Repository navigation
fix(ios): book every surface on the lit panel, not the first window - #4122
Conversation
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
Size Report
Startup median (7 runs, lower is better):
|
![cubic-dev-ai[bot]](https://avatars.githubusercontent.com/in/1082092?s=60&v=4){kind=small}
There was a problem hiding this comment.
All reported issues were addressed across 36 files
Reply with feedback, questions, or to request a fix.
View guided diff | Turn on auto-fix | Re-trigger cubic
Apex SentinelVerdict: Since the last review: 12 new · 3 still open · 11 no longer raised 15 finding(s): 7 minor, 8 nit. Verdict: approve. 🟡 Minor (7)
🔵 Nit (8)
Reply under an inline comment with Earlier reviewsReview of
|
| Where | Finding |
|---|---|
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerRequestPanel.swift:68 |
🆕 A lit-panel booking is never released when the next command starts, so the booking guard can be bypassed |
🟡 Minor (7)
| Where | Finding |
|---|---|
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerXCTestEventBridge.m:112 |
🆕 Lit-panel read failure on the SpringBoard/alert host reports APP_SCREEN_WINDOW_UNRESOLVED instead of the panel reason |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerRequestPanel.swift:96 |
🆕 elementSearchRoots refuses an empty lit-panel window list that elementSearchRoots' own doc treats as valid |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+CommandExecution.swift:33 |
🆕 Lit-panel booking is skipped only for screenshot; other capture-routing commands may pay a booking refusal before resolving their own surface |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+TextEntry.swift:589 |
boundTextEntryInputQuery picks the first matching root query by side-effecting .count, losing the fail-closed .element semantics |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+TextInputProbe.swift:96 |
probeTextInputs' empty-tree check still reads the whole app tree, not the lit-panel roots it searches |
packages/platform-apple/src/foldable/orientation.ts:81 |
🆕 A device whose fold facts or panel count are unknown is routed as single-panel, so it silently keeps the first-window route |
packages/platform-apple/src/foldable/orientation.ts:89 |
🆕 readFoldablePanelTable can cache a never-read panel count as single-panel for 5s when a short request is cut off mid-read |
🔵 Nit (6)
| Where | Finding |
|---|---|
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+Interaction.swift:200 |
Selector candidate hasTappableFrame is now only computed for the non-hittable fallback path |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+CommandDispatch.swift:434 |
🆕 A lit-panel recording keeps a captured panel value that dies with the request's Command |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+Interaction.swift:169 |
🆕 findElement(app:text:) loses the app-wide fail-open and can report the wrong first match across lit roots |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+Snapshot.swift:327 |
🆕 hard-coded display-ID sentinel in the new windowsOnLitDisplay helper |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+SynthesizedInteraction.swift:281 |
🆕 Touch visualization frame still uses app.frame's origin on a lit-panel request, while the tap anchors at the booked window |
packages/apple-runner/src/runner-command-traits.ts:100 |
🆕 keyboardDismiss scoped as a surface command for a capture reason it does not serve |
Review of d6aab77a · 03.11.2026 10:53 UTC
Apex Sentinel
Verdict: approve
Since the last review: 13 new · 3 still open · 11 no longer raised
16 finding(s): 9 minor, 7 nit. Verdict: approve.
🟡 Minor (9)
| Where | Finding |
|---|---|
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+TextInputProbe.swift:96 |
🆕 The probe's treeReadable check still reads the whole app, not the lit-panel roots it resolved |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerRequestPanel.swift:20 |
RunnerLitPanelState is read and written from several threads without synchronisation |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+Alert.swift:190 |
🆕 Unsettled: XCTest issue containment inside withBoundedInteractionIdleTimeoutIfSupported for a throwing activateElement |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+Interaction.swift:270 |
🆕 resolveRunnerWindow can hand callers CGRect.null, whose origin is +infinity, where the predecessor returned app.frame |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+Interaction.swift:270 |
🆕 resolveRunnerWindow's .null fallback for a lit-panel request does not prevent the app-frame fallback it claims to forbid |
packages/platform-apple/src/core/runner-client.ts:27 |
🆕 Surface commands sent through runApplePressSeries bypass the new lit-panel naming |
packages/platform-apple/src/core/runner-client.ts:27 |
🆕 A retried or resent surface command reuses a lit panel read before the retry |
packages/platform-apple/src/foldable/runner-panel.ts:60 |
🆕 The probe's time is charged to the request budget only when a panel is named |
packages/platform-apple/src/foldable/runner-panel.ts:64 |
The probe-budget refusal does not disclose dispatched even though the disclosure fixture says it never dispatched |
🔵 Nit (7)
| Where | Finding |
|---|---|
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerRequestPanel.swift:104 |
Lit-panel element search re-walks the whole (uncapped) window list on every read |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+Interaction.swift:200 |
🆕 hasTappableFrame is now skipped for candidates that are hittable or outside the expected point |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerRequestPanel.swift:74 |
🆕 bookRequestLitPanelSurface refuses an unreadable app bundle identity as APP_SCREEN_UNRESOLVED for every surface command |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+TextEntry.swift:589 |
🆕 boundTextEntryInputQuery picks its query by count, not by resolution |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerRequestPanel.swift:97 |
🆕 elementQueries' doc describes only element reads and omits text-input probes |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+Snapshot.swift:484 |
🆕 publishLitPanelRoot is applied in two different shapes for the same invariant |
packages/apple-runner/src/runner-command-traits.ts:100 |
🆕 ADR 0025's amendment table and the TS traits table agree, but nothing links them to the Swift booking rules |
Review of 096d77d2 · 03.11.2026 10:09 UTC
Apex Sentinel
Verdict: request_changes
Since the last review: 8 new · 6 still open · 10 no longer raised
14 finding(s): 1 major, 5 minor, 8 nit. Verdict: request_changes.
