@jaypie/llm 1.3.13 → 1.3.15

package/dist/esm/index.d.ts

@@ -525,7 +525,97 @@ interface LlmProgressEvent {
     usage?: LlmUsage;
 }
 type LlmProgressCallback = (event: LlmProgressEvent) => unknown | Promise<unknown>;
+/**
+ * Serializable snapshot of the request side of one operate() call.
+ * `input` is the raw, pre-interpolation input with multimodal parts preserved.
+ */
+interface LlmExchangeRequest {
+    cache?: LlmCache;
+    data?: NaturalMap;
+    effort?: LlmEffort;
+    explain?: boolean;
+    /** Format normalized to JSON Schema (Zod/Natural already rendered) */
+    format?: JsonObject;
+    input: string | LlmHistory | LlmInputMessage | LlmOperateInput;
+    instructions?: string;
+    /** Requested model (pre-fallback) */
+    model?: string;
+    placeholders?: {
+        input?: boolean;
+        instructions?: boolean;
+        system?: boolean;
+    };
+    providerOptions?: JsonObject;
+    system?: string;
+    temperature?: number;
+    /** Names of tools offered to the model */
+    tools?: string[];
+    turns?: boolean | number;
+    user?: string;
+}
+/**
+ * Serializable snapshot of the response side of one operate() call.
+ */
+interface LlmExchangeResponse {
+    content?: string | JsonObject;
+    error?: LlmError$1;
+    /** Turns this call added to history (tool calls/results included); never the resent history */
+    historyDelta: LlmHistory;
+    reasoning?: string[];
+    status: LlmResponseStatus;
+    stopReason?: string;
+    /** Per-model-call usage items (provider/model on each item) */
+    usage: LlmUsage;
+    /** Cumulative usage keyed `provider:model` */
+    usageTotals?: Record<string, LlmUsageItem>;
+}
+/**
+ * How the call was actually served after fallback and retry resolution.
+ */
+interface LlmExchangeResolution {
+    fallbackAttempts?: number;
+    fallbackUsed?: boolean;
+    /** Served model */
+    model?: string;
+    /** Served provider */
+    provider?: string;
+    /** Model-request retries within the served attempt */
+    retries?: number;
+}
+interface LlmExchangeTiming {
+    /** Milliseconds from start to settlement */
+    duration: number;
+    /** ISO 8601 */
+    startedAt: string;
+}
+/**
+ * One serializable request/response envelope per operate() settlement
+ * (success or failure). Contains no functions.
+ */
+interface LlmExchangeEnvelope {
+    /** Provider response id(s) per model turn, when available */
+    ids?: string[];
+    request: LlmExchangeRequest;
+    resolution: LlmExchangeResolution;
+    response: LlmExchangeResponse;
+    timing: LlmExchangeTiming;
+}
+type LlmExchangeCallback = (envelope: LlmExchangeEnvelope) => unknown | Promise<unknown>;
+/**
+ * Prompt-caching control for operate()/stream().
+ * - `true` / omitted → caching enabled at the default `"5m"` TTL
+ * - `false` / `0` → caching disabled
+ * - `"5m"` / `"1h"` → enabled at that TTL (TTL honored by Anthropic/OpenRouter;
+ *   other providers ignore it and cache with their own defaults)
+ */
+type LlmCache = boolean | 0 | "5m" | "1h";
 interface LlmOperateOptions {
+    /**
+     * Prompt caching for the stable request prefix (system prompt + tools).
+     * Enabled by default; pass `false`/`0` to opt out, or `"5m"`/`"1h"` to set
+     * the TTL. See {@link LlmCache}.
+     */
+    cache?: LlmCache;
     data?: NaturalMap;
     /**
      * Provider-neutral reasoning effort (lowest | low | medium | high | highest).
@@ -589,6 +679,12 @@ interface LlmOperateOptions {
     };
     instructions?: string;
     model?: string;
+    /**
+     * Fires once per operate() settlement (success or failure) with a fully
+     * serializable request/response envelope. Errors thrown by the callback
+     * are logged and never interrupt the call.
+     */
+    onExchange?: LlmExchangeCallback;
     /**
      * Receives lightweight progress events as the operate loop runs:
      * start, model_request, model_response, tool_call, tool_result,
@@ -628,6 +724,10 @@ interface LlmUsageItem {
     output: number;
     reasoning: number;
     total: number;
+    /** Prompt-cache tokens served from cache this call (billed at ~0.1x input) */
+    cacheRead?: number;
+    /** Prompt-cache tokens written to cache this call (billed at ~1.25x input) */
+    cacheWrite?: number;
     provider?: string;
     model?: string;
 }
@@ -635,6 +735,11 @@ type LlmUsage = LlmUsageItem[];
 interface LlmOperateResponse {
     content?: string | JsonObject;
     error?: LlmError$1;
+    /**
+     * Serializable exchange envelope for this call. Present only when
+     * `onExchange` was passed or exchange persistence is enabled.
+     */
+    exchange?: LlmExchangeEnvelope;
     /** Number of providers attempted (1 = primary only, >1 = fallback(s) used) */
     fallbackAttempts?: number;
     /** Whether a fallback provider was used instead of the primary */
@@ -680,6 +785,12 @@ declare class Llm implements LlmProvider {
     operate(input: string | LlmHistory | LlmInputMessage | LlmOperateInput, options?: Omit<LlmOperateOptions, "model"> & {
         model?: LlmModelOption;
     }): Promise<LlmOperateResponse>;
+    /**
+     * Stamp fallback resolution onto the envelope the operate loop attached to
+     * the response and deliver it to the caller's onExchange. Fires once per
+     * operate() settlement; callback errors are logged and never thrown.
+     */
+    private settleExchange;
     stream(input: string | LlmHistory | LlmInputMessage | LlmOperateInput, options?: Omit<LlmOperateOptions, "model"> & {
         model?: LlmModelOption;
     }): AsyncIterable<LlmStreamChunk>;
@@ -931,4 +1042,4 @@ declare class XaiProvider implements LlmProvider {
 }
 
