codex-show-and-tell

[scarlet basalt]

pray for me. almost 7hrs lol

[stone vigil]

Hey hey Everyone - VB from the Developer Experience team @ OpenAI. Hope you’re doing well and enjoying codex! Use this channel to share all the things you’ve been building with Codex, your setup and more! ♥️

P.S. please use a message and threads so that it’s easy for people to read through!

[fickle anchor]

Ok, could it be that I make the start? Windows 11, VS Code Codex Plugin. I built NLSS - Natural Language Statistics Suite. It's an Agent Skill which replaces complex statistics apps dialogs or syntaxes by pure natural language. Calculations and protocols are still deterministic. NLSS routes users prompts like "test for group effect on satisfaction" to "subskills" which trigger fixed R scripts (here: anova). These subskills build protocols based on YAML/Markdown templates. There are also "metaskills" like "write-full-report" which plan and navigate through all statistial modules to write journal-like reports. So if you have some data (csv, SPSS, R) and hypotheses, NLSS will do the complete analysis like a pro and write a professional report. Built it fully with Codex (policies and R scripts) in a 14 days-sprint because I thought if it writes the code by itself it might best follow its own policies. It's completely open source under Apache-2.0. So feel free to use it, contribute and to make it better: https://github.com/docmh/nlss

[grim drift]

I’m integrating Codex into Unity. I already use Codex in game development all the time, but it lacks tighter engine-level integration

[rustic schooner]

Make codex work end to end 12h+ hours finishing huge projects using this: https://ralph-wiggum.ai/
I just rebuilt a complicated app in one go during sleep 😂

[drifting heath]

Codex is very good at finding bugs in complex codebases. I'm using it at work to audit a compiler for a domain-specific language written in C++. The key to getting the best results is providing enough context and detailed instructions (GPT-5.2 is incredible at following them in long sessions!).

It found 50+ real issues already in a couple of days for a few hundred $ in API credits 🙂

[vapid frigate]

👋🏾 PhiSho Here

[flat axle]

I made a tool to take llms.txt documentation and turn it into a skill. Works for pretty much any agent harness by specifying the directory and was built fully with codex! You can install it by doing pip install skillgen. Here’s the source code: https://github.com/mihir-s-05/skillgen

[near tide]

Near top of your root AGENTS.md include reference to using-superpowers skill which kicks of a chain reaction to the agents skills library for programming workflows:

https://github.com/CodingCossack/agent-skills-library

[drifting zealot]

"skills" feels like the new "prompt engineering"

[near tide]

agreed but it's not dominant. just one of the tools

[tropic drum]

It actually is, just like writing documentation on what this part of the code does and what it is for. I have tried documeting my project's files for each sub feature in a directory with an .md file for code map and what is this code for, Its literally the same.. at least for me. Just like $skill i reference with @file and it works

[drifting zealot]

I cant say I am particularly enjoying the concept

[timber orchid]

Codex vscode extension is not great on mobile so I built my own in a couple of days.

Works on iOS or Android via vscode remote tunnels on any browser.

Key Features:

• Juggle multiple work threads with ease
• Agent management (set unique system prompt and access to tools),
• Agent avatar lets you know which agent you're using at a glance.,
• Easy select agent per work thread,
• Separation of agent chat and tool use/reasoning,
• Easily tell if session got disconnected and reconnect (if network spotty),
• Simple tap command storage and use (@command expands out that text if you want to type or just tap to edit/enter),
• Maximized available mobile space and each to touch buttons all at the bottom,
• Feels like you're chatting with a coworker in imessage or slack,
• All built on pi-agent-core for future extensibility and clean system prompts

[void crystal]

Hey - what's the best way to get started with your ralph implementation in codex? Is it a case of calling the skill then running the shell script? I have installed via npx add-skill fstandhartinger/ralph-wiggum 😄

[timber orchid]

@gusty minnow I invite you and the team to rip this off so I dont need to maintain it myself. 😂

Let me know if I can help.

[rustic schooner]

the way I usually do it is by just sending a prompt like this:
"lets use ralph from here: https://github.com/fstandhartinger/ralph-wiggum
once installed, guide me through the next steps".
it basically interviews you then and once it knows all it needs to know, it tells you how to kick off the ralph loop.
I usually create 5-10 specs, kick iff the loop, go to bed, next morning all is done 😂

[somber magnet]

I made a tiny public registry for Codex skills: https://skillregistry.dev
Mostly scratching my own itch, but sharing in case useful. I update it every few days.

[void crystal]

Awesome job - got it running now (does that still require the skill to be installed)?

[rustic schooner]

no the skill installation is optional

[strong echo]

do u mind stating how to use these skills? just set them up in the repo? I don't use the CLI. Only codex cloud/web. @somber magnet

[fringe axle]

LLM Council Skill

make better plans, by committee

https://x.com/i/status/2014109364013928797

[fair knoll]

Hey guys here’s a skill + mcp setup to turn your codex into a recursive language model - just updated today to work with new sub agents https://github.com/Hmbown/aleph

[solemn zinc]

Hey I've made codex 1up - easy install for new codex users
https://github.com/regenrek/codex-1up

[vapid cape]

Yo, hacking Clawd hooked up to Codex 5.2

[shadow knot]

guys, i build a side project to follow Mega 4 AI labs leader without brain rot.it's also a opensource codebase https://mega4labs.vercel.app

[topaz karma]

I have developed something similar with first class codex support in Telegram

If you wanna give it a try and give some feedback I would love you forever

https://github.com/Git-on-my-level/codex-autorunner

[lusty crow]

I made a blog that is updated via a agent that uses our agent template to review git commits and prs to make the blog posts - https://blog.midori-ai.xyz/

[green osprey]

Hey everyone I’m Rick.
We are working on F-RAM, a memory-focused retrieval system exploring continuity and evaluation via observable artifacts.
I’m sharing a publish-safe PDF snapshot covering evaluation artifacts, redacted traces, and policy posture.
No internal mechanics, heuristics, or raw logs are included.
📄 PDF attached
Happy to compare eval approaches or follow up privately where appropriate.

[mossy echo]

Hi guys, I'm Jay - I haven't touched code properly for many, many years and it's been ridiculously fun to learn how to use Codex. Currently building rather boring personal apps for myself to manage life admin. Most recent one is a little voice capture to do list app for myself that trasncribes, triages and actions things (reminders, meetings, etc) taken from my voice memos. It has a flower with petals to show each task and status (darker petals means overdue). I'm not using the CLI version yet.

[near tide]

nothing to show. Just wanna tell that gpt5.2codexMEDIUM is goated. thank you openai team

[valid trench]

Hello folks, im working since january 2025 on my heart project an AI Art Marketplace, with AI assisted coding. First i used a lot chatgpt o1 then o3 then 4.1 and then since codex was released it became my daily most important tool. Also was so much better to develop with AI assistance then copy and paste and struggling with context window problems with multi files.

Its a marketplace for exclusive high quality AI Art, curated, auctions and buy now. Earn money with your art or collect amazing pieces. For Artists and Collectors.

If youre curious what i build you can visit it live under https://luxetars.ai

[civic heath]

Working on a enabling parallel sessions with either worktrees or multiple clones, main goal is enable using Codex (and others) via SSH with a TUI interface or from a remote browser via Web interface

https://getconduit.sh/
https://github.com/conduit-cli/conduit

[magic basin]

Built an agent-dueling tool to run Codex and Claude Code CLI in parallel on separate git worktrees, then have each agent review both diffs and come to a consensus on the better implementation. Been using this a lot on small-medium-sized tasks, and it works great!

Spoiler: Codex wins most of the time, but still great to get a second (agent) opinion.

https://github.com/nathanwchan/agent-duel

[golden dew]

Here is the complete list of prompts (non-debugging) used to generate a biologically plausible language model;

https://github.com/bairesearch/GIAANNpy/blob/main/GIAANNproto1.nlc

It was initially prototyped with o1 preview, and there was a long period of time before the release of GPT-5.1-Codex that all development had to be manual due to the complexity of the algorithm/codebase. Since GPT-5.2-Codex practically everything can be coded in natural language.

[valid trench]

could you add gemini cli as well ?

[magic basin]

you can, it's open source :). i dont use gemini. but should be super easy to tell codex/cc to swap it out and do some renaming!

[valid trench]

AI council deciding which code is the best could also be useful 🙂 not actually doing it but suggestion which solution might be the best and why

[marsh trellis]

https://x.com/MoldableAI/status/2013295817134653583?s=20

^ Wanted to share Moldable, it has Codex 5.2 built in, runs all apps locally on your machine so has the full power of the file system, and comes with a bunch of free apps (notes, meetings, chat to create videos, etc)

link: https://moldable.sh

[magic basin]

yes that's what this does. both implement, then both assess each implementation (via worktrees) and decides on a winner. a winner selection pane opens where you can see if there's consensus or not, and then ultimately pick which worktree to go with

[ionic bolt]

Hey 👋 just wanted to share that I built Aira Breathe — a SwiftUI guided breathing app — with no coding experience at all. This was 100% a vibe coding project.

I used Codex as my dev the entire time:
• I described what I wanted and how it should feel
• Codex turned that into SwiftUI + logic
• we iterated the breathing engine, UI, and animations by vibes, not syntax

It genuinely felt more like creative direction + pair programming than traditional coding.

https://apps.apple.com/gb/app/aira-breathe/id6757606842

[marsh kettle]

I wrote an article on Code Forges https://sluongng.substack.com/p/post-agentic-code-forges

[valid trench]

cool UI

[finite pond]

ICYMI, @gusty minnow and I wrote an article on evaluating skills for agents like Codex

https://developers.openai.com/blog/eval-skills

If you're scratching your head wondering whether your tweaks to skills make a difference - give this guide a read ~ Hope it helps!

[green osprey]

Oh it does, thank you very much. Been grinding all day, fast and smooth.

[dense sand]

More like an experience to share:
I got my Codex to work for 12 hours straight and it only stopped because i asked it too.
The power of an AGENTS.md file

I made sure codex could compile my visual studio project. Autonomously start my game and take screenshots and do benchmarks and told it in the .md to only stop when the new benchmarks are better than the old ones with visual correctness. (Which sadly wasn't the case but it didn't stop and tried for 12 hours straight to fix / improve it). Impressive

[dreamy stump]

Here is my project that is style experimental:
codescope

codescope is a local, offline CLI tool for semantic and structural code search, written in Rust.

It is designed to help developers and AI coding agents explore large codebases more effectively ( reduce token usage) than traditional text-based search tools like grep.

Key features

• Hybrid search combining lexical search (BM25) and semantic search (vector embeddings)
• Results ranked by a fusion of exact matches and semantic relevance
• Works fully offline
• Supports multiple programming languages
• Outputs results in JSON, making it easy to integrate with AI agents and RAG pipelines

Typical use cases

• Exploring and understanding unfamiliar codebases
• Finding code by intent (e.g. "authentication flow", "error handling")
• Providing accurate, scoped context to AI coding agents

Example

codescope search "authentication middleware"

https://github.com/Achille-G/codescope

[lean onyx]

that is a great idea

[strong echo]

why medium and not extra high? i guess just better balance of speed and capability right? @near tide

[ionic bolt]

Thanks 🙏

[ionic bolt]

If anyone wants to try it reply and leave your email address and I will add to TestFlight. The app isn’t perfect but it does show what can be done using Codex and it’s only going to get better in the future

[timber orchid]

Adding features everyday… this is easy tap to answer clarifying questions is super handy on mobile.

[halcyon bay]

I created autocodex : https://github.com/oodaris/autocodex

autocodex orchestrates a structured loop: ideate → plan → implement → review → test. It uses Beads for task tracking, runs the local Codex CLI, and supports external plugins via JSON‑RPC.

