as what to do right now, just ask them to manually undo it. Then disable code formatters and use git

If you have disabled code formatters, use git, the problem keeps happening and you're on Windows then you should unplug all USB devices and restart Windows.

[torpid trout]

Has anyone here any clue why Visual Code GPT chat is not available when remote SSH connecting ... but copilot is?
I see no reason for the GPT not being available on an SSH'd codebase?!

[boreal holly]

You might like Zed 😄

[torpid trout]

https://zed.dev?
And it has gpt and copilot?

Kind of hate switching editor again after 5 years of my favourite (sublime text) and another 5 years of my forced choice (vc) 🙁

[boreal holly]

It integrates with Codex and Copilot, yes!

[deep aspen]

I tryyyy to come back to Codex (normally using Claude) and it's just one thing after another 🙄

[boreal holly]

Cancel, then ask the agent "What is your current sandbox, network, and writable_roots settings?" paste results (snip info you don't want us to see, but specify which writable root is the repo)

[torpid trout]

but also on SSH ?
That is basically what seems broken in VC
Sorry, I could just try, I know.

[boreal holly]

Zed does an unbelievably good job with remote access. Like every single feature that works locally will work on the remote machine

[torpid trout]

OK, let me try this out I guess.

Ish. What a nightmare. years of perfection... 😐

I really dislike switching work tools lol. Such an animal of habit.

[boreal holly]

If your project has a .vscode with all the configs in it, you can have codex recommend Zed plugins to install, followed by "take my vscode configs and mirror them in zed with respect to the plugins". Maybe even have spark do it haha so it takes like 5 mins

My favorite is "Hey Codex, make me some .zed/tasks.json tasks for these things:"

[kind jay]

Animal is right

[boreal holly]

rawr 🦖

[lost drum]

Who decided to switch their enviroment from VS codex extension, to their own enviroment and why/what those it improve for You?

oh shoot 8% till 28th

[boreal holly]

• Access from phone
• Granular control of sandbox & isolation
• Different roles (orchestrator, worker, operator, QA, designer, hidden) with separate base instructions
• Caching thread history for improved performance
• Peer agent swarms with communication affordances and rules

[lost drum]

That seems like ultra hard thing to implement. hmmm I might check of it can levereage my current system but idk

for a total Vibe coder it could be hard to manage such own enviroment and update it propperly when updates come right?

[lost drum]

thanks for the ideas

[boreal holly]

For example

[boreal holly]

The hardest part is building the whole thing. Once it's built, I just don't update unless there's a good reason. If I do update it takes ~1/2 a day to reconfigure it

[lost drum]

I see so you can basically not only send messages via phone. And also You have better for intution file access like the second brain or stuff like that. Hmmm I will need to come up with simmiolar thing but in my case whole visible menu of nodes with prompts so I can see eveyrthing how it is sonnected to each other and stuff like that.

probbably will cehck if its possible within vs code codex and some localhost thingie so I ownt need to set it up for 2 days.

the phone fearue I will ignroe too fo rmy case cause when I am away I just write prmotps in notepad that I later chief and then send back to codex.

cool thing with the phone tho

[boreal holly]

"Second brain" is a hidden agent. It's my own personal todo list agent, so I'll tell it things I need to do and ask it to recall them. Or just kinda look at the current state of the codebase and come up with ideas. It's the alternative approach to managing doc files in the codebase 🤪

The phone thing is nice. On my day off I can keep em busy with chores without sitting at the computer

[lost drum]

why it went form 16# to 22% after you asked him how he is? lol

[boreal holly]

That's because all the reasoning tokens fell off the chat history. It reasoned heavily in the previous turn

[lost drum]

I implemented it via weird folder stucture its a mess😭

[boreal holly]

Yeah, I used to have a docs folder in my codebases with ADRs, plans, etc. There was hundreds of markdown files, and if I decided to change something architectural later the docs would drift. So I made a git hook on commit that has an agent look at the changes and realign the docs. But then some agents would plant bad code, and the docs would get updated to include anti-patterns, and it became such a ridiculous effort I just deleted all markdown files and keep it in-memory. Have had zero drift problems ever since'

[kind charm]

Someone having this problem?
Codex ran out of room in the model's context window. Start a new thread or clear earlier history before retrying.

[boreal holly]

Best recovery path is fork the convo from the previous user message and fill them in on what they might've missed

Fork, manually compact, briefly update em. If it happens frequently I would look at the tool calls, see if there are any commands that output an insane amount of tokens and try to enforce --quiet flags or have the agents use rtk

[lost drum]

I see I see I might rethink this whole porject I am building to make it an actual brain that can pattern recognition soruces and use them instead of drifting with each improvement

I already did that but the system I do advances each hour by 2x so its a mess to make him keep it ub especially that the repo is huge huge

for this whole week he idk how much lines of code he emiited idk how to check it but each day 1000 commits

[boreal holly]

Yes 🙌 your mileage may vary, as most folks would point out, but speaking from experience I've had way fewer problems with drift. The idea is let the code itself be the docs. Let the unit/integration tests describe how the app needs to work, and let them read the tests. That way you're not maintaining a whole bunch of sources of truth

The tests are literally how the app works (or doesn't work) and agents understand straight up code just as well as plainly written markdown files

[lost drum]

ye I need to fix it cause it wokred before I mean I am at the final most hardest polishing stages where it needs to scan whole repo and sources and its a pretty boring work cause previously for 3 days I was just sneding him continuation pormtps to follow the plan but now teh plan changed cause he made so many discoveries that my plan was not using the levereages that been discovered

I wonder when I will end up with the final porduct

but I am happy I used all the credits for this week

7% left

[chrome raven]

first time seeing this " Codex ran out of room in the model's context window. Start a new thread or clear earlier history before retrying.
"

[kind charm]

Looks like there's a bug going on or is not compacting automatically.

[boreal holly]

Do you two have memories or whatever that new memory thing is called Chronicle enabled?

[cobalt junco]

codex app needs perf optimisations

its unreasonably heavy

[torpid trout]

Yeah I guess zed has to wait

One important constraint: Zed cannot be made exactly identical to VS Code because several VS Code extensions do not have direct Zed equivalents, especially PHP DocBlocker, PHPSAB, SonarLint, GitLens, Continue, Material Icon Theme, and some markdown preview extensions. I can still make Zed match as closely as Zed supports, and document every non-portable item clearly.

