fix(opencode): correct Bedrock Luna token usage

packages/opencode/src/session/session.ts

@@ -358,12 +358,21 @@ export const getUsage = (input: { model: Provider.Model; usage: Usage; metadata?
     ),
   )
 
+  // Bedrock GPT-5.6 Luna includes cached tokens in its input count, while the
+  // AI SDK normalizer adds the cache fields again. Remove that duplicate before
+  // splitting the prompt into cached and non-cached tokens below.
+  const duplicatedBedrockCache =
+    input.model.api.npm === "@ai-sdk/amazon-bedrock" && input.model.api.id.includes("openai.gpt-5.6-luna")
+      ? cacheReadInputTokens + cacheWriteInputTokens
+      : 0
+  const contextTokens = safe(inputTokens - duplicatedBedrockCache)
+
   // AI SDK v6 normalized inputTokens to include cached tokens across all providers
   // (including Anthropic/Bedrock which previously excluded them). Always subtract cache
   // tokens to get the non-cached input count for separate cost calculation.
-  const adjustedInputTokens = safe(inputTokens - cacheReadInputTokens - cacheWriteInputTokens)
+  const adjustedInputTokens = safe(contextTokens - cacheReadInputTokens - cacheWriteInputTokens)
 
-  const total = input.usage.totalTokens
+  const total = duplicatedBedrockCache ? safe(contextTokens + outputTokens) : input.usage.totalTokens
 
   const tokens = {
     total,
@@ -376,7 +385,6 @@ export const getUsage = (input: { model: Provider.Model; usage: Usage; metadata?
     },
   }
 
-  const contextTokens = inputTokens
   const costInfo =
     input.model.cost?.tiers
       ?.filter((item) => item.tier.type === "context" && contextTokens > item.tier.size)

packages/opencode/test/session/compaction.test.ts

@@ -9,6 +9,7 @@ import * as Stream from "effect/Stream"
 import { Config } from "@/config/config"
 import { LLM } from "../../src/session/llm"
 import { SessionCompaction } from "../../src/session/compaction"
+import { isOverflow } from "../../src/session/overflow"
 import { Token } from "@/util/token"
 import { Plugin } from "../../src/plugin"
 import { provideTmpdirInstance, TestInstance } from "../fixture/fixture"
@@ -62,6 +63,7 @@ function createModel(opts: {
   input?: number
   cost?: Provider.Model["cost"]
   npm?: string
+  apiID?: string
 }): Provider.Model {
   return {
     id: "test-model",
@@ -81,7 +83,7 @@ function createModel(opts: {
       input: { text: true, image: false, audio: false, video: false },
       output: { text: true, image: false, audio: false, video: false },
     },
-    api: { npm: opts.npm ?? "@ai-sdk/anthropic" },
+    api: { id: opts.apiID ?? "test-model", npm: opts.npm ?? "@ai-sdk/anthropic" },
     options: {},
   } as Provider.Model
 }
@@ -1704,6 +1706,35 @@ describe("SessionNs.getUsage", () => {
     expect(result.tokens.cache.read).toBe(200)
   })
 
+  test("does not double count cached input from Bedrock GPT-5.6", () => {
+    const model = createModel({
+      context: 200_000,
+      input: 200_000,
+      output: 64_000,
+      npm: "@ai-sdk/amazon-bedrock",
+      apiID: "us.openai.gpt-5.6-luna",
+    })
+    const result = SessionNs.getUsage({
+      model,
+      usage: usage({
+        inputTokens: 302_820,
+        outputTokens: 639,
+        totalTokens: 303_459,
+        cacheReadInputTokens: 132_976,
+        cacheWriteInputTokens: 18_433,
+      }),
+    })
+
+    expect(result.tokens).toEqual({
+      total: 152_050,
+      input: 2,
+      output: 639,
+      reasoning: 0,
+      cache: { read: 132_976, write: 18_433 },
+    })
+    expect(isOverflow({ cfg: {} as ConfigV1.Info, tokens: result.tokens, model })).toBe(false)
+  })
+
   test("handles anthropic cache write metadata", () => {
     const model = createModel({ context: 100_000, output: 32_000 })
     const result = SessionNs.getUsage({