chore: version packages
This PR was opened by the Changesets release GitHub action. When you're ready to do a release, you can merge this and the packages will be published to npm automatically. If you're not ready to do a release yet, that's fine, whenever you add more changesets to main, this PR will be updated.
Releases
@e2e-dev/acp@0.1.0
Minor Changes
- #960
ffed34eThanks @enekesabel! - Release: New package: runagent.actandagent.asserton your own coding agent over the Agent Client Protocol, signed in with the agent's own login.claudeCode()andcodex()start those agents with e2e's step tools as their only tools;acp({ command, args })starts any other ACP agent. Every action goes through the runner, so steps replay from the cache without the agent, and a tool of the agent's own that runs without asking fails the step withPOLICY_DENIED.
Patch Changes
- #1457
e3195e3Thanks @samanyudas! - A step that starts right after a step that timed out now gets a fresh agent when the old agent runs a tool of its own as it stops. Before, the step failed withMODEL_PROVIDER_FAILED("ACP connection closed"), because it reused the session that tool had closed.
@e2e-dev/browserstack@0.1.0
Minor Changes
- #1307
64432f1Thanks @malopezr7! - Release: New package:@e2e-dev/browserstack.browserstack()leases BrowserStack App Automate real Android and iOS devices for the mobile engine, one isolated agent-device daemon per worker slot. Anhttpsappis uploaded once per run, an attempt's video links BrowserStack's recording of the session, and a release fails the run when agent-device can't confirm it ended the session.
e2e@0.22.0
Minor Changes
- #1437
94e6e76Thanks @okwasniewski! - Breaking: a secret fill no longer turns off screenshots for the model.agent.assert(..., { vision: 'only' }),vision: true, thescreenshotand point tools,e2e mcp, ande2e explorefindings 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.pixelsTaintedis removed, and a new step'svisionDegradedis neverPIXEL_TAINTED.
Patch Changes
-
#1282
7518ae5Thanks @MarcosCamara01! - Adddrag_atto the agent's pixel tools, inagent.actande2e mcp, anddragAt(from, to)toExecutorActions: a drag from one screenshot point to another, for a shape on a canvas or a pin on a map. Liketap_at, the bare points go to the engine'sdragTopointer kind, where the model aimed; an engine without it drags by id when the points land on two different listed controls, likedrag. The replay cache records both points and replays each the way it replays atap_atpoint. -
#1447
82d5095Thanks @21prnv! - A replay no longer fails its end check on an alert that shows an ISO 8601 timestamp, such asSaved at 2026-10-24T15:41:27.955Z, when the time differs from the recording. -
#1381
d2ffb31Thanks @okwasniewski! -app.startRecording()andapp.stopRecording()record one span of the screen, such as a single transition, on the web and on devices.stopRecordingreturns the videos: a filepath, or aurlwhen a hosted browser or device keeps the recording. A start fails while the test'svideorecords, 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
startVideoandstopVideomay now be called more than once per attempt, never overlapping. -
#1387
4461e9dThanks @bheemreddy-samsara! - AddonBeforeActiontoagent.act, a hook that runs before each action, so a test can capture the screen an action started on. -
#1379
9e53c40Thanks @okwasniewski! - A mobile target withoutdeviceskips 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 atpreparenaming the holding sessions instead of failing on the first one. -
#1382
3a8d6f1Thanks @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
34287cfThanks @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
bfae325Thanks @samanyudas! -e2e runande2e explorerefuse 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
5aec484Thanks @TheTuso! - Ane2ecommand whose config or test file leaves a timer or socket open exits two seconds after it finishes, instead of staying open. -
#1354
6901a01Thanks @samanyudas! -toHaveCountand the list forms oftoHaveTextandtoContainTextread a locator under a missing frame as zero matches, sotoHaveCount(0)passes after an iframe closes instead of failing withLOCATOR_NOT_FOUND. -
#1401
ce249b6Thanks @kacperkapusciak! -e2e explorein a project with no config runs thee2e initwizard first instead of failing withCONFIG_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 thee2e explorecommand as typed. Without a terminal, as in CI or under a coding agent, nothing changes: no files are written and the error stays. -
#1334
6c8e71fThanks @21prnv! -locator.filter()takeshasNot, which drops matches whose subtree resolves a locator, as in Playwright. Third-party engines must handle the newhasNotfield of afilterexpression. -
#960
ffed34eThanks @enekesabel! - Export the built-in agent's action tools, verdict tool, and rules frome2e/agent:createGrammarTools,createVerdictTool,BASE_RULES, andVERDICT_RULES, plusisRuntimeHardStopand theScreenOutputtype, for a custom tool loop or executor that keeps the built-in actions and verdicts. -
#1389
6b04d18Thanks @kacperkapusciak! -e2e init --example with-angularstarts 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
