feat(server): server-controlled default streaming TTS model

The streaming speech provider settings page was seeding its model
picker with a hardcoded `volcengine/seed-tts-2.0`, which contradicts
the just-landed "streaming model catalog comes from the server"
contract. Now the default also comes from configKV.

- `UNSPEECH_UPSTREAM.streaming` gains `defaultModel?: string`.
- `GET /api/v1/audio/models/streaming` response gains `default: string
  | null` reading that field.
- admin slice apply: `streaming.defaultModel` survives key/URL
  rotation alongside `streaming.models`.
- `providerOfficialSpeechStreaming.listModels` stashes the value in a
  module-scope ref, exposed via `getDefaultStreamingModel()`.
- Streaming speech settings page reads the helper instead of the
  hardcoded id; falls back to the first server-returned model if the
  operator hasn't curated a default. `handleGenerateSpeech` no longer
  has a `seed-tts-2.0` backstop — a missing backend prefix in the
  selected model id now throws instead of silently defaulting.

apps/server/src/routes/admin/config/router/index.ts

@@ -91,6 +91,12 @@ const UnspeechSliceSchema = object({
     ),
     plaintextKey: pipe(string(), nonEmpty('streaming.plaintextKey is required'), maxLength(MAX_KEY_LENGTH)),
     keyEntryId: optional(pipe(string(), nonEmpty(), maxLength(200), NO_PIPE)),
+    models: optional(array(object({
+      id: pipe(string(), nonEmpty('streaming.models[].id is required'), maxLength(200)),
+      name: optional(pipe(string(), nonEmpty(), maxLength(200))),
+      description: optional(pipe(string(), nonEmpty(), maxLength(500))),
+    }))),
+    defaultModel: optional(pipe(string(), nonEmpty('streaming.defaultModel must not be empty'), maxLength(200))),
   })),
 })
 

apps/server/src/routes/openai/v1/index.ts

@@ -290,7 +290,6 @@ export function createV1Routes(
               logger.withError(err).warn('Failed to close stream writer')
             }
 
-            // Extract usage from final SSE data lines
             let usage: UsageInfo = {}
             try {
               const lines = tailBuffer.split('\n').filter(l => l.startsWith('data: ') && !l.includes('[DONE]'))
@@ -702,14 +701,9 @@ export function createV1Routes(
   }
 
   async function handleListTTSModels(_c: Context<HonoEnv>) {
-    // Surface the concrete TTS models the operator has configured plus the
-    // `auto` alias. Clients need real model ids to pass `?model=<id>` to
-    // `/audio/voices`, otherwise the voice catalog endpoint can never resolve
-    // anything beyond the DEFAULT_TTS_MODEL catalog — which is the bug that
-    // hid the Azure voices from the UI.
-    //
-    // `auto` is kept on top as an explicit "use the operator default" knob
-    // for clients that don't care which concrete model handles them.
+    // Surface the concrete TTS models the operator has configured. The UI
+    // should select an explicit model id so voice catalog requests stay
+    // model-scoped instead of hiding behind DEFAULT_TTS_MODEL.
     const config = await configKV.getOrThrow('LLM_ROUTER_CONFIG')
     // `LLM_ROUTER_CONFIG` is `optional()` at the schema, so its inferred type
     // tolerates `undefined`. `getOrThrow` already throws on missing entries,
@@ -717,29 +711,32 @@ export function createV1Routes(
     // a TS narrowing aid.
     const modelIds = Object.keys(config?.tts?.models ?? {}).sort()
     return Response.json({
-      models: [
-        { id: 'auto', name: 'Auto' },
-        ...modelIds.map(id => ({ id, name: id })),
-      ],
+      models: modelIds.map(id => ({ id, name: id })),
     })
   }
 
   async function handleListStreamingTTSModels(_c: Context<HonoEnv>) {
     const unspeech = await configKV.getOptional('UNSPEECH_UPSTREAM')
     const models = unspeech?.streaming?.models ?? []
+    // `available` is the operator-controlled visibility switch the client gates
+    // the streaming provider on. It tracks whether `UNSPEECH_UPSTREAM.streaming`
+    // is configured at all — not whether `models[]` happens to be empty — so an
+    // operator who has wired the upstream but not yet curated models still
+    // surfaces the provider rather than silently hiding it.
     return Response.json({
+      available: !!unspeech?.streaming?.baseURL,
       models: models.map(m => ({
         id: m.id,
         name: m.name ?? m.id,
         description: m.description,
       })),
+      default: unspeech?.streaming?.defaultModel ?? null,
     })
   }
 
   const chatGuard = configGuard(configKV, ['FLUX_PER_REQUEST'], 'Service is not available yet')
   const ttsGuard = configGuard(configKV, ['FLUX_PER_1K_CHARS_TTS'], 'TTS service is not available yet')
 
-  // 60 requests per minute per user for LLM completions
   const completionsRateLimit = rateLimiter({ max: 60, windowSec: 60, metrics: rateLimitMetrics, routeLabel: 'openai.completions' })
 
   // OpenAI-compatible surface (mounted at /api/v1/openai). Only routes that

apps/server/src/routes/openai/v1/route.test.ts

@@ -11,8 +11,6 @@ import { afterAll, describe, expect, it, vi } from 'vitest'
 import { createV1Routes } from '.'
 import { ApiError } from '../../../utils/error'
 
-// --- Mock helpers ---
-
 function createMockFluxService(flux = 100): FluxService {
   return {
     getFlux: vi.fn(async () => ({ userId: 'user-1', flux })),
@@ -165,8 +163,6 @@ function createTestApp(
 
 const testUser = { id: 'user-1', name: 'Test User', email: 'test@example.com' }
 
-// --- Tests ---
-
 describe('v1CompletionsRoutes', () => {
   const originalFetch = globalThis.fetch
 
@@ -319,12 +315,10 @@ describe('v1CompletionsRoutes', () => {
       const data = await res.json() as { id: string }
       expect(data.id).toBe('chatcmpl-1')
 
-      // Verify flux was debited via billingService
       expect(billingService.consumeFluxForLLM).toHaveBeenCalledWith(
         expect.objectContaining({ userId: 'user-1', amount: 1 }),
       )
 
-      // Verify upstream was called with correct URL and resolved model
       expect(globalThis.fetch).toHaveBeenCalledWith(
         'http://mock-gateway/chat/completions',
         expect.objectContaining({
@@ -406,13 +400,11 @@ describe('v1CompletionsRoutes', () => {
       )
 
       expect(res.status).toBe(500)
-      // Post-billing: no charge on failed requests
       expect(billingService.consumeFluxForLLM).not.toHaveBeenCalled()
     })
 
     it('should return 503 when config keys are missing', async () => {
       const configKV = createMockConfigKV()
-      // Override getOptional to return null for required keys
       configKV.getOptional = vi.fn(async () => null)
 
       const app = createTestApp(createMockFluxService(), configKV)
@@ -748,7 +740,7 @@ describe('v1CompletionsRoutes', () => {
   })
 
   describe('gET /api/v1/audio/models', () => {
-    it('exposes auto alias plus every configured tts model id', async () => {
+    it('exposes every configured tts model id', async () => {
       const app = createTestApp(
         createMockFluxService(),
         createMockConfigKV({
@@ -771,14 +763,13 @@ describe('v1CompletionsRoutes', () => {
 
       expect(res.status).toBe(200)
       const data = await res.json() as { models: { id: string, name: string }[] }
-      expect(data.models[0]).toEqual({ id: 'auto', name: 'Auto' })
-      expect(data.models.slice(1).map(m => m.id)).toEqual([
+      expect(data.models.map(m => m.id)).toEqual([
         'alibaba/cosyvoice-v2',
         'microsoft/v1',
       ])
     })
 
