@schlenkr/ragents 0.1.4 → 0.1.6

package/packages/ai/dist/types.d.ts

@@ -0,0 +1,541 @@
+import type { OpenRouterOptions } from "./api/ai-sdk.ts";
+import type { AssistantMessageDiagnostic } from "./utils/diagnostics.ts";
+import type { AssistantMessageEventStream } from "./utils/event-stream.ts";
+export type { AssistantMessageEventStream } from "./utils/event-stream.ts";
+export type KnownApi = "openai-completions";
+export type Api = KnownApi | (string & {});
+export type KnownProvider = "openrouter";
+export type ProviderId = KnownProvider | string;
+export type ThinkingLevel = "minimal" | "low" | "medium" | "high" | "xhigh" | "max";
+export type ModelThinkingLevel = "off" | ThinkingLevel;
+export type ThinkingLevelMap = Partial<Record<ModelThinkingLevel, string | null>>;
+export type ChatTemplateKwargValue = string | number | boolean | null | {
+    $var: "thinking.enabled" | "thinking.effort";
+    omitWhenOff?: boolean;
+};
+/** Token budgets for each thinking level (token-based providers only) */
+export interface ThinkingBudgets {
+    minimal?: number;
+    low?: number;
+    medium?: number;
+    high?: number;
+}
+export type CacheRetention = "none" | "short" | "long";
+export type Transport = "sse" | "websocket" | "websocket-cached" | "auto";
+/** Provider-scoped environment overrides. Values take precedence over process.env. */
+export type ProviderEnv = Record<string, string>;
+export type ProviderHeaders = Record<string, string | null>;
+export type SessionAffinityFormat = "openai" | "openai-nosession" | "openrouter";
+export interface ProviderResponse {
+    status: number;
+    headers: Record<string, string>;
+}
+export interface StreamOptions {
+    temperature?: number;
+    maxTokens?: number;
+    signal?: AbortSignal;
+    apiKey?: string;
+    /**
+     * Preferred transport for providers that support multiple transports.
+     * Providers that do not support this option ignore it.
+     */
+    transport?: Transport;
+    /**
+     * Prompt cache retention preference. Providers map this to their supported values.
+     * Default: "short".
+     */
+    cacheRetention?: CacheRetention;
+    /**
+     * Optional session identifier for providers that support session-based caching.
+     * Providers can use this to enable prompt caching, request routing, or other
+     * session-aware features. Ignored by providers that don't support it.
+     */
+    sessionId?: string;
+    /**
+     * Optional callback for inspecting or replacing provider payloads before sending.
+     * Return undefined to keep the payload unchanged.
+     */
+    onPayload?: (payload: unknown, model: Model<Api>) => unknown | undefined | Promise<unknown | undefined>;
+    /**
+     * Optional callback invoked after an HTTP response is received and before
+     * its body stream is consumed.
+     */
+    onResponse?: (response: ProviderResponse, model: Model<Api>) => void | Promise<void>;
+    /**
+     * Optional custom HTTP headers to include in API requests.
+     * Merged with provider defaults; caller values override default headers.
+     * On AWS Bedrock these are injected via a Smithy `build`-step middleware so
+     * they are covered by SigV4 signing; reserved headers (`x-amz-*`,
+     * `authorization`, `host`) are silently ignored to preserve SigV4 / bearer auth.
+     * A null value suppresses a provider/API default header with the same name.
+     */
+    headers?: ProviderHeaders;
+    /**
+     * HTTP request timeout in milliseconds for providers/SDKs that support it.
+     * For example, OpenAI and Anthropic SDK clients default to 10 minutes.
+     */
+    timeoutMs?: number;
+    /**
+     * WebSocket connect timeout in milliseconds for providers that support
+     * WebSocket transports. This covers the connection/open handshake only;
+     * stream idleness after connection uses timeoutMs.
+     */
+    websocketConnectTimeoutMs?: number;
+    /**
+     * Maximum retry attempts for providers/SDKs that support client-side retries.
+     * For example, OpenAI and Anthropic SDK clients default to 2.
+     */
+    maxRetries?: number;
+    /**