Required fixes
- Lit-panel bookings are reused across apps because
===compares the XCUIAutomationObject proxy, not the app
🟠 Major (1)
| Where | Finding |
|---|---|
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerRequestPanel.swift:44 |
🆕 Lit-panel bookings are reused across apps because === compares the XCUIAutomationObject proxy, not the app |
🟡 Minor (5)
| Where | Finding |
|---|---|
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerRequestPanel.swift:20 |
A lit-panel booking survives the next command when the request leaves main idle work behind |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+Interaction.swift:233 |
hasTappableFrame turns a booking refusal into a thrown error, changing what a duplicate-match refusal reports |
packages/platform-apple/src/foldable/orientation.ts:76 |
🆕 readFoldablePanelTable back-off can strand a genuinely multi-panel device on the historic route after a first failed read |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerRequestPanel.swift:64 |
🆕 elementSearchRoots throws away the distinction it just logged |
packages/platform-apple/src/foldable/runner-panel.ts:63 |
panel-table-unreadable can be reported for a panel table that answered, and the code reuses APP_SCREEN_UNRESOLVED |
🔵 Nit (8)
| Where | Finding |
|---|---|
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+Navigation.swift:107 |
🆕 backInApp ignores the booked lit panel: the back control is searched app-wide, so a dark-panel duplicate can be tapped |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerRequestPanel.swift:63 |
🆕 elementQueries re-walks the whole window subtree for every read a lit-panel request performs |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+Lifecycle.swift:54 |
captureRunnerFrame now reads the main-only requestLitPanel from off the main thread |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+Lifecycle.swift:52 |
🆕 captureRunnerFrame keeps reading the live requestLitPanel while the recording pump was pinned to its own panel |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+Snapshot.swift:296 |
windowsOnLitDisplay's generic parameter buys nothing and the nil-display rule is duplicated |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerXCTestEventBridge.m:240 |
🆕 Lit-panel gestureViewport reports a null frame when no window books on the lit panel |
packages/platform-apple/src/core/runner-client.ts:27 |
🆕 The lit-panel probe reads devicectl on the path of every tap on any device the fold fact admits |
packages/platform-apple/src/foldable/runner-panel.ts:57 |
A surface request whose panel read came back empty still sends the full timeoutMs |
Review of 34154226 · 03.11.2026 09:28 UTC
Apex Sentinel
Verdict: approve
Since the last review: 11 new · 5 still open · 14 no longer raised
16 finding(s): 9 minor, 7 nit. Verdict: approve.
🟡 Minor (9)
| Where | Finding |
|---|---|
packages/platform-apple/src/foldable/runner-panel.ts:52 |
🆕 Panel probe consumes the request timeout twice; the runner gets less budget than options still promises |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+Interaction.swift:267 |
A lit-panel request whose window cannot be booked yields a .null window frame into TapPointPolicy instead of a typed refusal |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+TextEntry.swift:590 |
🆕 boundTextEntryInputQuery drops the app-wide fallback and picks roots by query count, so a lit-panel request can lose an input the old code resolved |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerRequestPanel.swift:62 |
elementSearchRoots caches the lit windows for the whole request, so a window that vanishes mid-request stays a search root |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+CommandExecution.swift:34 |
🆕 backSystem and appSwitcher are refused with APP_SCREEN_OFF_LIT_PANEL for a gesture that never touches the app's windows |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+Interaction.swift:236 |
Lit-panel tap-point gate fails open when no window is booked, unlike the route it replaced |
packages/platform-apple/src/foldable/orientation.ts:33 |
🆕 A single failing 5s panel read routes a foldable's surface commands through the first-window path for 5s |
packages/platform-apple/src/foldable/runner-panel.ts:54 |
🆕 Panel-table probe silently overwrites the request's own timeoutMs and can refuse the request outright |
packages/platform-apple/src/foldable/runner-panel.ts:55 |
🆕 A request with no explicit timeoutMs gets no budget accounting for the panel read |
🔵 Nit (7)
| Where | Finding |
|---|---|
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerRequestPanel.swift:20 |
🆕 Lit-panel write is queued asynchronously, so a command can read the previous command's panel |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerXCTestEventBridge.m:313 |
Element search walks every window uncapped, discarding the #2995 issue cap the booking walk exists for |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerRequestPanel.swift:17 |
Lit-panel state lives on RunnerTests but is written by a queue hop and read from other threads' queued work without a documented single owner |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+CommandExecution.swift:429 |
🆕 The .type no-response fallback now shares a Result with the lit-panel refusal |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+Snapshot.swift:296 |
🆕 windowsOnLitDisplay and snapshotDisplayID are only ever exercised with their real arguments |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerXCTestEventBridge.h:43 |
🆕 AGENT_DEVICE_MAX_LIT_PANEL_WINDOWS is a bare define with no literal in ADR 0025 |
packages/platform-apple/src/foldable/orientation.ts:81 |
🆕 Lit-panel gate depends on device name matching, so an unnamed Duo silently keeps the buggy first-window route |
Review of 19b038ae · 03.11.2026 08:38 UTC
Apex Sentinel
Verdict: request_changes
Since the last review: 16 new · 3 still open · 11 no longer raised
19 finding(s): 2 major, 10 minor, 7 nit. Verdict: request_changes.
