Repository navigation
feat(mobile): improve profile screen accessibility (Phase 2) - #12129
Conversation
|
E2E verification for
中文: |
|
Guard integration verification for pushed
Exact debug APK SHA256: 中文验收已推送 早先 18 项真实界面断言及 200% 横屏前后截图保留为原始功能证据,不重复计入当前套件数量。不宣称实际 TalkBack 语音、实体设备或完整恢复验收。APK 精确 SHA256 及基础回归修复/迁移/不支持时拒绝加载的验收链接见上文。仍为依赖草稿,需上游审查/工作流授权;本地通过不等于上游 CI 通过。 |
b636409 to
d53ad5f
Compare
|
Please do not rebase or force-push to an active PR as it invalidates existing review comments. Note for future reference, the bots always squash all changes into a single commit automatically as part of the integration. 中文请勿对活跃的 PR 执行 rebase 或 force-push,因为这会使已有的评审评论失效。另外,供日后参考:作为集成流程的一部分,机器人始终会自动将所有改动压缩(squash)为单个提交。 |
|
September 18 post-merge rebase verification: #11722 has merged. I removed its already-merged history and rebased this slice onto Fresh checks on this exact head:
Modern runs required profile isolation. The legacy provider cannot establish supported browser-profile operation: unsupported profile/clear cases are skips, not passes; it also cannot exercise the intermediate provider with profile support but no completed clear. Earlier screenshots and extra interactive/adversarial scenarios remain prior evidence for identical code, not fresh runs. Combined-feature and physical-device acceptance are not claimed. The shared JavaScript check and its two reproduced upstream-baseline limitations are detailed in the September 18 report on #12130; this native slice does not change those JavaScript sources. A successful full repository build is not claimed. This one-time rebase was needed to remove the merged scaffold history and resolve the resulting conflicts. I have noted the force-push reminder; further review fixes will be additive commits. Please review the focused slice when convenient; I am available to address feedback. 中文:#11722 已合并,本切片已变基并重新验证;实现文件与之前发布的版本逐字节相同。Android 构建和 lint 通过,JVM 21 项通过;API36 为 11 通过/1 跳过,API26 为 8 通过/4 跳过,跳过项未计入通过。合并顺序仍依赖 #12121,其他功能切片未组合验收。完整 JS 构建/集成类型检查仍有在上游基线复现的问题,详见 #12130 的本次说明;不会声称全部检查通过。已记录避免 force-push 的提醒,后续评审修改将追加提交。请维护者评审,我会继续处理反馈。 |
Local emulator verification of
|
| Host | macOS 26.6 arm64, Zulu JDK 17.0.20, Gradle 8.2.1 wrapper (jar SHA-256 matches gradle-wrapper.jar.sha256), AGP 8.2.0, Android Emulator 37.1.11 |
| Devices | google_apis arm64 images: API 36 (WebView 133.0.6943.137) and API 26 (WebView 58.0.3029.125) |
| Arms | PR 50e097a6d2 · Parent = #12121 head a765ac1ecf. The tree diff from Parent to PR is exactly the accessibility commit d53ad5f12a (6 files, +270/−12). For the parent-arm test run, I copied the PR's ProfileAccessibilityDeviceTest.kt into the parent tree. |
| Main | origin/main 0adae3c254 has no packages/mobile-shell changes since merge-base 8b39d93a42. git merge-tree against main is clean. |
| Isolation | A concurrent session was using the same SDK. I ran a private adb server (port 5039) and put the emulators on ports 5600/5602, outside the default adb scan range. Every command was addressed with -s. No Gradle connected* task was used; APKs were installed with adb install and run with am instrument. |
1. Build, JVM tests, lint: matches the PR's claims
:app:testDebugUnitTest :app:assembleDebug :app:assembleRelease :app:lintDebug :app:assembleDebugAndroidTest all succeed. 21/21 JVM tests pass (ProfileVault 15, ProfileCipher 3, OriginPolicy 3). Lint: 0 errors, 6 GradleDependency warnings.
2. Committed device suites: match the PR's claims and CI
| Device | Result |
|---|---|
API 36, requireProfileIsolation=true |
11 passed, 1 assumption skip (oldProviderFailsClosedBeforeCreatingNamedProfileOrWebView), 0 failed |
API 26, requireProfileIsolation=false |
8 passed, 4 assumption skips (the MULTI_PROFILE-dependent cases), 0 failed |
The upstream Keystore and profiles (API 26/36) jobs on this head also ran 12 tests with the same skip pattern.
3. The new tests detect missing changes
Parent arm with the PR's test file (API 36): 5/5 new tests fail. Each failure message matches the change it covers:
| Test | Failure on parent |
|---|---|
profileActionsIdentifyTheirOwnConnection |
Expected native accessibility node not found |
editorFieldsExposeTheirVisibleLabelsAndKeepTheTokenMasked |
Input must identify its visible label |
validationIsPoliteAndTheUserCanCorrectTheAddress |
expected:<1> but was:<0> (liveRegion) |
storageFailureHasPoliteDescription |
expected:<1> but was:<0> (liveRegion) |
storageRecoveryIsScrollableAndResetCancellationPreservesData |
Expected native accessibility node not found (no ScrollView) |
Mutation matrix (API 36): 6/6 mutants killed. I reverted one production change at a time on the PR head. Each mutant fails exactly the one test meant to cover it:
Mutant (one-line revert in MainActivity.kt) |
Failing test(s) |
|---|---|
Drop Connect contentDescription |
profileActionsIdentifyTheirOwnConnection only |
Drop Delete contentDescription |
profileActionsIdentifyTheirOwnConnection only |
Drop caption.labelFor = id |
editorFieldsExpose… only |
| Drop editor error live region | validationIsPolite… only |
| Drop storage hint live region | storageFailureHasPoliteDescription only |
setContentView(content) instead of the ScrollView wrapper |
storageRecoveryIsScrollable… only |
Live-region events. I wrote a verification probe for this (not part of the PR; source is in the assets branch). It is an instrumentation test that records every TYPE_WINDOW_CONTENT_CHANGED event whose source node has liveRegion != 0 after each Save:
| Step | API 36 PR | API 36 parent | API 26 PR | API 26 parent |
|---|---|---|---|---|
| 1. invalid origin, then Save | 1 (live=1, error text) |
0 | 1 | 0 |
| 2. Save again, same input | 0 | 0 | 1 | 0 |
3. ftp://x (same message) |
0 | 0 | 1 | 0 |
| 4. empty name (new message) | 1 | 0 | 1 | 0 |
This shows accessibility services now receive events they can announce, and the parent sent none. Steps 2–3 differ between API levels; see minor note (b). I did not test TalkBack speech: the google_apis images do not include TalkBack.
