feat: Track cache tokens

src/client.ts

@@ -50,6 +50,8 @@ export interface SpanData {
     prompt_tokens?: number
     completion_tokens?: number
     tokens?: number
+    prompt_cached_tokens?: number
+    prompt_cache_creation_tokens?: number
     reasoning_tokens?: number
   }
   context?: {

src/event-processor.ts

@@ -609,12 +609,18 @@ export class EventProcessor {
     // Mark as processed
     state.processedLlmMessages.add(messageId)
 
-    // Extract token info
+    // Extract token info. OpenCode reports input tokens exclusive of cache
+    // buckets, so normalize prompt_tokens to Braintrust's inclusive convention
+    // by adding cache read/write tokens back in.
     const tokens = messageInfo.tokens as Record<string, unknown> | undefined
+    const cacheTokens = tokens?.cache as Record<string, unknown> | undefined
     const inputTokens = (tokens?.input as number) || 0
     const outputTokens = (tokens?.output as number) || 0
     const reasoningTokens = (tokens?.reasoning as number) || 0
-    const totalTokens = inputTokens + outputTokens + reasoningTokens
+    const cacheReadTokens = (cacheTokens?.read as number) || 0
+    const cacheWriteTokens = (cacheTokens?.write as number) || 0
+    const promptTokens = inputTokens + cacheReadTokens + cacheWriteTokens
+    const totalTokens = promptTokens + outputTokens + reasoningTokens
 
     // Extract model info
     const providerID = (messageInfo.providerID as string) || "unknown"
@@ -662,9 +668,11 @@ export class EventProcessor {
       metrics: {
         start: msToSeconds(time.created as number),
         end: msToSeconds(time.completed as number),
-        prompt_tokens: inputTokens,
+        prompt_tokens: promptTokens,
         completion_tokens: outputTokens,
         tokens: totalTokens,
+        prompt_cached_tokens: cacheReadTokens || undefined,
+        prompt_cache_creation_tokens: cacheWriteTokens || undefined,
         reasoning_tokens: reasoningTokens || undefined,
       },
       metadata: {

src/span-sink.ts

@@ -65,6 +65,9 @@ export interface SpanTree {
     prompt_tokens?: number
     completion_tokens?: number
     tokens?: number
+    prompt_cached_tokens?: number
+    prompt_cache_creation_tokens?: number
+    reasoning_tokens?: number
   }
   metadata?: Record<string, unknown>
   children: SpanTree[]

src/test-helpers.ts

@@ -297,7 +297,12 @@ export function messageCompleted(
   sessionID: string,
   messageID: string,
   options?: {
-    tokens?: { input: number; output: number }
+    tokens?: {
+      input: number
+      output: number
+      reasoning?: number
+      cache?: { read?: number; write?: number }
+    }
     model?: { providerID: string; modelID: string }
     time?: { created: number; completed: number }
   },
@@ -317,10 +322,13 @@ export function messageCompleted(
     path: { cwd: "/test", root: "/test" },
     cost: 0.001,
     tokens: {
-      input: options?.tokens?.input || 10,
-      output: options?.tokens?.output || 5,
-      reasoning: 0,
-      cache: { read: 0, write: 0 },
+      input: options?.tokens?.input ?? 10,
+      output: options?.tokens?.output ?? 5,
+      reasoning: options?.tokens?.reasoning ?? 0,
+      cache: {
+        read: options?.tokens?.cache?.read ?? 0,
+        write: options?.tokens?.cache?.write ?? 0,
+      },
     },
   }
   return {
@@ -379,6 +387,9 @@ export interface ExpectedSpan {
     prompt_tokens?: number
     completion_tokens?: number
     tokens?: number
+    prompt_cached_tokens?: number
+    prompt_cache_creation_tokens?: number
+    reasoning_tokens?: number
   }
   metadata?: Record<string, unknown>
   children?: ExpectedSpan[]
@@ -549,6 +560,21 @@ function spanMatchesSingle(actual: SpanTree, expected: ExpectedSpan): boolean {
       return false
     if (expected.metrics.tokens !== undefined && actual.metrics?.tokens !== expected.metrics.tokens)
       return false
+    if (
+      expected.metrics.prompt_cached_tokens !== undefined &&
+      actual.metrics?.prompt_cached_tokens !== expected.metrics.prompt_cached_tokens
+    )
+      return false
+    if (
+      expected.metrics.prompt_cache_creation_tokens !== undefined &&
+      actual.metrics?.prompt_cache_creation_tokens !== expected.metrics.prompt_cache_creation_tokens
+    )
+      return false
+    if (
+      expected.metrics.reasoning_tokens !== undefined &&
+      actual.metrics?.reasoning_tokens !== expected.metrics.reasoning_tokens
+    )
+      return false
   }
 
