@hiper2d/ai-agents 0.14.1 → 0.16.0

package/README.md

@@ -69,10 +69,10 @@ const { audio, costUSD } = await voice.speak({ text: 'Night falls.', voice: 'Kor
 const { text } = await voice.transcribe({ audio: recording, mimeType: 'audio/webm' });
 ```
 
-`openai` runs gpt-4o-mini-tts + Whisper, `google` runs Gemini 3.1 Flash TTS + Gemini 3.5
+`openai` runs gpt-4o-mini-tts + Whisper, `google` runs Gemini 3.8 Flash-Lite TTS + Gemini 3.5
 Transcribe (`VOICE_MODEL_CONSTANTS`, prices in `VOICE_MODEL_PRICING`). The `voiceStyle`
-direction works for both providers: OpenAI takes it as instructions, Gemini gets it folded
-into the prompt ("Say gravely: …").
+direction works for both providers: OpenAI takes it as instructions, Gemini as the part's
+`speechMetadata.style` (3.8 TTS reads the text verbatim, so no inline "Say X:" prefix).
 
 ### Images and portrait sheets (`@hiper2d/ai-agents/images`)
 

package/dist/index.d.mts

@@ -907,6 +907,8 @@ interface SpeechResult {
     audio: ArrayBuffer;
     costUSD: number;
     usage: SpeechUsage;
+    /** Gemini only: the styled request was safety-blocked, so the line was read without its style. */
+    styleDropped?: boolean;
 }
 interface TranscriptionRequest {
     audio: ArrayBuffer;
@@ -943,7 +945,7 @@ interface VoiceAgent {
 declare const VOICE_MODEL_CONSTANTS: {
     readonly OPENAI_TTS: "gpt-4o-mini-tts";
     readonly OPENAI_STT: "whisper-1";
-    readonly GOOGLE_TTS: "gemini-3.1-flash-tts-preview";
+    readonly GOOGLE_TTS: "gemini-3.8-flash-lite-tts";
     readonly GOOGLE_STT: "gemini-3.5-transcribe";
 };
 declare const SUPPORTED_VOICE_PROVIDERS: readonly VoiceProvider[];
@@ -1006,7 +1008,7 @@ declare class OpenAiVoiceAgent implements VoiceAgent {
 declare class GoogleVoiceAgent implements VoiceAgent {
     private readonly apiKey;
     readonly provider: "google";
-    readonly ttsModel: "gemini-3.1-flash-tts-preview";
+    readonly ttsModel: "gemini-3.8-flash-lite-tts";
     readonly sttModel: "gemini-3.5-transcribe";
     constructor(apiKey: string);
     speak(request: SpeechRequest): Promise<SpeechResult>;
@@ -1042,7 +1044,7 @@ declare function transcribeWithOpenAi(audioBuffer: ArrayBuffer, apiKey: string,
 interface GoogleTtsAudioOptions {
     /** e.g. "Kore", "Puck" */
     voiceName: string;
-    /** "mysteriously", "excitedly", or a longer direction */
+    /** "mysteriously", "excitedly", or a longer direction; sent as speechMetadata.style */
     voiceStyle?: string;
 }
 interface GoogleTtsResult {
@@ -1053,15 +1055,16 @@ interface GoogleTtsResult {
         inputTokens: number;
         outputTokens: number;
     };
+    /** True when the styled request was safety-blocked and the line was read without its style. */
+    styleDropped?: boolean;
 }
 /**
- * Gemini TTS has no instruction field: delivery is directed in the text itself
- * ("Say cheerfully: Have a wonderful day!" in the docs). A short style (1-3
- * words) becomes that "Say X:" prefix; a longer direction is used as written,
- * ending in the colon that separates it from the line to read. The same style
- * value feeds OpenAI's `instructions`, so one field serves both providers.
+ * Gemini 3.8 TTS reads the part text as a verbatim transcript: an inline
+ * "Say cheerfully:" prefix (the 3.1 convention) may be spoken aloud. Delivery
+ * direction goes in the part's `speechMetadata.style` instead, short adverb or
+ * longer sentence alike. Trailing punctuation is trimmed; blank means none.
  */
-declare function buildGoogleTtsPrompt(text: string, voiceStyle?: string): string;
+declare function normalizeTtsStyle(voiceStyle?: string): string | undefined;
 /** Wraps raw 16-bit mono PCM in a WAV header. */
 declare function pcmToWav(pcmData: Uint8Array, sampleRate?: number): ArrayBuffer;
 /**
@@ -1071,12 +1074,22 @@ declare function pcmToWav(pcmData: Uint8Array, sampleRate?: number): ArrayBuffer
  * difference between a diagnosable log line and "no audio data".
  */
 declare function describeEmptyTtsResponse(response: any): string;
+/**
+ * A safety block on the prompt or the candidate. The 3.1 preview model
+ * false-positived on volume directions ("Say quietly: <harmless line>" blocked
+ * 4 of 6 on 2026-09-24); 3.8 Flash-Lite with the style in speechMetadata passed
+ * the same lines 9 of 9, but the fallback below stays as a net.
+ */
+declare function isSafetyBlocked(response: any): boolean;
 /**
  * Core Gemini TTS call: text + API key in, WAV + token usage out.
  *
- * Uses generateContent rather than the newer Interactions API the docs show:
- * both serve the 3.1 TTS model (verified 2026-09-05), and this one is typed in
- * the SDK and reports usageMetadata, which billing needs.
+ * 3.8 TTS returns WAV with a RIFF header by default (3.1 returned headerless
+ * audio/l16); raw PCM is still wrapped if a response ever carries it.
+ *
+ * When a styled request comes back safety-blocked, the line is read once more
+ * without the style: a block on the direction should not silence the line. An
+ * unstyled block is not repeated.
