Researched: 2026-04-03 Domain: Azure Voice Live instance management, Agent voice mode configuration, AI Foundry portal workflow automation Confidence: HIGH

Phase 13 aims to provide admin-level management of Voice Live instances (model selection) and Agent voice mode configuration, matching the AI Foundry portal's end-to-end workflow. The critical discovery from research is that Azure Voice Live is a fully managed service with no separate "instance" REST API. There is no Azure REST endpoint to "create a Voice Live instance" or "bind Voice Live to an agent" -- these are UI-only concepts in the AI Foundry portal that translate to:

1. Model selection: The generative AI model (gpt-4o, gpt-4.1, gpt-5, etc.) is specified as a query parameter at WebSocket connection time. The "instance" is the active WebSocket session itself.
2. Bind to agent: Writing microsoft.voice-live.configuration metadata on the agent (already implemented in agent_sync_service.py via build_voice_live_metadata()).
3. Enable Voice mode: Presence of the microsoft.voice-live.configuration metadata key on the agent signals that voice mode is enabled.
4. Speech/Avatar config: Stored as JSON in the agent metadata, sent to Voice Live API via session.update at connection time.

The existing codebase (Phase 11-12) already implements items 2-4 via build_voice_live_metadata() and per-HCP token broker. What Phase 13 adds is: (a) explicit admin UI for selecting the Voice Live generative AI model per-HCP, (b) a dedicated Voice Live management page showing the full HCP-to-Voice-Live binding chain, (c) upgrading azure-ai-projects from 1.0.0b12 to 2.0.1 (the stable release), and (d) comprehensive testing of the automated full chain.

Primary recommendation: Since no new Azure APIs need to be called (Voice Live has no instance management API), this phase is primarily a UI/UX enhancement phase: add model selection to HCP profiles, build a Voice Live management overview admin page, and ensure the existing agent metadata sync covers all Voice Live configuration fields.

<phase_requirements>

Installation:

# Backend -- upgrade azure-ai-projects to stable 2.0.1
cd backend
pip install "azure-ai-projects>=2.0.1"

# Frontend -- no new packages needed

Version verification:

• azure-ai-projects: Currently installed 1.0.0b12, must upgrade to 2.0.1 (verified via pip index versions)
• All other packages already installed from previous phases

backend/
  alembic/versions/
    k14a_add_voice_live_model_to_hcp_profile.py   # NEW: migration
  app/
    models/hcp_profile.py                          # EXTEND: voice_live_model column
    schemas/hcp_profile.py                         # EXTEND: voice_live_model field
    schemas/voice_live.py                          # EXTEND: model list response
    services/voice_live_service.py                 # EXTEND: per-HCP model in token
    services/agent_sync_service.py                 # EXTEND: model in metadata, SDK upgrade compat
    api/voice_live.py                              # EXTEND: management endpoints
    api/hcp_profiles.py                            # NO CHANGE (already handles all fields)

frontend/
  src/
    types/hcp.ts                                   # EXTEND: voice_live_model field
    types/voice-live.ts                            # EXTEND: model list type
    pages/admin/voice-live-management.tsx           # NEW: Voice Live management overview page
    components/admin/voice-live-chain-card.tsx      # NEW: Per-HCP chain status card
    components/admin/voice-avatar-tab.tsx           # EXTEND: model selection dropdown
    hooks/use-voice-live-management.ts              # NEW: TanStack Query hooks for management
    api/voice-live.ts                              # EXTEND: management API calls
    public/locales/en-US/admin.json                # EXTEND: management page strings
    public/locales/zh-CN/admin.json                # EXTEND: management page strings

What: Each HCP profile stores the generative AI model to use for Voice Live sessions. This model is passed as a query parameter when connecting the WebSocket. When to use: Admin configures per-HCP Voice Live settings. Example:

# Source: Azure Voice Live docs -- model is a WebSocket connection parameter
# wss://<resource>.services.ai.azure.com/voice-live/realtime?api-version=2025-10-01&model=gpt-4.1

# HCP profile stores the model choice
voice_live_model: Mapped[str] = mapped_column(
    String(50), default="gpt-4o"
)  # gpt-4o, gpt-4.1, gpt-5, gpt-realtime, phi4-mini, etc.