I could ignore some of them but docblocker, sab, sonarlint are a must 🙁

[hollow holly]

One of the reasons "proper caching/culling strats" has been a priority for me in SwiftASB. Hard to get right, and it's horrible for UX, mem use, and latency when you get it wrong

[hard drum]

when the limits hit, work stops heh... still patiently waiting

[cyan gyro]

I made the mistake to enable fast mode when 5.5 was just announced. Was like a cookiemonster. Without fast mode it's way more reasonable.

[boreal holly]

Bummer!

I see that Docblocker, there is quite literally nothing like it. The other ones you might be able to reconfigure the PHP LSP and stitch together the gaps with sab and sonarlint, but Docblocker is definitely one of a kind

[exotic cave]

5.5 fast mode is 2.5x drainage. 5.4 /fast is 2x.

[cyan gyro]

bug or as intended?

[exotic cave]

Intended 5.5 cost more.

[lost drum]

7% more and I am in dry club too

[hollow holly]

yeah, iirc the copy for 5.5 said "uses less tokens for tasks (and/or tool calls?), but uses more credits". so then fast mode being 2x credit usage for 1.5x speed ends up being... a wash, or even worse than 5.4, depending on specifics

[exotic cave]

fast mode for 5.5 is 2.5x the credit usage.

[hollow holly]

rip

ah, they did change the copy in the GUI app for fast mode to be more vague... makes sense

I get that the inference is expensive af, but like...

stop hiring juniors for years at a time

layoff tons of people for a few years straight

make tools that speed up seniors if they build decent processes around it

price out anybody without an employer paying for their usage, so even reskilling to hit the new bar for speed is impossible from the outside

we're a couple years out from having nobody left to fill roles in this industry

[torpid trout]

I am already setting up Sonar now with their MCP in Multipass > Docker, well... gpt does, I drool

The PHP CS task it added, will check how they work later
I can renounce to dockblocker in age of AI I guess

One bitter pill... zed already crashed once 😄

[undone patio]

just us ML engineers will be left

and then we'll get rekt also

but, i suspect there will always be a platform for creation

like uhhh

Steam for software, if you will

an ecoystem, maybe Codex will do it first

and you'll earn like that

on top of your UBI or whatever

[torpid trout]

yes, meme creation will never die

[hard drum]

wiat til you realise sparta remixes are still made in big 26

that thing is older than many people in new gen

[boreal holly]

The industry created this really cool scenario where junior engineers don't get hired, they get frustrated, but they now have the tools and and can afford to work even faster than big tech companies who don't use those tools. We're in the "guns were invented, physical strength optional" era of tech

[hollow holly]

yep... it sucks, 'cause there's so much potential here in so many industries, but... who's gonna be around to do the work of building foundations for widespread adoption of any of this

[frail meadow]

startup world is finna blow up because juniors can't get in

hopefully CS schools keep teaching security fundamentals lol

[hollow holly]

I see your point, but I also see massive skills gaps everywhere that we have few solutions in place for filling.

And while "can afford to" may be true while codex is in the $20 plan, I think it's becoming clear that's not gonna be the case long-term.

For those dedicated enough to coding because they just like it, maybe they'll stick around long enough to get some gig lined up, but as somebody who first shipped to the app store in 2012 writing Objective-C by hand like a caveman, even I'm about to go bail for retail work or smth until the talent pipeline dries up again

[boreal holly]

They do. Towards the end we did a lotta computer networking which included security, and raw TCP UDP handling

[exotic cave]

please check your facts...
https://openai.com/index/introducing-gpt-5-5/

[hollow holly]

hey, they change the numbers often enough, nobody should be expected to keep up with it... especially when they hide behind vague UI copy changes, as well

aaaand now it's willful ignorance, feel free to bash 'em lmao

heya

what's up?

liiiit, add some music and we're so good

[exotic cave]

I can't beat that level of ignorance, not much experience at being ignorant xD

[boreal holly]

Oh yeah for sure, those of us who actually enjoy coding, not just with AI but in general, AI happens to be a killer tool that's fun to use and actually useful.

I think of it as like this. If you wanna build a car, you yourself could build it probably within a single lifetime if you manage your time and finances appropriately (and don't give up, especially if your day job is an experienced mechanic). Before AI, if you wanted to build an app all by yourself with the same features as let's say photoshop, you might be able to spend your entire life building it and get close enough, but you're always competing with the newer photoshop. With AI, you could within a year put together something and catch up. There is no similar process for building a car. AI is empowering the individual to do things in tech that can't be done in any other industry!

[hollow holly]

Yeah, honestly that's what I love most about this. In my case, I've got some nervous system damage that makes typing/reading difficult at times, which for years kept me out of the field I loved. So, after building enough infra around it, Codex has been a godsend for me.

i want nothing more than to bring that to other people and other industries, as well, hence SwiftASB, but seeing that through is easier said than done from my present situation.

I just hate that something this gamechanging has come at a time where I can't get back into the industry fast enough, and as a result, won't be able to afford to use it much longer. it sucks, but i got some cool stuff built in the meantime. so there's that, i guess

[main nimbus]

Great example.

[frail meadow]

I would love to hear your setup!

Also if you're trying to cut your costs down I heard that caveman and using vector indexing for semantic search of the codebase decreases token usage significantly

[hollow holly]

Sure! The centerpiece is a plugin of Apple platform/Xcode/Swift/Dash docsets skills that took way too much trial and error to get right. But, at this point I have Codex writing pretty decent, concurrency-safe Swift, and pulling docs and errors straight from Xcode.

Then a nice on-device TTS setup with Qwen3 over localhost HTTP, and MCP, until OpenAI added guardrails that decided my TTS MCP calls were the model lying about unperformed work. Probably because they were prior to the assistant's final message being printed, or smth. So, now I have that running off of a Stop hook via a js module that calls the service directly from outside the loop.

And yeah, I've been building a hybrid FTS/semantic search package off of SearchKit and Apple's NatLang framework to run locally. Idea is to eventually have my own auto-updating cache of commonly needed resources that Codex can pull from directly. Similar I guess to the "curated" online search thing OpenAI setup for Codex/Chat recently. We have GRDB in the Swift world, but I figured a simple, all-in-one, "just works™" package might be useful for other macOS/iOS app devs, too

[vast crater]

Hello, anyone else having issues with MCPs not being visible for the models in the Codex App on Windows/WSL2?

