@hiper2d/ai-agents 0.5.3 → 0.7.0

package/dist/index.d.mts

@@ -363,15 +363,14 @@ declare const LLM_CONSTANTS: {
     GEMINI_PRO: string;
     GEMINI_FLASH: string;
     GEMINI_LITE: string;
-    MISTRAL_LARGE: string;
     MISTRAL_MEDIUM: string;
     MISTRAL_SMALL: string;
-    MISTRAL_MAGISTRAL: string;
     GROK: string;
     KIMI: string;
     GLM: string;
     GLM_FLASH: string;
     FUGU_ULTRA: string;
+    FUGU_MAX: string;
     QWEN_MAX: string;
     QWEN_FLASH: string;
     MINIMAX: string;
@@ -482,7 +481,7 @@ declare const DEEPSEEK_PEAK_SCHEDULE: PeakPricing;
 declare const MODEL_PRICING: Record<string, ModelPricing>;
 /** True for hybrid thinking-only models — the ones whose effective output price is a known
  *  multiple of the sticker price. Always-on reasoning models (GPT-5, Gemini, Grok, Kimi,
- *  Fable, Magistral) also burn reasoning tokens, but their multiplier hasn't been measured. */
+ *  Fable) also burn reasoning tokens, but their multiplier hasn't been measured. */
 declare function isHybridThinkingModel(modelApiName: string): boolean;
 interface CostCalculationOptions {
     cacheHitTokens?: number;
@@ -529,19 +528,26 @@ declare function getProviderSignatureFields(aiType: string, signature?: string):
  * - Gemini 3.x thinkingLevel: minimal|low|medium|high (3.1 Pro and 3.7 Flash reject minimal)
  * - Z.AI GLM-5.3 / 5.3-Flash: low|high|max only
  * - DeepSeek V4: low|high|max (the API itself aliases medium → high)
- * - Sakana Fugu: high|xhigh
+ * - Sakana Fugu (verified live 2026-09-20): high|xhigh|max on every model ("low" → 400 naming
+ *   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)
- * Qwen accepts reasoning_effort but ignores it (thinking_budget is its knob); MiniMax, Kimi,
- * Grok and Mistral expose no effort parameter.
+ * - 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.
+ * Qwen accepts reasoning_effort but ignores it (thinking_budget is its knob); MiniMax, Kimi and
+ * Grok expose no effort parameter.
  */
 type OpenAIReasoningEffort = 'minimal' | 'low' | 'medium' | 'high' | 'xhigh';
 type AnthropicReasoningEffort = 'low' | 'medium' | 'high' | 'xhigh' | 'max';
 type GeminiReasoningEffort = 'minimal' | 'low' | 'medium' | 'high';
 type GlmReasoningEffort = 'low' | 'high' | 'max';
 type DeepSeekReasoningEffort = 'low' | 'high' | 'max';
-type FuguReasoningEffort = 'high' | 'xhigh';
+type FuguReasoningEffort = 'high' | 'xhigh' | 'max';
 type MetaReasoningEffort = 'minimal' | 'low' | 'medium' | 'high' | 'xhigh' | 'max';
+type MistralReasoningEffort = 'minimal' | 'low' | 'medium' | 'high' | 'xhigh';
 /** The shared scale, lowest first. */
 declare const REASONING_EFFORT_SCALE: readonly ReasoningEffort[];
 declare const OPENAI_REASONING_EFFORTS: readonly OpenAIReasoningEffort[];
@@ -551,6 +557,7 @@ declare const GLM_REASONING_EFFORTS: readonly GlmReasoningEffort[];
 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[];
 /** 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;
@@ -560,6 +567,7 @@ declare const toGlmEffort: (effort: ReasoningEffort) => GlmReasoningEffort;
 declare const toDeepSeekEffort: (effort: ReasoningEffort) => DeepSeekReasoningEffort;
 declare const toMetaEffort: (effort: ReasoningEffort) => MetaReasoningEffort;
 declare const toFuguEffort: (effort: ReasoningEffort) => FuguReasoningEffort;
+declare const toMistralEffort: (effort: ReasoningEffort) => MistralReasoningEffort;
 
