e2e@0.22.0

Minor Changes

• #1437 94e6e76 Thanks @okwasniewski! - Breaking: a secret fill no longer turns off screenshots for the model. agent.assert(..., { vision: 'only' }), vision: true, the screenshot and point tools, e2e mcp, and e2e explore findings keep working after a sign-in, in serial groups, and in tests that restore a saved session. Screenshots still mask secure fields, and secret values are still redacted from all text. A generic secret filled into a plain input, or a value the app shows outside a secure field, such as behind a show-password toggle, is in the screenshot and reaches your model provider, so use test credentials. StepExecutorContext.pixelsTainted is removed, and a new step's visionDegraded is never PIXEL_TAINTED.

Patch Changes

• #1282 7518ae5 Thanks @MarcosCamara01! - Add drag_at to the agent's pixel tools, in agent.act and e2e mcp, and dragAt(from, to) to ExecutorActions: a drag from one screenshot point to another, for a shape on a canvas or a pin on a map. Like tap_at, the bare points go to the engine's dragTo pointer kind, where the model aimed; an engine without it drags by id when the points land on two different listed controls, like drag. The replay cache records both points and replays each the way it replays a tap_at point.

• #1447 82d5095 Thanks @21prnv! - A replay no longer fails its end check on an alert that shows an ISO 8601 timestamp, such as Saved at 2026-10-24T15:41:27.955Z, when the time differs from the recording.

• #1381 d2ffb31 Thanks @okwasniewski! - app.startRecording() and app.stopRecording() record one span of the screen, such as a single transition, on the web and on devices. stopRecording returns the videos: a file path, or a url when a hosted browser or device keeps the recording. A start fails while the test's video records, so set { video: 'off' } on the test.

  test('the dialog opens', { video: 'off' }, async ({ app, screen }) => {
    await app.open();
    await app.startRecording({ name: 'dialog' });
    await screen.getByRole('button', 'Open').tap();
    const [video] = await app.stopRecording();
  });

  An engine's startVideo and stopVideo may now be called more than once per attempt, never overlapping.

• #1387 4461e9d Thanks @bheemreddy-samsara! - Add onBeforeAction to agent.act, a hook that runs before each action, so a test can capture the screen an action started on.

• #1379 9e53c40 Thanks @okwasniewski! - A mobile target without device skips a booted simulator or emulator that another agent-device session holds, such as one another tool is driving, and drives the free ones. When every booted device is held, the run fails at prepare naming the holding sessions instead of failing on the first one.

• #1382 3a8d6f1 Thanks @okwasniewski! - Bump agent-device to 0.21.31, zod to 4.6.7, @modelcontextprotocol/server to 2.4.0, and oxc to 0.155.0.

• #1233 34287cf Thanks @Varshith-Kali! - A run now exits with the cleanup error's code, usually 3, when a test's cleanup failed. The list reporter shows the cleanup failure in the summary and under the test for passed and flaky tests, so a CI job that goes red names the cleanup failure instead of reporting a green summary.

• #1351 bfae325 Thanks @samanyudas! - e2e run and e2e explore refuse an empty --reporter (--reporter '', --reporter ',', or --reporter "$REPORTERS" with the variable unset) as a usage error, exit 2, instead of running with no reporters and printing nothing.

• #1287 5aec484 Thanks @TheTuso! - An e2e command whose config or test file leaves a timer or socket open exits two seconds after it finishes, instead of staying open.

• #1354 6901a01 Thanks @samanyudas! - toHaveCount and the list forms of toHaveText and toContainText read a locator under a missing frame as zero matches, so toHaveCount(0) passes after an iframe closes instead of failing with LOCATOR_NOT_FOUND.