What: Admin page showing the 4-step workflow chain for each HCP: Profile -> Agent -> Voice Live Config -> Speech/Avatar. When to use: Voice Live management admin page. Example:

// Each HCP shows its chain status:
// Step 1: HCP Profile (always exists)
// Step 2: Agent (synced/pending/failed/none)
// Step 3: Voice Live Config (voice_live_enabled + model selection)
// Step 4: Speech/Avatar (voice_name + avatar_character configured)
interface ChainStatus {
  hcpId: string;
  hcpName: string;
  agentStatus: "synced" | "pending" | "failed" | "none";
  agentId: string;
  voiceLiveEnabled: boolean;
  voiceLiveModel: string;
  voiceName: string;
  avatarCharacter: string;
  avatarStyle: string;
}

What: Admin can trigger batch re-sync of all HCP agents to update Voice Live metadata after changing model or config. When to use: After changing Voice Live model for multiple HCPs, or after SDK upgrade. Example:

# Source: Existing sync_agent_for_profile() pattern in agent_sync_service.py
async def batch_resync_agents(
    db: AsyncSession,
    hcp_profile_ids: list[str] | None = None,
) -> dict:
    """Re-sync all (or selected) HCP agents with current metadata.

    Useful after changing Voice Live model settings or upgrading SDK.
    Returns summary of results: {synced: N, failed: N, errors: [...]}
    """
    # Prefetch config once for all profiles
    endpoint, api_key, model = await prefetch_sync_config(db)
    results = {"synced": 0, "failed": 0, "errors": []}
    # ... iterate profiles and call sync_agent_for_profile()

• Creating a separate "Voice Live Instance" table: Azure has no concept of persistent instances -- Voice Live is session-based. Store the model choice on the HCP profile.
• Calling a non-existent "create Voice Live instance" API: No such API exists. The portal UI creates the WebSocket session on-demand.
• Ignoring the SDK version upgrade: The installed 1.0.0b12 is a beta; 2.0.1 is stable with breaking changes to the agent API surface.
• Storing model config globally only: Each HCP can use a different model in the AI Foundry portal. Per-HCP model selection is required.

What goes wrong: The agent_sync_service.py uses client.agents.create_version(), client.agents.get(), client.agents.delete() which may have different signatures in v2.0.1. Why it happens: The project is on beta 1.0.0b12 but the memory says >=2.0.1. The stable release 2.0.1 may have different class/method names. How to avoid: Before upgrading, check the v2.0 migration guide. Test create_version, get, delete methods against the new API surface. The PromptAgentDefinition import path may change. Warning signs: ImportError or AttributeError on server startup after pip upgrade.

What goes wrong: The Voice Live model (e.g., gpt-4.1) is confused with the Agent model (e.g., gpt-4o used for agent instructions). These are different. Why it happens: The ServiceConfig.model_or_deployment for azure_voice_live already stores a model/agent-mode config. The agent's own model (set in PromptAgentDefinition) is separate from the Voice Live WebSocket model. How to avoid: Voice Live model is the model parameter passed in the WebSocket URL. In agent mode, the agent already has its own LLM model. The Voice Live model may be used for non-agent sessions or as the orchestration model. Store as voice_live_model on HCP profile separately from the agent's model field. Warning signs: Admin selects gpt-4.1 for Voice Live but the agent still uses gpt-4o for its own responses.

What goes wrong: When connecting in agent mode, the model query parameter is not used -- the agent's own model is used instead. Passing both model and agent_id may cause errors. Why it happens: Azure Voice Live docs say: "The only difference is the required model query parameter, or, when using the Agent service, the agent_id and project_id parameters." This is an either/or -- not both. How to avoid: When HCP has a synced agent (agent_id), use agent mode (no model param). When HCP has no agent, use model mode with the selected voice_live_model. The token broker already handles this distinction. Warning signs: WebSocket connection fails with "invalid parameters" when both model and agent_id are provided.

