@hiper2d/ai-agents 0.10.0 → 0.12.0

package/dist/index.d.mts

@@ -537,8 +537,14 @@ declare function getProviderSignatureFields(aiType: string, signature?: string):
  *   not on the shared scale and means "don't reason", which an agent expresses by omitting the
  *   field, so the exported slice starts at minimal. Only "high" is documented to return the
  *   thinking chunk; the intermediate levels are accepted but unverified.
- * Qwen accepts reasoning_effort but ignores it (thinking_budget is its knob); MiniMax, Kimi and
- * Grok expose no effort parameter.
+ * - 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.
+ * - 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';
@@ -548,6 +554,8 @@ type DeepSeekReasoningEffort = 'low' | 'high' | 'max';
 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[];
@@ -558,6 +566,8 @@ declare const DEEPSEEK_REASONING_EFFORTS: readonly DeepSeekReasoningEffort[];
 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;
@@ -568,6 +578,8 @@ declare const toDeepSeekEffort: (effort: ReasoningEffort) => DeepSeekReasoningEf
 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
@@ -1535,6 +1547,14 @@ declare class GrokAgent extends AbstractAgent {
      * Reasoning extraction and token accounting are identical to askWithZodSchema.
      */
     doAskText(messages: AIMessage[]): Promise<[string, string, TokenUsage$1?, string?]>;
+    /**
+     * A stable key per (game, bot) so xAI routes this conversation's requests to the same
+     * server. Without it a request often lands on a cache-cold machine and the whole prompt
+     * bills at full input price — xAI calls this out as the most common caching mistake.
+     * Falls back to the bot name alone outside a game (story generation, tests), which is
+     * still stabler than sending nothing.
+     */
+    private get promptCacheKey();
     private createResponse;
     /**
      * Converts game history to Responses API input items. The system instruction is
@@ -1722,4 +1742,4 @@ 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, 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, 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