4. Storage recovery at large text: fix confirmed
At 640×360 dp with 200% font, both arms show the whole storage page, so the clipping only appears in a shorter window. The configuration below reproduces it on standard Pixel 5 hardware:
- Parent: Reset connections shows only 57 of its 190 px, and swiping does nothing. The 57 px strip (about 16 dp) can still be tapped and opens the Reset dialog. So "unreachable" is more precisely: clipped to a strip below the 48 dp minimum touch target, with an unreadable label.
- PR: the page scrolls, and Reset shows at its full 190 px.
- PR, functional check at the same configuration: Reset → Cancel keeps the vault bytes at
01 02 03. Reset → confirm deletes the vault file and returns to the empty Connections list.
5. Finding: editor validation error is off-screen in landscape (predates the PR; in scope)
The error TextView is the last child of the dialog's ScrollView, below the token field, the Keep checkbox and the credential hint. Save does not scroll it into view. In landscape the dialog is too short to show it, so after Save nothing visible changes:
The probe recorded whether the error node was isVisibleToUser right after the first Save (API 36, Pixel 5):
| Configuration | PR head | PR + candidate fix |
|---|---|---|
| default portrait | visible | visible |
| default landscape | not visible | visible |
| 200% font, portrait | visible | visible |
| 200% font, landscape | not visible | visible |
| 200% font + display size Largest, portrait | not visible | visible |
| 200% font + display size Largest, landscape | not visible | visible |
The parent arm gives the same result in default landscape (not visible, and 0 live events), so this PR did not introduce it. The design doc for this slice does require checking "editor errors … with large text/landscape", but the PR's evidence only covers storage recovery. At large text, the error is only visible after scrolling to the end of the form:
The existing test doesn't catch this: validationIsPolite… uses find without isVisibleToUser. It passed (5/5) in the 200% + Largest portrait configuration, in the same run where the probe showed the error off-screen.
Candidate fix (5 added lines, MainActivity.kt editProfile):
fun showError(message: CharSequence?) {
error.text = message
error.post { error.requestRectangleOnScreen(Rect(0, 0, error.width, error.height), false) }
}
// …
} catch (invalid: IllegalArgumentException) { showError(invalid.message) }
catch (_: Exception) { showError(getString(R.string.save_failed)) }Test change (2 lines in validationIsPolite…): lock the activity to landscape, and require the error node to be visible:
it.onActivity { activity -> activity.requestedOrientation = android.content.pm.ActivityInfo.SCREEN_ORIENTATION_LANDSCAPE }
// …
val error = find { it.isVisibleToUser && it.text?.toString() == context.getString(R.string.changed_origin_credential) }I ran it on a portrait device like the CI emulator:
- PR head: only this test fails (
Expected native accessibility node not found). - PR + candidate fix: 5/5 pass.
- Full committed suite with the candidate fix and this test change applied: API 36: 13 OK (1 skip); API 26: 13 OK (4 skips). The 13 are the 12 committed tests plus my probe.
- Lint is still 0 errors.
Full diffs: candidate-fix-scroll-only.diff, test-suggestion.diff.
This is not a regression, so it doesn't have to block this slice. It's small, and it is the exact scenario the slice's design names, so I'd suggest including it before merge rather than in a follow-up.
Minor notes (non-blocking)
- (a) Duplicate names share one accessible name.
ConnectionProfile.createdoes not require unique names. With two profiles namedWork(https://a.example/andhttps://b.example/), the six buttons exposeConnect to Work/Edit Work/Delete Worktwice (checked on API 26). The origin label is visible, but it isn't in the descriptions. Either include the origin in the three strings, or reject duplicate names. - (b) On API 36, a repeated identical error does not fire a live-region event (steps 2–3 above; API 26 does fire). A screen-reader user who taps Save twice with the same mistake hears nothing the second time. Clearing and re-setting the text restores the event, but it produced two events per Save in my probe. I did not check how TalkBack de-duplicates those, so I'm not suggesting it as the fix.
- (c) Action pin.
reactivecircus/android-emulator-runner@324029e2…is the SHA of the annotated tag object for v2.34.0, not a commit. It peels to commit1dcd0090116d15e7c562f8db72807de5e036a4ed. Actions resolves it (CI is green), and a tag object is content-addressed, so it's still immutable. Most pin checkers and update bots expect the commit SHA, though. - Diff scan (the precheck flagged
prompt_injection:print_secrets): nothing in the diff reads secrets. The workflow runs onpull_requestwithcontents: read. The only interpolations are static matrix values, and the onlysudois the standard KVM udev rule.