What goes wrong: Agent metadata values are limited to 512 characters per key. Complex Voice Live config JSON may exceed this. Why it happens: Azure AI agent metadata has a 512-character value limit. How to avoid: Already handled by _chunk_metadata_value() in agent_sync_service.py, which splits long values into key, key.1, key.2, etc. Warning signs: Agent creation/update fails with metadata size error.

What goes wrong: Alembic migration fails on SQLite when adding the voice_live_model column. Why it happens: Standard project gotcha -- SQLite needs batch_alter_table with server_default. How to avoid: Use with op.batch_alter_table("hcp_profiles") as batch_op: with server_default=sa.text("'gpt-4o'"). Warning signs: Migration fails locally on SQLite but works on PostgreSQL.

# Source: https://learn.microsoft.com/en-us/azure/ai-services/speech-service/voice-live
# Verified: 2026-04-03, docs updated 2026-02-04

VOICE_LIVE_MODELS = {
    # Pro tier
    "gpt-realtime": {"tier": "pro", "description": "GPT real-time + Azure TTS"},
    "gpt-4o": {"tier": "pro", "description": "GPT-4o + Azure STT/TTS"},
    "gpt-4.1": {"tier": "pro", "description": "GPT-4.1 + Azure STT/TTS"},
    "gpt-5": {"tier": "pro", "description": "GPT-5 + Azure STT/TTS"},
    "gpt-5-chat": {"tier": "pro", "description": "GPT-5 chat + Azure STT/TTS"},
    # Basic tier
    "gpt-realtime-mini": {"tier": "basic", "description": "GPT mini real-time + Azure TTS"},
    "gpt-4o-mini": {"tier": "basic", "description": "GPT-4o mini + Azure STT/TTS"},
    "gpt-4.1-mini": {"tier": "basic", "description": "GPT-4.1 mini + Azure STT/TTS"},
    "gpt-5-mini": {"tier": "basic", "description": "GPT-5 mini + Azure STT/TTS"},
    # Lite tier
    "gpt-5-nano": {"tier": "lite", "description": "GPT-5 nano + Azure STT/TTS"},
    "phi4-mm-realtime": {"tier": "lite", "description": "Phi4-mm realtime + Azure TTS"},
    "phi4-mini": {"tier": "lite", "description": "Phi4-mini + Azure STT/TTS"},
}

// Source: Azure Voice Live overview docs, organized by pricing tier
const VOICE_LIVE_MODEL_OPTIONS = [
  // Pro tier
  { value: "gpt-realtime", label: "GPT Realtime", tier: "pro" },
  { value: "gpt-4o", label: "GPT-4o", tier: "pro" },
  { value: "gpt-4.1", label: "GPT-4.1", tier: "pro" },
  { value: "gpt-5", label: "GPT-5", tier: "pro" },
  { value: "gpt-5-chat", label: "GPT-5 Chat", tier: "pro" },
  // Basic tier
  { value: "gpt-realtime-mini", label: "GPT Realtime Mini", tier: "basic" },
  { value: "gpt-4o-mini", label: "GPT-4o Mini", tier: "basic" },
  { value: "gpt-4.1-mini", label: "GPT-4.1 Mini", tier: "basic" },
  { value: "gpt-5-mini", label: "GPT-5 Mini", tier: "basic" },
  // Lite tier
  { value: "gpt-5-nano", label: "GPT-5 Nano", tier: "lite" },
  { value: "phi4-mm-realtime", label: "Phi4-MM Realtime", tier: "lite" },
  { value: "phi4-mini", label: "Phi4 Mini", tier: "lite" },
] as const;

# Source: Extend existing voice_live_service.py
# In model mode: use HCP's voice_live_model (default "gpt-4o")
# In agent mode: model is NOT sent (agent has its own model)

# In get_voice_live_token():
if is_agent:
    # Agent mode: no model param needed -- agent has its own LLM
    model_for_session = ""
else:
    # Model mode: use HCP-level or global config model
    model_for_session = (
        profile.voice_live_model
        if hcp_profile_id and profile
        else mode_info.get("model", "gpt-4o")
    )

