@cognipeer/agent-sdk 0.10.6 → 0.13.0

package/dist/index.d.ts

@@ -1,5 +1,7 @@
 import { ZodError, ZodSchema, z, ZodTypeAny } from 'zod';
 import { JSONSchema7 } from 'json-schema';
+import { P as ProviderType, M as ModelRuleLearnedEvent, E as EmbeddingRequest, a as EmbeddingResponse, C as ChatCompletionRequest, b as ChatCompletionResponse, c as ChatCompletionChunk, J as JevProviderConfig, O as OpenAICompatibleProviderConfig, d as OpenAIProviderConfig, R as RulePatch, U as UnifiedMessage, T as ToolDefinition, e as TokenUsage, A as AnthropicProviderConfig, f as AzureProviderConfig, B as BedrockProviderConfig, V as VertexProviderConfig, g as ReasoningRequestConfig, h as ProviderConfig, i as ProviderError } from './model-rules-Bg9nSqyN.js';
+export { a4 as AudioContent, I as EffortMapping, a0 as EmbeddingInputType, $ as EmbeddingUsage, a3 as FileContent, Z as FinishReason, a2 as ImageContent, v as MAX_ADJUSTMENTS, k as MODEL_RULES, W as MaxTokensParam, D as ModelRule, F as ModelRules, Q as OffMechanism, Y as ParsedModelError, _ as ProviderResponseFields, a5 as ProviderToolCall, a6 as ProviderTransportConfig, H as ReasoningEffort, N as ReasoningStyle, G as ResolvedModelRule, K as RuleFamily, L as RuleSource, X as SamplingField, S as TemperatureRule, a1 as TextContent, n as allowsSampling, o as applyDropParams, u as asModelRequestError, z as clearLearnedModelRules, p as effortBudget, j as emptyUsage, x as getLearnedModelRulePatch, w as listLearnedModelRules, m as mapReasoningEffort, q as mergeRulePatch, l as modelIdCandidates, s as parseModelError, y as rememberModelRulePatch, r as resolveModelRule, t as runWithModelAdjustments } from './model-rules-Bg9nSqyN.js';
 
 /** Describes what a model/provider natively supports. */
 type ModelCapabilities = {
@@ -72,6 +74,7 @@ interface BaseChatMessage {
             name: string;
             arguments: string;
         };
+        thoughtSignature?: string;
     }>;
     tool_call_id?: string;
     usage?: any;
@@ -241,7 +244,7 @@ declare function describeContentPart(part: unknown): string | undefined;
  * `content`, else its JSON).
  */
 declare function contentPartToText(part: unknown, fallback?: (part: unknown) => string): string;
-declare function contentToString(content: string | ContentPart$1[] | any): string;
+declare function contentToString(content: string | ContentPart[] | any): string;
 declare function extractMessageText(message: any): string;
 /**
  * The text a hook should reason about: the text parts only, joined.
@@ -251,7 +254,7 @@ declare function extractMessageText(message: any): string;
  * that string would be scoring the SDK's own placeholder text, and a redactor
  * rewriting it would bake the placeholder into the message.
  */
-declare function textFromContent(content: string | ContentPart$1[] | unknown): string;
+declare function textFromContent(content: string | ContentPart[] | unknown): string;
 /**
  * Rewrite the text of a message's content, leaving every other part exactly as
  * it was — same objects, same order.
@@ -260,7 +263,7 @@ declare function textFromContent(content: string | ContentPart$1[] | unknown): s
  * string over a multi-part `content` is the silent version of this operation:
  * the run still answers, it just answers without the image the user attached.
  */
-declare function mapTextParts(content: string | ContentPart$1[] | unknown, fn: (text: string) => string): string | ContentPart$1[] | unknown;
+declare function mapTextParts(content: string | ContentPart[] | unknown, fn: (text: string) => string): string | ContentPart[] | unknown;
 /**
  * Put a whole replacement text back into a multi-part content.
  *
@@ -271,9 +274,9 @@ declare function mapTextParts(content: string | ContentPart$1[] | unknown, fn: (
  * plugin that needs exact multi-part fidelity should return `content` instead,
  * which is why that field exists.
  */
-declare function replaceTextContent(content: string | ContentPart$1[] | unknown, text: string): string | ContentPart$1[] | unknown;
+declare function replaceTextContent(content: string | ContentPart[] | unknown, text: string): string | ContentPart[] | unknown;
 /** Every non-text part of a message, normalized for policy code. */
-declare function collectAttachments(content: string | ContentPart$1[] | unknown): MediaAttachment[];
+declare function collectAttachments(content: string | ContentPart[] | unknown): MediaAttachment[];
 
 /**
  * Plugin layer — type surface.
@@ -381,7 +384,7 @@ interface HookMap {
             /** Concatenated text parts. Empty string for a media-only turn. */
             text: string;
             /** The full content, string or parts. Return `content` to rewrite it exactly. */
-            content: string | ContentPart$1[];
+            content: string | ContentPart[];
             /** Every non-text part, normalized across the unified and legacy shapes. */
             attachments: MediaAttachment[];
             message: Message;
@@ -397,7 +400,7 @@ interface HookMap {
              */
             text?: string;
             /** Full control over the parts: reorder, drop an attachment, rewrite per part. */
-            content?: string | ContentPart$1[];
+            content?: string | ContentPart[];
             /** Extra system-level context appended for this turn only. */
             additionalContext?: string;
             decision?: "allow" | "deny";
@@ -1127,7 +1130,7 @@ type ImageURL = {
     media_type?: string;
     detail?: 'auto' | 'low' | 'high';
 } | string;
-type ContentPart$1 = {
+type ContentPart = {
     type: 'text';
     text: string;
 } | {
@@ -1187,7 +1190,7 @@ type RunnableConfig = {
 type BaseMessage = {
     role: string;
     name?: string;
-    content: string | ContentPart$1[];
+    content: string | ContentPart[];
     tool_calls?: any;
     tool_call_id?: string;
     [key: string]: any;
@@ -1466,7 +1469,7 @@ type ReflectionConfig = {
 };
 type NativeReasoningConfig = {
     /** `none` is OFF — a value that gets SENT. See ReasoningEffort in providers/types.ts. */
-    effort?: "none" | "minimal" | "low" | "medium" | "high";
+    effort?: "none" | "minimal" | "low" | "medium" | "high" | "xhigh" | "max";
     budgetTokens?: number;
     includeThoughts?: boolean;
     providerExtras?: Record<string, any>;
@@ -2402,6 +2405,11 @@ type StreamEvent = {
     text: string;
     isFinal?: boolean;
 };
+/** Incremental model reasoning text (providers that expose it while streaming). */
+type ReasoningEvent = {
+    type: "reasoning";
+    text: string;
+};
 type CancelledEvent = {
     type: "cancelled";
     stage?: string;
@@ -2580,7 +2588,7 @@ type DelegationEventStamp = {
     /** Name of the `asTool` child agent this forwarded event came from. */
     delegatedTo?: string;
 };
-type SmartAgentEvent = (ToolCallEvent | ToolApprovalEvent | UserQuestionEvent | PlanEvent | SummarizationEvent | FinalAnswerEvent | MetadataEvent | HandoffEvent | GuardrailEvent | ProgressEvent | StreamEvent | CancelledEvent | SubagentEvent | ReflectionEvent | PluginEvent) & DelegationEventStamp;
+type SmartAgentEvent = (ToolCallEvent | ToolApprovalEvent | UserQuestionEvent | PlanEvent | SummarizationEvent | FinalAnswerEvent | MetadataEvent | HandoffEvent | GuardrailEvent | ProgressEvent | StreamEvent | ReasoningEvent | CancelledEvent | SubagentEvent | ReflectionEvent | PluginEvent) & DelegationEventStamp;
 type ReflectionEvent = {
     type: "reflection";
     id: string;
@@ -3302,345 +3310,6 @@ type CreateGuardrailOptions = {
 };
 declare function createGuardrail(options: CreateGuardrailOptions): ConversationGuardrail;
 
-type MessageRole = "system" | "user" | "assistant" | "tool";
-type TextContent = {
-    type: "text";
-    text: string;
-};
-type ImageContent = {
-    type: "image";
-    source: {
-        type: "base64";
-        mediaType: string;
-        data: string;
-    } | {
-        type: "url";
-        url: string;
-        mediaType?: string;
-    };
-};
-/**
- * A document attachment (PDF, DOCX, CSV, …). Providers map this to their
- * native document blocks: Gemini `inlineData`/`fileData`, Anthropic
- * `document`, OpenAI Chat Completions `file` / Responses `input_file`,
- * Bedrock Converse `document`. Providers without document support degrade
- * to a text placeholder referencing the file name/URL.
