Slightly Caffeinated

Chris Gmyr (00:00)
Hey, welcome back to the slightly caffeinated podcast. I'm Chris Gmyr

TJ (00:03)
I'm TJ Miller.

Chris Gmyr (00:05)
T J, so what's new in your world this week?

TJ (00:07)
man, this week has been crazy since our last recording because my wife had her double mastectomy. right? That that was upcoming.

Chris Gmyr (00:16)
Yes.

TJ (00:17)
Yeah. So that happened last Friday.

everything went really well. you know, day of both surgeons were super happy with how everything went. you know, my wife had a rough recovery that night in the hospital. but we got her out the next morning, and you know, we just had her post-surgical follow-up yesterday, and everybody's really happy with how it's going. so I

As good as could be expected so far, which has been a real blessing. Yeah.

Chris Gmyr (00:47)
Nice. Yeah, I'm glad everything

went well and she got back home pretty quickly and was

TJ (00:51)
Yeah.

Chris Gmyr (00:52)
resting at home. That's way better than sitting in a hospital for multiple days.

TJ (00:55)
Yeah.

Yeah. That was a real rough go that night. But yeah, she I so I took Friday off for because that was the day of surgery. It took Monday off just to like, you know, have have another day to ease into everything, which was helpful.

you know, just kind of spent Monday like making sure I was all caught up on like chores and everything. Cause with she she's got tons of like weight restrictions. She can't lift. She's got some mobility restrictions. So I just like wanted to push hard and get everything taken care of. So that was nice. and then yeah, other than that, man, I've just been hammering well, I've I've not been, I've had Claude hammering on

V three of Halcyon's sync engine. So we can talk about that a little bit later. That's been kind of fun to watch that happen and let that go. so yeah. Yeah, and then and then just dig jamming at work, building like a bunch of really cool stuff right now. so having having a ton of fun there. Yeah, how about you, man?

Chris Gmyr (01:48)
Nice. Yeah. that sounds great. around here, yeah, just been I don't know, I feel like another week of craziness. we had

TJ (01:55)
Sure, yeah.

Chris Gmyr (01:56)
my wife's birthday yesterday, so I took the day off, and just handled a lot of the kids' stuff so so she can kinda relax. But

TJ (02:04)
Awesome.

Chris Gmyr (02:04)
also had like a tree company here yesterday, like doing some trimming, so they were here like the majority of the day. to one of our many house projects that we're trying to tick off.

TJ (02:14)
Yeah.

Chris Gmyr (02:14)
And then yeah, just busy with work, a couple of the side projects, which we'll talk about later. and yeah, caught the snake on Monday.

TJ (02:23)
Yes Yes

Chris Gmyr (02:25)
so yeah, my wife and daughter were out when I was taking lunch and because I had the camera on, I got a notification and it's like, he he's out. And my wife was like, Whatever you do, just call the people, don't don't mess with it. Don't like I don't

need you

TJ (02:39)
Yeah.

Chris Gmyr (02:40)
like

bitten or whatever. just

TJ (02:43)
Right.

Chris Gmyr (02:44)
just don't be stupid. Leave it alone. And I thought about that for a minute.

And then I went outside with my bucket

TJ (02:51)
Yeah.

Chris Gmyr (02:51)
and got like a big stick and I have a a big green egg for smoking and grilling and s and things like that. And

TJ (02:59)
Yeah.

Chris Gmyr (03:00)
it it comes with a hook to clear out the charcoal underneath 'cause everything

TJ (03:04)
sure.

Chris Gmyr (03:04)
like settles down and it has a little hole opening in the front. So it has this like little hook scoop thing.

TJ (03:11)
Yeah.

Chris Gmyr (03:12)
and that's like the only like hook type thing that I could find like really easily. So I

Grab that stick bucket. and after I don't know, I felt like forever, but it was probably like two minutes that I was out there and wrangling this thing, and luckily it wasn't like as big as I thought it was, but able to kind of like pinch it a little bit together and drop it in the bucket, put the top on. And then I called

TJ (03:33)
Nice.

Chris Gmyr (03:33)
the wildlife guys and I'm like, it's in a bucket in the back, just come back and get it. I don't care about the bucket, keep it, just get it out of here.

TJ (03:40)
Just get the snake out of here.

Chris Gmyr (03:42)
So yeah, and then I had like a ton of meetings on Monday. So it was like ten minutes before that. So my like nerves were shot and just

TJ (03:49)
yeah.

