Skip to content

v0.10.0

Released November 11, 2026

v0.10.0 brings 123 pull requests, three days after v0.9.0. Highlights:

  • The libraries ship for .NET 11 and .NET 10.
  • [Serial] test classes on NUnit: the members share one session.
  • A hook before each agent action.
  • Project tools declared as commands in e2e.config.json, served by e2e mcp.
  • ToMatchSchema and Expect.Undefined.
  • Steadier mobile runs: a shut-down device boots when none is booted, and a leased device stays with its worker across attempts.
Terminal window
dotnet add package E2E --version 0.10.0
dotnet add package E2E.NUnit --version 0.10.0 # or E2E.XUnit.V3
dotnet tool update --global E2E.Cli

⚠️ This release has breaking changes. See Breaking changes before you upgrade.

Every library package ships net10.0 (the current LTS) and net11.0 (the current STS). src/Directory.Build.props sets both. The API is the same on each, and nothing was dropped: a net10.0 project keeps building.

Tests, samples, and apps stay on net10.0. The MAUI sample and the web benchmark target net11.0.

A project that runs on .NET 11 only:

<TargetFramework>net11.0</TargetFramework>

A test project that runs on both. dotnet test runs the suite once per framework:

<TargetFrameworks>net10.0;net11.0</TargetFrameworks>

Playwright’s install script follows the build output. For a net11.0 build:

Terminal window
pwsh bin/Debug/net11.0/playwright.ps1 install chromium
Package v0.9.0 v0.10.0
E2E, E2E.NUnit, E2E.XUnit.V3 net10.0 net10.0, net11.0
E2E.Mobile, E2E.Mobile.NUnit, E2E.Mobile.XUnit.V3, E2E.Mobile.DevFlow net10.0 net10.0, net11.0
E2E.WinApp, E2E.WinApp.NUnit, E2E.WinApp.XUnit.V3 net10.0 net10.0, net11.0
E2E.Acp, E2E.GitHub, E2E.Expo, E2E.BrowserStack, E2E.Limrun, E2E.Kernel net10.0 net10.0, net11.0
E2E.TestMu, E2E.TestMu.Mobile net10.0 net10.0, net11.0
E2E.Cli net10.0 net10.0

E2E.Cli stays on net10.0. dotnet run fails on a multi-targeted project unless you pass -f, so the tool stays on one framework. The project sets no RollForward. The machine that runs e2e needs the .NET 10 runtime. A .NET 11 SDK does not include it.

The MAUI guide builds net11.0-ios and net11.0-android. The iOS simulator build is bin/Debug/net11.0-ios/iossimulator-arm64/MyApp.app. .NET 11 raises Android’s minimum to API 24, and the workload needs the API 37 platform. The test project stays a plain net10.0 or net11.0 project, so it builds on a machine without the MAUI workload.

E2E.WinApp, E2E.WinApp.NUnit, and E2E.WinApp.XUnit.V3 add net11.0-windows7.0 beside net10.0-windows7.0 in AssetTargetFallback. The UI Automation library ships Windows assets only, so a project that restores it repeats both fallbacks and adds NU1701 to NoWarn. Windows guide

global.json pins SDK 11.0.100 with rollForward set to latestFeature, so a later 11.0 SDK is accepted for the build. That setting picks the SDK. It does not roll the e2e tool onto the .NET 11 runtime. The CI workflows install 10.0.x and 11.0.x.

[Serial] on an NUnit E2ETest class makes its members one group (#1489). They share one session, so the page one member left is still there for the next, and a field carries a value between them. The session starts with the first member. After a member fails, the later members are skipped. Give them [DependsOnTest] to fix the order. Do not combine it with [FixtureLifeCycle(LifeCycle.InstancePerTestCase)], or with tests running in parallel in the fixture.