Required fixes
- Coordinate taps/long-presses/drags on a lit-panel request anchor at the booked window's real origin while offsetting by the panel-box origin
- RunnerCopyLitPanelWindows ignores the AGENT_DEVICE_MAX_LIT_PANEL_WINDOWS cap the header documents, so every lit-panel element read walks all ~40 windows
🟠 Major (2)
| Where | Finding |
|---|---|
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+Interaction.swift:642 |
🆕 Coordinate taps/long-presses/drags on a lit-panel request anchor at the booked window's real origin while offsetting by the panel-box origin |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerXCTestEventBridge.m:319 |
🆕 RunnerCopyLitPanelWindows ignores the AGENT_DEVICE_MAX_LIT_PANEL_WINDOWS cap the header documents, so every lit-panel element read walks all ~40 windows |
🟡 Minor (10)
| Where | Finding |
|---|---|
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+Interaction.swift:637 |
A lit-panel tap that fails to book its window computes coordinates from CGRect.null origin |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerRequestPanel.swift:79 |
elementSearchRoots treats an empty lit-window list as a valid search root set, so a foldable can report ELEMENT_NOT_FOUND instead of the typed refusal |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerRequestPanel.swift:80 |
🆕 Lit-panel element-search roots are cached for the whole request, so a window opened by a mutation is never searched |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+Interaction.swift:239 |
hasTappableFrame can hand TapPointPolicy a null window frame on a lit-panel request |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+TextEntry.swift:555 |
🆕 A lit-panel refusal during text-entry confirmation is reported as a text-typing failure, not the typed panel code |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerXCTestEventBridge.m:260 |
🆕 A lit-panel request whose window list holds only vanished windows books nothing and is answered APP_SCREEN_WINDOW_UNRESOLVED instead of a panel-typed code |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerXCTestEventBridge.m:306 |
🆕 RunnerCopyLitPanelWindows walks the whole window list, ignoring the cap the header says exists for this walk |
packages/platform-apple/src/core/display-inventory.ts:418 |
🆕 appleLitPanel checks activeDisplay but then picks the lit panel from power |
packages/platform-apple/src/core/runner-client.ts:27 |
🆕 runAppleRunnerCommand now has two entry points with different behaviour: the new wrapper in runner-client.ts (whi… |
packages/platform-apple/src/foldable/orientation.ts:33 |
🆕 A first multi-panel request that finds no lit panel permanently disables lit-panel routing for that device |
🔵 Nit (7)
| Where | Finding |
|---|---|
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerRequestPanel.swift:43 |
🆕 RunnerResolveLitPanelWindow discards the window frame it resolved and recomputes a panel box |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+CommandDispatch.swift:436 |
🆕 Reading requestLitPanel for the recording clip breaks the state's main-thread-only rule |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+CommandExecution.swift:36 |
🆕 Lit-panel booking is skipped for .screenshot only via a command comparison, so the list of surface commands is duplicated in the runner |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+TextEntry.swift:594 |
🆕 `return (queries.count == 1 ? |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+TextInputProbe.swift:180 |
🆕 Duplicated input-type list in textInputQueries |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerXCTestEventBridge.m:255 |
🆕 Unplaced-window rule is stricter than the ADR text (refuses a full-panel window found before an unplaced one) |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerXCTestEventBridge.m:294 |
🆕 RunnerResolveLitPanelWindow's log line prints "application screen resolved" for a lit-panel booking |
Review of 9f5d9469 · 03.11.2026 01:51 UTC
Apex Sentinel
Verdict: request_changes
Since the last review: 13 new · 1 still open · 12 no longer raised
14 finding(s): 1 major, 5 minor, 8 nit. Verdict: request_changes.
Required fixes
- litPanelScreenFrame picks the panel's orientation from the window's aspect, so a partial window books an off-panel reference box
🟠 Major (1)
| Where | Finding |
|---|---|
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerRequestPanel.swift:63 |
🆕 litPanelScreenFrame picks the panel's orientation from the window's aspect, so a partial window books an off-panel reference box |
🟡 Minor (5)
| Where | Finding |
|---|---|
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+Lifecycle.swift:54 |
🆕 Recording pump reads requestLitPanel from shared state instead of the panel its recordStart booked |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerAppScreenCapture.swift:110 |
🆕 Off-lit-panel capture refusals get the wrong hint text |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerRequestPanel.swift:73 |
elementSearchRoots treats a read that answers with no lit windows as a successful empty search |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+Interaction.swift:236 |
🆕 hasTappableFrame names the booked window but passes onScreenWindowFrame |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+SnapshotExecution.swift:172 |
🆕 New runner error code is a bare string literal in the refusal |
🔵 Nit (8)
| Where | Finding |
|---|---|
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerRequestPanel.swift:73 |
🆕 Element-search issues on a lit-panel read are held and thrown away, not re-recorded |
packages/platform-apple/src/foldable/orientation.ts:33 |
🆕 A lit-panel request re-reads the panel table on every request; the 5s back-off only ever applies to a device whose panel count is still unknown |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+TextEntry.swift:592 |
🆕 boundTextEntryInputQuery reads XCUIElementQuery.count outside any issue containment |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerRequestPanel.swift:68 |
🆕 One over-long line in litPanelScreenFrame breaks the file's wrapping |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerRequestPanel.swift:90 |
🆕 elementQueries runs one full-subtree query per lit window for every read |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+Interaction.swift:642 |
🆕 A lit-panel tap anchors at an infinite offset when the booked window belongs to another app instance instead of refusing |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+Snapshot.swift:209 |
🆕 Snapshot root is published after the walk in one route and before it in the other |
packages/platform-apple/src/foldable/orientation.ts:89 |
🆕 A cancelled panel-table read for a known foldable is reported as panel-table-unreadable instead of a cancellation |
Review of 83add297 · 02.11.2026 22:33 UTC
Apex Sentinel
Verdict: approve
Since the last review: 10 new · 3 still open · 11 no longer raised
13 finding(s): 5 minor, 8 nit. Verdict: approve.
🟡 Minor (5)
| Where | Finding |
|---|---|
packages/platform-apple/src/core/runner-client.ts:32 |
🆕 runAppleRunnerCommand recomputes the request timeoutMs, so a read-only surface command is resent with the short remaining budget |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+CommandExecution.swift:429 |
🆕 typeCommand swallows a thrown lit-panel refusal into a generic 'type produced no response' error |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerRequestPanel.swift:65 |
elementSearchRoots returns empty array, not a refusal, when zero windows match the lit display |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+Snapshot.swift:299 |
🆕 The lit-panel snapshot has no fallback when a window snapshot reports no displayID |
packages/apple-runner/src/runner-command-traits.ts:121 |
🆕 home is excluded from RUNNER_COMMAND_RESOLVES_SURFACE although the ADR names a home-screen gesture as the bug being fixed |
🔵 Nit (8)
| Where | Finding |
|---|---|
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+Interaction.swift:634 |
Interaction coordinate: lit-panel branch falls through to the shared path on iOS, duplicating the resolved-window logic |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerXCTestEventBridge.m:328 |
RunnerCopyLitPanelWindows walks the whole window list, ignoring the AGENT_DEVICE_MAX_LIT_PANEL_WINDOWS cap it exists to bound |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+SnapshotExecution.swift:169 |
🆕 litPanelBackendPinRefusal hardcodes the tree backend's rawValue instead of comparing to the enum |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerRequestPanel.swift:29 |
🆕 setRequestLitPanel's main-queue ordering guarantee is void when the command starts on the main thread |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerRequestPanel.swift:48 |
🆕 Mid-request app switch silently replaces the booked lit-panel surface |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+CommandExecution.swift:25 |
🆕 executePrepared books the lit-panel surface for every command except screenshot (RunnerTests+CommandExecution.swi… |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+Snapshot.swift:331 |
🆕 The panel filter's display read is a second, looser implementation of the bridge's display-ID lookup |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+TextInputProbe.swift:22 |
🆕 Text-input queries are built outside the probe's issue scope, so containment rests on XCUITest query laziness |
Review of 72bc4411 · 02.11.2026 21:55 UTC
Apex Sentinel
Verdict: request_changes
Since the last review: 9 new · 5 still open · 6 no longer raised
14 finding(s): 1 major, 8 minor, 5 nit. Verdict: request_changes.
