Corvus Build And Test

Building and Testing Corvus.JsonSchema

Solution Files

All test projects use MSTest (MSTest.Sdk 4.2.2) with Microsoft Testing Platform (MTP). The global.json pins the MSTest.Sdk version and configures MTP as the test runner.

All test projects target both net10.0 and net481. CodeGenerator.Tests produces an empty assembly on net481 (the CLI tool is .NET 10 only); <IsTestProject>false</IsTestProject> on net481 plus --ignore-exit-code 8 in CI prevents it from failing the test run. Three source-generator test projects (JMESPath, Jsonata, JsonLogic) exclude their SourceGeneratorDiagnosticTests.cs file on net481 because those Roslyn-hosted tests require .NET Core reference assemblies (integration tests run on both TFMs). Running dotnet test --solution Corvus.Text.Json.slnx without -f runs tests on all applicable TFMs.

V4 projects live in src-v4/ and tests-v4/ and are included in the solution.

Build Commands

# Full build
dotnet build Corvus.Text.Json.slnx

# Build a specific project
dotnet build src\Corvus.Text.Json\Corvus.Text.Json.csproj

TreatWarningsAsErrors=true is set across all projects — any warning fails the build.

Pre-Commit Checks

Before every commit, verify these mandatory gates:

1. Warning-free build: dotnet build Corvus.Text.Json.slnx must report 0 Warning(s).
2. Code sample catalog: if any file under .github/, docs/, or skill/instruction files was modified (even incidentally), update and verify:

.\docs\update-code-sample-catalog.ps1 -UpdateFile <relative-path>   # for each changed file
.\docs\update-code-sample-catalog.ps1 -Check                         # must exit 0

CI runs the catalog check and fails the build if it is stale. This is the most commonly missed pre-commit gate.

Running Tests

Mandatory Filters

ALWAYS exclude failing and outerloop categories:

⚠️ Use Corvus types, not System.Text.Json

Test code and assertions must use Corvus.Text.Json types (JsonElement, JsonValueKind, ParsedJsonDocument<T>, etc.), not System.Text.Json equivalents. Do not add using System.Text.Json; to test files. The two namespaces share type names (JsonElement, Utf8JsonWriter, JsonWriterOptions) which cause ambiguity errors, and using STJ types in assertions would test the wrong library.

System.Text.Json is acceptable only for test data infrastructure — e.g., reading JSON fixture files with System.Text.Json.JsonDocument to enumerate test cases. In those cases, fully-qualify the STJ types (e.g., System.Text.Json.JsonElement).

⚠️ Prefer exact assertions

Always use Assert.AreEqual with the complete expected value. Do not use StringAssert.Contains, Assert.IsTrue(x.StartsWith(...)), or similar weak assertions for output verification — these mask bugs where the format changes but still contains the checked substring.

Acceptable exceptions:

• Error/exception message substring checks (e.g., StringAssert.Contains(ex.Message, "T0410"))
• Buffer-growth tests writing 15+ iterations in a loop (output is hundreds of bytes; Contains verifying data integrity is fine)
• Non-deterministic output that cannot be reproduced exactly

Technique for capturing exact values:

1. Write a temporary Assert.Fail(actualValue) or Console.WriteLine to capture the exact output
2. Or use a file-based app (.cs script) referencing the library project to call the API directly
3. Use raw string literals (""") for JSON expected values — \u002B, \n, \t are literal characters matching JSON content

// GOOD — exact assertion with raw string literal
Assert.AreEqual("""{"name":"Alice","age":30}""", json);

// BAD — weak assertion that passes even if output is wrong
StringAssert.Contains(json, "Alice");

# Run all tests (standard — all 21 test projects, both TFMs)
dotnet test --solution Corvus.Text.Json.slnx --filter "TestCategory!=failing&TestCategory!=outerloop"

# Run all tests on a specific TFM
dotnet test --solution Corvus.Text.Json.slnx -f net10.0 --filter "TestCategory!=failing&TestCategory!=outerloop"

