feat: support custom model info overrides

packages/types/src/__tests__/custom-model-info.test.ts

@@ -0,0 +1,75 @@
+import { applyCustomModelInfo, customModelInfoSchema, type ModelInfo } from "../model.js"
+import { providerIdentifiers, providerSettingsSchemaDiscriminated } from "../index.js"
+
+describe("custom model info", () => {
+	it("overlays only supported metadata and preserves provider-owned fields", () => {
+		const model: ModelInfo = {
+			maxTokens: 4096,
+			contextWindow: 8192,
+			supportsImages: false,
+			supportsPromptCache: false,
+			inputPrice: 0.1,
+			outputPrice: 0.2,
+			description: "Provider metadata",
+		}
+
+		expect(
+			applyCustomModelInfo(model, {
+				customModelInfo: {
+					contextWindow: 128_000,
+					maxTokens: 16_384,
+					supportsImages: true,
+					supportsPromptCache: true,
+				},
+			}),
+		).toEqual({
+			...model,
+			contextWindow: 128_000,
+			maxTokens: 16_384,
+			supportsImages: true,
+			supportsPromptCache: true,
+		})
+	})
+
+	it("does not synthesize model info without a valid context window", () => {
+		expect(
+			applyCustomModelInfo(undefined, {
+				customModelInfo: {
+					contextWindow: 0,
+					maxTokens: -1,
+					supportsImages: true,
+				},
+			}),
+		).toBeUndefined()
+	})
+
+	it("synthesizes safe defaults when only a context window is supplied", () => {
+		expect(
+			applyCustomModelInfo(undefined, {
+				customModelInfo: { contextWindow: 64_000, supportsImages: true },
+			}),
+		).toEqual({
+			maxTokens: undefined,
+			contextWindow: 64_000,
+			supportsImages: true,
+			supportsPromptCache: false,
+		})
+	})
+
+	it("rejects unsupported pricing fields in the persisted override schema", () => {
+		expect(customModelInfoSchema.safeParse({ contextWindow: 64_000, inputPrice: 1 }).success).toBe(false)
+		expect(
+			providerSettingsSchemaDiscriminated.safeParse({
+				apiProvider: providerIdentifiers.openrouter,
+				customModelInfo: { contextWindow: 64_000, outputPrice: 1 },
+			}).success,
+		).toBe(false)
+	})
+
+	it("rejects unsafe integer overrides", () => {
+		const unsafeInteger = Number.MAX_SAFE_INTEGER + 1
+
+		expect(customModelInfoSchema.safeParse({ contextWindow: unsafeInteger }).success).toBe(false)
+		expect(customModelInfoSchema.safeParse({ maxTokens: unsafeInteger }).success).toBe(false)
+	})
+})

packages/types/src/model.ts

@@ -183,6 +183,86 @@ export const modelInfoSchema = z.object({
 
 export type ModelInfo = z.infer<typeof modelInfoSchema>
 
+/**
+ * User-supplied metadata for a model whose discovered metadata is incomplete
+ * or unavailable. This is intentionally narrower than ModelInfo: prices and
+ * other accounting fields must remain provider-owned.
+ */
+const positiveSafeIntegerSchema = z
+	.number()
+	.int()
+	.positive()
+	.refine(Number.isSafeInteger, { message: "Expected a safe integer" })
+
+export const customModelInfoSchema = z
+	.object({
+		maxTokens: positiveSafeIntegerSchema.optional(),
+		contextWindow: positiveSafeIntegerSchema.optional(),
+		supportsImages: z.boolean().optional(),
+		supportsPromptCache: z.boolean().optional(),
+	})
+	.strict()
+
+export type CustomModelInfo = z.infer<typeof customModelInfoSchema>
+
+export type CustomModelInfoSettings = {
+	customModelInfo?: Partial<CustomModelInfo> | null
+}
+
+const isPositiveInteger = (value: unknown): value is number =>
+	typeof value === "number" && Number.isSafeInteger(value) && value > 0
+
+/**
+ * Applies the user metadata overlay without allowing invalid values to enter
+ * model arithmetic or cost/capability fields outside the supported override.
+ * When no discovered info exists, a context-window override is required to
+ * synthesize a usable ModelInfo.
+ */
+export const applyCustomModelInfo = (
+	info: ModelInfo | undefined,
+	settings: CustomModelInfoSettings | undefined,
+): ModelInfo | undefined => {
+	const override = settings?.customModelInfo
+
+	if (!override) {
+		return info
+	}
+
+	const validOverride: CustomModelInfo = {}
+
+	if (isPositiveInteger(override.contextWindow)) {
+		validOverride.contextWindow = override.contextWindow
+	}
+
+	if (isPositiveInteger(override.maxTokens)) {
+		validOverride.maxTokens = override.maxTokens
+	}
+
+	if (typeof override.supportsImages === "boolean") {
+		validOverride.supportsImages = override.supportsImages
+	}
+
+	if (typeof override.supportsPromptCache === "boolean") {
+		validOverride.supportsPromptCache = override.supportsPromptCache
+	}
+
+	if (info) {
+		return Object.keys(validOverride).length > 0 ? { ...info, ...validOverride } : info
+	}
+
+	if (!validOverride.contextWindow) {
+		return undefined
+	}
+
+	return {
+		maxTokens: undefined,
+		contextWindow: validOverride.contextWindow,
+		supportsImages: false,
+		supportsPromptCache: false,
+		...validOverride,
+	}
+}
+
 export type ModelRecord = Record<string, ModelInfo>
 