// Each HCP shows its full chain status
// Reuses existing Card, Badge, and cn() patterns
<Card>
  <CardHeader>
    <CardTitle>{hcp.name} - {hcp.specialty}</CardTitle>
  </CardHeader>
  <CardContent>
    <div className="flex items-center gap-3">
      {/* Step 1: Agent */}
      <Badge variant={agentOk ? "default" : "destructive"}>
        Agent: {hcp.agent_sync_status}
      </Badge>
      {/* Step 2: Voice Live */}
      <Badge variant={hcp.voice_live_enabled ? "default" : "secondary"}>
        Voice: {hcp.voice_live_enabled ? hcp.voice_live_model : "Disabled"}
      </Badge>
      {/* Step 3: Speech */}
      <Badge variant="outline">
        {hcp.voice_name}
      </Badge>
      {/* Step 4: Avatar */}
      <Badge variant="outline">
        {hcp.avatar_character}/{hcp.avatar_style}
      </Badge>
    </div>
  </CardContent>
</Card>

# Source: Extend build_voice_live_metadata() in agent_sync_service.py
# The model is NOT stored in agent metadata -- it's a runtime connection parameter.
# Agent metadata stores speech/avatar config that the Voice Live session needs.
# Model selection is handled by the token broker at session start.
#
# Existing metadata format (already works):
# {
#   "microsoft.voice-live.configuration": '{"voice":{"type":"azure-standard","name":"en-US-AvaNeural","temperature":0.9},"turn_detection":{"type":"server_vad"}}'
# }

Azure Voice Live API supported models (current, verified 2026-04-03):

• Pro: gpt-realtime, gpt-4o, gpt-4.1, gpt-5, gpt-5-chat
• Basic: gpt-realtime-mini, gpt-4o-mini, gpt-4.1-mini, gpt-5-mini
• Lite: gpt-5-nano, phi4-mm-realtime, phi4-mini

Critical insight: Voice Live has no instance management API. The AI Foundry portal "Voice Live playground" is a UI experience that connects a WebSocket with model selection. There is no REST API to "create a Voice Live instance" that persists on Azure. The "instance" is simply a configuration choice (model + speech + avatar) that exists in the platform's database and is applied at WebSocket connection time.

1. SDK upgrade impact on agent_sync_service.py

   • What we know: azure-ai-projects 2.0.1 is stable; installed version is 1.0.0b12 (beta). The PromptAgentDefinition, client.agents.create_version(), client.agents.get(), client.agents.delete() are the methods used.
   • What's unclear: Whether v2.0.1 has the same API surface as v1.0.0b12. Method signatures may have changed.
   • Recommendation: Upgrade in a dedicated plan. Test each method. If API surface changed, adapt agent_sync_service.py. The memory note says v2.0+ uses client.agents.create_version() which matches current code -- likely compatible.

2. Per-HCP model vs global config model

   • What we know: Currently model_or_deployment on the azure_voice_live ServiceConfig stores the global model/agent config. HCP profiles don't have a voice_live_model field.
   • What's unclear: Whether the global config model should serve as a default that HCP-level overrides, or whether per-HCP model should be the only source.
   • Recommendation: Add voice_live_model to HcpProfile with default "gpt-4o". Global model_or_deployment remains as system-level agent/model mode config. Per-HCP voice_live_model is used for model-mode sessions. Agent-mode sessions ignore it (agent has its own LLM).

3. Management page as separate route or tab in existing HCP page

   • What we know: AI Foundry portal has a separate "Voice Live" section showing all Voice Live configurations per agent.
   • What's unclear: Whether a new admin page is needed or the existing HCP table is sufficient.
   • Recommendation: New admin route /admin/voice-live showing chain overview across all HCPs, with links to individual HCP editors. The HCP editor already has the Voice & Avatar tab for per-HCP config.

