@hiper2d/ai-agents 0.18.0 → 0.19.1

package/dist/index.d.mts

@@ -293,6 +293,17 @@ declare class ModelInvalidResponseError extends ModelError {
  *              Google's non-configurable filter — no safetySettings change lifts it.
  *              `reason` = the blockReason or finishReason string.
  */
+/**
+ * The provider did not answer within the agent's `requestTimeoutMs`. The ask gives up
+ * instead of hanging until the host's serverless function is killed, which leaves no log
+ * line and no error for the player to retry from. Observed 2026-10-01: `deepseek-flash`
+ * accepted requests and never answered (no tokens, even streaming) while `deepseek-v4-pro`
+ * on the same key answered in 0.5s. Retryable — usually on another model.
+ */
+declare class ModelTimeoutError extends ModelError {
+    timeoutMs: number;
+    constructor(modelType: string, timeoutMs: number);
+}
 declare class ModelRefusalError extends ModelError {
     reason?: string;
     constructor(modelType: string, message?: string, reason?: string);
@@ -370,7 +381,6 @@ declare const LLM_CONSTANTS: {
     KIMI: string;
     GLM: string;
     GLM_FLASH: string;
-    FUGU_ULTRA: string;
     FUGU_MAX: string;
     QWEN_MAX: string;
     QWEN_FLASH: string;
@@ -388,6 +398,13 @@ declare const LLM_CONSTANTS: {
  * runaway one. Reasoning effort and thinking budgets are the knobs that change spend.
  */
 declare const DEFAULT_MAX_OUTPUT_TOKENS = 8192;
+/**
+ * Wall-clock limit on one ask, retries inside the provider SDK included. Past it the ask
+ * throws ModelTimeoutError. Measured 2026-10-01 over 7k production calls (14 days): 0.5%
+ * ran past 60s — mostly Grok turns and story generation. Callers that legitimately need
+ * longer raise `requestTimeoutMs` on the instance, like maxOutputTokens.
+ */
+declare const DEFAULT_REQUEST_TIMEOUT_MS = 60000;
 type ModelTag = 'very-fast' | 'fast' | 'slow' | 'very-slow' | 'extremely-slow' | 'cheap' | 'expensive';
 type ReasoningEffort = 'minimal' | 'low' | 'medium' | 'high' | 'xhigh' | 'max';
 /**
@@ -534,11 +551,11 @@ declare function getProviderSignatureFields(aiType: string, signature?: string):
  *   exactly that enum). The server default is xhigh for fugu-ultra and high for the rest.
  * - Meta Muse Spark (verified 2026-09-12): minimal|low|medium|high|xhigh|max ("none" → 400; max is
  *   Standard tier only)
- * - Mistral Small 4 / Medium 3.5 (verified 2026-09-18): the API enum is
- *   none|minimal|low|medium|high|xhigh, with reasoning OFF unless the field is sent. "none" is
- *   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.
+ * - Mistral Small 4 / Medium 3.5 / Large 4: the documented API enum is
+ *   none|minimal|low|medium|high|xhigh, but every model we run accepts only "none" and "high"
+ *   (each other value is a 400 "not supported for this model, supported values: [high, none]",
+ *   checked on all three 2026-10-06). Reasoning is OFF unless the field is sent; "none" means
+ *   "don't reason", which an agent expresses by omitting the field, so the slice is just high.
  * - 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.
@@ -1286,6 +1303,12 @@ declare abstract class AbstractAgent {
      */
     reasoningEffort?: ReasoningEffort;
     thinkingBudgetTokens?: number;
+    /**
+     * Wall-clock limit on one ask (DEFAULT_REQUEST_TIMEOUT_MS). Past it the ask throws
+     * ModelTimeoutError; the provider request is abandoned, not cancelled. Raise it per
+     * instance for calls that legitimately run long, like maxOutputTokens.
+     */
+    requestTimeoutMs: number;
     protected readonly instruction: string;
     /**
      * The instruction split on CACHE_TIER_MARKER: [shared static tier, per-bot tier].