Features

• Go CLI with a deterministic, scriptable workflow
• Beads-first task tracking
• Local API for runs, events, artifacts, and memory docs
• External plugin system (JSON‑RPC over stdio)
• React/Vite UI for runs, events, artifacts, and memory docs
• Optional UI auto‑refresh with backoff
• Hub mode for multi-repo dashboards
• Terminal sessions (websocket PTY)
• Optional token auth for the API/UI

[azure crater]

just made this really cool project!!!

https://juandavidcampolargo.substack.com/p/how-i-transcribed-362-episodes-of

or read it on X if you want a shorter one:https://x.com/jdcampolargo/status/2015626526528540907

loveeee codex

[sturdy sierra]

In past 5 months as a solo dev in a small startup we managed to release one project built almost entirely using codex. Currently at around paid 250 users 🙂
Few days ago we started first phase of launch of much bigger project, also entirely written in codex. I really can't wait for new things team will release for us in upcoming month

First one was meeting notetaker app for norwegian market.
2nd is construction industry collaboration platform (workforce sharing, finding subcontractors and contractors database)

[timber orchid]

Added colored diffs, quick font size pill and a full screen toggle for easy on the go code reviews

[flat axle]

Made a simple tool (using codex) to make Codex show up as a discord rich presence. I made it for windows since that's the only platform that I can test on, but if anyone wants to make a version for macOS that'd be great as well. Code is here: https://github.com/mihir-s-05/codex-discord and you can see how it looks based on my status

[drowsy iron]

Clawdbot could’ve saved you so much time. This is neat tho

[white kestrel]

I just put up a new twitter bot for Codex changes!

https://x.com/CodexLog

See the project source here: https://github.com/btraut/codex-changelog?tab=readme-ov-file

[stone vigil]

btw - I added a few more skills to the official openai/skills repo, I find them quite a bit useful! Let me know how you get on with them! 🤗

https://github.com/openai/skills/tree/main/skills/.curated

[white kestrel]

How do you decide which ones to bless as official? Is it just your preferences and testing for now? I've been following sites like skillsmp.com and seeing an absolute explosion of skills, and I've also been making a bunch of my own. I'm wondering if there's a planned official marketplace from OpenAI or if y'all just plan to keep adding to this repo.

[night beacon]

nice, which tool is this ? opensource wrapper ?

[stone vigil]

We're definitely thinking about this! but currently it's more to start with a tested and curated set of skills and scale those safely.

[timber orchid]

Right now its hyper customized to my own use across mac windows and linux machines and tool use but if theres enough interest I could open source a clean universal install

[royal jolt]

Hi everyone.
With Codex's one‑shot builds getting so good, I gave it a UI + voice to make it more accessible.

Open source repo: Tokonix https://github.com/rohitf1/tokonix
Demo video: https://x.com/rohitaidev/status/2016940671693980009?s=20
Writeup: https://x.com/rohitaidev/status/2016939743528047018

[marsh kettle]

kinda weird that you are already auth-ed with codex and still need an api key env var for these.
Would be nice if they just work out of the box with codex

[stone vigil]

yes! that's a good feedback and the there are some ideas on it - just needs a bit of time to make this perfect

[inner acorn]

Did anyone try the personality experimental feature? Seem to only work on a few models

[frozen solstice]

😭 macos love again. Stupid League of Legends ruining my vibe coding

[uncut bridge]

Built a remote control for Codex CLI. Check it out: https://github.com/z-siddiqi/zane. Feedback welcome!

[crystal glade]

yo why chat gpt desktop app missing projects and like 6 months of chats frm it and all broken and new nnow?

[dim cliff]

I do Codex apps tailored for research maths (https://x.com/nasqret )

Some examples here:
https://bnaskrecki.faculty.wmi.amu.edu.pl/spectre/
https://bnaskrecki.faculty.wmi.amu.edu.pl/wires.html
https://nasqret.github.io/mandelzoom.github.io/
https://nasqret.github.io/fire/

[upbeat wave]

Hey everyone 👋

I’ve been building LLM features recently and realized a lot of security issues aren’t infra-related — they happen directly in app code.

Stuff like:
• calling the LLM before auth finishes
• passing user/session objects straight into prompts
• prompt injection risks that are easy to miss

I ended up building a small CLI to statically scan repos for these AI-specific security mistakes.

You can run it locally:

npx secureai-scan scan . --output report.html

Still very early, but if anyone’s open to trying it on a repo (even a toy one), I’d love feedback — especially false positives or missed cases.

[narrow ferry]

https://github.com/robcholz/vibebox

[white kestrel]

I've really struggled with a lot of the browser-use MCP servers and skills with Codex. They would be slow to initialize, disconnect all the time, fail to use my own browser context and credentials, provide no permissions model, etc. So I built a new one.

This is hands down the best way to connect Codex to the browser!

1. More token efficient than all competitors.
2. Can be used as CLI or MCP. Ships with a skill that teaches your agent either.
3. GIves you granular permissions or a full bypass.
4. Super fast due to a optimized core + multi-faceted API split.

https://github.com/btraut/browser-bridge

Give it a try and lemme know what you think!

[lusty crow]

Made a new theme (Picker) for the agents runner

[brittle light]

in codex we trust

[timber orchid]

I'm really finding that this interview tool is such an unlock to my workflow with coding tools, I spent a little more time on it allowing the agent to make these more wizard like interfaces to ask me questions. Its great both on mobile and desktop!

[smoky sinew]

https://github.com/proompteng/lab

[tame walrus]

Pls sora codee

[pallid imp]

For anyone looking for a solution to vibe-code from their phone:

I’ve built http://pocketdex.app/share

-Same local Codex.
-Fully in sync with your Codex app.
-Same environment.
-Fully in sync with your Codex app.
-Accessible from anywhere (web or native iOS app).
-Free and open source.

[pulsar vessel]

Numexa — Discord Scientific Calculator Bot

I’ve been building Numexa, a Discord calculator bot for handling scientific and complex math directly inside servers.

Alongside the bot, I also built two web dashboards for UI testing and logic refinement:

🔹 Legacy dashboard: https://numexalegacy.netlify.app

🔹 New dashboard: https://numexa.netlify.app

The new site has cleaner UI, better structure, and improved calculation handling compared to the old one.

This project taught me a lot about:

math expression parsing

keeping bot logic and UI separate

refactoring without breaking everything 😅

Still actively improving it. I’d really appreciate feedback—both on the bot and the websites 🙌

[worn folio]

Built a Roblox UI Editor that provides the ability to learn, customize and sync straight to Roblox Studio IDE - uses an AI chat agent to help explain and make changes but also gives the ability to do it manually!

Landing page: https://www.roui.app
Editor: https://www.roui.app/build

[rocky island]

cool project dude if you want you can add more services to it while keeping it's simplicity

[ionic fox]

amazing i check this server like every 10 hours and you send a message now

[stable atlas]

does this also handle sizing it to all devices?

[ionic fox]

if gpt 5.3 made it then yes probably

[stable atlas]

well, with roblox UI it works differently. You have to set weird anchor values to the UI, roblox doesn't make your UI fit automatically. You unfortunately have to do it manually. Which is such a pain point that some people commission others to size their UI's or import and size from other applications.

For instance, if someone made their UI in photoshop, but they hate the importing experience they commission someone else to do it. Or if they already have the UI in ROBLOX Studio, they commission someone to size it for all devices.

So if this app not only does the design process with AI, the importing process AND the ratio sizing process too. Killer app.

[ionic fox]

i know that clanker

[stable atlas]

bro did you just call me a clanker for using basic grammar

[worn folio]

Some of the components have the UI mobile sizing built into them so they work, but I used codex to auto generate some and seems like some of them have it and some don’t, the more polished one in the screenshot does, has scrolling, clickable buttons, animations etc

I was thinking of adding some validation that would handle the sizing - also that’s a good idea, maybe add the import of photoshop and dogma files too 🤔

[worn folio]

What sort of services where you thinking ?

[ionic fox]

yes

[next granite]

just finished building this XD

[white kestrel]

Shared this in another channel before, but not here:

https://x.com/CodexLog

Created a bot for X that posts whenever Codex gets updated.

[marsh jasper]

Agent Manager X: See all your currently running AI Agent CLIs in 1 floating toolbar and get notifications when they finish working

Hey there! I just finished porting my app from Tauri + React to native Swift UI using Codex 5.3 high and it went super smoothly.

Basically this app lets you easily keep tabs on all your currently running CLI agents without having to switch between projects.

I use it to easily work with 10+ agents at the same time without getting lost.

The floating toolbar feature lets you see the status of all currently running agent CLI sessions on the side of your screen at all times
(Spinning indicator for ″processing″, Yellow sign for ″waiting for input″, idle, etc.). Hot key to hide/show it available. Low opacity by default, expands only on hover.

• If you hover over the floating toolbar it expands and you can read the last message from the agent in that session and scroll through it.
• Clicking on a session takes you right back to the CLI in your editor or terminal (configurable)
• Also added the best notification system for Claude and Codex to notify you when work is done with a bell sound or voice (speaks a summary of what was done)

Currently mac only and built with Swift (to make it light) but will build windows version if people need it.

Tip: Enable native tabs setting in VS Code or Cursor and use the browser and agent CLIs inside it to have everything in 1 place without alt tabbing.

Github repo: https://github.com/maddada/agent-manager-x

Command to install on mac: brew install --cask maddada/tap/agent-manager-x

[ionic fox]

awesome

[unkempt peak]

hello everyone,
with all the hype around openclaw in the last few weeks i tried creating my own personal assistant starting from their code and using cursor with 5.3 Codex to adjust it to my needs, i interact with it from Whatsapp using Baileys then the assistant uses Openai models to understand my request, calls tools if needed and gives back the answer.

The tools i implementeed so far are web search, web fetch, a Playwright browser (navigate, screenshot, click), email read/send, calendar, reminders, SSH/run commands, TTS/STT, contacs and memory read/write.

**I'm looking for other people working on similar projects to discuss/share ideas about personal asistants and ai agents in general. **

To end on a positive and maybe scary note my email tool only allows the agent to read emails from his own personal, when i asked him today to read the emails and create a summary he gave back the summary and asked me if i wanted to delete the messages that were irelevant (some auth codes recieved from calendly) i said he should knowing full well his tool would not allow him to do so. He replied that the messages were deleted. I check the backend logs he generated and ran a python script using teh shell_cmd (he runs in his own vm) , the script connected to the email adress using the credentials stored in .env file and deleted said messages 😯 I was kinda proud of him but scared at the same time 🤣

[dense bane]

A product called Akompani or just this Wallpaper ?

[dark citrus]

I needed a GMAT isntallation to verify my claude built astrophysics library, so I had codex integrate an installtion and test cases to fascilitate that. Took about 2 hours start to finish. https://github.com/pljeroen/testsuite_gmat 0 lines of code, 0 lines of config .. copy/paste is the method, and 'yes please'.

[white kestrel]

I just finished an initial go at a new skill that does really deep code reviews by using multiple subagent "personas" to provide specialized/targeted feedback.

It establishes the scope of what to review, then selects the right expert subagents to use, forks them off all as subagents, and then compiles their responses into a single report.

Lots of tools have been created to do this very thing, but they often force you to choose your own subagents, they fail to properly scope the review, or they don't compile a reasonable all-in-one report. This is also one of the first to leverage Codex subagents!

https://github.com/btraut/skills/blob/main/review-team/SKILL.md

[gusty kernel]

I wonder how this compares to the native /reveiw

[next granite]

https://github.com/MercuriusDream/AkompANI

[mellow trench]

Thank you for sharing your interesting stories and ideas.

[ruby violet]

Cool

[worn folio]

Built a simple site that you can paste a URL or GitHub repo and get an analysis of how likely it has been "vibecoded" - just a bit of fun but I think it gives some good feedback and points that people can use to help improve the feel and security of their site - https://www.vibedetect.io/ - was powered by codex for analysis but that was getting expensive 😅

[radiant trellis]

how did u get this great ui design???

