Here's my number

Matrix calls

I defeated my laziness, and now my Matrix homeserver supports Matrix Calls. It is based on WebRTC, and it supports fancy video calls as well.

It wasn't too difficult to set up, but if you are not attentive enough, LLMs will direct you somewhere else, away from the actual problem you might have with your setup. My main advice is: don't rush, just RTFD.

Is it complete?

I don't think there will be a point when I look at my homeserver and say, "OK, all the features are here, now I can just use it." No, it's not possible.

I'm grateful to all the open-source contributors from the Matrix ecosystem, even if I don't use their products. As for the products I do use, Tuwunel is going strong with new features. Commet is also my go-to desktop client at the moment, and Element X is my go-to client for iOS.

My next step is to replace that ugly Matrix registration page with a new, modern, cute OIDC flow. I will have to make a few decisions about the tech stack, and after that I will be ready to crack it the same way I did with calls.

The way of Vibing

I think I have the name for my agentic harness. It is already "kinda" usable, but I want to make sure it has only the really important features.

The features I already have might need some optimization. It's definitely something I will do before releasing it to the general public.

I watched a talk about Pi Agent and liked some ideas. By "liked", I mean some ideas were stolen for my CLI.

Lobotomites in 2026

Claude is extremely dumb. I understand Opus 4.7 is probably "different" from 4.6, but this flavour of different results in horrible decisions. 4.7 makes some abso-fucking-lutely DEGENERATE assumptions, and it's lazy all the time. Sometimes it decides to do something easy to miss. Code review is still a problem that I haven't completely solved.

GPT 5.5 rocks. I understand people have different experiences, and this experience is mine. GPT catches potential problems, follows instructions, and DOESN'T make crazy assumptions.

One might say "skill issue". First, maybe. Second, the regression is obvious. Opus 4.6, before lobotomization and even after lobotomization, behaved like it knew what to do. It was like talking to a software engineer. 4.7 is an "insert your swear of choice".

What do I have now?

Well, the first problem I started solving for my agent was task management. Jira, Trello, or anything else like that is medieval tech.

For now, I haven't really incorporated any of the modern Scrum/Agile features, but I think I might implement them later. Of course, it would only make sense if it helps with task visibility and prevents errors related to working with the task management system.

I don't know many people who liked all of this Scrum/Agile-adjacent BS. Ideally, you have a project manager, scrum master, product owner, etc. I believe that all the bookkeeping can be done by LLMs.

For people, it means they will have the opportunity to start working on something else, something more fun.

Conclusions

I will switch from my Claude Max subscription to an OpenAI Pro subscription. I am not the biggest fan of either company, but subsidized tokens are subsidized tokens.

The next step for my Matrix homeserver is a new registration flow.