# Run a single test class
dotnet test --solution Corvus.Text.Json.slnx --filter "FullyQualifiedName~ParsedJsonDocumentTests&TestCategory!=failing&TestCategory!=outerloop"

# Run a single test method
dotnet test --solution Corvus.Text.Json.slnx --filter "FullyQualifiedName~ParseValidUtf8BOM&TestCategory!=failing&TestCategory!=outerloop"

Test by Feature Area

# JSON Schema draft-specific tests (all in Corvus.Text.Json.Tests)
dotnet test --project tests\Corvus.Text.Json.Tests --filter "JsonSchemaTestSuite=Draft202012&TestCategory!=failing&TestCategory!=outerloop"

# Standalone evaluator tests (all in Corvus.Text.Json.Tests)
dotnet test --project tests\Corvus.Text.Json.Tests --filter "TestCategory~StandaloneEvaluatorTestSuite&TestCategory!=failing&TestCategory!=outerloop"

# Annotation tests (all in Corvus.Text.Json.Tests)
dotnet test --project tests\Corvus.Text.Json.Tests --filter "TestCategory~AnnotationTestSuite&TestCategory!=failing&TestCategory!=outerloop"

# JSONata conformance
dotnet test --project tests\Corvus.Text.Json.Jsonata.Tests --filter "TestCategory!=failing&TestCategory!=outerloop"

# JMESPath conformance
dotnet test --project tests\Corvus.Text.Json.JMESPath.Tests --filter "TestCategory!=failing&TestCategory!=outerloop"

# YAML conformance
dotnet test --project tests\Corvus.Text.Json.Yaml.Tests --filter "TestCategory!=failing&TestCategory!=outerloop"

# JSONPath conformance
dotnet test --project tests\Corvus.Text.Json.JsonPath.Tests --filter "TestCategory!=failing&TestCategory!=outerloop"

# JSONPath code-gen
dotnet test --project tests\Corvus.Text.Json.JsonPath.CodeGeneration.Tests --filter "TestCategory!=failing&TestCategory!=outerloop"

Key Test Projects (21 runnable)

Plus 6 supporting model/utility projects that generate types consumed by other tests.

Target Frameworks

• Libraries: net9.0;net10.0;netstandard2.0;netstandard2.1
• Tests: net10.0;net481 (all projects). CodeGenerator.Tests is an empty assembly on net481.
• Run a specific TFM: dotnet test -f net10.0 ...

Collecting Code Coverage

Use dotnet-coverage (Microsoft Code Coverage), not Coverlet. Coverlet 10.0.0 has a known instrumentation bug that reports 0% for many types despite tests exercising the code.

Full test suite coverage (all TFMs, merged automatically)

⚠️ CRITICAL: Always collect baseline/full coverage WITHOUT -f

NEVER pass -f net10.0 when collecting baseline or full coverage. This misses all #if !NET / #if NETSTANDARD2_0 code paths (unsafe pointer fallbacks, polyfills, etc.) and produces incomplete results. Omit -f entirely — both TFMs run and dotnet-coverage merges them automatically. Only use -f for targeted single-test-class verification during iterative improvement.

# 1. Build once
dotnet build Corvus.Text.Json.slnx

# 2. Collect coverage — all TFMs (dotnet-coverage merges automatically)
dotnet-coverage collect `
    --output TestResults\coverage.cobertura.xml `
    --output-format cobertura `
    -s dotnet-coverage.settings.xml `
    "dotnet test --solution Corvus.Text.Json.slnx --filter `"TestCategory!=failing&TestCategory!=outerloop`" --no-build"

All test projects target both net10.0 and net481. Running without -f executes tests on both TFMs, and dotnet-coverage produces a single merged Cobertura XML. This captures TFM-conditional code paths (e.g., #if NETSTANDARD2_0 polyfill branches, net481 fallback code) that a single-TFM run would miss.