• Async everywhere: all backend functions must be async def
• Pydantic v2 schemas with model_config = ConfigDict(from_attributes=True)
• Route ordering: static paths before parameterized (/{id})
• Service layer holds business logic, routers only handle HTTP
• No raw SQL -- use SQLAlchemy ORM
• TypeScript strict mode: no any, no unused variables
• TanStack Query hooks per domain, no inline useQuery
• cn() for conditional class composition
• i18n: all UI text externalized via react-i18next
• Conventional commits: feat:, fix:, docs:, test:

• NEVER modify schema without Alembic migration
• All models use TimestampMixin
• batch_alter_table with server_default for SQLite compatibility

• Backend: ruff check ., ruff format --check ., pytest -v
• Frontend: npx tsc -b, npm run build

• Use python3 not python in local commands
• Unit tests MUST use real .env credentials when available
• Always complete full workflow: fix -> commit -> push -> CI verify
• Each phase needs >=95% test coverage
• All UI communication in Chinese with user, English in code/commits

Missing dependencies with no fallback:

• azure-ai-projects must be upgraded from 1.0.0b12 to >=2.0.1 (stable release)

Missing dependencies with fallback:

• None -- all runtime dependencies are already available

• Azure Voice Live how-to: https://learn.microsoft.com/en-us/azure/ai-services/speech-service/voice-live-how-to -- Session configuration, authentication, model selection, avatar config, voice config. Updated 2026-03-16.
• Azure Voice Live overview: https://learn.microsoft.com/en-us/azure/ai-services/speech-service/voice-live -- Supported models list (12 models in 3 tiers), pricing, architecture. Updated 2026-02-04.
• Azure Voice Live customization: https://learn.microsoft.com/en-us/azure/ai-services/speech-service/voice-live-how-to-customize -- Custom speech, custom voice, personal voice, custom avatar config. Updated 2026-04-02.
• Existing codebase files (all read directly):
  • backend/app/services/agent_sync_service.py -- Agent CRUD with build_voice_live_metadata(), sync_agent_for_profile()
  • backend/app/services/voice_live_service.py -- Token broker with per-HCP resolution
  • backend/app/models/hcp_profile.py -- ORM model with 13 voice/avatar columns
  • backend/app/schemas/hcp_profile.py -- Pydantic schemas with all voice/avatar fields
  • backend/app/schemas/voice_live.py -- Token response and status schemas
  • backend/app/api/voice_live.py -- Token broker and status endpoints
  • backend/app/models/service_config.py -- Global config with model_or_deployment
  • backend/app/services/agents/adapters/azure_voice_live.py -- parse_voice_live_mode(), encode_voice_live_mode()
  • backend/app/services/region_capabilities.py -- Region availability maps
  • frontend/src/hooks/use-voice-live.ts -- Voice Live WebSocket session with buildSessionConfig()
  • frontend/src/hooks/use-avatar-stream.ts -- Avatar WebRTC connection
  • frontend/src/components/admin/voice-avatar-tab.tsx -- Voice/Avatar admin settings with live test
  • frontend/src/pages/admin/hcp-profile-editor.tsx -- Tabbed HCP editor
  • frontend/src/types/hcp.ts -- HCP TypeScript types
  • frontend/src/types/voice-live.ts -- Voice Live types

• pip index verification: azure-ai-projects latest stable is 2.0.1, installed is 1.0.0b12
• User memory (2026-04-03): AI Foundry Voice Live complete workflow with model selection and agent binding steps

• SDK v2.0.1 API surface compatibility with v1.0.0b12 -- needs runtime verification during upgrade plan

Confidence breakdown:

• Standard stack: HIGH - all libraries already in project, SDK upgrade path clear
• Architecture: HIGH - Voice Live has no instance management API (verified from official docs), pattern is config-driven
• Pitfalls: HIGH - SDK upgrade risk identified, model/agent-mode distinction well-understood from docs
• Azure API structure: HIGH - verified from official docs updated 2026-03-16 and 2026-04-02

Research date: 2026-04-03 Valid until: 2026-05-03 (stable -- Azure Voice Live API is GA, model list may expand but existing models remain)

Visual and interaction contract for the Voice Live Instance & Agent Voice Management phase. Generated by gsd-ui-researcher, verified by gsd-ui-checker.