  */
 declare function generateGoogleTtsAudio(text: string, apiKey: string, options: GoogleTtsAudioOptions): Promise<GoogleTtsResult>;
 
@@ -1791,4 +1804,4 @@ declare class MiniMaxAgent extends AbstractAgent {
     doAskText(messages: AIMessage[]): Promise<[string, string, TokenUsage$1?, string?]>;
 }
 
-export { type AIMessage, ANTHROPIC_REASONING_EFFORTS, API_KEY_CONSTANTS, AbstractAgent, type AgentActivityData, AgentFactory, type AgentLoggingConfig, type AnthropicReasoningEffort, type AnthropicTokenUsage, type ApiKeyMap, type BeforeAskHook, BotResponseError, BudgetController, type BudgetControllerOptions, BudgetExceededError, type BudgetLimit, type BudgetVerdict, CACHE_TIER_MARKER, ClaudeAgent, type CostCalculationOptions, DEEPSEEK_PEAK_SCHEDULE, DEEPSEEK_REASONING_EFFORTS, DEFAULT_LOGGING_CONFIG, DEFAULT_MAX_OUTPUT_TOKENS, type DeepSeekReasoningEffort, type DeepSeekTokenUsage, DeepSeekV2Agent, FUGU_REASONING_EFFORTS, FuguAgent, type FuguReasoningEffort, GEMINI_AUDIO_TOKENS_PER_SECOND, GEMINI_REASONING_EFFORTS, GLM_REASONING_EFFORTS, type GeminiReasoningEffort, GlmAgent, type GlmReasoningEffort, GoogleAgent, type GoogleSttOptions, type GoogleSttResult, type GoogleTokenUsage, type GoogleTtsAudioOptions, type GoogleTtsResult, GoogleVoiceAgent, Gpt5Agent, GrokAgent, type GrokTokenUsage, InMemorySpendStore, type JsonSchemaOptions, KIMI_REASONING_EFFORTS, KimiAgent, type KimiReasoningEffort, type KimiTokenUsage, type LLMModel, LLM_CONSTANTS, type LlmLogger, type LoggingConfig, MESSAGE_ROLE, META_REASONING_EFFORTS, MISTRAL_REASONING_EFFORTS, MODEL_PRICING, MetaAgent, type MetaReasoningEffort, type MetaTokenUsage, MiniMaxAgent, MistralAgent, type MistralReasoningEffort, type MistralTokenUsage, type Modality, ModelAuthenticationError, type ModelConfig, ModelError, ModelInvalidResponseError, ModelOverloadError, type ModelPricing, ModelQuotaExceededError, ModelRateLimitError, ModelRefusalError, type ModelTag, ModelUnavailableError, OPENAI_REASONING_EFFORTS, type OpenAIReasoningEffort, type OpenAITokenUsage, type OpenAiSttOptions, type OpenAiSttResult, type OpenAiTtsAudioOptions, type OpenAiTtsVoice, OpenAiVoiceAgent, type PeakPricing, type PricingUnit, type ProviderSchema, type TokenUsage as ProviderTokenUsage, type ProviderType, QwenAgent, REASONING_EFFORT_SCALE, type ReasoningEffort, SUPPORTED_VOICE_PROVIDERS, type SpeechRequest, type SpeechResult, type SpeechUsage, type SpendInput, type SpendKind, type SpendLedger, type SpendLedgers, type SpendStore, type SpendWindow, SupportedAiKeyNames, SupportedAiModels, type TokenPairUsage, type TokenUsage$1 as TokenUsage, type TranscriptionRequest, type TranscriptionResult, type TranscriptionUsage, VOICE_MODEL_CONSTANTS, VOICE_MODEL_PRICING, VOICE_PROVIDER_API_KEY, type VoiceAgent, VoiceAgentFactory, type VoiceModelPricing, type VoiceProvider, XAI_REASONING_EFFORTS, type XaiReasoningEffort, ZodSchemaConverter, applySpend, buildGoogleTtsPrompt, calculateAnthropicCost, calculateCost, calculateDeepSeekCost, calculateGeminiSttCost, calculateGeminiTtsCost, calculateGoogleCost, calculateGrokCost, calculateKimiCost, calculateMetaCost, calculateMistralCost, calculateModelCost, calculateOpenAICost, calculateOpenAiSttCost, calculateOpenAiTtsCost, clampReasoningEffort, cleanResponse, createCatalog, createVoiceAgent, describeEmptyTtsResponse, evaluateBudget, extractAnthropicTokenUsage, extractTokenUsageFromResponse$2 as extractAnthropicTokenUsageFromResponse, extractDeepSeekTokenUsage, extractTokenUsageFromResponse$6 as extractDeepSeekTokenUsageFromResponse, extractFirstJsonObject, extractGoogleTokenUsage, extractTokenUsageFromResponse$1 as extractGoogleTokenUsageFromResponse, extractGrokTokenUsage, extractTokenUsageFromResponse$4 as extractGrokTokenUsageFromResponse, extractKimiTokenUsage, extractTokenUsageFromResponse$5 as extractKimiTokenUsageFromResponse, extractMetaTokenUsage, extractTokenUsageFromResponse$3 as extractMetaTokenUsageFromResponse, extractMistralTokenUsage, extractTokenUsageFromResponse as extractMistralTokenUsageFromResponse, extractOpenAITokenUsage, extractTokenUsageFromResponse$7 as extractOpenAITokenUsageFromResponse, extractTokenUsage, extractUsageAndCalculateCost, firstRefusal, generateGoogleTtsAudio, generateOpenAiTtsAudio, generateSchemaInstructions, getModelConfigByApiName, getModelDisplayName, getModelProviderName, getModelTags, getProviderSignatureFields, isBudgetExceededError, isHybridThinkingModel, isInPeakWindow, isPeakBilling, isWeekendAt, ledgerSpend, logger, mergeThinking, modelHasTag, modelIsFast, needsPromptBasedSchema, parseAndValidateLlmJson, pcmToWav, periodEnd, periodKey, safeValidateResponse, setBeforeAskHook, setLlmLogger, stableHashHex, stripInlineThinking, supportsNativeJsonSchema, toAnthropicEffort, toDeepSeekEffort, toFuguEffort, toGeminiEffort, toGlmEffort, toKimiEffort, toMetaEffort, toMistralEffort, toOpenAIEffort, toXaiEffort, transcribeWithGemini, transcribeWithOpenAi, validateResponse };
