Repository navigation
refactor(ios): decide bridge-failure scope and warning once - #3914
Merged
Merged
Conversation
The capture-scoped bridge refusals were classified twice: once to decide whether the app generation's circuit opens, and again to pick the fallback warning sentence. A new refusal could keep the circuit and still publish a generation-wide warning, or the reverse (#3874, #3895). One classifier in snapshot-route-failure-policy.ts now returns a single disposition that the route consumes for both decisions; the standalone constants, the refusal set, and the duplicated warning branches are deleted.
The guest failure code is an open string. An object-literal table would answer a code that names an inherited property, so the lookup uses a map and every unknown code keeps the conservative generation default.
Member
Author
|
Validation receipts:
|
Contributor
Size Report
Startup median (7 runs, lower is better):
|
Member
Author
|
This PR is ready to merge. I reviewed 0a58d5b and found nothing that needs to change. All 16 checks are green at that commit, and there are no conflicts. The live simulator runs in the PR body have no attached logs, but the unchanged route tests cover this behavior-preserving refactor, so that is enough. I did not run the tests locally. Not blocking, and you can take or leave these: |
thymikee
added
the
ready-for-human
Contributor
|
thymikee
added a commit
that referenced
this pull request
Oct 22, 2026
The refusal wording belongs to the snapshot-route failure policy, not to this fix. Refs #3899
thymikee
added a commit
that referenced
this pull request
Oct 24, 2026
* origin/main: fix(android): bound the dumpsys reads behind dialog and foreground checks (#3936) feat(android): detect Fire OS devices as Android TV targets and document Fire TV support (#3923) docs(macos): document Electron/Chromium app automation through the accessibility tree (#3932) test(capture-kit): encode solid test PNGs unfiltered to stay inside the unit budget (#3933) fix: guard bounded lifecycle and claim recovery (#3911) fix(ios): clear a long text value wherever the caret lands (#3908) fix(snapshot): stop iOS 27 floating-bar containers from covering content (#3912) ci: cache and bound the Linux desktop dependency install (#3919) fix: refuse resource operations in bounded composites (#3910) refactor(ios): decide bridge-failure scope and warning once (#3914) refactor(capture-kit): encode post-gesture stabilization baseline provenance (#3915) perf: expect the persistent session on Android snapshot-raw (#3913) test(ios): hold the fill read-count pin to the route it measures (#3909) perf(ios): admit tap text-input probes from fresh target evidence (#3881) feat(perf): time every Android command against the fixture app (#3902) chore: mark 0.21.31 as released
thymikee
added a commit
that referenced
this pull request
Oct 25, 2026
* fix(ios): publish a windowed iPad app's capture in screen space (#3899) An iPad window that does not fill the screen reports its own screen box while its subtree arrives anchored at its own top-left. The capture published a screen-space viewport around window-relative rects: the fold clipped content to (w - x) x (h - y), a rect named no screenshot position, and a synthesized tap landed at the point minus the window origin. SnapshotContentTranslation reads the one vector from a carrier's (0,0)-anchored copy of the viewport box to the viewport itself, and SnapshotGeometrySpace.normalized translates the subtree it declares in the same pass that turns a quarter-turned subtree. CoordinateSpaceRotation now turns axes only; the digitizer origin stays where the display put it. Replayed from contracts/fixtures/window-coordinate-space.json in both languages. Refs #3899 * revert: keep the snapshot-route warning owned by #3914 The refusal wording belongs to the snapshot-route failure policy, not to this fix. Refs #3899 * fix(ios): read a placed window's snapshot frames at their screen place (#3899) XCUIElement.frame reports a placed window's content in screen space, but XCUIElementSnapshot.frame reports it in the window's own content space, so the text-input probe saw two spaces on an iPad Windowed Apps window and refused a fill whose point was on the screen. The probe, the focus tap, and the typing reader now read the same SnapshotContentTranslation vector the capture reads, from the two live boxes that name it. Refs #3899 * fix(ios): keep caret repositioning and keyboard evidence decision stable on a placed window [ #3899 ] * fix(ios): translate only snapshot frames and pin the unmeasured landscape case [ #3899 ] * refactor(ios): keep the window-content translation with the capture that runs it [ #3899 ] * refactor(ios): make ResolvedTextEntryInput own the probe-side screen frame (#3899) The readiness retap re-read the input as a bare XCUIElementSnapshot and tapped its raw center through tapAt, which takes screen points, so on a windowed iPad the retap landed outside the window. The live window lookup also ran whole-app queries four to six times per fill on every iOS device. Every snapshot frame that becomes a dispatched point now passes through the one content translation, held on ResolvedTextEntryInput.frameOnScreen. The vector is read once per command and stored on TextEntryTarget.contentTranslation; the probe matching, focus tap, refresh point, caret tap and readiness retap all read that one property. The threaded translation: parameters and the per-read lookups are gone, and the once-per-command keyboard read in the typing command is reverted to the base behavior with the retap