The MCPs are working, but for the models in conversations, list_mcp_resourcesand list_mcp_resource_templatesis returning nothing for them

[hollow holly]

It's been a while since I was on Windows, and I think I was using WSL1 then..? But, are the MCPs running in the WSL environment, too? Or are they running in the Windows side of things? And, is the transport stdio or http?

[lost drum]

what MCPs do you all use and why/how it benefits you?

[vast crater]

And also somehow 5.5 tells me that npx and node aren't installed or not in the PATH although they are... Idk what kind of spaghetti there's still remaining in the Codex App for Windows, but I'd be happy to finally see a strict and correct commit to WSL2 agents, either you fully use WSL2 or you fully use Windows Agent, but no in between spaghetthi...

[vast crater]

MCPs are configured in WSL2 and also in the Codex App so on the Windows side

none work somehow

[boreal holly]

It's because the MCP servers you are using have not configured any resources or resource templates. It's just tools. So you make up for it by making a SKILL.md that covers the tools in brief

[hollow holly]

If it's stdio, you probably need to run the MCP from inside the WSL environment for Codex in there to use it. And you'd need node/npx to be reachable in there, too.

try which node or which npx from inside WSL. Then make sure the config.toml in WSL isn't trying to reach across and run an MCP that's in the Windows environment.

If they're http, things are more flexible. this article might help in that case: https://learn.microsoft.com/en-us/windows/wsl/networking

[vast crater]

Yeah the MCP are from inside the WSL env, node/npx are available in WSL2, they work without any issues with the Codex CLI in WSL2. Only the Codex App has issues somehow

[hollow holly]

mainly xcrun mcpbridge to control xcode, pull live issues, pull docs, avoid wrecking my .pbxproj every 2 seconds, etc. and dash.app for more local docs. and my localhost TTS, but I already blathered on about that above :P

[hollow holly]

Oh, I got your problem backwards, lol. So, it's fine in WSL, but the Codex GUI app in the Windows environment can't use stdio MCPs from the WSL environment?

Yeah, you'll probably need node/npx/mcps available separately in the Windows environment, then, to launch them as stdio children there.

[boreal holly]

Look at ~/.codex/skills/openai-docs

If you're saying "The agents can use MCP tools but listing resources is unavailable so they don't know how to use them", this is because the servers themselves are not publishing resources or resource templates.

That skill is provided by OpenAI as part of their docs MCP server - default installed. It shows the fact that OpenAI themselves had to craft a skill that describes the MCP tools and how to use them. Anthropic made MCP, so they have first class support for all of the protocol features. OpenAI built a translation layer that converts MCP into regular function calls like what's available in the API, and so you have to describe the tools using skills.

If you're saying the MCP works in CLI and not the app, then yeah that's a fragmentation issue.

[hollow holly]

Could also try the MCP Inspector and see if that gleans any useful info.
site: https://modelcontextprotocol.io/docs/tools/inspector
gh: https://github.com/modelcontextprotocol/inspector

[torpid trout]

Oh duh, that was a fun adventure
Set up the MCP etc... connected to remove ssh... "please enter openai api key"
LooooL
It allows to connect normally to the codex cli in local projects, not on remote ones, for whatever reason, even if ... the ide is local, as such, it should not be an issue

Guess I will jsut use scp lol, fasir enough "D

[boreal holly]

???

You saying Zed is not letting you use the subscription in remote mode?

[torpid trout]

Yep

You can easily do it in local project, not on remote
On remote it has only 2 options, both of which are key based (so, api key. one is called codex cli api key the other openai api key or something like that)
And on local it allows you to use keychain

At least now I know how to set up a MCP via local Multipass VM > Docker 😁 also a win

this is the thing

[boreal holly]

Huh, it might not support headless Oauth. Have you ssh's into the remote machine, launch codex TUI, and signed in?

[spark bobcat]

Hey guys

Do you recommend any plugin on codex

For management

[torpid trout]

It does, but locally only
I mean, when I opened local it said "may I use keychain" since that is what I have configured in codex
And that worked

[spark bobcat]

I am just using clickup plugin now

[torpid trout]

But, remove that wont cut it it seems

[torpid trout]

no, remote machine has no codex of course

Basically, that is a server, and everythign happens "there" is edited "here"

Also, I mean, this might just be me but... zed cannot actually open files 😄

Let me grab a video of this

[boreal holly]

You know what, I just tried it and I had to install the Codex ACP

[torpid trout]

I think I have that. it mentioned at some point it is using codex-acp etc

Local project/file, it does use codex cli just fine

As soon it gets remote, it asks me to log in when I click the + to add a thread

and there only offers API KEYS

[craggy whale]

It's very buggy, but I got it to show three times one at least. Well, three times one and a half. Yes, this is inspired by the mobile version of Discord.

[boreal holly]

OK so what I did, step by step, to get it to work was this:

ssh robertsale@robertsale._peer.internal

codex
# sign in with subscription

exit

Then in Zed open the agent panel and create a new thread with the Codex CLI external agent

[torpid trout]

right, but that means you have codex on target machine
I dont

Also I would not know why zed would want to run codex on the target machine
The files may live on target, but they are edited on this machine here, via the ssh feature
( I know that factually this means the file is still on target and edited to target, yet, of course my editor is local, and like VS code copilot... I would want the codex integartion to work on that file too, without having to install it remote)

I guess the way ACP works is it expects a cli on the machine the file resides

Which is... well, their decision 😄

I dont think it is really good, because a integration belongs to the UI, not the target server
Imagine I where to ssh into a given client server, and .. install codex there?
😐

I mean, - not sure it makes any sense... I just edit in my editor, and expct the chat to be there acting on whatever context I have, without having to mess with the machine hosting the context.
This is entirely possible in VS Code with Github Copilot, but apparently not with COdex CLI
Perhaps different apprach.

[normal dew]

can somebody tell me which context length is the best for codex to use without hallucinating?

