Benchmarks

Two views are published here. The first measures shuck against shellcheck on whole real-world shell repositories — the situation a developer hits in CI on a changed-files run. The second measures the same comparison on individual fixtures, which is closer to the editor / language-server feedback loop.

Compare tools within the same snapshot. Absolute numbers across different machines are useful for rough orientation only.

On real-world repos

Each row below is a single linter invocation over every shell script in the corresponding open-source repository. Both tools see the same filtered file list (no .zsh sources, no git hook samples) and lint with all rules enabled.

Total across 9 repos: shuck 3.58 s vs shellcheck 139.51 s = 38.94x speedup over 7,671 files / 614,131 lines.

Captured May 3, 2026, 4:29 AM on Apple M5 Max local snapshot. shuck 0.0.31 vs shellcheck 0.11.0.

On a single file

The fixtures below are individual shell scripts checked into the benchmark crate. Sub-second numbers here are the relevant signal for editor integrations and pre-commit hooks: shuck completing under ~50 ms makes lint-on-keystroke practical without debouncing. Two datasets are shown — a checked-in Apple M5 Max snapshot captured from a local run, and a Linux CI snapshot regenerated during the GitHub Pages deploy for the latest published release.

Snapshot overview

Methodology

• The runner is hyperfine with 3 warmups and 10 measured runs per case.
• shuck is measured with check --no-cache so the results reflect parsing and linting work rather than cache reuse.
• The comparison command is shellcheck --enable=all --severity=style <fixture> on the same input.
• --ignore-failure is intentional. These fixtures contain lint findings, so non-zero exit codes are expected and the benchmark is measuring runtime rather than success state.
• Each fixture is benchmarked independently, and the all case benchmarks one invocation over the entire vendored corpus.

Reproducing results

Refresh the checked-in local snapshot

make bench-macro-site-local

Generate a CI-style dataset manually

./scripts/benchmarks/setup.sh hyperfine shellcheck
./scripts/benchmarks/run.sh
python3 ./scripts/benchmarks/export_website_data.py \
  --repo-root . \
  --bench-dir .cache \
  --output website/generated/benchmarks/ci-latest.json \
  --dataset-id ci-latest \
  --dataset-name "GitHub Actions latest release snapshot" \
  --dataset-description "Generated during the website deploy workflow for the latest published release." \
  --environment-kind ci \
  --environment-label "GitHub Actions ubuntu-latest"

Refresh the repo-corpus snapshot

make bench-repo-corpus

Benchmark corpus

The benchmark corpus is vendored into the repository so every run uses the same inputs. The current corpus contains 7 files, 25,515 lines, and 857.7 KiB of shell source.