Codex 0.107.0 is here, and with it: FORKS🍴 This minor update includes the ability to fork your current chat history directly with custom multi agents with just a quick prompt. "Fork this session using n subagent(s)" This unlocks some interesting use-cases where you can launch multiple agents that retain... all of your chat history up to that point, each in their own independent session. For example, you've been discussing architecture with Codex for 20 minutes, talked through constraints, reviewed existing code, settled on a model. Now you want to see two different implementation strategies side by side. Fork lets both children start from that full shared understanding. With a normal spawn, you'd have to re-explain the entire context to each child, and they'd inevitably miss nuance from the earlier conversation - which is basically a form of compaction/compression. With something like this, you could also just tell it to fork itself 5 times, and implement the same UI/frontend 5 different times in their own worktree to see which one you like the most. Other potential uses: - branching after reviews - try and rollback patterns using worktrees - parallel debugging strategies Otherwise, this is a minor update with improvements to the app server, realtime transcription, and memory. Enjoy!show more

desperately need an /undo command.

It's actually really good about undoing things but you got to be really careful that it doesn't run git commands But I totally agree with you

Can you tell it what task must do each one and what reasoning level use? Like, one do security review, one for UX review and another one for best practices skills.

You sure can. You need to define custom agent roles in your config file first, and then you can have it call any combo you want. Codex is incredibly dynamic now. Enable multi agents in /experimental

Nice. Going to build a skill to build / customize the config files, so that way I can your pre-built agent roles to hire/fire agents on spot. Hope that I don’t ended up firing half the agents 😂

haha nice! i actually built one already

Any subagent roles catalog site? Like but for roles.

forking mid-session is huge. been doing this manually with Claude Code—copy context, spin up parallel agents. native support cuts ~5min of setup per branch

fork preserving full chat history is the right move. i run a multi-agent system and the 're-explain everything' tax on each new spawn was killing my flow. does fork keep tool call results or just the conversation?

it keeps the whole session

Does fork automatically creates worktree?

No but you canjust ask it to

When will the remote control be available?

my guess is less than 4 weeks.

I have burned 5% of my Pro quote in 1 hour. This is dangerous! :D

just 5%? you're not trying hard enough 😂😂

Im trying to contain myself! It is Monday still!

LOL nah, let it rain tokens

Just give us teams (put the fries in the bag)

😅 you can make teams now (i love fries)

Oh for some reason I thought you were on the codex team 🤦‍♂️

But you can make teams, its not quite as straightforward as Claude, but I could teach you how to do it. [agents] max_depth = 2 (or more if you're brave) then you can define teams and launch them by calling a "manager" (which would be a lvl1 subagent) which calls a "team" of defined subagents (which would be depth level 2 agents). You can definitely do this.

Mmmm but can they talk to each other? 👉👈 this is more like the shit rolls down hill model

yes they can talk to eachother, but they dont really communicate that often no matter what harness you use. even in claude, the parent agent gives the subagent a prompt and the subagent does the work, and returns a summary there's no meaningful 2 way conversation in either case the only time communication might happen in the sense that you're probably thinking is if the subagents hit a blocker.

That's really cool. Any indication of when this will be available in the IDE extension?

It's probably available now. Works in the App You can test it like this

🤩 I definitely have some use cases for this, thank you 😊

rock on man. get after it.

Will do 💪

@badlogicgames Codex implemented /tree except it's prompted?

Realtime transcription has been really nice! I've been using it a lot lately.

@Snixtp It's so nice. I have had transcription hotkeys for a while, so its nice not to have to do that anymore

Yup!

Can you give me best setup for codex spark

Read this!

Session forking for parallel implementation strategies is the killer use case here. A/B testing different architectural approaches with shared context is something I've wanted for months.

lol I wanted to ask for this feature yesterday and today it is already implemeted

@ulanyde That is fasssssst

this feature is insane 🤯 can't wait to try it out!

context forking should also help the main agent use subagents potentially much more reliably

I honestly dont think they're related but i have found that codex calls them SUPER reliably now. What are you running into exactly?

no issues, i just think this can help with multilayer subagents (eg sub-sub-sub-subagents) to be more aligned

fasho.

Wow my favourite feature from the pi coding harness is now available in codex 🥳

pi is amazing.

This is incredible. Great list of use case ideas.

Thanks :) That's my goal with these breakdowns is to really explain what's going on under the hood. Just saying "forks now available" isn't enough to visualize the utility.

They want me to use codex app but keep releasing all the juicy features in the CLI... Do these translate to codex app? Can I tell codex app the same thing since it's using my CLI under the hood?

@slopwareindy Yeah this works in the app too!

Oh interesting. When do you think rewind will be coming to codex app? It's so annoying loading the CLI just to do it and then having to restart codex app for it be recognized in the app.

wow

Should be very positive for Windows users that use Windows Terminal as well. A fix for the diff background issue we had on Windows Terminal is in. On top of that, you can add your custom themes and we'll support their custom diff backgrounds as well!

can we use subagents with different models yet

Yes you can

Hey there!

Hey hey!

I have been forking manually, like a handoff. Not with subagents. I assumed it creates a new worktree and the code doesn’t merge back into the main thread.

It doesn't merge it back into the main thread but that's different than a work tree

hope they make my idea: while agent doing your task, it will ask you many questions as it can and it will continue to loop and do what you picked from the options

That is more or less what plan mode does now, but it's not building it in real time. Ideally you really want to zoom out and look at the whole plan before you go and start implementing things. That said you could still do exactly what you're asking about right now through the use of skills they don't need to implement anything for you to accomplish this. You can turn on the ask question tool and build a skill to do exactly that

available in Codex app?

Should be, I'll test it later for you. But you could just turn on multi-agents and try it if you want. The litmus test how you can know it's working is just ask it what the first message in the thread says, it should be able to tell you even though it doesn't show the history

can you please make ctrl+v an image from clipboard to the chatbox a thing ? It's the only thing that's stopping me using codex scaffold.

@ReyArtAge Hey Rey, I don't work there, but you can already do this for quite a long time :)

Yeah its really useful!

For this to work, should I have the "Multi - Agents" feature enabled?