Rubber Duck Thursday!

D. Hey. Hey. Hey, baby. Dip. Hey y'all. Hey y'all. Welcome to Robbo Duck Thursday. I am Cadesha. Welcome, welcome, welcome to the stream. Um, thanks so much for joining. Yes, just on time. We just started today. We started at um 30 minutes on the hour today instead of right on the dot. And I am so excited to be here with you guys. As usual, I have to turn down the music because it it's so nice. It gets me distracted. So, I'm just going to switch it out a little bit. But how are you doing? Wow. How are you doing? Let me know how you're doing in the chatty chat. Let me know what you're working on, what you're building, um how you're feeling. Like I feel like I feel like so much is going on in the industry and it's impossible to keep up. But you know that's just the nature of engineering and tech and it's just a matter of you know trying to keep up as much as you can and trying to do things with the trying to do things and trying to try the new tools as they come out. But yeah, how are you doing? How's it going? I'm doing great. It's been it's been a long week. Uh but today I want to chat with you about the things that we've been shipping at GitHub. Maybe build something. Let me know your project ideas in the chat so we could, you know, possibly do something together. And yeah, we're just going to be vibing. I see Amy says, "Anybody vibe coding yet?" Is anyone vibe coding? Let us know. Let us know what you're vibing. Let us know what you're building. I am personally obsessed with vibe coding. Like, I can't stop talking about it. Um, I just love building these little tools. Now, shipping is another is another um concern, but awesome beans. So, I'm just gonna jump right into it. I'm just gonna jump right into it. Is anybody not vibe coding? That is the real question. Hilar. I'd love to know what skills, what agents, and what fun things you have in your environment, you know, that you have configured in your environment and what it kind of looks like because I've noticed that different developers and different engineers kind of like prefer different tools and different, you know, different like um plugins for their stack. So, that's pretty cool. So, let me just I'm going to pop you guys as usual over to my other screen and yeah, let's get into the stream. Let's get into the show. So, we're going to start with the change log here. Let's see what we've been doing at GitHub. One second here. Oh yeah, and there's this blog post I wanted to talk to you guys about because it's so cool. Um, so let's pull up the change log and the bloggy here. Okay, let me share my screen. Let's share. Yeah, let me know what you guys are building. What's your vibe coding or what's your vibe engineering? Like you know what you're building with agents. Have you been using um like Open Claw and those other open- source AI agents? Like things are moving so fast. It's actually wild. It's actually wild. Let me turn off this music here cuz it's a little distracting for me. Okay, music paused. Let's see. I prove everything with math, not vibe. Uh, so Ryan, are you using any AI agents um in your, you know, endeavor to prove things with math? I'm I'm curious here. How are you building these days? Okay, awesome. Okay, so chat is going working on an AI linkbook web app. That sounds cool. Tell me more about that. What is it? What is What does it do? Agentic engineering. Yes, I know my engineers love their um semantically correct phrasing. I prefer the specd driven approach over vibe coding. Yep, I like that approach as well. All right, so let's get into the change log. I'm just going to pop this up here. Boop. And then let's change. Okie do. Let me Okay, so let's see what we've been doing this week at GitHub. Um, so Copilot Cloud Agent fields add usage metrics. Okay. Uh do you know what copilot cloud agent is? So copilot cloud agent is essentially copilot is essentially github copiloton on.com. Let me just go here. So here and then if I go to the agents tab you'll see that we have a whole bunch of a whole bunch of things here. And that's essentially Copilot Cloud Agent. Um, one second here. I'm kind of moving slow today. My bad. O, okay. And so that's essentially Copilot Cloud Agent. So you can send tasks to GitHub copilot um right here and it will run it for you. You have access to you know all the latest and greatest models and I believe you can also use claude and codecs on um within copilot cloud agent. So that's pretty cool. And then you can also set up your own custom agents here. So if you wanted to say add an accessibility feature and you wanted to use I don't know codecs on github.com you can and this is all included in your GitHub copilot subscription so I think that's pretty cool and then you have access to all these models auto mode has been um something I've been exploring lately. Have you tried auto mode um both in the CLI and on.com? So that's what copilot cloud agent is. It's essentially using GitHub Copilot in the cloud or like on github.com. So pretty cool. Pretty cool here. And I know I know the team has been doing also some really good work there to make sure that it works properly for