whale-code 6.5.11 → 6.6.0

package/dist/cli/services/__tests__/agent-loop-tools-maxturns.test.js

@@ -1,486 +0,0 @@
-/**
- * Tests for agent-loop.ts — max turns enforcement
- */
-
-import { describe, it, expect, vi, beforeEach } from "vitest";
-
-// ============================================================================
-// Mocks — must be declared before imports
-// ============================================================================
-
-vi.mock("../local-tools.js", () => ({
-  LOCAL_TOOL_DEFINITIONS: [{
-    name: "bash",
-    description: "Run shell commands",
-    input_schema: {
-      type: "object",
-      properties: {
-        command: {
-          type: "string"
-        }
-      }
-    }
-  }, {
-    name: "read_file",
-    description: "Read a file",
-    input_schema: {
-      type: "object",
-      properties: {
-        path: {
-          type: "string"
-        }
-      }
-    }
-  }],
-  executeLocalTool: vi.fn(async (name, _input) => ({
-    success: true,
-    output: `local-tool-result:${name}`
-  })),
-  isLocalTool: vi.fn(name => ["bash", "read_file"].includes(name))
-}));
-vi.mock("../interactive-tools.js", () => ({
-  INTERACTIVE_TOOL_DEFINITIONS: [{
-    name: "ask_user_question",
-    description: "Ask user",
-    input_schema: {
-      type: "object",
-      properties: {
-        question: {
-          type: "string"
-        }
-      }
-    }
-  }],
-  executeInteractiveTool: vi.fn(async () => ({
-    success: true,
-    output: "user answered"
-  }))
-}));
-vi.mock("../config-store.js", () => ({
-  loadConfig: vi.fn(() => ({
-    access_token: "test-token",
-    refresh_token: "test-refresh",
-    proxy_url: "https://proxy.test"
-  })),
-  resolveConfig: vi.fn(() => ({
-    storeId: "store-123",
-    proxyUrl: "https://proxy.test"
-  })),
-  getProxyUrl: vi.fn(() => "https://proxy.test")
-}));
-vi.mock("../auth-service.js", () => ({
-  getValidToken: vi.fn(async () => "test-jwt-token"),
-  refreshSession: vi.fn(async () => ({
-    success: true,
-    config: {
-      access_token: "refreshed-token"
-    }
-  }))
-}));
-vi.mock("../server-tools.js", () => ({
-  isServerTool: vi.fn(() => false),
-  loadServerToolDefinitions: vi.fn(async () => []),
-  executeServerTool: vi.fn(async () => ({
-    success: true,
-    output: "server-result"
-  })),
-  getServerStatus: vi.fn(() => ({
-    connected: false
-  })),
-  setServerToolContext: vi.fn()
-}));
-vi.mock("../telemetry.js", () => ({
-  nextTurn: vi.fn(() => 1),
-  createTurnContext: vi.fn(() => ({
-    traceId: "trace-1",
-    spanId: "span-1",
-    conversationId: "conv-1",
-    userId: "user-1",
-    userEmail: "test@test.com"
-  })),
-  logSpan: vi.fn(),
-  generateSpanId: vi.fn(() => "span-gen-1"),
-  flushCliSpans: vi.fn()
-}));
-vi.mock("../error-logger.js", () => ({
-  captureError: vi.fn(),
-  addBreadcrumb: vi.fn()
-}));
-vi.mock("../agent-events.js", () => ({
-  AgentEventEmitter: vi.fn(),
-  setGlobalEmitter: vi.fn(),
-  clearGlobalEmitter: vi.fn()
-}));
-vi.mock("../mcp-client.js", () => ({
-  mcpClientManager: {
-    getTools: vi.fn(() => []),
-    isMcpTool: vi.fn(() => false),
-    callTool: vi.fn(async () => ({
-      success: true,
-      output: "mcp-result"
-    }))
-  }
-}));
