@hiper2d/ai-agents 0.11.0 → 0.13.0

package/dist/index.d.mts

@@ -540,8 +540,11 @@ declare function getProviderSignatureFields(aiType: string, signature?: string):
  * - xAI Grok 4.7 (docs.x.ai/developers/model-capabilities/text/reasoning, 2026-09-22):
  *   low|medium|high|xhigh, default high, sent on the Responses API as `reasoning: {effort}`.
  *   Chat Completions has no equivalent. Grok 4.6 and earlier had no knob at all.
- * Qwen accepts reasoning_effort but ignores it (thinking_budget is its knob); MiniMax and Kimi
- * expose no effort parameter.
+ * - Moonshot Kimi K3 (platform.kimi.ai/docs/guide/use-reasoning-effort, 2026-09-22):
+ *   low|high|max, default max, top-level on Chat Completions. Moonshot shipped the lower
+ *   levels after launch; K3 accepted only "max" when this agent was written.
+ * Qwen accepts reasoning_effort but ignores it (thinking_budget is its knob); MiniMax exposes
+ * no effort parameter.
  */
 type OpenAIReasoningEffort = 'minimal' | 'low' | 'medium' | 'high' | 'xhigh';
 type AnthropicReasoningEffort = 'low' | 'medium' | 'high' | 'xhigh' | 'max';
@@ -552,6 +555,7 @@ type FuguReasoningEffort = 'high' | 'xhigh' | 'max';
 type MetaReasoningEffort = 'minimal' | 'low' | 'medium' | 'high' | 'xhigh' | 'max';
 type MistralReasoningEffort = 'minimal' | 'low' | 'medium' | 'high' | 'xhigh';
 type XaiReasoningEffort = 'low' | 'medium' | 'high' | 'xhigh';
+type KimiReasoningEffort = 'low' | 'high' | 'max';
 /** The shared scale, lowest first. */
 declare const REASONING_EFFORT_SCALE: readonly ReasoningEffort[];
 declare const OPENAI_REASONING_EFFORTS: readonly OpenAIReasoningEffort[];
@@ -563,6 +567,7 @@ declare const FUGU_REASONING_EFFORTS: readonly FuguReasoningEffort[];
 declare const META_REASONING_EFFORTS: readonly MetaReasoningEffort[];
 declare const MISTRAL_REASONING_EFFORTS: readonly MistralReasoningEffort[];
 declare const XAI_REASONING_EFFORTS: readonly XaiReasoningEffort[];
+declare const KIMI_REASONING_EFFORTS: readonly KimiReasoningEffort[];
 /** Clamps `effort` to the nearest level in `allowed` (by rank on the shared scale, ties go up). */
 declare function clampReasoningEffort<T extends ReasoningEffort>(effort: ReasoningEffort, allowed: readonly T[]): T;
 declare const toOpenAIEffort: (effort: ReasoningEffort) => OpenAIReasoningEffort;
@@ -574,6 +579,7 @@ declare const toMetaEffort: (effort: ReasoningEffort) => MetaReasoningEffort;
 declare const toFuguEffort: (effort: ReasoningEffort) => FuguReasoningEffort;
 declare const toMistralEffort: (effort: ReasoningEffort) => MistralReasoningEffort;
 declare const toXaiEffort: (effort: ReasoningEffort) => XaiReasoningEffort;
+declare const toKimiEffort: (effort: ReasoningEffort) => KimiReasoningEffort;
 
 /**
  * Unified token usage utilities for all AI providers
@@ -1388,6 +1394,15 @@ declare class Gpt5Agent extends AbstractAgent {
     private readonly logTemplates;
     private readonly errorMessages;
     constructor(name: string, instruction: string, model: string, apiKey: string, temperature: number, enableThinking?: boolean, agentLoggingConfig?: AgentLoggingConfig);
+    /**
+     * `reasoning: {effort}`, sent only when the catalog names a level. Every current OpenAI
+     * model defaults to "medium" when the field is omitted — verified 2026-09-22 by calling
+     * astra/sol/terra/luna with no reasoning param and reading the effort the API echoes back,
+     * since OpenAI documents the default for Sol, Luna and the 5.6 family but not for Astra.
+     * The catalog pins all four to medium so the level is explicit in our request rather than
+     * inherited from a default OpenAI can change under us.
+     */
+    private reasoningParams;
     /**
      * Structured output method using Zod with OpenAI's Responses API
      * This provides better schema handling and runtime validation
@@ -1736,4 +1751,4 @@ 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, KimiAgent, 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, 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, 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 };