you. Okay, I'm seeing some some questions in the comments that I can't personally answer. Um I know that there is a blog post about the reason for the um the student program getting a little cut. So you can definitely go on our blog and see and read more about the decision to do that. Okay, let's see what else we have. I also saw recently that the team shipped that you can assign dependabot um like pull requests to GitHub copilot on.com. I thought that was incredible because I don't know about you, but I find it so hard to keep up with all the version changes and like all the PRs that Depend would open and assigning just being able to assign it to Copilot Chef's Kiss. Um, I thought that was a really good Easter egg that the team shipped. All right, so better debugging. Okay, let's see what's uh what are what's new actually. So, okay, you can now view and manage agent sessions from issues and projects. And so, this is new as of what? Yesterday, what's today's 23rd? This is new as of today. So, you can now view and manage agent sessions from issues and projects. Oh, that's pretty cool. Session, let's see how it looks. This is what it looks like. Ah, so there's a new session running tab here that you click and it surfaces right here in the issue. That's pretty good. Like instead of having to go back to the agents tab, it just surfaces right there. Pretty neat. I think that's pretty neat. Session assign panel in projects as well. Ah, so you click on the agent session and it also surfaces on the side panel. I think that's pretty good. Have you have you ever tried um using copilot in your PRs for like PR reviews or you know building uh features for you running tests um assigning those depend I'm curious how you're using GitHub copiloton.com would love to learn. But this is a this is a cool ship that the team released this week. Actually today released today. Usage telemetry C++ code intelligence forget CLI public preview code stuff copilot auto model selection. Oh what's this? So this was recently as well. So auto model selection is now generally available in the CLI for all plans. With auto, copilot chooses the most efficient model on your behalf. Auto is dynamic, giving you reliable access to your favorite models while mitigating rate limits. Uh rate limits in the industry is, you know, really hot right now. Um here at GitHub, we're not the only one trying to control um the cost. So using auto mode is a really good option here. See which model was used directly. Stay in control. Respect your policies. Pretty cool. So you can use auto mode now in copilot CLI. Um you can manage agent skills as well. And of course Opus 4.7. Have you tried Opus 4.7? There's been so many new ships. It's been so hard to keep up. So now there's a GH skill. Oh. Oh, with the GitHub CLI. Okay, there's now there's a GH skill command that makes it easy to discover, install, and manage publish agent skills from the from GitHub repositories. Oh, update the CLI to version 2.9.0. Then discover and install skills interactively. GH skill install Da da gskll install. Okay. GH skill search MCP apps. Very cool. So now you can use the GitHub CLI, not the Copilot CLI, but the GitHub CLI to discover, install, and engage with GitHub with um skill agent skills. Skills are automatically installed to the correct directory for your agent host. You can target a specific agent or scope with flags. Very cool. So, GH install GitHub. Awesome. Da da documentation writer agent cloud code scope user. Pretty cool. Did you did you read about this? Oh, sorry. Let me see what the comments are saying here. Copilot comments on PRs are great. I'm actually I'm actually I love that you love that. I love it all the time. I I use it all the time. I mean, and it's it actually catches things that I'm just like, "Wow, thank you so much for surfacing that for me." These aren't available on the free plan, are they? When you say these, what do you mean? Please consider interact. I like to pick models manually because the lower powered models are just bad. Can auto actually pick a decent model. We can test it out today to see um what model it's surfacing for like different type of tasks. So, put in the chat what type of task you want me to test with the auto mode and we can just do it right here right now together. Okay. So, let's let's respond to bronze quest like bronze question. Um, hello Francesco from Italy. Thank you for joining. So the question is the he's saying that he likes to pick models on manually because the lower powered models are just bad. So can auto mode actually pick a decent model? So let's test it out together. Let me pull up let me pull up a project here that we can use and let's give it a go. Let me just pull up everything that I need here. I have this release videos. Can I use this project? One second. Listen to some music as I um No, hold on. Let me find a project for you so we can test this together. Let's see. I could I could use the um what's that project called again? Chapter Smith. Yeah, I can use that one. That's always a good one to pull from. All righty. Yeah, I think I can use this project here. Yeah. Okay. Yeah. Let's use Chapter