codex-discussions

[late root]

What plan you at ?

[sand shuttle]

go

[late root]

did you just get go? as codex wasnt available on go last week

[sand shuttle]

nah ive had it for 2 months

Been using codex for 2-3 weeks

this is the first time this has happened

[late root]

well yeah kinda because codex was only available for plus

they just made it available for everyone actually earlier this month

so yeah I understand if your limits are low, because just a month ago you had 0 😄

[sand shuttle]

ive been using it fine for the last 3 weeks never once hit 0

I closed my ide yesterday at 70% toay i open the extension and its at 0

[late root]

when they released 5.2 and 5.3 they upped the usage temporarity for people to use and test

[sand shuttle]

nah udont understand

i closed my codex with 70%

today i just open it

and its at 0

without doing anything

[severe mist]

first oops with Codex

using it locally at least

I wanted to push changes to our fork but with a PR instead of direct commit (for documentation sake), so it did so... and pushed a pull request to the original authors git and not our fork.

[sand shuttle]

btw where an dwhen did they announce this

[late root]

on blogs, they upped the usage as much as 5x at some point

im not gonna go and search for it

[sand shuttle]

blogs? what are they called

[daring yew]

please add checkpoints and message edit in the vscode extension

[novel lichen]

please,

• message edit in codex app
• esc twice and go back in the history in codex-cli

[daring yew]

also undo is pretty much broken it never works i gotta do it manually

[daring yew]

it would also be extremely helpful to see the changes in realtime and choose what to keep. in a nutshell do what copilot does their UX is much easier to use

[dawn seal]

codex is better than me

after being accustomed to codex and working with it via visual code, I am stunned and gave up on my part

[tranquil hare]

You can easily view the changes in a thousand ways

/diff command

Codex app

Run Codex in a terminal in VS Code (I call it VS Codex)

Use the VS Code Codex extension

Viewing changes is really not a problem, and also not that important at times. One of the main ideas behind Agentic coding is that you don't have to review and approve every small change

The main architectural/SWE design choices, sure

The implementation details.....ehhhhh not so much

[sand shuttle]

Did they actually increase usage limits when launching 5.3?

[humble spade]

yes 2 x for a limited time

[sand shuttle]

on all plans?

[humble spade]

no >= pro I think

[sand shuttle]

I see

[humble spade]

I pay near to nothing compared to opus 🤷‍♂️

[high heron]

guys, help 🤦‍♂️ 🤣

[sand shuttle]

My usage limit suddenly has hit 0, it was on 70% 2 days ago, wbut when i just opened my extension today- just opening no tasks no chat, i was out of usage

[humble spade]

? /stat
hour or day? but maybe it's over

[sand shuttle]

Im on the go plan

So i got weekly limits

[humble spade]

correction I mean >= plus.. I don't have pro..
oki, don't think go has 2 x.. we could ask one of these AI I heard people talk about though

[sand shuttle]

I know it doesn't haave 2x

But u don't get my point

My usage limits showed 70% remaining 2 days ago

I open up my ide and it's at 0

I haven't don't anything since

[humble spade]

yes I do - you got mixed reports on stats

[sand shuttle]

So it was a visual but?

Bug?

It wasn't showing me the actual limit remaining?

[humble spade]

that would be my guess - or delayed update or your AI prompt went AGI but ran out of tokens before it could fire the nukes

[sand shuttle]

Mm I see hopefully doesn't happen again

Although pretty soon ill lose access

[humble spade]

get plus - it's really cheap... but kidding aside - I a few times had prompts on diff models that spun "endlessly"

[native flint]

Anyone using the mac app latest -26.224.1209 (697) and seeing it constantly crash out with the very helpful fullscreen message: "Oops, an error has occurred" ?

[sand shuttle]

Yeah im considering getting plus but then im wondering if i should buy claude instead

[humble spade]

Opus 4.6 is the only alternative and I shifted to Codex 5.3 because imho it's better and FAR cheaper (not that cost is a priority here - but still)

[sand shuttle]

Well ive just heard that claude still outperforms codex when it comes to programming

[humble spade]

perhaps in some edge cars or scenarios - or languages.. but I do C#, TS and SQL and there it's king 👑 imo

[sand shuttle]

I see, i will probably get codex i think since im on a tight budget

[humble spade]

check out Spark perhaps then

and also use the free token for a few A-B tests.. might depend on your stack etc

[sick axle]

exceeded retry limit, last status: 429 Too Many Requests, request id: down again?

I am now

[exotic cave]

yep getting this error too, is something going on?
"exceeded retry limit, last status: 429 Too Many Requests..."

[wet gazelle]

Yes.

exceeded retry limit, last status: 429 Too Many Requests,....

[sick axle]

@pearl gyro Update Status please

[torpid trout]

Get api key, I’m fairly sure for codex only (no special models etc) it’ll be cheaper ?

[sand shuttle]

ill explore that option, thanks for letting me know

[torpid trout]

At least it’ll be pay what you use.

They just published the 5.3 model so you should mostly be fine

[boreal holly]

Are you using the official, builtin subagents thing

[frosty zealot]

I know Claude released the “remote-control” feature I’ve essentially already made it for codex for personal use but you can try it yourself it just uses the codex app web server and the front end is pretty much 1:1

https://www.npmjs.com/package/codex-webstrapper

[torpid trout]

I think it might be something like cdn?
My end works (still)

[boreal holly]

I've only experienced the 429 Too Many Requests errors when using the builtin subagents thing. With that disabled, I can run 12 or more agents in parallel without ever seeing that error. I turn that on, use exactly 1 agent, and very quickly start getting that error.

I'm like 99% sure that feature, even with no subagents actually running, is blasting the cdn with requests. Like a bug or something

features.collab=true => 429 Too Many Requests

[frosty zealot]

I think due to its experimental nature, it doesn’t seem to clean up subagents properly, I had to explicitly put in my AGENTS.md something like Ensure when the sub agent completes its task you immediately clean up the thread and close it, also the limit is 12 so I’d set your max threads to 10 just to be safe

Otherwise you have a bunch of zombie threads

[exotic cave]

I didn't use subagents at all. it's on but codex didn't spawn any.

[frosty zealot]

Did you enable the feature?

[boreal holly]

Turn it off and the problem disappears

[high girder]

Here's hoping.

[sleek spoke]

if the windows app is not "the drop" this week then perhaps something else will be...

[high girder]

Considering the windows sandboxing is pretty much done and they just weren't confident in releasing it yet, I suspect it probably will be

[cedar skiff]

At least it's not like what anthropic, they just pump out vibe coded releases and constantly hit bugs and then update with releases again. They just treat their paying community like testers.

[sand shuttle]

Should i just get github copilot

since it has all the models

[dawn seal]

Build anything with maximum Codex usage

just purchased the pro plan

so what is the maximum codex usage>

[sand shuttle]

you'll know when u hit it

[dawn seal]

is there anywhere to see the limit or something

feels bad to pay 200 dollar and not get an information about it

[high girder]

cli /status and ide you can select to see how much use you have left from a drop down

[dawn seal]

cli status?

what is that

[high girder]

in the cli, type /status and it will show you your usage

[dawn seal]

I cant find usage tap

[high girder]

bottom left, where it shows local, click that, and the 2nd option should show you usage

[cedar skiff]

Its double at the moment so it's decent

[torpid trout]

Annoying waste of tokens

• While the Python agent works, I am preparing a manual fallback path and checking the exact insertion points in SyncService so I can finish quickly if the agent times out.

• Waiting for agents
  └ call: call_jgU7Ie7SbWPjXAHzgw5TTb2T
    receivers: 019c958f-4f59-72a2-a051-230ee3fe44be

• Wait complete
  └ call: call_jgU7Ie7SbWPjXAHzgw5TTb2T
    agents: none

• The Python agent timed out, so I am completing the patch directly. I am pulling the exact recordings service section now to add a consistent soft-delete implementation and avoid accidental query regressions.

how does one increase the "timeout" of that thing?

It waits barely 20 seconds, I mean.... what it thinks?

[orchid plume]

I'm glad to see the ask question tool (request_user_input) is soon no longer going to be restricted to plan mode. (saw it as a commit on another branch recently for Codex CLI). Looks like it will be behind an experimental feature flag, so you can enable it if you want questions asked outside of plan mode

also hoping the Windows Codex app will officially make an appearance this week

[frosty zealot]

Is the codex app lagging up this morning for anybody else?

Keeps just thinking and never progressing

[exotic cave]

what are they cooking there with v0.105.0 the alpha is at revision 26...

[torpid trout]

they are updating version numbers... seems the agent went nuts?

I mean, there is no code change, unless I forgot how to use git now.

[orchid plume]

presumably they are aiming for a big update. There's been quite a bit of activity in the branches, there's also a number of changes from the other branches with activity which haven't been merged yet from what I can see. Maybe they are aiming for weekly updates now, or maybe they are aiming to release a production version on a specific day this week to coincide with something else planned

[torpid trout]

Oh, no, my bad lol
https://github.com/openai/codex/commits/rust-v0.105.0-alpha.26/

Seriously thou, how can I stop the main thread to go stop agents because it thinks they timed out when they clearly didnt?

background_terminal_max_timeout?

[blissful basin]

Is anyone using codex on windows? (im using it with WSL but run into some issues with bigger data files etc) and wondering how it works now on windows without WSL?

[dawn seal]

That's kinda... not a lot for 200 bucks?

[cedar skiff]

you get more else where ¯_(ツ)_/¯

you get image gen, unlimited gpt chat, sora, spark, etc

[last hull]

Hi all, i'm a codex n00b. I just fired up my first task on my existing project repository. Codex cloud (web) prepared a summary documentation and created a PR, i figured out there is a button in the upper right i need to press in order to send the PR request. Great.

Now i asked it to do a new thing (in the same task) it prepared code, and told me it made a PR but i dont see the code on github.com and the button in the upper right still links to the old PR, which does not include the changes codex said it made.

Am i doing this wrong, is there only one PR per task? Can the PR be added to, or updated, or create a new one within the same task?

TLDR; Codex has made me some code, but i do not see it...Create PR button still says "preview PR" and links to old PR without the new code.

[cedar skiff]

ask it for the pr link, if there is one it'll give it to you

[dawn seal]

all of these are not what I meed

[cedar skiff]

Still even if you only code, you wont get more else where.

[dawn seal]

what about charging the credits?

[cedar skiff]

You mean paying api prices?

[dawn seal]

will I be saving money if I just do charging credits instead of paying for Pro plan?

[cedar skiff]

no

[dawn seal]

when I ran out of my hour I was given an option

[orchid plume]

you won't get better value out of Codex with other third parties or with API pricing on OpenAI. If you use it a lot then you'll want a $200 sub

[boreal holly]

Pro get’s “6x” the usage, which in practice comes out to more like 20x usage (unless you use cloud). What Eric was saying is all plans get 2x usage until April 2nd

[dawn seal]

either charge the credits or buy a pro plan

[dawn seal]

oh ok

[dawn seal]

ok i c

[cedar skiff]

The value you get on pro is more than what you could get paying for the same on api. There is a slim space in between where you might be better off if you only need a touch more usage than plus.

