@hiper2d/ai-agents 0.34.1 → 0.36.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/dist/index.d.mts CHANGED
@@ -378,6 +378,18 @@ declare const API_KEY_CONSTANTS: {
378
378
  readonly MINIMAX: "MINIMAX_API_KEY";
379
379
  readonly META: "META_API_KEY";
380
380
  };
381
+ /**
382
+ * Providers whose API ENFORCES a JSON schema on the answer (askWithZodSchema sends the
383
+ * schema natively and the provider constrains the output to it). The rest only get the
384
+ * schema described in the prompt: Grok, GLM and Fugu run in "JSON object" mode (valid JSON,
385
+ * any shape), DeepSeek and MiniMax have no JSON mode on these calls at all. Meta is listed
386
+ * although its json_schema runs with strict: false (the shape is guided, not guaranteed).
387
+ * Hosts use this to decide where a schema can replace a plain-text ask: on most enforcing
388
+ * providers the prompt cache is keyed on the output format, so a host that mixes text and
389
+ * schema calls loses the cache between them (measured 2026-11-12 on OpenAI, Mistral and
390
+ * Gemini: 0% cached on a text reply → schema vote, 77-83% with one schema on both).
391
+ */
392
+ declare const SCHEMA_ENFORCING_API_KEYS: ReadonlySet<string>;
381
393
  declare const SupportedAiKeyNames: Record<string, string>;
382
394
  declare const LLM_CONSTANTS: {
383
395
  CLAUDE_FABLE: string;
@@ -551,6 +563,8 @@ declare function getProviderSignatureFields(aiType: string, signature?: string):
551
563
  metaEncryptedReasoning?: string;
552
564
  openaiEncryptedReasoning?: string;
553
565
  };
566
+ /** Whether the provider behind a catalog model enforces a JSON schema on the answer (see SCHEMA_ENFORCING_API_KEYS). Unknown ids: false. */
567
+ declare function enforcesSchema(modelId: string): boolean;
554
568
 
555
569
  /**
556
570
  * Per-provider reasoning-effort vocabularies.
@@ -565,7 +579,7 @@ declare function getProviderSignatureFields(aiType: string, signature?: string):
565
579
  * Verified against provider docs 2026-08-30:
566
580
  * - OpenAI GPT-5.x: minimal|low|medium|high|xhigh
567
581
  * - Anthropic adaptive thinking (Fable 5 / Opus 4.8 / Sonnet 5): low|medium|high|xhigh|max
568
- * - Gemini 3.x thinkingLevel: minimal|low|medium|high (3.1 Pro and 3.7 Flash reject minimal)
582
+ * - Gemini 3.x thinkingLevel: minimal|low|medium|high (3.1 Pro, 4 Argon and 4 Hydrogen reject minimal)
569
583
  * - Z.AI GLM-5.3 / 5.3-Flash: low|high|max only
570
584
  * - DeepSeek V4: low|high|max (the API itself aliases medium → high)
571
585
  * - Sakana Fugu (verified live 2026-09-20): high|xhigh|max on every model ("low" → 400 naming
@@ -1920,4 +1934,4 @@ declare class MiniMaxAgent extends AbstractAgent {
1920
1934
  doAskText(messages: AIMessage[]): Promise<[string, string, TokenUsage$1?, string?]>;
1921
1935
  }
1922
1936
 
1923
- 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 ClaudeCacheMode, type CostCalculationOptions, DEEPSEEK_PEAK_SCHEDULE, DEEPSEEK_REASONING_EFFORTS, DEFAULT_LOGGING_CONFIG, DEFAULT_MAX_OUTPUT_TOKENS, DEFAULT_REQUEST_TIMEOUT_MS, 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, MINIMAX_REASONING_EFFORTS, MISTRAL_REASONING_EFFORTS, MODEL_PRICING, MetaAgent, type MetaReasoningEffort, type MetaTokenUsage, MiniMaxAgent, type MiniMaxReasoningEffort, MistralAgent, type MistralReasoningEffort, type MistralTokenUsage, type Modality, ModelAuthenticationError, type ModelConfig, ModelError, ModelInvalidResponseError, ModelOverloadError, type ModelPricing, ModelQuotaExceededError, ModelRateLimitError, ModelRefusalError, type ModelTag, ModelTimeoutError, 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, type RetryHook, type RetrySettledFn, 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, setRetryHook, stableHashHex, stripInlineThinking, supportsNativeJsonSchema, toAnthropicEffort, toDeepSeekEffort, toFuguEffort, toGeminiEffort, toGlmEffort, toKimiEffort, toMetaEffort, toMiniMaxEffort, toMistralEffort, toOpenAIEffort, toXaiEffort, transcribeWithGemini, transcribeWithOpenAi, validateResponse };