package/dist/index.d.ts

@@ -540,8 +540,11 @@ declare function getProviderSignatureFields(aiType: string, signature?: string):
  * - xAI Grok 4.7 (docs.x.ai/developers/model-capabilities/text/reasoning, 2026-09-22):
  *   low|medium|high|xhigh, default high, sent on the Responses API as `reasoning: {effort}`.
  *   Chat Completions has no equivalent. Grok 4.6 and earlier had no knob at all.
- * Qwen accepts reasoning_effort but ignores it (thinking_budget is its knob); MiniMax and Kimi
- * expose no effort parameter.
+ * - Moonshot Kimi K3 (platform.kimi.ai/docs/guide/use-reasoning-effort, 2026-09-22):
+ *   low|high|max, default max, top-level on Chat Completions. Moonshot shipped the lower
+ *   levels after launch; K3 accepted only "max" when this agent was written.
+ * Qwen accepts reasoning_effort but ignores it (thinking_budget is its knob); MiniMax exposes
+ * no effort parameter.
  */
 type OpenAIReasoningEffort = 'minimal' | 'low' | 'medium' | 'high' | 'xhigh';
 type AnthropicReasoningEffort = 'low' | 'medium' | 'high' | 'xhigh' | 'max';
@@ -552,6 +555,7 @@ type FuguReasoningEffort = 'high' | 'xhigh' | 'max';
 type MetaReasoningEffort = 'minimal' | 'low' | 'medium' | 'high' | 'xhigh' | 'max';
 type MistralReasoningEffort = 'minimal' | 'low' | 'medium' | 'high' | 'xhigh';
 type XaiReasoningEffort = 'low' | 'medium' | 'high' | 'xhigh';
+type KimiReasoningEffort = 'low' | 'high' | 'max';
 /** The shared scale, lowest first. */
 declare const REASONING_EFFORT_SCALE: readonly ReasoningEffort[];
 declare const OPENAI_REASONING_EFFORTS: readonly OpenAIReasoningEffort[];
@@ -563,6 +567,7 @@ declare const FUGU_REASONING_EFFORTS: readonly FuguReasoningEffort[];
 declare const META_REASONING_EFFORTS: readonly MetaReasoningEffort[];
 declare const MISTRAL_REASONING_EFFORTS: readonly MistralReasoningEffort[];
 declare const XAI_REASONING_EFFORTS: readonly XaiReasoningEffort[];
+declare const KIMI_REASONING_EFFORTS: readonly KimiReasoningEffort[];
 /** Clamps `effort` to the nearest level in `allowed` (by rank on the shared scale, ties go up). */
 declare function clampReasoningEffort<T extends ReasoningEffort>(effort: ReasoningEffort, allowed: readonly T[]): T;
 declare const toOpenAIEffort: (effort: ReasoningEffort) => OpenAIReasoningEffort;
@@ -574,6 +579,7 @@ declare const toMetaEffort: (effort: ReasoningEffort) => MetaReasoningEffort;
 declare const toFuguEffort: (effort: ReasoningEffort) => FuguReasoningEffort;
 declare const toMistralEffort: (effort: ReasoningEffort) => MistralReasoningEffort;
 declare const toXaiEffort: (effort: ReasoningEffort) => XaiReasoningEffort;
+declare const toKimiEffort: (effort: ReasoningEffort) => KimiReasoningEffort;
 
 /**
  * Unified token usage utilities for all AI providers
@@ -1388,6 +1394,15 @@ declare class Gpt5Agent extends AbstractAgent {
     private readonly logTemplates;
     private readonly errorMessages;
     constructor(name: string, instruction: string, model: string, apiKey: string, temperature: number, enableThinking?: boolean, agentLoggingConfig?: AgentLoggingConfig);
+    /**
+     * `reasoning: {effort}`, sent only when the catalog names a level. Every current OpenAI
+     * model defaults to "medium" when the field is omitted — verified 2026-09-22 by calling
+     * astra/sol/terra/luna with no reasoning param and reading the effort the API echoes back,
+     * since OpenAI documents the default for Sol, Luna and the 5.6 family but not for Astra.
+     * The catalog pins all four to medium so the level is explicit in our request rather than
+     * inherited from a default OpenAI can change under us.
+     */
+    private reasoningParams;
     /**
      * Structured output method using Zod with OpenAI's Responses API
      * This provides better schema handling and runtime validation
@@ -1736,4 +1751,4 @@ 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, KimiAgent, 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, 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, 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 };

