Repository navigation
Antigravity UI/UX polish: review fixes and phone support - #373
Conversation
The shared Button sets pointer-events: none when disabled, so the title tooltip on the disabled "Remove runtime" button never appeared, and keyboard users could not focus it to find out. The reason is now visible text next to the button, linked with aria-describedby. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Install, Cancel and Remove each unmount (or are disabled) as soon as the runtime state changes, which dropped keyboard focus to the page body in the middle of the setup dialog. Focus now moves to whichever control replaced the one the user was on: Cancel while installing, Install or Retry after a cancel or removal, Remove runtime after an install, or the section itself when nothing is left to act on. Escape inside the inline remove confirmation now backs out of the confirmation instead of closing the whole dialog. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…client Activity that Google Antigravity runs itself (delete_file, move_file, web_fetch, agent_subagents, agent_tool and the fresh-session notice) had no verb pair, so rows read "Delete file x" both while running and when done, unlike "Writing" → "Wrote". The Mac, iOS and Android verb maps gain the same pairs, and each client's test expects the same lines. Agent commands carried no detail, so their rows read only "Running" or "Ran". They now read "Ran a command"; the raw command line is still never persisted. The persisted fresh-session label is unchanged so older phone builds, which show the label as-is, still read naturally. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Approval cards for Antigravity reads, fetches and other tools fell back
to the raw tool name in lowercase ("read file needs approval", "web
fetch needs approval") next to "Edit file needs approval". Deletes and
moves were presented as edits.
- The card label map covers read_file, delete_file, move_file,
web_fetch and agent_tool, and any unknown tool name now falls back
to sentence case.
- Permission requests carry whether a file change deletes or moves, so
the card and its summary say "Delete file" / "wants to delete files".
Phone approval cards use generic copy and the summary line, so they
need no change.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
useHarnessStatus swallowed a failed status read, which left the setup dialog on "Checking the Google Antigravity runtime…" forever. The hook now reports the error and a retry, and the section shows what went wrong with a "Try again" button. Any later status broadcast clears the error. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The runtime section is shared by every ACP harness, but its copy hard-coded "Google's own agent runtime", "dl.google.com" and "download from Google", and the setup dialog said "runs Google's agent". A second harness would have shown Google's name and host. - Harness definitions declare a publisher; Antigravity's is "Google". - The installer reports the host of the pinned archive for this computer, and the status projection carries both to the renderer. - The renderer accepts only a bare host name and a short publisher, and the copy says "its own agent runtime" when either is missing rather than guessing. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
While verifying, extracting or validating, the runtime section drew a full-width bar at half opacity (an inline style outside the token system), which read as a finished download. Like the update bars in About and the sidebar, the bar now appears only with a real percentage; the summary line names the current phase. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The setup dialog said "Install the runtime above before signing in" whenever sign-in was unavailable, including on computers the runtime does not support (no Install button exists), while an update was needed (the button says Update), mid-install, and while the status was still loading. The hint now says what is actually missing in each state, and nothing while loading. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The polite live region read the raw phase names ("validating",
"activating") while the visible summary said "Starting it once to
confirm it works" and "Finishing". Both now use the same phase copy.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Before installing, the summary line ("Not installed. 111 MB download
from Google.") and the install disclosure right below it both gave the
size. The summary now just says "Not installed."; the disclosure keeps
the size, source and free space next to the Install button.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Install (and Update or Retry) is the only step forward in the setup dialog until the runtime exists, but it used the neutral filled variant. It now uses the accent variant, as primary actions do elsewhere (for example "Update and restart" in About). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Antigravity fell back to a letter tile next to real logos in the model picker, provider settings and onboarding, and to a "G" monogram on the phones. - Mac and iOS bundle Google's one-colour Antigravity glyph as a themed mark (viewBox fitted to the glyph so it sits at the same optical size as the other marks). - Android, which draws monograms for every provider, uses Antigravity's brand blue. - The iOS test now also checks that every supported slug has a bundled logo. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The model picker's accessible description reads "Provider X, Deployment local, Inputs …", but Antigravity entries said "Agent running on this computer", breaking the pattern. The visible tag and spoken phrase now come from one shared helper: "· Agent" and "Deployment agent on this computer". Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…edule When the app default model was Google Antigravity, an unpinned scheduled task showed "If no app default is available, this task cannot run." But a default is available and works in chats; it just can't run unattended. The warning now names the default and says why: it runs only in chats open on this computer, so pin a provider. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…kers
The Assistant dock gate added earlier is on useAssistantChat, which no
UI renders, so Assistant chats still offered Google Antigravity in
their model picker. Picking it failed only after sending ("works only
in ordinary desktop chats, not here").
