fix: omit non-positive max_tokens for custom models

Custom models used sane-default metadata with maxTokens: -1 (unlimited),
which Fireworks rejects with '400 max_tokens must be non-negative'.

OpenAI API spec does not specify -1 as special value and explicitly allows
max tokens to be omitted. Remove max tokens from the sane-default metadata.

Author: p12tic

packages/types/src/providers/openai.ts

@@ -930,8 +930,9 @@ export const openAiNativeModels = {
 	},
 } as const satisfies Record<string, ModelInfo>
 
+// `maxTokens` is intentionally omitted so that no max output token limit is sent to the
+// provider by default.
 export const openAiModelInfoSaneDefaults: ModelInfo = {
-	maxTokens: -1,
 	contextWindow: 128_000,
 	supportsImages: true,
 	supportsPromptCache: false,

src/api/providers/__tests__/base-openai-compatible-provider.spec.ts

@@ -325,6 +325,7 @@ describe("BaseOpenAiCompatibleProvider", () => {
 			expect(model.id).not.toBe("test-model")
 			// Falls back to sane default metadata so the rest of the pipeline works.
 			expect(model.info).toEqual(openAiModelInfoSaneDefaults)
+			expect(model.info).not.toHaveProperty("maxTokens")
 		})
 
 		it("sends the custom model id verbatim to the API", async () => {
@@ -343,6 +344,18 @@ describe("BaseOpenAiCompatibleProvider", () => {
 			)
 		})
 
+		it("omits max_tokens for a custom model", async () => {
+			mockCreate.mockImplementationOnce(() => asyncStreamFrom([]))
+
+			const customHandler = new TestOpenAiCompatibleProvider("test-api-key", {
+				apiModelId: "some/custom-model-not-in-list",
+			})
+
+			await collectStream(customHandler.createMessage("system prompt", []))
+
+			expect(mockCreate.mock.calls[0][0].max_tokens).toBeUndefined()
+		})
+
 		it("completePrompt sends the custom model id verbatim to the API", async () => {
 			mockCreate.mockResolvedValueOnce({ choices: [{ message: { content: "ok" } }] })
 

src/api/providers/__tests__/fireworks.spec.ts

@@ -98,6 +98,29 @@ describe("FireworksHandler", () => {
 		expect(model.info).toEqual(expect.objectContaining(fireworksModels[testModelId]))
 	})
 
+	it("should omit max_tokens for a custom model", async () => {
+		const handlerWithCustomModel = new FireworksHandler({
+			apiModelId: "accounts/fireworks/models/deepseek-v4p1-flash",
+			fireworksApiKey: "test-fireworks-api-key",
+		})
+
+		await collectStream(handlerWithCustomModel.createMessage("system prompt", []))
+
+		expect(mockCreate.mock.calls[0][0].model).toBe("accounts/fireworks/models/deepseek-v4p1-flash")
+		expect(mockCreate.mock.calls[0][0].max_tokens).toBeUndefined()
+	})
+
+	it("should still send a positive max_tokens for a known model", async () => {
+		const handlerWithModel = new FireworksHandler({
+			apiModelId: "accounts/fireworks/models/kimi-k2-instruct",
+			fireworksApiKey: "test-fireworks-api-key",
+		})
+
+		await collectStream(handlerWithModel.createMessage("system prompt", []))
+
+		expect(mockCreate.mock.calls[0][0].max_tokens).toBe(16384)
+	})
+
 	it.each([
 		{
 			modelId: "accounts/fireworks/models/glm-5p1" as const,

src/api/providers/__tests__/lmstudio.spec.ts

@@ -239,7 +239,7 @@ describe("LmStudioHandler", () => {
 			const modelInfo = handler.getModel()
 			expect(modelInfo.id).toBe(mockOptions.lmStudioModelId)
 			expect(modelInfo.info).toBeDefined()
-			expect(modelInfo.info.maxTokens).toBe(-1)
+			expect(modelInfo.info.maxTokens).toBeUndefined()
 			expect(modelInfo.info.contextWindow).toBe(128_000)
 		})
 	})