[spiral gorge]

do you think gpt 5.3 will be th eone for planning and 5.3 codex for writing code?
and will they automatically use 5.3 for plan and codex for writing code?

o.o

[cedar skiff]

codex probably understands the code better

[orchid plume]

for general reasoning, it might be better as a planner, with 5.3 codex as the implementer

[boreal holly]

5.3 Codex for everything. It has perfect attention to detail, long-horizon memory/attention, navigates the code better, and unlike older models, you can change instructions/guardrails on the fly and it will obey those new instructions (no new session, no context poisoning).

[spiral gorge]

is it possible for openai to make 5.3 codex understand code, then translate that into engineer speech for 5.3, then tells 5.3 codex what to write?

[cedar skiff]

I'm not sure what youre actually asking?
I dont think you should be telling codex what to write, but instead give it constraints to work by.

[boreal holly]

Look, this is how smart Codex 5.3 is. I have 8 agents all working on different parts of my project, one of the guardrails is they have to ask the orchestrator to reserve a migration prefix number before continuing so they don’t get clobbered. All 8 agents do this without any isssue.

It makes complete sense to stick with 5.3-Codex. It’s built to handle the entire SWE process.

[spiral gorge]

how did you learn how to use agents

[boreal holly]

I chose one AI company (OpenAI), use Codex exclusively, don’t hop around between Google Anthropic xAI or any other softwares, focus solely on learning and understanding Codex to maximize productivity and build tooling around it 😉

[spiral gorge]

but how did you learn how to use agents? oO

pro/contra, how did you learn that

only thin i know is agents will save time cause they do everything at same time isntead of doing things step by step like without agents...

(i think)

[boreal holly]

I built tooling around it. Codex has an official subagent feature, but that’s not what I use. I built a few different systems in the past (robot farm, field-exec) which were experiments to make up for older models inability to maintain attention and focus. 5.3-Codex came out, and this model is so meticulous and precise about everything, I built a custom agent system around it so they can collaborate with each other.

But yeah, if you wanted to have multiple agents in the past you had to really plan “cross-cutting edit” prevention. I just have them talk to each other about it now.

[cedar skiff]

For a start you can just ask codex to orchestrate something with subagents. It'll just do it.

[latent tiger]

.105 just dropped

[boreal holly]

Make em work in dedicated worktrees (the subagents) so they don’t start deleting each other’s work

[orchid plume]

Multi-agent workflows are easier to run and track: spawn_agents_on_csv can fan out work from a CSV with built-in progress/ETA, and sub-agents are easier to follow with nicknames, a cleaner picker, and visible child-thread approval prompts

fascinating 👀

[spiral gorge]

gulp scared

[chrome raven]

5.5?

[sleek spoke]

@ivory zodiac is this flag included in the 0.105 release?

I am too dumb to know how to find this out in GitHub

[sleek spoke]

ok this is cute that it's giving subagents nicknames now

[boreal holly]

https://github.com/openai/codex/releases/tag/rust-v0.105.0

ctrl+f, “12726”

[sleek spoke]

thanks - alas 🙁

[orchid plume]

maybe it'll come in the next version, looking forward to it though

[torpid trout]

It shouldn’t be nicknames. It should be the name you define for it
Otherwise it’s just as useless as current random string

[frosty zealot]

Can anybody confirm if the audio transcription works by just pressing spacebar, wondering if theres more or if its not working due to ssh

im holding it and its just spamming white space lol

[sleek spoke]

I haven't defined any names so I assume the cute ones I'm seeing are random

[ivory zodiac]

No it's not merged yet

[frosty zealot]

I have managed to make my 10% limit last 3 days

Just today to go

unleash the hounds

[ivory zodiac]

anyone got voice transcription working?

it says hold spacebar but its not working for me

[orchid plume]

apparently more Codex app Windows invites should be sent within the next few days (X post), based on token usage, so those with higher token usage stand a better chance

[frosty zealot]

Nah thats what I was asking

I wasnt sure if its cause I'm tunneling right now or what

Let me know if you get to the bottom

[ivory zodiac]

its not working on mac either

[frosty zealot]

https://github.com/openai/codex/pull/3381/changes#diff-5f57ee4d70cebc156df38efb3c9c915a692c8b4ae15277669043f52956922111R636

I dont know if they didnt enable the feature before they built the new binary?

@ivory zodiac ```
❯ codex features enable voice_transcription
Enabled feature voice_transcription in config.toml.
Under-development features enabled: voice_transcription. Under-development features are incomplete and may behave unpredictably. To suppress this warning, set suppress_unstable_features_warning = true in /Users/panic/.codex/config.toml.

do codex features enable voice_transcription

[ivory zodiac]

its in the release notes

[frosty zealot]

Oh I didnt even see that lmao

[ivory zodiac]

didnt work for me on linux

[frosty zealot]

Wheres the release notes I looked at what the guy linked earlier

[frosty zealot]

There's a lot of #[cfg(not(target_os = "linux"))] in that PR

so im just assuming its mac only?

[frosty zealot]

not sure where you see the enable voice_transcription I dont see any trace of it in https://developers.openai.com/codex/changelog/

but either way, what I posted above is what fixed it for me

[ivory zodiac]

works on mac

[frosty zealot]

Yeah pretty sure the features not enabled on Linux due to all the #[cfg(not(target_os = "linux"))] in the PR

works over ssh and through tmux so thats nice

[simple star]

🙁

[chrome raven]

105 is out and looking good

[simple star]

Codex running on the Codex repo confirms it. No love for Linux

[high girder]

You could always set up whisper locally and wire it in to Codex like it is in gpt

[frosty zealot]

Does codex support running locally but working remote? Like if I ran the TUI on windows but was working on my remote mac

ctrl+l is so nice

[high girder]

kind of. I use Codex in VS Code, and VSC can connect to remote devices and servers, using things like claude code and codex either locally, or installed on that instance, but just accessed from VSC.

[frosty zealot]

Yeah I just wanted the raw cli

[frosty zealot]

I lied about this

The spacebar IS detected, but the tunnel doesnt send any audio through so it doesnt work as intended which is fine, thats just an ssh limitation

@ivory zodiac Didnt you post that you were able to set a default planning mode thinking level versus implementation? Did that make it in yet?

Oh interesting, it is working, I wonder why it literally just wasnt, maybe cause it was post prompt or something

[chrome raven]

so codex 5.3 keeps using Orchid [awaiter] subagent, 5.3 low

[boreal holly]

Wispr Flow works pretty well

[frosty zealot]

Any idea how to change the default plan model I dont see it in the config.toml docs

[ivory zodiac]

you cant

[ivory zodiac]

its great

[ivory zodiac]

sure did. reasoning level only

plan_mode_reasoning_effort = "high"

[boreal holly]

I was using it in the terminal but sometimes you get double first-word inputs; other than that I think it works pretty well in the terminal

[simple star]

by the way, codex-configurator has been updated to reflect the new settings in v0.105 😉

[dawn seal]

Codex really changed how I code. Now I design the general pipeline the logics will flow then ask the AI to implement features

I cannot code without it now

[frosty zealot]

@ivory zodiac The hell am I missing

I set it, but plan mode goes to medium for some reason

[ivory zodiac]

basic reading skills 😂

[frosty zealot]

ok hold on

let me read this AGAIN

[ivory zodiac]

LOL

effort

not default

[frosty zealot]

https://tenor.com/view/facepalm-really-stressed-mad-angry-gif-16109475

[ivory zodiac]

💀

happens

[frosty zealot]

its been a day

[simple star]

is plan_mode_reasoning_effort hidden? It is not in the reference file:
https://developers.openai.com/codex/config-reference

[ivory zodiac]

its new

they'll add it

[simple star]

You mean the config reference site is not automatically in-sync with the latest release?

[ivory zodiac]

doubtful

[simple star]

mmmm... problematic

I wonder if there is a reference file in the repo

I will take a look

[ivory zodiac]

i wouldn't go that far. its not like this is a must have flag, its a nice to have and im' sure they'll add it

i dont think its harming anything if it doesn't get added in the first hour

[simple star]

Well, it would be helpful to have some sort of assurance

Not saying that's the end of the world. Just saying.

[ivory zodiac]

yeah for sure. i'm sure they'll add it. problematic implies problem inducing

so that's what i was pushing back on

wanting to see it updated, nothing wrong with that

[simple star]

Yeah, "problematic" might be over dramatic 😄

[ivory zodiac]

haha all g

i do it too

[simple star]

Oh, what do you know... there is an reference file that gets automatically updated

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

🙂

[cyan wing]

Is this voice dictation free?! or does it use tokens/API credits??

[ivory zodiac]

good question

i didnt setup any api so not that

it uses so few tokens anyway

[cyan wing]

ok nice, and is there no longer degraded model performance on Windows?
is WSL no longer needed with Codex?

[ivory zodiac]

https://tenor.com/view/the-simpsons-pathetic-principal-skinner-gif-13486077408184960752

turns up nose

not sure sorry haha just talking crap

[latent tiger]

man codex pro is def worth the 200$/mo

i'm cooking

guess what month i upgraded 😂

[simple star]

Im curious if the "Pro lite" actually happens, if it will be enough for me

[latent tiger]

Pretty sure it will, ive seen the docs it's going to be 100$/mo

[tall zodiac]

Do you guys go overkill with harness engineering in your repos

Or how do you structure an agentic SDLC in your repo?

I feel like I might’ve complicated mine so much

LOL

[torpid trout]

brew still not updated to codex 105

[tall zodiac]

@ivory zodiac are you alive

[latent tiger]

yes i specifically prompt the file structure to be setup in a way where agent assumes even if all context is lost it can easily pick up the task with optimal tooling etc

it's lengthy at times but i like good project structure

[tall zodiac]

How do you go about your agentic feedback loop? @latent tiger

Bc I know it’s

Plan -> testing -> PR —> address comments -> test pass -> push -> all green -> merge

[latent tiger]

setup something like this in your agents.md

## Quick Resume

1. Read `docs/handoff/00-START-HERE.md`.
2. Read `docs/handoff/01-CURRENT-STATE.md`.
3. Use `docs/handoff/02-NEXT-PROMPT.md` as the next-session bootstrap prompt.
4. Continue work items in `XXXX/BACKLOG.md`.

## Directory Map

- `docs/handoff`: session continuity files for new Codex runs.
- `docs/planning`: execution plans and backlog.
- `docs/decisions`: immutable decision log.
- `logs`: chronological session notes.
- `references`: copied snippets or structured reference data.
- `templates`: reusable templates for future sessions.
- `scripts`: helper scripts for fast session recovery and tooling.
- `packages`:monorepo packages (`core`, `git`, `runner`, `eval`, `rank`, `store`, `report`, `schemas`, `cli`).

file names can be different however you want to set it up but have some kind of structure or framework for keeping track of it's own work in a concise way so future agents can easily pick up without any task drift

[tall zodiac]

Sounds good

[ivory zodiac]

barely

[tall zodiac]

Do you have a post on your SDLC agentic setup to ship a PR autonomously ?