• #1401 ce249b6 Thanks @kacperkapusciak! - e2e explore in a project with no config runs the e2e init wizard first instead of failing with CONFIG_NOT_FOUND. The exploration starts right after when nothing is left to do by hand: dependencies installed, a model with its key set or its sign-in done, and a web target, whose URL init names. Otherwise init lists the steps left, ending with the e2e explore command as typed. Without a terminal, as in CI or under a coding agent, nothing changes: no files are written and the error stays.

• #1334 6c8e71f Thanks @21prnv! - locator.filter() takes hasNot, which drops matches whose subtree resolves a locator, as in Playwright. Third-party engines must handle the new hasNot field of a filter expression.

• #960 ffed34e Thanks @enekesabel! - Export the built-in agent's action tools, verdict tool, and rules from e2e/agent: createGrammarTools, createVerdictTool, BASE_RULES, and VERDICT_RULES, plus isRuntimeHardStop and the ScreenOutput type, for a custom tool loop or executor that keeps the built-in actions and verdicts.

• #1389 6b04d18 Thanks @kacperkapusciak! - e2e init --example with-angular starts a new project from the new Angular example: a standalone component greeting app with three locator tests and two agent tests, run against the Angular dev server.

• #1398 e087ed2 Thanks @samanyudas! - e2e init --help no longer says it leaves existing files untouched. It says an existing config and test files stay as they are, while package.json and .gitignore get only what they are missing.

• #1338 0d3e403 Thanks @kacperkapusciak! - e2e init lets you select both the web and mobile engines. The engine question is now a multiselect, and --engine takes a comma-separated list (--engine web,mobile). With both, the config gets a target for each, and the examples go in tests/web.e2e.ts and tests/mobile.e2e.ts, each test with requires so the other target skips it. Selecting no engine replaces the None option.

• #1463 7372b0a Thanks @okwasniewski! - agent.assert and agent.waitFor take params, like agent.act. The judge reads {name} with its value, a unique() value included. A Secret reaches the judge by name only, so check a secret's value on screen with a locator assertion.

  const name = unique(`acme-${Date.now()}`);
  await agent.act('create a company named {name}', { params: { name } });
  await agent.assert('the table shows company {name}', { params: { name } });

• #1316 6b41857 Thanks @kacperkapusciak! - e2e mcp sessions have a run_code tool. Pass it test statements as a test body writes them, such as await screen.getByLabel('Email').fill('ada@example.com'), and the session runs them on its screen and app with a test's waits and matchers. It stops at the first failure and returns each statement's result, then what changed on screen. The statements that pass, except reads, join the test_lines draft. A project tool named run_code now fails the config load with INVALID_CONFIG. The bundled skill tells the coding agent to run the lines no action printed, such as assertions, with run_code before it writes them into the test.

• #1315 3d2c362 Thanks @kacperkapusciak! - e2e mcp's open_session takes a test, named like an e2e run positional (tests/checkout.e2e.ts:12), and runs that one test in the session before handing it over: its hooks, test.extend fixtures, and body, as e2e run runs them. The session then stays on the screen where the test passed or failed, and the opening text leads with the outcome: the error code and message, the failing line with the code around it, the URL, the nodes nearest a locator that matched nothing, and the steps that ran. afterEach hooks and fixture teardown run at close_session. The test's own steps print no test line; the session's lines start with the agent's first action, and test_lines returns them as the lines that continue the test. Agent steps replay and record in the replay cache as on a run's first attempt. A name that selects several tests fails with TEST_AMBIGUOUS, and a serial group member, a test that starts from a setup test's session, or a test that runs as several agents fails with SESSION_TEST_UNSUPPORTED. The bundled skill tells the coding agent to fix a failing test this way instead of clicking through the flow again.

• #1191 9bc7c32 Thanks @TheTuso! - Agent navigate to a full URL on the app now records it without the origin, so cached steps replay when the app runs on another port or host with the same app.identity. Re-record older entries to pick this up.

• #1037 a481fd3 Thanks @rudycelekli! - locator.filter() takes hasNotText, which drops matches whose text contains a string or matches a RegExp, as in Playwright. Third-party engines must handle the new hasNotText field of a filter expression.