Smith here. I'm just going to keep it on this branch. All right. So, this is like a Twitch stream. No, literally. Uh, it literally is like a Twitch stream. We do stream on Twitch, YouTube, LinkedIn, and I think we're also streaming to the Twitter community as So, yeah, it's literally like a Twitch stream. So, I want to go back to that question so we can test auto mode together. So, let me just surface it again. I'm going to share my screen. Oh, did I? Let me just reshare here. Screen. I want to share the full screen. Perfect. All right. So, you should be seeing my um my terminal here. And then you release So, this is what we were talking about. And so, let's test it out together. I'm trying to find the the comment that I had. I should have saved it. Thank you. Because I could have find it. So, this is what we are attempting to answer today. Can auto mode choose a decent model and you know I'm fully aware that decent model is going to be subjective like based on what use the user likes. So like I also like to choose models manually. So I'm also curious to see if auto mode will be able to pick a model that I would have probably selected for the task. But I think the pro the um I think the goal of auto mode is to solve the use one model for every task thing that I see a lot of engineers do. So I know that we all love Opus, right? We love Opus. We love Opus 4.6. Love her. Now it's 4.7. We love Codeex 5.4. But are you switching your models depending on the the task that you're you're working on? Right. So, like if you're doing something like some really deep debugging, pull for Opus, right? Because it's excellent. But if you're doing some some like front-end work where you're like you're updating designs or like you're updating something, Gemini is actually very good at front-end work, right? So, like you want to make sure that you're using the right model according to the task. If you're doing a quick patch on documentation or even like a quick fix on a method, you can use Sonnet. you can use one of the sonnet models for that because it's very good and even um pulling into open-source models because in uh copilot CLI you can actually bring your own keys including bringing in your own local model keys and use it with copilot CLI. So you don't have to use or like it's not recommended to use the same model for every single task. So I think that's what auto mode is really trying to to solve here. So, let's get into it. Okay, so I have this project here. It's called Chapter Smith. And essentially, it allows creators or, you know, people like me who make a lot of videos and post it on YouTube. So, it allows us to generate um YouTube chapters for our videos automatically. And so, that's what this project here is doing. So, it's nothing super complicated. Here it is in VS Code. It is a Nex.js project. Nothing crazy going on, right? So, the f I guess the first thing I need to do is to figure out what I want to add because transparently I haven't been in here in a few weeks in a few weeks. So, I guess my first question is I'm going to ask Copilot what I can possibly add in here to improve the application. Let's just go full screen so I can toggle back and forth here. Okay, that's a lot nicer. So, let's uh invoke co-pilot. I'm going to just do allow all or you know the yellow mode. I just want this to be local right now. Uh my MCP servers aren't loading. That's okay. I'll look at that later. I honestly I have too many MCP servers and too many like I have 67 agents. Why? Why do I need so many agents? Why do I have 233 skills? The primary reason I have so many is because I do a lot of demos. And so just having a lot of tools by my side is good, but you don't need to have so many, right? It it just adds bloat. So let's do model and let's go to auto. And now we're in auto mode. And it looks like the first model selected is Quad Haiku 4.5. But we haven't given it a task yet. So I'm going to switch into plan mode here. Oh, let me actually hide this comment so you can see what I'm doing. I want to make sure that you can see what I'm doing. Uh, will there be a list technical live session? So, usually Alexander, we do like to get into the weeds in these streams and just have a bit of technical fun. Um, let me know what you would like to see and I can bring it back to the team and maybe we will do it. Let me know. We do have a stream called Open Source Fridays where we talk to maintainers about what they're building. But that's also relatively technical. So let me know what you would like to see and we can see what we can do. Okay. So let's test auto mode. Okay. So I'm going to as you see I selected auto mode here. So like I'm going to hit shift tab and I'm going going to go into plan mode in copilot CLI and I'm going to say okay I'm going to say like I want to make some nothing crazy. I'm not going to do any crazy prompting here to this app. What suggestions do you have that I can implement? um create a markdown file in the root with your suggestions that I can review. So essentially I'm asking it to take a look at the project and figure out okay what improvements can I make to this app and the reason I'm asking