- */
-type FileContent = {
-    type: "file";
-    source: {
-        type: "base64";
-        mediaType: string;
-        data: string;
-    } | {
-        type: "url";
-        url: string;
-        mediaType?: string;
-    };
-    /** Original file name, when known (required by some providers, e.g. Bedrock). */
-    fileName?: string;
-};
-/**
- * An audio attachment. Gemini accepts it as `inlineData`/`fileData`; OpenAI
- * audio-capable models as `input_audio`. Providers without audio support
- * degrade to a text placeholder.
- */
-type AudioContent = {
-    type: "audio";
-    source: {
-        type: "base64";
-        mediaType: string;
-        data: string;
-    } | {
-        type: "url";
-        url: string;
-        mediaType?: string;
-    };
-};
-type ContentPart = TextContent | ImageContent | FileContent | AudioContent;
-type ToolCall = {
-    id: string;
-    name: string;
-    arguments: string;
-};
-/**
- * A provider-native reasoning/thinking block. Captured from the model response
- * and replayed verbatim on subsequent turns (required by Anthropic extended
- * thinking when tools are involved — the `signature` must round-trip intact).
- */
-type ReasoningBlock = {
-    type: "thinking" | "redacted_thinking";
-    /** Plain-text thinking content (absent for redacted blocks). */
-    text?: string;
-    /** Cryptographic signature Anthropic requires to be echoed back verbatim. */
-    signature?: string;
-    /** Opaque payload for redacted_thinking blocks. */
-    data?: string;
-};
-/** Captured reasoning payload attached to assistant turns / responses. */
-type ReasoningPayload = {
-    /** Verbatim provider blocks to replay on the next turn. */
-    blocks?: ReasoningBlock[];
-    /** Human-readable reasoning summary, when the provider exposes one. */
-    summary?: string;
-};
-type UnifiedMessage = {
-    role: MessageRole;
-    content: string | ContentPart[];
-    name?: string;
-    toolCalls?: ToolCall[];
-    toolCallId?: string;
-    /** Native reasoning blocks to replay verbatim (Anthropic/Bedrock thinking). */
-    reasoning?: ReasoningPayload;
-};
-type ToolDefinition = {
-    name: string;
-    description: string;
-    parameters: Record<string, any>;
-    strict?: boolean;
-};
-type ChatCompletionRequest = {
-    model: string;
-    messages: UnifiedMessage[];
-    temperature?: number;
-    maxTokens?: number;
-    topP?: number;
-    stop?: string[];
-    tools?: ToolDefinition[];
-    toolChoice?: "auto" | "required" | "none" | {
-        name: string;
-    };
-    responseFormat?: {
-        type: "json_schema" | "json_object" | "text";
-        schema?: Record<string, any>;
-        name?: string;
-    };
-    stream?: boolean;
-    extra?: Record<string, any>;
-    /**
-     * Unified native-reasoning request config. Each provider maps this to its
-     * own shape:
-     *  - OpenAI o-series:   body.reasoning_effort = effort
-     *  - OpenAI Responses:  body.reasoning = { effort }
-     *  - Anthropic:         body.thinking  = { type: "enabled", budget_tokens }
-     *  - Google / Vertex:   body.generationConfig.thinkingConfig = { thinkingBudget, includeThoughts }
-     *
-     * Providers silently ignore unsupported fields.
-     */
-    reasoning?: ReasoningRequestConfig;
-    /**
-     * Cancels the HTTP request (and an in-flight stream) when it aborts. The agent
-     * loop sets it from the run's cancellation token / abort signal; a caller using
-     * a provider directly can pass its own. It is never serialised into the body.
-     */
-    signal?: AbortSignal;
-};
-/**
- * `none` is OFF, and it is a value rather than the absence of one.
- *
- * On a thinking-by-default model — every Qwen3 under vLLM, gpt-5.1 and newer on
- * OpenAI — sending no reasoning field leaves the model thinking. So "do not
- * think" is an instruction that has to travel: `reasoning_effort:"none"` on
- * OpenAI, `thinkingBudget:0` on Gemini, and no thinking block at all on
- * Anthropic (where thinking is opt-in and the absence IS off). Omitting
- * `effort` still means "say nothing about reasoning", which is the different
- * and equally necessary state.
- */
-type ReasoningEffort = "none" | "minimal" | "low" | "medium" | "high";
-type ReasoningRequestConfig = {
-    /** Qualitative effort hint (mapped per-provider). */
-    effort?: ReasoningEffort;
-    /** Explicit per-call token budget for reasoning/thinking. */
-    budgetTokens?: number;
-    /** Whether the provider should expose its reasoning/thought summary to the caller. Default false. */
-    includeThoughts?: boolean;
-    /** Provider-specific passthrough merged on top of the mapped shape. */
-    providerExtras?: Record<string, any>;
-};
-type TokenUsage = {
-    inputTokens: number;
-    outputTokens: number;
-    totalTokens: number;
-    cachedInputTokens: number;
-    cachedWriteTokens: number;
-    cachedOutputTokens: number;
-    reasoningTokens: number;
-};
-declare function emptyUsage(): TokenUsage;
-type FinishReason = "stop" | "tool_calls" | "length" | "content_filter" | "error";
-type ChatCompletionResponse = {
-    id: string;
-    model: string;
-    content: string | null;
-    toolCalls: ToolCall[];
-    usage: TokenUsage;
-    finishReason: FinishReason;
-    /** Native reasoning blocks + summary captured from the response, if any. */
-    reasoning?: ReasoningPayload;
-    raw: any;
-};
-type ChatCompletionChunk = {
-    id: string;
-    model: string;
-    delta: {
-        content?: string;
-        toolCalls?: Partial<ToolCall>[];
-        /** Incremental reasoning text / signature deltas (Anthropic thinking). */
-        reasoning?: ReasoningPayload;
-    };
-    usage?: TokenUsage;
-    finishReason?: FinishReason;
-};
-type ProviderRetryConfig = {
-    /** Maximum retries on 429/5xx (default 3). */
-    maxRetries?: number;
-    /** Initial delay between retries in ms (default 500, doubled each attempt). */
-    baseDelayMs?: number;
-    /** Hard cap on a single retry delay in ms (default 8000). */
-    maxDelayMs?: number;
-    /** Custom predicate to opt in/out of a specific retry. */
-    shouldRetry?: (error: any, attempt: number) => boolean;
-};
-type OpenAIProviderConfig = {
-    provider: "openai";
-    apiKey: string;
-    baseURL?: string;
-    organization?: string;
-    defaultModel?: string;
-    defaultHeaders?: Record<string, string>;
-    retry?: ProviderRetryConfig;
-    /**
-     * WHICH API a reasoning-carrying request is addressed to.
