Skip to content

Merge upstream v2.6.0 into tb-mods - #1

Merged
jochen-testingbot merged 152 commits into
tb-modsfrom
tb-mods-merge-v2.6.0
May 25, 2026
Merged

jochen-testingbot merged 152 commits into
tb-modsfrom
tb-mods-merge-v2.6.0

Conversation

@jochen-testingbot

Copy link
Copy Markdown
Collaborator

Summary

Resyncs tb-mods with mobile-dev-inc/Maestro v2.6.0 (151 upstream commits, 672 files, ~28k/34k lines). Adopts the new MaestroDriverLib Swift package and PermissionButtonFinder architecture while keeping all 12 fork-specific iOS / TestingBot mods intact.

Baseline snapshot for rollback: tag tb-mods-pre-v2.6.0-merge.

Conflict resolutions

File Resolution
gradle.properties, maestro-cli/gradle.properties Bumped to 2.6.0-tb (kept the -tb suffix pattern)
maestro-client/.../maestro-app.apk Took upstream's apk. We never modified maestro-android sources, so upstream's artifact is authoritative
XCTestDriverClient.kt Combined upstream's transportCall() socket-timeout latching with our XCTest request/response info logging. Our AppNotRunning vs AppCrash split is preserved (Errors.kt already had both types)
ScreenSizeHelper.swift Kept upstream's effectiveWidth/effectiveHeight landscape-aware dimension handling AND our defensive .faceUp/.faceDown/.unknown/@unknown default branches (avoids fatalError when device is flat)
SetPermissionsRequest.swift Switched to import MaestroDriverLib's PermissionValue (removed the local enum)
MaestroDriverLib/.../PermissionValue.swift Extended upstream's enum with our location-specific (.always, .inuse, .never) and photos-specific (.limited) cases
SystemPermissionHelper.swift Kept all 16 of our XCUIElement-based dialog handlers (permissionDialogs table, handleInterruption, autoAllowPermissions, alert/sheet/static-text matching). Adopted upstream's new (appHierarchy: AXElement, foregroundApp:) async signature. Upstream's PermissionButtonFinder is wired as a defensive last-resort fallback for notifications

Files NOT touched by the merge (intentional)

  • iOS physical-device runner zips under driver-iphoneos/ — kept the re-signed versions (upstream did not touch this path since fork point)
  • All 12 fork-specific commits — preserved on the tb-mods parent of the merge
  • CLAUDE.md, maestro-ios/README.md — fork-only files, kept

Action required before deploying

The iOS XCTest runner zips at driver-iphoneos/ reflect our pre-merge Swift code. For upstream's PermissionButtonFinder, transportCall, and MaestroDriverLib changes to take effect on physical devices, the runner must be rebuilt and re-signed:

./maestro-ios-xctest-runner/build-maestro-ios-runner.sh
./maestro-ios-xctest-runner/resign-maestro-ios-runner.sh /path/to/your.mobileprovision

Then deploy the new ~/.maestro/maestro-iphoneos-driver-build.zip to test machines.

Test plan

  • ./gradlew compileKotlin — passes
  • ./gradlew test — note: 5 XCTestDriverClientTest tests fail with BindException on port 22087 due to a pre-existing upstream issue with hardcoded test ports (reproduces on a clean v2.6.0 test file)
  • Build the iOS runner with build-maestro-ios-runner.sh
  • Re-sign with resign-maestro-ios-runner.sh
  • Smoke a physical-iOS flow against a registered UDID
  • Smoke a simulator flow
  • Verify --driver-host-port parallel-execution path still works (custom session-manager logic)
  • Verify location/photos/bluetooth permission handling (the extended PermissionValue cases)

Rollback

If anything goes wrong: git reset --hard tb-mods-pre-v2.6.0-merge

Co-Authored-By: Claude Opus 4.7 (1M context)

Fl0p and others added 30 commits February 5, 2026 10:42
* Show hierarchy and check-syntax commands in help

Remove the hidden flag from hierarchy and check-syntax commands
to make them visible in the maestro --help output. These are useful
commands that users should be aware of.

Changes:
- PrintHierarchyCommand: Remove hidden = true
- CheckSyntaxCommand: Remove hidden = true

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
* feat: Add speed parameter support for Flutter web scrolling

* Animated scroll for flutter.

* removed console.logs

* removed debugging code

* removed not required code.

* in-progress working on comments

* feat: Add coordinate-based swipe support for Flutter web

Refactored smooth scroll implementation to support both direction-based
and coordinate-based scrolling for Flutter web apps. The swipe(start, end)
method now correctly converts touch coordinates to wheel events for Flutter.

