/**
 * Port of the Mode-Latency-Benchmark provider table and request builders
 * (`upstream/Mode-Latency-Benchmark/src/model_latency_benchmark/core.py`,
 * commit 8123838, Apache-2.0), plus the percentile statistics the plugin
 * aggregates from measured call durations.
 *
 * The upstream tool probes providers over HTTP; the plugin reuses its
 * `BenchmarkResult` vocabulary (ttft / total time / tokens / error) and adds
 * `latencyStats` for per-model percentile aggregation over the llm/stream
 * wrapper.
 *
 * @module dsh-budget/estimate/latency-stats
 */
/** Authentication style of one upstream provider config. */
export type ProviderAuthType = 'bearer' | 'query_param';
/** One provider config entry, verbatim from upstream core.py. */
export interface ProviderSpec {
    /** Display name. */
    name: string;
    /** HTTP endpoint. */
    endpoint: string;
    /** Environment variable carrying the API key. */
    envKey: string;
    /** Default model id. */
    model: string;
    /** Authentication style. */
    authType: ProviderAuthType;
    /** Terminal color tag. */
    color: string;
    /** Terminal icon. */
    icon: string;
}
/** The upstream provider table (4 entries), verbatim. */
export declare const PROVIDERS: Readonly<Record<string, ProviderSpec>>;
/** One benchmark outcome (upstream BenchmarkResult vocabulary). */
export interface BenchmarkResult {
    /** Provider table key. */
    providerId: string;
    /** Provider display name. */
    providerName: string;
    /** Time to first token in seconds; null when never reached. */
    ttft: number | null;
    /** Total response time in seconds; null on failure. */
    totalTime: number | null;
    /** Chunk/token count observed. */
    tokens: number;
    /** Failure detail; null on success. */
    error: string | null;
    /** Whether the run completed. */
    success: boolean;
}
/**
 * Build the request headers for one provider, upstream style.
 *
 * @param provider - provider config (reads `authType` only).
 * @param apiKey - provider API key (used only for bearer auth).
 * @returns the header map.
 */
export declare function buildRequestHeaders(provider: Readonly<ProviderSpec>, apiKey: string): Record<string, string>;
/**
 * Build the request body for one provider, upstream style.
 *
 * @param provider - provider config.
 * @param prompt - the user prompt.
 * @returns the JSON body object.
 */
export declare function buildRequestBody(provider: Readonly<ProviderSpec>, prompt: string): Record<string, unknown>;
/**
 * Build the request URL for one provider, upstream style (query-param auth
 * appends the key).
 *
 * @param provider - provider config.
 * @param apiKey - provider API key (used only for query-param auth).
 * @returns the request URL.
 */
export declare function buildRequestUrl(provider: Readonly<ProviderSpec>, apiKey: string): string;
/** Percentile statistics over one duration sample window (milliseconds). */
export interface LatencyStats {
    /** Sample count. */
    count: number;
    /** Arithmetic mean (ms). */
    mean: number;
    /** Minimum (ms). */
    min: number;
    /** Maximum (ms). */
    max: number;
    /** 50th percentile (ms). */
    p50: number;
    /** 95th percentile (ms). */
    p95: number;
    /** 99th percentile (ms). */
    p99: number;
    /** Population standard deviation (ms). */
    stdev: number;
}
/**
 * Aggregate latency statistics over a sample window. Deterministic and pure:
 * callers own the sample retention policy.
 *
 * @param samples - durations in milliseconds (any order, may be empty).
 * @returns the statistics; every field is 0 for an empty window.
 */
export declare function latencyStats(samples: readonly number[]): LatencyStats;
/**
 * Format a millisecond duration the way the upstream tool displays it.
 *
 * @param ms - duration in milliseconds.
 * @returns `123ms` for sub-second values, `1.23s` otherwise.
 */
export declare function formatMs(ms: number): string;
//# sourceMappingURL=latency-stats.d.ts.map