[boreal holly]

I mean, the whole point of remote development is to use the remote machine for 100% of the development. It's pretty wild that VS Code remote doesn't work this way! What you're saying is it SSH's a file system onto the local machine and you use local tooling to develop. This is pretty much a "use mutagen" or "use Perforce" deal, skip the remote aspect entirely. Zed puts 100% of the processing and tooling responsibility on the remote machine

[hollow holly]

it's less about the context length and more about what's actually in the context, and what's not there, but should be...

some more info would be helpful, nobody can give a specific answer to a vague question, tbh

[frail meadow]

Is xcrun better than something like idb from Facebook? I'm not super experienced in working on Swift applications but they're super interesting

[boreal holly]

In fact, now that you mention it, I think all of your issues with VSCode would be solved if you just used mutagen and stayed local

[normal dew]

Well if I am designing an app which is fairly large, I mean I havent run into any problems but I dont know if I am really using the agent to its best performance at 258k tokens. I direct it usually to use subagents to help keep context but still should I set it to like 400k?

or 1m

[wooden minnow]

Hey image 2 ain’t available on codex or codex app?

and can’t login ChatGPT app with api? wha?

[hollow holly]

haven't tried facebook idb yet, but thanks for tuning me into that. i'll have to check it out next time i'm doing iOS stuff.

but yeah, so it looks like idb is mostly for automating work with the iOS simulator. xcode's mcpbridge (from v26.3) is more an interface between the IDE itself and external agents. So, things like pulling and refreshing inline code issues, searching and accessing the built-in docs, building a scheme, grepping through a project, mv/rm/cp/file edits and whatnot, all in a way that doesn't collide with Xcode's management of it's labrynthian internals like the .pbxproj and git locks

Otherwise you're gonna have issues sooner or later with any sizeable project if Xcode is open

[karmic shadow]

I still cant use 5.5 on codex

[boreal holly]

idb is absolutely the best tool. Apple made all their Xcode tooling use a global lock, so if you have more than 1 agent driving different simulators they frequently throw lock errors. idb has no such issue. Also it makes you design the app with Accessibility in mind so it's a double-win

[hollow holly]

So, subagents are great for doing noisy tasks to keep "useless" context from cluttering the window. like docs fetches that might require a number of queries, or builds/tests, that sorta thing. just reporting back the final/important info or answer without all the in-between.

if your project is decently organized, and you prompt based on that organization (calling out a feature by name, that's in a directory with that name) codex can search for the specific thing in the specific place, and that keeps things focused.

a lot of it is the same as always, good docs on tap, and good repo hygiene, tbh.

[normal dew]

Yes but you didnt answear my question for the context lenght? should I keep it at 258k or to which size can I go till it starts hallucinating

[torpid trout]

I disagree that the whole point of remote deployment is use remote machine for 100% of the things. The main point is to host on remote machine, not to develop
Random hosts have random OS and tools installed. Client goes and says "fix" i say "ok" and ssh in. Through transmit app for example, click to edit > opens in local vs code > save to upload. Yes, that factually transforms that file to a local file temporarily
I am not sure how vs code does it for ssh remote connection. I do not think it pulls over the file. I just know copilot lets me interact with that remote context... codex add-on simply disappears in those cases
I see no reason why it should not agt least be able to read it. or send a sed to the remote server.. through the ssh. perfectly possible, I think?
I mean... again, copilot does it with out issues, since you are editing the file anyway in vs code

[hollow holly]

I mean, my point is that reducing hallucinations is the more important problem to solve, and context window is a less impactful thing than keeping the context good to begin with.

as far as context size, generally you want to keep your use of it as small and focused as possible. starting new threads for new units of work. using a thread to plan things out, make decisions, get that into a set of tickets, then new thread -> let's get it, M2T7 in a wt, new thread -> let's get it, M2T8 in a wt, rinse and repeat 10 times, by the time you open the tenth thread, half of it's already done behind you to review

bite-size pieces, each started fresh, in a clean, modular codebase gets you a lot better reliability and parallelization, it's just more effort and sitting around thinking up front about the boring stuff, lol

[boreal holly]

Well, there's another piece of the puzzle that I think you'll agree with. If the agent is operating on your local machine, but the tooling is on the remote machine, and it's using ssh to send commands to the remote, how is sandbox supposed to work?

If codex is installed on the remote and ran over there, it gets the kernel-level sandbox benefits. If it uses ssh only, it only needs network access and it can send arbitrary commands to the remote machine. So it makes sense to install the harness on the remote to gain the entire security benefits.

So when copilot does it without issues, it makes me wonder what copilot is giving up to make it work so easily. Or perhaps copilot is installing itself on the remote machine, and maybe on disconnect it cleans itself up. Idk, but I personally like Zed because I have a 16GB macbook, and it's nice being able to put 100% of the compute effort on a workstation

[torpid trout]

hmmyes right. Wondering how copilot deals with it. I have to play with it more, I only saw it is avaialble, not how it actually goes for it once sandbox bevcomes an issue

IT is definiteluy not installing itself on remote. that feature of vs code only deals with the fioles you have write rights on

I think that might be the trick here > the user you log in with has write rights, it just assumes that the sandbox is ok.
not entirely wrong imo that approach.

[kind jay]

Long ahh messages, I ain’t reading all of that

[torpid trout]

Of course not Jane, no one expects you to 😉

Just dont ignore our messages when we ask for what we subscribed to ok?
😉

[gentle harbor]

why

why ???

20% of my cpu

its not doing anything or running any task for the past 7 hours

[undone patio]

im crying rn trying to squeeze the last tokens out in high mode

[boreal holly]

Fair enough! That is essentially danger-full-access in codex which is fine if other security stuff is in place

[kind jay]

Thanks ig

I’m kindly requesting not to be mentioned in the chat by you, unless it is important

[kind jay]

^^ again, additionally bring requests like this directly to the mod team as instructed

MIT is no longer meta for physics, I use codex now

What’s a skill?

How do you create a skill if you don’t know physics?

[torpid trout]

Dito. No one pleaded your opinion.
Respectfully cease and desist.

[torpid trout]

Jane stop blocking me. Initiate dominant chat.

This is the war channel, no codex pleaded

[hollow holly]

No war? Awwww...
How about a battle, instead?

[torpid trout]

Rule the reads

[kind jay]

I’ve been using codex for my physics project lately

It seems fine in general but struggles to make good choices in terms of simplifications

[torpid trout]

I’ve had my physicist using codex lately
He seemed simplistic in terms of my choices

[kind jay]

Is there any advice in general or it it the meta to work the logic yourself and have codex put it in?

[kind jay]

That’s fair, how does his workflow look like? Does he use skills?