+     * Maximum delay in milliseconds to wait for a retry when the server requests a long wait.
+     * If the server's requested delay exceeds this value, the request fails immediately
+     * with an error containing the requested delay, allowing higher-level retry logic
+     * to handle it with user visibility.
+     * Default: 60000 (60 seconds). Set to 0 to disable the cap.
+     */
+    maxRetryDelayMs?: number;
+    /**
+     * Optional metadata to include in API requests.
+     * Providers extract the fields they understand and ignore the rest.
+     * For example, Anthropic uses `user_id` for abuse tracking and rate limiting.
+     */
+    metadata?: Record<string, unknown>;
+    /**
+     * Provider-scoped environment values. These take precedence over process.env for
+     * provider configuration such as regional settings, endpoint placeholders, and
+     * proxy variables.
+     */
+    env?: ProviderEnv;
+}
+export type ProviderStreamOptions = StreamOptions & Record<string, unknown>;
+/**
+ * Maps known APIs to their full provider-specific stream option types.
+ * Type-only imports from API implementation modules are erased at emit, so
+ * this is tree-shake safe.
+ */
+export interface ApiOptionsMap {
+    "openai-completions": OpenRouterOptions;
+}
+/**
+ * Full stream options for an API. Known APIs resolve to their concrete option
+ * type; custom API strings fall back to the generic shape.
+ */
+export type ApiStreamOptions<TApi extends Api> = TApi extends keyof ApiOptionsMap ? ApiOptionsMap[TApi] : StreamOptions & Record<string, unknown>;
+/**
+ * The uniform stream contract of an API implementation module: every module
+ * under `src/api/` exports exactly `stream` and `streamSimple`, so the module
+ * itself satisfies this interface. Lazy wrappers (`lazyApi()`) and provider
+ * factories pass these around as values. This is the untyped dispatch shape;
+ * per-API option typing lives on the implementation modules themselves and on
+ * `Provider.stream()` via `ApiStreamOptions`.
+ */
+export interface ProviderStreams {
+    stream(model: Model<Api>, context: Context, options?: StreamOptions): AssistantMessageEventStream;
+    streamSimple(model: Model<Api>, context: Context, options?: SimpleStreamOptions): AssistantMessageEventStream;
+}
+export interface SimpleStreamOptions extends StreamOptions {
+    reasoning?: ThinkingLevel;
+    /** Custom token budgets for thinking levels (token-based providers only) */
+    thinkingBudgets?: ThinkingBudgets;
+}
+export type StreamFunction<TApi extends Api = Api, TOptions extends StreamOptions = StreamOptions> = (model: Model<TApi>, context: Context, options?: TOptions) => AssistantMessageEventStream;
+export interface TextSignatureV1 {
+    v: 1;
+    id: string;
+    phase?: "commentary" | "final_answer";
+}
+export interface TextContent {
+    type: "text";
+    text: string;
+    textSignature?: string;
+}
+export interface ThinkingContent {
+    type: "thinking";
+    thinking: string;
+    thinkingSignature?: string;
+    /** When true, the thinking content was redacted by safety filters. The opaque
+     *  encrypted payload is stored in `thinkingSignature` so it can be passed back
+     *  to the API for multi-turn continuity. */
+    redacted?: boolean;
+}
+export interface ImageContent {
+    type: "image";
+    data: string;
+    mimeType: string;
+}
+export interface VideoContent {
+    type: "video";
+    data: string;
+    mimeType: string;
+}
+export interface FileContent {
+    type: "file";
+    data: string;
+    mimeType: string;
+    filename: string;
+}
+export type UserAttachment = ImageContent | VideoContent | FileContent;
+export type UserContent = TextContent | UserAttachment;
+export type InputModality = "text" | "image" | "video" | "file" | "audio";
+export interface ToolCall {
+    type: "toolCall";
+    id: string;
+    name: string;
+    arguments: Record<string, any>;
+    thoughtSignature?: string;
+}
+export interface Usage {
+    input: number;
+    output: number;
+    cacheRead: number;
+    cacheWrite: number;
+    /** Subset of `cacheWrite` written with 1h retention. Only Anthropic reports this split. */
