Resolve a Testability Obstacle

Introduce the smallest behavior-preserving seam needed to test a specific C# behavior, then add deterministic tests that prove both the behavior and the seam. The production edit is a means to the requested test, not an invitation to redesign adjacent code.

When to Use

• A requested test would otherwise read/write the real filesystem.
• Behavior depends on the current time, delay, random value, environment, console, process, or another ambient dependency.
• The user explicitly permits or requests a safe production seam.
• Existing tests cannot control a dependency without process-global mutation.

When Not to Use

• The dependency is already injected or passed as an argument. Write tests with a fake through the existing seam using code-testing-agent.
• The user wants a repository-wide testability audit. Use detect-static-dependencies.
• The user wants wrappers generated but not call sites/tests changed. Use generate-testability-wrappers.
• The user requests a broad mechanical migration. Use migrate-static-to-wrapper, then generate tests separately.
• The user already selected an existing replacement such as TimeProvider or IFileSystem and asks to migrate call sites to it. Use migrate-static-to-wrapper, which also updates affected tests.
• The code is not C#/.NET.

Inputs

Workflow

Step 1: Prove the obstacle

Read the target production path and its existing tests. Identify the exact ambient operation preventing a deterministic test and the behavior that must remain unchanged. Do not run a repository-wide static scan for a single-class request.

If an adequate seam already exists, stop refactoring and use it. This skill adds no value when a fake can already be supplied.

Step 2: Select the smallest safe seam

Choose by dependency and repository constraints:

The scoped AsyncLocal<T> rule applies to every static API that must retain its public static shape — clocks, filesystem access, environment lookups, identity generation, and randomness. The scope captures and restores the previous value; never implement Dispose() as an unconditional assignment to null. Store the provider/value itself in AsyncLocal<T>. Do not put a mutable Stack<T>, list, or other shared mutable collection in the slot: child execution contexts can inherit the same object and corrupt each other's nesting. When the provider itself is mutable (for example an in-memory store or fake time provider), establish a fresh provider inside each parallel flow rather than mutating one inherited instance from a parent context.

Constructor injection is the default for instance classes. Reuse the repository's DI and naming conventions, but do not add a DI container to a class library just to satisfy this workflow.

For a static class or a public API that cannot change, use a scoped ambient seam only when constructor/parameter injection is impossible. The override must:

• flow across await (AsyncLocal<T>, not [ThreadStatic]);
• return IDisposable and restore the previous value, including nested scopes;
• default to the real production dependency;
• avoid a process-global mutable fake that makes tests non-parallel.

Use built-in fake-time-aware overloads instead of inventing an IDelay wrapper:

Test delayed behavior by starting the operation, proving it is incomplete, advancing FakeTimeProvider, then awaiting it. Never wait for wall-clock time.

For a nested ambient override, each scope owns the value that was active when it started. Dispose scopes in LIFO order with using (which emits try/finally) or an explicit finally; disposing the inner scope restores the outer value, never an unconditional null. For an environment-backed static API, use this shape:

public static class FeatureFlags
{
    private static readonly AsyncLocal<Func<string, string?>?> s_environment = new();

    public static bool IsEnabled(string name)
    {
        var reader = s_environment.Value;
        var value = reader is null
            ? Environment.GetEnvironmentVariable(name)
            : reader(name);

        return string.Equals(value, "true", StringComparison.OrdinalIgnoreCase);
    }

    public static IDisposable OverrideEnvironment(Func<string, string?> reader)
    {
        ArgumentNullException.ThrowIfNull(reader);

        var previous = s_environment.Value;
        s_environment.Value = reader;
        return new RestoreScope(() => s_environment.Value = previous);
    }

    private sealed class RestoreScope : IDisposable
    {
        private Action? _restore;

        public RestoreScope(Action restore)
        {
            _restore = restore;
        }

        public void Dispose() =>
            Interlocked.Exchange(ref _restore, null)?.Invoke();
    }
}

The exception test must observe the outer value after the exception has escaped the inner using scope but before the outer scope is disposed:

using var outer = FeatureFlags.OverrideEnvironment(_ => "true");
Assert.True(FeatureFlags.IsEnabled("Preview"));