Changes:
- Add smoothScrollFlutterByDelta() for explicit delta values
- Update swipe() to detect Flutter and dispatch wheel events
- Keep smoothScrollFlutter() as a wrapper for direction-based scrolling
* added tags & properties in
- html report
- junit report

* fix: id if flow name has spaces
…inc#2970)

* fix ansi not printing complex sub flows

* pr fix

* pr fix

* pr fix

* add ut
fix: timeouts creating errors in compilation
* Bump datafaker to v2.5.3

* Bump graaljs to 24.2.0 (This is minor. I didn't want to tackle v25)

* Bump log4j to v2.25.3

* Bump selenium to 4.40.0
…onName predictable (mobile-dev-inc#3005)

* Fix e2e tests following changes in demo_app

* Make inputRandomPersonName predicatable

Right now it firstname and lastname, possibly preceded by a prefix and possibly followed by a suffix. I don't think that's the right behaviour for reliable tests. This solves that predictability that our e2e was expecting.
* Add web support for clearState

* Tidy tests

---------

Co-authored-by: Dan Caseley <dan@mobile.dev>
…ile-dev-inc#3034)

make assertScreenshot work without specifying the extension
* feat: added deviceLocale and depreciated localeutils

* feat: added devicelocale and used that

* removed duplicate platform

* update code to use new device configurations

* added deviceCatalogtest

* depreciate locale test

* fix platform import

* fix test

* fix: added cloud devices from response

* added sealed MaestroDeviceConfiguration

* add more fields to DeviceConfiguration

* fix: device catalog device

* fix: tests

* added web device launch

* moved platform inside resolve and avoided try/catch

* fix: pick recommended device to take defaults directly

* improved device list response.

* fix device catalog test

* use api for device

* feat: added deviceLocale and depreciated localeutils

* feat: added devicelocale and used that

* removed duplicate platform

* update code to use new device configurations

* added deviceCatalogtest

* depreciate locale test

* fix platform import

* fix test

* fix: added cloud devices from response

* added sealed MaestroDeviceConfiguration

* add more fields to DeviceConfiguration

* fix: device catalog device

* fix: tests

* added web device launch

* moved platform inside resolve and avoided try/catch

* fix: pick recommended device to take defaults directly

* improved device list response.

* fix device catalog test

* use api for device

* fix: added new flags to device start, and depreaciated --os-version

* removed the default list

* fix test

* reverted os flag discription

* fix: added deviceConfiguration to available device

* moved orientation and locale inside device

* change type name

* removed locale utils test

* fix tests

* moved getMacOSArchitecture to client

* fix: added image test + handled CPU_Architecture enum differently

* monor fix
mobile-dev-inc#3050)

feat: make all options for resolve to be passed explisitly + added os version
…nc#3055)

* typed request in devicecatalog.resolve

* fix: devicespec type
amanjeetsingh150 and others added 29 commits May 11, 2026 14:44
* Route xctest runner logs into the per-run debug output

Each iOS xctest log producer (XCRunnerCLIUtils.runXcTestWithoutBuild,
LocalSimulatorUtils.launchUITestRunner,
LocalIOSDeviceController.launchRunner) now takes the destination
directory as an explicit parameter at the call site. LocalXCTestInstaller
threads a required logsDir field through to all three producers, and
MaestroSessionManager points the installer at
<run-debug-dir>/xctest_runner_logs/ so the runner log lands inside the
same per-run bundle as commands.json, screenshots, and maestro.log.

Drops the static XCRunnerCLIUtils.logDirectory singleton, its
AppDirsFactory-backed rotation, and the unused clearLogs() method.
Consumers own the path and its lifecycle now -- the existing
TestDebugReporter.deleteOldFiles sweep already covers the new location.

* ci: drop redundant xctest_runner_logs artifact uploads

The xctest runner log now lands in
~/.maestro/tests/<run>/xctest_runner_logs/ as part of the per-run debug
output, which is already uploaded under the maestro-root-dir-ios
artifact (path: ~/.maestro). The dedicated steps that re-uploaded
~/Library/Logs/maestro/xctest_runner_logs are no longer needed.

Leaves the disabled test-ios-xctest-runner job's upload alone -- that
job runs the runner directly via test-maestro-ios-runner.sh, not through
LocalXCTestInstaller, so its log path is unrelated to this change.

* Place xctest runner log at the run-dir root

Hand the run debug directory directly to LocalXCTestInstaller so the
xctest_runner_<date>.log file lands alongside maestro.log,
commands-*.json, and screenshots, instead of in a nested
xctest_runner_logs/ subdirectory.

* ci: read xctest log from per-run debug output in preflight verifier

The script globbed ~/Library/Logs/maestro/xctest_runner_logs/*.log,
which is empty now that the iOS driver writes its log into the per-run
debug-output dir (~/.maestro/tests/<run>/xctest_runner_*.log). Update
the glob to walk the new tree (honoring XDG_STATE_HOME, matching
EnvUtils.xdgStateHome) and fail with a clear message if no log matched.
…le-dev-inc#3264)

* fix(env): fail fast with helpful message when env value is null

A YAML entry like `KEY:` (no value) deserializes to null inside
Map<String, String> via Kotlin generic erasure and later crashes
DefineVariablesCommand.evaluateScripts with an opaque Kotlin
null-intrinsics error. Validate at withEnv and surface the offending
key(s) instead.