Not verified
- TalkBack speech, physical devices, WebView content, and Activity/process restoration. The PR also does not claim these.
- Windows and Linux hosts. Linux is covered by upstream CI.
Rig (probe test, mutator, visibility matrix) is at wenshao/qwen-code@assets-pr12129.
中文说明
50e097a6d2 本地模拟器验证(macOS arm64,API 36 + API 26)
结论:无障碍切片的行为与描述一致,没有发现回归。四处生产改动都不可缺少,而且各自有专门的测试覆盖。还剩一个在本切片范围内的问题,它在本 PR 之前就存在,但值得在这里一起修:横屏时编辑器的校验错误出现在对话框可视区域以下,默认字体下也是如此,视力正常的用户点 Save 后看不到任何反馈。我为它实测了一个 5 行修复和一个 2 行测试改动(见下文)。合并顺序仍须排在 #12121 之后,#12121 目前仍未合并。
PR 的测试平台表里 macOS 为“未测试”,本次验证补上了这一平台。
环境
| 主机 | macOS 26.6 arm64、Zulu JDK 17.0.20、Gradle 8.2.1 wrapper(jar 的 SHA-256 与 gradle-wrapper.jar.sha256 一致)、AGP 8.2.0、Android Emulator 37.1.11 |
| 设备 | google_apis arm64 镜像:API 36(WebView 133.0.6943.137)与 API 26(WebView 58.0.3029.125) |
| 对照臂 | PR 50e097a6d2;父版本 = #12121 头 a765ac1ecf。父版本到 PR 的树差异恰好就是无障碍提交 d53ad5f12a(6 个文件,+270/−12)。跑父版本臂的测试时,我把 PR 的 ProfileAccessibilityDeviceTest.kt 拷进了父版本源码树。 |
| main | origin/main 0adae3c254 自 merge-base 8b39d93a42 以来没有改动 packages/mobile-shell;对 main 做 git merge-tree 无冲突。 |
| 隔离 | 另一个并发会话在共用同一套 SDK。我单独起了一个 adb server(端口 5039),模拟器放在 5600/5602(不在 adb 默认扫描范围内),所有命令都带 -s 指定设备。没有用 Gradle 的 connected* 任务,而是 adb install 安装 APK、am instrument 运行测试。 |
1. 构建 / JVM 测试 / lint:与 PR 声明一致
:app:testDebugUnitTest :app:assembleDebug :app:assembleRelease :app:lintDebug :app:assembleDebugAndroidTest 全部成功。JVM 测试 21/21 通过(ProfileVault 15、ProfileCipher 3、OriginPolicy 3)。lint 0 错误,6 条 GradleDependency 警告。
2. 已提交的设备测试套件:与 PR 声明及 CI 一致
- API 36(
requireProfileIsolation=true):11 通过,1 个 assumption 跳过(oldProviderFailsClosedBeforeCreatingNamedProfileOrWebView),0 失败。 - API 26(
requireProfileIsolation=false):8 通过,4 个 assumption 跳过(依赖 MULTI_PROFILE 的用例),0 失败。 - 上游
Keystore and profiles (API 26/36)作业在该提交上同样跑了 12 项,跳过模式一致。
3. 新测试能发现改动缺失
- 父版本代码 + PR 的测试文件(API 36):5 项新测试全部失败,每条失败信息都对应它所覆盖的改动(见英文表格)。
- 变异矩阵(API 36):6/6 个变异体全部被杀。 我在 PR 头上每次只回退一处生产改动,每个变异体都只让它对应的那一条测试失败。
- live region 事件。 为此我写了一个验证用探针(不属于本 PR,源码在 assets 分支):一个 instrumentation 测试,记录每次 Save 之后、来源节点
liveRegion != 0的TYPE_WINDOW_CONTENT_CHANGED事件。PR 臂在错误出现时会派发live=1的事件,父版本一次都没有。这说明无障碍服务现在能收到可播报的事件;API 26 与 API 36 在重复错误上的差异见小建议 (b)。我没有测试 TalkBack 实际语音:google_apis镜像里没有 TalkBack。
4. 大字体下的存储恢复页:修复已确认
在 640×360 dp、200% 字体下,两臂都能完整显示存储页,可见裁剪只在更矮的窗口里出现。下面这个配置在标准 Pixel 5 硬件上就能复现(见图 01:Pixel 5、显示大小“最大”(560 dpi)、200% 字体、横屏):
- 父版本:“Reset connections”190 px 高,只露出 57 px,滑动也没有反应。这条 57 px(约 16 dp)的残边仍然能点中并打开 Reset 对话框,所以“无法访问”更准确的说法是:被裁成低于 48 dp 最小触控尺寸的窄条,标签也看不清。
- **PR:**页面可以滚动,Reset 完整显示 190 px。
- **PR 在同一配置下的功能检查:**Reset → 取消后 vault 字节仍为
01 02 03;Reset → 确认后删除 vault 文件,回到空的 Connections 列表。
5. 发现:横屏下编辑器校验错误在屏幕外(PR 之前就有,属于本切片范围)
错误 TextView 是对话框 ScrollView 的最后一个子项,排在 token 输入框、Keep 复选框和凭据提示之下,而点 Save 后不会把它滚进可视区。横屏时对话框高度不够显示它,所以点 Save 后画面没有任何可见变化(图 02)。
在 6 种配置下测了第一次 Save 后错误节点是否 isVisibleToUser(API 36、Pixel 5),PR 头在以下 4 种配置里都看不见错误:默认横屏、200% 横屏、200%+最大显示竖屏、200%+最大显示横屏。候选修复在 6 种配置里全部可见。
父版本在默认横屏下结果相同(看不见,live 事件为 0),所以这不是本 PR 引入的问题。不过本切片的设计文档明确要求检查“大字体/横屏下的编辑器错误”,而 PR 的证据只覆盖了存储恢复页。大字体下,错误只有滚到表单最底才看得到(图 03)。
现有测试发现不了这个问题:validationIsPolite… 用的是 find,没有检查 isVisibleToUser。在 200%+最大显示竖屏配置下它照样通过(5/5),而同一次运行里探针显示错误就在屏幕外。
**候选修复(在 MainActivity.kt 的 editProfile 里新增 5 行):**设置错误文本后,用 requestRectangleOnScreen 把错误滚进可视区(代码见英文部分)。
**测试改动(validationIsPolite… 里 2 行):**把 activity 锁定为横屏,并要求错误节点可见。我在与 CI 模拟器一样的竖屏设备上跑了:
- PR 头:只有这一条失败(
Expected native accessibility node not found)。 - PR + 候选修复:5/5 通过。
- 加上候选修复和这个测试改动后跑完整套件:API 36 13 项通过(1 跳过),API 26 13 项通过(4 跳过)。13 项是 12 项已提交测试加上我的探针。
- lint 仍为 0 错误。
这不是回归,不必因此阻塞本切片。但改动很小,而且正是本切片设计点名的场景,所以我建议在合并前一起改,而不是留给后续 PR。
小建议(不阻塞)
- (a) 重名配置共用同一个无障碍名称。
ConnectionProfile.create不要求名称唯一。两个都叫Work的配置(https://a.example/和https://b.example/)的 6 个按钮,会各出现两次Connect to Work/Edit Work/Delete Work(在 API 26 上检查)。origin 标签虽然可见,但不在描述里。可以把 origin 加进这三条文案,或者拒绝重名。 - (b) API 36 上重复出现的相同错误不再派发 live region 事件(API 26 会派发)。屏幕阅读器用户犯同一个错误、连点两次 Save,第二次什么也听不到。先清空再设置文本能恢复事件,但在我的探针里每次 Save 会产生两个事件。我没有核实 TalkBack 如何对它们去重,所以不把这个作为修复建议。
- (c) Action 固定引用。
reactivecircus/android-emulator-runner@324029e2…是 v2.34.0 附注标签对象的 SHA,不是提交 SHA,剥皮后对应提交1dcd0090116d15e7c562f8db72807de5e036a4ed。Actions 能解析它(CI 是绿的),标签对象也是内容寻址的,所以依然不可变。但多数固定引用检查工具和依赖更新机器人期望的是提交 SHA。 - diff 扫描(precheck 曾以
prompt_injection:print_secrets拦截):diff 中没有任何读取 secrets 的内容。workflow 以pull_request触发,权限为contents: read,插值只有静态的 matrix 值,唯一的sudo是标准的 KVM udev 规则。
未验证
- TalkBack 语音、实体设备、WebView 内容、Activity/进程恢复。PR 本身也没有声称验证过这些。
- Windows 与 Linux 主机。Linux 由上游 CI 覆盖。
装置(探针测试、变异脚本、可见性矩阵)在 wenshao/qwen-code@assets-pr12129。
|
@qwen-code /triage |
|
@qwen-code /resolve |
|
Qwen Code did not run conflict resolution for this request. PR #12129 comes from the fork |
|
