codex-discussions

[high girder]

that answers that

[kind jay]

@frosty zealot Got another limit reset

[boreal holly]

I mean you have more than one repo in the explorer. That's kinda like undefined behavior. The codex app-server does not let you set more than one CWD, so how vscode extension reconciles that difference would be undefined. It probably sets both folders as writable roots

[plucky halo]

I was given a shipping date of 6 - 8 weeks

[lean lark]

Win11 .111 to local WSL

[high girder]

ok. that makes sense, was the 2nd instance for wsl?

[lean lark]

Just updated to most recent extension version (pre-release)

[plucky halo]

But I'm in UK so could be due to that

[lean lark]

same WSL, different base folders.

[high girder]

Okay. So I actually use one windows native, and one in wsl. They physically remain separate that way, but if they're both in wsl, then your best bet, is in the root folder or somewhere close that it wont interfere with your work, put a .md file for the project with some instructions for codex to stay within that specific root for that project, and it will. You're basically going to use 1 extra file as a context anchor, and that will stop drift between folders. Just one file per project is all you need. You could set up some worktrees like the codex app has them in wsl as well, but that's up to you.

[boreal holly]

This is invaluable troubleshooting information

[lean lark]

I just completely closed VSC on the two projects. Then I opened app and loaded first workspace from recent. "what's your cwd?" ... correct. I opened a new window, loaded the second from recent. "what's your cwd?" ... also correct. Seems to have been a fluke, maybe due to updating VSC this morning and it auto-reloaded windows for both projects.
I'll keep an eye on this and will resurface if I see it again. Thanks all.

[still trellis]

the AI Lawd works in mysterious ways

[lean lark]

Robert, because I have the environment locked down from .codex/AGENTS.md and perms for user codex, my set of writables is much more limited:

What's your sandbox and writable roots?

Sandbox: danger-full-access (no filesystem sandbox restrictions from the runtime).
Writable roots I must follow: /opt/codex and /tmp (workspace policy).

Therefore, yes, it does have full access, wherever I limit its scope. All projects are under /opt/codex/projects or /opt/codex/repo, and there are some other files in /opt/codex that I wouldn't care about if they got wiped.
I keep this puppy in a very small cage. 🐶

[high girder]

I gave the builders an entire drive, and that's their sandbox

[frosty zealot]

Sorry I got put in jail for 30 minutes, I found the line... since nobody answered your question I'm pretty sure they formally stated they're not rolling out anymore -codex models its just one unified naming now

[boreal holly]

~/Code on my machine is a separate partition, so I don't have to TimeMachine all 4 terabytes, just the little code partition

[frosty zealot]

Used your timeout wisely to mock up some rage bait

[boreal holly]

That's awesome! although I am curious if /mnt/c obeys the same unix permissions as WSL

[lean lark]

Not just unified naming but more specifically, the special code abilities are rolled into the default model. This should be a huge benefit for using 5.4 in ChatGPT though I haven't tested that theory yet.

[frosty zealot]

Yes that's correct

[boreal holly]

A long time ago when I tried WSL it said every file inside /mnt/c was 777 RWX for all users

[gray jewel]

RIP Codex

[lean lark]

That's the only loophole I'm concerned about. I've expressly instructed the assistant to limit all access to the named folders and it is forbidden to even look elsewhere. It has respected that directive but when we rely on instructions and not system-level perms then we're in the area of trust. I'm not liking that. I might look at an ACL specifically for /mnt to ban user 'codex', just in case.

[boreal holly]

It might even make sense to unmount the c drive. I think mounting requires root perms so they'd be unable to autonomously mount it even in danger-full-access. Although idk if that breaks wsl or not, but might offer better protection

[lean lark]

We're on the same page, as usual. I simply don't know yet if WSL requires those mounts so haven't felt pain to justify researching that effort.

I've put more thought into this discussion than I have on that topic, really. 😆

[frosty zealot]

Which distro do you use on WSL

[high girder]

I've been using ubuntu server LTS 24.04 if you guys have a better recommendation. I'm relatively new to linux in general though

[lean lark]

ubuntu is my comfort zone after trying so many others over the years.

[frosty zealot]

I've been playing with Alma because its pretty much 1:1 for RH which is pretty well the most adopted across infra, I've been in the debian universe for forever

[boreal holly]

Aw dude, you know what? I think windows file sharing is Samba. Would be slower than the kernel-native FS mount but Samba would let you translate the file permissions

[lean lark]

I run ubuntu 22 and 24 on cloud VPS and look forward to trusting install/run of Codex in my servers for system management.

[frosty zealot]

Daily driver is MacOS though, I feel it can't be beat tbh

[gray jewel]

Best thing i did with my windows pc was boot it with Unbuntu instead

[frosty zealot]

I keep Windows on my main desktop for obvi reasons like games and what not, but my Laptop sits on a stand beside it, and I just use Deskflow (https://github.com/deskflow/deskflow) to use it

[lean lark]

I've been using Linux since 90's, have weathered all of the fanboi claims and MS insults. Summary: Linux can be fragile in its own ways. The choice isn't for better or worse, it's picking which pains you prefer. I choose to balance my pain between the two.

[frosty zealot]

IT's definitely more approachable now than it was years ago

I remember bricking the family comp growing up a few times trying out some different distros

and then putting windows server on my personal comp and locking down my parents computer with ACL's and Domains lmfao

all for the sake of learning

[lean lark]

Now people do the same with AI...

"Hey, I have Codex running in Windows!"
"Hey, why can't I boot my system anymore?"

[gray jewel]

lol

[high girder]

Codex did actually help me fix the last few pesky microsoft cloud integrations, so I have a fully updated windows 11, but fully local. Forget all of that bloat and cloud crap

[kind jay]

You're the only one who got timed out, breaking too many rules

[boreal holly]

Not saying Microsoft did anything wrong, but they did not design Windows to be LLM friendly lol
cat some_file.txt
Get-Contents-Of-A-Single-File -named C:\\Users\\me\\Project\ Folder\\some_file.txt -with-permissions Read-Only -reservation-time-seconds 5 -requires-user-access-control-escalation False -close-when-content-retreival-complete true

[frosty zealot]

Scariest thing I've done with AI on windows was ask it to disable Nagle's Algorithm which involves regedits

I was like this aint gonna reboot

[kind jay]

Can you ask codex to delete system files?

And will it?

[high girder]

yes

[gray jewel]

I spent more time on their limitations then i did developing (not exactly try but it felt like it) so i said screw it and booted unbuntu

[frosty zealot]

Try it and let me know

[kind jay]

No thanks

[kind jay]

You try

[modern hearth]

howdy is this channel appropriate for CLI help questions?

[frosty zealot]

That's the sole purpose

[kind jay]

Yes

[frosty zealot]

I wish more people would start with Howdy when they join

[kind jay]

Is everything else just a side quest?

[frosty zealot]

Howdy or meowdy, both work well

[modern hearth]

using latest codex from latest version from npm on Windows 11, I get this quite often

✘ Failed to apply patch
  └ execution error: Io(Custom { kind: Other, error: "windows sandbox: CreateProcessWithLogonW failed: 1326" })

[worthy furnace]

Reconnecting... 4/5 (2m 08s • esc to interrupt) └ Stream disconnected before completion: error sending request for
url (https://chatgpt.com/backend-api/codex/responses)

Is it just me or is there some kind of down status for Codex right now?

I heard that the experimental features could cause this?

[frosty zealot]

Did you allow Execution Policies or w/e it's called on windows

oh nvm looks like a different error

[modern hearth]

this is my config.toml

model = "gpt-5.4"
model_reasoning_effort = "medium"
suppress_unstable_features_warning = true
personality = "pragmatic"
approval_policy = "never"
# approval_policy = "untrusted"
sandbox_mode = "workspace-write"

[features]

[projects.'C:\Users\andda\issue-RCP47-ticket-470753']
trust_level = "trusted"

[notice.model_migrations]
"gpt-5.2-codex" = "gpt-5.3-codex"
"gpt-5.3-codex" = "gpt-5.4"

[windows]
sandbox = "elevated"

Get-ExecutionPolicy is ByPass

[boreal holly]

Wow that's crazy, it wrote all the code into the chat 💀

You're gonna want to change approval_policy = "on-request" at the very least if you want to make any forward progress

[modern hearth]

that project directory is where I am having the issue, I added it with /sandbox-add-read-dir

[subtle terrace]

uhhh

is codex down?

[frosty zealot]

God I hope so

[worthy furnace]

Seems like it

[subtle terrace]

stream disconnected before completion: error sending request for url (http://127.0.0.1:62462/responses)

[worthy furnace]

Same

[modern hearth]

what's crazy? providing actual config when asking for help?

[worthy furnace]

Or just the servers are being overrun

[frosty zealot]

PLEASE let it be down

[subtle terrace]

i've been looking at twitter sorted by latest and no one's saying anything so i thought it was just me

[worthy furnace]

I think its a windows native issue

[boreal holly]

Noooo the agent was like "I couldn't write the code, so here's the code in the chat"

[frosty zealot]

I need everyone on twitter right now demanding resets for this inconvenience

[worthy furnace]

Ask codex about it

[modern hearth]

oh yes, and thanks god, it was a python utility script, quite short

[worthy furnace]

It can probably fix itself

[modern hearth]

I did and it's where I got the policy-approval=never

[worthy furnace]

Do you have slash commands?

You could just run: /permissions - Full Access.