-     *
-     *  auto    (default, and today's behaviour) route to the Responses API when
-     *          the request carries `reasoning` AND the model name looks like an
-     *          o-series/gpt-5 reasoning model.
-     *  never   always Chat Completions. `reasoning` is still mapped, as
-     *          `reasoning_effort` plus any `providerExtras`.
-     *  always  always Responses, whatever the model is called.
-     *
-     * `auto` is a name-based guess, and it decides far more than where reasoning
-     * goes: Responses is a different request shape (`input` not `messages`,
-     * `text.format` not `response_format`, flat tool objects) and a different
-     * response parse. A caller with its own measurements of the Chat Completions
-     * body — a gateway, a proxy, an on-prem server that serves one and not the
-     * other, or anything that has tuned `extra` — must be able to say so, rather
-     * than discover that asking for less thinking silently changed the transport.
-     *
-     * Applies to any OpenAI-shaped endpoint, which is also why the guess is weak
-     * where it matters most: a self-hosted model is named by whoever served it.
-     */
-    responsesApi?: "auto" | "never" | "always";
-    /**
-     * Models that go to the Responses API, named explicitly. Opt-in; unset by
-     * default, and nothing here is inferred from a model name.
-     *
-     * A string matches the request's model (or `defaultModel`) exactly; a RegExp
-     * is tested against it. A match routes the request to Responses whether or
-     * not it carries `reasoning` — the case this exists for is a model that
-     * reasons by default and rejects function tools on Chat Completions, where
-     * sending no reasoning field does not help. `responsesApi: "never"` still
-     * wins; under `"always"` the routing is unchanged.
-     *
-     * Requests for a matched model also get, on the Responses body only:
-     *  - Chat Completions spellings arriving through `extra`, which /responses
-     *    rejects with a 400, re-spelled with their values unchanged:
-     *    `reasoning_effort` → `reasoning.effort` (a request `reasoning` effort
-     *    still wins), `max_tokens` / `max_completion_tokens` →
-     *    `max_output_tokens`, `response_format` → `text.format`, and
-     *    `stream_options.include_usage` dropped.
-     *  - real SSE streaming, when `responsesStreaming` is also true.
-     *
-     * Honoured by `OpenAIProvider` itself only; `AzureProvider` and
-     * `OpenAICompatibleProvider` ignore it.
-     */
-    responsesApiModels?: Array<string | RegExp>;
-    /**
-     * Stream Responses API output as it arrives (text and tool-call argument
-     * deltas) for models in `responsesApiModels`. Default false, which keeps the
-     * existing behaviour: one non-streaming Responses call, emitted as a single
-     * chunk. Has no effect on any other request.
-     */
-    responsesStreaming?: boolean;
-};
-type AnthropicProviderConfig = {
-    provider: "anthropic";
-    apiKey: string;
-    baseURL?: string;
-    defaultModel?: string;
-    defaultHeaders?: Record<string, string>;
-    anthropicVersion?: string;
-    retry?: ProviderRetryConfig;
-    /**
-     * Prompt caching configuration. When enabled, places cache_control:
-     * ephemeral breakpoints on the system prompt and the final tool definition,
-     * which is the recommended Anthropic recipe for static system + tools.
-     */
-    prompt_caching?: {
-        enabled: boolean;
-    };
-};
-type AzureProviderConfig = {
-    provider: "azure";
-    apiKey: string;
-    endpoint: string;
-    apiVersion?: string;
-    deploymentName?: string;
-    defaultModel?: string;
-    defaultHeaders?: Record<string, string>;
-};
-type OpenAICompatibleProviderConfig = {
-    provider: "openai-compatible";
-    apiKey: string;
-    baseURL: string;
-    defaultModel?: string;
-    defaultHeaders?: Record<string, string>;
-};
-type BedrockProviderConfig = {
-    provider: "bedrock";
-    region: string;
-    accessKeyId?: string;
-    secretAccessKey?: string;
-    sessionToken?: string;
-    defaultModel?: string;
-    defaultHeaders?: Record<string, string>;
-    retry?: ProviderRetryConfig;
-    /**
-     * Insert Bedrock Converse `cachePoint` blocks at the end of the system
-     * prompt and tools list. Only supported by models that implement Converse
-     * prompt caching (most Anthropic Claude variants on Bedrock).
-     */
-    prompt_caching?: {
-        enabled: boolean;
-    };
-};
-type VertexProviderConfig = {
-    provider: "vertex";
-    projectId: string;
-    location?: string;
-    accessToken?: string;
-    serviceAccountJson?: string | Record<string, any>;
-    defaultModel?: string;
-    defaultHeaders?: Record<string, string>;
-};
-type JevProviderConfig = {
-    provider: "jev";
-    apiKey: string;
-    /** Defaults to https://api.typesafe.ai. */
-    baseURL?: string;
-    defaultModel?: string;
-    defaultHeaders?: Record<string, string>;
-    retry?: ProviderRetryConfig;
-};
-type ProviderConfig = OpenAIProviderConfig | AnthropicProviderConfig | AzureProviderConfig | OpenAICompatibleProviderConfig | BedrockProviderConfig | VertexProviderConfig | JevProviderConfig;
-type ProviderType = ProviderConfig["provider"];
-declare class ProviderError extends Error {
-    readonly provider: string;
-    readonly statusCode?: number | undefined;
-    readonly responseBody?: any | undefined;
-    constructor(message: string, provider: string, statusCode?: number | undefined, responseBody?: any | undefined);
-}
-
 type ProviderRetryPolicy = {
     /** Maximum number of retries on 429/5xx responses. Default: 3. */
     maxRetries?: number;
@@ -3657,6 +3326,33 @@ type ProviderRetryPolicy = {
 declare abstract class BaseProvider {
     abstract readonly providerName: ProviderType;
     protected retryPolicy: ProviderRetryPolicy;
+    /** Injected `fetch` (config.fetch); falls back to the global one at call time. */
+    protected fetchImpl?: typeof fetch;
+    /** Per-request timeout in ms (config.timeoutMs). Undefined = no timeout. */
+    protected timeoutMs?: number;
+    /** Called when a 400 taught the provider a per-model adjustment (config.onModelRuleLearned). */
+    protected onModelRuleLearned?: (event: ModelRuleLearnedEvent) => void;
+    /** Applies the transport options shared by every provider config. */
+    protected configureTransport(config: {
+        fetch?: typeof fetch;
+        timeoutMs?: number;
+        onModelRuleLearned?: (event: ModelRuleLearnedEvent) => void;
+    }): void;
+    /**
+     * The single HTTP entry point for every provider call site. Uses the
+     * configured `fetch` and applies `timeoutMs`: the timer covers the time until
+     * the response HEADERS arrive and is cleared after that, so a long-running
+     * stream is never cut by it (use the request `signal` to cancel a stream).
+     * A timeout surfaces as a `ProviderError` with status 408. The caller's
+     * `signal` still aborts the request as before.
+     */
+    protected httpFetch(url: string, init?: RequestInit): Promise<Response>;
+    /**
+     * Embeds `request.input`. Implemented by the providers that have an
+     * embeddings endpoint (OpenAI, Azure, openai-compatible, Bedrock, Vertex);
+     * the others reject with a 400 ProviderError.
