pi-ext

import { mkdtempSync, rmSync, writeFileSync } from "node:fs";
import { tmpdir } from "node:os";
import path from "node:path";
import {
	type Api,
	type AssistantMessage,
	isContextOverflow,
	type Model,
} from "@earendil-works/pi-ai";
import { afterEach, describe, expect, it, vi } from "vitest";
import { createTestSession, type TestSession } from "../../../test/harness";
import { createGratis } from "../index.ts";
import {
	anthropicText,
	contextOverflow,
	googleText,
	liveModel,
	openAIText,
	type Recorded,
	type Reply,
	type Responder,
	rateLimited,
	startUpstream,
} from "./fake-upstream.ts";

type Registry = {
	refresh(options: {
		allowNetwork?: boolean;
		providers?: string[];
	}): Promise<unknown>;
	getAll(): Model<Api>[];
	getAvailable(): Model<Api>[];
	find(provider: string, id: string): Model<Api> | undefined;
	streamSimple(
		model: Model<Api>,
		context: unknown,
	): { result(): Promise<AssistantMessage> };
	getApiKeyForProvider(provider: string): Promise<string | undefined>;
};

const KILO_MODELS = "/kilo/api/gateway/models";
const KILO_CHAT = "/kilo/api/gateway/chat/completions";
const OPENROUTER_MODELS = "/openrouter/api/v1/models";
const OPENROUTER_CHAT = "/openrouter/api/v1/chat/completions";
const OPENROUTER_MESSAGES = "/openrouter/api/v1/messages";
const GOOGLE = "/google/v1beta/models/";

const openrouterCatalog = {
	data: [
		liveModel("qwen/qwen3.8-27b:free"),
		liveModel("anthropic/claude-trial:free"),
		liveModel("openrouter/free", {
			context_length: 200_000,
			top_provider: { context_length: null, max_completion_tokens: null },
		}),
		liveModel("qwen/qwen3.8-27b", {
			pricing: { prompt: "0.0000002", completion: "0.0000008" },
		}),
	],
};

const kiloCatalog = {
	data: [
		liveModel("minimax/minimax-m3:free"),
		liveModel("vendor/paid-model", {
			pricing: { prompt: "0.000001", completion: "0.000002" },
		}),
		liveModel("vendor/fake-free:free", {
			pricing: { prompt: "0.1", completion: "0" },
		}),
		liveModel("vendor/no-tools:free", { supported_parameters: ["max_tokens"] }),
		liveModel("openrouter/free", {
			context_length: 200_000,
			top_provider: { context_length: null, max_completion_tokens: null },
		}),
		liveModel("kilo-auto/free", {
			context_length: 256_000,
			top_provider: { context_length: 256_000, max_completion_tokens: 32_768 },
		}),
	],
};

let session: TestSession | undefined;
let upstream: Awaited<ReturnType<typeof startUpstream>> | undefined;
let agentDir: string | undefined;

afterEach(async () => {
	session?.dispose();
	session = undefined;
	await upstream?.close();
	upstream = undefined;
	if (agentDir) rmSync(agentDir, { recursive: true, force: true });
	agentDir = undefined;
});

interface Setup {
	respond: Responder;
	env?: Record<string, string>;
	auth?: Record<string, unknown>;
	firstResponseDeadlineMs?: number;
	routingStatePath?: string;
}

async function setup({
	respond,
	env = {},
	auth,
	firstResponseDeadlineMs,
	routingStatePath,
}: Setup) {
	upstream = await startUpstream(respond);
	agentDir = mkdtempSync(path.join(tmpdir(), "gratis-agent-"));
	if (auth)
		writeFileSync(path.join(agentDir, "auth.json"), JSON.stringify(auth));
	session = await createTestSession({
		extensionFactories: [
			createGratis({
				resolveUrl: upstream.resolveUrl,
				...(firstResponseDeadlineMs === undefined
					? {}
					: { firstResponseDeadlineMs }),
				...(routingStatePath === undefined ? {} : { routingStatePath }),
			}),
		],
		env: { PI_CODING_AGENT_DIR: agentDir, ...env },
	});
	const context = session.session.extensionRunner.createContext();
	const registry = context.modelRegistry as Registry;
	await registry.refresh({ allowNetwork: true, providers: ["grts"] });
	return { registry, context, upstream };
}

const grtsIds = (registry: Registry) =>
	registry
		.getAll()
		.filter((model) => model.provider === "grts")
		.map((model) => model.id);

const ask = (registry: Registry, id: string) => {
	const model = registry.find("grts", id);
	if (!model) throw new Error(`missing grts/${id}`);
	return registry
		.streamSimple(model, {
			messages: [{ role: "user", content: "hi", timestamp: Date.now() }],
		})
		.result();
};

const kiloOnly: Responder = (request) => {
	if (request.path === KILO_MODELS) return { status: 200, json: kiloCatalog };
	if (request.path === KILO_CHAT) return openAIText("kilo says hi");
	return undefined;
};

const chats = (requests: Recorded[], prefix: string) =>
	requests.filter(
		(request) => request.method === "POST" && request.path.startsWith(prefix),
	);

describe("gratis keyless kilo", () => {
	it("registers only free kilo models plus auto when no keys are set", async () => {
		const { registry } = await setup({ respond: kiloOnly });

		expect(grtsIds(registry)).toEqual([
			"auto",
			"kilo/kilo-auto/free",
			"kilo/openrouter/free",
			"kilo/minimax/minimax-m3:free",
		]);
		expect(
			registry.getAvailable().filter((model) => model.provider === "grts"),
		).toHaveLength(4);
	});

	it("labels every model free and data-caveated", async () => {
		const { registry } = await setup({ respond: kiloOnly });

		for (const model of registry
			.getAll()
			.filter((entry) => entry.provider === "grts")) {
			expect(model.cost).toEqual({
				input: 0,
				output: 0,
				cacheRead: 0,
				cacheWrite: 0,
			});
			expect(model.name).toContain("prompts may be logged");
		}
	});

	it("advertises published virtual-model parameters and falls back only when absent", async () => {
		const { registry } = await setup({ respond: kiloOnly });

		expect(registry.find("grts", "kilo/kilo-auto/free")).toMatchObject({
			contextWindow: 256_000,
			maxTokens: 32_768,
		});
		expect(registry.find("grts", "kilo/openrouter/free")).toMatchObject({
			contextWindow: 200_000,
			maxTokens: 8_192,
		});
	});

	it("stays invisible when the kilo catalog is unreachable and no key is set", async () => {
		const { registry } = await setup({
			respond: () => ({ status: 503, json: { error: "down" } }),
		});

		expect(grtsIds(registry)).toEqual([]);
	});

	it("yields no kilo models and no error for an empty free pool", async () => {
		const { registry } = await setup({
			respond: (request) =>
				request.path === KILO_MODELS
					? { status: 200, json: { data: [liveModel("vendor/paid")] } }
					: undefined,
		});