Merge conflicts with This PR conflicts with Since "Allow edits from maintainers" is not enabled on this PR, I could not push the resolution directly. Instead I merged current To update this PR, run on your local branch: git checkout feat/android-phase2-accessibility
git pull --ff-only https://github.com/QwenLM/qwen-code.git merge-fix/pr-12129
git push origin feat/android-phase2-accessibilityResolution approach: kept Alternatively, enable "Allow edits from maintainers" and I will push the merge directly. 与 此 PR 与 由于此 PR 未开启 "Allow edits from maintainers",我无法直接推送解决结果,改为将当前 请在本地分支执行: git checkout feat/android-phase2-accessibility
git pull --ff-only https://github.com/QwenLM/qwen-code.git merge-fix/pr-12129
git push origin feat/android-phase2-accessibility解决方式:保留 或者开启 "Allow edits from maintainers",我可以直接推送合并结果。 |
Head branch was pushed to by a user without write access
|
The current |
|
Please rebase — this PR carries ~130 commits that are already merged into The Files tab shows 1,230 files, +147,021/−51,284, but the actual feature delta is only 6 files, +278/−14. Verified locally against the current head ( Same root cause as #12130: the branch merged Please rebase onto current |
|
Hi @jabrailkhalil 👋 Thank you so much for your contributions! Really appreciate the effort on the mobile shell accessibility and functionality improvements 🙌 I noticed that all four of your open PRs are currently blocked by merge conflicts — the fork branch has fallen ~506 commits behind main, resulting in 400+ file conflicts across the entire repo. This makes a normal rebase/merge impractical:
The good news is that your actual changes are nicely scoped to @wenshao Since these are well-reviewed mobile-only changes, would it be possible for a maintainer to cherry-pick or manually port the relevant commits onto fresh branches from current main? That would unblock all four PRs without asking the contributor to resolve hundreds of unrelated conflicts. Happy to help identify the exact commits/files if needed! 🚀 |
Head branch was pushed to by a user without write access
983f471 to
27276d9
Compare
…omment The comment recorded six jobs breaching the 100ms smaller-corpus bound with three breaching on all three vitest attempts. Re-enumerating the two sources that reported them gives eight distinct jobs, five breaching on all three attempts: the review sweep's four plus job 109188400510, and the three from @jabrailkhalil's report (#12127, #12129, #12130). All eight are verified directly against the jobs API as `Test (ubuntu-latest, Node 22.x)`, `conclusion: failure`, on `GitHub Actions NNN...` runners labelled `ubuntu-latest`. The 110-177ms range is unchanged. The per-attempt counts are as reported by those two sources -- the Actions logs for those runs have expired, so that half is attributed rather than re-measured. Comment only; no behaviour change. Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
…tency gate (QwenLM#13094) * test(core): treat GitHub-hosted lanes as shared in the recall scan latency gate * test(core): narrow the hosted-lane recall scan relaxation to the row it was measured on Widening the single SHARED_CI boolean made a GitHub-hosted lane inherit the ECS pool's contention-sized bounds wholesale: the warm gate flipped from median-of-5 < 50ms to best-of-5 < 1000ms, and every cold row went to 10x. Only the 1000-topic cold row was measured over its bound there — run 36640349705 reports `recall-scan-latency.test.ts (2 tests | 1 failed)` with all three retries failing at ceiling 125, so the warm test and the 200/500 cold rows were green under the strict bounds they keep here. Split the switch into POOL_CI and HOSTED_CI so the hosted arm relaxes only that row, to the measured worst case (226ms) plus the same ~20% slack the strict 1000-topic bound already uses = 275ms. That still reddens the second CST parse's ~+48% (~335ms), which the pool's contention-sized 1250ms would have swallowed. The pool and strict arms keep their ceilings and statistic unchanged, and the pool key stays the `ecs-qwen-` prefix ci.yml uses for QWEN_SKIP_LATENCY_BUDGETS. HOSTED_CI reads RUNNER_ENVIRONMENT, Actions' documented lane discriminator and the signal this repo's other gates already use, with the RUNNER_NAME display-name prefix kept beside it rather than load-bearing alone. Both gates now carry the detected lane, the classifier flags and the ceiling in force in the assertion message, because `silent: true` in packages/core/vitest.config.ts discards the console tables this file's header calls the artifact worth reading; `vitest/valid-expect` defaults to one argument, hence the local disable in the helper. Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com> Patrol-Run: qwen-pr-closeout/jmunyl8iwhf * test(core): relax the hosted-lane recall scan bound on every cold row Round-2 review (R2-1) and an independent report from @jabrailkhalil both show GitHub-hosted `Test (ubuntu-latest)` jobs breaching the 100ms bound the 200/500-topic cold rows keep, not only the 125ms bound on the 1000-topic row. On main the off-pool ceilings are warm=50, cold 200/500=100 and cold 1000=125, so `expected X to be less than 100` can only come from a smaller-corpus cold row: six hosted jobs measured 110-177ms there, three of them on all three vitest attempts. Restricting the relaxation to `topicCount >= 1000` therefore left those jobs red, and the comment recording "the 200/500 cold rows stay green under theirs" was an over-generalization from a single run. Treat the hosted lane as one lane class: every cold row gets the measured 2.2x multiplier (220ms against a 177ms worst breach, 275ms against a 227ms worst sample). The warm-cache gate is untouched, so the round-1 narrowing that mattered is preserved. Also record the fleet span rather than one runner as the calibration basis, correct the `~1.8x` figure to the baseline-relative basis the neighbouring `~5x` uses, and pin the lane arms by value so the hosted arm executes somewhere -- this PR's own CI job lands on the pool. Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com> * test(core): correct the hosted-lane breach count in the calibration comment The comment recorded six jobs breaching the 100ms smaller-corpus bound with three breaching on all three vitest attempts. Re-enumerating the two sources that reported them gives eight distinct jobs, five breaching on all three attempts: the review sweep's four plus job 109188400510, and the three from @jabrailkhalil's report (QwenLM#12127, QwenLM#12129, QwenLM#12130). All eight are verified directly against the jobs API as `Test (ubuntu-latest, Node 22.x)`, `conclusion: failure`, on `GitHub Actions NNN...` runners labelled `ubuntu-latest`. The 110-177ms range is unchanged. The per-attempt counts are as reported by those two sources -- the Actions logs for those runs have expired, so that half is attributed rather than re-measured. Comment only; no behaviour change. Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com> --------- Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
|
@yiliang114 Thank you for checking the branch drift. That conflict situation has since been resolved: GitHub currently reports #12127, #12129, #12130, and #12247 as @doudouOUC Thank you for the independent approval of #12129. I am keeping the converged feature scope unchanged; the Suggestion-level follow-ups remain in #13111. This does not mean every CI gate is green: #12129 still has a failed Ubuntu unit-test check, and #12130 has failed device/unit checks that remain to be rerun or resolved. #12130 also still has a formal |
|
October 3 integration follow-up: I merged current main At this head, debug, unsigned release, and instrumentation APK builds, all 21 JVM tests, and Android lint pass locally on Windows/JDK 17/SDK 34. No emulator or physical device is connected locally, so I am not claiming a new device pass. Fresh upstream CI is now required; the previous head's Feishu/unit failure is not being waved away as a pass. The converged Suggestion-level follow-ups remain tracked in #13111 rather than widening this PR again. @yiliang114 The conflicts introduced by the microphone merge have now been addressed on this branch. Please review the current merge state and checks when they settle. Current head: Current-head CI update: Android Mobile Shell has now passed, including native build/unit/lint and the API 26/API 36 device jobs. This is upstream device evidence on 中文说明当前 head 的 CI 更新:Android Mobile Shell 已通过,包括原生构建/单元测试/lint 及 API 26/API 36 设备任务。这是 |
|
@qwen-code /triage |
Round-3 local verification of PR #12129 — head
|
| # | Finding | Severity | Status at a2929967 |
|---|---|---|---|
| R1 (main) | Editor validation error off-screen in landscape | Critical-ish (pre-existing, in scope) | Fixed — re-measured: error isVisibleToUser=true in all 3 probed configurations on head vs false in all 3 on base (§3) |
| R1-1 | Fixture could overwrite real/legacy/unreadable/interrupted data | Critical (round-1 blocker) | Fixed, still holds — 4 plant kinds → 5 skips each, files byte-identical (§5) |
| R2-5 | token.contentDescription = previous?.token mutant survived |
Suggestion (test gap) | Fixed — mutant now killed by the window-wide token scan; failure message carries no token (§4) |
| R2-6 | Persisted-save assertion printed identical toString() on both sides |
Suggestion (diagnosis) | Fixed — per-field assertions; mutant now fails with The browser identity must be preserved expected:<…> but was:<…> (§4) |
| R2-8 | Destructive-Cancel mutant passed 5/10 runs (race) | Suggestion (test gap) | Fixed — with the committed drain, the mutant fails deterministically (Recovery page still shown after cancelling Reset not found) (§4) |
