fix(ai): route scene generation through configured providers

backend/unispeaking-server/src/main/java/com/unispeaking/component/scene/CustomSceneGenerator.java

@@ -92,20 +92,17 @@ public CustomSceneDefinition generate(
 		String normalizedInput = requiredInput(sceneInput);
 		String prompt = buildPrompt(normalizedInput, currentPreference, profile);
 		BusinessException lastFailure = null;
-		List<String> models = List.of(
-				AiProviderRegistry.QWEN_LLM_FLASH,
-				AiProviderRegistry.QWEN_LLM_PLUS);
-		for (int index = 0; index < models.size(); index++) {
+		int maximumAttempts = 2;
+		for (int index = 0; index < maximumAttempts; index++) {
 			int attempt = index + 1;
-			String modelId = models.get(index);
 			String attemptPrompt = attempt == 1
 					? prompt
 					: prompt + "\n\nA prior generation attempt did not satisfy the JSON contract. "
 							+ "Return a corrected JSON object only.";
 			try {
 				long llmStartedAt = System.nanoTime();
 				String content = providerRegistry.executeLlmTask(
-						modelId,
+						null,
 						attemptPrompt,
 						null,
 						LlmResponseFormat.JSON_OBJECT);
@@ -118,9 +115,8 @@ public CustomSceneDefinition generate(
 				catch (BusinessException exception) {
 					if ("CUSTOM_SCENE_LLM_RESPONSE_INVALID".equals(exception.code())) {
 						LOGGER.warn(
-								"custom scene LLM response rejected sceneId={} model={} attempt={} llmMs={} parseMs={} responseChars={}",
+								"custom scene LLM response rejected sceneId={} route=default attempt={} llmMs={} parseMs={} responseChars={}",
 								sceneId,
-								modelId,
 								attempt,
 								llmMillis,
 								elapsedMillis(parseStartedAt),
@@ -129,22 +125,20 @@ public CustomSceneDefinition generate(
 					throw exception;
 				}
 				LOGGER.info(
-						"custom scene LLM completed sceneId={} model={} attempt={} llmMs={} parseMs={}",
+						"custom scene LLM completed sceneId={} route=default attempt={} llmMs={} parseMs={}",
 						sceneId,
-						modelId,
 						attempt,
 						llmMillis,
 						elapsedMillis(parseStartedAt));
 				return definition;
 			}
 			catch (BusinessException exception) {
 				lastFailure = exception;
-				if (index + 1 < models.size()) {
+				if (index + 1 < maximumAttempts) {
 					LOGGER.warn(
-							"custom scene LLM falling back sceneId={} failedModel={} nextModel={} code={}",
+							"custom scene LLM retrying configured route sceneId={} attempt={} code={}",
 							sceneId,
-							modelId,
-							models.get(index + 1),
+							attempt,
 							exception.code());
 					continue;
 				}

backend/unispeaking-server/src/main/java/com/unispeaking/infrastructure/ai/qwen/QwenLlmProvider.java

@@ -39,6 +39,8 @@ public class QwenLlmProvider extends LlmProvider {
 	private final ObjectMapper objectMapper;
 	private final String apiKey;
 	private final URI endpoint;
+	private String workspaceId = "";
+	private String region = "";
 	private final String model;
 	private final Duration readTimeout;
 	private final int maxResponseBytes;
@@ -63,6 +65,8 @@ public QwenLlmProvider(
 				model,
 				positiveDuration(readTimeoutSeconds, "Qwen LLM read timeout"),
 				maxResponseBytes);
+		this.workspaceId = trim(workspaceId);
+		this.region = trim(region);
 	}
 
 	public QwenLlmProvider(
@@ -131,7 +135,8 @@ private AiProviderResponse<String> callForContent(
 		if (prompt.isBlank()) {
 			throw nonRetryableFailure("INVALID_LLM_PROMPT", "LLM task prompt is required");
 		}
-		requireHttpsEndpoint();
+		URI requestEndpoint = resolveEndpoint();
+		requireHttpsEndpoint(requestEndpoint);
 
 		try {
 			Map<String, Object> body = new LinkedHashMap<>();
@@ -142,7 +147,7 @@ private AiProviderResponse<String> callForContent(
 				body.put("response_format", Map.of("type", "json_object"));
 			}
 			HttpRequest httpRequest = HttpRequest.newBuilder()
-					.uri(endpoint)
+					.uri(requestEndpoint)
 					.timeout(readTimeout)
 					.header("Authorization", "Bearer " + credential)
 					.header("Content-Type", "application/json")
@@ -225,19 +230,27 @@ private Object parseContent(String content) {
 		}
 	}
 