-    it('returns only the auto alias when no tts models are configured', async () => {
+    it('returns an empty list when no tts models are configured', async () => {
       const app = createTestApp(
         createMockFluxService(),
         createMockConfigKV({
@@ -793,7 +784,7 @@ describe('v1CompletionsRoutes', () => {
 
       expect(res.status).toBe(200)
       const data = await res.json() as { models: { id: string, name: string }[] }
-      expect(data.models).toEqual([{ id: 'auto', name: 'Auto' }])
+      expect(data.models).toEqual([])
     })
 
     it('should return 401 when unauthenticated', async () => {
@@ -805,7 +796,7 @@ describe('v1CompletionsRoutes', () => {
   })
 
   describe('gET /api/v1/audio/models/streaming', () => {
-    it('returns the operator-configured streaming model catalog', async () => {
+    it('returns the operator-configured streaming model catalog + default', async () => {
       const app = createTestApp(
         createMockFluxService(),
         createMockConfigKV({
@@ -818,6 +809,7 @@ describe('v1CompletionsRoutes', () => {
                 { id: 'volcengine/seed-tts-2.0', name: 'Volcengine Seed-TTS 2.0', description: 'TTS 2.0' },
                 { id: 'volcengine/seed-tts-1.0' },
               ],
+              defaultModel: 'volcengine/seed-tts-2.0',
             },
           },
         }),
@@ -829,11 +821,37 @@ describe('v1CompletionsRoutes', () => {
       )
 
       expect(res.status).toBe(200)
-      const data = await res.json() as { models: { id: string, name: string, description?: string }[] }
+      const data = await res.json() as { available: boolean, models: { id: string, name: string, description?: string }[], default: string | null }
+      expect(data.available).toBe(true)
       expect(data.models).toEqual([
         { id: 'volcengine/seed-tts-2.0', name: 'Volcengine Seed-TTS 2.0', description: 'TTS 2.0' },
         { id: 'volcengine/seed-tts-1.0', name: 'volcengine/seed-tts-1.0' },
       ])
+      expect(data.default).toBe('volcengine/seed-tts-2.0')
+    })
+
+    it('returns default: null when operator has not set a streaming default', async () => {
+      const app = createTestApp(
+        createMockFluxService(),
+        createMockConfigKV({
+          UNSPEECH_UPSTREAM: {
+            restBaseURL: 'http://unspeech.local:5933',
+            streaming: {
+              baseURL: 'wss://unspeech.local',
+              keys: [{ id: 'k1', ciphertext: 'enc' }],
+              models: [{ id: 'volcengine/seed-tts-2.0', name: 'Vol' }],
+            },
+          },
+        }),
+      )
+
+      const res = await app.fetch(
+        new Request('http://localhost/api/v1/audio/models/streaming', { method: 'GET' }),
+        { user: testUser } as any,
+      )
+
+      const data = await res.json() as { default: string | null }
+      expect(data.default).toBeNull()
     })
 
     it('returns an empty list when UNSPEECH_UPSTREAM is unset', async () => {
@@ -845,11 +863,12 @@ describe('v1CompletionsRoutes', () => {
       )
 
       expect(res.status).toBe(200)
-      const data = await res.json() as { models: unknown[] }
+      const data = await res.json() as { available: boolean, models: unknown[] }
+      expect(data.available).toBe(false)
       expect(data.models).toEqual([])
     })
 
-    it('returns an empty list when streaming subtree has no models', async () => {
+    it('reports available: true with empty models when streaming subtree has no models', async () => {
       const app = createTestApp(
         createMockFluxService(),
         createMockConfigKV({
@@ -869,7 +888,8 @@ describe('v1CompletionsRoutes', () => {
       )
 
       expect(res.status).toBe(200)
-      const data = await res.json() as { models: unknown[] }
+      const data = await res.json() as { available: boolean, models: unknown[] }
+      expect(data.available).toBe(true)
       expect(data.models).toEqual([])
     })
 

apps/server/src/services/adapters/config-kv.test.ts

@@ -21,8 +21,6 @@ describe('configKVService', () => {
     service = createConfigKVService(redis as any)
   })
 
-  // --- get ---
-
   it('get should throw 503 when key is not set', async () => {
     await expect(service.getOrThrow('FLUX_PER_1K_CHARS_TTS'))
       .rejects
@@ -43,8 +41,6 @@ describe('configKVService', () => {
     expect(redis.get).toHaveBeenCalledWith(configRedisKey('FLUX_PER_REQUEST'))
   })
 
-  // --- getOptional ---
-
   it('getOptional should return schema default when key has one', async () => {
     const value = await service.getOptional('FLUX_PER_REQUEST')
     expect(value).toBe(5)
@@ -62,7 +58,35 @@ describe('configKVService', () => {
     expect(value).toBe(200)
   })
 
-  // --- set ---
+  it('getOptional should throw CONFIG_INVALID when Redis contains malformed JSON', async () => {
+    // ROOT CAUSE:
+    //
+    // If an operator edits config:LLM_ROUTER_CONFIG directly with invalid JSON,
+    // JSON.parse used to throw SyntaxError through the request handler and log
+    // it as an unhandled 500.
+    //
+    // We fixed this by translating stored config parse/validation failures into
+    // a stable API error at the configKV boundary.
+    redis._store.set(configRedisKey('LLM_ROUTER_CONFIG'), '{"llm":{}')
+
+    await expect(service.getOptional('LLM_ROUTER_CONFIG'))
+      .rejects
+      .toMatchObject({
+        statusCode: 503,
+        errorCode: 'CONFIG_INVALID',
+      })
+  })
+
+  it('getOptional should throw CONFIG_INVALID when Redis contains schema-invalid JSON', async () => {
+    redis._store.set(configRedisKey('FLUX_PER_REQUEST'), JSON.stringify('5'))
+
+    await expect(service.getOptional('FLUX_PER_REQUEST'))
+      .rejects
+      .toMatchObject({
+        statusCode: 503,
+        errorCode: 'CONFIG_INVALID',
+      })
+  })
 
   it('set should write value to Redis with prefix', async () => {
     await service.set('FLUX_PER_REQUEST', 10)

apps/server/src/services/adapters/config-kv.ts

@@ -1,6 +1,7 @@
 import type Redis from 'ioredis'
 import type { InferOutput } from 'valibot'
 
+import { errorMessageFrom } from '@moeru/std'
 import { any, array, boolean, check, nonEmpty, number, object, optional, parse, picklist, pipe, record, regex, string } from 'valibot'
 
 import { createServiceUnavailableError } from '../../utils/error'
@@ -71,6 +72,7 @@ export const streamingTtsUpstreamSchema = object({
     })),
     [],
   ),
+  defaultModel: optional(string()),
 })
 
 export const unspeechUpstreamSchema = object({
@@ -158,7 +160,19 @@ type ConfigDefinitions = {
 type ConfigKey = keyof ConfigDefinitions
 
 function parseValue<K extends ConfigKey>(key: K, raw: string): ConfigDefinitions[K] {
-  return parse(ConfigEntrySchemas[key], JSON.parse(raw)) as ConfigDefinitions[K]
+  try {
+    return parse(ConfigEntrySchemas[key], JSON.parse(raw)) as ConfigDefinitions[K]
+  }
+  catch (error) {
+    throw createServiceUnavailableError(
+      'Service configuration is invalid',
+      'CONFIG_INVALID',
+      {
+        key,
+        message: errorMessageFrom(error) ?? 'Unknown config parse error',
+      },
+    )
+  }
 }
 
 function serializeValue<K extends ConfigKey>(key: K, value: ConfigDefinitions[K]): string {

apps/server/src/services/adapters/tts/azure.ts

@@ -228,7 +228,7 @@ export const azureAdapter: TtsAdapter = {
     if (!ctx.keyPlaintext)
       throw createServiceUnavailableError('azure tts key not configured', 'AZURE_TTS_NOT_CONFIGURED')
 
-    const url = `${ctx.unspeechBaseURL.replace(/\/+$/, '')}/api/voices?backend=microsoft&region=${encodeURIComponent(ctx.region)}`
+    const url = `${ctx.unspeechBaseURL.replace(/\/+$/, '')}/api/voices?provider=microsoft&region=${encodeURIComponent(ctx.region)}`
 
     let response: Response
     try {

apps/server/src/services/adapters/tts/dashscope-cosyvoice.ts

@@ -109,7 +109,7 @@ export const dashscopeCosyvoiceAdapter: TtsAdapter = {
     // (unspeech/pkg/backend/alibaba/voices.go `//go:embed voices.json`),
     // so this call is in-memory on unspeech's side and only crosses a TCP
     // hop. No upstream credential is required.
-    const url = `${ctx.unspeechBaseURL.replace(/\/+$/, '')}/api/voices?backend=alibaba`
+    const url = `${ctx.unspeechBaseURL.replace(/\/+$/, '')}/api/voices?provider=alibaba`
 
     let response: Response
     try {

apps/server/src/services/adapters/tts/index.test.ts

@@ -61,7 +61,7 @@ describe('getAdapter', () => {
 })
 
