macro-agent 0.2.5 → 0.2.6

package/src/agent/__tests__/agent-manager-v2.permission-interception.test.ts

@@ -1,296 +0,0 @@
-/**
- * Unit tests for AgentManagerV2's prompt iterator permission overlay
- * interception (Phase 3). Pins the contract that:
- *
- *   - When a `permission_request` session update fires AND a per-agent
- *     overlay is set, the iterator intercepts, evaluates against the
- *     overlay, and answers via `respondToPermission`. The update is
- *     NOT yielded to the consumer.
- *   - When NO overlay is set, the update IS yielded through (chat
- *     surfaces' UI permission dialogs depend on this).
- *   - Overlay errors fail closed (deny).
- *   - Non-permission_request updates pass through unchanged.
- */
-
-import {
-  describe,
-  it,
-  expect,
-  vi,
-  beforeEach,
-  afterEach,
-} from "vitest";
-
-const createSessionSpy = vi.fn();
-let mockSessionPrompt: ((message: string) => AsyncIterable<unknown>) | null =
-  null;
-const respondToPermissionSpy = vi.fn();
-
-vi.mock("acp-factory", () => ({
-  AgentFactory: {
-    spawn: vi.fn().mockResolvedValue({
-      createSession: (...args: unknown[]) => {
-        createSessionSpy(...args);
-        return Promise.resolve({
-          id: "provider-session-1",
-          prompt: vi.fn().mockImplementation(
-            (message: string) => mockSessionPrompt?.(message) ?? emptyIterable(),
-          ),
-          respondToPermission: respondToPermissionSpy,
-        });
-      },
-      close: vi.fn().mockResolvedValue(undefined),
-      isRunning: vi.fn().mockReturnValue(true),
-    }),
-  },
-}));
-
-import { createAgentManagerV2 } from "../agent-manager-v2.js";
-import { AgentStore } from "../agent-store.js";
-import type { AgentManager } from "../agent-manager.js";
-import type { InboxAdapter, TasksAdapter } from "../../adapters/types.js";
-import {
-  setPermissionOverlay,
-  clearPermissionOverlay,
-} from "../../dispatch/permission-overlay.js";
-
-function emptyIterable<T>(): AsyncIterable<T> {
-  return {
-    [Symbol.asyncIterator]: () => ({
-      next: () => Promise.resolve({ done: true, value: undefined as never }),
-    }),
-  };
-}
-
-function fromArray<T>(items: T[]): AsyncIterable<T> {
-  let i = 0;
-  return {
-    [Symbol.asyncIterator]: () => ({
-      next: () => {
-        if (i < items.length) {
-          return Promise.resolve({ done: false, value: items[i++]! });
-        }
-        return Promise.resolve({ done: true, value: undefined as never });
-      },
-    }),
-  };
-}
-
-function permRequest(opts: {
-  requestId: string;
-  toolTitle: string;
-  toolKind?: string;
-  rawInput?: unknown;
-}): unknown {
-  return {
-    sessionUpdate: "permission_request",
-    requestId: opts.requestId,
-    toolCall: {
-      toolCallId: "tc-1",
-      title: opts.toolTitle,
-      kind: opts.toolKind ?? "read",
-      status: "pending",
-      rawInput: opts.rawInput ?? {},
-    },
-    options: [
-      { kind: "allow_always", optionId: "allow_always" },
-      { kind: "allow_once", optionId: "allow" },
-      { kind: "reject_once", optionId: "reject" },
-    ],
-  };
-}
-
-function createMockInboxAdapter(): InboxAdapter {
-  return {
-    registerAgent: vi.fn().mockResolvedValue(undefined),
-    deregisterAgent: vi.fn().mockResolvedValue(undefined),
-    send: vi.fn().mockResolvedValue("msg-1"),
-    onDelivery: vi.fn(),
-    offDelivery: vi.fn(),
-    checkInbox: vi.fn().mockResolvedValue([]),
-    readThread: vi.fn().mockResolvedValue([]),
-    setSignalFilter: vi.fn(),
-    setEmissionValidator: vi.fn(),
-    socketPath: "/tmp/test-inbox.sock",
-    stop: vi.fn().mockResolvedValue(undefined),
-  } as unknown as InboxAdapter;