fixed. * docs(ios): name the probe-side owner and the screen-space dispatch basis (#3899) Amends ADR 0004 to name ResolvedTextEntryInput as the probe-side owner of the content vector, cites the measured Add-button rect at (659,321), records that the runner republishes a windowed capture in screen space where the bridge-refusal warning said it was offset, and names screen space plus the app.frame origin assumption in the interactionCoordinate comment. * refactor(ios): cut the screen-space comments and the duplicated frame offset (#3899) The vector on the target replaces the offset arithmetic in textEntryRefreshPoint, which read the snapshot frame and added the vector itself. One owner now: ResolvedTextEntryInput.frameOnScreen. The comments that restated the capture's rule at every call site are cut to the constraint the code cannot show. SnapshotCoordinateSpace.swift ships in the npm package without comment stripping, so its prose lives in ADR 0004 and the golden table instead; the same rule text is removed from the TypeScript decoder's foot note. * refactor(runner): make the text-entry content vector a non-defaulted tri-state frameOnScreen becomes optional and carries the #3060 rule: it names no box without a snapshot or under an unknown content space, so the four consumer guards collapse into one nil check. liveContentTranslation returns TextEntryContentTranslation (known/none/unknown) instead of CGSize?, so a failed window read can no longer look like full screen and send the readiness retap to a raw window-local point. The vector now rides on TextInputAtPoint, cutting one parameter from the probe signatures. retapAt skips the point on unknown; probe matching and the focus tap fall back to element.frame. Adds the unknown-space retap unit test. * feat(runner): log the readiness retap point and its skips * fix(runner): take the probe vector from the resolved window box alone Reply to review 5642729478: - liveContentTranslation builds the zero-origin reported box from the resolved window's size, so an app.frame that unions transformed subtrees cannot silently cost a placed window its vector (P1 thread 1). - Synthesized dispatch turns about the display's oriented box, never a placed window's smaller box; landscape under Windowed Apps stays unmeasured and the ADR says so (P1 thread 2). - requireTextInputAtPoint passes the live vector (P2 thread 3). - ADR 0004 distinguishes the refusing bridge from the converting runner and states the focus-tap/retap refusal rule (P2 threads 4-5). - The initial focus tap dispatches the caller's requested point when the read names no screen box; the live element.frame is never dispatched, and focusTapPoint is pinned by a unit test (P2 thread 7). * refactor(runner): shrink the windowed-iPad screen-space change to its rule Keep the accepted design (one SnapshotContentTranslation rule in normalized, ResolvedTextEntryInput.frameOnScreen, a non-defaulted known/none/unknown vector read once per command, rotation about the sole display) and cut what did not serve it: cosmetic renames and reformatting, comments that restated code, the one-caller focusTapPoint helper, TextInputAtPoint's copy of the target's vector, the duplicated sole-display math, duplicate unit tests and fixture rows, and most of the ADR 0004 amendment prose. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * test(snapshot): pin translated hittability and the offset rule's empty-box guards Review follow-ups to the shrink: assert a translated node's hittability, add the two fixture rows that discriminate the positive-finite guards, keep windowContentCases names unique, read the display size from the fixed coordinate space, and correct ADR 0004's vector-read and gap wording. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(runner): restore the discriminating window-content rows and the retap skip reason Put back the moved-window, keyboard, and SpringBoard rows in windowContentCases, and log one skip line carrying its reason when the readiness retap names no screen point. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
One classifier,
bridgeFailureDisposition, now decides both whether a bridge failure opens the generation's circuit and which fallback-warning sentence the capture carries. Previously two owners classified the same failure —opensGenerationCircuitfor scope, string comparisons against duplicated constants for wording — and they could drift apart (#3874 and #3895 each had to teach both owners the same lesson). Behavior is unchanged for every reachable failure; unknown guest codes keep the conservative generation default.Touched files: 3. No scope expansion. Production lines:
snapshot-route.ts622 → 599, new policy module 54 — net +31 against the plan's aspirational 30–70-line reduction; the classification data moved out of the route without shrinking, in exchange for one owner of the decision. Gross diff 246 lines including tests.Validation
Tested SHA:
0a58d5bdab.pnpm typecheck,pnpm lint,pnpm formatpass;snapshot-route.test.tsunchanged, all route/policy/observability cases pass.check:affected --run --headreceipt recorded per head in comments.989804f11f^(pre-fix(ios): scope bridge ownership refusals to the capture #3874), the window case FAILs onba953a8066^(pre-feat(ios): route iPad Simulator snapshots through the host AX bridge #3895), all pass at head.warnings: null. iPad Stage Manager windowed — the refusal warning repeated identically across 6+ captures, zerocircuit-disabledin the session log.CI obligations: standard lanes. Unresolved risks: none known.
🤖 Generated with Claude Code