 describe('dashscopeCosyvoiceAdapter.getVoiceCatalog', () => {
-  it('calls unspeech with backend=alibaba (no Bearer)', async () => {
+  it('calls unspeech with provider=alibaba (no Bearer)', async () => {
     const adapter = getAdapter('dashscope-cosyvoice')
     const fetchImpl = vi.fn(async () => new Response(JSON.stringify({
       voices: [{ id: 'longxiaochun_v2', name: 'Longxiaochun v2' }],
@@ -75,7 +75,7 @@ describe('dashscopeCosyvoiceAdapter.getVoiceCatalog', () => {
 
     expect(voices).toEqual([{ id: 'longxiaochun_v2', name: 'Longxiaochun v2' }])
     const [calledUrl, init] = (fetchImpl as unknown as { mock: { calls: [string, RequestInit][] } }).mock.calls[0]
-    expect(calledUrl).toBe('http://unspeech.local/api/voices?backend=alibaba')
+    expect(calledUrl).toBe('http://unspeech.local/api/voices?provider=alibaba')
     const headers = (init.headers ?? {}) as Record<string, string>
     expect(headers.Authorization).toBeUndefined()
   })
@@ -92,7 +92,7 @@ describe('dashscopeCosyvoiceAdapter.getVoiceCatalog', () => {
 })
 
 describe('volcengineAdapter.getVoiceCatalog', () => {
-  it('calls unspeech with backend=volcengine and forwards adapterParams.model as ?model=', async () => {
+  it('calls unspeech with provider=volcengine and forwards adapterParams.model as ?model=', async () => {
     const adapter = getAdapter('volcengine')
     const fetchImpl = vi.fn(async () => new Response(JSON.stringify({
       voices: [{ id: 'zh_female_x', name: 'X' }],
@@ -106,7 +106,7 @@ describe('volcengineAdapter.getVoiceCatalog', () => {
 
     expect(voices).toEqual([{ id: 'zh_female_x', name: 'X' }])
     const [calledUrl] = (fetchImpl as unknown as { mock: { calls: [string, RequestInit][] } }).mock.calls[0]
-    expect(calledUrl).toBe('http://unspeech.local/api/voices?backend=volcengine&model=seed-tts-2.0')
+    expect(calledUrl).toBe('http://unspeech.local/api/voices?provider=volcengine&model=seed-tts-2.0')
   })
 
   it('omits ?model= when adapterParams.model is not set', async () => {
@@ -118,7 +118,7 @@ describe('volcengineAdapter.getVoiceCatalog', () => {
       fetchImpl,
     })
     const [calledUrl] = (fetchImpl as unknown as { mock: { calls: [string, RequestInit][] } }).mock.calls[0]
-    expect(calledUrl).toBe('http://unspeech.local/api/voices?backend=volcengine')
+    expect(calledUrl).toBe('http://unspeech.local/api/voices?provider=volcengine')
   })
 })
 
@@ -140,7 +140,7 @@ describe('azureAdapter.getVoiceCatalog', () => {
     expect(voices).toEqual([{ id: 'en-US-AvaMultilingualNeural', name: 'Ava' }])
     expect(fetchImpl).toHaveBeenCalledTimes(1)
     const [calledUrl, init] = (fetchImpl as unknown as { mock: { calls: [string, RequestInit][] } }).mock.calls[0]
-    expect(calledUrl).toBe('http://unspeech.local:5933/api/voices?backend=microsoft&region=eastasia')
+    expect(calledUrl).toBe('http://unspeech.local:5933/api/voices?provider=microsoft&region=eastasia')
     const headers = init.headers as Record<string, string>
     expect(headers.Authorization).toBe('Bearer subscription-key-XYZ')
   })

apps/server/src/services/adapters/tts/volcengine.ts

@@ -120,7 +120,7 @@ export const volcengineAdapter: TtsAdapter = {
     // further by `compatible_models` — adapterParams.model is the operator-
     // configured resource id (e.g. `seed-tts-2.0`).
     const url = new URL(`${ctx.unspeechBaseURL.replace(/\/+$/, '')}/api/voices`)
-    url.searchParams.set('backend', 'volcengine')
+    url.searchParams.set('provider', 'volcengine')
     const apiResourceId = typeof ctx.adapterParams?.model === 'string'
       ? ctx.adapterParams.model
       : undefined

apps/server/src/services/domain/admin/router-config/index.ts

@@ -96,6 +96,10 @@ export interface UnspeechSliceInput {
     plaintextKey: string
     /** @default 'volcengine-prod-1' */
     keyEntryId?: string
+    /** Operator-curated streaming models exposed to the frontend picker. */
+    models?: Array<{ id: string, name?: string, description?: string }>
+    /** Server-curated default streaming model id. */
+    defaultModel?: string
   }
 }
 
@@ -262,7 +266,8 @@ export function buildUnspeechSlice(input: UnspeechSliceInput, envelope: Envelope
         baseURL: input.streaming.upstreamURL,
         keys: [{ id: keyEntryId, ciphertext }],
         adapterParams: {},
-        models: [],
+        models: input.streaming.models ?? [],
+        defaultModel: input.streaming.defaultModel,
       },
     },
   }
@@ -451,18 +456,22 @@ export function createAdminRouterConfigService(deps: AdminRouterConfigDeps) {
       nextRouterConfig = buildNextRouterConfig(input.mode, existing, llmTtsSlices)
     }
 
-    // Step 3: build the next UNSPEECH_UPSTREAM. Streaming `models` carry the
-    // operator-curated catalog and must survive key/URL rotation, so we read
-    // existing and graft them onto the new value when the slice's streaming
+    // Step 3: build the next UNSPEECH_UPSTREAM. Streaming `models` +
+    // `defaultModel` are operator-curated and must survive key/URL rotation,
+    // so we graft them from the existing entry when the slice's streaming
     // subtree is set (otherwise there's nothing to merge into).
     let nextUnspeech: UnspeechUpstream | undefined
     if (unspeechSlice) {
       const existing = await deps.configKV.getOptional('UNSPEECH_UPSTREAM')
       const newValue = unspeechSlice.value
-      if (newValue.streaming && existing?.streaming?.models?.length) {
+      if (newValue.streaming && existing?.streaming) {
         nextUnspeech = {
           ...newValue,
-          streaming: { ...newValue.streaming, models: existing.streaming.models },
+          streaming: {
+            ...newValue.streaming,
+            models: existing.streaming.models?.length ? existing.streaming.models : newValue.streaming.models,
+            defaultModel: existing.streaming.defaultModel ?? newValue.streaming.defaultModel,
+          },
         }
       }
       else {

apps/server/src/services/domain/admin/router-config/tests/admin-router-config.test.ts

@@ -356,11 +356,23 @@ describe('createAdminRouterConfigService', () => {
         streaming: {
           upstreamURL: 'wss://unspeech.example/v1/audio/speech/stream',
           plaintextKey: 'volc',
+          models: [
+            { id: 'volcengine/seed-tts-2.0', name: 'Seed-TTS 2.0', description: 'Low-latency streaming TTS' },
+          ],
+          defaultModel: 'volcengine/seed-tts-2.0',
         },
       }],
     })
 
     expect(kv.store.has('UNSPEECH_UPSTREAM')).toBe(true)
+    expect(kv.store.get('UNSPEECH_UPSTREAM')).toMatchObject({
+      streaming: {
+        models: [
+          { id: 'volcengine/seed-tts-2.0', name: 'Seed-TTS 2.0', description: 'Low-latency streaming TTS' },
+        ],
+        defaultModel: 'volcengine/seed-tts-2.0',
+      },
+    })
     expect(kv.store.has('LLM_ROUTER_CONFIG')).toBe(false)
     expect(result.invalidatedKeys).toEqual(['UNSPEECH_UPSTREAM'])
     expect(captured.map(p => JSON.parse(p.payload).key)).toEqual(['UNSPEECH_UPSTREAM'])

apps/server/src/services/domain/llm-router/router.ts

@@ -1,6 +1,7 @@
 import type { Buffer } from 'node:buffer'
 
 import type Redis from 'ioredis'
+import type { Voice } from 'unspeech'
 
 import type { GatewayMetrics } from '../../../otel'
 import type { EnvelopeCrypto } from '../../../utils/envelope-crypto'
@@ -176,6 +177,7 @@ export function createLlmRouterService(options: CreateLlmRouterServiceOptions) {
   const logger = useLogger('llm-router').useGlobalConfig()
   const fetchImpl = options.fetchImpl ?? globalThis.fetch
   const configLoader = createConfigLoader({ configKV: options.configKV, ttlMs: options.configCacheTtlMs })
+  const ttsVoiceCatalogLoads = new Map<string, Promise<Voice[]>>()
 
