Documentation
¶
Overview ¶
Command coverdiff reports changed lines the tests never executed.
It exists because of a measured failure, not a metric: an entire layer of eviction handling landed across several commits with zero coverage, and a later audit proved that deleting any of it shipped green through the full validation sequence. Overall coverage percentages would not have said so — the repo's number barely moved — but "which of the lines I just changed did no test run" would have, at commit time.
Advisory by design: it prints and exits zero. Some changed lines are legitimately uncoverable here (CLI wiring, error branches that need a cloud account), so a hard gate would grow a threshold-and-exclusion system that is its own maintenance. The contract is CLAUDE.md's: run it before committing, read what it prints, and either add the test or know why not.
Usage:
go test ./... -coverprofile=.cover.out go run ./tools/coverdiff -profile .cover.out -base HEAD
-base is any git rev; HEAD means "the working tree's uncommitted changes", HEAD~1 reviews the last commit.