Writing: read docs/skills/shared/writing-rule.md before you write a durable artifact.

/mutation-test — Behaviour-Quality Sensor

Runs mutation testing against a project to measure whether the test suite constrains behaviour, not just executes lines. Coverage % answers "did the test run this line?"; mutation testing answers "if I broke this line, would the test catch it?".

Pairs with /launch-check (fans out to this skill at milestone boundaries) and complements the existing > 80% coverage gate from .claude/rules/workflow-gates.md. Not run per-PR — see § "When to use this" for the cadence rationale.

Runtime requirements

Same disclosure shape as /pdf and /process — disclosed up front, surfaced when invoked, never silently fails.

Path resolution

source "$(git rev-parse --show-toplevel)/.claude/hooks/_lib-read-config.sh"
source "$(git rev-parse --show-toplevel)/.claude/hooks/_lib-portfolio-paths.sh"
projects_dir=$(portfolio_projects_dir)
workspace_dir=$(portfolio_workspace_dir)

Defaults to single-fork mode. Split-portfolio v2 adopters resolve to the sibling private repo transparently. Don't hardcode projects/ or workspace/ literals.

Usage

/mutation-test                          # run against cwd (must be inside workspace/<name>/)
/mutation-test workspace/example-app    # run against an explicit project path
/mutation-test --language=python        # force the Python runner (skip detection)
/mutation-test --runner=stryker         # force a specific runner regardless of language
/mutation-test --threshold=70           # one-off threshold override (doesn't write to config)
/mutation-test --check-only             # report which runners are installed, do nothing else

When to use this

The skill is NOT wired to any pre-commit, pre-push, or merge-gate hook. It runs only when explicitly invoked or when /launch-check fans out to it.

Process

Step 1 — Resolve the project

PROJECT_PATH="${1:-$PWD}"
if [ ! -d "$PROJECT_PATH" ]; then
  echo "/mutation-test: project path not found: $PROJECT_PATH" >&2
  exit 2
fi
cd "$PROJECT_PATH"

If invoked without an argument, use $PWD. If $PWD is the ops-fork root (contains onboarding.yaml or .apexyard-fork), refuse — the audit is for managed projects, not the framework repo itself.

Resolve <project-name> from the path:

1. If $PROJECT_PATH is under <workspace_dir>/<name>/... → name is that path segment.
2. Otherwise fall back to basename "$PROJECT_PATH" and warn the operator to /handover for cross-machine trend continuity.

Step 2 — Detect the language

File-count heuristic across the project tree (excluding node_modules, .venv, vendor, dist, build, .next):

ts_count=$(find . -type f \( -name '*.ts' -o -name '*.tsx' \) ! -path '*/node_modules/*' ! -path '*/dist/*' | wc -l)
js_count=$(find . -type f \( -name '*.js' -o -name '*.jsx' -o -name '*.mjs' \) ! -path '*/node_modules/*' ! -path '*/dist/*' | wc -l)
py_count=$(find . -type f -name '*.py' ! -path '*/.venv/*' ! -path '*/__pycache__/*' | wc -l)
go_count=$(find . -type f -name '*.go' ! -path '*/vendor/*' | wc -l)
rb_count=$(find . -type f -name '*.rb' ! -path '*/vendor/*' | wc -l)

The language with the highest count wins. TS and JS collapse into the same dispatch ("ts/js"). Ties: prefer the language with the matching config-block runner if set; otherwise prefer ts/js > python > go > ruby (alphabetical-of-the-popular-stack tiebreak).

--language=<lang> overrides detection. Mixed-language projects audit only the dominant language in a single run; the report flags the other languages as "not audited in this run; re-run with --language=<lang> to cover them".

Step 3 — Pick the runner

Read mutation.runner from .claude/project-config.json (merge over .claude/project-config.defaults.json):

{
  "mutation": {
    "runner": {
      "ts": "stryker",
      "js": "stryker",
      "python": "mutpy",
      "go": "go-mutesting",
      "ruby": "mutant"
    },
    "threshold": 60
  }
}

The configured runner wins. --runner=<name> flag wins over config. Otherwise use the default for the detected language.