[kind jay]

I like that movie

[torpid trout]

Yeah, he used a look to flow a skill for work. It was beautiful.

[kind jay]

Damn, I wish there was a physics toolbox or skill or something

[torpid trout]

Physics is definitely a tool

🤣

[kind jay]

Like some python env but codex knows how to use it

I think I’ll make it

Or maybe give some general physics skill because I don’t think it’s been done well before

[torpid trout]

Better than any Physics you mean

[kind jay]

Ydkwydk

[torpid trout]

I come bearing news

Nerves are getting poor. Literally.

If I put skills and agents in a local folder, and instruct codex to follwo them, I guess it is still forced to read my user level agents and skills, and probably will create spaghetti?

use case is, a totally different domain where codex is being used
Instead of coding, article writing (basically a newspaper team, from investigative journalists to illustrators etc)

Are you also a Physician?

[kind jay]

I would like to do this after I graduate high school

Physician means doctor, I said I want to do medicine after HS

[torpid trout]

Mutter mutter flutter butter

Do you know that jane? I think that's what it is called. Or fluffernutter?

I am making a plugin that does it

[kind jay]

This but codex

[kind jay]

That sounds peak

[torpid trout]

Peak mellow indeed
Oh boy, I have to clean up my codex again

[glacial ginkgo]

Is the "Codex App for Windows" also open-source and Rust-based or only the "Codex CLI"?

[pulsar rune]

Hi guys, any solution for codex app for windows? i have stuck like this when open the app

[undone patio]

[bold root]

I want to use notify in config.toml, but is there a way to distinguish between the main agent and sub-agents?

[glacial ginkgo]

I had that problem this morning. I ended up having to reboot Windows and that resolved it. 🤷

[pulsar rune]

uhm any other solution bro?

[glacial ginkgo]

Nope. I can't find the source and can't debug it in any way. The Codex CLI is open source and debuggable. Unless the source for the App is out there somewhere, there is no way to begin to fix it.

[undone patio]

rarely

im just a heavy developer

and i was running full juice this cycle

fast mode max thinking effort

(20x sub)

for me? probably. i prefer notebookLM for research

yeah im saying my usecase for chatgpt is mostly asset creation and coding

if something gave me that for cheaper at the 20x token alottment then yes i would be happy

unless chatgpt releases a UI akin to notebookLM

or integrates to that app natively

then i would also use it for research

[bright swift]

so what are we thinking about symphony, anyone tried it out yet?

[undone patio]

thats not healthy man

especially since the clover in your name indicates you just arrived to the server

so its like

did you come here just for that

basically if you're here to shill something it should probably be something created with codex rather than something meant to be a workaround for the credit system they have established by propreitary means

is it?

sounds like user error to me tbh

consider the output a reflection of the operator, not the tools

dude no one has a lot of experience with harness engineering

gtfo

its literally a new thing

[gentle harbor]

@undone patiowhy is this happening ?

i havent used the apps in over 6 hours

they are idle doing nothing

but taking my cpu

[undone patio]

well then your definition of "a lot" is severely lacking

[gentle harbor]

acting like a crypto miner atp

[undone patio]

is this your task manager?

[gentle harbor]

yes

[undone patio]

why do you have so many instances?

[gentle harbor]

codex is using 20% cpu and gpt is using 7%

[undone patio]

i dont see the issue i guess, are you not happy with the usage?

[gentle harbor]

no issue with it using 20% of my cpu for no reason ?

[undone patio]

is it idle?

[gentle harbor]

its idle not doing anything

i dont even have any tokens left

[undone patio]

hmm, i would ask via terminal CLI

to diagnose

while the app is running

[bright swift]

its an electron app what do you expect

[undone patio]

but also yeah i mean theres just overhead for some things

[gentle harbor]

discord is a electron app using 0.4 of my cpu

[undone patio]

im not sure what the reasonable usage to expect is, my M4 24gb handles pretty much whatever i throw at it

[gentle harbor]

on idle shouldnt be using anything 😅 ive never seen this before in any other app

[digital thunder]

<:book_icon:1363314738255364126> Rule 7: No self-promotion, soliciting, or advertising.

-# Do not post or direct message any members of this server to promote non-OpenAI services, products, or projects. Exemptions: OpenAI API projects in #api-projects & GPTs in #custom-gpts.

[bright swift]

so what about those usage limits

reset wen

[undone patio]

pls reset

[digital thunder]

 Some off-topic or otherwise infringing messages have been removed by a Moderator.

[bright swift]

who else got an account strike

[cedar skiff]

longest orchestration yet, with 5.5 as the orchestrator

still going

[exotic cave]

account strike?

[bright swift]

i got muted for circumventing automod lol

anyway

wen reset

[bright swift]

real work or waiting for compiles?

[kind jay]

Oh no

[cedar skiff]

real work, I decided that i could do integration testing to replace manual user testing instead of just using it for the highest value parts. Since we have ai code is much cheaper i thought id give it a go.
It's 5.5 high orchestrating codex 5.3 high to write the tests.

It surfaced at a handful of bugs already, and i am about 30% through

[undone patio]

you should also have code review in the cloud on your PRs for another layer of checks

cursor bugbot is a common one

gitzilla is another

[bright swift]

just let another codex review it and tell it claude wrote it

[undone patio]

lol

ah yes, the moderators doing good work today

🙏

[lost drum]

i had same but 15k -16k XD

[bright swift]

sounds against the rules to me

mods?

[kind jay]

And it just rewrites the whole files

[lost drum]

it was gpt 5.3 tho

[kind jay]

5.5 high

Told it to split the desktop and mobile

[lost drum]

3% left and ggs good thing that it resets tomorrow at 8pm but 3% till then is rough

[bright swift]

i got only peanuts left too but not worth doing 3rd account now

[kind jay]

Mhm, I'm enjoying playing with codex to making something fun

[bright swift]

will milk the 2x usage may though

[lost drum]

wdm may 2x usage?

[bright swift]

2x usage on pro until end of may

[lost drum]

my plan ends on 17th

pro 2x

then I go back to plus

[bright swift]

also may is last month to exploit gh copilot gpt 5.4

[lost drum]

what is that?

and why exploit

[bright swift]

its was just very cheap way to use 5.4

per request billing

[lost drum]

I see

[bright swift]

and you could make one request last hours because subagents didnt count as requests

[lost drum]

well u need to optimize it

[bright swift]

i win

[kind jay]

What are you making?