+export { type AIMessage, ANTHROPIC_REASONING_EFFORTS, API_KEY_CONSTANTS, AbstractAgent, type AgentActivityData, AgentFactory, type AgentLoggingConfig, type AnthropicReasoningEffort, type AnthropicTokenUsage, type ApiKeyMap, type BeforeAskHook, BotResponseError, BudgetController, type BudgetControllerOptions, BudgetExceededError, type BudgetLimit, type BudgetVerdict, CACHE_TIER_MARKER, ClaudeAgent, type CostCalculationOptions, DEEPSEEK_PEAK_SCHEDULE, DEEPSEEK_REASONING_EFFORTS, DEFAULT_LOGGING_CONFIG, DEFAULT_MAX_OUTPUT_TOKENS, type DeepSeekReasoningEffort, type DeepSeekTokenUsage, DeepSeekV2Agent, FUGU_REASONING_EFFORTS, FuguAgent, type FuguReasoningEffort, GEMINI_AUDIO_TOKENS_PER_SECOND, GEMINI_REASONING_EFFORTS, GLM_REASONING_EFFORTS, type GeminiReasoningEffort, GlmAgent, type GlmReasoningEffort, GoogleAgent, type GoogleSttOptions, type GoogleSttResult, type GoogleTokenUsage, type GoogleTtsAudioOptions, type GoogleTtsResult, GoogleVoiceAgent, Gpt5Agent, GrokAgent, type GrokTokenUsage, InMemorySpendStore, type JsonSchemaOptions, KIMI_REASONING_EFFORTS, KimiAgent, type KimiReasoningEffort, type KimiTokenUsage, type LLMModel, LLM_CONSTANTS, type LlmLogger, type LoggingConfig, MESSAGE_ROLE, META_REASONING_EFFORTS, MISTRAL_REASONING_EFFORTS, MODEL_PRICING, MetaAgent, type MetaReasoningEffort, type MetaTokenUsage, MiniMaxAgent, MistralAgent, type MistralReasoningEffort, type MistralTokenUsage, type Modality, ModelAuthenticationError, type ModelConfig, ModelError, ModelInvalidResponseError, ModelOverloadError, type ModelPricing, ModelQuotaExceededError, ModelRateLimitError, ModelRefusalError, type ModelTag, ModelUnavailableError, OPENAI_REASONING_EFFORTS, type OpenAIReasoningEffort, type OpenAITokenUsage, type OpenAiSttOptions, type OpenAiSttResult, type OpenAiTtsAudioOptions, type OpenAiTtsVoice, OpenAiVoiceAgent, type PeakPricing, type PricingUnit, type ProviderSchema, type TokenUsage as ProviderTokenUsage, type ProviderType, QwenAgent, REASONING_EFFORT_SCALE, type ReasoningEffort, SUPPORTED_VOICE_PROVIDERS, type SpeechRequest, type SpeechResult, type SpeechUsage, type SpendInput, type SpendKind, type SpendLedger, type SpendLedgers, type SpendStore, type SpendWindow, SupportedAiKeyNames, SupportedAiModels, type TokenPairUsage, type TokenUsage$1 as TokenUsage, type TranscriptionRequest, type TranscriptionResult, type TranscriptionUsage, VOICE_MODEL_CONSTANTS, VOICE_MODEL_PRICING, VOICE_PROVIDER_API_KEY, type VoiceAgent, VoiceAgentFactory, type VoiceModelPricing, type VoiceProvider, XAI_REASONING_EFFORTS, type XaiReasoningEffort, ZodSchemaConverter, applySpend, calculateAnthropicCost, calculateCost, calculateDeepSeekCost, calculateGeminiSttCost, calculateGeminiTtsCost, calculateGoogleCost, calculateGrokCost, calculateKimiCost, calculateMetaCost, calculateMistralCost, calculateModelCost, calculateOpenAICost, calculateOpenAiSttCost, calculateOpenAiTtsCost, clampReasoningEffort, cleanResponse, createCatalog, createVoiceAgent, describeEmptyTtsResponse, evaluateBudget, extractAnthropicTokenUsage, extractTokenUsageFromResponse$2 as extractAnthropicTokenUsageFromResponse, extractDeepSeekTokenUsage, extractTokenUsageFromResponse$6 as extractDeepSeekTokenUsageFromResponse, extractFirstJsonObject, extractGoogleTokenUsage, extractTokenUsageFromResponse$1 as extractGoogleTokenUsageFromResponse, extractGrokTokenUsage, extractTokenUsageFromResponse$4 as extractGrokTokenUsageFromResponse, extractKimiTokenUsage, extractTokenUsageFromResponse$5 as extractKimiTokenUsageFromResponse, extractMetaTokenUsage, extractTokenUsageFromResponse$3 as extractMetaTokenUsageFromResponse, extractMistralTokenUsage, extractTokenUsageFromResponse as extractMistralTokenUsageFromResponse, extractOpenAITokenUsage, extractTokenUsageFromResponse$7 as extractOpenAITokenUsageFromResponse, extractTokenUsage, extractUsageAndCalculateCost, firstRefusal, generateGoogleTtsAudio, generateOpenAiTtsAudio, generateSchemaInstructions, getModelConfigByApiName, getModelDisplayName, getModelProviderName, getModelTags, getProviderSignatureFields, isBudgetExceededError, isHybridThinkingModel, isInPeakWindow, isPeakBilling, isSafetyBlocked, isWeekendAt, ledgerSpend, logger, mergeThinking, modelHasTag, modelIsFast, needsPromptBasedSchema, normalizeTtsStyle, parseAndValidateLlmJson, pcmToWav, periodEnd, periodKey, safeValidateResponse, setBeforeAskHook, setLlmLogger, stableHashHex, stripInlineThinking, supportsNativeJsonSchema, toAnthropicEffort, toDeepSeekEffort, toFuguEffort, toGeminiEffort, toGlmEffort, toKimiEffort, toMetaEffort, toMistralEffort, toOpenAIEffort, toXaiEffort, transcribeWithGemini, transcribeWithOpenAi, validateResponse };