• #1215 d8c597b Thanks @Varshith-Kali! - A structured model call that returns no parseable output (for example an empty length finish) now fails as MODEL_OUTPUT_INVALID instead of MODEL_PROVIDER_FAILED.

• #1328 1c4e906 Thanks @DeryFerd! - e2e login on Windows no longer opens the sign-in URL through a command interpreter. The URL comes from the provider's own response in the device flow, and cmd /c start read a |, &, or ^ in it as command syntax, so a URL that carried one ran a command of its own. The URL is now handed to the default handler for its scheme as a single argument, and only an http: or https: URL is opened; any other scheme is printed for you to open yourself.

• #1399 92409e5 Thanks @samanyudas! - The built-in agent's instructions now say that tap_at, hover_at, and drag_at act on the bare points when the engine supports it, and fall back to the listed controls by id only when it does not. They previously said every point tool acts on a listed control by its id.

• #1357 9837386 Thanks @amoscicki! - A replayed step whose last action starts a navigation now waits up to 15 seconds for the recorded destination, instead of handing off to the agent after about 5 seconds.

• #1056 dc8f56d Thanks @rudycelekli! - Report a test's duration from the attempt that decided it when a retry is interrupted or cannot start.

• #1256 2bc00e9 Thanks @arun2728! - e2e/runner exports explore(), the counterpart of e2e explore that run() is to e2e run. It takes the project (cwd, and config as a file path or an E2EConfig value), the goal and its budgets (goal, target, agent, session, maxSteps, timeoutMs), interruptSignal, and onEvent, and resolves with the run's result plus the exploration record (explore.ended, explore.steps, explore.findings) without exiting the caller. Cancellation awaits teardown. The notices e2e explore prints before the run starts, such as the target it chose, arrive through onEvent as run notices. It shares run()'s reservation, so overlapping calls reject with RUNNER_BUSY, and each exploration that returns a report sends one e2e_run_completed event with command: "runner.explore".

• #1312 9d884f5 Thanks @kacperkapusciak! - The bundled agent skill now has the coding agent write a test while it walks the flow over e2e mcp, instead of exploring first and writing the test from memory. It creates the test file before it opens the session and adds each step as it takes it. A step is an agent.act goal, written in the words on screen before the agent walks its path, or, for an exact value, the Test line: the action printed. When the outcome shows, the agent locates it and adds the expect. The e2e mcp server's instructions, e2e mcp --help, and the MCP server and coding agents docs say the same.

• #1372 0cc79f1 Thanks @BiBoyang! - Report rejected recording writes as not-written.

• #1452 5b6fd01 Thanks @samanyudas! - A multi-line unique() value now replays from the cache. The screen shows it on one line, so the end check now looks for it on one line too. Before, the replay missed: REPLAY_STALE (end-mismatch) with --strict-cache, a hand-off to the agent without it. An older runner reads an entry recorded with this fix as invalid-entry.

• #1394 e9e8c50 Thanks @kacperkapusciak! - e2e upgrade updates the project's copies of the agent skill in .agents/skills/e2e and .claude/skills/e2e to the newly installed version, instead of telling you to run e2e init. It also updates a stale copy when the packages are already up to date, and --dry-run names the copies it would update.

• #958 85e92f9 Thanks @kuishou68! - upload and setInputFiles accept a button that opens a file chooser, such as one over a hidden file input. The agent's upload tool says so, so the model can pick the button it sees.

• #1429 3a841bb Thanks @KrzysztofMoch! - A web() target takes app.permissions with 'grant' states, granted to every page of the attempt: permissions: { 'clipboard-read': 'grant', 'clipboard-write': 'grant' } lets the app use the clipboard in headless Chromium. Names are checked against the browser's at config load; 'deny' and 'reset' are refused on web.