Repository navigation
Releases
1.0.2
Changed
- A stable tag publishes
server.jsonto the MCP registry.publish.ymlhas anmcp-registryjob after the GitHub Release that signs in through GitHub OIDC and publishesio.github.anilcancakir/tapir, so the registry listing follows each release instead of a manual publish. (.github/workflows/publish.yml)
Fixed
-
tapir says where
--cdp-portgives uptapir_evaluate.--cdp-portmakes dowel run flutter on theweb-serverdevice, which has no VM Serviceevaluate, yet the quickstart, the skill and the README recommended the flag with no caveat, so an agent that followed them calledtapir_evaluateand got an SDK error. Thetapir_evaluatedescriptor, skill rule 16 and its MCP reference, the MCP tool reference, the quickstart, the README andllms.txtnow say it, andtapir:doctorprints anEvaluate: unavailableinfo line for any session that records acdpPort. The error text itself is dowel's (1.0.4). Fixes #91. (lib/src/tapir_dowel_provider.dart,lib/src/commands/tapir_doctor_command.dart,skills/tapir/,doc/) -
tapir_dismiss_modals,tapir_reset_overlaysandtapir_navigateno longer freeze the app over a modal that guards its exit. go_router refuses the pop of aGoRoutewithonExitand decides in a microtask, so the synchronouspopUntilre-popped the same route forever and the page stopped answering until a reload. Modals are now popped one at a time; a guarded one is popped once and given three frames to leave. One that leaves is counted and dismissal carries on; one whose guard refuses stays open, is not counted and ends dismissal on its Navigator. A call that arrives while another dismissal is waiting on a guarded modal joins it and gets the same count, so the modal is popped once.tapir_navigatethen reaches its adapter as before, and its URL check still decidesnavigated. (lib/src/extensions/ext_modal_router.dart,lib/src/tapir_dowel_provider.dart,skills/tapir/,doc/) -
tapir:doctoraccepts theMagicDevtools.installPost()wrapper as Magic wiring. An app onmagic_devtools1.x that callsMagicDevtools.installPre()andMagicDevtools.installPost()gotMagic detected but MagicTapirIntegration missing; wire it with tapir:install, because the Magic-init check looked only forMagicTapirIntegration.install, whichinstallPostcalls itself. EitherMagicTapirIntegration.install()orMagicDevtools.installPost()now reads asMagic-stack detected, integration wired;installPre()alone, or neither call, still gets the advice. (lib/src/commands/tapir_doctor_command.dart,doc/commands/tapir-doctor.md) -
tapir_set_checkboxwaits for a control that flips after a round trip. A switch the app updates only once its PATCH answers still showed the old state when the handler read it back right after the tap, so the call answeredverified: falseand the CLI exited 1 although the toggle worked and persisted. While the read differs from the requested value it is now repeated every 50 ms for up to 1.5 s, with no second tap; a control that flips at once answers at once, and one that ignores the tap still readsverified: falseafter the window. The response shape is unchanged. (lib/src/extensions/ext_checkbox.dart,lib/src/tapir_dowel_provider.dart,skills/tapir/,doc/mcp/tool-reference.md) -
tapir:installleaves an app wired throughMagicDevtools.installPre()andinstallPost()alone. The installer looked only forTapirPlugin.install()andMagicTapirIntegration.install(), which themagic_devtoolswrapper calls itself, so it added a secondTapirPluginblock, aMagicTapirIntegrationblock, an unusedpackage:tapir/tapir.dartimport and aWidgetsFlutterBinding.ensureInitialized()line (an app whose zone wrapper owns the binding did not want one) to an app that was already wired.installPre()now stands for the tapir block, its import and the binding line,installPost()for the Magic block; an app that calls onlyinstallPre()still gets the Magic block. The wrapper counts only when the resolvedmagic_devtoolsshipslib/tapir.dart: a release that predates tapir has the same wrapper names but installs the Dusk integrations, so the installer still wires tapir and warns to upgrade.tapir:doctorapplies the same rule toinstallPost(). (lib/src/commands/tapir_install_command.dart,lib/src/commands/tapir_doctor_command.dart,doc/commands/tapir-install.md)
1.0.1
Changed
- The pub.dev topic
mcp-serverismcp. Only tapir and crumbs used the first; the second is shared by many MCP packages, so the listing is easier to find. (pubspec.yaml)
Fixed