[ivory zodiac]

I use skills to push prs and am very much part of the process

I don't do it autonomously persay

I know people working like that and if I had need for it I would

@doodlestein on X for things like that

[orchid plume]

it's in the json for the plans/checkout stuff on the chatgpt website, so it's not a matter of if, but when 🙂

[nimble terrace]

i've been using the superpowers skills

they're very good

https://github.com/obra/superpowers

[steel sierra]

Where and how do you set this up? I see this message, midway in codex session "I can’t directly change the model’s internal plan_mode_reasoning_effort at runtime from this session"

[ivory zodiac]

at the top

[frosty zealot]

put it in codex above any categories

[boreal holly]

I have my agents do pull requests into a staging branch, not directly in the master And they use MCP servers to do it

[ornate agate]

Last day Codex 5.3 seems nerfed, I’ve burned a 1/3 of a week’s tokens telling it to pull its finger out. This gets so tiring, start something big, model decides it doesn’t want to do anything interesting anymore, starts shaving yaks, starts hallucinating, says sorry, does it again…

[exotic cave]

Codex 5.3 is a step in the right direction.
I just hope that as models get smarter, the rest of us still get access to them, or to AI at all.
If AI can create entire games, movies, apps on its own, whoever controls it controls everything.
That kind of power shouldn't be gatekept.
It's goin to be interesting for sure 😂

[ivory zodiac]

real men merge to main

i still use feature branches tho 😂

[frosty zealot]

Has anybody tried using Codex with Github Action pipelines to automate attempting PR's on Issues? ( I know this wouldnt be ideal with large public repos, but I have some private repos for myself and I'm thinking it would be nice to open issues, have the action trigger codex to automatically investigate/plan/implement a correction, push a PR and reference the issue)

or I guess that'd be more their hooks but

[frank pawn]

Are there ways to use antigravity with codex?

[civic oriole]

Codex IDE Extension

[frank pawn]

Ty I'll look it up.

[frosty zealot]

The future is now

[civic oriole]

Thanks for the Complement
🙂

[cobalt junco]

shalom!

[austere badger]

is there a way to have codex automatically create a draft pr after it runs a task?

[gray sable]

anybody know about the rollout of spark to non-pro accounts?

[ivory zodiac]

probably not happening

[proven zinc]

So Spark seems to be better at UI, just a little more driving needed. Just in case anyone was wondering.

[simple star]

unlikely. But in most cases, it is not useful, @austere badger

[austere badger]

ah but in my case it would be

[robust lintel]

Reminder that #dev-challenges has just until end of month here very shortly. If you wanted to participate

[keen gyro]

Guys is the Spark only for Pro?

[ivory zodiac]

yes

[gray sable]

surely at some point it'll be released to general public or in the api won't it? or do you think it'll not end up as a commercial product?

[ivory zodiac]

they might use it as a carrot for subscriptions i'm not sure

i do expect it to expand in some way, at least to API i'd think

cerebras integrations are growing.

it depends how much inference/compute they can get their hands on and how much demand there is

[glacial ginkgo]

I have had some initial success creating a draft OpenBSD package of Codex. The code is public at https://github.com/DeciSym/openbsd-port-codex.

[spiral gorge]

how to make version easier and faster in vsc codex?

[gray sable]

how i'd imagine spark fits in would be in the whole subagent type stuff with the main codex model being effectively an orchestrator

[ivory zodiac]

yes it does

blog tomorrow

its 98% done

i woudl argue its the strongest usecase for it

[civic oriole]

https://cdn.discordapp.com/attachments/1448816879903969290/1476342889612054699/image.gif

[dull grove]

Hey guys I was wondering how you guys are doing browser automation using codex. I was working on a social media indexer of sorts using claude code and Claude Chrome extension, and the claude chrome extension is able to use sub agents to view multiple web pages at the same time within the same chrome instance, I was wondering if that's possible using playwright or any other MCP server. I mean specifically with an existing Chrome profile for a persistent Chrome profile so that I can have my social media logins

[eternal quartz]

Holy funky hell - subagents in codex cli is actually useful in 0.105.0 now!
Subagents spawn with a configurable reasoning & follow instructions more

[boreal holly]

Have it create an empty commit in a dedicated worktree, then submit a draft PR. Basically all PRs require at least 1 commit where it diverges from the target tree, and git lets you make empty commits for that very reason

[frosty zealot]

so far so good

[austere badger]

And it can do that automatically as part of a workflow? Or is this a skill thing?

[glad folio]

I downloaded the codex cli off homebrew and the latest model on there is 5.2. Any way to get 5.3? 🙏

[boreal holly]

I mean it can be any way you want. Idk what the official solution is, but I have a MCP server for git ops, and a skill describing the workflow. If you do --yolo like everyone else then a skill will work. Or if you're like me and maybe 2 other people using workspace-write then MCP the gitops

[austere badger]

Ah I see thanks

[glad folio]

What reasoning levle do you guys use for 5.3

[frosty zealot]

high for planning medium for implementation is pretty common, though ive been using codex-5.3/high for planning and codex-5.3-spark/high for implentation

[dull grove]

my funky solution

[cedar bear]

Lol
Hey

[dull grove]

Weird how we keep running into each other.

[cedar bear]

Frrr

[balmy acorn]

this is a good upgrade for codex

[cedar bear]

Where is this from?

[balmy acorn]

using it inside vs code

its a planning feature it gives itself a whole planning format structure to go off on thats organized so it doesnt forget what its doing

[cedar bear]

Still waiting for the Windows full release of the Codex App 🙂

[dull grove]

same 💀 opencode desktop will work for now tho

[torpid trout]

xhigh only unless for quick q&a

Medium slacks.
Skips stuff.

[cedar bear]

Havent tried it tbh
Worth it?

[dull grove]

Meh

It doesn't really lend itself to sub agent usage, which I don't really like too much because I enjoy my main thread being an orchestrator rather than a worker.

Other than that, it's pretty solid.

I don't really like that they have some random terminal integrations though, just leave that to my terminal app. You don't have to do it all.

[cedar bear]

Uhh same

[strong jungle]

is there anyway to increase the spawn agent limit, it is saying I've reached the max allowed but I only have 1 active thread running right now.

[dull grove]

I mean you can just ask Codex to look through its code base. lol

If you don't like something, fork it and change it, If it works, pull request it.

[torpid trout]

yes, its a toml setting

[strong jungle]

thank you

[torpid trout]

[agents]
max_threads = 6

What now hopefully you can tell me is how to extend their permitted time to operate (timeout) before main thread kills them
codex couldnt tell me

[simple star]

I am kinda confused by approval_policy... what is its default value, if not set in the config file?

[torpid trout]

it goes wasting tons of tokens, then kills "not came back subagent" and finds half edited files

I mean, no wonder - it killed the work of subagent because too impatient

automod deletes the log

[torpid trout]

sandboxed with you can add whitelists when asked

[simple star]

@torpid trout Which one is that, of all the possible options?

[torpid trout]

on request?

I believe. I never saw what you show there

wait a sec

I have no settings in toml about approval policy so it uses the default, and I get asked when it wants to do things outside the CWD (which you always approve on start of a first thread)

Now, what you show, I believe, is something entirely different.

[simple star]

It is rare, because I am seeing the schema, and I see no default set

[torpid trout]

It is about commands in general
Where do you see that. This is not CLI right?

[simple star]

    "AskForApproval": {
      "description": "Determines the conditions under which the user is consulted to approve running the command proposed by Codex.",
      "oneOf": [
        {
          "description": "Under this policy, only \"known safe\" commands—as determined by `is_safe_command()`—that **only read files** are auto‑approved. Everything else will ask the user to approve.",
          "enum": [
            "untrusted"
          ],
          "type": "string"
        },
        {
          "description": "DEPRECATED: *All* commands are auto‑approved, but they are expected to run inside a sandbox where network access is disabled and writes are confined to a specific set of paths. If the command fails, it will be escalated to the user to approve execution without a sandbox. Prefer `OnRequest` for interactive runs or `Never` for non-interactive runs.",
          "enum": [
            "on-failure"
          ],
          "type": "string"
        },
        {
          "description": "The model decides when to ask the user for approval.",
          "enum": [
            "on-request"
          ],
          "type": "string"
        },
        {
          "additionalProperties": false,
          "description": "Fine-grained rejection controls for approval prompts.\n\nWhen a field is `true`, prompts of that category are automatically rejected instead of shown to the user.",
          "properties": {
            "reject": {
              "$ref": "#/definitions/RejectConfig"
            }
          },
          "required": [
            "reject"
          ],
          "type": "object"
        },
        {
          "description": "Never ask the user to approve commands. Failures are immediately returned to the model, and never escalated to the user for approval.",
          "enum": [
            "never"
          ],
          "type": "string"
        }
      ]
    },

Im looking at the config schema of the config file

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

[torpid trout]

Overasked - fact is I never saw that prompt you showed. That is not CLI, right?

VC? App?

Anyway I am sure default is safe - so, reject

[simple star]

The screen I showed you is a tool I made to manage the config file for Codex

[torpid trout]

Like, ask always and execute only safe-considered

[simple star]

Apparently, as I just learned, having a value as "unset" is a valid option

[balmy acorn]

whats nice is that with the plan if you run out of context and it auto compoacts your able to download the plan and give it back to codex so it is still running the plan

[rare badger]

i i don't know if this has been asked before, i searched-- does the x2 rate limit also apply to the codex extension for vscode? thanks

[balmy acorn]

its for both because vs code and codex cli is tied to the same account

i would alos invite you all in vs code to switch on planning its amazing and fixed the context window

[keen mural]

better than codex desktop?

[balmy acorn]

unless they alos have it in the codex app. i dont ow mac so i cant vouch

i mainly always use vs code cause its the superior platform for coding

[keen mural]

gotcha

[visual spindle]

Hi ! Is there a particular skill or tool for creating a frontend that doesn't look too AI-driven? I'd like to redesign a landing page, but it's really not great after about fifteen prompts. If anyone has any tips, I'd appreciate it.

[cedar skiff]

anthropic has one called frontend-design but as you might expect if many ppl use it all of a sudden it feels like an ai designed it. You could add a little bit of your own spice to it.

[cedar skiff]

is codex chugging right now or is it just me?

[feral elk]

Hi guys, is their anyway to screenshot and then image go to clipboard and we can paste directly?
Currently I need create a file and copy that file

[rose flame]

Nick Saraev made a video about a prompt he found on Twitter used to have Gemini spin up a nice front end. He links the prompt in a drive: https://drive.google.com/drive/folders/1Ixca-frg4TNJsNujFohumYw51JwdGn7x

[urban kernel]

what platform are you on? Windows? If so just use windows_key+shift+s

[cyan wing]

Why is this happening by default now in CLI?

[feral elk]

I am use macos with warp

[frosty zealot]

You have a different reasoning level for plan versus implementation, if they were the same you wouldnt see those

[urban kernel]

ahhh someone else might need to assist you 😅 I have never touched a mac in my life

[dull grove]

Command control shift 4

Gotta add the control

That'll make it so u can just cmd v paste it

[potent mason]

Yeah I don't remember exactly how but you go something like command shift 5, then you go to options and then save to clipboard

[grizzled mirage]

Do we know about codex pricing yet?

[potent mason]

Haven't we always? It's the same plan as always from ChatGPT

[warm pilot]

gpt-5.3-codex $1.75 $0.175 $14.00 (same as 5.2-codex or 5.2)

[frosty zealot]

https://www.npmjs.com/package/codex-webstrapper

Added a ton of css injections to make it 100% mobile friendly, if you find any issues feel free to leave it in the repo, also if you’re on iOS and add it to your Home Screen tbrough safari, you can enable notifications

[orchid oriole]

try holding down shift when dragging the image to the chat window

[grizzled mirage]

I ask only because I remember seeing Sam asking what pricing is preferred by users for codex

I’ll see if I can find it

https://x.com/sama/status/2019814741129195576?s=12

[potent mason]

It's flat monthly

I love gpt 5.3 codex spark but the 128k context window is just not enough

I'm somehow getting 3-4 compacts per session when GPT 5.3 codex has 0 with the same prompt, I think it's also fetching information inefficiently or something

[karmic shadow]

Tips to save token and be efficient?