 export type RouterModels = Record<DynamicProvider | LocalProvider, ModelRecord>

packages/types/src/provider-settings.ts

@@ -2,6 +2,7 @@ import { z } from "zod"
 
 import {
 	modelInfoSchema,
+	customModelInfoSchema,
 	openAiCodexServiceTierSchema,
 	reasoningEffortSettingSchema,
 	verbosityLevelsSchema,
@@ -72,6 +73,28 @@ export type DynamicProvider = (typeof dynamicProviders)[number]
 export const isDynamicProvider = (key: string): key is DynamicProvider =>
 	dynamicProviders.includes(key as DynamicProvider)
 
+/**
+ * Providers that accept user-supplied `customModelInfo` overrides in the
+ * settings UI. This is a strict subset of `dynamicProviders` — only providers
+ * whose model metadata can be manually adjusted by the user.
+ *
+ * NOTE: Of these, only vercel-ai-gateway and zoo-gateway apply the overlay
+ * inside `RouterProvider.resolveModelInfo()`. The others (openrouter, requesty,
+ * unbound) apply it in their own overridden `getModel()` methods.
+ */
+export const customModelInfoProviders = [
+	providerIdentifiers.openrouter,
+	providerIdentifiers.requesty,
+	providerIdentifiers.unbound,
+	providerIdentifiers.vercelAiGateway,
+	providerIdentifiers.zooGateway,
+] as const
+
+export type CustomModelInfoProvider = (typeof customModelInfoProviders)[number]
+
+export const isCustomModelInfoProvider = (key: string): key is CustomModelInfoProvider =>
+	customModelInfoProviders.includes(key as CustomModelInfoProvider)
+
 /**
  * LocalProvider
  *
@@ -185,6 +208,7 @@ const baseProviderSettingsSchema = z.object({
 	reasoningEffort: reasoningEffortSettingSchema.optional(),
 	modelMaxTokens: z.number().optional(),
 	modelMaxThinkingTokens: z.number().optional(),
+	customModelInfo: customModelInfoSchema.nullish(),
 
 	// Model verbosity.
 	verbosity: verbosityLevelsSchema.optional(),

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

@@ -77,6 +77,17 @@ describe("KenariHandler", () => {
 			expect(result.info.supportsPromptCache).toBe(false)
 		})
 
+		it("does not apply gateway-only custom metadata overrides", async () => {
+			const handler = new KenariHandler({
+				...mockOptions,
+				customModelInfo: { contextWindow: 100_000, maxTokens: 10_000 },
+			})
+			const result = await handler.fetchModel()
+
+			expect(result.info.contextWindow).toBe(1_048_576)
+			expect(result.info.maxTokens).toBe(32_768)
+		})
+
 		it("falls back to the default model id when none is configured", async () => {
 			const handler = new KenariHandler({ kenariApiKey: "test-key" })
 			const result = await handler.fetchModel()

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

@@ -18,6 +18,7 @@ import { Anthropic } from "@anthropic-ai/sdk"
 import OpenAI from "openai"
 
 import { OpenRouterHandler } from "../openrouter"
+import { getModelEndpoints } from "../fetchers/modelEndpointCache"
 import { Package } from "../../../shared/package"
 import { makeApiHandlerOptions } from "../../../test-utils/api"
 import { asyncStreamFrom, collectStream } from "../../../test-utils/stream"
@@ -102,13 +103,22 @@ vitest.mock("../fetchers/modelCache", () => ({
 	}),
 }))
 
+vitest.mock("../fetchers/modelEndpointCache", () => ({
+	getModelEndpoints: vitest.fn().mockResolvedValue({}),
+}))
+
 describe("OpenRouterHandler", () => {
 	const mockOptions = makeApiHandlerOptions({
 		openRouterApiKey: "test-key",
 		openRouterModelId: "anthropic/claude-sonnet-4",
 	})
 