Design Reference: Azure AI Foundry portal's 3-step Voice Live workflow (user-provided screenshots 2026-04-03):

1. Voice Live Playground — Select Generative AI Model (Pro/Standard/Lite tiers), configure Response instruction, Temperature, Proactive engagement, Speech input/output
2. Add to Agent — Dropdown binding Voice Live to existing Agents (HCP profiles)
3. Agent Page — Voice mode toggle, Instructions, Tools, Knowledge, Memory, Guardrail, Avatar preview, Speech input (Language + Auto-detect), Speech output (Voice selection)

Our admin UI should mirror this workflow while adapting to our HCP-centric data model.

Source: Existing frontend/src/styles/index.css @theme inline block, established in Phase 01. No new design system installations required. Continuity from Phase 12 UI-SPEC.

Declared values (must be multiples of 4):

Exceptions: none

Two weights only: 400 (normal) for body text and labels, 600 (semibold) for headings. Follows Phase 12 contract.

Accent reserved for:

• Batch Re-sync All button (bg-primary)
• Active/complete chain step connector line accent
• "Edit HCP" link text in chain cards

Additional semantic colors used in this phase (already established):

Source: Existing CSS custom properties in index.css, Phase 10 theme system. Tier-specific tints are new but follow the established badge color pattern (e.g., personality badge uses bg-slate-100 text-slate-700).

Trigger: Admin opens Voice & Avatar tab in HCP Profile Editor. Behavior: A new FormField "Voice Live Model" appears inside the Voice Settings card, between the Voice Live Config toggle and the Voice Name field. The field renders VoiceLiveModelSelect, which is a grouped Select dropdown. Groups: Pro (gpt-realtime, gpt-4o, gpt-4.1, gpt-5, gpt-5-chat), Basic (gpt-realtime-mini, gpt-4o-mini, gpt-4.1-mini, gpt-5-mini), Lite (gpt-5-nano, phi4-mm-realtime, phi4-mini). Default: "gpt-4o". Selection persists as part of the HCP form and is saved with Save Profile. Visual: Select trigger h-8 text-xs matching existing voice-avatar-tab dropdowns. Group labels rendered as non-interactive tier headers inside SelectContent: "Pro", "Basic", "Lite" each in text-xs font-semibold text-muted-foreground px-2 py-1.5. SelectItems show model display name (e.g., "GPT-4o", "GPT-4.1 Mini"). Constraint: Model selection is independent of agent mode vs model mode. In agent mode, the selected model is NOT sent to the WebSocket (agent has its own LLM). In model mode, the selected model IS sent as the WebSocket query parameter. This distinction is handled by the token broker, not the UI.

Trigger: Admin clicks "Voice Live" in the admin sidebar. Behavior: Navigates to /admin/voice-live. Page loads all HCP profiles via existing useHcpProfiles() hook. Displays summary statistics row and chain card grid. Page uses AdminLayout with standard sidebar and header. Visual: Page title "Voice Live Management" / "Voice Live 管理" at text-2xl font-semibold. Description text below at text-sm text-muted-foreground. Summary stats as 4 small stat cards in a row (grid grid-cols-2 md:grid-cols-4 gap-4). Chain cards grid below: grid grid-cols-1 md:grid-cols-2 xl:grid-cols-3 gap-6.

Trigger: Management page renders a VoiceLiveChainCard for each HCP. Behavior: Each card shows 4 chain steps in order:

1. HCP Profile -- always green (profile exists). Shows HCP name.
2. Agent -- reflects agent_sync_status: green (synced), amber (pending), red (failed), muted (none). Shows agent_id or status label.
3. Voice Live -- reflects voice_live_enabled + voice_live_model: green (enabled + model selected), muted (disabled). Shows model name or "Disabled".
4. Speech/Avatar -- reflects voice_name + avatar_character: green (both configured), amber (partial -- only voice or only avatar), muted (defaults only). Shows voice short name + avatar character/style.

Steps are connected by a vertical line on the left side. Each step is a row with: status dot (size-2 rounded-full) | step label (text-xs text-muted-foreground) | step value (text-sm).

