@hiper2d/ai-agents 0.28.1 → 0.29.0

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
@@ -580,7 +580,9 @@ declare function getProviderSignatureFields(aiType: string, signature?: string):
580
580
  * alias for thinking_budget (low = 4,096, medium = 16,384, xhigh = 262,144 tokens) and the
581
581
  * two are mutually exclusive — sending both is a 400. QwenAgent sends the budget instead,
582
582
  * because 1,024 is tighter than the lowest level effort can express.
583
- * MiniMax exposes no effort parameter.
583
+ * - MiniMax M3.1 Flash (platform.minimax.io text-chat-openai reference, 2026-10-27):
584
+ * low|medium|high|xhigh|max, default max, top-level `reasoning_effort` on Chat Completions;
585
+ * "none" → 400. Only M3.1 Flash honours it — M3 and M2.x ignore the field.
584
586
  */
585
587
  type OpenAIReasoningEffort = 'minimal' | 'low' | 'medium' | 'high' | 'xhigh';
586
588
  type AnthropicReasoningEffort = 'low' | 'medium' | 'high' | 'xhigh' | 'max';
@@ -592,6 +594,7 @@ type MetaReasoningEffort = 'minimal' | 'low' | 'medium' | 'high' | 'xhigh' | 'ma
592
594
  type MistralReasoningEffort = 'minimal' | 'low' | 'medium' | 'high' | 'xhigh';
593
595
  type XaiReasoningEffort = 'low' | 'medium' | 'high' | 'xhigh';
594
596
  type KimiReasoningEffort = 'low' | 'high' | 'max';
597
+ type MiniMaxReasoningEffort = 'low' | 'medium' | 'high' | 'xhigh' | 'max';
595
598
  /** The shared scale, lowest first. */
596
599
  declare const REASONING_EFFORT_SCALE: readonly ReasoningEffort[];
597
600
  declare const OPENAI_REASONING_EFFORTS: readonly OpenAIReasoningEffort[];
@@ -604,6 +607,7 @@ declare const META_REASONING_EFFORTS: readonly MetaReasoningEffort[];
604
607
  declare const MISTRAL_REASONING_EFFORTS: readonly MistralReasoningEffort[];
605
608
  declare const XAI_REASONING_EFFORTS: readonly XaiReasoningEffort[];
606
609
  declare const KIMI_REASONING_EFFORTS: readonly KimiReasoningEffort[];
610
+ declare const MINIMAX_REASONING_EFFORTS: readonly MiniMaxReasoningEffort[];
607
611
  /** Clamps `effort` to the nearest level in `allowed` (by rank on the shared scale, ties go up). */
608
612
  declare function clampReasoningEffort<T extends ReasoningEffort>(effort: ReasoningEffort, allowed: readonly T[]): T;
609
613
  declare const toOpenAIEffort: (effort: ReasoningEffort) => OpenAIReasoningEffort;
@@ -616,6 +620,7 @@ declare const toFuguEffort: (effort: ReasoningEffort) => FuguReasoningEffort;
616
620
  declare const toMistralEffort: (effort: ReasoningEffort) => MistralReasoningEffort;
617
621
  declare const toXaiEffort: (effort: ReasoningEffort) => XaiReasoningEffort;
618
622
  declare const toKimiEffort: (effort: ReasoningEffort) => KimiReasoningEffort;
623
+ declare const toMiniMaxEffort: (effort: ReasoningEffort) => MiniMaxReasoningEffort;
619
624
 
620
625
  /**
621
626
  * Unified token usage utilities for all AI providers
@@ -1858,4 +1863,4 @@ declare class MiniMaxAgent extends AbstractAgent {
1858
1863
  doAskText(messages: AIMessage[]): Promise<[string, string, TokenUsage$1?, string?]>;
1859
1864
  }
1860
1865
 
1861
- 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, 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, 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, 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, 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 };
1866
+ 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, 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, 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, toMiniMaxEffort, toMistralEffort, toOpenAIEffort, toXaiEffort, transcribeWithGemini, transcribeWithOpenAi, validateResponse };
package/dist/index.d.ts CHANGED
@@ -580,7 +580,9 @@ declare function getProviderSignatureFields(aiType: string, signature?: string):
580
580
  * alias for thinking_budget (low = 4,096, medium = 16,384, xhigh = 262,144 tokens) and the
581
581
  * two are mutually exclusive — sending both is a 400. QwenAgent sends the budget instead,
582
582
  * because 1,024 is tighter than the lowest level effort can express.