this question is because I want us to work on something together. So I just want copilot to like surface stuff that I can do right now with you. Um but this is this is interesting because right now it's using Claude Haiku 4.5. How many of you have used Claude Haiku 4.5? I've never I've never used it transparently. I've never used it transparently, but let's let's see what Haiku comes up with, right? Uh because we don't need to use the same model for everything. Like I am I am obsessed with like Codeex 5.4 and Opus 4.6. Um I'm still testing on 4.7 since it just came out. Uh and so like I don't need to use the same model for everything. You know what I mean? Because models are good at different things. I also like to use Gemini uh when whenever I'm doing like the front-end design work as well. Okay, let's see your questions while Copilot is running here. So, how reliable is the auto mode by picking the right model for a specific task? That's what we're testing right now. Um, Raphael, since this just came to the copilot CLI, I haven't tested it yet, so we're testing it together. So, right now, it chose uh to use the Claude Haiku 4.5 model as the first model. That was the first model that surfaced when I selected the auto mode. And I believe that's the model that's that's being used right now to get this work done. So yeah, let's see what happens. Oh no, sorry that was my phone. Okay, so let's see what we have here and then I'll go into some more questions. So what types of improvements are you most interested in? That's a great question. This is a question I would get from Oasus 4.6. fix. Um, if you've ever used Opus 4.6 in plan mode, this is something that it would ask all the time, right? So, it's nice to see that I'm getting the same type of questions, right? I can suggest improvements across multiple categories like features, new functionality, performance, UX design, code quality, reliability, or deployment operations. I mean, how about all of them? Like, give me a comprehensive list. Honestly, honestly, give me a comprehensive list because I have not shipped this project, y'all. I don't know why. I have this bad habit of like building tools for myself and not shipping it. Like, I just build it. I the adrenaline is gone and I just forget about it. And I need to stop doing that. So, this is a project I need to actually ship so that I can use it, so that people can use it because it's actually pretty good. It's actually pretty good. Okay, so Raphael, this is what we're testing right now. So let's see what other questions we have. Do you find that using the terminal to prompt is better than using the using the VS code track plugins or is it just personal preference for me dangerous? It's personal preference, right? So I still like going to VS Code and using the agent mode here because I can use agent mode, ask mode, and plan mode in um Copilot chat as well. It's like the same experience. It's just a matter of per of preference right now. Um the CLI team does ship every single day like very very often. So sometimes features in Copilot CLI is not in VS Code at the same time. So um that's primarily why I use Copilot CLI, but I still go to Copilot chat to get a lot of work done. It's just honestly it's just a matter of preference. We just want to make sure that we meet you where you work best. So if you like to work in the terminal, we got you. If you like to work in your editor, we got you. And if you want to, you know, test out working on github.com with copilot cloud agent, we got you there as well. You know what I mean? Like we just want to make sure that you work um you can work where you work best and where you prefer to work. So I wouldn't say it's better or worse. It's It's like you have access to all the models and all the things. Um, so yeah. Okay. So, it looks like Haiku is still doing the stuff here. Let's see what other questions. Uh, what time is open source Friday streams? Yeah, so we stream open source Fridays every Friday at I'm Central. So, it's at like 100 p.m. 1:00 p.m. Eastern. 1:00 p.m. Eastern every Friday. And there's a session scheduled for tomorrow with I believe it's Andrea who'll be doing it. So, definitely stay tuned. So, every Friday we talk to maintainers about the apps that they're maintaining. So, it can be like huge maintainers, smaller maintainers, and we're always looking for new projects to feature. So, if you're interested in being on Open Source Fridays, definitely go to the Open Source Friday repository, look at the requirements, and submit your project, and we'll review it. Okay. So, let's see what we have and then I'll go through some more questions. There are a lot of questions here. Um, let's see. Okay. So, look at this. We have a comprehensive improvement analysis that was done by Haiko 4.5. Um, exit plan mode. Yeah, I want to exit plan mode. Um because I actually want to see what this looks like. Oh, I think I actually have the project open. So, improvements.mmd. Perfect. Hopefully, you can see this. Okay, let me zoom in one more. And let's close this side panel here so we can just focus. All right, so this document outlines prioritized improvement