AgentEval: Instructions file for GitHub Copilot

AgentEval Test Guidelines

Test Naming Convention

Use: MethodName_StateUnderTest_ExpectedBehavior

[Fact]
public async Task HaveCalledTool_WhenToolWasCalled_ShouldPass()

Test Structure

• Tests mirror src/ folder structure in tests/AgentEval.Tests/
• Use xUnit with [Fact] and [Theory] attributes
• Each public class/method should have corresponding tests

Using FakeChatClient

For metrics that call LLMs, use FakeChatClient to avoid API calls:

var fakeClient = new FakeChatClient("""{"score": 95, "explanation": "Good"}""");
var metric = new FaithfulnessMetric(fakeClient);
var result = await metric.EvaluateAsync(context);

Creating ToolUsageReport for Tests

var report = new ToolUsageReport(new List<ToolCallRecord>
{
    new() { Name = "SearchTool", CallId = "call-1", Result = "found" },
    new() { Name = "ProcessTool", CallId = "call-2", Result = "done" }
});

Creating PerformanceMetrics for Tests

var metrics = new PerformanceMetrics
{
    TotalDuration = TimeSpan.FromSeconds(2.5),
    TimeToFirstToken = TimeSpan.FromMilliseconds(250),
    PromptTokens = 100,
    CompletionTokens = 50,
    ModelUsed = "gpt-4o"
};

Assertion Tests Pattern

When testing assertions, expect ToolAssertionException or PerformanceAssertionException:

[Fact]
public void HaveCalledTool_WhenToolNotCalled_ShouldThrow()
{
    var report = new ToolUsageReport([]);
    
    var ex = Assert.Throws<ToolAssertionException>(() =>
        report.Should().HaveCalledTool("MissingTool"));
    
    Assert.Contains("MissingTool", ex.Message);
    Assert.NotNull(ex.Expected);
    Assert.NotNull(ex.Actual);
}

Multi-Target Framework Testing

Tests run on net8.0, net9.0, net10.0 - ensure compatibility. Use #if NET8_0 sparingly.