feat(friendli): fetch models from /v1/models

packages/types/src/__tests__/provider-identifiers.test.ts

@@ -114,6 +114,7 @@ describe("provider identifiers", () => {
 			providerIdentifiers.nanogpt,
 			providerIdentifiers.ioIntelligence,
 			providerIdentifiers.kimiCode,
+			providerIdentifiers.friendli,
 		])
 		expect(localProviders).toEqual([providerIdentifiers.ollama, providerIdentifiers.lmstudio])
 		expect(internalProviders).toEqual([providerIdentifiers.vscodeLm, providerIdentifiers.githubCopilot])

packages/types/src/provider-settings.ts

@@ -102,6 +102,7 @@ export const dynamicProviders = [
 	providerIdentifiers.nanogpt,
 	providerIdentifiers.ioIntelligence,
 	providerIdentifiers.kimiCode,
+	providerIdentifiers.friendli,
 ] as const
 
 export type DynamicProvider = (typeof dynamicProviders)[number]

src/api/providers/fetchers/__tests__/friendli.spec.ts

@@ -0,0 +1,168 @@
+// npx vitest run api/providers/fetchers/__tests__/friendli.spec.ts
+
+import axios from "axios"
+
+import { openAiModelInfoSaneDefaults } from "@roo-code/types"
+
+import { getFriendliModels } from "../friendli"
+
+vi.mock("axios")
+
+describe("getFriendliModels", () => {
+	it("maps chat models from /v1/models", async () => {
+		vi.mocked(axios.get).mockResolvedValueOnce({
+			data: {
+				data: [
+					{
+						id: "zai-org/GLM-5.3",
+						mode: "chat",
+						description: "GLM",
+						context_length: 1_048_576,
+						max_completion_tokens: 131_072,
+						pricing: { input: "0.00000126", output: "0.00000396", input_cache_read: "0.000000234" },
+						input_modalities: ["text", "image"],
+						reasoning_options: [
+							{ type: "effort", values: ["low", "high", "max", "ultracode"] },
+							{ type: "budget_tokens", min: -1, max: 1_048_576 },
+						],
+					},
+					{
+						id: "google/gemma-4-31B-it",
+						mode: "chat",
+						pricing: { input: "0.00000014", output: "0.0000004" },
+					},
+					{ id: "zai-org/GLM-5.2", mode: "chat", input_modalities: ["text"] },
+					{ id: "some/embedding", mode: "embedding" },
+				],
+			},
+		})
+
+		const signal = new AbortController().signal
+		const models = await getFriendliModels({ signal })
+
+		expect(axios.get).toHaveBeenCalledWith("https://api.friendli.ai/serverless/v1/models", { signal })
+		expect(Object.keys(models)).toEqual(["zai-org/GLM-5.3", "google/gemma-4-31B-it", "zai-org/GLM-5.2"])
+		expect(models["zai-org/GLM-5.3"]).toEqual({
+			maxTokens: 131_072,
+			contextWindow: 1_048_576,
+			supportsImages: true,
+			supportsPromptCache: true,
+			supportsMaxTokens: true,
+			inputPrice: expect.closeTo(1.26),
+			outputPrice: expect.closeTo(3.96),
+			cacheReadsPrice: expect.closeTo(0.234),
+			supportsReasoningEffort: ["low", "high", "max"],
+			reasoningEffort: "high",
+			description: "GLM",
+		})
+		expect(models["google/gemma-4-31B-it"].supportsReasoningEffort).toBeUndefined()
+		expect(models["google/gemma-4-31B-it"].supportsPromptCache).toBe(false)
+		// A model without context_length gets the sane default window.
+		expect(models["google/gemma-4-31B-it"].contextWindow).toBe(openAiModelInfoSaneDefaults.contextWindow)
+		// Missing input_modalities and a text-only list both mean no image input.
+		expect(models["google/gemma-4-31B-it"].supportsImages).toBe(false)
+		expect(models["zai-org/GLM-5.2"].supportsImages).toBe(false)
+	})
+
+	it("accepts null reasoning_options and keeps the catalog", async () => {
+		vi.mocked(axios.get).mockResolvedValueOnce({