tapir_wait_for_network_idlerefuses apollIntervalMsbelow 100 as aninvalid_param. The 100 ms floor was anassert, which a profile build strips: therepollIntervalMs: 0never advanced the elapsed counter, so the wait never ended. The handler checks the floor itself now and answers-32602. (lib/src/extensions/ext_wait_find.dart)- The network-idle wait is described as the code behaves. The tool description, the
tapir:wait_for_network_idlehelp, the tool reference, the skill andllms.txtsay that the count is zero unless an adapter or a reader the app installs reports one:pendingHttpCountReaderis a public function an app can replace with its own counter, crumbs is one option and not a requirement, andMagicHttpFacadeAdapter.pendingCountcomes frommagic_devtools, not from crumbs. The call matches afterceil(idleMs / pollIntervalMs)consecutive zero samples (the first sample counts, so the defaults match after about 400 ms and a window of one poll interval or less matches on the first sample; a request that starts and finishes between two samples is not seen), not on a confirmed contiguous idle window.timeoutMsis checked after a sample that did not complete the window, so a window that completes at or after it still returnsmatched: true(with the defaults,timeoutMs: 300matches at about 400 ms) and only a deadline reached earlier is atimeouterror. They no longer say the call returns immediately without crumbs or thattimeoutMsis the maximum total wait. (lib/src/tapir_dowel_provider.dart,lib/src/commands/tapir_wait_for_network_idle_command.dart,skills/tapir/,llms.txt,doc/mcp/tool-reference.md) - The network-idle notes no longer name a crumbs adapter that does not exist. The skill said crumbs's
DioHttpAdapterreports 0, but 1.0.0 of crumbs removed it andpendingCounthas no default body. Only an adapter that implementspendingCount(MagicHttpFacadeAdapter, inmagic_devtools) moves the count, and an interceptor that only callsCrumbsStore.recordHttpcounts nothing. (skills/tapir/,llms.txt) - One install command. tapir depends on the Flutter SDK, so the README,
llms.txtand the skill sayflutter pub add tapir, as the installation page already did. - The
install.yamlmessage and the installation page count the Wind contract right. The contract carries seven keys (className,breakpoint,brightness,platform,states,bgColor,textColor), not six, and the Magic enrichers are listed as examples because there are more than the five named. (install.yaml,doc/getting-started/installation.md) - The docs say which flags print the Wind fields.
tapir:snapprints thewind:block only with--includeEnrichers, andtapir:observereturns justbreakpointandstatesunless you pass--includeEnrichers=full. The block appears under each W-widget that has aclassName(notWAnchor,WBreakpoint,WindAnimationWrapperorWKeyboardActions), withbgColorandtextColoronly when they resolve. The installation page, the README and thetapir:snappage (which still said six fields) now say so. (doc/getting-started/installation.md,README.md,doc/commands/tapir-snap.md) - One startup figure for the fast wrapper. The install message and the installation page quoted about 110 ms and about 50 ms; both say about 0.2 s, which is what
./bin/dowelmeasures on a cached build (0.17 to 0.35 s), and the "~3s" fordart run tapirsays 0.4 to 2.7 s (warm and first run), the figures dowel's docs use. (install.yaml,doc/getting-started/installation.md,lib/src/commands/tapir_install_command.dart) - The tables in the command pages are left-aligned. The 18 pages under
doc/commands/used the unaligned|---|separator; they use|:---|like the rest of the docs. (doc/commands/) - The tables in the README,
ARCHITECTURE.mdand the skill are left-aligned too. (README.md,ARCHITECTURE.md,skills/tapir/)
1.0.0
Added
TapirPlugin.isDisabled. Whether the build set--dart-define=TAPIR_DISABLE=1(ortrue,yes), read once from the define.install()returned early with no signal, so an integration went on installing its watchers and perf hooks around a tapir that registered nothing; it can now check the same flaginstall()does. (lib/src/tapir_plugin.dart)ext.tapir.fillcarries the type step'schecksblock. It copied onlyeffect, so a fill whose receives-events check wasindeterminateorskippedread as a clean pass;checksis now present on exactly the responses where the type step stamped it. (lib/src/extensions/ext_fill.dart)tapir:doctorasks the running app. The TAPIR_DISABLE, enricher and semantics rows read the CLI process (its own define, a count hardcoded to 0, a semantics state hardcoded to on), so the one ERROR row could never fail. With a session they now callext.tapir.boot_idover the recorded VM Service URI: no answer forext.tapir.