fix(ai): send max_tokens to DeepSeek APIs

Auto-detect DeepSeek in generated and runtime compatibility metadata so built-in and custom models use the supported token limit field.

packages/ai/CHANGELOG.md

@@ -11,6 +11,7 @@
 
 - Fixed upstream request buffer limit failures to trigger automatic assistant retries.
 - Fixed OpenAI Responses function and custom tool calls to preserve namespaces during streaming, proxying, and replay ([#7709](https://github.com/earendil-works/pi/issues/7709)).
+- Fixed built-in and custom DeepSeek API models to send output limits through the supported `max_tokens` field.
 
 ## [0.84.1] - 2026-08-07
 

packages/ai/scripts/generate-models.ts

@@ -643,11 +643,18 @@ function detectOpenAICompletionsCompat(model: Model<"openai-completions">): Open
 		isCloudflareAiGateway ||
 		isAntLing;
 
+	const isDeepSeek = provider === "deepseek" || baseUrl.includes("deepseek.com");
 	const useMaxTokens =
-		baseUrl.includes("chutes.ai") || isMoonshot || isCloudflareAiGateway || isTogether || isNvidia || isAntLing || isZai;
+		baseUrl.includes("chutes.ai") ||
+		isDeepSeek ||
+		isMoonshot ||
+		isCloudflareAiGateway ||
+		isTogether ||
+		isNvidia ||
+		isAntLing ||
+		isZai;
 
 	const isGrok = provider === "xai" || baseUrl.includes("api.x.ai");
-	const isDeepSeek = provider === "deepseek" || baseUrl.includes("deepseek.com");
 	const isOpenRouterDeveloperRoleModel =
 		isOpenRouter && (model.id.startsWith("anthropic/") || model.id.startsWith("openai/"));
 	const cacheControlFormat =

packages/ai/src/api/openai-completions.ts

@@ -1475,8 +1475,10 @@ function detectCompat(model: Model<"openai-completions">): ResolvedOpenAIComplet
 		isCloudflareAiGateway ||
 		isAntLing;
 
+	const isDeepSeek = provider === "deepseek" || baseUrl.includes("deepseek.com");
 	const useMaxTokens =
 		baseUrl.includes("chutes.ai") ||
+		isDeepSeek ||
 		isMoonshot ||
 		isCloudflareAiGateway ||
 		isTogether ||
@@ -1485,7 +1487,6 @@ function detectCompat(model: Model<"openai-completions">): ResolvedOpenAIComplet
 		isZai;
 
 	const isGrok = provider === "xai" || baseUrl.includes("api.x.ai");
-	const isDeepSeek = provider === "deepseek" || baseUrl.includes("deepseek.com");
 	const isOpenRouterDeveloperRoleModel =
 		isOpenRouter && (model.id.startsWith("anthropic/") || model.id.startsWith("openai/"));
 	const cacheControlFormat = provider === "openrouter" && model.id.startsWith("anthropic/") ? "anthropic" : undefined;

packages/ai/test/openai-completions-tool-choice.test.ts

@@ -1429,6 +1429,47 @@ describe("openai-completions tool_choice", () => {
 		}
 	});
 
+	it("sends max_tokens for built-in and custom DeepSeek API models", async () => {
+		const customModel = {
+			...localOpenAICompletionsModel,
+			id: "custom-deepseek-model",
+			name: "Custom DeepSeek Model",
+			provider: "custom-deepseek",
+			baseUrl: "https://api.deepseek.com",
+		} satisfies Model<"openai-completions">;
+		const nativeModels = [
+			getModel("deepseek", "deepseek-v4-flash")!,
+			getModel("deepseek", "deepseek-v4-pro")!,
+		] as const;
+		const cases = [...nativeModels, customModel] as const;
+
+		for (const model of nativeModels) {
+			expect(model.compat?.maxTokensField).toBe("max_tokens");
+		}
+
+		for (const model of cases) {
+			let payload: unknown;
+
+			await streamSimple(
+				model,
+				{
+					messages: [{ role: "user", content: "Hi", timestamp: Date.now() }],
+				},
+				{
+					apiKey: "test",
+					maxTokens: 123,
+					onPayload: (params: unknown) => {
+						payload = params;
+					},
+				},
+			).result();
+
+			const params = (payload ?? mockState.lastParams) as { max_tokens?: number; max_completion_tokens?: number };
+			expect(params.max_tokens).toBe(123);
+			expect(params.max_completion_tokens).toBeUndefined();
+		}
+	});
+
 	it("sends max_tokens for Z.AI completions models", async () => {
 		const cases = [getModel("zai", "glm-5-turbo")!, getModel("zai", "glm-5.2")!] as const;