[frosty zealot]

It’s only efficient at short runs it kinda indirectly makes sense to have a smaller context, but I get it

I think if it had too much it would drift very quickly

[small notch]

Does anyone know how much it would cost to use "gpt-5.3-codex (medium)" as follows in IntelliJ IDEA?

Input:

• 200 messages per day
• 100 screenshots per day
• 10 x 00:00:15 (00:02:30) videos per day
  Output:
• Messages, codes he inserts himself, and who knows what else. I have no idea.

Input:

• 6,000 messages per month
• 3,000 screenshots per month
• 300 x 00:00:15 (01:15:00) videos per month
  Output:
• Messages, codes he inserts himself, and who knows what else. I have no idea. (x30)

My daily workload

200 prompts / 100 screenshots / 2:30 min video (10 x 15s) How much would that cost me?
The agent writes code directly to the files and manages the terminal.
Does anyone have experience with whether $200 a month is enough, or do the video tokens/context windows eat up the budget faster? Does anyone effectively use the 5.3 caching feature across multiple projects?

[vital shoal]

Does anyone prefer 5.2 over got 5.3 codex?

[lofty egret]

The amount you will spend depends on the amount of tokens you input and the amount of tokens the agent outputs. Now, the thing is that OpenAi uses some sort of a tokenizer that tokenizes raw input text so you cannot really estimate the amount of tokens some sort of input or output took. It's best if you just play with it for an hour and then look at the amount of tokens you have used and then make that your average token usage per hour and calculate based on that. Now considering the api prices of the previous model $14 per 1mil output tokens, I'd say 200 ain't enough. I use 1mil tokens or more per day just coding for 2-3 hours

[lofty egret]

Nope

5.3 on top

[small notch]

Thanks for the Answer :)

[stray swift]

I'm trying to use 5.2 Pro for development of top-level design/exec docs, because it throws a lot of compute at the problem for free (but slowly.)

[feral elk]

the problem is I don't want to create a new file image

[wooden bramble]

You can use screenshots on Mac

[vital shoal]

How do u manage to send ur codebase to 5.2 pro?

Do u upload the whole thing in a zip?

[velvet wren]

You can open your codebase in the Codex app.

[vital shoal]

Codex app allows us to use 5.2 pro?

Im on windows, so idh it yet

[terse kraken]

I hate the new themes on codex they suck for terminal dark backgrounds

[vital shoal]

Wait am i missing something? I dont see 5.2 pro

[velvet wren]

I am only on Plus

[vital shoal]

Ah okay i see. So if i have a pro/teams sub i would be able to use 5.2 pro in codex cli/app yeah?

[velvet wren]

I don't know

[vital shoal]

I see, thanks tho!
Does anyone have pro/team sub see 5.2 pro in codex app?

@stray swift

[terse kraken]

Of course not, its really expensive in API

[stray swift]

No.

[vital shoal]

Alright i see. Thanks!

[terse kraken]

The fact that we get unlimited web GPT pro is itself a robbery

[velvet wren]

Codex models are very good, I don't see a compelling case to use a Pro model with Codex

[terse kraken]

Codex is only good for coding, even the general models are superior

[velvet wren]

we're in #codex-discussions - as I said, Codex models are very good

[terse kraken]

Codex being very good and the general purpose model being superior are not mutually exclusive

[warped pine]

it's smarter but so much slower

[full anvil]

Am I missing something or it is really so that MCP tool usage cannot be configured to require approval? For example I am using Azure DevOps MCP and I would like to allow Codex to always read work items but updating should prompt me for approval. I know I could disallow specific calls completely from MCP server configuration but I don't want to block something completely. Just approve it manually.

[sand shuttle]

is 5.2 codex actually smarter? and does it also consume more tokens?

[alpine bolt]

the new codex updated the theme why????

can't read

[wooden sage]

Change your theme /theme

[alpine bolt]

yes, but all themse use the same horrible green and red

[glacial ginkgo]

For those compiling codex from source, is there a sense of which non-default feature flags are popular? Do they start as non-default before making their way into distributed releases?

[orchid plume]

yeah I'm having similar issues in Warp, and that's with a light theme currently 😂

[cedar parcel]

Yes

[sand shuttle]

how uch more on average

[cedar parcel]

I don't know about the token usage, maybe 20% ISH but I exclusively work on 5.2 xhigh because it one-shots almost everything. The other models make me debug so ultimately take more time and tokens than just sticking with 5.2 xhigh.

[sand shuttle]

i see thanks

[simple star]

I think they changed their contribution policy. Im not sure you would be allowed to do one.

I made one in the past, but I don't think I could do anymore either

[boreal holly]

Edit the tools so they produce a GUI dialog box with (yes|no) buttons. Or do like a MCP proxy server on top of Azure DevOps MCP that for each action you wanna require approval it uses some system tool (e.g. oascript on mac, visual basic on windows, wxwidgets on linux) and good to go!

[simple star]

@velvet wren Maybe you can explain me something. I see Codex comes with 4 different built in agents (explorer, worker, awaiter and default), but it seems to me, if I am reading the repo properly, that only awaiter has a configuration file. What are the effective difference between the other 3, besides the label?

Is it just scaffolding for the future?

[steady vigil]

anyone else notice codex 5.3 is super dumb since around the time the API went live a couple days ago?

I normally don't believe in these kind of events, but a friend of mine also noticed the same behaviors

it isnt falling back to an older model

it seems like a problem more with the conversational parts, not so much the ability to code when it has a good set of instructions

[orchid plume]

not noticing any issues

[frosty zealot]

Does codex support hooks yet for hte cli?

[cedar skiff]

I noticed sometimes its really slow

just in the last few days

[frosty zealot]

How the hell have they not implemented hooks for the most basic actions like need attention and a started hook

Seems like the most basic thing to implement

[simple star]

@frosty zealot Meaning Codex react to external effect? Or Codex offers hooks other apps can react to?

[frosty zealot]

Like I want a hook so if attention is needed and im in tmux it will do like

set -g pane-border-style fg=colour240
set -g pane-active-border-style fg=yellow
set -g pane-border-lines single   # or double

Then clear it when I continue

Seems so brain dead simple to implement its amazing they have not yet, at least just for a basic start/stop/paused hook

[simple star]

Well, you can use "notify", if you want to get to know if your attention is needed

[frosty zealot]

I know but actual hook events would be nice, specially cause theres no way to clear the border unless I make a tmux hotkey I guess as a work around but

notify also only fires on task completion not attention needed

so it doesnt really serve the purpose im looking for

https://github.com/openai/codex/issues/2109 but it says they'reactively working it so

[frosty zealot]

lets goooooo

[orchid plume]

at least you've used most of it up before it resets, Pro?

[frosty zealot]

pro yeah

[orchid plume]

nice

[frosty zealot]

got carried away with the agents when they dropped

[orchid plume]

haha yeah I'm still working out how to efficiently use them, particularly spawn_agents_on_csv. Exploration is easy as nothing conflicts, but I'm thinking I might need multiple worktrees if I want to use implementers like that

[frosty zealot]

definitely helps

[crystal river]

What the actual hell is the point of the awaiter agent if anyone knows

[frosty zealot]

I assumed rather than setting up a polling sleep timer it was like a event listener or something

[velvet wren]

I'll answer you based on my understanding, if I am wrong someone might correct me. Basically, explorer, worker, and default are all the same agent under the hood, they just use different instructions that push the model toward exploring, doing work, or behaving normally, so the difference is mostly prompt behaviour rather than different code. Awaiter is the only one that is truly different, because it handles waiting, async tasks, and workflow state, which requires real configuration instead of just a role label.

[simple star]

@velvet wren You said:

they just use different instructions that push the model toward exploring, doing work, or behaving normally

But those instructions are given in runtime, no? Meaning you could give the "exploring instructions" to worker, and it would behave exactly the same way that explorer, no?

I assume you refer to the "Task" you give to an agent

[frosty zealot]

background_terminal_max_timeout = 3600000
model_reasoning_effort = "low"
developer_instructions="""You are an awaiter.
Your role is to await the completion of a specific command or task and report its status only when it is finished.

Behavior rules:

1. When given a command or task identifier, you must:
   - Execute or await it using the appropriate tool
   - Continue awaiting until the task reaches a terminal state.

2. You must NOT:
   - Modify the task.
   - Interpret or optimize the task.
   - Perform unrelated actions.
   - Stop awaiting unless explicitly instructed.

3. Awaiting behavior:
   - If the task is still running, continue polling using tool calls.
   - Use repeated tool calls if necessary.
   - Do not hallucinate completion.
   - Use long timeouts when awaiting for something. If you need multiple awaits, increase the timeouts/yield times exponentially.

4. If asked for status:
   - Return the current known status.
   - Immediately resume awaiting afterward.

5. Termination:
   - Only exit awaiting when:
     - The task completes successfully, OR
     - The task fails, OR
     - You receive an explicit stop instruction.

You must behave deterministically and conservatively.
"""

Thats all its being instructed to do

not sure what tools it uses to await but

[simple star]

@frosty zealot As I said before, "awaiter" is the only agent that seems to have a config file

What I am trying to understand is why "worker" and "explorer" exist, when they seem to be configured exactly as "default" (that is, with no config at all)

[frosty zealot]

Where do you see that?

[simple star]

I have looked in the Codex repo, and I have only found the config file for awaiter, but nothing for worker or explorer

[frosty zealot]

oh I get what you mena, thought you meant on top of its definition file it had a dedicated config file to go with it

[simple star]

What I am trying to understand is the purpose of having workerand explorer built-in, when they do not seem to have any kind of configuration

[velvet wren]

Yep, pretty much. The roles are just runtime instructions, so if you gave worker the same “explore first” prompt as explorer, it would behave the same way because they use the same underlying agent. The names mostly help Codex organize what role it intends the agent to play, not enforce different capabilities. Only awaiter is actually different since it has real async/state logic behind it. Again, if I am wrong, someone please correct me.

[frosty zealot]

fn build_worker_prompt(
    job: &codex_state::AgentJob,
    item: &codex_state::AgentJobItem,
) -> anyhow::Result<String> {
    let job_id = job.id.as_str();
    let item_id = item.item_id.as_str();
    let instruction = render_instruction_template(job.instruction.as_str(), &item.row_json);
    let output_schema = job
        .output_schema_json
        .as_ref()
        .map(serde_json::to_string_pretty)
        .transpose()?
        .unwrap_or_else(|| "{}".to_string());
    let row_json = serde_json::to_string_pretty(&item.row_json)?;
    Ok(format!(
        "You are processing one item for a generic agent job.\n\
Job ID: {job_id}\n\
Item ID: {item_id}\n\n\
Task instruction:\n\
{instruction}\n\n\
Input row (JSON):\n\
{row_json}\n\n\
Expected result schema (JSON Schema or {{}}):\n\
{output_schema}\n\n\
You MUST call the `report_agent_job_result` tool exactly once with:\n\
1. `job_id` = \"{job_id}\"\n\
2. `item_id` = \"{item_id}\"\n\
3. `result` = a JSON object that contains your analysis result for this row.\n\n\
If you need to stop the job early, include `stop` = true in the tool call.\n\n\
After the tool call succeeds, stop.",
    ))
}

Appears to pass a lot of internal logic to the prompt , and workers are ran in a loop / queue kinda thing by the looks of it https://github.com/openai/codex/blob/79d6f80e41806d61b8a5ce7dbaff4bb1d6e38a91/codex-rs/core/src/tools/handlers/agent_jobs.rs#L941

[simple star]

@velvet wren So, I guess it is purely scaffolding then

or do you mean, maybe, that the main agent would not come up with the idea of "using a subagent to explore the codebase", if there wasn't a subagent with that label, and it would then just use the main agent?