Chris Gmyr (03:50)
like it took me like probably two hours to like settle back down a little bit, but got it in the bucket, got it out. and then when I came down before I started dinner, we have this like copper mesh stuff that we use for various things. So I took a bunch of that and just stuffed it in the hole that he was like going into.

so hopefully that'll keep it patched for a little bit and no other critters we'll we'll try and get in there

TJ (04:14)
Yeah.

Chris Gmyr (04:14)
until we get like a an actual affix. But my wife was not happy about it, but at least the snake is gone. I called her like as I was heading up to my meeting and I'm like, don't be mad.

TJ (04:24)
Yep.

Chris Gmyr (04:24)
But I got

the snake and she's like, you are an idiot.

TJ (04:28)
But it worked, I got it.

Chris Gmyr (04:29)
It did work and it wasn't it wasn't too bad. so I wouldn't call myself a a snake capturing expert, but yeah, at least got a good story out of it.

TJ (04:36)
For sure, for sure, man.

geez. And then yeah, my son starts school next week too, right? Like you've got you got that coming up as well, right? Or you're already back, right?

Chris Gmyr (04:42)
We're nope, we already we already started, yeah.

TJ (04:46)
Yeah, yeah, that's that's coming up as well. So yeah, just a little chaos, I think, this weekend running around doing back to school, clothes shopping at the last minute. And and then

Chris Gmyr (04:56)
Yeah.

TJ (04:57)
his my son's birthday is the third. and then

We have his birthday party on the fifth. I don't know if I did I tell you we're I I bought a pack for him and his friends to go to Laser Tag. So it's like

a

Chris Gmyr (05:09)
yeah.

TJ (05:10)
big laser tag place. They got an arcade too. So the like the party pack came with like a bunch of credits and laser tag and stuff. So that's gonna be a a total riot. Yeah. It'll be good. Yeah.

Chris Gmyr (05:18)
Nice, that'll be fun. Jealous. sweet.

So yeah, any new coffee stuff over the last week?

TJ (05:26)
I'm back on the Kirkland stuff. But at least it's not the terrible dark roast that my wife got. It's their like Kirkland artisan blend, which is actually not so bad. so we're on that. I still I haven't gotten around to like looking into the logistics of the coffee from Australia yet. It's on it's on my list, but I've got I've got some other financial priorities to take care of. So I I would love to get some more of that though. I don't know how about you, man.

Chris Gmyr (05:53)
yeah, did some co-working on Tuesday at the Turkish coffee spot. So got some Turkish coffee and a pour over, which was really good. it was one of the ones from counterculture, because they have counterculture there. I don't remember

TJ (06:08)
Nice.

Chris Gmyr (06:09)
which one off the top of my head. didn't take a picture of it or anything, but yeah, it was super good. So yeah, changing it up a little bit. I did get my new package in the mail for

the subscription order, which has like a a one off purchase of the honey processed one. So we didn't open

TJ (06:23)
Mm-hmm.

Chris Gmyr (06:24)
that one yet. probably over the next couple of days we'll probably open that. So hopefully by the time we record next, have a report for that. But yeah, that's about

TJ (06:31)
Ye y

Chris Gmyr (06:32)
it for me.

TJ (06:32)
yeah, I'd love to hear about that. So for sure fill us in.

Chris Gmyr (06:35)
Sweet. Yeah. Well, next up, I know we've been talking about this offline a little bit because we've both been cranking on the Claude subscriptions

TJ (06:45)
Yeah.

Chris Gmyr (06:45)
recently, for sure. And yeah, they keep on reminding us that the limits are going to be changed on Monday. So ha yeah. Yep,

TJ (06:53)
Right. That it's that that fifty percent additional limit, right?

Chris Gmyr (06:57)
it was fifty percent. They added it, I don't remember a couple months back, to I don't know.

feels like patchover like bandage, you know, something else that they did at the time. but now that last date is on Monday. So we're recording on Thursday. have a couple more days of the increased limits. And it's a little unclear of what they're going to do with it. we talked a little bit like offline about thinking that they're gonna extend it or change them somehow. but a lot of people online are pretty upset about the

potential risk of the decrease and also all the V five models have been a little little janky and overly verbose and not as good as they probably should be for the token usage and all that, which I know that we've run into too, so

TJ (07:43)
yeah. Like I from what I really have gathered, like I've seen so many people over the last like week to two weeks like be posting that they're like dropping their Claude subscription because they're super dissatisfied with the Opus V and like how the direction that that's heading. It's like super verbose, super wordy, like.

really doing some like kind of weird stuff coding wise. it's it's not been a pleasurable experience. And so I've seen a lot of people like just moving because of that alone. I've definitely heard some rumblings about like, well, what happens with the limits? And honestly, I'm kind of there too. As much as I am like all bought into Claude and Anthropic, like if these limits shift, I

I'm probably gonna at the very least drop from the 200 plan to the $100 plan and then probably increase my spend on OpenAI side. because I I just don't see how how they get away with like they're they're gonna blow up their market share, I think if they pull these limits back with how poorly Opus V has been received.

