Copilot CLI update: chronicle, plugins, and fleet mode | GitHub Checkout

Look at the CLI work. Oh, it's going. Look at the CLI. It is going. You could do /tasks there and see all of these things and see how they're doing. Welcome, Ryan. Thank you for being a GitHub check out. Yeah, thanks for having me. Excited to be back. Copilot CLI has become my main driver. I've been using it. There is so much that you can do in it. It is hard for me to keep up. There are loads of functionality that's being added each and every day. If you're like me, you like to be on pre-release builds, so you get the newest things that drop immediately. And also, you know, it's always great to make sure that you have the experimental mode turned on. You could do /experimental show to see what experiments we're currently running. There's some really neat stuff in here. Like, for example, we have a sub-agent that can manage your MCP servers and agents and skills. That's a great way to make sure that, you know, when it comes to all of the ways that you can extend the Copilot CLI, that it's going to, you know, help you manage all those plugins that you're installing, all those extensions that you're installing. I love plugins because they allow you to combine MCP servers and skills and custom instructions and custom agent definitions all under one roof. Because if you are an end user, it might be very confusing as to when the best time is to use an MCP server, when the best time is to use a custom agent, or when it is to build a skill. And so, what we've done is we've made the plugin system. But if I do plugin marketplace marketplace browse, awesome Copilot, we can take a look at some of the really cool plugins that folks have built here. And there are plugins for everything from I like this one, double-check three-layer verification pipeline for AI output. So, it makes sure that the AI isn't putting any kind of hallucinations or nonsense in there. You know, if you use .NET, there's a bunch of skills that are good for you to be running in .NET. And what's great is I don't need to worry about whether the people who made this .NET plugin decided to use skills or custom agents or MCP servers. I just run one command to plugin install that, and then I get the benefit of all of that in my sessions. If you are creating extensions, you know, extensibility concepts for the CLI that you want other users to use, I highly recommend packaging it as a as a plugin instead of as, you know, a mishmash of of skills and whatnot. All right, can you show us this in action? Maybe you didn't know that plugins were a thing. Maybe you didn't know about all of the slash commands that we are adding all of the time. Also, you can press a question mark here to kind of see all of the different keyboard shortcuts and and slash commands that we've added over time cuz there's just so so many of them and we're adding more all of the time. But I can't keep track of all of those. Even as a PM on the team, it's really hard for me to keep track. And I want to know that I'm using the Copilot CLI to the best of its ability and I'm getting all of the wonderful goodness that it comes with. That's where the chronicle command comes in. So, this is going to require that you are in experimental mode right now. But what that's going to do is Copilot's going to have a maintain an SQLite database of every single session I've ever had with it. It's going to remember every single prompt. It's going to remember every single time that we debugged something and it wasn't on the right track. It's going to remember, you know, just how I speak to it and the kinds of things that we that we go over. And it all stays local on your machine. But I can use that to figure out how well I'm using all of the functionality of the tool. So, I can do things like chronicle and then I can say improve. And so, what chronicle improve is going to do is go through all of my session logs. And it's going to pick out places where, you know, maybe I can be a little bit better with my prompting, right? Maybe, "Yo, hey, if you're a little more forceful with your prompting, you know, you might get this done a little bit more efficiently." Or it might look at my Copilot instructions and look at all of our sessions and say, "You know, there's there's something that you keep reminding me of every single time. Every single session you're reminding me to do this and do that. Like, why don't you put it in your Copilot instructions file? Or, hey, you execute this action over and over and over again. You do the same thing. You ask me to do the same thing. Why don't you build a skill for it?" It's noticed that I haven't created PRs for me a lot. I keep asking it to update the PR format to use the same format as other PRs in this repo. Because on github.com, we have we have formats, right? And we have and we have templates. And whenever we're making a PR, people are using the templates. When it's just making a raw PR from the client, it doesn't know about that. And so, it wants me to document the expected PR format in our, you know, in in our Copilot instructions. Oh. Yeah. And so, like, you know, if I do that, then, you know, then it will know that that's the way that PR should be formatted and then it won't, you know, prompt me to do it every single