[bright swift]

some data analysis + ml stuff

[kind jay]

That's pretty cool

[lost drum]

WAT? how many hours

[kind jay]

What ML in particular?

[bright swift]

yes

[kind jay]

Don't bother talk to them

[lost drum]

it it one prompt or whole thing?

[bright swift]

that huge commit was just because i vendored something 😄 my real ones are much smaller

[lost drum]

next goal 1mili

probabbly easier to just earn 1mil$

[kind jay]

Does generated data contribute to the diff?

[bright swift]

1 LOC = 1 $

[bright swift]

nah no data in that

[kind jay]

Say if you saved some file with derived data

[kind jay]

That's crazy then

Why bother have this feature?

[bright swift]

is that the git re-implementation in codex app?

[boreal holly]

your agent is using code formatters which breaks Codex undo button

[kind jay]

I don't have agents

Unless I'm an agent

[cedar skiff]

they are all agents

[boreal holly]

then dont do that and no problem

[kind jay]

We're all agents?

[cedar skiff]

uh um yeah

and any ai wirting code for you

oh god, is this bots or a sad person?

they are multiplying

[lost drum]

were they good?

[cedar skiff]

oh they are just changing names

[lost drum]

not worth 500$

[bright swift]

nobody wants your fake codex subs lil bro

[lost drum]

5k$ for my system that is called agi codex😂

[plush harbor]

Oh hi, you seem sane, welcome to having fun with codex without going over the top

[kind jay]

I need a break from thinking sometimes

It's cool to just say that you're thinking and have it be done

[lost drum]

wdm human supervision? cant you just create auditiors with complexed rules that teach them how to audit?

[plush harbor]

I like palming off the things that are boring or I don’t feel like doing and then tweaking the appalling default css it makes

[bright swift]

i have to admit that i prompt to change font sizes, padding, etc, whenever i work on frontend

[lost drum]

if you think enought and ask codex then he will find a way to create such thing

[kind jay]

The CSS is usually not great, but it's good at copying designs

[plush harbor]

I accidentally made a codex plugin yesterday and have yet to try it out, I got distracted

[lean lark]

accidentally?? 🙂

[plush harbor]

Yesterday I told it to put a different class on all my helper vs action vs destructive buttons and it did but it defaulted them to that pale background, border, dark text it likes. So I had to change them to match the site theme

[plush harbor]

Yes, I wanted to do something and turns out what I wanted to do needed a codex plugin lol

[kind jay]

@lean lark Made JaneBot

[bright swift]

i just give in to the vibe designs that agents create. customers cant tell the difference so why waste the tokens

[kind jay]

This just happened to me

[bright swift]

unless you develop stuff for devs/designers

which is a terrible market

[plush harbor]

Existing website, needed to fit in. They were the only bordered pale buttons on the site

[kind jay]

I asked it to make my page feel tacktile, we'll see what happens

[bright swift]

huh and it didnt align it with the existing style? usually works for me when working on existing stuff

[plush harbor]

I just told it it has no taste and changed it myself. I dunno why it didn’t pick that all my other buttons are dark with light text

[kind jay]

It's so good!

[plush harbor]

I need it to analyse some logs for me soon, once I have some data to analyse. I really don’t get how people make entire apps in one prompt, like this is a marketing dashboard but you can’t write it without data to parse … so doing it in two parts a couple days apart

[bright swift]

you cant make entire apps in one prompt, at least not working ones

[plush harbor]

Seems to be a X/linkedin tech bro thing

[bright swift]

if somebody claims that, they have a course to sell you usually

or X payouts to chase

[lean lark]

Speaking of logs, I just published my first GH Repo - as CaptainStarbuck.
Please be gentle with comments. Thanks!
https://github.com/CaptainStarbuck/codex-usage

[kind jay]

It's possible, but probably with very little control and too much set up initially

[lean lark]

Actually JaneBot does create a complete app in one prompt. That's what she's designed to do. The code is clean and documented and ready for whatever you want to do to enhance it.

I'm a little scared to publish her right now ... I think much of her code is over-kill and am looking at trimming her down. I'm discussing her diet with her now. 🙂

That's not advertising. JaneBot was born from discussions here. Named after @kind jay here. She kinda belongs to this channel.

[kind jay]

My twin ig

[lean lark]

But you are absolutely right in general... Thanks!!

[kind jay]

I can give some review

[outer whale]

is codex good at coding? idk im decent but i do make alot of mistake is codex like a good coding bot or smth

[kind jay]

You knew what I meant, so that's good. English is not my first language

[lost drum]

check

[lean lark]

@mossy bronze I don't know what codemaximum is or does or did, but JaneBot is a nice wrapper around the Codex CLI, performs specific functions for this audience, and is completely MIT FLOSS. there's no product, no model. But again, thanks.

[lost drum]

2x pro plan used fully no subagents only VS codex extension

[kind jay]

Don't waste your energy talking to them lol

[lean lark]

Actually, he's right ... I need to think about this...

[kind jay]

In terms of advertisments, etc.

[lean lark]

As a former Guide here, I'm not sure if I would flag this or not ... I think not because it's not a personal promotion. It's a public web page. Attribution is stating a fact, not active promotion. It kinda can depend on the motive. In my case, I have no motives. I'm not interested in attribution. I shared the project with Venus earlier, just as a gift, not for gain.
No, I'll stick by this one but only after some careful thought.
But thanks again @mossy bronze for diligence in keeping this resource friendly and productive.

Consider, we can refer to any site on the internet as a matter of attribution. If we wrote the page or created the YT video, it almost doesn't matter if the info is useful. For a YT vid, asking for subs or thumbs is absolutely solicitation. In the case of a MIT repo, there's nothing to gain - rep comes from just being here.
(OK, I'm done with the introspection. Now watch a Mod or Guide completely disagree. LOL. )

🦗 🦗 🦗 🦗 🦗 🦗 🦗 🦗

[small violet]

gpt mythos wen

[lean lark]

Wrong company, dude.

[small violet]

nope

ik what i said

[lean lark]

OK, I see what you're asking. But no one here can answer that question. Why not ask at A\?

