/**
 * @license
 * Copyright 2026 Steven Roussey <sroussey@gmail.com>
 * SPDX-License-Identifier: Apache-2.0
 */
import type { AiProviderRunFn, AiSessionContext, CacheCheckpointTaskInput, CacheCheckpointTaskOutput, ChatMessage, ToolDefinition } from "@workglow/ai";
import type { OpenAiModelConfig } from "./OpenAI_ModelSchema";
/**
 * Merges a checkpoint prefix into the unified generation input: the prefix's
 * messages come first, the caller's tail follows (its `messages`, or its
 * `prompt` lifted into a user message — the shared message builders only fall
 * back to `prompt` when the message list is empty), the prefix system prompt
 * applies when the call carries none (an explicit `""` falls through too —
 * suppressing the warmed instructions would diverge the replayed prefix), and
 * the effective `tools` are the caller's when it declares any, else the
 * prefix's — tools precede the conversation in the serialized request, so a
 * tools-warmed prefix consumed without them never shares the warm-up's cached
 * prefix. Returns undefined when the session has no prefix so plain calls
 * take the unmodified path.
 *
 * OpenAI's prompt caching is automatic and keyed on the request's literal
 * token prefix, so replaying identical prefix content is both the correctness
 * path and the cache-hit path — no per-request cache annotations exist.
 */
export declare function mergeOpenAICheckpointPrefix(session: AiSessionContext | undefined, input: {
    readonly messages?: readonly unknown[] | undefined;
    readonly systemPrompt?: string | undefined;
    readonly prompt?: unknown;
    readonly tools?: readonly ToolDefinition[] | undefined;
}): {
    messages: readonly ChatMessage[];
    systemPrompt: string | undefined;
    tools: readonly ToolDefinition[] | undefined;
} | undefined;
/**
 * Warm-up run-fn for `["cache.checkpoint"]` on OpenAI. Sends the prefix once
 * (minimal `max_output_tokens` — the Responses API floor is 16) so the
 * server-side automatic prompt cache is populated before consumers arrive.
 *
 * The warm-up is advisory: OpenAI caches long prompt prefixes on its own and
 * may evict at any time, so consumption never depends on this call having
 * succeeded — consumers always replay the full prefix content.
 * {@link finalizeResponsesRequest} derives the `prompt_cache_key` from the
 * request's model + instructions + tools, so this warm-up and every consumer
 * replaying the same prefix converge on the same key without coordination.
 */
export declare const OpenAI_CacheCheckpoint_Stream: AiProviderRunFn<CacheCheckpointTaskInput, CacheCheckpointTaskOutput, OpenAiModelConfig>;