   /**
    * Run one upstream's key list in order, returning either:
@@ -652,44 +654,55 @@ export function createLlmRouterService(options: CreateLlmRouterServiceOptions) {
       }
     }
 
-    const unspeechBaseURL = (await options.configKV.getOrThrow('UNSPEECH_UPSTREAM')).restBaseURL
+    const existingLoad = ttsVoiceCatalogLoads.get(cacheKey)
+    if (existingLoad != null)
+      return existingLoad
 
-    // Live providers (Azure) need the decrypted Azure subscription key + region;
-    // static-catalog providers (alibaba, volcengine) ignore both. The router
-    // decrypts unconditionally so the adapter doesn't have to know which
-    // category it's in — adapters that don't need creds just won't read them.
-    const region = typeof upstream.adapterParams?.region === 'string'
-      ? upstream.adapterParams.region
-      : undefined
+    const load = (async () => {
+      const unspeechBaseURL = (await options.configKV.getOrThrow('UNSPEECH_UPSTREAM')).restBaseURL
 
-    const keyEntry = upstream.keys[0]
-    const plaintext = slice.model.provider === 'azure'
-      ? options.envelopeCrypto.decryptKey(keyEntry.ciphertext, { modelName, keyEntryId: keyEntry.id })
-      : undefined
+      // Live providers (Azure) need the decrypted Azure subscription key + region;
+      // static-catalog providers (alibaba, volcengine) ignore both. The router
+      // decrypts unconditionally so the adapter doesn't have to know which
+      // category it's in — adapters that don't need creds just won't read them.
+      const region = typeof upstream.adapterParams?.region === 'string'
+        ? upstream.adapterParams.region
+        : undefined
 
-    try {
-      const voices = await adapter.getVoiceCatalog({
-        keyPlaintext: plaintext,
-        region,
-        adapterParams: upstream.adapterParams ?? {},
-        unspeechBaseURL,
-        fetchImpl,
-      })
+      const keyEntry = upstream.keys[0]
+      const plaintext = slice.model.provider === 'azure'
+        ? options.envelopeCrypto.decryptKey(keyEntry.ciphertext, { modelName, keyEntryId: keyEntry.id })
+        : undefined
 
-      // Cache only on success — failure responses must NOT be persisted or
-      // the next admin reconfigure would have to wait out the TTL even after
-      // fixing credentials.
-      const ttl = options.ttsVoiceCacheTtlSeconds ?? ttsVoicesCacheTtl(slice.model.provider)
-      await options.redis.set(cacheKey, JSON.stringify(voices), 'EX', ttl)
-        .catch((err) => {
-          logger.withError(err).withFields({ cacheKey }).warn('failed to write tts voices cache')
+      try {
+        const voices = await adapter.getVoiceCatalog({
+          keyPlaintext: plaintext,
+          region,
+          adapterParams: upstream.adapterParams ?? {},
+          unspeechBaseURL,
+          fetchImpl,
         })
 
-      return voices
-    }
-    finally {
-      plaintext?.fill(0)
-    }
+        // Cache only on success — failure responses must NOT be persisted or
+        // the next admin reconfigure would have to wait out the TTL even after
+        // fixing credentials.
+        const ttl = options.ttsVoiceCacheTtlSeconds ?? ttsVoicesCacheTtl(slice.model.provider)
+        await options.redis.set(cacheKey, JSON.stringify(voices), 'EX', ttl)
+          .catch((err) => {
+            logger.withError(err).withFields({ cacheKey }).warn('failed to write tts voices cache')
+          })
+
+        return voices
+      }
+      finally {
+        plaintext?.fill(0)
+      }
+    })().finally(() => {
+      ttsVoiceCatalogLoads.delete(cacheKey)
+    })
+
+    ttsVoiceCatalogLoads.set(cacheKey, load)
+    return load
   }
 
   /**

apps/server/src/services/domain/llm-router/tests/router.test.ts

@@ -754,5 +754,48 @@ describe('createLlmRouterService', () => {
       // adapter's `Error & { status }` was read as undefined.
       expect(fallbackCalls[0][1]).toMatchObject({ reason: '401' })
     })
+
+    it('listTtsVoices deduplicates concurrent cold-cache upstream fetches per model', async () => {
+      // ROOT CAUSE:
+      //
+      // Azure voice catalogs are cached after a successful fetch, but concurrent
+      // cold-cache requests used to miss Redis together and each hit unspeech's
+      // microsoft voices endpoint. That can amplify one settings-page open into
+      // several Azure voices/list calls and trigger upstream 429.
+      //
+      // We fixed this by sharing the in-flight catalog load for the same
+      // provider/model cache key. Failures are still returned to every caller and
+      // are not cached.
+      const { config, crypto } = makeTtsConfig({
+        upstreams: [{ baseURL: 'https://az.example', keyIds: ['kA1'], adapterParams: { region: 'eastasia' } }],
+      })
+
+      let resolveFetch!: () => void
+      const fetchImpl = vi.fn(() => new Promise<Response>((resolve) => {
+        resolveFetch = () => resolve(happyResponse({
+          voices: [{ id: 'en-US-AvaMultilingualNeural', name: 'Ava' }],
+        }))
+      }))
+
+      const router = createLlmRouterService({
+        configKV: makeConfigKV(config),
+        envelopeCrypto: crypto,
+        gatewayMetrics: null,
+        fetchImpl,
+        redis: makeRedisStub(),
+      })
+
+      const first = router.listTtsVoices('tts-test')
+      const second = router.listTtsVoices('tts-test')
+
+      await vi.waitFor(() => {
+        expect(fetchImpl).toHaveBeenCalledTimes(1)
+      })
+      resolveFetch()
+
+      const [firstVoices, secondVoices] = await Promise.all([first, second])
+      expect(firstVoices.map(voice => voice.id)).toEqual(['en-US-AvaMultilingualNeural'])
+      expect(secondVoices.map(voice => voice.id)).toEqual(['en-US-AvaMultilingualNeural'])
+    })
   })
 })

packages/stage-pages/src/pages/settings/modules/speech.vue

@@ -87,13 +87,13 @@ function syncOpenAICompatibleSettings() {
 
 onMounted(async () => {
   await providersStore.loadModelsForConfiguredProviders()
-  await speechStore.loadVoicesForProvider(activeSpeechProvider.value)
+  speechStore.ensureActiveSpeechModel()
+  await speechStore.loadVoicesForProvider(activeSpeechProvider.value, activeSpeechModel.value || undefined)
   syncOpenAICompatibleSettings()
 })
 
 watch(activeSpeechProvider, async (newProvider, oldProvider) => {
   await providersStore.loadModelsForConfiguredProviders()
-  await speechStore.loadVoicesForProvider(newProvider)
 
   // Reset model and voice when switching providers (but not on initial load)
   if (oldProvider !== undefined && oldProvider !== newProvider) {
@@ -102,12 +102,18 @@ watch(activeSpeechProvider, async (newProvider, oldProvider) => {
     activeSpeechVoice.value = undefined
   }
 
+  // Re-seed the streaming default model after the reset above so its voices
+  // load model-scoped (the server only returns recommended voices for an
+  // explicit ?model=). No-op for other providers / when a model is selected.
+  speechStore.ensureActiveSpeechModel()
+  await speechStore.loadVoicesForProvider(newProvider, activeSpeechModel.value || undefined)
+
   syncOpenAICompatibleSettings()
 })
 
 watch(activeSpeechModel, async () => {
   if (activeSpeechProvider.value) {
-    await speechStore.loadVoicesForProvider(activeSpeechProvider.value)
+    await speechStore.loadVoicesForProvider(activeSpeechProvider.value, activeSpeechModel.value || undefined)
   }
 })
 

packages/stage-pages/src/pages/settings/providers/speech/official-provider-speech-streaming.vue

@@ -4,7 +4,7 @@ import {
   ProviderSettingsLayout,
   SpeechPlayground,
 } from '@proj-airi/stage-ui/components'
-import { streamingSynthesize } from '@proj-airi/stage-ui/libs'
+import { getDefaultStreamingModel, streamingSynthesize } from '@proj-airi/stage-ui/libs'
 import { useAuthStore } from '@proj-airi/stage-ui/stores/auth'
 import { useSpeechStore } from '@proj-airi/stage-ui/stores/modules/speech'
 import { useProvidersStore } from '@proj-airi/stage-ui/stores/providers'
@@ -23,18 +23,19 @@ const { isAuthenticated, credits, needsLogin } = storeToRefs(authStore)
 
 const providerId = 'official-provider-speech-streaming'
 const providerMetadata = computed(() => providersStore.getProviderMetadata(providerId))
-const defaultModel = 'volcengine/seed-tts-2.0'
 
 const providerConfig = computed(() => providersStore.getProviderConfig(providerId))
 