1937
+ 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 ClaudeCacheMode, type CostCalculationOptions, DEEPSEEK_PEAK_SCHEDULE, DEEPSEEK_REASONING_EFFORTS, DEFAULT_LOGGING_CONFIG, DEFAULT_MAX_OUTPUT_TOKENS, DEFAULT_REQUEST_TIMEOUT_MS, 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, MINIMAX_REASONING_EFFORTS, MISTRAL_REASONING_EFFORTS, MODEL_PRICING, MetaAgent, type MetaReasoningEffort, type MetaTokenUsage, MiniMaxAgent, type MiniMaxReasoningEffort, MistralAgent, type MistralReasoningEffort, type MistralTokenUsage, type Modality, ModelAuthenticationError, type ModelConfig, ModelError, ModelInvalidResponseError, ModelOverloadError, type ModelPricing, ModelQuotaExceededError, ModelRateLimitError, ModelRefusalError, type ModelTag, ModelTimeoutError, 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, type RetryHook, type RetrySettledFn, SCHEMA_ENFORCING_API_KEYS, 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, enforcesSchema, 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, setRetryHook, stableHashHex, stripInlineThinking, supportsNativeJsonSchema, toAnthropicEffort, toDeepSeekEffort, toFuguEffort, toGeminiEffort, toGlmEffort, toKimiEffort, toMetaEffort, toMiniMaxEffort, toMistralEffort, toOpenAIEffort, toXaiEffort, transcribeWithGemini, transcribeWithOpenAi, validateResponse };
package/dist/index.d.ts CHANGED
@@ -378,6 +378,18 @@ declare const API_KEY_CONSTANTS: {
378
378
  readonly MINIMAX: "MINIMAX_API_KEY";
379
379
  readonly META: "META_API_KEY";
380
380
  };
381
+ /**
382
+ * Providers whose API ENFORCES a JSON schema on the answer (askWithZodSchema sends the
383
+ * schema natively and the provider constrains the output to it). The rest only get the
384
+ * schema described in the prompt: Grok, GLM and Fugu run in "JSON object" mode (valid JSON,
385
+ * any shape), DeepSeek and MiniMax have no JSON mode on these calls at all. Meta is listed
386
+ * although its json_schema runs with strict: false (the shape is guided, not guaranteed).
387
+ * Hosts use this to decide where a schema can replace a plain-text ask: on most enforcing
388
+ * providers the prompt cache is keyed on the output format, so a host that mixes text and
389
+ * schema calls loses the cache between them (measured 2026-11-12 on OpenAI, Mistral and
390
+ * Gemini: 0% cached on a text reply → schema vote, 77-83% with one schema on both).
391
+ */
392
+ declare const SCHEMA_ENFORCING_API_KEYS: ReadonlySet<string>;
381
393
  declare const SupportedAiKeyNames: Record<string, string>;
382
394
  declare const LLM_CONSTANTS: {
383
395
  CLAUDE_FABLE: string;
@@ -551,6 +563,8 @@ declare function getProviderSignatureFields(aiType: string, signature?: string):
551
563
  metaEncryptedReasoning?: string;
552
564
  openaiEncryptedReasoning?: string;
553
565
  };
566
+ /** Whether the provider behind a catalog model enforces a JSON schema on the answer (see SCHEMA_ENFORCING_API_KEYS). Unknown ids: false. */
567
+ declare function enforcesSchema(modelId: string): boolean;
554
568
 
555
569
  /**
556
570
  * Per-provider reasoning-effort vocabularies.