@@ -1306,6 +1329,11 @@ declare abstract class AbstractAgent {
      */
     askWithZodSchema<T>(zodSchema: z.ZodSchema<T>, messages: AIMessage[]): Promise<[T, string, TokenUsage$1?, string?]>;
     askText(messages: AIMessage[]): Promise<[string, string, TokenUsage$1?, string?]>;
+    /**
+     * Races the provider call against requestTimeoutMs. A hung provider (accepts the request,
+     * never answers) otherwise runs until the host kills the function with no error to show.
+     */
+    private withTimeout;
     private stampDuration;
     /** Failed calls carry their duration too — a 35s provider stall that errors is still signal. */
     private stampErrorDuration;
@@ -1804,4 +1832,4 @@ declare class MiniMaxAgent extends AbstractAgent {
     doAskText(messages: AIMessage[]): Promise<[string, string, TokenUsage$1?, string?]>;
 }
 
-export { type AIMessage, ANTHROPIC_REASONING_EFFORTS, API_KEY_CONSTANTS, AbstractAgent, type AgentActivityData, AgentFactory, type AgentLoggingConfig, type AnthropicReasoningEffort, type AnthropicTokenUsage, type ApiKeyMap, type BeforeAskHook, BotResponseError, BudgetController, type BudgetControllerOptions, BudgetExceededError, type BudgetLimit, type BudgetVerdict, CACHE_TIER_MARKER, ClaudeAgent, type CostCalculationOptions, DEEPSEEK_PEAK_SCHEDULE, DEEPSEEK_REASONING_EFFORTS, DEFAULT_LOGGING_CONFIG, DEFAULT_MAX_OUTPUT_TOKENS, type DeepSeekReasoningEffort, type DeepSeekTokenUsage, DeepSeekV2Agent, FUGU_REASONING_EFFORTS, FuguAgent, type FuguReasoningEffort, GEMINI_AUDIO_TOKENS_PER_SECOND, GEMINI_REASONING_EFFORTS, GLM_REASONING_EFFORTS, type GeminiReasoningEffort, GlmAgent, type GlmReasoningEffort, GoogleAgent, type GoogleSttOptions, type GoogleSttResult, type GoogleTokenUsage, type GoogleTtsAudioOptions, type GoogleTtsResult, GoogleVoiceAgent, Gpt5Agent, GrokAgent, type GrokTokenUsage, InMemorySpendStore, type JsonSchemaOptions, KIMI_REASONING_EFFORTS, KimiAgent, type KimiReasoningEffort, type KimiTokenUsage, type LLMModel, LLM_CONSTANTS, type LlmLogger, type LoggingConfig, MESSAGE_ROLE, META_REASONING_EFFORTS, MISTRAL_REASONING_EFFORTS, MODEL_PRICING, MetaAgent, type MetaReasoningEffort, type MetaTokenUsage, MiniMaxAgent, MistralAgent, type MistralReasoningEffort, type MistralTokenUsage, type Modality, ModelAuthenticationError, type ModelConfig, ModelError, ModelInvalidResponseError, ModelOverloadError, type ModelPricing, ModelQuotaExceededError, ModelRateLimitError, ModelRefusalError, type ModelTag, ModelUnavailableError, OPENAI_REASONING_EFFORTS, type OpenAIReasoningEffort, type OpenAITokenUsage, type OpenAiSttOptions, type OpenAiSttResult, type OpenAiTtsAudioOptions, type OpenAiTtsVoice, OpenAiVoiceAgent, type PeakPricing, type PricingUnit, type ProviderSchema, type TokenUsage as ProviderTokenUsage, type ProviderType, QwenAgent, REASONING_EFFORT_SCALE, type ReasoningEffort, SUPPORTED_VOICE_PROVIDERS, type SpeechRequest, type SpeechResult, type SpeechUsage, type SpendInput, type SpendKind, type SpendLedger, type SpendLedgers, type SpendStore, type SpendWindow, SupportedAiKeyNames, SupportedAiModels, type TokenPairUsage, type TokenUsage$1 as TokenUsage, type TranscriptionRequest, type TranscriptionResult, type TranscriptionUsage, VOICE_MODEL_CONSTANTS, VOICE_MODEL_PRICING, VOICE_PROVIDER_API_KEY, type VoiceAgent, VoiceAgentFactory, type VoiceModelPricing, type VoiceProvider, XAI_REASONING_EFFORTS, type XaiReasoningEffort, ZodSchemaConverter, applySpend, 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 };
+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 };

package/dist/index.d.ts

@@ -293,6 +293,17 @@ declare class ModelInvalidResponseError extends ModelError {
  *              Google's non-configurable filter — no safetySettings change lifts it.