 /**
  * Unified token usage utilities for all AI providers
@@ -641,8 +649,10 @@ declare function extractAnthropicTokenUsage(response: any): TokenUsage | null;
 declare function extractGoogleTokenUsage(response: any): TokenUsage | null;
 /**
  * Mistral-specific token usage extraction
- * Mistral SDK uses camelCase (promptTokens, completionTokens, totalTokens)
- * and reasoning tokens may be in additionalProperties for Magistral models
+ * Mistral SDK uses camelCase (promptTokens, completionTokens, totalTokens). Reasoning tokens
+ * are not itemised today (they sit inside completionTokens, observed 2026-09-18 on Small 4 and
+ * Medium 3.5 with reasoning_effort: high); the additionalProperties probe below stays in case
+ * a reasoning_tokens field ever appears.
  */
 declare function extractMistralTokenUsage(response: any): TokenUsage | null;
 
@@ -1417,12 +1427,36 @@ declare class GoogleAgent extends AbstractAgent {
     private handleGeminiError;
 }
 
+/**
+ * Mistral agent (Mistral SDK chat.complete) for the hybrid Small 4 / Medium 3.5 models.
+ *
+ * Reasoning is off on the API by default and switched on per request with `reasoning_effort`
+ * (docs.mistral.ai/studio/conversations/reasoning). SDK 1.x has no typed field for it, so it is
+ * injected into the wire body by the same beforeRequest hook that adds `prompt_cache_key`.
+ * With it set, the assistant message comes back as content chunks — a `thinking` chunk (a list
+ * of text chunks holding the trace) followed by the `text` answer — and this works together
+ * with json_schema structured output (verified live 2026-09-18 on both models). Without the
+ * field the reply is a plain string and no trace exists.
+ *
+ * Multi-turn: Mistral asks that the full assistant message, thinking chunk included, be
+ * replayed into history ("stripping reasoning traces degrades performance"). The trace is
+ * plain text, so it rides on `AIMessage.thinking` — no provider signature field — and
+ * convertToMistralMessages rebuilds each prior assistant turn as [thinking, text] chunks.
+ *
+ * Reasoning tokens are counted inside completion_tokens; no separate reasoning_tokens field
+ * arrives, so cost accounting on the output rate already covers the trace.
+ */
 declare class MistralAgent extends AbstractAgent {
     private readonly client;
     private get defaultParams();
     private readonly logTemplates;
     private readonly errorMessages;
     constructor(name: string, instruction: string, model: string, apiKey: string, enableThinking?: boolean, agentLoggingConfig?: AgentLoggingConfig);
+    /**
+     * History in Mistral's shape. A prior assistant turn that carries a stored trace is
+     * replayed as [thinking, text] chunks, as Mistral's multi-turn guidance asks; every other
+     * turn is a plain string.
+     */
     private convertToMistralMessages;
     private processReply;
     private processStructuredReply;
@@ -1438,9 +1472,8 @@ declare class MistralAgent extends AbstractAgent {
      */
     doAskWithZodSchema<T>(zodSchema: z.ZodSchema<T>, messages: AIMessage[]): Promise<[T, string, TokenUsage$1?, string?]>;
     /**
-     * Plain-text ask: no schema appended, no responseFormat. Note that Magistral
-     * reasoning models only return thinking traces when responseFormat is NOT
-     * json_object, so unlike askWithZodSchema this path can surface thinking content.
+     * Plain-text ask: no schema appended, no responseFormat. Reasoning (and the trace) works
+     * the same way as on the schema path.
      */
     doAskText(messages: AIMessage[]): Promise<[string, string, TokenUsage$1?, string?]>;
 }
@@ -1620,7 +1653,6 @@ declare class FuguAgent extends AbstractAgent {
     private readonly errorMessages;
     constructor(name: string, instruction: string, model: string, apiKey: string, enableThinking?: boolean, agentLoggingConfig?: AgentLoggingConfig);
     private convertToOpenAIMessages;
-    private logRawUsageForCalibration;
     private extractThinkingAndUsage;
     private prependSystemInstruction;
     doAskWithZodSchema<T>(zodSchema: z.ZodSchema<T>, messages: AIMessage[]): Promise<[T, string, TokenUsage$1?, string?]>;
@@ -1690,4 +1722,4 @@ 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, MODEL_PRICING, MetaAgent, type MetaReasoningEffort, type MetaTokenUsage, MiniMaxAgent, MistralAgent, 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, 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, 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 };