* fix(env): surface bad key name in parser error summary

Use a typed EnvVariableMissingValueError so MaestroFlowParser.wrapException
can render "Environment variable has no value: KEY at file:line:col"
instead of the generic "Parsing Failed" summary.
…ed (mobile-dev-inc#3262)

warn when deprecated --android-api-level/--ios-version are used

Print a runtime warning pointing users to --device-os when either flag
is passed to maestro cloud. Previously the flags were marked deprecated
and hidden from help, but used silently with no feedback.
… tracking (mobile-dev-inc#3138)

feat: enable parallel iOS simulator execution

iOS simulators share the host's localhost, causing port collisions when
multiple Maestro processes target different sims simultaneously. Session
tracking was per-platform, so two processes on different devices would
interfere with each other's sessions.

Changes:
- Per-device session tracking: SessionStore keys are now
  "{platform}_{deviceId}_{sessionId}" instead of "{platform}_{sessionId}"
- Add --driver-host-port CLI flag for explicit XCTest server port
- Auto-select available ports with isPortAvailable() check
- Refactor SessionStore from singleton to injectable class (DI)
- Add shouldCloseSession(platform, deviceId) for per-device shutdown
  instead of global activeSessions().isEmpty()
- Add cross-process file locking to KeyValueStore (~/.maestro/sessions)
- Append PID to debug log directory to prevent parallel race
- Enable useJUnitPlatform() in maestro-cli (was missing)
- Add SessionStoreTest with 8 tests covering isolation and lifecycle

Verified: 3 iOS simulators + Android emulator running simultaneously,
all passing. Both --driver-host-port (explicit) and auto-port-selection
work correctly.

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-authored-by: Proksh Luthra <35415752+proksh@users.noreply.github.com>
…-dev-inc#3235)

* feat: make JUnit XML id and classname attributes configurable

Add optional `id` and `classname` keys to the flow YAML config section.
When present, they override the corresponding JUnit XML <testcase>
attributes; when omitted, behavior is unchanged (all default to flow name).

Closes mobile-dev-inc#3151

* refactor: move id/classname into properties as junitId/junitClassname

JUnit-specific test case attributes (id, classname) are now configured
via the existing properties map instead of top-level config fields.
This avoids confusing users of HTML reports with JUnit-only concepts
and makes the purpose of these values explicit through naming.

Properties values are now script-evaluated, enabling ${...} expressions.
The JUnit reporter extracts junitId/junitClassname for XML attributes
and filters them from <property> elements to avoid redundancy.

---------

Co-authored-by: Balint Dezsi <balint.dezsi@sap.com>
Co-authored-by: Dan Caseley <dan@mobile.dev>
…state (mobile-dev-inc#3283)

The maestro-driver-ios.zip and maestro-driver-iosUITests-Runner.zip
under driver-iPhoneSimulator/Debug-iphonesimulator/ were rebuilt and
committed unintentionally as part of mobile-dev-inc#3138 (db366b0). That PR's logic
changes don't require updated driver binaries, so restoring the zips
to their previous content.

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* Add links to Maestro Cloud run and App binary ID to HTML report.

* Restrict cloud footer metadata to cloud reports

Remove cloud metadata merging from local `maestro test` summary aggregation and only render the HTML footer when a cloud upload URL is present.

Co-authored-by: Cursor <cursoragent@cursor.com>

---------

Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Dan Caseley <dan@mobile.dev>
Signed-off-by: Ash Wu <hsatac@gmail.com>
* Make hierarchy command JSON only

* Backpeddle on `--output` - its not needed

* Revert change to PickDeviceView - it's overkill and no longer used

* Don't duplicate options. Fix this later.
* MCP visualizer

Browser-based visualizer for MCP-driven sessions: a session-aware
McpVisualizerDriver wraps the underlying driver and streams screen
state, commands, and run progress to a Ktor server. The Vite React UI
renders the device, command log, and hardware controls, and forwards
touch/key/button input back to the simulator. Includes mock-backend
support for offline UI iteration.

* Address PR review on MCP visualizer

- Make /api/device/input an opaque proxy to simulator-server's stdin so
  the server doesn't need to know the protocol grammar; reject newlines
  and oversize bodies. The TS client owns the format.
- Reject /api/* requests whose Sec-Fetch-Site isn't same-origin/none/
  absent so a drive-by browser page can't silently drive the simulator
  via no-cors fetches against 127.0.0.1.
- Coalesce concurrent DeviceStream.start() callers behind a Mutex so the
  SSE MaestroConnected handler and POST /api/device/start don't spawn
  duplicate simulator-servers. A separate input Mutex serializes stdin
  writes; a JVM shutdown hook is the safety net for any process that
  escapes a start/close race.
- Fire the visualizer-link hint only on the first tool response of a
  session instead of every response, to avoid bloating outputs and
  fighting the agent's task.
… UNAVAILABLE (mobile-dev-inc#3290)

The Android driver opens an in-process gRPC channel against the maestro-driver
app on the device. Today the channel has no retry policy, so a transient
"No route to host" or stream-reset on the pod-local network bubbles up as a
StatusRuntimeException and fails the entire job — the backend then re-queues
the run on a different worker, costing ~30-60s per occurrence.