  *              `reason` = the blockReason or finishReason string.
  */
+/**
+ * The provider did not answer within the agent's `requestTimeoutMs`. The ask gives up
+ * instead of hanging until the host's serverless function is killed, which leaves no log
+ * line and no error for the player to retry from. Observed 2026-10-01: `deepseek-flash`
+ * accepted requests and never answered (no tokens, even streaming) while `deepseek-v4-pro`
+ * on the same key answered in 0.5s. Retryable — usually on another model.
+ */
+declare class ModelTimeoutError extends ModelError {
+    timeoutMs: number;
+    constructor(modelType: string, timeoutMs: number);
+}
 declare class ModelRefusalError extends ModelError {
     reason?: string;
     constructor(modelType: string, message?: string, reason?: string);
@@ -370,7 +381,6 @@ declare const LLM_CONSTANTS: {
     KIMI: string;
     GLM: string;
     GLM_FLASH: string;
-    FUGU_ULTRA: string;
     FUGU_MAX: string;
     QWEN_MAX: string;
     QWEN_FLASH: string;
@@ -388,6 +398,13 @@ declare const LLM_CONSTANTS: {
  * runaway one. Reasoning effort and thinking budgets are the knobs that change spend.
  */
 declare const DEFAULT_MAX_OUTPUT_TOKENS = 8192;
+/**
+ * Wall-clock limit on one ask, retries inside the provider SDK included. Past it the ask
+ * throws ModelTimeoutError. Measured 2026-10-01 over 7k production calls (14 days): 0.5%
+ * ran past 60s — mostly Grok turns and story generation. Callers that legitimately need
+ * longer raise `requestTimeoutMs` on the instance, like maxOutputTokens.
+ */
+declare const DEFAULT_REQUEST_TIMEOUT_MS = 60000;
 type ModelTag = 'very-fast' | 'fast' | 'slow' | 'very-slow' | 'extremely-slow' | 'cheap' | 'expensive';
 type ReasoningEffort = 'minimal' | 'low' | 'medium' | 'high' | 'xhigh' | 'max';
 /**
@@ -534,11 +551,11 @@ declare function getProviderSignatureFields(aiType: string, signature?: string):
  *   exactly that enum). The server default is xhigh for fugu-ultra and high for the rest.
  * - Meta Muse Spark (verified 2026-09-12): minimal|low|medium|high|xhigh|max ("none" → 400; max is
  *   Standard tier only)
- * - Mistral Small 4 / Medium 3.5 (verified 2026-09-18): the API enum is
- *   none|minimal|low|medium|high|xhigh, with reasoning OFF unless the field is sent. "none" is
- *   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.
+ * - Mistral Small 4 / Medium 3.5 / Large 4: the documented API enum is
+ *   none|minimal|low|medium|high|xhigh, but every model we run accepts only "none" and "high"
+ *   (each other value is a 400 "not supported for this model, supported values: [high, none]",
+ *   checked on all three 2026-10-06). Reasoning is OFF unless the field is sent; "none" means
+ *   "don't reason", which an agent expresses by omitting the field, so the slice is just high.