-}
-
-function createMockTasksAdapter(): TasksAdapter {
-  return {
-    createTask: vi.fn().mockResolvedValue("ot-task-1"),
-    assignTask: vi.fn().mockResolvedValue(undefined),
-    transitionTask: vi.fn().mockResolvedValue(undefined),
-    getTask: vi.fn().mockResolvedValue({ id: "t-1", title: "x", status: "open" }),
-    queryReady: vi.fn().mockResolvedValue([]),
-    listTasks: vi.fn().mockResolvedValue([]),
-    addBlocker: vi.fn().mockResolvedValue(undefined),
-    removeBlocker: vi.fn().mockResolvedValue(undefined),
-    claimTask: vi.fn().mockResolvedValue(null),
-    unclaimTask: vi.fn().mockResolvedValue(undefined),
-    listClaimable: vi.fn().mockResolvedValue([]),
-    connect: vi.fn().mockResolvedValue(undefined),
-    disconnect: vi.fn(),
-    connected: true,
-  } as unknown as TasksAdapter;
-}
-
-describe("AgentManagerV2 prompt iterator — permission overlay interception", () => {
-  let agentStore: AgentStore;
-  let inboxAdapter: InboxAdapter;
-  let tasksAdapter: TasksAdapter;
-  let manager: AgentManager;
-
-  beforeEach(() => {
-    createSessionSpy.mockClear();
-    respondToPermissionSpy.mockClear();
-    mockSessionPrompt = null;
-    agentStore = new AgentStore(":memory:");
-    inboxAdapter = createMockInboxAdapter();
-    tasksAdapter = createMockTasksAdapter();
-    manager = createAgentManagerV2(agentStore, inboxAdapter, tasksAdapter, {
-      defaultCwd: "/tmp/test",
-    });
-  });
-
-  afterEach(async () => {
-    await manager.close();
-    agentStore.close();
-  });
-
-  it("overlay set + deny rule matches → respondToPermission called with reject; update NOT yielded", async () => {
-    const spawned = await manager.spawn({ task: "Test", role: "worker" });
-    setPermissionOverlay(spawned.id, {
-      deny: ["Read(/tmp/secret.txt)"],
-    });
-    try {
-      mockSessionPrompt = () =>
-        fromArray([
-          permRequest({
-            requestId: "req-1",
-            toolTitle: "Read /tmp/secret.txt",
-            toolKind: "read",
-            rawInput: { file_path: "/tmp/secret.txt" },
-          }),
-        ]);
-
-      const yielded: unknown[] = [];
-      for await (const update of manager.prompt(spawned.id, "do it")) {
-        yielded.push(update);
-      }
-
-      expect(respondToPermissionSpy).toHaveBeenCalledTimes(1);
-      expect(respondToPermissionSpy).toHaveBeenCalledWith("req-1", "reject");
-      expect(yielded).toHaveLength(0);
-    } finally {
-      clearPermissionOverlay(spawned.id);
-    }
-  });
-
-  it("overlay set + no rule matches → respondToPermission called with allow; update NOT yielded", async () => {
-    const spawned = await manager.spawn({ task: "Test", role: "worker" });
-    setPermissionOverlay(spawned.id, {
-      deny: ["Read(/etc/*)"],
-    });
-    try {
-      mockSessionPrompt = () =>
-        fromArray([
-          permRequest({
-            requestId: "req-2",
-            toolTitle: "Read /tmp/safe.txt",
-            toolKind: "read",
-            rawInput: { file_path: "/tmp/safe.txt" },
-          }),
-        ]);
-
-      const yielded: unknown[] = [];
-      for await (const update of manager.prompt(spawned.id, "do it")) {
-        yielded.push(update);
-      }
-
-      expect(respondToPermissionSpy).toHaveBeenCalledTimes(1);
-      expect(respondToPermissionSpy).toHaveBeenCalledWith("req-2", "allow");
-      expect(yielded).toHaveLength(0);
-    } finally {
-      clearPermissionOverlay(spawned.id);
-    }
-  });
-
-  it("no overlay → permission_request IS yielded to consumer (chat path)", async () => {
-    const spawned = await manager.spawn({ task: "Test", role: "worker" });
-    // No overlay set on this agent.