Required fixes
- The lit-panel window walk reads every window's frame outside the text-input probe's issue containment
🟠 Major (1)
| Where | Finding |
|---|---|
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerRequestPanel.swift:60 |
The lit-panel window walk reads every window's frame outside the text-input probe's issue containment |
🟡 Minor (8)
| Where | Finding |
|---|---|
packages/platform-apple/src/core/runner-client.ts:27 |
🆕 Every surface command on a known foldable now pays a fresh devicectl/guest panel probe, up to 5s of its own timeoutMs |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+Interaction.swift:232 |
🆕 hasTappableFrame now fails open on a lit-panel request whose window cannot be booked |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+TextEntry.swift:536 |
lit-panel refusals swallowed by try? fall back to app-wide reads that may be on the dark panel |
packages/platform-apple/src/foldable/orientation.ts:66 |
readFoldablePanelTable keeps a permanent single-panel verdict from a failed first read on a device whose panel count was never read |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+Snapshot.swift:206 |
🆕 Collapsed-tab fallback nodes are still gathered from the whole application tree on a lit-panel snapshot |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerXCTestEventBridge.m:302 |
🆕 RunnerCopyLitPanelWindows ignores the AGENT_DEVICE_MAX_LIT_PANEL_WINDOWS cap its own constant documents |
packages/platform-apple/src/core/runner-client.ts:22 |
litPanel is added only by the platform-apple runner-client wrapper, not at the apple-runner contract entry |
packages/platform-apple/src/foldable/runner-panel.ts:33 |
🆕 A cancelled panel-table read is reported as panel-table-unreadable |
🔵 Nit (5)
| Where | Finding |
|---|---|
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerRequestPanel.swift:60 |
elementSearchRoots caches the lit-panel window list for the whole request, so a mid-request fold answers from the dark panel |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerRequestPanel.swift:22 |
🆕 New RunnerLitPanelFact fields are decoded but never validated against the units the runner compares them in |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+Interaction.swift:629 |
🆕 interactionCoordinate's #if os(iOS) block was reopened, so the lit-panel branch silently changes the macOS path structure |
packages/apple-runner/src/runner-command-traits.ts:95 |
🆕 RUNNER_COMMAND_RESOLVES_SURFACE is positioned between the traits table and the accessors that read it |
packages/platform-apple/src/__tests__/interactor-foldable-runner-requests.test.ts:33 |
🆕 The new test file's leading comment is a single five-line sentence |
Review of 0c2a02ed · 02.11.2026 19:44 UTC
Apex Sentinel
Verdict: request_changes
Since the last review: 6 new · 5 still open · 103 no longer raised
11 finding(s): 1 major, 6 minor, 4 nit. Verdict: request_changes.
Required fixes
- Lit-panel capture treats an unreadable window list as
APP_SCREEN_WINDOW_UNRESOLVED, which falls back to SpringBoard — the dark-panel capture the ADR says must fail
🟠 Major (1)
| Where | Finding |
|---|---|
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerXCTestEventBridge.m:247 |
Lit-panel capture treats an unreadable window list as APP_SCREEN_WINDOW_UNRESOLVED, which falls back to SpringBoard — the dark-panel capture the ADR says must fail |
🟡 Minor (6)
| Where | Finding |
|---|---|
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+TextEntry.swift:536 |
🆕 probeTextInputs / resolveTextEntryInput now throw away a lit-panel refusal, reading it as no input |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+CommandExecution.swift:30 |
Lit-panel booking now gates system-surface commands (backSystem, appSwitcher, keyboardDismiss) on the session app having a lit-panel window |
packages/platform-apple/src/foldable/orientation.ts:81 |
The panel-table back-off caches a transient failure as "single-panel" and silences the lit panel for 5s |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerRequestPanel.swift:60 |
🆕 Lit-panel window list is cached for the whole request, so a window that appears mid-request is invisible to later reads |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+SnapshotExecution.swift:172 |
🆕 A stale private-ax backend pin now aborts every Apple settle/selector capture on a foldable |
packages/platform-apple/src/core/runner-client.ts:27 |
host-bound runAppleRunnerCommand callers bypass lit-panel booking entirely |
🔵 Nit (4)
| Where | Finding |
|---|---|
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerXCTestEventBridge.m:318 |
RunnerCopyLitPanelWindows walks every window with no cap, the hazard AGENT_DEVICE_MAX_LIT_PANEL_WINDOWS exists to bound |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerRequestPanel.swift:60 |
🆕 A selector/point read of an app whose windows are all off the lit panel reports "no match" instead of APP_SCREEN_OFF_LIT_PANEL |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+SnapshotExecution.swift:172 |
🆕 litPanelBackendPinRefusal message can claim a pin to the same tree backend it requires |
packages/platform-apple/src/core/runner-client.ts:42 |
🆕 Exhausted panel-read budget is turned into a 1 ms runner timeout instead of a typed refusal |
Review of 04bd1109 · time not recorded
Apex Sentinel
Verdict: request_changes
Since the last review: 14 new · 1 still open · 93 no longer raised
15 finding(s): 1 major, 7 minor, 7 nit. Verdict: request_changes. reached max rounds (1).