| R2-2 / R2-4 | After a full suite run, retiredBrowsers=1 residue makes the accessibility class skip on reused installs |
Minor, deferred by author | Stands (re-measured) — full run leaves profiles=0 retiredBrowsers=1 fileBytes=84; rerun of the class: 5 skips, 0 fails. CI unaffected (fresh AVD per job); the PR body explicitly defers fixture hardening to a follow-up |
| R2-7 | Nested-ScrollView mutant survives | — | Unchanged; round 2 refuted the example (not a regression; suggested fix breaks CI geometry). Not re-measured — no related code changed |
| R2-3 | PR wording: JVM vs device test split | Nit | Wording in the current body is accurate (21 JVM tests; device tests in the emulator lanes) |
| Minor (a) | Duplicate profile names share accessible names | Note | Stands — out of this slice's declared scope |
| Minor (b) | API 36 dedupes a repeated identical live-region text | Note | Stands — re-measured: first error fires 1 polite event, identical re-sets fire 0 (§3) |
| Minor (c) | android-emulator-runner pinned to annotated-tag SHA | Nit | Stands — workflow untouched by this PR |
1. Central claim and delta since round 2
Central claim: the slice improves native profile accessibility (identifiable actions, labeled inputs, polite validation, scrollable recovery) without changing credential behavior, and its tests guard those properties.
git diff 983f471bb1 a29299677c -- packages/mobile-shell/app/src/main/java/com/qwen/mobileshell/MainActivity.kt shows zero changes to any accessibility hunk (contentDescription / labelFor / liveRegion / showError / ScrollView) — the two merges from main added the microphone-consent and file-picker code that came with main, and nothing else in the PR's scope. The PR-owned test files changed exactly by the round-2 follow-ups: localized expected descriptions, window-wide token scans, per-field persisted assertions, and the post-Cancel drain. The Gradle wrapper jar SHA-256 matches gradle-wrapper.jar.sha256 in both trees, and the PR diff touches no build script, wrapper, workflow, or properties file.
2. Gates at head
| Gate | Result |
|---|---|
:app:assembleDebug :app:assembleRelease :app:testDebugUnitTest :app:lintDebug :app:assembleDebugAndroidTest |
BUILD SUCCESSFUL (1 m 13 s) |
| JVM unit tests | 21/21 (OriginPolicy 3, ProfileCipher 3, ProfileVault 15), 0 failures |
lintDebug |
0 errors, 6 GradleDependency warnings (pre-existing dependency-version notices) |
| Upstream CI at this head | all non-skipped checks pass, including Keystore and profiles (API 26) and (API 36) device lanes |
3. A/B load-bearing proof (API 36 emulator, CI-like 320×640 @ 160 dpi, requireProfileIsolation=true)
Base control = worktree at 1a5aae80d9 + the PR's three test files + the 3 string resources (compile-enabling only; no production wiring). Head = a2929967.
| Cell | Base | Head |
|---|---|---|
| Full device suite | — | 56 pass, 1 assumption skip (oldProviderFailsClosed…, expected on a current provider), 0 fail (57 committed tests) |
ProfileAccessibilityDeviceTest |
5/5 FAIL as expected, each message naming the missing change | 5/5 pass |
EmptyProfileFixtureTest |
10/10 pass (fixture safety is pre-existing main behavior) | 10/10 (in full-suite run) |
Base failure messages: …not found: Connect to Accessibility Alpha, Input must identify its visible label, …not found: Visible changed-origin validation message, Storage failure must be a polite live region expected:<1> but was:<0>, …not found: Scrollable storage recovery page.
Editor-error visibility probe (round-2 ProbeR2Test, invalid origin + one Save; CI-like screen):
| Configuration | base 1a5aae80 |
head a2929967 |
|---|---|---|
| default font, landscape | isVisibleToUser=false, 0 live events |
visible, 1 polite live event |
| 200% font, landscape | false, 0 | visible, 1 |
| 200% font, portrait | false (bounds y=660 below window 616), 0 | visible, 1 |
Recovery page probe (200% font, landscape, 4 real swipes): base has no ScrollView (scrollViews=[]) and Reset stays clipped; head's ScrollView is scrollable and Reset remains fully visible (bounds 220–276 of a 296-high window). Storage hint live region: head 1, base 0.
Live-region events match round 2: head fires 1 polite event on the first error; steps 2–4 re-set the identical changed-origin text (the changed-origin guard fires before name validation, so the empty-name case is never reached with an invalid address present) and API 36's framework dedup suppresses all three; base fires 0 events at every step. Note: round 1's probe recorded 1 event for the empty-name step; the save-path code is byte-identical across the round-1, round-2 and current heads (checked via git show 50e097a6d2…:MainActivity.kt), so that September reading is a probe-level difference, not a code change.