Chris Gmyr (08:52)
Yeah, yeah, hundred percent. and with all the projects that I've been working on lately, like I've been right up to the limits and basically saying like, okay, like last night, even at eleven o'clock they turned over and it was like, Okay, gotta dial down the work so we don't hit the limits to actually like get through, you know, the evening on the kind of autonomous runs and things like that. So I've been making a bunch of shifts too.

in I don't know, maybe like a month ago I upgraded to the two hundred a month max plan and

TJ (09:26)
Same.

Chris Gmyr (09:26)
also have the hundred dollar codecs plan. so very underutilized still on the codecs, but like still hitting or getting close to hitting like the limits on a weekly basis for Claude. granite like that's up until recently the only thing that I've been really like churning on. but yeah, if those

those limits come back down and that cuts my week in half. Like that's I don't know. That's not gonna be a good thing.

TJ (09:51)
No, no. And yeah, I've I've I spent so much time over this week just at my wit's end with Opus. even even dropping down to like four eight didn't feel really all that good either. yeah, just I I don't know, man. I can't and I I just I can't justify that kind of spend anymore, like for for that level of quality. So I mean, I'll we'll we'll see. I think it's gonna be

I think I think it's they have they have some interesting choices to make and I I I see them either extending it or doing something else in place of it because dude, they're just gonna bleed market share at that point.

Chris Gmyr (10:31)
Yeah, yeah. So

TJ (10:34)
Yeah, we'll see. So my so my I I think we lay out some predictions here. I think I think they do they either extend it or they do something similar on or before Monday.

Chris Gmyr (10:46)
Yeah. I I really want them to lock in the increased limits. So say like we're not we're we're just locking in like the fifty percent more and just giving that to everyone.

TJ (10:55)
That's just the that's just the new limit, yeah.

Chris Gmyr (10:58)
yeah. Hopefully they they keep the higher limit and just lock that in and say, that that's it, or

That and they have like a a point release of all the five models, which I don't know about that. We'll see.

TJ (11:12)
I've I I

have been hearing some rumblings about probably like 5.1 coming out. somebody said somebody caught like some change in the web UI. And they're like, last time they made a change in the web UI like this, there was a new model release that came shortly after. So I I think I'm definitely picking up.

around the community that like there's there's some a fair amount of people starting to guess that they're gonna be like dropping a five point one, especially with all the complaints. Like if they do some changes in like post training, you know, and then you know get get some of this verboseness and some of this like weird stuff that it's doing under control. I could see them maybe doing that and having a five point one release soon.

Chris Gmyr (11:59)
Yeah. Yeah. Definitely. And yeah, I've tried the concise output config and it barely it it seems like it barely does anything. If it does a lot, I'd be surprised what like the normal output is 'cause I've had mine changed for quite a while. but later in the the context and the conversation it it just loses it anyway. So it

TJ (12:21)
Yeah,

I feel like it loses it. And it's not even just the verboseness, it's that it like introduces all this like weird technical jargon and like just makes whatever it's saying like so convoluted and like hard to digest. And I wonder I wonder if that has anything to do with

the watermarking. I I wonder if the watermarking system, because you you you know about the yeah, so like the watermarking system, I wonder if that is causing it to have like this insane output that is just so hard to even read and understand. I

Chris Gmyr (13:03)
Interesting.

TJ (13:04)
I had, I did some like Aaron Francis posted some like prompt that was like send this to your agent to like analyze all your conversations, tell you like what

like how you how you work, what you're good at, what you're not good at. And like it was so funny. It like went through all the like a bunch of conversation history. And it one of the things it mentioned is like, yeah, you constantly are like self-correcting it to like speak plainly and like can be concise.

Chris Gmyr (13:31)
Yeah.

TJ (13:32)
And I'm like, yeah, because that's been terrible. we can talk about this a little bit more later, but how I use it right now is I I was using

Hive with Opus as the lead and then spinning up Fable Opus sonnet workers as needed, and then having Codex do adversarial review. I switched, I was just getting so frustrated with it. And Codex was finding all sorts of stuff that like should it should not have found. Like Claude should have done a better job. so I switched to Fable as the lead instantly better. Like I so, so much better.

I I'm so frustrated that I didn't do that in the first place, but I was like, I don't want to blow up my fable limits. But like that was such a better experience.