package/dist/index.d.ts

@@ -907,6 +907,8 @@ interface SpeechResult {
     audio: ArrayBuffer;
     costUSD: number;
     usage: SpeechUsage;
+    /** Gemini only: the styled request was safety-blocked, so the line was read without its style. */
+    styleDropped?: boolean;
 }
 interface TranscriptionRequest {
     audio: ArrayBuffer;
@@ -943,7 +945,7 @@ interface VoiceAgent {
 declare const VOICE_MODEL_CONSTANTS: {
     readonly OPENAI_TTS: "gpt-4o-mini-tts";
     readonly OPENAI_STT: "whisper-1";
-    readonly GOOGLE_TTS: "gemini-3.1-flash-tts-preview";
+    readonly GOOGLE_TTS: "gemini-3.8-flash-lite-tts";
     readonly GOOGLE_STT: "gemini-3.5-transcribe";
 };
 declare const SUPPORTED_VOICE_PROVIDERS: readonly VoiceProvider[];
@@ -1006,7 +1008,7 @@ declare class OpenAiVoiceAgent implements VoiceAgent {
 declare class GoogleVoiceAgent implements VoiceAgent {
     private readonly apiKey;
     readonly provider: "google";
-    readonly ttsModel: "gemini-3.1-flash-tts-preview";
+    readonly ttsModel: "gemini-3.8-flash-lite-tts";
     readonly sttModel: "gemini-3.5-transcribe";
     constructor(apiKey: string);
     speak(request: SpeechRequest): Promise<SpeechResult>;
@@ -1042,7 +1044,7 @@ declare function transcribeWithOpenAi(audioBuffer: ArrayBuffer, apiKey: string,
 interface GoogleTtsAudioOptions {
     /** e.g. "Kore", "Puck" */
     voiceName: string;
-    /** "mysteriously", "excitedly", or a longer direction */
+    /** "mysteriously", "excitedly", or a longer direction; sent as speechMetadata.style */
     voiceStyle?: string;
 }
 interface GoogleTtsResult {
@@ -1053,15 +1055,16 @@ interface GoogleTtsResult {
         inputTokens: number;
         outputTokens: number;
     };
+    /** True when the styled request was safety-blocked and the line was read without its style. */
+    styleDropped?: boolean;
 }
 /**
- * Gemini TTS has no instruction field: delivery is directed in the text itself
- * ("Say cheerfully: Have a wonderful day!" in the docs). A short style (1-3
- * words) becomes that "Say X:" prefix; a longer direction is used as written,
- * ending in the colon that separates it from the line to read. The same style
- * value feeds OpenAI's `instructions`, so one field serves both providers.
+ * Gemini 3.8 TTS reads the part text as a verbatim transcript: an inline
+ * "Say cheerfully:" prefix (the 3.1 convention) may be spoken aloud. Delivery
+ * direction goes in the part's `speechMetadata.style` instead, short adverb or
+ * longer sentence alike. Trailing punctuation is trimmed; blank means none.
  */
-declare function buildGoogleTtsPrompt(text: string, voiceStyle?: string): string;
+declare function normalizeTtsStyle(voiceStyle?: string): string | undefined;
 /** Wraps raw 16-bit mono PCM in a WAV header. */
 declare function pcmToWav(pcmData: Uint8Array, sampleRate?: number): ArrayBuffer;
 /**
@@ -1071,12 +1074,22 @@ declare function pcmToWav(pcmData: Uint8Array, sampleRate?: number): ArrayBuffer
  * difference between a diagnosable log line and "no audio data".
  */
 declare function describeEmptyTtsResponse(response: any): string;
+/**
+ * A safety block on the prompt or the candidate. The 3.1 preview model
+ * false-positived on volume directions ("Say quietly: <harmless line>" blocked
+ * 4 of 6 on 2026-09-24); 3.8 Flash-Lite with the style in speechMetadata passed
+ * the same lines 9 of 9, but the fallback below stays as a net.
+ */
+declare function isSafetyBlocked(response: any): boolean;
 /**
  * Core Gemini TTS call: text + API key in, WAV + token usage out.
  *
- * Uses generateContent rather than the newer Interactions API the docs show:
- * both serve the 3.1 TTS model (verified 2026-09-05), and this one is typed in
- * the SDK and reports usageMetadata, which billing needs.
+ * 3.8 TTS returns WAV with a RIFF header by default (3.1 returned headerless
+ * audio/l16); raw PCM is still wrapped if a response ever carries it.
+ *
+ * When a styled request comes back safety-blocked, the line is read once more
+ * without the style: a block on the direction should not silence the line. An
+ * unstyled block is not repeated.
