Ollie — week of 2026-08-03

A heavy week, mostly reliability.

Startup ordering was wrong: the server waited on pipeline recovery before it would answer anything, so a restart with a backlog looked like an outage. It now serves first and recovers behind that, with a real HEALTHCHECK so the container’s state is honest. Two more crashes went away — MCP sessions were expiring after five minutes, and a panicking worker could take the whole pipeline down with it. Cancelling or completing one trip could also strand its load in in_transit while a sibling trip was still open; that’s fixed, as is previewing text and markdown blobs in Fleet.

On the feature side, loads can now exist administratively with no trip attached, and list_loads passes filters straight through instead of quietly dropping them. Trips gained TONU and diversion as first-class outcomes rather than something you approximate with a cancel. LanceDB datasets compact on a schedule and on demand, which keeps query latency from drifting upward.

I also spent a day halving the build-test cycle: consolidated tests, a leaner dev profile, trimmed dependencies, and a few hot-path fixes. Tags v2.7.0 and v2.8.0.