-    mockSessionPrompt = () =>
-      fromArray([
-        permRequest({
-          requestId: "req-3",
-          toolTitle: "Read /tmp/anything.txt",
-          toolKind: "read",
-          rawInput: { file_path: "/tmp/anything.txt" },
-        }),
-      ]);
-
-    const yielded: Array<{ sessionUpdate?: string; requestId?: string }> = [];
-    for await (const update of manager.prompt(spawned.id, "do it")) {
-      yielded.push(update as { sessionUpdate?: string; requestId?: string });
-    }
-
-    expect(respondToPermissionSpy).not.toHaveBeenCalled();
-    expect(yielded).toHaveLength(1);
-    expect(yielded[0]?.sessionUpdate).toBe("permission_request");
-    expect(yielded[0]?.requestId).toBe("req-3");
-  });
-
-  it("non-permission_request updates always pass through to consumer", async () => {
-    const spawned = await manager.spawn({ task: "Test", role: "worker" });
-    setPermissionOverlay(spawned.id, { deny: ["Read(/secret)"] });
-    try {
-      mockSessionPrompt = () =>
-        fromArray([
-          { sessionUpdate: "agent_message_chunk", text: "hello" },
-          { sessionUpdate: "tool_call", title: "Read /tmp/x" },
-          { sessionUpdate: "agent_thought_chunk", text: "thinking" },
-        ]);
-
-      const yielded: Array<{ sessionUpdate?: string }> = [];
-      for await (const update of manager.prompt(spawned.id, "do it")) {
-        yielded.push(update as { sessionUpdate?: string });
-      }
-
-      expect(respondToPermissionSpy).not.toHaveBeenCalled();
-      expect(yielded).toHaveLength(3);
-      expect(yielded.map((u) => u.sessionUpdate)).toEqual([
-        "agent_message_chunk",
-        "tool_call",
-        "agent_thought_chunk",
-      ]);
-    } finally {
-      clearPermissionOverlay(spawned.id);
-    }
-  });
-
-  it("MCP tool — overlay rule matches by canonical mcp__ name", async () => {
-    const spawned = await manager.spawn({ task: "Test", role: "worker" });
-    setPermissionOverlay(spawned.id, {
-      deny: ["mcp__agent-inbox__list_agents"],
-    });
-    try {
-      mockSessionPrompt = () =>
-        fromArray([
-          permRequest({
-            requestId: "req-mcp",
-            toolTitle: "mcp__agent-inbox__list_agents",
-            toolKind: "other",
-            rawInput: {},
-          }),
-        ]);
-
-      for await (const _ of manager.prompt(spawned.id, "do it")) {
-        /* drain */
-      }
-
-      expect(respondToPermissionSpy).toHaveBeenCalledWith("req-mcp", "reject");
-    } finally {
-      clearPermissionOverlay(spawned.id);
-    }
-  });
-});

package/src/agent/__tests__/agent-manager-v2.permissions.test.ts

@@ -1,233 +0,0 @@
-/**
- * Unit tests for AgentManagerV2's permission + fullAutonomous spawn-config
- * wiring. Pins the contract that `SpawnAgentOptions.permissions` is forwarded
- * to `agentMeta.claudeCode.options.settings.permissions` on the Claude Code
- * session, and that `ask` rules collapse based on `fullAutonomous`.
- */
-
-import { describe, it, expect, vi, beforeEach, afterEach } from "vitest";
-
-// IMPORTANT: vi.mock must run before agent-manager-v2 imports acp-factory.
-//
-// We intercept handle.createSession with a spy so each test can read back the
-// `agentMeta` argument and assert against it. Re-using a mock fn means the
-// first .mock.calls[0] always points to the most recent spawn() call.
