penguin-harness/packages/web/e2e/subagent.spec.mjs

/**
 * The subagent (run_subagent) card survives a **reload**: the parent Trace only stores a single
 * session_meta pointer for the child session; the server's readMessages expands the child Trace
 * from that pointer and fills in the origin chain, and the frontend reattaches the child session
 * to the tool card from that. Also verifies the child session's title is generated by the model
 * from **the child session's own conversation** (the mock only replies with "Subagent TODO
 * summary" when the title request sees the child's own answer), and that discovery-based
 * backfill must not collapse it to NULL.
 *
 * Standalone spec: shares one server with chat.spec.mjs, so it registers its own user here
 * (not the first user; registration auto-provisions a `project-<8hex>`), independent of
 * chat.spec's execution order.
 */
import { test, expect } from "@playwright/test";
import { provisionAndLogin } from "./auth.mjs";

const BASE = process.env.BASE_URL;
const MOCK = process.env.MOCK_URL;
const U = "subuser";
const P = "password123";

test("subagent card survives a reload; child session title is generated from its own conversation", async ({
  page,
}) => {
  await provisionAndLogin(page.request, U, P);

  // Account creation auto-provisions a default Project; must use it (the frontend selects it
  // too) — don't create another one.
  const projects = await (await page.request.get(`${BASE}/api/projects`)).json();
  expect(projects.projects, "auto-provisioned project").toHaveLength(1);
  const projectId = projects.projects[0].projectId;

  const put = await page.request.put(`${BASE}/api/projects/${projectId}/models`, {
    data: {
      defaultModel: { provider: "custom", modelId: "claude-4-8" },
      models: [
        {
          provider: "custom",
          modelId: "claude-4-8",
          apiKey: "sk-mock",
          baseUrl: MOCK,
          contextWindow: 200000,
        },
      ],
    },
  });
  expect(put.ok(), "put models").toBeTruthy();

  const agentId = "default_agent";
  // Approval defaults to allow-all: child sessions inherit the parent's approval mode, no
  // manual approval needed.
  const sessRes = await page.request.post(
    `${BASE}/api/projects/${projectId}/agents/${agentId}/sessions`,
    { data: { provider: "custom", modelId: "claude-4-8" } },
  );
  expect(sessRes.ok(), `create session: ${await sessRes.text()}`).toBeTruthy();
  const sess = await sessRes.json();
  const sessionId = sess.session.sessionId;

  await page.goto(`${BASE}/chat/${sessionId}`);
  const ta = page.getByPlaceholder(/输入消息/);
  await ta.waitFor();
  await ta.fill("run a subagent");
  await page.getByRole("button", { name: "发送" }).click();

  // Parent's final answer (turn 2): the whole turn has ended. Nothing changes with streaming
  // after this, so assertions are deterministic (turn end collapses the "Reasoning & tools"
  // group, and child session end collapses the child session card).
  await expect(page.getByText("Command finished; the result looks as expected.")).toBeVisible();

  /** Expand the "Reasoning & tools" group → expand the child session card → child session body visible. */
  const openSubagent = async () => {
    await page
      .getByRole("button", { name: /运行完毕/ })
      .first()
      .click();
    await expect(page.getByText("run_subagent").first()).toBeVisible();
    const card = page.getByRole("button", { name: /子会话/ }).first();
    await expect(card).toBeVisible();
    await card.click();
    await expect(page.getByText("Subagent report: 3 TODOs").first()).toBeVisible();
  };

  // Under live streaming, the child session card is nested inside the run_subagent tool card.
  await openSubagent();

  // --- DOM order: the subagent expansion renders BELOW the tool's own content ---
  // Expand the run_subagent tool card itself so its arguments block is in the DOM, then assert
  // the child-session card comes after it in document order (it used to render between the tool
  // header and the expanded arguments/output).
  const toolRow = page.locator("button[aria-expanded]").filter({ hasText: "run_subagent" }).first();
  await toolRow.click();
  await expect(toolRow).toHaveAttribute("aria-expanded", "true");
  const argsPre = page.locator("pre", { hasText: '"prompt"' }).first();
  await expect(argsPre).toBeVisible();
  const subagentBtn = page.getByRole("button", { name: /子会话/ }).first();
  const cardFollowsArgs = await argsPre.evaluate(
    (pre, card) => Boolean(pre.compareDocumentPosition(card) & Node.DOCUMENT_POSITION_FOLLOWING),
    await subagentBtn.elementHandle(),
  );
  expect(cardFollowsArgs, "subagent expansion renders below the tool arguments/output").toBe(true);

  // --- Must still hold after reload (the fix's goal: parent Trace's child session pointer + server-side expansion) ---
  await page.reload();
  await openSubagent();

  // --- Child session title: generated by the model from the child session's own conversation (async, poll until persisted) ---
  const childOf = async () => {
    const list = await (
      await page.request.get(`${BASE}/api/projects/${projectId}/agents/${agentId}/sessions`)
    ).json();
    return list.sessions.find((s) => s.sessionId !== sessionId) ?? null;
  };
  await expect
    .poll(async () => (await childOf())?.title ?? null, { timeout: 10000 })
    .toBe("Subagent TODO summary");
  const child = await childOf();

  // --- The child's OWN Trace session_meta records source=subagent: written by core's spawn
  // site, the single source of truth (the server's registration fallback cannot mask this —
  // it never writes the child Trace), and what the derived list source ultimately rests on. ---
  const childMessages = await (
    await page.request.get(`${BASE}/api/sessions/${child.sessionId}/messages`)
  ).json();
  const childMeta = childMessages.messages.find(
    (m) => m.type === "session_meta" && !m.origin?.length,
  );
  expect(childMeta, "child trace session_meta").toBeTruthy();
  expect(childMeta.payload.source).toBe("subagent");

  // --- Sidebar: the child session (source=subagent) nests inside the collapsed "Subagents"
  // folder (per-origin folders sit parallel to "Archived" within the same temp-workspace
  // group). Reload first so the sidebar list carries the persisted title/source, and the
  // folder is back to its default collapsed state. ---
  await page.reload();
  const sidebar = page.getByRole("complementary");
  const subagentFolder = sidebar.getByRole("button", { name: "子智能体（1）" });
  await expect(subagentFolder, "collapsed Subagents folder").toBeVisible();
  // Collapsed by default: the child row is not rendered until the folder is expanded.
  await expect(sidebar.getByText("Subagent TODO summary")).toHaveCount(0);
  await subagentFolder.click();
  await expect(sidebar.getByText("Subagent TODO summary")).toBeVisible();
});