version: "1.0.0" name: cve-known-vulnerabilities description: Detect known CVEs in Maven, NuGet, or npm dependencies using the GitHub Security Advisories API user-invocable: false

Local CVE Dependency Assessment

Audit dependencies without modifying application files. Do not call an assessment or CVE MCP tool.

Inputs

• workspace-path: Absolute project root.
• output-path: Required JSON result path supplied by the assessment catalog.
• scan-scope: direct (default) or all.

1. Detect Ecosystem

• Maven/Gradle project → advisory ecosystem maven.
• .sln, .slnx, or project files → nuget.
• package.json → npm.
• If none apply, write [] to output-path and stop.

2. Collect Dependencies

Prefer project wrappers and structured command output. Never edit manifests or lock files.

Maven/Gradle

• Maven: run ./mvnw dependency:list -DoutputAbsoluteArtifactId=true; add -DexcludeTransitive=true for direct scope. Fall back to mvn, then static pom.xml parsing.
• Gradle: run ./gradlew dependencies --configuration runtimeClasspath; fall back to gradle, then static build.gradle/build.gradle.kts parsing.
• Coordinate format: groupId:artifactId:version.

NuGet

• Run dotnet list package --format json; add --include-transitive for all scope.
• If no solution exists at the root, target each discovered project explicitly.
• Fall back to PackageReference, Directory.Packages.props, and packages.config parsing.
• Coordinate format: PackageName@version.

npm

• Run npm list --json; add --depth=0 for direct scope.
• Fall back to package-lock.json, yarn.lock, pnpm-lock.yaml, then package.json.
• Coordinate format: package-name@version.

For each dependency retain its manifest path and 1-based declaration line when directly declared; use line 0 for transitive dependencies.

3. Query GitHub Advisories

Batch coordinates in groups of at most 30. Use gh api when available:

/advisories?ecosystem=<maven|nuget|npm>&affects=<url-encoded-coordinates>&per_page=100

Follow pagination. If gh is unavailable, use the GitHub REST API with GITHUB_TOKEN when present. Never use general web search for vulnerability conclusions.

If one batch fails, record the error and continue remaining batches. A failed query must not silently become a clean scan: return a wrapper with status: partial and the failed batch details instead of [].

4. Normalize Results

• Ignore withdrawn advisories.
• Use cve_id, falling back to ghsa_id.
• Deduplicate the same advisory across dependency batches.
• Map severity: critical/high → mandatory; medium → optional; low/unknown → potential.
• Group all affected dependencies for one advisory into one finding.
• Include advisory URL, vulnerable range, first patched version, and manifest evidence.

Output

On a complete scan, write a flat JSON array to output-path; use [] only when every query succeeded and no vulnerability was found:

[
  {
    "id": "CVE-2024-0001",
    "name": "Advisory summary",
    "status": "FOUND",
    "category": "CVE",
    "severity": "mandatory",
    "storyPoint": 1,
    "evidence": {
      "files": ["pom.xml:42"],
      "explanation": "Advisory URL, affected coordinates, vulnerable range, and patched version"
    }
  }
]

When any advisory batch fails, write:

{
  "input_name": "CVE dependency assessment",
  "analysis_method": "GitHub Security Advisories API",
  "status": "partial",
  "result": {
    "finding": "CVE scan completed with one or more failed advisory batches",
    "confidence": "low",
    "evidence": ["<failure details>"],
    "values": ["<all successfully returned FOUND findings>"]
  }
}

Write the output exactly once using a temporary file followed by atomic rename. Return only the output path and concise counts to the parent assessment.