-vi.mock("../hooks.js", () => ({
-  loadHooks: vi.fn(() => []),
-  runBeforeToolHook: vi.fn(async () => ({
-    allow: true
-  })),
-  runAfterToolHook: vi.fn(async () => ({})),
-  runSessionHook: vi.fn(async () => {})
-}));
-vi.mock("../../../shared/agent-core.js", () => {
-  class MockLoopDetector {
-    resetTurn = vi.fn();
-    getSessionStats = vi.fn(() => ({
-      totalErrors: 0,
-      failedStrategies: 0,
-      consecutiveFailedTurns: 0
-    }));
-  }
-  return {
-    LoopDetector: MockLoopDetector,
-    COMPACTION_TRIGGER_TOKENS: 167_000,
-    COMPACTION_TOTAL_BUDGET: 2_000_000,
-    DEFAULT_SESSION_COST_BUDGET_USD: 25,
-    AGENT_DEFAULTS: {
-      maxTurns: 200,
-      temperature: 1.0,
-      maxConcurrentTools: 7,
-      enableDelegation: true,
-      maxDurationMs: 900_000,
-      maxDurationMsWorkflow: 120_000,
-      enableModelRouting: true,
-      subagentMaxTokens: 32000,
-      subagentMaxTurns: 25,
-      subagentTemperature: 1.0,
-      subagentMaxRetries: 3,
-      maxContinuations: 10,
-      subagentMaxContinuations: 5,
-      maxMessageChars: 100_000,
-      maxToolResultChars: 2048,
-      compactionTriggerTokens: 167_000,
-      compactionTotalBudget: 2_000_000,
-      sessionCostBudgetUsd: 25.00,
-      teammateMaxTotalTurns: 50,
-      teamTimeoutMs: 600_000,
-      workerStallMs: 180_000,
-      teamReviewMaxTokens: 1024,
-      apiTimeoutMs: 90_000,
-      toolTimeoutMs: 60_000,
-      toolResultTruncateChars: 500_000
-    },
-    emitCostWarningIfNeeded: vi.fn(),
-    getCompactionConfig: vi.fn(() => ({
-      isNative: true,
-      triggerTokens: 167_000,
-      totalBudget: 2_000_000
-    })),
-    getContextManagement: vi.fn(() => ({
-      betas: [],
-      config: {
-        edits: []
-      }
-    })),
-    resolveAgentLoopConfig: vi.fn(() => ({
-      maxTurns: 200,
-      temperature: 1.0,
-      maxTokens: 32000,
-      maxConcurrentTools: 7,
-      enableDelegation: true,
-      maxDurationMs: 900_000,
-      enableModelRouting: true,
-      maxMessageChars: 100_000,
-      contextOverrides: {},
-      subagentMaxTokens: 32000,
-      subagentMaxTurns: 25,
-      subagentTemperature: 1.0,
-      maxContinuations: 10,
-      subagentMaxContinuations: 5,
-      subagentMaxRetries: 3,
-      maxToolResultChars: 2048,
-      defaultsUsed: []
-    }))
-  };
-});
-vi.mock("../agent-config.js", () => ({
-  loadCLIAgentConfig: vi.fn(async () => null),
-  resetAgentConfigCache: vi.fn()
-}));
-const mockStreamResult = {
-  text: "Hello from Claude",
-  toolUseBlocks: [],
-  thinkingBlocks: [],
-  thinkingTokens: 0,
-  citations: [],
-  usage: {
-    inputTokens: 100,
-    outputTokens: 50,
-    cacheCreationTokens: 0,
-    cacheReadTokens: 0
-  },
-  compactionContent: null,
-  contextManagementApplied: false,
-  stopReason: "end_turn"
-};
-vi.mock("../../../shared/sse-parser.js", () => ({
-  parseSSEStream: vi.fn(() => async function* () {/* empty */}()),
-  processStreamWithCallbacks: vi.fn(async () => ({
-    ...mockStreamResult
-  })),
-  collectStreamResult: vi.fn(async () => ({
-    text: "summary",
-    usage: {
-      inputTokens: 10,
-      outputTokens: 5,
-      cacheCreationTokens: 0,
-      cacheReadTokens: 0
-    }
-  }))
-}));