suggestions across multiple categories. So, let's see. Am I going to click? Okay, great. Let's see. So, features, history, dark mode, reorder, editing, UI, batch processing, auto detect, chapter length. Oh, I like that. Progress indicator. So that's like a UX code quality add testing error boundary because this was a personal project. I I really didn't do any testing, you know. Don't don't come for me. How many of you add tests when you're building for yourself, you know? So copilot here with clot haiku is really showing showing the things that I did not do. Performance caching of course caching memorization streaming responses reliability rate limiting. Again, I did not I did not do that because it was just for me. It's it's local. It's local. It's not shipped yet. Um, health monitoring, retry logic. So, these are actually very very good suggestions on things that we can do. And I didn't have to use heavy Opus 4.6 to do it. Now, I think that's pretty good, right? I think I think Haiku did a pretty good job. efficient, cost effective, and you know, didn't have to use any of my premium requests here. So, chapter editing UI let users edit timestamps, titles, and descriptions before exporting. Currently, they're locked after generation. Ah, so I can show you how this app works just so you have some better context Probably should do npm install. I probably should start using bun, right? I probably should start using bun. But, you know, honestly, old habits die hard. I've been using npm for what, like 7, 8 years? Like, that's a long time. Wow, that took a while to start, huh? All right, so I want you to come over here. And you know what's interesting? I actually had a completely different stream planned for today. So kudos to you for, you know, telling me what you want. So this is the project. So essentially, it allows me to get chapters. And you know what chapters are? You know on YouTube whenever you go into the description and you see those, you know, if you goes at 000.5, you get this. That's what a chapter is. And the creator has to do that by hand. It's manual. It's it it's not autogenerated. It's manual. So, if I do a YouTube video, I can hit generate chapters and hopefully she works. And she works perfect. And then I get chapters here. So, what the suggestion is is to allow people to come in here and edit the edit the titles because right now it's just like the chapter is generated and I can copy it or I can export it as text and that's it. So, I think that's a pretty good feature suggest suggestion. So, let's let's try to do it with um auto mode. What do you think? I think that would be something good to do for a feature like that. Like if you wanted to add the ability to edit something in your UI, you know, like it's going to be it's going to involve like state management, all that stuff. What model would you use? What model would you use? I would love to know. Tell me in the comments. Okay. Tell me in the comments what model you would use. Okay. Okay. So, let's let me look at the comments and then I'll I'll do the thing. Hold on. what do you like more about Opus versus Sonnet? Honestly, I think Opus is is just such a high thinker um of a model. It it it's really able to just like think about it, think about even things that I miss um whenever I ask it to do something. So, I really like that. I find that Sonnet just does what you tell it to do without really like looking at the nuances of things, if that makes sense. So I really do like that about Opus and that's why I use it a lot for higher level work. Now with Sonnet, it just does what I ask it to do, which is excellent, right? So like if I'm making a small change on a method or if I want to, you know, implement some documentation, I'm going to pull for Sonnet. You know what I'm saying? Because it's going to do what I ask it to do with specific instructions. So I think because Opus is uh such a high reasoning, high thinking model, I I do like to use it a lot. But let's see what auto mode is going to suggest for this feature that we're going to implement together. Okay, I'll do like one or two more questions and then we can keep going. Let me know if you're enjoying this session. I'm having a good time. Well, let me know if you're having a good time. That's the most important thing. And also, let me know what you think of my project. Do you like it? Like, I think it's pretty cool. you know, like I like building applications for creators and so I'm going to ship it. I'm going to ship it, guys. Don't don't worry. I'm going to ship it. I'm going to add test. I'm going to add observability, analytics, all the stuff. Don't you know it's it's it will come together. It will come together. DIY projects, test, and pro Brian tests are the feedback loop that AI needs. I agree, Brian. I agree. Recordings will be available or not if I want to see later. Yes, sessions are recorded. You can watch on LinkedIn, YouTube, uh Twitch and also Twitter. So, let's keep going. So, let's implement the feature where you where we allow users to edit the title and also the description before exporting. Okay. I'm going to come back here and I actually want to