+			data: {
+				data: [
+					{ id: "vendor/fetched-only", reasoning_options: null },
+					{ id: "zai-org/GLM-5.1", reasoning_options: [{ type: "toggle" }] },
+				],
+			},
+		})
+
+		const models = await getFriendliModels()
+
+		expect(Object.keys(models)).toEqual(["vendor/fetched-only", "zai-org/GLM-5.1"])
+		expect(models["vendor/fetched-only"].supportsReasoningEffort).toBeUndefined()
+		expect(models["vendor/fetched-only"].supportsReasoningBinary).toBeUndefined()
+	})
+
+	it("maps toggle-only reasoning to supportsReasoningBinary without inventing an effort", async () => {
+		vi.mocked(axios.get).mockResolvedValueOnce({
+			data: {
+				data: [
+					{
+						id: "zai-org/GLM-5.1",
+						reasoning_options: [{ type: "toggle" }, { type: "budget_tokens", min: -1, max: 202_752 }],
+					},
+					{
+						id: "zai-org/GLM-5.2",
+						reasoning_options: [{ type: "toggle" }, { type: "effort", values: ["high", "max"] }],
+					},
+				],
+			},
+		})
+
+		const models = await getFriendliModels()
+
+		expect(models["zai-org/GLM-5.1"].supportsReasoningBinary).toBe(true)
+		expect(models["zai-org/GLM-5.1"].supportsReasoningEffort).toBeUndefined()
+		expect(models["zai-org/GLM-5.1"].reasoningEffort).toBeUndefined()
+		// A toggle next to effort levels keeps the effort selector.
+		expect(models["zai-org/GLM-5.2"].supportsReasoningBinary).toBeUndefined()
+		expect(models["zai-org/GLM-5.2"].supportsReasoningEffort).toEqual(["high", "max"])
+	})
+
+	it("defaults to the first supported effort when high is not offered", async () => {
+		vi.mocked(axios.get).mockResolvedValueOnce({
+			data: {
+				data: [
+					{
+						id: "vendor/no-high",
+						reasoning_options: [{ type: "effort", values: ["ultracode", "low", "medium", "max"] }],
+					},
+				],
+			},
+		})
+
+		const models = await getFriendliModels()
+
+		expect(models["vendor/no-high"].supportsReasoningEffort).toEqual(["low", "medium", "max"])
+		expect(models["vendor/no-high"].reasoningEffort).toBe("low")
+	})
+
+	it("keeps reserved upstream IDs as own keys", async () => {
+		vi.mocked(axios.get).mockResolvedValueOnce({
+			data: {
+				data: [
+					{ id: "__proto__", context_length: 1000 },
+					{ id: "toString", context_length: 2000 },
+				],
+			},
+		})
+
+		const models = await getFriendliModels()
+
+		expect(Object.getPrototypeOf(models)).toBeNull()
+		expect(Object.keys(models)).toEqual(["__proto__", "toString"])
+		expect(models["__proto__"].contextWindow).toBe(1000)
+		expect(models["toString"].contextWindow).toBe(2000)
+		expect(models["valueOf"]).toBeUndefined()
+	})
+
+	it("rethrows cancellation instead of returning an empty record", async () => {
+		const controller = new AbortController()
+		controller.abort()
+		vi.mocked(axios.get).mockRejectedValueOnce(new Error("canceled"))
+
+		await expect(getFriendliModels({ signal: controller.signal })).rejects.toMatchObject({ name: "AbortError" })
+	})
+
+	it("returns an empty record on an invalid response", async () => {
+		vi.spyOn(console, "error").mockImplementation(() => {})
+		vi.mocked(axios.get).mockResolvedValueOnce({ data: { invalid: true } })
+
+		expect(await getFriendliModels()).toEqual({})
+	})
+
+	it("returns an empty record on request failure", async () => {
+		vi.spyOn(console, "error").mockImplementation(() => {})
+		vi.mocked(axios.get).mockRejectedValueOnce(new Error("Network error"))
+
+		expect(await getFriendliModels()).toEqual({})
+	})
+})