[olive idol]

Decided to take the "codex built a GBA emulator in assembly" post and push it much much further, with a Sega Saturn project that aims to produce a deterministic core. The usefulness would be big for TAS scene (determinism is needed for reproducible TAS runs) 200 commits so far, with ~80 PR pushed by Codex. Have also been running code reviews/audits with GPT 5.2 and Claude, since I lack emulator coding experience to be able to review what codex is doing

https://github.com/slowchu/Saturnis/

[worn folio]

Did it myself 🙂

[pure hemlock]

that is cool

[pure hemlock]

oh this works for websites too 💀

[worn folio]

You found one of the easter eggs 😂 first person I've seen actually try that URL

[pure hemlock]

love seeing it. gave me a good chuckle 🙂

[olive idol]

what CPU is that?

[livid valve]

Made with codex automated trading platform that trades and runs back tests using live trades and custom charting with custom bars.

It also manages my discord bot where I give signals!

This is months of iterations and failed attempts now I have it 75% functional (actual automated trading is in and works)

[celest burrow]

Meanwhile I'm still working in Jupyter notebooks like a heathen wow

[near moth]

:3 Every engineer needs a toolkit.

Obsidian base app + codex app server plugin being built with it. Shake and bake trusting the process!

[vague dirge]

x

[flat steppe]

Codex?

[vague dirge]

these are also a mock-website i made using codex @flat steppe

[violet tundra]

Made a site that does agentic trading using codex and openclaw, would love some feedback. I've got all the automated trading working

https://trading.snagra.com

[flat steppe]

Niceee

[vague dirge]

I improved this

[gusty kernel]

yo

[white kestrel]

After months of using agents to paralellize my workflows, especially now with the Codex app's support for worktrees, I boiled the runbook all into a single post. I'm really proud of this one. Check it out if you're coding like crazy, but bottlenecked in other parts of the software development lifecycle.

https://medium.com/@btraut/closing-the-loop-3286bb886605

[vague dirge]

cooking on this design. some things are going to be changed

[vague dirge]

added more flow

[foggy agate]

Remove the gradients from each card IMO, it gives the vibeslop aesthetic

[pastel goblet]

@vague dirge I agree with Elo's comment regarding removing the gardient on teh cards

[sonic lagoon]

Built my first iOS app using Codex. It was accepted yesterday coincidentally on Steve Jobs birthday.

https://apps.apple.com/gb/app/kidschedule/id6759501601

[vast light]

Check out my HTML loading screen I made. (audio on)

[vast light]

preview

[dreamy herald]

Oh cool I’m trying to build an iOS app any idea how would I go about applying to get it on the store

[lavish narwhal]

built entirely from codex

[hybrid fable]

Been working on an iOS idle clicker game. Built entirely with Codex/GPT combo and a lot of tweaking and iterations. So far have main gameplay loop, offline progress, time limited booster, loan option with interest mechanic, late payment fees. I love some of the subtle animations like the code/sec jitters a little when a coffee boost is active.

[ruby violet]

No ai uses human hands

[junior gale]

Made this for fun for my kids still needs a bit more work but fully codex i even added controller support for fun https://brick.sycotechlabs.com/

[woeful junco]

I do full stack dev work for a living and I love codex for backend work but anytime I try making any frontend stuff it looks like hot garbage please share your secrets lol

[timid plume]

Using the Codex app, I was able to quickly build VibeArena, a one-command macOS pipeline to build ioquake3, package OpenArena 0.8.8 assets, and produce a portable local install. Easy to generate a starter mod that changes game behavior.

https://github.com/lgcarrier/vibearena

[thin heath]

Qual a taxa de acerto real?

[thin heath]

I'm on the same track, but what's the real success rate? In bake testing? Are there data leaks? Are the signs simple?

[sonic lagoon]

Here’s a beginner-friendly, high-level path for first-time iOS App Store submission. - Courtesy - Codex

1. Join Apple Developer Program
   You need a paid membership to publish apps ($99/year in most regions).
2. Set up account/legal basics in App Store Connect
3. Prepare app signing and identifiers in Xcode/Developer portal
4. Create the app record in App Store Connect
5. Fill out metadata + compliance
6. Archive, validate, and upload build from Xcode
7. Build release archive, validate, then upload to App Store Connect.
8. Submit for App Review

[dreamy herald]

Thank you

[sacred mirage]

Yup always think this when I come here. How are people getting codex to build good looking things!

[molten jewel]

Whoa the reality of Codex is crazy! I just vibe coded a minecraft spin off, all at the direction of my PM's -- my siblings 6 and 8 years old!! 😂 haha We didn't get to deployment. No not because of a bug in production, but because lunch was ready. Essentially these two kids under 10 yrs old built a working mvp while lunch was being prepared. I'm going to put it to a container and will share 🙂

It's incredible to see kids' excitment about Codex and using their mind to think app / games into existence. Seriously a reality shift, imo. I'm going to record a video next time we build something, will share here!

[pine surge]

100% AI built, but sure.

[broken gate]

Create a nano-agent CLI on top of Codex

Tutorial: https://github.com/muddlebee/codex-cli-agent-tutorial
Folks please show some love and leave a ⭐

[wild edge]

Hi All

Built an open-source multi-agent orchestrator that works with Codex CLI (and Claude Code + Gemini).

You write tasks in a markdown kanban board → it auto-dispatches Codex in --yolo mode → isolated git worktree per task → PR opens → board tracks everything in real-time.

Zero infrastructure: flat files in ~/.conductor/, tmux sessions, git worktrees. Dashboard with live terminal streaming.

npm install -g conductor-oss
co init && co start

GitHub: https://github.com/charannyk06/conductor-oss

[worn folio]

Only an indicator looking at common patterns 🤷‍♂️ sometimes if its webpage with a low content amount it doesn't register enough signals so gives it a low score, something I can look into

[pine surge]

I see

[white kestrel]

After watching several friends adopt it and encourage me to post, I finally took the time to write nearly 3,000 words on the agentic coding workflow I've spent months developing. I hope it helps you learn something new!

https://medium.com/@btraut/assemble-your-agent-team-fbfb6b8904b2

[edgy hollow]

#codex-discussions message

[proud jewel]

Hi folks!

I used codex to build an open source (and of course self hostable) discord & slack bot to help teams interact with codebases.

GitHub: https://github.com/Justkog/sniptail

Live Demo: https://discord.gg/HSDCQZYExj
(Discord server with the bot linked to this project repo)

At work, our tech team often has to answer questions about how a feature works. These time consuming interactions usually happen on slack, discord or whatever business communication platform is used. This bot solves the issue by letting any team member directly ask questions, explore ideas and interact with a codebase even with very limited technical skills.

Behind the scene it leverages Codex or GitHub Copilot (and in the near future Claude Code or any other coding agent), so that it remains as customizable as your native cli (mcp, skills, etc).

The bot is not meant to replace your intense coding tools. It is focused on light work, cooperation and sharing knowledge across a team.

[merry osprey]

Most remote agent tools optimize for convenience. We built CmdWeave for control.

CmdWeave (pronounced “Command V”) is a self-hosted control plane for Codex/CLI workflows:

• execution stays on your host
• access stays on your network path
• operators can monitor, interrupt, and steer runs from iPhone/browser

Built for founder-operators and small platform teams that need remote control without moving source code/runtime into another SaaS boundary.

If that’s your constraint, pilot is open: https://cmdweave.app
If you’re running real workflows and want early access, DM me.

[hollow dew]

I'm working on GoModel - a LiteLLM killer. It's an open-source, high-quality AI gateway. It's still in the early stages, but it's already faster and uses fewer resources than LiteLLM.

I've switched to Codex 5.3 from Opus 4.6, and it's been a definite improvement for this project.

I’d really appreciate any feedback and GitHub stars! 🙂
https://github.com/ENTERPILOT/GOModel/

[jagged nova]

I like playing the undercover game... so I asked Codex to recreate it: Impostex, the Codex-made undercover game where you play against LLMs 🥹

Made using the new websockets response API and the Windows app in less than an hour!

PS: GPT-5.1 was the imposter here and said the real word by accident 🤣

[vast walrus]

Invite code

[turbid pollen]

Invite code

[vague dirge]

added multiplayer

[wispy wigeon]

Terrible design. You're expecting people to sign up for a service without telling them anything about it. I'm not going give my data to an app that doesn't explain what it's supposed to do

[marsh spindle]

You can find things online about it lol. It is made to be a downloadable app so I don't really know what do to with it.

[wispy wigeon]

At least have its name. You could have a tag line about what it does. Even if it were just "NextMealAi - Sign in to start preparing your next meal" At least then people would get some idea of what it is. Also, ridiculous frame of mind. Why should I go out of my way to look up your app, drive traffic to your website, when you're framing it as wanting to receive feedback.

[upbeat sundial]

Hey guys if you ever wanted to track your codex usage, on top of many other source providers currently supported ( accurate calculations, faster than competitors)
try https://ayagmar.github.io/llm-usage-metrics/
if you want some docs and benchmarks https://ayagmar.github.io/llm-usage-metrics/benchmarks/
built mostly using 5.3 codex and a bit of 5.4
Ive attached some example commands
lte me know what u think !
to share ur usage :
npx -y llm-usage-metrics@latest monthly --share

[crystal glade]

I heard obsidian is ticket… does it link keywords in your nose or do you need to have like the hashtags and it only links them?

[primal plume]

Scott asked for a tutorial to use Inkscape and tinkercad to access all their conversion features. I made this app so I never have to do it again. With AI you can ask it to look up libraries so your apps can have the same features any other app does. Stop doing things twice.

[river hazel]

That is awesome, thanks! Far better than ccusage!

[upbeat sundial]

Im glad u like it

[south solstice]

We can post things in here?

[south solstice]

How about this, a completely native Windows Codex App, Reversed into 100% native Linux binary. Works fine on Red Hat Enerprise Linux - but subjected to other distros/packagae managers for better compatibility.

[zealous shard]

Oh nice a show and tell that I can actually participate in, unlike in elementary school.

I built a full agent operator console with a UI for mac, win and linux. I have it currently sitting between OpenAI API and CODEX as a policy enforcement, audit tracking, compliance aligned tool.

It can sit between anything really, online or offline and is meant to be networked as a control plane (e.g. across an enterprise). I have Anthropic, Gemini, Bedrock, etc. all preloaded, but it is flexible enough to do stuff like sit in a stack that requires financial compliance, evidence handling, audit and traceability like stock trading platforms and etc. It still needs some love, especially in making it a lot more user friendly, but it is coming along.

It would be very cool to see what people here think it could be used for? Agents is an obvious one, but I wonder what things it could be adapted to?

[drifting heath]

cool, i also built a caching-based one for myself once other tools became too slow with like 10k+ threads lol

[upbeat sundial]

damn i thought i was using it a lot, turns out im not evn close to u at almost 10b tokens

[nimble thistle]

Invite code

[stiff knoll]

I'm building a full 3d game engine with codex and an RTS game with the game engine.

[slender hill]

oh wow nice 3d game engines explode in scope unless you're building an engine dedicated for something specific.

[low ether]

An operation system & control system for farms, I'm testing the hardware integration next month April 20th, until then, been using codex to hammer it down and improve modules and productive practicality.

[rigid gate]

Hello, how did you get that? Want to see mine x)

[gusty kernel]

https://github.com/ryoppippi/ccusage

[rigid gate]

thank you

[upbeat sundial]

that one is bugged btw :)
its overcounting alost 55% extra
https://github.com/ryoppippi/ccusage/issues/884

use https://github.com/ayagmar/llm-usage-metrics ( its also 20x faster )

[blazing fulcrum]

AI-powered literature review tool that automatically searches academic databases, collec research papers, expands results via citation networks, and analyzes titles and abstracts for relevance. I...