  * - 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.
@@ -1286,6 +1303,12 @@ declare abstract class AbstractAgent {
      */
     reasoningEffort?: ReasoningEffort;
     thinkingBudgetTokens?: number;
+    /**
+     * Wall-clock limit on one ask (DEFAULT_REQUEST_TIMEOUT_MS). Past it the ask throws
+     * ModelTimeoutError; the provider request is abandoned, not cancelled. Raise it per
+     * instance for calls that legitimately run long, like maxOutputTokens.
+     */
+    requestTimeoutMs: number;
     protected readonly instruction: string;
     /**
      * The instruction split on CACHE_TIER_MARKER: [shared static tier, per-bot tier].
@@ -1306,6 +1329,11 @@ declare abstract class AbstractAgent {
      */
     askWithZodSchema<T>(zodSchema: z.ZodSchema<T>, messages: AIMessage[]): Promise<[T, string, TokenUsage$1?, string?]>;
     askText(messages: AIMessage[]): Promise<[string, string, TokenUsage$1?, string?]>;
+    /**
+     * Races the provider call against requestTimeoutMs. A hung provider (accepts the request,
+     * never answers) otherwise runs until the host kills the function with no error to show.
+     */
+    private withTimeout;
     private stampDuration;
     /** Failed calls carry their duration too — a 35s provider stall that errors is still signal. */
     private stampErrorDuration;
@@ -1804,4 +1832,4 @@ declare class MiniMaxAgent extends AbstractAgent {
     doAskText(messages: AIMessage[]): Promise<[string, string, TokenUsage$1?, string?]>;
 }
 
-export { type AIMessage, ANTHROPIC_REASONING_EFFORTS, API_KEY_CONSTANTS, AbstractAgent, type AgentActivityData, AgentFactory, type AgentLoggingConfig, type AnthropicReasoningEffort, type AnthropicTokenUsage, type ApiKeyMap, type BeforeAskHook, BotResponseError, BudgetController, type BudgetControllerOptions, BudgetExceededError, type BudgetLimit, type BudgetVerdict, CACHE_TIER_MARKER, ClaudeAgent, type CostCalculationOptions, DEEPSEEK_PEAK_SCHEDULE, DEEPSEEK_REASONING_EFFORTS, DEFAULT_LOGGING_CONFIG, DEFAULT_MAX_OUTPUT_TOKENS, type DeepSeekReasoningEffort, type DeepSeekTokenUsage, DeepSeekV2Agent, FUGU_REASONING_EFFORTS, FuguAgent, type FuguReasoningEffort, GEMINI_AUDIO_TOKENS_PER_SECOND, GEMINI_REASONING_EFFORTS, GLM_REASONING_EFFORTS, type GeminiReasoningEffort, GlmAgent, type GlmReasoningEffort, GoogleAgent, type GoogleSttOptions, type GoogleSttResult, type GoogleTokenUsage, type GoogleTtsAudioOptions, type GoogleTtsResult, GoogleVoiceAgent, Gpt5Agent, GrokAgent, type GrokTokenUsage, InMemorySpendStore, type JsonSchemaOptions, KIMI_REASONING_EFFORTS, KimiAgent, type KimiReasoningEffort, type KimiTokenUsage, type LLMModel, LLM_CONSTANTS, type LlmLogger, type LoggingConfig, MESSAGE_ROLE, META_REASONING_EFFORTS, MISTRAL_REASONING_EFFORTS, MODEL_PRICING, MetaAgent, type MetaReasoningEffort, type MetaTokenUsage, MiniMaxAgent, MistralAgent, type MistralReasoningEffort, type MistralTokenUsage, type Modality, ModelAuthenticationError, type ModelConfig, ModelError, ModelInvalidResponseError, ModelOverloadError, type ModelPricing, ModelQuotaExceededError, ModelRateLimitError, ModelRefusalError, type