Skip to content

v0.9.0

Released November 8, 2026

v0.9.0 brings 225 pull requests, one week after v0.8.0. Highlights:

  • Run agent steps on Claude Code or Codex, with their own login.
  • Post the test results as a pull request comment from GitHub Actions.
  • Rerun only what failed.
  • Check the setup with e2e doctor before a run.
  • Get a ready-to-run test project from e2e init.
  • Web locators that behave like Playwright’s, backed by hundreds of ported Playwright tests.
Terminal window
dotnet add package E2E --version 0.9.0
dotnet add package E2E.NUnit --version 0.9.0 # or E2E.XUnit.V3
dotnet tool update --global E2E.Cli

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

E2E.Acp runs Agent.ActAsync and Agent.AssertAsync on Claude Code or Codex over the Agent Client Protocol. The agent signs in with its own login, so you need no model provider. The runner records every action, so steps still replay from the cache without the agent. The agent gets only the step tools: its own file, shell, and web tools are off, and a tool of its own fails the step with POLICY_DENIED.

Terminal window
dotnet add package E2E.Acp
npm install -D @agentclientprotocol/claude-agent-acp # Node must be on the PATH
public abstract class AppTest : E2ETest
{
private static readonly IStepExecutor Claude = AcpExecutor.ClaudeCode(new AcpAgentOptions { Model = "sonnet" });
protected override IStepExecutor? AgentExecutor => Claude;
}

Use AcpExecutor.Codex() for Codex, or AcpExecutor.Create(command, args) for any other agent that speaks ACP. Your coding agent guide

E2E.GitHub posts the results as one pull request comment from GitHub Actions, writes the job summary, and updates the same comment on a rerun. Failures come with their step, the line to look at, and links to the evidence.

Terminal window
dotnet add package E2E.GitHub --version 0.9.0

Reporters now come from e2e.config.json:

{ "reporters": ["E2E.GitHub.GitHubReporter, E2E.GitHub"] }

Give the workflow pull-requests: write and pass GITHUB_TOKEN: ${{ github.token }} to the test step. Pull request comments guide

E2E_LAST_FAILED=1 reruns the tests the last run did not pass. Every other test skips itself. The GitHub comment folds the rerun into the previous run, so it still shows the whole suite.

Terminal window
dotnet test
E2E_LAST_FAILED=1 dotnet test

Or override LastFailed on your E2ERunFixture. Test filters guide

e2e doctor checks that e2e explore and e2e mcp can start here, without starting anything: the config, each target’s engine (browser installed, devices, app path), and the model each agent resolved to. --check-model calls each model once, and --json prints a report for scripts.

Terminal window
e2e doctor
e2e doctor --check-model

e2e doctor guide

e2e init now creates the test project too: a .csproj with the packages at the tool’s version, and an ExampleTests.cs. In a folder that already has a project, it adds only the missing package references.

Terminal window
e2e init MyApp.E2ETests # NUnit
e2e init MyApp.E2ETests --framework xunit --install
e2e init --example TodoMvc my-sample # start from a sample

Coding agents guide

Accounts and several targets in the config

Section titled “Accounts and several targets in the config”

credentials declares test accounts. The password is a secret and is redacted. Credentials.User(name) reads the account in a test, and e2e explore and e2e mcp sign in with them. targets can list several entries, each with its own name.

{
"targets": [
{ "name": "web", "platform": "web", "app": { "url": "http://127.0.0.1:4173" } },
{ "name": "ios", "platform": "ios", "app": { "bundleId": "com.example.app" } }
],
"credentials": { "admin": { "username": "admin@example.com" } }
}

Leave the password out of the committed file and set E2E_USER_ADMIN_PASSWORD, which wins over the config. Pick a target with E2EFixture.TargetName, OpenHostOptions.Target, or --target.

Sign in once and restore the session in other tests, as NUnit could in v0.8.0.

public sealed class Dashboard : E2ETest
{
[E2ESetup("authenticate as admin", Saves = ["admin"])]
public async Task Authenticate()
{
var admin = Credentials.User("admin");
await App.OpenAsync("/login");
await Screen.GetByLabel("Username").FillAsync(admin.Username);
await Screen.GetByLabel("Password").FillAsync(admin.Password);
await Screen.GetByRole("button", "Sign in").TapAsync();
await SaveSessionAsync("admin");
}
[Fact, E2ESession("admin")]
public async Task Opens_the_dashboard_directly() => await App.OpenAsync("/dashboard");
}

Test filters also work on xUnit v3 and on the mobile and Windows adapters.

Expect.That(value) checks plain values, with Jest-style matchers and partial matches:

Expect.That(links, "the playground lists its pages").ToBeGreaterThanOrEqual(2);
Expect.That(payload).ToEqual(new { Id = Expect.StringMatching("^usr_"), CreatedAt = Expect.Any(typeof(DateTime)) });

Value checks

  • Vision: a screenshot tool and a pixel mode, so the agent can act on a canvas, a chart, or a screen the engine cannot read.
  • Point tools: tap_at, hover_at, drag_at, type_at, press_at, select_at.
  • More verbs: long_press, right_click, hover, drag, dismiss_keyboard, and resize_viewport.
  • Exploration: e2e explore runs up to 100 steps and 2 hours, digs into edge cases once the main flows are covered, and prints through the list reporter.
Terminal window
e2e cache ls # one row per recorded step
e2e cache stats # directory, entries, size
e2e cache clear

E2E_NO_CACHE=1 runs every step live for one run, to try a new model.