*warns that the app was built withTAPIR_DISABLEor never callsTapirPlugin.install()and fails the semantics row, a tree that is off fails it, and the enricher row prints the app's count. With no session, or an app that does not answer, each row says skipped. An app that registers noext.tapir.*is asked again once, a second later, since a hot restart drops every extension until tapir registers them again andext.tapir.boot_idregisters last; only the second answer fails the row.ext.tapir.boot_idanswers two more keys for it,semanticsEnabledandenrichers. Breaking: the doctor can now exit 1 where its error row could not fail before, so a CI step that runs it against an app without tapir installed starts failing. (lib/src/commands/tapir_doctor_command.dart,lib/src/extensions/ext_boot.dart)tapir:installandtapir:doctorname whatfluttersdk_duskleft behind. An oldbin/dispatcher.dartmade the installer skipdowel installin silence. Both now list abin/dispatcher.dartthat importsfluttersdk_artisan,bin/fsa,lib/app/commands/_index.g.dart, app commands still written against theArtisan*classes and afluttersdkserver in.mcp.json, then the commands to run (dart run dowel install,dart run dowel commands:refresh,dart run tapir tapir:install); the doctor reports them as a warning. (lib/src/commands/tapir_install_command.dart,lib/src/commands/tapir_doctor_command.dart)- Two error types,
refusedandinvalid_param. A refusal (a ref that is not a text field, a container holding several checkboxes, a perf call with no session open) answeredtype: unexpected, and most bad input answered the generic extension error.refusedanswers JSON-RPC-32001;invalid_paramandmissing_paramanswer the standard invalid-params-32602; every other type keeps-32000.TapirErrorEnvelope.codeis the one place the mapping lives. (lib/src/utils/error_envelope.dart)
Changed
- Breaking: tapir requires
dowel ^1.0.0. The^0.0.21range cannot resolve next to dowel 1.0.0 or the Magic packages that depend on it. dowel 1.0.0 trims its barrel and renames the install results, so thehide Errorimports are gone. (pubspec.yaml) - Breaking:
set_checkboxruns the actionability gate. It tapped wherever the control was, so an offscreen or covered checkbox reportedverified: falsewith no reason. It now gates the control it taps like the other tapping verbs: one below the fold is scrolled into view, and one that is disabled, covered or moving is refused with the gate's reason (obscured by other widget, ...) and its envelope type. It takescheckStableandcheckReceivesEventslike them,tapir:set_checkboxincluded as--[no-]checkStableand--[no-]checkReceivesEvents, and stampscheckswhen the gate could not confirm. A script no longer has to scroll the control into view first, and handles a refusal where it used to readverified: false. (lib/src/extensions/ext_checkbox.dart,lib/src/tapir_dowel_provider.dart,lib/src/commands/tapir_set_checkbox_command.dart) - Breaking:
set_checkboxreports the settledvalue.valueechoed the request, so a control whose owner ignored the tap answeredvalue: true, effect.after: false, verified: false. It is now the state read back after the tap, equal toeffect.after. (lib/src/extensions/ext_checkbox.dart) - Breaking: refusals and bad input change their error type and code. The sites that refuse a target now answer
refused(-32001): the ambiguous or statelessset_checkboxref, the text verbs on a ref that is not a text field,fillandclearwith no field under the ref,focuswith nothing focusable,perf_endandperf_insightandperf_tracewithout the session they need, andsemantics_holdrelease outside a session. The sites that reject a value now answerinvalid_param(-32602):perf_beginmode,scrolldelta,screenshotrect,semantics_holdaction andset_checkboxvalue. Everymissing_paramanswers-32602instead of-32000. A caller that branched onunexpectedor on-32000for these reads the new type. (lib/src/extensions/) - Breaking: the console and exception readers take a request object.
recentLogsReaderwas({int limit, String? minLevel})andrecentExceptionsReader({int limit}), so a filter added later would have broken every host that assigns an exact-shape closure. Each is nowList<Map<String, Object?>> Function(ConsoleRequest)andFunction(ExceptionsRequest), with the same fields and defaults; both request classes are exported. A host writesrecentLogsReader = (ConsoleRequest request) => ...and readsrequest.limit. (lib/src/extensions/ext_console.dart,lib/src/extensions/ext_exceptions.dart,lib/tapir.dart) - Breaking: the barrel exports named symbols only.
