Mutation Adequacy

Mutation Adequacy Skill

Overview

The mutation-adequacy review dimension is the adequacy backstop for the relaxed verification mix (verification ladder slice 3, R5). When the planner ships the cheaper testing strategy — strict types + inline invariants + one PBT + one acceptance test instead of granular per-behavior red-green — mutation adequacy is the machine guard against the vacuous tests (and vacuous PBT properties) that omission could otherwise admit. It scores whether the test suite actually kills injected mutants: the strongest signal that tests fail for the right reason.

This dimension gates the HIGH risk tier only, at the /review boundary only — it is the review-phase counterpart to the delegation-time check_test_adequacy (R3) gate. The coupling to the high tier is policy data in review-contract.ts (the dimension name = this skill's folder name); you never decide tier membership in this skill. The gate is registered as a required D1 review dimension (adaptLadderGate(MUTATION_GATE_NAME, 'D1', handleMutationAdequacy)) — a structural participation fact, distinct from severity (below).

Two orthogonal axes, two default severities. The mutation score stays advisory by default: a sub-threshold mutationScore surfaces survivor follow-ups and warns, but does not block, unless an explicit review.gates['mutation-adequacy'] config override raises its severity. A 100% score is neither expected nor required (equivalent mutants exist).

The NoCoverage axis is a second, independent check DR-6 added, with its own real, functioning enforcement path: under review.mutationEnforcement: block, the block-mode guard (workflow/guards.ts Check 4b) fails the review → synthesize transition whenever the diff-scoped noCoverage count exceeds the configured maxNoCoverage budget (default 0 — zero uncovered changed lines tolerated), independent of the score. This is live code, not a deferred plan — but it is opt-in, the same way the score's severity is: review.mutationEnforcement defaults to advisory, and mutation-adequacy's per-gate severity default is deliberately advisory too (so a project's dimension-level defaults never silently re-block a demoted gate). A project flips NoCoverage enforcement on with one config line (review.mutation-enforcement: block); honor whichever way it's resolved, never hardcode "advisory" when recording this dimension's result.

The CI side is separate and narrower: a diff-scoped mutation gate runs on pull-request events, but it runs in observe mode unconditionally — it logs its verdict and always exits 0, regardless of any .exarchos.yml setting. It is not yet a blocking CI backstop; the flip is deferred, pending a clean runner verdict on the full server suite. Never represent the CI gate as a merge blocker — where NoCoverage enforcement is live, it is on the review path above, not in CI.

Note: advisory (where it applies) refers to whether a sub-threshold score or an over-budget NoCoverage count actually stops the transition. The dimension entry is required on the high tier regardless of either axis's resolved severity (the all-reviews-passed guard fails if the mutation-adequacy review is absent), so run and record it every time.

Triggers

Activate this skill when:

• review reaches the quality stage on a HIGH-tier feature
• The review contract lists mutation-adequacy in the required reviews for this workflow
• You need to assess whether the (possibly relaxed) test mix actually kills mutants

Do not activate for medium/low-tier work, or outside the /review boundary — those paths do not require this dimension.

Execution

Step 1: Run the diff-scoped mutation gate

Invoke the action against the review/PR base ref. It runs the resolved mutation command diff-scoped (Stryker --since, cargo-mutants --in-diff, mutmut path restriction — resolved from the toolchains SoT, never composed by hand), so the run completes in < minutes, not the full-tree time budget.

exarchos:exarchos_orchestrate({
  action: "mutation-adequacy",
  featureId: "<featureId>",
  base: "<review/PR base ref>",      // e.g. "main" — reuse the same base the review diff uses
  worktreePath: "<optional worktree>", // 'auto' resolves the calling delegation's worktree
  operationId: "<optional idempotency key>"
  // scope defaults to "diff"; do NOT pass scope:"full" here (see Anti-Patterns)
})

The action emits mutation.executing_started / mutation.executed (liveness, INV-10) and a foldable gate.executed carrying mutationScore (INV-1) automatically. Do not hand-emit these events.

Step 2: Read the carrier

The action returns the fixed carrier (data). The shape is stable regardless of pass/fail/degrade:

Degrade signals to recognize (each returns passed: true so the gate never blocks closed-with-error):

• skipped: true + reason — no mutation runner resolved. Report the remediation; do not treat as a failure.
• warning + a warnings[] entry — the runner produced no parseable report. Note it; do not throw.
• deferred: true + scope: 'full' — you (incorrectly) requested full scope; re-run with the default diff scope.

See references/reading-the-carrier.md for the full carrier and report shape.

Step 3: Turn survivors into kill-this-mutant follow-ups (INV-12)

The action already maps each surviving and NoCoverage mutant to a next_actions entry of the form "write a test that kills <file>:<line>". Surface these as concrete, actionable review findings — each one names a specific assertion the suite is missing:

• A survived mutant means a test exercises that line but asserts nothing strong enough to detect the mutation. The follow-up: add an assertion that distinguishes the mutated behavior.
• A NoCoverage mutant means no test touches that line at all. The follow-up: add a test that exercises it, then assert on the observable behavior.

Record these as issues with category: "test-quality" so they ride the same review-report contract as the other dimensions. Do not invent generic "improve coverage" advice — quote the file:line the action surfaced.

Step 4: Apply the verdict — two axes, each with its own resolved severity

The score axis stays advisory (warning) by default. A sub-threshold mutationScore:

• surfaces the survivor follow-ups (Step 3),
• warns in the review report,
• does not block the review → synthesize transition — unless an explicit review.gates['mutation-adequacy'] config override raises its severity. Honor the resolved severity, do not hardcode it. See references/advisory-threshold.md for why the default is a soft threshold and how to calibrate it from the score trend.

The NoCoverage axis is independent, with its own real block-mode enforcement path: under review.mutationEnforcement: block, if the folded noCoverage count exceeds the configured maxNoCoverage budget (default 0), the block-mode guard fails the transition regardless of the score. Like the score, it is opt-in (review.mutationEnforcement defaults to advisory) — resolve and honor the actual configured mode, don't assume either axis is on or off. Record both axes honestly in the review report — do not fold a blocking NoCoverage failure into advisory-score language, and do not describe either axis as enforced by CI (the diff-scoped CI wiring runs in observe mode; see Anti-Patterns).

Required Output Format

Record the dimension result on the review state. The review key MUST be the kebab-case dimension name (it equals this skill's folder name):

exarchos:exarchos_workflow({ action: "update", featureId: "<id>", updates: {
  reviews: { "mutation-adequacy": {
    status: "pass",            // "pass" is honest when both axes are advisory (the default); if either
                                // axis's severity is configured to block, report the real result instead
    summary: "mutationScore 0.62 (threshold 0.40); 3 survivors surfaced as kill-test follow-ups",
    issues: [
      { severity: "MEDIUM", category: "test-quality", file: "src/foo.ts", line: 42,
        description: "surviving mutant — no assertion distinguishes the mutated branch",
        required_fix: "write a test that kills src/foo.ts:42" }
    ]
  } }
}})

A passing-value status (pass | passed | approved | fixes-applied, case-insensitive) is required for the all-reviews-passed guard — a flat string is silently ignored and blocks the transition.

Anti-Patterns

References

• references/reading-the-carrier.md — reading the carrier and the Stryker mutation-testing-report-schema.
• references/advisory-threshold.md — why the threshold is a soft default, and how to calibrate it.