  */
 declare function generateGoogleTtsAudio(text: string, apiKey: string, options: GoogleTtsAudioOptions): Promise<GoogleTtsResult>;
 
@@ -1791,4 +1804,4 @@ declare class MiniMaxAgent extends AbstractAgent {
     doAskText(messages: AIMessage[]): Promise<[string, string, TokenUsage$1?, string?]>;
 }
 
-export { type AIMessage, ANTHROPIC_REASONING_EFFORTS, API_KEY_CONSTANTS, AbstractAgent, type AgentActivityData, AgentFactory, type AgentLoggingConfig, type AnthropicReasoningEffort, type AnthropicTokenUsage, type ApiKeyMap, type BeforeAskHook, BotResponseError, BudgetController, type BudgetControllerOptions, BudgetExceededError, type BudgetLimit, type BudgetVerdict, CACHE_TIER_MARKER, ClaudeAgent, type CostCalculationOptions, DEEPSEEK_PEAK_SCHEDULE, DEEPSEEK_REASONING_EFFORTS, DEFAULT_LOGGING_CONFIG, DEFAULT_MAX_OUTPUT_TOKENS, type DeepSeekReasoningEffort, type DeepSeekTokenUsage, DeepSeekV2Agent, FUGU_REASONING_EFFORTS, FuguAgent, type FuguReasoningEffort, GEMINI_AUDIO_TOKENS_PER_SECOND, GEMINI_REASONING_EFFORTS, GLM_REASONING_EFFORTS, type GeminiReasoningEffort, GlmAgent, type GlmReasoningEffort, GoogleAgent, type GoogleSttOptions, type GoogleSttResult, type GoogleTokenUsage, type GoogleTtsAudioOptions, type GoogleTtsResult, GoogleVoiceAgent, Gpt5Agent, GrokAgent, type GrokTokenUsage, InMemorySpendStore, type JsonSchemaOptions, KIMI_REASONING_EFFORTS, KimiAgent, type KimiReasoningEffort, type KimiTokenUsage, type LLMModel, LLM_CONSTANTS, type LlmLogger, type LoggingConfig, MESSAGE_ROLE, META_REASONING_EFFORTS, MISTRAL_REASONING_EFFORTS, MODEL_PRICING, MetaAgent, type MetaReasoningEffort, type MetaTokenUsage, MiniMaxAgent, MistralAgent, type MistralReasoningEffort, type MistralTokenUsage, type Modality, ModelAuthenticationError, type ModelConfig, ModelError, ModelInvalidResponseError, ModelOverloadError, type ModelPricing, ModelQuotaExceededError, ModelRateLimitError, ModelRefusalError, type ModelTag, ModelUnavailableError, OPENAI_REASONING_EFFORTS, type OpenAIReasoningEffort, type OpenAITokenUsage, type OpenAiSttOptions, type OpenAiSttResult, type OpenAiTtsAudioOptions, type OpenAiTtsVoice, OpenAiVoiceAgent, type PeakPricing, type PricingUnit, type ProviderSchema, type TokenUsage as ProviderTokenUsage, type ProviderType, QwenAgent, REASONING_EFFORT_SCALE, type ReasoningEffort, SUPPORTED_VOICE_PROVIDERS, type SpeechRequest, type SpeechResult, type SpeechUsage, type SpendInput, type SpendKind, type SpendLedger, type SpendLedgers, type SpendStore, type SpendWindow, SupportedAiKeyNames, SupportedAiModels, type TokenPairUsage, type TokenUsage$1 as TokenUsage, type TranscriptionRequest, type TranscriptionResult, type TranscriptionUsage, VOICE_MODEL_CONSTANTS, VOICE_MODEL_PRICING, VOICE_PROVIDER_API_KEY, type VoiceAgent, VoiceAgentFactory, type VoiceModelPricing, type VoiceProvider, XAI_REASONING_EFFORTS, type XaiReasoningEffort, ZodSchemaConverter, applySpend, buildGoogleTtsPrompt, calculateAnthropicCost, calculateCost, calculateDeepSeekCost, calculateGeminiSttCost, calculateGeminiTtsCost, calculateGoogleCost, calculateGrokCost, calculateKimiCost, calculateMetaCost, calculateMistralCost, calculateModelCost, calculateOpenAICost, calculateOpenAiSttCost, calculateOpenAiTtsCost, clampReasoningEffort, cleanResponse, createCatalog, createVoiceAgent, describeEmptyTtsResponse, evaluateBudget, extractAnthropicTokenUsage, extractTokenUsageFromResponse$2 as extractAnthropicTokenUsageFromResponse, extractDeepSeekTokenUsage, extractTokenUsageFromResponse$6 as extractDeepSeekTokenUsageFromResponse, extractFirstJsonObject, extractGoogleTokenUsage, extractTokenUsageFromResponse$1 as extractGoogleTokenUsageFromResponse, extractGrokTokenUsage, extractTokenUsageFromResponse$4 as extractGrokTokenUsageFromResponse, extractKimiTokenUsage, extractTokenUsageFromResponse$5 as extractKimiTokenUsageFromResponse, extractMetaTokenUsage, extractTokenUsageFromResponse$3 as extractMetaTokenUsageFromResponse, extractMistralTokenUsage, extractTokenUsageFromResponse as extractMistralTokenUsageFromResponse, extractOpenAITokenUsage, extractTokenUsageFromResponse$7 as extractOpenAITokenUsageFromResponse, extractTokenUsage, extractUsageAndCalculateCost, firstRefusal, generateGoogleTtsAudio, generateOpenAiTtsAudio, generateSchemaInstructions, getModelConfigByApiName, getModelDisplayName, getModelProviderName, getModelTags, getProviderSignatureFields, isBudgetExceededError, isHybridThinkingModel, isInPeakWindow, isPeakBilling, isWeekendAt, ledgerSpend, logger, mergeThinking, modelHasTag, modelIsFast, needsPromptBasedSchema, parseAndValidateLlmJson, pcmToWav, periodEnd, periodKey, safeValidateResponse, setBeforeAskHook, setLlmLogger, stableHashHex, stripInlineThinking, supportsNativeJsonSchema, toAnthropicEffort, toDeepSeekEffort, toFuguEffort, toGeminiEffort, toGlmEffort, toKimiEffort, toMetaEffort, toMistralEffort, toOpenAIEffort, toXaiEffort, transcribeWithGemini, transcribeWithOpenAi, validateResponse };