-// Model picker. Pulled from the provider's `extraMethods.listModels` —
-// today that's two hard-coded Volcengine variants, but the picker stays
-// generic so adding ICL / other backends later doesn't need UI changes.
+// Model picker. The catalog and the default model id both come from the
+// server's `/api/v1/audio/models/streaming` response (operator-controlled
+// via `UNSPEECH_UPSTREAM.streaming`); no client-side hardcoded defaults so
+// adding ICL / other backends doesn't need a UI release.
 const providerModels = computed(() => providersStore.getModelsForProvider(providerId))
 const modelsLoading = computed(() => providersStore.isLoadingModels[providerId] || false)
+const serverDefaultModel = ref<string | null>(null)
 const model = computed({
   get(): string {
-    return (providerConfig.value?.model as string | undefined) ?? defaultModel
+    return (providerConfig.value?.model as string | undefined) ?? serverDefaultModel.value ?? ''
   },
   set(val: string) {
     providerConfig.value.model = val
@@ -57,8 +58,13 @@ async function loadVoices() {
 
 onMounted(async () => {
   await providersStore.fetchModelsForProvider(providerId)
-  if (!providerConfig.value.model)
-    providerConfig.value.model = defaultModel
+  // `getDefaultStreamingModel()` is populated by the provider's listModels()
+  // (just ran via fetchModelsForProvider). If the operator hasn't curated a
+  // default server-side, fall back to the first model the server returned
+  // so the picker always has something selected.
+  serverDefaultModel.value = getDefaultStreamingModel() ?? providerModels.value[0]?.id ?? null
+  if (!providerConfig.value.model && serverDefaultModel.value)
+    providerConfig.value.model = serverDefaultModel.value
   await loadVoices()
 })
 
@@ -76,11 +82,18 @@ watch(model, async () => {
 // per-preview because there's no LLM token stream here — we just send
 // one `text` frame containing the static preview prompt.
 async function handleGenerateSpeech(input: string, voiceId: string, _useSSML: boolean): Promise<ArrayBuffer> {
-  const requestedModel = model.value || defaultModel
+  const requestedModel = model.value
+  if (!requestedModel)
+    throw new Error('No streaming TTS model selected and server returned no default')
   // `model` looks like `volcengine/seed-tts-2.0`. The trailing path is
-  // forwarded as Volcengine's `api_resource_id` so the upstream knows
-  // which model variant to use; matches the wiring in `Stage.vue`.
-  const apiResourceId = requestedModel.includes('/') ? requestedModel.split('/', 2)[1] : 'seed-tts-2.0'
+  // forwarded as Volcengine's `api_resource_id` so the upstream knows which
+  // model variant to use; matches the wiring in `Stage.vue`. We require the
+  // `<backend>/<resource>` shape and refuse anything else — silently picking
+  // a fallback resource id hides config drift.
+  const slashIndex = requestedModel.indexOf('/')
+  if (slashIndex < 0)
+    throw new Error(`Streaming model id missing backend prefix: ${requestedModel}`)
+  const apiResourceId = requestedModel.slice(slashIndex + 1)
   const result = await streamingSynthesize({
     model: requestedModel,
     voice: voiceId,

packages/stage-ui/src/components/scenes/Stage.vue

@@ -32,7 +32,7 @@ import { useIOTraceBridge } from '../../composables/use-io-trace-bridge'
 import { initIOTracer } from '../../composables/use-io-tracer'
 import { useSpeechPipelineAnalytics } from '../../composables/use-speech-pipeline-analytics'
 import { Emotion, EMOTION_EmotionMotionName_value, EMOTION_VRMExpressionName_value, EmotionThinkMotionName } from '../../constants/emotions'
-import { getDefinedProvider } from '../../libs/providers/providers'
+import { getDefaultStreamingModel, getDefinedProvider } from '../../libs/providers/providers'
 import { createStageTtsSession } from '../../libs/speech/tts-session'
 import { useAudioContext, useSpeakingStore } from '../../stores/audio'
 import { useBackgroundStore } from '../../stores/background'
@@ -575,8 +575,17 @@ function buildStreamingSnapshot(): StreamingSessionSnapshot | null {
   const voiceId = activeSpeechVoice.value?.id
   if (!voiceId)
     return null
-  const sessionModel = (activeSpeechModel.value as string | undefined) || 'volcengine/seed-tts-2.0'
-  const apiResourceId = sessionModel.includes('/') ? sessionModel.split('/', 2)[1] : 'seed-tts-2.0'
+  // Resolve the concrete streaming model id. The active speech model is only
+  // valid here when it carries the `<backend>/<api_resource_id>` shape the ws
+  // upstream expects — the HTTP TTS `auto` alias (and an empty selection after
+  // a provider switch) must NOT reach the bridge, so fall back to the
+  // server-curated default instead of a hardcoded id. Returns null (segmenter
+  // fallback) when neither resolves, rather than guessing a resource id.
+  const activeModel = activeSpeechModel.value as string | undefined
+  const sessionModel = activeModel?.includes('/') ? activeModel : getDefaultStreamingModel()
+  if (!sessionModel?.includes('/'))
+    return null
+  const apiResourceId = sessionModel.split('/', 2)[1]
   // TTS 2.0 / ICL 2.0 ship subtitles asynchronously relative to audio
   // (per the wire spec), so chunk-on-sentence-end would drop frames.
   // Buffer the entire session and decode at session.finished instead.

packages/stage-ui/src/composables/use-auth-provider-sync.ts

@@ -1,6 +1,7 @@
 import { nextTick } from 'vue'
 
 import { initializeAuth } from '../libs/auth'
+import { getStreamingTtsAvailable } from '../libs/providers'
 import { useAuthStore } from '../stores/auth'
 import { useConsciousnessStore } from '../stores/modules/consciousness'
 import { useHearingStore } from '../stores/modules/hearing'
@@ -16,6 +17,14 @@ const AUTH_ACTIVATED_PROVIDERS: Array<{ id: string, module: 'consciousness' | 's
   { id: 'official-provider-speech', module: 'speech' },
 ]
 
+// The streaming TTS provider is NOT in the static list above because its
+// visibility is operator-controlled: `UNSPEECH_UPSTREAM.streaming` may be
+// unconfigured server-side. It's bootstrapped separately (see
+// `syncStreamingSpeechProvider`) — probed on sign-in, then force-configured
+// only when the server reports it available, mirroring how the HTTP TTS
+// provider uses `forceProviderConfigured` but gating it on a server signal.
+const STREAMING_SPEECH_PROVIDER_ID = 'official-provider-speech-streaming'
+
 /**
  * Glue layer: uses auth lifecycle hooks to activate/deactivate
  * official providers. Providers themselves know nothing about auth.
@@ -60,7 +69,7 @@ export function useAuthProviderSync() {
         case 'speech':
           if (!speechStore.activeSpeechProvider || speechStore.activeSpeechProvider === 'speech-noop') {
             speechStore.activeSpeechProvider = id
-            speechStore.activeSpeechModel = 'auto'
+            speechStore.activeSpeechModel = ''
           }
           break
         case 'hearing':
@@ -85,8 +94,54 @@ export function useAuthProviderSync() {
     catch (err) {
       console.error('error loading models for official providers', err)
     }
+
+    await syncStreamingSpeechProvider()
   })
 