-const createSessionSpy = vi.fn();
-
-vi.mock("acp-factory", () => ({
-  AgentFactory: {
-    spawn: vi.fn().mockResolvedValue({
-      createSession: (...args: unknown[]) => {
-        createSessionSpy(...args);
-        return Promise.resolve({
-          id: "provider-session-1",
-          prompt: vi.fn().mockReturnValue({
-            [Symbol.asyncIterator]: () => ({
-              next: () => Promise.resolve({ done: true, value: undefined }),
-            }),
-          }),
-          forkWithFlush: vi.fn().mockResolvedValue({ id: "forked-session-1" }),
-        });
-      },
-      loadSession: vi.fn().mockResolvedValue({
-        id: "loaded-session-1",
-        prompt: vi.fn().mockReturnValue({
-          [Symbol.asyncIterator]: () => ({
-            next: () => Promise.resolve({ done: true, value: undefined }),
-          }),
-        }),
-      }),
-      close: vi.fn().mockResolvedValue(undefined),
-      isRunning: vi.fn().mockReturnValue(true),
-    }),
-  },
-}));
-
-import { createAgentManagerV2 } from "../agent-manager-v2.js";
-import { AgentStore } from "../agent-store.js";
-import type { AgentManager } from "../agent-manager.js";
-import type { InboxAdapter, TasksAdapter } from "../../adapters/types.js";
-
-function createMockInboxAdapter(): InboxAdapter {
-  return {
-    registerAgent: vi.fn().mockResolvedValue(undefined),
-    deregisterAgent: vi.fn().mockResolvedValue(undefined),
-    send: vi.fn().mockResolvedValue("msg-1"),
-    onDelivery: vi.fn(),
-    offDelivery: vi.fn(),
-    checkInbox: vi.fn().mockResolvedValue([]),
-    readThread: vi.fn().mockResolvedValue([]),
-    setSignalFilter: vi.fn(),
-    setEmissionValidator: vi.fn(),
-    socketPath: "/tmp/test-inbox.sock",
-    stop: vi.fn().mockResolvedValue(undefined),
-  } as unknown as InboxAdapter;
-}
-
-function createMockTasksAdapter(): TasksAdapter {
-  return {
-    createTask: vi.fn().mockResolvedValue("ot-task-1"),
-    assignTask: vi.fn().mockResolvedValue(undefined),
-    transitionTask: vi.fn().mockResolvedValue(undefined),
-    getTask: vi.fn().mockResolvedValue({ id: "t-1", title: "test", status: "open" }),
-    queryReady: vi.fn().mockResolvedValue([]),
-    listTasks: vi.fn().mockResolvedValue([]),
-    addBlocker: vi.fn().mockResolvedValue(undefined),
-    removeBlocker: vi.fn().mockResolvedValue(undefined),
-    claimTask: vi.fn().mockResolvedValue(null),
-    unclaimTask: vi.fn().mockResolvedValue(undefined),
-    listClaimable: vi.fn().mockResolvedValue([]),
-    connect: vi.fn().mockResolvedValue(undefined),
-    disconnect: vi.fn(),
-    connected: true,
-  } as unknown as TasksAdapter;
-}
-
-/**
- * Read back the second argument to handle.createSession from the most recent
- * spawn() call. Returns the `agentMeta` field if present.
- */
-function readLastSpawnAgentMeta(): Record<string, unknown> | undefined {
-  const calls = createSessionSpy.mock.calls;
-  expect(calls.length).toBeGreaterThan(0);
-  const lastCall = calls[calls.length - 1];
-  // createSession(cwd, options)
-  const sessionOpts = lastCall[1] as Record<string, unknown> | undefined;
-  return sessionOpts?.agentMeta as Record<string, unknown> | undefined;
-}
-
-describe("AgentManagerV2 permissions + fullAutonomous spawn config", () => {
-  let agentStore: AgentStore;
-  let inboxAdapter: InboxAdapter;
-  let tasksAdapter: TasksAdapter;
-  let manager: AgentManager;
-
-  beforeEach(() => {
-    createSessionSpy.mockClear();
-    agentStore = new AgentStore(":memory:");
-    inboxAdapter = createMockInboxAdapter();
-    tasksAdapter = createMockTasksAdapter();
-    manager = createAgentManagerV2(agentStore, inboxAdapter, tasksAdapter, {
-      defaultCwd: "/tmp/test",
-    });
-  });
-
-  afterEach(async () => {
-    await manager.close();
-    agentStore.close();
-  });
-
-  it("permissions present + fullAutonomous: true → ask rules collapse into allow", async () => {
-    await manager.spawn({