@@ -565,7 +579,7 @@ declare function getProviderSignatureFields(aiType: string, signature?: string):
565
579
  * Verified against provider docs 2026-08-30:
566
580
  * - OpenAI GPT-5.x: minimal|low|medium|high|xhigh
567
581
  * - Anthropic adaptive thinking (Fable 5 / Opus 4.8 / Sonnet 5): low|medium|high|xhigh|max
568
- * - Gemini 3.x thinkingLevel: minimal|low|medium|high (3.1 Pro and 3.7 Flash reject minimal)
582
+ * - Gemini 3.x thinkingLevel: minimal|low|medium|high (3.1 Pro, 4 Argon and 4 Hydrogen reject minimal)
569
583
  * - Z.AI GLM-5.3 / 5.3-Flash: low|high|max only
570
584
  * - DeepSeek V4: low|high|max (the API itself aliases medium → high)
571
585
  * - Sakana Fugu (verified live 2026-09-20): high|xhigh|max on every model ("low" → 400 naming
@@ -1920,4 +1934,4 @@ declare class MiniMaxAgent extends AbstractAgent {
1920
1934
  doAskText(messages: AIMessage[]): Promise<[string, string, TokenUsage$1?, string?]>;
1921
1935
  }
1922
1936
 
1923
- 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 ClaudeCacheMode, type CostCalculationOptions, DEEPSEEK_PEAK_SCHEDULE, DEEPSEEK_REASONING_EFFORTS, DEFAULT_LOGGING_CONFIG, DEFAULT_MAX_OUTPUT_TOKENS, DEFAULT_REQUEST_TIMEOUT_MS, 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, MINIMAX_REASONING_EFFORTS, MISTRAL_REASONING_EFFORTS, MODEL_PRICING, MetaAgent, type MetaReasoningEffort, type MetaTokenUsage, MiniMaxAgent, type MiniMaxReasoningEffort, MistralAgent, type MistralReasoningEffort, type MistralTokenUsage, type Modality, ModelAuthenticationError, type ModelConfig, ModelError, ModelInvalidResponseError, ModelOverloadError, type ModelPricing, ModelQuotaExceededError, ModelRateLimitError, ModelRefusalError, type ModelTag, ModelTimeoutError, 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, type RetryHook, type RetrySettledFn, 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, setRetryHook, stableHashHex, stripInlineThinking, supportsNativeJsonSchema, toAnthropicEffort, toDeepSeekEffort, toFuguEffort, toGeminiEffort, toGlmEffort, toKimiEffort, toMetaEffort, toMiniMaxEffort, toMistralEffort, toOpenAIEffort, toXaiEffort, transcribeWithGemini, transcribeWithOpenAi, validateResponse };
1937
+ 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 ClaudeCacheMode, type CostCalculationOptions, DEEPSEEK_PEAK_SCHEDULE, DEEPSEEK_REASONING_EFFORTS, DEFAULT_LOGGING_CONFIG, DEFAULT_MAX_OUTPUT_TOKENS, DEFAULT_REQUEST_TIMEOUT_MS, 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, MINIMAX_REASONING_EFFORTS, MISTRAL_REASONING_EFFORTS, MODEL_PRICING, MetaAgent, type MetaReasoningEffort, type MetaTokenUsage, MiniMaxAgent, type MiniMaxReasoningEffort, MistralAgent, type MistralReasoningEffort, type MistralTokenUsage, type Modality, ModelAuthenticationError, type ModelConfig, ModelError, ModelInvalidResponseError, ModelOverloadError, type ModelPricing, ModelQuotaExceededError, ModelRateLimitError, ModelRefusalError, type ModelTag, ModelTimeoutError, 