/**
 * Inference capacity probe (3.0.0 foundation).
 *
 * Optional backend telemetry adapter. Physical capacity belongs to the backend,
 * not to any agent/assignment/worker. For llama.cpp-compatible servers we can
 * observe `total_slots` from `/props` and per-slot `is_processing` from `/slots`
 * reliably. When a backend does not expose these, the caller reports
 * NOT_OBSERVED instead of guessing.
 *
 * Credentials are resolved per-probe and never persisted by the scheduler.
 */
import type { SharedInferenceResource } from "./types.js";
export interface InferenceCapacityObservation {
    capacity?: number;
    busySlots?: number;
    contextWindow?: number;
    observedAtMs: number;
}
export type InferenceCapacityProbeResult = {
    status: "observed";
    observation: InferenceCapacityObservation;
} | {
    status: "not_observed";
    reason: string;
};
export interface InferenceCapacityProbe {
    probe(resource: SharedInferenceResource): Promise<InferenceCapacityProbeResult>;
}
export interface LlamaCppCapacityProbeOptions {
    /** Resolve the bearer key at probe time (never persisted). */
    apiKey?: string | (() => Promise<string | undefined>);
    fetchImpl?: typeof fetch;
    timeoutMs?: number;
    now?: () => number;
}
/**
 * llama.cpp server `/props` + `/slots` observation.
 *
 * `/props` exposes `total_slots` (the authoritative physical slot count) and
 * `default_generation_settings.n_ctx` (context window). `/slots` exposes each
 * slot's `is_processing` flag, from which busy slots are derived.
 */
export declare class LlamaCppCapacityProbe implements InferenceCapacityProbe {
    private readonly _apiKey?;
    private readonly _fetchImpl;
    private readonly _timeoutMs;
    private readonly _now;
    constructor(options?: LlamaCppCapacityProbeOptions);
    probe(resource: SharedInferenceResource): Promise<InferenceCapacityProbeResult>;
    private _get;
}
//# sourceMappingURL=capacity-probe.d.ts.map