billion-context 0.1.111-pr.777.621 → 0.1.111-pr.783.626

package/dist/index.js

@@ -54644,7 +54644,7 @@ function usageTotals(protocol, usage) {
   }
   if (protocol === "openai") {
     const prompt = num4(usage["prompt_tokens"]);
-    const cached = num4(usage["prompt_tokens_details"]?.["cached_tokens"]);
+    const cached = num4(usage["prompt_tokens_details"]?.["cached_tokens"]) ?? num4(usage["prompt_cache_hit_tokens"]);
     return {
       total: prompt !== void 0 ? promptInputTotal("openai", prompt, cached) : void 0,
       cached
@@ -55873,6 +55873,12 @@ function stripFinishReasonChunk(buf) {
     return buf;
   }
 }
+function openaiCachedTokens(u2) {
+  const pd = u2.prompt_tokens_details;
+  if (typeof pd?.cached_tokens === "number") return pd.cached_tokens;
+  if (typeof u2.prompt_cache_hit_tokens === "number") return u2.prompt_cache_hit_tokens;
+  return void 0;
+}
 function createOpenaiAdapter(requestBody, clientSystem, absorbName) {
   const model = requestBody.model ?? "unknown";
   let responseId = `chatcmpl-proxy-${Date.now()}`;
@@ -56062,12 +56068,11 @@ function createOpenaiAdapter(requestBody, clientSystem, absorbName) {
         if (!choice) {
           if (parsed.usage) {
             const u2 = parsed.usage;
-            const pd = u2.prompt_tokens_details;
             yield {
               kind: "usage",
               inputTokens: typeof u2.prompt_tokens === "number" ? u2.prompt_tokens : void 0,
               outputTokens: typeof u2.completion_tokens === "number" ? u2.completion_tokens : void 0,
-              cachedTokens: typeof pd?.cached_tokens === "number" ? pd.cached_tokens : void 0
+              cachedTokens: openaiCachedTokens(u2)
             };
             if (sawRealToolCall) {
               yield { kind: "meta", chunk: rawBuf };
@@ -56082,12 +56087,11 @@ function createOpenaiAdapter(requestBody, clientSystem, absorbName) {
           const hadToolCalls = [...pending.values()].some((tc) => tc.name.length > 0 || tc.id.length > 0);
           yield* settleToolCalls();
           const u2 = parsed.usage;
-          const pd = u2?.prompt_tokens_details;
           yield {
             kind: "usage",
             inputTokens: typeof u2?.prompt_tokens === "number" ? u2.prompt_tokens : void 0,
             outputTokens: typeof u2?.completion_tokens === "number" ? u2.completion_tokens : void 0,
-            cachedTokens: typeof pd?.cached_tokens === "number" ? pd.cached_tokens : void 0
+            cachedTokens: u2 ? openaiCachedTokens(u2) : void 0
           };
           if (sawRealToolCall) {
             const chunk = rawBuf;
@@ -58551,7 +58555,9 @@ function usageFromSseEvent(obj) {
     return {
       inputTokens: num3(usage["prompt_tokens"]),
       outputTokens: num3(usage["completion_tokens"]),
-      cachedTokens: num3(usage["prompt_tokens_details"]?.["cached_tokens"])
+      // DeepSeek-style upstreams report KV-cache hits as top-level
+      // prompt_cache_hit_tokens instead of the standard details field (#779).
+      cachedTokens: num3(usage["prompt_tokens_details"]?.["cached_tokens"]) ?? num3(usage["prompt_cache_hit_tokens"])
     };
   }
   return void 0;
@@ -59060,7 +59066,8 @@ async function pipePluginJson(stream2, res, session, protocol) {
         applyUsageSample(session, {
           inputTokens: input,
           outputTokens: num3(usage["completion_tokens"]) ?? num3(usage["output_tokens"]),
-          cachedTokens: num3(usage["prompt_tokens_details"]?.["cached_tokens"]) ?? num3(usage["input_tokens_details"]?.["cached_tokens"]) ?? num3(usage["cache_read_input_tokens"])
+          cachedTokens: num3(usage["prompt_tokens_details"]?.["cached_tokens"]) ?? num3(usage["input_tokens_details"]?.["cached_tokens"]) ?? num3(usage["cache_read_input_tokens"]) ?? // #779: DeepSeek-style top-level field (openai wire)
+          num3(usage["prompt_cache_hit_tokens"])
         }, protocol);
         markDirty(session);
       }