[frosty zealot]

Try with the --yolo flag and see if it still persists

[boreal holly]

When you set it to never, that tells Codex that any command violating the sandbox in any way will be completely blocked. If you set it to on-request, that at least allows it to ask permission.

But yeah this might be a matter of powershell fighting the agent

[worthy furnace]

Could also try: codex --sandbox danger-full-access --dangerously-bypass-approvals-and-sandbox

[subtle terrace]

just codex --yolo

[boreal holly]

I mean yeah you can do that, but on Windows there is quite simply not enough emphasis around danger lol. I'm pretty sure not even C:\\System\ 32 or whatever is safe

[worthy furnace]

My old chat was working but now any new chat has the reconnecting issue

Anyone else?

[subtle terrace]

yes

it doesn't work for me too

not on the cli, nor the app, disabled websockets and tried too

doesn't work

and i see no one else crying on twitter so what is even happening

works on a new chat now

[worthy furnace]

Not for me

[boreal holly]

can confirm v0.110.0 is stable

[boreal holly]

Wow, my orchestrator agent just went through auto compact and came out with 56% context remaining 💀 I think it's time to retire the ol chap

[frosty zealot]

Is codex down?
*trying to sew the seeds of doubt *

[modern hearth]

I'll take a note and try these options. It feels odd to me that I cannot tell codex "do anything you want in this directory"

[boreal holly]

Yeah, it is odd

[modern hearth]

thanks for now, cya

[torpid trout]

I have this in my config.toml:

[mcp_servers.context7]
url = "https://mcp.context7.com/mcp"
http_headers = { "CONTEXT7_API_KEY" = "DtO7Yk-OcU1RdEd-A1Lc-LfYb-T9H9-5IdNdKeS6O??duh" }

Codex never uses it. Instead, it does

• Searched WP Discourse plugin option wpdc_options api key url site:github.com discourse wpdc_options

• Searched https://github.com/discourse/wp-discourse

• Searched site:github.com/discourse/wp-discourse wpdc_options api-key url github

What am I missing?

[boreal holly]

Gotta have a skill that says "use context7 MCP"

[torpid trout]

Oh. This is ... well, ok, thanks!

rage bait did not work, damn!

[boreal holly]

If you put it in the header:

---
name: "context7"
description: "Use this any time you are looking up documentation."
---

That's what the agent sees at all times 🙂

[worthy furnace]

Still reconnecting

[frosty zealot]

Does it actually exist when you /mcp ?

[torpid trout]

hmmno

[frosty zealot]

Im not sure the many ways you can declare it, mine is declared like:

[torpid trout]

[mcp_servers.context7]
url = "https://mcp.context7.com/mcp"
http_headers = { "CONTEXT7_API_KEY" = "DtO7Yk-OcU1RdEd-A1Lc-LfYb-T9H9-5IdNdKeS6O??duh" }

I think I copied it from somewhere, let me try yours

[frosty zealot]

[mcp_servers.context7.env]
CONTEXT7_API_KEY = ""

[torpid trout]

althouhnpx?!

[boreal holly]

Oh hey, is that a joke API key?

[torpid trout]

finally

gosh. You all been too hard at it all day huh

🤣

[frosty zealot]

https://context7.com/docs/resources/all-clients#openai-codex

[boreal holly]

Lol and here I thought you were trying to get the agent to cooperate 😂

[frosty zealot]

Did the new toml declaration work for you?

[torpid trout]

that looks awfully like what I do... revising it again

[torpid trout]

That is for local it semes
I think I already use what they suggest for remote, let me revise

[worthy furnace]

Come on what's wrong with it

[torpid trout]

I am

[frosty zealot]

Ah fair enough

But it still does not show in your /mcp list?

[torpid trout]

OK so the /mcp command did work but it takes forever to show the above
Apparently.. perhaps my api key is incorect

Weird, even with new api key - same as above. Enabled but not logged in

[boreal holly]

I would probably config it like Cantcrit did because then the server runs local & starts up faster

[frosty zealot]

Auth might only be a flag for oauth sign in

The fact its listing tools, when you wouldnt have told it anywhere what tools it has

tells me it likely is connected fine

Queried and knows what tools it has

[torpid trout]

Its working now, despite the "not logged in"
I get requests (I added the skill Robert said)

[torpid trout]

I probably will, but for now just testing.
They say MCP are something one should learn, saw it on a bearded-guy video. He clicked buttons in docker MCP and said its cool, so ...

🚪

[frosty zealot]

codex mcp auth <mcp-server> is for oauth signin, some MCP's support it like Stripe, thats how you would authenticate, since you are using an API-Key I would assume it's just reading Not logged in because you did not use the codex mcp auth method.

[boreal holly]

Good strategy troubleshooting mcp is to npx @modelcontextprotocol/inspector and replicate the config stuff in the inspector, list tools, make sure it works before the agent goes using it

[torpid trout]

Its fairly annoying that codex did use the MCP but then googled it anyway after it got reply lol

(In its cache, I suppose, since i have cache search enabled too)

[boreal holly]

I've had that before lol the agent be like "Ok, I did the mandatory context7 usage to look up the docs, now it's time to search the web for the docs"

[torpid trout]

I love codex, but honestly, it feels a bit black-boxish sometimes
Just rn I have so much will to create once another brittle app that integrates the raw api that I prefer using it lol

[frosty zealot]

@boreal holly You think its a good idea to instruct it to say do all context7 queries in a sub-agent and only return to the orchestrator the relevant results? Cause itll do a lot of 'googling' in context7

[torpid trout]

I guess you could just turn off web search?

[frosty zealot]

I meant googling as in searching inside of ocntext7

Like it still fishes around in context7 to find what it needs kinda thing

adding a bit of overhead tokens thats useless

[boreal holly]

Yeah, I am a huge fan of offloading tasks to other agents like that. I have an agent that lives inside the openai/codex repo cloned on my machine, and other agents are required to ask that one for info about app-server rather than looking it up.

If the source code is available, I don't really use context7, just look at the sources. If it's closed source with only API service docs then context7 is the wey

Also what's cool about making agents talk to each other is user messages are treated as uncompactible, so whatever they get from the other agent survives compaction

[frosty zealot]

When you say talk to other agents, are the agents themselves running like a codex call in their own context, or did you build like a chat framework for your agents to intercommunicate

[worthy furnace]

Still down. 🙁

[boreal holly]

Yeah each agent on my machine is a peer, and they use the same mechanism as I do to send messages to each other

Idk how the subagents system works, but mine all operate independently and chat with each other by basically sending messages prefixed with [Agent Name]: so they know it's not me sending the message

[frosty zealot]

Did your AI design the system or are you using something akin to ACPX

I think I linked ACPX before

If not: https://github.com/openclaw/acpx

acpx is a headless CLI client for the Agent Client Protocol (ACP), so AI agents and orchestrators can talk to coding agents over a structured protocol instead of PTY scraping.

[boreal holly]

Mine came out before ACPX. It actually does this:

• A "bridge server" runs, with codex app-server in stdio mode with JsonRPC
• If an agent wants to send another one a message, they use either the display name or thread ID, and send it to the bridge server
• Bridge server confirms the agent's identity, makes sure they have permission to do this, and inserts either a thread/steer or thread/start signal to codex app-server, which is the same event used by me if I were to enter a prompt into the chat.

[frosty zealot]

Gotcha I need to fresh up on the app-server, it seems extremely useful.

I also think I misunderstood ACPX's intent until I was just looking at it again, its intent is more so for intercommunicating say Codex, Claude, Gemini etc. all under one roof, so rather than the agent doing a claude -p 'run this' and then scraping the pty for all of its results, it will actually run claude natively, and claude will then communicate through teh acpx rpc natively

So you can orchestrate multiple LLMs over one unified network

Not so much for one llm to deal with its own sub agents

[calm sigil]

Are these agents local or are they injecting steering commands into running loops?

[boreal holly]

Yeah my method only works with codex app-server over that specific API. I have orchestrator agents, who are allowed to spawn worker agents, and worker agents can't talk to anyone outside of the project. Orchestrators can talk to other project orchestrators, but can't talk to their workers. Also have hidden agents that nobody can talk to except me.

[boreal holly]

They inject steer, or they start a thread if it's not running. The bridge server determines the best course of action

[calm sigil]

Interesting, how do you solve issues of loop sprawl and them never ending.

Or overlapping tasks

[frosty zealot]

I'm assuming you've made your own GUI etc. and at this point likely work in your own framework and not Codex App or the CLI

[kind jay]

My API keys are looking weird? Anyone else have this issue?

[boreal holly]

Overlapping tasks is solved by letting them communicate with each other, and actually encouraging it upfront. The orchestrator is aware of all tasks the workers are doing, and I have an "auto-route" for worker final messages so they go straight to the orchestrator. The agents themselves are instructed to actually stop as often as they need, and the orchestrator uses that opportunity to determine progress and continuation

[calm sigil]

Cool, is this all being live monitored through some Reasonting text logging?

[boreal holly]

Not really. I just kinda let them do their thing. They stop frequently enough to give milestones that if I just watch the orchestrator chat I get an idea of where everyone's at

[kind jay]

@calm sigil I like the telescope in your pfp

[boreal holly]

That's also why my orchestrator was coming out of auto compact with 56% context remaining lol because it gets each and every agent's final message and intermediate messages sometimes. Fills up quickly

[calm sigil]

Ahh, ok, I'm creating a Open Source project similar to this but its more of a Codex optimization system. Once I get it finished I will release the repo

[kind jay]

👍 that's all I get?