		expect(grtsIds(registry)).toEqual([]);
	});

	it("streams a pinned kilo model anonymously under the grts identity", async () => {
		const { registry, upstream } = await setup({ respond: kiloOnly });

		const message = await ask(registry, "kilo/kilo-auto/free");

		expect(message.stopReason).toBe("stop");
		expect(message.content).toEqual([{ type: "text", text: "kilo says hi" }]);
		expect(message.provider).toBe("grts");
		expect(message.model).toBe("kilo/kilo-auto/free");
		const [chat] = chats(upstream.requests, KILO_CHAT);
		expect(chat?.body?.model).toBe("kilo-auto/free");
		expect(chat?.headers.authorization).toBeUndefined();
	});

	it("never forces reasoning off on router models whose upstream may require it", async () => {
		const { registry, upstream } = await setup({
			respond: (request) =>
				request.path === KILO_MODELS
					? {
							status: 200,
							json: {
								data: [
									liveModel("openrouter/free", {
										supported_parameters: ["tools", "reasoning"],
									}),
								],
							},
						}
					: kiloOnly(request),
		});

		await ask(registry, "kilo/openrouter/free");

		const body = chats(upstream.requests, KILO_CHAT)[0]?.body ?? {};
		expect(body).not.toHaveProperty("reasoning");
	});

	it("sends an optional kilo key when one is set", async () => {
		const { registry, upstream } = await setup({
			respond: kiloOnly,
			env: { GRATIS_KILO_TOKEN: "kg-value" },
		});

		await ask(registry, "kilo/kilo-auto/free");

		expect(chats(upstream.requests, KILO_CHAT)[0]?.headers.authorization).toBe(
			"Bearer kg-value",
		);
	});

	it("reads the native kilo key when no gratis kilo key is set", async () => {
		const { registry, upstream } = await setup({
			respond: kiloOnly,
			env: { KILO_API_KEY: "kn-value" },
		});

		await ask(registry, "kilo/kilo-auto/free");

		expect(chats(upstream.requests, KILO_CHAT)[0]?.headers.authorization).toBe(
			"Bearer kn-value",
		);
	});
});

const everyBackend: Responder = (request) => {
	if (request.path === KILO_MODELS) return { status: 200, json: kiloCatalog };
	if (request.path === OPENROUTER_MODELS)
		return { status: 200, json: openrouterCatalog };
	if (request.path === KILO_CHAT) return openAIText("kilo says hi");
	if (request.path === OPENROUTER_CHAT) return openAIText("openrouter says hi");
	if (request.path.startsWith(OPENROUTER_MESSAGES))
		return anthropicText("anthropic shape says hi");
	if (request.path.startsWith(GOOGLE)) return googleText("gemini says hi");
	return undefined;
};

const backendsOf = (registry: Registry) => [
	...new Set(
		grtsIds(registry)
			.filter((id) => id !== "auto")
			.map((id) => id.slice(0, id.indexOf("/"))),
	),
];

describe("gratis keyed backends", () => {
	it("shows only the keyed backend next to kilo for one native key", async () => {
		const { registry } = await setup({
			respond: everyBackend,
			env: { GROQ_API_KEY: "gq-value" },
		});

		expect(backendsOf(registry)).toEqual(["groq", "kilo"]);
		expect(grtsIds(registry)).toContain("groq/openai/gpt-oss-120b");
	});

	it("registers every keyed backend with its curated catalog", async () => {
		const { registry } = await setup({
			respond: everyBackend,
			env: {
				GEMINI_API_KEY: "gm-value",
				OPENROUTER_API_KEY: "on-value",
				GROQ_API_KEY: "gq-value",
				MISTRAL_API_KEY: "ms-value",
			},
		});

		expect(backendsOf(registry)).toEqual([
			"google",
			"groq",
			"mistral",
			"openrouter",
			"kilo",
		]);
		expect(registry.find("grts", "mistral/devstral-latest")?.api).toBe(
			"mistral-conversations",
		);
		expect(registry.find("grts", "google/gemini-3.8-flash")?.api).toBe(
			"google-generative-ai",
		);
	});

	it("lets GRATIS_OPENROUTER_TOKEN override the native OpenRouter key", async () => {
		const { registry, upstream } = await setup({
			respond: everyBackend,
			env: {
				GRATIS_OPENROUTER_TOKEN: "og-value",
				OPENROUTER_API_KEY: "on-value",
			},
		});

		const message = await ask(registry, "openrouter/qwen/qwen3.8-27b:free");

		expect(message.content).toEqual([
			{ type: "text", text: "openrouter says hi" },
		]);
		expect(
			chats(upstream.requests, OPENROUTER_CHAT)[0]?.headers.authorization,
		).toBe("Bearer og-value");
	});

	it("keeps only free OpenRouter models, virtual router first", async () => {
		const { registry } = await setup({
			respond: everyBackend,
			env: { OPENROUTER_API_KEY: "on-value" },
		});

		expect(
			grtsIds(registry).filter((id) => id.startsWith("openrouter/")),
		).toEqual([
			"openrouter/openrouter/free",
			"openrouter/qwen/qwen3.8-27b:free",
			"openrouter/anthropic/claude-trial:free",
		]);
	});

	it("yields no OpenRouter models and no error for zero free variants", async () => {
		const { registry } = await setup({
			respond: (request) =>
				request.path === OPENROUTER_MODELS
					? { status: 200, json: { data: [liveModel("vendor/paid")] } }
					: everyBackend(request),
			env: { OPENROUTER_API_KEY: "on-value" },
		});

		expect(backendsOf(registry)).toEqual(["kilo"]);
	});

	it("falls back to the static OpenRouter router when discovery fails", async () => {
		const { registry } = await setup({
			respond: (request) =>
				request.path === OPENROUTER_MODELS
					? { status: 500, json: { error: "down" } }
					: everyBackend(request),
			env: { OPENROUTER_API_KEY: "on-value" },
		});

		expect(
			grtsIds(registry).filter((id) => id.startsWith("openrouter/")),
		).toEqual(["openrouter/openrouter/free"]);
	});

	it("registers a backend from a key in Pi's credential store", async () => {
		const { registry, upstream } = await setup({
			respond: everyBackend,
			auth: {
				openrouter: { type: "api_key", key: "os-value" },
			},
		});
		// Pi emits session_start while creating the session; lookup runs in the background.
		await expect
			.poll(() => backendsOf(registry))
			.toEqual(["openrouter", "kilo"]);
		const message = await ask(registry, "openrouter/openrouter/free");
		expect(message.stopReason).toBe("stop");
		expect(
			chats(upstream.requests, OPENROUTER_CHAT)[0]?.headers.authorization,
		).toBe("Bearer os-value");
	});