export 'src/tapir_plugin.dart'andexport 'src/ref_registry.dart'exported everything, including the semantics handle, the boot id minting, the install counter and the ref registry's test helpers. Each export now names its symbols, and those internals moved out of the classes tosrc/:TapirPlugin.releaseSemantics,acquireSemantics,semanticsReleased,mintBootId,installCountandresetSemanticsForTesting, andRefRegistry.registerForTesting,disposeGroupandactiveRefs, are no longer public.TapirPlugin.aiTestDisableEnvValue, mutable and only used by tests, is gone; readTapirPlugin.isDisabled. (lib/tapir.dart,lib/src/tapir_plugin.dart,lib/src/ref_registry.dart) - Breaking: four CLI commands take their MCP tool's name, and
dblclickisdouble_click.tapir:waitistapir:wait_for,tapir:modalistapir:dismiss_modals,tapir:resizeistapir:resize_viewportandtapir:deviceistapir:device_profile(their substrate routes follow:dowel:tapir:resize_viewport,dowel:tapir:device_profile). Besideright_clickandtriple_click,ext.tapir.dblclick,tapir:dblclickandtapir_dblclickareext.tapir.double_click,tapir:double_clickandtapir_double_click. No alias is kept: rename them in scripts, allow lists andtapir:runsteps. (lib/src/commands/,lib/src/extensions/ext_pointer.dart,lib/src/tapir_dowel_provider.dart)
Removed
- Breaking:
ext.tapir.find_by_label. No command, MCP tool, doc or skill called it, andext.tapir.findcovers it;ext.tapir.find_by_textstays for the perf runner. (lib/src/extensions/ext_wait_find.dart)
Fixed
tapir:installresolves amagic_devtoolspath dependency outside the project. The walk to.dart_tool/package_config.jsonwent through a file path that kept a trailing separator, so the../../magic_devtoolspub writes forpath: ../magic_devtoolsresolved one level too deep and the Magic wiring was skipped with a "no lib/tapir.dart" warning. The walk stays on directory URIs, and everyrootUriresolves through a file path, so a percent-encoded or absolutefile:URI works too. (lib/src/commands/tapir_install_command.dart)set_checkboxtaps a bare switch, not its labelled container. ASwitchorCheckboxis no semantics boundary, so insideSemantics(container: true, label: ...)its flag lands on the container's node and the tap went to the container's centre, which can sit beside the control. The tap now goes to the render object that set the flag. (lib/src/extensions/ext_checkbox.dart)tapir:hot_reload_and_snaprecognises every reload line. It matchedin \d+ms, which misses the web runner'sReloaded application in ...and any reload of a second or more, printed as1,234ms; such a reload waited 10 s and exited 1 although it worked. (lib/src/commands/tapir_hot_reload_and_snap_command.dart)set_checkboxrefuses a value that is not a boolean. Anything butfalseor0read as true, so"flase"ticked the box; onlytrue,false,1and0are accepted, and the rest is refused asinvalid_param. (lib/src/extensions/ext_checkbox.dart)- A target half off screen with no
Scrollableis tapped on its visible part. The gate passes a rect that still overlaps the viewport, and the pointer then went to its centre outside the view and reached nothing. The dispatch point, and the gate's hit probe, are the centre of the rect clipped to the visible viewport. (lib/src/utils/actionability_gate.dart) - The
FakeCdpServerlifecycle test no longer flakes. It asserted thatstart()andstop()each took under 100 ms, which failed under a loaded full run; it now bounds both at 10 s to catch a hang and keeps the port re-bind check. Test-only. (test/src/cdp/fake_cdp_server_test.dart) tapir:doctorfinishes when a file the leftovers scan reads cannot be read. An unreadable.mcp.jsonthrew out of the scan and aborted the whole run. The row now warnscould not inspect <path>: ...instead, and never readsnone foundafter a scan that stopped. (`lib/...
0.0.24
Changed
- Breaking, nothing is migrated automatically. In a consumer project: (1) swap the dependencies with
flutter pub remove fluttersdk_dusk fluttersdk_artisan && flutter pub add tapir doweland rename the imports and theDusk*identifiers; (2) follow the dowel 0.0.21 migration (move.artisan/to.dowel/, delete.dowel/plugins.json), deletebin/dispatcher.dart,bin/fsaandlib/app/_plugins.g.dart(the installer keeps an existingbin/dispatcher.dart, and the old one importsfluttersdk_artisan, sobin/no longer analyses) and rundart run tapir tapir:installagain; (3) renamedusk_*totapir_*in any saved MCP allow list, prompt or script (a Claude Code allow-list entrymcp__fluttersdk__dusk_*becomesmcp__dowel__tapir_*, since the server key is nowdowel); (4) delete the stalefluttersdkserver entry from.mcp.json(dowel leaves entries an earlier release wrote alone) and from every client you registered by hand (claude mcp remove fluttersdk, the VS Code or Claude Desktop settings), since they run./bin/fsa mcp:serve, which step (2) deletes; then rundart run tapir mcp:installand registerdowelwith the per-client commands in the README; (5) in a Magic app, remove theimport 'package:magic_devtools/dusk.dart'line and theMagicDuskIntegration.install()call that the old installer added tolib/main.dart(MagicDuskIntegrationis not aDusk*name that step (1) renames), and check thatlib/main.dartcallsMagicTapirIntegration.install()aftertapir:installhas run (the installer adds that call once the installedmagic_devtoolsshipstapir.dart; see the Magic bullet below), without adding a second one. - Renamed from