+  // Bootstrap the streaming TTS provider from the server's availability signal.
+  // Probing populates `getStreamingTtsAvailable()` (and the default model /
+  // voices) via the provider's listModels(). The availability override drives
+  // the provider's presence in the available/configured lists (and thus the
+  // settings card + picker); force-configure makes it selectable. It is never
+  // set as the active speech provider — the HTTP TTS provider stays default.
+  async function syncStreamingSpeechProvider() {
+    if (providersStore.getProviderMetadata(STREAMING_SPEECH_PROVIDER_ID) == null)
+      return
+
+    await providersStore.fetchModelsForProvider(STREAMING_SPEECH_PROVIDER_ID)
+
+    const available = getStreamingTtsAvailable()
+    providersStore.setProviderAvailabilityOverride(STREAMING_SPEECH_PROVIDER_ID, available)
+
+    if (available) {
+      providersStore.forceProviderConfigured(STREAMING_SPEECH_PROVIDER_ID)
+      // The speech-module watcher skips voice loading for streaming until it's
+      // confirmed configured (avoids a pre-probe request on reload), so when
+      // streaming is the persisted active provider, load its voices now that
+      // it's confirmed available.
+      if (speechStore.activeSpeechProvider === STREAMING_SPEECH_PROVIDER_ID) {
+        speechStore.ensureStreamingDefaultModel()
+        await speechStore.loadVoicesForProvider(STREAMING_SPEECH_PROVIDER_ID, speechStore.activeSpeechModel || undefined)
+      }
+      return
+    }
+
+    providersStore.setProviderUnconfigured(STREAMING_SPEECH_PROVIDER_ID)
+    // `setProviderUnconfigured` blanks `validatedCredentialHash`, which makes
+    // the speech-module reset watcher skip its own clear. So when the server
+    // now reports streaming unavailable on an authenticated reload (no logout
+    // event fires), clear a stale active streaming selection here.
+    clearActiveStreamingSelection()
+  }
+
+  function clearActiveStreamingSelection() {
+    if (speechStore.activeSpeechProvider !== STREAMING_SPEECH_PROVIDER_ID)
+      return
+    speechStore.activeSpeechProvider = ''
+    speechStore.activeSpeechModel = ''
+    speechStore.activeSpeechVoiceId = ''
+  }
+
   authStore.onLogout(() => {
     hasSynced = false
 
@@ -94,6 +149,15 @@ export function useAuthProviderSync() {
       providersStore.setProviderUnconfigured(id)
     }
 
+    // Streaming TTS is bootstrapped outside AUTH_ACTIVATED_PROVIDERS, so reset
+    // it explicitly. `setProviderUnconfigured` blanks `validatedCredentialHash`,
+    // which makes the speech-module watcher skip its own reset (it guards
+    // against racing initial validation), so clear the active selection here
+    // too when streaming was the active provider.
+    clearActiveStreamingSelection()
+    providersStore.setProviderUnconfigured(STREAMING_SPEECH_PROVIDER_ID)
+    providersStore.setProviderAvailabilityOverride(STREAMING_SPEECH_PROVIDER_ID, false)
+
     // Reset active provider/model if they belong to an auth-activated provider
     for (const { id, module } of AUTH_ACTIVATED_PROVIDERS) {
       switch (module) {

packages/stage-ui/src/libs/providers/providers/index.ts

@@ -34,6 +34,11 @@ import './cloudflare-workers-ai'
 import './azure-ai-foundry'
 import './official'
 
+export {
+  getDefaultStreamingModel,
+  getStreamingTtsAvailable,
+} from './official'
+
 export {
   getDefinedProvider,
   listProviders,

packages/stage-ui/src/libs/providers/providers/official/index.ts

@@ -2,7 +2,7 @@ import type { Ref, WatchSource } from 'vue'
 
 import type { ModelInfo, VoiceInfo } from '../../../../stores/providers'
 
-import { watch } from 'vue'
+import { ref, watch } from 'vue'
 import { z } from 'zod'
 
 import { getAuthToken } from '../../../../libs/auth'
@@ -13,11 +13,34 @@ import { createOfficialAudioProvider, createOfficialOpenAIProvider, OFFICIAL_ICO
 export const OFFICIAL_SPEECH_PROVIDER_ID = 'official-provider-speech'
 export const OFFICIAL_SPEECH_STREAMING_PROVIDER_ID = 'official-provider-speech-streaming'
 
-// Locale → voice id map recommended by the server. Populated by listVoices()
-// from the /audio/voices response's `recommended` field so the auto-pick can
-// prefer a curated default per locale. Falls back to language + first-voice
-// matching when the server returns no recommendations.
-let recommendedVoicesByLocale: Record<string, string> = {}
+// Locale → voice id map recommended by the server, keyed by provider id.
+// Populated by each speech provider's listVoices() from the response's
+// `recommended` field so the auto-pick can prefer a curated default per
+// locale. Keyed per provider because the HTTP and streaming providers have
+// independent catalogs and recommendation buckets. Falls back to language +
+// first-voice matching when the server returns no recommendations.
+const recommendedVoicesByProvider: Record<string, Record<string, string>> = {}
+
+// Server-curated default streaming model id, populated by the streaming
+// provider's listModels(). Pages that need to seed an initial model selection
+// read this via getDefaultStreamingModel() instead of hardcoding an id.
+let defaultStreamingModelId: string | null = null
+
+export function getDefaultStreamingModel(): string | null {
+  return defaultStreamingModelId
+}
+
+// Operator-controlled visibility switch for the streaming provider. The server
+// reports it via `/api/v1/audio/models/streaming` (`available`), and the
+// auth-activation glue gates `forceProviderConfigured` on this so the provider
+// only surfaces when `UNSPEECH_UPSTREAM.streaming` is configured server-side.
+// Reactive so the providers store re-derives configured speech providers when
+// the probe resolves after sign-in.
+const streamingTtsAvailable = ref(false)
+
+export function getStreamingTtsAvailable(): boolean {
+  return streamingTtsAvailable.value
+}
 
 const officialConfigSchema = z.object({})
 
@@ -107,9 +130,8 @@ export const providerOfficialSpeech = defineProvider({
     listVoices: async (_config, _provider, model): Promise<VoiceInfo[]> => {
       // Voice catalogs are model-scoped on the server side. Pass the active
       // model through so Azure / cosyvoice / future provider voices route to
-      // the right adapter. `auto` defers to the server's DEFAULT_TTS_MODEL,
-      // but it MUST be sent explicitly — an absent `model` is treated as a
-      // client bug and returns 400.
+      // the right adapter. If model discovery has not completed yet, keep the
+      // legacy `auto` request as a startup fallback.
       const target = model && model.length > 0 ? model : 'auto'
       const url = new URL(`${SERVER_URL}/api/v1/audio/voices`)
       url.searchParams.set('model', target)
@@ -137,7 +159,7 @@ export const providerOfficialSpeech = defineProvider({
       // Refresh the server-side recommendation map. Done here rather than
       // threading it through the return value because the auto-pick watcher
       // lives in this module and reads the same singleton.
-      recommendedVoicesByLocale = (data.recommended && typeof data.recommended === 'object') ? data.recommended : {}
+      recommendedVoicesByProvider[OFFICIAL_SPEECH_PROVIDER_ID] = (data.recommended && typeof data.recommended === 'object') ? data.recommended : {}
 
       if (!Array.isArray(data.voices))
         throw new Error('audio voices upstream returned malformed body')
@@ -155,10 +177,9 @@ export const providerOfficialSpeech = defineProvider({
           // NOTICE: deliberately dropping `compatible_models`. The official
           // provider resolves voices through the server's /audio/voices?model=
           // endpoint, which already returns only voices valid for the active
-          // model (or the DEFAULT_TTS_MODEL when the client's selection is the
-          // 'auto' alias). Re-applying the client-side filter on top would
-          // zero out the list because upstream compatibility ids never match
-          // 'auto'. See packages/stage-pages/.../speech.vue filter predicate.
+          // model. Re-applying the client-side filter on top can zero out the
+          // list when upstream compatibility ids differ from AIRI's router ids.
+          // See packages/stage-pages/.../speech.vue filter predicate.
           languages: Array.isArray(v.languages) ? v.languages : [],
         }
       })
@@ -214,17 +235,26 @@ export const providerOfficialSpeechStreaming = defineProvider({
   extraMethods: {
     listModels: async (): Promise<ModelInfo[]> => {
       // Streaming TTS catalog is operator-controlled via configKV
-      // (`STREAMING_TTS_MODELS`). The wire `model` field uses the
-      // `<backend>/<api_resource_id>` shape unspeech expects (see
-      // `unspeech/docs/wire-protocols/audio-speech-stream-v1.md`); the server
-      // returns whatever the operator put there, no client-side defaults.
+      // (`UNSPEECH_UPSTREAM.streaming`). Wire shape uses `<backend>/<api_resource_id>`
+      // (see `unspeech/docs/wire-protocols/audio-speech-stream-v1.md`); the
+      // server returns whatever the operator put there, no client-side
+      // defaults. `default` (when set) seeds initial model selection via
+      // {@link getDefaultStreamingModel}.
+      // Reset the operator-driven signals up front so a failed/aborted probe
+      // leaves the provider hidden rather than stuck on a stale "available".
+      streamingTtsAvailable.value = false
+      defaultStreamingModelId = null
+
       const res = await globalThis.fetch(`${SERVER_URL}/api/v1/audio/models/streaming`, { headers: authHeaders() })
       if (!res.ok)
         throw new Error(`streaming models upstream ${res.status}: ${await res.text().catch(() => '')}`.slice(0, 256))
 