go to the doc just so I can just like copy the request. This one it's definitely a high impact, right? So, so I can do two things here. I can go to copilot and I can say I can like add the improvements file here. So if you do at improvements, if you can spell, and then you do tab, that just adds the uh documents straight into the prompt, right? And I can say, take a look at the chapter editing UI suggestion and um create a plan to implement it. And I can start in plan mode to get this done. But I feel like, you know, that's just a lot of work. But I just wanted to show you that you can do that. I'm just going to copy pasta right now. So, I'm going to I'm going to leave this here, but I just wanted to know that you can do that. I'm going to do shift tab, go into plan mode. If you saw like down here, switch to plan mode. If I do shift tab again, then it's autopilot where copilot CLI will just run until it's finished. And then this is just like regular ask mode. So, I'm going to go into plan mode here. And I'm going to say um create a plan if I can spell at this point at this like point like in time. Typos are is such a core part of my brand. It's crazy. Create a plan to implement the following. And I'm just going to paste. Boop. Okay. Okay, so auto mode is still on claw haiku 4.5 you know I think the release said that it showed us hold on let's go back to the release so how it works auto is dynamic giving you reliable access to your favorite models while m while mitigating rate limits it routes to models like GPZ 5.4 5.3 codec sonnet 4.6 six and high school 4.5 based on your plan and policies. Oh, the model auto will route to will change over time. Let's see. For chapter editing, which interaction pattern would you prefer? I think I want inline editing. Click to edit fields directly. I think that would be um good. How should edited changes be saved? Save after every change automatically. Save with a manual save changes button. Save on blur. Okay. I actually like a manual save changes button because I like the feel of clicking save changes. It just tells me that it's changed. Even though like saving automatically is also good or saving on blur like once the field loses focus is also good. But I I like the button. I I just like a button. Call me old school. Should the editing UI include undo redo functionality? Oh yes. These are great questions. You know, plan mode is so good at helping you really think through planning um your feature or your implement like whatever you're doing. It's really good at helping you uh think through whatever you're trying to build So, I wonder if this model is going to change So far, it's been Claude Haiku 4.5, right? so Krishna, that's a very long message. Um, it's it's hard to read that message in the chat. Ooh, Henrique, this is a very good question. Did you have a risk matrix save as a MD beforehand or are you trusting each model's individual assessment? Honestly, I do not have a risk matrix save as a MD beforehand. What does that look like? I I haven't implemented that before. I'm curious. Things are changing so fast and it's hard to keep up. So, right now I'm trusting the models individual assessment. Um, I do have like agent skills and I do have um custom model, but I'm not using any custom agents right now. But I do have skills that it can use, but so far it's just been the model doing its thing. Always a good time. I'm glad. Okay. I find sonnet so far very easy to use. Good. All righty. So, we have a plan. So, I'm just going to go to accept and accept plan and build on autopilot in the interest of time here. And it looks like we're still using the same model. I wonder when the model changes, you know. I'm not sure. Oh, thank you Bronn for answering. Krishna's question. So Krishna, I couldn't really I can't really read it. It's it's too long to see here, but I really like Bron's um response here. So Krishna, the advice is to create a portfolio with project that show you uh that show what you know. This will set you apart from the crowd in your job search. Just make sure the projects you share actually solve a real problem. That is very good advice. In this day and age, it's a matter of what, you know, you have to show what you can do. The the competition is is pretty high. And so, being able to show you what you can do in a portfolio is going to be very good. And also lucky for you, we actually just launched an episode on GitHub for beginners on using GitHub GitHub pages to um build and deploy your portfolio. So I'm going to show you that while while it's building here. Oh, I can just come here. Yes. So, if you go to GitHub's YouTube page, go under videos, you will see here, who is that? It's me. I'm kidding. But you can you can learn how to get started with GitHub pages. And in the example, we actually deployed or we actually did a developer portfolio site. So, it's actually very timely. Krishna, I would encourage you to go forth and watch this tutorial to learn how to, you know, build and deploy your your portfolio site using GitHub pages for free. That's the operative word here, free. Uh we also have courses uh to teach you how to use Copilot CLI. So if