fix: extract thoughtsTokenCount for Gemini thinking models

- Add reasoning_tokens extraction from Gemini usageMetadata.thoughtsTokenCount
  in worker's getDetailedUsage()
- Add reasoningTokens to GoogleBodyProcessor output in Jawn
- Fix promptTokens calculation when no caching is present
  (was returning undefined instead of raw promptTokenCount)

This fixes incorrect token counting and cost calculations for Gemini
thinking models (2.5 Flash, 3.x Pro) where thoughtsTokenCount was not
being stored separately.

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Author: colegottdank

valhalla/jawn/src/lib/shared/bodyProcessors/__tests__/googleBodyProcessor.test.ts

@@ -118,4 +118,58 @@ describe("GoogleBodyProcessor", () => {
       promptCacheReadTokens: 13644,
     });
   });
+
+  it("handles Gemini thinking model responses with thoughtsTokenCount", async () => {
+    const body = {
+      candidates: [
+        {
+          content: {
+            role: "model",
+            parts: [
+              {
+                text: "25 * 47 = 1175",
+              },
+            ],
+          },
+          finishReason: "STOP",
+          avgLogprobs: -0.24105726576781206,
+        },
+      ],
+      usageMetadata: {
+        promptTokenCount: 15,
+        candidatesTokenCount: 359,
+        totalTokenCount: 1035,
+        thoughtsTokenCount: 661,
+        trafficType: "ON_DEMAND",
+        promptTokensDetails: [
+          {
+            modality: "TEXT",
+            tokenCount: 15,
+          },
+        ],
+        candidatesTokensDetails: [
+          {
+            modality: "TEXT",
+            tokenCount: 359,
+          },
+        ],
+      },
+      modelVersion: "gemini-2.5-flash",
+      createTime: "2026-02-24T06:59:17.723195Z",
+      responseId: "RUydafuRLIuolu8P8YbxgAg",
+    };
+
+    const { usage } = await parse(body);
+
+    expect(usage).toEqual({
+      totalTokens: 1035,
+      promptTokens: 15,
+      // thoughts + candidates (661 + 359)
+      completionTokens: 1020,
+      // reasoningTokens should be extracted separately
+      reasoningTokens: 661,
+      heliconeCalculated: false,
+      promptCacheReadTokens: undefined,
+    });
+  });
 });

valhalla/jawn/src/lib/shared/bodyProcessors/googleBodyProcessor.ts

@@ -70,19 +70,23 @@ export class GoogleBodyProcessor implements IBodyProcessor {
       ?.promptTokenCount;
     const cachedContentTokens = usageMetadataItem?.usageMetadata
       ?.cachedContentTokenCount;
+    // If there's no caching, return raw promptTokenCount
+    // If there's caching, subtract cached tokens from prompt tokens
     const adjustedPromptTokens =
-      promptTokens !== undefined && cachedContentTokens !== undefined
-        ? promptTokens - cachedContentTokens
+      promptTokens !== undefined
+        ? promptTokens - (cachedContentTokens ?? 0)
         : undefined;
 
+    const thoughtsTokenCount = usageMetadataItem?.usageMetadata?.thoughtsTokenCount ?? 0;
+    const candidatesTokenCount = usageMetadataItem?.usageMetadata?.candidatesTokenCount ?? 0;
+
     return ok({
       processedBody: parsedResponseBody,
       usage: {
         totalTokens: usageMetadataItem?.usageMetadata?.totalTokenCount,
         promptTokens: adjustedPromptTokens,
-        completionTokens:
-          (usageMetadataItem?.usageMetadata?.thoughtsTokenCount ?? 0) +
-          (usageMetadataItem?.usageMetadata?.candidatesTokenCount ?? 0),
+        completionTokens: thoughtsTokenCount + candidatesTokenCount,
+        reasoningTokens: thoughtsTokenCount > 0 ? thoughtsTokenCount : undefined,
         heliconeCalculated: false,
         promptCacheReadTokens: cachedContentTokens,
       },

worker/src/lib/dbLogger/DBLoggable.ts

@@ -532,6 +532,7 @@ export class DBLoggable {
           promptTokenCount?: number;
           candidatesTokenCount?: number;
           cachedContentTokenCount?: number;
+          thoughtsTokenCount?: number;
         };
       };
       const usageMetadata = response.usageMetadata;
@@ -540,6 +541,7 @@ export class DBLoggable {
         prompt_tokens: usageMetadata?.promptTokenCount,
         completion_tokens: usageMetadata?.candidatesTokenCount,
         prompt_cache_read_tokens: usageMetadata?.cachedContentTokenCount,
+        reasoning_tokens: usageMetadata?.thoughtsTokenCount,
       };
     }