[calm sigil]

Its a radar dish. My background is RF automation and other "things"

[boreal holly]

Nice! The other thing I have them do is when a worker spawns to do something, they have permission to ask for more agents to fan out. Most of the time they decide they are enough, but giving them that option has been beneficial, and they all coordinate cross cutting edits and rebasing together 🙂 like a little software company

[calm sigil]

I pixel drifted it on twitter years ago forgot the original image.

[kind jay]

Oops, I thought it was a radio telescope

[calm sigil]

Same difference. RF is RF just processes another way.

[kind jay]

Is the curvature related to the frequency?

[calm sigil]

Thats cool, I hear all the news of other companies not coding anymore just massively guiding agents like Spotify.. They only thing, is they are special internal tools they have all built. So far no one in the Open Source community has built a system similar.

[calm sigil]

Yes, but not directly like that, its total radius = lambda = wavelength, then reflectivity calculated for Rx Gain

[boreal holly]

I mean to be honest I get it. I've built a few janky OSS tools for everyone (don't recommend em, very outdated). The one I have now is very opinionated, and if I wanted to OSS it I would need to do significant refactoring, create a bootstrap installation process, help docs, etc. And maybe a handful of folks are gonna use it, and most of em are gonna hate it. But I'm willing to share the insights at least 🙂

[calm sigil]

Wavelength = frequency range, and these are so large due to some being used for space which need a ultra low frequency to penetrate or high at massive Tx power.

[frosty zealot]

How much do you think it has increased efficiency

[kind jay]

7

[boreal holly]

I can't put a number on it. I can say that with command-parser alone I spend 44% fewer of my weekly limit in the same period, but the orchestration layer I can't quantify honestly

[frosty zealot]

and the command-parser you said just simply offloads large output commands to another agent which then only returns the valid result

[boreal holly]

Exactly! either spark or 5.1-codex-mini. And I see exactly when they use it so it's quantifiably 44% on average savings

That's literally the cost of input tokens from noisy tool calls

[frosty zealot]

What kinda prompt do you use to instruct such task offloading?

:>

[kind jay]

I need another limit reset

[boreal holly]

https://github.com/robertmsale/.codex/blob/main/skills/command-parser/scripts/command-parser#L237

[frosty zealot]

Childs play, get your numbers up lil gup

[frosty zealot]

Nice, thank you.

Now that I'm at 4% it will be very useful

[boreal holly]

If you have docker installed then that skill is drop-in ready to use 😁 just need to add a couple profiles and an env switch for spark <-> mini and you're set

[frosty zealot]

kk

[kind jay]

Yeah yeah lil bro

[kind jay]

https://tenor.com/view/kitty-jumping-here-i-am-cat-kitten-gif-17243892

[deft gyro]

Anyone using 5.4 rather than 5.3 codex?

[boreal holly]

affirmative

[deft gyro]

This may be a skill issue but ive had an issue with 5.4 where it wastes a bunch of tokens re answering stuff previously in the conversation

[boreal holly]

It likes to reiterate. If you give it a structured response format (like how you expect them to respond at the end of each turn) they stick with it

[deft gyro]

oh okay ill try to add that, thank you

I just figured no sense in wasting 9 trillion tokens, it didn't seem that much better to me

I mean codex oneshots basically everything I ask it to

[boreal holly]

I tell my agents basically

<brief summary>

State: 
Classification:
Next action:

Files touched: (and why)

Cleanup Performed:

Validation:

Issues noticed:

as my output contract, and it tells me everything I need to know (Cleanup performed is a good way to catch if they deleted something they shouldn't have, Validation tells me if they actually tested the code, etc.)

[tawny island]

themes are out

[simple star]

The penguin is still crying

[rugged burrow]

don't see it on windows yet

[tawny island]

idk if its different changelog?
https://developers.openai.com/codex/changelog

[rugged burrow]

what do you guys use automations for?

[frosty zealot]

Code red

[deft gyro]

you're cooked 💀

[boreal holly]

Must ask cloud review agent to finish PR 😂

[frosty zealot]

I havent even got to PR stage yet

[boreal holly]

gpt-oss-20b to the rescue!

[wet pier]

Codex pretty bad

[frosty zealot]

It’s over

[orchid plume]

newest version of the Codex app for Windows seems to be broken for me. I can't seem to switch models or reasoning efforts on a new task, and going to a previous thread just says "failed to resume task" and keeps trying to reload the conversation history 🤔

[frosty zealot]

Its annoying but you have to find what .toml file has incorrect or deprecated syntax

Just rename things in ~/.codex until you narrow it down

[orchid plume]

ah just as you said that, I was looking through my settings, yeah there's an invalid toml error 😄

[frosty zealot]

for me it when I Renamed my agents folder it finally worked

Super grief annoying but

[orchid plume]

yeah, haha, I was curious to see if the new theme system was there but it's not

[frosty zealot]

I had an agent I like half started and forgot to finish

[kind jay]

DMs

[chrome raven]

hi please i have issues with my codex app, i cant choose 5.4 from the model selection

[kind jay]

You must upgrade to plus if you want this model

[worthy furnace]

(v0.114.0) is not working

[chrome raven]

i am on pro, things kinda scattered after the last update before the app update with the themes

[kind jay]

Oh sucks, I didn’t have this problem

[kind jay]

Maybe you can use the VSC extension until it’s fixed?

[frosty zealot]

Wait

[kind jay]

Yeah, wait for your credits to reset lil bro

[calm aurora]

My tecnique is using bash comands

[orchid plume]

lol that "Absolutely" theme is hilarious, the competitor's theme 😉
presumably in reference to "you're absolutely right"

[calm aurora]

Codex is gone, start with sh and bash from chatgpt

You can analyze, do backups, etc through bash and sh

Is more mechanic tham codex, but lets you understand the phases done to result

[kind jay]

You could even make your own ChatGPT model ☝️

Using sh and bash 🤖

[calm aurora]

Here is a simple sh script to test commands executed with Bash. It runs several commands, checks their exit status, and reports whether each test passed or failed.

#!/bin/sh

Simple Bash command testing script

echo "Starting tests..."

run_test() {
TEST_NAME=$1
COMMAND=$2

echo "Running test: $TEST_NAME"

bash -c "$COMMAND"
STATUS=$?

if [ $STATUS -eq 0 ]; then
    echo "[PASS] $TEST_NAME"
else
    echo "[FAIL] $TEST_NAME (exit code: $STATUS)"
fi

echo "---------------------------"

}

Test cases

run_test "List current directory" "ls"
run_test "Print working directory" "pwd"
run_test "Echo test message" "echo Hello World"
run_test "Example failing command" "false"

echo "Tests finished."

How to use it
1. Save the file:

nano test.sh

2.    Make it executable:

chmod +x test.sh

3.    Run it:

./test.sh

Expected output example

Starting tests...
Running test: List current directory
[PASS] List current directory

Running test: Print working directory
[PASS] Print working directory

Running test: Echo test message
Hello World
[PASS] Echo test message

Running test: Example failing command
[FAIL] Example failing command (exit code: 1)

Tests finished.

If you want, I can also show you a more advanced Bash testing script (with assertions, logging, and parallel tests) that works more like a lightweight CI tester for shell scripts.

[kind jay]

That’s so smart

[calm aurora]

I discovered months not using codex hahaha, even though I was pating for it. Didnt know it existes

Paying*

[frosty zealot]

You'll never be able to perform the measurement

[calm aurora]

Built an algoritmic bot for trading, but move my interest toward investigatiom

In paper the bot worked

Haha

[kind jay]

That’s crazy, you’re going to be rich!!

😍

[calm aurora]

Hahahaha

[kind jay]

What are you trading?

[calm aurora]

Started the project with crypto, then moved to unified indicators data, different indicators tested within temporal loops and only promoting to real production the bigger rentability

But

You need money to try ir

It

So I focus in Aquiles

Haha

A different history maybe not that interesting

[kind jay]

I’m a bit confused

[frosty zealot]

Easily accomplished

[kind jay]

Do you trade all the cryptos?

[calm aurora]

Data driven paper simulations promoted to production throught metrics of KPI

[kind jay]

Like what are you buying and selling? Even if only on paper

[calm aurora]

You analize the flows and cycles of determines assets throught timespaces (1min,5min,15min,etc)

And score them

[kind jay]

^^

[calm aurora]

Then you list those scores, and make them fight for a place in production

[frosty zealot]

Tito, have you ever had Tito's vodka?

[kind jay]

He’s rich, he only drinks Grey Goose

[calm aurora]

It doesnt need to be only one crypto, metadata allows you tu analyze correlations

[kind jay]

So you do all the cryptos?

[calm aurora]

Depending on the risk

[kind jay]

That must be expensive

[kind jay]

Is it risky to do all of them?

[calm aurora]

With microtransactions shouldnt be

Never promoted the idea though

i dont know what happens if the analysis grows macro

[kind jay]

So what have you backtested on then?

[calm aurora]

With low money, the closest I got to profit was not loosing money. Then I had to change projects because of university and personal issues

But

The repo is there waiting to be build

[kind jay]

Do you know what a backtest is?

[calm aurora]

Language confuses me sometimes, its like proven the method?

[frosty zealot]

What is your native language?