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, type RetryHook, type RetrySettledFn, SCHEMA_ENFORCING_API_KEYS, 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, enforcesSchema, 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, setRetryHook, stableHashHex, stripInlineThinking, supportsNativeJsonSchema, toAnthropicEffort, toDeepSeekEffort, toFuguEffort, toGeminiEffort, toGlmEffort, toKimiEffort, toMetaEffort, toMiniMaxEffort, toMistralEffort, toOpenAIEffort, toXaiEffort, transcribeWithGemini, transcribeWithOpenAi, validateResponse };
package/dist/index.js CHANGED
@@ -80,6 +80,7 @@ __export(src_exports, {
80
80
  OpenAiVoiceAgent: () => OpenAiVoiceAgent,
81
81
  QwenAgent: () => QwenAgent,
82
82
  REASONING_EFFORT_SCALE: () => REASONING_EFFORT_SCALE,
83
+ SCHEMA_ENFORCING_API_KEYS: () => SCHEMA_ENFORCING_API_KEYS,
83
84
  SUPPORTED_VOICE_PROVIDERS: () => SUPPORTED_VOICE_PROVIDERS,
84
85
  SupportedAiKeyNames: () => SupportedAiKeyNames,
85
86
  SupportedAiModels: () => SupportedAiModels,
@@ -109,6 +110,7 @@ __export(src_exports, {
109
110
  createCatalog: () => createCatalog,
110
111
  createVoiceAgent: () => createVoiceAgent,
111
112
  describeEmptyTtsResponse: () => describeEmptyTtsResponse,
113
+ enforcesSchema: () => enforcesSchema,
112
114
  evaluateBudget: () => evaluateBudget,
113
115
  extractAnthropicTokenUsage: () => extractAnthropicTokenUsage,
114
116
  extractAnthropicTokenUsageFromResponse: () => extractTokenUsageFromResponse6,
@@ -988,6 +990,15 @@ var API_KEY_CONSTANTS = {
988
990
  MINIMAX: "MINIMAX_API_KEY",
989
991
  META: "META_API_KEY"
990
992
  };
993
+ var SCHEMA_ENFORCING_API_KEYS = /* @__PURE__ */ new Set([
994
+ API_KEY_CONSTANTS.OPENAI,
995
+ API_KEY_CONSTANTS.ANTHROPIC,
996
+ API_KEY_CONSTANTS.GOOGLE,
997
+ API_KEY_CONSTANTS.MISTRAL,
998
+ API_KEY_CONSTANTS.QWEN,
999
+ API_KEY_CONSTANTS.MOONSHOT,
1000
+ API_KEY_CONSTANTS.META
1001
+ ]);
991
1002
  var SupportedAiKeyNames = {
992
1003
  [API_KEY_CONSTANTS.OPENAI]: "OpenAI",
993
1004
  [API_KEY_CONSTANTS.ANTHROPIC]: "Anthropic",
@@ -1227,14 +1238,27 @@ var SupportedAiModels = {
1227
1238
  tags: ["expensive"]
1228
1239
  },
1229
1240
  [LLM_CONSTANTS.GEMINI_FLASH]: {
1230
- // Repointed 3.6 → 3.7 (2026-08-13) → 3.8 (2026-09-02); stable picker id, same pattern as gpt.
1231
- // 3.7 rejected thinkingLevel 'minimal' (low|medium|high only), unlike 3.5/3.6 — 3.8 untested
1232
- // on 'minimal', so keep the pin at 'medium' or above.
1233
- displayName: "Gemini 3.8 Flash",
1234
- modelApiName: "gemini-3.8-flash",
1241
+ // Repointed 3.6 → 3.7 (2026-08-13) → 3.8 (2026-09-02) → Gemini 4 Hydrogen (2026-11-12);
1242
+ // stable picker id, same pattern as gpt — the slot keeps the name `gemini-flash` although
1243
+ // Google dropped "Flash" from the model's name. Id verified against GET /v1beta/models
1244
+ // (1M in / 65K out). Accepts low|medium|high, rejects 'minimal' (400, probed — as did
1245
+ // 3.7 and 3.8 Flash).