time. It sees that I'm always referencing issues and other pieces of customer feedback and everything when I'm writing these PRs and I'm always asking it to do that as well. And so, it's like, "Well, you know, maybe I should add an instruction to when creating GitHub issues or PRs, include any URLs from the conversation as links in the body." That way those make sure those go in there. Each of those it will it can add to my Copilot instructions if I wanted to. This is a way to make sure that your context and your agent instructions are always up to date and always making sure that, you know, you're using the tool to the best of its ability. Chronicle improve is about improving the instructions and the context that the agent has. Chronicle tips is more about me. It's more about how am I going through and talking to the agent and how am I engaging with the tool. It's saying that I'm frequently just pasting GitHub issue URLs and PR links and docs pages and asking the agent to analyze them. And it's reminding me that, "Hey, there's a command to do that. It's called /research." /research is awesome cuz it will kind of go and do a deep dive across web searches. It will look across GitHub and do semantic searches across, you know, our platform for open source repos that might be doing similar things to you. What I've gotten to the habit of doing is as I'm kind of winding down at the end of the day or the end of the week, it's a Friday afternoon, I'll run chronicle tips and I'll, you know, just kind of see what it says and try to incorporate that into what I do next week. Listen, this is all really solid feedback, Ryan. Like, and things that you wouldn't think about. Like, you're deep in these work sessions and you're not thinking about how you can actually, like, do a self-healing of sorts about the way that you're working with Copilot CLI. Yeah, self-healing is an excellent way to think about it. And I like that we have both the chronicle improve command to help improve instruction files and context that help your entire team and working in your repo more effectively. But then we also have chronicle tips that is more focused on you and making sure that you know how to use these tools the best that you can. Because the agent is keeping track of all of your sessions, it's much more easy to become a power user and to use these tools much more deeply and much more effectively. Can I ask you about security, like, cuz you mentioned it's a spins up as a SQL database that is local to your machine. Yeah. This is just indexing all of the session logs that are already on your computer. And so, they're not going anywhere. They're in they're in your home/.copilot directory. You know, this is where all your config and all of your session logs and all of your history is stored for the Copilot CLI in this directory. You don't have to worry about your team, for example, or your boss being able to see what the output of chronicle tips would be for you. Fantastic. Let's keep my shame in the privacy and comfort of between me and Copilot CLI. Yeah. Now, that SQLite database, it doesn't just get used when you run the the chronicle command. If you say, "Hey, we were talking about this last week." Or like, "Hey, we tried this approach before and it didn't work." It will still dynamically use the tool to go read the database and go back and see, you know, what happened and use the tips from there. So, you know, you don't have to even remember to run chronicle to get some of the benefit of reasoning over your your history. Amazing. This is epic. Now, let's see how our background agents are doing from before. We have our Gemini 3 Opus 4.6 and GPT 5.4 refactoring agents running. Something else that we shipped somewhat recently is when you have your background agents, you can do /tasks to see what is going on. And so, you can see that it's been running for 6 minutes, almost 7 minutes. And I can click enter on it to actually kind of go inside of the session and see and see what it's doing. Sometimes they get into their own little loops and you just want to you want to stop them. Once each of those returns, our main agent is going to synthesize all those recommendations together, like it's saying, and give me one report to to read. And, you know, obviously what's great about sending it to multiple models, each one might have a slightly different perspective on things. You know, I've heard that a lot of folks think that Gemini is better at front end and better at design. So, it's possible that for the front end parts of this project, Gemini might have, you know, the better recommendations. And, you know, sometimes all three agents are going to come up with the same problems. That's a high signal that we should probably do something about that. I love Claude Opus 4.6 a lot. It's my favorite model. And that's the one that I'm using normally on high effort level here. Sometimes when I'm spending, you know, hours-long session with Opus writing code, I know that there's some blind spots in it, right? Cuz I'm in the same session with the same agent context. And sometimes we get caught on the minutia a little bit. You know, the agent starts to just latch on to some concept and it just won't let go. And, you know, this is how you lead to over-optimizing for certain paths. And