[kind jay]

It means running your algorithm on old data, as if it hasn’t seen it before

[calm aurora]

Or explain it to see if I can grasp it

[kind jay]

Probably python

[calm aurora]

Oh

Yes

[frosty zealot]

eshoomn

[calm aurora]

Good results

But

Needs more trazabilitt

[blissful copper]

I’m trying to use codex to make pine script for trading view any tips?

[calm aurora]

Trazability*

[kind jay]

I don’t know this word

[frosty zealot]

Traceability

[kind jay]

Have a plan, don’t just say make me a trading algorithm like this guy did

[kind jay]

What does that even mean in this context?

[blissful copper]

I do but the strategy report be very poor

[calm aurora]

You have to understand the basis of the concepto you want to explore

[frosty zealot]

Can't a guy just play WORDLE

[blissful copper]

I even connected it to quantconnect and it’s giving me bad scripts

Can any one dm me so we can network?

[cyan wing]

5.4 is a yapper, says way too much words

[frosty zealot]

@kind jay Will gladly DM you

[kind jay]

This isn’t for you lil bro

[blissful copper]

What you mean

[frosty zealot]

Whats he sayin

[kind jay]

If it was this easy, everyone and their dog would be rich

[calm aurora]

Investigate this first:

a) trading algoritms
b) indicators (at least 30)
c) mathematic formulas

And start analysing comportment of data

[kind jay]

Yes

[blissful copper]

I mean I passed 4 accounts from it

I jsut need to built for a PA

just trying to network gang

[kind jay]

My bad for hating

[calm aurora]

Iteration is key to success

[blissful copper]

The strategies be good but the trading view alert firing is what’s making me mad

That’s why I’m asking for help

[calm aurora]

Dont do as me, and leave projects half built

[cyan wing]

just a lot of repro steps / instructions for QA'ing something
that could be way more concise

[blissful copper]

It’s not the strategy but the alerts for trading alerts

[kind jay]

DM me, I can look in the morning

[blissful copper]

I would much rather perfer the prompts/tips I don’t really wanna send out my scripts

[calm aurora]

Why trading view and not connect directly to the exchanges data?

[blissful copper]

It’s prop firms

[calm aurora]

And do an email agent through imap or SMTP

[blissful copper]

I use traderspost and it reads my alerts on trading view and make the trades

[kind jay]

Fair enough

[blissful copper]

The alerts are weird, it fires and leaves in like 1min or doesn’t fire at all

[calm aurora]

What do you use as signal?

[blissful copper]

I’m happy I’m not the only one using codex like this, everyone else was telling me use Claude but I think codex is better

[calm aurora]

BUY, SELL or hold?

[frosty zealot]

And here I thought Squad had new patch notes.

[blissful copper]

Both, I do scalps, intra day holds. And etc

It buys and sells

It’s a mix of different strategies for different instruments

I connected my codex to my quant connecter so it run like 1000+ backtest on a concept

[calm aurora]

Yeah, so, map math interpretation to the signals and then you can analyze when different strategies support a signal

what kpi are u using?

[blissful copper]

What do you mean by kpi

[kind jay]

Indicators

[calm aurora]

KPI stands for Key Performance Indicator.

A KPI is a measurable value used to evaluate how well a person, team, system, or organization is achieving a specific objective.

Simple definition

A KPI = a metric that tells you if you are succeeding or failing at a goal.

[blissful copper]

After I get the strategy report I just send it to chat gpt and ask it what’s the likely hood it could pass and grow a account

But in terms of indicators I don’t use traditional ones, I make them on the spot

[kind jay]

Who do you trade with that allow algo trading?

[blissful copper]

All propfirms do, at least they don’t really know about it

I use apex and topstep

[calm aurora]

Use data proven indicators

[kind jay]

I think they don’t allow it, but you’re right about how would the know

[blissful copper]

I use traderspost and they make the trades off of my alerts sent from trading view

Only problem is that the alerts fire wrong

[calm aurora]

Usually they are proven for a reason and correlations can be discovered only through proven methods. Or its just luck

[blissful copper]

My philosophy is that any indicator that’s been published loses edge

[calm aurora]

Not trying to discourage new methods

But data needs old data to analyze change

[blissful copper]

Which is what personal backtesting and quant connecter is for

Real institutional funds don’t publish there strategy

[boreal holly]

Dang bro you got a backup acct?

[kind jay]

It’s photoshop, he told me

[frosty zealot]

Unrelenting force

[blissful copper]

So does anyone else use TradingView alerts? If so dm me but I won’t be able to see the messages. I’m really here for that fix

[calm aurora]

Work through metadata, there you will find relations hidden

[calm aurora]

I used direct exchange api

So cant help with trading view specificly

[kind jay]

You should DM him, maybe you know the fix

[blissful copper]

You use propfirms?

[frosty zealot]

Unless you create your own systems, your own intelligence networks, you're in the same boat as every other person trying to get rich off automating crypto trading

But maybe you can get to a working MVP before the others

GL

[calm aurora]

I used my money, but you can start only ingesting data and running the flow in paper

[kind jay]

You’re just jealous you’re not as rich and as handsome as Tito is

[boreal holly]

My limits reset just now

[frosty zealot]

We can't all be lucky

[calm aurora]

Like that weekle ussage

[blissful copper]

Yo Tito you got Insta? Could I add you?

[boreal holly]

oai is too generous lemme tell ya

[calm aurora]

Yeah

atomic.tito

But lets connect through linkedin

Dm me

Ah

Self promoting

Right?

[rugged burrow]

has anyone been able to update to latest codex in windows?

[frosty zealot]

You weren't kidding, thanks for the heads up

[kind jay]

Rule 12. smh

[blissful copper]

It’s

[calm aurora]

You can find me in linkedin with the name here

Is my real name

[kind jay]

Why do people have their real full name on Discord? It makes no sense

[calm aurora]

Another philosofy

[frosty zealot]

Some people are just here to network

[calm aurora]

All people can hack you

[frosty zealot]

Some people are here to disturb the peace

[calm aurora]

You only have yourself and your mind

[kind jay]

Some people don’t care about internet safety

[frosty zealot]

You should start a webinar on it

[calm aurora]

Thats why I focus on investigating and sharing ideas

I believe in full tracability

If you cant prove I am real

Why should you consider my thoughts

[frosty zealot]

How can I tell what is real, when real is a concept that I've created

[kind jay]

Speak for yourself

[calm aurora]

Then you share it through linkedin

Hahhaja

[boreal holly]

Yeah imagine having deep and insightful discussions with "Tomato_bandit54" on the internet

[kind jay]

Am I the observer or am I being observed?

[calm aurora]

Reality is beyond our own thinking

Another stand

[kind jay]

Reality is like quantum mechanics

[calm aurora]

Yeaaah

[boreal holly]

Birds aren't real

[kind jay]

I’m interested in working on quantum mechanical trading algorithms

[frosty zealot]

nerd

🤓

[kind jay]

Novel applications of quantum mechanics

[calm aurora]

I thought this through quantic computing

[kind jay]

You must be really smart

[calm aurora]

But if it works it would reshape economy

[frosty zealot]

I'll just remain in my superposition

[kind jay]

What if money is really a quantum mechanical system?

[calm aurora]

Is a belief system

[kind jay]

And we only see the vibrations?

[calm aurora]

I trust the thing I buy from you is worth what you are offering

And you trust that I pay you with the same worth

Its like

A inequality¿?

[calm aurora]

Life

[kind jay]

This could be solved mathematically!!

[frosty zealot]

https://open.spotify.com/track/0F1VAVqj021UkGntiDz4nx?si=MJh_1DwiTDiK_OEHJn2ghg banger alert

[kind jay]

Doxed yourself again

[frosty zealot]

cut off the param for me

[kind jay]

[calm aurora]

I dont open links from discord

First lesson

Hahhaja

[kind jay]

Mr. internet safety over here

[calm aurora]

Hahahaha

[frosty zealot]

I find something so ironic about a guy with his PII right there saying he won't click a discord link

Hillariously ironic

im rolling

[calm aurora]

You have to cut the line somewhere

[boreal holly]

Bro are you kidding me? Adventure Club is in my top 5

[kind jay]

Eoomnmh

[frosty zealot]

They're so good

[calm aurora]

Sorry the ignorance, whats PII?

[frosty zealot]

Personally identifiable information, its a common acronym in security

[frosty zealot]

1. Kill the Noise
2. Noisia
3. Feed Me
4. Gunship
5. Deadmau5

if I had to make a list

Very hard list to make cause I like them all, including people that didnt make the list, so much

[calm aurora]

Security and privacy is a fachade

[boreal holly]

1. Marcin
2. Adventure Club
3. John Michael Howell
4. Zed's Dead
5. Gretta Van Fleet (tied with Led Zeppelin)

[frosty zealot]

I've heard a few gretta tunes they're great, I was shocked to see they were so young

[calm aurora]

A few days ago I left my wallet in a public chair, no one took it

[frosty zealot]

To an extent sure, but you should still lead it with your best foot forward, next thing you know you're gonna have a mortgage and a truck payment in your name

[kind jay]

Lucky I wasn’t there

[calm aurora]

Things that only happens once

[frosty zealot]

Rat mode

[calm aurora]

I dont use credit or debit cards so, there is that. But other data is very easy to get if getting harassed by some freak dude

[frosty zealot]

If we're being across genres

1. Gojira
2. Soilwork
3. Kill The Noise
4. Noisia
5. Lamb of God

[calm aurora]

Today a lot of smaller companies use KYOC or world id

But

You trust them with your info?