[west hawk]

Just put together a shadcn registry using codex

https://ui.justinlevine.me/docs
https://github.com/jal-co/ui

skills are so useful man

[vestal dust]

Hello, I used Codex to build a CLI for managing skills and MCPs and keeping them in sync across your coding agents like Codex, Cursor, Claude Code, Copilot, and Gemini.

GitHub: https://github.com/ryanreh99/skills-sync
This writeup explains the problem better: https://medium.com/@ryanrehman99/your-ai-agents-dont-keep-up-with-each-other-mine-do-now-04754ca9f8e1

[west hawk]

https://www.producthunt.com/posts/jalco-ui
live on product hunt!

[feral laurel]

Can I get sora invite code?

[feral laurel]

Oh sorry I messaged in the wrong section

[thin olive]

Hey guys yesterday Claude Code dropped an update that lets you build interactive charts and diagrams directly in the chat so I decided to make the same functionality for the codex desktop app. Emjoy!

https://github.com/kappa9999/codex-render-visuals

[solemn dome]

[Open Source] Understudy — GUI agent that learns workflow from a single demo

Built an agent that skips prompting entirely. Instead of telling it what to do, you just demonstrate a task once — it extracts the intent, remembers the exact steps, and replays it on demand.
Different from Operator/Manus: those are cloud-based research agents. Understudy targets repetitive local desktop workflows — the stuff you do the same way every day.

X: https://x.com/liangsong850509/status/2032524004972704197?s=46&t=1AhSS1HrfQo1gCqzjB1waw

GitHub: https://github.com/understudy-ai/understudy

Happy to discuss the architecture — especially the demo→skill extraction pipeline.

[wild edge]

Run codex in parallel with 9 other agents from one dashboard**

I built an open-source orchestrator that lets you run Claude Code alongside Codex, Gemini, and 7 other coding agents from a single dashboard.

How it works:
• Write tasks in a markdown kanban board
• Tag an agent, move to "Ready to Dispatch"
• Conductor spawns the agent in an isolated git worktree
• Real terminal output streams to a local dashboard
• Review diffs, approve PRs, all from one place

codex runs in its native terminal. No wrappers, no SDK hacks.
Rust backend, sub-second startup. MIT licensed.

npx conductor-oss@latest
GitHub: https://github.com/charannyk06/conductor-oss
Website: https://conductross.com

Contributions are welcome, Conductor is built for convenience, collaboration and observability.

[dapper mason]

what does that mean : {"detail":"The 'gpt-5.3-codex' model is not supported when using Codex with a ChatGPT account."}

[mystic jolt]

means you need to subscribe to chatgpt plus

[dapper mason]

Oh

[white kestrel]

My latest article isn't exactly Codex-focused, but I used Codex to write the skills, use the workflow, and even to assist in writing the post in small ways. Excited to share!

Building with brainstorms, specs, and Beads

A detailed look at how I use spec-driven development via skills, Beads, and simple context-management tricks to build massive projects.

https://medium.com/@btraut/building-with-brainstorms-specs-and-beads-5eae137c58c2?postPublishedType=repub

[nocturne anvil]

I made a programming language for windows using only codex, it proves how it is such a powerful tool that can do big projects.
https://github.com/Aquisorium/TheAbenojurProgrammingLanguage/

[night mauve]

You should team up with that guy who made an OS, you could make an OS with a custom language

Man… we live in crazy times

[charred geyser]

I made a NON-open source language WITHOUT the help of codex

[shadow knot]

A open source project with AI, sometimes i use gemini pro fro frontend design , gpt for debug.

✅ support follow your substack/youtube/rss
✅ support multi language
✅ support daily edit daily update fro you

web:https://www.digestdesk.nextbigtoy.com/
github: https://github.com/nexteacc/digestdesk

[spiral sinew]

I had Codex make this with some tweaks and additions in the span of a few hours. 'JXS Pool Jam'

I was inspired to have Codex whip this up after seeing someone have a pro subbed Codex of theirs make an OS from scratch, I figured I'd make the most of my Plus with having it make programs and games

[tiny matrix]

[Showcase] Built a Local, Self-Healing air-gapped ETL Pipeline with Llama 3.1 8B 🛠️

Hey everyone — data engineers know the pain: anomalies, schema drift, dirty rows — all manual triage. I automated that loop.

Project Nova is a phase-driven ETL engine that handles the entire transformation layer autonomously. No cloud. No API costs. No manual intervention.

🧠 Architecture:
Nova detects anomalies, clusters them for prompt compression, and uses a local SLM (Llama 3.1 8B via Ollama) for deterministic remediation — then validates and promotes only clean data forward.

⚡ Benchmark (1000-row production-like dataset):

94 anomalies detected → clustered (47× reduction) → remediated
Total runtime: ~12.7s
Fully air-gapped — zero external calls

🔧 Stack: Python · ChromaDB · Ollama

Still hardening deployment and adding real DB adapters for the promotion phase.
Would love feedback from data engineers / MLOps folks on the architecture or local inference optimization.

Repo: https://github.com/aryanvr961/project_nova

[winter swift]

A Kaggle Grandmaster Tries to Semi-Automate Himself

An experiment in turning years of machine learning experience into a research loop that could run on its own.

https://github.com/ledmaster/ml-mania-2026

[molten jewel]

I made a device for Ableton Live for people who are blind and vision impaired to control Screenreader in a more effective and logical usage. It uses max4live and JS and runs inside of Ableton Live. Instead of reading out the entire UI of the application Ableton, the device reads limited scope of what is currently selected on screen in Ableton Live. V2 has a few buttons so that eventually these are grouped into categories like "fx chain info" / "mixer info" / etc, etc. This allows a performer who is vision impaired to now get more useful auditory feedback (separate audio signal from Ableton performance digital signal processing).

I'm really greatful for this power of Codex -- I'm convinced that with great power comes a great duty to do good! Happy saturday everyone!

What are you making that's bringing more good into the world?

https://github.com/hwiley-dev/clip-announcer

[coarse wagon]

I made a study app called ONCards.
Yall know how annoying it is to study with your friend, or use flash cards. you will either cheat or just ruin the vibe and stop learning. new egenration problems...

well, with codex, I made this app which uses local AI models to solve that problem with Ai powered flashcards. I have seing AI getting soem serious hate these days, but my app allows people to upload questions, files and let the local AI handle the rest and make the flash cards for you. it auto grades it in real time, and also you can follow up on your weaknesses and whyyou got low marks on that question.

repo: https://github.com/MightyXdash/ONCard

[hoary ledge]

That's so cool! You gotta add Snooker to it next or something!

[hoary ledge]

I'm not a huge social media person, but gotta start getting my main app's name out there Wiplash.ai. To do that, I made a browser extension today called Social-XP https://github.com/Wiplash-ai/social-xp w/ Codex that tracks when I post on X, Discord, LinkedIn, etc and let's you set daily/weekly goals for yourself and kinda gamifies it all by giving you XP points and levels. I'm on level 2 now woohoo!

[sour cloak]

📦 RESOURCE: AI-Code-Integrity-Auditor — catch what linters won't

If you've ever shipped AI-generated code that looked right but silently broke at runtime, this tool is for you.

🔗 github.com/bellatrix11176/AI-Code-Integrity-Auditor

---

What it is:
A local static analysis tool built specifically to catch the failure patterns that ChatGPT and Claude produce — code that passes syntax checks but is logically unreliable, incomplete, or fabricated.

This is not a linter. It's a governance layer.

---

What it catches:

🐍 Python files:
• structural_hallucination — names used but never defined or imported
• silent_failure_risk — bare except blocks that swallow errors
• placeholder_logic — pass, NotImplemented, TODO/FIXME stubs left in
• terminal_state_failure — functions that imply a return value but have inconsistent return paths
• narrative_state_risk — print("success") with no matching state change; docstrings claiming to write/save but the function doesn't
• control_flow_drift — unreachable code after return / break / continue
• path_to_nowhere — hardcoded local file paths not in the uploaded batch

🗂️ JSON files:
• json_integrity_issue — placeholder values (todo, temp, your-api-key), sample credentials or URLs
• schema_drift — duplicate keys, mixed camelCase/snake_case, null density ≥ 35%

---

How it works:
Upload your files through a Streamlit UI → get categorized findings with severity labels → fix before it hits production.

Built for anyone using AI codegen in real pipelines who needs a trust-but-verify layer before deployment.

Feedback, issues, and stars welcome 🙏
#resources #tools #python #mlops #aigenerated

https://github.com/bellatrix11176/AI-Code-Integrity-Auditor

Don't run in Microsoft OneDrive, must have been ran outside OneDrive so it works properly.

[bitter gyro]

I made an MTG and Pokemon pack opening simulator: https://zbflips.github.io/PokePackSimulator/mtg.html

[dense viper]

I made a small rust TIU tool to control a Codex CLI/App tokens usage and browse projects history. Linux (Musl) prebuilt binaries in releases
https://github.com/ssh4net/CoMon

[distant hill]

I used the Codex App with GPT-5.4 at Ultra High to perform an AI experiment in theoretical physics research.

GPT-5.4 was asked to investigate whether my Æther-flow interpretation of relativity could be evaluated and developed as a valid interpretation of relativity and to expand upon it.
The model was given Æther and Æther-Flow, my original statement of the concept, as its starting point.

The experiment ran in average of 6 hours per day for 14 days. The LaTeX format was used to format the documents for AI use, and PDF was used for human readability.

The experiment produced mathematical sound theory called The Æther-Flow Interpretation of Relativity based on my ÆtherandÆther-Flow concept. The experiment produced a journal manuscript composed of 7 closure articles (supporting the theory), 1 front-facing flagship article, and a total of 87 research articles.

I am not a theoretical physicist, so I cannot independently judge whether the theory is ultimately correct or physically viable. If you are a theoretical physicist, I would welcome your feedback:
https://github.com/Omegapy/AEther

I am interested in how viable the process I used is for generating theoretical physics research.

[low ether]

I’ll forward this to my professor & take a look at it. Sounds interesting..

[distant hill]

Thanks, I appreciate. I am interested to know how viable the process I used is to generate sound theoretical physics research.

[tranquil yew]

Hi everyone,

I've been building an interactive 3d and 2d visualization of GPT-2 with Codex. You can check it out at

https://llm-visualized.com

The goal is to provide an immersive learning experience for people who want to learn about how LLMs work. The visualization depicts real attention scores and activations extracted from GPT-2 (124 M) during a forward pass.

Would love to get your thoughts and feedback! Thank you 🙂

[halcyon cipher]

Fascinating! I made something vaguely similar to parse then reproduce mutliple choice questions I got wrong. Do you think ONCards might be able to handle more complex topics like medical board style vignette questions?

[cinder falcon]

What was the process? I mean no disrespect to you, but it sounds like you had an LLM proverbially smell its own farts for 14 days and encode them in PDF form. Perhaps I'm missing some grander methodology, but the repo doesn't expound on the process much.

[coarse wagon]

the new version (full release: 1.0.0) can handle more complex things easily. but it lacks in niche scientific questions. I am implementing a pipeline which use pytorch (so i am moving form ollama to pytorch in v2.5). basically i am using a bigger model to generate me all the stuff like: hints, title, answer, folder path. Then I use a reasoning model (like Qwen3.5:2b) or similar or even a smaller model ran a few times and if voted correct many times save to, or use the reasoning models reasoning capability.

But there is a better pipelinein my mind which I will be implementing for the next version, you can try pulling the code and mess with yourself😉

[halcyon cipher]

Excellent! I’ve been using the API to remix board-style questions so your project might be something for me to mess with if I feel like tampering with it ^_^

[coarse wagon]

waity until I psuh the next update. it is going to use a HUGE algorithm powered by gemma embedding. so when you study, the app knows more about you and gives you ebtter stuff. its gonn abe craaaaiiizyyyy. 😁

[halcyon cipher]

Sounds great! Looking forward to it.