Empirical 14d sample from production worker logs: 109 INFRA_ERRORs from
StatusRuntimeException on this channel, 85% of them in callViewHierarchy
(an idempotent read).

Enabling channel-level retry with a bounded service config covers all calls
on the channel without per-callsite changes:

- Transparent retries fire automatically when the RPC never reached the server
  (e.g. couldn't open the connection). Safe even for non-idempotent calls
  because the server never saw the request.
- Configured retries fire on UNAVAILABLE returned by the server, bounded by
  maxAttempts=4 with 0.5s→2s exponential backoff.

If retries are exhausted, behaviour matches today: the StatusRuntimeException
bubbles up and the existing INFRA_ERROR classification re-queues the job.

Tests use an in-process gRPC server to verify the policy fires for
UNAVAILABLE, stops at maxAttempts, and leaves non-retryable codes alone.
…bled for XCTest (mobile-dev-inc#3291)

feat(http-client): opt-in retry on transient 5xx and IOException

Wires an OkHttp application interceptor into HttpClient.build that retries
calls on transport-level IOException or HTTP 5xx responses, up to a bounded
maxAttempts with exponential backoff. 4xx responses pass through unchanged
because they are deterministic client errors.

Enabled for XCTestDriverClient by default: empirical 14d sample from production
worker logs shows 117 INFRA_ERRORs from XCUITestServerError (5xx with
"kAXErrorInvalidUIElement" and similar transient state from the XCTest
runner while the UI is updating). With retry on the HTTP client, the rest of
the runner observes a single logical call and the job no longer needs to be
re-queued on a different worker just to absorb a transient hiccup.

If retries are exhausted, the last response or IOException is surfaced
unchanged, so existing classification (AppCrash, OperationTimeout,
UnknownFailure → INFRA_ERROR) still fires as a safety net.

Retry events are recorded on the existing `http.client.retries` micrometer
counter (kind=http_5xx or exception class name) so we can observe how often
the new path is firing post-deploy.

Other callers of HttpClient.build default to retryOnTransientFailure = false
to keep behaviour identical for them.

Test plan:
- New TransientFailureRetryInterceptorTest covers: 5xx-then-200 succeeds,
  5xx exhausts at maxAttempts, 4xx is not retried, IOException retries
  bounded, IOException recovers if next attempt succeeds, retry disabled is
  no-op, 2xx makes exactly one call.
- Updated XCTestDriverClientTest's AppCrash parameterized case to enqueue
  three 500 responses (matches the new default maxAttempts) so the
  classification assertion still passes.
…obile-dev-inc#3284)

fix(web): catch Throwable around DevTools setup so CDP service errors don't escape

WebDriver.open() wrapped the optional DevTools setup in `catch (e: Exception)`,
but Selenium's CDP version-selection throws `ServiceConfigurationError` (a
subclass of `Error`, not `Exception`) when no matching `CdpInfo` provider is on
the classpath. The error escaped and surfaced as an INFRA_ERROR on the worker
side. Widen the catch to `Throwable` — the block already documents that any
failure here should be swallowed because DevTools support is optional.
…le-dev-inc#3278)

* fix(orchestra): consistent sourceDescription for runScript

runScript previously stored the resolved absolute script path in
sourceDescription, while runFlow / retry store the literal YAML `file:`
field. The path was doing double duty: display label and the anchor
GraalJsEngine.setCurrentScriptDir uses to resolve relative file lookups
(e.g. multipartForm filePath). That made the user-facing description
drift with the workspace root (`Run /var/folders/.../workspace.../foo.js`
when YAML is read from an extracted temp dir, vs `Run /foo.js` from a
zip filesystem).

Split the two concerns:

- RunScriptCommand gains scriptDir (absolute parent of the script,
  populated by YamlFluentCommand). The JS engine uses it directly to
  anchor relative file lookups; sourceName / sourceDescription is no
  longer derived from a path.
- sourceDescription now mirrors runScript.file (the literal YAML field),
  matching runFlow and retry. originalDescription is plain
  "Run $sourceDescription", same shape as the other two.

scriptDir defaults to null so commands assembled outside the YAML
parser keep working, and FAIL_ON_UNKNOWN_PROPERTIES=false absorbs the
new field across rolling deploys.

* fix(orchestra): consistent sourceDescription for runScript

runScript previously stored the resolved absolute script path in
sourceDescription, while runFlow / retry store the literal YAML `file:`
field. The path was doing double duty: display label and the anchor
GraalJsEngine.setCurrentScriptDir uses to resolve relative file lookups
(e.g. multipartForm filePath). That made the user-facing description
drift with the workspace root (`Run /var/folders/.../workspace.../foo.js`
when YAML is read from an extracted temp dir, vs `Run /foo.js` from a
zip filesystem).

