fix: max tokens when using models like opus with providers other than anthropic (#4307)

3 files changed, 101 insertions, 3 deletions

diff --git a/packages/opencode/src/provider/transform.ts b/packages/opencode/src/provider/transform.ts
index e578d806f..668f30412 100644
--- a/packages/opencode/src/provider/transform.ts
+++ b/packages/opencode/src/provider/transform.ts
@@ -176,7 +176,7 @@ export namespace ProviderTransform {
   }
 
   export function maxOutputTokens(
-    providerID: string,
+    npm: string,
     options: Record<string, any>,
     modelLimit: number,
     globalLimit: number,
@@ -184,7 +184,7 @@ export namespace ProviderTransform {
     const modelCap = modelLimit || globalLimit
     const standardLimit = Math.min(modelCap, globalLimit)
 
-    if (providerID === "anthropic") {
+    if (npm === "@ai-sdk/anthropic") {
       const thinking = options?.["thinking"]
       const budgetTokens = typeof thinking?.["budgetTokens"] === "number" ? thinking["budgetTokens"] : 0
       const enabled = thinking?.["type"] === "enabled"
diff --git a/packages/opencode/src/session/prompt.ts b/packages/opencode/src/session/prompt.ts
index f4dc84be5..ae6aff220 100644
--- a/packages/opencode/src/session/prompt.ts
+++ b/packages/opencode/src/session/prompt.ts
@@ -345,7 +345,7 @@ export namespace SessionPrompt {
           maxRetries: 0,
           activeTools: Object.keys(tools).filter((x) => x !== "invalid"),
           maxOutputTokens: ProviderTransform.maxOutputTokens(
-            model.providerID,
+            model.npm ?? "",
             params.options,
             model.info.limit.output,
             OUTPUT_TOKEN_MAX,
diff --git a/packages/opencode/test/provider/transform.test.ts b/packages/opencode/test/provider/transform.test.ts
new file mode 100644
index 000000000..e6080d54c
--- /dev/null
+++ b/packages/opencode/test/provider/transform.test.ts
@@ -0,0 +1,98 @@
+import { describe, expect, test } from "bun:test"
+import { ProviderTransform } from "../../src/provider/transform"
+
+const OUTPUT_TOKEN_MAX = 32000
+
+describe("ProviderTransform.maxOutputTokens", () => {
+  test("returns 32k when modelLimit > 32k", () => {
+    const modelLimit = 100000
+    const result = ProviderTransform.maxOutputTokens("@ai-sdk/openai", {}, modelLimit, OUTPUT_TOKEN_MAX)
+    expect(result).toBe(OUTPUT_TOKEN_MAX)
+  })
+
+  test("returns modelLimit when modelLimit < 32k", () => {
+    const modelLimit = 16000
+    const result = ProviderTransform.maxOutputTokens("@ai-sdk/openai", {}, modelLimit, OUTPUT_TOKEN_MAX)
+    expect(result).toBe(16000)
+  })
+
+  describe("azure", () => {
+    test("returns 32k when modelLimit > 32k", () => {
+      const modelLimit = 100000
+      const result = ProviderTransform.maxOutputTokens("@ai-sdk/azure", {}, modelLimit, OUTPUT_TOKEN_MAX)
+      expect(result).toBe(OUTPUT_TOKEN_MAX)
+    })
+
+    test("returns modelLimit when modelLimit < 32k", () => {
+      const modelLimit = 16000
+      const result = ProviderTransform.maxOutputTokens("@ai-sdk/azure", {}, modelLimit, OUTPUT_TOKEN_MAX)
+      expect(result).toBe(16000)
+    })
+  })
+
+  describe("bedrock", () => {
+    test("returns 32k when modelLimit > 32k", () => {
+      const modelLimit = 100000
+      const result = ProviderTransform.maxOutputTokens("@ai-sdk/amazon-bedrock", {}, modelLimit, OUTPUT_TOKEN_MAX)
+      expect(result).toBe(OUTPUT_TOKEN_MAX)
+    })
+
+    test("returns modelLimit when modelLimit < 32k", () => {
+      const modelLimit = 16000
+      const result = ProviderTransform.maxOutputTokens("@ai-sdk/amazon-bedrock", {}, modelLimit, OUTPUT_TOKEN_MAX)
+      expect(result).toBe(16000)
+    })
+  })
+
+  describe("anthropic without thinking options", () => {
+    test("returns 32k when modelLimit > 32k", () => {
+      const modelLimit = 100000
+      const result = ProviderTransform.maxOutputTokens("@ai-sdk/anthropic", {}, modelLimit, OUTPUT_TOKEN_MAX)
+      expect(result).toBe(OUTPUT_TOKEN_MAX)
+    })
+
+    test("returns modelLimit when modelLimit < 32k", () => {
+      const modelLimit = 16000
+      const result = ProviderTransform.maxOutputTokens("@ai-sdk/anthropic", {}, modelLimit, OUTPUT_TOKEN_MAX)
+      expect(result).toBe(16000)
+    })
+  })
+
+  describe("anthropic with thinking options", () => {
+    test("returns 32k when budgetTokens + 32k <= modelLimit", () => {
+      const modelLimit = 100000
+      const options = {
+        thinking: {
+          type: "enabled",
+          budgetTokens: 10000,
+        },
+      }
+      const result = ProviderTransform.maxOutputTokens("@ai-sdk/anthropic", options, modelLimit, OUTPUT_TOKEN_MAX)
+      expect(result).toBe(OUTPUT_TOKEN_MAX)
+    })
+
+    test("returns modelLimit - budgetTokens when budgetTokens + 32k > modelLimit", () => {
+      const modelLimit = 50000
+      const options = {
+        thinking: {
+          type: "enabled",
+          budgetTokens: 30000,
+        },
+      }
+      const result = ProviderTransform.maxOutputTokens("@ai-sdk/anthropic", options, modelLimit, OUTPUT_TOKEN_MAX)
+      expect(result).toBe(20000)
+    })
+
+    test("returns 32k when thinking type is not enabled", () => {
+      const modelLimit = 100000
+      const options = {
+        thinking: {
+          type: "disabled",
+          budgetTokens: 10000,
+        },
+      }
+      const result = ProviderTransform.maxOutputTokens("@ai-sdk/anthropic", options, modelLimit, OUTPUT_TOKEN_MAX)
+      expect(result).toBe(OUTPUT_TOKEN_MAX)
+    })
+  })
+})