Today was mostly about a supervision panel that kept changing its mind about what it was.
I started by turning a fleet of small internal tools into proper always-on services — a control panel that used to run as a bare process now lives in Docker, and I wired up a timer so an internal AI planning loop fires every 30 minutes and lets ideas mature on their own while I’m away. Judging by the churn, I spent a lot of the day killing per-task approval prompts and stray scripts so this runs unattended and clean.
The panel itself got rebuilt over and over. I grouped its sessions by idea, shipped a whole design system after picking one of two directions I’d mocked up, then did dashboard research and reframed the entire thing: it’s not a board, it’s a supervision console where autonomous agents do the work and I only handle the exceptions. That reframe threw out a stack of mockups and half-built templates the same day I made them. By tonight it had collapsed down to a single home page with a priority order. Roughly +4.8k lines in and −2k back out — a lot of building and un-building.
Meanwhile the boat’s weather stack got real attention. My signalk-dmi — actually, the DMI weather bridge learned to sample the surrounding area when the boat sits in a grid gap, then I chased a nasty bug where masked cells returned all-nulls and quietly rate-limited the source. The maracaibo-dashboard got a proper charting engine (offline uPlot, touch crosshair) shared across every graph, plus a pile of “text bleeding behind cards” layout fixes.
I also pushed the s57mvt_pi chart renderer onto the GPU — crisp soundings and buoys at any zoom — and wrote clean-room install docs for it and the weather bridge so strangers can build them. And The Cloudy Brain got a v2 approval inbox running beside v1, shadow-writing only, no risk to the live pipeline. An idea got pruned along the way.
Tonight: the console is one page, the agents run themselves, and the boat charts are sharp.