Mini Cyberdeck

By 2 September I had added eight keys under the screen, four across and two down. For a rapid prototype, I cut holes for them and a window for the screen into the plastic box that came with a two-inch display, and photographed it that day. All eight share one analogue pin, each at its own voltage, and I built the ladder from the resistors I had on hand: a 1k where a 10k should go, then two 5k in series when that read wrong. The firmware learns what each key reads, so the exact resistor values and their tolerance don’t matter, as long as every key reads differently. It switches off any key that reads too close to another and keeps the rest, and it leaves the whole pad off if fewer than three can be told apart. The first time I ran it, it refused to arm because the ladder I’d built read wrong. The launcher takes the pad’s shape, eight tiles in four by two, so each key opens the tile in its place.

0.6ssourcemain/buttons.c:61-62 at b60d541: LONG_PRESS_MS 600, LONG_MAX_MS 1500.between a press and a hold

The device reads Claude Code’s state from the terminal itself. The bell it rings when it stops for a person means NEEDS YOU; braille spinner glyphs, or output in the last 2.5 seconds, mean WORKING. Put down, the standby face says which in large type, and a NEEDS YOU lights the screen even untouched. Press Reply, or double-tap the glass, and the sheet covers the top half only, so the question keeps painting underneath. The first reply menu covered the whole screen and hid the prompt it existed to answer; the design panel I asked for on day three caught it.

On 12 September I asked for the recorder to feed Hermes, the agent on the Pi. A note recorded in VOICE goes over the tailnet, is transcribed by Whisper through OpenRouter, and Hermes answers on the device. It worked, and that night I asked why Hermes did so little with the notes: its prompt told it to file each one and reply in a single line. Now Hermes works out whether a note is something to keep, a question or a request, uses its tools to do what the note asks, and answers for a 320 by 240 screen: the outcome first, and never a question back. On 24 September the recorder and the notes list became one VOICE app at my ask, and you decide what a take is for after you’ve made it.

The first update over the air, on 2 September, failed with a screen telling me to put the file in the home directory on the Pi. The file was there. The real cause was memory: the update opened a second SSH connection beside the live session, and its first read needed a 16 KB block when the largest free block of internal RAM was 15 KB. So the failure screen names the step that failed. An update now streams into a spare slot, is checked before anything is erased, and is marked good only once it reaches the Pi; if it never does, the next restart puts the old version back.

I’d used the same board as a camera in April 2025, in a different project: an image-recognition device for experimenting with code and LLM calls. On 3 September the cyberdeck got a camera too. That night the key pad died whenever the camera had been used. The firmware handed pin 7 back and forth between the two at five places in the code, and by the next afternoon Claude Code had put the fault down to my wiring and stopped. Then my meter read 20 ohms between pin 7 and ground while the keys were dead, which is what a chip holding the pin low looks like, and unplugging the camera’s ribbon brought the key pad back, though the stored levels had to be learned again. So I dropped the camera. Separately, the analogue microphone, which shared the key pad’s converter, had been swapped in code earlier that afternoon for a digital one on three pins of its own, and I wired it that night.

14pins freedsourceFirmware commit 03784aa, 4 September 2026: 'fifteen pins - 2 4 6 8 9 10 11 12 13 14 15 16 17 21'. The list holds 14 pins; the commit's 'fifteen' also counts GPIO7, which stays with the key pad.