package/dist/index.js

@@ -55,6 +55,7 @@ __export(src_exports, {
   Gpt5Agent: () => Gpt5Agent,
   GrokAgent: () => GrokAgent,
   InMemorySpendStore: () => InMemorySpendStore,
+  KIMI_REASONING_EFFORTS: () => KIMI_REASONING_EFFORTS,
   KimiAgent: () => KimiAgent,
   LLM_CONSTANTS: () => LLM_CONSTANTS,
   MESSAGE_ROLE: () => MESSAGE_ROLE,
@@ -161,6 +162,7 @@ __export(src_exports, {
   toFuguEffort: () => toFuguEffort,
   toGeminiEffort: () => toGeminiEffort,
   toGlmEffort: () => toGlmEffort,
+  toKimiEffort: () => toKimiEffort,
   toMetaEffort: () => toMetaEffort,
   toMistralEffort: () => toMistralEffort,
   toOpenAIEffort: () => toOpenAIEffort,
@@ -1059,10 +1061,16 @@ var SupportedAiModels = {
   // GPT-6 Astra (2026-09-03): OpenAI's frontier tier above Sol. No `none` reasoning effort;
   // temperature/top_p are rejected — Gpt5Agent sends neither, so the same agent serves it.
   // The catalog temperature is only carried for the agent constructor signature.
+  // All four OpenAI entries carry reasoningEffort: 'medium' as of 2026-09-22. That is
+  // OpenAI's own current default, established by MEASUREMENT rather than docs: each model was
+  // called with no reasoning param and the effort it echoes back read off the response, since
+  // OpenAI documents the default for Sol, Luna and the 5.6 family but never states Astra's.
+  // Pinning it changes no behaviour today; it stops the default moving under us silently.
   [LLM_CONSTANTS.GPT_ASTRA]: {
     displayName: "GPT-6 Astra",
     modelApiName: "gpt-6-astra",
     apiKeyName: API_KEY_CONSTANTS.OPENAI,
+    reasoningEffort: "medium",
     hasThinking: true,
     temperature: 1,
     tags: ["expensive"]
@@ -1077,6 +1085,7 @@ var SupportedAiModels = {
     displayName: "GPT-6 Sol",
     modelApiName: "gpt-6-sol",
     apiKeyName: API_KEY_CONSTANTS.OPENAI,
+    reasoningEffort: "medium",
     hasThinking: true,
     temperature: 1,
     tags: ["expensive"]
@@ -1085,6 +1094,7 @@ var SupportedAiModels = {
     displayName: "GPT-5.6 Terra",
     modelApiName: "gpt-5.6-terra",
     apiKeyName: API_KEY_CONSTANTS.OPENAI,
+    reasoningEffort: "medium",
     hasThinking: true,
     temperature: 1,
     tags: ["fast", "expensive"]
@@ -1093,6 +1103,7 @@ var SupportedAiModels = {
     displayName: "GPT-6 Luna",
     modelApiName: "gpt-6-luna",
     apiKeyName: API_KEY_CONSTANTS.OPENAI,
+    reasoningEffort: "medium",
     hasThinking: true,
     temperature: 1,
     tags: ["fast", "cheap"]
@@ -1142,6 +1153,7 @@ var SupportedAiModels = {
     modelApiName: "grok-4.7",
     apiKeyName: API_KEY_CONSTANTS.GROK,
     hasThinking: true,
+    reasoningEffort: "high",
     temperature: 0.7
   },
   // Mistral models. Two hybrid entries since 2026-09-18: reasoning is off by default on the
@@ -1190,6 +1202,9 @@ var SupportedAiModels = {
     hasThinking: true,
     // Temperature is omitted from the request: kimi-k3 rejects any value other than 1.
     // Speed samples: 17s (2026-08-04) and 28.9s (2026-08-05) — graded into the >25s tier.
+    // Pinned 'high' 2026-09-22, down from Moonshot's 'max' default: K3 had no levels at
+    // launch and the agent hardcoded max, so this is the first time the knob exists.
+    reasoningEffort: "high",
     tags: ["very-slow", "expensive"]
   },
   // Z.AI models — thinking-only entry (non-thinking variant retired 2026-08-05)
@@ -1704,6 +1719,7 @@ var FUGU_REASONING_EFFORTS = ["high", "xhigh", "max"];
 var META_REASONING_EFFORTS = ["minimal", "low", "medium", "high", "xhigh", "max"];
 var MISTRAL_REASONING_EFFORTS = ["minimal", "low", "medium", "high", "xhigh"];
 var XAI_REASONING_EFFORTS = ["low", "medium", "high", "xhigh"];
+var KIMI_REASONING_EFFORTS = ["low", "high", "max"];
 function clampReasoningEffort(effort, allowed) {
   const rank = REASONING_EFFORT_SCALE.indexOf(effort);
   let best = allowed[0];
@@ -1726,6 +1742,7 @@ var toMetaEffort = (effort) => clampReasoningEffort(effort, META_REASONING_EFFOR
 var toFuguEffort = (effort) => clampReasoningEffort(effort, FUGU_REASONING_EFFORTS);
 var toMistralEffort = (effort) => clampReasoningEffort(effort, MISTRAL_REASONING_EFFORTS);
 var toXaiEffort = (effort) => clampReasoningEffort(effort, XAI_REASONING_EFFORTS);
+var toKimiEffort = (effort) => clampReasoningEffort(effort, KIMI_REASONING_EFFORTS);
 