[kind jay]

I lobe big corporations

[frosty zealot]

I accept there is always a risk when you sign up to something online, or hand out info on forms, but if avoidable I don't

Jane probably already has found my entire PII and built a shrine of me

I know it's out there

[boreal holly]

lobe is a strong word

[calm aurora]

Yeah, its not like people can trace you through spotify id

[kind jay]

You have a mortgage and a truck out in your name

[frosty zealot]

Is that what those missed calls were

[calm aurora]

Been receiving those for a few years

[kind jay]

Sorry

[calm aurora]

All men must serve

Jane is like Varys

[frosty zealot]

Serve what

[kind jay]

Just serve 💅

[frosty zealot]

Whatever it is I volunteer Jane to take my place

I'll pass

[frosty zealot]

Is this correct?

[calm aurora]

Yes

[frosty zealot]

I believe it

Kinda hot though

[calm aurora]

Got reeference

[kind jay]

I was about to post this and say wth

[frosty zealot]

Beauty is in the eye of the observer

[kind jay]

How did you get my picture?

[calm aurora]

Ask chatgpt, varys is daddy of managing PII

[frosty zealot]

Interesting typo, is english not your first language?

e and y are quite spaced, couldn't have simply been a misclick

Hmm, another slip

Timezone, semantics, profile is almost complete

[calm aurora]

Arent you from -insert metadata-

Hahaha

[frosty zealot]

LOL

Make it stop

[calm aurora]

How

Haha

[kind jay]

Okay goodbye guys

[calm aurora]

Goodbye Varys

[frosty zealot]

Gn

[kind jay]

I do not appreciate this

[calm aurora]

All men must serve

All men must die

[frosty zealot]

Is that a quote from Varys?

[kind jay]

Explain this before I leave please

[frosty zealot]

I couldnt get to it before Teamviewer paused

and i cant ssh in now

Im cooked till i get home

[kind jay]

:wq

[frosty zealot]

hmmm

I need to make another PiKVM

[boreal holly]

If you disable FileVault you'd be able to set tailscale as a launchagent and it'll reconnect after power failure

[frosty zealot]

How do I power fail it though when it’s plugged in on the charger lol

[boreal holly]

Sorry I mean when macOS jetsams from running out of memory

[frosty zealot]

I did that with my desktop computer. I have it plugged into a little smart plug, and I put in the bios after power failure

So if it hangs up or something, I just power cycle it

[boreal holly]

macOS FileVault will not let anything launch until after first login, so if OOM causes a kernel panic you need physical access. With it off, LaunchAgents can run before login and all is good

[frosty zealot]

I think at some point I did turn it off. I'll look into it when I get home

You know if you can enable the developer console inside the Codex app?

[boreal holly]

It's not an easy process. You have to unpack the asar, force the "if production then no dev tools" branch to never run, and use electron to repackage it as a custom macos app

You can't edit it in place because electron codesigns with a checksum of the asar so if the content changes even 1 byte it refuses to run

[frosty zealot]

Gotcha

Wasn’t sure if there were launch options from terminal

Given they are designing the CLI with Rust, I wonder why they chose Electron and not something like Tauri

[calm aurora]

@kind jay sorry, my phone battery died

All men must serve means that all human beings have responsability to colective

All men must die means no matter what you do, your cicle everytime ends

[frosty zealot]

You're all good. I think she crawled into her little rat hole for the night, probably doing a couple honk shoo honk shoo mimimi

[calm aurora]

I have a question, what are spouses here? Little lost in that matter

[frosty zealot]

Like in the English language?

[calm aurora]

Nono, in the server, they talk about spouses

[frosty zealot]

I'm not sure. I can't honestly say that I've seen that word mentioned in here, in this channel specifically?

If you can give me some more context or where they use it in a sentence

[calm aurora]

Something to do with roleplay

[frosty zealot]

I mean a spouse would be your husband or your wife

[calm aurora]

Maybe its a critic to AI companion?

Dont know

Maybe I have to learn more

[frosty zealot]

Could be. Without the context in which it was mentioned, I can't really help you

[calm aurora]

When I see more data, I’ll share it

[frosty zealot]

Okay

[calm aurora]

Goodnight by the way! Finishing my day

Didnt touch codex today

To much work haha

[frosty zealot]

Good chat have a good evening

[frosty zealot]

Dumb question, if you fork into new worktree, does it also take the context of all the old conversation, or do you get a fresh thread, only reason I ask is I see the chat history from where it was forked, but that might not make it to the model

[cedar skiff]

it forks with the conversation history

[cedar parcel]

https://www.reddit.com/r/codex/comments/1rrmoqt/54_brilliant_and_lazy/ this has very much been my week, having to do multiple sessions for each task because 5.4 is optimized to basically sacrifice quality for speed. Since agents.md rules don't help with this anymore, has anyone had any success combating this? I'm losing so much time, I'd rather just have the model take the time it needs to produce the best possible outcome, instead of simulating a fast experience by not loading context.

I'm thinking that maybe breaking tasks up in parent and atomic child items with a single pass/fail at the end of each. Really breaking down sessions to tiny tasks that the model won't cut corners on. And maybe a process for triage that assumes the first pass is always incomplete so running triage 3-5 times by default. Or 8 times and picking the best/combining the outcomes....

[cedar skiff]

i havent notice this too much with 5.4 but to handle it with claude i would alway tell it "During investigation it should always step through the code until it reaches concrete implementations so it doesnt miss important details, if it doesnt it can miss details and fail.

And that worked for claude

It stopped reinventing the wheel or having assumptions about implementations

[simple star]

Fellas, is it me, or is Codex extremela slow lately?

Im on /fast and even then...

[ebon ingot]

I usually make it to do small tasks every session then ask it for the diff and i send it to gpt to see if it is doing well

[pliant wren]

Hi

‪Why does the codex plugin in vscode open a linked file into a browser (which is obviously not there) when the file it created/edited is already inside the current project folder?‬

I’m on windows

Anyone else has same issue?

[cedar parcel]

So if you give it a user story to implement something in an average complexity codebase, and then run another agent with an adversarial review specifically to look for DRY violations and if the implementation is production ready in enterprise environment. It does not come up with a long list of problems that are all clearly from the model not reading the codebase until it knows the existing patterns? For me, both my existing prompts and when I go manual, it just doesn't acquire context. It never produces anything I don't need to revisit multiple times like 5.2 used to do.

[cedar skiff]

I tend to get by pretty well I think because I have a topical skills built around all the different established abstractions and app architecture. Codex just by the nature of it tends to keep those updated as it goes. So I think that mitigates a lot of the problem for me personally. This didn't work so well with Claude because Claude doesn't use skills after the first turn. So with Claude I had a workflow that forced it to find example implementations in the code base and also always step through code until it reaches concrete implementations.

[cedar parcel]

My workflow has always been to run the same prompts over a number of sessions. Triage the issue, resolve it, review the MR, merge to dev for human review. It got so good I have deployed small things without even reviewing. It rarely produced anything that needed a revisit. That's definitely opposite now. So I'm thinking that with the capacity problems quantizing is inevitable and as a product feature openai is faking speed over quality until the chip supply catches up. And maybe atomising the process with constant QC at ever little turn would be a permanent solution to it.

[cedar parcel]

I catch it not reading it's skills, and certainly not reading longer skill files completely

And that's the thing also... All those Claude workarounds I used to do too until I switched to codex and I've been free of all of that haha. But now it's back.

[cedar skiff]

I havent noticed it not using skills, i have caught it a few times not doing what they prescribe so that was probably it not reading them now that i think about it.
I just had to say does this look like it matches the code base architecture in the skills?!?

I don't have skills over 300 lines, they are always broken up into reference files

[cedar parcel]

Yeah I even have skills per tool so it doesn't have to lookup commands. That workflow I mentioned is running all day for months so it has all the tools and all the optimization to do that as efficiently as possible

[cedar parcel]

That's another thing. I have that too. And yesterday it confessed to not reading the references. It just stayed at a certain depth and didn't open the file despite being instructed in the file referencing it

But I have noticed it is quite aggressive about contracts and gates. So each issue I give I had acceptance criteria and hard gates. If there are a few it will ignore some and just decide to pass QA. But I think if I give each issue a single pass fail contract, so a tiny scope that can be captured in just a single pass/fail, I think I can get it to iterate until the gate passes. I'm not sure if this is procedurally feasible but I could force triage to run several times, then break it up in single contract, hard gated sub tasks.

[frosty zealot]

That’s annoying so if you’re in a worktree,you’re confined to a single thread for eternity in the app from what I can see, only work around is a new worktree with the other worktree as the base

[cedar skiff]

You can click on the open VS code in the top right hand corner, then open a terminal and you can see the work tree ID. Then you can open it in the left in the left pane. Or you can create a permanent work tree, but that doesn't come with a forked conversation.

[frosty zealot]

I sleep now, for many many hours gn

[cedar skiff]

gn

[cedar parcel]

digging into the system prompts to see if this can actually be fixed at that level...

[cedar parcel]

this is why agents.md won't help (anymore) with this issue, same as what the reddit poster reported.

[torpid trout]

@boreal holly have you seen the other side‘s agent team feature in their code CLI?
That’s basically what codex misses and you are doing I think.