	it("streams google-shaped entries through the Google API", async () => {
		const { registry, upstream } = await setup({
			respond: everyBackend,
			env: { GEMINI_API_KEY: "gm-value" },
		});

		const message = await ask(registry, "google/gemini-3.8-flash");

		expect(message.stopReason).toBe("stop");
		expect(message.content).toEqual([{ type: "text", text: "gemini says hi" }]);
		const [request] = chats(upstream.requests, GOOGLE);
		expect(request?.path).toContain("gemini-3.8-flash:streamGenerateContent");
		expect(request?.headers["x-goog-api-key"]).toBe("gm-value");
	});

	it("never asks newer Gemini Flash models for the unsupported MINIMAL thinking level", async () => {
		const { registry, upstream } = await setup({
			respond: everyBackend,
			env: { GEMINI_API_KEY: "gm-value" },
		});

		await ask(registry, "google/gemini-3.8-flash");

		const [request] = chats(upstream.requests, GOOGLE);
		const config = JSON.stringify(request?.body?.generationConfig ?? {});
		expect(config).not.toMatch(/MINIMAL/i);
	});

	it("streams anthropic-shaped entries through the Anthropic API", async () => {
		const { registry, upstream } = await setup({
			respond: everyBackend,
			env: { OPENROUTER_API_KEY: "on-value" },
		});
		const id = "openrouter/anthropic/claude-trial:free";

		expect(registry.find("grts", id)?.api).toBe("anthropic-messages");
		const message = await ask(registry, id);

		expect(message.stopReason).toBe("stop");
		expect(message.content).toEqual([
			{ type: "text", text: "anthropic shape says hi" },
		]);
		expect(chats(upstream.requests, OPENROUTER_MESSAGES)[0]?.body?.model).toBe(
			"anthropic/claude-trial:free",
		);
	});
});

const GROQ_CHAT = "/groq/openai/v1/chat/completions";
const GROQ_CANDIDATES = [
	"groq:openai/gpt-oss-120b",
	"groq:llama-3.3-70b-versatile",
	"groq:openai/gpt-oss-20b",
	"groq:llama-3.1-8b-instant",
];

/** Kilo and Groq fakes whose chat replies depend on the upstream model. */
function cascade(
	kilo: (model: string) => Reply,
	groq: (model: string) => Reply = () => openAIText("groq says hi"),
): Responder {
	return (request) => {
		if (request.path === KILO_MODELS) return { status: 200, json: kiloCatalog };
		const model = String(request.body?.model);
		if (request.path === KILO_CHAT) return kilo(model);
		if (request.path === GROQ_CHAT) return groq(model);
		return undefined;
	};
}

const chatModels = (requests: Recorded[]) =>
	requests
		.filter((request) => request.method === "POST")
		.map((request) => `${request.path.split("/")[1]}:${request.body?.model}`);

describe("gratis auto router", () => {
	it("advertises the tightest candidate context and output limits", async () => {
		const { registry } = await setup({
			respond: cascade(() => openAIText("ok")),
			env: { GROQ_API_KEY: "gq-value" },
		});

		expect(registry.find("grts", "auto")).toMatchObject({
			contextWindow: 131_072,
			maxTokens: 8_192,
			cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
		});
	});

	it("tries the virtual kilo routers first", async () => {
		const { registry, upstream } = await setup({
			respond: cascade(() => openAIText("kilo says hi")),
		});

		const message = await ask(registry, "auto");

		expect(message.content).toEqual([{ type: "text", text: "kilo says hi" }]);
		expect(message.model).toBe("auto");
		expect(chatModels(upstream.requests)).toEqual(["kilo:kilo-auto/free"]);
	});

	it("completes on the next kilo candidate after upstream saturation, then skips the saturated one", async () => {
		const { registry, upstream } = await setup({
			respond: cascade((model) =>
				model === "kilo-auto/free" ? rateLimited : openAIText("router says hi"),
			),
		});

		const first = await ask(registry, "auto");
		const second = await ask(registry, "auto");

		expect(first.content).toEqual([{ type: "text", text: "router says hi" }]);
		expect(second.stopReason).toBe("stop");
		// kilo-auto/free cools on its own: the second request costs no round trip to it.
		expect(chatModels(upstream.requests)).toEqual([
			"kilo:kilo-auto/free",
			"kilo:openrouter/free",
			"kilo:openrouter/free",
		]);
	});

	it("moves to the next backend when groq is exhausted and then skips the cooling backend", async () => {
		const { registry, upstream } = await setup({
			respond: cascade(
				() => openAIText("kilo says hi"),
				() => rateLimited,
			),
			env: { GROQ_API_KEY: "gq-value" },
		});

		const first = await ask(registry, "auto");
		const second = await ask(registry, "auto");

		expect(first.stopReason).toBe("stop");
		expect(first.content).toEqual([{ type: "text", text: "kilo says hi" }]);
		expect(second.content).toEqual([{ type: "text", text: "kilo says hi" }]);
		expect(chatModels(upstream.requests)).toEqual([
			...GROQ_CANDIDATES,
			"kilo:kilo-auto/free",
			"kilo:kilo-auto/free",
		]);
	});

	it("advances on context overflow without cooling while a next backend remains", async () => {
		const { registry, upstream } = await setup({
			respond: cascade(
				() => openAIText("kilo says hi"),
				() => contextOverflow,
			),
			env: { GROQ_API_KEY: "gq-value" },
		});

		const first = await ask(registry, "auto");
		await ask(registry, "auto");

		expect(first.stopReason).toBe("stop");
		expect(first.content).toEqual([{ type: "text", text: "kilo says hi" }]);
		// Groq was not cooled: the second request tries it again.
		expect(
			chatModels(upstream.requests).filter((hop) => hop.startsWith("groq:")),
		).toHaveLength(8);
	});

	it("treats already-normalized overflow text as overflow", async () => {
		const { registry, upstream } = await setup({
			respond: cascade(
				() => openAIText("kilo says hi"),
				() => ({
					status: 400,
					json: { error: { message: "context_length_exceeded: upstream" } },
				}),
			),
			env: { GROQ_API_KEY: "gq-value" },
		});

		const first = await ask(registry, "auto");
		await ask(registry, "auto");

		expect(first.content).toEqual([{ type: "text", text: "kilo says hi" }]);
		expect(
			chatModels(upstream.requests).filter((hop) => hop.startsWith("groq:")),
		).toHaveLength(8);
	});

	it("reports overflow Pi can compact when every backend overflows", async () => {
		const { registry } = await setup({
			respond: cascade(
				() => contextOverflow,
				() => contextOverflow,
			),
			env: { GROQ_API_KEY: "gq-value" },
		});

		const message = await ask(registry, "auto");

		expect(message.stopReason).toBe("error");
		expect(message.errorMessage).toMatch(/^context_length_exceeded: /);
	});