+    cacheWrite1h?: number;
+    /**
+     * Reasoning/thinking tokens, when the provider reports them. This is a subset of
+     * `output`: `output` already includes these tokens. Set to a number (possibly 0) by
+     * providers that expose a reasoning breakdown; left undefined by providers that don't.
+     */
+    reasoning?: number;
+    totalTokens: number;
+    cost: {
+        input: number;
+        output: number;
+        cacheRead: number;
+        cacheWrite: number;
+        total: number;
+    };
+}
+export type StopReason = "stop" | "length" | "toolUse" | "error" | "aborted";
+export interface UserMessage {
+    role: "user";
+    content: string | UserContent[];
+    timestamp: number;
+}
+export interface AssistantMessage {
+    role: "assistant";
+    content: (TextContent | ThinkingContent | ToolCall)[];
+    api: Api;
+    provider: ProviderId;
+    model: string;
+    responseModel?: string;
+    responseId?: string;
+    diagnostics?: AssistantMessageDiagnostic[];
+    usage: Usage;
+    stopReason: StopReason;
+    errorMessage?: string;
+    timestamp: number;
+}
+export interface ToolResultMessage<TDetails = any> {
+    role: "toolResult";
+    toolCallId: string;
+    toolName: string;
+    content: (TextContent | ImageContent)[];
+    details?: TDetails;
+    /**
+     * Names from `Context.tools` that became available after this result.
+     * Providers with native deferred tool loading use this as the load point;
+     * other providers ignore it and use `Context.tools` normally.
+     */
+    addedToolNames?: string[];
+    isError: boolean;
+    timestamp: number;
+}
+export type Message = UserMessage | AssistantMessage | ToolResultMessage;
+import type { TSchema } from "typebox";
+export interface Tool<TParameters extends TSchema = TSchema> {
+    name: string;
+    description: string;
+    parameters: TParameters;
+}
+export interface Context {
+    systemPrompt?: string;
+    messages: Message[];
+    tools?: Tool[];
+}
+/**
+ * Event protocol for AssistantMessageEventStream.
+ *
+ * Streams should emit `start` before partial updates, then terminate with either:
+ * - `done` carrying the final successful AssistantMessage, or
+ * - `error` carrying the final AssistantMessage with stopReason "error" or "aborted"
+ *   and errorMessage.
+ */
+export type AssistantMessageEvent = {
+    type: "start";
+    partial: AssistantMessage;
+} | {
+    type: "text_start";
+    contentIndex: number;
+    partial: AssistantMessage;
+} | {
+    type: "text_delta";
+    contentIndex: number;
+    delta: string;
+    partial: AssistantMessage;
+} | {
+    type: "text_end";
+    contentIndex: number;
+    content: string;
+    partial: AssistantMessage;
+} | {
+    type: "thinking_start";
+    contentIndex: number;
+    partial: AssistantMessage;
+} | {
+    type: "thinking_delta";
+    contentIndex: number;
+    delta: string;
+    partial: AssistantMessage;
+} | {
+    type: "thinking_end";
+    contentIndex: number;
+    content: string;
+    partial: AssistantMessage;
+} | {
+    type: "toolcall_start";
+    contentIndex: number;
+    partial: AssistantMessage;
+} | {
+    type: "toolcall_delta";
+    contentIndex: number;
+    delta: string;
+    partial: AssistantMessage;
+} | {
+    type: "toolcall_end";
+    contentIndex: number;
+    toolCall: ToolCall;
+    partial: AssistantMessage;
+} | {
+    type: "done";
+    reason: Extract<StopReason, "stop" | "length" | "toolUse">;
+    message: AssistantMessage;
+} | {
+    type: "error";
+    reason: Extract<StopReason, "aborted" | "error">;
+    error: AssistantMessage;
+};
+/**
+ * Compatibility settings for OpenAI-compatible completions APIs.
+ * Use this to override URL-based auto-detection for custom providers.
+ */
+export interface OpenAICompletionsCompat {
+    /** Whether the provider supports the `store` field. Default: auto-detected from URL. */
+    supportsStore?: boolean;
+    /** Whether the provider supports the `developer` role (vs `system`). Default: auto-detected from URL. */
+    supportsDeveloperRole?: boolean;
+    /** Whether the provider supports `reasoning_effort`. Default: auto-detected from URL. */