Required fixes
- A stale litPanel on the session app's command hard-refuses taps for a single-panel (or non-foldable) target served in place
🟠 Major (1)
| Where | Finding |
|---|---|
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+CommandExecution.swift:16 |
🆕 A stale litPanel on the session app's command hard-refuses taps for a single-panel (or non-foldable) target served in place |
🟡 Minor (7)
| Where | Finding |
|---|---|
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+SnapshotCapturePlan.swift:165 |
🆕 A raw or backend-pinned snapshot on a foldable is not narrowed to the tree tier, yet still filters every window by displayID — so it can always refuse with APP_SCREEN_UNRESOLVED |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+Interaction.swift:261 |
🆕 Lit-panel route silently falls back to app.frame when the booking fails, the fallback the amendment says it removes |
packages/platform-apple/src/core/runner-client.ts:32 |
The panel probe is added in front of the runner timeout instead of taken from it, so a surface command can spend 2x timeoutMs |
packages/platform-apple/src/foldable/orientation.ts:32 |
🆕 Panel-table back-off is shared with the rotation route, so one failed probe also disables guest-HID routing for 5s |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerRequestPanel.swift:32 |
🆕 Panel booking is keyed to no app, so a later read for a different process reuses the first app's window |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+CommandDispatch.swift:431 |
🆕 A fold during a recording makes every pump frame refuse on the lit panel, and the pump drops them silently: recordStop reports success on an empty clip |
packages/platform-apple/src/foldable/orientation.ts:69 |
🆕 A lit-panel read that fails on a caller deadline backs off every surface command for 5s, even while rotations keep succeeding |
🔵 Nit (7)
| Where | Finding |
|---|---|
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+Snapshot.swift:214 |
🆕 Snapshot sets the root rect to the panel's native box while its windows keep turned frames |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerRequestPanel.swift:32 |
🆕 Booked lit-panel XCUIElement is reused across the whole request; a vanished window later records an XCTest issue (#2995) |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+Models.swift:222 |
🆕 New RunnerLitPanelFact decodes non-numeric JSON fields as 0, silently taking the single-panel route |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+Snapshot.swift:207 |
🆕 Unused let on the lit-panel snapshot path |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+SnapshotCapturePlan.swift:165 |
🆕 plan shadows its own parameter in the new lit-panel override |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerXCTestEventBridge.h:38 |
🆕 RunnerLitPanel and the panel-fact struct duplicate the same three fields with two spellings |
apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerXCTestEventBridge.m:191 |
🆕 RunnerLitPanelWindowScreen logs on every call with no mute and reads two KVC keys per window |
![apex-sentinel-staging[bot]](https://avatars.githubusercontent.com/in/5196457?s=60&v=4){kind=small}
There was a problem hiding this comment.
Apex Sentinel: 19 new findings on 37db43f. The verdict and the full list are in the summary comment.
37db43f to
563e4e8
Compare
There was a problem hiding this comment.
All reported issues were addressed across 9 files (changes from recent commits).
Reply with feedback, questions, or to request a fix.
View guided diff | Turn on auto-fix | Re-trigger cubic
There was a problem hiding this comment.
Apex Sentinel: 19 new findings on 563e4e8. The verdict and the full list are in the summary comment.
|
At 563e4e8 the lit-panel fix is not ready: two device-facing paths changed at this head have no live run, and several earlier problems still stand. CI is green (20 checks, none failing), and there are no conflicts. Before merge, the Duo evidence below must be posted and the open P1 threads fixed. The graft was rewritten (it no longer throws, it copies The snapshot-tree rewrite (root republish, off-panel surface drop, SpringBoard window graft) is a second concept on top of the surface-window fix. It spans Not blocking, and you can take or leave these: (1) Could the PR be smaller? Send These open threads still apply and need a fix at their owners: the abort on the devicectl path that parks the device for 5s (#4122 (comment)), the incomplete-table lit panel (#4122 (comment)), the single-slot panel read (#4122 (comment)), the main-thread graft watchdog (#4122 (comment)), the off-panel refusal gap (#4122 (comment)) and the cap of 8 (#4122 (comment)). The P2s at #4122 (comment) (synthesized dispatch display), #4122 (comment) (runbook lines removed), #4122 (comment) (absent panel wording), #4122 (comment) (probe before litPanel check), #4122 (comment) (requestId cancellation) and #4122 (comment) (selection=3 overreach) also hold, as do about 25 lower-priority threads and nits. Please resolve these threads, which no longer apply: #4122 (comment) (readText is now in the set and has a fixture), #4122 (comment) (the no-panel walk matches main's loop), #4122 (comment) (the graft is non-throwing now), #4122 (comment) (readText is in the set), #4122 (comment) (reportedCount is written before the cap), #4122 (comment) (sliceLeft is taken after the resolve), #4122 (comment) (seen is copied, rollback discards it), #4122 (comment) (redundant branch removed), #4122 (comment) (factCount matches the ADR), #4122 (comment) (absent litPanel takes the historic branch), #4122 (comment) (the probe runs once before the retry loop), #4122 (comment) (same sliceLeft ordering), #4122 (comment) (the catch rolls back every graft node), #4122 (comment) (each attribute is read once), #4122 (comment) (home uses pressHomeButton) and #4122 (comment) (all producers use the wrapper). I could not run a Duo simulator and saw only the PR-body log excerpts, not the raw runner.log. I did not run the Swift unit tests or |
563e4e8 to
32e56a4
Compare
There was a problem hiding this comment.
All reported issues were addressed across 18 files (changes from recent commits).
Reply with feedback, questions, or to request a fix.
View guided diff | Turn on auto-fix | Re-trigger cubic
There was a problem hiding this comment.
Apex Sentinel: 15 new findings and 1 moved to a new line on 32e56a4. The verdict and the full list are in the summary comment.
32e56a4 to
5f2be24
Compare
There was a problem hiding this comment.
All reported issues were addressed across 8 files (changes from recent commits).
Reply with feedback, questions, or to request a fix.
View guided diff | Turn on auto-fix | Re-trigger cubic
There was a problem hiding this comment.