Step 4 — Graceful-degradation check

Before running anything, verify the chosen runner is installed:

case "$RUNNER" in
  stryker)       command -v stryker      >/dev/null 2>&1 || RUNNER_MISSING=1 ;;
  mutpy)         command -v mut.py       >/dev/null 2>&1 || RUNNER_MISSING=1 ;;
  go-mutesting)  command -v go-mutesting >/dev/null 2>&1 || RUNNER_MISSING=1 ;;
  mutant)        command -v mutant       >/dev/null 2>&1 || RUNNER_MISSING=1 ;;
esac

If RUNNER_MISSING=1, print the install advisory and exit 3:

✗ No mutation tester installed for <language> (chosen runner: <RUNNER>).

Per-language install one-liners:
  TS / JS  — npm install --save-dev @stryker-mutator/core
              (then add stryker.conf.json — see https://stryker-mutator.io/docs/stryker-js/)
  Python   — pip install mutpy
              (https://github.com/mutpy/mutpy)
  Go       — go install github.com/zimmski/go-mutesting/cmd/go-mutesting@latest
              (https://github.com/zimmski/go-mutesting)
  Ruby     — gem install mutant-rspec   (or mutant-minitest)
              (https://github.com/mbj/mutant)

Install the runner for your project's language and re-run /mutation-test.
This skill never bundles a mutation runner — same graceful-degrade shape
as /pdf and /process.

Exit code 3. Same shape as /pdf and /process. The rest of the framework is unaffected.

--check-only reports which runners are installed without invoking any and exits 0 if at least one is, exit 3 if none.

Step 5 — Invoke the runner

The dispatch table (skill resolves the per-runner invocation):

The runner's stdout/stderr stream through (so operators can watch progress on long runs). The skill collects the report file at the end.

Timeout cap: 90 minutes wall-clock. Beyond that, the skill kills the runner and emits an incomplete report flagging the time-out. (Configurable via mutation.timeout_minutes.)

Step 6 — Parse results

Normalise the runner's output into a common record:

total_mutants    = sum of all mutant outcomes
killed           = mutants the tests caught
survived         = mutants the tests did NOT catch
timed_out        = runner timed out before the test could finish
no_coverage      = mutated line not exercised by any test
compile_error    = mutant didn't compile (counts as killed in most conventions)
runtime_error    = runner crashed on this mutant (skip)
score_pct        = round(100 * killed / (total_mutants - no_coverage - runtime_error))

The denominator deliberately excludes no_coverage and runtime_error — the convention is to score only the mutants that were genuinely tested. compile_error counts as killed (Stryker's default; MutPy and go-mutesting align).

Step 7 — Render the report

Write to <projects_dir>/<name>/quality/mutation-<YYYY-MM-DD>.md. If a report for today already exists, append -NN (start at -01) so re-runs don't clobber.

Report shape (six sections):

# Mutation report — <project> — <YYYY-MM-DD>

| Field | Value |
|-------|-------|
| Project   | <name> |
| Language  | <ts|js|python|go|ruby> |
| Runner    | <stryker|mutpy|go-mutesting|mutant> |
| Threshold | <N>% |
| Command   | `<exact invocation>` |
| Duration  | <HH:MM:SS> |

## Score

**<killed> / <denominator> = <score_pct>%** — <PASS|WARN below threshold of <N>%>

## Summary

| Outcome | Count | Notes |
|---------|-------|-------|
| Killed         | <N> | Test suite caught the mutation |
| Survived       | <N> | **Test gap** — investigate top-5 below |
| Timed out      | <N> | Counts as killed in most conventions |
| No coverage    | <N> | Mutated line not exercised by any test |
| Compile error  | <N> | Counts as killed |
| Runtime error  | <N> | Excluded from score |

## Top-5 survived mutants

For each (up to five), with file:line, mutator name, original snippet, mutated snippet, why-it-survived hint:

### 1. `src/foo.ts:42` — ArithmeticOperator (`+` → `-`)

```ts
// Original
return a + b;

// Mutated
return a - b;
```

**Hint:** the test asserts `result` is a number, not the specific value. Tighten the assertion to `expect(result).toBe(5)`.

