fix(usage): correct cache rate denominator and total tokens fallback

calculateCacheRate now uses inputTokens + cachedTokens as denominator
instead of totalTokens (which excludes cache read tokens). Also fix
extractTotalTokens fallback to not include cachedTokens.

Author: NGLSL

src/components/usage/RequestEventsDetailsCard.tsx

@@ -91,11 +91,15 @@ const extractMetric = (value: unknown): number | null => {
   return value;
 };
 
-const calculateCacheRate = (cachedTokens: number, totalTokens: number): number | null => {
-  if (totalTokens <= 0) {
+// 缓存命中率 = 缓存读取token / (输入token + 缓存读取token)
+// 分母用 inputTokens + cachedTokens 而不是 totalTokens，因为 totalTokens 不含缓存token，
+// 在缓存量很大时（如 147K cached vs 797 total），用 totalTokens 做分母会导致超过 10000% 的荒谬百分比
+const calculateCacheRate = (cachedTokens: number, inputTokens: number): number | null => {
+  const denominator = inputTokens + Math.max(cachedTokens, 0);
+  if (denominator <= 0) {
     return null;
   }
-  return Math.max(cachedTokens, 0) / totalTokens;
+  return Math.max(cachedTokens, 0) / denominator;
 };
 
 const getExportCacheRate = (value: number | null): string => {
@@ -209,7 +213,7 @@ export function RequestEventsDetailsCard({
         );
         const firstByteLatencyMs = extractFirstByteLatencyMs(detail.first_byte_latency_ms);
         const latencyMs = extractLatencyMs(detail);
-        const cacheRate = calculateCacheRate(cachedTokens, totalTokens);
+        const cacheRate = calculateCacheRate(cachedTokens, inputTokens);
         const chunkCount = extractMetric(detail.chunk_count);
         const responseBytes = extractMetric(detail.response_bytes);
         const apiResponseBytes = extractMetric(detail.api_response_bytes);

src/utils/usage.ts

@@ -711,15 +711,13 @@ export function extractTotalTokens(detail: unknown): number {
   if (typeof tokens.total_tokens === 'number') {
     return tokens.total_tokens;
   }
+  // total_tokens 缺失时的兜底计算：只累加 input + output + reasoning，
+  // 不加 cachedTokens，因为缓存读取是已缓存的旧 token，不是本次消耗的新 token
   const inputTokens = typeof tokens.input_tokens === 'number' ? tokens.input_tokens : 0;
   const outputTokens = typeof tokens.output_tokens === 'number' ? tokens.output_tokens : 0;
   const reasoningTokens = typeof tokens.reasoning_tokens === 'number' ? tokens.reasoning_tokens : 0;
-  const cachedTokens = Math.max(
-    typeof tokens.cached_tokens === 'number' ? Math.max(tokens.cached_tokens, 0) : 0,
-    typeof tokens.cache_tokens === 'number' ? Math.max(tokens.cache_tokens, 0) : 0
-  );
 
-  return inputTokens + outputTokens + reasoningTokens + cachedTokens;
+  return inputTokens + outputTokens + reasoningTokens;
 }
 
 /**