I made my claude code talk to the claude web sessions and vice versa. No copying around anymore needed.

Long time no see, I hope you're doing fine. I built a live updating HUD (heads up display for my claude code). Which means: NOW, my scientific researcher can run in Claude Code AND I get automatically updated what it's doing. Funnily enough: It does NOT need any tokens. That's the restrictions I built it with. Then console at the right top as you see, can be opened at ANY time, it doesn't influence the rest. BUT, it turned out to be VERY helpful. Sections in the Current Output: - Headline with Date - Calculations with subsections - Intersession Bus: Since I made my claude code instances talk to each other The Intersession bus shows how many messages AND which went back and forth between different claude code sessions, termux is my Android Phone. P.S.: Finally I got time to post some things on Saturday again, at least for some minutes. :-) The update AUTOMATICALLY changes, depending on what needs to be displayed.

Yes. It was token off mid session I was in past night here. Around 3am CET. I literally was executing one statement in claude code, the next it was like: HINT: no need to watch the whole thing. Feel free to skim through. This model cannot be found. So I had to go back to opus 4.8. What I was working on amongst other things is this: An empathic voice agent that has real empathy. The empathy engine already exist since two years. Now I am building in push to speak, etc.

@Nicole McCain AI I was managing now to rebuild something that's working incredibly well in roughly 15 minutes. But have to ask my architecture to detach from some components it used. P.S. Holy shit. Now it even surprised me. My architecture added something on top. Hold down RIGHT CTRL speak into any text field. It works. Even here in the browser. So now I got two versions. Haha. I'll definitely merge them.

@Nicole McCain AI A working version of my own . running fully on my own machine, came together in under an hour today. So my estimation was accurate. 1-2 days. If you have a better os with programming skills etc. like I mentioned before... some hours.

Finally got some time again (occasionally, at least 😄). What I've been building is one system, not four tricks: → a local, self-owned voice model — runs on my machine, I own the whole stack → multiple Claude Code instances that talk to each other over a shared channel they split the work, review each other, and catch each other's mistakes, with no human relaying messages → a background monitor listening for that cross-talk → and when something lands, it's spoken aloud —> I can hear the system think It's the instrument for the scientific research I run inside Claude. The payoff isn't any one piece it's that the system switches between research and engineering mid-flow without dropping the thread: a scientific question that pops up during a build gets resolved in place instead of stalling it and an engineering need that surfaces during research gets handled the same way. Research unlocking development <-> development unlocking research ALL in one running loop. What it's got you covered on: knowing whether a result is real or you're fooling yourself. It deflated one of my own findings that looked like a discovery, then handed back the smaller, real thing hiding underneath. What it can't do yet and I'll say it plainly: originate genuine novelty that doesn't just reduce to something already known. Not there yet. The system is strong at checking, grounding, and sharpening; originating true novelty is the open problem. Naming that gap precisely is half the work.