[Serial]
public sealed class WorkspaceWizardTests : E2ETest
{
private string _chosenPlan = "";
[Test]
public async Task Step_1_name_the_workspace()
{
await App.OpenAsync("/wizard");
await Screen.GetByLabel("Workspace name").FillAsync("Rocketry");
await Screen.GetByRole("button", "Next").TapAsync();
}
[Test, DependsOnTest(nameof(Step_1_name_the_workspace))]
public async Task Step_2_pick_a_plan()
{
await Screen.GetByLabel("Plan").SelectOptionAsync("Pro");
_chosenPlan = "Pro";
await Screen.GetByRole("button", "Next").TapAsync();
}
}

ActOptions.OnBeforeAction runs before each action of Agent.ActAsync, including actions the replay cache replays (#1436). The step waits for it. Use it to capture the screen the action started on. The event has Index (1-based, the same index the action’s OnAction event gets) and Source (ActionSource.Model, or ActionSource.Replay for a cached action). It does not wait for the screen to settle. A mutating project tool runs it. A read-only tool does not. If it throws, the step fails before the action runs, with a message that starts with agent.act onBeforeAction failed, and the recording is left alone.

The OnAction event that follows reports the recorded action summary (#1428).

await Agent.ActAsync("open settings and turn on dark mode", new ActOptions
{
OnBeforeAction = e => App.ScreenshotAsync($"before {e.Index}"),
});

agents.<name>.tools declares project tools as commands (#1484). Each command runs in the project root. The model’s input is JSON on stdin, stdout is the answer, and stderr is printed (e2e mcp redacts it). A non-zero exit, or two minutes without an exit, fails the call. e2e mcp serves the tool beside the session’s tools. A tool written in code is still an AgentTool on the fixture.

description, mutates, and command (the program, then its arguments) are required. inputSchema is a JSON schema object. Leave it out and the tool takes an object with no properties. platforms limits where the tool is offered. mutates: true takes a slot of the step’s action budget, and the replay cache cannot replay past the call.

A name the agent or an MCP session already uses is INVALID_CONFIG (#1517): observe, done, the action verbs, complete_step, locate, run_code, test_lines, start_recording, stop_recording, and report_finding.

{
"agents": {
"default": {
"tools": {
"seed_cart": {
"description": "Seed the cart with a SKU",
"mutates": true,
"command": ["node", "tools/seed-cart.mjs"],
"inputSchema": {
"type": "object",
"properties": { "sku": { "type": "string" } },
"required": ["sku"],
"additionalProperties": false
}
}
}
}
}
}

ToMatchSchema checks a plain value against a schema you write as a StandardSchemaV1<TOutput> delegate (#1518). There is no Standard Schema library in .NET. The delegate returns StandardSchemaV1Result<TOutput>.Success(output) or .Failure(issues), and each StandardSchemaV1Issue carries its path. A match returns the output, defaults applied. A mismatch lists every issue by its path. A negated check returns null. A schema that validates asynchronously cannot be passed. A null schema is INVALID_ARGUMENT.

StandardSchemaV1<Dictionary<string, object?>> User = value =>
{
if (value is Dictionary<string, object?> record && record.GetValueOrDefault("id") is int id)
{
return StandardSchemaV1Result<Dictionary<string, object?>>.Success(
new Dictionary<string, object?> { ["id"] = id, ["role"] = record.GetValueOrDefault("role") ?? "member" });
}
return StandardSchemaV1Result<Dictionary<string, object?>>.Failure(
[new StandardSchemaV1Issue("expected a number", ["id"])]);
};
var user = Expect.That<object?>(new Dictionary<string, object?> { ["id"] = 1 }).ToMatchSchema(User);

Expect.Undefined stands for JavaScript’s undefined, which C# has no value for (#1402). null is only null. ToBeDefined passes for null. ToBeUndefined passes only for Expect.Undefined. ToEqual ignores a member that holds Expect.Undefined. Value checks

When no device is named and none is booted, prepare boots a shut-down simulator or emulator (#1567). agent-device picks one: on iOS an iPhone over an iPad over an Apple TV, on Android a phone or tablet over a TV, then the newest OS, then name order. A booted device another agent-device session holds is still skipped. When every booted device is held, prepare still fails with ENGINE_FAILURE.