+    supportsReasoningEffort?: boolean;
+    /** Whether the provider supports `stream_options: { include_usage: true }` for token usage in streaming responses. Default: true. */
+    supportsUsageInStreaming?: boolean;
+    /** Which field to use for max tokens. Default: auto-detected from URL. */
+    maxTokensField?: "max_completion_tokens" | "max_tokens";
+    /** Whether tool results require the `name` field. Default: auto-detected from URL. */
+    requiresToolResultName?: boolean;
+    /** Whether a user message after tool results requires an assistant message in between. Default: auto-detected from URL. */
+    requiresAssistantAfterToolResult?: boolean;
+    /** Whether thinking blocks must be converted to text blocks with <thinking> delimiters. Default: auto-detected from URL. */
+    requiresThinkingAsText?: boolean;
+    /** Whether all replayed assistant messages must include an empty reasoning_content field when reasoning is enabled. Default: auto-detected from URL. */
+    requiresReasoningContentOnAssistantMessages?: boolean;
+    /** Format for reasoning/thinking parameter. "openai" uses reasoning_effort, "openrouter" uses reasoning: { effort }, "deepseek" uses thinking: { type } plus reasoning_effort when supported, "together" uses reasoning: { enabled } plus reasoning_effort when supported, "zai" uses thinking: { type }, "qwen" uses top-level enable_thinking: boolean, "qwen-chat-template" uses chat_template_kwargs.enable_thinking and preserve_thinking, "chat-template" uses configurable chat_template_kwargs, "string-thinking" uses top-level thinking: string, and "ant-ling" uses reasoning: { effort } only when the mapped effort is non-null. Default: "openai". */
+    thinkingFormat?: "openai" | "openrouter" | "deepseek" | "together" | "zai" | "qwen" | "chat-template" | "qwen-chat-template" | "string-thinking" | "ant-ling";
+    /** Kwargs to send as `chat_template_kwargs` when `thinkingFormat` is `chat-template`. Use `{ "$var": "thinking.enabled" }` or `{ "$var": "thinking.effort" }` for host-controlled thinking values. */
+    chatTemplateKwargs?: Record<string, ChatTemplateKwargValue>;
+    /** OpenRouter-compatible routing preferences sent as the `provider` request field. */
+    openRouterRouting?: OpenRouterRouting;
+    /** Vercel AI Gateway routing preferences. Only used when baseUrl points to Vercel AI Gateway. */
+    vercelGatewayRouting?: VercelGatewayRouting;
+    /** Whether z.ai supports top-level `tool_stream: true` for streaming tool call deltas. Default: false. */
+    zaiToolStream?: boolean;
+    /** Whether the provider supports the `strict` field in tool definitions. Default: true. */
+    supportsStrictMode?: boolean;
+    /** Cache control convention for prompt caching. "anthropic" applies Anthropic-style `cache_control` markers to the system prompt, last tool definition, and last user/assistant text content. */
+    cacheControlFormat?: "anthropic";
+    /** Whether to send session-affinity data from `options.sessionId`. Default: false. */
+    sendSessionAffinityHeaders?: boolean;
+    /** Provider-specific deferred tool serialization mode. */
+    deferredToolsMode?: "kimi";
+    /** Session-affinity header format: `openai` sends `session_id`, `x-client-request-id`, and `x-session-affinity`; `openai-nosession` sends `x-client-request-id` and `x-session-affinity`; `openrouter` sends `x-session-id`. Does not affect the `prompt_cache_key` body param, which is governed by cache retention. Default: auto-detected. */
+    sessionAffinityFormat?: SessionAffinityFormat;
+    /** Whether the provider supports long prompt cache retention (`prompt_cache_retention: "24h"` or Anthropic-style `cache_control.ttl: "1h"`, depending on format). Default: true. */
+    supportsLongCacheRetention?: boolean;
+}
+/** Compatibility settings for OpenAI Responses APIs. */
+export interface OpenAIResponsesCompat {
+    /** Whether the provider supports the `developer` role (vs `system`). Default: true. */
+    supportsDeveloperRole?: boolean;
