fix: 修复缓存

src/services/api/claude.ts

@@ -2907,7 +2907,12 @@ async function* queryModel(
     provider: getAPIProvider(),
     input: convertMessagesToLangfuse(messagesForAPI, systemPrompt),
     output: convertOutputToLangfuse(newMessages),
-    usage: { input_tokens: usage.input_tokens, output_tokens: usage.output_tokens },
+    usage: {
+      input_tokens: usage.input_tokens,
+      output_tokens: usage.output_tokens,
+      cache_creation_input_tokens: usage.cache_creation_input_tokens,
+      cache_read_input_tokens: usage.cache_read_input_tokens,
+    },
     startTime: new Date(startIncludingRetries),
     endTime: new Date(),
     completionStartTime: ttftMs > 0 ? new Date(start + ttftMs) : undefined,

src/services/langfuse/__tests__/langfuse.test.ts

@@ -284,6 +284,48 @@ describe('Langfuse integration', () => {
       }))
       expect(mockRootEnd).toHaveBeenCalled()
     })
+
+    test('includes cache tokens in usageDetails when provided', async () => {
+      process.env.LANGFUSE_PUBLIC_KEY = 'pk-test'
+      process.env.LANGFUSE_SECRET_KEY = 'sk-test'
+      const { createTrace, recordLLMObservation } = await import('../tracing.js')
+      const span = createTrace({ sessionId: 's1', model: 'claude-3', provider: 'firstParty' })
+      mockStartObservation.mockClear()
+      mockRootUpdate.mockClear()
+      recordLLMObservation(span, {
+        model: 'claude-3',
+        provider: 'firstParty',
+        input: [],
+        output: [],
+        usage: { input_tokens: 10000, output_tokens: 50, cache_creation_input_tokens: 2000, cache_read_input_tokens: 7000 },
+      })
+      expect(mockRootUpdate).toHaveBeenCalledWith(expect.objectContaining({
+        usageDetails: {
+          input: 19000, // 10000 + 2000 + 7000
+          output: 50,
+          cache_read: 7000,
+          cache_creation: 2000,
+        },
+      }))
+    })
+
+    test('omits cache fields when not provided', async () => {
+      process.env.LANGFUSE_PUBLIC_KEY = 'pk-test'
+      process.env.LANGFUSE_SECRET_KEY = 'sk-test'
+      const { createTrace, recordLLMObservation } = await import('../tracing.js')
+      const span = createTrace({ sessionId: 's1', model: 'claude-3', provider: 'firstParty' })
+      mockRootUpdate.mockClear()
+      recordLLMObservation(span, {
+        model: 'claude-3',
+        provider: 'firstParty',
+        input: [],
+        output: [],
+        usage: { input_tokens: 100, output_tokens: 20 },
+      })
+      expect(mockRootUpdate).toHaveBeenCalledWith(expect.objectContaining({
+        usageDetails: { input: 100, output: 20 },
+      }))
+    })
   })
 
   describe('recordToolObservation', () => {

src/services/langfuse/tracing.ts

@@ -68,7 +68,12 @@ export function recordLLMObservation(
     provider: string
     input: unknown
     output: unknown
-    usage: { input_tokens: number; output_tokens: number }
+    usage: {
+      input_tokens: number
+      output_tokens: number
+      cache_creation_input_tokens?: number
+      cache_read_input_tokens?: number
+    }
     startTime?: Date
     endTime?: Date
     completionStartTime?: Date
@@ -109,11 +114,17 @@ export function recordLLMObservation(
       gen.otelSpan.setAttribute(LangfuseOtelSpanAttributes.TRACE_USER_ID, userId)
     }
 
+    // Anthropic splits input into uncached + cache_read + cache_creation.
+    // Langfuse's "input" should be the total prompt tokens so cost calc is correct.
+    const cacheRead = params.usage.cache_read_input_tokens ?? 0
+    const cacheCreation = params.usage.cache_creation_input_tokens ?? 0
     gen.update({
       output: params.output,
       usageDetails: {
-        input: params.usage.input_tokens,
+        input: params.usage.input_tokens + cacheCreation + cacheRead,
         output: params.usage.output_tokens,
+        ...(cacheRead > 0 && { cache_read: cacheRead }),
+        ...(cacheCreation > 0 && { cache_creation: cacheCreation }),
       },
     })