ModelTag, ModelUnavailableError, OPENAI_REASONING_EFFORTS, type OpenAIReasoningEffort, type OpenAITokenUsage, type OpenAiSttOptions, type OpenAiSttResult, type OpenAiTtsAudioOptions, type OpenAiTtsVoice, OpenAiVoiceAgent, type PeakPricing, type PricingUnit, type ProviderSchema, type TokenUsage as ProviderTokenUsage, type ProviderType, QwenAgent, REASONING_EFFORT_SCALE, type ReasoningEffort, SUPPORTED_VOICE_PROVIDERS, type SpeechRequest, type SpeechResult, type SpeechUsage, type SpendInput, type SpendKind, type SpendLedger, type SpendLedgers, type SpendStore, type SpendWindow, SupportedAiKeyNames, SupportedAiModels, type TokenPairUsage, type TokenUsage$1 as TokenUsage, type TranscriptionRequest, type TranscriptionResult, type TranscriptionUsage, VOICE_MODEL_CONSTANTS, VOICE_MODEL_PRICING, VOICE_PROVIDER_API_KEY, type VoiceAgent, VoiceAgentFactory, type VoiceModelPricing, type VoiceProvider, XAI_REASONING_EFFORTS, type XaiReasoningEffort, ZodSchemaConverter, applySpend, 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 };
+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 };

package/dist/index.js

@@ -43,6 +43,7 @@ __export(src_exports, {
   DEEPSEEK_REASONING_EFFORTS: () => DEEPSEEK_REASONING_EFFORTS,
   DEFAULT_LOGGING_CONFIG: () => DEFAULT_LOGGING_CONFIG,
   DEFAULT_MAX_OUTPUT_TOKENS: () => DEFAULT_MAX_OUTPUT_TOKENS,
+  DEFAULT_REQUEST_TIMEOUT_MS: () => DEFAULT_REQUEST_TIMEOUT_MS,
   DeepSeekV2Agent: () => DeepSeekV2Agent,
   FUGU_REASONING_EFFORTS: () => FUGU_REASONING_EFFORTS,
   FuguAgent: () => FuguAgent,
@@ -72,6 +73,7 @@ __export(src_exports, {
   ModelQuotaExceededError: () => ModelQuotaExceededError,
   ModelRateLimitError: () => ModelRateLimitError,
   ModelRefusalError: () => ModelRefusalError,
+  ModelTimeoutError: () => ModelTimeoutError,
   ModelUnavailableError: () => ModelUnavailableError,
   OPENAI_REASONING_EFFORTS: () => OPENAI_REASONING_EFFORTS,
   OpenAiVoiceAgent: () => OpenAiVoiceAgent,
@@ -889,6 +891,14 @@ var ModelInvalidResponseError = class extends ModelError {
     this.truncated = truncated;
   }
 };
+var ModelTimeoutError = class extends ModelError {
+  timeoutMs;
+  constructor(modelType, timeoutMs) {
+    super(`${modelType} did not respond within ${Math.round(timeoutMs / 1e3)}s`, modelType);
+    this.name = "ModelTimeoutError";
+    this.timeoutMs = timeoutMs;
+  }
+};
 var ModelRefusalError = class extends ModelError {
   reason;
   constructor(modelType, message = `${modelType} refused to answer (stop_reason: refusal)`, reason) {
@@ -980,7 +990,6 @@ var LLM_CONSTANTS = {
   KIMI: "kimi",
   GLM: "glm",
   GLM_FLASH: "glm-flash",
-  FUGU_ULTRA: "fugu-ultra",
   FUGU_MAX: "fugu-max",
   // Qwen (QwenCloud/DashScope). Stable picker ids without the version, matching the gpt/gemini
   // pattern, so future repoints don't orphan persisted ids.
@@ -992,6 +1001,7 @@ var LLM_CONSTANTS = {
   MUSE_SPARK: "muse-spark"
 };
 var DEFAULT_MAX_OUTPUT_TOKENS = 8192;
+var DEFAULT_REQUEST_TIMEOUT_MS = 6e4;
 var SupportedAiModels = {
   // Claude Fable 5.1 (2026-09-02, was Fable 5) - frontier reasoning model. Thinking is always on (no non-thinking variant).