4. Mutation matrix (API 36, head tree, one mutation at a time)
Pristine control: 5/5 pass. Then 11 single-point mutants, each killed by exactly its covering test:
| Mutant | Killed by |
|---|---|
| drop Connect contentDescription | profileActionsIdentifyTheirOwnConnection |
| drop Edit contentDescription | same |
| drop Delete contentDescription | same |
drop caption.labelFor |
editorFieldsExposeTheirVisibleLabelsAndKeepTheTokenMasked |
| drop editor-error live region | validationIsPoliteAndTheUserCanCorrectTheAddress |
| drop storage-hint live region | storageFailureHasPoliteDescription |
drop ScrollView wrapper (setContentView(content)) |
storageRecoveryIsScrollableAndResetCancellationPreservesData |
drop requestRectangleOnScreen (showError) |
validationIsPolite… (visibility assertion, landscape-locked) |
R2-5: token.contentDescription = previous?.token |
editorFieldsExpose… — window-wide scan; message contains no token (checked: 0 occurrences in the log) |
| R2-6: browserId rotates on every save | validationIsPolite… — The browser identity must be preserved expected:<cd58936b…> but was:<407d2744…> |
R2-8: Cancel runs reset() + loadProfiles() |
storageRecovery… — deterministic kill (round 2: flaky 5/10) |
5. Fixture safety on real storage (R1-1 re-check, head, API 36)
| Planted on an empty install | Result | Files |
|---|---|---|
| profile saved through the real Keystore vault | 5 skips: Refusing to overwrite saved connections or retired browser data |
vault md5 unchanged |
legacy qwen_profiles prefs |
5 skips: Refusing to overwrite legacy connection data |
prefs md5 unchanged |
| unreadable vault bytes | 5 skips: Refusing to overwrite an unreadable connection vault |
md5 unchanged |
leftover connection-profiles.v1.bak |
5 skips: Refusing to overwrite an interrupted connection-vault write |
md5 unchanged |
6. Standing known issues (none blocking; all disclosed in earlier rounds)
- R2-2/R2-4 residue (re-measured at head, §status table): affects reused local installs only; CI runs a fresh AVD per job and the class order puts
EmptyProfileFixtureTestfirst. The PR body explicitly defers fixture hardening and reused-install retired-profile handling to a follow-up. Acceptable as scoped. - Minor notes (a) duplicate-name accessible labels, (b) API 36 live-region dedup, (c) annotated-tag action pin: unchanged, non-blocking.
reviewDecisionstill shows CHANGES_REQUESTED from the bot's round-1 review; its only Critical (R1-1) is verified fixed in rounds 2 and 3. Clearing it needs a maintainer approval/dismissal.
Not covered
- API 26 local re-run: the image download was impractical from this network (sdkmanager + proxied curl both stalled at ~2 MB). The API-26-specific code paths at this head are byte-identical to the round-2-verified head (production accessibility code unchanged; merges touched microphone/file-picker paths that are capability-gated off below WebView 111), and the upstream
Keystore and profiles (API 26)lane is green at this head. Round-2's API 26 results carry forward on that basis. - TalkBack speech (google_apis images ship no TalkBack), physical devices, x86/x86_64 CI images (I used arm64), Windows/Linux hosts, Activity/process restoration, WebView content of a live connection. The PR does not claim these.
- R2-7's scrollability pinning under enlarged text (round 2 already adjudicated the suggested fix as CI-breaking).
Methodology
Host: macOS arm64, JDK 17 (qwen-verify-toolchain/jdk17), Gradle 8.2.1 wrapper (SHA-256 verified against the committed checksum in both trees), Android Emulator with google_apis;android-36;arm64-v8a (WebView 133.0.6943.137, both MULTI_PROFILE and DELETE_BROWSING_DATA present → requireProfileIsolation=true). AVD created CI-like (no --device: 320×640 @ 160 dpi). Private adb server (port 6037), emulator on port 5630 outside the default scan range; APKs installed with adb install -r -t and run with am instrument (never Gradle connected*); animations disabled. Base arm = base worktree + PR test files + the 3 compile-enabling string resources. Mutation rows: pristine restore → one textual mutation (each anchor verified to occur exactly once) → incremental rebuild → pm clear → install → run the accessibility class; per-test status codes read from the instrumentation stream (-2 = failure, -4 = assumption skip). Fixture arms: probe-planted data on a fresh install, md5 snapshots before/after the class run. Probes are the round-2 rig (ProbeR2Test), rebuilt at both trees. Rig, per-row logs and raw outputs are in the artifact directory (tmp/pr12129-verify-20261005-180118/).
Evidence captures
Local maintainer verification by @wenshao — evidence for the merge decision; not a GitHub review approval.
|
Released in v0.25.1. |









What this PR does
Improves native connection actions, editor labels, polite validation and scrollable storage recovery. Validation errors are brought into view after Save. Test fixtures preserve existing credentials and refuse unsafe setup.
Current head:
a29299677c567bd6423da118e34306e888109c53. Based on main1a5aae80d9; #11722, #12121 and #12126 are merged. This is an independently reviewable Phase 2 slice.Why it's needed
Large text and landscape could hide validation or recovery controls, and assistive technology could not reliably distinguish profile actions. Tests must also detect token exposure and confirm persisted edits without replacing real user data.