Assistant and Bot chats now leave agent-backed providers out of the
picker. If a chat already has one selected, the composer says why it
can't send ("Google Antigravity runs only in ordinary desktop chats,
not in Assistant chats. Choose another model.") instead of failing
the turn.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Mac leaves agent-backed providers such as Google Antigravity out of the phone's model catalog, so opening a chat Antigravity answered on the Mac silently resolved to the host's default model. A reply from the phone then went to a different model with no explanation. iOS and Android now show a note above the composer: "This chat used Google Antigravity, which runs only on your Mac. Replies from here use Gemini Flash." The note disappears once the chat is on a model the phone can use. Bot chats, whose model is fixed, are unaffected. The agent-provider list mirrors the Mac's ACP_HARNESS_PROVIDER_IDS. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
very-hermes-bot
left a comment
very-hermes-bot
left a comment
There was a problem hiding this comment.
Inline review — 1 finding(s) anchored to the diff. See the pinned summary comment for the overview.
Hermes Review BotConfidence: 5 Engine: SummaryAntigravity UI/UX polish resolves accessibility, focus management, and surface-scoping feedback from PR #371, while introducing cross-client mobile parity for desktop-started agent chats on iOS and Android. The setup dialog replaces the disabled button's hidden tooltip with accessible reason text, adds layout-effect focus recovery across dynamic action button unmounts, and traps Escape to dismiss the inline removal confirmation without closing the outer dialog. Agent approvals and timeline activity rows now support verb pairs across desktop and mobile, mask raw command execution strings to protect secrets, and enforce sentence casing for unknown tools. Model selection explicitly excludes agent harnesses from Assistant and Bot chats, informs mobile users when a desktop-only agent model cannot answer from a phone, and guards scheduled task configurations against unattended agent defaults. Maintainers should double-check that the capture-phase Confidence Score: 5/5Fully traced across backend handlers, renderer components, focus/keyboard lifecycle, IPC contracts, and native iOS/Android implementations with verified test coverage. 📁 Important Files Changed
FindingsNo findings. Sequence DiagramsequenceDiagram
autonumber
actor User
participant Dialog as HarnessRuntimeSection
participant IPC as Main Process (IPC)
participant Installer as AcpRuntimeInstaller
User->>Dialog: Press Enter on "Install Google Antigravity"
Dialog->>IPC: providers:harness:install
Dialog->>Dialog: Set acting=true; button disables
IPC->>Installer: Start download
Installer-->>IPC: Status change: installing (downloading)
IPC-->>Dialog: providers:harness:changed
Dialog->>Dialog: useLayoutEffect recovers focus -> "Cancel installation"
Installer-->>IPC: Download complete (installed)
IPC-->>Dialog: providers:harness:changed
Dialog->>Dialog: Cancel unmounts -> useLayoutEffect shifts focus to "Remove runtime"
User->>Dialog: Press Enter on "Remove runtime"
Dialog->>Dialog: confirmRemove=true -> focus shifts to "Keep"
User->>Dialog: Press Escape
Dialog->>Dialog: Window capture listener intercepts Escape, calls setConfirmRemove(false)
Dialog->>Dialog: Keep unmounts -> useLayoutEffect restores focus to "Remove runtime"
[]
|
![pullfrog[bot]](https://avatars.githubusercontent.com/in/1768019?s=60&v=4){kind=small}
There was a problem hiding this comment.