   [LLM_CONSTANTS.CLAUDE_FABLE]: {
@@ -1152,16 +1162,20 @@ var SupportedAiModels = {
   },
   // Always-on reasoning — no non-thinking sibling. Bumped 4.6 → 4.7 on 2026-09-22 (id verified
   // against GET /v1/models). Same price, same 200K extended-context boundary; the context
-  // window is 500K. NOTE 4.7 newly accepts `reasoning_effort` (low | medium | high | xhigh,
-  // default high) where 4.6 had no knob at all. GrokAgent still sends none, so behaviour is
-  // unchanged — but Grok is the app's most expensive model per turn (94% of its output is
-  // reasoning), so a lower pin is the obvious lever to measure before anything else.
+  // window is 500K. 4.7 accepts `reasoning: {effort}` (low | medium | high | xhigh, default
+  // high) where 4.6 had no knob at all; GrokAgent sends the pin below.
+  //
+  // Pinned medium 2026-10-06. At high, production day summaries ran ~2 min and the werewolf
+  // day-2 fixture measured votes 27-55s and summaries 44-55s (1.5-3k output tokens) — too
+  // close to the 60s per-call timeout. At medium the same prompts took votes 5-8s and
+  // summaries 23-25s (3-5x fewer output tokens) with the same vote target and summaries just
+  // as specific.
   [LLM_CONSTANTS.GROK]: {
     displayName: "Grok 4.7",
     modelApiName: "grok-4.7",
     apiKeyName: API_KEY_CONSTANTS.GROK,
     hasThinking: true,
-    reasoningEffort: "high",
+    reasoningEffort: "medium",
     temperature: 0.7
   },
   // Mistral models. Two hybrid entries since 2026-09-18: reasoning is off by default on the
@@ -1180,6 +1194,13 @@ var SupportedAiModels = {
   // returns `mistral-medium-3-5`, `mistral-medium-3`, `mistral-medium-latest` and
   // `magistral-medium-latest` as one entry — so Magistral seats had silently become Medium
   // seats already. Consumers map both retired ids onto the surviving pair.
+  //
+  // Mistral Large 4 (`mistral-large-4`, $0.68/$2.09, $0.07 cached) was added in 0.19.0 and
+  // removed in 0.19.1 (2026-10-06/07): too slow to play. Like Small/Medium it accepts only
+  // reasoning_effort high|none, and unlike them it reasons when the field is omitted. At high
+  // a realistic day-2 vote took 25-154s (half past the 60s timeout) and its parallel story
+  // batches finished ~100s apart (Mistral served them one at a time: ~6 min per story). Off,
+  // it answers in ~1s with no trace. Re-test before re-adding.
   [LLM_CONSTANTS.MISTRAL_MEDIUM]: {
     displayName: "Mistral Medium 3.5",
     // The docs' explicit 3.5 id. It was `mistral-medium-3` until 2026-09-18, which Mistral
@@ -1264,9 +1285,9 @@ var SupportedAiModels = {
   // real against $0.85 tracked, a 5.7x undercharge. It is a router with no published price,
   // so the rate is not even guaranteed stable, and its cache hit rate was 9.3% — effectively
   // zero, since every hit came from a duplicate call seconds apart rather than turn-to-turn
-  // prefix reuse. Ultra costs the same and is predictable.
+  // prefix reuse.
   //
-  // reasoningEffort is PINNED on both entries because the server default differs per model:
+  // reasoningEffort was PINNED on both entries because the server default differs per model:
   // fugu-ultra defaults to `xhigh` (console.sakana.ai/models), fugu-max to `high`. Until
   // 2026-09-20 the agent sent no effort at all, so every Ultra turn ran at xhigh. Measured that
   // day on a 2k-token three-day game prompt (two samples each, chat completions, json mode):
@@ -1276,17 +1297,17 @@ var SupportedAiModels = {
   //   max   xhigh   41s /  47s   $0.02
   // `max_tokens` is not a latency lever for Ultra: Sakana applies it to the final response only,
   // the orchestrator "still uses maximum token limit". Effort is the only knob.