-vi.mock("../../../shared/api-client.js", () => ({
-  callServerProxy: vi.fn(async () => new ReadableStream()),
-  callTranscribe: vi.fn(async () => "transcribed text"),
-  buildAPIRequest: vi.fn(() => ({
-    maxTokens: 16384,
-    betas: ["context-management-2025-06-27"],
-    contextManagement: {
-      edits: []
-    },
-    thinking: undefined
-  })),
-  buildSystemBlocks: vi.fn((prompt, cost) => [{
-    type: "text",
-    text: prompt,
-    cache_control: {
-      type: "ephemeral"
-    }
-  }, {
-    type: "text",
-    text: cost
-  }]),
-  prepareWithCaching: vi.fn((tools, messages) => ({
-    tools,
-    messages
-  })),
-  trimGeminiContext: vi.fn(msgs => msgs),
-  trimOpenAIContext: vi.fn(msgs => msgs),
-  requestProviderCompaction: vi.fn(async () => null)
-}));
-vi.mock("../../../shared/constants.js", () => ({
-  MODELS: {
-    HAIKU: "claude-haiku-4-5-20251001",
-    SONNET: "claude-sonnet-4-6",
-    OPUS: "claude-opus-4-6"
-  },
-  getProvider: vi.fn(() => "anthropic")
-}));
-vi.mock("../../../shared/tool-dispatch.js", () => ({
-  dispatchTools: vi.fn(async (blocks, executor, opts) => {
-    const results = [];
-    for (const block of blocks) {
-      const start = Date.now();
-      opts?.onStart?.(block.name, block.input);
-      const res = await executor(block.name, block.input);
-      opts?.onResult?.(block.name, res.success, res.output, Date.now() - start);
-      results.push({
-        type: "tool_result",
-        tool_use_id: block.id,
-        content: res.output,
-        is_error: !res.success
-      });
-    }
-    return {
-      results,
-      bailOut: false,
-      bailMessage: undefined
-    };
-  }),
-  buildAssistantContent: vi.fn(({
-    text,
-    toolUseBlocks
-  }) => {
-    const blocks = [];
-    if (text) blocks.push({
-      type: "text",
-      text
-    });
-    for (const t of toolUseBlocks) blocks.push({
-      type: "tool_use",
-      id: t.id,
-      name: t.name,
-      input: t.input
-    });
-    return blocks;
-  })
-}));
-vi.mock("../memory-manager.js", () => ({
-  loadMemory: vi.fn(async () => "memory content"),
-  addMemory: vi.fn(async () => {}),
-  removeMemory: vi.fn(async () => {}),
-  listMemories: vi.fn(async () => [])
-}));
-vi.mock("../git-context.js", () => ({
-  refreshGitContext: vi.fn(async () => "git context"),
-  resetGitContext: vi.fn()
-}));
-vi.mock("../claude-md-loader.js", () => ({
-  loadClaudeMd: vi.fn(async () => "claude md"),
-  reloadClaudeMd: vi.fn(async () => "claude md"),
-  resetClaudeMdCache: vi.fn()
-}));
-vi.mock("../permission-modes.js", () => ({
-  setPermissionMode: vi.fn(),
-  getPermissionMode: vi.fn(() => "normal"),
-  isToolAllowedByPermission: vi.fn(() => true)
-}));
-vi.mock("../model-manager.js", () => ({
-  setModel: vi.fn(),
-  setModelById: vi.fn(),
-  getModel: vi.fn(() => "claude-sonnet-4-6"),
-  getModelShortName: vi.fn(() => "sonnet"),
-  estimateCostUsd: vi.fn(() => 0.005)
-}));
-vi.mock("../session-persistence.js", () => ({
-  saveSession: vi.fn(async () => {}),
-  loadSession: vi.fn(async () => null),
-  listSessions: vi.fn(async () => []),
-  findLatestSessionForCwd: vi.fn(async () => null)
-}));
-vi.mock("../system-prompt.js", () => ({
-  buildSystemPrompt: vi.fn(async () => "You are a helpful AI assistant.")