something that I really like to do every now and then is say, "Let's refactor our code to make it simpler and easier for humans and agents to work with and maintain." Now, I could just ask Opus to do that. But I want to get opinions from some of the other models. So, I'll say, "Spin up background sub-agents using Opus 4.6, GPT 5.4, and Gemini 3 and collect all of their refactoring recommendations." I love how you can be very specific about what models you want to can kind of get those sanity checks from. I like to use, you know, like, the best one available from each of our providers that we have, right? So, each one is independently looking at my full code base and is going to produce refactoring recommendations. But, yeah, you know, there's nothing stopping you from doing this with more models as well, but you look, if you're going to use Opus and GPT-5.4, you probably don't want to use, you know, the minis and Haiku as well, I think Right, right. You get pretty good uh feedback there. All right, so our three background agents have completed and our main agent has uh sorted them into three tiers. Uh so, here tier one, these are the ones that all three models said we should probably fix. And so, hey, there's loads of duplicate code. We have a big monolithic app.py. We have a god class that's 1,200 lines. We should probably extract that and decompose it. And it was telling me what the effort level is and what the impact is going to be on the maintenance as well, which is quite nice. Here are things that at least two of the models found. Only Gemini found that if we use the state pattern, the command pattern for our main simulation loop and dispatching actions, uh this will make things a little bit easier to find. There are big three that are uh at the core that are the core maintainability bottleneck. And it wants us to begin with the quick wins. A lot of them could probably be done in parallel, right? So, if I just say go ahead, it's just going to you know, by default, it's just going to go down the line one by one by one and do all of them. But, if I want them to all be done in parallel, we have our fleet mode. And not only do I want it to do all of these things in parallel, but I also just want it to keep going until it's all done, right? I want to be able to go up, go have lunch, and when I come back, everything's been done. And so, I'm going to go into autopilot mode. The way I got here was with a shift tab, so you go past plan mode and into autopilot mode. And so, what autopilot is going to be mean is agent, keep going in a loop until you finish all the tasks that I assigned you. And then if I do {slash} fleet, that's going to have a bunch of parallel sub agents that are all going to um kind of it's going to compose a to-do list and then it's going to see how it can parallelize things and make sure that it's parallelizing things across multiple sub agents. So, I could say, let's implement the quick wins and plan out each of the bigger three. Excellent. Ryan, could you have told fleet what model to use as well? In general, whenever the agent is going to go and spin up a a sub agent, if you say use this model for this task, it is pretty good at at doing that on its own. And you know, again, that's just really one of the superpowers that you have when you're using when you're using GitHub Copilot. Um you know, if you're if you're using Claude code, uh you know, by default, you just have Claude's models. If you're using Codex, well, by default, you only have OpenAI's models. We bring you the best of the best. We bring you the models as quickly as we can. You know, usually on the same day as they're released. Rest assured that with your Copilot subscription, you always have access to the latest and most powerful models. There's 15 to do items here, so quite a lot of stuff. Nine of them can just be done in parallel, which is great. And then three of them are going to be done after the quick wins. What's great is I don't have to watch it and I can just leave it here and then go get lunch. So, I think I think that's what I'm going to go do. I always forget about fleet and it's like so incredible. In kind of the power of cuz as we saw with the analysis that the different models said, like maybe if it is a more meaty front-end task, I want to assign that to Gemini because that's what I had the best experience. So, having the choice that I can pick the right model for the job is incredible. What's next? What's coming up for Well, you know, I can actually make use of a of a of an in-progress feature that um actually will allow me to keep track of what the agent is doing when I go have lunch. Being able to connect to remotely running Copilot sessions is something that we're looking at right now and something that I uh think I might try out uh when I go have lunch to see how this task is going. Oh, I can't wait for that to land for everybody. That's going to be really huge. And that was an exciting look at all the new features of Copilot CLI. Have you checked it out? Have you tried out Chronicle? Let me know in the comments. Leave a comment. I'll be sure to reply to you. Please don't forget to like, subscribe to this channel so you never miss another dev tip or feature update. Push those changes to main and I'll catch you on the next release.