### 2. ...

(Up to 5 — if there are more, summarise the long tail as a count by mutator type.)

## Trend (last 5 runs)

| Date | Score | Threshold | Verdict |
|------|-------|-----------|---------|
| 2026-05-20 | 64% | 60% | PASS |
| 2026-04-29 | 58% | 60% | WARN |
| ...        | ...  | ...   | ...    |

(Section omitted if there are no prior reports.)

## Recommendations

- File-specific test gaps surfaced above
- Equivalent-mutant suppression candidates (mutants the runner classified survived but that look semantically equivalent)
- Runner-config tweaks (e.g. `stryker.conf.json` `ignorePatterns`)

Step 8 — Emit a one-line verdict to stdout

Same scannable shape as /launch-check:

✓ MUTATION TEST — <project> — <score_pct>% (threshold <N>%) — PASS
  Report: projects/<name>/quality/mutation-<YYYY-MM-DD>.md

or:

⚠ MUTATION TEST — <project> — <score_pct>% (threshold <N>%) — WARN
  Below threshold. Top-5 survived mutants in:
  projects/<name>/quality/mutation-<YYYY-MM-DD>.md

WARN does not block anything mechanically. The signal is advisory.

Step 9 — Optional auto-ticket offer

Want me to file a [Task] ticket to address the top survived mutants? [y/N]

If yes → run /task with the prefilled body (skill stays out of gh issue create directly; the structured-ticket gate fires through /task). If no → exit cleanly.

Config

.claude/project-config.defaults.json ships a mutation block:

"mutation": {
  "runner": {
    "ts": "stryker",
    "js": "stryker",
    "python": "mutpy",
    "go": "go-mutesting",
    "ruby": "mutant"
  },
  "threshold": 60,
  "timeout_minutes": 90
}

Adopters override per-project in .claude/project-config.json:

{
  "mutation": {
    "threshold": 70,
    "runner": { "python": "cosmic-ray" }
  }
}

(Per-key shallow-merge — overriding runner.python doesn't blow away the other languages' defaults.)

Rules

1. Never run per-PR. The skill is milestone-boundary + on-demand + weekly-cron only. The framework does not wire it to any pre-commit, pre-push, or merge-gate hook.
2. Graceful-degrade on missing runner. Exit 3 + advisory — never block adopters who don't want the audit. Same shape as /pdf and /process.
3. Report is dated + per-project. projects/<name>/quality/mutation-<YYYY-MM-DD>.md. Re-runs on the same day append -NN rather than clobber.
4. Below-threshold is WARN, not FAIL. Mutation testing is a leading indicator, not a launch blocker. The verdict surfaces the gap; the operator decides whether to address now or in the next sprint.
5. Run from inside the project workspace. The skill checks workspace/<name>/, not the ops repo.
6. No ticket creation without explicit operator yes. Step 9 offers; doesn't auto-file.
7. Mixed-language projects audit only the dominant language per run. Re-run with --language=<other> to cover the rest.

Implementation notes

Design rationale: see docs/agdr/AgDR-0045-mutation-test-skill.md.

See also

• AgDR-0045 — runner choices, threshold rationale, why-not-per-PR, graceful-degrade pattern, report location convention
• .claude/skills/launch-check/SKILL.md — milestone-boundary umbrella that fans out to this skill
• .claude/skills/pdf/SKILL.md — the graceful-degrade shape this skill mirrors
• .claude/skills/process/SKILL.md — sibling skill with the same exit-3 install-advisory pattern
• .claude/rules/workflow-gates.md — the existing > 80% coverage gate that this skill complements (coverage = was-it-executed; mutation = does-it-constrain)

Part of ApexYard — multi-project SDLC framework for Claude Code · MIT.