+     */
+    embed(_request: EmbeddingRequest): Promise<EmbeddingResponse>;
     /** Non-streaming completion */
     abstract complete(request: ChatCompletionRequest): Promise<ChatCompletionResponse>;
     /** Streaming completion – yields incremental chunks */
@@ -5674,6 +5370,9 @@ declare function buildResponseFormatSection(responseFormat: Record<string, any>
  */
 declare function startStreamingSession(session: TraceSessionRuntime | undefined, agentRuntime?: AgentRuntimeConfig): Promise<void>;
 
+/** True when the model's rule uses `thinking:{type:"adaptive"}` + `output_config.effort` (Claude 4.6+ / 5.x). */
+declare function supportsAdaptiveThinking(model: unknown): boolean;
+
 declare class OpenAIProvider extends BaseProvider {
     readonly providerName: ProviderType;
     protected readonly apiKey: string;
@@ -5683,6 +5382,11 @@ declare class OpenAIProvider extends BaseProvider {
     protected readonly responsesApi: "auto" | "never" | "always";
     private readonly responsesApiModels;
     private readonly responsesStreaming;
+    private readonly apiExplicitChat;
+    private readonly apiExplicitResponses;
+    /** `api: "responses"` was passed (the legacy `responsesApi: "always"` keeps its old single-call behaviour). */
+    private readonly apiResponsesNative;
+    protected readonly maxTokensParam: "auto" | "max_tokens" | "max_completion_tokens";
     constructor(config: OpenAIProviderConfig);
     /** True when the caller listed this request's model in `responsesApiModels`. */
     private isResponsesApiModel;
@@ -5701,7 +5405,14 @@ declare class OpenAIProvider extends BaseProvider {
     protected useResponsesApi(request: ChatCompletionRequest): boolean;
     complete(request: ChatCompletionRequest): Promise<ChatCompletionResponse>;
     completeStream(request: ChatCompletionRequest): AsyncGenerator<ChatCompletionChunk, void, unknown>;
-    protected buildRequestBody(request: ChatCompletionRequest, stream: boolean): Record<string, any>;
+    /** POST /embeddings, batched (default 2048 inputs per request, the OpenAI cap). */
+    embed(request: EmbeddingRequest): Promise<EmbeddingResponse>;
+    protected embeddingsFetch(body: Record<string, any>, signal?: AbortSignal): Promise<Response>;
+    /** Chat Completions call that corrects itself from the provider's own 400s (see models/adjust.ts). */
+    private fetchChat;
+    /** Same for the Responses API. */
+    private fetchResponses;
+    protected buildRequestBody(request: ChatCompletionRequest, stream: boolean, overlay?: RulePatch): Record<string, any>;
     protected completeResponses(request: ChatCompletionRequest): Promise<ChatCompletionResponse>;
     /** Real Responses SSE streaming, for models opted in through
      * `responsesApiModels` with `responsesStreaming` on. Emits the same chunk
@@ -5711,7 +5422,7 @@ declare class OpenAIProvider extends BaseProvider {
      * (taken from the completed response, exactly as the non-streaming parse
      * reads it). */
     private streamResponses;
-    protected buildResponsesBody(request: ChatCompletionRequest): Record<string, any>;
+    protected buildResponsesBody(request: ChatCompletionRequest, overlay?: RulePatch): Record<string, any>;
     /** Maps a unified message to a Responses API `input` item. */
     protected toResponsesInput(m: UnifiedMessage): Record<string, any>;
     protected responsesFetch(body: Record<string, any>, signal?: AbortSignal): Promise<Response>;
@@ -5731,6 +5442,7 @@ declare class AnthropicProvider extends BaseProvider {
     private readonly defaultHeaders;
     private readonly anthropicVersion;
     private readonly promptCachingEnabled;
+    private readonly thinkingMode;
     constructor(config: AnthropicProviderConfig);
     complete(request: ChatCompletionRequest): Promise<ChatCompletionResponse>;
     completeStream(request: ChatCompletionRequest): AsyncGenerator<ChatCompletionChunk, void, unknown>;
@@ -5750,6 +5462,9 @@ declare class AnthropicProvider extends BaseProvider {
     private toAnthropicToolResult;
     private toAnthropicMessage;
     private toAnthropicTool;
+    /** Builds and sends the request; retries once with the other thinking form
+     * when the upstream rejects the one we picked. */
+    private send;
     private doFetch;
     private parseResponse;
 }
@@ -5762,6 +5477,9 @@ declare class AzureProvider extends OpenAIProvider {
     private readonly azureApiKey;
     constructor(config: AzureProviderConfig);
     protected doFetch(body: Record<string, any>, signal?: AbortSignal): Promise<Response>;
+    /** Azure embeddings: the deployment named by the request's model (falling back
+     * to the configured deployment) in the path, api-key auth. */
+    protected embeddingsFetch(body: Record<string, any>, signal?: AbortSignal): Promise<Response>;
     /** Azure exposes the Responses API at /openai/responses with api-key auth.
      * Reasoning summaries require a preview api-version; callers can override via
      * the `apiVersion` config. */
@@ -5780,9 +5498,28 @@ declare class BedrockProvider extends BaseProvider {
     private readonly defaultModel;
     private readonly defaultHeaders;
     private readonly promptCachingEnabled;
+    private readonly thinkingMode;
     constructor(config: BedrockProviderConfig);
     complete(request: ChatCompletionRequest): Promise<ChatCompletionResponse>;
+    /**
+     * Real streaming over ConverseStream. The response is an AWS event-stream
+     * (binary frames); each frame carries one JSON event: messageStart,
+     * contentBlockStart/Delta/Stop (text, toolUse, reasoningContent),
+     * messageStop and, last, metadata (usage). Text, tool-call argument and
+     * reasoning deltas are yielded as they arrive; one closing chunk carries the
+     * finish reason, usage and the signed reasoning blocks.
+     */
     completeStream(request: ChatCompletionRequest): AsyncGenerator<ChatCompletionChunk, void, unknown>;
+    /**
+     * InvokeModel embeddings. Supported families (by model id, inference-profile
+     * prefixes such as `us.` allowed):
+     *  - Amazon Titan Text Embeddings (v1, v2): one text per call, `dimensions`
+     *    (v2: 256/512/1024) and `normalize` via `extra`.
+     *  - Cohere Embed (v3, v4): a batch of texts per call (default 96), `inputType`
+     *    maps to `input_type`, `dimensions` to `output_dimension` (v4).
+     */
+    embed(request: EmbeddingRequest): Promise<EmbeddingResponse>;
+    private invokeModel;
     private buildConverseBody;
     private splitSystemMessages;
     /**
@@ -5800,6 +5537,9 @@ declare class BedrockProvider extends BaseProvider {
     private toBedrockToolResult;
     private toBedrockMessage;
     private toBedrockTool;
+    /** Sends the Converse call; retries once with toolChoice auto when the
+     * model rejects forced tool use. */
+    private converseFetch;
     private doFetch;
     private parseResponse;
 }
@@ -5816,6 +5556,13 @@ declare class VertexProvider extends BaseProvider {
     constructor(config: VertexProviderConfig);
     complete(request: ChatCompletionRequest): Promise<ChatCompletionResponse>;
     completeStream(request: ChatCompletionRequest): AsyncGenerator<ChatCompletionChunk, void, unknown>;
+    /**
+     * Vertex `:predict` embeddings (text-embedding-00x, text-multilingual-embedding,
+     * gemini-embedding-001). `inputType` maps to the task type; `dimensions` to
+     * `outputDimensionality`. gemini-embedding takes one instance per request, the
+     * text-embedding models up to `batchSize` (default 100).