@@ -537,8 +537,14 @@ declare function getProviderSignatureFields(aiType: string, signature?: string):
  *   not on the shared scale and means "don't reason", which an agent expresses by omitting the
  *   field, so the exported slice starts at minimal. Only "high" is documented to return the
  *   thinking chunk; the intermediate levels are accepted but unverified.
- * Qwen accepts reasoning_effort but ignores it (thinking_budget is its knob); MiniMax, Kimi and
- * Grok expose no effort parameter.
+ * - 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.
+ * - 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';
@@ -548,6 +554,8 @@ type DeepSeekReasoningEffort = 'low' | 'high' | 'max';
 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[];
@@ -558,6 +566,8 @@ declare const DEEPSEEK_REASONING_EFFORTS: readonly DeepSeekReasoningEffort[];
 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;
@@ -568,6 +578,8 @@ declare const toDeepSeekEffort: (effort: ReasoningEffort) => DeepSeekReasoningEf
 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
@@ -1535,6 +1547,14 @@ declare class GrokAgent extends AbstractAgent {
      * Reasoning extraction and token accounting are identical to askWithZodSchema.
      */
     doAskText(messages: AIMessage[]): Promise<[string, string, TokenUsage$1?, string?]>;
+    /**
+     * A stable key per (game, bot) so xAI routes this conversation's requests to the same
+     * server. Without it a request often lands on a cache-cold machine and the whole prompt
+     * bills at full input price — xAI calls this out as the most common caching mistake.
+     * Falls back to the bot name alone outside a game (story generation, tests), which is
+     * still stabler than sending nothing.
+     */
+    private get promptCacheKey();
     private createResponse;
     /**
      * Converts game history to Responses API input items. The system instruction is
@@ -1722,4 +1742,4 @@ 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, 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, 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,
@@ -83,6 +84,7 @@ __export(src_exports, {
   VOICE_MODEL_PRICING: () => VOICE_MODEL_PRICING,
   VOICE_PROVIDER_API_KEY: () => VOICE_PROVIDER_API_KEY,
   VoiceAgentFactory: () => VoiceAgentFactory,
+  XAI_REASONING_EFFORTS: () => XAI_REASONING_EFFORTS,
   ZodSchemaConverter: () => ZodSchemaConverter,
   applySpend: () => applySpend,
   buildGoogleTtsPrompt: () => buildGoogleTtsPrompt,
@@ -160,9 +162,11 @@ __export(src_exports, {
   toFuguEffort: () => toFuguEffort,
   toGeminiEffort: () => toGeminiEffort,
   toGlmEffort: () => toGlmEffort,
+  toKimiEffort: () => toKimiEffort,
   toMetaEffort: () => toMetaEffort,
   toMistralEffort: () => toMistralEffort,
   toOpenAIEffort: () => toOpenAIEffort,
+  toXaiEffort: () => toXaiEffort,
   transcribeWithGemini: () => transcribeWithGemini,
   transcribeWithOpenAi: () => transcribeWithOpenAi,
   validateResponse: () => validateResponse
@@ -1140,6 +1144,7 @@ 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
@@ -1188,6 +1193,9 @@ 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)
@@ -1701,6 +1709,8 @@ var DEEPSEEK_REASONING_EFFORTS = ["low", "high", "max"];
 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];
@@ -1722,6 +1732,8 @@ var toDeepSeekEffort = (effort) => clampReasoningEffort(effort, DEEPSEEK_REASONI
 var toMetaEffort = (effort) => clampReasoningEffort(effort, META_REASONING_EFFORTS);
 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) {
@@ -3959,6 +3971,16 @@ ${schemaDescription}`;
       throw new Error(this.errorMessages.apiError(error));
     }
   }
+  /**
+   * A stable key per (game, bot) so xAI routes this conversation's requests to the same
+   * server. Without it a request often lands on a cache-cold machine and the whole prompt
+   * bills at full input price — xAI calls this out as the most common caching mistake.
+   * Falls back to the bot name alone outside a game (story generation, tests), which is
+   * still stabler than sending nothing.
+   */
+  get promptCacheKey() {
+    return this.gameId ? `${this.gameId}:${this.name}` : this.name;
+  }
   createResponse(input, jsonMode) {
     return this.client.responses.create({
       model: this.model,
@@ -3969,9 +3991,16 @@ ${schemaDescription}`;
       // ever starts truncating — measured turns peak far below the shared default.
       max_output_tokens: this.maxOutputTokens,
       // We manage conversation state ourselves; encrypted reasoning is only
-      // returned for unstored responses.
+      // returned for unstored responses. Grok 4.7 returns reasoning.encrypted_content
+      // on the Responses API whether or not `include` asks for it; we keep asking so
+      // the behaviour is identical if the model is ever pointed back at 4.6.
       store: false,
       include: ["reasoning.encrypted_content"],
+      prompt_cache_key: this.promptCacheKey,
+      // Grok 4.7 is the first Grok with an effort knob (4.6 had none) and defaults to
+      // "high". Only send it when the catalog asks for a level, so an unset entry keeps
+      // the provider default rather than us silently picking one.
+      ...this.reasoningEffort ? { reasoning: { effort: toXaiEffort(this.reasoningEffort) } } : {},
       ...jsonMode ? { text: { format: { type: "json_object" } } } : {}
     });
   }
@@ -4079,8 +4108,8 @@ 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
@@ -5357,6 +5386,7 @@ var AgentFactory = class {
   Gpt5Agent,
   GrokAgent,
   InMemorySpendStore,
+  KIMI_REASONING_EFFORTS,
   KimiAgent,
   LLM_CONSTANTS,
   MESSAGE_ROLE,
@@ -5385,6 +5415,7 @@ var AgentFactory = class {
   VOICE_MODEL_PRICING,
   VOICE_PROVIDER_API_KEY,
   VoiceAgentFactory,
+  XAI_REASONING_EFFORTS,
   ZodSchemaConverter,
   applySpend,
   buildGoogleTtsPrompt,
@@ -5462,9 +5493,11 @@ var AgentFactory = class {
   toFuguEffort,
   toGeminiEffort,
   toGlmEffort,
+  toKimiEffort,
   toMetaEffort,
   toMistralEffort,
   toOpenAIEffort,
+  toXaiEffort,
   transcribeWithGemini,
   transcribeWithOpenAi,
   validateResponse