Snow-White ships a Claude Code skill that teaches Claude how to read a Snow-White JUnit XML report and turn failing criteria into concrete test improvements — without you having to explain the report format every time.

The skill lives in the repository at .apm/skills/snow-white/ and is published as an APM (Agent Package Manager) package, so you can install and update it like any other dependency.

Why this matters for agentic development

When an AI coding agent writes both the implementation and the tests that exercise it, a green test suite stops being independent evidence — it can just as easily mean the agent’s tests agree with the agent’s code, not that either actually matches the API contract you specified. Snow-White sits outside that loop: it correlates real OpenTelemetry traces from the running application against the OpenAPI spec on file, so coverage results stay a ground-truth signal no agent can satisfy by construction.

The --agentic flag on snow-white calculate is the direct integration point: instead of human-readable progress logs, it prints one line of JSON with the pass/fail result and a per-criterion breakdown of what’s still undercovered — built for a coding agent (or this skill) to parse and act on directly, without scraping a report format meant for people.

What it does

Once installed, Claude automatically applies the skill whenever it encounters a Snow-White JUnit XML report, a quality gate failure, or a request like “fix the failing coverage criteria.” It:

  1. Parses the JUnit XML structure Snow-White emits (testsuites → testsuite per API → testcase per criterion).
  2. Detects correlation failures (tests="0") and walks through the onboarding checklist instead of treating it as a test gap.
  3. Looks up each failing criterion against the Quality Gate Criteria hierarchy, so it fixes the highest-level failing criterion first instead of duplicating work on children that will pass automatically once the parent does.
  4. Proposes integration tests — real HTTP calls with the OTEL agent attached, since Snow-White correlates traces, not mocks — matching your existing test style.
  5. Summarizes what was fixed, what criteria that resolves, and what’s left.

The package bundles references/criteria-hierarchy.md alongside the skill, so Claude can look up the full criteria hierarchy offline rather than fetching this site.

Installation

Install the APM CLI, then add the skill to your project’s apm.yml:

dependencies:
  apm:
    - bbortt/snow-white/.apm/skills/snow-white#v1.10.0

apm install

Pin a tag from the releases page rather than tracking main. APM packaging ships from the first release containing apm.yml; the example above is illustrative.

The /.apm/skills/snow-white path segment matters. Without it APM resolves the whole Snow-White monorepo (~18 MB) into apm_modules/ to deploy two files; with it, only the skill directory is fetched.

APM deploys the skill to the directory your harness expects — .claude/skills/snow-white/ for Claude Code, and .agents/skills/snow-white/ for the harnesses that share that location. Run apm update to pick up a newer release.

Manual copy

If you would rather not add APM, copy the two files directly:

mkdir -p .claude/skills/snow-white/references
curl -Lo .claude/skills/snow-white/SKILL.md \
  https://raw.githubusercontent.com/bbortt/snow-white/main/.apm/skills/snow-white/SKILL.md
curl -Lo .claude/skills/snow-white/references/criteria-hierarchy.md \
  https://raw.githubusercontent.com/bbortt/snow-white/main/.apm/skills/snow-white/references/criteria-hierarchy.md

Claude Code picks up skills automatically from .claude/skills/ — no further configuration needed. The next time Claude sees a Snow-White JUnit report in your conversation, it will apply the workflow above.