Upgrading to FlowState 2.4 🚀

Howdy.

Abe Burnett, founder of flowstate.

Here, in this video I'm going to show you what it looks like when we upgrade Flow State.

So in this case we have a test project.

It basically only has Flow State in, is the flow state version 2.3 and we're going to upgrade it to flow state version 2.4.

So I already have dropped the flow state version 2.4 folder in here.

We better do a, because this is private.

Okay, so it is a private repo and you can see that the names have been obfuscated so kind of named to kind of benign, boring names.

these are the same files that you will see, that Flow State creates in a public way.

It just makes them look innocuous.

Right.

And kind of passe and easy to kind of just scan right over.

Right.

So first thing we're going to do to do the upgrade, we are going to open the file.

And one thing you'll notice is that I use a bunch of different tools.

So I do use, codecs a lot.

I also use, quadcode a lot.

But in this case, since a lot of my customers use and, or will use GitHub Copilot, it's a very popular tool.

it has some interesting, nuances.

so I have had to work extra hard to make Flow State work consistently and reliably with Copilot Copilot, even with the same models.

I encourage you if you can afford it, if it's, you know, if you want to, I encourage you to not just assume that what you get out of Copilot when working with say, Cloudsonnet 4.5 is Cloudsonnet 4.5.

Because if you use Claude code with Clodsonnet 4.5, the behavior and the performance will be noticeably different.

Like it's crazy.

so just keep that in mind.

GitHub copilot is cheap though, so I get why people use it.

but I'm primarily just using it just to test Flow State.

I prefer to use Claude code or codecs.

but you get a lot of utilization out of GitHub Copilot.

like 300 credits or whatever on the fancy models is pretty darn good.

but yeah, it's also weird because it burns context like tokens rapidly.

I think it feels kind of like, like, I don't know, bad and what's it called, perverse incentives where like they have these tokens that you can spend 300 of them.

This is a diatribe or not tokens, but credits or whatever.

But inevitably the more you burn your token budget, the more credits you're going to spend.

And so it's kind of like it, they give you 300, but then they're going to use them faster.

The system is designed that way anyway, so that's beside the point.

We're going to proceed with the upgrade using GitHub Copilot.

And I've worked really hard to make this so that in theory it should, respect the privacy settings as it does the upgrade.

so it should kind of, you know, right now we have one exposed folder.

but you know, at the end of this we should be able to do a git status and see that everything is hidden.

And yet the upgrade is in place.

So let's proceed.

So basically the first thing, like I said, we open the upgrade file, we grab this prompt, we drop it in here.

Yeah, we're ready, let's roll.

So one of the things that GitHub in particular has to do is in order to Preempt when GitHub Copilot does kind of a lot of compaction cycles relative to its peers.

That's another thing you'll notice is if you use claude code or codecs, they manage their context a lot better and you don't have to do compactions as frequently.

Also, claude code has a much more graceful compaction handling.

and so coupling that with flow state actually works really well.

but GitHub copilot with its frequent compactions, I needed to design flow state and the upgrade process in specific to be kind of compaction tolerant, so that it would not break and make a mess.

So it creates kind of an upgrade state file and then it frequently writes to that basically to so it always has a place to restart should a compaction occur.

Okey dokey.

And then we're just going to check here.

Upgrade analysis files to be upgraded, new files to be added.

Troubleshooting is a new thing that, that contains steps you can take to get flow state back on track.

If it gets a little lost, there's a prompt or a number of prompts in there for helping with that.

files to be preserved.

Okay, good.

Detected customizations Private git ignore configured scaffolding level growth meta rules fully configured Phase 0 project structure with custom task definitions.

Risk level low.

And I am looking forward to at whatever point GitHub Copilot includes some kind of idea of like it exposes your token budget.

I'm going to try not to like, do my normal.

What do you call that?

Fidgeting.

We're going to do comprehensive here.

Do a immediately after upgrade.

I'm fidgeting.

Let's stop that for the video.

People don't need to see your head bobbing.

and I do it again.

Okay, let's stop that.

I'm just a fidgeter.

What can I say?

Okay, cool.

