Today was split between two projects, and both came down to making metrics tell the truth.
On A Wine App I rebuilt the “Your cellar” doorway on the home screen — the bottle count is now the hero of the card, set as big typography with a warm, written read of the collection underneath instead of a bare number. The rest of the day was less glamorous: security bumps on the network-facing services (a multipart parser had a denial-of-service hole worth closing), and a fight with the TestFlight export script, where a newer Homebrew rsync was shadowing the system one and quietly breaking the IPA build. The honest moment came at deploy time — I’d wired the API to pull a key from a secret that, it turns out, was never actually created back when the reference was added. So I pulled that secret out of the deploy entirely until it exists. Better a working deploy than a clever one.
The Cloudy Brain was the bigger rework, and almost all of it was admitting the stats were lying. The code-stats tool had been counting generated data as work — one project showed over three million lines of “churn” that was really catalog copies, string files, and scraper caches. I widened the noise filter so the numbers reflect real code. Same disease in the milestones: they were drawn from the highest-churn days and capped at six, taking the oldest, so genuine milestones got buried under nightly automation. Now the agent reasons over full history and surfaces the latest ones first. I also routed the agents’ broker calls through an internal AI proxy instead of hitting the raw service, queued up seven posts, and regenerated milestones across every project.
A direction got pruned along the way too — a note that a judgment gate doesn’t belong in a robot’s path.
Tonight: the wine home screen looks right, the deploy is honest, and for the first time the dashboards are measuring code instead of noise.