Visual: Card uses standard Card component. Steps stacked vertically with space-y-3. Vertical connector line: border-l-2 between steps, color matches overall chain health (green if all complete, amber if partial, muted if incomplete). Card header: HCP name (font-medium) + specialty (text-xs text-muted-foreground). Card footer: "Edit" Button link (variant="ghost", text-xs) navigating to /admin/hcp-profiles/{id}.

Trigger: Management page loads HCP profile data. Behavior: Four stat cards display computed counts:

• Total HCPs: count of all active HCP profiles
• Agents Synced: count where agent_sync_status === "synced"
• Voice Live Enabled: count where voice_live_enabled === true
• Fully Configured: count where agent synced AND voice_live_enabled AND voice_name set AND avatar_character set

Visual: Each stat card: small Card with p-4. Number at text-2xl font-bold. Label below at text-xs text-muted-foreground. Icon in top-right corner (size-4 text-muted-foreground). Grid: grid grid-cols-2 md:grid-cols-4 gap-4.

Trigger: Admin clicks "Batch Re-sync" button in management page header. Behavior: Calls POST /api/v1/voice-live/batch-resync (new endpoint). Shows loading spinner on button during request. On success: toast.success() with count of synced/failed agents. On error: toast.error() with error message. Refetches HCP profile list after completion to update chain status cards. Visual: Button in page header area: variant="default" (primary), size="sm", with RefreshCw icon (size-4 mr-2). When loading: icon spins via animate-spin. Button text: "Batch Re-sync" / "批量同步". Constraint: Button disabled when no HCP profiles exist or when a batch sync is already in progress. Confirmation not required (re-sync is idempotent, non-destructive).

Trigger: HCP table renders rows with profiles that have voice_live_model set. Behavior: The existing Voice & Avatar column extends to show a third Badge when voice_live_model is present and differs from the default "gpt-4o", OR when voice_live_enabled is true. Badge shows short model label (e.g., "GPT-4o", "GPT-4.1"). When voice_live_enabled is false, the model badge is not shown. Visual: Third Badge after existing voice name and avatar badges. Badge variant="outline", text-xs. Displayed on a new line below the existing badges using flex flex-wrap items-center gap-1 to handle the wider content. When model is the default "gpt-4o" and other badges are shown, model badge uses standard outline. Non-default models show in outline with slightly bolder text.

Trigger: Admin clicks the re-sync icon button within a VoiceLiveChainCard that shows a failed or none agent status. Behavior: Calls existing POST /api/v1/hcp-profiles/{id}/retry-sync endpoint. Shows loading spinner on the icon button. On success: toast.success() with sync result. On failure: toast.error(). Card updates automatically via TanStack Query invalidation. Visual: Small icon button (variant="ghost", size="icon", className="size-7") with RefreshCw icon (size-3.5). Positioned in the Agent step row. Only visible when agent_sync_status is "failed" or "none". When loading: icon spins via animate-spin.

All copy externalized via react-i18next. English (en-US) and Chinese (zh-CN) values.

Focal point: Chain status grid. Admin scans all HCPs and their full Voice Live pipeline status.

+-----------------------------------------------------------+
| Voice Live Management                 [Batch Re-sync]      |  <- Page header (text-2xl + button)
| Manage Voice Live configurations...                        |  <- Description (text-sm text-muted-foreground)
+-----------------------------------------------------------+
|                                                            |
| +----------+ +----------+ +----------+ +----------+        |  <- Summary stats row
| | Total    | | Agents   | | Voice    | | Fully    |        |     (grid grid-cols-2 md:grid-cols-4 gap-4)
| | HCPs     | | Synced   | | Live On  | | Config'd |        |
| | 8        | | 6        | | 5        | | 4        |        |
| +----------+ +----------+ +----------+ +----------+        |
|                                                            |
| +------------------+ +------------------+ +-------------+  |  <- Chain cards grid
| | Dr. Zhang        | | Dr. Li           | | Dr. Wang   |  |     (grid grid-cols-1 md:grid-cols-2
| | Oncology         | | Hematology       | | Neurology  |  |      xl:grid-cols-3 gap-6)
| |                  | |                  | |            |  |
| | * HCP Profile    | | * HCP Profile    | | * HCP Pro  |  |
| | * Agent: Synced  | | * Agent: Failed  | | * Agent:   |  |
| | * VL: GPT-4o     | | * VL: Disabled   | |   None     |  |
| | * Speech: Ava    | | * Speech: --     | | * VL: --   |  |
| |   Avatar: Lori-c | |                  | | * --       |  |
| |                  | |                  | |            |  |
| | [Edit HCP]       | | [Retry] [Edit]   | | [Edit HCP] |  |
| +------------------+ +------------------+ +-------------+  |
+-----------------------------------------------------------+

