| name | test-first-development |
| description | Mandatory test-first workflow for changes under `src/` in ClrProfiler. Covers CLR EventListener parsing and correlation, bounded-channel delivery, timer sampling, profiler and tracker lifecycle, callback error handling, statistics values, Datadog or logger metric projection, multi-target builds, and regression verification with TUnit. |
Test-First Development
Use this skill for every task that adds or modifies code under src/. Skip it only for changes limited to documentation, configuration, generated artifacts, or the .agents directory.
Also apply performance-requirements when changing per-event, per-sample, or per-metric paths.
Red-green workflow
- Add the smallest test that demonstrates the missing or incorrect observable behavior.
- Run that test against the current production code and confirm it fails for the expected reason. A compile failure is acceptable for a new API.
- Implement the minimum production change that makes the test pass.
- Re-run the targeted test, then its containing test project.
- Run the full Release build and test suite used by CI.
- For an observable API or usage change, update
README.md in the same change.
For a behavior-preserving performance refactor, first prove the relevant tests pass and capture a repeatable allocation or throughput baseline. Treat that baseline as red, then require behavioral parity after the change.
Select the test project
- Core listeners, timers, statistics,
ProfilerTracker, event integrity, and lifecycle: tests/ClrProfiler.UnitTest/ClrProfiler.UnitTest.csproj
- Datadog adapter, logger adapter, tag and metric mapping, and
ClrTracker: tests/CleProfiler.DatadogTracing.UnitTest/CleProfiler.DatadogTracing.UnitTest.csproj
The CleProfiler.DatadogTracing.UnitTest directory contains an existing CleProfiler spelling. Use the path as it exists; do not rename it as an incidental cleanup.
Write TUnit tests
This repository uses TUnit on Microsoft Testing Platform. It is not xUnit or NUnit, and VSTest options do not apply.
- Mark tests with
[Test], not [Fact].
- Supply cases with
[Arguments(...)], not [Theory]/[InlineData].
- Assertions are fluent and must be awaited:
await Assert.That(actual).IsEqualTo(expected). Common forms include .IsTrue(), .Contains(...), .DoesNotContain(...), .HasSingleItem(), and .Count().IsEqualTo(n). HasSingleItem() returns the item, so var result = await Assert.That(list).HasSingleItem(); is the idiomatic single-result assertion.
- Serialize tests that touch process-wide state with
[NotInParallel] on the class.
- Inside a test,
TestContext.Current!.Execution.CancellationToken is the ambient token for WaitAsync and similar bounded waits.
See tests/ClrProfiler.UnitTest/EventListenerDataIntegrityTest.cs and tests/CleProfiler.DatadogTracing.UnitTest/MetricTagProjectionTest.cs for the established style.
Run tests
Filter with --treenode-filter using a /assembly/namespace/class/method path. VSTest's --filter is rejected outright by the test executable (Unknown option '--filter'), and passing it through dotnet test silently selects zero tests and still exits non-zero, which reads like a pass if the summary is not checked. --logger "console;verbosity=normal" fails the same way.
dotnet test tests/ClrProfiler.UnitTest/ClrProfiler.UnitTest.csproj -c Release --treenode-filter "/*/*/ProfilerLifecycleTest/*"
dotnet test tests/CleProfiler.DatadogTracing.UnitTest/CleProfiler.DatadogTracing.UnitTest.csproj -c Release --treenode-filter "/*/*/DatadogTracingUnitTest/*"
Wildcards apply per path segment, so /*/*/*AllocationTest/* selects every matching class. A comma-separated list and (A|B) alternation are not supported; use a wildcard that covers both classes, or run them separately.
Always read the summary counts, not just the exit code. A filter that matches nothing reports total: 0.
Finish with the sequence .github/workflows/build.yaml runs:
dotnet build -c Release
dotnet test -c Release --no-build
dotnet pack -c Release
When the net9.0 runtime is missing locally
The test projects target net9.0 only. dotnet test then fails to launch with Framework 'Microsoft.NETCore.App', version '9.0.0' not found and reports Zero tests ran, and DOTNET_ROLL_FORWARD does not help through the dotnet test driver. Run the built test executable directly instead:
dotnet build -c Release
$env:DOTNET_ROLL_FORWARD = "LatestMajor"
tests/ClrProfiler.UnitTest/bin/Release/net9.0/ClrProfiler.UnitTest.exe --treenode-filter "/*/*/ProfilerLifecycleTest/*"
tests/CleProfiler.DatadogTracing.UnitTest/bin/Release/net9.0/CleProfiler.DatadogTracing.UnitTest.exe
Add --output Detailed to see per-test results and any Console.WriteLine diagnostics; the default output shows only the summary.
Do not work around this by adding target frameworks to the test projects. CI provisions the runtime and runs dotnet test normally.
Test the right boundary
- Prefer deterministic tests through public APIs or existing
internal seams exposed with InternalsVisibleTo.
- For raw CLR payload parsing, feed synthetic event names, timestamps, and payloads into the narrow listener seam. Do not require real GC or scheduler timing when a deterministic payload answers the question.
- Use real runtime events only for integration behavior that cannot be represented synthetically. Put such tests in the non-parallel collection when they alter process-wide GC, EventSource, DogStatsD, port, or tracker state.
- Use loopback or fakes for metric transport. Bound all waits with
CancellationTokenSource, WaitAsync, or an equivalent timeout.
- Do not use arbitrary
Task.Delay as the primary assertion mechanism. Signal completion with TaskCompletionSource created with RunContinuationsAsynchronously.
- Assert callback count, order, payload fields, and error forwarding explicitly. Avoid assertions that only prove that no exception was thrown.
Required regression coverage
Choose the applicable cases; do not add unrelated tests.
Event listeners and channels
- Recognized and ignored event names, including versioned names.
- Missing, null, differently typed, malformed, and out-of-range payload fields.
- Correct statistics fields and tick-to-duration conversion.
- Event order and behavior at channel capacity; make intentional loss policy explicit.
- Callback exception reporting followed by successful processing of a later event.
- Reader cancellation and continued reading across
Stop/Restart.
GC correlation
- A normal start/end pair.
- Overlapping foreground and background collections completed in a different order.
- An end observed without its start.
- Stale or colliding bounded correlation state when the implementation makes that possible.
- Suspend/restart correlation independently from collection start/end correlation.
Timers and lifecycle
- Independent state for multiple listeners or trackers.
- Idempotent repeated lifecycle calls.
- Stop/restart behavior without spawning duplicate readers or timers.
- Cancellation, reset, and disposal ownership.
- A disposed timer or tracker cannot recreate owned resources.
Metric adapters
- Exact metric name, type, value, and tags for every affected statistics variant.
- Enum/reason mapping, unknown values, and unit conversion.
- Logger and Datadog projections remain aligned where they intentionally expose the same metrics.
- Cache behavior does not leak mutable arrays or grow from arbitrary input.
Multi-target and compatibility checks
The core library targets net8.0, net9.0, and net10.0; the Datadog adapter targets net8.0 and net9.0; tests currently run on net9.0. A passing test target does not prove all library targets compile, so always run the solution build after changing production code.
Because the test projects declare a single target framework, dotnet test -f net8.0 fails with NETSDK1005 rather than running anything. Verify other targets with dotnet build -c Release over the solution.
Keep ClrProfiler zero-dependency. Do not move adapter packages into the core project to simplify a test.