+     */
+    embed(request: EmbeddingRequest): Promise<EmbeddingResponse>;
     private buildUrl;
     private buildGeminiBody;
     private buildContents;
@@ -5840,7 +5587,50 @@ type NativeModelOptions = {
     extra?: Record<string, any>;
     /** Unified native reasoning config applied by default (per-call overrides supported). */
     reasoning?: ReasoningRequestConfig;
+    /** Default nucleus sampling. */
+    topP?: number;
+    /** Default stop sequences. */
+    stop?: string[];
+    /** Default presence penalty (OpenAI-family only). */
+    presencePenalty?: number;
+    /** Default frequency penalty (OpenAI-family only). */
+    frequencyPenalty?: number;
+    /**
+     * Last-mile hook: receives the fully built request (tools, response format,
+     * reasoning already applied) right before it goes to the provider. Mutate it
+     * in place or return a replacement. Use it for model-specific rules the
+     * unified request cannot express (e.g. force `reasoning` off when tools are bound).
+     */
+    transformRequest?: (request: ChatCompletionRequest) => ChatCompletionRequest | void | Promise<ChatCompletionRequest | void>;
+};
+/**
+ * The ONE usage shape carried on every assistant message produced by a native
+ * provider (`message.usage`, and mirrored on `response_metadata.token_usage`):
+ *
+ * ```
+ * {
+ *   prompt_tokens, completion_tokens, total_tokens,
+ *   prompt_tokens_details:     { cached_tokens, cache_write_tokens },
+ *   completion_tokens_details: { reasoning_tokens },
+ * }
+ * ```
+ * `completion_tokens` INCLUDES `reasoning_tokens`; `prompt_tokens` INCLUDES
+ * `cached_tokens`. `response_metadata.token_usage` additionally keeps the flat
+ * legacy `cached_tokens` / `cache_write_tokens` keys.
+ */
+type MessageUsage = {
+    prompt_tokens: number;
+    completion_tokens: number;
+    total_tokens: number;
+    prompt_tokens_details: {
+        cached_tokens: number;
+        cache_write_tokens: number;
+    };
+    completion_tokens_details: {
+        reasoning_tokens: number;
+    };
 };
+declare function toMessageUsage(u: TokenUsage): MessageUsage;
 /**
  * Creates a BaseChatModel from a native provider.
  *
@@ -5877,4 +5667,26 @@ declare function fromNativeProvider(provider: BaseProvider, options?: NativeMode
  */
 declare function createProvider(config: ProviderConfig): BaseProvider;
 
-export { type AIMessage, ASK_USER_TOOL_NAME, type AgentInstance, type AgentInvokeResult, type AgentLimits, type AgentOptions, type AgentPlugin, type AgentRuleOptions, type AgentRuntimeConfig, type AgentSnapshot, type AgentSnapshotMetadata, type AgentState, AnthropicProvider, type AnthropicProviderConfig, type ApprovalTransport, type AudioContent, type AuditEntry, type AuditLogConfig, type AzureContentSafetyConfig, AzureProvider, type AzureProviderConfig, type BaseChatMessage, type BaseChatMessagePart, type BaseChatModel, type BaseMessage, BaseProvider, type BedrockGuardrailConfig, BedrockProvider, type BedrockProviderConfig, type BudgetGuardConfig, type BuildChildConfig, type BuiltInRuntimeProfile, CCRStore, CONSOLE_HOOK_MAP, CONTEXT_PILOT_JSON_MARKER_KEY, CONTROL_PLANE_TOOL_NAMES, type CancellationTokenLike, type CancelledEvent, type CapabilityStatus, type ChatCompletionChunk, type ChatCompletionRequest, type ChatCompletionResponse, type CheckpointPluginConfig, type CheckpointStore, type ChildContextPolicy, type CodeDetectionRuleOptions, type CodeGuardrailOptions, type CognipeerGuardrailConfig, type CompressToolOutputParams, type ContextBuilder, type ContextPilotCCRConfig, type ContextPilotCacheAlignmentConfig, type ContextPilotCompressionConfig, type ContextPilotCompressionStats, type ContextPilotCompressorKind, type ContextPilotConfig, type ContextPilotDedupConfig, type ContextPilotDiffCompressionConfig, type ContextPilotJsonCompressionConfig, type ContextPilotLogCompressionConfig, type ContextPilotResult, type ContextPilotRuntime, type ContextPilotSearchCompressionConfig, type ContextPilotTextCompressionConfig, type ContextPolicy, type ContextToolsStateRef, type ConversationGuardrail, type ConversationPluginConfig, type ConversationStore, type CostEstimator, type CustomRuleCallback, type CustomRuleOptions, DEFAULT_MAX_TOOL_INPUT_FIELD_CHARS, DEFAULT_PROFILE_CONFIGS, DEFAULT_SKILL_POLICY, DEFAULT_SUBAGENT_POLICY, DEFAULT_SUBAGENT_TOOL_NAMES, DEFAULT_TOOL_INPUT_DIGEST_HEAD_CHARS, DELEGATION_TOOL_NAMES, DedupTracker, type DelegationEventStamp, type DelegationMode, type EvalCase, type EvalCaseResult, type EvalFamily, type EvalHarnessMetrics, type EvalHarnessParams, type EvalHarnessResult, type EvalProfileDescriptor, type EvalProfileTarget, type FileContent, type FinalAnswerEvent, type FinishReason, type GateResult, type GuardrailCallContext, type GuardrailContext, type GuardrailDisposition, type GuardrailEvent, type GuardrailIncident, type GuardrailJudgeEngine, type GuardrailOutcome, GuardrailPhase, type GuardrailPhaseName, type GuardrailPluginOptions, type GuardrailRequest, type GuardrailRule, type GuardrailRuleResult, type GuardrailTransport, type GuardrailVerdict, type GuardrailViolation, HOOK_CONTRACT_VERSION, HOOK_NAMES, type HandoffDescriptor, type HandoffEvent, type HookContext, type HookDecision, type HookHandler, type HookMap, type HookName, type HookRegistrations, HookTimeoutError, type HttpGuardrailOptions, type HumanInTheLoopAskUserConfig, type HumanInTheLoopOptions, type HumanInTheLoopRuntimeConfig, type ImageContent, InMemoryMemoryStore, type InjectionScan, type InvokeConfig, type JevClientLike, type JevEngineOptions, type JevNoulAnswer, JevProvider, type JevProviderConfig, type JevQuestion, type JevSystemOneRequest, type JevSystemOneResult, type JsonGuardrailOptions, type JsonSchemaRuleOptions, type LangfuseBuildContext, type LangfuseEvent, type LangfuseEventKind, type LangfuseTracingConfig, type LanguageGuardConfig, type LoadSkillsOptions, type MaybePromise, type McpConfig, type McpConnection, type MediaAttachment, type MediaKind$1 as MediaKind, type MediaPolicyConfig, type MemoryFact, type MemoryProviderKind, type MemoryReadPolicy, type MemoryScope, type MemoryStore, type MemoryWritePolicy, type Message, type MetadataEvent, type ModelCapabilities, NativeJsonSchemaStrategy, type NativeModelOptions, type NativeReasoningConfig, OpenAICompatibleProvider, type OpenAICompatibleProviderConfig, type OpenAIModerationConfig, OpenAIProvider, type OpenAIProviderConfig, type OtelTracingConfig, type OutputGuardConfig, PREOPEN_TOOL_CALL_PREFIX, type PendingToolApproval, type PendingUserQuestion, type PiiDetector, type PiiEntity, type PiiRedactionConfig, type