Feed

Builder 0.10.1 update:

• Fast async file search: type @query -> Tab to autocomplete.
• Token counter improvements: counts context much more accurately, which makes it feel like you have ~+15% more context window.
• Compact, rollback, and supervisor no longer invalidate caches, saves a lot of money.
• You can pick a different model for supervisor right during onboarding
• You can skip sign in, e.g. for using local models
• Better system prompt for more frequent subagent runs + more comments in the code
• Lots of fixes for when events don’t show up in history
• Fix for 400 when compact interrupted mid tool call.

TIL: If you use Codex with dangerously skip permissions access mode, you should also set approval policy to never, not to anything else, because all approval policies except never include conflicting system instructions! In config.toml: approval_policy = "never"

ksrc 0.5.1

• You complained that models don’t use ksrc on their own, so I reworked the prompt, making it way more aggressive.
• Now it’s Apache licensed so it can be used in enterprise. I didn’t know AGPL imposed restrictions even on binaries.
• Brew now publishes the binary instead of building from scratch
• Now even if your project or gradle files won’t build, ksrc will work in partial mode instead of throwing an error.

Codex users need to update the skill manually.

Published my FlowMVI skill for AI agents:

https://opensource.respawn.pro/FlowMVI/integrations/ai-agents

Also documented at that link how agents can read the library docs. If you install this skill, your agents will be able to create new plugins on their own, will stop making compilation errors when working with the library, and will use the right existing APIs instead of inventing their own stuff. I’ve been using it in Respawn for a while now - Codex immediately started cranking out all sorts of plugins, pipelines, event buses with FlowMVI. Seems to be vibing with the lib 😗.

JetBrains just announced a new IDE called Fleet. I dug into the internals of the installer and runtime of this IDE: It’s just a fork of JetBrains Fleet with some small UI changes and a redesign for agentic programming. Written in Kotlin + Compose Multiplatform. Skiko is used for rendering.

And I have another wonderful piece of news for you. With Kotlin 2.3 you’ll have to turn off configure on demand if you have a JS target, significantly slowing down configuration time. This is just… again, I have no words. The Kotlin team had a bug, and they decided to fix it by making it a build error until you disable configure on demand. A key feature that speeds up Gradle configuration. https://youtrack.jetbrains.com/issue/KT-40822. 🤡

Updated dependencies in Respawn. The new AGP version deprecated the old android plugin. And the new plugin has no build types, which means no BuildConfig.DEBUG flag. No flavors, no codegen at all, you can’t do resource placeholders, and even tests aren’t enabled by default. Now I have to rip out my global isDebuggable flag that I use in hundreds and hundreds of places, and a simple Android Gradle Plugin upgrade, as always, has turned into a 2000-line migration, with Codex already trying to deal with it for two and a half hours. I have no words. Make SURE that before updating to AGP 9.0 (it’s ALREADY in beta now), you remove ALL usages of BuildConfig from all modules except :app. In 9.0 the project just won’t build anymore.

The second cool news is that last week a Junie CLI version came out, meaning another console programming agent. I’ll be testing it on the weekend and will report back to you, but first an article on Codec vs Claude. I’m extremely curious what a TUI written in Kotlin looks like and how it works.

Perplexity for Android came out today - or rather I got the email yesterday evening. I’m not sure yet if I should switch to it, because I’ll lose Google sync and password management. But I think it’s really good news. I’ll test if it has an agent mode and report back.

NEVER use MutableStateFlow.update function in your Kotlin code. This function will execute an infinite loop that compares old state to the new one, and if they don’t match, ALL code that you provided in the lambda will be re-executed. Trust me, you don’t want this for code that charges the user’s credit card.

Compose team have deprecated resource accessors for the Compose Gradle plugin in 1.10-beta. I never understood why they needed them. I guess it was for convenience and to look cool, but now I have to fix 50+ deprecations across 5+ projects. Why deprecate them? Well, time to enslave Claude again ig 🤷🏻.

https://github.com/JetBrains/compose-multiplatform/pull/5462

In release 1.10.0-beta01 of the Compose Multiplatform, JetBrains have added ability to draw their Compose UI below the UIKit UI, which means that now we can implement liquid glass components as normal native views without all the bugs and hacks that previously we had to use like inverting the relationship. This should make liquid glass adoption incredibly easy now.

https://github.com/JetBrains/compose-multiplatform-core/pull/2501

Guys, exciting update. As you can already see, I started making small posts more frequently. That’s because I made a new separate section on my website called Feed. All posts are now translated to English and Russian automatically. And you can view them in two versions on the website or subscribe to them using the RSS feed. They are also now published to Twitter, Telegram, and more automatically for me. So creating content has never been easier for me. Every time I have a piece of wisdom for you, I can share it within a minute.

Recent Claude update added a plan sub-agent, but the plan sub-agent is just a duplicate of explore sub-agent. This is obviously a bug, already filed on GitHub and will be fixed soon. In the meantime, I renamed my feature planner agent to be named just plan so that it overrides the broken plan sub-agent with my own custom agent, which works better anyway.