Important
Two verified P2 issues need addressing before merging.
Reviewed changes Reviewed all 50 changed files and 17 commits at 5a99b01d against stacked base feature/antigravity-aiden-integration-0631d1 (31eb003e): desktop runtime setup, activity/approvals, provider presentation, Assistant/Bot gates, schedule warnings, and native phone notices.
- [P2] Runtime focus recovery misses asynchronous transitions at
harness-runtime-section.tsx:128-150. With the production section and shared Dialog mounted in Chromium, asynchronous completion or failure while Cancel installation has focus leaves focus on the outer dialog rather than Remove runtime or Retry installation. Radix's MutationObserver redirects focus before the passive effect executes, and itslostpredicate then excludes the dialog container. Perform replacement-control focus synchronization before the observer, such as in a layout effect, and add mounted async completion/failure coverage rather than only testing the target mapping. - [P2] Extracted warning copy breaks a registered test at
scheduled-task-editor.tsx:615.npx tsx --test renderer/components/scheduled-task-editor.test.tsxreports 2 passed and 1 failed because line 32 still requiresNo provider pinned.in the component source, although the text now comes from the helper. Replace the obsolete source assertion with rendered warning coverage for absent and Antigravity defaults, not another production-source grep.
Verification: 107 focused TypeScript tests, 91 Android JVM tests, iOS release-policy/asset checks, and 10 CI-registry tests passed. The focus issue was reproduced in Chromium. Native iOS XCTest was not run locally on Linux. Hosted CI was still in progress at the last check.
Inline review publication encountered GitHub API errors, including an internal GraphQL failure with request ID 7808:1E1A50:13C63F1:3EE219B:6AC67AB6; findings are included in the review body to preserve the full feedback.
gpt-6.1-sol ｜ 𝕏
…review-6e4d31 # Conflicts: # android/app/src/main/java/sbtbiswas/AidenOnTheGo/features/shared/AidenProviderIcon.kt # android/app/src/test/java/sbtbiswas/AidenOnTheGo/features/shared/AidenProviderIconTest.kt
When an install failed or finished in the background while Cancel had focus, Radix's focus scope moved focus from the removed button to the dialog container in a MutationObserver callback. That ran before the passive effect, which then saw focus on the dialog (not the page body) and left it there instead of moving to Retry or Remove runtime. Focus recovery now runs in a layout effect, in the same task as the DOM commit, and treats focus on any container around the section as lost. A new Electron e2e drives the real renderer and dialog with stubbed harness IPC: async progress, async failure and async completion each land focus on the replacement control, and Escape backs out of the remove confirmation. It fails against the previous passive effect. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…copy The provider warning's text now comes from scheduledTaskProviderGuardrail, so the source-grep assertion for "No provider pinned." in scheduled-task-editor.test.tsx failed. The warning is a small exported component, and the test renders it for no default, an Antigravity default, a usable default and a pinned provider. The test also checks that the warning has no border, ring or outline. This replaces the guardrail source greps rather than adding more. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
very-hermes-bot
left a comment
very-hermes-bot
left a comment
There was a problem hiding this comment.
Inline review — 1 finding(s) anchored to the diff. See the pinned summary comment for the overview.
There was a problem hiding this comment.
✅ No new issues found.
Reviewed changes Reviewed the changes since 5a99b01d, including both prior review fixes and the adaptations after merging main.
- Moved runtime focus recovery into a layout effect. Verified asynchronous transitions focus Cancel, Retry, and Remove before the dialog's focus scope intervenes, while preserving focus outside the runtime section.