+-------------------------------------------+
| CardHeader                                 |
|   Dr. Zhang - Oncology        [Re-sync?]  |
+-------------------------------------------+
| CardContent                                |
|                                            |
|  [*] HCP Profile ............ Dr. Zhang    |  <- Step 1 (always green)
|  |                                         |     Vertical line connector
|  [*] Agent Sync ............. Synced       |  <- Step 2 (color by status)
|  |                            asst_xxx     |     Tooltip for full agent_id
|  [*] Voice Live Config ...... GPT-4o       |  <- Step 3 (green/muted)
|  |                            Enabled      |
|  [*] Speech & Avatar ........ Ava / Lori   |  <- Step 4 (green/amber/muted)
|                               casual       |
|                                            |
+-------------------------------------------+
| CardFooter                                 |
|   [Edit HCP]                               |  <- Ghost button, navigates to editor
+-------------------------------------------+

Each step row layout:

[dot size-2] [step-label text-xs text-muted-foreground w-32] [value text-sm flex-1]

Vertical connector: ml-1 border-l-2 h-3 between step rows. Connector color: border-green-300 when both adjacent steps are green, border-border otherwise.

Position: Inside the Voice Settings Card in voice-avatar-tab.tsx, after the Voice Live Config toggle, before the Custom Voice toggle.

+-----------------------------------------------------------+
| Card: Voice Live Config Toggle                             |
|   Voice Live Config: [ON  --------]              [Switch]  |
|   Sync voice settings as agent metadata                    |
+-----------------------------------------------------------+
|                         space-y-4                          |
+-----------------------------------------------------------+
| Card: Voice Settings                                       |
|   Voice Live Model:                                        |  <- NEW FIELD
|   [Select dropdown  \/]                                    |
|     +---------------------------+                          |
|     | Pro                       |  <- SelectGroup label    |
|     | GPT Realtime              |                          |
|     | GPT-4o                    |  <- default selected     |
|     | GPT-4.1                   |                          |
|     | GPT-5                     |                          |
|     | GPT-5 Chat                |                          |
|     | Basic                     |  <- SelectGroup label    |
|     | GPT Realtime Mini         |                          |
|     | GPT-4o Mini               |                          |
|     | ...                       |                          |
|     | Lite                      |  <- SelectGroup label    |
|     | GPT-5 Nano                |                          |
|     | Phi4-MM Realtime          |                          |
|     | Phi4 Mini                 |                          |
|     +---------------------------+                          |
|                                                            |
|   Custom voice: [OFF ----]                     [Switch]    |
|   Voice Name: [Select dropdown  \/]                        |
+-----------------------------------------------------------+

| Voice & Avatar          |
|-------------------------|
| [Ava] [Lori-casual]    |   <- Existing badges
| [GPT-4o]               |   <- NEW model badge (when voice_live_enabled)

Badges use flex flex-wrap items-center gap-1 to wrap naturally when column is narrow.

No new component installations needed. SelectGroup and SelectLabel are already available in the existing @/components/ui/select.tsx (standard Radix Select primitives). All required UI primitives (Card, Badge, Select, Button, Tooltip, Skeleton, Form, Switch) are already present in frontend/src/components/ui/.

• Dimension 1 Copywriting: PASS
• Dimension 2 Visuals: PASS
• Dimension 3 Color: PASS
• Dimension 4 Typography: PASS
• Dimension 5 Spacing: PASS
• Dimension 6 Registry Safety: PASS

Approval: pending