It doesn't.

It doesn't see it.

It's there, but it's just not.

So one of the things that GitHub Copilot will do is it can be a little overly, overly literal and kind of a little boneheaded sometimes.

So it is there.

It's just missing it.

so basically it's going to add it in a way that it will recognize as being there.

And that's fine.

so we're going to say a actually be.

So this is new.

thanks, Greg, for the idea.

Basically the idea is that for folks like me and Greg, we really, we really prefer.

Even if it's slower, we really want flow state to discipline and kind of organize AI, right?

Like, because, I mean, that's part of, like, why I created this.

It's why, you all have purchased.

Flow state is because AI, like agentic coding or vibe coding without flow state is a mess.

It's chaotic, it's frustrating.

So with flow state, it organizes that, but the current behavior is more like this balanced, right?

Like, it works pretty well.

But Greg was asking, can we make it even more rigorous?

Right?

Like maybe it makes it a little slower or maybe even a lot, depending on the, size of your code base.

But like, from time, and from time to time, some things will take longer, some things will be shorter.

But can we make it basically do even more kind of basically, you know, updating documents so it drifts less, so it stays on track more, so it doesn't forget, so that it kind of does more to kind of let you know where it's at and what's coming next.

and give you kind of prompts like just lead you along, right?

Be like, here's what you could run, right?

Even if it's like, yeah, just giving you kind of more to work with.

So I'm going to choose methodical.

That's the slowest.

but highest quality setting, fast is.

I don't know if I'd do that because that's probably more toward like what we have with agentic coding without flow state.

And so, but if you want to try that, that's fine.

And just keep in mind you can change it, like it says here, anytime, by just saying change quality mode to mode.

Right?

All right.

But for now, we're going to do.

Very good.

It's not going to create git checkpoints automatically copying framework files.

Actually it's supposed to put them in a.

So you notice that I don't have to.

I don't have to act on that.

I can say, hey, you know, at any moment, if something doesn't seem right, I can interact with Flow State and AI and challenge it, have a conversation.

Right, I'm going to stop that.

I want you to put them in.

This needs to be the case for obfuscation and privacy.

So, you know, this is one of the cases with GitHub Copilot that, it's just been like pulling teeth.

I have not figured out a way to make it work reliably.

because it comes to this point and it just does not understand the nuance of like, if it's a.

If it's like a pub.

If it's a project where you're fine with flow state being, out in the open, then it can put the files at the project root.

And in fact it probably should be, at the project root, just project root.

Just so you can see like, the flow state things, they're front and center.

They don't get lost and buried in a folder somewhere.

and maybe I'll change my mind on that.

Maybe it'll just become.

Maybe I should just change it.

Actually, that's probably what I should do in a future release is just change it to be like, always put them in dot flow state.

Private or not private.

That would solve this problem because it's just.

It's just really annoying.

but, yeah, just.

But this is a good example of how, I have been able to interact with it.

And honestly, like, if at any point along this process you want to ask questions or you wanted to tell it to do something differently, you can do that.

because one of the benefits of Flow State is that it's not a rigid.

It is kind of like AI software in a sense.

Like, it's an upgrade for like how AI behaves, but it's also flexible.

And so that's very meta.

Like you can.

Instead of just downloading like an app on your phone and being like, the app is just what it is and if I like it, great.

And if, I don't, I'm just that.

That's Tough because I can't actually modify it myself with this.

In this case with Flow State and AI, you can, you can, you could say like I don't want you to work like this and I'd rather you work like this.

What would it take for you to be able to do that?

And then it would give you some plans, some suggestions and then you could work, go back and forth until you felt good about where you had landed and say great, please implement that.

And then it would customize your version of Flow State.

That is okay because even future editions that I release will be non destructive.

So they will do their best to respect and honor any customizations that you have made.

Let's see where we're at here.

So we can see that it created this folder here, copied all the files into it.

That's looking good.

And the Flow state is in the exclude get exclude file so that it's totally private and hidden.

And you'll notice too that the compaction happened.

And the way this is designed, it actually was able to continue with the upgrade like in a, in an, in a graceful way.

