fix(ai): send top_k and drop penalties for xAI Responses (#58868)

Author: rekram1-node

packages/ai/src/protocols/xai-responses.ts

@@ -31,6 +31,11 @@ const XAIResponsesHostedToolItem = Schema.Union([
   ),
 ])
 
+const Body = Schema.Struct({
+  ...OpenResponses.OpenResponsesBody.fields,
+  top_k: Schema.optional(Schema.Number),
+})
+
 const adapter = {
   id: ADAPTER,
   name: NAME,
@@ -53,7 +58,14 @@ const fromRequest = Effect.fn("XAIResponses.fromRequest")(function* (request: LL
       message:
         "xAI requires explicit compaction through LLMClient.compact; automatic context management is not supported",
     })
-  return yield* OpenResponses.fromRequestWithAdapter(request, adapter)
+  const body = yield* OpenResponses.fromRequestWithAdapter(request, adapter)
+  // xAI Responses does not support penalties (they are ignored), but does support top_k.
+  return {
+    ...body,
+    presence_penalty: undefined,
+    frequency_penalty: undefined,
+    top_k: request.generation?.topK,
+  }
 })
 
 const HOSTED_TOOLS = {
@@ -83,7 +95,7 @@ const step = (state: OpenResponses.ParserState, input: OpenResponses.Event) => {
 export const protocol = Protocol.make({
   id: ADAPTER,
   body: {
-    schema: OpenResponses.OpenResponsesBody,
+    schema: Body,
     from: fromRequest,
   },
   stream: {

packages/ai/test/provider/xai-responses.test.ts

@@ -73,6 +73,22 @@ describe("xAI Responses route", () => {
     }),
   )
 
+  it.effect("sends top_k and omits unsupported penalties", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          prompt: "Hello",
+          generation: { topK: 20, presencePenalty: 0.5, frequencyPenalty: 0.5 },
+        }),
+      )
+
+      expect(prepared.body.top_k).toBe(20)
+      expect(prepared.body.presence_penalty).toBeUndefined()
+      expect(prepared.body.frequency_penalty).toBeUndefined()
+    }),
+  )
+
   it.effect("parses xAI reasoning summaries", () =>
     Effect.gen(function* () {
       const response = yield* LLMClient.generate(LLM.request({ model, prompt: "Think" })).pipe(