-      const data = await res.json() as { models?: { id: string, name?: string, description?: string }[] }
+      const data = await res.json() as { available?: boolean, models: { id: string, name?: string, description?: string }[], default?: string | null }
       if (!Array.isArray(data.models))
-        throw new Error('streaming models upstream returned malformed body')
+        throw new Error('streaming models upstream missing models[]')
+
+      streamingTtsAvailable.value = data.available === true
+      defaultStreamingModelId = typeof data.default === 'string' && data.default.length > 0 ? data.default : null
 
       return data.models.map(m => ({
         id: m.id,
@@ -235,8 +265,8 @@ export const providerOfficialSpeechStreaming = defineProvider({
     },
     listVoices: async (_config, _provider, model): Promise<VoiceInfo[]> => {
       // Streaming voices live behind a dedicated endpoint
-      // (`/audio/voices/streaming`) because they come from a separate
-      // configKV entry (`STREAMING_TTS_UPSTREAM`) than the HTTP TTS
+      // (`/audio/voices/streaming`) because they come from the
+      // `UNSPEECH_UPSTREAM.streaming` configKV subtree rather than the HTTP TTS
       // `?model=...` lookup. The server proxies to unspeech's
       // `/api/voices?provider=volcengine`, which ships an embed-time
       // catalogue without requiring credentials.
@@ -264,7 +294,14 @@ export const providerOfficialSpeechStreaming = defineProvider({
           languages?: { code: string, title: string }[]
           preview_audio_url?: string
         }[]
+        recommended?: Record<string, string>
       }
+
+      // Mirror the HTTP provider: stash the server's per-locale recommendations
+      // so setupOfficialSpeechAutoPick can seed a curated default voice when
+      // the streaming provider becomes active.
+      recommendedVoicesByProvider[OFFICIAL_SPEECH_STREAMING_PROVIDER_ID] = (data.recommended && typeof data.recommended === 'object') ? data.recommended : {}
+
       if (!Array.isArray(data.voices))
         throw new Error('streaming voices upstream returned malformed body')
 
@@ -318,10 +355,13 @@ function lookupRecommendedVoiceId(locale: string, map: Record<string, string>):
   return undefined
 }
 
-// NOTICE: Only the official speech provider auto-configures a default voice
-// after login. Third-party providers leave voice selection to the user. The
-// target locale is derived from the UI locale on each run — we don't persist
-// it, since that was the root of the cross-provider filter drift bug.
+const AUTO_PICK_PROVIDER_IDS = new Set([OFFICIAL_SPEECH_PROVIDER_ID, OFFICIAL_SPEECH_STREAMING_PROVIDER_ID])
+
+// NOTICE: Only the official speech providers (HTTP + streaming) auto-configure
+// a default voice after login. Third-party providers leave voice selection to
+// the user. The target locale is derived from the UI locale on each run — we
+// don't persist it, since that was the root of the cross-provider filter
+// drift bug.
 export function setupOfficialSpeechAutoPick(ctx: {
   activeSpeechProvider: Ref<string>
   activeSpeechVoiceId: Ref<string>
@@ -329,12 +369,12 @@ export function setupOfficialSpeechAutoPick(ctx: {
   uiLocale: WatchSource<string> | Ref<string>
 }) {
   watch([ctx.availableVoices, ctx.activeSpeechProvider], ([voices, provider]) => {
-    if (provider !== OFFICIAL_SPEECH_PROVIDER_ID)
+    if (!AUTO_PICK_PROVIDER_IDS.has(provider))
       return
     if (ctx.activeSpeechVoiceId.value)
       return
 
-    const providerVoices = voices[OFFICIAL_SPEECH_PROVIDER_ID]
+    const providerVoices = voices[provider]
     if (!providerVoices?.length)
       return
 
@@ -356,7 +396,7 @@ export function setupOfficialSpeechAutoPick(ctx: {
     //   3) any English voice (en-US, then en-*) — broadest comprehensible
     //      fallback when the user's locale has no coverage at all
     //   4) alphabetical first voice, as a last resort
-    const recommendedId = lookupRecommendedVoiceId(targetLocale, recommendedVoicesByLocale)
+    const recommendedId = lookupRecommendedVoiceId(targetLocale, recommendedVoicesByProvider[provider] ?? {})
     const speaksLocale = (v: VoiceInfo, code: string) => (v.languages || []).some(l => l.code === code)
     const match = (recommendedId && providerVoices.find(v => v.id === recommendedId))
       || providerVoices.find(v => speaksLocale(v, targetLocale))

packages/stage-ui/src/stores/modules/speech.test.ts

@@ -1,8 +1,22 @@
-import { describe, expect, it } from 'vitest'
+import { createPinia, setActivePinia } from 'pinia'
+import { beforeEach, describe, expect, it, vi } from 'vitest'
 
-import { toSignedPercent } from './speech'
+import { OFFICIAL_SPEECH_PROVIDER_ID, OFFICIAL_SPEECH_STREAMING_PROVIDER_ID } from '../../libs/providers/providers/official'
+import { useProvidersStore } from '../providers'
+import { toSignedPercent, useSpeechStore } from './speech'
+
+vi.mock('vue-i18n', () => ({
+  useI18n: () => ({
+    locale: { value: 'en-US' },
+    t: (_key: string, fallback?: string) => fallback ?? _key,
+  }),
+}))
 
 describe('speech store helpers', () => {
+  beforeEach(() => {
+    setActivePinia(createPinia())
+  })
+
   it('formats positive percentages with a plus sign', () => {
     expect(toSignedPercent(25)).toBe('+25%')
   })
@@ -15,4 +29,53 @@ describe('speech store helpers', () => {
   it('formats zero as 0%', () => {
     expect(toSignedPercent(0)).toBe('0%')
   })
+
+  /**
+   * @example
+   * await speechStore.loadVoicesForProvider(OFFICIAL_SPEECH_STREAMING_PROVIDER_ID, 'volcengine/seed-tts-2.0')
+   */
+  it('does not load streaming voices before server availability is confirmed', async () => {
+    const providersStore = useProvidersStore()
+    const speechStore = useSpeechStore()
+    const listVoices = vi.fn(async () => [])
+    const metadata = providersStore.providerMetadata[OFFICIAL_SPEECH_STREAMING_PROVIDER_ID]
+    metadata.capabilities.listVoices = listVoices
+    providersStore.providerRuntimeState[OFFICIAL_SPEECH_STREAMING_PROVIDER_ID].isConfigured = false
+
+    const voices = await speechStore.loadVoicesForProvider(
+      OFFICIAL_SPEECH_STREAMING_PROVIDER_ID,
+      'volcengine/seed-tts-2.0',
+    )
+
+    expect(voices).toEqual([])
+    expect(listVoices).not.toHaveBeenCalled()
+  })
+
+  /**
+   * @example
+   * speechStore.ensureActiveSpeechModel()
+   */
+  it('resets stale streaming model when the regular official speech provider is active', () => {
+    const providersStore = useProvidersStore()
+    const speechStore = useSpeechStore()
+    speechStore.activeSpeechProvider = OFFICIAL_SPEECH_PROVIDER_ID
+    speechStore.activeSpeechModel = 'volcengine/seed-tts-2.0'
+    speechStore.activeSpeechVoiceId = 'zh_female_x'
+    speechStore.activeSpeechVoice = {
+      id: 'zh_female_x',
+      name: 'X',
+      provider: OFFICIAL_SPEECH_STREAMING_PROVIDER_ID,
+      languages: [],
+    }
+    providersStore.providerRuntimeState[OFFICIAL_SPEECH_PROVIDER_ID].models = [
+      { id: 'microsoft/v1', name: 'microsoft/v1', provider: OFFICIAL_SPEECH_PROVIDER_ID },
+      { id: 'alibaba/cosyvoice-v2', name: 'alibaba/cosyvoice-v2', provider: OFFICIAL_SPEECH_PROVIDER_ID },
+    ]
+
+    speechStore.ensureActiveSpeechModel()
+
+    expect(speechStore.activeSpeechModel).toBe('microsoft/v1')
+    expect(speechStore.activeSpeechVoiceId).toBe('')
+    expect(speechStore.activeSpeechVoice).toBeUndefined()
+  })
 })

packages/stage-ui/src/stores/modules/speech.ts

@@ -2,6 +2,7 @@ import type { SpeechProviderWithExtraOptions } from '@xsai-ext/providers/utils'
 
 import type { VoiceInfo } from '../providers'
 