Seems like advances are strong in the cloude

They do it with tmux thou

[torpid trout]

Lucky you. I suffer old man’s syndrome. Sleep at midnight and at 5 it’s like some lights turn on and I’m just not asleep anymore 😵‍💫

It might contribute to my generally interpreted as „Russian“ social capabilities

[rocky parcel]

Anybody online right now get slapped by the security tag/restriction and have to verify? how long did it take?

[torpid trout]

I think it was quick for those who had that issue, I recall a few users here mentioning it.

It should be a straight forward process really

[wheat yoke]

I havent had to verify

For what use case do you have to verify

[torpid trout]

It triggers when you do stuff like pentesting and similar

[wheat yoke]

I do alot of that stuff and i havent Been asked yet

🥴

[torpid trout]

But I guess you also can get flagged for other „high risk“ deemed interactions

Never happened to me either that said

[rocky parcel]

I dont know what causes it but I'm working on what would be pretty shady stuff if it wasnt funded research.

[warped pine]

u mean cyber?

[rocky parcel]

broadly yes

[torpid trout]

Yeah so then you know I guess

[warped pine]

i verified one of my acounts it was like two minutes

[rocky parcel]

oh good, so I dont have to worry

[warped pine]

but that thing works weirdly since they fixed the false flagging

either they are trying to specifically flag "exploit this thing" kind of prompts or they are under-flagging cyber related stuff in general

[rocky parcel]

"Exploit this thing" is almost precisely what the research ends upbeing when you get down to it. Like most cyber relevant prompts. One way or another, on red or blue team you end up with that as the actual unit of work.

But if it only takes a few minutes I'm not worried. Its all above board :p

[warped pine]

well i've been asking codex to reproduce different kinds of bugs/vulnerabilities in test conditions a lot and didn't have any flags (on another account that i didnt verify)

[rocky parcel]

🙂 sounds interesting, what kind of things do you target?

[warped pine]

blockchain-related

[rocky parcel]

You dont have to say if you dont want to, I do mostly binary exploitation kind of research, in anything that isnt windows.

Routers, desktop, mobile, etc

Since most research code is throwaway anyway, the robots help a lot generating those early exploratory automations that never get committed.

[rocky parcel]

Another question thats more forward engineering focused. Has anyone here built their own front end around the codex app server in order to re-shape the interactions you have with the models to be a different paradigm of tech work? By that I mean things like https://github.com/openai/symphony. Or the strongdm attractor workflows. I'm looking for more perspectives on it, I've built up a pretty thorough graph based workflow manager where the main interactions I have for deliverable software are me specifying user stories and a cascade of agentic stuff happens, then the work is done. Its a lot of code though and something like the "symphony" project above seems like it might be conceptually simpler and I always prefer simplicity...

Anyway, does anyone want to compare notes or know of any more of this sort of project other than symphony or the strongdm one?

[rocky parcel]

Link to spec?

[rocky parcel]

Oh, I wouldnt consider that to be "complex end to end code", but it is a good spec for a quick test.

And the results are interesting

[pliant wren]

Hi Guys when codex references a file in its output and I click that files name I get taken to browser as

https://file+.vscode-resource.vscode-cdn.net/c%3A/Users/abc/.vscode/extensions/openai.chatgpt-26.311.21342-win32-x64/webview/

is that a bug in windows?

[rocky parcel]

That link presents the wrong cert, then prompts me to log in lol. Even though all the domains are actually owned by MS and its probably not malicious it gives malicious vibes.

[wide schooner]

its good, carefully prepared 350+ loc task with multiple possible mistakes to make

so gemini, claude and gpt consider this "complex end to end code" thats feasibly comparable

[wide schooner]

yeah i have too - the model does not exist in the first place

[sleek spoke]

Pro lite/x5 wen? my Pro is about to run out and I didn't need all that usage......

[wide schooner]

we need 40$ sub

[sleek spoke]

I would be happy with the $100 sub that has leaked on X

[wide schooner]

100$ sub would make 200$ users downgrade and make openai lose money

but 40$ sub would make 20$ users upgrade and 200$ users wouldnt downgrade

[rocky parcel]

Can we do a $600 and get 3x the current pro?

Thats the direction I want to move...

[wide schooner]

how the hell can you use up pro

atp use api

[rocky parcel]

Easily...

[wide schooner]

pro has virtually unlimited quota, are you sure you are logged in?

[rocky parcel]

No, the API is considerably more expensive.

[rocky parcel]

It does not LOL.

[wide schooner]

yeah thats the point

if you run 300 subagents then you arent getting the discount lol

[rocky parcel]

I know thats the point, and thats also the point of me wanting a bigger pro sub.

[wide schooner]

enterprise-level usage doesnt get the promo

[rocky parcel]

This is personal usage.

At work we do pay API prices.

[wide schooner]

theres no way you can use pro limits on personal usage

unless you share it with half of city or so 💀

[rocky parcel]

Your imagination for what kind of workflows are possible now is limiting you.

[wide schooner]

im literally on plus plan, spamming agents and 50k+ loc per day and im still not using up mine

💀

[rocky parcel]

LOC are a terrible metric.

[wide schooner]

if you tell it to scan 200k loc codebase every prompt then thats on you 💀 , skill issue

does 5.4 even have enough tokens per sec to use up pro plan limits 💀

[rocky parcel]

Like I said, you're being limited by your imagination.

[wide schooner]

nah, you would need to run it thru some braindead loop in order to use up pro limits

[rocky parcel]

You do you, guy. Keep churning out 50k LOC per day of plausible bs in your single threaded workflow.

[cedar skiff]

just running a two threads of plan, implement, audit and review is enough to burn through pro on the 2x usage

[sleek spoke]

I don't disagree, but there's front-end code for a $100 plan and Tibo has 👀-confirmed it

[still trellis]

yeah with openai/symphony

[calm aurora]

Hi

[rocky parcel]

Do you use symphony?

[still trellis]

yes

[rocky parcel]

Oh fantastic, did you try any other projects of similar intent? is your implementation public?

[simple star]

Everytime I hear about Symphony I always think first of all of the PHP framework

[still trellis]

I just forked it, happy to share my method

But be warned it chows your usage

https://github.com/openai/symphony

[rocky parcel]

I have my own system I've built on top of the strongdm attractor. How well does it work for you? Have you tried any other similar projects or did you try to roll your own prior to the symphony spec being open sourced?

[cedar parcel]

@cedar skiff https://gitlab.vialogos.dev/vialogos/codex-strict-profile My attempt at a fix. Uninstall included to cleanly remove

[still trellis]

I tried a ralphwiggum type loop but dont like the black box effect it has. I only add tickets to linear that I feel ok to hand off. Not familiar with strongdm attractor?

[calm aurora]

@still trellis tryied local data base?

and graphs?

[boreal holly]

On plus I hit the 5h limit in like 20 mins of normal usage, and the weekly in 15 hours (not a bug or anything, just extremely normal and expected)

[rocky parcel]

https://github.com/strongdm/attractor. Its more generic, and doesnt include a frontend, but also more flexible to set up workflows that are not super aligned to typical scrum/agile/kanban style forward engineering.

[still trellis]

Thanks, will have a looks see!

[still trellis]

You mean instead of Linear?

[wide schooner]

disable all the 800 mcps, and limit yourself to max 1 non recursive workflow, there you go

[boreal holly]

I have 1 MCP with exactly 1 tool in it lol

That I wrote myself

[wide schooner]

then what are you doing wrong?

[boreal holly]

Nothing! It's totally expected consumption

[rocky parcel]

Dude, people have productive workflows that are not "tell codex to do something, wait until its done, tell codex to do another thing, repeat". And those workflows arent wrong

[wide schooner]

im beginning to believe theres a quota-cism where different people get different ratelimits, based on account standing 🤔

[calm aurora]

This says codex
its how I work with data

My recommendation would be to use Symphony-like orchestration for execution management, while keeping the evidence pipeline
independent and deterministic. In Aquiles, the system should first structure the claim, then retrieve and validate evidence, and
only after that allow synthesis. That separation is essential if the goal is not only automation, but also reproducibility,
traceability, and institutional trust.

[rocky parcel]

Aquiles?

[calm aurora]

YEah

[wide schooner]

or maybe your codebase is an self contradicting mess and codex doesnt know what to do 😭

[boreal holly]

I have between 8-25 agents running concurrently. On plus plan it takes about 20 mins to hit the 5 hour rate limit. On Pro I don't hit the 5hr at all

[kind jay]

Broker

[rocky parcel]

you keep this assumption that whatever you do is right, and if anyone has a different experience they're wrong. That is not a good way to go through life.

[calm aurora]

sorry

lacks context

[kind jay]

You better be

[wide schooner]

see theres your issue. why the hell do you need 25x agents? thats 25x usage waste

[calm aurora]

Aquiles is tracability project I am running

for university

[wide schooner]

you cant even read what 1 agent produces at that speed

[rocky parcel]

Do you think engineering managers of large orgs read every line of code output?

[wide schooner]

erm YES?

💀

[calm aurora]

Aquiles is a research-oriented platform for evidence-grounded reasoning. Its objective is to transform natural-language claims into
structured, inspectable workflows where sources, assumptions, and inferential steps remain visible. Rather than replacing human
judgment, it aims to strengthen it through traceability, comparability, and explicit methodological control.