+    /** Session-affinity header format: `openai` sends `session_id` and `x-client-request-id`; `openai-nosession` sends `x-client-request-id`; `openrouter` sends `x-session-id`. Does not affect the `prompt_cache_key` body param, which is governed by cache retention. Default: auto-detected. */
+    sessionAffinityFormat?: SessionAffinityFormat;
+    /** Whether the provider supports `prompt_cache_retention: "24h"`. Default: true. */
+    supportsLongCacheRetention?: boolean;
+    /** Whether the model supports client-executed tool search for deferred tools. Default: false. */
+    supportsToolSearch?: boolean;
+}
+/** Compatibility settings for Anthropic Messages-compatible APIs. */
+export interface AnthropicMessagesCompat {
+    /**
+     * Whether the provider accepts per-tool `eager_input_streaming`.
+     * When false, the Anthropic provider omits `tools[].eager_input_streaming`
+     * and sends the legacy `fine-grained-tool-streaming-2025-05-14` beta header
+     * for tool-enabled requests.
+     * Default: true.
+     */
+    supportsEagerToolInputStreaming?: boolean;
+    /** Whether the provider supports Anthropic long cache retention (`cache_control.ttl: "1h"`). Default: true. */
+    supportsLongCacheRetention?: boolean;
+    /**
+     * Whether to send the `x-session-affinity` header from `options.sessionId`
+     * when caching is enabled. Required for providers like Fireworks that use
+     * session affinity for prompt cache routing (requests to the same replica
+     * maximize cache hits).
+     * Default: false.
+     */
+    sendSessionAffinityHeaders?: boolean;
+    /**
+     * Whether the provider supports Anthropic-style `cache_control` markers on
+     * tool definitions. When false, `cache_control` is omitted from tool params.
+     * Some Anthropic-compatible providers (e.g., Fireworks) do not support this
+     * field on tools and may reject or ignore it.
+     * Default: true.
+     */
+    supportsCacheControlOnTools?: boolean;
+    /**
+     * Whether the model accepts the Anthropic `temperature` request field.
+     * Claude Opus 4.7+ rejects non-default temperature values.
+     * Default: true.
+     */
+    supportsTemperature?: boolean;
+    /**
+     * Whether to force adaptive thinking (`thinking.type: "adaptive"` plus
+     * `output_config.effort`) regardless of the model id. Built-in models that
+     * require adaptive thinking set this in generated metadata. Custom
+     * Anthropic-compatible providers can set this to `true` for any model whose
+     * upstream requires the adaptive format. Set to `false` to
+     * opt out on overridden built-in models.
+     * Default: false.
+     */
+    forceAdaptiveThinking?: boolean;
+    /** Whether to replay empty thinking signatures as `signature: ""` instead of converting thinking to text. Default: false. */
+    allowEmptySignature?: boolean;
+    /**
+     * Whether the provider supports deferred tools loaded by `tool_reference`
+     * blocks in tool results. Default: true for first-party Anthropic models
+     * except Haiku and models older than Claude 4.5; false for other providers.
+     */
+    supportsToolReferences?: boolean;
+}
+/**
+ * OpenRouter provider routing preferences.
+ * Controls which upstream providers OpenRouter routes requests to.
+ * Sent as the `provider` field in the OpenRouter API request body.
+ * @see https://openrouter.ai/docs/guides/routing/provider-selection
+ */
+export interface OpenRouterRouting {
+    /** Whether to allow backup providers to serve requests. Default: true. */
+    allow_fallbacks?: boolean;
+    /** Whether to filter providers to only those that support all parameters in the request. Default: false. */
+    require_parameters?: boolean;
+    /** Data collection setting. "allow" (default): allow providers that may store/train on data. "deny": only use providers that don't collect user data. */
+    data_collection?: "deny" | "allow";
+    /** Whether to restrict routing to only ZDR (Zero Data Retention) endpoints. */
+    zdr?: boolean;
+    /** Whether to restrict routing to only models that allow text distillation. */
+    enforce_distillable_text?: boolean;
+    /** An ordered list of provider names/slugs to try in sequence, falling back to the next if unavailable. */