-	private void requireHttpsEndpoint() {
-		String host = endpoint == null || endpoint.getHost() == null
+	private URI resolveEndpoint() {
+		String effectiveWorkspaceId = ProviderCredentialOverride.currentOr("workspaceId", workspaceId);
+		if (!effectiveWorkspaceId.isBlank()) {
+			return buildEndpoint(effectiveWorkspaceId, region);
+		}
+		return endpoint;
+	}
+
+	private void requireHttpsEndpoint(URI requestEndpoint) {
+		String host = requestEndpoint == null || requestEndpoint.getHost() == null
 				? ""
-				: endpoint.getHost().toLowerCase(java.util.Locale.ROOT);
-		if (endpoint == null
-				|| !endpoint.isAbsolute()
-				|| !"https".equalsIgnoreCase(endpoint.getScheme())
+				: requestEndpoint.getHost().toLowerCase(java.util.Locale.ROOT);
+		if (requestEndpoint == null
+				|| !requestEndpoint.isAbsolute()
+				|| !"https".equalsIgnoreCase(requestEndpoint.getScheme())
 				|| !host.endsWith(".maas.aliyuncs.com")
-				|| endpoint.getUserInfo() != null
-				|| endpoint.getPort() != -1
-				|| !"/compatible-mode/v1/chat/completions".equals(endpoint.getPath())
-				|| endpoint.getRawQuery() != null
-				|| endpoint.getRawFragment() != null) {
+				|| requestEndpoint.getUserInfo() != null
+				|| requestEndpoint.getPort() != -1
+				|| !"/compatible-mode/v1/chat/completions".equals(requestEndpoint.getPath())
+				|| requestEndpoint.getRawQuery() != null
+				|| requestEndpoint.getRawFragment() != null) {
 			throw retryableFailure(
 					"QWEN_LLM_ENDPOINT_INVALID",
 					"Qwen LLM endpoint must be the trusted Aliyun compatible-mode URL");

backend/unispeaking-server/src/main/java/com/unispeaking/provider/AiProviderRegistry.java

@@ -115,9 +115,7 @@ public AiProviderRegistry(
 			@Value("${AI_PROVIDER_ROUTE_TTS:}")
 			String ttsRoute,
 			@Value("${AI_PROVIDER_ROUTE_TRANSCRIPTION:}")
-			String transcriptionRoute,
-			@Value("${AI_QINIU_MODELS_ENABLED:false}")
-			boolean qiniuModelsEnabled) {
+			String transcriptionRoute) {
 		this(
 				realtimeProviders,
 				llmProviders,
@@ -129,8 +127,7 @@ AiCapability.REALTIME, parseRoute(realtimeRoute),
 						AiCapability.LLM, parseRoute(llmRoute),
 						AiCapability.SCORING, parseRoute(scoringRoute),
 						AiCapability.TTS, parseRoute(ttsRoute),
-						AiCapability.TRANSCRIPTION, parseRoute(transcriptionRoute)),
-				qiniuModelsEnabled);
+						AiCapability.TRANSCRIPTION, parseRoute(transcriptionRoute)));
 	}
 
 	public AiProviderRegistry(
@@ -155,30 +152,6 @@ public AiProviderRegistry(
 			List<TtsProvider> ttsProviders,
 			List<TranscriptionProvider> transcriptionProviders,
 			Map<AiCapability, List<String>> configuredRoutes) {
-		this(
-				realtimeProviders,
-				llmProviders,
-				scoringProviders,
-				ttsProviders,
-				transcriptionProviders,
-				configuredRoutes,
-				true);
-	}
-
-	AiProviderRegistry(
-			List<RealtimeProvider> realtimeProviders,
-			List<LlmProvider> llmProviders,
-			List<ScoringProvider> scoringProviders,
-			List<TtsProvider> ttsProviders,
-			List<TranscriptionProvider> transcriptionProviders,
-			Map<AiCapability, List<String>> configuredRoutes,
-			boolean qiniuModelsEnabled) {
-		if (!qiniuModelsEnabled) {
-			realtimeProviders = withoutProvider(realtimeProviders, "qiniu");
-			llmProviders = withoutProvider(llmProviders, "qiniu-maas");
-			configuredRoutes = withoutQiniuModels(configuredRoutes);
-			LOGGER.info("Qiniu AI models are disabled; RTI and MaaS adapters will not be registered");
-		}
 		this.realtimeProviders = registerProviders(realtimeProviders, AiCapability.REALTIME);
 		this.llmProviders = registerProviders(llmProviders, AiCapability.LLM);
 		this.scoringProviders = registerProviders(scoringProviders, AiCapability.SCORING);
@@ -191,29 +164,6 @@ public AiProviderRegistry(
 		this.models = List.copyOf(modelDefinitions.values());
 	}
 