-}));
-vi.mock("../background-processes.js", () => ({
-  listProcesses: vi.fn(() => []),
-  listBackgroundAgents: vi.fn(() => [])
-}));
-
-// ============================================================================
-// Imports — after all mocks
-// ============================================================================
-
-import { runAgentLoop, resetSessionState } from "../agent-loop.js";
-import { getValidToken } from "../auth-service.js";
-import { isToolAllowedByPermission, getPermissionMode } from "../permission-modes.js";
-import { estimateCostUsd } from "../model-manager.js";
-import { loadHooks } from "../hooks.js";
-import { processStreamWithCallbacks } from "../../../shared/sse-parser.js";
-import { callServerProxy } from "../../../shared/api-client.js";
-import { isLocalTool } from "../local-tools.js";
-import { isServerTool } from "../server-tools.js";
-import { mcpClientManager } from "../mcp-client.js";
-import { captureError } from "../error-logger.js";
-import { loadConfig } from "../config-store.js";
-function makeCallbacks(overrides = {}) {
-  return {
-    onText: vi.fn(),
-    onToolStart: vi.fn(),
-    onToolResult: vi.fn(),
-    onUsage: vi.fn(),
-    onDone: vi.fn(),
-    onError: vi.fn(),
-    onAutoCompact: vi.fn(),
-    ...overrides
-  };
-}
-
-// ============================================================================
-// TESTS — max turns
-// ============================================================================
-
-describe("agent-loop — budget enforcement & max turns", () => {
-  beforeEach(() => {
-    vi.clearAllMocks();
-    resetSessionState();
-    mockStreamResult.text = "Hello from Claude";
-    mockStreamResult.toolUseBlocks = [];
-    mockStreamResult.thinkingBlocks = [];
-    mockStreamResult.thinkingTokens = 0;
-    mockStreamResult.usage = {
-      inputTokens: 100,
-      outputTokens: 50,
-      cacheCreationTokens: 0,
-      cacheReadTokens: 0
-    };
-    mockStreamResult.compactionContent = null;
-    mockStreamResult.contextManagementApplied = false;
-    mockStreamResult.stopReason = "end_turn";
-    vi.mocked(processStreamWithCallbacks).mockReset();
-    vi.mocked(processStreamWithCallbacks).mockResolvedValue({
-      ...mockStreamResult
-    });
-    vi.mocked(callServerProxy).mockReset();
-    vi.mocked(callServerProxy).mockResolvedValue(new ReadableStream());
-    vi.mocked(captureError).mockReset();
-    vi.mocked(isToolAllowedByPermission).mockReturnValue(true);
-    vi.mocked(getPermissionMode).mockReturnValue("normal");
-    vi.mocked(isLocalTool).mockImplementation(name => ["bash", "read_file"].includes(name));
-    vi.mocked(isServerTool).mockReturnValue(false);
-    vi.mocked(mcpClientManager.isMcpTool).mockReturnValue(false);
-    vi.mocked(loadHooks).mockReturnValue([]);
-    vi.mocked(estimateCostUsd).mockReturnValue(0.005);
-    vi.mocked(getValidToken).mockResolvedValue("test-jwt-token");
-    vi.mocked(loadConfig).mockReturnValue({
-      access_token: "test-token",
-      refresh_token: "test-refresh",
-      proxy_url: "https://proxy.test"
-    });
-  });
-  it("stops loop when budget exceeded", async () => {
-    vi.mocked(estimateCostUsd).mockReturnValue(10.0);
-    let callCount = 0;
-    vi.mocked(processStreamWithCallbacks).mockImplementation(async () => {
-      callCount++;
-      if (callCount === 1) {
-        return {
-          ...mockStreamResult,
-          text: "",
-          toolUseBlocks: [{
-            id: "tu-1",
-            name: "bash",
-            input: {
-              command: "echo"
-            }
-          }],
-          stopReason: "tool_use"
-        };
-      }
-      return {
-        ...mockStreamResult
-      };
-    });
-    const callbacks = makeCallbacks();
-    await runAgentLoop({
-      message: "Expensive",
-      conversationHistory: [],
-      callbacks,
-      maxBudgetUsd: 0.001
-    });
-    expect(callbacks.onError).toHaveBeenCalled();
-    const errorMsg = vi.mocked(callbacks.onError).mock.calls[0][0];
-    expect(errorMsg).toContain("Budget exceeded");
-  });
-  it("respects maxTurns option", async () => {
-    vi.mocked(processStreamWithCallbacks).mockImplementation(async () => ({
-      ...mockStreamResult,
-      text: "processing",
-      toolUseBlocks: [{
-        id: "tu-1",
-        name: "bash",
-        input: {
-          command: "echo"
-        }
-      }],
-      stopReason: "tool_use"
-    }));
-    vi.mocked(estimateCostUsd).mockReturnValue(0);
-    const callbacks = makeCallbacks();
-    await runAgentLoop({
-      message: "Loop",
-      conversationHistory: [],
-      callbacks,
-      maxTurns: 3
-    });
-    expect(vi.mocked(processStreamWithCallbacks).mock.calls.length).toBe(3);
-  });
-});