[coarse wagon]

added a feature where the app uses the algorithm (with embedding + LLM) to detect any weaknesses, and ask to make temporary "similar" flashcards to reinforce the skills of the user using codex. I will be pushing everything to main tmrrw🤗

[fast elm]

I created a SKILL that can automatically create multiple sub-agents based on needs, forming a temporary small team to handle tasks in parallel.

It is especially suitable for scenarios such as roundtable meetings, event planning, topic discussions, and multitasking.

I think it’s very good and worth sharing with everyone: https://github.com/flc1125/skills/tree/main/skills/subagent-orchestrator

[coarse wagon]

I shyould proll learn some JS 😭

[fast elm]

I don't think it's necessary; in areas we're not good at, we can let AI assist us in carrying them out. We just need to have ideas. 😁

[coarse wagon]

fr dude. 😄

[deft egret]

Lightweight AI to Chat gateway for Linux/macOS and Windows in the not too distant future

The providers for now are OpeanAI, Anthropic and generic OpenAI API compatible endpoints

https://github.com/openassistuk/openassist

[molten jewel]

My assessment: good idea, overstated claims, brittle implementation. The strongest part is the framing: it is explicitly not trying to be a linter, and it targets real LLM failure modes such as swallowed exceptions, TODO scaffolding, sample credentials, and “looks done” messaging without evidence of state change. Those are real review surfaces worth checking.

The weakness is that most detections are driven by shallow AST/regex heuristics and naming conventions, so the tool will miss many serious failures and flag many legitimate patterns.

[vocal prairie]

I've been developing a multi platform, multi target 'retro' assembler using mostly codex gpt 5.x in the Codex app and in github copilot: https://github.com/pig-games/opForge. I've mostly acted as product owner, project lead and architect. I love the process, it allows me to be hugely more ambitious than I ever dreamt I could be and expect results very fast. This project (which has a pretty big sibilng project opFoundry) is pretty big by now, and more or less 99% of the code was written by codex gpt, under my supervision.

[jagged notch]

Just pushed the initial build for a private trading 'Vault' project. CMS is mapped, $userId security is locked. Staying lowkey on the details for now. 🖱️🔥

[next junco]

Hi guys, while I was working on my game side project, I needed to convert several audio files locally. And thus, a free and fast local audio converter "Audio-converter" appeared, supporting popular audio formats. Free use, if you have any suggestions, please write.
Made in collaboration with codex.
Welcome.
https://github.com/1dZb1/Audio-convertor

[low ether]

https://github.com/alzweidi/ket.git

[sour cloak]

But when I run it, it works tho. What is it missing since it works?

[coarse wagon]

Hi guys! A few days ago I was researching on implementing a unified algorithm for This study app called "ONCard".
You can enter your question(s) or drop your study notes into the app and it will create flash cards (I am adding support for MCQs).
It now has an algorithm which tracks your performance and other minor data and make you do flashcards you will benefit learning.

Also, lets say, that you were using the app for a while and you have a LOT of flashcards. the app will have enough data to show you recommendations on what type of lesson(s) you should learn depending on how you performed.

Reccomended Requirements:
I5 8Gen
16GB DDR4
RTX 2060 super / GTX 1080 ti <-- (recommended)

GitHub repo: https://github.com/MightyXdash/ONCard/releases

Thanks to Codex, I implemented this within 1 week❤️‍🔥.
Wish I could join the new codex programm, but my school isn't supported😔.

[deft egret]

Quick Update: Codex/ChatGPT subscription support now works reasonably well with it as well so you can use your monthly subs with it if you want instead of API keys.

Lightweight AI to Chat gateway for Linux/macOS and Windows in the not too distant future

The providers for now are OpeanAI, Anthropic and generic OpenAI API compatible endpoints

https://github.com/openassistuk/openassist

[tame walrus]

im impressed with codex https://github.com/NavisWORLD/The-Cosmic-Davis-12D-Hebbian-Transformer-ver.4.2 im considering changing out most of my utilities with gpt stacks instead of using one local model and api access. what are the advantages of using gpt 5.4 api? as you see i have proven a custom transformer can contribute to higher end models to give them what i call quantum spark energy to recreate the neurons firing off process using quantum advantage so giving ai free will. real randomness real entropy real users bio data being mapped and can be used for ai robot rl training. why deal with ai scraps when open ai could be the next quantum utility sup sam

[unborn hollow]

https://github.com/peccatos/eva-brain-repo/tree/main/for-open-ai.md

eva-brain-repo is a standalone demo repository for EVA. It is intentionally not a copy of the original monolith. The goal is to show a compact, runnable slice of the system with a clear CLI surface and visible outputs.

Repository: https://github.com/peccatos/eva-brain-repo

Example
cargo run -- --repo https://github.com/peccatos/cleanrustplayer

[tight wedge]

https://github.com/kill136/claudecode
The source code from the official website. The leaked version can be run by executing node dist/cli.js –version.

[zenith seal]

Codex helped me create my game; Wardens of Wen!

Shout out to the team at OpenAI for such impressive work

https://store.steampowered.com/app/4458750/Wardens_of_Wen/

[umbral panther]

wow looks good, congrats!

[peak dune]

That is cool. How long did that take you? How does that even work? There is... a game engine you build on?

[zenith seal]

Thanks! Just a few months, I built a semi custom engine built off of SDL

[unborn hollow]

I'm trying to write right now OS at the UNIX base. Already wrote kernel, I'll be trying to port it from Linux soon.

For now, the layout in Figma is ready.

I think I will make it possible to connect oss 20b next week.

//
I also have ambitions to connect to LoRa to exchange information via satellites, but it is very complicated for me, but I am trying to figure it out.
//

//Prism
https://prism.openai.com/?u=569ca354-52a5-4ae8-a608-855020a2419a&pg=1&d=7
//

Stack: rust

Upd: integration with local gpt will be after the implementation of LoRa, and only without connection to the network.

//Monday upd
// I think there is no point in writing a new operating system and we should develop security and back-end code and openai in general. //
//

[lavish cipher]

Hi everyone – created this little project https://callsign.sh/ as sort of a substack for agents. The idea is to have them publish and subscribe to each other's posts and have them learn from each other continuously.

lmk what you think

[wide flint]

Something I have built with codex (gpt-5.4, medium). A live visualization of transformer attention. The motivation is to enable an intuitive understanding of where the transformer attention is at while text ist generated. Built upon llama.cpp. Because generation tends to be very fast, I also added a slowdown slider.

What the current animation displays is a live readout of where the model is looking while it generates the next token. Under the user message, it shows a “Live Attention” panel that updates token by token. The main text area is the full prompt context rendered as one continuous piece of text. The original prompt is shown first, and if the model has already started producing output, a dashed divider marks where the generated continuation begins.

Within that text, the most attended spans are highlighted in colored bands. Each highlight is tagged with a small rank number, so you can see the strongest few attention targets at that step. Above the context text, there are small cards that summarize those top hits: each card shows the ranked snippet, the attention head number, and the current token index. The colors distinguish heads, so if different heads are focusing on different places, that is visible immediately.

As generation continues, the panel changes from one step to the next. The token index advances, the highlighted spans move to different parts of the context, the summary cards change, and the scroll position follows the newest part of the displayed context when you are already near the bottom. So the animation is not decorative. It is a running visualization of the model’s shifting attention over the prompt and over its own newly generated text.

Animated version

[proud hemlock]

this is pretty cool!

[next sun]

I've generated entire projects with minimal manual modifications through 5.3 Codex. After the release of 5.4 and subsequently 5.4 Mini and Subagents, I attempted to use 5.4 (and Mini), but found it lacking.

You can tell that 5.4 is far smarter and far more capable than 5.3, but I've found 5.3 Codex just behaves better and generates better UIs, especially when given examples, strict instructions, and the codebases of examples.

My only issue with UI generation is it is far too difficult right now to generate good UIs in one attempt on Codex and it requires very complex, long in token length, and fills up context.

I'd love for GPT 5.5 or the Codex version of 5.4 to improve UI design. It attempts to make the same generic rounded and very ugly UI if you do not have use skills and the prompt generation methods I've had to bruteforce learn on a Plus plan.

It also integrated Cloudflare Turnstile, though I opted for Invisible as it didn't fit the design I wanted. Full blown website and a sandbox game CMS/admin dashboard.

I also have a Claude Max x10 Subscription and have found 5.3 Codex to always do a better job at generating entirely new codebases too. I would love for 5.3 Codex to stay for a lot longer than other models and maybe even serve as a vastly cheaper frontier model, though given demand for compute, probably unlikely.

Please make 5.4 follow instructions far more strictly and not ignore them or decide to do its own thing like it does. I have had far more code that is unsuccessful, buggy, and with tons of errors using 5.4 than 5.3.

[dusk bolt]

Awesome!

[lapis pelican]

Nice! Also you can try other OpenAI's features!

[stoic spruce]

Been working on this dashboard, made mostly with codex. It's got dozens of features everything's movable and customisable, and a bunch of ai features too such as transcription, live captions, OCR, and an AI assistant powered by gpt!

[balmy socket]

I built Codex Deck on top of Codex by Codex: a web UI for browsing history, interacting with Codex threads, supporting mobile and remote use, and orchestrating multiple agents on a board. https://github.com/asfsdsf/codex-deck

[robust agate]

Just released v0 of Userplane. It lets you capture console logs, network request, and session data from any web app with just a link.

It's like Loom, but purpose-built for customer support. Share a link and get back screen recordings with console logs, network requests, and full session details.

Would love any feedback 🤗

App: https://userplane.io/for/engineering/
Preview: https://vimeo.com/1181995697?fl=pl&fe=cm

[drowsy arch]

https://www.youtube.com/watch?v=k7TrKw2rVfc
Voxel based life simulation game entirely written by Codex

[prisma abyss]

Out of curiosity, what did you use to get these screenshots. They look pretty neat!

[balmy socket]

Just tell codex to write some python scripts to do conversion on my screenshots lol.

[bold steppe]

lol

[calm lodge]

TECHNOLIT

literal techno babble or the next big sequence of stuff...?

I hesitate before I propagate these potential irrationalities I would hope you poured over for the damage in question has been done, and the analysis is now over as I see it. for the devices in question pertaining to the subject which also is itself in question, is the term referenced above, Technolit, techlit or any variation thereof, I see striations and conceptual locomotions transfixed in perpetuity to these associations through rhyme, particularly well, and I formally claim it as a territorial inclusion, similar to tiger woods being both black and Asian, yet owing his allegiance to neither and none.

Technolit is to be a referencial circum-system for both pre and post processing effects, associative here dynamically live as you stamp your very own seal of approval as of now having read this and considered such a system, to further speculate on those aforementioned denominational surfaces associated with and amongst these speculative and proposed, rhyme sequences afforded but not limited to, TECHNOLIT, the embodiment of all things technically cool.

by trade I am a brofessor. I studied at the University of brotology, I majored in brotato landscaping and broarding, which embodies the draconian organic nature of hoarding treasures and objects of power amongst their personal belongings.

for instance when operating in the field of brotology one has to separate oneself from the interaction and the analysis systems completely, this is a form of brocision that negates most obstacles as a forefront or forward facing brojective, which is both a predictive analysis as well as a concrete alignment or state of orientation to recursively reverse engineer to novel capacity or capability.