Chris Gmyr (14:16)
Yeah, yeah, a hundred percent. After after we were talking the other day and I saw you made that change, I tried it on new sessions moving forward and it's been significantly better, at least on the the lead side,

TJ (14:28)
Yes.

Chris Gmyr (14:29)
and making like better decisions and which model to use and just the work output, the reviews that the lead is kind of just double checking. Like I

TJ (14:37)
Mm-hmm.

Chris Gmyr (14:37)
have like review skills and other things going on, like on the worker side, but still I had the lead kind of look into

Making sure that it's it's going in the right direction and things were done correctly, as per, you know, the plan or the board or anything like that. So yeah, it's been way better as Fable. And I've been like hesitant to use it too, because I don't want to, you know, blow out the limits or anything like that. But it seems like it's been especially only running Fable as the lead, it's pretty low usage comparatively.

TJ (15:09)
Mm-hmm.

Chris Gmyr (15:09)
So I mean, just to keep everything on track, it's it's a lot worth it.

TJ (15:13)
Yeah, I a hundred percent. Like I I really have been kicking myself for not have starting the workload that I started with Fable as the lead. But yeah, it was it was just I couldn't I could not read another status update from Opus. Like it just made my brain spaghetti. So yeah.

Chris Gmyr (15:33)
Yeah,

yeah, totally. So I'm gonna pull this up because it's very much related to what we've been talking about.

TJ (15:40)
Yeah.

Chris Gmyr (15:41)
so cool. So we have some predictions for next week. We'll check in and see what happens there. but on the hive front, and with all the limits and we were talking about codecs, I've been wanting to utilize the codex subscription more besides just through like adversarial reviews and the counselor skill in workflow.

which is Aaron Francis's. so I was doing counselors with Opus and Seoul, which has been fantastic and catching a lot of things, you know, as well. But I'm like, I want the workers to utilize the codex subscription because I kept on checking out the usage and codex was at like one, two percent or something like that for the week, even after doing like a bunch of work. so I'm like, this is kind of ridiculous. I want to adjust things. So

Now Hive has an option to utilize Claude and Codecs. all the harnesses, the models, all of that is good to go. So what I've been working on switching is moving all the like documentation and the skills and everything else like that from the dot claude directories into dot agents and then simlinking back to claude.

to have like more the industry standard and just having like the one-offs for Claude to read. and then kind of moving away from the built-in skills for like code-review and also the simplify skills, which I use all the time, and kind of building up my own set of skills that I can use kind of cross agent and harness. and then in the steps for the lead to pick.

whatever sort of model to choose for the work. there's like a whole new table on like pick the harness first based on like these categories and the the type of work that needs to be done. And then you pick the model from there. so everything is like config based. So in one project you can like only be using Claude. In another project you can be using just Codex. In another project you can be using both. And the way the hooks work is that it'll read

all the extracted profile assets and figure out like which agents and harnesses and stuff like that that you're using. So that table is dynamic based on the project config that you have. So

TJ (17:54)
No, nice.

Chris Gmyr (17:55)
it's been working like really well. and I have to dial in the the custom skills and the table a little bit more. but just seeing the lead of like, hey, this would be great for you know, Codex Luna on

high or something like that. And then this other one, this is still on Claude Opus and just seeing it make those decisions and then being able to call out to Codex and then have like another pane pop up and have just all the things that hive already does and the messages going back and forth and changing context and all that has been like super awesome to see. So working on the the profile assets a little bit more and getting those solidified but

Like the the plumbing and the usage within a hive is like ready to go. So yeah.

TJ (18:42)
Yeah, dude. I I'm loving Hive. I think that's I think getting like Codecs integrated at a yeah, instead of having it to like have to shell out to the Codex CLI, it's I think it's gonna be really nice to have that integrated into Hive as a like first party citizen for sure. Yeah.

Chris Gmyr (19:01)
Yeah. So

that's been that's been a a fun project. But also very frustrating on the cloud side because I've had to like pull in codecs docs and do like my own searching because whatever it was doing to say like, codex can do this but not that, or like here's five things that codex can't do. It's like, I don't I don't think that's actually true. You know what? Can you search

for stuff and it's like, nope, verified it, can't do it. I'm like, but what about this doc like directly and link to like the agent skills or something like that in codex? It's like, yeah, my bad.

TJ (19:39)
that's and that is exactly the kind of stuff that was like making me infuriated with Claude. Like telling me it like this limitation exists. And I'm like, no, it doesn't. you're absolutely right. It doesn't. What what are you doing? Like Codex

Chris Gmyr (19:54)
Yeah. Yeah. It was so bad.