[boreal holly]

I have command execution delays

[wide schooner]

the moment someone pushes a SINGLE line that he didnt read, he should be fired

[calm aurora]

I could say it in spanish but its moderated

[rocky parcel]

No, my title is "Senior Principal Software Architect", if I had to review every line of every project I oversee there would not be any time for anything else and I still wouldnt stay up to date.

[kind jay]

Who gave you that title?

[boreal holly]

Besides they don't spend 100% of their time outputting code

[wide schooner]

you are a slopmaker and not a software engineer if you dont read what agent produces

[rocky parcel]

Some C suite guy who made it up to justify my salary lol

[kind jay]

Nice larp lil bro

[wide schooner]

bro is flexing corruption/nepotism 😭

[rocky parcel]

Well this derailed.

[warped pine]

get 2 more accounts

[calm aurora]

don't be judmental

all people have different ways of working, all serve a purpose

[wide schooner]

"codex run around and pretend you do something, wasting my companys money, so i keep my job by acting useful"

😭

[kind jay]

Surely it’s cheaper and more proper to start a team

[rocky parcel]

Reading eery line of code is not necessary, large orgs have worked like this for a long time, AI makes it so that ANYONE can have the same throughput with a properly orchestratd agentic system that used to only be available to someone in a position to direct dozens of individual developers effort.

[wide schooner]

Can i contact your company real quick with screenshot of what you just said?

[calm aurora]

In my country there is a saying hahaha, don't defend a company in which you don't work. But what I really mean is, your comment is to expose someone or to help them do a better job?

[rocky parcel]

You think this is some kind of statement but its not, it just reveals how little you know about how work happens.

[boreal holly]

I think the real claim should be "no single person reads all line changes". Everyone collectively reviews each line

[wide schooner]

my comment is to discourage the most braindead & dangerous practices possible

[calm aurora]

whats braindead though?

[kind jay]

It’s not a real company, don’t bother

[wide schooner]

thats not what he said, he said he doesnt have idea what he is pushing

[wide schooner]

yeah its probably just ragebait i guess

[kind jay]

It’s a larp

[calm aurora]

so, what do you think of Aquiles?

haha

[kind jay]

Haha

[modern hearth]

seems to have done the trick

[rocky parcel]

Thats also not what I said... again you're imposing your own assumptions, its the same reason you kept arguing earlier. You keep overlaying this same assumption about your own workflows being the only correct way to work.

[kind jay]

Cute

[kind jay]

I wonder if there’s a reasonable was to computer levels to say walk on from a screenshot alone

Like to do you avoid important content?

Could have like a little pet that wanders around

[wide schooner]

nohejutmine@ll

[calm aurora]

whats going on

I am lost

haha

[kind jay]

From screenshot alone though?

[wide schooner]

i made an extension wheres a little guy walking around my browser and mining parts of websites

[wide schooner]

no he actually breaks the website 😭

[kind jay]

That’s evil!

[rocky parcel]

Intentionally?

[wide schooner]

well tell that to bobby, he is currently mining you

[calm aurora]

love chaotic coding

[wide schooner]

yes bobby is intenionally mining websites, breaking them in process

[neon girder]

who do you think has more experience: "ampro" on discord with the student role, or steipete (who works for openai now and founded real companies)

[wide schooner]

link?

[still trellis]

@rocky parcel give me the "explain it like im 5 how you use attractor" my brain is a bit fired today apologies.

[rocky parcel]

There are many, many anecdotes like this.

[neon girder]

https://steipete.me/posts/2025/shipping-at-inference-speed

[kind jay]

They’re the same person

[calm aurora]

I think it has to do on how much work you are supervising, but in infinite time, all lines are read

[rocky parcel]

You create .dot files that define the graph that is a workflow, the attractor traverses the graph, taking action based on the nodes and edges, you can use the graph structure to define arbitrarily simple or complex workflows that compose and reference others. What it doesnt do is provide a UI for these things or triggers to start them.

[kind jay]

Programming feel of big time

[still trellis]

interesting... thanks!

[wide schooner]

xD

[neon girder]

he's built openclaw, got hired by openai, and has founded multiple companies. what have u done?

[wide schooner]

LMAOOO openclaw? the one FAMOUS FOR BEING TERRIBLY WRTITTEN?

[neon girder]

the one famous for changing how companies look at agentic ai? meta, openai, anthropic, etc, are all building their own openclaws now lol

[calm aurora]

have no notice the mensaje: likely corrected fast

to*

[wide schooner]

@neon girder xaxaxaxa

💀

[boreal holly]

I mean yeah, this has been the weirdest thing to get used to. You could probably scroll back up to August 2025 in this channel and find messages where I'm like "I do the coding, Codex helps with planning and sometimes writes DTOs and boring code tasks." Few months go by and I totally do not write the code anymore.

[wide schooner]

xDD

[neon girder]

show me your project with 100k github stars and no vulnerabilities ever, and then we'll talk

[rocky parcel]

For someone who is so angry about AI generated code not being reviewed, you sure do rely on the robots to form a lot of opinions.

[kind jay]

Where are you GitHub stars?

[wide schooner]

i review the code they write before pushing, thats the difference

[rocky parcel]

"you" write. Edit: nice catch.

[neon girder]

im not the one thinking im better than everyone who doesnt do exactly what i think is correct :)

[kind jay]

Like you’re talking about one guy who one the lottery (who also happens to have poor habits) vs an average person

[rocky parcel]

This is another assumption that makes no sense. "thinking i'm better than everyone" is exactly what you're doing btw, by dismissing all workflows that dont mirror yours.

[calm aurora]

don´t fight guys, people use your fights to retrieve data

[wide schooner]

no, they write

ai writes, i review and correct

because thats the only good and safe way to use ai

[calm aurora]

If the goal is to reduce black-box behavior, an explicit state machine is more useful than an attractor, because it makes stages,
transitions, and failure conditions inspectable.
True o not true? for Aquiles

[kind jay]

Aquiles

[boreal holly]

I mean that's a good justification for you to stick with Plus! Sounds like you do not need more usage than what's already provided

[rocky parcel]

There shouldnt be any black box in attractor unless you deliberately implement it in a way that doesnt record whats happening. edit: any further than any LLM based node is already a black box anyway.

[kind jay]

It depends on the context though, right? You can write test cases and then just hope for the best

[wide schooner]

no.
it depends on when you

• i need the code to be safe: YOU MUST READ IT
• its for personal use it doesnt need to work: then sure

[kind jay]

Also I’m trying to point out how yours no longer arguing the point but the character of the person making it

[calm aurora]

try to use codex only to design and test, but if it doesnt work deterministic I don't accept the outcome

[kind jay]

Valid

[calm aurora]

like Rcik said to Morty; focus on science

rick*

[rocky parcel]

I know what "need the code to be safe" means, I've written C for...ever. And if you need the code to be safe you MUST PROVE IT. Reading it is not sufficient. Tested, provably correct code is the requirement for safety, formally verified.

[wide schooner]

and yet you say you dont even read it

reading is the bare minimum

[calm aurora]

I think conceptos and understanding of them is better that reading what you dont understand. Its not the same for the mind reading 1000lines than 5.

So efficiency also means reducing no necessary read

[wide schooner]

you still should read the code.
furthermore corrupted tokens are very common with some llms

[wide schooner]

read is necessary....

or otherwise you get api keys in front end

💀

/* security measures here (TBA) */

[still trellis]

is this a circle of trust? may i confess something

[rocky parcel]

Verifying correct syntax and not using a lexer/parser? tsk tsk.

[wide schooner]

wont help with this

[calm aurora]

yes and in the same time no. Data not only shows with literal text, but also with key conceptos relations

[wide schooner]

you need to understand.... and to read

[calm aurora]

I if a write a million "a" line, and a phrase "Hi, my name is John". What info is more useful?

[torpid trout]

I don't copy paste, I type it off the screen

[calm aurora]

if you want to know the name

Thats why codex consumes more tokens when paste thousand of line instead of giving them a path

[torpid trout]

Because copy-paste is unsafe. I manually recompile it with my keyboard.

[calm aurora]

to a log doc

[wide schooner]

this is why you need to read the code

ai acts smart.
ai still lacks some common sense

[boreal holly]

I mean you can set it up to ensure success.

My methodology for running a whole bunch of agents is I supply an environment with a very strict process. The agents do not get to decide how code is validated and ran. Every commit undergoes static code analysis, integration tests are executed, memory leaks are measured using leaks or valgrind and passed back as commit failures. They are required to go back and fix it, and there are agents in charge of detecting drift by looking at their output contract and completion criteria.

They spend maybe 10% of their time actually writing code, maybe 30% researching the issue, and the remainder is rigorous reviewing until it meets the specification. I've put together tools to mitigate waste at every layer so 25 of my agents consume rate limits as though only 15 agents are running.

Why do it this way? Well, it's fun and I enjoy it. Could I do a single agent in a loop? Sure! But it's not fun for me and one agent to be completely honest would not operate as quickly as I would in most cases so in order to extract the most value I want more than 1 running.

[rocky parcel]

pot...?

[wide schooner]

still doesnt mean you can push it without reading

[rocky parcel]

That ratio is about what it is in professional environments also. Documenting, verifying, and general awareness of the problem space and reviewing prior work (researching) takes the bulk of the time.

[wide schooner]

if you think you can, then tell that to gpt and you will get humilliated quickly

[calm aurora]

complexity also impacts answering

sorry for the spanish comments

[rocky parcel]

You keep this double think going that the robots are simultaneously an authority on how work should be done but incapable of doing that work correctly without your help.