+export { type AIMessage, ANTHROPIC_REASONING_EFFORTS, API_KEY_CONSTANTS, AbstractAgent, type AgentActivityData, AgentFactory, type AgentLoggingConfig, type AnthropicReasoningEffort, type AnthropicTokenUsage, type ApiKeyMap, type BeforeAskHook, BotResponseError, BudgetController, type BudgetControllerOptions, BudgetExceededError, type BudgetLimit, type BudgetVerdict, CACHE_TIER_MARKER, ClaudeAgent, type CostCalculationOptions, DEEPSEEK_PEAK_SCHEDULE, DEEPSEEK_REASONING_EFFORTS, DEFAULT_LOGGING_CONFIG, DEFAULT_MAX_OUTPUT_TOKENS, type DeepSeekReasoningEffort, type DeepSeekTokenUsage, DeepSeekV2Agent, FUGU_REASONING_EFFORTS, FuguAgent, type FuguReasoningEffort, GEMINI_AUDIO_TOKENS_PER_SECOND, GEMINI_REASONING_EFFORTS, GLM_REASONING_EFFORTS, type GeminiReasoningEffort, GlmAgent, type GlmReasoningEffort, GoogleAgent, type GoogleSttOptions, type GoogleSttResult, type GoogleTokenUsage, type GoogleTtsAudioOptions, type GoogleTtsResult, GoogleVoiceAgent, Gpt5Agent, GrokAgent, type GrokTokenUsage, InMemorySpendStore, type JsonSchemaOptions, KIMI_REASONING_EFFORTS, KimiAgent, type KimiReasoningEffort, type KimiTokenUsage, type LLMModel, LLM_CONSTANTS, type LlmLogger, type LoggingConfig, MESSAGE_ROLE, META_REASONING_EFFORTS, MISTRAL_REASONING_EFFORTS, MODEL_PRICING, MetaAgent, type MetaReasoningEffort, type MetaTokenUsage, MiniMaxAgent, MistralAgent, type MistralReasoningEffort, type MistralTokenUsage, type Modality, ModelAuthenticationError, type ModelConfig, ModelError, ModelInvalidResponseError, ModelOverloadError, type ModelPricing, ModelQuotaExceededError, ModelRateLimitError, ModelRefusalError, type ModelTag, ModelUnavailableError, OPENAI_REASONING_EFFORTS, type OpenAIReasoningEffort, type OpenAITokenUsage, type OpenAiSttOptions, type OpenAiSttResult, type OpenAiTtsAudioOptions, type OpenAiTtsVoice, OpenAiVoiceAgent, type PeakPricing, type PricingUnit, type ProviderSchema, type TokenUsage as ProviderTokenUsage, type ProviderType, QwenAgent, REASONING_EFFORT_SCALE, type ReasoningEffort, SUPPORTED_VOICE_PROVIDERS, type SpeechRequest, type SpeechResult, type SpeechUsage, type SpendInput, type SpendKind, type SpendLedger, type SpendLedgers, type SpendStore, type SpendWindow, SupportedAiKeyNames, SupportedAiModels, type TokenPairUsage, type TokenUsage$1 as TokenUsage, type TranscriptionRequest, type TranscriptionResult, type TranscriptionUsage, VOICE_MODEL_CONSTANTS, VOICE_MODEL_PRICING, VOICE_PROVIDER_API_KEY, type VoiceAgent, VoiceAgentFactory, type VoiceModelPricing, type VoiceProvider, XAI_REASONING_EFFORTS, type XaiReasoningEffort, ZodSchemaConverter, applySpend, calculateAnthropicCost, calculateCost, calculateDeepSeekCost, calculateGeminiSttCost, calculateGeminiTtsCost, calculateGoogleCost, calculateGrokCost, calculateKimiCost, calculateMetaCost, calculateMistralCost, calculateModelCost, calculateOpenAICost, calculateOpenAiSttCost, calculateOpenAiTtsCost, clampReasoningEffort, cleanResponse, createCatalog, createVoiceAgent, describeEmptyTtsResponse, evaluateBudget, extractAnthropicTokenUsage, extractTokenUsageFromResponse$2 as extractAnthropicTokenUsageFromResponse, extractDeepSeekTokenUsage, extractTokenUsageFromResponse$6 as extractDeepSeekTokenUsageFromResponse, extractFirstJsonObject, extractGoogleTokenUsage, extractTokenUsageFromResponse$1 as extractGoogleTokenUsageFromResponse, extractGrokTokenUsage, extractTokenUsageFromResponse$4 as extractGrokTokenUsageFromResponse, extractKimiTokenUsage, extractTokenUsageFromResponse$5 as extractKimiTokenUsageFromResponse, extractMetaTokenUsage, extractTokenUsageFromResponse$3 as extractMetaTokenUsageFromResponse, extractMistralTokenUsage, extractTokenUsageFromResponse as extractMistralTokenUsageFromResponse, extractOpenAITokenUsage, extractTokenUsageFromResponse$7 as extractOpenAITokenUsageFromResponse, extractTokenUsage, extractUsageAndCalculateCost, firstRefusal, generateGoogleTtsAudio, generateOpenAiTtsAudio, generateSchemaInstructions, getModelConfigByApiName, getModelDisplayName, getModelProviderName, getModelTags, getProviderSignatureFields, isBudgetExceededError, isHybridThinkingModel, isInPeakWindow, isPeakBilling, isSafetyBlocked, isWeekendAt, ledgerSpend, logger, mergeThinking, modelHasTag, modelIsFast, needsPromptBasedSchema, normalizeTtsStyle, parseAndValidateLlmJson, pcmToWav, periodEnd, periodKey, safeValidateResponse, setBeforeAskHook, setLlmLogger, stableHashHex, stripInlineThinking, supportsNativeJsonSchema, toAnthropicEffort, toDeepSeekEffort, toFuguEffort, toGeminiEffort, toGlmEffort, toKimiEffort, toMetaEffort, toMistralEffort, toOpenAIEffort, toXaiEffort, transcribeWithGemini, transcribeWithOpenAi, validateResponse };

package/dist/index.js

@@ -87,7 +87,6 @@ __export(src_exports, {
   XAI_REASONING_EFFORTS: () => XAI_REASONING_EFFORTS,
   ZodSchemaConverter: () => ZodSchemaConverter,
   applySpend: () => applySpend,
-  buildGoogleTtsPrompt: () => buildGoogleTtsPrompt,
   calculateAnthropicCost: () => calculateAnthropicCost,
   calculateCost: () => calculateCost,
   calculateDeepSeekCost: () => calculateDeepSeekCost,
@@ -140,6 +139,7 @@ __export(src_exports, {
   isHybridThinkingModel: () => isHybridThinkingModel,
   isInPeakWindow: () => isInPeakWindow,
   isPeakBilling: () => isPeakBilling,
+  isSafetyBlocked: () => isSafetyBlocked,
   isWeekendAt: () => isWeekendAt,
   ledgerSpend: () => ledgerSpend,
   logger: () => logger,
@@ -147,6 +147,7 @@ __export(src_exports, {
   modelHasTag: () => modelHasTag,
   modelIsFast: () => modelIsFast,
   needsPromptBasedSchema: () => needsPromptBasedSchema,
+  normalizeTtsStyle: () => normalizeTtsStyle,
   parseAndValidateLlmJson: () => parseAndValidateLlmJson,
   pcmToWav: () => pcmToWav,
   periodEnd: () => periodEnd,
@@ -1240,10 +1241,15 @@ var SupportedAiModels = {
     // Live 2026-08-30 (one sample each): day-2 vote 11.8s, 15-character story 56.2s.