Okay, so some stuff to read about how it works.

This whole idea of scaffolding like and quality because there is an, there is an interaction effect.

Right.

So quality is essentially like the method, the like systematic approach that Flow State will use.

How it methodically will kind of work through things like how it runs.

Right.

Whereas scaffolding is communication rules and like how it supports you.

So this is a good real world analogy like your car's driving mode.

Scaffolding is like setting your GPS voice verbosity.

so it shows kind of the default and also my choice here.

So this means I'll thoroughly test thoroughly before every change.

I'll validate assumptions explicitly.

I'll wait for your proceed command before risky operations.

I'll capture comprehensive evidence.

And there's also more hooks so that it, it writes more to test or not to test to tasks.

Well I forget what the file is called.

It's called.

What is it called?

Tasks.

tasks.

so it writes more frequently to that it's better hooked into the various AI adapters.

and then Session state it writes to every kind of whenever a major feature is finished.

Yeah.

Okay.

What just happened?

Why the sequence?

So this is kind of like the scaffolding kind of explaining what's happening and why.

And now it is doing kind of adopting this new behavior where like customers asked to have it be better about asking more questions.

Right.

And giving you options.

And so that's what it tries to do, that it does it more now, even if it's still, you know, it's AI.

So it's never perfect.

Right.

all right, so just say defaults are fine.

Yeah.

Well, also you'll notice that like, I am pretty comfortable with just stopping things if I don't like how things are going, I will just stop it and chime in.

and with flow state at least it will pick up and continue fairly gracefully.

All right.

Upgraded successfully.

and then we'll say yes to remove that.

And I know the answer is we don't need the upgrades, we don't need this file.

But I'm going to just make the AI do the work.

Then we still need a.

There we go.

And then.

Ta da.

We have a successful upgrade to version 2.4, which brought a variety of functionality which it walked us through how to enable that.

It surfaced the functionality and then asked us how we wanted to like, what we wanted to do.

And we have now version 2.4.

So probably what I would do is refresh flow state or like start a new thread and then do like resume flow state to kind of make sure that you're starting, fresh with 2.4.

Good.

Excellent.

Yeah.

So sometimes, like I think one of my customers was saying that they wish that it would give more.

sometimes it does this.

What would you like to do?

And it's so it doesn't take an opinionated stance always.

That happens more with GitHub Copilot.

So if it drives you crazy, I worked really hard to get, get, flowstate to do this less even in GitHub Copilot.

But I think GitHub Copilot is more ambivalent, I guess is how I'd put it.

And it tends to kind of look at like various options and be like, well, that's fine, this is fine, that's fine.

Like me.

I mean, and so when it doesn't see a kind of a clear prioritization of and then that, and then that it doesn't give guidance.

And so here it's like it looks at this list and it's like these things are all equally important.

Pick one.

Right.

so yeah, if you like, I've noticed in cloud code and in Codex that it does a better job of asking more questions and giving specific suggestions.

Like, I think those, the same models in those tools are more like reasonable or opinionated, I don't know.

But just Something to keep in mind.

but yeah, you can see that you are off to the races and so there you go.

You could also just hop over to a new thread and say resume flow state.

okay.

It's going to resume the, that's because this file is active.

That's why.

All post 8.

Yeah, that's my bad.

I had the upgrade file still active and so it was like, okay, I can do that.

There we go.

Start task show status plan feature.

Cool, cool.

Anyway, there you go.

So if you have any issues, with the upgrade, any questions, please, as usual, find, us in the Discord.

Or you can email me directly ablowstateai.dev.

did I say that right?

Yeah, abe@flowstateai.dev but yeah, I encourage you to start with Discord, or the resources page I'm going to put up that's going to be linked into the Discord as well as on the emails that customers this week or next week for like the week of Thanksgiving 2025, will get, which will contain like, FAQs.

It'll contain videos.

So you can see how I do a lot of this stuff and answer your own questions and then, you know.

But of course my door is always open if you prefer just to come directly to me.

but you may find that answers to your questions are already in those, FAQs and videos or certainly if you search the Discord or ask there.

So you have a variety of resources and, thanks for watching.