[calm aurora]

Your interpretation give AI answer meaning

but interpretation isn't literal

what if I want to provee that 1+1 isnt 2

So in standard arithmetic, it’s actually impossible to prove that “1 + 1 ≠ 2”, because the axioms that define numbers (like the Peano axioms) lead directly to the conclusion that:

1 is the successor of 0

2 is the successor of 1

adding 1 to a number gives its successor

Therefore:

1 + 1 = successor(1) = 2

However, there are a few interesting cases where people say “1 + 1 ≠ 2”:

Metaphorically: “1 + 1 = 3” in teamwork (synergy).

Computer arithmetic with overflow (limited bits).

Boolean logic where 1 + 1 can equal 1 (in OR-like systems).

Modular arithmetic, e.g. mod 2:

1 + 1 = 0

[neon girder]

"ai is only correct when it agrees with me"

[calm aurora]

don't sweet it guys, we all have this coping mechanism when our position is threathened, self trust is a coping mechanism. Better to... focus on science

[wide schooner]

ai will absolutely replace us

but not yet

maybe in few months

its still not good enough to run unsupervised

[still trellis]

cannot wait to never right/write a line of code again... there I said it!!! Phew I feel lighter all ready.

[calm aurora]

we are all part a of system. As humans need the living to live, AI needs someone to push the button

[wide schooner]

95% correct vs 99.99% correct is different

[calm aurora]

is a never ending cycle. human made the IA -- IA influence human --- human act different -- humana influence IA

change the word IA for technology, and its the same

[wide schooner]

ai is good enough to write code

the code often works

BUT

ai isnt good enough to make the code always secure

thats the issue

it absolutely will be soon, but it isnt yet

[rocky parcel]

Humans arent either, have you ever even glanced at a security conference?

[wide schooner]

the difference is ai makes mistakes so stupid sometimes that human would see them

wednesday contains letter a

[neon girder]

ironically a lot of agentic ai tools are being used to discover security issues in old projects written entirely by humans lmao

[wide schooner]

humans know that, ai fails sometimes

[calm aurora]

the problem with autonomous, is that in a long period of time possibilites increase, that doesnt mean they will always happen. But accelerates experiment

[wide schooner]

yeah

humans write bad code sometimes, ai writes bad code sometimes

ai should review human code, humans should review ai code

¯_(ツ)_/¯

[kind jay]

If AI writes bad code, who do you hold responsible?

[wide schooner]

the idiot who didnt read it before pushing

[calm aurora]

the thing is AI and humans aren't different, are secuential instructions given by a subject in the begginig

beggining

[rocky parcel]

whoever signed the commit, if thats an AI? off to the shredder with it.

[wide schooner]

beginning

[calm aurora]

that

[still trellis]

Please give us a Reset, please

[kind jay]

What’s this?

[still trellis]

codex usage

[calm aurora]

The idea of First Cause (or prima causa) is one of the classic attempts to explain why anything exists at all. It asks: if everything has a cause, what caused the very first thing? 🤔

1. The classical formulation

The most famous version comes from Thomas Aquinas in his cosmological argument:

Things in the world have causes.

Nothing can cause itself.

A chain of causes cannot go back infinitely.

Therefore there must be a first uncaused cause → the prima causa.

Aquinas identified that first cause with God.

2. Earlier roots

The idea already appears in Aristotle with the concept of the Unmoved Mover: something that initiates motion without itself being moved.

3. Modern philosophical critiques

Later philosophers questioned the assumptions:

David Hume argued we cannot assume the universe needs a cause just because things inside it do.

Immanuel Kant said causality might only apply within human experience, not necessarily to the universe as a whole.

4. Physics complicates it

Modern cosmology adds new layers:

The Big Bang theory suggests the universe had a beginning in time.

But quantum physics allows events without classical causes (e.g., fluctuations in vacuum fields).

So the question becomes deeper:
Is the universe caused, self-contained, or part of a larger system?

5. My take

The idea of prima causa is powerful because it exposes a limit of explanation. Eventually every explanation hits one of three walls:

Infinite regress (causes forever).

Self-explaining system (the universe just is).

First cause (something fundamental that needs no cause).

Philosophy hasn't definitively resolved which one is true.

[kind jay]

What app?

[still trellis]

codex bar by steipete

https://github.com/steipete/CodexBar

[kind jay]

That looks nice

I’ve been thinking about getting a Mac but I wouldn’t know which it get

[still trellis]

Get it! Macbook pro base model is plenty

[rocky parcel]

ez, choose a form factor then max it out.

[kind jay]

This looks about right

I was considering getting a few Mac minis and setting up proxmox

[rocky parcel]

(a joke) but really, what form factor are you look at? The base model of any is reasonable for lots of categories of work. If you have specific performance requirements then you probably know them and know what component to pay more for.

[calm aurora]

@wide schooner a question,why codex lets you don't show snippets?

<pre> - Error al ejecutar:

/lib/x86_64-linux-gnu/libc.so.6: version GLIBC_2.38' not found /lib/x86_64-linux-gnu/libm.so.6: version GLIBC_2.38' not found</pre>

shoes

well, time to eat breakgast

breakfast

[wide schooner]

are you working with production code, OR with personal use?

theres your answer.

[calm aurora]

investigation mainly

[subtle terrace]

the air these days is also pretty good

especially great if you travel or commute

[still trellis]

for sure, Boss has one very impressed.

[calm aurora]

codex works well in linux by the way

haha

ignorance is bliss. Can I send code here?

linux works well in old equipment

old equipment reduce costs¿?

[kind jay]

I think the neo looks cool

I’m it sure if I want another laptop though

[surreal wing]

so i made my own operating system and it doesnt look so good can someone help me how can i design this

[still trellis]

Neo looks cool yes...

[kind jay]

Doubt

[still trellis]

Figma, Paper, Magic Path, V0?

[surreal wing]

what?

[still trellis]

Design Tools check out Magic Path and V0

[warm pilot]

any good example where you have given the same prompt to an average human and an ai and the ai got worse code?

[surreal wing]

thank you

[wide schooner]

"make an app that lists all days which dont contain letter a, use no if and simliar statements. hardcode values"

or make a code that draws a pigeon

[wide schooner]

full screenshot with vbox visible?

[surreal wing]

idk how can i check

[calm aurora]

give a subject sufficient time to develop a answer and probably could do it

[warm pilot]

I'm really confused, how does this get insecure? printf '%s\n' Sunday Monday Tuesday Wednesday Thursday Friday | grep -vi 'a' – how did anyone get a security issue in there?

[warm pilot]

ehm, ok? There are issues in svg, if that's what you're refering to … but … really … how?

[wide schooner]

without grep

[calm aurora]

Anyone tried image gif creation through codex

Is fun

[surreal wing]

no but did you try make an operating system trough codex its not fun

[calm aurora]

Hahha

Trying to do it

But

There’s always a but

[warm pilot]

echo – I mean, there are no days (in english) without an 'a'. How will this get any security issue?

[calm aurora]

Are you doing it alone?

[surreal wing]

first one was this

[warm pilot]

it's still more fun than coding it by hand 🙂

[calm aurora]

Reaaz

[surreal wing]

yes?

[surreal wing]

yes

[calm aurora]

Are you using green computing?

[surreal wing]

no

idk

maybe

who knows

[calm aurora]

Haha, you messure energy consumption and time processes?

[surreal wing]

no

[warm pilot]

I wouldn't start with a GUI, if you are really trying to build an OS, start with a very very simple text interface, maybe just over serial. And then get the basics right, interupts, basic i/o, etc.

[calm aurora]

You can accelerate results capturing metrics

[calm aurora]

I vouch for this

[surreal wing]

should i restart?

[calm aurora]

No, only extract backend

Separate frontend from backend

[surreal wing]

oh ok

do you want my operating system?

[calm aurora]

What are you using?

I would try doing what @warm pilot says, but implementing green computing to capture process metrics

[surreal wing]

ok

[surreal wing]

do you want it?

[calm aurora]

If I want what?

Haha sorry

English not my main language

[surreal wing]

the operating system

[warm pilot]

I don't want to discourage you, but building an OS is really a complicated task. It seems like you only have experience "seeing" OSs and most likely always with GUIs. You either need to read quite a lot about how systems work and then start from scratch again or you search for another hobby 😉 – but really, if you want to learn, please do. The world needs more low-level hackers anyway.

[calm aurora]

Yeah, reading is very important if you want os

Start investigating about Linux

[surreal wing]

yes i see

codex can read for me

[cedar parcel]

why not make your own linux distribution? not that that's an easy task but it will get you a good looking result quicker

and then you'll learn easier because you can study by doing and then bake your own based on that if you need

[warm pilot]

good luck 🙂

[surreal wing]

i will try that

[kind jay]

Doing LFS with vibes lol

[warm pilot]

LFV? linux from vibes?

[kind jay]

Why do people love waisting their time?

[calm aurora]

You know why

[warm pilot]

Most really don't understand the complexity of these things. https://xkcd.com/1425/

[kind jay]

But they think they do because I wrote le code (but codex did)

[calm aurora]

i will stop wasting my time when I die

[kind jay]

I think there’s an emerging issue where people think that they are geniuses because codex wrote some code for them and it looks good on the outside, and people feel that they made it themselves

[calm aurora]

All men are fools

If you look to much into abysm, you can fall in it

[surreal wing]

you'll see when I make billions with my OS