package/dist/index.d.ts

@@ -363,15 +363,14 @@ declare const LLM_CONSTANTS: {
     GEMINI_PRO: string;
     GEMINI_FLASH: string;
     GEMINI_LITE: string;
-    MISTRAL_LARGE: string;
     MISTRAL_MEDIUM: string;
     MISTRAL_SMALL: string;
-    MISTRAL_MAGISTRAL: string;
     GROK: string;
     KIMI: string;
     GLM: string;
     GLM_FLASH: string;
     FUGU_ULTRA: string;
+    FUGU_MAX: string;
     QWEN_MAX: string;
     QWEN_FLASH: string;
     MINIMAX: string;
@@ -482,7 +481,7 @@ declare const DEEPSEEK_PEAK_SCHEDULE: PeakPricing;
 declare const MODEL_PRICING: Record<string, ModelPricing>;
 /** True for hybrid thinking-only models — the ones whose effective output price is a known
  *  multiple of the sticker price. Always-on reasoning models (GPT-5, Gemini, Grok, Kimi,
- *  Fable, Magistral) also burn reasoning tokens, but their multiplier hasn't been measured. */
+ *  Fable) also burn reasoning tokens, but their multiplier hasn't been measured. */
 declare function isHybridThinkingModel(modelApiName: string): boolean;
 interface CostCalculationOptions {
     cacheHitTokens?: number;
@@ -529,19 +528,26 @@ declare function getProviderSignatureFields(aiType: string, signature?: string):
  * - Gemini 3.x thinkingLevel: minimal|low|medium|high (3.1 Pro and 3.7 Flash reject minimal)
  * - Z.AI GLM-5.3 / 5.3-Flash: low|high|max only
  * - DeepSeek V4: low|high|max (the API itself aliases medium → high)
- * - Sakana Fugu: high|xhigh
+ * - Sakana Fugu (verified live 2026-09-20): high|xhigh|max on every model ("low" → 400 naming
+ *   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)
- * Qwen accepts reasoning_effort but ignores it (thinking_budget is its knob); MiniMax, Kimi,
- * Grok and Mistral expose no effort parameter.
+ * - 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.
+ * Qwen accepts reasoning_effort but ignores it (thinking_budget is its knob); MiniMax, Kimi and
+ * Grok expose no effort parameter.
  */
 type OpenAIReasoningEffort = 'minimal' | 'low' | 'medium' | 'high' | 'xhigh';
 type AnthropicReasoningEffort = 'low' | 'medium' | 'high' | 'xhigh' | 'max';
 type GeminiReasoningEffort = 'minimal' | 'low' | 'medium' | 'high';
 type GlmReasoningEffort = 'low' | 'high' | 'max';
 type DeepSeekReasoningEffort = 'low' | 'high' | 'max';
-type FuguReasoningEffort = 'high' | 'xhigh';
+type FuguReasoningEffort = 'high' | 'xhigh' | 'max';
 type MetaReasoningEffort = 'minimal' | 'low' | 'medium' | 'high' | 'xhigh' | 'max';
+type MistralReasoningEffort = 'minimal' | 'low' | 'medium' | 'high' | 'xhigh';
 /** The shared scale, lowest first. */
 declare const REASONING_EFFORT_SCALE: readonly ReasoningEffort[];
 declare const OPENAI_REASONING_EFFORTS: readonly OpenAIReasoningEffort[];
@@ -551,6 +557,7 @@ declare const GLM_REASONING_EFFORTS: readonly GlmReasoningEffort[];
 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[];
 /** 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;
@@ -560,6 +567,7 @@ declare const toGlmEffort: (effort: ReasoningEffort) => GlmReasoningEffort;
 declare const toDeepSeekEffort: (effort: ReasoningEffort) => DeepSeekReasoningEffort;
 declare const toMetaEffort: (effort: ReasoningEffort) => MetaReasoningEffort;
 declare const toFuguEffort: (effort: ReasoningEffort) => FuguReasoningEffort;
+declare const toMistralEffort: (effort: ReasoningEffort) => MistralReasoningEffort;
 
