refactor(web): share one setup wizard and finish Antigravity setup from model pickers

[rynfar]

Setup flows used three different stepper implementations with repeated dialog chrome, and window capture brought back a second copy (ui/wizard-steps) beside the shared wizard from the welcome wizard port. Separately, #455 left out upstream's provider-setup link, and Pylon's Antigravity UI port (#440) left out the client pieces that make Antigravity usable from the picker: a provider that still needs setup could not be set up from the model picker, the internal default model marker could show up raw, the Settings refresh never reloaded account models, and every server restart showed a sign-in banner for an account that was still signed in.

What changed

Shared wizard components (#10832). Welcome, Add provider, Publish repository, Pylon Connect and SnapShot setup now use WizardPopup, WizardHeader, WizardSteps, WizardPanel and WizardFooter. The duplicate ui/wizard-steps.tsx is deleted; the Connect dialog's switch to WizardSteps from #8103 is carried into the shared stepper. Publish repository keeps its completed-step summaries.

Provider-setup link (#9348 client hunks, #9754 project defaults hook-up). When an enabled provider with integrated setup (Antigravity, or any provider whose server reports setup.canInstall/canAuthenticate) is not installed, signed out, or has no usable model:

• the model picker shows the environment's status message with Open provider setup (or Set up when several need it) and keeps that instance visible in the sidebar;
• the provider status banner above a thread shows the same link;
• the composer shows Open provider settings in place of the picker when no provider resolves (for the requested instance, or the first provider with setup when the thread is not locked), and when the thread's bound instance is disabled and no enabled, available continuation peer could take the thread (with such a peer, the picker stays so the user can switch).

Each opens Settings → Providers on the owning device with that instance selected, via /settings/providers?environmentId=…&instanceId=…. A cold-opened link waits for the device list; a device that is not in the loaded list shows "Device unavailable", while a saved device that is disconnected keeps its usual "Could not connect to this device" state. A removed instance says so instead of opening another account. Pickers wired: composer, Settings → Projects machine default and project default model, text generation model, source control writer model.

Status messages keep the environment's own error (for example a subscription error) and point integrated-setup providers to setup instead of "Sign in via the CLI".

Antigravity client completion (#9348, #9647).

• The antigravity-default marker resolves to the account catalog's default model in the picker, the trigger label ("Choose model" until a catalog exists) and the composer's model state. Antigravity never borrows another account's model or a static default, keeps an unavailable saved model visible, and no longer lists or offers custom models in Settings → Providers (they were never sent).
• The composer blocks Antigravity sends with a reason when it is not installed, signed out, has no model chosen, has not loaded its catalog, or the saved model left a ready catalog. After a server restart (installed, auth not checked yet) a thread with a chosen model sends and session startup checks the saved sign-in; a new thread or default-model selection, which has no model until the catalog loads, is asked to refresh models rather than to choose from an empty list.
• The status banner stays hidden in that restart state, so it no longer invites a needless re-sign-in; install failures, errors and confirmed sign-outs still show.
• Settings → Providers refresh and the mobile model refresh send refreshModels, so the server reloads account model catalogs.

Sources

Upstream SHA	Upstream PR	Outcome
3faeee49ac67dfd9534369f1e1c627c0356b75ac	#10832	Adopted
d487dfbf46be344e818725be70ee04be2436bfb4	#9647	Adopted (banner guard and send-block change; its Antigravity doc wording lands with the docs restructure PR)
06336460c9988f29c71e839c4c9c840c4552e077	#9348	Client port completed here (setup link, route target, status messages, default-model marker, send blocks, refreshModels); server and earlier UI pieces came through the Antigravity ports
9f40b2f563c662b43887b11ff99c466fe871c1af	#9754	Completes the project-defaults onOpenProviderSetup hunk excluded from #455