+import { errorMessageFrom } from '@moeru/std'
 import { useLocalStorageManualReset } from '@proj-airi/stage-shared/composables'
 import { refManualReset } from '@vueuse/core'
 import { generateSpeech } from '@xsai/generate-speech'
@@ -11,7 +12,7 @@ import { useI18n } from 'vue-i18n'
 import { toXml } from 'xast-util-to-xml'
 import { x } from 'xastscript'
 
-import { setupOfficialSpeechAutoPick } from '../../libs/providers/providers/official'
+import { getDefaultStreamingModel, OFFICIAL_SPEECH_PROVIDER_ID, OFFICIAL_SPEECH_STREAMING_PROVIDER_ID, setupOfficialSpeechAutoPick } from '../../libs/providers/providers/official'
 import { useProvidersStore } from '../providers'
 
 export function toSignedPercent(value: number): string {
@@ -88,6 +89,13 @@ export const useSpeechStore = defineStore('speech', () => {
       return []
     }
 
+    // Streaming provider visibility is server-driven and only confirmed after
+    // the auth probe force-configures it. Keep the gate at the public loader so
+    // pages cannot bypass it and issue `/voices/streaming` while unavailable.
+    if (provider === OFFICIAL_SPEECH_STREAMING_PROVIDER_ID && !providersStore.configuredProviders[provider]) {
+      return []
+    }
+
     isLoadingSpeechProviderVoices.value = true
     speechProviderError.value = null
 
@@ -102,7 +110,7 @@ export const useSpeechStore = defineStore('speech', () => {
     }
     catch (error) {
       console.error(`Error fetching voices for ${provider}:`, error)
-      speechProviderError.value = error instanceof Error ? error.message : 'Unknown error'
+      speechProviderError.value = errorMessageFrom(error) ?? 'Unknown error'
       return []
     }
     finally {
@@ -115,12 +123,64 @@ export const useSpeechStore = defineStore('speech', () => {
     return availableVoices.value[provider] || []
   }
 
+  function clearVoiceSelection() {
+    activeSpeechVoiceId.value = ''
+    activeSpeechVoice.value = undefined
+  }
+
+  // Streaming TTS voices are model-scoped: the server only returns recommended
+  // voices for an explicit `?model=`. Ensure the active model is a valid
+  // streaming model id so voice loading gets the right recommendations (parity
+  // with the HTTP provider's auto-pick). Reseeds the server-curated default
+  // both when no model is selected AND when `activeSpeechModel` still holds a
+  // stale id from a previously-active provider (the global model ref is shared
+  // across providers, and the per-surface reset may not have run yet). No-op
+  // for non-streaming providers.
+  function ensureStreamingDefaultModel() {
+    if (activeSpeechProvider.value !== OFFICIAL_SPEECH_STREAMING_PROVIDER_ID)
+      return
+    const streamingModels = providersStore.getModelsForProvider(OFFICIAL_SPEECH_STREAMING_PROVIDER_ID)
+    const hasValidSelection = !!activeSpeechModel.value && streamingModels.some(m => m.id === activeSpeechModel.value)
+    if (hasValidSelection)
+      return
+    // Replace an empty/stale (non-streaming) selection with the server default.
+    // When no default can be resolved yet (catalog not loaded), clear it to ''
+    // so callers pass `undefined` (server returns the full streaming catalog)
+    // rather than forwarding a stale non-streaming model id as `?model=`.
+    const nextModel = getDefaultStreamingModel() ?? streamingModels[0]?.id ?? ''
+    if (activeSpeechModel.value === nextModel)
+      return
+    activeSpeechModel.value = nextModel
+    // The previously-selected voice belonged to the stale/empty model context,
+    // so drop it; auto-pick re-picks a recommended voice for the new model.
+    clearVoiceSelection()
+  }
+
+  function ensureActiveSpeechModel() {
+    ensureStreamingDefaultModel()
+
+    if (activeSpeechProvider.value !== OFFICIAL_SPEECH_PROVIDER_ID)
+      return
+
+    const models = providersStore.getModelsForProvider(OFFICIAL_SPEECH_PROVIDER_ID)
+    if (!models.length)
+      return
+
+    const hasValidSelection = !!activeSpeechModel.value && models.some(m => m.id === activeSpeechModel.value)
+    if (hasValidSelection)
+      return
+
+    activeSpeechModel.value = models[0]?.id ?? ''
+    clearVoiceSelection()
+  }
+
   // Watch for provider changes and load voices
   watch(activeSpeechProvider, async (newProvider) => {
-    if (newProvider) {
-      await loadVoicesForProvider(newProvider)
-      // Don't reset voice settings when changing providers to allow for persistence
-    }
+    if (!newProvider)
+      return
+    ensureActiveSpeechModel()
+    await loadVoicesForProvider(newProvider, activeSpeechModel.value || undefined)
+    // Don't reset voice settings when changing providers to allow for persistence
   }, {
     // REVIEW: should we always load voices on init? What will happen when network is not available?
     immediate: true,
@@ -159,7 +219,8 @@ export const useSpeechStore = defineStore('speech', () => {
   )
 
   onMounted(() => {
-    loadVoicesForProvider(activeSpeechProvider.value).then(() => {
+    ensureActiveSpeechModel()
+    loadVoicesForProvider(activeSpeechProvider.value, activeSpeechModel.value || undefined).then(() => {
       if (activeSpeechVoiceId.value) {
         activeSpeechVoice.value = availableVoices.value[activeSpeechProvider.value]?.find(voice => voice.id === activeSpeechVoiceId.value)
       }
@@ -173,6 +234,10 @@ export const useSpeechStore = defineStore('speech', () => {
     uiLocale: locale,
   })
 
+  watch(providerModels, () => {
+    ensureActiveSpeechModel()
+  })
+
   watch([activeSpeechVoiceId, availableVoices], ([voiceId, voices]) => {
     if (voiceId) {
       // For OpenAI Compatible, create a custom voice object (no voices available from API)
@@ -334,6 +399,8 @@ export const useSpeechStore = defineStore('speech', () => {
     speech,
     loadVoicesForProvider,
     getVoicesForProvider,
+    ensureStreamingDefaultModel,
+    ensureActiveSpeechModel,
     generateSSML,
     resetState,
   }

packages/stage-ui/src/stores/providers.ts

@@ -2275,6 +2275,19 @@ export const useProvidersStore = defineStore('providers', () => {
   const providerValidationInFlight = new Map<string, Promise<boolean>>()
   const providerRevalidationLoops = new Map<string, { resume: () => void }>()
 
+  // Server-driven availability overrides for providers whose visibility can
+  // only be decided at runtime from the backend (e.g. the streaming TTS
+  // provider, which exists only when `UNSPEECH_UPSTREAM.streaming` is
+  // configured server-side). A `false` entry hides the provider from the
+  // available lists regardless of its static `isAvailableBy`; an absent entry
+  // means no override. Written by the auth-sync glue after it probes the
+  // server. Reactive so the available/configured provider lists re-derive.
+  const providerAvailabilityOverrides = ref<Record<string, boolean>>({})
+
+  function setProviderAvailabilityOverride(providerId: string, available: boolean) {
+    providerAvailabilityOverrides.value = { ...providerAvailabilityOverrides.value, [providerId]: available }
+  }
+
   const configuredProviders = computed(() => {
     const result: Record<string, boolean> = {}
     for (const [key, state] of Object.entries(providerRuntimeState.value)) {
@@ -2686,9 +2699,17 @@ export const useProvidersStore = defineStore('providers', () => {
   }
 
   const availableProvidersMetadata = computedAsync<ProviderMetadata[]>(async () => {
+    // Spread-read the overrides synchronously so this re-runs when a
+    // server-driven availability flips: computedAsync uses watchEffect, which
+    // only tracks reactive reads before the first `await` — the per-provider
+    // `isAvailableBy()` below runs after one, so reads inside it aren't tracked.
+    const overrides = { ...providerAvailabilityOverrides.value }
     const providers: ProviderMetadata[] = []
 
     for (const provider of allProvidersMetadata.value) {
+      if (overrides[provider.id] === false)
+        continue
+
       const p = getProviderMetadata(provider.id)
       const isAvailableBy = p.isAvailableBy || (() => true)
 
@@ -2786,6 +2807,7 @@ export const useProvidersStore = defineStore('providers', () => {
     resetProviderSettings,
     forceProviderConfigured,
     setProviderUnconfigured,
+    setProviderAvailabilityOverride,
     availableProvidersMetadata,
     allChatProvidersMetadata,
     allAudioSpeechProvidersMetadata,