Merge pull request #3914 from code-yeongyu/fix/max-output-tokens

fix: prevent non-positive maxOutputTokens from leaking to SDK

src/shared/model-settings-compatibility.test.ts

@@ -615,6 +615,18 @@ describe("resolveCompatibleModelSettings", () => {
     expect(result.changes).toEqual([])
   })
 
+  test("#given desired.maxTokens is 0 #then maxTokens is dropped", () => {
+    const result = resolveCompatibleModelSettings({
+      providerID: "openai",
+      modelID: "gpt-5.4",
+      desired: { maxTokens: 0 },
+      capabilities: { maxOutputTokens: 128_000 },
+    })
+
+    expect(result.maxTokens).toBeUndefined()
+    expect(result.changes).toEqual([])
+  })
+
   // Passthrough: undefined desired values produce no changes
   test("no-op when desired settings are empty", () => {
     const result = resolveCompatibleModelSettings({

src/shared/model-settings-compatibility.ts

@@ -175,6 +175,10 @@ export function resolveCompatibleModelSettings(
   }
 
   let maxTokens = input.desired.maxTokens
+  if (maxTokens !== undefined && maxTokens <= 0) {
+    maxTokens = undefined
+  }
+
   if (
     maxTokens !== undefined &&
     input.capabilities?.maxOutputTokens !== undefined &&