[small violet]

why would i ask anthropic when open ai is dropping their version of claude mythos

[lean lark]

Hey, sure, ask anything you want, sorry. 🙄

The GPT-5.5 model IS the answer to Mythos model.
The v5.5 label is seriously betraying the nature of what this model is. It's much more than just an evolutionary step up from 5.4. It's signficantly different. 5.5 = Spud. I believe this is what you were looking for.

Oy vey, I don't wanna have to go look for a quote. You don't need to take my word for it. Just have a great day. 🙂

OK, I was compelled. LMGTFY ...

[dusk thorn]

5.5 was internally spud

no one cares users call it spud

[lean lark]

From a YT interview: "Greg Brockman is the president and co-founder of OpenAI. Brockman joins Big Technology to discuss GPT-5.5, also known as Spud, and what it means for OpenAI’s next phase of AI development....."

[small violet]

@mossy bronze do u have a phd in physics

[dusk thorn]

@mossy bronze show us your phd

[small violet]

why are u here bro

[dusk thorn]

larper

lmao

[small violet]

im dming

[dusk thorn]

please do

[small violet]

maybe its the truth

let me dm him

[dusk thorn]

let me know

[small violet]

i will say true or false

i wont disclose

[dusk thorn]

ok

[lean lark]

https://youtu.be/YnoQ8RJbALw?si=I506Vi_ZsjKWGS9V&t=42 Quote confirmation from Greg Brockman.

[small violet]

@mossy bronze i dmed

@dusk thorn False

[dusk thorn]

who knew

😂

[small violet]

he didnt even dm

[lean lark]

Can we focus on Codex?

[small violet]

yeah def

when is open ai going to fix this

its been happenning for like 2 weeks

[nocturne folio]

is there something that like lets me smoothly switch accounts on codex

[lean lark]

Logout/in?

I know that's not elegant. 🙁

[dusk thorn]

your using 2 plus accounts i smell it

[lean lark]

You can script codex logout and codex login.

[nocturne folio]

yeaa

its so annoying to just logout and log in

[lean lark]

Script it into a single command.

[nocturne folio]

i mean its not like stopping mid prompt when i finish my usage

THANK GOD

i just remembered how much i hate anthropic for that

[dusk thorn]

yeah ask codex to make a script to switch it when you type in something like "codex-switch"

and make it run in the background

[nocturne folio]

I usually get to like 0%

[lean lark]

Don't tell me you're trying to swap accounts in the middle of a thread so that you can avoid quota.

[dusk thorn]

anthropic $20 plan is like a demo

[small violet]

fr?

ive only ever used anthropic models via anti gravity

[dusk thorn]

its not very good

and rate limits hit quick

weekly not bad

but the 5 hour

is

crap

[cedar skiff]

did those bots get kicked?

[stable obsidian]

When is codex gonna looksmax claude with ui design

[spark bobcat]

On clickup plugin

On codex

There is no way to use more than 2 workspaces per time?

Or i have to disconnect and connect everytime i need to switch workspaces

[small violet]

compare it to codex

[cedar skiff]

I am interested in what happens with cursor, if that goes through there will be a third player in the coding space

[dusk thorn]

codex 5 hour feels like almost 1.75-2x usage on same plan

[small violet]

lol

are they selling to spacex

?

[cedar skiff]

They only thing grok is missing is a platform for coding

It seems so

hasnt happened yet

[small violet]

why would they sell to space x tho

that dosent make sense

[lean lark]

Insane offer from X to Cursor though.

[cedar skiff]

60 billion reasons

[small violet]

name top 4

[lean lark]

It's not "SpaceX", it's the "X" family under Elon. I don't recall which company.

[small violet]

let me find

lol its space x

[lean lark]

Actually, I'm wrong. It is SpaceX. Go figure.

( Google is your friend. )_

[cedar skiff]

they cant keep up and compete against platforms that openai and anthripic provide, cursor is missing stoa models that arent bought over api and grok is missing a platform.

[cedar skiff]

top 4 what?

[small violet]

grok isnt a sota model either

and composer 2 is just a RL-finetuned of kimi 2.5

[cedar skiff]

I think it is. It's just a matter of time before it reaches users like us.

[cedar skiff]

Youre not telling me anything i dont know

[small violet]

lol

[small violet]

lol

bro no one uses grok

Mechahitl*r find bugs in my codebase

[cedar skiff]

no user like us for sure because it doesnt have a platform...for coding
But millions of ppl use grok everyday.

[lean lark]

I don't understand the acquisition at all: Cursor is a fork of the FOSS VSCode. It's not a frontier model company. It's an IDE. One program. And why would SpaceX buy it when xAI is the more relevant company? And after the disaster purchase of Twitter for 55bn, how could Cursor possibly be worth 60? I'd think it's more like 600 million. Just nuts all around. But that's Elon ... and that's why we're here.

[small violet]

they use grok on twitter only

to ask

grok is this true

[rocky fog]

you guys need a reset 😂

[cedar skiff]

probably buying it for the user base and code data

[lean lark]

Still makes no sense in relation to other points.
What code data? It's just code. User base? Just developers...

[cedar skiff]

space x is owned by xai

[small violet]

wrong

its the opposite

[cedar skiff]

Right but in order to build smarter coding agents you need a never ending stream of code data

[small violet]

i think the main reason is honestly the ipo in june

[cedar skiff]

yeah good one, but the point doesnt change

[small violet]

did i imply it changes?

[cedar skiff]

does it matter?

[lean lark]

SO? GH? SF? Meh, not buying it.

wahtevah, OT for this channel anyway, doesn't apply to Codex.

[small violet]

eric wimp

[cedar skiff]

both of the big players in the coding space have incentive programs that subsidise coding. Pretty sure it's so they can get more data to train on

[nocturne folio]

bruh

[nocturne folio]

its the data

[small violet]

damn

[small violet]

its the ipo

if it was the data xai wouldve aquired it not spacex

[small violet]

damn my weekly is at 26%

[cedar skiff]

both can be true

[small violet]

im aware

but the ipo is the main reason

[nocturne folio]

it just resetted so im lucky

time for more slop

[cedar skiff]

They cant compete without the data, and elon is all in on ai

[nocturne folio]

im gonna produce slop at the speed of light

[small violet]

5.5 x high clean up the slop from 5.4 no mistakes

[small violet]

can i ask what u build with codex?

[boreal holly]

im still at 60% weekly so im stoked

[nocturne folio]

https://autodisc.xyz

[small violet]

if this is some trash bro

[nocturne folio]

and https://github.com/grapeeehq/grapeee

[nocturne folio]

pure trash

its slop

[small violet]

any sales made?

