compose-unbounded-list-audit

Find and fix Jetpack Compose screens that render an unbounded, growing list eagerly (forEach inside a scrolling Column instead of LazyColumn), and prove the LazyColumn conversion by content rather tha

SKILL.md

GitHub

Auditing Compose screens for eager unbounded lists

A fast data layer proves nothing about the UI. observeHistory() returning in 76ms at 1,000 sessions still freezes the screen if the composable does history.forEach { Row(...) } inside a verticalScroll(...) Column: every row is composed whether or not it is on screen.

1. Find the eager lists

Lazy containers are the safe ones; the risk is forEach over data that grows without bound (history, sessions, entries, attempts, measurements, feed rows).

for p in pathlib.Path("app/src/main/kotlin/.../ui").rglob("*.kt"):
    src = p.read_text(encoding="utf-8")
    lazy = len(re.findall(r"LazyColumn|LazyVerticalGrid|LazyRow", src))
    eager = [m.group(1) for m in re.finditer(r"(\w+)\.forEach(?:Indexed)?\s*\{", src)]
    growing = [e for e in eager if any(k in e.lower() for k in
              ("history","sessions","entries","attempts","measurements","feed","rows","stats"))]

Then confirm per hit whether the data is capped upstream (observeRecent(limit), .take(n)) — many are, and those need no change. Only an uncapped source inside a non-lazy container is a defect.

2. Pick the right fix per screen

  • Screen must show everything (a full log) → convert to LazyColumn with stable keys. Group headers become their own item(key = ...); rows become items(list, key = { it.id }).
  • Screen is a summary (a dashboard's recent activity) → cap it with .take(n) and rely on the "see all" entry point already beside it. Cheaper and keeps the screen a simple Column.

3. Prove the conversion by CONTENT

A LazyColumn refactor compiles happily while emitting nothing — the failure mode is silent, so the mutation must be too:

// seed through the app's own live repository, then drive the real screen
val label = db.sessionDao().byId(sessionId)!!.label
assertTrue("the log did not render the session labelled \"$label\"",
    compose.onAllNodesWithText(label, substring = true).fetchSemanticsNodes().isNotEmpty())

Mutate by replacing the collection in items(...) with emptyList() and confirm the test names the missing row. Asserting a generic marker is weaker AND error-prone: a row showing XP via a pill renders "+120" and "XP" as separate text nodes, so a " XP" substring never matches. Assert the seeded row's own label — it is both stronger and unambiguous.

Traps

  • The instrumented gate uninstalls the app, so a device sweep run straight after one drives an empty launcher. Reinstall first.
  • Driving the UI to create fixture data fights whatever state the device is in (a rest day, a half-finished session). Seeding through the repository in the test is deterministic; driving taps is not.