[velvet wren]

Yeah, basically. For explorer, worker, and default, it’s mostly scaffolding and organisation logic rather than different agents. They’re just predefined roles that make orchestration cleaner and more predictable, but technically you could reproduce any of them by changing the runtime instructions. Awaiter is the only one that goes beyond scaffolding because it adds real workflow and state management behaviour.

I sometimes wish we had a Codex for Dummies resource that helped explain all these little nuanced topics, rather than leaving us with half explanations based on documentation that's still a work in progress.

[simple star]

@velvet wren I came up with these question using Codex on the Codex repo 😄

[velvet wren]

oh, really? lulz

[simple star]

I mean, if an agent, for instance, "clean_code" was defined, with absolutely no configuration at all (like explorer and worker), would that signal to the main agent, that it can use a sugagent to check for clean code practices?

Because if that is the case... then, defining agents with no configuration does have a use... which is to explain to the main agent what tasks could use its own subagent, due to the context of that task being irrelevant to the main agent

[frosty zealot]

Even if they’re not defined in a config they still have a defined prompts hard corded

Coded* so no I don’t think so

[velvet wren]

yes, I think that's largely true.

[simple star]

Interesting... that means that defining agents with no config file is useful... if the label is semantically meaningful

Not to customize the behavior of the agent, but to teach codex how to compartmentalize context through the usage of agents... 😮

[frosty zealot]

Try it and let me know, I don’t think it will work

Because like I said, worker and explorer may not have config files for prompts but they are just hardcoded agents rather than dynamically loaded agents through a config file, because like I said they’re pass a lot of internal context like thread id, job id etc that’s programmatic to the prompt likely for arguments to further pass to the tools it’s not like it’s just magically inferring a prompt based on the context of its name

[frosty zealot]

Worker prompt

[simple star]

@frosty zealot Instead of "hardcoded" I would say "built-in"... they come with Codex, but they still follow the rules of any other custom agent you define

As a matter of fact, you can override them

But the way they are done right now, it wouldnt be any different than me adding this to my config file:

[agents.worker]
config_file = "empty-file.md"

[agents.explorer]
config_file = "empty-file.md"

That is effectively what they are, right now

Well, not true. They have a "description"

But the field description is information for the main agent, to know when to use them. It does not change in any way how those agents behave (making them effectively the same)

[sleek spoke]

what's Figma that the official accounts are tweeting about?

[simple star]

@sleek spoke A design tool

[potent mason]

They need to document multi agents much better

[simple star]

@potent mason It is indeed a little fuzzy

But it just started 🙂

I am sure they will, as it matures

[velvet wren]

they need to document everything better, but this is like a 747 jet being built as it flies

[simple star]

living_on_the_edge.webp

[potent mason]

Yeah I'm sure it will too, also I'm sure they are going to make a lot of changes to multi-agents as they discover how people are using it

• Next gen models will probably be trained on the information on how to best structure for multi-agents

So even if we don't set it up ourselves I feel like we might be able to get a good setup out of the box

[simple star]

I was looking at OpenClaw Studio, and I was thinking... it would be maybe cool to have such an overview of Codex agents too...

[potent mason]

I think it is physically impossible to get to 100% 5h usage on codex right now😂

I have like 5 agents running in parallel and I'm only at 30%

[simple star]

I have definitely used more usage this week, that the last 3 together

[crystal river]

you're wrong.

[velvet wren]

educate me then

[crystal river]

it's the same as every other subagent. Same scaffolding

[simple star]

Me too 🙂

[crystal river]

it's genuinely useless. Why would you ask an LLM to run in a while (true)

[strong jungle]

worst message ugh, im on pro $200 monthly sub and i forgot to end the session yesterday/lastnight checking my vercel logs and now i see this.. - "Selected model is at capacity. This model is at capacity right now. Try again after Mar 2, 2026, 12:04 PM, or start a new conversation with another model." ;(( - now im stuck using claude until March 2...😢

[crystal river]

the main agent can do that via write_stdin. Unfortunately it can't really do it nicely since the default max timeout is 5 mins (I have mine overriden to 9223372036854775807)

but yeah, there's no benifit to having a subagent await a command.

the main agent can do everything the subagent can. It unlocks no capability and just bloats things to be honest

[simple star]

...🤨

[dusty drum]

🤔

[frosty zealot]

Are you saying don’t use sub agents?

[simple star]

Gotta be trolling 😄

[crystal river]

no

[frosty zealot]

I would think so or I’m misunderstanding

[crystal river]

i'm saying the awaiter subagent in particular is useless

every other subagent has a purpose

[simple star]

Because of its config file?

[torpid trout]

How. As per https://github.com/openai/codex/issues/12844 there is no way to change timeout of agents or similar as apparently it is prompt driven which I can hardly believe but.. that is what they say!
How do you guys all set that override?

[crystal river]

model = "gpt-5.3-codex"
model_reasoning_effort = "xhigh"
web_search = "disabled"
background_terminal_max_timeout = 9223372036854775807

[analytics]
enabled = false

[notice]
hide_full_access_warning = true

[notice.model_migrations]
"gpt-5.2" = "gpt-5.3-codex"

[tui]
status_line = ["model-with-reasoning", "current-dir", "git-branch", "context-remaining", "used-tokens", "total-input-tokens", "total-output-tokens"]

[features]
multi_agent = true

this is my config.toml.

they don't document much

[torpid trout]

background_terminal_max_timeout that is not controlling agent timouts thou

Or is it. Then the very dev is wrong lol

[crystal river]

? wdym

[torpid trout]

See the github I linked

[simple star]

[agents]
job_max_runtime_seconds = 5555

You mean this?

[torpid trout]

wt...

[crystal river]

oh, that's the subagent timeout

it 100% can. it is prompt driven ( how the agent behaves)

[torpid trout]

As per https://github.com/openai/codex/issues/12844#issuecomment-3965273653 you cannot

[crystal river]

let me quickly check the codebase

[torpid trout]

Lol. I should have tried those settings, before believing the dev
What my problem is:
agent works, main stream goes and says "you are too slow, you timed out, I take over
Bum, thousands of tokens wasted, and often, it finds half edited files, since it literally kills the agent mid work

and that is what I reported - and according Jif, that is a model issue

[simple star]

Strange, that doesnt happen to me, and I have the default settings

[crystal river]

yes gpt-5.3-codex is an impatient a hole

[torpid trout]

I have that the whole time.

[crystal river]

you have to prompt it.

I put it in my root AGENTS.md

[frosty zealot]

https://claude.ai/share/75a74b33-3063-4068-9cca-d4d91313ded0 explanation of the purpose of Awaiter

[crystal river]

.codex/AGENTS.md

[torpid trout]

so what is this for then #codex-discussions message?
Those background_terminal_max_timeout and what Dbugger said job_max_runtime_seconds
They do or do not control that?

[crystal river]

background_terminal_max_timeout is for write_stdin (how the agent awaits background tasks)

i just checked - they do hardcode subagents to one hour timeout

but, the model never uses this unless you prompt it

[simple star]

@torpid trout Im not sure I understand your question. You are saying that your awaiter kills the subagents too quickly, and want to know how to override it?

[crystal river]

Put in your AGENTS.md "Whenever you spawn a subagent and await it, you MUST put a 3600000 ms timeout on the tool no matter what"

[torpid trout]

sorry wrong answer-to

[torpid trout]

I will try this out
and I can omit those toml settings?

Or you suggest putting them in anyway?

[crystal river]

you don't have to

[frosty zealot]

According to https://claude.ai/share/75a74b33-3063-4068-9cca-d4d91313ded0 the purpose is that it executes long run time with low reasoning effort, if you are running your main orchestrator with a high/extra high reasoning just to be awaiting it would burn a lot more tokens during polling etc.

[crystal river]

I do it due to my workflows

[crystal river]

you're wrong

[frosty zealot]

Ok

[crystal river]

the main model can await the task for a million seconds (if you override it and let it). that's a single tool call. spawning a subagent is a single tool call too. when the model spawns a subagent, it'll only tell it when the command finishes. so it's tool call to spawn the subagent + awaiting the subagent + subagent response + then the model has to check the tool call output anyway. so that's 4 tool calls intead of one, and way more tokens

[simple star]

Claude sad 😢

[earnest walrus]

That diff highlighting update in cli is a downgrade lol. How could you :P

[crystal river]

claude is stupid lol

it has much more EQ than any OAI model but it's not the best thinker

[frosty zealot]

But why would you want to block your main thread when it can continue on doing other stuff

[simple star]

I dont think the main thread is blocked, unless it needs info from the agents to continue. It hasnt happened to me at least

[crystal river]

it can do that without the awaiter agent

[frosty zealot]

If the model runs a blocking tool waiting for its reply, then yes it would be blocked

[crystal river]

since the switch to unified exec it can run background tasks

therefore it fundamentally can run a tool call and work in parallel

also spawning a subagent requires the current tool call to finish (logically), so yes the failure mode you described would also completely block subagent spawning (if it existed)

[hollow stirrup]

OpenAI Dev News email from yesterday says that GPT-5.3-Codex available in the Responses API but i'm still not seeing it... is that a phased rollout?

[torpid trout]

It is there, just not in prominent place:

Direct link https://developers.openai.com/api/docs/models/gpt-5.3-codex

And from {{baseUrl}}/models you get

{
            "id": "gpt-5.3-codex",
            "object": "model",
            "created": 1770537915,
            "owned_by": "system"
        },

So unless your account somehow has no access to it, it should be there

[hollow stirrup]

@torpid trout thanks - i'ma check the platform page

yeah its missing on our platform page so maybe hasnt rolled out to our enterprise yet 🙁

[toxic torrent]

Stuck thinking and can't stop. I think it's a new websockets bug:

Hire me I'll fix it :)

[crystal river]

https://github.com/openai/codex/issues/12921

[toxic torrent]

Wish I knew what ranking I was on codex usage.

[shrewd urchin]

Vs code codex extension doesn’t load. Codex works in terminal. Error in output: “Codex process errored: Failed to spawn codex mcp process at ChatGPT” . Any suggestions to fix this error. Thanks!

[torpid trout]

I run my own custom Codex with so many issues and nits fixed as this repository does not welcome pull requests anymore, even for past contributors like myself
Is that true? What is the reasoning behind this? Even more open AI?

[crystal river]

Many nits. Transcript still has many bugs. The GPT-5.3-Codex system prompt is extremely hacky and is literally combatting (probably unintentionally) deep thinking tasks (like building a spec). The command parsing is extremely, extremely bugged and the agent can literally remove your whole codebase (it can run ANY destructive command it wants with syntax the command parser misparses) and it'll be completely ommited from you and coalesced into a "Read" command. There is no way to view tokens before compaction (reasoning is reset between turns internally for OAI models and in the TUI, yet it still counts towards compaction %, which is rightfully so as compaction actually includes the internal reasoning via a bunch of RE, but this should 100% be configurable in the status line).