Full suite coverage runs ~150K tests across 21 test projects × 2 TFMs and takes 30–45 minutes.

Single-TFM coverage (when needed)

For targeted debugging, you can collect coverage for a single TFM:

dotnet-coverage collect `
    --output TestResults\coverage-net10.0.cobertura.xml `
    --output-format cobertura `
    -s dotnet-coverage.settings.xml `
    "dotnet test --solution Corvus.Text.Json.slnx -f net10.0 --filter `"TestCategory!=failing&TestCategory!=outerloop`" --no-build"

Note: single-TFM runs miss TFM-conditional branches. Use the all-TFM approach above for accurate coverage baselines.

Single test class coverage

dotnet-coverage collect `
    --output TestResults\mytest.cobertura.xml `
    --output-format cobertura `
    -s dotnet-coverage.settings.xml `
    "dotnet test --solution Corvus.Text.Json.slnx -f net10.0 --filter `"FullyQualifiedName~MyTestClass&TestCategory!=failing&TestCategory!=outerloop`" --no-build"

Key points

• The dotnet-coverage.settings.xml in the repo root filters coverage to published library assemblies only (18 assemblies including Corvus.Numerics)
• Output is a single Cobertura XML — running without -f automatically merges both TFMs
• Always build before collecting: dotnet build Corvus.Text.Json.slnx first, then --no-build in the test command
• When comparing before/after, always use the same approach (preferably all-TFM)
• The Cobertura XML uses full Windows paths in filename attributes — use os.path.basename() or equivalent when parsing

⚠️ Do NOT use Coverlet

--collect:"XPlat Code Coverage" (Coverlet) reports 0% coverage for many types including ref structs, static classes, and even regular sealed classes. This was verified by running the same tests with both tools — dotnet-coverage correctly reported 65–92% coverage for types that Coverlet reported as 0%.

Coverage settings file

The dotnet-coverage.settings.xml file controls which assemblies are instrumented and which source files are excluded. It includes all published V5 library assemblies plus V4 code generation assemblies. If you add a new published assembly, add a corresponding <ModulePath> entry.

Source exclusions configured in the settings file:

• src-v4/Corvus.Json.ExtendedTypes/Corvus.Json/GeneratedCoreTypes/ — V4 CLI-generated core types (~144 files) that inflate the denominator without meaningful coverage value
• *.g.cs files under obj/ — Roslyn source-generator output (regex generators, JSON schema generators, etc.)
• SR.cs and *.Designer.cs — auto-generated resource string files that are not meaningfully testable

When parsing Cobertura XML manually, apply the same exclusions: skip <class> entries whose filename contains GeneratedCoreTypes, has an obj directory segment, ends with SR.cs, or ends with .Designer.cs. Failure to exclude these will significantly undercount coverage for packages with resource files (e.g., Validator has 130 untestable lines in SR.cs + Strings.Designer.cs).

Parsing Cobertura XML

The Cobertura XML has <class> elements inside <package> elements. Each <class> has a filename attribute and <line> children with number, hits, and optional condition-coverage attributes.

Important: Partial classes and compiler-generated closures (<>c) appear as separate <class> entries for the same file. When computing per-file coverage, aggregate across all <class> entries that share the same filename:

import xml.etree.ElementTree as ET, os

def get_coverage_by_file(xmlfile):
    tree = ET.parse(xmlfile)
    root = tree.getroot()
    results = {}
    for cls in root.iter('class'):
        fn = cls.get('filename', '')
        basename = os.path.basename(fn)
        if basename not in results:
            results[basename] = {'covered': set(), 'total': set()}
        for l in cls.findall('.//line'):
            num = int(l.get('number', 0))
            results[basename]['total'].add(num)
            if int(l.get('hits', 0)) > 0:
                results[basename]['covered'].add(num)
    return results

Use sets (not counts) to avoid double-counting lines that appear in multiple <class> entries.