I also added some CI tests in Github actions that auto detect dev server slowdown accumulating over time, and they run on every PR automatically. Hopefully this can guard us against such issues in the long run.
Was chasing around a memory leak last night in Stati dev server. Turns out it was the CSS watcher all along. Replacing the watcher glob with an explicit list of files resolved this.
Well, I might have built a implementation loop orchestrator for GitHub Copilot. Still in early development and without polish, but damn it works great.
It's starting to look like and work like the real deal. Still a lot of features planned.
Good thing is that the tool can build itself without any issues! So that's nice.
Did a major refactor, extended configuration options and refined agent prompts. Now includes 40 specialized agents and orchestrators. Spawns subagents to be cost effective.
Context engineering is soo important if you want to have a tool that can pass tasks automatically between different agents.
So starting with an empty context is great and all, but don't forget to add the right context to start as well.
Monitoring improved a lot, and the script as well. Might need some more love, but feels pretty usable for now.
Let's see what a few more iterations result in.
I finally built my own Ralph loop script. Spawning sub-agents is good and all, but it has its limitations as you can run out of orchestrator context on bigger chunks of work.
Running my own script is much cleaner as a new context is guaranteed every run.
It's pretty barebones right now, but it gets the job done.
Right side is a batch script with the loop and all the logic, left side is a node monitoring tool that shows you the latest logs so you can keep track.
- And now the fully automated implementation with review rounds runs since 8 this morning.
This would mean you can build a fairly complex game from idea to MVP within 24 hours. Wild.
I did a personal record in agentic dev today.
- GDD with AI help was done in 2 hours, a fully automated implementation plan was generated in another 2.
- Task split, deep dive and review rounds fully automated were 6 hours. (62 tasks 60k+ lines)
- And now..