TJ (19:56)
came back with some review at some point and it was like, yeah, like half this feature isn't implemented. Like the tests exist, but like there's no implementation for it. Like, how, how did you do that? Yeah, it yeah.

Chris Gmyr (20:08)
Yeah.

It's ridiculous. So

TJ (20:10)
geez. No. So

I'm really excited about that because I've been like I said a minute ago, like my workflow has been having Codex do the adversarial reviews. But it would be even if I stick with that workflow of like Claude is the main and you know, I think it'll be really nice instead of having to just like shell out for those reviews to like actually have a proper pane open and it do its its thing inside of Codex. I think that'll be great. I'm not

I'm not looking forward to all of the tool chain juggling between Claude's files, agent files, and all of that. I'm not excited about having to dive into that, but hopefully that shouldn't be too much of a pain.

Chris Gmyr (20:51)
Yeah, I think everything has been working really well as far as like the Claude MD versus Agent D. I have those simlinked together. So right now I'm defaulting to the industry standard, which is agent everything, and then simlinking to Claude, except for Claude rules. From what I can tell, like Codex doesn't actually have that or like they're they have rules, but it's like a programmatic set of rules to do something

TJ (21:20)
interesting.

Chris Gmyr (21:20)
Also it's

it's not it's not like Claude's rules, which I think are amazing.

TJ (21:24)
Yeah.

Chris Gmyr (21:26)
so I haven't crossed that boundary yet, but in my instructions right now it says like if you need to do work that directly ties to rule files that can't be injected into codecs, like just use Claude. so there's a little bit more

routing to do before it spins up a worker, but until, you know, I figure out something to do with the rules across harnesses, then it just is what it is right now.

TJ (21:54)
Man, that is so disappointing because I I have for a long time made heavy use of rules and it is a I I feel like it's a fundamental piece to maintaining the kind of code quality and architecture that like I want to have. So it's a little frustrating that that doesn't exist, but yeah.

I wonder

Chris Gmyr (22:17)
I

TJ (22:18)
it makes me it makes me wonder about using

'Cause it 'cause that is a harness thing. That's not a provider thing. Like, so I wonder if like open code has support for has support for rule files. And then you could just use codecs through open code and like r maybe your rule files apply. I don't know. I have not looked into it, but that that could be interesting. If it does, I don't know.

Chris Gmyr (22:40)
Yeah. That's a good call. Yeah. My other

Yeah, I'll have to check that out. My other idea was to do like a a rules skill or something like that. So almost like a a prereq to your review step. because we want like the implement implementation done before a review, right? So running like a rules skill would scan the rules that you have and then

see what you have kind of dirty in the branch and then see if any of those like basically tell the school the skill how to interpret the rules and then see if any of the rules get triggered based on what you have in your branch currently and then do like fine connect the files to the rules, execute the rules into like the contacts window and then execute on whatever it is that the rules say and then commit and then do your

review after that. So that was my tentative thought to pull the rules.

TJ (23:40)
That's

that's super logical. I think that would be that I think that'd be i outside of looking into if open code supports it or not, like that would be the next play for sure. I think

Chris Gmyr (23:52)
Yeah, yeah.

TJ (23:53)
that's probably the right way to go.

Chris Gmyr (23:55)
Sweet. Yeah. So I'll look into that too. But codecs baked in, gotta dial in the extractable and configurable kind of tools after that. But yeah, it's been super solid. So excited to see it out in the wild and what happens.

TJ (24:10)
Yeah, man.

Chris Gmyr (24:12)
cool. But yeah, I wanna go over to some of your projects. So we've

been work the sync work for

TJ (24:17)
Yeah. Yeah,

so I I think this is like leads perfectly into talking about Halcyon's Sync works. So I have a working implementation of Halcyon's Sync Engine, and I modeled it exactly after Obsidian's Sync engine. however I was running into just some

like weird edge cases that I didn't like and some like I just didn't like the handling of it. And so I did a bunch of research last Saturday with Opus and Fable and had it also like brainstorm with codex. And I'm like, look, there's gotta be a better way of doing this. And so we decided to take a more git style approach than

Obsidian sync approach because obsidian when it runs into conflict. So let's say on my work machine, I create a README MD file inside Obsidian. And then offline, I also create a README MD on my like my personal machine. Once sync comes enabled, like one of those is going to lose and obsidian's

way of handling that is just kind of throwing their hands up in the air and going, well, go back through document history and like resolve your own conflict. I just I don't like that. And I think that plays out especially poorly for Headless Sync because that's something that I definitely found that I wanted. Like I've been starting to use IRS a little bit more again.