+    order?: string[];
+    /** List of provider names/slugs to exclusively allow for this request. */
+    only?: string[];
+    /** List of provider names/slugs to skip for this request. */
+    ignore?: string[];
+    /** A list of quantization levels to filter providers by (e.g., ["fp16", "bf16", "fp8", "fp6", "int8", "int4", "fp4", "fp32"]). */
+    quantizations?: string[];
+    /** Sorting strategy. Can be a string (e.g., "price", "throughput", "latency") or an object with `by` and `partition`. */
+    sort?: string | {
+        /** The sorting metric: "price", "throughput", "latency". */
+        by?: string;
+        /** Partitioning strategy: "model" (default) or "none". */
+        partition?: string | null;
+    };
+    /** Maximum price per million tokens (USD). */
+    max_price?: {
+        /** Price per million prompt tokens. */
+        prompt?: number | string;
+        /** Price per million completion tokens. */
+        completion?: number | string;
+        /** Price per image. */
+        image?: number | string;
+        /** Price per audio unit. */
+        audio?: number | string;
+        /** Price per request. */
+        request?: number | string;
+    };
+    /** Preferred minimum throughput (tokens/second). Can be a number (applies to p50) or an object with percentile-specific cutoffs. */
+    preferred_min_throughput?: number | {
+        /** Minimum tokens/second at the 50th percentile. */
+        p50?: number;
+        /** Minimum tokens/second at the 75th percentile. */
+        p75?: number;
+        /** Minimum tokens/second at the 90th percentile. */
+        p90?: number;
+        /** Minimum tokens/second at the 99th percentile. */
+        p99?: number;
+    };
+    /** Preferred maximum latency (seconds). Can be a number (applies to p50) or an object with percentile-specific cutoffs. */
+    preferred_max_latency?: number | {
+        /** Maximum latency in seconds at the 50th percentile. */
+        p50?: number;
+        /** Maximum latency in seconds at the 75th percentile. */
+        p75?: number;
+        /** Maximum latency in seconds at the 90th percentile. */
+        p90?: number;
+        /** Maximum latency in seconds at the 99th percentile. */
+        p99?: number;
+    };
+}
+/**
+ * Vercel AI Gateway routing preferences.
+ * Controls which upstream providers the gateway routes requests to.
+ * @see https://vercel.com/docs/ai-gateway/models-and-providers/provider-options
+ */
+export interface VercelGatewayRouting {
+    /** List of provider slugs to exclusively use for this request (e.g., ["bedrock", "anthropic"]). */
+    only?: string[];
+    /** List of provider slugs to try in order (e.g., ["anthropic", "openai"]). */
+    order?: string[];
+}
+export interface ModelCostRates {
+    input: number;
+    output: number;
+    cacheRead: number;
+    cacheWrite: number;
+}
+export interface ModelCostTier extends ModelCostRates {
+    /** Use this tier for requests whose total input usage exceeds this token count. */
+    inputTokensAbove: number;
+}
+export interface ModelCost extends ModelCostRates {
+    /** Request-wide pricing tiers. The highest matching input threshold applies to the full request. */
+    tiers?: ModelCostTier[];
+}
+export interface Model<TApi extends Api> {
+    id: string;
+    name: string;
+    api: TApi;
+    provider: ProviderId;
+    baseUrl: string;
+    reasoning: boolean;
+    /**
+     * Maps thinking levels to provider/model-specific values.
+     * Missing keys use provider defaults. null marks a level as unsupported.
+     */
+    thinkingLevelMap?: ThinkingLevelMap;
+    input: InputModality[];
+    cost: ModelCost;
+    contextWindow: number;
+    maxTokens: number;
+    headers?: Record<string, string>;
+    /** Compatibility overrides for OpenAI-compatible APIs. If not set, auto-detected from baseUrl. */
+    compat?: TApi extends "openai-completions" ? OpenAICompletionsCompat : TApi extends "openai-responses" | "openai-codex-responses" ? OpenAIResponsesCompat : TApi extends "anthropic-messages" ? AnthropicMessagesCompat : never;
+}
+//# sourceMappingURL=types.d.ts.map