-	beforeEach(() => clearAllMocks())
+	beforeEach(() => {
+		clearAllMocks()
+		// Reset getModelEndpoints to its default empty-object return so per-test
+		// overrides (e.g. specific-provider test) don't leak into subsequent tests.
+		vitest.mocked(getModelEndpoints).mockResolvedValue({})
+	})
 
 	it("initializes with correct options", () => {
 		const handler = new OpenRouterHandler(mockOptions)
@@ -140,14 +150,75 @@ describe("OpenRouterHandler", () => {
 			})
 		})
 
+		it("applies custom metadata before deriving request parameters", async () => {
+			const handler = new OpenRouterHandler({
+				...mockOptions,
+				customModelInfo: {
+					contextWindow: 100_000,
+					maxTokens: 10_000,
+					supportsImages: false,
+					supportsPromptCache: false,
+				},
+			})
+
+			const result = await handler.fetchModel()
+
+			expect(result.info.contextWindow).toBe(100_000)
+			expect(result.info.maxTokens).toBe(10_000)
+			expect(result.info.supportsImages).toBe(false)
+			expect(result.info.supportsPromptCache).toBe(false)
+			expect(result.maxTokens).toBe(10_000)
+		})
+
+		it("applies custom metadata to a discovered specific-provider endpoint", async () => {
+			vitest.mocked(getModelEndpoints).mockResolvedValue({
+				"test-provider": {
+					contextWindow: 128_000,
+					maxTokens: 16_384,
+					supportsImages: true,
+					supportsPromptCache: true,
+				},
+			})
+
+			const handler = new OpenRouterHandler({
+				...mockOptions,
+				openRouterSpecificProvider: "test-provider",
+				customModelInfo: { contextWindow: 100_000, maxTokens: 10_000 },
+			})
+
+			const result = await handler.fetchModel()
+
+			expect(result.info.contextWindow).toBe(100_000)
+			expect(result.info.maxTokens).toBe(10_000)
+		})
+
+		it("synthesizes metadata for an unlisted configured model", async () => {
+			const modelId = "provider/unlisted-model"
+			const handler = new OpenRouterHandler({
+				...mockOptions,
+				openRouterModelId: modelId,
+				customModelInfo: {
+					contextWindow: 100_000,
+					maxTokens: 10_000,
+				},
+			})
+
+			const result = await handler.fetchModel()
+
+			expect(result.id).toBe(modelId)
+			expect(result.info.contextWindow).toBe(100_000)
+			expect(result.info.maxTokens).toBe(10_000)
+			expect(result.maxTokens).toBe(10_000)
+		})
+
 		it("returns default model info when options are not provided", async () => {
 			const handler = new OpenRouterHandler({})
 			const result = await handler.fetchModel()
 			expect(result.id).toBe("anthropic/claude-sonnet-4.5")
 			expect(result.info.supportsPromptCache).toBe(true)
 		})
 
-		it("honors custom maxTokens for thinking models", async () => {
+		it("clamps maxTokens to 20% of context window for thinking models", async () => {
 			const handler = new OpenRouterHandler(
 				makeApiHandlerOptions({
 					openRouterApiKey: "test-key",

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

@@ -159,6 +159,45 @@ describe("RequestyHandler", () => {
 			})
 		})
 
+		it("applies custom metadata before deriving request parameters", async () => {
+			const handler = new RequestyHandler({
+				...mockOptions,
+				customModelInfo: {
+					contextWindow: 100_000,
+					maxTokens: 10_000,
+					supportsImages: false,
+					supportsPromptCache: false,
+				},
+			})
+
+			const result = await handler.fetchModel()
+
+			expect(result.info.contextWindow).toBe(100_000)
+			expect(result.info.maxTokens).toBe(10_000)
+			expect(result.info.supportsImages).toBe(false)
+			expect(result.info.supportsPromptCache).toBe(false)
+			expect(result.maxTokens).toBe(10_000)
+		})
+
+		it("synthesizes metadata for an unlisted configured model", async () => {
+			const modelId = "provider/unlisted-model"
+			const handler = new RequestyHandler({
+				...mockOptions,
+				requestyModelId: modelId,
+				customModelInfo: {
+					contextWindow: 100_000,
+					maxTokens: 10_000,
+				},
+			})
+
+			const result = await handler.fetchModel()
+
+			expect(result.id).toBe(modelId)
+			expect(result.info.contextWindow).toBe(100_000)
+			expect(result.info.maxTokens).toBe(10_000)
+			expect(result.maxTokens).toBe(10_000)
+		})
+
 		it("returns default model info when options are not provided", async () => {
 			const handler = new RequestyHandler({})
 			const result = await handler.fetchModel()