/** * Kernel test harness — ports of the idealized reference model tests. * * Tests the ControllerKernel with in-memory adapters to validate * core control-plane semantics independently of HTTP transport, * SQLite persistence, or process management. */import { describe, expect, test, beforeEach } from "bun:test";import type { RecipeId } from "../types/brand";import type { Recipe } from "../modules/lifecycle/types";import { ControllerKernel } from "./kernel";import { EVENT_TYPE, RUN_STATUS } from "./contracts";import { createTestKernelParts, type TestKernelParts } from "./testing"; function makeRecipe(overrides: {  id: string;  name: string;  served_model_name: string;  model_path: string;  backend: Recipe["backend"];  port?: number;}): Recipe {  return {    id: overrides.id as RecipeId,    name: overrides.name,    served_model_name: overrides.served_model_name,    model_path: overrides.model_path,    backend: overrides.backend,    port: overrides.port ?? 8000,    host: "0.0.0.0",    env_vars: null,    tensor_parallel_size: 1,    pipeline_parallel_size: 1,    max_model_len: 4096,    gpu_memory_utilization: 0.9,    kv_cache_dtype: "auto",    max_num_seqs: 16,    trust_remote_code: false,    tool_call_parser: null,    reasoning_parser: null,    enable_auto_tool_choice: false,    quantization: null,    dtype: null,    python_path: null,    extra_args: {},    max_thinking_tokens: null,    thinking_mode: "off",  };} // ---------------------------------------------------------------------------// Unit tests (from idealized unit.test.mjs)// --------------------------------------------------------------------------- describe("kernel unit tests", () => {  let parts: TestKernelParts;   beforeEach(() => {    parts = createTestKernelParts();  });   test("provider router resolves managed models by id or served model name", () => {    const recipe = makeRecipe({ id: "phi4", name: "Phi-4", served_model_name: "phi-4", model_path: "/models/phi-4", backend: "vllm" });    parts.recipes.upsert(recipe);     const byId = parts.providers.resolve("phi4");    expect(byId.kind).toBe("managed");    if (byId.kind === "managed") {      expect(String(byId.recipe.id)).toBe("phi4");      expect(byId.resolvedModel).toBe("phi-4");    }     const byServed = parts.providers.resolve("phi-4");    expect(byServed.kind).toBe("managed");    if (byServed.kind === "managed") {      expect(String(byServed.recipe.id)).toBe("phi4");      expect(byServed.resolvedModel).toBe("phi-4");    }  });   test("provider router resolves explicit external provider/model references", () => {    const resolved = parts.providers.resolve("anthropic/claude-3-7-sonnet");    expect(resolved.kind).toBe("external");    if (resolved.kind === "external") {      expect(resolved.provider).toBe("anthropic");      expect(resolved.resolvedModel).toBe("claude-3-7-sonnet");      expect(resolved.baseUrl).toContain("anthropic");    }  });   test("node registry rejects overlapping layer allocations for the same model key", () => {    parts.cluster.register("node-a", "10.0.0.1");    parts.cluster.register("node-b", "10.0.0.2");     const first = parts.cluster.setAllocation("node-a", "mixtral", 0, 15);    expect(first.startLayer).toBe(0);    expect(first.endLayer).toBe(15);     expect(() => parts.cluster.setAllocation("node-b", "mixtral", 10, 20)).toThrow(/overlaps existing/);     const nonOverlapping = parts.cluster.setAllocation("node-b", "mixtral", 16, 31);    expect(nonOverlapping.startLayer).toBe(16);    expect(nonOverlapping.endLayer).toBe(31);  });}); // ---------------------------------------------------------------------------// Integration tests (from idealized integration.test.mjs)// --------------------------------------------------------------------------- describe("kernel integration tests", () => {  let parts: TestKernelParts;  let kernel: ControllerKernel;   beforeEach(() => {    parts = createTestKernelParts();    kernel = new ControllerKernel(parts);  });   test("activating a new model aborts in-flight runs for the previous active local model", async () => {    const phi4 = makeRecipe({ id: "phi4", name: "Phi-4", served_model_name: "phi-4", model_path: "/models/phi-4", backend: "vllm" });    const qwen3 = makeRecipe({ id: "qwen3", name: "Qwen-3", served_model_name: "qwen-3", model_path: "/models/qwen-3", backend: "sglang" });    kernel.registerRecipe(phi4);    kernel.registerRecipe(qwen3);     const session = kernel.createSession("switch test");    await kernel.activateModel("phi4");     const run = parts.chats.startRun({      sessionId: session.id,      requestedModel: "phi4",      resolvedModel: "phi-4",      provider: "local",      status: RUN_STATUS.RUNNING,    });     await kernel.activateModel("qwen3");     const updatedRun = parts.chats.getRun(run.id);    expect(updatedRun?.status).toBe(RUN_STATUS.ABORTED);     const eventTypes = parts.events.list().map((e) => e.type);    expect(eventTypes).toContain(EVENT_TYPE.MODEL_SWITCHED);    expect(eventTypes).toContain(EVENT_TYPE.CHAT_RUN_ABORTED);  });   test("proxy completion