fluttersdk_dusktotapir. The package, the repository (anilcancakir/tapir), the barrels (package:tapir/tapir.dartandpackage:tapir/cli.dart), everyDusk*class (nowTapir*, andDuskArtisanProviderisTapirDowelProvider), the executable and the file names carry the new name. Also renamed: theext.dusk.*VM Service extensions (nowext.tapir.*), the MCP tools (dusk_*istapir_*), the CLI commands (dusk:*istapir:*, and the substrate commandsartisan:dusk:*aredowel:tapir:*), the environment variables (DUSK_*isTAPIR_*), the perf zone key (#fluttersdk_interactionis#tapir_interaction) and the docs URLs (https://anilcancakir.com/tapir). - Built on
dowelandwind_kit_diagnostics_contracts.fluttersdk_artisanis nowdowel ^0.0.21andfluttersdk_wind_diagnostics_contractsiswind_kit_diagnostics_contracts ^1.3.0. The.mcp.jsonserver key written bymcp:installisdowel. - The Magic step of
tapir:installneeds amagic_devtoolsbuilt on tapir. When the consumer's pubspec namesmagic_devtoolsand the resolved release shipspackage:magic_devtools/tapir.dart, the installer wiresMagicTapirIntegrationfrom it. Themagic_devtoolsreleases that still shipMagicDuskIntegrationhave no such library, so with one of them the installer skips the Magic lines and warns; upgrademagic_devtoolsto a release built on tapir and rerundart run tapir tapir:install(see Fixed). - The floor is Dart 3.13 and Flutter 3.47. Flutter 3.50 deprecates
package:flutter/material.dartin favour ofmaterial_ui, which needs that floor; the tests and the example app now build onmaterial_ui, the package itself imports neither, and the repository is formatted with the tall style that floor enables. A consumer on an older SDK stays on the lastfluttersdk_duskrelease.
Fixed
tapir:set_checkboxreads the checked state from Semantics. It used to test the widget type (is Checkbox,is Switchfromflutter/material), so a control frommaterial_ui,wind_kitor any hand-builtSemantics(checked: ...)read as "unknown, assume unchecked", and setting a checked box to checked toggled it off. The state now comes from the checked or toggled flag of the ref's Semantics node, which every library sets. A ref without a checked or toggled state is refused instead of assumed unchecked, and a mixed checkbox is always tapped. (lib/src/extensions/ext_checkbox.dart)set_checkboxtaps the control it read and refuses an ambiguous container. When the state came from a descendant node (a labelled container that does not merge), the tap still went to the centre of the ref's own rect, which could be a sibling control or nothing: a container holding three switches read the first and flipped the second. The tap now goes to the centre of the descendant whose state was read, and a ref whose own node has no checked or toggled flag while several descendants do is refused with an error naming the ambiguity. A ref whose own node carries a flag (a merging tile included) is read as that control, as before. (lib/src/extensions/ext_checkbox.dart)tapir:set_checkboxexits 1 when the effect is not verified. It printedCheckbox e5 set to trueand exited 0 without readingeffect.verified, so a tap that missed looked like a success. It now prints the value read back after the tap and exits 1, notes when the box started mixed;--jsonand--includeSnapshotstill print the whole response and exit 0. The--includeSnapshothelp no longer promises a snapshot the response never carries. (lib/src/commands/tapir_set_checkbox_command.dart)tapir:installwires Magic only when the installedmagic_devtoolsshipstapir.dart. The pubspec entry alone decided it, so amagic_devtoolsrelease that predates tapir got an import ofpackage:magic_devtools/tapir.dartand alib/main.dartthat did not compile. The installer now resolves the package through the nearest.dart_tool/package_config.jsonat or above the project (a pub workspace member uses the one at the workspace root; an unreadable or invalid one counts as unresolved) and, withoutlib/tapir.dart, skips the Magic lines and warns to upgrade and rerun.tapir:doctorfollows the same check, so it no longer advises wiring a class the app cannot import. (lib/src/commands/tapir_install_command.dart,lib/src/commands/tapir_doctor_command.dart)
Removed
- The
FluttersdkDuskArtisanProvideralias inpackage:tapir/cli.dart. The class dowel'splugins:refreshgenerates for this package,TapirDowelProvider, is the provider itself, so the alias is gone.