583
- * MiniMax exposes no effort parameter.
583
+ * - MiniMax M3.1 Flash (platform.minimax.io text-chat-openai reference, 2026-10-27):
584
+ * low|medium|high|xhigh|max, default max, top-level `reasoning_effort` on Chat Completions;
585
+ * "none" → 400. Only M3.1 Flash honours it — M3 and M2.x ignore the field.
584
586
  */
585
587
  type OpenAIReasoningEffort = 'minimal' | 'low' | 'medium' | 'high' | 'xhigh';
586
588
  type AnthropicReasoningEffort = 'low' | 'medium' | 'high' | 'xhigh' | 'max';
@@ -592,6 +594,7 @@ type MetaReasoningEffort = 'minimal' | 'low' | 'medium' | 'high' | 'xhigh' | 'ma
592
594
  type MistralReasoningEffort = 'minimal' | 'low' | 'medium' | 'high' | 'xhigh';
593
595
  type XaiReasoningEffort = 'low' | 'medium' | 'high' | 'xhigh';
594
596
  type KimiReasoningEffort = 'low' | 'high' | 'max';
597
+ type MiniMaxReasoningEffort = 'low' | 'medium' | 'high' | 'xhigh' | 'max';
595
598
  /** The shared scale, lowest first. */
596
599
  declare const REASONING_EFFORT_SCALE: readonly ReasoningEffort[];
597
600
  declare const OPENAI_REASONING_EFFORTS: readonly OpenAIReasoningEffort[];
@@ -604,6 +607,7 @@ declare const META_REASONING_EFFORTS: readonly MetaReasoningEffort[];
604
607
  declare const MISTRAL_REASONING_EFFORTS: readonly MistralReasoningEffort[];
605
608
  declare const XAI_REASONING_EFFORTS: readonly XaiReasoningEffort[];
606
609
  declare const KIMI_REASONING_EFFORTS: readonly KimiReasoningEffort[];
610
+ declare const MINIMAX_REASONING_EFFORTS: readonly MiniMaxReasoningEffort[];
607
611
  /** Clamps `effort` to the nearest level in `allowed` (by rank on the shared scale, ties go up). */
608
612
  declare function clampReasoningEffort<T extends ReasoningEffort>(effort: ReasoningEffort, allowed: readonly T[]): T;
609
613
  declare const toOpenAIEffort: (effort: ReasoningEffort) => OpenAIReasoningEffort;
@@ -616,6 +620,7 @@ declare const toFuguEffort: (effort: ReasoningEffort) => FuguReasoningEffort;
616
620
  declare const toMistralEffort: (effort: ReasoningEffort) => MistralReasoningEffort;
617
621
  declare const toXaiEffort: (effort: ReasoningEffort) => XaiReasoningEffort;
618
622
  declare const toKimiEffort: (effort: ReasoningEffort) => KimiReasoningEffort;
623
+ declare const toMiniMaxEffort: (effort: ReasoningEffort) => MiniMaxReasoningEffort;
619
624
 
620
625
  /**
621
626
  * Unified token usage utilities for all AI providers
@@ -1858,4 +1863,4 @@ declare class MiniMaxAgent extends AbstractAgent {
1858
1863
  doAskText(messages: AIMessage[]): Promise<[string, string, TokenUsage$1?, string?]>;
1859
1864
  }
1860
1865
 
1861
- 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, 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, 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, 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, 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 };
1866
+ 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, 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, 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, toMiniMaxEffort, toMistralEffort, toOpenAIEffort, toXaiEffort, transcribeWithGemini, transcribeWithOpenAi, validateResponse };