And I wanted to have access to my markdown wiki and that's gonna be something that's synced through Halcyon. And so I I need to that's Iris sits on a server, so I needed a headless sync. And it's like, well, headless sync can't automatically make those merge decisions. So I'm like, well, I gotta like figure out a better merge engine for like all of this stuff. So I spent

a good chunk of Saturday, like going back and forth, having it brainstorm with Codecs. We came up with a new plan. I had it also, I can't remember where I saw this, but like I also prompted it after we built out the the strategy and like the overall plan. I had it do an implementation plan, but how I prompted it was

And I saw this from somebody else. Like I prompted it specifically to say, lower power models will be working on this. So like write your implementation plan for a lower-powered model to tackle this. and that I think I think that definitely resulted in like a

more thorough implementation plan, better calling out of like edge cases or like things that might come up. so we had a pretty solid plan. and so I jumped into Hive, started up a lead and the way I have this set up is I've got a Halcyon directory and inside of that has all the sub-related repositories. The Rust Core, the iOS app, the desktop app.

the Laravel backend for handling like sync and payments and then the marketing website, the documentation website. So it's all just kind of like in this like subdirectory. So I pop a lead open in there, pointed at the plan, and I'm like, yeah, like here's the the work strategy of like

You're the lead, started with Opus as the lead. you spin up fable opus sonnet workers like as you see fit for like fit the best model to the task, and then have codex adversarially review everything. So

I've had that running on implementation since Saturday night and it is still going. I checked right before we hopped on and I had just crossed the 30 million output tokens mark. and I'm like all I'm just under 18 million input tokens. and it's it's just wild that it's I've had it running almost nonstop outside of earlier this week. I did stop the whole process.

And switch to Fable as the lead. and that has been infinitely better for the whole process because it was it I I don't know, I woke up I started it Saturday night, I woke up Sunday morning and like I had told it before I went to bed. I'm like, my expectation is that like this is done and I'm checking it in the morning. Cause that's really what I thought. I woke up in the morning and it had barely done like

Four or five tasks because it just kept getting into these like weird review loops, and then Codex was like every time Codex ran a review, it was finding like huge issues with the entire system. And so I'm like, this is terrible. now Fable as the lead.

Like the reviews are coming back much more sensibly. But like for a little while there, like every time it was like, Yeah, I have Codex Review, it was like, no, these are huge issues. even like I was saying, like it it found something that was tested, but like only half the implementation was actually done. Like just weird, weird stuff like that. but after switching to Fable, like Hive has just been an absolute

dream to work with because it fits my mental model like perfectly for how I want this to go. so I did a I think it was last night or no it was this morning I sent a like a summary to the the group chat of just like, yeah, here's where it's at. I think it's like 85-ish percent complete with the sync system. it built out all of the headless implementation

And I think after it's done with the last of the the few implementation details of sync, it then just has to do the integration into the iOS and desktop app. I think it's already integrated into the headless sync. and the way that I I built that is there's a Rust core that handles all of this complexity. That Rust core is what gets pulled into

The desktop app and the iOS app. So like I don't have to worry about these different implementations across these things. It's all one shared core engine for all of it.

Chris Gmyr (30:34)
Yep. That's a good move.

TJ (30:35)
so I'm my my expectation is that like hopefully it's done sometime this weekend, and hopefully I test it out. at some point like midweek, I'm like, I'm having doubts. I'm like, how is it taking?

This long to like, I I kind of expected it to be done overnight. Like, I didn't know I'd be running Claude for a week straight on this problem. So, and and so I got I got kind of frustrated at one point and was like, do we need to like, are we overcomplicating this? Do we need to like, is this a solved problem? Like end-to-end encrypted Git style sync diffing like.

That just just feels like a solved problem. And I pointed it at Obsidian again. And it was like, look, Obsidian just bails out of these complexities. It like throws its hands up in the air and just forces the user to like go back through document history to like restore bits and pieces, copy and paste stuff around. And like what you are building is way more in-depth than than what they're doing. So maybe stick with that plan.

So I'm sticking with it. I really, I really hope after all of this, like after over 30 million tokens, a weath week worth of work, like I really am hoping that I just turn it on and it works. Like if it if it is a mess, I'm gonna be so distraught. Like

Chris Gmyr (32:07)
Yeah, I hear you there. It's been quite the struggle. And then adding in like the the V five issues and Opus stuff has just been gnarly.

TJ (32:16)
Yeah.