0.0.23
Added
- Dusk wakes a hidden or minimised app instead of driving a frozen one. A hidden, minimised or occluded desktop window moves the lifecycle to
hiddenand turns frame production off, so a tap reported "no observable change", a snapshot went stale and a screenshot returned the last cached frame. Everyext.dusk.*verb now deliversAppLifecycleState.inactiveand then, once a frame has run,resumedthrough theflutter/lifecyclechannel, and schedules a forced frame when frames are off and the lifecycle ishiddenorpaused(neverdetached); a response that forced carriesframesForced: true. It is re-applied on every call because the OS flips the lifecycle back tohiddenon its next occlusion event. The wake counts only when a frame actually runs after it (otherwise the previous lifecycle is delivered again, unless the engine sent one of its own meanwhile), and it is never attempted in the app on web, where Chrome withholds animation frames from a hidden tab; the CLI sendsPage.bringToFrontthere instead (see the web recovery entry below). Verified on macOS with the window hidden throughosascript(dusk:tap,dusk:snap,dusk:screenshotanddusk:navigateall work) and on web with a second tab in front (warnings returned, nothing claimed as forced). The perf verbs,boot_id,close_app,console,exceptions,semantics_holdandevaluateobserve the engine as it is and do not force. The app's lifecycle observers seeresumedwhile this happens;--dart-define=DUSK_FORCE_FRAMES=false(kDuskForceFrames) turns it off. (lib/src/utils/frame_sync.dart,lib/src/utils/dusk_register.dart,lib/src/utils/dusk_response.dart,lib/src/extensions/,doc/reference/frame-production.md) dusk:snapanddusk:observesay when a control is off screen or covered. A ref line may end with[offscreen](the widget lies outside the viewport) and/or[occluded](a hit test at its centre reaches something painted above it), after[ref=eN]; a healthy screen prints neither. Eachext.dusk.observecandidate gainsinViewportandoccluded;isVisibleis unchanged. (lib/src/utils/viewport_visibility.dart,lib/src/extensions/ext_snapshot.dart,lib/src/extensions/ext_observe.dart)dusk:snap --grepreads an invalid regex as text. A label such asContainer Sample (MP4, H.264is an unterminated group and used to crash the snapshot; it now matches as a literal substring and the payload carriesgrepLiteral: true. A valid regex keeps its regex meaning. (lib/src/extensions/ext_snapshot.dart)- An
e<N>ref survives a rebuild. Snapshots bind a label-and-index query and the match count to every labelled ref (RefBinding, withRefRegistry.bind,lookupBindingandrebind, additive statics). When the element behind a ref has left the tree, the action relocates it only if the live tree has the same number of matches as at snapshot time, and the response lists the ref underreresolved. With a different count the olddefunct (element no longer mounted)error is unchanged. A ref is also pinned to the route it was minted on (RefBinding.route, held weakly): a same-label control on another route is never a match, so a ref taken on a page you have since left fails asdefunctinstead of acting on the next page. Known limit: the query has no role, so a node of another role with the same label in the same position on the same route could match. (lib/src/ref_registry.dart,lib/src/extensions/ext_pointer.dart,lib/src/extensions/ext_snapshot.dart) dusk:screenshotinfers the format from--output..png,.jpgand.jpegpick the format when--formatis not given, so a.pngfile no longer holds JPEG bytes. An explicit--formatwins and a mismatch with the extension warns on stderr. (lib/src/commands/dusk_screenshot_command.dart)dusk:run/dusk_run: several verbs in one call.dusk:run "tap --ref=e5" "fill --ref=q1 --text=\"a b\"" "snap"takes each step as a shell-tokenised verb line (a bare verb gets thedusk:prefix), parses it with that verb's own options and runs it on one shared VM connection. The output is one line per step ([1/3] tap --ref=e8 ok: Tapped e8, the first output line cut to 120 characters), then the last step's output. It stops at the first failing step withstep N failed (exit C): <step>and that exit code;--continueruns every step and exits with the first non-zero code. A nesteddusk:run, an unknown verb and a verb that does not boot connected (dusk:resize,dusk:device,dusk:install,dusk:doctor,dusk:perf_campaign,dusk:perf_compare) fail as their step, as does any step that throws (it exits 3); a step that ends the app (dusk:close_app,dusk:perf_run) belongs last. The MCP tool takes{steps: string[], continue?: boolean}and routes asartisan:dusk:run, the seventh substrate-routed tool. Between steps use--text,--keyor q-refs, since e-refs belong to one snapshot. 