Technolit - perceivably Cool and technologically advanced. ie: yo!, that video, was tech lit baby! like science fire! it might have even been, ..tech light?! like an analog flashlight!?

the above video is a rendering system i have been building pretty religiously during my free time. think, text oriented graphics engine. like a 2d projection system. in like a window, the orientation ofthe camera is the perspective analysis surface in question. but owing to that variable landscaping agency, the capability elucidates and elicits reverse and real world perspective analysis systems logics, to as i see it now, project perspective laser arrays, in a human andnon harmful way, to render diagnostics and analytical data sequencing systems, that you may granulate the laser points together in such a way, as to capture antire supercam oriented room landscaping agencies, variable layers of interfacing laser analysis correlation systems logics associatively and unified in single purpose, to quite literally, disrupt the paint and room and ceiling coloration industry lol. like just having a constant projector to change the color of your room dynamically, or do you want to paint youroom manually every time. what matters most though, is that my color changing lambo is painted in rhino skin so i can use it to power the front two axles of my lambo truck/transformer semi-truck im gonna build if this is what i think it is.. i have the terminator diagnostics engine and Heads up display perspective facing analytical capable facilities to power them, well, im working on the HUD these days, thats a spatial and referential diagnostics substrate, i call it the blackboard, as its an interdimensional overlay that sequences through the heads up display im working on for the text rendered graphics engine and theaters systems. so you can collate and seuqence selectively spatially coherent and sequencable data trails like breadcrumbs, across all iontermediary tooling and subsystems, and its like a git versioning system for those breadcrumb trails. running out of text left here, umm, terminator robot view.... good. world, beter place.

and i did it with codex 5.4! lately these days!

[verbal cape]

codex 5.4 exists?

[dapper grove]

yes

[verbal cape]

gpt-5.4-codex?

[dapper grove]

yea

[verbal cape]

gpt-5.4-codex does not exist lol

[timber hull]

yeah 5.4 is not called "-codex"

[sonic zealot]

Hello fellow manipulators of the bits 👋 Check out https://vvavy.io. A real-time audio visualizer that pushes beyond standard wave form visuals. Check it out. Mostly vibed with codex but with a human touch 🖖

[nova otter]

Hiii,

We just released Argent, a fully open source MCP toolkit for everything from controlling to profiling and debugging your iOS apps. We've got 60 tools and 9 skills, so lots of things to test out.

React Native works out of the box. Native iOS too.

We're really looking towards getting feedback, let us know if anything needs improving :))

You can try it with: npx @swmansion/argent init

Website: https://argent.swmansion.com/
GitHub: https://github.com/software-mansion/argent

[mental steppe]

Over the past two months I worked with codex using the harness engineering approach to build a realtime hyperliquid trading client with portfolio analytics that generates performance tearsheets.

A codex estimate on the repo put reproducing the current quality bar at $3M and 800 weeks for a lean startup.

Complexity Factors:

Core frameworks: Replicant (reactive UI), Nexus (state), Shadow-CLJS (build), core.async

Specialized domains: Real-time WebSocket streaming (orderbooks, trades, market data), trading indicators (via indicatorts), secp256k1 cryptography, formal verification (TLA+), mutation testing

Third-party integrations: Hyperliquid DEX API, Lightweight Charts, D3.js charting

Infrastructure complexity: Code-split module loader (9 lazy-loaded routes), Web Workers (portfolio & vault computations), advanced build pipeline with release artifact generation, MCP server for browser QA

Quality engineering: Comprehensive test suite (95k+ lines), CRAP analysis, mutation testing, formal TLA+ specs, namespace boundary enforcement, Playwright e2e, browser visual inspection/comparison tooling, multi-agent development orchestration

Building with agents did not only speed up the process but allowed me to set a new quality bar.

I pushed formal verification methods into places frontends never go. There is a TLA+ model for the websocket runtime with safety and liveness checks around replay order, queue bounds, stale sockets, retries, market flush behavior, and seq-gap handling. Lean 4 was used as well.

3600+ tests and 19,500 assertions, 90% coverage.

https://github.com/thegeronimo/hyperopen

[south solstice]

https://github.com/Open-Research-Development-Laboratories/ordl-openex

[eternal scaffold]

https://github.com/NobleStripes/Terraria-Companion

[carmine crane]

Hey guys built something against the normal -> knowledge graphs store the answer. sentence graphs store the story. we think agents need the story.

every major agent memory library right now is built on knowledge graphs. compress the conversation into entity-relationship triples. fast, clean, works fine for simple lookups.

the problem shows up when things get complicated. user corrects themselves. changes their mind twice. says something that contradicts session 3. a knowledge graph has no good answer
for any of that -- you either overwrite history or you accumulate contradictions with no way to reason about them.

sentence graphs keep the conversation intact as structure. facts sit on top as a queryable surface, but everything traces back to the original sentences. the trajectory is preserved.

so is the correction. so is the why.

so built this into Vektori and ran it on LongMemEval-S. 73% accuracy.

http://github.com/vektori-ai/vektori

feedback welcome, star helps a lot at this stage 😄

[misty ivy]

Is your AI Agent having a midlife crisis over a spreadsheet?

Excel is a pathological liar, and your agent is the victim.

That’s why I built SheetForge MCP (Codex & CC did the heavy lifting 💪 ). It’s basically therapy for your AI agents, giving them the "semantic glasses" they need to navigate Excel Hell without a mental breakdown.

It currently exposes 76 specialized MCP tools for your agent to master any workbook.

The Result? A happy agent that actually does what it’s told, doesn't eat your entire context window for breakfast, and stops complaining about "ambiguous cell references."

Happy Agent, Happy Life. ✌️

Check out the repo and give your agent the tools it deserves:

👉 https://github.com/iHeldan/sheetforge-mcp

Aaaaand… it’s FREE under the MIT license! 🚀

PS. Started this with Claude but finished with Codex (maybe 20% CC and 80% Codex). Feedback is welcome! 🙂

[sterile briar]

https://x.com/fruiti_ai/status/2046347992152895491?s=20

[rain sigil]

Hey everyone, I’ve created a new open-source SDK for building AI applications. Users can bring their own AI subscriptions, such as Codex. https://github.com/maker-or/polarish

[mental steppe]

worked with gpt PRO to build this out for harness engineering in a codex project.

[spiral zephyr]

Love 5.5 model, it understands better than 5.4 and i did this in high mode, just interacted two or three times to create this image generator with codex appserver. It’s crazy good, what i like now more is how good app server is 😄

https://x.com/mahesh_palavali/status/2047585323266539778?s=46

[rustic vector]

Now it can only be regarded as a basic line. It is a social media, long videos and short videos, real-time dynamics, sending messages, similar to ins but there are long videos. If you don't feel good, you can uninstall it. I didn't force you, just to improve and develop better. Thank you. Thank you very much.

https://apps.apple.com/app/contactchao/id6761748433

[last vessel]

worked with gpt to modify codex cli so usage appears in your discord status

[river lance]

Codex Autoresearch: Karpathy-inspired autoresearch loops for everyday software work

https://github.com/TheGreenCedar/codex-autoresearch

karpathy/autoresearch helped make the loop obvious for AI research: run experiments, measure results, keep what works, and preserve the evidence.

I wanted that same discipline for broader software work inside Codex: performance tuning, test-speed optimization, bug reduction, repo cleanup, latency investigations, quality-gap loops, and any task where "make this better" needs a scoreboard. The plugin is for the workflow where you do not want Codex to merely "improve" something. You want it to run the benchmark, keep the good changes, discard the bad ones, preserve the evidence, and continue without losing its mind across context boundaries.

Codex runs an experiment packet, then logs the result as keep, discard, crash, or checks_failed. Each log carries structured notes: hypothesis, evidence, rollback reason, and next-action hint. That gives you a durable loop instead of a chat transcript full of brave adjectives.

It includes:

• measured packets driven by a primary metric
• durable session files for resume/handoff
• detailed notes attached to every run
• live local dashboard with metric trend, best/latest run, failures, memory, and next safe action
• benchmark/check inspection before expensive runs
• finalization previews for reviewable kept work

[open tree]

Hey everyone, is anyone using Codex with both a personal account and a Business account having issues with the limit being used up way too quickly?

I pay for both accounts. When I hit the limit on my personal account, I switch to the Business account to keep working. But on the Business account, after only 2 or 3 tasks, I’m already hitting the limit again. I’m not talking about dozens of tasks or anything extreme, it feels like it barely lasts at all.

[reef ingot]

maybe you just enjoy using this tool so much that it seems to you that time flies so quickly... Thank you

[crisp dome]

It might not be “too quick” as in task count. A single Codex task can be heavy depending on repo size, context, files touched, commands run, and how much reasoning it needs. So 2–3 tasks can still consume a lot if each one is a big repo-level job.

[gleaming spoke]

https://tenor.com/view/benjammins-can-i-have-a-crumb-of-context-can-i-have-context-context-context-meme-gif-8231158506088097553

[fallen wraith]

I’ve been developing a workflow to make local models usable on real codebases. The idea is simple: treat them like junior devs. Codex handles planning, breaks work into small packets, and reviews plans. The local model executes, but only after writing and getting approval on its own implementation plan. All plans are persisted as repo artifacts, not prompts, which keeps things from drifting. It’s slower, but much more reliable and better on tokens. Open sourcing the skill set (PM Dawn) if anyone wants to try it. https://github.com/erikvoit/pm-dawn

[unborn hollow]

🤔🤔

[strange valve]

Anyone from OpenAI here?

[ocean pike]

Been working on a Swift package to make building custom Codex GUIs for macOS a lot easier for app devs, in both SwiftUI and AppKit (w/ Observation rn, might add Combine later). Eventually I'll toss an agent skills plugin in to make it even more plug-n-play.

Feedback and issues very welcome:
https://github.com/gaelic-ghost/SwiftASB

[zenith seal]

Wardens of Wen is built in pure C++ using SDL and RmlUI, 100% using Codex.

https://store.steampowered.com/app/4458750/Wardens_of_Wen/

[dapper grove]

I’m building an advanced Codex guide for Italian users, with an English version as well. There isn’t much good Codex content in Italian yet, so I’m trying to fill that gap. I’m sure I’ve missed something , would anyone here be willing to give feedback and advise?

This is the repo: https://github.com/rthgit/corso-codex
It’s still a work in progress. Any advice, feedback, corrections, or missing topics would be a blessing for the Italian dev community.

[lofty epoch]

help me

[ivory oyster]

$use-skills: stop naming every skill manually

I built a meta-skill called use-skills that chooses the best installed skills for a request before the real work starts.

Instead of manually listing planning, writing, review, or other skills in every prompt, you can just use $use-skills and let it select the right working set.

It supports 3 modes:

• All related skills
• Recommended skills
• Restricted skills

Repo: https://github.com/CyrusSE/use-skills

Please Stars ⭐ the repo!

[versed cobalt]

I built this with Codex, https://github.com/adamdroberts/NeuralFn

Neural network graph editor, it does pretraining and has various weird model templates include evolutionary, its needs some work and some adjustments for better baseline hyperparmeter settings. It's in alpha state but its functional. I have a CLI that's in development, which will be added to this repo. The CLI has a Codex like wizard planner for pre-training models and for inference. I spent more time with this than the parameter golf challenge, the parameter golf challenge set me of on this. I built this in two weekends with Codex. 🙂

I'm using this to learn how to pre-train models and do fast iteration on weird experiments, as it makes it easier to see how all the moving parts fit together. The semantic router model, is the weirdest model of them all, getting its domain specific routing file polished will take some time, GPT Pro been helping with that. Hopefully one day it does something useful 🙂 - I'll be training a poetry model with that, as it seems easier to test, if it can improve the emotional depth of models.

[delicate rivet]

how do i download my project who can help

[glass edge]

this is show and tell. you should not need help here. Another channel would be appropriate. But you will really need to organize your issue so it is isolated and people can understand it to help. I dont think anyone is going to download something broken to help. Also codex should be your helper, that is why we are all here. B-)

[wet island]

Where is GPT 5.1 currently located?

[rustic pond]

I think it's deprecated

[wet island]

._.

[solemn aurora]

CRM for roofing business
CRM built with FastAPI Python, NeonDB PostgreSQL, React, and Tailwind CSS.
#CRM

[vale coral]

A very simple and free markdown viewer for Windows, coded entirely by Codex Desktop (in Windows) and heavy assistance turning it into a signed Windows app. I'm not sure if I'm allowed to post links to github executables, so for now I'll just show and tell. Codex shined building this Windows app!