Split the two concerns:

- RunScriptCommand gains scriptDir (absolute parent of the script,
  populated by YamlFluentCommand). The JS engine uses it directly to
  anchor relative file lookups; sourceName / sourceDescription is no
  longer derived from a path.
- sourceDescription now mirrors runScript.file (the literal YAML field),
  matching runFlow and retry. originalDescription is plain
  "Run $sourceDescription", same shape as the other two.

scriptDir defaults to null so commands assembled outside the YAML
parser keep working, and FAIL_ON_UNKNOWN_PROPERTIES=false absorbs the
new field across rolling deploys.

* updated comment
…-dev-inc#3285)

* chore(deps): bump selenium to 4.43.0 and CDP bindings to v145

Selenium 4.40.0 bundled CDP support up to Chrome 144, so any worker on
Chrome 145+ failed to load a matching CdpInfo provider and threw
ServiceConfigurationError at DevTools setup. Selenium 4.43.0 bundles
v145/v146/v147 CDP bindings, restoring support for current Chrome stable.

- libs.versions.toml: selenium + selenium-devtools → 4.43.0, artifact
  suffix v142 → v145 (the lowest version bundled by 4.43.0).
- Updated three import sites (WebDriver, CdpWebDriver, WebScreenRecorder)
  from v144 to v145 packages.

Companion to mobile-dev-inc#3284 (which makes the unmatched-provider case non-fatal
going forward).

* chore(deps): bump CDP bindings to v147

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…obile-dev-inc#3275)

* CI: Fix gradle config to prevent android sentinel failures or iOS dependency failures

* Add CI job to test building on all 3 OSes, and test version output

Disable Maestro CLI analytics to not pollute the test

Fix Windows path setting in GHA

Merge e2e testing together

Make CLI tests local-first

* Extra test, improved usage and formatting

* Fix Windows tests that output to stderr

* Additional tests, deal with ANSI in bash
…#3297)

* fix(mcp-visualizer): support physical Android devices

The visualizer always launched simulator-server with the `android`
subcommand (emulator-only), which panicked on physical Android serials
with "emulator with id <serial> should be running". And even with the
right subcommand it could not find its screen-sharing agent resources
because installSimulatorServer only unpacked the binary, not its
sibling `resources/` tree.

- Plumb a typed StreamDeviceType (android | android_device | ios)
  through DeviceStreamTarget, MaestroConnected, and McpMaestroSession
  so the launch site picks the correct subcommand based on
  Device.DeviceType at the source. Browser carries the field opaquely
  in /api/device/start. Invalid combinations (BROWSER, real iOS) are
  unrepresentable.
- Unpack the entire deps/simulator-server/<platform>/ classpath
  subtree into ~/.maestro/deps/ via a new Unpacker.unpackTree, so
  screen-sharing-agent.jar + per-ABI .so files land next to the
  binary. Uses NIO FileSystems.newFileSystem to share one walker
  across jar and filesystem resources. Verifies all declared
  executableEntries were present so missing-binary packaging bugs
  fail loudly.

* chore(deps): bump simulator-server to f4b4f058

Required by the visualizer fix above: this release adds the
`android_device` subcommand for physical Android and ships the
screen-sharing-agent.jar + per-ABI .so files alongside the binary.

* fix(mcp-visualizer): flush SSE headers on stream open

Ktor 2.3.x's respondBytesWriter buffers the response until the first
write. /api/events/stream has no initial payload, so the browser sees
ERR_EMPTY_RESPONSE on the EventSource and never reconnects, hiding
the entire command log and tap-animation pipeline. The bug has been
latent since the visualizer was introduced in mobile-dev-inc#3288 — exposed now
that physical Android lets us exercise the events path.

Write an SSE comment line on stream open — the standard pattern to
flush headers without producing a data event the client will
misinterpret.

* refactor(mcp): consolidate stream-device-type mapping in createSession

Per pedro18x review: the EMULATOR/REAL→ANDROID/ANDROID_DEVICE mapping
lived in two places (createAndroidSession and StreamDeviceType.forConnected)
and could drift if a new DeviceType lands in one without the other.

Resolve the stream type once at createSession's entry, bail with a
clear unsupported message if forConnected returns null, and pass the
resolved value into the per-platform builders. forConnected becomes
the sole authority. The "real iOS unsupported" check is now subsumed
by the same null branch.
* Fix web date input text entry

* Add web date input e2e flow

---------

Co-authored-by: Leonardo Viveiros <leo@wooshair.com>
Co-authored-by: Dan Caseley <dan@mobile.dev>
…bile-dev-inc#3300)

* fix(android-webview): preserve hostname in DummyDns so devtools augmentation works

`androidWebViewHierarchy: devtools` was silently returning empty webview
hierarchies. OkHttp resolves the request URL via DummyDns, then builds
`InetSocketAddress(InetAddress, port)` from the result. That constructor
copies `addr.getHostName()` into the holder — for `getLoopbackAddress()`
this is "localhost", erasing the webview socket name. `AdbSocket.connect()`
then reads `endpoint.hostString` and calls
`dadb.open("localabstract:localhost")` instead of the real
`localabstract:webview_devtools_remote_*`. The IOException is swallowed by
the catch blocks in `getWebViewInfos`/`getWebViewTreeNodes`, so the
hierarchy is never augmented and webview-only elements can't be found.