- Added mounted Electron interaction coverage. The new test exercises install progress, failure, completion, and Escape from the removal confirmation using controllable harness IPC.
- Replaced the obsolete schedule-warning assertion. The test now renders the production warning for missing and agent-backed defaults, and checks that usable defaults and pinned providers show no warning.
- Adapted Android Antigravity artwork to the shared logo renderer. Added the generated VectorDrawable, registered its provider mapping, and updated native attribution notices.
Verification passed: 33 focused renderer tests, focused Android JVM suites, all 43 generated Android logo checks, iOS release-policy/artwork checks, desktop and e2e TypeScript checks, and focused lint. A Chromium probe with the production runtime section and shared Radix dialog confirmed the focus fix and Escape behavior. Native iOS XCTest and full Electron e2e were not run on this Linux runner; hosted CI on this head is still incomplete.
gpt-6.1-sol ｜ 𝕏
… copy In an Assistant or Bot chat, a signed-out Google Antigravity selection showed "needs an API key. Add one in Settings → Providers." The missing-credential check ran before the agent-surface reason, and Antigravity signs in to an account rather than taking a key. The composer now asks acpHarnessReadinessMessage first. For an agent model it says the chat type can't host it (signing in wouldn't help), or otherwise asks the user to sign in. Other providers keep the existing API-key copy. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
There was a problem hiding this comment.
✅ No new issues found.
Reviewed changes Reviewed the composer-readiness fix in 8a69115c since the prior review at 69f181ce, with the full PR diff as context.
- Prioritized agent-surface restrictions. A signed-out Antigravity selection in an Assistant or Bot chat now explains that the chat type cannot host it before offering authentication guidance.
- Corrected sign-in guidance. Ordinary desktop chats now request Antigravity account sign-in rather than an API key, while other providers retain their existing authentication copy and send-readiness checks.
- Added behavioral coverage. The shared-helper test checks restriction precedence, signed-out and configured agent states, and preservation of non-agent behavior.
Verification passed: 22 focused harness/runtime/schedule tests, desktop TypeScript checking, and lint on all three incrementally changed files. Electron e2e and native mobile suites were not rerun for this desktop-only delta. Hosted CI on this head is still incomplete; this review does not establish merge readiness.
gpt-6.1-sol ｜ 𝕏
"custom GLM effort selector persists and sends every supported thinking mode" (#384) clicks levels that are revealed only while the control is hovered or focused. Playwright's click scrolls the target into view, which can move the group out from under the stationary pointer. The group then collapses, the radio jumps back, and the click times out with the radiogroup intercepting pointer events. The race is on main too: locally it failed 2/10 at the merge base and 1/10 on this branch, and it failed every attempt on CI shard 2 for this PR. The test now focuses the level first. Focus-within keeps the group expanded and never changes the level, so the click is still the action under test. With the fix: 20/20 locally. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
There was a problem hiding this comment.
✅ No new issues found.
Reviewed changes Reviewed 4044f405 since the prior Pullfrog review at 8a69115c, with the complete PR diff as context.
- Stabilized thinking-level clicks. Added focus and visibility checks before clicking each option, keeping the hover-revealed selector open during scrolling without forcing clicks or changing timeouts or retries.
- Preserved behavioral coverage. Confirmed that focus alone does not select a level and that the test retains checked-state assertions, request-payload checks for all five configured modes, relaunch persistence, and capability-reset coverage.