1246
+ // Pinned 'low' (3.8 Flash ran 'medium'): Hydrogen is twice the price and thinks about
1247
+ // twice as much per level. Werewolf day-2 vote fixture, 3 runs a level, 2026-11-12:
1248
+ // Hydrogen low 3.4-3.9s 211-263 thinking tokens 0.58-0.60¢
1249
+ // Hydrogen medium 5.4-6.7s 528-761 0.81-0.98¢
1250
+ // Hydrogen high 6.4-7.8s 839-1175 1.03-1.29¢
1251
+ // 3.8 Flash medium 5.1-7.1s 203-315 0.29-0.34¢
1252
+ // 3.8 Flash low 2.1-2.3s 0 0.21¢
1253
+ // So 'low' is the level that thinks as much as the model it replaces did, a little
1254
+ // faster, at twice the cost instead of three times; every run named the same vote target
1255
+ // with a reason as specific. No `temperature` (see the Pro note). No speed tag until it
1256
+ // has 100+ production turns (3.8 Flash had none: p50 9.9s).
1257
+ displayName: "Gemini 4 Hydrogen",
1258
+ modelApiName: "gemini-4-hydrogen",
1235
1259
  apiKeyName: API_KEY_CONSTANTS.GOOGLE,
1236
1260
  hasThinking: true,
1237
- reasoningEffort: "medium"
1261
+ reasoningEffort: "low"
1238
1262
  },
1239
1263
  [LLM_CONSTANTS.GEMINI_LITE]: {
1240
1264
  displayName: "Gemini 3.5 Flash Lite",
@@ -1742,13 +1766,12 @@ var MODEL_PRICING = {
1742
1766
  extendedContextThresholdTokens: 2e5
1743
1767
  },
1744
1768
  [SupportedAiModels[LLM_CONSTANTS.GEMINI_FLASH].modelApiName]: {
1745
- // 3.8 Flash (2026-09-02) launched at the same rates as 3.7. 3.7's launch pricing was
1746
- // scheduled to double to $1.50/$7.50/$0.15 on 2027-01-01 (ai.google.dev pricing page,
1747
- // fetched 2026-08-13) — ACTION NEEDED then: re-check whether 3.8 follows and update.
1748
- // Cache storage cost ($0.50 / 1M tokens per hour) is not tracked here — the
1749
- // schema only models per-token call costs, not time-based storage.
1750
- inputPrice: 0.75,
1751
- outputPrice: 3.75,
1769
+ // Gemini 4 Hydrogen (ai.google.dev pricing page, Standard tier, read 2026-11-12): one
1770
+ // flat tier, twice the input and output rates of the 3.8 Flash it replaced
1771
+ // ($0.75 / $3.75) with the same $0.075 cached rate (5% of input, as on Argon). Implicit
1772
+ // caching only — explicit caching is not supported, so there is no storage cost.
1773
+ inputPrice: 1.5,
1774
+ outputPrice: 7.5,
1752
1775
  cacheHitPrice: 0.075
1753
1776
  },
1754
1777
  [SupportedAiModels[LLM_CONSTANTS.GEMINI_LITE].modelApiName]: {
@@ -1911,6 +1934,10 @@ function getProviderSignatureFields(aiType, signature) {
1911
1934
  }
1912
1935
  return {};
1913
1936
  }
1937
+ function enforcesSchema(modelId) {
1938
+ const config = SupportedAiModels[modelId];
1939
+ return !!config && SCHEMA_ENFORCING_API_KEYS.has(config.apiKeyName);
1940
+ }
1914
1941
 
1915
1942
  // src/reasoning-effort.ts
1916
1943
  var REASONING_EFFORT_SCALE = ["minimal", "low", "medium", "high", "xhigh", "max"];
@@ -5882,6 +5909,7 @@ var AgentFactory = class {
5882
5909
  OpenAiVoiceAgent,
5883
5910
  QwenAgent,
5884
5911
  REASONING_EFFORT_SCALE,
5912
+ SCHEMA_ENFORCING_API_KEYS,
5885
5913
  SUPPORTED_VOICE_PROVIDERS,
5886
5914
  SupportedAiKeyNames,
5887
5915
  SupportedAiModels,
@@ -5911,6 +5939,7 @@ var AgentFactory = class {
5911
5939
  createCatalog,
5912
5940
  createVoiceAgent,
5913
5941
  describeEmptyTtsResponse,
5942
+ enforcesSchema,
5914
5943
  evaluateBudget,
5915
5944
  extractAnthropicTokenUsage,
5916
5945
  extractAnthropicTokenUsageFromResponse,