package/dist/index.js CHANGED
@@ -61,6 +61,7 @@ __export(src_exports, {
61
61
  LLM_CONSTANTS: () => LLM_CONSTANTS,
62
62
  MESSAGE_ROLE: () => MESSAGE_ROLE,
63
63
  META_REASONING_EFFORTS: () => META_REASONING_EFFORTS,
64
+ MINIMAX_REASONING_EFFORTS: () => MINIMAX_REASONING_EFFORTS,
64
65
  MISTRAL_REASONING_EFFORTS: () => MISTRAL_REASONING_EFFORTS,
65
66
  MODEL_PRICING: () => MODEL_PRICING,
66
67
  MetaAgent: () => MetaAgent,
@@ -167,6 +168,7 @@ __export(src_exports, {
167
168
  toGlmEffort: () => toGlmEffort,
168
169
  toKimiEffort: () => toKimiEffort,
169
170
  toMetaEffort: () => toMetaEffort,
171
+ toMiniMaxEffort: () => toMiniMaxEffort,
170
172
  toMistralEffort: () => toMistralEffort,
171
173
  toOpenAIEffort: () => toOpenAIEffort,
172
174
  toXaiEffort: () => toXaiEffort,
@@ -1445,20 +1447,29 @@ var SupportedAiModels = {
1445
1447
  thinkingBudgetTokens: 1024,
1446
1448
  tags: ["slow", "cheap"]
1447
1449
  },
1448
- // MiniMax M3 (OpenAI-compatible endpoint, 1M context). Thinking-only entry: M3's `thinking`
1449
- // param defaults to adaptive (it decides per-request how much to think) and can be disabled,
1450
- // making it hybrid for cost purposes. The agent always sends `reasoning_split: true` so
1451
- // thinking arrives in `reasoning_content` instead of as `<think>` tags inside the answer.
1452
- // Note: unlike Qwen, M3 has NO thinking-budget parameter — adaptive is the only throttle.
1453
- // Speed from the 2026-08-05 live measurement (single sample): 25.3s → the >25s tier.
1454
- // Temperature: MiniMax range is [0,2], default 1.
1450
+ // MiniMax M3.1 Flash (OpenAI-compatible endpoint, 1M context). `minimax` repointed 2026-10-27
1451
+ // from MiniMax-M3 ($0.30/$1.20 after the permanent 50% off, 25s single-sample turn, 60s
1452
+ // timeouts in prod on 2026-10-25 and 10-27) to MiniMax-M3.1-Flash: half the price
1453
+ // ($0.15/$0.60), and the first MiniMax model with an effort dial — low|medium|high|xhigh|max,
1454
+ // provider default max, so it is pinned. Thinking is always on (`thinking: disabled` → 400),
1455
+ // thinking arrives in `reasoning_content` because the agent sends `reasoning_split: true`
1456
+ // (M3.1 rejects false). Temperature and top_p are documented as ignored ("not supported
1457
+ // yet", always 1 / 0.95), so the entry sets none and the agent sends none.
1458
+ // Live 2026-10-27 at medium: the all-providers sweep ran the schema ask in 3.1s, prose in
1459
+ // 1.7s and the 8-character scene (363 output tokens) in 2.4s — ~150 tok/s, the fastest
1460
+ // provider in the sweep that day (Kimi Highspeed's scene takes 8s); a 3.2k-token game
1461
+ // prompt with a two-sentence JSON reply took 1.9-2.1s at every effort level vs 3.5s for M3.
1462
+ // Tagged `fast` rather than `very-fast`: the sweep straddles the 3s line and production
1463
+ // turns carry 10k+ tokens of context. Launched 2026-09-27 as Token-Plan-only (the
1464
+ // 2026-10-02 probe got `unknown model` on the pay-as-you-go key); on the key since at
1465
+ // least 2026-10-27.