Assert.Throws<InvalidOperationException>(() =>
{
    using var inner = FeatureFlags.OverrideEnvironment(_ => "false");
    Assert.False(FeatureFlags.IsEnabled("Preview"));
    throw new InvalidOperationException("test");
});

Assert.True(FeatureFlags.IsEnabled("Preview"));

Also overlap two async flows that each establish a fresh override and assert that each flow sees only its own value. Parallel-only tests do not catch the common "dispose sets null" bug. Do not mutate process environment variables in these tests; the scoped reader is the deterministic input. Choose an outer value different from the production fallback so clearing the slot cannot accidentally pass the restoration assertion.

Step 3: Preserve behavior and API shape

Keep the production change mechanical:

• Wrap only members used by the target behavior.
• Default implementations delegate directly to the original API.
• Preserve exceptions, path handling, time zone, and DateTime.Kind.
• Keep existing public signatures unless the user explicitly permits an API change.
• Do not move business logic into the wrapper or fix unrelated production bugs.

Deterministic serialized text is a deliberate exception to preserving ambient platform formatting. If the user asks for exact reproducible output across platforms, use the format's explicit separator (use literal \n when none is specified) and assert that literal content. Keep Environment.NewLine only when platform-native output is part of the existing contract.

For time replacements:

• DateTime.UtcNow -> timeProvider.GetUtcNow().UtcDateTime
• DateTime.Now -> timeProvider.GetLocalNow().LocalDateTime
• DateTimeOffset.UtcNow -> timeProvider.GetUtcNow()
• DateTimeOffset.Now -> timeProvider.GetLocalNow()

Step 4: Keep production defaults wired

Update every composition root or constructor call affected by the seam. Production must still use real time/filesystem/etc. by default. If the project uses DI, register the default implementation with the lifetime matching repository conventions. If it does not use DI, compose explicitly; do not introduce a container.

Build the affected production project before writing tests. A compile failure here is a seam problem, not a test problem.

Step 5: Write deterministic tests

Use the repository's existing test project. If none exists, invoke scaffold-dotnet-test-project first.

Tests must supply controlled dependencies:

• fixed/advanced time rather than wall-clock waiting;
• an in-memory fake filesystem or hand-rolled fake rather than temp/real files;
• no environment mutation, external process, console input, or network.

Assert the requested business result and at least one interaction/state observable that proves the fake dependency drove the path. Include a production-default test only when it can remain deterministic; never touch the real filesystem merely to prove the adapter delegates.

Choose the narrowest seam that supports the behavior. A single File.WriteAllText call can be an injected Action<string, string> with a real default; do not create an interface, implementation, friend-assembly setting, and extra project wiring unless repository conventions or multiple operations justify them.

Do not add InternalsVisibleTo merely to reach a constructor-injected delegate or other seam that the test project can already supply. Friend-assembly access is justified only when the chosen minimum seam must remain internal and the exact test assembly is known.

Step 6: Verify the complete path

Run the affected production build, targeted test project, and repository-level test command. Re-read the diff and confirm:

1. every production change is required by the seam;
2. no real ambient resource is used by the new tests;
3. current-time semantics and public behavior are preserved;
4. existing tests were not replaced or duplicated.

Inspect the test summary, not only the exit code. Zero discovered tests, a build without the requested test run, or any failing/erroring test means the task is incomplete. Fix discovery/execution and rerun before reporting success.

Output Contract

Provide a compact Requirement | Evidence table. Cite the production seam, production default wiring, exact test names, and passing commands. If a package restore or build blocks validation, report that blocker rather than claiming the tests pass.

Validation

• The original obstacle was concrete and in the requested path.
• An existing seam was reused when available.
• The new abstraction exposes only members required by the target behavior.
• Production defaults still delegate to the original dependency.
• Time conversions preserve local/UTC and DateTime.Kind semantics.
• Static ambient overrides are async-safe, scoped, nested, and reversible.
• New tests use fixed/in-memory dependencies and no real I/O or wall clock.
• Production build and targeted/repository tests pass with at least one requested test discovered.

Common Pitfalls