-  [LLM_CONSTANTS.FUGU_ULTRA]: {
-    displayName: "Sakana Fugu Ultra",
-    modelApiName: "fugu-ultra",
-    apiKeyName: API_KEY_CONSTANTS.FUGU,
-    hasThinking: false,
-    reasoningEffort: "high",
-    tags: ["extremely-slow", "expensive"]
-  },
+  //
+  // Fugu Ultra (`fugu-ultra`, $5/$30, $0.50 cached; ×~2 above 272K) was REMOVED in 0.19.1
+  // (2026-10-07): too slow and too expensive for what it played. At `high` a day-2 vote took
+  // 39s and a turn 65-78s (past the 60s request timeout), and as Game Master its four parallel
+  // story batches ran past 240s each. It also billed "orchestration" tokens (its internal expert
+  // calls) on top of the visible counts; FuguAgent still folds those in, in case Max ever
+  // reports them. Consumers map the id to Max.
   // Fugu Max (2026-09-20, fugu-max → fugu-max-v1.0): Sakana's "largest pool of models on the
-  // cost–performance Pareto frontier" — the everyday Fugu at 5x less than Ultra's output rate
-  // and a fraction of its latency. It reasons (reasoning_tokens 80-500 a turn) but, like
+  // cost–performance Pareto frontier" — 5x less than Ultra's output rate and a fraction of its
+  // latency; the only Fugu since 0.19.1. Live 2026-10-06: day-2 vote 11s, 15-character story
+  // pipeline 100s. It reasons (reasoning_tokens 80-500 a turn) but, like
   // Ultra, never returns reasoning_content, hence hasThinking: false. No orchestration
   // tokens observed on any call.
   [LLM_CONSTANTS.FUGU_MAX]: {
@@ -1593,21 +1614,10 @@ var MODEL_PRICING = {
     extendedContextCacheHitPrice: 1,
     extendedContextThresholdTokens: 2e5
   },
-  // Sakana Fugu models (console.sakana.ai/pricing, read 2026-09-20). Base `fugu` was retired
-  // 2026-08-04 — it had no published price and measured out at these same ultra rates, so it
-  // has no pricing entry. fugu-ultra: above 272K context the rates roughly double. Ultra also
-  // reports "orchestration" tokens (its internal expert calls) in prompt_tokens_details /
-  // completion_tokens_details, billed at these same input/output rates — FuguAgent folds them
-  // into the token counts before pricing.
-  [SupportedAiModels[LLM_CONSTANTS.FUGU_ULTRA].modelApiName]: {
-    inputPrice: 5,
-    outputPrice: 30,
-    cacheHitPrice: 0.5,
-    extendedContextInputPrice: 10,
-    extendedContextOutputPrice: 45,
-    extendedContextCacheHitPrice: 1,
-    extendedContextThresholdTokens: 272e3
-  },
+  // Sakana Fugu (console.sakana.ai/pricing, read 2026-09-20). Base `fugu` (retired 2026-08-04)
+  // and `fugu-ultra` (removed 0.19.1) have no entries. Any "orchestration" tokens reported in
+  // prompt_tokens_details / completion_tokens_details bill at these same rates — FuguAgent
+  // folds them into the token counts before pricing.
   // fugu-max: flat rates, no context tiers.