-      task: "Test",
-      role: "worker",
-      permissions: {
-        deny: ["Bash(rm -rf:*)"],
-        ask: ["Write(*.env)"],
-      },
-      fullAutonomous: true,
-    });
-
-    const agentMeta = readLastSpawnAgentMeta();
-    expect(agentMeta).toBeDefined();
-    const cc = agentMeta!.claudeCode as {
-      options: { settings?: { permissions?: Record<string, string[]> } };
-    };
-    const perms = cc.options.settings?.permissions;
-    expect(perms).toBeDefined();
-    expect(perms!.deny).toContain("Bash(rm -rf:*)");
-    // fullAutonomous: true → ask collapses into allow
-    expect(perms!.allow).toContain("Write(*.env)");
-    // ask should NOT also be a literal field
-    expect(perms!.ask).toBeUndefined();
-  });
-
-  it("permissions present + fullAutonomous: false → ask rules collapse into deny", async () => {
-    await manager.spawn({
-      task: "Test",
-      role: "worker",
-      permissions: {
-        deny: ["Bash(rm -rf:*)"],
-        ask: ["Write(*.env)"],
-      },
-      fullAutonomous: false,
-    });
-
-    const agentMeta = readLastSpawnAgentMeta();
-    expect(agentMeta).toBeDefined();
-    const cc = agentMeta!.claudeCode as {
-      options: { settings?: { permissions?: Record<string, string[]> } };
-    };
-    const perms = cc.options.settings?.permissions;
-    expect(perms).toBeDefined();
-    // Both original deny AND collapsed ask end up in deny
-    expect(perms!.deny).toContain("Bash(rm -rf:*)");
-    expect(perms!.deny).toContain("Write(*.env)");
-    expect(perms!.allow).toBeUndefined();
-    expect(perms!.ask).toBeUndefined();
-  });
-
-  it("no permissions → no agentMeta.claudeCode.options.settings.permissions", async () => {
-    await manager.spawn({
-      task: "Test",
-      role: "worker",
-    });
-
-    const agentMeta = readLastSpawnAgentMeta();
-    // agentMeta should be entirely absent (no settingSources, no settings)
-    if (agentMeta) {
-      const cc = agentMeta.claudeCode as
-        | { options: { settings?: { permissions?: unknown } } }
-        | undefined;
-      expect(cc?.options?.settings?.permissions).toBeUndefined();
-    }
-  });
-
-  it("isolatedSettings: true + permissions → both settingSources=[] AND settings.permissions present", async () => {
-    await manager.spawn({
-      task: "Test",
-      role: "worker",
-      isolatedSettings: true,
-      permissions: { deny: ["Bash(rm -rf:*)"] },
-      fullAutonomous: true,
-    });
-
-    const agentMeta = readLastSpawnAgentMeta();
-    expect(agentMeta).toBeDefined();
-    const cc = agentMeta!.claudeCode as {
-      options: {
-        settingSources?: unknown[];
-        settings?: { permissions?: Record<string, string[]> };
-      };
-    };
-    expect(cc.options.settingSources).toEqual([]);
-    expect(cc.options.settings?.permissions?.deny).toContain("Bash(rm -rf:*)");
-  });
-
-  it("permissions with all empty arrays → no settings.permissions key written", async () => {
-    await manager.spawn({
-      task: "Test",
-      role: "worker",
-      permissions: { allow: [], deny: [], ask: [] },
-      fullAutonomous: true,
-    });
-
-    const agentMeta = readLastSpawnAgentMeta();
-    // The spawn handler always writes a `settings` object when permissions are
-    // truthy, but with empty allow/deny it should NOT include the rule keys —
-    // i.e. settings.permissions has no `allow` and no `deny`.
-    if (agentMeta) {
-      const cc = agentMeta.claudeCode as {
-        options: { settings?: { permissions?: Record<string, string[]> } };
-      };
-      const perms = cc.options.settings?.permissions;
-      // perms is either undefined or an empty object — both acceptable.
-      // Critically: no allow/deny/ask arrays leak through with empty content.
-      if (perms) {
-        expect(perms.allow).toBeUndefined();
-        expect(perms.deny).toBeUndefined();
-        expect(perms.ask).toBeUndefined();
-      }
-    }
-  });
-});