Verification passed: e2e TypeScript checking, lint on the changed spec, collection of its three Playwright tests, and seven focused thinking-control/provider-thinking tests. Full Electron execution requires macOS and was not run on this Linux runner. Hosted CI on 4044f405 is still incomplete; this review does not establish merge readiness.
gpt-6.1-sol ｜ 𝕏
…-upgrades-compare-42da8d Main claimed Aiden Remote contract revision 25 for the Bots rework (#377), so the phone simulator viewer moves to revision 26: the constant, fixture, desktop/iOS/Android revision assertions, the mobile fixture gate, and the simulator-scoped docs and OpenAPI text. Also fix main's chat-pane, which still read the removed bot query after #377 and #373 crossed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2 participants
Summary
Follow-up to #371 (now merged into
main).titletooltip that never appears (the sharedButtonsetspointer-events: nonewhen disabled).aria-describedby.accent, as the dialog's primary action.useAssistantChat, which no UI renders. Assistant chats still offered Antigravity, and the turn only failed after sending.Assets: Google's official Antigravity one-colour glyph, with its viewBox fitted to the glyph so it sits at the same optical size as the other provider marks.
Contracts:
AcpHarnessStatusgains optionalpublisherandruntime.downloadHost. It is desktop IPC only; there is no Aiden Remote protocol change.Test plan
npm run test:acp: 127/127 (afternpm run build:worktree-file-io; a stale native helper from before feat: Google Antigravity provider over ACP, with reusable ACP harness primitives #371 failsclient-files.test.ts)npx tsc --noEmit,npm run type-check:e2e, eslint on every changed TS file, CI registry testnode scripts/run-ci-tests.mjs --lane renderer: all jobs pass. That coversrenderer-other1956/1956 and the preserved CLI-package, generative-UI Chromium and iOS release-policy suites.runtime-subagents: 2711/2711.core-git: 2519/2520 (see the flake below).npm ci && npm run build && npm test, 79/79 (the bundle now reachesrenderer/shared/acp-harness.tsthroughprovider-deployment.ts)./gradlew :app:testDebugUnitTest :app:lintDebug :app:compileDebugAndroidTestKotlin: 422/422 unit tests, lint cleanAidenChatTestson the iPhone 18 Pro simulator, Xcode 27 beta: 261/261. This is simulator evidence, not device acceptance.tests/e2e/providers-settings.spec.ts, 4/4. The new test drives the real setup dialog with stubbed harness IPC through async progress, failure and completion, and through Escape in the remove confirmation. It fails against the earlier passive-effect focus recovery.main: Android provider logos come frommain's VectorDrawable generator (--checkreports all 43 up to date), andAidenProviderIconTest,AidenModelPreferenceTestandAidenChatTestpassFlake, existing on
mainand unrelated to this branch:main/services/external-editors.test.ts, "Linux Zed aliases preserve PATH precedence, executable guards and literal workspace argv".SyntaxError: Unexpected end of JSON input.core-gitlane runs; passes on its own (16/16).writeFileSync.🤖 Generated with Claude Code
CI failures on
8a69115c(run 37821395118)Each failed job was rerun once, as the repo's rules allow.
tests/e2e/custom-model-options.spec.ts, "custom GLM effort selector persists and sends every supported thinking mode" (from Add per-model effort settings for custom connections #384).locator.clicktimed out and the radiogroup intercepted pointer events. It failed on both attempts and both retries.mainmerge base and 1/10 on this branch.4044f405): the test focuses the level first. Focus keeps the control open and never changes the level, so the click is still what's tested. 20/20 locally after the fix.AidenChatTests.testChatOpenFetchesTheProgressSnapshotOnceAndReloadsDoNotDuplicateIt()failed on the hosted iPhone 17 Pro simulator, with app-launch failures in the same job.AidenChatTests262/262 on the iPhone 18 Pro simulator (Xcode 27 beta).CI failure on
4044f405(run 37834947647)RootViewPicker$RootViewWithoutFocusException, because the emulator window never gained focus:AidenChatProgressUiTest.nestedAgentBackRestoresParentThenRosterAfterRecreationAidenChatProgressUiTest.chatInspectorOwnerRestoresAfterDelayedNegotiationAndRosterHydrationAidenChatTests.testForeignRunResponseStaysVisibleUntilAFailedTranscriptReadRecovers()failed after 63.8 s, with "Failed to launch app" from the hosted simulator in the same job.