src/api/providers/fetchers/friendli.ts

@@ -0,0 +1,86 @@
+import axios from "axios"
+import { z } from "zod"
+
+import {
+	type ModelInfo,
+	type ReasoningEffortExtended,
+	openAiModelInfoSaneDefaults,
+	reasoningEffortsExtended,
+} from "@roo-code/types"
+
+import { parseApiPrice } from "../../../shared/cost"
+import { throwIfAborted } from "../utils/abort-signal"
+
+const friendliModelSchema = z.object({
+	id: z.string(),
+	mode: z.string().optional(),
+	description: z.string().optional(),
+	context_length: z.number().optional(),
+	max_completion_tokens: z.number().optional(),
+	pricing: z
+		.object({
+			input: z.string().optional(),
+			output: z.string().optional(),
+			input_cache_read: z.string().optional(),
+		})
+		.optional(),
+	input_modalities: z.array(z.string()).optional(),
+	reasoning_options: z.array(z.object({ type: z.string(), values: z.array(z.string()).optional() })).nullish(),
+})
+
+type FriendliModel = z.infer<typeof friendliModelSchema>
+
+const friendliModelsResponseSchema = z.object({ data: z.array(friendliModelSchema) })
+
+const parseFriendliModel = (model: FriendliModel): ModelInfo => {
+	const efforts = model.reasoning_options
+		?.find((option) => option.type === "effort")
+		?.values?.filter((v): v is ReasoningEffortExtended =>
+			(reasoningEffortsExtended as readonly string[]).includes(v),
+		)
+	// A toggle without effort levels (e.g. GLM-5.1, Gemma) is an on/off switch, not an effort scale.
+	const toggleOnly = !efforts?.length && !!model.reasoning_options?.some((option) => option.type === "toggle")
+
+	return {
+		maxTokens: model.max_completion_tokens,
+		contextWindow: model.context_length ?? openAiModelInfoSaneDefaults.contextWindow,
+		supportsImages: model.input_modalities?.includes("image") ?? false,
+		supportsPromptCache: model.pricing?.input_cache_read !== undefined,
+		supportsMaxTokens: true,
+		inputPrice: parseApiPrice(model.pricing?.input),
+		outputPrice: parseApiPrice(model.pricing?.output),
+		cacheReadsPrice: parseApiPrice(model.pricing?.input_cache_read),
+		...(efforts?.length && {
+			supportsReasoningEffort: efforts,
+			reasoningEffort: efforts.includes("high") ? "high" : efforts[0],
+		}),
+		...(toggleOnly && { supportsReasoningBinary: true }),
+		description: model.description,
+	}
+}
+
+/** Fetches the public (no auth) Friendli model catalog. */
+export async function getFriendliModels(opts?: { signal?: AbortSignal }): Promise<Record<string, ModelInfo>> {
+	// Null prototype: upstream IDs such as "__proto__" or "toString" must stay plain keys.
+	const models: Record<string, ModelInfo> = Object.create(null)
+
+	try {
+		const response = await axios.get("https://api.friendli.ai/serverless/v1/models", { signal: opts?.signal })
+		const result = friendliModelsResponseSchema.safeParse(response.data)
+
+		if (!result.success) {
+			console.error(`Friendli models response is invalid: ${JSON.stringify(result.error.format())}`)
+			return models
+		}
+
+		for (const model of result.data.data) {
+			if (model.mode && model.mode !== "chat") continue
+			models[model.id] = parseFriendliModel(model)
+		}
+	} catch (error) {
+		throwIfAborted(opts?.signal)
+		console.error(`Error fetching Friendli models: ${error instanceof Error ? error.message : String(error)}`)
+	}
+
+	return models
+}