   if (expected.input !== undefined) {
@@ -630,6 +656,38 @@ export function getDiff(actual: SpanTree | null, expected: ExpectedSpan, path =
         `${path}.metrics.completion_tokens: expected ${expected.metrics.completion_tokens}, got ${actual.metrics?.completion_tokens}`,
       )
     }
+    if (
+      expected.metrics.tokens !== undefined &&
+      actual.metrics?.tokens !== expected.metrics.tokens
+    ) {
+      diffs.push(
+        `${path}.metrics.tokens: expected ${expected.metrics.tokens}, got ${actual.metrics?.tokens}`,
+      )
+    }
+    if (
+      expected.metrics.prompt_cached_tokens !== undefined &&
+      actual.metrics?.prompt_cached_tokens !== expected.metrics.prompt_cached_tokens
+    ) {
+      diffs.push(
+        `${path}.metrics.prompt_cached_tokens: expected ${expected.metrics.prompt_cached_tokens}, got ${actual.metrics?.prompt_cached_tokens}`,
+      )
+    }
+    if (
+      expected.metrics.prompt_cache_creation_tokens !== undefined &&
+      actual.metrics?.prompt_cache_creation_tokens !== expected.metrics.prompt_cache_creation_tokens
+    ) {
+      diffs.push(
+        `${path}.metrics.prompt_cache_creation_tokens: expected ${expected.metrics.prompt_cache_creation_tokens}, got ${actual.metrics?.prompt_cache_creation_tokens}`,
+      )
+    }
+    if (
+      expected.metrics.reasoning_tokens !== undefined &&
+      actual.metrics?.reasoning_tokens !== expected.metrics.reasoning_tokens
+    ) {
+      diffs.push(
+        `${path}.metrics.reasoning_tokens: expected ${expected.metrics.reasoning_tokens}, got ${actual.metrics?.reasoning_tokens}`,
+      )
+    }
   }
 
   if (expected.error !== undefined && !nameMatches(actual.error, expected.error)) {

src/tracing.test.ts

@@ -70,6 +70,44 @@ describe("Event to Span Transformation", () => {
     )
   })
 
+  it("tracks OpenCode cache read/write tokens on LLM spans", async () => {
+    const sessionId = "ses_cache"
+    const messageId = "msg_cache"
+
+    await assertEventsProduceTree(
+      session(
+        sessionId,
+        sessionCreated(sessionId),
+        chatMessage("Use the cached context"),
+        textPart(sessionId, messageId, "Done."),
+        messageCompleted(sessionId, messageId, {
+          tokens: { input: 10, output: 5, cache: { read: 3, write: 2 } },
+        }),
+        sessionIdle(sessionId),
+      ),
+      {
+        span_attributes: { name: "OpenCode: test-project", type: "task" },
+        children: [
+          {
+            span_attributes: { name: "Turn 1", type: "task" },
+            children: [
+              {
+                span_attributes: { name: "anthropic/claude-3-haiku", type: "llm" },
+                metrics: {
+                  prompt_tokens: 15,
+                  completion_tokens: 5,
+                  tokens: 20,
+                  prompt_cached_tokens: 3,
+                  prompt_cache_creation_tokens: 2,
+                },
+              },
+            ],
+          },
+        ],
+      },
+    )
+  })
+
   it("stores modelID on turn metadata from chat.message hook", async () => {
     const sessionId = "ses_turn_model"
     const messageId = "msg_turn_model"

src/tracing.ts

@@ -457,12 +457,18 @@ export function createTracingHooks(
           // Mark as processed
           state.processedLlmMessages.add(messageId)
 
-          // Extract token info
+          // Extract token info. OpenCode reports input tokens exclusive of cache
+          // buckets, so normalize prompt_tokens to Braintrust's inclusive
+          // convention by adding cache read/write tokens back in.
           const tokens = messageInfo.tokens as Record<string, unknown> | undefined
+          const cacheTokens = tokens?.cache as Record<string, unknown> | undefined
           const inputTokens = (tokens?.input as number) || 0
           const outputTokens = (tokens?.output as number) || 0
           const reasoningTokens = (tokens?.reasoning as number) || 0
-          const totalTokens = inputTokens + outputTokens + reasoningTokens
+          const cacheReadTokens = (cacheTokens?.read as number) || 0
+          const cacheWriteTokens = (cacheTokens?.write as number) || 0
+          const promptTokens = inputTokens + cacheReadTokens + cacheWriteTokens
+          const totalTokens = promptTokens + outputTokens + reasoningTokens
 
           // Extract model info
           const providerID = (messageInfo.providerID as string) || "unknown"
@@ -524,9 +530,11 @@ export function createTracingHooks(
             metrics: {
               start: msToSeconds(time.created as number),
               end: msToSeconds(time.completed as number),
-              prompt_tokens: inputTokens,
+              prompt_tokens: promptTokens,
               completion_tokens: outputTokens,
               tokens: totalTokens,
+              prompt_cached_tokens: cacheReadTokens || undefined,
+              prompt_cache_creation_tokens: cacheWriteTokens || undefined,
               reasoning_tokens: reasoningTokens || undefined,
             },
             metadata: {