     tags: ["cheap"]
   },
-  // Sakana Fugu models — OpenAI-compatible. They reason internally (and bill it as
-  // "orchestration" tokens), but never surface reasoning to us: responses come back with
-  // reasoning_tokens: 0 and no reasoning_content. So hasThinking is false — there's no
-  // thinking content to show and no user-facing thinking toggle. Single entry per model.
+  // Sakana Fugu models — OpenAI-compatible. They reason internally, but Chat Completions
+  // (what FuguAgent uses) never returns the text: reasoning_tokens are reported and billed,
+  // there is no reasoning_content. So hasThinking is false. Single entry per model.
+  // The Messages and Responses APIs DO return reasoning (thinking blocks with an empty
+  // signature / reasoning items, some with encrypted_content), but probed 2026-09-24 on
+  // fugu-max it isn't worth switching for, since we keep reasoning only to replay it: the
+  // returned text is often a vague status summary or empty rather than the actual
+  // reasoning, and replayed reasoning reaches the model distorted (a planted "4817" came
+  // back as "4829" on both APIs, presumably rewritten by the orchestrator).
   //
   // Base `fugu` was RETIRED 2026-08-04. It was carried as a cheap everyday option at an
   // assumed $1/$3, but reconciling token logs against the Sakana balance showed it actually
@@ -1926,8 +1932,9 @@ function extractTokenUsageFromResponse8(response) {
 var VOICE_MODEL_CONSTANTS = {
   OPENAI_TTS: "gpt-4o-mini-tts",
   OPENAI_STT: "whisper-1",
-  // ai.google.dev/gemini-api/docs/speech-generation
-  GOOGLE_TTS: "gemini-3.1-flash-tts-preview",
+  // ai.google.dev/gemini-api/docs/speech-generation — Google's named replacement
+  // for gemini-3.1-flash-tts-preview (2026-09-24: ~6 s vs ~15 s per line, ~4x cheaper)
+  GOOGLE_TTS: "gemini-3.8-flash-lite-tts",
   // ai.google.dev/gemini-api/docs/transcribe — Interactions API only (see google-stt.ts)
   GOOGLE_STT: "gemini-3.5-transcribe"
 };
@@ -1939,9 +1946,9 @@ var VOICE_PROVIDER_API_KEY = {
 var VOICE_MODEL_PRICING = {
   [VOICE_MODEL_CONSTANTS.OPENAI_TTS]: { pricePerMillionCharacters: 15 },
   [VOICE_MODEL_CONSTANTS.OPENAI_STT]: { pricePerMinute: 6e-3 },
-  // Measured 2026-09-05: ~32 audio tokens per second of speech, so a 15-second
-  // line is ~$0.01 — about 3-5x an OpenAI line of the same length.
-  [VOICE_MODEL_CONSTANTS.GOOGLE_TTS]: { textInputPricePerM: 1, audioOutputPricePerM: 20 },
+  // ~32 audio tokens per second of speech, so a 15-second line is ~$0.003.
+  // Google doubles both rates to $1 / $12 on 2027-01-01 — update then.
+  [VOICE_MODEL_CONSTANTS.GOOGLE_TTS]: { textInputPricePerM: 0.5, audioOutputPricePerM: 6 },
   // ~25 audio tokens per second in, ~175 text tokens per minute out: ≈ $0.005/min.