you're used if you're new to Copilot CLI and Copilot CLI is the agent I'm using right here. So if you're looking for more of a beginner perspective, you can go here and we have a copilot CLI for beg beginners session also running on our YouTube channel. We share a lot of goodies on our YouTube channel. So definitely check it out. Okay, definitely check it out. So cool, right? Look at it. GitHub GitHub's YouTube looks so good, you know? So many smiling faces. Okay, I'm getting distracted again. Let's go back here. So, auto mode is still on Claude Haiku 4.5, right? So, this is the model that we've been using throughout the entire entire session, which is interesting to me. Oh, look at that. Good job on the tutorial for static GitHub pages. Awesome. I'm happy that you like it. you know like for get up for beginners this season we uh I really wanted to partner with engineers in the company and they came on and they you know presented and it it's been great. I'm happy that you enjoy it and the template is also pretty good as well. So the template is also available for you to use. let's see what Copilot's done so far. Exit plan mode with a summary. It's going to start implemented. So now it's in autopilot mode. So it's just going to run until it's finished. It's editing the chapter list prop. Create validation utilities. It's using the current chapter history hook that's there. I was creating a hook directory. Wasn't there already a hook directory? Interesting. Great. Now, let me mark the phase a to-dos as done. Okay, let's see what it's actually doing. Where did it put the um Did it create a plan dock? I don't see one hook directory. It's editing that one. So, so far it's not going into like any file that it's not supposed to. It looks like it's it's doing like really targeted work here that I would expect from honestly a higher tier model. cuz you know sometimes models can just go off and do whatever they want to do. But it seems like Haiku has been very focused on the work and only touching the files that need to um edited here. So editable chapter file. I created this brand new component here to edit a chapter. It's using the format time from the utils. It's using all the validations that we have. That's very good. So, this shows that it actually understands the codebase, the existing codebase that we had that we have to in to um what am I trying to say to add this feature? We have our interface of course, TypeScript stuff. Timestamp, start, time, title, description. These are all the things that's going to be edited. Um, added a use effect here. I mean, honestly, looking at this code, what do you think? What do you think? Tell me before I tell you what I think. What do you think? Right. I'm like going through the code here and I'm like by any chance. Did you use copilot to use /fleet for implementation? I didn't I didn't I didn't feel that there was a need to use um to like spawn sub agents. I didn't feel that that there was a need to use fleet mode to implement this feature because it's it's just like one feature. So, I didn't I didn't think that was necessary. Haiku seems to be slower than Sonnet. that is one thing I will say. Um, and I'm actually happy that you pointed that out, Amy. It is taking a little bit longer to implement the feature. Um, but it's actually doing a pretty good job. I I I must say it has this handle description change, validate, feel, handle, timestamp, change, start, edit. Um, the code is pretty well organized as well. Handle field, save. Look at that. Don't exit mode validation fails. some really good comments handle can say it's thinking about key down just monitoring I don't know and then it like you know surface here and then we have the button to save you know I am it's taking a little longer but speed isn't always you know the best you know like speed isn't always the best all the time sometimes how are you reviewing you know the work that agents do. Are you reviewing the work that agents do? I usually use copilot code review on GitHub to get the review done and then like you know go back and and have the agent update stuff. But how are you doing it? Because I know a lot of people a lot of developers aren't reviewing the work that's being done because they're focused on moving fast, moving fast. Slow down, you know, and and still keep your your practices. But I do agree that yeah, it is it is taking a little bit longer because yeah, we're almost time. It's almost time to end the stream and it's still going. So I know code looks okay, but run it to really know ex I know I'm waiting for it to finish to see um if we can run it. If I say hurry up, do you think it will do something? there was a error login call somewhere. Now it's creating a test. It's better than me. You know, there's a transcript route issue. Huh? Hold on. What's going on? The body is defined at line 90 inside the try block. The issues that is being referenced in the catch block where it's out of scope. What are you trying to fix here? There's an issue in the export route. Let me fix. So look at this. Now it's gone to the export route. And this is what I mean when I say sometimes the modules I