kotlin-test-file-overwrite-guard
Add test coverage to a Kotlin/Gradle repo without silently overwriting existing tests, and keep fixtures from asserting premises a grep got wrong. Use before creating a test file, or when a new test f
SKILL.mdskills/managed/kotlin-test-file-overwrite-guard/SKILL.md
GitHubAdding tests without destroying tests
Two failure modes that both produce confident, wrong work. Observed in a Kotlin/Gradle Android repo; neither is repo-specific.
1. Never write to a path you have not confirmed is new
Symptom: a commit whose message says add coverage shows deletions in the diffstat, and the suite total rises by fewer tests than were written.
1 file changed, 76 insertions(+), 44 deletions(-) # a "new" file cannot delete
unit: 174 -> 177 # but 7 tests were added
Rules:
- Before writing, answer does this exist? with
git log --oneline -3 -- <path>, not with a heuristic or a coverage script written minutes earlier. - Path exists: read it, then edit. Write is only for genuinely new files.
- A detector with known false positives (it called an obviously-tested file untested) also has false negatives. Distrust it in both directions.
Recovery if it already happened:
git show <commit>^:<path> > .tmp/prev.kt # the overwritten version
grep -c "@Test" .tmp/prev.kt # what was destroyed
Merge rather than choose: old tests often assert something a rewrite misses
(e.g. only measured sites appear in the map - a latest.size assertion).
Verify each recovered test name is back, then sweep the session:
for c in <commits>; do git show --numstat --format="" $c | awk '$2>0 {print $3}'; done
2. Make the fixture prove its own premise
A test assuming "this name is not in the catalogue/seed/enum" will one day fail for an unrelated reason, and the failure looks exactly like a production bug.
A case-sensitive grep of one file is not proof of absence. Catalogues are often seeded from more than one place (a static list and a feature module).
Assert the premise against the real source of truth:
assertEquals("the fixture name must be unknown here", null, dao.byName(foreign))Prefer that over a "surely absent" literal: the literal works today and rots when the catalogue grows; the assertion cannot.
Name enum members explicitly (
MeasurementSite.WAIST) rather than positionally (entries.first()), which silently re-targets on reorder.
Diagnose in the assertion, not in the source
When a new test fails, put the facts in the assertion message and run once instead of reading production code for theories:
assertTrue("restored=$restoredNames stored=\"${row?.name}\"", restored.contains(foreign))
The mismatch between what the fixture wrote and what the database stored is
usually the whole answer. In the observed case the archive said Ring Muscle-Up
while the stored row read Ring Muscle-up, proving case-insensitive matching
had found a pre-existing row and the production code was correct all along.