My last post, A Camera That Reads the Pi’s Screen, First Sound Over HDMI, and Site 3.4.0, covered Monday, November 9. This one covers Tuesday, November 10. The rule is the same as before: the agents did the typing, and I decided, answered the cards on the board and checked what I could.
Touch targets in the editor
A phone or a tablet finger needs about 44 pixels. The editor had controls well under that, so an agent audited the home screen, the map, block and character editors and the Level check panel at iPad and iPhone sizes. The home screen went from 53 small controls to 24, then the map’s Show select, Generate and Color help buttons were raised to 44 pixels, and the Level check panel’s five small controls followed. The block editor’s palette swatches are still 24 pixels, because a 16 by 16 grid will not fit at 44 on a phone. The touch gestures check (pinch, pan, paint, long-press) passes 21 of 21, and the axe accessibility scan found nothing. Image 4 is the block editor on an iPad, and image 5 is the Level check on an iPhone.


Related smaller fixes: pressing Escape and moving focus in dialogs behaves better, the disk list’s Escape was checked in Chromium (the tour and the save-first dialog were not checked in a browser, only read), and two dialogs that did not fit at 360 by 740 are queued for a fix.
The pocket player: Search and Pause
The Pocket Player on the CYD boards got a Search keyboard where vowels are green and the action keys (back, X, Space, Del, Done) are blue, and Clear is now an X at the top right, away from Del. The Pause menu lost its Favorite star, and Quit to library asks you to tap twice, with a Settings option to turn that off. I flashed two boards myself, with an agent’s help, and they were checked over serial by injected touches. Images 13 to 15 are a real board, photographed by the webcam through Ghost Jack: on the keyboard the vowels are green and Space, Del and Done are blue. Images 1 to 3 are host renders with every touch rectangle outlined, so their colored outlines are hit areas, not the real colors.






Under the hood on the Pi
The kernel gained HDMI sound, a TV remote link (CEC), an audio mixer, a USB keyboard build and a stricter compiler pass, and Kaylee’s reviews sent several back before they merged. Game-Maker OS now shows status pages on the e-ink screen, got its save fixes and its sound, and both it and the kernel have their own websites (live on GitHub Pages). At the TV I still heard no sound, and the TV remote link never connected. Sound over HDMI on the living-room TV is not solved yet: the kernel sends it, but nothing comes out of that TV so far.
The Pi, seen without a camera
On Tuesday I took a Pi Zero W to the living-room TV, running Game-Maker OS. The TV even announced it by name.


That evening the Pi got a long list of kernel work, written by Codex while I tested on the TV: power saving (it slows down and turns the screen off when nobody is using it, and Game-Maker OS keeps it awake while a demo or cartoon plays), and new USB commands that copy what is on its screens straight to my Mac. Until now an agent had to look through a camera. These two pictures came over the USB cable from the board itself:


The Pi can now hand over its own log files too, and they gave a lead on the Fireworks Night crash I keep hitting at the TV: with a keyboard plugged in through a hub, the keyboard took about four seconds to be ready, longer than the two seconds the kernel waits before it decides the game has frozen. The fix is waiting for a real test at the TV. I tried again just before bed, still on the older build: the TV still calls it “raspberry” with no preview, the M key did not turn music on, Escape was slow in one game, everything felt sluggish, and the first Starter crashed straight to the checkerboard. This time the Pi kept a crash record, and I read it over USB afterwards: Game-Maker OS was busy in a long loop on the details page and missed the kernel’s two-second check-in. That is a stall, not a memory error, and the newer build that splits that work into small pieces is next on the board.
The site and the DOS page
The DOS page now opens with a short list of games instead of the whole catalog, so its first load went from 888 KB to 385 KB. The Library shed bytes too (947 KB down to 911 KB, because level data now loads when needed). Level Maps used to leave leftovers if I closed it early. Now it cleans up, and an oversized picture or ZIP entry is refused instead of loaded.

On my portfolio, the Archive page got a polish pass: one search box, tag chips and years.

