Show sync activity per user (#7)

[beygorghor]

Closes #7.

What changes

A Users entry in the left nav opens a page listing every user in the directory with, over a
switchable window of 7 days (the default), 30 days or 3 months: their number of syncs, total
submissions, distinct devices and last sync. Every column header sorts, and clicking it again
reverses. A user with no activity in the window keeps their row — zeros, an em-dash, dimmed —
because a supervisor looking for who went quiet needs to see them.

How it works

One SQL aggregate per window returns one row per user (41 of them); the browser re-sorts that
array in place, so only a change of window goes back to the server.

Two details in src/features/user-activity/api/queries.ts carry the whole query:

• The window predicate sits in the on clause of a left join, not in a where. In a where it
  would drop exactly the inactive users the ticket is about. This is Django's FilteredRelation.
• Every aggregate is cast ::int. Postgres returns count() as bigint, and pg hands bigint
  back as a string — measured: "6066" against 6066. Without the cast the table sorts "9"
  above "10" and nothing errors.

The window travels as an enum (7d / 30d / 90d); the server turns it into a cutoff date. The
query takes that date as a parameter rather than calling now(), so the tests pin the window
instead of racing the clock.

No migration: every column already exists.

New ADR

0015 — Cast aggregates to int in the SQL. It records the cast, rejects the alternative (a
global pg.types.setTypeParser), and states the uncomfortable part: PGlite returns an uncast
count() as a number, so no test can catch a missing cast. Review is the only check. ADR 0003
anticipated PGlite lacking something Neon has; this is the mirror image, where the test stays green.

What was checked

• 57 tests pass, up from 44 (pnpm test), pnpm exec tsc --noEmit clean.
• The query tests run on in-process Postgres and cover: a user who never synced, a user whose syncs
  all predate the window, repeat syncs from one device counted once, alphabetical order.
• The table tests run in jsdom on props: default order, sorting and reversing, the em-dash, quiet
  users last in both directions, the empty state.
• Against the real database: 41 users returned, 342 syncs over 7 days against 6,066 over 90.

One round of review found two real defects, both fixed here:

• The sort test's fixture made syncCount order and alphabetical order identical, so a table that
  never left the username column passed it. New fixture (10 and 9) disagrees with the alphabet
  and with text ordering; mutating the comparator now fails it.
• Switching the window unmounted the table and silently reset the sort. Fixed with
  placeholderData: keepPreviousData.

Worth knowing

• The seed has no quiet user. Every one of the 41 users has synced within 7 days, so the dimmed
  row will not appear in the preview. ADR 0006 makes districts go silent, not users. The behaviour
  is covered by tests; it is just not demonstrable on seeded data.
• Follow-ups deliberately not done here: asserting ::int in the compiled SQL (needs the builder
  exposed separately from the function that runs it), and moving relativeDays out of both table
  components into src/ui/ (it would edit device-syncs while other pairs are in it).

Spec: docs/superpowers/specs/2026-09-18-user-activity-design.md ·
Plan: docs/superpowers/plans/2026-09-18-user-activity.md

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
wee-app	Ready	Preview	Sep 18, 2026 8:56am UTC