   [SupportedAiModels[LLM_CONSTANTS.FUGU_MAX].modelApiName]: {
     inputPrice: 2,
@@ -1735,7 +1745,7 @@ var GLM_REASONING_EFFORTS = ["low", "high", "max"];
 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 MISTRAL_REASONING_EFFORTS = ["high"];
 var XAI_REASONING_EFFORTS = ["low", "medium", "high", "xhigh"];
 var KIMI_REASONING_EFFORTS = ["low", "high", "max"];
 function clampReasoningEffort(effort, allowed) {
@@ -2403,6 +2413,12 @@ var AbstractAgent = class {
    */
   reasoningEffort;
   thinkingBudgetTokens;
+  /**
+   * Wall-clock limit on one ask (DEFAULT_REQUEST_TIMEOUT_MS). Past it the ask throws
+   * ModelTimeoutError; the provider request is abandoned, not cancelled. Raise it per
+   * instance for calls that legitimately run long, like maxOutputTokens.
+   */
+  requestTimeoutMs = DEFAULT_REQUEST_TIMEOUT_MS;
   instruction;
   /**
    * The instruction split on CACHE_TIER_MARKER: [shared static tier, per-bot tier].
@@ -2437,7 +2453,7 @@ var AbstractAgent = class {
     if (beforeAskHook) await beforeAskHook(this);
     const startedAt = Date.now();
     try {
-      const [result, thinking, usage, signature] = await this.doAskWithZodSchema(zodSchema, messages);
+      const [result, thinking, usage, signature] = await this.withTimeout(this.doAskWithZodSchema(zodSchema, messages));
       return [result, thinking, this.stampDuration(usage, startedAt), signature];
     } catch (error) {
       this.stampErrorDuration(error, startedAt);
@@ -2448,13 +2464,25 @@ var AbstractAgent = class {
     if (beforeAskHook) await beforeAskHook(this);
     const startedAt = Date.now();
     try {
-      const [content, thinking, usage, signature] = await this.doAskText(messages);
+      const [content, thinking, usage, signature] = await this.withTimeout(this.doAskText(messages));
       return [content, thinking, this.stampDuration(usage, startedAt), signature];
     } catch (error) {
       this.stampErrorDuration(error, startedAt);
       throw error;
     }
   }
+  /**
+   * Races the provider call against requestTimeoutMs. A hung provider (accepts the request,
+   * never answers) otherwise runs until the host kills the function with no error to show.
+   */
+  withTimeout(call) {
+    const timeoutMs = this.requestTimeoutMs;
+    let timer;
+    const timeout = new Promise((_, reject) => {
+      timer = setTimeout(() => reject(new ModelTimeoutError(this.model, timeoutMs)), timeoutMs);
+    });
+    return Promise.race([call, timeout]).finally(() => clearTimeout(timer));
+  }
   stampDuration(usage, startedAt) {
     return usage ? { ...usage, durationMs: Date.now() - startedAt } : usage;
   }
@@ -5480,7 +5508,6 @@ var AgentFactory = class {
       case LLM_CONSTANTS.GLM_FLASH:
         return new GlmAgent(name, instruction, model.modelApiName, key, model.temperature, shouldEnableThinking);
       // Sakana Fugu models — always-on reasoning, no temperature (ignored by the model)
-      case LLM_CONSTANTS.FUGU_ULTRA:
       case LLM_CONSTANTS.FUGU_MAX:
         return new FuguAgent(name, instruction, model.modelApiName, key, shouldEnableThinking);
       // Qwen models — thinking-only (enable_thinking always sent)
@@ -5520,6 +5547,7 @@ var AgentFactory = class {
   DEEPSEEK_REASONING_EFFORTS,
   DEFAULT_LOGGING_CONFIG,
   DEFAULT_MAX_OUTPUT_TOKENS,
+  DEFAULT_REQUEST_TIMEOUT_MS,
   DeepSeekV2Agent,
   FUGU_REASONING_EFFORTS,
   FuguAgent,
@@ -5549,6 +5577,7 @@ var AgentFactory = class {
   ModelQuotaExceededError,
   ModelRateLimitError,
   ModelRefusalError,
+  ModelTimeoutError,
   ModelUnavailableError,
   OPENAI_REASONING_EFFORTS,
   OpenAiVoiceAgent,