   [VOICE_MODEL_CONSTANTS.GOOGLE_STT]: { audioInputPricePerM: 2, textOutputPricePerM: 12 }
 };
@@ -2048,12 +2055,10 @@ var import_genai = require("@google/genai");
 var AUDIO_TOKENS_PER_SECOND = 32;
 var SAMPLE_RATE = 24e3;
 var PCM_BYTES_PER_SECOND = SAMPLE_RATE * 2;
-function buildGoogleTtsPrompt(text, voiceStyle) {
+var WAV_HEADER_BYTES = 44;
+function normalizeTtsStyle(voiceStyle) {
   const style = voiceStyle?.trim().replace(/[:.!,;\s]+$/, "");
-  if (!style) return text;
-  const isShort = style.split(/\s+/).length <= 3 && !/[.!?,;]/.test(style);
-  return isShort ? `Say ${style}: ${text}` : `${style}:
-${text}`;
+  return style || void 0;
 }
 function pcmToWav(pcmData, sampleRate = SAMPLE_RATE) {
   const numChannels = 1;
@@ -2097,28 +2102,51 @@ function describeEmptyTtsResponse(response) {
   if (!parts.length) bits.push("no parts");
   return bits.join(", ");
 }
+function isSafetyBlocked(response) {
+  const finish = response?.candidates?.[0]?.finishReason;
+  return finish === "SAFETY" || finish === "PROHIBITED_CONTENT" || !!response?.promptFeedback?.blockReason;
+}
+function findAudioPart(response) {
+  const parts = response?.candidates?.[0]?.content?.parts ?? [];
+  const inline = parts.find((part) => part.inlineData?.mimeType?.startsWith("audio/") && part.inlineData?.data)?.inlineData;
+  return inline ? { data: inline.data, mimeType: inline.mimeType } : void 0;
+}
 async function generateGoogleTtsAudio(text, apiKey, options) {
   const client = new import_genai.GoogleGenAI({ apiKey });
-  const response = await client.models.generateContent({
+  const request = (style2) => client.models.generateContent({
     model: VOICE_MODEL_CONSTANTS.GOOGLE_TTS,
-    contents: [{ parts: [{ text: buildGoogleTtsPrompt(text, options.voiceStyle) }] }],
+    contents: [{ parts: [style2 ? { text, speechMetadata: { style: style2 } } : { text }] }],
     config: {
       responseModalities: ["AUDIO"],
       speechConfig: { voiceConfig: { prebuiltVoiceConfig: { voiceName: options.voiceName } } }
     }
   });
-  const candidate = response.candidates?.[0];
-  const parts = candidate?.content?.parts ?? [];
-  const audioPart = parts.find((part) => part.inlineData?.mimeType?.startsWith("audio/"));
-  if (!audioPart?.inlineData?.data) {
-    throw new Error(`No audio data in Google TTS response (${describeEmptyTtsResponse(response)})`);
-  }
-  const pcmData = new Uint8Array(Buffer.from(audioPart.inlineData.data, "base64"));
+  const style = normalizeTtsStyle(options.voiceStyle);
+  let response = await request(style);
+  let styleDropped = false;
+  let blockedInputTokens = 0;
+  if (!findAudioPart(response) && style && isSafetyBlocked(response)) {
+    blockedInputTokens = response?.usageMetadata?.promptTokenCount ?? 0;
+    response = await request(void 0);
+    styleDropped = true;
+  }
+  const audioPart = findAudioPart(response);
+  if (!audioPart) {
+    throw new Error(`No audio data in Google TTS response (${describeEmptyTtsResponse(response)}${styleDropped ? ", also without the style" : ""})`);
+  }
+  const bytes = new Uint8Array(Buffer.from(audioPart.data, "base64"));
+  const isWav = /^audio\/(x-)?wav/i.test(audioPart.mimeType);
+  const audio = isWav ? bytes.buffer.slice(bytes.byteOffset, bytes.byteOffset + bytes.byteLength) : pcmToWav(bytes);
+  const pcmBytes = isWav ? Math.max(0, bytes.length - WAV_HEADER_BYTES) : bytes.length;
   const usageMetadata = response.usageMetadata ?? {};
-  const inputTokens = usageMetadata.promptTokenCount ?? 0;
+  const inputTokens = (usageMetadata.promptTokenCount ?? 0) + blockedInputTokens;
   const reportedOutput = usageMetadata.candidatesTokenCount;
-  const outputTokens = reportedOutput && reportedOutput > 0 ? reportedOutput : Math.ceil(pcmData.length / PCM_BYTES_PER_SECOND * AUDIO_TOKENS_PER_SECOND);
-  return { audio: pcmToWav(pcmData), usage: { inputTokens, outputTokens } };
+  const outputTokens = reportedOutput && reportedOutput > 0 ? reportedOutput : Math.ceil(pcmBytes / PCM_BYTES_PER_SECOND * AUDIO_TOKENS_PER_SECOND);
+  return {
+    audio,
+    usage: { inputTokens, outputTokens },
+    ...styleDropped ? { styleDropped } : {}
+  };
 }
 
 // src/voice/google-stt.ts
@@ -2149,11 +2177,11 @@ var GoogleVoiceAgent = class {
   ttsModel = VOICE_MODEL_CONSTANTS.GOOGLE_TTS;
   sttModel = VOICE_MODEL_CONSTANTS.GOOGLE_STT;
   async speak(request) {
-    const { audio, usage } = await generateGoogleTtsAudio(request.text, this.apiKey, {
+    const { audio, usage, styleDropped } = await generateGoogleTtsAudio(request.text, this.apiKey, {
       voiceName: request.voice,
       voiceStyle: request.voiceStyle
     });
-    return { audio, costUSD: calculateGeminiTtsCost(usage), usage };
+    return { audio, costUSD: calculateGeminiTtsCost(usage), usage, ...styleDropped ? { styleDropped } : {} };
   }
   async transcribe(request) {
     const { text, durationSeconds, usage } = await transcribeWithGemini(request.audio, this.apiKey, { mimeType: request.mimeType });
@@ -4909,7 +4937,11 @@ ${schemaDescription}`;
         const params = {
           ...this.defaultParams,
           messages: openAIMessages,
-          ...this.thinkingParams()
+          ...this.thinkingParams(),
+          response_format: {
+            type: "json_schema",
+            json_schema: ZodSchemaConverter.toOpenAIJsonSchema(zodSchema, "response_schema")
+          }
         };
         completion = await this.client.chat.completions.create(params);
       } catch (apiError) {
@@ -5518,7 +5550,6 @@ var AgentFactory = class {
   XAI_REASONING_EFFORTS,
   ZodSchemaConverter,
   applySpend,
-  buildGoogleTtsPrompt,
   calculateAnthropicCost,
   calculateCost,
   calculateDeepSeekCost,
@@ -5571,6 +5602,7 @@ var AgentFactory = class {
   isHybridThinkingModel,
   isInPeakWindow,
   isPeakBilling,
+  isSafetyBlocked,
   isWeekendAt,
   ledgerSpend,
   logger,
@@ -5578,6 +5610,7 @@ var AgentFactory = class {
   modelHasTag,
   modelIsFast,
   needsPromptBasedSchema,
+  normalizeTtsStyle,
   parseAndValidateLlmJson,
   pcmToWav,
   periodEnd,