 /**
  * Unified token usage utilities for all AI providers
@@ -641,8 +649,10 @@ declare function extractAnthropicTokenUsage(response: any): TokenUsage | null;
 declare function extractGoogleTokenUsage(response: any): TokenUsage | null;
 /**
  * Mistral-specific token usage extraction
- * Mistral SDK uses camelCase (promptTokens, completionTokens, totalTokens)
- * and reasoning tokens may be in additionalProperties for Magistral models
+ * Mistral SDK uses camelCase (promptTokens, completionTokens, totalTokens). Reasoning tokens
+ * are not itemised today (they sit inside completionTokens, observed 2026-09-18 on Small 4 and
+ * Medium 3.5 with reasoning_effort: high); the additionalProperties probe below stays in case
+ * a reasoning_tokens field ever appears.
  */
 declare function extractMistralTokenUsage(response: any): TokenUsage | null;
 
@@ -1417,12 +1427,36 @@ declare class GoogleAgent extends AbstractAgent {
     private handleGeminiError;
 }
 
+/**
+ * Mistral agent (Mistral SDK chat.complete) for the hybrid Small 4 / Medium 3.5 models.
+ *
+ * Reasoning is off on the API by default and switched on per request with `reasoning_effort`
+ * (docs.mistral.ai/studio/conversations/reasoning). SDK 1.x has no typed field for it, so it is
+ * injected into the wire body by the same beforeRequest hook that adds `prompt_cache_key`.
+ * With it set, the assistant message comes back as content chunks — a `thinking` chunk (a list
+ * of text chunks holding the trace) followed by the `text` answer — and this works together
+ * with json_schema structured output (verified live 2026-09-18 on both models). Without the
+ * field the reply is a plain string and no trace exists.
+ *
+ * Multi-turn: Mistral asks that the full assistant message, thinking chunk included, be
+ * replayed into history ("stripping reasoning traces degrades performance"). The trace is
+ * plain text, so it rides on `AIMessage.thinking` — no provider signature field — and
+ * convertToMistralMessages rebuilds each prior assistant turn as [thinking, text] chunks.
+ *
+ * Reasoning tokens are counted inside completion_tokens; no separate reasoning_tokens field
+ * arrives, so cost accounting on the output rate already covers the trace.
+ */
 declare class MistralAgent extends AbstractAgent {
     private readonly client;
     private get defaultParams();
     private readonly logTemplates;
     private readonly errorMessages;
     constructor(name: string, instruction: string, model: string, apiKey: string, enableThinking?: boolean, agentLoggingConfig?: AgentLoggingConfig);
+    /**
+     * History in Mistral's shape. A prior assistant turn that carries a stored trace is
+     * replayed as [thinking, text] chunks, as Mistral's multi-turn guidance asks; every other
+     * turn is a plain string.
+     */
     private convertToMistralMessages;
     private processReply;
     private processStructuredReply;
@@ -1438,9 +1472,8 @@ declare class MistralAgent extends AbstractAgent {
      */
     doAskWithZodSchema<T>(zodSchema: z.ZodSchema<T>, messages: AIMessage[]): Promise<[T, string, TokenUsage$1?, string?]>;
     /**
-     * Plain-text ask: no schema appended, no responseFormat. Note that Magistral
-     * reasoning models only return thinking traces when responseFormat is NOT
-     * json_object, so unlike askWithZodSchema this path can surface thinking content.
+     * Plain-text ask: no schema appended, no responseFormat. Reasoning (and the trace) works
+     * the same way as on the schema path.
      */
     doAskText(messages: AIMessage[]): Promise<[string, string, TokenUsage$1?, string?]>;
 }
@@ -1620,7 +1653,6 @@ declare class FuguAgent extends AbstractAgent {
     private readonly errorMessages;
     constructor(name: string, instruction: string, model: string, apiKey: string, enableThinking?: boolean, agentLoggingConfig?: AgentLoggingConfig);
     private convertToOpenAIMessages;
-    private logRawUsageForCalibration;
     private extractThinkingAndUsage;
     private prependSystemInstruction;
     doAskWithZodSchema<T>(zodSchema: z.ZodSchema<T>, messages: AIMessage[]): Promise<[T, string, TokenUsage$1?, string?]>;
@@ -1690,4 +1722,4 @@ 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, MODEL_PRICING, MetaAgent, type MetaReasoningEffort, type MetaTokenUsage, MiniMaxAgent, MistralAgent, 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, 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, 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 };