Development

mise install
mise run ci    # fmt/lint/test; see [CLAUDE.md](https://github.com/illumination-k/mutrim/blob/main/CLAUDE.md)

Tasks are grouped by MISE_ENV: base (formatting, GitHub Actions), golang (go test, golangci-lint, govulncheck) and bazel (gazelle, buildifier, bazel test //...). For example, MISE_ENV=bazel mise run ci:bazel runs only the Bazel checks.

mise run bench runs every benchmark and summarizes it with benchstat; given a git ref it first runs the same benchmarks at that ref in a temporary worktree and compares the two, which is how a performance change should be judged:

mise run bench origin/main                  # working tree vs main, 6 runs each
BENCH=Greedy PKGS=./minimize mise run bench # one benchmark, no comparison
go test ./mutator -run '^$' -bench . -benchpkg go/types    # an extra package to generate for

BENCH (the -bench regexp), COUNT (runs per benchmark) and PKGS narrow it; the raw results stay in .bench/.

Package	Benchmark	Measures
mutator	BenchmarkGenerate	mutant generation with the go/types pre-filter, per mutant
mut	BenchmarkSite	one schemata site: identity build, another mutant active, traced, traced from all goroutines
runner	BenchmarkRun	a whole run of the schemata fixture without its looping mutants, 1 and GOMAXPROCS jobs
minimize	BenchmarkGreedy	the set cover, from 100 tests × 1000 sites to 3000 × 30000
minimize	BenchmarkCompose	building the matrix the cover runs on, same sizes

mise run dogfood (tools/dogfood.sh) runs the schemata pipeline on mutrim’s own packages and writes report.json and minimize.json per package under .mutrim/. It takes a while: the runner tests build and execute test binaries, and every mutant reruns them. Pass package directories to limit it, for example tools/dogfood.sh criteria minimize. Each phase’s wall time (gen, build, run, minimize) lands in .mutrim/timings.tsv, so it is also the end-to-end benchmark.