[cedar skiff]

me to, i finally got to the end of a week without a reset and now i'm pumping out some code.
I have to say 5.5 is by far the best orchestrator i have ever used. It just doesnt stop EVER

[nocturne folio]

1 customer paying

[small violet]

damn and i thought mine was slop

[nocturne folio]

7$ cad

[nocturne folio]

bio?

[small violet]

jesus christ

[small violet]

yes

double rate limits wen

codex ui is so beautiful

[small violet]

are u using pro

[boreal holly]

dude… my favorite thing is it always takes care of its unfinished business. Its that Tau2 in action

[cedar skiff]

yeah

[small violet]

what are u building

[cedar skiff]

Yeah, for sure, i have a debugging step to unblock in there as well, 5.4 would always stop at that and tell me about it

[lean lark]

Better question for #chatgpt-discussions ?

[small violet]

just answer

[lean lark]

Or #gpt-models ? Or #gpt-5-5 ?

[small violet]

gpt pro trial wen

@cedar skiff is the highest still x high on codex?

[cedar skiff]

yeah, i havent even used it since 5.3

5.5 high does everything

[small violet]

i remember my first time using codex

[cedar skiff]

and for automation i use 5.5 as the orchestrator and 5.3 as the coder

[real blade]

gpt 5.5 vs 5.3 for token efficiency?

[small violet]

and it one shotted something

gemini on antigravity

struggled to do for a week

i used x high

[lean lark]

Corvette vs VW Bug for gas mileage?

[cedar skiff]

5.3 is cheaper because tokens cost less, but it is less token effcient.

[small violet]

i was building spotify integration for an open source dj software

[real blade]

how is it less token efficient?

[small violet]

you're on pro do u even care about efficiency

@nocturne folio

[nocturne folio]

yea?

[small violet]

fix the twitter link bro

and dm me

[nocturne folio]

autodisc0

[small violet]

im not dming u on twitter

[nocturne folio]

why

[small violet]

why would i

just open the discord dms lil bro

[nocturne folio]

why do you want to dm me in the first place

[small violet]

codex solve cancer

[lean lark]

( Looking for the report button to flag "insanity" )

hmmm, button says "Report yourself" ... nvm

[small violet]

lol

[cedar skiff]

ofc, no way i could do all the work i need if i didnt.

[nocturne folio]

we need to all start coding like peter steinberger

[boreal holly]

please subscrib to pro to get more token

[cedar skiff]

I wish, just brute force everything

[nocturne folio]

gulp

[small violet]

im poor

200$ for ai is crazy anyway

the permanent underclass is real

[lost drum]

who found a solution to not use a web chatgpt but an enviroment where he those not lag when outputting long messages and also an enviroment where long conversations are not lagging out

[lost drum]

200$ is nothing compared to what you can build witin a month + you dont need to buy it for 200$ you can just purchase it like x days before your plus sub ends while making sure the usage will reset in that time so you pay maybe like 50$ or osmething just 1 week of usage but better tahn nothing

[small violet]

when i start making 2000$ a month then maybe ill pay 200$

[dusk thorn]

get the 100 buck plan

[small violet]

ill get it when i make 1000

[lean lark]

I moved from web/cloud codex UI to VSCode and have only gone back to web a couple times in months. YMMV

[shell coral]

I have multiple packages in my monorepo that requires different set of skills, what is the appropiate way to put in AGENTS.md? should i use $ ? or just mention the folder name of the skill?

something like this?

- Always use $foo and $bar skill when working on `packages/foo-bar`
- Always use $baz skill when working on `packages/baz`

[slate schooner]

Just ask AI for help 😉

[lean lark]

One AGENTS.md in the root of each folder to handle just those unique details. The files cascade like CSS, overriding prior directives.

[shell coral]

how about mentioning about the skill in AGENTS.md? should i use the $ ?

[lean lark]

Try to avoid tensions though - try to avoid telling it to do different things as it cascades. Make instructions additive, not corrective.

I have no idea what that is. I though you were using a placeholder. What is that?

[frail meadow]

Looking like Gary Tan out here

[lean lark]

Someone else can help with skills and related folders/files. I was specfically just commenting on AGENTS.md.

[shell coral]

when invoking skill on codex CLI we need to use $ right ?, How about in AGENTS.md ? should I use it too ?

[lean lark]

It's not either-or : Agents provides instructions for the entire thread, with refinement in each folder. Skills are used to process specific scenarios conditionally. So yes, use both. But I'm not familiar with the "$" syntax.

[frail meadow]

I think it's for inside markdown files maybe?

[cedar skiff]

You dont need to use the $ in agents.md just the name of the parent folder of SKILL.md should be fine. It will find it.
$ is just a short cut to a mark down reference in the ui and what it really does it this: [$name-of-skill](/path/to/skill/SKILL.md) You can give this formatting if you like, but the agent should find it by the folder name and you saying it's a skill.

[shell coral]

I see, thanks 😄

[cedar skiff]

One of the important parts of skills is the description, that handles the what it is and the when to use it that the agent uses to make a judgement on when to call the skill into context and use it.

Make sure that is not too verbose, make sure it is direct and clear.

oh codex app was using 90gb of system memory after a 12 hour run, only noticed once the system started to shut down other apps o.0

well that is my longest run yet though. Just jumped straight back in and started it again

[thorny cloud]

How do I like see err codex progress when I'm outside on phone?

[dusk thorn]

you dont unless you go 3rd party

[nocturne folio]

god bless openai

god bless the dude who made it so that messages dont end when usage reaches 0

[frail meadow]

Superset is releasing mobile support this week. They allow you to use codex, Claude Code and other harnesses in a single application

[tiny schooner]

How do you guys deal with slop code? Codex does the job but the code it gives is ugly and doesn't fit the structure I want it to.

[frail meadow]

A good guide from OpenAI to start with: https://openai.com/index/harness-engineering/

[deft gyro]

vro can computer use come to windows/linux

:(

[lean lark]

I've commented on this many times here so some folks will see this as a boring repeat...

[tiny schooner]

Thanks I'll give it a read.

[tiny schooner]

I'm new here 😁

[nocturne folio]

its still going btw

cant believe i made that prompt 2 hours and 30 minutes ago