	it("fails only when every backend is cooling and lists them", async () => {
		const { registry } = await setup({
			respond: cascade(
				() => rateLimited,
				() => ({
					status: 429,
					json: { error: { message: "You exceeded your current quota" } },
				}),
			),
			env: { GROQ_API_KEY: "gq-value" },
		});

		const first = await ask(registry, "auto");
		const second = await ask(registry, "auto");

		expect(first.stopReason).toBe("error");
		expect(first.errorMessage).toContain("kilo exhausted (rate)");
		expect(first.errorMessage).toContain("groq exhausted (quota)");
		expect(second.errorMessage).toContain("kilo cooling");
		expect(second.errorMessage).toContain("groq cooling");
	});

	it("makes a cooled backend available again after its rate window", async () => {
		vi.useFakeTimers({ toFake: ["Date"] });
		try {
			let kiloDown = false;
			const { registry, upstream } = await setup({
				respond: cascade(
					() => (kiloDown ? rateLimited : openAIText("kilo says hi")),
					() => rateLimited,
				),
				env: { GROQ_API_KEY: "gq-value" },
			});
			await ask(registry, "auto");
			const groqHops = () =>
				chatModels(upstream.requests).filter((hop) => hop.startsWith("groq:"))
					.length;
			expect(groqHops()).toBe(4);

			// Groq is now demoted behind Kilo, but no longer cooling: it is tried when Kilo fails.
			kiloDown = true;
			vi.setSystemTime(Date.now() + 60_000 + 1);
			await ask(registry, "auto");

			expect(groqHops()).toBe(8);
		} finally {
			vi.useRealTimers();
		}
	});

	it("replays same-model history with its native reasoning state", async () => {
		const { registry, upstream } = await setup({ respond: kiloOnly });
		const id = "kilo/kilo-auto/free";
		const model = registry.find("grts", id);
		if (!model) throw new Error("missing model");
		const previous = await ask(registry, id);
		previous.content = [
			{
				type: "thinking",
				thinking: "plan",
				thinkingSignature: JSON.stringify([
					{ type: "reasoning.text", text: "plan", index: 0 },
				]),
			},
			{ type: "text", text: "kilo says hi" },
		];

		await registry
			.streamSimple(model, {
				messages: [
					{ role: "user", content: "hi", timestamp: 1 },
					previous,
					{ role: "user", content: "again", timestamp: 2 },
				],
			})
			.result();

		const replay = chats(upstream.requests, KILO_CHAT).at(-1)?.body
			?.messages as Record<string, unknown>[];
		expect(replay[1]?.reasoning_details).toEqual([
			{ type: "reasoning.text", text: "plan", index: 0 },
		]);
	});

	it("moves on when a backend reports upstream high demand", async () => {
		const { registry } = await setup({
			respond: cascade(() => ({
				status: 503,
				json: {
					error: {
						code: 503,
						message: "This model is currently experiencing high demand.",
						status: "UNAVAILABLE",
					},
				},
			})),
			env: { GROQ_API_KEY: "gq-value" },
		});

		const message = await ask(registry, "auto");

		expect(message.content).toEqual([{ type: "text", text: "groq says hi" }]);
	});

	it("surfaces other upstream errors instead of cascading", async () => {
		const { registry, upstream } = await setup({
			respond: cascade(
				() => openAIText("kilo says hi"),
				() => ({
					status: 401,
					json: { error: { message: "invalid credentials" } },
				}),
			),
			env: { GROQ_API_KEY: "gq-value" },
		});

		const message = await ask(registry, "auto");

		expect(message.stopReason).toBe("error");
		expect(message.errorMessage).toContain("invalid credentials");
		expect(chatModels(upstream.requests)).toEqual(["groq:openai/gpt-oss-120b"]);
	});

	it("moves past a hop that streams nothing before its deadline", async () => {
		const { registry, upstream } = await setup({
			respond: cascade(
				() => openAIText("kilo says hi"),
				() => ({ hang: true }),
			),
			env: { GROQ_API_KEY: "gq-value" },
			firstResponseDeadlineMs: 150,
		});
		const started = Date.now();

		const first = await ask(registry, "auto");
		const second = await ask(registry, "auto");

		expect(first.content).toEqual([{ type: "text", text: "kilo says hi" }]);
		expect(second.content).toEqual([{ type: "text", text: "kilo says hi" }]);
		// One silent hop marks Groq congested: no waiting on its siblings, then Groq cools.
		expect(chatModels(upstream.requests)).toEqual([
			"groq:openai/gpt-oss-120b",
			"kilo:kilo-auto/free",
			"kilo:kilo-auto/free",
		]);
		expect(Date.now() - started).toBeLessThan(2_000);
	});

	it("skips a listed model the account cannot use instead of failing", async () => {
		const notFound: Reply = {
			status: 404,
			json: {
				status: 404,
				title: "Not Found",
				detail: "Function 'abc': Not found for account 'xyz'",
			},
		};
		const { registry, upstream } = await setup({
			respond: cascade(
				() => openAIText("kilo says hi"),
				(model) =>
					model === "openai/gpt-oss-120b"
						? notFound
						: openAIText("groq says hi"),
			),
			env: { GROQ_API_KEY: "gq-value" },
		});

		const first = await ask(registry, "auto");
		const second = await ask(registry, "auto");

		expect(first.content).toEqual([{ type: "text", text: "groq says hi" }]);
		expect(second.content).toEqual([{ type: "text", text: "groq says hi" }]);
		expect(chatModels(upstream.requests)).toEqual([
			"groq:openai/gpt-oss-120b",
			"groq:llama-3.3-70b-versatile",
			"groq:llama-3.3-70b-versatile",
		]);
	});

	it("still surfaces a pinned model's not-found error", async () => {
		const { registry } = await setup({
			respond: cascade(
				() => openAIText("kilo says hi"),
				() => ({
					status: 404,
					json: { error: { message: "model not found" } },
				}),
			),
			env: { GROQ_API_KEY: "gq-value" },
		});

		const message = await ask(registry, "groq/openai/gpt-oss-120b");

		expect(message.stopReason).toBe("error");
		expect(message.errorMessage).toContain("model not found");
	});

	it("completes in the last-resort pass when every backend is cooling", async () => {
		let recovered = false;
		const { registry, upstream } = await setup({
			respond: cascade(
				() => rateLimited,
				() => (recovered ? openAIText("groq is back") : rateLimited),
			),
			env: { GROQ_API_KEY: "gq-value" },
		});
		const failed = await ask(registry, "auto");
		expect(failed.stopReason).toBe("error");
		const before = upstream.requests.length;
		recovered = true;

		const message = await ask(registry, "auto");

		expect(message.content).toEqual([{ type: "text", text: "groq is back" }]);
		expect(chatModels(upstream.requests.slice(before))).toEqual([
			"groq:openai/gpt-oss-120b",
		]);
	});