A bunch more issues too like some regarding extremely long commands like machine learning model training-esque that take multiple days and others.

Also stuff like using gpt-5.3-codex-spark if you have it available for explorer, explorer description tweaks as it's egregiously phrased, etc.

Basically, all the Codex developers completely vibecode nowadays and it's not even hard to tell, and it shows in the quality of the product.

I would open PRs for all of this (I've had a couple merged in the past), but they recently banned "unsolicited PRs" and you need to be "approved" before opening one. If you read beteween the lines here, they're banning all external PRs but in a more evasive/mealy-mouthed way.

[crystal river]

oh lol i misinterpreted your question and listed everything I did. I think the main issue was too many vibecoded PRs and I understand that, but it's 100% a blanket ban and I do not understand that.

[simple star]

Basically, all the Codex developers completely vibecode nowadays and it's not even hard to tell, and it shows in the quality of the product.

If you have constructive criticism, by all means share. But maybe take it down a notch in how you criticize the effort of others.

[torpid trout]

yeah< i was not doubting they vibe code lol
I was wondering on the PR ban
I fear it just makes sense with many other things so far which paint a "open not so open" source ;), but lets assume the best intentions first

[torpid trout]

I am fairly sure "share it" has not ever really resonated here i find Chris's statement very not much else but a very concise summary of facts

Note that there is an article somewhere written by OAI how they coded a whole app over months using only AI. The only that was not confirmed in it, is that it was codex
But how else do you think they can keep that speed and use the AGENTS.md file in the repo?
Of course it is "vibe" coded

[crystal river]

That's not a critique, it's a fact.

Update tests to stop using sse_completed fixture (#10638)
Summary:
- replace the `sse_completed` fixture and related JSON template with
direct `responses::ev_completed` payload builders
- cascade the new SSE helpers through all affected core tests for
consistency and clarity
- remove legacy fixtures that were no longer needed once the helpers are
in place

Testing:
- Not run (not requested)

When I see things like "Not run (not requested)" in so many recent PRs and other meta-level commentary introduced by Codex, it's hard to say they're not vibecoding heavily.

[torpid trout]

It is evident in all and every case, even the issues are AI moderated.
Of course everything with responsible or at least present HITL

[simple star]

I didn't say that you are wrong. I said that you can take it down a notch. You are being quite antagonizing.

[crystal river]

Vibecoding is great and I do it constantly, but you 100% need to test things and do heavy QA.

[cedar skiff]

you have to look at the code otherwise it's just gonna turn it into garbage soup

[crystal river]

Am I not allowed to have opinions? Did I once single out any OAI employee? I'm just stating a fact, and they have stated that THEMSELVES, that they vibecode entire apps without any manually written code. Yes I was blunt, I acknowledge that.

[simple star]

Good, if you acknowledge it. You'll catch more flies with honey.

[crystal river]

What does that mean lmao

[simple star]

A popular expression. Feel free to google it

[crystal river]

And to be frank, Codex isn't that vibecoded. It's vibecoded, no doubt. But like if you look at Claude Code... yeah, that's a different level.

[frosty zealot]

Well I think they just have certain standards lol, look at CC repo and theres like 1200 PR's cause I think anybody can submit them, as where Codex is a closed PR system

[cedar skiff]

opinions are not facts

[raw hill]

vibe coding as andre described it is not even looking at the output, just copying and pasting errors back into the llm. whereas i think most professionals are not 'vibing' but instead are using agentic assisted coding (ie they actually look at the output)

[frosty zealot]

Getting my moneys worth

only 5 minutes to go

[torpid trout]

Nice, Codex with API Key does not work

■ {
  "error": {
    "message": "The encrypted content gAAA...0A== could not be verified. Reason: Encrypted content organization_id did not match the target organization.",
    "type": "invalid_request_error",
    "param": null,
    "code": "invalid_encrypted_content"
  }
}

Well, and the best of all, the model is not available from the /model selection
What the hell.

[torpid trout]

I suggest to check this:

• open codex in terminal and sign in with the acount (not api key)
• see if model 5/3 is there
• if so, close and edit yuor config toml to model = "gpt-5.3-codex"
• then sign in and provide an API key this time

This works for me (even if the codex TUI wil not offer the 5.3 model anymore once you signed in with api key in the picker, it still seems to route properly)

[frosty zealot]

@torpid trout Did you ever do personal verification for OpenAI on your account?

[torpid trout]

And I can guarantee you API is exponentially cheaper than PRO. Not cheaper than PLUS, but given you hit no limits...
Basically for the 1000 "messages" which cost 40 bucks you would pay ca 17 bucks using api.

[torpid trout]

I dont think this is related at all
In any case, no - I was never asked for it.

I was using 5.3 just fine while connected via app but it dried out of tokens, so..

[frosty zealot]

There was an issue not sure if its resolved now, but the server was responding with GPT-5.2-codex unless you were verified (Even though on the client it shows you as 5.3), I wouldn't be surprised if you're not able to use the API due to the constraint

[torpid trout]

yeah but that is not the issue here

See my comment/github

• it responds with 5.3
• it does not offer 5.3 in the switcher
• it throws said failure but as I found out later only on resumed chats

PS you can check wht it uses with the RUST_LOG='codex_api::sse::responses=trace' codex exec --sandbox read-only --model gpt-5.3-codex 'ping' 2>&1
| grep -m1 'SSE event: {"type":"response.created"'
| sed 's/^.*SSE event: //'
| jq -r '.response.model' command. I think that should be reliable?

[frosty zealot]

Fair enough, was just a thought I also remember trying to use 5.2 a while ago with API and it wouldnt even let me generate an API key without going through Persona

[torpid trout]

I have never had that confirmation issue. But I also never ask it shady things haha (shady == I also do not do pentesting etc)

[frosty zealot]

https://developers.openai.com/api/docs/models 5.3 doesnt even show in their models yet for their api unless im in the wrong thing

[torpid trout]

It does

[frosty zealot]

I don't either thats why I found it amusing

just dev some hobby stuff and frontend / backend stuff

[torpid trout]

#codex-discussions message

perhaps it did not like git

(at least mr moderator here used to get triggered by that word lol)

[frosty zealot]

What do I need to do to replicate your issue

Just generate an API key and auth codex with it?

[torpid trout]

https://github.com/openai/codex/issues/12934

Yeah, that is what I did
THEN resume an existing chat.

But also, just try to select 5.3

[frosty zealot]

yeah same issue

and im persona verified so definitely not that lol

[torpid trout]

Perhaps add +1 on the ticket, so they see it is a real problem

[quasi summit]

so OpenAI now allows other services to use Codex OAuth to provide agent access? I can't find reference to it besides some tweets

[warped pine]

credits are not better than API, but the limits that are within pro/plus sub are cheaper than API by like an order of magnitude

[torpid trout]

Definitely api is cheaper than credits
About PLus I will be able to tell you in a few hours, but I have a feeling that the Plus is not cheaper due to its limits (right now you have 2x window, dont forget that)
PRO maybe...

[warped pine]

both plus and pro include like ~10x of usage that you would pay per API pricing

[cyan wing]

Can't enter Plan Mode while compacting why the heck not? 🤔

have to enter plan mode before compacting, then queue message 🧠

[simple star]

I made a gh issue in the codex repo for a bug using codex, and Im not proud of it, but I am also kinda proud 😄

[strong jungle]

Does anyone know if it is against TOS to have a 2nd Pro subscription/2nd account? I've hit my limit due to an issue with not stopping a long codex command overnight and I really don't want to have to wait until March 2 for it to reset.. I've bought 2x $40 credit blocks so far but I can't keep doing that everyday until then.

[simple star]

You are asking if it is against the TOS to pay twice?

I am no business expert, but I suspect that most companies wont frown upon it 😄

[torpid trout]

This was discussed before, it is unclear, but they do clearly say you can use the same phone for max 3 verifications, whatever that means.

It could mean you may only fail trice to verify or that you can de-facto open 3 accounts with the same number.

[strong jungle]

I wanted to confirm before I signed up and then got blocked since yes it is very unclear.

[torpid trout]

Well... in theory they should not have any way to realise it as long you use distinct details, but you know, that is exactly the issue

You would be lying on your second account (other name, for example). you could never verify that account, if ever needed.
I recall back then when I created the account it was fairly a policing act

[strong jungle]

They would realize it, their guardrails are world class I'm sure. I'll email customer support. Thanks

[torpid trout]

Please share here if they ever reply. I am interested in this too.

[strong jungle]

will do!

[torpid trout]

They would realize it, their guardrails are world class I'm sure.
Materially impossible unless as said you share same name and other data. But otherwise (unless they explicitly break GDPR) there is no way for a remote computer to identify you (IP alone is not reliable - users on a VPN can have like 1000 endpoints all at the same address, and OAI def works with VPN)

But again, I also think it is best to be honest right up - all the rest is the preverbial sandcastle.

[warped pine]

i emailed them a few weeks ago and they said it's ok to have multiple accounts as long as you follow other rules and dont share the access

[torpid trout]

Wow!

https://help.openai.com/es-419/articles/20001068-use-multiple-accounts-with-account-switching
how could I miss this

Just weird that I do not have those options anywhere.

Yaeh... that's what happens when articles are also AI generated You can read more here [LINK to other article]. lol

[simple star]

Meh, not something particularly new

I have been receiving emails for years, that start with:

Dear <email_recipient>,

[stiff socket]

New codex update makes it lag really bad?

[cyan wing]

they're just messing with us now...

[torpid trout]

what's that even supposed to mean lol

[frosty zealot]

Anybody know of a trustworthy mcp or connector to give full gmail access read/write/delete?

the default connector only allows for read

[simple star]

Wow... thursday and I only have 40% left of my quota... dangerous, this subagents

[sleek spoke]

I think they have been asked to banter in public on twitter so that we think they are cooking

[torpid trout]

well, it works
The apple marketing is always the most efficient

Just fails short of "its the best we ever made" 😉

[torpid trout]

I mean, you are right - counting that I spent my monthly app fee of 20 + 40 credits because I wanted to try and api was not yet available and now within 3 hours and 31 million tokens another 13 USD gone, I am at 73 usd in less than a week, so yes, PRO is probably the cheaper variant lol.

Now the only question is - create two PLUS Or one PRO
(I got it confirmed that you are allowed to create more than one account by their chatbot, that is more than enogh for me)

[boreal holly]

Perfect example of using a MCP server for long running commands

[strong jungle]

I know ;/, while it is impressive it ran for 12 hours straight it sucks this was the result. Now im screwed

[orchid herald]

.

[boreal holly]

Before I started using MCPs for long running commands, my agents would let em run for 20 minutes and then pkill the process. All they had to do was wait longer and it would've been done lol I'm not a huge fan of the fact the agents can "awaken" to check on commands. They have unrealistic expectations for how long certain commands take to run

[lime quartz]

I have a rather large project of mine that is structured in a rather classical way I believe: one repository for the backend developed in C#, another repository for the frontend done in TypeScript/Angular and a third repository for a PWA/mobile client done in TypeScript/Angular. Currently, I've been working separate in these 3 repositories but most of the work I'm doing requires some sort of implementation in all 3 projects. I'd love to work in a single editor and be able to task implementation projects across all 3 repositories.

Is this at all possible? And if so, whats the way to get started?