1455
1466
  [LLM_CONSTANTS.MINIMAX]: {
1456
- displayName: "MiniMax M3",
1457
- modelApiName: "MiniMax-M3",
1467
+ displayName: "MiniMax M3.1 Flash",
1468
+ modelApiName: "MiniMax-M3.1-Flash",
1458
1469
  apiKeyName: API_KEY_CONSTANTS.MINIMAX,
1459
1470
  hasThinking: true,
1460
- temperature: 1,
1461
- tags: ["very-slow", "cheap"]
1471
+ reasoningEffort: "medium",
1472
+ tags: ["fast", "cheap"]
1462
1473
  },
1463
1474
  // Meta Muse Spark 1.3 (added 2026-09-12) on Meta's own Model API — Standard tier, i.e.
1464
1475
  // the private model id (the `-contributor` id is a quarter of the price but Meta trains
@@ -1775,18 +1786,14 @@ var MODEL_PRICING = {
1775
1786
  outputPrice: 0.47,
1776
1787
  cacheHitPrice: 0.016
1777
1788
  },
1778
- // MiniMax M3. Rates from platform.minimax.io/docs/guides/pricing-paygo (2026-08-05, USD,
1779
- // "permanent 50% off" already applied): ≤512k and >512k input tiers. Caching is automatic
1780
- // (≥512 input tokens), hits reported in prompt_tokens_details.cached_tokens; no write fee
1781
- // for M3.
1789
+ // MiniMax M3.1 Flash. Rates from platform.minimax.io/docs/guides/pricing-paygo (2026-10-27,
1790
+ // USD): one tier across the 1M window (M3 had ≤512k / >512k tiers; the page lists none for
1791
+ // M3.1 Flash). Caching is automatic (≥512 input tokens), hits reported in
1792
+ // prompt_tokens_details.cached_tokens; no write fee listed.
1782
1793
  [SupportedAiModels[LLM_CONSTANTS.MINIMAX].modelApiName]: {
1783
- inputPrice: 0.3,
1784
- outputPrice: 1.2,
1785
- cacheHitPrice: 0.06,
1786
- extendedContextInputPrice: 0.6,
1787
- extendedContextOutputPrice: 2.4,
1788
- extendedContextCacheHitPrice: 0.12,
1789
- extendedContextThresholdTokens: 512e3
1794
+ inputPrice: 0.15,
1795
+ outputPrice: 0.6,
1796
+ cacheHitPrice: 0.015
1790
1797
  },
1791
1798
  // Meta Muse Spark 1.3, Standard tier. Rates from ai.developer.meta.com/docs/pricing-rate-limits
1792
1799
  // (2026-09-12): no long-context premium at any point of the 1M window; reasoning tokens bill
@@ -1816,7 +1823,7 @@ var HYBRID_THINKING_API_NAMES = /* @__PURE__ */ new Set([
1816
1823
  // hybrid by the same definition: we force reasoning on, so they carry the multiplier.
1817
1824
  SupportedAiModels[LLM_CONSTANTS.QWEN_MAX].modelApiName,
1818
1825
  SupportedAiModels[LLM_CONSTANTS.QWEN_FLASH].modelApiName,
1819
- SupportedAiModels[LLM_CONSTANTS.MINIMAX].modelApiName,
1826
+ // MiniMax left the set 2026-10-27: M3.1 Flash cannot turn thinking off (M3 could).
1820
1827
  // Mistral reasons only when asked (`reasoning_effort`), and we always ask — hybrid by the
1821
1828
  // same definition as Qwen. Added 2026-09-18 when the Mistral entries went thinking-on.