	it("returns the upstream error for a pinned model that 429s", async () => {
		const { registry, upstream } = await setup({
			respond: cascade(() => rateLimited),
			env: { GROQ_API_KEY: "gq-value" },
		});

		const message = await ask(registry, "kilo/minimax/minimax-m3:free");

		expect(message.stopReason).toBe("error");
		expect(message.errorMessage).toContain("Rate limit exceeded");
		expect(chatModels(upstream.requests)).toEqual([
			"kilo:minimax/minimax-m3:free",
		]);
	});

	it("normalizes unrecognized overflow errors so Pi compacts", async () => {
		const { registry } = await setup({
			respond: cascade(() => ({
				status: 400,
				json: {
					error: {
						message:
							"Requested 300000 tokens but this free model's context window is 262144.",
					},
				},
			})),
		});
		const message = await ask(registry, "kilo/minimax/minimax-m3:free");
		expect(isContextOverflow(message)).toBe(false);

		const runner = session?.session.extensionRunner;
		const [handler] = runner.extensions[0].handlers.get("message_end");
		const result = await handler(
			{ type: "message_end", message },
			runner.createContext(),
		);

		expect(result?.message?.errorMessage).toMatch(/^context_length_exceeded: /);
		expect(isContextOverflow(result.message)).toBe(true);
	});
});

const NVIDIA_MODELS = "/nvidia/v1/models";
const NVIDIA_CHAT = "/nvidia/v1/chat/completions";
const ZAI_CHAT = "/zai/api/paas/v4/chat/completions";
const HETZNER_MODELS = "/hetzner/api/v1/models";
const HETZNER_CHAT = "/hetzner/api/v1/chat/completions";

const nvidiaLive = {
	data: [
		{ id: "z-ai/glm-5.3", object: "model" },
		{ id: "moonshotai/kimi-k2.6", object: "model" },
		{ id: "nvidia/nemotron-3-ultra-550b-a55b", object: "model" },
		{ id: "vendor/unknown-model", object: "model" },
	],
};
const hetznerLive = {
	data: [
		{ id: "Qwen/Qwen3.6-35B-A3B-FP8", max_model_len: 131_072 },
		{ id: "Vendor/New-Model", max_model_len: 65_536 },
	],
};

const addedBackends: Responder = (request) => {
	if (request.path === KILO_MODELS) return { status: 200, json: kiloCatalog };
	if (request.path === NVIDIA_MODELS) return { status: 200, json: nvidiaLive };
	if (request.path === HETZNER_MODELS)
		return request.headers.authorization === "Bearer hz-value"
			? { status: 200, json: hetznerLive }
			: { status: 401, json: { error: "unauthorized" } };
	if (request.path === NVIDIA_CHAT) return openAIText("nvidia says hi");
	if (request.path === ZAI_CHAT) return openAIText("zai says hi");
	if (request.path === HETZNER_CHAT) return openAIText("hetzner says hi");
	return undefined;
};

const idsOf = (registry: Registry, backend: string) =>
	grtsIds(registry).filter((id) => id.startsWith(`${backend}/`));

describe("gratis added backends", () => {
	it("keeps only curated free NVIDIA models that are live, in curated rank", async () => {
		const { registry } = await setup({
			respond: addedBackends,
			env: { NVIDIA_API_KEY: "nv-value" },
		});

		expect(backendsOf(registry)).toEqual(["nvidia", "kilo"]);
		expect(idsOf(registry, "nvidia")).toEqual([
			"nvidia/z-ai/glm-5.3",
			"nvidia/moonshotai/kimi-k2.6",
		]);
	});

	it("falls back to the curated NVIDIA list when the live list is unreachable", async () => {
		const { registry } = await setup({
			respond: (request) =>
				request.path === NVIDIA_MODELS
					? { status: 503, json: { error: "down" } }
					: addedBackends(request),
			env: { NVIDIA_API_KEY: "nv-value" },
		});

		expect(idsOf(registry, "nvidia")).toEqual([
			"nvidia/moonshotai/kimi-k3",
			"nvidia/z-ai/glm-5.3",
			"nvidia/moonshotai/kimi-k2.6",
			"nvidia/z-ai/glm-5.3-flash",
			"nvidia/nvidia/nemotron-3.5-lightning-30b-a3b",
			"nvidia/openai/gpt-oss-20b",
		]);
	});

	it("streams a pinned NVIDIA model with the NVIDIA key", async () => {
		const { registry, upstream } = await setup({
			respond: addedBackends,
			env: { GRATIS_NVIDIA_TOKEN: "nvg-value", NVIDIA_API_KEY: "nv-value" },
		});

		const message = await ask(registry, "nvidia/z-ai/glm-5.3");

		expect(message.content).toEqual([{ type: "text", text: "nvidia says hi" }]);
		const [chat] = chats(upstream.requests, NVIDIA_CHAT);
		expect(chat?.body?.model).toBe("z-ai/glm-5.3");
		expect(chat?.headers.authorization).toBe("Bearer nvg-value");
	});

	it("registers exactly the free Z.ai Flash models on the general endpoint", async () => {
		const { registry, upstream } = await setup({
			respond: addedBackends,
			env: { ZAI_API_KEY: "zai-value" },
		});

		expect(idsOf(registry, "zai")).toEqual([
			"zai/glm-4.7-flash",
			"zai/glm-4.5-flash",
			"zai/glm-4.6v-flash",
		]);
		const message = await ask(registry, "zai/glm-4.7-flash");

		expect(message.content).toEqual([{ type: "text", text: "zai says hi" }]);
		const posts = upstream.requests.filter((r) => r.method === "POST");
		expect(posts.map((r) => r.path)).toEqual([ZAI_CHAT]);
		expect(posts.some((r) => r.path.includes("/coding/"))).toBe(false);
		expect(posts[0]?.headers.authorization).toBe("Bearer zai-value");
	});

	it("follows Hetzner's keyed live list, keeping curated metadata for known ids", async () => {
		const { registry, upstream } = await setup({
			respond: addedBackends,
			env: { HETZNER_INFERENCE_API_KEY: "hz-value" },
		});

		expect(idsOf(registry, "hetzner")).toEqual([
			"hetzner/Qwen/Qwen3.6-35B-A3B-FP8",
			"hetzner/Vendor/New-Model",
		]);
		expect(
			registry.find("grts", "hetzner/Qwen/Qwen3.6-35B-A3B-FP8"),
		).toMatchObject({
			contextWindow: 131_072,
			maxTokens: 65_536,
			input: ["text", "image"],
		});
		expect(registry.find("grts", "hetzner/Vendor/New-Model")).toMatchObject({
			contextWindow: 65_536,
			maxTokens: 8_192,
			input: ["text"],
		});
		const discovery = upstream.requests.find((r) => r.path === HETZNER_MODELS);
		expect(discovery?.headers.authorization).toBe("Bearer hz-value");