-	private static <T extends AbstractAiProvider> List<T> withoutProvider(
-			List<T> providers,
-			String providerId) {
-		return providers.stream()
-				.filter(provider -> !provider.providerId().equalsIgnoreCase(providerId))
-				.toList();
-	}
-
-	private static Map<AiCapability, List<String>> withoutQiniuModels(
-			Map<AiCapability, List<String>> configuredRoutes) {
-		if (configuredRoutes == null || configuredRoutes.isEmpty()) return Map.of();
-		Map<AiCapability, List<String>> filtered = new EnumMap<>(AiCapability.class);
-		configuredRoutes.forEach((capability, route) -> filtered.put(
-				capability,
-				route.stream()
-						.map(AbstractAiProvider::normalizeModelId)
-						.filter(modelId -> !modelId.equals(QINIU_REALTIME_PLUS))
-						.filter(modelId -> !modelId.equals(QINIU_MAAS_QWEN_PLUS))
-						.filter(modelId -> !modelId.equals(QINIU_MAAS_DEEPSEEK_FLASH))
-						.toList()));
-		return Map.copyOf(filtered);
-	}
-
 	public List<AiModelDefinition> models() {
 		AiRuntimeConfiguration runtime = runtimeConfiguration();
 		if (!runtime.databaseBacked()) return models;
@@ -426,9 +376,16 @@ public byte[] generateSpeechAudioBytes(
 			String token,
 			String voice) {
 		if (modelId == null || modelId.isBlank()) {
-			throw new BusinessException(
-					"AI_TTS_MODEL_REQUIRED",
-					"A voice-specific TTS request requires an explicit model");
+			return unboxAudio(invokeRouteWithResult(
+					context,
+					AiCapability.TTS,
+					id -> {
+						AiProviderResponse<byte[]> measured = getTtsProvider(id)
+								.generateSpeechAudioMeasured(
+										text, credential(id, token), voice);
+						return new AiProviderResponse<>(boxAudio(measured.response()),
+								measured.providerRequestId(), measured.usage());
+					}).response());
 		}
 		return invokeExplicitMeasured(context, AiCapability.TTS, modelId,
 				id -> getTtsProvider(id).generateSpeechAudioMeasured(
@@ -444,6 +401,9 @@ public String executeLlmTask(
 			String prompt,
 			String token,
 			LlmResponseFormat responseFormat) {
+		if (modelId == null || modelId.isBlank()) {
+			return executeLlmTaskRouted(prompt, token, responseFormat).response();
+		}
 		return invokeExplicitMeasured(
 				automaticContext("llm"),
 				AiCapability.LLM,
@@ -855,7 +815,14 @@ private AiModelConfiguration modelConfigurationForLedger(String modelId, AiCapab
 
 	private AiInvocationContext automaticContext(String businessScene) {
 		AiInvocationContext scoped = AiInvocationContexts.current();
-		if (scoped != null) return scoped;
+		if (scoped != null) {
+			return new AiInvocationContext(
+					UUID.randomUUID(),
+					scoped.userId(),
+					scoped.sessionId(),
+					scoped.businessScene(),
+					scoped.routeKey());
+		}
 		String userId = null;
 		try {
 			if (authService != null) userId = authService.currentUserIdOrNull();
@@ -877,11 +844,16 @@ private static long elapsedMillis(long startedNanos) {
 	}
 
 	private boolean shouldFailOver(BusinessException exception) {
+		String code = exception.code() == null ? "" : exception.code();
+		// Authentication and account-policy failures cannot be retried against the
+		// same Qiniu model, but they must not prevent the configured route fallback.
+		if ("QINIU_MAAS_LLM_REQUEST_FAILED".equals(code)) {
+			return true;
+		}
 		Boolean classifiedRetryable = AbstractAiProvider.retryable(exception);
 		if (classifiedRetryable != null) {
 			return classifiedRetryable;
 		}
-		String code = exception.code() == null ? "" : exception.code();
 		return !code.startsWith("INVALID_")
 				&& !code.startsWith("UNSUPPORTED_")
 				&& !code.endsWith("_INTERRUPTED")

backend/unispeaking-server/src/main/java/com/unispeaking/service/scene/CustomSceneService.java

@@ -105,7 +105,7 @@ public byte[] synthesizeSpeech(String sceneId, String text, String model) {
 		}
 		UserProfile profile = profileService.getProfile(definition.userId());
 		byte[] audio = providerRegistry.generateSpeechAudioBytes(
-				AiProviderRegistry.QWEN_TTS,
+				model,
 				text.strip(),
 				null,
 				profile == null ? null : profile.voiceId());