irradiate

Fast mutation testing for Python, written in Rust. Built for CI.

Code coverage tells you which lines ran. Mutation testing tells you which lines are actually tested. irradiate makes small changes to your code — flipping < to <=, swapping + with -, replacing True with False — and checks whether your tests catch each one. If they don't, that's a gap.

Quick start

pip install irradiate

# Test only functions changed in your PR
irradiate run --diff main

That's it. irradiate finds your src/ and tests/, generates mutants for the changed code, and reports which ones survived.

Add to CI in 3 lines

- uses: nwyin/irradiate@v0
  with:
    diff: origin/main
    fail-under: "80"

This runs mutation testing on every PR, fails if the score drops below 80%, and posts inline annotations on surviving mutants.

Example output

$ irradiate run --diff main
Generating mutants...
  done in 3ms (14 mutants across 1 files)
Running stats + validation...
  done in 195ms
Running mutation testing (14 mutants, 10 workers)...

Mutation testing complete (14 mutants in 0.1s, 175 mutants/sec)
  Cache hits: 0
  Cache misses: 12
  Killed:    11
  Survived:  1
  No tests:  2
  Score:     91.7%

Survived mutants:

  number_mutation (1):
    simple_lib/__init__.py:6  replaced `0` with `1`  [simple_lib.x_add__irradiate_3]

Why irradiate

• Fast — pre-warmed pytest workers with fork-per-mutant execution. Pytest starts once. Tests run many times.
• 38+ mutation operators — arithmetic, comparison, boolean, string methods, return values, exception types, regex patterns, decorator removal, and more.
• Full decorator support — @lru_cache, @app.route, custom decorators all mutated via source-patching. decorator_removal tests whether decorators matter.
• Type-check filter — --type-checker mypy skips mutants caught by static analysis (~35% fewer test runs on typed codebases).
• Incremental — --diff main tests only functions changed since a git ref.
• Cached — content-addressed results survive rebases, branch switches, and touch. Share across CI with remote cache sync hooks.
• CI-native — --fail-under for gating, GitHub Actions annotations, JSON/HTML reports, composite action.
• Drop-in — works with any pytest project.

Install

pip install irradiate

Requires Python 3.10+ with pytest installed. See the installation guide for more options.

Usage

# Test functions changed since main (the CI use case)
irradiate run --diff main

# Run on the full codebase
irradiate run

# Sample 10% of mutants for fast feedback
irradiate run --sample 0.1

# Generate reports
irradiate run --report json   # Stryker mutation-testing-report-schema v2
irradiate run --report html   # self-contained HTML report

# Fail CI if score is below threshold
irradiate run --fail-under 80

# Explore results
irradiate results
irradiate show module.x_func__irradiate_1

Configuration

Configure via [tool.irradiate] in pyproject.toml:

[tool.irradiate]
paths_to_mutate = "src"
tests_dir = "tests"
do_not_mutate = ["**/generated/*", "**/vendor/*"]
pytest_add_cli_args = ["-x", "--tb=short"]

Source paths can also be passed as positional arguments: irradiate run src/mylib. All settings can be overridden via CLI flags. Run irradiate run --help for the full list.

Features

Mutation operators (38 categories)

Arithmetic, comparison, boolean, augmented assignment, unary, string mutation/emptying, number literals, constant replacement, lambda bodies, return values, assignments, default arguments, argument removal, method swaps, dict kwargs, exception types, match/case removal, condition negation, condition replacement, statement deletion, keyword swap, loop mutation, ternary swap, slice index removal, regex pattern mutations (11 operators: anchor removal, charclass negation, shorthand negation, quantifier removal/change, lookaround negation, alternation removal, and more).

Functions can be excluded with # pragma: no mutate.

Execution model

By default, workers fork after pytest collection. Each mutant runs in an isolated child process with no restart overhead. For projects with complex test infrastructure, --isolate runs each mutant in a fresh subprocess instead. --verify-survivors re-tests survivors in isolate mode after the main run to catch false negatives from warm-session state leakage.

Incremental mode (--diff)

Only mutate functions touched by a git diff. Uses git merge-base to compare against the divergence point, so --diff main does the right thing on feature branches.

Reporting

Terminal output groups survived mutants by operator. --report json writes Stryker mutation-testing-report-schema v2, compatible with the Stryker Dashboard. --report html generates a self-contained report using mutation-testing-elements. On GitHub Actions, irradiate auto-emits ::warning annotations on survived mutants and writes a Markdown step summary.

Caching

Content-addressable cache keyed on SHA-256 of function body, test IDs, and operator. Survives rebases, branch switches, and touch (mtime-based caches don't). Use --no-cache to force a full re-run.

Decorator support

@property, @classmethod, and @staticmethod are handled natively via a descriptor-aware trampoline. Other decorated functions are currently skipped; a source-patching fallback is planned (#13).

Sampling (--sample)

Test a random subset of mutants for fast CI feedback. Academic research shows 5-10% random sampling gives 99% R² correlation with the full mutation score.

• --sample 0.1: test 10% of mutants
• --sample 100: test exactly 100 mutants
• --sample-seed 42: override RNG seed (default: 0 for reproducibility)

Sampling is operator-stratified, so every mutation category is proportionally represented.

CI integration

The GitHub Actions composite action (shown above) auto-detects the CI environment, emits inline ::warning annotations on survived mutants, and writes a Markdown step summary. See the CI integration guide for advanced configuration, caching, and non-GitHub setups.

Performance tuning

Parallelism defaults to CPU count (--workers N to override). Workers are recycled when RSS exceeds --max-worker-memory N MB. --covered-only skips mutants with no test coverage. --no-stats skips coverage collection when you want to test all mutants against all tests. Per-mutant timeout defaults to 10x baseline (--timeout-multiplier N).

How it compares to mutmut

Guides

• Quick start — install, run, interpret results
• CI integration — GitHub Actions, caching, gating
• Understanding results — what mutation scores mean
• Surviving mutants — what to do when mutants survive
• Performance tuning — workers, sampling, --covered-only
• Configuration — pyproject.toml reference
• Comparison with mutmut — detailed feature comparison

Acknowledgments

irradiate's trampoline architecture and mutation operator design are informed by mutmut. The naming convention is partially compatible with mutmut to ease migration.

License

MIT