		const message = await ask(registry, "hetzner/Qwen/Qwen3.6-35B-A3B-FP8");
		expect(message.content).toEqual([
			{ type: "text", text: "hetzner says hi" },
		]);
	});

	it("falls back to the static Hetzner list when the live list is unreachable", async () => {
		const { registry } = await setup({
			respond: (request) =>
				request.path === HETZNER_MODELS
					? { status: 500, json: { error: "down" } }
					: addedBackends(request),
			env: { GRATIS_HETZNER_TOKEN: "hz-value" },
		});

		expect(idsOf(registry, "hetzner")).toEqual([
			"hetzner/Qwen/Qwen3.6-35B-A3B-FP8",
			"hetzner/Qwen3.8-27B",
		]);
	});

	it("never calls Hetzner or lists it without a key", async () => {
		const { registry, upstream } = await setup({ respond: addedBackends });

		expect(backendsOf(registry)).toEqual(["kilo"]);
		expect(upstream.requests.some((r) => r.path.startsWith("/hetzner/"))).toBe(
			false,
		);
	});

	it("orders every backend by auto priority when all are keyed", async () => {
		const { registry } = await setup({
			respond: (request) =>
				request.path === OPENROUTER_MODELS
					? { status: 200, json: openrouterCatalog }
					: addedBackends(request),
			env: {
				GEMINI_API_KEY: "gm-value",
				NVIDIA_API_KEY: "nv-value",
				OPENROUTER_API_KEY: "on-value",
				GROQ_API_KEY: "gq-value",
				MISTRAL_API_KEY: "ms-value",
				ZAI_API_KEY: "zai-value",
				HETZNER_INFERENCE_API_KEY: "hz-value",
			},
		});

		expect(backendsOf(registry)).toEqual([
			"google",
			"zai",
			"groq",
			"mistral",
			"nvidia",
			"openrouter",
			"hetzner",
			"kilo",
		]);
	});

	it("cools a Z.ai backend reporting insufficient balance and moves on", async () => {
		const { registry, upstream } = await setup({
			respond: (request) => {
				if (request.path === KILO_MODELS)
					return { status: 200, json: { data: [] } };
				if (request.path === ZAI_CHAT)
					return {
						status: 429,
						json: { error: { code: "1113", message: "Insufficient balance" } },
					};
				return addedBackends(request);
			},
			env: {
				ZAI_API_KEY: "zai-value",
				HETZNER_INFERENCE_API_KEY: "hz-value",
			},
		});

		const first = await ask(registry, "auto");
		const second = await ask(registry, "auto");

		expect(first.content).toEqual([{ type: "text", text: "hetzner says hi" }]);
		expect(second.content).toEqual([{ type: "text", text: "hetzner says hi" }]);
		expect(chats(upstream.requests, ZAI_CHAT)).toHaveLength(2);
	});
});

const statusOf = async () => {
	const runner = session?.session.extensionRunner;
	const command = runner.extensions[0].commands.get("gratis");
	await command.handler("", runner.createContext());
	return String(session?.events.uiCallsFor("notify").at(-1)?.args[0]);
};

describe("gratis served model", () => {
	it("names the model a virtual router resolved to, under the requested model", async () => {
		const { registry } = await setup({
			respond: (request) =>
				request.path === KILO_CHAT
					? openAIText("router says hi", "dots-studio/dots-3-note-preview:free")
					: kiloOnly(request),
		});

		const pinned = await ask(registry, "kilo/kilo-auto/free");
		const auto = await ask(registry, "auto");

		expect(pinned).toMatchObject({
			model: "kilo/kilo-auto/free",
			responseModel: "kilo/dots-studio/dots-3-note-preview:free",
		});
		expect(auto).toMatchObject({
			model: "auto",
			responseModel: "kilo/dots-studio/dots-3-note-preview:free",
		});
	});

	it("leaves responseModel unset when a pinned model served itself", async () => {
		const { registry } = await setup({
			respond: (request) =>
				request.path === GROQ_CHAT
					? openAIText("groq says hi", "openai/gpt-oss-120b")
					: kiloOnly(request),
			env: { GROQ_API_KEY: "gq-value" },
		});

		const message = await ask(registry, "groq/openai/gpt-oss-120b");

		expect(message.model).toBe("groq/openai/gpt-oss-120b");
		expect(message.responseModel).toBeUndefined();
	});

	it("falls back to the hop model when the transport reports none", async () => {
		const { registry } = await setup({
			respond: (request) =>
				request.path === KILO_MODELS
					? { status: 200, json: { data: [] } }
					: everyBackend(request),
			env: { GEMINI_API_KEY: "gm-value" },
		});

		const message = await ask(registry, "auto");

		expect(message.responseModel).toBe("google/gemini-3.8-flash");
	});
});

describe("gratis status command", () => {
	it("reports backends, missing keys, and that nothing ran yet", async () => {
		await setup({ respond: kiloOnly, env: { GROQ_API_KEY: "gq-value" } });

		const text = await statusOf();

		expect(text).toContain("Backends: groq 6, kilo 3");
		expect(text).toContain(
			"No key: google, zai, mistral, nvidia, openrouter, hetzner",
		);
		expect(text).toContain("Cooling: none");
		expect(text).toContain("Last: no grts request yet");
	});

	it("reports the last auto route, what it skipped, and cooling backends", async () => {
		const { registry } = await setup({
			respond: cascade(
				() => openAIText("kilo says hi", "dots-studio/dots-3:free"),
				() => rateLimited,
			),
			env: { GROQ_API_KEY: "gq-value" },
		});

		await ask(registry, "auto");
		const text = await statusOf();

		expect(text).toMatch(
			/Last: auto → kilo\/dots-studio\/dots-3:free · ok · 5 hops · \d+s ago/,
		);
		expect(text).toContain("Skipped: groq exhausted (rate)");
		expect(text).toMatch(/Cooling: groq \(1m left\)/);
	});

	it("reports a failed pinned route as an error without a served model", async () => {
		const { registry } = await setup({
			respond: cascade(() => rateLimited),
		});

		await ask(registry, "kilo/minimax/minimax-m3:free");
		const text = await statusOf();

		expect(text).toMatch(
			/Last: kilo\/minimax\/minimax-m3:free → kilo\/minimax\/minimax-m3:free · error · 1 hop/,
		);
	});
});

describe("gratis performance-aware routing", () => {
	it("cuts a hop sooner when its model usually starts fast", async () => {
		let groqCalls = 0;
		const { registry } = await setup({
			respond: cascade(
				() => openAIText("kilo says hi"),
				() => (++groqCalls <= 3 ? openAIText("groq says hi") : { hang: true }),
			),
			env: { GROQ_API_KEY: "gq-value" },
			firstResponseDeadlineMs: 4_000,
		});
		for (let i = 0; i < 3; i++) await ask(registry, "auto");
		const started = Date.now();