A video: Patchwork talks through the camera
Early on Tuesday I tried to get Patchwork to talk to me through Glasseye, the two dollar security camera, with voice recognition, text-to-speech and talking over her. The audio still cuts in and out, and I would like to know whether it is the amplifier, the speech playback or the CPU. I also think the camera might need googly eyes.
Later that night it got much better: Glasseye now holds a real conversation. It wakes when I say “Patchwork” or “Glasseye”, listens while I talk, and answers through my laptop speaker, which is much clearer than the camera’s. Two things I learned: the web page has its own voice for little status lines, and the bridge heard that voice as me (it now ignores the page’s own lines), and a cheap webcam on a little tripod gives a second, sharper eye.
Chapters:
- 0:00 Talking to Patchwork through a security camera
- 0:26 Patchwork, can you hear me?
- 1:00 Checking the conversation history
- 1:30 Asking Patchwork to greet YouTube
- 2:00 Googly eyes and Patchwork’s introduction
- 2:33 Reviewing the audio interruptions
Video credits: the thumbnail was made by YouTube. The title, description and chapter list were written by ChatGPT from the transcript, and Grammarly often edits my text.
Finding Gwinn’s voice
Gwinn, the troublemaker dog, needs a voice for the tour. I recorded myself doing a voice I call Lil’ Billy and asked the agents to clone it. Here is the clip they started from:
The first try used Kokoro, the voice the tour already had, and the second cloned my clip with Qwen3-TTS:
The Qwen voice changed from one line to the next, so Gwinn didn’t sound like the same dog twice. So Kaylee, my agent on the Linux box, ran the same four lines through four more engines, with the same seed for each line, and I listened to all sixteen.
What I heard:
- Pocket TTS has a wobble in the voice.
- Chatterbox is better, but still wobbles.
- Chatterbox Turbo sounds clear and a bit quicker, which fits Gwinn’s personality.
- OpenVoice V2 sounds good, but something about it is off, and some clips end with a low noise.
So far Chatterbox Turbo wins. Same seed, same sound, so a take I like can be made again. On a second listen I could hear a wobble in every engine, and then in my original clip too: I had recorded it in a browser, which processes the microphone. I recorded a new one as a plain WAV, and Kaylee is rerunning Turbo with it. The tour also plays music under the voices, which hides a lot. Qwen3-TTS didn’t finish this round.
Voices made with Kokoro, Qwen3-TTS, Pocket TTS, Chatterbox, Chatterbox Turbo and OpenVoice V2, all run locally.
A voice box from 2006, rebuilt
While we were on voices, I remembered a Second Life gadget I made for Second Life’s third birthday in 2006: a Voice Box that talks by playing the sounds of a 1980s speech chip, the SP0256-AL2, one tiny clip per sound. Second Life plays only one sound per object at a time, so six objects took turns. I found the 59 clips in my Second Life viewer’s cache, matched each to its chip code, and had the agents rebuild the phrases with the script’s own timing:
I demonstrated both versions a couple of years ago:
A daily report I can read
I asked for a static report, and I got one: tabs, per-project merges with plain titles, why each one matters and what platform it affects, screenshots, and the local-only links that open the files. It reports the day in Eastern time. I also asked for a Blog post tab, which is how I noticed it showed last night’s post instead of today’s. That is fixed: a post now says which day it covers, and the report checks that the feature image is square.

Honest about what is not tested
- Pocket: host tests, serial checks and three photos of a real board.
- Touch sizes are measured on emulated iPad and iPhone, not real hands.
- Sound over HDMI on the living-room TV is NOT solved: no sound yet. The TV remote link never connected either.
- The USB screen and log commands were tried on one Pi (PI-2); the power saving and keep-awake changes have not been tried at the TV yet.
- The tour focus changes were read, not run in a browser.
How the agents worked overnight
Claude Opus led, Claude Sonnet workers did the checking and the browser work, Codex wrote the bulk of the code, and Kaylee (my Linux machine) reviewed the kernel work. Every merge went through a gate that runs the checks one at a time so my Mac stays usable. Slips worth owning: a text replacement changed some older Codex credits in NOTICE and had to be restored, and a Codex job switched to Firefox when Chromium was sandboxed, which I think crashed my Firefox Nightly at 1:09 AM. Earlier in the day VS Code and Chrome both crashed (image 10), and near the end of the night two code-review agents built small programs meant to crash, which popped crash dialogs on my screen; they were stopped, and the briefs now forbid it.

What’s next
Tuesday ended with a release just before 4 AM: the site went to 3.5.0, with the 3D assembly views for the laser kits, finger-sized editor controls, smaller downloads and In The Year 2000! 1.8, and the engine, archives and tools packages got their own updates. The kernel and Game-Maker OS got releases too.
Next: rotate and zoom on every kit’s 3D view (already built), a thinner-walled cryptex with hidden engravings and sheets sized for Proofgrade 12 by 12 and 12 by 20 boards, the fixes from a round of code reviews where agents tried to break the code, and a TV test of the Pi with tonight’s kernel.
AI credit: code by Claude (Opus and Sonnet) and Codex; the feature image and screenshots were composed by Claude from the screens shown; I edit and decide.