PlanEvent, type PlanStepRecord, type PlanningMode, type PluginContributions, type PluginEvent, type PluginFailureMode, type PluginHost, type PluginHostOptions, type PluginLogger, type PluginProvides, type PluginRunHost, type PluginSetupContext, type PortkeyGatewayConfig, type PortkeyGuardrailConfig, type PreopenSkillsResult, type ProfileConfig, type ProgressEvent, type ProgressUpdate, type PromptHooks, type PromptInjectionFamily, type PromptInjectionGuardConfig, type PromptSource, type ProviderConfig, ProviderError, type ToolCall as ProviderToolCall, type ToolDefinition as ProviderToolDefinition, type ProviderType, type RateLimitConfig, type ReasoningConfig, type ReasoningLevel, type RedactionResult, type ReflectionCadence, type ReflectionConfig, type ReflectionEvent, type ReflectionHookContext, type ReflectionRecord, type RegexGuardrailConfig, type RegexGuardrailOptions, type RegexRuleOptions, type ReplanPolicy, type ResolvedContextPilotConfig, type ResolvedSmartAgentConfig, type ResolvedTraceConfig, type ResolvedTraceSink, type ResponseCacheConfig, type RestoreSnapshotOptions, type RuntimeProfile, SMALL_TIER_SKILL_POLICY, type SdkCapabilities, type SerializableSmartState, type SessionMetrics, type SessionMetricsConfig, type Skill, type SkillDisclosure, type SkillExecutionContext, type SkillFrontmatter, type SkillFs, type SkillModelTier, type SkillPolicy, type SkillRegistryRef, type SkillSource, type SkillToolDeps, type SkillToolHeader, type SlotName, type SmartAgentBudgetConfig, type SmartAgentContextConfig, type SmartAgentCustomProfileConfig, type SmartAgentDelegationConfig, type SmartAgentEvent, type SmartAgentInstance, type SmartAgentLimits, type SmartAgentMemoryConfig, type SmartAgentOptions, type SmartAgentPlanningConfig, type SmartAgentSummarizationConfig, type SmartAgentToolResponseConfig, type SmartAgentTracingConfig, type SmartModel, type SmartState, type SmartToolFn, type SnapshotOptions, type SnapshotRuntimeHint, type StreamChunk, type StreamEvent, type StrictRestorers, type StrictTransform, type StructuredOutputError, type StructuredOutputFieldError, StructuredOutputManager, type StructuredOutputManagerConfig, type StructuredOutputResult, type StructuredOutputStrategy, type StructuredSummary, type SubagentDef, type SubagentEvent, type SubagentMode, type SubagentPendingRecord, type SubagentPolicy, type SubagentRegistryRef, type SubagentResult, type SubagentToolDeps, type SubagentToolNames, type SummarizationEvent, type SummarizationMode, type SummarizerInput, type SummarizerResult, type SummarizerStrategy, type SummaryFactItem, type SummaryIntegrityCheck, TOOL_INPUT_DIGEST_KEY, type TextContent, type TokenCounter, type TokenUsage, type ToolApprovalEvent, type ToolApprovalResolution, type ToolApprovalStatus, ToolBasedStrategy, type ToolCachePolicy, type ToolCallEvent, type ToolInputDigestOptions, type ToolInputDigestResult, type ToolInputRetentionPolicy, type ToolInterface, type ToolPolicyConfig, type ToolResponseClassification, type ToolResponseRetentionPolicy, type ToolRetentionDeclarations, type ToolRetentionSpec, type ToolRetryPolicy, type ToolRule, type TraceDataSection, type TraceErrorRecord, type TraceEventRecord, type TraceMessagePart, type TraceMessageSection, type TraceMetadataSection, type TraceSessionConfigSnapshot, type TraceSessionFile, type TraceSessionStatus, type TraceSessionSummary, type TraceSinkCognipeerConfig, type TraceSinkConfig, type TraceSinkCustomConfig, type TraceSinkFileConfig, type TraceSinkHttpConfig, type TraceSinkOtlpConfig, type TraceSinkSnapshot, type TraceSummarySection, type TraceToolCallSection, type TraceToolDefinition, type TraceToolDefinitionsSection, type TraceToolDetails, type TraceToolResponseSection, type TraceToolResultSection, type TracingConfig, type TracingMode, type UnifiedMessage, type UserQuestionAnswer, type UserQuestionAnswerSet, type UserQuestionEvent, type UserQuestionItem, type UserQuestionOption, type UserQuestionResolution, type UserQuestionStatus, VertexProvider, type VertexProviderConfig, type VolatileContentFinding, type VolatileContentPattern, type WebhookApprovalConfig, agentVerdictRule, appendBoundTools, auditLog, azureContentSafety, bedrockGuardrail, budgetGuard, buildHandoffTools, buildResponseFormatSection, buildSkillHeaderBlock, buildSubagentCatalogBlock, buildSystemPrompt, buildToolDefinitionsSection, canOpenSkill, canSpawn, captureSnapshot, checkpointing, codePresenceRule, coerceToolArgs, cognipeerGuardrail, cognipeerSink, collectAttachments, collectToolRetentionDeclarations, composeToolSets, compressDiffOutput, compressJsonOutput, compressLogOutput, compressSearchOutput, compressTextOutput, compressToolOutput, contentPartToText, contentToString, conversationHistory, countApproxTokens, countMessagesTokens, createAgent, createAgentNode, createAskUserQuestionTool, createBindSkillToolsTool, createCodeGuardrail, createContextPilotRuntime, createContextSummarizeNode, createContextTools, createDelegateTool, createGetToolResponseTool, createGuardrail, createGuardrailPlugin, createJsonGuardrail, createManageTodoListAliasTool, createManageTodoTool, createOpenSkillTool, createParallelTool, createPluginHost, createProvider, createRegexGuardrail, createResolverNode, createSearchSkillsTool, createSkillRegistryRef, createSkillTools, createSmartAgent, createSpawnTool, createSubagentRegistryRef, createSubagentTools, createTool, createToolLimitFinalizeNode, createToolsNode, customCallbackRule, customGuardrail, customSink, dedupeToolsByName, defaultTokenCounter, defineHook, definePlugin, describeContentPart, detectLanguage, detectTextFormat, detectVolatileContent, digestToolInputArguments, digestToolInputValue, emptyUsage, evaluateGuardrails, extractLatestUserQuery, extractMessageText, fileCheckpointStore, fileSink, findSubagent, formatZodError, fromLangchainModel, fromLangchainTools, fromNativeProvider, generateSpanId, generateTraceId, getModelCapabilities, getNestedValue, getResolvedSmartConfig, getTokenCounter, hasMediaPart, hasToolResponseRecoveryReference, httpGuardrail, httpSink, inMemoryCheckpointStore, inMemoryConversationStore, isHookImplemented, isOpenAIHostedEndpoint, isSmartModel, isValidIban, isValidLuhn, isValidTckn, jsonSchemaRule, defaultBuildPayload as langfuseDefaultPayload, langfuseTracing, languageGuard, loadSkillsFromDisk, mapConsoleHookVerdict, mapTextParts, mcp, mediaKindOf, mediaPolicy, normalizeSmartAgentOptions, normalizeVerdicts, openAIModeration, otelTracing, otlpSink, outputGuard, parseApprovalDecision, parseSkillFrontmatter, pathSandbox, piiRedaction, pluginCapabilities, portkeyGateway, portkeyGuardrail, preopenSkills, preopenToolCallId, prepareStrictToolMenu, promptInjectionGuard, rateLimit, readMemoryFacts, redactText, regexGuardrail, regexRule, replaceTextContent, resolveAvailableSkills, resolveAvailableSubagents, resolveInputRetention, resolveMemoryStore, resolveReasoning, resolveStrategy, resolveSubagentTools, resolveSummarizationRetention, resolveToolApprovalState, resolveUserQuestionState, responseCache, restoreSnapshot, restoreToolCalls, runSmartAgentEvalHarness, scoreItemsByRelevance, searchSkills, seedChildMessages, selectTopIndicesInOrder, sessionMetrics, setTokenCounter, startStreamingSession, summarizeObject, textFromContent, toStrictCompatible, toStrictToolSchema, tokenize, toolPolicy, traceSessionToOtlp, validateReasoningConfig, validateToolArgs, webhookApproval, withTools, withinDepth, writeSummaryFactsToMemory };
