fix(catalog): omit Responses penalties on all first-party xAI models

xAI's /v1/responses rejects presence/frequency penalties for every Grok
model, not only reasoners. Gate supportsPenaltyAndStopParams on isXaiHost
so xai/grok-2 no longer serializes presence_penalty.

docs/provider-endpoint-constraints.md

@@ -242,6 +242,7 @@ Both the paid API-key provider (`xai` / `XAI_API_KEY`) and SuperGrok OAuth
 
 - omit `reasoning.effort` unless the model is on the Grok effort-capable allowlist
 - omit `reasoning.summary` (the host rejects it; do not fall back to `"auto"`)
+- omit presence/frequency penalties (`/v1/responses` rejects them for every Grok model)
 - include `reasoning.encrypted_content` on the request
 - replay encrypted reasoning items on later turns
 

packages/ai/CHANGELOG.md

@@ -6,6 +6,7 @@
 
 - Stopped treating `XAI_API_KEY` as SuperGrok (`xai-oauth`) sign-in for availability, so paid-key-only setups default to `xai/grok-4.5` instead of the zero-cost SuperGrok catalog path.
 - Omitted unsupported `reasoning.summary` on paid xAI Responses requests (`xai/grok-4.5`), matching SuperGrok, so a thinking level no longer serializes `summary: "auto"`.
+- Omitted presence/frequency penalties on all first-party xAI Responses models, including non-reasoning ids such as `xai/grok-2`.
 
 ## [17.3.4] - 2026-08-14
 

packages/ai/test/xai-oauth-effort-strip.test.ts

@@ -121,6 +121,16 @@ describe("xAI OAuth Responses reasoning payload (regression)", () => {
 		expect(params.temperature).toBe(0.2);
 	});
 
+	test("paid xai/grok-2 omits presence_penalty on non-reasoning Responses models", () => {
+		const grok2 = getBundledModel<"openai-responses">("xai", "grok-2");
+		if (!grok2) throw new Error("xai/grok-2 must be in bundled models.json");
+
+		const { params } = buildParams(grok2, singleUserContext, { presencePenalty: 0.4, temperature: 0.2 }, undefined);
+
+		expect(params).not.toHaveProperty("presence_penalty");
+		expect(params.temperature).toBe(0.2);
+	});
+
 	test("paid xai/grok-4.5 clamps minimal reasoning effort to low", () => {
 		const grok45 = getBundledModel<"openai-responses">("xai", "grok-4.5");
 		if (!grok45) throw new Error("xai/grok-4.5 must be in bundled models.json");

packages/catalog/CHANGELOG.md

@@ -19,6 +19,7 @@
 - Marked first-party xAI Responses hosts (`xai` and `xai-oauth`) as not supporting `reasoning.summary`, so paid `xai/grok-4.5` effort requests omit the unsupported field instead of sending `summary: "auto"`.
 - Removed unsupported `xhigh` (and `max`) thinking tiers from first-party Grok Responses catalog rows; leftover `xhigh`/`max` requests clamp to `high`.
 - Stopped baking `reasoningEffortMap` on first-party xAI catalog rows that omit `reasoning.effort` (`omitReasoningEffort: true`).
+- Suppressed presence/frequency penalties on every first-party xAI Responses model, including non-reasoning ids such as `grok-2`; xAI's `/v1/responses` marks those fields unsupported.
 
 ## [17.3.4] - 2026-08-14
 

packages/catalog/src/compat/openai.ts

@@ -722,9 +722,9 @@ export function buildOpenAIResponsesCompat(spec: OpenAIResponsesSpecLike): Resol
 		// OpenAI proprietary reasoning models (o-series, gpt-5+) reject explicit
 		// temperature/top_p/… with a 400 on every serving host (#5606).
 		supportsSamplingParams: !isOpenAISamplingRestrictedModelId(id),
-		// xAI reasoning models 400 on presence/frequency penalties and stop
-		// (https://docs.x.ai/developers/model-capabilities/text/reasoning).
-		supportsPenaltyAndStopParams: !(isXaiHost && reasoningCapable),
+		// xAI `/v1/responses` rejects presence/frequency penalties for every
+		// model, not only reasoners (https://docs.x.ai/developers/rest-api-reference/inference/chat).
+		supportsPenaltyAndStopParams: !isXaiHost,
 		thinkingFormat,
 		reasoningDisableMode: resolveReasoningDisableMode(thinkingFormat),
 		omitReasoningEffort: false,

packages/catalog/src/types.ts

@@ -365,8 +365,9 @@ export interface OpenAICompat {
 	supportsSamplingParams?: boolean;
 	/**
 	 * Whether presence/frequency penalties and stop sequences may be sent.
-	 * xAI reasoning models reject `presencePenalty`, `frequencyPenalty`, and
-	 * `stop` with a 400. When unset, auto-detected. Default: true.
+	 * First-party xAI `/v1/responses` rejects penalty fields for every model.
+	 * xAI reasoning models also reject them (and `stop`) on chat completions.
+	 * When unset, auto-detected. Default: true.
 	 */
 	supportsPenaltyAndStopParams?: boolean;
 	/** Always send a max-token field when the caller did not provide one. Default: auto-detected (Kimi-family models derive TPM limits from max_tokens). */

packages/catalog/test/build.test.ts

@@ -272,12 +272,14 @@ describe("xAI Responses reasoning-effort suppression", () => {
 		expect(oauth.supportsReasoningSummary).toBe(false);
 	});
 
-	it("keeps penalty and stop params on non-reasoning paid xAI models", () => {
-		const compat = buildOpenAIResponsesCompat({
-			...grokResponsesSpec("grok-4-fast-non-reasoning", "xai"),
+	it("suppresses penalty params on every first-party xAI Responses model", () => {
+		const reasoning = buildOpenAIResponsesCompat(grokResponsesSpec("grok-4.5", "xai"));
+		const nonReasoning = buildOpenAIResponsesCompat({
+			...grokResponsesSpec("grok-2", "xai"),
 			reasoning: false,
 		});
-		expect(compat.supportsPenaltyAndStopParams).toBe(true);
+		expect(reasoning.supportsPenaltyAndStopParams).toBe(false);
+		expect(nonReasoning.supportsPenaltyAndStopParams).toBe(false);
 	});
 
 	it("omits effort for paid xai models off the Grok allowlist", () => {