package/packages/ai/dist/utils/diagnostics.d.ts

@@ -0,0 +1,13 @@
+export interface DiagnosticErrorInfo {
+    name?: string;
+    message: string;
+    stack?: string;
+    code?: string | number;
+}
+export interface AssistantMessageDiagnostic {
+    type: string;
+    timestamp: number;
+    error?: DiagnosticErrorInfo;
+    details?: Record<string, unknown>;
+}
+//# sourceMappingURL=diagnostics.d.ts.map

package/packages/ai/dist/utils/error-body.d.ts

@@ -0,0 +1,25 @@
+export declare const MAX_PROVIDER_ERROR_BODY_CHARS = 4000;
+export interface NormalizedProviderError {
+    /** HTTP status code, when one could be extracted from the SDK error object. */
+    status?: number;
+    /** Raw HTTP body reason, already trimmed and truncated to the cap. */
+    body?: string;
+    /** `error.message`, or `safeJsonStringify(error)` for a non-`Error` throw. */
+    message: string;
+    /** True when `message` already contains the body (no separate body to add). */
+    messageCarriesBody: boolean;
+}
+export declare function normalizeProviderError(error: unknown): NormalizedProviderError;
+/**
+ * Compose a display string from a normalized error. When the message already
+ * carries the body (Anthropic / `@google/genai` happy path) or no body/status
+ * was extracted, the message is returned unchanged. Otherwise the status and
+ * body are surfaced, with an optional provider prefix.
+ *
+ * - no prefix: `"<status>: <body>"`
+ * - prefix:    `"<prefix> (<status>): <body>"`
+ */
+export declare function formatProviderError(norm: NormalizedProviderError, prefix?: string): string;
+export declare function truncateErrorText(text: string, maxChars: number): string;
+export declare function safeJsonStringify(value: unknown): string;
+//# sourceMappingURL=error-body.d.ts.map

package/packages/ai/dist/utils/estimate.d.ts

@@ -0,0 +1,17 @@
+import type { Context, Message, Usage, UserContent } from "../types.ts";
+export interface ContextUsageEstimate {
+    /** Estimated total context tokens. */
+    tokens: number;
+    /** Tokens reported by the most recent applicable assistant usage block. */
+    usageTokens: number;
+    /** Estimated tokens after the most recent applicable assistant usage block. */
+    trailingTokens: number;
+    /** Index of the applicable message that provided usage, or null when none exists. */
+    lastUsageIndex: number | null;
+}
+export declare function calculateContextTokens(usage: Usage): number;
+export declare function estimateTextTokens(text: string): number;
+export declare function estimateTextAndImageContentTokens(content: string | UserContent[]): number;
+export declare function estimateMessageTokens(message: Message): number;
+export declare function estimateContextTokens(context: Context | readonly Message[]): ContextUsageEstimate;
+//# sourceMappingURL=estimate.d.ts.map

package/packages/ai/dist/utils/event-stream.d.ts

@@ -0,0 +1,21 @@
+import type { AssistantMessage, AssistantMessageEvent } from "../types.ts";
+export declare class EventStream<T, R = T> implements AsyncIterable<T> {
+    private queue;
+    private waiting;
+    private done;
+    private finalResultPromise;
+    private resolveFinalResult;
+    private isComplete;
+    private extractResult;
+    constructor(isComplete: (event: T) => boolean, extractResult: (event: T) => R);
+    push(event: T): void;
+    end(result?: R): void;
+    [Symbol.asyncIterator](): AsyncIterator<T>;
+    result(): Promise<R>;
+}
+export declare class AssistantMessageEventStream extends EventStream<AssistantMessageEvent, AssistantMessage> {
+    constructor();
+}
+/** Factory function for AssistantMessageEventStream (for use in extensions) */
+export declare function createAssistantMessageEventStream(): AssistantMessageEventStream;
+//# sourceMappingURL=event-stream.d.ts.map

package/packages/ai/dist/utils/headers.d.ts

@@ -0,0 +1,2 @@
+export declare function headersToRecord(headers: Headers): Record<string, string>;
+//# sourceMappingURL=headers.d.ts.map

package/packages/ai/dist/utils/json-parse.d.ts

@@ -0,0 +1,16 @@
+/**
+ * Repairs malformed JSON string literals by:
+ * - escaping raw control characters inside strings
+ * - doubling backslashes before invalid escape characters
+ */
+export declare function repairJson(json: string): string;
+export declare function parseJsonWithRepair<T>(json: string): T;
+/**
+ * Attempts to parse potentially incomplete JSON during streaming.
+ * Always returns a valid object, even if the JSON is incomplete.
+ *
+ * @param partialJson The partial JSON string from streaming
+ * @returns Parsed object or empty object if parsing fails
+ */
+export declare function parseStreamingJson<T = Record<string, unknown>>(partialJson: string | undefined): T;
+//# sourceMappingURL=json-parse.d.ts.map