Apex Sentinel: 16 new findings on 5f2be24. The verdict and the full list are in the summary comment.
|
I reviewed 5f2be24. The earlier findings from the 563e4e8 review are mostly fixed, but two defects remain, and the required Smoke Tests check is red. The macOS phase of Smoke Tests (run 41220653930) fails in the runner unit suite: 390 tests, 2 assertion failures, both in testSurfaceWindowWalkRefusesAWindowOnTheLitPanelWhenTheListWasCutShort (:478 and :482). That test exercises the new guard, so this failure comes from the diff. The iOS phase passed, with flaky retries on 01-settings.ad and gesture-pan-duration.ad. I could not run a Duo simulator, and I did not run the Swift tests locally. The new guard at RunnerXCTestEventBridge.m:348 can never run. It sits inside The truncation downgrade at :340 also stops The evidence the earlier review asked for is still missing (RunnerTests+Snapshot.swift:355). The PR body has no Could a smaller design remove all of this? The design question in the earlier review got no reply, and this delta adds more patches around the same concepts: a second truncation downgrade (dead, as above), a bypass flag on the retry memo, an optional panel parameter that falls back to the slot, and an Not blocking, and fine to take or leave: the delta joined Open threads: these still apply and should stay open, P1 #4122 (comment) (one panel slot) and #4122 (comment) (main-thread graft), P2 #4122 (comment) (missing xcrun), #4122 (comment) (guard cannot run) and #4122 (comment) (no-panel wording). About 40 lower-priority threads also still apply. These do not apply at 5f2be24 and can be resolved. Fixed in code: #4122 (comment) #4122 (comment) #4122 (comment) #4122 (comment) #4122 (comment) #4122 (comment) #4122 (comment) #4122 (comment) #4122 (comment) #4122 (comment) #4122 (comment) #4122 (comment) #4122 (comment) #4122 (comment) #4122 (comment) #4122 (comment) #4122 (comment) #4122 (comment) #4122 (comment). Answered or benign: the graft does not throw and rollback drops its keys #4122 (comment) #4122 (comment) #4122 (comment) #4122 (comment) #4122 (comment) #4122 (comment) #4122 (comment) #4122 (comment); readText, screenshot, home and the wrapper are already covered #4122 (comment) #4122 (comment) #4122 (comment) #4122 (comment) #4122 (comment) #4122 (comment) #4122 (comment) #4122 (comment); the count and else-if comments no longer match the code #4122 (comment) #4122 (comment) #4122 (comment) #4122 (comment); replay, probe and abort paths are correct #4122 (comment) #4122 (comment) #4122 (comment) #4122 (comment); the no-panel walk stops at the first measurable window #4122 (comment) #4122 (comment) #4122 (comment). |
5f2be24 to
1fbcbfa
Compare
There was a problem hiding this comment.
Apex Sentinel: 15 new findings and 1 moved to a new line on 1fbcbfa. The verdict and the full list are in the summary comment.
|
This is a follow-up on the earlier review (#4122 (comment)), now at 1fbcbfa. The two failing unit assertions are fixed, and a truncated or unplaceable lit-panel window now refuses instead of booking the strip. I checked this by reading the fixture and resolver, not by running the Swift runner suite. Smoke Tests is still running. Its macOS phase runs the resolver this change touches, so its result counts for this PR. There are no conflicts. One defect remains. The new refusal at https://github.com/callstack/agent-device/blob/1fbcbfa/apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerXCTestEventBridge.m#L341 still turns an off-lit-panel walk into unresolvedScreen when a window is unplaceable or the list is truncated. A SpringBoard on the Duo names 39-43 windows against the cap of 8, so the fallback is always truncated. selectObservedScreenCapture only moves on for unresolvedWindow or offLitPanel, so an app-side downgrade skips the system surface, and a SpringBoard-side downgrade ends in APP_SCREEN_UNRESOLVED. On the Duo, APP_SCREEN_OFF_LIT_PANEL can then never reach the host from the fallback. A capture, screenshot, keyboard or recording sample that SpringBoard could serve fails with "did not report the display it is shown on" and a misleading retry hint. The rule: a walk that found nothing on the lit panel must still let the caller ask the system surface, and only the final answer may weaken its claim. The smallest fix is to uncap the panel walk (stop at the first exact panel-box match, otherwise read the whole list). That removes listTruncated and AGENT_DEVICE_MAX_PANEL_SURFACE_WINDOWS. Keep returning offLitPanel from the walk, and apply the unplaceable downgrade only to the fallback result in selectObservedScreenCapture. Please add a unit test with more than 8 windows, all off the lit panel, that asserts the selectObservedScreenCapture result. Invert testSurfaceWindowWalkReportsADisplayRefusalWhenAnUnreadableWindowRemains to match. The PR body still has no live output for the graft, the tree republish or readText by coordinate (https://github.com/callstack/agent-device/blob/1fbcbfa/apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests%2BSnapshot.swift#L355). After the fix above, please run Not blocking: these can be taken or left. The new unplaceable-window refusal has no walk case that combines ScreenlessWindowFixture with innerPanelStripWindow (RunnerTests+XCTestEventBridgeTests.swift:467), so one case with [ScreenlessWindowFixture(), innerPanelStripWindow()] expecting .unresolvedScreen and a nil window would cover it. Open threads from other reviewers. These still apply: the one-slot overwrite in command dispatch (#4122 (comment)), the main-thread graft read (#4122 (comment)), the ADR 0025 wording (#4122 (comment)), and the missing xcrun in device-foldable.md (#4122 (comment)). 41 lower-priority threads also still apply, and eight new ones from this commit are not traced yet. These do not apply and can be resolved: 4431955384 (graft does not throw out of the tier), 4431955390 (readText is already in the surface-resolving list), 4431955397 and 4432325679 (reported= is the pre-cap count, so listTruncated is live), 4431955428 (the else-if is gone), 4431955430 (placed counts only display-named windows), 4431955435 (replay path is correct), 4431955437 (probe path is correct), 4431955445, 4432076445, 4432076450 and 4432325666 (rotation route fixed, retry memo no longer gates rotation), 4432076435 and 4432325683 (fixed: observedFrame is chosen.frame), 4432309422 (fixed: guard runs before the failure mapping), 4432325668 and 4432465773 (a truncated or unplaceable litDisplayWindow now refuses), 4432325690 and 4432485202 (no-panel walk stops at the first measurable window), 4432485215 (abort path is correct), 4431955402, 4432076412, 4432076421, 4432076426, 4432076431 and 4432325670 (benign), 4431955408 (rollback drops its keys), and 4432076437, 4432076439, 4432076443, 4432325687, 4432485192 and 4432485206 (already covered). All of these are under #4122 (comment). Before merge, please fix the off-lit-panel downgrade with its test, post the Duo snapshot and readText output, and get Smoke Tests green. |
There was a problem hiding this comment.