didn't ask you to fix the export route. And I'm gonna say that here Oh, I have a typo. See, I tell you typos are a part of my brain. You're absolutely right. I got sidetracked. Sir, that's not what I asked you to do. So, these are pre-existing issues not related to the chapter editing. Let me focus back on just editing and revert those unrelated changes. Yeah. So, I was giving it praise too soon. Um because I didn't I didn't I didn't ask you to fix the export feature. That's for another time. Um so it it did not elucinate. What it did was it was trying to do something that I didn't ask it to do. So I guess there's a there's an issue in the export feature. Um, which may very well be true. I haven't there are no tests in this repository. Um, but like it was just like going and doing something I didn't ask it to do. So, Cassidy, what are you doing here? Everyone say hello to Cassidy. You know she's on vacation. but yeah, I was I wasn't I was like, "Oh, hi cool. You're doing a good job." But I didn't ask you to do that, buddy. You know, um, and yeah, you know, it is it is slow. Um, Amy, I do agree. It is taking a lot longer to implement this edit feature. And the edit, it's not it's nothing crazy. It's nothing crazy. Oh, Amy says, "Hi, Cassidy." Okay. So, do you have an agents or instruction files in this project? Could that be guiding the model selection? So, I don't believe the agents.md file or the um do I even have those in here? You know, I'm showing all my cards on on what I do when I'm like not doing something to teach. So, I don't have an agents.mmd file, but I do have a copiloted instructions file, which I'm happy that I do. And so, it doesn't have anything crazy in here, but the instructions file isn't guiding the model selection. I think based on, let me just go back here. based on your plan and policy. So, it's it's a little vague. The models auto I wonder if there's more in the documentation on how the models are selected. What is I think this is the docs. Okay. Enabling access auto when you select auto from the list of other models. Auto model selection chooses from supported models. When using auto the model use for each terminal you can change the model. You can change the model compiler usage generation reference. You may find the different monforment. Yeah. Um, we'd have to read up more about it, right? Like how is it actually choosing the models in auto mode? We need to have like more blogs and docs and videos about it just so we have some more education out there. But I don't believe that the um instruction files are affecting are affecting that. But friends, we didn't finish and you know the time has come. The time has come. The time has come. Let me create a comprehensive test document showing what we built. Okay, let me go back to the UI to see if it actually if it's actually So friends, the lesson here is the lesson is um haiku. Oh no. Are you kidding me? Okay, so it looks like there was an error. So let me see if it will run again. Okay. Okay. Reload. Okay. All right. Okie do. Look at that. There's a save button. Oh, did it save? Save. Oh, I can edit but it doesn't it doesn't save, you know. But I can edit but it doesn't save uh test. Save, you know, but at least the edit stays. You know what I'm saying? So, what is the lesson here, folks? What's the lesson here? Oh no. What is the lesson? This was not a good demo. Um, this is what happens when you do stuff live, you know. Um, my lesson is that Haiku 4.5 is a no. Is a no for me. Um, you know, Braun, I don't know if I would just chalk it up to auto model. You know, I wouldn't I wouldn't say that. You know, I think it's I I Oh, no. That's how you know it's real. I don't even know what to say, but And Haiku is still running. And I'm just going to stop I'm just going to stop you there, buddy. I'm just going to stop you there. I hope you learned something today. I hope the session was informative. I wasn't specific enough. So, it's my fault. You know, maybe I wasn't. you know, maybe maybe I wasn't specific enough, you know. So, yeah, I think the lesson here is haiku for planning, not for implementing. But how can we control what model auto mode selects, you know, because I I I was curious to see if auto mode would have switched the model when it was time to implement, but it kept the same model for the implementation as well. I thought the the research and the planning was very good from Haiku, but I was expecting it to switch over to something like Codeex for example. anyway guys, that is that is what we have for today. Um do not, you know, make sure you use auto mode. Test it out for yourself. Yeah, I was very surprised that it didn't that it didn't switch for the implementation as well. But um Salvi, that is that's all we have for today. I will see you next week. You know, this is how you know it's live. This was fun though. This was a fun test. Um next week we can test it out I do agree. I love having all the different options, all the multimodel, auto mode, all the different ways to use it. It's, you know, we want to meet you where you are. But I'm going to end here, guys. So, bye. That's what Cassidy usually does. But