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

This commit is contained in:
Aiden Cline
2026-10-24 23:23:21 -05:00
committed by GitHub
Unverified
parent b7321f7fcc
commit b367fdd168
2 changed files with 30 additions and 2 deletions
+14 -2
View File
@@ -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: {
@@ -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(