42 commands and 40 tools now. (lib/src/commands/dusk_run_command.dart,lib/src/dusk_artisan_provider.dart,doc/commands/dusk-run.md)--snapshot=full|diff|noneondusk:tap,fill,type,navigateandscroll, and asnapshotenum ondusk_tap,dusk_fillanddusk_navigate. It overridesincludeSnapshot.diffreturns only the lines that changed against the last unfiltered snapshot (adusk:snapor an earlier action) in unified-diff form: a changed row is- <old line>then+ <new line>, an added row+, a removed row-, then a(K of N lines changed)footer, and the payload addssnapshotMode: "diff". The mode is read case-insensitively, and an unknown one comes back assnapshotErrorin the payload rather than as a silently missing snapshot. With no baseline yet the full text comes back. A hot restart clears the baseline, and a filtered snap (grep,within,depth,interactiveOnly) or an enricher walk never records it. (lib/src/utils/snapshot_diff.dart,lib/src/extensions/ext_snapshot.dart, the action verbs and commands)dusk:snap --jsonprints the whole envelope. The tree alone droppedgrepLiteral,warningsandframesForced; text mode is unchanged. (lib/src/commands/dusk_snap_command.dart)dusk:press_key --modifiersis applied.control(orctrl),shift,altandmeta, in any case, as a comma list on the CLI and an array over MCP, are held down before the key and released after it, even when delivery throws, so exact-match shortcuts such as Ctrl+S and Shift+Tab fire. An unknown name is an error before anything is pressed. (lib/src/extensions/ext_text_input.dart,lib/src/utils/key_press.dart,lib/src/commands/dusk_press_key_command.dart)- A tap on a merge boundary goes to the control the ref names (#60). A
MergeSemanticsboundary taps at its own centre, which can belong to another handler (a full-frame stage that pauses on any tap) while the gate still saysconfirmed. When exactly one labelled control owns the merged label (aSemanticswith a label that is part of it and anonTaporbutton) and a hit test at its centre reaches it,dusk:tapsends the pointer there and the payload addsretargeted {from, to, ownerRect}(withcentreHandled: truewhen the centre has a tap handler of its own, so the stage or row it would have tapped is named); the summary readsTapped e7 on its control at (x, y), not the merged centre (x, y). With several candidates, or an owner that is covered or off screen, the tap stays at the centre and the payload addstapPointOutsideOwner {ownerRect, point}with one stderr line.retarget: false(CLI--no-retarget) taps the centre as before. Snapshot and observe say so before any tap: atapOwner: {"x","y","w","h"}sub-line under the ref and atapOwnercandidate field, both present only when the centre misses the control, withcentreHandled: trueadded when the centre has its own handler andretargets: falsewhen a tap cannot go to the control. The gate and its reasons are unchanged;hover,dblclick,right_click,triple_clickanddragstill use the centre. (lib/src/utils/tap_owner.dart,lib/src/extensions/ext_pointer.dart,lib/src/extensions/ext_snapshot.dart,lib/src/extensions/ext_observe.dart,lib/src/commands/dusk_tap_command.dart) effect.targetsays which field a text verb wrote.dusk:fill,dusk:typeanddusk:clearaddtarget {matchKind: overlap|nearest|element, label, rect}to thetexteffect, and the CLI prints a warning when the match wasnearest. (lib/src/utils/effect_report.dart,lib/src/extensions/ext_text_input.dart,lib/src/commands/dusk_fill_command.dart,lib/src/commands/dusk_type_command.dart)- The CLI recovers a hidden web tab.
snap,observe,tap,fill,type,navigate,scrollandpress_keycall the extension throughcallExtensionVisible. When the answer carries a webwarningsblock, or the call times out (the VM Service gives up on a hidden tab that never answers) and the tab reportshidden, and the session has acdpPort, the CLI sends CDPPage.bringToFrontonce (it raises the Chrome window). Read-only verbs (snap,observe) are then called again once; a verb that acts is not repeated, since its gesture may already have fired, and a stderr notice says to rundusk:snap. A CDP failure leaves the original result and exit code; without acdpPortnothing changes. (lib/src/cdp/page_visibility.dart,lib/src/commands/web_visibility.dart)
Changed
- Every MCP tool schema is closed, and lists the knobs its handler reads. artisan validates each
tools/callagainst the schema, so a key the schema did not list was either let through to a handler that ignored it (a typo answered with a clean response) or, when the schema was incomplete, impossible to pass. All 40 schemas now declareadditionalProperties: false, and the keys the handlers read but the schemas o...