Switch DummyDns to `InetAddress.getByAddress(hostname, loopback bytes)` so
the original hostname round-trips through to `endpoint.hostString`.

Regression in mobile-dev-inc#3171 (Apr 14) — the prior `.dadb(dadb)` HTTP client routed
`localabstract.<socket>.adb` hosts via its own resolver, which preserved
the socket name; the new SocketFactory + DummyDns pair did not.

Adds a unit test that exercises the OkHttp address-constructor path.

* fix: added e2e test
…bile-dev-inc#3299)

* chore(cli): remove bundled Maestro Studio in favor of desktop app

The `studio` command no longer launches the bundled web Studio. It now
prints a notice with the OS-appropriate download link for the Maestro
Studio desktop app (.dmg / .exe / .AppImage). The `:maestro-studio:server`
Gradle dependency is dropped from the CLI so the Studio server is no
longer shipped inside the CLI artifact. The deprecation message in
`query` is updated to point users at the same download links.

* chore: delete maestro-studio module

Follow-up to 3f9ba8f, which removed the CLI's bundled Studio. The
:maestro-studio:server and :maestro-studio:web Gradle modules are no
longer referenced by any build target, so drop the directory along
with the orphan settings.gradle.kts includes and the module table
entry in CONTRIBUTING.md.
…dev-inc#3298)

* refactor(mcp-visualizer): extract McpVisualizerOrchestra factory

Move the Orchestra-with-visualizer-callbacks wiring out of RunTool into
its own file so the run path stays focused on flow execution.

* refactor(mcp-visualizer): use identity-keyed command IDs

Switch callIds from `$flowId:$index` to a process-scoped counter keyed
by MaestroCommand instance identity. This lets pending and runtime
events upsert into the same visualizer row by sharing a callId for the
same command reference — groundwork for publishing the full set of
commands as pending up front before execution begins.

* feat(mcp-visualizer): show pending commands up front via flow snapshots

Replace the per-command event stream with a `maestro.flow_state` snapshot
of the full command list, published on every change. The backend seeds
every top-level command with a `pending` status at `onFlowStart` so the
visualizer can render the whole flow before execution begins, then
re-publishes a fresh snapshot on each status transition.

If the flow aborts (onCommandFailed rethrows), trailing pending rows are
swept to `skipped` so they don't sit unresolved.

The frontend becomes a pure projection: a single reducer merges incoming
snapshots into its row map by commandId, with prior runs' rows
preserved automatically because their commandIds never reappear in
later snapshots.

* refactor(mcp-visualizer): move nested-command rendering decision to frontend

Backend now publishes every command in the tree — top-level and nested
— with a depth field, instead of filtering nested ones via a depth-stack
in onCommandStart/Complete. Seeding walks CompositeCommand.subCommands()
recursively at flow start. Status callbacks just flip statuses by
identity; no more outer-only gating.

Frontend filters to depth === 0 at render time, so behavior is identical
to before: only top-level commands appear, each treated atomically.
Future nested-rendering work (indent, collapsible tree, etc.) lives
entirely in the frontend.

* chore(mcp-visualizer): add design sandbox at /?demo

Extract a VisualizerLayout component (commands panel + device pane,
including bg-tint and ring driven by row statuses) so both the live App
and a static demo can render the same chrome with their own data. The
demo lives in src/demo.tsx and is loaded dynamically only when
`import.meta.env.DEV && ?demo` is set, so prod builds drop it via
dead-code elimination.

Visit http://127.0.0.1:5173/?demo (after `npm run dev`) to iterate on
panel rendering against a fixed set of rows covering every status. To
remove the sandbox, delete src/demo.tsx and the dev-only branch at the
bottom of main.tsx.

* feat(mcp-visualizer): dark mode via prefers-color-scheme

Add `dark:` Tailwind variants throughout the visualizer panel, device
frame, no-stream card, and hardware rail. Tailwind v4's `dark:`
variant follows `prefers-color-scheme`, which the browser ties to the
OS color preference.

Note: this is OS-preference-based, not IDE-theme-based. VS Code/Cursor
Simple Browser does not pass IDE theme through to embedded pages —
that's only available to first-party extension webviews via
`vscode-light`/`vscode-dark` body classes and `var(--vscode-*)` CSS
variables. For users who sync OS and IDE themes (the common case), the
visualizer effectively follows the IDE; for others it will track the
OS preference.