A leased device stays with its worker across the attempts that run on that slot (#1552). Disposing an attempt leaves the leased device booted under its session. The run’s finish closes those sessions, then releases the leases. A provider recording whose stop failed stays marked on the worker and is finished before the next attempt records.

  • A test that restores a saved session runs the [E2ESetup] test that saves it first, once per run (#1488).
  • E2EConfig.WithTarget(name, engine) puts an IEngine on a named target, in place of the target of that name or after the others when none has it (#1476). The config digest includes that engine’s declaration (#1469).
  • Engine authors outside the E2E assembly can call the shared helpers: KeyGrammar, ExpressionResolution.ResolveExpression, TextRules.MatchesText, Urls, Time, Timing.RaceAbortAsync, Secret.IsSecret, and JsonValueValidation.ValidateJsonValue. IEngine.ValidateApp runs once per target at config load (#1530). Writing an engine
  • An upload path sample:photo.png, sample:photo.jpg, sample:document.pdf, sample:notes.txt, sample:table.csv, or sample:empty.txt is a file e2e ships. e2e mcp serves upload when the engine declares EngineCapabilities.Upload, and test_lines prints SetInputFilesAsync for a project file (#1539).
  • A click on a checkbox or radio goes through that control’s own label (#1569).
  • open_session loads the test assembly again on every call, so a rebuilt test runs as it is now. test accepts file:line, the line of the attributes or the signature (#1482).
  • After a test times out, teardown runs under a budget of its own, so the session can still drive the app. open_session {test} opens on the screen the body left (#1492).
  • AssertAsync, WaitForAsync, and ExtractAsync show the judge where the screen is: a path on the app’s origin, the whole address off it (#1541).
  • Device.SetPermissionAsync(string?, string?) takes the wire names, such as "camera" and "grant". An unknown name or state is INVALID_ARGUMENT before any device command (#1499).
  • Device.UnstableRaw accepts apps.uninstall, command.hardwareButton (power, volume-up, volume-down), and settings.update for text-size and animations (Android only). Anything else throws NotSupportedException. TestMu refuses settings and some of these calls with UNSUPPORTED_OPERATION (#1475).
  • A test’s swipe of the screen waits out the transition the previous action started (#1551). A test’s swipe or scroll sets postGestureStabilization to false, and the next action on a control checks that the control stopped moving (#1568).
  • Device prepare logs how long each step took, one line per device, such as boot, the iOS runner, and opening Settings. A step that fails or times out is in that line (#1545).
  • Docs: GitLab CI/CD (#1463) and the Appium migration guide (#1443).

Alongside those, the release ports upstream tests the port was missing, including the web benchmark suites, observation and trace checks, value expectations, and the device-provider lease contract. The full changelog lists every commit.

  • LocatorAction.PressSequentially is removed. Locator.PressSequentiallyAsync is focus, then the keyboard. A custom engine keeps focus in ActionKinds, declares EngineCapabilities.Keyboard, and implements TypeTextAsync so the app gets key events. An engine that declares no focus or no keyboard is refused with UNSUPPORTED_CAPABILITY before any node resolves (#1429). Writing an engine
  • IStepCache gains bool Writable { get; }. Write returns int?: the size in bytes, or null when the entry was not written (#1430).
  • ToBeDefined passes for null. ToBeUndefined passes only for Expect.Undefined (#1402).
  • Expect.Poll throws ConfigurationException (INVALID_CONFIG) for a negative Timeout or a non-positive Interval (#1401). A Timeout of zero sets no deadline: the poll runs until the caller’s cancellation token ends. A token that cannot end is INVALID_CONFIG too.
  • An engine secret it did not declare fails with SECRET_UNAVAILABLE. EngineStartOptions.ResolveSecret resolves only a name listed in IEngine.Secrets, and throws ConfigurationException for any other (#1407).