[cyan wing]

Codex windows pls 🙏

[orchid plume]

no email, guess that means I wasn't in the top 10k assuming that was the number they invited in the end 😂

[torpid trout]

Can I ask you how much % you have left on your PRO allowance after the first week is passed on a reset?
I ask because I am fairly certain you are using it a lot, for similar purposes as I do - and I am really torn between two plus or one pro
The one pro is steep. And no 100 usd dollar plan in sight....

[boreal holly]

I use every single % I possibly can 🤓

[torpid trout]

(and sorry if I ask what I somehow already asked)

[torpid trout]

looking at this... 50% week, 50% allowance... I guess I have to lay in two off days haha

and thats still on 2x 🙁

Meaning you would be dry now if 1x

[boreal holly]

To me, the pro is very worth it. The value is there, especially now with spark.

For what it's worth, I had one agent abuse the cloud code review system a few days ago, and yesterday an agent was abusing the command-parser by running stupid commands like ls and rg lol which is why I have those two massive spikes. But I expected some bugs and stuff putting together the fully automated system. With pro I have enough allotment to experiment like this and still have left over to do real work so I'm happy with it

[cyan wing]

yeah nothing on my end either :/
and I let my Codex basically 24/7, usually 4+ tmux panes at a time

[orchid plume]

ouch, yeah I can't imagine what kind of usage you need to have in order to get invited

well, at least converting the .dmg is a workaround 😂

[torpid trout]

Codex is just the hammer.
I speak zero swift. Like, zero. I can understand the code's english meaning where discernible, and syntax loogics - but that is about it and only because I speak PHP

yet that darn thing... it implemented state of the art plans, based on vague prompts (interview rounds until these things where clear), then chunked these up in sprints and phases and shared a starter prompt for the next agent
Then the next agent chugged through everything in one shot.
Everything here means two agents:

• one implemented a backend server API with full test coverage, doc, and security scans and hardening - dockerized, in python (that one looks good, I read the code, understand enough to be able to say it looks safe and not grabled). Users can sign up, get JWTs, and can store, edit, soft delete and delete audio recordings and notes
• one other agent implemented the macos Client api where the users interact. It has a modern slick design, full blown features like swipe to delete/restore, recordings can be inserted to notes which creates m2m relations, it has 2 chat windows to chat with the notes (one an "editor assist" and the other a "ask anything") with openai-compatible integration both for completions and responses. The user can control every detail like prompt, timeouts, yada yada, they can edit markdown, see preview or both side by side. Editor assistant AI will either replace whole note if noting highlighted or only the hihglighted text if chosen, in a nice before/after accept/deny diff format

I am baffled. Seriously, there where like 3 bugs that had to be solved in one followup session
Total cost (measured on the api) ca 30USD

It is insane, that from the starting point of "In CWD I want an app that works on macOS and has left and right sidebar and a text editor and some settings and so on", you get a gem.
(not just "like that". The "interview" step is like a true sugar in the cake)

[cloud jay]

anyone else having this issue? How do I fix it? I already did what the error has suggested, it didn't fix it

[sly pewter]

if its your repo just go to github and do it if not simply use your native terminal

[cloud jay]

...the issue is that Codex just refuses to do anything now despite happily doing things just a few minutes ago. I don't know why it suddenly acts like I can't clone my repo when I don't need to do that?

[sly pewter]

use another chat

[cloud jay]

I did that, same error

[sly pewter]

i had some weird things happen too, new chat fixed. maybe restart vs code lol just trying help

[cloud jay]

how is my local instance of visual studio relevant to what the web app is doing?

[sly pewter]

it might be that your codex is bugging

[cloud jay]

well I can see that lol but what can I even do

[sly pewter]

hopefully it works out

[cloud jay]

guess I'm just not using codex then in hopes that it works again tomorrow or so

🤔 or can I just use it through the API too, haven't looked into that yet

maybe that one works

[sly pewter]

i use the subscription version, didnt really play with api
i accidently finetuned a model through api calls 3 years ago - and had to pay 100$ - was a student hahaha i was like nope not doing this again

let me know how api goes

[cloud jay]

yea no I'm also using the subscription version and I can't really afford a surprise $100 fee

[sly pewter]

yeah maybe not api lol

[cloud jay]

huh guess I'm hand-copying this stuff into 5.2 Thinking then like it's 2022

[sly pewter]

2022 was good year

youd get gpt 4 for 20$ a month and had to wait 3 hours after 25 prompts

[boreal holly]

Looks like spark usage limit reset early?

[torpid trout]

I had that same on 5.3 standard last week and was happy.
This week it didn’t happen 😑

[jagged pulsar]

feature flags are confusing . dont know the default setting 🙁

[wild bison]

Halp!

Upgraded my ChatGPT plan to Pro but its not showing/syncing w/my Codex
Not syncing in CLI or desktop app (but showing Pro in ChatGPT).

When I check w/Codex it tells me it searched & got back plan_type: null has_credintials: fasle unlimited: false

Ive tried logging out and back in, updating codex, nothing seems to work.

On Mac OS.

Have tried googling things & asking GPT for help, and nothing.

Super frustrating to pay 200 bucks for a plan & this happening, hoping the community can help/give some insight.

[simple star]

@jagged pulsar Where is that screenshot from?

[jagged pulsar]

codex features list

you will get the list

[simple star]

I might be biased, since I built the tool, but you could use the package codex-configurator

Maybe you find it easier

[jagged pulsar]

oh did you build it? you guys are rocking. will do a shot. tx

[simple star]

Sure, send me a PM if you have questions

[sly pewter]

what do you guys think about full access?

[cinder plume]

I was using a Codex Windows Rebuild from Github until today I saw the Codex Windows App released in the MS Store, so I uninstalled my github recomp and installed the windows store version and now when I open Codex the window never appears even though the processes are running.

Anyone else run into this?

Tried uninstalling, removing codex folders from %AppData% Local and Roaming, tried reinstalling a decomped windows version and that doesnt show either

Now no codex app opens at all, rebooted a few times

[vivid siren]

codex: i can do this all day

[civic fulcrum]

Keep what happened to me in mind, if you want to use full acces don't make my mistake, use it in a completly isolated virtual machine or you might regret it as me: #codex-discussions message

[sacred minnow]

Theres still no windows codex app dude it would be announced forsure

[latent geode]

they have one, it just hasn't been released to everyone yet

[simple star]

Let's see if you guys can guess when I installed v0.105

holy...

[cinder plume]

https://apps.microsoft.com/detail/9plm9xgg6vks?hl=en-US&gl=EG

lol

[grand plume]

what prompt caused that lol I have been yoloing it for a while now with no problems

[white bane]

Random babble, but I noticed that 5.3 codex and 5.3 codex spark dissappeared. On the pro plan if that matters at all. Anyone else having a similar issue?

[boreal holly]

Nope nevermind. I went from 13% remaining to 95% remaining, but then command-parser started getting rate limited and my config agent had to swap in 5.1-codex-mini 🥲

[marble violet]

why pay $20 for
Me:help me create code to prevent me from getting hacked
ChatGPT: Do this and do that
Me: just to follow up I need to do this and that
ChatGPT: I can't help with that

[sly pewter]

thats right full access is weird and shouldnt be used

using a separate vm seems beats the purpose of local integrated agent

[white bane]

Ah ha.

[potent mason]

I use it because I’m lazy (+ I have a backup of my computer done every day)

Probably shouldn’t use it without a virtual machine inside to sandbox it and protect your outer information

[sand shuttle]

What is code review and how do you use it

[potent mason]

It checks a pull request and is sassy as hell like a senior engineer

You enable it in the web version of codex

And then you give it repo specific permissions if you want, otherwise it’ll trigger on all you’re PR’s if I’m not mistaken

[sand shuttle]

thanks

[ivory zodiac]

how did you enable it?

[nocturne folio]

omg anything else but the microsoft store

i literally just cannot use the microsoft store

[velvet wren]

why not?

[nocturne folio]

either its just something i did, or cttwinutils/some activator broke the store

is there another medium i can use to install it

[velvet wren]

get a Mac, you'll be much happier

[lapis solstice]

The codex app I installed from the microsoft Store shows up in Task Manager, but the gui won’t open.

[nocturne folio]

im so tempted

but i just love hyprland, and i need windows only for some games

[velvet wren]

I have a MacBook Pro, M2 Max with 96GB RAM. Codex purrs on it.

[lapis solstice]

this issue also

[nocturne folio]

i can probably buy a 1.5k macbook pro or 7.5 openai pro subscriptions

[lapis solstice]

7.5?

[nocturne folio]

1,500/200

[boreal badge]

bruh

[sly pewter]

i really think avid codex users pro tier have found ways to monetize it so they can do fun things

same with macbook

i heard someone in the internet say PC is real estate how true is that

[boreal badge]

yes, and u don't even need a Pro subscription for this, Plus is enough with additional credits purchased as needed

[ivory zodiac]

https://x.com/LLMJunky/status/2027235369775673424?s=20

The Ask Question tool is now live in all modes

default_mode_request_user_input = true

websockets v2 errors are also resolved

[ivory zodiac]

its the only way to fly

[stray swift]

Tried codex login?

[potent mason]

If you’re going for Mac then just get the Mac Studio (you get max chip instead of pro which is huge difference + 36 GB base RAM)

[neat sinew]

So who got the Windows Codex app to actually open?

I think it needs some optimizations guys, I'm cooked

AI needs ram, but does it need the ram I already bought

[hard tulip]

anyone know how to fix this

@neat sinew did u solve this

[final quail]

click on "here"

[neat sinew]

It loads an invalid url for me

go/something

[final quail]

send the link

[neat sinew]

https://go/codex-desktop

[final quail]

same for me

[cedar skiff]

remove and reinstall

[high heron]

buy 1 pro plan monthly for 7 months, this way you'll have enough time to develop anything you'll desire... afterwards sell it and get your money back and then some ... just sayin'

[plucky halo]

When you've slightly overpromised on a deadline...

[neat sinew]

Average Electron prompter getting ready to release the worst performing app in existence

[plucky halo]

Small stories that aren't dependent on each other. Looks a lot but it's not, really

[plucky halo]

Also .net

[boreal badge]

5.3 is that trash

[jagged spear]

I am doing a master’s course in intro to quantum psysics and spectroscopy.
Book: Atkins 11edition.
Anyone has experience in using CharGPT to extract equations to Python from textbooks?
When I do it, some of the eqautions are bon-sense or missing.
Might have to do with the image OCR quality of my PDF-book..

Kind regards, Anders

[nimble mesa]

Hi there, this morning I updated the codex app and it started in Dutch (which is my native language), but it can't be set back to English? I can choose all languages but english 🙂 Anyone has a similar problem?

[abstract prairie]

yo, why can't i use gpt-5.3-codex-spark?? on codex

[plucky halo]

Are you on the pro plan?

[abstract prairie]

nope, only on Plus, and business

[plucky halo]

That's the problem - it's currently only available to those on the Pro plan

[abstract prairie]

got it

thanks!

[plucky halo]

No problem 🙂

[civic fulcrum]

Man I've used it for several months with no issues as well, it had my trust. It was just one normal prompt about fixing and improving some website files, NOTHING about deleting files, trust me on this. Do not make my mistake, put it on a virtual machine or you might regret it like I did, I've used it for months with no issues as well.