* chore(mcp-visualizer): add @types/react and @types/react-dom

* style(mcp-visualizer): float command panel and redesign row visuals

Panel sits as a floating, translucent card centered next to the device
pane and stretches to the device's height. Rows render with circle status
icons (glyph inside), a faint vertical connector tinted by status, and a
polished sky highlight for the running row.

* fix(mcp-visualizer): publish SSE events on a single-threaded dispatcher

Snapshot updates fan out through scope.launch on Dispatchers.IO, which
lets two rapid publishes run on different threads and reach the SSE
client's write mutex out of order. The frontend's snapshot merge is a
blind overwrite by commandId, so an older snapshot arriving last
regresses statuses (PENDING flickers in after STARTED, etc.).

Route event delivery through Dispatchers.IO.limitedParallelism(1) so
snapshots are written in submission order. Split deviceStream.start off
to its own launch on the unconstrained scope so its 30s blocking call
doesn't head-of-line block subsequent snapshots.

* feat(mcp-visualizer): syntax highlight YAML rows and quiet the scrollbar

YAML in each command row now renders with light syntax coloring (indigo
keys, emerald strings, violet numbers, muted punctuation). Falls back to
plain text when the line doesn't match the expected key:value shape.

Also pins the list scrollbar to a low-opacity neutral across default,
hover, and active states so the browser can't swap it for the system
highlight color.

* style(mcp-visualizer): make running row read at a glance

Opaque sky background (sky-200 light, sky-900 dark) with a thick ring
and saturated shadow halo so the active command is unmistakable. Skip
the connector line inside the running row body — it would otherwise
draw a thin pending-colored stripe across the highlight — but keep the
short segment in the gap below so the flow is still visually linked.

* fix(mcp-visualizer): scroll the running command into view, not the list bottom

Pending rows now extend past the running one (seeded up front via the
flow snapshot), so scrolling to scrollHeight buried what's actually
executing. Use findLast to locate the running row and scrollIntoView
with block: nearest so the panel only moves when needed.

* style(mcp-visualizer): tone down running glow and refine YAML typography

Light-mode running row uses a softer shadow, YAML keys render semibold
while values render light for clearer hierarchy, and the device ring
uses brighter sky-500 in light mode.

* style(mcp-visualizer): self-host Inter and JetBrains Mono via fontsource

Drops the dependency on whatever fonts the host happens to ship so the
visualizer looks the same on macOS, Linux, and Windows.

* fix(mcp-visualizer): render optional-missing commands as skipped, not warned

Orchestra reports an optional command whose target isn't on screen as
WARNED, but to the user that's an expected miss, not a problem worth
an amber exclamation. Map it to SKIPPED so it reads as a quiet dash.

* style(mcp-visualizer): let the command panel grow with the viewport

The command panel was pinned to 320px regardless of available space.
Now it grows via flex-1 up to a 64rem group cap, keeps a 320px floor
so it doesn't collapse on narrow screens, and the [panel + device]
group stays centered (including when the panel is collapsed to its
icon strip).

* style(mcp-visualizer): add a bit of breathing room on the command panel's left edge

* chore(mcp-visualizer): bump react and react-dom to 19 to match @types

The package was pinned to react@^18.3 while the type packages were
already on @types/react@^19, which would bite anyone running tsc
even though vite build skips type-checking. Aligning runtime with
the types is the safer direction since 19 is current.
mobile-dev-inc#3301)

* refactor(mcp-viewer): rename MCP Visualizer to Maestro Viewer

Renames the mcp-visualizer module, Kotlin package, classes, CLI flags
(--no-viewer, --viewer-port), and stderr ready marker. User-facing
strings now refer to "Maestro Viewer".

* feat(mcp-viewer): add open_maestro_viewer tool

Exposes the running Maestro Viewer URL to MCP agents. The response
tells the agent to surface the URL to the user rather than shelling
out to a launcher, since most host environments (Claude Code, Cursor,
Codex) have a preferred embedded preview surface.

* test(mcp): cover open_maestro_viewer in MCP tool tests

Adds open_maestro_viewer to expected_tool_list so run_mcp_tool_tests.sh
no longer fails its discovery check, and asserts the isError path by
running the MCP server with --no-viewer.
…inc#3179)

* fix: throw correct test error in case of invalid timeout

* fix: updated all timeout to use parseTimeoutMs and added test

* updated error message for non number timeout
…mobile-dev-inc#3303)

* feat(analytics): attach sanitized commandStringUsed to PostHog events

Adds a CommandArgsSanitizer that produces a redacted representation of
the CLI invocation (sensitive flag values replaced with <REDACTED>,
positional file paths dropped) and exposes it as a super-property on
the Analytics singleton. The merge in convertEventToEventData attaches
it to every event the CLI emits, so future flag-level deprecation
questions become a PostHog query instead of per-flag instrumentation.

Sensitive flags: --api-key/--apiKey, -e/--env, --name,
--app-binary-id/--appBinaryId. Both alias casings included.

