How are you using the Codex app on Windows?

[matheuspeppers]

Until now, my entire AI agent coding workflow has always been in WSL2, using tools installed within WSL and keeping the project in the WSL file system itself. I use command-line applications such as Opencode, Claude Code, Codex CLI, and Copilot CLI.

Since the Codex app for Windows was released, I was very excited because my biggest bottleneck so far has always been having a more fluid workflow for working with multiple agents using work trees. The Codex app would solve this problem for me, as working with work trees through it is a simple and fast process.

However, the workflow of the Codex application using WSL has been terrible. First, it is slower when the project is in the WSL file system. And if I move the project to the Windows file system and use the integrated terminal with the agent configured to run on WSL, it gets a little faster, but anything I do in VS Code becomes extremely slow.

It has been very difficult to find a configuration that allows me to have a smooth and reliable workflow with AI agents. Ideally, in the same project, I would not be limited to using only Codex, but could also use Claude Code and Copilot, which are the ones I use the most.

An alternative would be to migrate everything to native Windows, but I have never tested that. I’ve heard that it’s not as good because AI tools tend to make more mistakes with Windows terminal commands, as they are more complex. I also know that the ideal scenario would be to use macOS, but that’s not my current situation, and I can’t migrate to it right now.

So, I wanted to ask you: how have you been working with the Codex app for Windows, especially when working with working trees? Are you using a fully native Windows setup? Are your projects on WSL or the Windows file system?

I would like to understand your settings so that I can arrive at a configuration that allows me to have a smooth and reliable workflow using the Codex application, and still be able to use Codex CLI, Claude Code, and Copilot CLI in the same project if necessary.

[EricGT]

I have not tried using Codex app for Windows so can only infer that if it works with Codex or Claude Code via the CLI then it will also work correctly using VSCode extensions for Codex and/or Claude Code.

If that is true then consider using VSCode with the extensions so as not to need the CLI.

For coding my setup is

• Windows 11
• VSCode
• VSCode entensions for Codex and Claude Code (Wish I could disable co-pilot, it just gets in the way)
• Git
• VSCode workspaces
• WSL or WSL 2 or such is not used.

Yes the generated Windows commands can be problematic but often the LLM get it correct. Updating the memory files such as AGENTS.md and CLAUDE.md works, but I also prefer to just create a MEMORY.md file and use that.

What makes it even hard for the LLMs is that I often do not setup the environment PATH variable to find needed tools which makes it hard the the LLM to find them often failing. By listing the full path in a memory file, the problem is often fixed.

I may use a LLM CLI once a week if absolutely needed.

HTH

[winsock]

I run it natively on Linux, Red Hat Enterprise Linux.

[amogus]

I just tried to use the native Codex app for Windows myself, and the result in this usecase was no good. This project I’m working on uses devcontainers and the devcontainers are Linux dependent. Ok, usually that’s no problem, I clone the project on wsl and work in there with vscode or jetbrains IDEs. But now I would need to open the project in the Codex app but it just freezes when trying to open a folder inside wsl.

My takeaway is that it’s not possible to use the Codex app if you need your project to live on WSL. If I’m wrong, I would be happy to learn the workaround, thanks.

[vb]

I started using the Windows app a few days ago and it has already become my main chat interface for coding projects on Windows. I would not want to be without it now, but it is still not quite there yet.

I still do the actual coding through the CLI because I have not been able to get past a permissions bug. In practice, I have to approve PowerShell commands manually, which feels odd and is a real blocker.

It is very useful for sending tasks to the cloud and getting a best-of-four style result when I am out of my depth and need a few additional perspectives.

It also feels much better to copy and paste reviews and plans from 5.4 Pro into Codex Chat than to deal with the current CLI workflow.

What I do not understand is why folders are still not a basic feature. After only a few days, I am already back to using complicated naming schemes with multiple variables just to keep track of my chats, and that does not scale when there are too many of them.

Overall, I am a fan, but my heart is still in the terminal, where I can control permissions the way I need.