e087ed2Thanks @samanyudas! -e2e init --helpno longer says it leaves existing files untouched. It says an existing config and test files stay as they are, whilepackage.jsonand.gitignoreget only what they are missing. -
#1338
0d3e403Thanks @kacperkapusciak! -e2e initlets you select both the web and mobile engines. The engine question is now a multiselect, and--enginetakes a comma-separated list (--engine web,mobile). With both, the config gets a target for each, and the examples go intests/web.e2e.tsandtests/mobile.e2e.ts, each test withrequiresso the other target skips it. Selecting no engine replaces the None option. -
#1463
7372b0aThanks @okwasniewski! -agent.assertandagent.waitFortakeparams, likeagent.act. The judge reads{name}with its value, aunique()value included. ASecretreaches 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
6b41857Thanks @kacperkapusciak! -e2e mcpsessions have arun_codetool. Pass it test statements as a test body writes them, such asawait screen.getByLabel('Email').fill('ada@example.com'), and the session runs them on itsscreenandappwith 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 thetest_linesdraft. A project tool namedrun_codenow fails the config load withINVALID_CONFIG. The bundled skill tells the coding agent to run the lines no action printed, such as assertions, withrun_codebefore it writes them into the test. -
#1315
3d2c362Thanks @kacperkapusciak! -e2e mcp'sopen_sessiontakes atest, named like ane2e runpositional (tests/checkout.e2e.ts:12), and runs that one test in the session before handing it over: its hooks,test.extendfixtures, and body, ase2e runruns 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.afterEachhooks and fixture teardown run atclose_session. The test's own steps print no test line; the session's lines start with the agent's first action, andtest_linesreturns 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 withTEST_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 withSESSION_TEST_UNSUPPORTED. The bundled skill tells the coding agent to fix a failing test this way instead of clicking through the flow again. -
#1191
9bc7c32Thanks @TheTuso! - Agentnavigateto 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 sameapp.identity. Re-record older entries to pick this up. -
#1037
a481fd3Thanks @rudycelekli! -locator.filter()takeshasNotText, which drops matches whose text contains a string or matches a RegExp, as in Playwright. Third-party engines must handle the newhasNotTextfield of afilterexpression. -
#1215
d8c597bThanks @Varshith-Kali! - A structured model call that returns no parseable output (for example an emptylengthfinish) now fails asMODEL_OUTPUT_INVALIDinstead ofMODEL_PROVIDER_FAILED. -
#1328
1c4e906Thanks @DeryFerd! -e2e loginon 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, andcmd /c startread 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 anhttp:orhttps:URL is opened; any other scheme is printed for you to open yourself. -
#1399
92409e5Thanks @samanyudas! - The built-in agent's instructions now say thattap_at,hover_at, anddrag_atact 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
9837386Thanks @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
dc8f56dThanks @rudycelekli! - Report a test's duration from the attempt that decided it when a retry is interrupted or cannot start. -
#1256
2bc00e9Thanks @arun2728! -e2e/runnerexportsexplore(), the counterpart ofe2e explorethatrun()is toe2e run. It takes the project (cwd, andconfigas a file path or anE2EConfigvalue), the goal and its budgets (goal,target,agent,session,maxSteps,timeoutMs),interruptSignal, andonEvent, 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 noticese2e exploreprints before the run starts, such as the target it chose, arrive throughonEventasrunnotices. It sharesrun()'s reservation, so overlapping calls reject withRUNNER_BUSY, and each exploration that returns a report sends onee2e_run_completedevent withcommand: "runner.explore". -
#1312
9d884f5Thanks @kacperkapusciak! - The bundled agent skill now has the coding agent write a test while it walks the flow overe2e 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 anagent.actgoal, written in the words on screen before the agent walks its path, or, for an exact value, theTest line:the action printed. When the outcome shows, the agent locates it and adds theexpect. Thee2e mcpserver's instructions,e2e mcp --help, and the MCP server and coding agents docs say the same. -
#1372
0cc79f1Thanks @BiBoyang! - Report rejected recording writes as not-written. -
#1452
5b6fd01Thanks @samanyudas! - A multi-lineunique()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 asinvalid-entry. -
#1394
e9e8c50Thanks @kacperkapusciak! -e2e upgradeupdates the project's copies of the agent skill in.agents/skills/e2eand.claude/skills/e2eto the newly installed version, instead of telling you to rune2e init. It also updates a stale copy when the packages are already up to date, and--dry-runnames the copies it would update. -
#958
85e92f9Thanks @kuishou68! -uploadandsetInputFilesaccept a button that opens a file chooser, such as one over a hidden file input. The agent'suploadtool says so, so the model can pick the button it sees. -
#1429
3a841bbThanks @KrzysztofMoch! - Aweb()target takesapp.permissionswith'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.
