import type { EnvVar } from "../internals/envfile.js";
import type { SizingResult } from "./sizing.js";
/**
 * The local-inference env block from the blueprint's "Local Inference Engine
 * Configuration Profiles". The four static keys are emitted verbatim; the one
 * computed key (`OLLAMA_NUM_PARALLEL`) carries the sizing engine's parallel
 * count — Ollama's documented variable for that exact metric.
 *
 * Order is fixed so the managed block regenerates byte-identically on re-run.
 */
/** Static keys whose values come straight from the blueprint (not host-derived). */
export declare const STATIC_OLLAMA_ENV: readonly EnvVar[];
/** The scope name used for the aih-managed env region in the shell profile. */
export declare const HARDWARE_SCOPE = "hardware";
/**
 * Build the ordered env vars for the inference engine: the four static blueprint
 * keys followed by the computed `OLLAMA_NUM_PARALLEL`. `llamacpp` shares the same
 * tuning surface, so the block is identical regardless of engine today; the
 * parameter is kept for forward compatibility and intent.
 */
export declare function inferenceEnv(sizing: SizingResult): EnvVar[];
/**
 * Human-readable summary of the profiled host and the derived budget. Emitted as
 * a `doc` action: the memory ceiling and thread pool are per-request runtime
 * options in Ollama (not server env vars), so they are surfaced as guidance
 * rather than invented environment keys.
 */
export declare function profileDoc(profile: {
    totalRamGb: number;
    cpuCores: number;
    gpuName: string;
    vramGb: number;
    backend: string;
}, sizing: SizingResult, modelSizeGb: number): string;