[thinktank]

I’m having trouble getting everything to work properly. I dunno if it’s just a bug that’s being solved, and I should note I haven’t read all of the Codex documentation yet.

My first-look impression is that it’s pretty cool. I genuinely think there’s a lot of potential with those Automations.

Universality

I understand the overall goal with Codex is having a universal way to work with code, and like @vb says, it’s close!

I think the single most confusing point is that each separate entry point has a place for my system prompts and stuff. I have a developer prompt on the web, and in the window app, are there separate prompts for each, the VSCode, and CLI as well? … Are all of these different prompts being stacked upon each other when I call something?

It would be great if there were a way, or a command like the MCP servers for your whole Codex account, I’d similarly like my account wide prompts, and account wide skills / subagents would also be extra wonderful.

Sandbox Permissions and Planning Mode

I dunno if this is just happening today (3/20/26) but I’ve been having permissions issues across codex entry points.

I started a thread last night in planning mode on the cloud, which did what I’d normally expect, with a list of suggested tasks to open in their separate threads. When I went to prompt it to continue from the thread, codex complained (a lot) that the permissions weren’t set properly. I was confused until I realized it must be talking about the planning mode.

Planning Mode on the Cloud

Also, Panning on the cloud and locally behave differently. On the cloud, you get a list of executable tasks that opens up in a new thread.

I’ve had issues with this behavior a couple of times. Like, what if I just want you to go down your own planned list and execute (the behavior from VS Code)?

Though, you know, now it would be pretty cool if you could assign subagents to that task.

Subagent Launch

I’m not sure if this is all related to a bug or set-up errors (very likely), but I just clicked the “try the subagent” button in the windows app. You get a prompt suggestion, “use a subagent to do such and such.”

I tried it, and got a bunch of confusing errors:
”””
That’s a cool setup, but I can’t actually do it from this run for two separate reasons:

Hard limits in this environment

1. I do not have a subagent-launch tool here.
   I can use the tools exposed in this session, but there is no facility available to spawn independent GPT-5.4 mini subagents.

2. This run is still constrained to read-only QA.

“““

So, on the one hand, there could be just bugs floating around; on the other, the documentation shows some default subagents (which i’d like to know more about); but doesn’t say anything about enabled tool use.

The custom subagents look pretty dope, and I’d definitely want Account-wide ones that I can call everywhere

Otherwise, I’m pretty excited for where the windows app is going.

[thinktank]

Hey everyone! I did a deep dive into the Codex docs and stuff and found the bug:

config.toml

Here’s my window’s app config compared to my ~\.codex\config.toml file.

See? “Sandbox settings” is not being overridden by the default config.toml already in ~\.codex\. But, were I to change the setting in the Windows App, they DO change the config.toml.

So, when I opened a repo which I’m using in all the Codex entry points, the Windows App overrode my “workspace-write” preference, and it threw a wrench in my whole weekend!

Naming Conventions!

Also, while we’re here, can I please encourage y’all to be consistent about naming conventions?

If Personalization == AGENTS.md in ~\.codex, it would be VERY USEFUL to know that.

… The config has a helpful link to the local file, why not good ol’ personalization.

I should note that the docs are similarly unclear about this matter.

My cloud-based prompts, like my cloud system prompt, isn’t appearing anywhere in my ~\.codex, which is definitely the source of deep confusion with regard to system prompts. It would be most excellent if there were a way to sync all of those files, prompts, agent definitions, preferences. There’s a way to sync “agents”… but what does that mean? I’d really like to be able to sync my entire user profile.

Applications

This bug is going to cause a lot of problems downstream for anyone thinking they’re going to build a codex app somewhere, then have a client install it on their windows app for day-to-day use. (The way you detect skills is EXCELLENT!)

I also think better on boarding for the Windows App is called for upon install: perhaps the Apps (windows, apple) are more “customer facing” by default—a bridge between the scary IDEs, and the more friendly ChatGPT interface. Which means, don’t assume the users know anything.