@e2e-dev/decision@0.3.1
Patch Changes
-
#1382
3a8d6f1Thanks @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. -
#1437
94e6e76Thanks @okwasniewski! - Breaking: a secret fill no longer turns off screenshots for the model.agent.assert(..., { vision: 'only' }),vision: true, thescreenshotand point tools,e2e mcp, ande2e explorefindings 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.pixelsTaintedis removed, and a new step'svisionDegradedis neverPIXEL_TAINTED.
@e2e-dev/github@0.4.4
Patch Changes
- #1438
d181e26Thanks @okwasniewski! - Ane2e explorerun posts its own pull request comment. Before, a job that rane2e runand thene2e explorereplaced the test results comment with the exploration's. Test run comments keep their marker, so the comments already on open pull requests are still updated. An exploration comment an earlier version posted keeps the test run's marker: the nexte2e exploreposts a new comment beside it, and a test run in the same job takes it back. A--last-failedrerun no longer folds in an exploration's report that shares its output directory.
@e2e-dev/limrun@0.2.3
Patch Changes
-
#1420
512eed7Thanks @okwasniewski! - Bump agent-device to 0.21.33. -
#1458
450be72Thanks @kacperkapusciak! - Update agent-device and @agent-device/testmu to 0.21.35. An unscoped boot picks the newest of several shut-down devices, and on iOS, pressing return with nothing focused keeps the runner session. -
#1382
3a8d6f1Thanks @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. -
#1449
abb10cfThanks @bheemreddy-samsara! - Update agent-device to 0.21.34. Hosted WebDriver fills replace the confirmed focused field instead of appending, and an unsupported active-element lookup refuses before tapping. The pre-tap focus probe no longer reads the previous field's geometry.
@e2e-dev/mobile@0.12.5
Patch Changes
-
#1379
9e53c40Thanks @okwasniewski! - A mobile target withoutdeviceskips 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 atpreparenaming the holding sessions instead of failing on the first one. -
#1420
512eed7Thanks @okwasniewski! - Bump agent-device to 0.21.33. -
#1458
450be72Thanks @kacperkapusciak! - Update agent-device and @agent-device/testmu to 0.21.35. An unscoped boot picks the newest of several shut-down devices, and on iOS, pressing return with nothing focused keeps the runner session. -
#1382
3a8d6f1Thanks @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. -
#1380
c3c4238Thanks @okwasniewski! - Add the unstabledevice.unstable_raw, which may change or go away in any release: the test's agent-device session name, the device's id and name, and an agent-deviceclientbound to that session, for what thedevicemethods don't cover. Client calls aren't recorded as steps, replayed from the agent cache, or redacted. -
#1334
6c8e71fThanks @21prnv! -locator.filter()takeshasNot, which drops matches whose subtree resolves a locator, as in Playwright. Third-party engines must handle the newhasNotfield of afilterexpression. -
#1410
b309bcdThanks @bheemreddy-samsara! - Re-read mobile fill and clear targets before typing, including cached replay. Missing inputs, changed security state, or incomplete captures retry after a fresh observation. -
#1449
abb10cfThanks @bheemreddy-samsara! - Update agent-device to 0.21.34. Hosted WebDriver fills replace the confirmed focused field instead of appending, and an unsupported active-element lookup refuses before tapping. The pre-tap focus probe no longer reads the previous field's geometry. -
#1427
75ae2abThanks @kacperkapusciak! - Read the placeholder an empty Compose Multiplatform field folds into its iOS label as the field's placeholder, so an empty field no longer reads as holding the example text.fillandcleartype into an unclassified iOS element that holds a value or only text, the way a custom-drawn Compose Multiplatform or Flutter field can read. -
#1297
bdd7315Thanks @shamilovtim! - A test step on a control the keyboard moves now waits for it to land instead of tapping where it was. Moving-target refusals are retried when agent-device sent no input. -
#1307
