The protocol is simple in a way that is not especially forgiving. Each request starts with a one-byte operation code followed by whatever data that operation needs. The reply contains only that operation’s response: no common header, transaction number, length, checksum, or error envelope. Read the wrong number of bytes once and the leftovers can look like the answer to the next request.

One documented RTCK measurement made the risk unusually concrete. On my J-Link EDU Mini V2, the request went out, no reply came back, and the probe stayed wedged after I closed it and reset the USB device. It recovered only after I unplugged and reconnected it.

The complete reconstruction is now this Trace. Ostiole uses it for the J-Link support that exists today, and there is enough protocol in it for another implementation to start without first reconstructing the same conversation from old clients and a probe on the bench.

It was a truly bizarre experience deliberately smearing my fingers with oil and then touching my phone screen. Try it! Rules are made up!

I spent longer than warranted looking at the old CD jewel cases in Matelsky’s photos, partly out of nostalgia and partly because they turn out to be the medium. Matelsky turns the highlights moving across a greasy fingerprint into curves for a pen plotter to scratch into the plastic, creating an image whose apparent depth comes into view under a point light as you move your head.

Matelsky puts a pen plotter and old CD jewel cases back to work together, a delightfully useless collision of dead technologies for the pure fun of making holograms.

Hardware-debugging tools are happy to do exactly what I asked instead of what I intended, then hand me enough hexadecimal rope to hang myself trying to work out what happened. An AI gets the same two chances. It can hallucinate the script and the result.

I want to optimize for a different target, one where AIs compose exactly the tools they need. Ostiole’s libraries should have small jobs and clear boundaries. The simulators should let an AI try them without a bench. The documentation should point it at the library that already owns an operation before it goes rummaging through the whole repository.

Eventually, I expect Ostiole to inspect, control, and program whole systems, with debugger and MCP interfaces using the same libraries instead of growing their own hardware stacks. That is a long way from reading a Cortex-M identity, but building it is the fun part—and somewhere along the way an AI will use Ostiole to debug a chip as part of doing something genuinely useful.

I write because I like the mental state created by writing. It calms and quiets me, but that is not why I write.

I read this while getting jon.dev ready to launch, so it seemed apropos. Rands has an answer. His post helped me think about mine.

The format here owes Daring Fireball an obvious debt. John credits Jason Kottke with the idea of mixing link posts and full articles in one chronological stream. He shamelessly copied it. I’m shamelessly copying him.

I don’t know exactly what this will turn into. I’m glad you’re here to find out with me.