Chris Gmyr (32:17)
I've just yeah. On some of the my projects, which we're gonna talk about more next time, making a landing page and the amount of tests that it made for a landing page that isn't even like public launched yet or whatever, and it's finding like

I made this test because, like, it's a landing page. It's super important for the public to, you know, be like making sure nothing is broken. It's like this is a super simple landing page. I don't need like a hundred and fifty different tests and like twelve hundred assertions based on like little shadows and text sizes and things like that. I'm like, we're gonna be iterating on this thing for the foreseeable future. It's like

It doesn't make any sense to maintain these tests. I'm like, kill off everything except for like the most valuable ten percent of it. It's just like how did it get this bad without it just

TJ (33:08)
Yeah, I I I yeah.

Chris Gmyr (33:13)
it's crazy.

TJ (33:13)
I saw somebody post on Twitter yesterday or the maybe the day before, like a very similar situation where it was like, and I've seen this in my test suites too, and I I'm not really sure how I want to deal with it yet. But yeah, like tests, like making sure the button's red. Like why is that a test? Or like you're testing that like the button text says this, like that is.

not valuable. And I am so sure that there are so many terrible tests that it's written over the last week inside of sync. And I'm I know I'm gonna have to go back and like come up with some skill or some guidance or something to like make sure tests are actually valuable and not like button text is this. Like insanity stuff like that. So

I'm seeing that a lot too. And yeah, landing pages, everything. So if I'm seeing that on landing pages, I can't believe the kind of disgusting tests that are in the Halcyon repos right now.

Chris Gmyr (34:14)
Yeah, yeah, exactly. And I feel like that's what it's been churning on like so much. you know, without guidance is just tests and just the the non feature work implementation. and I've seen that in hive too with like code comments. I know like code comments have been like a big one too.

TJ (34:33)
Paragraphs for our config value.

Chris Gmyr (34:36)
I started looking through like some of the source code. this was probably like last week or the week before or something like that. I'm like, this is

Again, this is crazy. It's like there's like ten thousand line file to for for one of the mechanisms. And it's like you look at the actual code and I had it count the lines of code versus lines of comments. It's like it's crazy amount. Like there's barely any code in this thing for the amount of comments that you have. So I had it basically delete all the comments.

And limit it to like a sentence or two max if you did need a comment. and then put in a test to then test the amount of comments in the code base. And which it flags it, anything over like a 5% ratio for of like comments to code gets flagged and CI fails. So it has to like rewrite them or move it into something else, like either the agents or like Claude, MD.

a rule, a full-blown like document or something somewhere besides like a code comment. It's like you cannot add a code comment. You have to put it somewhere else if it's actually that important. And it's just like I never in my whole what 26 years of doing engineering and programming never ever have I thought that I was gonna make a test to

Test that the ratio of comments in a code base. Like this is ridiculous.

TJ (36:04)
This is crazy. And and a slight tangent, don't get me started on CI right now. Like I know you and I are like just I I'm out of CI credits. Like I blew through them all because I know partially because I'm building two Mac apps, like iOS and Mac and and I know that's like build at like 10X or something. but I

Chris Gmyr (36:26)
Yep. Same. Same same.

TJ (36:29)
I I don't even know what to do. Like I think I think what I'm gonna probably do is wire up my laptop to be a test runner. And so if I'm like actively working on pushing on stuff, like it can be a GitHub runner on its own and like at at the very least, pull all of the Mac specific jobs and like push them to my laptop instead. because yeah, like

running out of your actions limit halfway through a month is terrible.

Chris Gmyr (36:55)
Yeah, yeah. I ran out of mine too. because originally it had two different versions of Mac OS running. And like you said, the actions run at 10x for Mac compared to like Ubuntu or Linux. so basically what I moved to is like a single canary run once a week for the Mac in CI for GitHub. And I just leverage it locally. Like it runs all the test suite locally for Mac because I'm on Mac.

but yeah, it is just way too expensive.

TJ (37:25)
That's

what I might do too, is like that's the other thought I had is like disabling the Mac test suites and linting in CI and then only leaving the the build step on version tag in CI so that it can that it will when I tag a release, it will do the release build, attach it to the release, push it to R2, so it can be distributed.

Chris Gmyr (37:49)
Mm-hmm.

TJ (37:50)
So I I'll I might leave those and then just kind of like have rules inside of you know, Claude that like before you push these test suites have to be ran locally.

Chris Gmyr (37:59)
Yep. Yep. Exactly. So probably doesn't help with like if you wanted to test multiple versions of macOS, you probably still need like those custom runners, but

TJ (38:09)
Yeah.

Chris Gmyr (38:10)
I don't know. For mine, at least for Hive, like I'm just like, Yeah, if it runs fine locally, I mean that's fine for me.

TJ (38:16)
Yeah.

