ExAthena.Coverage (ExAthena v0.19.0)

Copy Markdown View Source

Decide whether a run's tests actually executed the code it changed.

A green test suite proves a test exists, not that it covers the change: a test written for a private helper passes while the feature it was meant to cover is never executed at all.

Why zero, and never a threshold

Percentages lie for declarative code: a module whose body is compile-time DSL reports full coverage merely for being loaded, while a module full of real branches reports a low number for being partly tested. A threshold is therefore both unfair to the first and meaningless for the second. "Never executed at all" is the only claim a coverage table supports unambiguously.

Scope

The parser targets the NN.NN% | Module.Name table that mix test --cover prints, read out of a worker's report rather than run by the harness (tools execute commands; the harness only sees their output). Projects in other ecosystems simply produce no parseable data — reported as :no_data, which callers should treat as "ask for coverage", not "fail".

Summary

Functions

modules_in(path)

Modules a source file defines, in source order.

parse(text)

Module-name => percentage, parsed from a coverage table anywhere in text.

unexercised(changed_files, transcript, cwd)

Which of changed_files no test executed, given a worker's transcript.

Functions

modules_in(path)

@spec modules_in(String.t()) :: [String.t()]

Modules a source file defines, in source order.

parse(text)

@spec parse(String.t() | nil) :: %{required(String.t()) => float()}

Module-name => percentage, parsed from a coverage table anywhere in text.

The Total row is dropped: it is a summary, not a module.

unexercised(changed_files, transcript, cwd)

@spec unexercised([String.t()], String.t() | nil, String.t()) ::
  {:ok, [String.t()]} | :no_data

Which of changed_files no test executed, given a worker's transcript.

Returns :no_data when the transcript carries no coverage table at all — the caller cannot conclude anything, and should ask for one rather than treat silence as failure. Files defining no module (assets, config) are skipped: coverage has nothing to say about them.

A module absent from the report counts as unexercised — it was never even loaded.