* feat(analytics): narrow sanitizer to credential-only redaction

Drop --name and --app-binary-id/--appBinaryId from SENSITIVE_FLAGS.
Neither is a credential: --app-binary-id is an opaque server-generated
ID (and is already sent explicitly as the app_binary_id event property
where needed), and --name is a human-readable upload label. The
sanitizer now only redacts actual secrets (--api-key/--apiKey and
-e/--env).
…3305)

Vision models reject images whose longest side exceeds 2000px, which breaks the take_screenshot MCP tool mid-session. Downscale screenshots with bicubic resampling so the longest side stays within that limit, encode JPEG at 0.9 quality, guard against decode failures, and surface the exception type in error messages.

The 2000px cap is a hardcoded constant rather than a tool parameter: vision models share the same limit and the calling agent has no signal to vary it per call.

Fixes mobile-dev-inc#2952
* chore: prepare release 2.6.0

* chore: add screenshot downscale entry to 2.6.0 changelog

* Update CHANGELOG.md

Co-authored-by: Dan Caseley <dan@mobile.dev>

* Update CHANGELOG.md

Co-authored-by: Dan Caseley <dan@mobile.dev>

* Update CHANGELOG.md

Co-authored-by: Dan Caseley <dan@mobile.dev>

* Update CHANGELOG.md

Co-authored-by: Dan Caseley <dan@mobile.dev>

* Update CHANGELOG.md

Co-authored-by: Dan Caseley <dan@mobile.dev>

* Update CHANGELOG.md

Co-authored-by: Dan Caseley <dan@mobile.dev>

* Update CHANGELOG.md

Co-authored-by: Dan Caseley <dan@mobile.dev>

* Update CHANGELOG.md

Co-authored-by: Dan Caseley <dan@mobile.dev>

* Update CHANGELOG.md

Co-authored-by: Dan Caseley <dan@mobile.dev>

* Update CHANGELOG.md

Co-authored-by: Dan Caseley <dan@mobile.dev>

* Update CHANGELOG.md

Co-authored-by: Dan Caseley <dan@mobile.dev>

* Update CHANGELOG.md

Co-authored-by: Dan Caseley <dan@mobile.dev>

* Update CHANGELOG.md

Co-authored-by: Dan Caseley <dan@mobile.dev>

* Update CHANGELOG.md

Co-authored-by: Dan Caseley <dan@mobile.dev>

* Update CHANGELOG.md

Co-authored-by: Dan Caseley <dan@mobile.dev>

* Update CHANGELOG.md

Co-authored-by: Dan Caseley <dan@mobile.dev>

* Update CHANGELOG.md

Co-authored-by: Dan Caseley <dan@mobile.dev>

* Update CHANGELOG.md

Co-authored-by: Dan Caseley <dan@mobile.dev>

* Fix formatting error that Dan introduced!

---------

Co-authored-by: Dan Caseley <dan@mobile.dev>
Reconciles tb-mods with mobile-dev-inc/Maestro v2.6.0. Conflict
resolution notes:

- gradle.properties / maestro-cli/gradle.properties: bumped to 2.6.0-tb
  to follow the established -tb suffix convention.
- maestro-client/.../maestro-app.apk: took upstream's apk; we never
  modified the maestro-android sources so the upstream-built artifact
  is authoritative.
- XCTestDriverClient.kt: combined upstream's transportCall() wrapping
  for socket timeouts with our existing XCTest request/response info
  logging. Our AppNotRunning vs AppCrash split in processResponse is
  preserved.
- ScreenSizeHelper.swift: kept upstream's effectiveWidth/effectiveHeight
  for landscape-aware physicalScreenSize handling AND our defensive
  branches for .faceUp/.faceDown/.unknown/@unknown default (avoids
  fatalError when device is flat).
- SetPermissionsRequest.swift: switched to importing PermissionValue
  from MaestroDriverLib instead of redefining it locally. Extended the
  upstream enum (in MaestroDriverLib/Models/PermissionValue.swift) with
  our location-specific (.always/.inuse/.never) and photos-specific
  (.limited) cases.
- SystemPermissionHelper.swift: kept all 16 of our XCUIElement-based
  permission dialog handlers (permissionDialogs table, handleInterruption,
  autoAllowPermissions, fallback alert/sheet/static-text matching) and
  adopted upstream's new (appHierarchy: AXElement, foregroundApp:) async
  signature. Upstream's PermissionButtonFinder is wired as a defensive
  last-resort fallback for notifications.

iOS physical-device runner zips under driver-iphoneos/ kept (re-signed
versions). The iOS XCTest runner must be rebuilt and re-signed for the
upstream Swift changes (PermissionButtonFinder integration, transport
fixes) to take effect on device. See maestro-ios/README.md and
resign-maestro-ios-runner.sh.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
@jochen-testingbot
jochen-testingbot merged commit 41db7b9 into tb-mods May 25, 2026
13 of 16 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.