All reported issues were addressed across 28 files
Reply with feedback, questions, or to request a fix.
View guided diff | Turn on auto-fix | Re-trigger cubic
There was a problem hiding this comment.
Apex Sentinel: 10 new findings and 1 moved to a new line on bf68d70. The verdict and the full list are in the summary comment.
There was a problem hiding this comment.
All reported issues were addressed across 34 files
Reply with feedback, questions, or to request a fix.
View guided diff | Turn on auto-fix | Re-trigger cubic
There was a problem hiding this comment.
All reported issues were addressed across 3 files (changes from recent commits).
Reply with feedback, questions, or to request a fix.
View guided diff | Turn on auto-fix | Re-trigger cubic
There was a problem hiding this comment.
Apex Sentinel: 14 new findings and 1 moved to a new line on 04bd110. The verdict and the full list are in the summary comment.
|
Most of the earlier findings (#4122 (comment)) are fixed:
The host still fails open when the panel read is unresolved (orientation.ts#L84). After a devicectl failure, or a read cut short by the 5 s clamp, The live evidence the earlier review asked for is still absent for the paths in RunnerAppScreenCapture.swift#L72. It is unknown whether a Duo home-screen Not blocking, and fine to take or leave: the snapshot booking in One P1 thread still applies, on the shared panel slot being overwritten: #4122 (comment). 14 lower-priority threads still apply. The earlier threads below no longer apply and can be resolved. Fixed at this head: #4122 (comment), #discussion_r4431929847, #discussion_r4431929864, #discussion_r4431929869, #discussion_r4431929873, #discussion_r4431929878, #discussion_r4431929881, #discussion_r4431929892, #discussion_r4431929899, #discussion_r4431929909, #discussion_r4432062431, #discussion_r4432309418, #discussion_r4432309422, #discussion_r4432309424, #discussion_r4432309434, #discussion_r4432309440, #discussion_r4432465773 (same PR; the number is the discussion id). Graft code deleted: #discussion_r4431929859, #discussion_r4431929888, #discussion_r4431955384, #discussion_r4431955402, #discussion_r4431955408, #discussion_r4431955423, #discussion_r4432062434, #discussion_r4432076412, #discussion_r4432076421, #discussion_r4432076426, #discussion_r4432325670, #discussion_r4432325675, #discussion_r4432325696, #discussion_r4432325699, #discussion_r4432485194, #discussion_r4432747462, #discussion_r4432747466. Old resolver, surface or tree code deleted: #discussion_r4431955410, #discussion_r4431955413, #discussion_r4431955420, #discussion_r4431955424, #discussion_r4431955428, #discussion_r4431955430, #discussion_r4432076415, #discussion_r4432076431, #discussion_r4432076435, #discussion_r4432076457, #discussion_r4432076464, #discussion_r4432325677, #discussion_r4432325683, #discussion_r4432325692, #discussion_r4432325705, #discussion_r4432325711, #discussion_r4432325717, #discussion_r4432325719, #discussion_r4432485199, #discussion_r4432485217, #discussion_r4432747436, #discussion_r4432747450, #discussion_r4432747453, #discussion_r4432747458, #discussion_r4432747460. Back-off is not read on those paths, or abort handling is correct: #discussion_r4431955392, #discussion_r4431955445, #discussion_r4431955449, #discussion_r4432076445, #discussion_r4432076450, #discussion_r4432325666, #discussion_r4432485192, #discussion_r4432485208, #discussion_r4432485211, #discussion_r4432485215, #discussion_r4432747448, #discussion_r4433338782. No-panel route unchanged from main, or the booking refuses before acting: #discussion_r4431929885, #discussion_r4431929905, #discussion_r4431955390, #discussion_r4431955397, #discussion_r4431955435, #discussion_r4431955437, #discussion_r4432076417, #discussion_r4432076437, #discussion_r4432076439, #discussion_r4432076460, #discussion_r4432325668, #discussion_r4432325679, #discussion_r4432325687, #discussion_r4432325690, #discussion_r4432485188, #discussion_r4432485197, #discussion_r4432485201, #discussion_r4432485202, #discussion_r4432485204, #discussion_r4432485206, #discussion_r4432485220, #discussion_r4432485223, #discussion_r4432747440, #discussion_r4432747449, #discussion_r4433338770, #discussion_r4433338775, #discussion_r4433338811, #discussion_r4433338818. |
04bd110 to
dd5236e
Compare
…comments Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…earch The bound-input identifier query and the in-app back control searched the whole application, so on an unfolded Duo a same-identifier input or back button on the dark panel could answer. Both now search the element search's roots, and its refusal propagates: text-entry resolution throws instead of answering nil, and boundTextEntryInputIsGone answers "cannot prove gone". One owner returns the request's remaining budget, so the panel read and the runner timeout share one clock. The unreachable hasTappableFrame guard is removed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…trip On the unfolded Duo, an app's window is 871x669 beside the inner panel's status strip, and the booking used that window frame as the synthesized dispatch reference. The orientation flip then mapped points against 871 instead of the 951-point screen, so taps landed 80 points off, and a fill's focus tap missed. The booked frame is now the panel's box turned like the window, at its origin, which is unchanged for full-panel windows. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… reads The synthesized commit wait's observe closure and the replacement clear's read swallowed a lit-panel refusal with `try?`, so the wait timed out or the clear reported TEXT_INPUT_CLEAR_NOT_OBSERVED. Both now throw, and the refusal answers the request. The clear read also keeps main's answer for an input that is simply gone. The bridge exports one lit-panel resolver that returns the window, frame and screen, and lit-panel comments are trimmed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The booked box was inferred from the window: its origin and whether it was wider than tall. A window that the inner panel's status strip pushes to a non-zero origin would then book a box 80 points past the screen, and every tap, reference and snapshot root would land off. The box is now the panel's native portrait box laid out in the app's interface orientation, at the screen origin, through the existing soleScreenReferenceFrame. The XCUICoordinate anchor keeps the window's own origin. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…nslation The previous commit booked the screen box, which broke the placed-window translation: XCTest reports a window at (0,80) from its own top-left, and the snapshot shifts those frames into screen space only when the viewport is that window's frame. The booking is the window frame again, as for a placed iPad window (#3899), and on a lit-panel request the synthesized turn is about the lit panel's screen in the interface orientation instead of the window's shape. That replaces the aspect guess and the window origin inference. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…e walk cap The rotation-basis case now books the launched app's window under a lit panel taller than it and checks the synthesized reference and turn box, so a wiring regression fails it. The one-caller screen-frame helper is inlined. The window cap's doc names the booking walk it bounds; the element-search copy stays uncapped so every lit window is searched. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… cap Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
This branch now conflicts with main. Please rebase onto main and resolve the conflicts; the review at 0684301 (collapsed-tab candidates and the overlapping dark-window case) still applies to the rebased head. |
0684301 to
4d6f523
Compare
|
Rebased onto main Fixed the remaining collapsed-tab read. Regression proof at rebased baseline Native evidence on iPhone Duo Fresh iOS unit-runner, iOS production, and macOS production builds passed.