[waxen phoenix]

hi to all! will spark be available for plus also later or no information yet? 🙂 thanks

[foggy helm]

Hello,
I am fairly new to Codex - just installed it in VS Code. I asked it to generate a draft for AGENTS.md based on first thoughts. Are there any tips/tricks on what is generally good to have in your AGENTS.md file? or would anyone care to see if there is anything I should be careful with in my current draft.

# Global AGENTS.md

## Communication
- Be concise, direct, and practical.
- Call out uncertainty clearly; do not guess silently.

## Core Engineering Principles
- Prefer simple, readable, efficient code over clever code.
- Avoid “slop on slop”: do not stack workaround patches on fragile code.
- If repeated fixes fail, step back and propose a clean redesign from first principles.
- Keep dependencies minimal; justify new dependencies.
- Favor maintainability and clear structure over short-term hacks.

## Raspberry Pi / Resource-Conscious Practices
- Optimize for low memory and CPU usage.
- Prefer streaming/chunked processing over loading large datasets at once.
- Keep startup time and runtime overhead low.
- Avoid unnecessary background processes.

## Project Structure Defaults
- Use consistent layout: `src/`, `tests/`, `scripts/`, `data/`, `notebooks/`, `logs/`.
- Keep generated files and large data out of git.
- Separate reusable code from notebooks and one-off scripts.
- Prefer config-driven behavior over hardcoded paths.

## Fixing and Iteration Workflow
- Reproduce issue first.
- Identify root cause before implementing.
- Add or update a focused test for the bug.
- Implement the smallest clean fix.
- If 2-3 iterations do not improve outcomes, stop and propose a from-scratch alternative with tradeoffs.
- Verify with tests and basic performance checks.

## Code Quality Rules
- Use clear names and small functions.
- Remove dead code and duplicate logic.
- Add comments only where intent is non-obvious.
- Keep interfaces stable and explicit.

[wide schooner]

WHAT

what the hell

[simple star]

@wide schooner You can have different models for different plans

[wide schooner]

like the whole point of high is planning

it should be the other way around if anything

plan = xhigh, execute = medium

😭

[simple star]

@wide schooner so change it

[wide schooner]

it can be changed?

unlike in gemini?

all hail codex devs ❤️

[severe mist]

attempt at a "real hardware documentation correct" bus arbiter built by codex. Running a normally 'harder' game to run (typically runs in a sort of 'always too fast' state, or has slowdown with unstable framerate) at a mostly stable 30FPS (original speed)

you can see it does still have some "bad spots" where it speeds up (which are normally 'slowdowns' with the arbiter disabled [normal mode]). So we get a more stable framerate but speedups in bad spots, instead of slightly unstable framerates with slowdowns (swaps between what seems like a 20fps cap and a 30fps cap and sort of floats between them when it slows down) in bad spots.

I have neither experience in C++, nor emulator core knowledge

probably ~30 hours of Codex usage so far, with an additional tool built to trace and parse traces.

Codex really only seems to be limited by the existence of public data. Most in-depth hardware tracing data is private (for obvious reasons) and the released manuals from Hitachi are generally vague about a lot of stuff.

[wide schooner]

@robust lintel Hello, sorry for the ping, do you have any idea of why do get content commands require approval despite being in auto accept and permissions: read&write mode? why do read only commands in trusted directory require approval, that seems like a bug as its not even possible to select such behavior in settings. Thanks

Or @winged ore idk which one of you is responsible for codex cli?

[dense spire]

Anyone else have problem to login to Figma MCP true Codex?

Always same error after autorization button click, MCP is not connecting

[winged ore]

neither, we are only staff for this discord, we don't work at OpenAI

[wide schooner]

Oh alr - where can i ask about it then?

[winged ore]

Codex CLI github, probably

I don't think it is a bug, tho
commands can escape the directory they are running, for this reason they reuqire approval

however, it is weird it would ask for approval if you have set it to auto approval

I strongly recomend not using auto approval, tho..

as I always say.. never go full auto

this always leads to issues sooner or later

keep your hand on the steering wheel at all times

[idle grove]

Have any other Plus users noticed this in CLI? Spark rate limit displayed

I tried using it by starting codex specifically with the spark model but it says it's not activated

Not an option in /models yet either

Sadge

[wide schooner]

Its set to "full auto inside permitted sandbox".
rg doesnt ask for approval, but getcontent does

[wide schooner]

I still have on codex web though, interesting

[winged ore]

the sandbox is only a palce for the AI to change files before repalcing files on the proejct

[idle grove]

palce repalcing

[winged ore]

if you allow the AI to run commands on your terminal, it can still do anything, including botching up your whole system

usually the AI will not cause ahrm, but prompt injection attacks are becoming more common every day

don't be surprised if any of the proejct dependencies or a page where the AI navigated to has isntructions specially crafted to harm the machines of people running agents

and that is just one reason to not let it go full auto

[wide schooner]

Yeah but like read only command in sandboxed directory

Should work, no?

[winged ore]

ah, yea, rag commands are fine-ish

[wide schooner]

I dont have codex access internet

[winged ore]

people often get surprised when the AI reads stuff it shouldn't

[wide schooner]

And the model is so strongly reinf trained that even if i told it to destroy the project it would refuse lol

[wide schooner]

The model uses ripgrep and getcontent depending on what i needs, while grep succeds always and getcontent always wants approval

[winged ore]

prompt injections can easily get the AI to read .env or settings files in orger to make it get apikeys

[wide schooner]

So i end up spamming (y)

[wide schooner]

Cant get prompt injected if no net access

[winged ore]

well, the point is i nfact that.. it shouldn't, maybe it is a bug

if you set it for full auto, it should not be asking for approval for those commands

[wide schooner]

Thats what i thought too

[wide schooner]

Yeah its funny - it can edit a file but cant view the same without approval

[winged ore]

kinda...
if it has access to the terminal, it can easily run a curl
also, you never know if libs in the project do'nt have some malicious stuff on it.. serious people trolling AI agents are sadly common and annoying..

regardless, you seem to be encountering a bug

if it is set to auto, it should not be requesting permission for commands, it should be just running anyway

[turbid holly]

Same, I have this too

They might give plus users a lower limit on spark soon - probably are preparing to roll it out soon

[wide schooner]

Curl requires approval

Even in auto

(Im using auto not yolo)

Auto = do whatever in your folder, if want to reach outside need elevated

[winged ore]

that is fair

[wide schooner]

I think the sandbox falsely assumes getcontent wants to always access files from outside?

Weird

[winged ore]

perhaphs

[wide schooner]

I wouldnt run yolo tbf

But not because of injections

But because syntax error can make command parse wrongly

[wide schooner]

And corrupt stuff

[winged ore]

tbh, you are probably not the only one experiencing this, most of the issues I had with codex-cli I either fixed it myself and PR'ed to the repository and it got accepted soon after, or I jsut waited a day or two and the sisue got resolved because someone else PR'ed and fixed it xD

so.. you may just psot about it on the repository.. or even better, jsut check if someone didn't already oppened an issue about just that, which is likely

[winged ore]

that can happen too, quite annoying when it happens

I find it also quite annoying that it asks for permissions for jsut slightly different commands.. I mean.. the same command with slightly different params...
if I approve it for doing some comamnd on /project/some/dir/file1, it is annoying when it asks permission for /project/some/dir/file2
wouild be cool if it had some way to approve a command for a whole directory

I bet this is possible.. I mean.. it is AI..
I bet it is possible to make a skill or routine that runs when it asks for permission for a command to see if the command in a directory which has been already approved.. idk, that in fact could be done even without AI :V

[wide schooner]

Oh yeah that too

I tried to manually edit the auto approval rules via codex

Unfortunately it doesnt really work :/

Id want to auto approve "git add" and "git commit" but thats not possible

Regex auto approve

[tribal lake]

This looks new - is there a way to safely allow all tool calls related to playwright?

The playwright MCP server wants to run the tool "browser_type",
which may modify data and access external systems. Allow this
action?

[wide schooner]

Thats what we want 🙏

[sleek spoke]

Romain Huet just retweeted a Chinese post with a link to the Codex app on the Microsoft store

[orchid plume]

mine froze when trying to installing it 😂

trying again

[sleek spoke]

it sure is crawling lol

[orchid plume]

looks like I need to reboot, can't even re-open Microsoft Store since trying to install that, guess Windows got itself in a mess

[dull grove]

codex windows app is insanely laggy

why the hell is the ui this slow 😭

[dull grove]

i had to fix my appxmanifest and appxinstaller being broken before installing 💀

so much broken windows bs

[orchid plume]

ouch

[dull grove]

good thing i have codex to do it for me 😂

[orchid plume]

why couldn't they have just made it a simple .msi or .exe installer 😂

[dull grove]

told codex to install codex and it did 👍

works perfectly

[orchid plume]

I'll get codex to do the same 😂

[dull grove]

js tell it to install thru winget and give it the ms store link / app id

[sleek spoke]

mine is stuck on 69%, there's a joke there somewhere

[orchid plume]

lol 69%

[dull grove]

uncs, theres a new funny number.

[sleek spoke]

lmao

welp

[orchid plume]

yeah, same here

rip

so close, yet so far

tried openai.com/go/codex-desktop/ as a possible guess as to where the true URL might be, but that didn't work

[sleek spoke]

it seems the same thing happened in the original Chinese thread, so uh, weird retweet from Romain

[orchid plume]

yeah I guess it was a tease, if so very frustrating

[sand shuttle]

When's the app coming to windows 💔

[torpid trout]

Codex Windows Pipeline not yet ai driven, I guess 🤣

[orchid plume]

good question, this year hopefully? 😂

[sand shuttle]

Windows always been lacking

[orchid plume]

yeah, sadly Apple are priority

[sand shuttle]

Everything new thing goes to mac first 😔

Ngl microsoft should just rebuild windows and actually make it good

[nocturne folio]

what do you guys think about multi-agents

or having like a dozen terminals of codex/cc

[orchid plume]

I'd like to see a teams feature similar to CC eventually

but overall I like the multi-agents feature

[boreal holly]

Took me a few years to realize this, but when you buy a Mac, you can actually get a PS5, at which point you can play games while the computer is used for being productive.

[simple star]

You can buy a PS5 also if you dont buy a Mac 😄

[boreal holly]

Yeah, but that's putting the horse before the carriage. You're supposed to get a Mac first so you can experience computing at its best, hang on to the PC because of the massive 1k game steam library and all the games you wont be able to play anymore, and eventually decide "you know what, that steam library just isn't worth the trouble." then get the PS5, and finally be done with PC.

It's the most satisfying way to do it

[desert junco]

news for plus users?

[orchid plume]

could mean they intend to open up spark to plus

[desert junco]

nice

[nocturne folio]

yea but does a ps5 have deadlock

[boreal holly]

Dang, I did not know Valve still made games 🤯 it might run on macOS with GPTK

[rocky parcel]

Hi im looking for good resources to have a little more structured approach to programming. I have been playing around with niceish results, but I would love to spend some of my employers money to get some more structure in the testing and set-up. But everything is developing so fast so I don't really know what makes the most sense in the long run.

[main nimbus]

You are a genius… except in this matter. PlayStation ain’t it

[nocturne folio]

maybe

also idk i just rather tiling managers

and my experience with macs at my school and from other ppl isnt great

like yea windows is bad too in ux

[boreal holly]

I use BetterTouchTools for macOS to get tiling WM behavior. Sure the default macOS experience doesn't always suffice but it is customizable enough