fix(provider): improve prompt caching

packages/llm/src/protocols/bedrock-converse.ts

@@ -436,21 +436,19 @@ const mapFinishReason = (reason: string): FinishReason => {
   return "unknown"
 }
 
-// AWS Bedrock Converse reports `inputTokens` (inclusive total) with
-// `cacheReadInputTokens` and `cacheWriteInputTokens` as subsets. Pass
-// the total through and derive the non-cached breakdown. Bedrock does
-// not break reasoning out of `outputTokens` for any current model.
+// AWS reports inputTokens separately from cache reads and writes.
+// Bedrock does not break reasoning out of outputTokens for current models.
 const mapUsage = (usage: BedrockUsageSchema | undefined): Usage | undefined => {
   if (!usage) return undefined
   const cacheTotal = (usage.cacheReadInputTokens ?? 0) + (usage.cacheWriteInputTokens ?? 0)
-  const nonCached = ProviderShared.subtractTokens(usage.inputTokens, cacheTotal)
+  const inputTokens = usage.inputTokens === undefined ? undefined : usage.inputTokens + cacheTotal
   return new Usage({
-    inputTokens: usage.inputTokens,
+    inputTokens,
     outputTokens: usage.outputTokens,
-    nonCachedInputTokens: nonCached,
+    nonCachedInputTokens: usage.inputTokens,
     cacheReadInputTokens: usage.cacheReadInputTokens,
     cacheWriteInputTokens: usage.cacheWriteInputTokens,
-    totalTokens: ProviderShared.totalTokens(usage.inputTokens, usage.outputTokens, usage.totalTokens),
+    totalTokens: ProviderShared.totalTokens(inputTokens, usage.outputTokens, usage.totalTokens),
     providerMetadata: { bedrock: usage },
   })
 }

packages/llm/test/provider/bedrock-converse.test.ts

@@ -269,6 +269,39 @@ describe("Bedrock Converse route", () => {
     }),
   )
 
+  it.effect("adds cache reads and writes to Bedrock input usage", () =>
+    Effect.gen(function* () {
+      const body = eventStreamBody(
+        ["messageStart", { role: "assistant" }],
+        ["contentBlockDelta", { contentBlockIndex: 0, delta: { text: "Hello" } }],
+        ["contentBlockStop", { contentBlockIndex: 0 }],
+        ["messageStop", { stopReason: "end_turn" }],
+        [
+          "metadata",
+          {
+            usage: {
+              inputTokens: 5,
+              outputTokens: 2,
+              totalTokens: 12,
+              cacheReadInputTokens: 3,
+              cacheWriteInputTokens: 2,
+            },
+          },
+        ],
+      )
+      const response = yield* LLMClient.generate(baseRequest).pipe(Effect.provide(fixedBytes(body)))
+
+      expect(response.usage).toMatchObject({
+        inputTokens: 10,
+        nonCachedInputTokens: 5,
+        cacheReadInputTokens: 3,
+        cacheWriteInputTokens: 2,
+        outputTokens: 2,
+        totalTokens: 12,
+      })
+    }),
+  )
+
   it.effect("assembles streamed tool call input", () =>
     Effect.gen(function* () {
       const body = eventStreamBody(