0.0.22
Fixed
dusk:observelists a merged control once. The walk emitted every node that carries an action, andMergeSemanticskeeps each absorbed descendant in the tree as its own node flaggedisMergedIntoParent, so anMSSwitchrow came back as a 324x44 boundary plus a 44x24 duplicate and anMSButton"Kaydet" as 79x44 plus 79x36, while the platform sees one node. A merged node now yields no candidate and noqNhandle; the walk still descends through it, the candidate shape is unchanged, and the boundary is still listed. The candidate count can drop for a screen with merged controls, and a script that held a candidate by position has to re-observe; adusk:perf_run{role, name, index}target with an index above 0 now counts the same nodesdusk:snaplists. The per-label counter still counts merged nodes, because the find resolver does, so thematchIndexof every surviving handle is unchanged. (lib/src/extensions/ext_observe.dart)
0.0.21
Fixed
dusk:snaplists a merged control once. The walk printed every semantics node that carries an action, andMergeSemanticskeeps each absorbed descendant in the tree as its own node flaggedisMergedIntoParent, so a magic_starterMSSwitchrow showed up as abutton "X" [ref=e1]followed by a secondbutton "X" [ref=e2]while the platform sees one node. A merged node is no longer printed and gets no ref; the walk still descends through it, and the line shape, the ref syntax and the textbox collapse are unchanged. Refs after a removed duplicate shift down by one each, so a script that hard-coded ref numbers has to re-snap. AWSelectthat still prints twice is the wind side (two real platform nodes), fixed influttersdk_wind1.9.3. (lib/src/extensions/ext_snapshot.dart)
0.0.20
Changed
- The artisan floor names this batch's release.
fluttersdk_artisanmoves^0.0.19to^0.0.20. The old range already admitted 0.0.20, so a freshpub getresolves nothing differently; what changes is that the floor names the release this package is verified against. artisan 0.0.20 is not breaking, and a consumer inherits one behaviour from it: adusk:*command run while the app is still booting, beforeext.dusk.*is registered, now prints one retry line and exits 75 instead of an RPC stack and exit 3. The requirements tables indoc/getting-started/installation.mdanddoc/getting-started/index.mdand the stack lines inCLAUDE.mdand.github/copilot-instructions.mdquote the new floor. (pubspec.yaml,doc/getting-started/,CLAUDE.md,.github/copilot-instructions.md)
0.0.19
Changed
- The artisan floor names this batch's release.
fluttersdk_artisanmoves^0.0.17to^0.0.19. The old range already admitted 0.0.19, so a freshpub getresolves nothing differently; what changes is that the floor names the release this package is verified against. Neither artisan 0.0.18 nor 0.0.19 is breaking: 0.0.18 adds installer operations dusk does not use, and 0.0.19 widens artisan'sxmlconstraint to admit 7.x, which a consumer now inherits. The requirements tables indoc/getting-started/installation.mdanddoc/getting-started/index.mdand the stack lines inCLAUDE.mdand.github/copilot-instructions.mdquote the new floor. (pubspec.yaml,doc/getting-started/,CLAUDE.md,.github/copilot-instructions.md) - The example app moves
cupertino_iconsto^2.0.0. It touches only the unpublishedexample/app, not the package. (#53,example/pubspec.yaml,example/pubspec.lock)
0.0.18
Fixed
dusk:press_key/dusk_press_keyreaches the focused widget. The key was handed toHardwareKeyboard.instance.handleKeyEvent, which updates the pressed-key state and calls the keyboard's global handlers and nothing else, soFocus.onKeyEvent,ShortcutsandCallbackShortcutsnever heard it, whatever the tool description promised: threeArrowDownpresses on an app whose focusedFocusScopezaps channels on the arrows changed nothing. Both halves of the press now enter through the binding'sonKeyData, the handler real key data reaches, marked synthesized so each is dispatched at once, which delivers toHardwareKeyboardand to the focus tree alike. Every key also carries its own physical key now, where every press used to report the Enter key's.dusk:reset_overlayspressed its Escape the same broken way and now goes through the same helper, so aShortcuts-bound dismiss hears it. Two limits are documented on the helper: a real key whose raw message is still pending holds the press back until the next real key, and on an embedder that sends only raw key messages (none of Flutter's own) a press before the first real key fixes the transit mode to key data and real keys stop reaching the focus tree until a restart, so such an embedder must not be driven with it;reset_overlaysstays best-effort there, catching the debug assert as well as a missing handler. (lib/src/utils/key_press.dart,lib/src/extensions/ext_text_input.dart,lib/src/extensions/ext_modal_router.dart)
Added
dusk:press_key --key=<letter or digit>. A single ASCII letter or digit presses the key that carries it (Gandgboth press the G key with the characterg), so a shortcut bound to a letter can be driven; text still goes throughdusk:type, since a text field takes its text from the text input channel. Anything else outside the named keys is still refused. (lib/src/extensions/ext_text_input.dart,lib/src/commands/dusk_press_key_command.dart,lib/src/dusk_artisan_provider.dart,doc/mcp/tool-reference.md)