64432f1Thanks @malopezr7! - Add@e2e-dev/mobile/leasefor device-provider packages:agentDeviceLeases()leases devices from a cloud agent-device runs sessions on, one isolated daemon per worker slot, and fails a release that agent-device can't confirm ended the session. -
#1413
eeacc8aThanks @samanyudas! - A mobile action on a node that a newer snapshot replaced finds the node again instead of acting on a different element. -
#1300
4af1aa8Thanks @shamilovtim! - Re-resolve a locator when its target changes during the transition wait, instead of tapping a stale native reference that may now identify another control. -
#1037
a481fd3Thanks @rudycelekli! -locator.filter()takeshasNotText, which drops matches whose text contains a string or matches a RegExp, as in Playwright. Third-party engines must handle the newhasNotTextfield of afilterexpression. -
#1421
46df78dThanks @okwasniewski! - Theopen_appagent tool returns to the target's app when called without an app, so the agent can recover after a system sheet or another app takes the foreground. Device errors no longer end with agent-device's CLI advice, such as "Run apps to discover...". -
#1421
46df78dThanks @okwasniewski! -press('Enter')on a field taps it first when it is not focused, as a character key already did. Before, Enter went to whatever held focus, and on iOS failed when the keyboard was down.
@e2e-dev/testmu@0.0.3
Patch Changes
-
#1420
512eed7Thanks @okwasniewski! - Bump agent-device to 0.21.33. -
#1458
450be72Thanks @kacperkapusciak! - Update agent-device and @agent-device/testmu to 0.21.35. An unscoped boot picks the newest of several shut-down devices, and on iOS, pressing return with nothing focused keeps the runner session. -
#1382
3a8d6f1Thanks @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. -
#1449
abb10cfThanks @bheemreddy-samsara! - Update agent-device to 0.21.34. Hosted WebDriver fills replace the confirmed focused field instead of appending, and an unsupported active-element lookup refuses before tapping. The pre-tap focus probe no longer reads the previous field's geometry. -
#1307
64432f1Thanks @malopezr7! - A device release now fails the run when agent-device can't confirm it ended the TestMu AI session: a forced daemon stop, a lost shutdown report, or a daemon that exited during the run. Needs@e2e-dev/mobile0.12.5 or later.
@e2e-dev/web@0.15.4
Patch Changes
-
#1334
6c8e71fThanks @21prnv! -locator.filter()takeshasNot, which drops matches whose subtree resolves a locator, as in Playwright. Third-party engines must handle the newhasNotfield of afilterexpression. -
#1048
7f4475cThanks @rudycelekli! - Addbrowser.waitForDownloads(trigger, { count })to collect several files from one trigger.waitForDownloadnow also bounds completion and saving by its timeout. -
#1037
a481fd3Thanks @rudycelekli! -locator.filter()takeshasNotText, which drops matches whose text contains a string or matches a RegExp, as in Playwright. Third-party engines must handle the newhasNotTextfield of afilterexpression. -
#1044
555e226Thanks @rudycelekli! - Addroute.fetch()to read the upstream response inside a route handler before deciding it. -
#1288
8ce2c65Thanks @DeryFerd! - Screenshots no longer change the inline style of editable elements to hide the text caret, so a server-rendered app hydrating while a capture runs no longer reports a React attribute mismatch on its inputs and textareas. A capture runs with the caret hidden in the frames it can reach, and the stylesheet is taken back from every frame afterwards. -
#1046
5879bfcThanks @rudycelekli! - Add selective cookie clearing with validated name, domain and path filters, preserving other browser storage. -
#958
85e92f9Thanks @kuishou68! -uploadandsetInputFilesaccept a button that opens a file chooser, such as one over a hidden file input. The agent'suploadtool says so, so the model can pick the button it sees. -
#1429
3a841bbThanks @KrzysztofMoch! - Aweb()target takesapp.permissionswith'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. -
#1448
cdf910aThanks @KrzysztofMoch! - Screenshots no longer fail withENGINE_FAILURE("closed-shadow root tracking is unavailable") on a page that replaces the globalWeakMap, as some polyfills do. On such pages, locators find elements inside closed shadow roots again, and CDP reconnects no longer fail.