Skip to content

ADR-0009: 100% coverage on the server and on all Kotlin logic; screens held by line

Status: Accepted (2026-09-30), revised 2026-10-01

Context

The owner's standard is 100% line and branch coverage. The server and the pure-Kotlin modules meet it directly. The Android app module was first left unmeasured, on the belief that JaCoCo over Robolectric describes the harness more than the tests. Measuring it showed otherwise: the numbers are accurate, and they found real gaps. The activity and the settings store had no test at all, the SMS receiver's delivery path was never run, and Retry silently dropped a stalled page-turn request.

What JaCoCo cannot measure fairly is the branch count of Compose code. The Compose compiler wraps every composable in branches for skipping and restarting that no test steers.

Decision

  • The server is held at 100% line and branch coverage by coverage.py with fail_under = 100. No pragma: no cover exists, and a test fails if one appears.
  • The Kotlin :protocol and :core modules, which hold the codec, reassembly, the conversation state machine and the Markdown parser, are held at 100% line and branch by a JaCoCo floor that check enforces. Kotlin-generated synthetic classes are excluded.
  • In :app, everything outside the ui package (the controller, storage, settings, the SMS gateway and receiver, the resend worker, the activity) is held at 100% line and branch.
  • The Compose screens in ui are held at 99% of lines and no branch floor. The lines no test can reach are the closing braces of exhaustive whens, where Kotlin emits a throw for a case that cannot exist; they are four lines in 457.
  • The two browser scripts (the dashboard's and the site's) are tested in Node from the server suite. They are small enough that every path has a named test; no coverage tool is run over them.

Why

A floor is only honest where the tool measures the code and not its compiler. Lines of a screen are that; branches of a screen are not. Replacing the exhaustive whens with else branches would reach 100% of lines and lose the compiler's check that every screen and every kind of block is handled, which is the worse trade.

What it costs

Screen tests must reach every line, so a new control needs a test that scrolls to it and uses it. The app's tests run under Robolectric and take about three minutes.

A bug that only a real device shows can still hide behind a passing suite. The on-device acceptance run in docs/ACCEPTANCE.md exists for exactly that.

What would change our minds

A coverage tool that understands Compose's generated branches, which would let the screens carry a branch floor too.