Testing Agents

Objective: test behavior without network calls or local shell execution.

import assert from "node:assert/strict";
import { createHarnessSessionStore } from "@harness-kernel/core/runner";
import { MemorySessionStorage } from "@harness-kernel/core/runner/storage";
import { NoopSandbox } from "@harness-kernel/core/runner/sandbox";
import { agent } from "../src/agent.js";
import { EchoProvider } from "./support/echo-provider.js";

for (const mode of agent.modes) mode.toolApproval = "deny";

const store = await createHarnessSessionStore({
  agent: { definition: agent },
  providers: [new EchoProvider()],
  defaultModel: "echo/basic",
  storage: new MemorySessionStorage(),
  sandbox: new NoopSandbox(),
});

try {
  const result = await store.send("test-session", "hello");
  assert.equal(result.agentKey, "my-agent");
  assert.match(result.answer, /hello/);
} finally {
  await store.close();
}

Use session.waitForEvent() to assert lifecycle behavior:

const session = await store.getOrCreate("events");
const run = session.stream("trigger a tool");
await session.waitForEvent(TurnEndEvent, { timeoutMs: 5000 });
await run.result;

Fatal and recoverable errors can be tested through their canonical codes:

await assert.rejects(session.send("fail"));
assert.equal(session.getStatus().lastError?.code, "model.failed");

const failed = session.getEvents({ event: RunFailedEvent });
assert.equal(failed.at(-1)?.payload.error.code, "model.failed");

For recoverable tool failures, assert AgentToolResult.isError and the tool error code instead of expecting the run to fail.

For package-level validation, this repo also runs an external-consumer smoke test with npm pack. That catches missing package exports and subpath regressions:

pnpm build
pnpm test:consumer:packed

Boundary note: tests can replace all runtime infrastructure while keeping the same agent package.

API: Sessions and Model Providers.