One lease-protected push. Earlier live evidence keeps its original commit attribution. Fresh hosted coverage/native/integration checks and reviews remain required. No merge or labels. |
There was a problem hiding this comment.
Apex Sentinel: 4 new findings and 1 moved to a new line on 4d6f523. The verdict and the full list are in the summary comment.
|
Current-review follow-up at Confirmed correctness fix: Regression proof on iPhone Duo (fixtures), iOS 27.1, The test exercises the real Objective-C exception catcher and native booking owner with fault-injection window objects on the simulator. No natural live window-read exception is claimed. No manual app session or daemon was opened; the test device was shut down and the temporary checkout removed. Logs and results were retained in the parent checkout's Other current findings were checked against this head:
No replies to Apex threads were posted. Exact-head affected-gate evidence follows below. Prior live evidence stays attributed to its original SHA. Final exact-head validation: 22 selected checks: 11 passed, 10 skipped, 1 deduplicated. Passed: format, lint, typecheck, layering, fallow, native-dupes, build, vitest-related, unit, xctest-selection, packaged-runner-swift. Related Vitest: 1,059 passed, 0 failed, 0 skipped. Unit Vitest: 17,156 passed, 0 failed, 1 skipped. No failed modules or unhandled errors. The provider-integration check was deduplicated into related tests. Coverage, native iOS/macOS, web smoke, Linux evidence and replay lanes retain their receipt skip reasons (GitHub authority or parked manual workflows); they are not local passes. Publication uses the validated commit with an explicit lease against |
|
The code looks good at 4d6f523. The collapsed-tab read now goes through the proven-lit query owner, so the earlier query thread is fixed, and the newer inline threads on this head either no longer apply or describe behavior the ADR requires. All 19 checks pass and there are no conflicts. The published Duo run covers the scoped query and acquisition at this head. The overlapping dark-window case is covered by a native fixture that reaches the changed production functions, but it was not reproduced live or through the daemon. This review covers 4d6f523. The newer head ec80eec adds a refusal for an unread lit-panel window frame with a unit test, and the next pass will look at it. |
|
Package-growth assessment at
Vercel deployment is ready. No source changes were needed for these reports. Hosted checks/reviews remain pending; this is not a merge-ready claim. |
|
Quality and review-thread pass completed at Independent read-only review of the change since All six current inline findings now have code-linked replies and are resolved:
No source edit, push, or test rerun was warranted for these dispositions. GitHub reports MERGEABLE / CLEAN at this head: 17 checks successful, one cache-matrix check skipped, none failed or pending. The earlier exact-head affected receipt The implementation and current review concerns have no identified blocker in this pass. Final human approval and merge remain the maintainer's decision; no merge or labels were performed. |
|
The code looks good at ec80eec. The new commit refuses an unread lit-panel window frame through the same booking owner, and the collapsed-tab query fix from the earlier review is still in place. All checks pass and there are no conflicts. The inline threads on this head either no longer apply or describe behavior the ADR requires, so they can be resolved. The live evidence covers the Duo query, booking and daemon delivery, and the new refusal is covered by your native fault-injection run; a live cross-display overlap was not reproduced. |
thymikee
added
the
ready-for-human
Summary
Foldable surface commands name the lit panel and book a window by app bundle and request UUID. Capture, interaction, selectors and text probes reuse proven-lit roots. Typed refusals prevent outer-app fallback.
Required window-read exceptions now return
APP_SCREEN_UNRESOLVED; genuinely absent windows retainAPP_SCREEN_WINDOW_UNRESOLVEDand are not asked for frames.70 files in the full PR; this round changes two (+18/-1). No linked issue supplied.
Validation
Head
ec80eecb1e0251d51d93aa68071c88577f1ea7ba; basec8ad8fb1257787ce0973cf74271cb9a4fafb51d9.4d6f523plus the new native test fails five assertions across unreadable-only, unreadable-before-lit and unreadable-before-dark lists. At head, six selected native tests pass, including absent-window and ordinary booking controls.pnpm check:affected --run --head ec80eecbpassed; receipt.tmp/validation/run-RD6UYM/receipt.json,exactCommit=true,checkoutChanged=false; 11 passed, 10 skipped, 1 deduplicated. Vitest: 18,215 passed, 1 skipped, 0 failed.4d6f523; earlier live evidence retains its attribution. The exception regression uses native fault injection on iOS 27.1 Duo. No natural exception is claimed.Hosted checks at this head: 17 succeeded, one optional matrix lane skipped, none failed. Human review approval remains outstanding.