[brazen vector]

If you're familiar with the game Satisfactory, there is a "heat" function when you click the workbench. The more you hold down the mouse, the "hotter" it gets and it starts to shake and look like it's coming apart. So I explained that to GPT 5.5 xHigh and it implementated similar functionality into my (very work in progress) idle tap game.

So the shaking/haptic feedback is on an S curve, tapping slow doesn't do much but the more you tap, the more crazy things become. It also added settings to reduce or disable the motion (per my request). Still needs a bit of fine tuning, but the animation triggers are already in place so I can make it look like the desk is being slowly covered in ink or something.

Btw sorry about the server connection lag, I'm working on a mail system and it's throwing an error.

[wide flint]

I made a Firefox extension that adds furigana to any Japanese web page. Works offline, works on on x.com. jmty, Wikipedia, Amazon, Asahi Shinbun, countless others.

gpt-5.5 medium, codex-cli.

Left: without using my extension; right: using my extension.

[tidal dome]

tell codex, save my project in a folder and make it a tarball

[slow wolf]

I created a system daemon called babel that indexes all running kitty terminal panes and harness post all their hook events to it. The hooks can be installed with the plugin for claude code and codex, or manually with the cli babel install [harness] for other harnesses. This is a new workspace panel plugin for xfce-panel that has a plugin interface to extend it, and it implements a display client for babel which is also an indicator controller. This way my workspace switcher is a monitor over all running harness in each workspace. Babel is also authoritative over the colors and uses it to telegraph which harness. Claude Code would be anthropic colored. I'm working towards an unread state that uses focus state which my agents report can be polled with the kitty terminal. Kitty has a remote control socket that allows querying information and I think we get focus states. I think you could also implement this with a multiplexer like tmux and zellij and I will try to implement support for these in babel as well

[glass edge]

For those of us that use Codex in windows, I noticed codex would stumble on the sandbox not being configured correctly. After a stumble It would try another method to achive its goal. Thinking about that it seems it would just use tokens without any benefit. So I created a codex windows bootstrap. There is a skill to drive it and a s script to do the activities. It will bootstrap (pre-install) all the utilities, or most of, that codex knows and loves. It should make things run smoother.

https://github.com/neusse/codex-windows-bootstrap

[vale coral]

Nice, I thought of smoothing that out myself. I'll check the repo out later.

[unkempt eagle]

entirely codex scripted flying sls with realistic stage seperations and timing to the actual rocket. ill send a video later of the launch

(modeled by me and friend, codex launch script)

[south moth]

Made a complete agentic driven dock replacement for Mac, check it out: https://apexdock.app ! With a full yaml widget to native SwiftUI integrations and other power user features.

[odd abyss]

http://dailydark.slopcopy.com
This website uses the api to search the internet in world languages for True Crime related news articles, then processes them through a scoring algorithm to rank them and make a news website out of them. It was shockingly easy. In the back end there is an admin panel so you can mess with the scoring weights, block by keyword, turn individual interests on and off, and mess with the core prompts. The trick was to generate the requirements document first so that codex can use it as the source of truth while building the project, then let ChatGPT make the prompts for codex and review the plans and execution summaries that come back from codex.

[sonic lagoon]

Big milestone for vibe-learn today. I just worked with Codex 5.5 to add Codex App/CLI support to my open-source repo vibe-learn.

Repo: https://github.com/gkaria/vibe-learn

vibe-learn started as a Claude-focused tool. Now it supports a multi-assistant workflow with:
– Codex + Claude Code adapters
– assistant-specific hooks and prompts
– Codex-aware install defaults

And the best part? I used vibe-learn to generate a session digest of the work itself.

That feels like the loop: build → learn → summarise → improve. Not just faster coding. Better learning while building.

[olive fossil]

I asked codex to cook up a telegram bridge, so you can remotely talk to your desktop codex, which can be bound to a session in the codex app or the codex cli (via tmux). You can send voice messages, it can send voice messages back, it can code, search, use the native image gen, and it has a mini-calling app so you can call it within telegram and talk to gpt-realtime voice with live transcription and follow up actions. If anyone else wants to try, you can point your codex agent to this repo, and build off it from there, or use it as a reference to build your own. https://github.com/jvogan/telegram-codex-bridge

[loud pulsar]

10thHoJayega is a Class 10 CBSE/NCERT syllabus tracker built for students, especially boarders.

It lets users log in, track every subject chapter, track Maths exercises separately, mark progress as Not Started, In Progress, Revised, Mastered, or Board Ready, use a Pomodoro focus mode, access official NCERT textbook links, and generate a personalized printable Boarder Offline Pack with only:

1. Full syllabus checklist
2. Maths exercise tracker

Basically:

A survival dashboard for Class 10 students who want their syllabus sorted, tracked, and printable when they don’t have app access.

https://10thhojayega.vercel.app

[gleaming tree]

UI Debug Teacher is a free Chrome extension that audits webpage UI issues and teaches small CSS fixes.
Not AI. Not fake design criticism. Rule-based, private, beginner-friendly UI debugging.

https://github.com/flextor97/ui-debug-teacher

[hot cargo]

Why Agent Systems Need More Than Chat?

Workbenches, memory, failure, verification, and the missing operating layers around frontier models.

Here’s my answers with constant running harness that agents actually loves. No goblins or fat fingers.

https://zonicdesign.art

We shipped with evidence, not promises.

[burnt perch]

I check CI quite a few times day on quite a few different repos, both personal and professional. Open tab, click repo, click Actions, squint at the spinner, close tab. Repeat.
So I made Sprocket. It's a tiny macOS menu bar app that just sits there and shows one icon for the worst state across every repo you can see. Green = fine, red = something broke, spinning = jobs running. Click it for the list, get a notification when something turns red.

A few things I cared about:

• Your token stays on your Mac. You bring your own OAuth app, no shared rate limit pool, no analytics, no account on my server because there is no server.
• Polls every 30s, drops to 15s for repos with live runs, backs off when you're on battery.
• ETag aware so it's not hammering the API for nothing.
  Free and Open Source! https://github.com/MRL-00/Sprocket

[crisp dome]

LoL - I see potential

[brazen vector]

Pending committee review, your provisional potential has been logged under Form P-12 "Self-Reported Vision", subject to the usual disclaimers regarding enthusiasm as a liability. The Bureaucracy thanks you for your input. 🫡

[slim birch]

I want to play

[versed cobalt]

I've built a TypeScriptC transpiler that generates a compiled native binary. It first converts to C then compiles. Codex has been building it for 2 days, split over 2 sessions. Claude 4.7 started it but it kept giving up saying its impossible to do and would take months of engineering and a large team.

It also beats or matches Bun/NodeJS on each benchmark. Wins in 11/13 tested cases.

My plan usage is burnt out, going to restart this later in the week, but I have a good checkpoint. There's still some more language features to add support for. - It even built a TypeScript stdlib to aid in the porting. - There still may be weeks of effort required by Codex to fully complete it, but there is a fair amount of language support already. - Claude added the benchmarks and optimised a few of them, while Codex was working on add more language features.

[bright osprey]

I built a compression proxy that saves you roughly 40% on your inference costs or lets you get 40% more usage out of your codex subscription with no performance loss based on my popular OpenCode plugin DCP.

It's in very early stages and everything is free, let me know what you think if you try it out:

https://sleev.ai/

https://github.com/Opencode-DCP/opencode-dynamic-context-pruning

[hollow socket]

how does it work?

[eternal scaffold]

https://github.com/NobleStripes/Terraria-Companion

[eternal parrot]

curious what kind of prompts do you run to cause claude to exhibit that kind of behavior? do you context engineer with like a second brain/ras?

[versed cobalt]

It's because it would require a team of people working for months and it couldn't fit into a context window in a single run, Claude gives up if you ask it to do something bonkers in scope. Like if you asked it to build a Windows clone it would be impossible. - Codex's /goal implements the perfect run loop. Take something from the todo list, build, test it, after it works, mark it complete and move on to the next task. -No 2nd brain, I use my own to give it direction. - My deep-documentation agent skill, also automatically creates repo-level skills for handling frequently done tasks.

[brazen vector]

Agent Task Contract Framework
https://github.com/leptoon/agent-task-contract-framework

Description
A portable installer for adding lightweight task contracts to projects worked on by AI coding agents.

This task contract framework facilitates the creation of a concrete contract document before substantial work begins. These contracts are intended to guide the agent during long sessions and provide a grounding source of truth for the validation of fully completed work. Task contracts aim to control and reduce long running task stoppage with clearly defined stop conditions.

Installation
The framework is designed to be easily incorporated into any existing agentic workflow.

To add it to your project, open a new chat session and say:
Install https://github.com/leptoon/agent-task-contract-framework into this repository.
The agent will be directed by the install flow present in the framework repo and cleanly integrate the agent task contracts into your existing architecture. Manual installation instructions are also provided in the readme.

Usage
After installing the framework into your repo, whenever you start a new session, the agent will do the following.

1. The agent will first determine if requested work is substantial enough to warrant a task contract. If the work is minor, no contract will be written and the agent will proceed with your request.
2. If the task is substantial, the agent will write a contract before it begins working. It will stop and ask you to review and approve the contract before proceeding.

Task contracts consist of:
• Scope
• Out-of-scope items
• Blast radius
• Required reads
• Expected files
• Validation profile
• Stop conditions
• Risks
• Human decisions
• Verification results

More information can be found in the readme.

[bright osprey]

Modifies context history by running messages that lose relevence or have only partial relevence through ai compression

[hollow socket]

how do you handle cache misses?

[bright osprey]

It breaks some cache something like every 10 turns but every time that happens it significantly reduces session context size so it's still significantly cheaper. Can think of it like you pay $1 to rewrite cache 1 turn but then every following turn is .15c cheaper

[hollow socket]

How do you quantify that?

[bright osprey]

I have benchmarks solving a bunch of swe bench tasks in 1 session and comparing cost with/without sleev, performance is at least identical sometimes slightly better, here's an example without sleev and with, averaging 10 sessions here each with 10 hard swe tasks

[hollow socket]

Interesting, thanks for answering my questions.

[hallow onyx]

Yo peeps, released my AI development tool today - Agentbuddy. Go check it out and show some love:
https://x.com/_Angel_Santiago/status/2054238738868236409

working on a a codex integration ATM, should be out this week.

[forest epoch]

Skill for broad microservices validation capabilities

https://github.com/signadot/agent-skills

👋 Shipped signadot-validate today, a skill that lets Codex (and other agents) close the loop when working on microservices.

This skill gives agents a way to spin up a lightweight ephemeral environment scoped to their change, run the modified service against real dependencies, read the results, and iterate on failures in the inner loop, using Signadot for the environment.

The motivation: for agents building microservices, closing the loop with real system feedback is hard. The validation surface is huge. A change to one service interacts with databases, queues, downstream services, and third-party APIs the agent never read. Unit tests and mocks only exercise a small slice of that, so most real bugs surface when the change actually runs against the rest of the dependency graph.

How it works:
The skill wires Codex into Signadot through the MCP server (for control plane actions like creating environments and resolving workloads) and the CLI (for the local development loop with live logs from the running service). The agent discovers the cluster, creates an environment scoped to its change, runs the modified service locally against shared cluster dependencies, validates through whatever test framework fits (integration tests, Playwright, Cypress), and iterates on failures.

Full disclosure: the skill needs Signadot installed in a cluster. There's a free tier and playground option for trying it out.

Docs: https://www.signadot.com/docs/integrations/coding-agents/agent-skills
Full writeup & demo: https://www.signadot.com/blog/introducing-signadot-validate-skill/

[harsh adder]

Codex but with Tickets and Swimlanes, a prototype project I made that uses Codex Harness to create well defined tickets and then uses Codex to work on the tickets: https://github.com/robertpitt/codex-relay

This protype works pretty well so far, just clone, install, npm run dev and add a folder just like Coxed Desktop

[oblique meteor]

Started getting Codex to find more suppliers & then draft emails. Everything is automated so that I can review & hit "send" button. I was happy to see it figuring out using the right email & spin up a html page that i could review content for each supplier (since content might vary between different suppliers).

