diff --git a/packages/ai/src/protocols/xai-responses.ts b/packages/ai/src/protocols/xai-responses.ts index 94eb7e670d6..0680f87d954 100644 --- a/packages/ai/src/protocols/xai-responses.ts +++ b/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: { diff --git a/packages/ai/test/provider/xai-responses.test.ts b/packages/ai/test/provider/xai-responses.test.ts index 69909cf0a3b..161cddb69e1 100644 --- a/packages/ai/test/provider/xai-responses.test.ts +++ b/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(