1822
1829
  SupportedAiModels[LLM_CONSTANTS.MISTRAL_SMALL].modelApiName
@@ -1889,6 +1896,7 @@ var META_REASONING_EFFORTS = ["minimal", "low", "medium", "high", "xhigh", "max"
1889
1896
  var MISTRAL_REASONING_EFFORTS = ["high"];
1890
1897
  var XAI_REASONING_EFFORTS = ["low", "medium", "high", "xhigh"];
1891
1898
  var KIMI_REASONING_EFFORTS = ["low", "high", "max"];
1899
+ var MINIMAX_REASONING_EFFORTS = ["low", "medium", "high", "xhigh", "max"];
1892
1900
  function clampReasoningEffort(effort, allowed) {
1893
1901
  const rank = REASONING_EFFORT_SCALE.indexOf(effort);
1894
1902
  let best = allowed[0];
@@ -1912,6 +1920,7 @@ var toFuguEffort = (effort) => clampReasoningEffort(effort, FUGU_REASONING_EFFOR
1912
1920
  var toMistralEffort = (effort) => clampReasoningEffort(effort, MISTRAL_REASONING_EFFORTS);
1913
1921
  var toXaiEffort = (effort) => clampReasoningEffort(effort, XAI_REASONING_EFFORTS);
1914
1922
  var toKimiEffort = (effort) => clampReasoningEffort(effort, KIMI_REASONING_EFFORTS);
1923
+ var toMiniMaxEffort = (effort) => clampReasoningEffort(effort, MINIMAX_REASONING_EFFORTS);
1915
1924
 
1916
1925
  // src/pricing/token-usage-utils.ts
1917
1926
  function extractTokenUsage(response) {
@@ -5229,10 +5238,10 @@ var MiniMaxAgent = class extends AbstractAgent {
5229
5238
  get defaultParams() {
5230
5239
  return {
5231
5240
  model: this.model,
5232
- temperature: this.temperature,
5241
+ ...this.temperature !== void 0 ? { temperature: this.temperature } : {},
5233
5242
  stream: false,
5234
- // MiniMax deprecates max_tokens in favor of max_completion_tokens (M3 max is 512K,
5235
- // far above anything a turn needs).
5243
+ // MiniMax deprecates max_tokens in favor of max_completion_tokens (the M-series max is
5244
+ // 512K, far above anything a turn needs).
5236
5245
  max_completion_tokens: this.maxOutputTokens
5237
5246
  };
5238
5247
  }
@@ -5254,7 +5263,9 @@ var MiniMaxAgent = class extends AbstractAgent {
5254
5263
  thinkingParams() {
5255
5264
  return {
5256
5265
  thinking: { type: this.enableThinking ? "adaptive" : "disabled" },
5257
- reasoning_split: true
5266
+ reasoning_split: true,
5267
+ // Provider default is max; the catalog pins M3.1 Flash to medium. Older models ignore it.
5268
+ ...this.enableThinking && this.reasoningEffort ? { reasoning_effort: toMiniMaxEffort(this.reasoningEffort) } : {}
5258
5269
  };
5259
5270
  }
5260
5271
  convertToOpenAIMessages(messages) {
@@ -5667,7 +5678,7 @@ var AgentFactory = class {
5667
5678
  case LLM_CONSTANTS.QWEN_MAX:
5668
5679
  case LLM_CONSTANTS.QWEN_FLASH:
5669
5680
  return new QwenAgent(name, instruction, model.modelApiName, key, model.temperature, shouldEnableThinking);
5670
- // MiniMax M3 — adaptive thinking (the model decides per-request)
5681
+ // MiniMax M3.1 Flash — always-on adaptive thinking, effort from the catalog
5671
5682
  case LLM_CONSTANTS.MINIMAX:
5672
5683
  return new MiniMaxAgent(name, instruction, model.modelApiName, key, model.temperature, shouldEnableThinking);
5673
5684
  // Meta Muse Spark — Responses API with encrypted reasoning replay, effort from the catalog
@@ -5719,6 +5730,7 @@ var AgentFactory = class {
5719
5730
  LLM_CONSTANTS,
5720
5731
  MESSAGE_ROLE,
5721
5732
  META_REASONING_EFFORTS,
5733
+ MINIMAX_REASONING_EFFORTS,
5722
5734
  MISTRAL_REASONING_EFFORTS,
5723
5735
  MODEL_PRICING,
5724
5736
  MetaAgent,
@@ -5825,6 +5837,7 @@ var AgentFactory = class {
5825
5837
  toGlmEffort,
5826
5838
  toKimiEffort,
5827
5839
  toMetaEffort,
5840
+ toMiniMaxEffort,
5828
5841
  toMistralEffort,
5829
5842
  toOpenAIEffort,
5830
5843
  toXaiEffort,