 export { BedrockProvider, ErrorCategory, FireworksProvider, GoogleProvider as GeminiProvider, GoogleProvider, JaypieToolkit, constants as LLM, Llm, LlmError, LlmMessageRole, LlmMessageType, LlmProgressEventType, LlmQuotaError, LlmRateLimitError, LlmStreamChunkType, LlmTransientError, LlmUnrecoverableError, OpenRouterProvider, Toolkit, XaiProvider, extractReasoning, isLlmOperateInput, isLlmOperateInputContent, isLlmOperateInputFile, isLlmOperateInputImage, jsonSchemaToNaturalSchema, naturalSchemaToJsonSchema, toolkit, tools };
-export type { LlmEffort, LlmErrorOptions, LlmFallbackConfig, LlmHistory, LlmInputContent, LlmInputContentFile, LlmInputContentImage, LlmInputContentText, LlmInputMessage, LlmMessageOptions, LlmModelOption, LlmOperateInput, LlmOperateInputContent, LlmOperateInputFile, LlmOperateInputImage, LlmOperateOptions, LlmOperateResponse, LlmOptions, LlmProgressCallback, LlmProgressEvent, LlmProgressToolCall, LlmProvider, LlmStreamChunk, LlmStreamChunkDone, LlmStreamChunkError, LlmStreamChunkText, LlmStreamChunkToolCall, LlmStreamChunkToolResult, LlmTool };
+export type { LlmCache, LlmEffort, LlmErrorOptions, LlmExchangeCallback, LlmExchangeEnvelope, LlmExchangeRequest, LlmExchangeResolution, LlmExchangeResponse, LlmExchangeTiming, LlmFallbackConfig, LlmHistory, LlmInputContent, LlmInputContentFile, LlmInputContentImage, LlmInputContentText, LlmInputMessage, LlmMessageOptions, LlmModelOption, LlmOperateInput, LlmOperateInputContent, LlmOperateInputFile, LlmOperateInputImage, LlmOperateOptions, LlmOperateResponse, LlmOptions, LlmProgressCallback, LlmProgressEvent, LlmProgressToolCall, LlmProvider, LlmStreamChunk, LlmStreamChunkDone, LlmStreamChunkError, LlmStreamChunkText, LlmStreamChunkToolCall, LlmStreamChunkToolResult, LlmTool };