Wanderlist

Day plans and packing lists for the trips you take again.

You keep the trips you take again in one place — Thanksgiving at your parents', a conference in New York, Lisbon in March. Each one builds its own packing list from your dates and the weather where you're going that month, and shows how many bikes are free right now at the docks nearest the places you're staying. Afterwards you say what you didn't need and what you wished you'd brought, and the next time you take that trip, it all starts already corrected.

Built for Design, Build, Ship · MPCS 51238. Assignment 3 designed the flows and the interface; Assignment 4 (Full Stack App) put accounts, a database and one external API underneath them, which is what makes the list follow you from the laptop you planned on to the phone beside the suitcase. Assignment 5 (Live Data System) added a worker that runs around the clock and a page that updates itself.

Live: https://wanderlist2.vercel.app

The live layer — Assignment 5

A trip knows where you're going and when. That already drove the packing list, through fifteen years of weather normals for your dates. Now it drives two more things.

Where you're staying. You name the places that matter on this trip — the hotel's street, the neighbourhoods you're deciding between — and they're saved on the trip, so you say it once and they're there after you sign out and on your next device. It's a list, because comparing two neighbourhoods is half of what live counts are good for.

When you need things. Every inventory row can say always, or when it's cold / hot / wet, or on weekends / long trips. Those aren't labels: buildPackingList reads them, so tagging the umbrella "when it rains" once puts it on every wet trip without you re-adding it. The vocabulary is closed because each tag has to be something the app can decide from the dates and the climate normals it already has.

Live bikes. For each of those places, the nearest docks and how many bikes and empty spaces each has right now, arriving without a refresh.

There was briefly a generated day-by-day itinerary here, and it was cut. It needed a hand-written list of places per city, so it worked for six cities and silently did nothing everywhere else, and the only keyless way to generate places anywhere returns things like "Bombing of Barcelona". Naming where you're staying works in any city, and it is a better anchor than a city centre, which is nobody's hotel.

GBFS feed  ──every 30s──▶  Edge Function  ──only changed rows──▶  Postgres  ──Realtime──▶  Page

Drawn before the worker was written, and it has held: every box is still one hop and nothing in the chain can be skipped. Two dashed paths on the chart are the parts that are easy to miss — station_follows is what decides which systems get fetched at all, and worker_runs is the only thing that lets the page tell stale from down.

Why bikeshare

Ambient data beats episodic data. Anything tied to one person's trip — their flight, their game — only moves while that trip is happening, and would be a still photograph when somebody looks at it. Bikes come and go in Chicago at 2am.

Measured before committing to it, not assumed: 2,078 Chicago stations, 72 of which changed across two calls 120 seconds apart — 3.5%. The feed advertises ttl: 60.

Every city it offers, probed first

bike_systems began as twelve rows typed into a migration, which meant coverage was bounded by my typing rather than by what exists — a trip to Toronto was told "no bikeshare here", which wasn't true.

The opposite also failed. Letting people type any city and looking it up on demand sent someone to Pittsburgh and showed them "No docked bikeshare in Pittsburgh" — false twice over, and the reason why is the interesting part below. A destination now has to be a city where both halves of the app work, and the list of those is derived rather than typed: scripts/verify-systems.mjs probes all 1,277 systems in MobilityData's registry, end to end, and stamps the survivors. 242 cities, 34 countries.

The cost is stated rather than hidden: you can no longer make a trip to a village in the Alps, and that worked before. It is chosen because a packing list that works everywhere beside a live panel that works nowhere in particular is the worse thing to hand someone.

Six gates, every one of them added because a real system failed it in a way that reached the screen:

Gate	Rejected	Found by
Publishes station_information + station_status	321	Free-floating fleets publish neither
At least 20 located stations	391	Three painted parking squares is not a system
Some station reports free docks	136	Dott Paris: 12,667 stations, dock counts on none
Some station has a bike in it right now	16	Denver's system shut in 2020; its 119 docks still report sixteen empty spaces, permanently
City name geocodes within 100 km of its own stations	40	The weather half — see below
The name is a place, not a country	8	Switzerland's system has 12,563 stations and "Switzerland" is not a destination

The weather gate is the one worth explaining. /api/climate geocodes the destination string, so the test is whether that same call lands where the stations are. It throws out registry rows that were never cities — Mecklenburg-Vorpommern, a bare DE — and, less obviously, real cities whose names resolve somewhere else entirely: Cornwall GB to Cornwall, Pennsylvania, 5,691 km away; Brest FR to Brest, Belarus. Losing Cornwall is a real cost, taken because the alternative is packing someone for the wrong continent.

以上為 README 的開頭部分。其餘內容請閱讀原文。 檢視完整 README ↗