Pylon adaptations

• Branding kept in the wizards: PylonMark and "Set up Pylon" in the welcome wizard header (as WizardHeader identity), Pylon Connect copy, Pylon SnapShots/capture helper copy and GNOME window-picker guidance.
• Add provider keeps Pylon behavior: multiple-instance blocking (disabled drivers, Next/Add disabled with the reason) and Prime instance/ACP/maintenance guidance and home validation.
• GitActionsControl keeps useThreadShell and the "Unable to open repository" toast.
• Composer selection: upstream derives the setup target in resolveComposerProviderSelection; Pylon's routing lives in resolveComposerInstanceSelection (sessions, drains, continuation transitions). It now reports unavailableInstanceId (the locked instance first, a disabled session-bound instance, never the empty draft placeholder), and resolveComposerProviderSettingsAction decides when the settings button replaces the picker. Explicit unavailability (for example Prime on native Windows) keeps the picker so another provider can be chosen.
• Send blocks join Pylon's existing send-disabled chain after bound-provider conflicts and attachment blocks, and do not apply while answering a pending question.
• Picker and banner: Prime's single-model provider message, locked continuation compatibility checks, and the explicit-unavailability banner are unchanged.
• Model selection (Pylon fix; upstream has the defect): #9348 made an instance whose listed models are all hidden resolve to no model, which let driver-level fallbacks pick another instance's model or provider (a custom Claude router instance sending the default Claude model, a Claude instance getting a Codex model when default Claude is disabled, text generation using the first catalog model). Hiding is a picker preference, so resolution stays on the instance's own default; Antigravity still resolves to nothing outside its account catalog. Continue on another account falls back to the target account's default model and shows a toast when the account reports no model.
• Empty Antigravity catalog (Pylon fix; upstream has the defect): with no model chosen and no catalog, the composer asks to refresh models instead of "Choose an Antigravity model", which the user cannot act on.
• Mobile: only the refresh change. Upstream removed mobile provider setup in #9721 (already adopted in feat(mobile): add Android agent notifications and Material You theming #456). Upstream's mobile Antigravity send blocks (isModelSelectionUnavailable alerts in the thread composer and restores in the outbox drain) are not ported. Pylon mobile already holds disabled, uninstalled, signed-out, errored and unavailable providers through the shared provider admission check on new tasks and queued sends, and the server rejects a vanished model at turn start. Upstream's extra catalog-membership test would hold every Antigravity send while the catalog is empty after a restart, and every antigravity-default selection (the marker is only an alias), which contradicts #9647. Revisit when upstream fixes those cases or mobile gains marker resolution and catalog reload on its send path.

Verification

• vp test run: 818 web tests in 32 files and 2 mobile tests pass. Web files: AddProviderInstanceDialog (environment, logic), AddProviderInstanceWizardSteps, SnapShotSettings, SnapShotSetupDialog.logic, providerMultipleInstances, GitActionsControl.logic, ChatView.logic, ModelPickerContent (ts, tsx), ProviderModelPicker, ProviderStatusBanner, ProviderSettingsPanel (environment, logic), ProviderInstanceCard, ProviderModelsSection, settingsSearch, composerInstanceSelection, ProjectSettingsPanel.logic, SettingsPanels.logic, modelSelection, composerDraftStore, providerInstances, ThreadHandoff.logic, ExpandedImagePreview, composerAttachmentFiles, SnapShotCoordinator, PreviewView, useHandleNewThread, attachmentUploadQueue, composerProviderState, composer-logic. Mobile: provider-catalog-refresh.
• New tests: setup-aware status messages and banner link; the hidden restart banner; shouldOfferModelPickerSetup; the default-marker resolution and trigger label; Antigravity model state and custom-model isolation (selection and settings card); upstream's OpenCode and Antigravity catalog-gap describe.each; send-block reasons including the empty catalog; panel target, cold-open loading and removed-instance states; unavailableInstanceId for a disabled binding with and without a continuation peer; the settings action decision; all-hidden fallbacks for a custom Claude instance, a disabled default Claude instance and text generation; handoff with a hidden or empty target. The three all-hidden scenario tests fail without the fallback fix.
• vp run -F @t3tools/web typecheck, vp run -F @t3tools/mobile typecheck and vp run -F @t3tools/client-runtime typecheck: 0 error TS.
• vp lint and vp fmt --check on every changed file: 0 warnings, 0 errors, formatted.
• Conflict resolution check: each wizard caller differs from upstream's post-#10832 file only by Pylon's pre-existing divergence.
• No local client pass was run; the orchestrator does one integrated pass. Upstream evidence: #10832 before, after, step navigation recording; setup link and Antigravity picker UI in #9348.

User docs for the setup link and model refresh land with the docs restructure PR (#482), which rewrites the Antigravity guide and merges after this PR.

Ported by Claude Opus 5 in Claude Code.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
pylon-marketing	Ready	Preview	Sep 11, 2026 3:11pm UTC

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	14.1 KiB	13.9 KiB	−164 B (−1.1%)	15.1 KiB	✅
Codex	Thread snapshot wire	7.2 KiB	7.2 KiB	+5 B (+0.1%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.9 KiB	6.7 KiB	−169 B (−2.4%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	58.8 KiB	58.0 KiB	−910 B (−1.5%)	66.4 KiB	✅
Codex	Live turn messages	10	8	−2 (−20.0%)	21	✅
Claude	Total thread wire	14.1 KiB	13.9 KiB	−172 B (−1.2%)	15.1 KiB	✅
Claude	Thread snapshot wire	7.2 KiB	7.2 KiB	−1 B (−0.0%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.9 KiB	6.7 KiB	−171 B (−2.4%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	59.7 KiB	58.8 KiB	−928 B (−1.5%)	66.4 KiB	✅
Claude	Live turn messages	10	8	−2 (−20.0%)	21	✅

Baseline: 51545d7 · PR result: 9503961 · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 115.6 KiB
• Claude decoded thread snapshot: 116.3 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.