Reviewer Test Plan
How to verify
At large font scale in landscape, save an invalid address: the error must be visible and politely announced, then a corrected address must save. Check localized profile-specific actions, visible input labels and masked token fields. Scan text, descriptions and hints throughout both list and editor accessibility trees for the synthetic token. Verify the saved name, origin, credential and browser identity. Cancel Reset on a synthetic corrupt-vault screen and verify the original bytes. Existing real, legacy, retired or interrupted profile data must prevent destructive test setup.
Evidence (Before & After)
The new follow-up replaces hardcoded English descriptions with localized resources, scans the whole profile-list window for token leaks, and checks persisted fields separately so failures identify the broken field without printing a credential. The earlier upstream review contains prior-revision API 26/36 evidence. Current device assertions compile locally; local device execution is unavailable.
Tested on / Environment
Windows/JDK 17/Gradle 8.2.1/compile SDK 34: debug, unsigned release and instrumentation APKs compile; 21 JVM tests pass. Android lint: 0 errors, 7 existing warnings. Commands:
assembleDebug assembleRelease assembleDebugAndroidTest testDebugUnitTest lintDebug. No connected emulator or physical device was available. New-head upstream CI is required; previous-head green jobs do not establish this revision.Shared Ubuntu latency investigation, with exact revisions and limits. The unchanged assertion also fails on inspected current main in an isolated Windows comparison; that does not establish the Ubuntu root cause or make CI green.
Risk & Scope
Full Web Shell/TalkBack and physical-device acceptance remain separate. Fixture hardening and reused-install retired-profile handling remain follow-ups. The conservative data guard stays intact; a skipped populated installation is not permission to wipe its vault or browser data.
Design: English / 简体中文.
Linked Issues
Refs #11704, #11722, #12121, #12126.
中文说明
本 PR 的改动
改善原生连接操作、编辑器标签、礼貌校验提示及可滚动存储恢复页;保存后将错误滚动到可见位置。测试夹具保护已有凭据并拒绝不安全初始化。
当前提交:
a29299677c567bd6423da118e34306e888109c53。基于 main1a5aae80d9;#11722、#12121、#12126 已合并。本 PR 为可独立审阅的 Phase2 切片。为什么需要
大字体和横屏可能隐藏校验或恢复控件,辅助技术难以区分连接操作。测试需发现 token 泄露并确认保存结果,同时保护真实数据。
审阅者测试计划
验证方法
大字体横屏下保存无效地址,确认错误可见且礼貌播报,修正后可保存。检查本地化连接操作、可见字段标签及掩码 token;扫描列表与编辑器整个可访问性树的文本、描述和提示。核对保存后的名称、来源、凭据及浏览器身份。取消合成损坏 vault 的重置后字节保持不变。真实、旧版、退休或未完成写入的数据必须阻止破坏性初始化。
前后证据
本次用资源替代硬编码英文操作描述,扫描整个连接列表窗口的 token 泄露,并逐字段断言保存结果,失败时不输出凭据。先前上游评审的 API26/36 结果属于先前版本。当前设备断言已编译,本地没有可用设备执行。
测试平台及环境
Windows/JDK17/Gradle8.2.1/compile SDK34:debug、未签名 release 及 instrumentation APK 编译成功,21 项 JVM 通过;lint 为0错误、7项既有警告。运行
assembleDebug assembleRelease assembleDebugAndroidTest testDebugUnitTest lintDebug。本地无模拟器或真机;当前提交需新的上游 CI,旧版绿色不代表本版已通过。共同 Ubuntu 延迟调查及精确版本/限制:未修改的断言在所检查当前 main 的隔离 Windows 对照也失败,但尚未确认 Ubuntu 根因,也不表示 CI 已绿。
风险与范围
完整 Web Shell/TalkBack 和真机验收另行进行。夹具加固、复用安装的退休 profile 处理明确延后。保守数据保护保持不变;跳过含数据安装不表示允许删除 vault 或浏览器数据。
设计:English / 简体中文。
关联问题
关联 #11704、#11722、#12121、#12126。
Current-revision Android CI — 2026-09-29
The PR merge ref containing
2e4b2f2a1725192ef5790f87b03897c31409c9a2has now passed both upstream emulator jobs: API 26 (39 tests started, 4 explicit skips) and API 36 (39 started, 1 explicit skip). Both logs end inBUILD SUCCESSFUL. Skips belong to capability/provider-dependent profile tests; no accessibility test is listed as skipped. This supplies emulator execution evidence for the latest device-test changes, beyond local APK compilation. The separate Android build/unit/lint job and general repository checks were still queued at this check; no full-CI or physical-device pass is claimed.中文:包含当前提交
2e4b2f2的 PR 合并引用已通过上游 API26/36 模拟器作业;各启动39项,分别显式跳过4项和1项,日志均为BUILD SUCCESSFUL。跳过项属于能力/提供程序相关的 profile 测试,未列出 accessibility 测试跳过。此结果补充本地仅编译后的新版设备测试执行证据。独立 Android 构建/单测/lint 与全仓检查当时仍排队,不声称全 CI 或真机验收通过。October 3 integration status
Current main
1a5aae80d9is integrated after #12127 merged. Only the README conflict required resolution; accessibility code and fixtures are unchanged. On Windows/JDK 17/SDK 34, debug, unsigned release and instrumentation APK builds, 21 JVM tests and Android lint pass. No device is connected locally; current-head upstream device/CI checks and review remain required. Nonblocking Suggestions remain in #13111.中文:#12127 合入后,本次集成 main
1a5aae80d9。仅解决 README 冲突,无障碍代码和夹具保持不变。Windows/JDK 17/SDK 34 上 debug、未签名 release、instrumentation APK 构建、21 项 JVM 测试和 Android lint 通过。本地没有连接设备,仍需当前提交的上游设备/CI 检查及审查。非阻断 Suggestions 保留在 #13111。