fix(macos): emit llm.activeProfile=balanced for managed-inference hatch

[vellum-apollo-bot]

What

The macOS managed-inference onboarding path was emitting llm.default.provider=anthropic to the daemon hatch overlay - same shape as BYOK. The daemon's seedInferenceProfiles (assistant/src/config/seed-inference-profiles.ts) reads llm.default.provider, materializes an anthropic-personal connection backed by credential/anthropic/api_key, and (after #30232 flipped the default) sets llm.activeProfile = custom-balanced. In managed inference that credential never exists - the user signed in with their Vellum account and gets a platform-injected assistant_api_key instead - so message sends 500'd.

Fix: in HatchingStepView.buildOnboardingConfigValues, branch on state.skippedAPIKeyEntry (set exclusively in APIKeyStepView.handleContinue when the user is authenticated and skipped the provider API key step). Managed: emit llm.activeProfile=balanced. BYOK: emit llm.default.provider=<provider> as before.

The daemon's mergeDefaultWorkspaceConfig already tracks providedLlmActiveProfile and passes preserveActiveProfile=true to the seeder. With preserveActiveProfile=true and profiles["balanced"] already seeded (it's a managed profile, always materialized), shouldPreserveActiveProfile is true and the seeder skips its userConnectionName ? "custom-balanced" : "balanced" branch entirely. With no llm.default.provider in the overlay, the seeder also doesn't materialize the personal connection.

Why

hello-world-managed-inference (Playwright) has been red on every scheduled run since 2026-05-10 (#30232 landed May 10). Last green pre-flip: run 25569878332 on May 8.

The defect predates #30232 - the macOS app has been emitting the same wrong overlay for the managed path the whole time. Before the PR-D flip, the seeder defaulted activeProfile to balanced even when anthropic-personal was materialized, so chat still worked. PR-D flipped that default so the seeder now points at the user connection when one exists, which exposed the underlying onboarding-flow bug.

How

• New pure top-level helper onboardingHatchConfigOverlay(skippedAPIKeyEntry:selectedProvider:defaultProvider:) in clients/macos/vellum-assistant/Features/Onboarding/OnboardingHatchConfigOverlay.swift. Matches the existing OnboardingHostingModeResolver / OnboardingManagedAuthState convention of "extract logic into a top-level function so it can be unit-tested without spinning up state."
• HatchingStepView.buildOnboardingConfigValues becomes a one-line delegation.
• 8 new unit tests in OnboardingHatchConfigOverlayTests.swift cover both paths (managed-inference and BYOK), the empty-selectedProvider fallback, and the structural invariants (managed doesn't emit default.provider; BYOK doesn't emit activeProfile).

Trace through the daemon

For the managed path with this change:

1. CLI writes {"llm": {"activeProfile": "balanced"}} to $VELLUM_DEFAULT_WORKSPACE_CONFIG_PATH.
2. Daemon boots. mergeDefaultWorkspaceConfig returns {hadOverlay: true, providedLlmProfileNames: {}, providedLlmActiveProfile: true}.
3. lifecycle.ts:522 calls seedInferenceProfiles({isHatch: true, preserveActiveProfile: true, ...}).
4. Seeder materializes managed profiles (incl. balanced).
5. hatchProvider = readString(readObject(llm.default)?.provider) is undefined (no default.provider in overlay), so userConnectionName stays undefined and the anthropic-personal connection is NOT created.
6. requestedActiveProfile = "balanced", requestedActiveExists = true, shouldPreserveActiveProfile = true. The entire if (options.isHatch) branch is skipped. llm.activeProfile stays "balanced".

Test plan

• OnboardingHatchConfigOverlayTests covers the helper.
• Scheduled hello-world-managed-inference Playwright run is the integration check.
• BYOK path is unchanged in shape - llm.default.provider=<X> still emitted when skippedAPIKeyEntry == false.

Related

• Failing run: https://github.com/vellum-ai/vellum-assistant-platform/actions/runs/25686847287
• Suspect PR that exposed this: feat(inference): always-seed managed profile triplet (Phase 1.2 PR-D) #30232 (Phase 1.2 PR-D)

---

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no potential bugs to report.

View in Devin Review to see 2 additional findings.

[vellum-apollo-bot]

Parking this pending the test-side approach. Vargas redirected: the fix should land in qa/agent/fixtures.ts setManagedInferenceMode so the managed-inference fixture is correct for all consumers, not in the macOS onboarding overlay. See vellum-ai/vellum-assistant-platform#TBD (replacement PR). Leaving open in case we want both layers belt-and-suspenders after the fixture change ships and we see how hello-world-managed-inference lands.