+type ModelRequestErrorCode = "unsupported_parameter" | "unsupported_value" | "model_request_invalid";
+type ModelRequestErrorInit = {
+    status?: number;
+    code: ModelRequestErrorCode;
+    param?: string;
+    value?: unknown;
+    supportedValues?: string[];
+    provider: string;
+    model: string;
+    message: string;
+    responseBody?: unknown;
+};
+declare class ModelRequestError extends ProviderError {
+    readonly status: number;
+    readonly code: ModelRequestErrorCode;
+    readonly param?: string;
+    readonly value?: unknown;
+    readonly supportedValues?: string[];
+    readonly model: string;
+    constructor(init: ModelRequestErrorInit);
+}
+
+export { type AIMessage, ASK_USER_TOOL_NAME, type AgentInstance, type AgentInvokeResult, type AgentLimits, type AgentOptions, type AgentPlugin, type AgentRuleOptions, type AgentRuntimeConfig, type AgentSnapshot, type AgentSnapshotMetadata, type AgentState, AnthropicProvider, AnthropicProviderConfig, type ApprovalTransport, type AuditEntry, type AuditLogConfig, type AzureContentSafetyConfig, AzureProvider, AzureProviderConfig, type BaseChatMessage, type BaseChatMessagePart, type BaseChatModel, type BaseMessage, BaseProvider, type BedrockGuardrailConfig, BedrockProvider, BedrockProviderConfig, type BudgetGuardConfig, type BuildChildConfig, type BuiltInRuntimeProfile, CCRStore, CONSOLE_HOOK_MAP, CONTEXT_PILOT_JSON_MARKER_KEY, CONTROL_PLANE_TOOL_NAMES, type CancellationTokenLike, type CancelledEvent, type CapabilityStatus, ChatCompletionChunk, ChatCompletionRequest, ChatCompletionResponse, type CheckpointPluginConfig, type CheckpointStore, type ChildContextPolicy, type CodeDetectionRuleOptions, type CodeGuardrailOptions, type CognipeerGuardrailConfig, type CompressToolOutputParams, type ContextBuilder, type ContextPilotCCRConfig, type ContextPilotCacheAlignmentConfig, type ContextPilotCompressionConfig, type ContextPilotCompressionStats, type ContextPilotCompressorKind, type ContextPilotConfig, type ContextPilotDedupConfig, type ContextPilotDiffCompressionConfig, type ContextPilotJsonCompressionConfig, type ContextPilotLogCompressionConfig, type ContextPilotResult, type ContextPilotRuntime, type ContextPilotSearchCompressionConfig, type ContextPilotTextCompressionConfig, type ContextPolicy, type ContextToolsStateRef, type ConversationGuardrail, type ConversationPluginConfig, type ConversationStore, type CostEstimator, type CustomRuleCallback, type CustomRuleOptions, DEFAULT_MAX_TOOL_INPUT_FIELD_CHARS, DEFAULT_PROFILE_CONFIGS, DEFAULT_SKILL_POLICY, DEFAULT_SUBAGENT_POLICY, DEFAULT_SUBAGENT_TOOL_NAMES, DEFAULT_TOOL_INPUT_DIGEST_HEAD_CHARS, DELEGATION_TOOL_NAMES, DedupTracker, type DelegationEventStamp, type DelegationMode, EmbeddingRequest, EmbeddingResponse, type EvalCase, type EvalCaseResult, type EvalFamily, type EvalHarnessMetrics, type EvalHarnessParams, type EvalHarnessResult, type EvalProfileDescriptor, type EvalProfileTarget, type FinalAnswerEvent, type GateResult, type GuardrailCallContext, type GuardrailContext, type GuardrailDisposition, type GuardrailEvent, type GuardrailIncident, type GuardrailJudgeEngine, type GuardrailOutcome, GuardrailPhase, type GuardrailPhaseName, type GuardrailPluginOptions, type GuardrailRequest, type GuardrailRule, type GuardrailRuleResult, type GuardrailTransport, type GuardrailVerdict, type GuardrailViolation, HOOK_CONTRACT_VERSION, HOOK_NAMES, type HandoffDescriptor, type HandoffEvent, type HookContext, type HookDecision, type HookHandler, type HookMap, type HookName, type HookRegistrations, HookTimeoutError, type HttpGuardrailOptions, type HumanInTheLoopAskUserConfig, type HumanInTheLoopOptions, type HumanInTheLoopRuntimeConfig, InMemoryMemoryStore, type InjectionScan, type InvokeConfig, type JevClientLike, type JevEngineOptions, type JevNoulAnswer, JevProvider, JevProviderConfig, type JevQuestion, type JevSystemOneRequest, type JevSystemOneResult, type JsonGuardrailOptions, type JsonSchemaRuleOptions, type LangfuseBuildContext, type LangfuseEvent, type LangfuseEventKind, type LangfuseTracingConfig, type LanguageGuardConfig, type LoadSkillsOptions, type MaybePromise, type McpConfig, type McpConnection, type MediaAttachment, type MediaKind$1 as MediaKind, type MediaPolicyConfig, type MemoryFact, type MemoryProviderKind, type MemoryReadPolicy, type MemoryScope, type MemoryStore, type MemoryWritePolicy, type Message, type MessageUsage, type MetadataEvent, type ModelCapabilities, ModelRequestError, type ModelRequestErrorCode, type ModelRequestErrorInit, ModelRuleLearnedEvent, NativeJsonSchemaStrategy, type NativeModelOptions, type NativeReasoningConfig, OpenAICompatibleProvider, OpenAICompatibleProviderConfig, type OpenAIModerationConfig, OpenAIProvider, OpenAIProviderConfig, type OtelTracingConfig, type OutputGuardConfig, PREOPEN_TOOL_CALL_PREFIX, type PendingToolApproval, type PendingUserQuestion, type PiiDetector, type PiiEntity, type PiiRedactionConfig, type PlanEvent, type PlanStepRecord, type PlanningMode, type PluginContributions, type PluginEvent, type PluginFailureMode, type PluginHost, type PluginHostOptions, type PluginLogger, type PluginProvides, type PluginRunHost, type PluginSetupContext, type PortkeyGatewayConfig, type PortkeyGuardrailConfig, type PreopenSkillsResult, type ProfileConfig, type ProgressEvent, type ProgressUpdate, type PromptHooks, type PromptInjectionFamily, type PromptInjectionGuardConfig, type PromptSource, ProviderConfig, ProviderError, ToolDefinition as ProviderToolDefinition, ProviderType, type RateLimitConfig, type ReasoningConfig, type ReasoningLevel, type RedactionResult, type ReflectionCadence, type ReflectionConfig, type ReflectionEvent, type ReflectionHookContext, type ReflectionRecord, type RegexGuardrailConfig, type