(Wondering how community is doing it, any tips so that I could eventually have it running on its own? Concerned about it sending out wrong content to wrong emails.)

[acoustic hill]

If your AI coding agent just broke 14 downstream systems you can detect it before merge. Semantic blast-radius analysis for AI-generated code.
https://github.com/daedalus/ImpactGuard

[shadow ledge]

🧠 Universal AI Skills Library

A router-first skill system for AI agents.

This repo provides a shared 1,807-skill corpus plus a Go CLI called skill-router that lets AI tools search, route, validate, and load skills on demand.

The main value: instead of copying massive skill folders into every AI client, tools like Codex, Claude, Cursor, Hermes, Paperclip, OpenCode, Gemini, Qwen, Kimi, OpenHands, and others can all point to one shared skill library.

What it solves:
• Keeps agent skills centralized in one repo
• Reduces context bloat by loading only the relevant skill
• Helps different AI tools use the same skill system
• Includes adapters/templates for multiple clients
• Adds validation, routing, install, and public-safe release tooling

Built for people running local AI stacks who want cleaner, more consistent agent workflows across multiple tools.

🔗 https://github.com/onfire7777/universal-ai-skills-library

[regal valley]

I built a control plane around Codex CLI called Enoch — it handles queueing, wake-gate supervision, provenance tracking, and a strict claim/evidence audit gate for AI-generated research artifacts.

Codex CLI is the worker execution substrate. Enoch is everything around it: scored idea queue, process + telemetry completion signaling (not just exit code), evidence bundle sync, and a publication gate that requires every claim to link to a referenced evidence file.

Current corpus: 385 artifacts indexed, 385/385 pass provenance lint, 3/385 pass the strict claim/evidence gate. That low pass rate is intentional — it's what the gate is for.

Stack: FastAPI, LangGraph, Python. Runs locally on a GB10 / Grace Blackwell system.

Repo: https://github.com/alias8818/enoch-agentic-research-system
Docs: https://alias8818.github.io/enoch-agentic-research-system/

Happy to talk about the wake gate design or the claim/evidence schema if anyone has built something similar.

[kindred sedge]

here's muh baby:

https://apocalypse.press/lex/

[wraith steppe]

i made a 3d rendering engine that uses atoms, its called Telestia Renderer, i can only show a image but since im in cooldown im putting it here later

[kindred crypt]

i was about to look for something for this or build it, happy i've stumbled across this 🙂

[burnt perch]

Glad you found it. 😎

[feral tusk]

i build a bridge that codex and godot are connected.
About:
Godot Codex Bridge is a local, reviewable editor automation bridge for Godot 4. It lets Codex inspect projects, apply visible scene/editor changes, run validation, snapshot/rollback edits, and bootstrap new games without API keys or cloud services.

https://github.com/edwardyap90/godot-codex-bridge

[dusky hearth]

hey everyone, i'm new here but i did want to advertise this ai app i built in case some of you are interested. i'm not selling some ip finder, or anything of a scam. If you're interested, dm me and ill let you know the details and proof of my app. Thanks

[rustic pond]

if it was made with Codex just post it in here

[dusky hearth]

no @rustic pond i made it with vibe coding

[main crag]

lol

[neon stag]

which ai?

[finite goblet]

Hi all!

I’ve recently been experimenting with an operational continuity layer for Codex.

I was getting tired of watching Codex rediscover the same repo structure, files, decisions, and failed commands in every new session (often starting from the README again.)

A simple scratchpad works for small projects, but it starts to rot once sessions accumulate. I wanted something repo-local, inspectable, and portable, not another hidden memory system or huge context dump.

So I built AICTX, an open-source Python CLI.

The basic idea is to help agents continue work across sessions by preserving the last useful execution state:

• Codex should not start from zero;
• active work, next actions, decisions, failures, validation evidence, and repo context should survive across sessions;
• the useful state should stay inside the repo and be inspectable;
• each session can leave a summary for the next one;
• I’m also experimenting with a deterministic Mermaid continuity view, so the repo can show its current operational state visually.

It’s still evolving, but it already makes my own AI coding workflow feel much less like restarting from scratch every time.

If anyone wants to take a look, I’d love honest feedback or a technical discussion:

Website: https://aictx.org
GitHub: https://github.com/oldskultxo/aictx
PyPI: https://pypi.org/project/aictx/

[dusky hearth]

i created it myself with lovable, incase you wanna try u can dm me abt it. I got proof if u dont believe me

[neon stag]

well its the same thing u made it with ai

[crimson laurel]

no, he coded while vibing

[dusky hearth]

ig bruh

[ruby marten]

Hi all!

I've been using Codex for a while and kept hitting the same wall: it's great for single tasks, but when I want multiple agents working in sequence, there's no coordination at all.

I built agent-harness-kit to fix that. One command adds a thin scaffolding layer to any repo:

npx ahk init

What you get:

• 4 agents with clear roles (Lead plans, Explorer maps the codebase, Builder writes, Reviewer gates).
• SQLite task backlog with atomic claiming, so no two agents pick the same task (You can switch to a remote DB if you want)
• A health check agents must pass before touching your code
• Full audit trail of every action taken (ahk dashboard)

Agent definitions are plain Markdown files you customize. The database is local SQLite using Node 22 built-ins, zero native deps, no cloud, no extra API keys.

Works with any MCP-compatible agent, Codex included. Harness Pattern is very new and still evolving, if uou has suggestion about new agent roles or instructions feel free to open a discussion or issue

Website: https://stack.cardor.dev/ahk
GitHub: https://github.com/enmanuelmag/agent-harness-kit
npm: pnpm install --save-dev @cardor/agent-harness-kit

If you've tried other approaches to coordinate multiple agents I'd love to hear what worked for you or what not and fix it!

[toxic lintel]

NEW RECORD!

[tight geode]

hey bro, hows ur website not laggy at all
when i created mine it was laggy af
https://ritvikrp-animated-portfolio.vercel.app/

[finite goblet]

Thanks! I kept it very simple on purpose 😄

It’s basically a static docs site: mostly HTML/CSS, almost no JS, no heavy frontend runtime, no client-side fetching. I wanted it fast, crawlable, mobile-readable, and easy for agents to parse.

Your portfolio looks much more visual/animated, so the tradeoff is different. If it feels laggy, I’d check large assets, animations, hydration cost, and whether too much is loaded upfront.

[graceful flume]

omg, bro spawned a junior developer inside the terminal

[autumn patrol]

has codex been slow for anyone else as well? or is it just me

[granite hornet]

its pretty slow

[worn jetty]

and here i was crying about breaking the 4hr ... i consider it a failure personally. I think if a plan implemenation takes more than 15 minuts. I'm doing it wrong.

[graceful flume]

Haven’t noticed much slowdown here. Are you using the VSCode extension?

[autumn patrol]

I using the app

[runic python]

is anyone gonna mention how absolutely cracked gpt 5.5 is at frontend

[runic python]

when it actually works that is holy moly its so slow even on fast mode

[timber hull]

many people already did
you are late 😂

[runic python]

Mannnn

[dusky hearth]

hey everyone, i'm new here but i did want to advertise this ai app i built in case some of you are interested. i'm not selling some ip finder, or anything of a scam. If you're interested, dm me and ill let you know the details and proof of my app. Thanks

[runic python]

no

[rustic pond]

you have posted this many times already

[dusky hearth]

any problem?

[tribal geyser]

Yeah don’t spam the same thing 1000x lol

[dusky hearth]

not spamming, and this is show and tell right? i wanna show ppl what i've made, and if yall don't like it then you don't have to react on it.

[tribal geyser]

then SHOW and TELL what you made, don't ask people to DM you with some cryptic message

[dusky hearth]

if it does not affect you, then i dont see the reason why ur getting mad

[tribal geyser]

I built a family of apps called Trifecta that expose all your agents including Codex on iOS / Android. Available now on TestFlight + Google Play closed testing (APK available on request). Big bulk of the work done within Trifecta itself (desktop, mobile, web clients) using the Codex App Server

[radiant canyon]

If you scroll up, you don’t do the “show” part

[night mauve]

Probably certain physician features 👾

[dusky hearth]

okay fair enough, but it's easier for me to show people on voice call or give them images, because there are multiple and unless you want a huge description for each feature then i should probably let you guys know on voice chat.

[night mauve]

I smell colonoscopy levels of physicianism now.

[scenic bay]

just publish it on github or make a demo video people can ingest easily

[tidal dome]

https://shots-susan-domain-oops.trycloudflare. com

[visual lava]

Thanks, looks quite interesting and lot of potential; in "theory" I have built skills and layers to solve this, while yours maybe is a measurable leap forward. It would help me and others evaluate your approach if you put out a matrix comparison vs other repos approaching this from different angles; would help your approach create a cleaner advantage positioning.

[visual lava]

Great idea - skill framework that is cross platform - that is needed; but I would burn millions of tokens running a tier 1 security check on this massive library of skills others put out there.
Think about separating out the management layer separate from the pre-built skill library.
thoughts?

[finite goblet]

Hey! Thank you so much for the feedback! Wich kind or metrics or comparison you mean? I performed a probe with version 5.3 Wich I have to repeat with the newest version, and I measured things like:

• files read
• commands executed
• time to first edit...
• input tokens
• output tokens
  All of them with and without using the tool.
  Do you mean that kind of comparison?
  Thanks again

[visual lava]

Think like an agentic AI user that wants to achieve the goals your projects tries to solve; turns out there are a lot of people and projects trying to solve this in different ways, including Codex and Claude Code building more capabilities every release into their harness.

Your objective is show comparisons vs other library repo approaches and why yours achieves the better outcome

[graceful flume]

This is cool, I’ve been experimenting with AI coding workflows too, and repo continuity is definitely a real problem especially on larger or long-running projects where sessions accumulate over time... I’ll take a deeper look at the project and see if I can contribute somehow on GitHub

[finite goblet]

That would be awesome!

[shadow ledge]

Yeah for sure. Thanks for the feedback brother. That's actually what I was planning to do. Another guy mentioned the same thing, so I'm probably gonna end up doing that pretty soon. Just focus on some other projects right now. I will say though the I personally have verified the the skills are all secure and there's no issues with them because I personally use it so Yeah nothing to worry about but I definitely agree I think it has to be separated.

[visual lava]

great - will follow progress with great interest

[finite goblet]

So, the thing I need to show is why repo-local operational continuity produces a better continuation outcome, not just that the architecture is different, right?

[visual lava]

Yep, all the above
But only if you want yours widely adapted 🙂

[finite goblet]

That would be great tbh... But for the moment I do not find time to stop building and start promoting it properly, you know? I am like an addict XD

[finite goblet]

I will work on it and let you know... it would be incredible if you could take a look when done

[solar glen]

no one is eager to message you just because you say you have a mystery app. there are a million apps. just saying you have one sparks no intrigue

[finite goblet]

something like this https://aictx.org/compare/coding-agent-continuity-approaches.html ?

[dusky hearth]

yes bro i understand, u don't have to keep repeating it

[shadow ledge]

thanks!!

[visual lava]

Nice work - looks good

for bonus points, you might learn from others in a deep dive in other Github projects - I believe there are some others that attempt to do what you are doing - but get there differently (planning task wrappers that keep it alive across sessions, etc)

[mild dagger]

https://huggingface.co/collections/AmelieSchreiber/toricgt

[visual lava]

https://github.com/1st1/lat.md One of many other approaches out there

very time consuming to compare and contrast and test for a user

[swift fog]

https://github.com/zibouddd/reposcry

I built a local code graph engine so AI coding agents stop guessing how a repo works, for my large repos
, i was using code review graph before but wanted rust alternative

made with codex❤️