/**
 * Bridges MCP `sampling/createMessage` server requests to Codeep's host LLM.
 *
 * MCP servers that opt into the `sampling` capability send the host
 * (Codeep) a request to generate a completion on their behalf — usually
 * because they want LLM reasoning without their own provider keys. Per
 * spec, the host is free to refuse, swap models, or strip context; we
 * forward the messages to the active provider via `chat()` and return
 * just the assistant text.
 *
 * Notes on the bridge surface:
 *   - We strip image content (provider matrix varies; safer to skip than
 *     surprise the model). A future iteration can route images through
 *     the vision integration in mcpIntegration.ts.
 *   - We respect `params.modelPreferences.hints[].name` as an *advisory*
 *     model override — only if the user has the provider for it
 *     configured; otherwise we stay on the active model.
 *   - We honour `temperature`, `maxTokens`, `stopSequences` only where
 *     the underlying chat() path supports them (today: none — chat() uses
 *     the agent's configured temperature/maxTokens). Pass-through hooks
 *     are wired so the spec contract is honoured if/when chat() grows
 *     those knobs.
 *
 * Cost guard: every sampling request bills the user's active provider, so
 * a misbehaving server can drain credits. We enforce a per-server rate
 * limit (≥1 s spacing) and a per-process cap, and surface every accepted
 * request on stderr so the user can see what's happening.
 */
import type { SamplingCreateMessageParams, SamplingCreateMessageResult } from './mcpClient.js';
/** Reset the per-server counters. Called on session boundaries. */
export declare function resetSamplingBudget(): void;
export declare function handleMcpSamplingRequest(params: SamplingCreateMessageParams, serverName?: string): Promise<SamplingCreateMessageResult>;