Web locators that behave like Playwright’s

Section titled “Web locators that behave like Playwright’s”

Hundreds of Playwright’s own tests (selectors, locators, clicks, fill, keyboard, frames, routes, dialogs, downloads, screenshots) now run against the web engine. The fixes they drove:

  • GetByText and hasText match the DOM text, and GetByLabel matches aria-label, aria-labelledby, and <label> text.
  • Browser.Locator("body"), and any CSS selector, matches elements the agent’s tree does not list.
  • Browser.EvaluateAsync survives pages that replace globals such as Array or JSON.
  • BrowserStack App Automate real devices (E2E.BrowserStack). BrowserStack guide
  • e2e mcp: start_recording and stop_recording, type_secret with configured accounts, redaction of what user code prints, and registration for Codex in .codex/config.toml.
  • The list reporter shows live progress, the code frame and video under a failure, agent step lines, and rows for skipped, warned, and repeated tests.
  • E2E.Acp: run steps on an ACP coding agent, with Claude Code and Codex presets (#1000, #1004)
  • E2E.GitHub: the pull request comment reporter (#962); reporters from e2e.config.json (#969, #979)
  • --last-failed reruns (E2E_LAST_FAILED), folded into the GitHub comment (#999, #1010)
  • e2e doctor, with mobile targets and model checks (#988, #996)
  • e2e init scaffolds the test project: --framework, --install, --example (#1027)
  • e2e cache ls, stats, clear; E2E_NO_CACHE and E2E_CACHE_DIR_OVERRIDE (#1374, #1376, #1335, #1379)
  • Config: several targets (#937), credentials (#941, #945, #1263), artifacts and code-only stores (#942), output (#946), failOnSkippedFailure and maxDuration, carried and checked only (#1338); secret and credential providers (#1317)
  • Saved sessions (#944) and test filters (#964) on xUnit v3; filters on the mobile and Windows adapters (#990)
  • Value expectations: Expect.That(value) (#1373)
  • Agent: screenshot tool and pixel mode (#951), point tools (#959, #1013), long_press, right_click, hover, drag (#960), dismiss_keyboard (#1026), resize_viewport (#1045, #1047), the #id listing and diff results (#1034)
  • Explore: up to 100 steps and 2 hours, edge-case lenses (#1043); printed through the list reporter (#1042, #1031)
  • MCP: start_recording and stop_recording (#936), node ids and diff results (#939), type_secret with configured credentials (#952), redacts what user code prints (#1274), Codex registration (#1275)
  • BrowserStack device provider (#967)
  • List reporter: live window and console output (#1024), code frame and video rows (#963), agent step lines (#991), goldens and closing rows (#961, #1030)
  • Engine keyboard: typeText, pressKey, keyboard point tools (#992)
  • Run events, traces, and the AI trace path for reporters (#997); markdown report on the public Report tree (#957)
  • BROWSER_INSTALL_FAILED error code (#1252)
  • Web parity with Playwright: text and label matching on the DOM (#1089), CSS selectors list any element (#1033, #1040), frames (#1070, #1086), clicks and tap position (#1084, #1112), file inputs and routes (#1058), screenshots and masks (#1111, #1115), aborted navigations (#1113), EvaluateAsync on clobbered globals (#1116)
  • Replay cache: relocation by position and placeholder (#1267, #1289), dates with ordinal suffixes (#980, #1358), typed-value and secret-fill checks on replay (#1257, #1333, #1371), truncated traces (#1321)
  • Locator failures list the candidates each filter step left (#1015); timeout messages name the operation (#1119)
  • Flaky test fixes across timing, ports, and processes (#956, #1011, #1029, #1049, #1128, #1134, #1136, #1139)
  • Releases section (#931), sidebar restructure (#932), CodeRabbit accuracy fixes (#933)
  • Bug bash guide (#947), migration guides (#948, #1061), writing an engine (#1280)
  • References: errors (#1291, #1365), environment variables (#1255), agentic or deterministic tests (#1266), timeouts (#1392)
  • E2E.Eas is now E2E.Expo. Update the package reference and using E2E.Eas;. EasSimulators and EasSimulatorsOptions are now ExpoSimulators and ExpoSimulatorsOptions, and the provider name is expo-simulators. There are no aliases (#982, #987)
  • Reporters come from e2e.config.json. [assembly: E2EReporter] is removed. List reporters in reporters (#969)
  • WaitForAsync timeouts fail with LOCATOR_NOT_FOUND, not TIMEOUT (#938)
  • InputValueAsync and IsCheckedAsync throw ACTION_FAILED on an element that cannot hold a value or a checked state, instead of returning "" or false (#1114)
  • Browser.EvaluateAsync returns plain JSON only. A class instance such as a DOMRect fails with INVALID_ARGUMENT. Return rect.toJSON() (#1116)
  • Web text matching follows the DOM. GetByText no longer matches a placeholder or the control a label names. Use GetByLabel or GetByPlaceholder (#1089)
  • E2E.XUnit.V3 needs xunit.v3.extensibility.core 4.0.1 or later (#944)
  • ReportExploreFinding.ArtifactId is an artifact id, not a file path (#966)