		const message = await ask(registry, "auto");

		expect(message.content).toEqual([{ type: "text", text: "kilo says hi" }]);
		const waited = Date.now() - started;
		expect(waited).toBeGreaterThanOrEqual(2_900);
		expect(waited).toBeLessThan(3_800);
	});

	it("tries a repeatedly failing model after its healthy siblings", async () => {
		vi.useFakeTimers({ toFake: ["Date"] });
		try {
			const { registry, upstream } = await setup({
				respond: cascade(
					() => openAIText("kilo says hi"),
					(model) =>
						model === "openai/gpt-oss-120b"
							? rateLimited
							: openAIText("groq says hi"),
				),
				env: { GROQ_API_KEY: "gq-value" },
			});
			for (let i = 0; i < 3; i++) {
				await ask(registry, "auto");
				vi.setSystemTime(Date.now() + 61_000);
			}
			const before = upstream.requests.length;

			const message = await ask(registry, "auto");

			expect(message.content).toEqual([{ type: "text", text: "groq says hi" }]);
			expect(chatModels(upstream.requests.slice(before))).toEqual([
				"groq:llama-3.3-70b-versatile",
			]);
		} finally {
			vi.useRealTimers();
		}
	});

	it("shows measured models in the status command", async () => {
		const { registry } = await setup({ respond: kiloOnly });

		await ask(registry, "auto");
		const text = await statusOf();

		expect(text).toMatch(
			/Models:\n {2}kilo\/kilo-auto\/free · first token \d+\.\ds · 0% fail · 1 hop/,
		);
	});
});

const gratisCommand = async (args: string) => {
	const runner = session?.session.extensionRunner;
	const command = runner.extensions[0].commands.get("gratis");
	await command.handler(args, runner.createContext());
	const call = session?.events.uiCallsFor("notify").at(-1);
	return { text: String(call?.args[0]), level: call?.args[1] };
};
const completions = (prefix: string) =>
	session?.session.extensionRunner.extensions[0].commands
		.get("gratis")
		.getArgumentCompletions(prefix);

describe("gratis manual cooling", () => {
	const kiloRouter: Responder = (request) =>
		request.path === KILO_CHAT
			? openAIText("kilo says hi", "dots-studio/dots-3:free")
			: kiloOnly(request);

	it("asks for a model when nothing has answered yet", async () => {
		await setup({ respond: kiloRouter });

		const reply = await gratisCommand("cool");

		expect(reply.level).toBe("warning");
		expect(reply.text).toContain("name one: /gratis cool <model>");
	});

	it("cools the candidate that answered last, naming what it answered as", async () => {
		const { registry, upstream } = await setup({ respond: kiloRouter });
		await ask(registry, "auto");

		const reply = await gratisCommand("cool");
		const before = upstream.requests.length;
		await ask(registry, "auto");

		expect(reply).toEqual({
			level: "info",
			text: "grts/auto skips kilo/kilo-auto/free for 60 min (it answered as kilo/dots-studio/dots-3:free; gratis can only skip kilo/kilo-auto/free itself)",
		});
		expect(chatModels(upstream.requests.slice(before))).toEqual([
			"kilo:openrouter/free",
		]);
		expect((await gratisCommand("")).text).toContain(
			"Manually cooled: kilo/kilo-auto/free (60m left)",
		);
	});

	it("keeps a manually cooled model out of the last-resort pass", async () => {
		const { registry, upstream } = await setup({
			respond: cascade((model) =>
				model === "kilo-auto/free" ? openAIText("unwanted") : rateLimited,
			),
		});
		await gratisCommand("cool kilo/kilo-auto/free 30");

		const message = await ask(registry, "auto");
		await ask(registry, "auto");

		expect(message.stopReason).toBe("error");
		expect(chatModels(upstream.requests)).not.toContain("kilo:kilo-auto/free");
	});

	it("ends manual cooldowns on uncool", async () => {
		const { registry, upstream } = await setup({ respond: kiloRouter });
		await gratisCommand("cool kilo/kilo-auto/free");

		const one = await gratisCommand("uncool kilo/kilo-auto/free");
		await ask(registry, "auto");
		await gratisCommand("cool kilo/openrouter/free");
		await gratisCommand("cool kilo/minimax/minimax-m3:free");
		const all = await gratisCommand("uncool");
		const none = await gratisCommand("uncool");

		expect(one.text).toBe("kilo/kilo-auto/free is back in grts/auto");
		expect(chatModels(upstream.requests)).toEqual(["kilo:kilo-auto/free"]);
		expect(all.text).toBe("2 manually cooled models are back in grts/auto");
		expect(none).toEqual({
			level: "warning",
			text: "No model is manually cooled",
		});
	});

	it("still answers when the cooled model is pinned", async () => {
		const { registry } = await setup({ respond: kiloRouter });
		await gratisCommand("cool kilo/kilo-auto/free");

		const message = await ask(registry, "kilo/kilo-auto/free");

		expect(message.content).toEqual([{ type: "text", text: "kilo says hi" }]);
	});

	it("rejects unknown models, auto, and bad durations", async () => {
		await setup({ respond: kiloRouter });

		expect(await gratisCommand("cool kilo/nope")).toEqual({
			level: "warning",
			text: "kilo/nope is not a current grts model",
		});
		expect((await gratisCommand("cool auto")).text).toBe(
			"auto is not a current grts model",
		);
		for (const minutes of ["0", "1.5", "abc", "1441"])
			expect(
				(await gratisCommand(`cool kilo/kilo-auto/free ${minutes}`)).text,
			).toBe("Minutes must be a whole number from 1 to 1440");
		expect((await gratisCommand("warm")).text).toBe(
			"Usage: /gratis [cool [model] [minutes] | uncool [model]]",
		);
	});

	it("autocompletes subcommands, models, and manually cooled models", async () => {
		await setup({ respond: kiloRouter });
		await gratisCommand("cool kilo/openrouter/free");

		expect(await completions("c")).toEqual([{ value: "cool", label: "cool" }]);
		expect(await completions("cool kilo/k")).toEqual([
			{ value: "cool kilo/kilo-auto/free", label: "kilo/kilo-auto/free" },
		]);
		expect(await completions("uncool ")).toEqual([
			{ value: "uncool kilo/openrouter/free", label: "kilo/openrouter/free" },
		]);
		expect(await completions("cool kilo/kilo-auto/free 3")).toBeNull();
	});
});