RegexGuardrailOptions, type RegexRuleOptions, type ReplanPolicy, type ResolvedContextPilotConfig, type ResolvedSmartAgentConfig, type ResolvedTraceConfig, type ResolvedTraceSink, type ResponseCacheConfig, type RestoreSnapshotOptions, RulePatch, type RuntimeProfile, SMALL_TIER_SKILL_POLICY, type SdkCapabilities, type SerializableSmartState, type SessionMetrics, type SessionMetricsConfig, type Skill, type SkillDisclosure, type SkillExecutionContext, type SkillFrontmatter, type SkillFs, type SkillModelTier, type SkillPolicy, type SkillRegistryRef, type SkillSource, type SkillToolDeps, type SkillToolHeader, type SlotName, type SmartAgentBudgetConfig, type SmartAgentContextConfig, type SmartAgentCustomProfileConfig, type SmartAgentDelegationConfig, type SmartAgentEvent, type SmartAgentInstance, type SmartAgentLimits, type SmartAgentMemoryConfig, type SmartAgentOptions, type SmartAgentPlanningConfig, type SmartAgentSummarizationConfig, type SmartAgentToolResponseConfig, type SmartAgentTracingConfig, type SmartModel, type SmartState, type SmartToolFn, type SnapshotOptions, type SnapshotRuntimeHint, type StreamChunk, type StreamEvent, type StrictRestorers, type StrictTransform, type StructuredOutputError, type StructuredOutputFieldError, StructuredOutputManager, type StructuredOutputManagerConfig, type StructuredOutputResult, type StructuredOutputStrategy, type StructuredSummary, type SubagentDef, type SubagentEvent, type SubagentMode, type SubagentPendingRecord, type SubagentPolicy, type SubagentRegistryRef, type SubagentResult, type SubagentToolDeps, type SubagentToolNames, type SummarizationEvent, type SummarizationMode, type SummarizerInput, type SummarizerResult, type SummarizerStrategy, type SummaryFactItem, type SummaryIntegrityCheck, TOOL_INPUT_DIGEST_KEY, type TokenCounter, TokenUsage, type ToolApprovalEvent, type ToolApprovalResolution, type ToolApprovalStatus, ToolBasedStrategy, type ToolCachePolicy, type ToolCallEvent, type ToolInputDigestOptions, type ToolInputDigestResult, type ToolInputRetentionPolicy, type ToolInterface, type ToolPolicyConfig, type ToolResponseClassification, type ToolResponseRetentionPolicy, type ToolRetentionDeclarations, type ToolRetentionSpec, type ToolRetryPolicy, type ToolRule, type TraceDataSection, type TraceErrorRecord, type TraceEventRecord, type TraceMessagePart, type TraceMessageSection, type TraceMetadataSection, type TraceSessionConfigSnapshot, type TraceSessionFile, type TraceSessionStatus, type TraceSessionSummary, type TraceSinkCognipeerConfig, type TraceSinkConfig, type TraceSinkCustomConfig, type TraceSinkFileConfig, type TraceSinkHttpConfig, type TraceSinkOtlpConfig, type TraceSinkSnapshot, type TraceSummarySection, type TraceToolCallSection, type TraceToolDefinition, type TraceToolDefinitionsSection, type TraceToolDetails, type TraceToolResponseSection, type TraceToolResultSection, type TracingConfig, type TracingMode, UnifiedMessage, type UserQuestionAnswer, type UserQuestionAnswerSet, type UserQuestionEvent, type UserQuestionItem, type UserQuestionOption, type UserQuestionResolution, type UserQuestionStatus, VertexProvider, VertexProviderConfig, type VolatileContentFinding, type VolatileContentPattern, type WebhookApprovalConfig, agentVerdictRule, appendBoundTools, auditLog, azureContentSafety, bedrockGuardrail, budgetGuard, buildHandoffTools, buildResponseFormatSection, buildSkillHeaderBlock, buildSubagentCatalogBlock, buildSystemPrompt, buildToolDefinitionsSection, canOpenSkill, canSpawn, captureSnapshot, checkpointing, codePresenceRule, coerceToolArgs, cognipeerGuardrail, cognipeerSink, collectAttachments, collectToolRetentionDeclarations, composeToolSets, compressDiffOutput, compressJsonOutput, compressLogOutput, compressSearchOutput, compressTextOutput, compressToolOutput, contentPartToText, contentToString, conversationHistory, countApproxTokens, countMessagesTokens, createAgent, createAgentNode, createAskUserQuestionTool, createBindSkillToolsTool, createCodeGuardrail, createContextPilotRuntime, createContextSummarizeNode, createContextTools, createDelegateTool, createGetToolResponseTool, createGuardrail, createGuardrailPlugin, createJsonGuardrail, createManageTodoListAliasTool, createManageTodoTool, createOpenSkillTool, createParallelTool, createPluginHost, createProvider, createRegexGuardrail, createResolverNode, createSearchSkillsTool, createSkillRegistryRef, createSkillTools, createSmartAgent, createSpawnTool, createSubagentRegistryRef, createSubagentTools, createTool, createToolLimitFinalizeNode, createToolsNode, customCallbackRule, customGuardrail, customSink, dedupeToolsByName, defaultTokenCounter, defineHook, definePlugin, describeContentPart, detectLanguage, detectTextFormat, detectVolatileContent, digestToolInputArguments, digestToolInputValue, evaluateGuardrails, extractLatestUserQuery, extractMessageText, fileCheckpointStore, fileSink, findSubagent, formatZodError, fromLangchainModel, fromLangchainTools, fromNativeProvider, generateSpanId, generateTraceId, getModelCapabilities, getNestedValue, getResolvedSmartConfig, getTokenCounter, hasMediaPart, hasToolResponseRecoveryReference, httpGuardrail, httpSink, inMemoryCheckpointStore, inMemoryConversationStore, isHookImplemented, isOpenAIHostedEndpoint, isSmartModel, isValidIban, isValidLuhn, isValidTckn, jsonSchemaRule, defaultBuildPayload as langfuseDefaultPayload, langfuseTracing, languageGuard, loadSkillsFromDisk, mapConsoleHookVerdict, mapTextParts, mcp, mediaKindOf, mediaPolicy, normalizeSmartAgentOptions, normalizeVerdicts, openAIModeration, otelTracing, otlpSink, outputGuard, parseApprovalDecision, parseSkillFrontmatter, pathSandbox, piiRedaction, pluginCapabilities, portkeyGateway, portkeyGuardrail, preopenSkills, preopenToolCallId, prepareStrictToolMenu, promptInjectionGuard, rateLimit, readMemoryFacts, redactText, regexGuardrail, regexRule, replaceTextContent, resolveAvailableSkills, resolveAvailableSubagents, resolveInputRetention, resolveMemoryStore, resolveReasoning, resolveStrategy, resolveSubagentTools, resolveSummarizationRetention, resolveToolApprovalState, resolveUserQuestionState, responseCache, restoreSnapshot, restoreToolCalls, runSmartAgentEvalHarness, scoreItemsByRelevance, searchSkills, seedChildMessages, selectTopIndicesInOrder, sessionMetrics, setTokenCounter, startStreamingSession, summarizeObject, supportsAdaptiveThinking, textFromContent, toMessageUsage, toStrictCompatible, toStrictToolSchema, tokenize, toolPolicy, traceSessionToOtlp, validateReasoningConfig, validateToolArgs, webhookApproval, withTools, withinDepth, writeSummaryFactsToMemory };