with switch_on_request activates the requested managed model", async () => {    const phi4 = makeRecipe({ id: "phi4", name: "Phi-4", served_model_name: "phi-4", model_path: "/models/phi-4", backend: "vllm" });    kernel.registerRecipe(phi4);     const session = kernel.createSession("chat test");    const result = await kernel.proxyCompletion({      sessionId: session.id,      model: "phi4",      activationPolicy: "switch_on_request",      messages: [{ role: "user", content: "Explain KV cache eviction simply." }],    });     expect(result.resolution.kind).toBe("managed");    expect(result.completion.model).toBe("phi-4");    expect(kernel.getActiveRuntime()?.recipeId).toBe("phi4");    expect(result.run?.status).toBe(RUN_STATUS.COMPLETED);     const messages = parts.chats.listMessages(session.id);    expect(messages.length).toBe(1); // user message appended by kernel     const snapshot = parts.metrics.snapshot();    expect(snapshot.requestsTotal).toBe(1);  });   test("load_if_idle keeps the current active model when a different managed model is requested", async () => {    const phi4 = makeRecipe({ id: "phi4", name: "Phi-4", served_model_name: "phi-4", model_path: "/models/phi-4", backend: "vllm" });    const qwen3 = makeRecipe({ id: "qwen3", name: "Qwen-3", served_model_name: "qwen-3", model_path: "/models/qwen-3", backend: "sglang" });    kernel.registerRecipe(phi4);    kernel.registerRecipe(qwen3);     await kernel.activateModel("phi4");     const result = await kernel.proxyCompletion({      model: "qwen3",      activationPolicy: "load_if_idle",      messages: [{ role: "user", content: "hello" }],    });     expect(result.resolution.kind).toBe("managed");    expect(result.resolution.resolvedModel).toBe("phi-4");    expect(kernel.getActiveRuntime()?.recipeId).toBe("phi4");  });}); // ---------------------------------------------------------------------------// End-to-end test (from idealized e2e.test.mjs)// --------------------------------------------------------------------------- describe("kernel e2e test", () => {  test("full control-plane scenario: recipes, proxy, metrics, jobs, downloads, distributed", async () => {    const parts = createTestKernelParts();    const kernel = new ControllerKernel(parts);     const phi4 = makeRecipe({ id: "phi4", name: "Phi-4", served_model_name: "phi-4", model_path: "/models/phi-4", backend: "vllm", port: 8000 });    const llama3 = makeRecipe({ id: "llama3", name: "Llama-3.3", served_model_name: "llama-3.3", model_path: "/models/llama-3.3", backend: "llamacpp", port: 8081 });    kernel.registerRecipe(phi4);    kernel.registerRecipe(llama3);     const session = kernel.createSession("e2e session");     const download = await kernel.startDownload("phi4", "/data/models/phi4");    expect(download.status).toBe("downloading"); // kernel starts download, doesn't auto-complete     const job = await kernel.createJob("voice_assistant_turn", { transcript: "What is active right now?" });    expect(job.status).toBe("running"); // kernel starts job     const nodeA = kernel.registerNode("node-a", "10.1.0.10", { gpu: "L40S" });    const nodeB = kernel.registerNode("node-b", "10.1.0.11", { gpu: "L40S" });    kernel.heartbeatNode(nodeA.nodeId);    kernel.heartbeatNode(nodeB.nodeId);    const allocation = kernel.setAllocation("node-a", "phi4", 0, 31);    expect(allocation.nodeId).toBe("node-a");     const completionResult = await kernel.proxyCompletion({      sessionId: session.id,      model: "phi4",      activationPolicy: "switch_on_request",      messages: [        { role: "system", content: "Answer tersely." },        { role: "user", content: "Which model is serving me?" },      ],    });     expect(completionResult.completion.model).toBe("phi-4");    expect(kernel.getActiveRuntime()?.recipeId).toBe("phi4");     const externalResult = await kernel.proxyCompletion({      sessionId: session.id,      model: "anthropic/claude-3-7-sonnet",      messages: [{ role: "user", content: "Say hello." }],    });    expect(externalResult.resolution.kind).toBe("external");    expect(externalResult.completion.provider).toBe("anthropic");     const snapshot = kernel.snapshot();    expect(snapshot.activeRuntime?.recipeId).toBe("phi4");    expect(snapshot.metrics.requestsTotal).toBe(2);    expect(snapshot.jobs.length).toBe(1);    expect(snapshot.downloads.length).toBe(1);    expect(snapshot.nodes.length).toBe(2);    expect(snapshot.allocations.length).toBe(1);     const eventTypes = new Set(snapshot.events.map((e) => e.type));    expect(eventTypes.has(EVENT_TYPE.RECIPE_UPSERTED)).toBe(true);    expect(eventTypes.has(EVENT_TYPE.DOWNLOAD_UPDATED)).toBe(true);    expect(eventTypes.has(EVENT_TYPE.JOB_UPDATED)).toBe(true);    expect(eventTypes.has(EVENT_TYPE.NODE_UPDATED)).toBe(true);    expect(eventTypes.has(EVENT_TYPE.TOPOLOGY_UPDATED)).toBe(true);    expect(eventTypes.has(EVENT_TYPE.PROXY_COMPLETED)).toBe(true);     const evicted = await kernel.evictActiveModel();    expect(evicted?.recipeId).toBe("phi4");    expect(kernel.getActiveRuntime()).toBeUndefined();  });});