const ZAI_PRICING_PATH = "/zaidocs/guides/overview/pricing.md";
const MODELS_DEV_PATH = "/modelsdev/api.json";
const freeRow = (name: string) => `| ${name} | Free | Free | Free | Free |`;
const pricingPage = (...rows: string[]) =>
	[
		"### Text Models",
		"| Model | Input | Cached Input | Cached Input Storage | Output |",
		"| :- | :- | :- | :- | :- |",
		"| GLM-4.7 | $0.6 | $0.11 | Limited-time Free | $2.2 |",
		...rows,
	].join("\n");
const modelsDev = {
	zai: {
		api: "https://api.z.ai/api/paas/v4",
		models: {
			"glm-9-flash": {
				name: "GLM 9 Flash",
				cost: { input: 0, output: 0 },
				limit: { context: 300_000, output: 64_000 },
				modalities: { input: ["text"] },
			},
			"glm-9-pro": {
				name: "GLM 9 Pro",
				cost: { input: 2, output: 8 },
				limit: { context: 300_000, output: 64_000 },
			},
		},
	},
};

describe("gratis fresh catalogs", () => {
	it("lists Google models from Pi's catalog with Pi's thinking levels", async () => {
		const { registry, context } = await setup({
			respond: kiloOnly,
			env: { GEMINI_API_KEY: "gm-value" },
		});
		const pi = context.modelRegistry.find("google", "gemini-3.7-flash");

		const ids = idsOf(registry, "google");

		expect(ids[0]).toBe("google/gemini-3.8-flash");
		expect(ids).toContain("google/gemini-3.6-flash");
		expect(ids.some((id) => /latest|preview/.test(id))).toBe(false);
		expect(
			registry.find("grts", "google/gemini-3.7-flash")?.thinkingLevelMap,
		).toEqual(pi?.thinkingLevelMap);
	});

	it("shows a model added to Pi's catalog without a gratis update", async () => {
		const { registry, context } = await setup({
			respond: kiloOnly,
			env: { GROQ_API_KEY: "gq-value" },
		});
		expect(idsOf(registry, "groq")).not.toContain("groq/brand-new-model");

		context.modelRegistry.registerProvider("groq", {
			baseUrl: "https://api.groq.com/openai/v1",
			apiKey: "gq-value",
			api: "openai-completions",
			models: [
				{
					id: "brand-new-model",
					name: "Brand New",
					reasoning: false,
					input: ["text"],
					cost: { input: 0.1, output: 0.1, cacheRead: 0, cacheWrite: 0 },
					contextWindow: 131_072,
					maxTokens: 16_384,
				},
			],
		});

		expect(idsOf(registry, "groq")).toEqual(["groq/brand-new-model"]);
		expect(registry.find("grts", "groq/brand-new-model")?.cost.input).toBe(0);
	});

	it("follows Z.ai's pricing page, confirming unknown free ids on models.dev", async () => {
		const { registry, upstream } = await setup({
			respond: (request) => {
				if (request.path === ZAI_PRICING_PATH)
					return {
						status: 200,
						text: pricingPage(
							freeRow("GLM-4.7-Flash"),
							freeRow("GLM-9-Flash"),
							freeRow("GLM-9-Pro"),
						),
					};
				if (request.path === MODELS_DEV_PATH)
					return { status: 200, json: modelsDev };
				return kiloOnly(request);
			},
			env: { ZAI_API_KEY: "zai-value" },
		});

		expect(idsOf(registry, "zai")).toEqual([
			"zai/glm-4.7-flash",
			"zai/glm-9-flash",
		]);
		expect(registry.find("grts", "zai/glm-9-flash")).toMatchObject({
			baseUrl: "https://api.z.ai/api/paas/v4",
			contextWindow: 300_000,
		});
		expect(
			upstream.requests.filter((request) => request.path === MODELS_DEV_PATH),
		).toHaveLength(1);
	});

	it("lists no Z.ai models when the page lists none as Free", async () => {
		const { registry } = await setup({
			respond: (request) =>
				request.path === ZAI_PRICING_PATH
					? { status: 200, text: pricingPage() }
					: kiloOnly(request),
			env: { ZAI_API_KEY: "zai-value" },
		});

		expect(idsOf(registry, "zai")).toEqual([]);
	});

	it("keeps the last confirmed Z.ai list when the page becomes unreachable", async () => {
		let pageUp = true;
		const { registry } = await setup({
			respond: (request) =>
				request.path === ZAI_PRICING_PATH
					? pageUp
						? { status: 200, text: pricingPage(freeRow("GLM-4.5-Flash")) }
						: { status: 503, json: { error: "down" } }
					: kiloOnly(request),
			env: { ZAI_API_KEY: "zai-value" },
		});
		expect(idsOf(registry, "zai")).toEqual(["zai/glm-4.5-flash"]);

		pageUp = false;
		await registry.refresh({ allowNetwork: true, providers: ["grts"] });

		expect(idsOf(registry, "zai")).toEqual(["zai/glm-4.5-flash"]);
	});
});

describe("gratis learning across time", () => {
	it("keeps a twice-rejected model behind its siblings after the cooldowns end", async () => {
		vi.useFakeTimers({ toFake: ["Date"] });
		try {
			const { registry, upstream } = await setup({ respond: kiloOnly });
			await gratisCommand("cool kilo/kilo-auto/free");
			await gratisCommand("cool kilo/kilo-auto/free");
			vi.setSystemTime(Date.now() + 61 * 60_000);

			await ask(registry, "auto");

			expect(chatModels(upstream.requests)).toEqual(["kilo:openrouter/free"]);
		} finally {
			vi.useRealTimers();
		}
	});

	it("still applies a demotion learned before a restart", async () => {
		const stateDir = mkdtempSync(path.join(tmpdir(), "gratis-routing-"));
		const routingStatePath = path.join(stateDir, "routing.json");
		const respond = cascade(
			() => openAIText("kilo says hi"),
			(model) =>
				model === "openai/gpt-oss-120b"
					? rateLimited
					: openAIText("groq says hi"),
		);
		vi.useFakeTimers({ toFake: ["Date"] });
		try {
			const first = await setup({
				respond,
				env: { GROQ_API_KEY: "gq-value" },
				routingStatePath,
			});
			for (let i = 0; i < 3; i++) {
				await ask(first.registry, "auto");
				vi.setSystemTime(Date.now() + 61_000);
			}
			await session?.session.extensionRunner.emit({
				type: "session_shutdown",
				reason: "quit",
			});
			session?.dispose();
			session = undefined;
			await upstream?.close();
			upstream = undefined;

			const second = await setup({
				respond,
				env: { GROQ_API_KEY: "gq-value" },
				routingStatePath,
			});
			// Wait for the background load, then send exactly one request: a second
			// request would skip the model through its fresh 429 cooldown anyway.
			await expect
				.poll(statusOf)
				.toMatch(/groq\/openai\/gpt-oss-120b · .*demoted/);
			const before = second.upstream.requests.length;

			await ask(second.registry, "auto");

			expect(chatModels(second.upstream.requests.slice(before))[0]).toBe(
				"groq:llama-3.3-70b-versatile",
			);
		} finally {
			vi.useRealTimers();
			rmSync(stateDir, { recursive: true, force: true });
		}
	});
});