Yeah. I want to touch on this like real quick because I also you you you had you had this on here last week too of like no more open claw. I have been staring at Grokbot and Hermes and definitely not open claw. but I have found myself

back in Iris land and leveraging Iris a bit more after all the inspiration that I've kind of been seeing with Grokbot. That like I'm wondering how much work that would be to kind of put something sort of similar inside of Iris, where it could have like, you know, cross conversation. Like you could take a

spin up a new tab of iris and like they could talk to each other. And so you could have like different skills loaded and pinned in each one. they could have like different roles to do things, but

Yeah, I'm kinda I'm kinda back in this like potentially agent world, like personal agent world again.

Chris Gmyr (39:15)
Yep. Yeah. Well, yeah, I guess I can briefly mention it. We got a couple more minutes.

TJ (39:21)
Yeah.

Chris Gmyr (39:21)
hopefully the the open claw won't take too long. But yeah, ended up dumping open claw. I've had some really well, first of all, I haven't really used it a ton because what I've been doing is just like leaving my personal laptop open and just going on the Claude mobile app and that has been good enough, or just if it isn't open and running.

I'll just leave myself a little note in Todoist or Obsidian or something like that to look into or you know, save a a podcast or, you know, an article that I want, you know, and import it into Obsidian or something or to take action on it. so I really haven't been using the open claw system. I've been trying to keep it up to date. So I'll check in like every couple of weeks and be like open claw update, which has consistently messed something up every single time that I do an update.

TJ (40:10)
That's that is what

I've been hearing is every people have moved away from it in favor of either Grokbot or Harmes because every update like broke the system.

Chris Gmyr (40:19)
Yeah,

yeah, it just crazy. And one of the one of the updates actu what went through, like I fixed up a couple things and then kind of signed out out of it and didn't pay attention to it like a week or two after I logged back in. And it was basically a warning like every 10 minutes that the obsidian like headless CLI sync was failing. And it's like I

didn't get any sort of notification. Like you're connected to Discord. Like that's where I get notifications through. So this thing has been failing for over a week and I knew nothing about it until I actually logged back into the open claw UI. It's like this this is just the dumbest thing ever.

TJ (40:58)
Yeah.

Chris Gmyr (40:58)
And like I tried to like I got that fixed. I did like another update. I'm like, what am I doing here? 'Cause I got

TJ (41:04)
Yeah.

Chris Gmyr (41:05)
everything local too. I I wasn't

using it as much because like over over the last handful of months I keep on getting like API bills from Anthropic because that's I have Anthropic, I have Google and OpenAI connected to it for different things. And there's only like two little cron jobs running and nothing else really going on. And I wasn't using it. So it's like why am I getting a bill of like twenty to forty dollars, you know, every two weeks or something like that.

And so I asked that, I'm like, what is going on that like this and that? And Anthropic wasn't even like the default model. So it's like, yeah, these things were cached and never, you know, flushed out. So they're still connected and you still got like heartbeats going and stuff like that. I'm like, I no, we should have turned all this stuff off like in the last like big kind of refactor that I went through and did it. So you're telling me it hasn't, and that's why so it's like, okay, just take anthropic out, just use the Google one.

And then I got even more more often bills from Google for their API. I'm like, what is going on? And it's like, okay, just I did like a backup of the whole thing. I turned off the VPC on DigitalOcean just like I wanted

TJ (42:17)
Play yeah.

Chris Gmyr (42:18)
to the yeah, the droplet, just in case I wanted to turn it back on. But did like a full backup, turned it off. I'm like I like nope.

TJ (42:25)
Wash

your hands of it, yeah.

Chris Gmyr (42:27)
All

all done. So I think the next step is just I haven't really missed it the last couple of weeks that I've had it off. So I think I'm just gonna turn it like delete it, save the money,

TJ (42:38)
Yeah.

Chris Gmyr (42:39)
use use other things, or just like make notes for myself instead of having like an always on just agent who actively tries to make my life harder.

TJ (42:48)
You're right. Yeah.

Chris Gmyr (42:50)
So yes,

no more no more open claw for me. maybe I'll check out some of the other systems, but yeah, I've just been busy enough on on my own with the couple of projects and just leaving, you know, little notes for myself for whenever I'm back at the computer. So it's been fine. So no more open claw for me.

TJ (43:04)
Yeah. Yeah, makes sense, man. All right, on that note, you want to wrap

Chris Gmyr (43:08)
Yeah, let's wrap up. Thanks for listening to the Slightly Caffeinated podcast. Show notes and all the links and social channels are down below and also available at Slightly Caffeinated.fm. If you have any questions for us or content suggestion, go to Ask a Question page on our site and we'll feature it in an upcoming episode. Thank you all for listening and we'll catch you next week.

TJ (43:26)
See ya.