import type { AssistantUsageMetric, CacheUsageTotals } from "./types.js";

/**
 * Prompt-cache TTL. Anthropic's explicit cache has a documented 5-minute
 * default; OpenAI-style implicit prefix caches (openai-completions /
 * openai-responses and their many compatible providers) evict on the order
 * of 5-10+ minutes with no committed bound, so applying the 5-minute
 * constant to them both mislabels healthy idle gaps as cache expiry and
 * hides real causes behind a wrong tag.
 */
export const CACHE_TTL_MS = 5 * 60 * 1000;
export const IMPLICIT_CACHE_TTL_MS = 10 * 60 * 1000;

/** TTL that applies to a model's API protocol. */
export function cacheTtlMsForApi(api: string | undefined): number {
  return api === "anthropic-messages" ? CACHE_TTL_MS : IMPLICIT_CACHE_TTL_MS;
}

/** True when the gap to the previous turn exceeds the applicable cache TTL. */
export function isIdleExpired(idleMs: number | undefined, ttlMs: number = CACHE_TTL_MS): boolean {
  return idleMs !== undefined && idleMs > ttlMs;
}

/**
 * Canonical cache-hit % formula.
 *
 * Denominator = full prompt size sent on the turn.
 * Anthropic-style: input excludes newly-cached tokens, which arrive in
 * cacheWrite — so both must be included.
 * OpenAI-style: cacheWrite is 0, so this is backwards-compatible.
 */
export function computeCacheHitPercent(input: number, cacheRead: number, cacheWrite: number): number {
  const denominator = input + cacheRead + cacheWrite;
  if (denominator <= 0) return 0;
  return (cacheRead / denominator) * 100;
}

export function emptyTotals(): CacheUsageTotals {
  return {
    input: 0,
    output: 0,
    cacheRead: 0,
    cacheWrite: 0,
    totalTokens: 0,
    assistantMessages: 0,
  };
}

export function addToTotals(totals: CacheUsageTotals, message: AssistantUsageMetric): void {
  totals.input += message.input;
  totals.output += message.output;
  totals.cacheRead += message.cacheRead;
  totals.cacheWrite += message.cacheWrite;
  totals.totalTokens += message.totalTokens;
  totals.assistantMessages += 1;
}