 // src/pricing/token-usage-utils.ts
 function extractTokenUsage(response) {
@@ -2500,6 +2517,17 @@ ${instruction}`);
       apiKey
     });
   }
+  /**
+   * `reasoning: {effort}`, sent only when the catalog names a level. Every current OpenAI
+   * model defaults to "medium" when the field is omitted — verified 2026-09-22 by calling
+   * astra/sol/terra/luna with no reasoning param and reading the effort the API echoes back,
+   * since OpenAI documents the default for Sol, Luna and the 5.6 family but not for Astra.
+   * The catalog pins all four to medium so the level is explicit in our request rather than
+   * inherited from a default OpenAI can change under us.
+   */
+  reasoningParams() {
+    return this.reasoningEffort ? { reasoning: { effort: toOpenAIEffort(this.reasoningEffort) } } : {};
+  }
   /**
    * Structured output method using Zod with OpenAI's Responses API
    * This provides better schema handling and runtime validation
@@ -2520,6 +2548,7 @@ ${instruction}`);
           input,
           max_output_tokens: this.maxOutputTokens,
           prompt_cache_key: this.promptCacheKey,
+          ...this.reasoningParams(),
           text: {
             format: (0, import_zod2.zodTextFormat)(schemaToSend, "response_schema")
           }
@@ -2604,7 +2633,8 @@ ${instruction}`);
         instructions: this.instruction,
         input,
         max_output_tokens: this.maxOutputTokens,
-        prompt_cache_key: this.promptCacheKey
+        prompt_cache_key: this.promptCacheKey,
+        ...this.reasoningParams()
       });
       const content = response.output_text;
       if (!content) {
@@ -4100,8 +4130,8 @@ var KimiAgent = class extends AbstractAgent {
       model: this.model,
       stream: false,
       max_tokens: this.maxOutputTokens,
-      // Moonshot's only accepted level; "max" is not in the OpenAI SDK's ReasoningEffort union.
-      reasoning_effort: "max"
+      // "max" is not in the OpenAI SDK's ReasoningEffort union, hence the cast.
+      reasoning_effort: toKimiEffort(this.reasoningEffort ?? "max")
     };
   }
   // Log message templates
@@ -5378,6 +5408,7 @@ var AgentFactory = class {
   Gpt5Agent,
   GrokAgent,
   InMemorySpendStore,
+  KIMI_REASONING_EFFORTS,
   KimiAgent,
   LLM_CONSTANTS,
   MESSAGE_ROLE,
@@ -5484,6 +5515,7 @@ var AgentFactory = class {
   toFuguEffort,
   toGeminiEffort,
   toGlmEffort,
+  toKimiEffort,
   toMetaEffort,
   toMistralEffort,
   toOpenAIEffort,