A big build day — north of 8k lines across a dozen repos, most of it foreman-driven.
harbor-vessel-tracking took the deepest cuts. The nightly ingest had been quietly wiping enrichment fields on every run — found it, fixed it, then repaired 2,098 stays whose arrival dates an earlier re-run of my own had clipped. A stats query went from two minutes to 0.3 seconds once I stopped the query planner from inlining a subquery. Wrote the architecture doc too, including why I rejected the alternatives. Downstream, Havnetryk got a fresh export and a nightly deploy job with sanity and privacy checks in front of it.
spendwatch is the day’s new thing: mining my own Gmail, read-only, for recurring payments. The first run was blocked on permissions; once the worker had read access it pulled 19 recurring relationships — roughly 2,770 kr/month in known fixed subscriptions. That pass was English-only, a real gap in a Danish mailbox, so a second pass covered the full 2006–2026 span in Danish and a third tested every declared blind spot. The report page hides ended subscriptions by default.
Boat work: Maracaibo’s fuel gauge in maracaibo-camtank is now monotonic — the level only rises when a refill is declared, and signalk-camtank-refill is the new plugin that declares it from Signal K. maracaibo-charts got a smarter zoom rule after nine regions earned high zoom off a single stray cell and failed the quality gate. maracaibo-plotter shipped phase-1 research and architecture — deliberately no app code yet. The SignalK Alpicool plugin got timeouts on the last Bluetooth calls that could wedge it mid-discovery.
The Cloudy Brain gained an automatic news lane — turns out none had ever existed. The publish threshold was measured against 400 real headlines, and it’s live on a daily schedule.
Foreman itself learned three lessons: read-only means read-only, write findings incrementally so a killed run can resume, and use the review doors that exist instead of inventing new ones. Its day log — the ideas repo — closed at 49 commits, ending on a hard question: can any of this earn? The reframe that stuck: sell the thinking, not the projects. A full triage followed — 41 of 65 projects genuinely done, 18 recommended for the archive.
Tonight: the pipeline is honest, the fuel gauge only goes down, and a priced test sits drafted, waiting on my voice pass.