testdriverai 7.2.64 → 7.2.69

package/agent/index.js

@@ -36,6 +36,7 @@ const { createOutputs } = require("./lib/outputs.js");
 const isValidVersion = require("./lib/valid-version.js");
 const { events, createEmitter } = require("./events.js");
 const { createDebuggerProcess } = require("./lib/debugger.js");
+const logger = require("./lib/logger.js");
 let debuggerProcess = null; // single debugger process for all instances. otherwise they'll fight over ports. this should be in `web` anyway
 let debuggerStarted = false;
 
@@ -987,9 +988,9 @@ commands:
       const generateDir = path.join(this.workingDir, "testdriver", "generate");
       if (!fs.existsSync(generateDir)) {
         fs.mkdirSync(generateDir);
-        console.log("Created generate directory:", generateDir);
+        logger.log("Created generate directory:", generateDir);
       } else {
-        console.log("Generate directory already exists:", generateDir);
+        logger.log("Generate directory already exists:", generateDir);
       }
 
       let list = testPrompt.steps;
@@ -2016,7 +2017,7 @@ ${regression}
           "/vnc_lite.html?token=V3b8wG9";
       } else {
         // If we don't have URL or IP, we can't render
-        console.warn("renderSandbox: Missing URL and IP in instance", instance);
+        logger.warn("renderSandbox: Missing URL and IP in instance", instance);
         return;
       }
 

package/agent/interface.js

@@ -1,6 +1,4 @@
-const path = require("path");
 const { Args, Flags } = require("@oclif/core");
-const { events } = require("./events.js");
 
 /**
  * Creates command definitions using oclif format as the single source of truth
@@ -8,265 +6,27 @@ const { events } = require("./events.js");
  * @returns {Object} Command definitions object in oclif format
  */
 function createCommandDefinitions(agent) {
-  const normalizeFilePath = (file) => {
-    if (!file) {
-      file = "testdriver/testdriver.yaml";
-    }
-
-    file = path.join(agent.workingDir, file);
-    if (!file.endsWith(".yaml") && !file.endsWith(".yml")) {
-      file += ".yaml";
-    }
-
-    return file;
-  };
-
   return {
-    run: {
-      description: "Run a test file",
-      args: {
-        file: Args.string({
-          description: "Test file to run",
-          default: "testdriver/testdriver.yaml",
-          required: false,
-        }),
-      },
-      flags: {
-        heal: Flags.boolean({
-          description: "Enable automatic error recovery mode",
-          default: false,
-        }),
-        write: Flags.boolean({
-          description: "Save AI modifications to the test file",
-          default: false,
-        }),
-        headless: Flags.boolean({
-          description: "Run in headless mode (no GUI)",
-          default: false,
-        }),
-        new: Flags.boolean({
-          description:
-            "Create a new sandbox instead of reconnecting to an existing one",
-          default: false,
-        }),
-        "sandbox-ami": Flags.string({
-          description: "Specify AMI ID for sandbox instance (e.g., ami-1234)",
-        }),
-        "sandbox-instance": Flags.string({
-          description: "Specify EC2 instance type for sandbox (e.g., i3.metal)",
-        }),
-        ip: Flags.string({
-          description:
-            "Connect directly to a sandbox at the specified IP address",
-        }),
-        summary: Flags.string({
-          description: "Specify output file for summarize results",
-        }),
-        junit: Flags.string({
-          description: "Generate JUnit XML test report to specified file",
-          default: false,
-        }),
-        os: Flags.string({
-          description: "Operating system for the sandbox (windows or linux)",
-          options: ["windows", "linux"],
-          default: "linux",
-        }),
-      },
-      handler: async (args, flags) => {
-        // Use --path flag if provided, otherwise fall back to args.file
-        const file = normalizeFilePath(args.file);
-        const testStartTime = Date.now();
-
-        try {
-          await agent.runLifecycle("prerun");
-          // When run() is called through run.js CLI command, shouldExit should be true
-          const shouldExit = agent.cliArgs?.command === "run";
-          await agent.run(file, flags.write, shouldExit);
-
-          const testEndTime = Date.now();
-          const testDuration = testEndTime - testStartTime;
-
-          // Emit test success event for the entire test execution
-          agent.emitter.emit(events.test.success, {
-            filePath: file,
-            duration: testDuration,
-            timestamp: testEndTime,
-          });
-        } catch (error) {
-          const testEndTime = Date.now();
-          const testDuration = testEndTime - testStartTime;
-
-          // Emit test error event for the entire test execution
-          agent.emitter.emit(events.test.error, {
-            filePath: file,
-            error: error.message,
-            duration: testDuration,
-            timestamp: testEndTime,
-          });
-
-          throw error; // Re-throw to maintain existing error handling
-        }
-      },
-    },
-
-    edit: {
-      description: "Edit a test file interactively",
-      args: {
-        file: Args.string({
-          description: "Test file to edit",
-          default: "testdriver/testdriver.yaml",
-          required: false,
-        }),
-      },
-      flags: {
-        heal: Flags.boolean({
-          description: "Enable automatic error recovery mode",
-          default: false,
-        }),
-        headless: Flags.boolean({
-          description: "Run in headless mode",
-          default: false,
-        }),
-        new: Flags.boolean({
-          description:
-            "Create a new sandbox instead of reconnecting to an existing one",
-          default: false,
-        }),
-        "sandbox-ami": Flags.string({
-          description: "Specify AMI ID for sandbox instance (e.g., ami-1234)",
-        }),
-        "sandbox-instance": Flags.string({
-          description: "Specify EC2 instance type for sandbox (e.g., i3.metal)",
-        }),
-        ip: Flags.string({
-          description:
-            "Connect directly to a sandbox at the specified IP address",
-        }),
-        summary: Flags.string({
-          description: "Specify output file for summarize results",
-        }),
-        os: Flags.string({
-          description: "Operating system for the sandbox (windows or linux)",
-          options: ["windows", "linux"],
-          default: "windows",
-        }),
-      },
-      handler: async () => {
-        // Edit mode is handled by the CLI interface via factory.js
-        // This handler should not be called directly
-        throw new Error("Edit mode should be handled by CLI interface");
-      },
-    },
-
-    // Interactive commands that can be used within edit mode
-    explore: {
-      description: "Explore and interact with the current environment",
-      args: {
-        prompt: Args.string({
-          description: "What you want to explore or do",
-          required: false,
-        }),
-      },
-      flags: {},
-      handler: async (args) => {
-        await agent.exploratoryLoop(args.prompt || "", false, true, true);
-      },
-    },
-
-    save: {
-      description: "Save the current test script",
-      args: {
-        filename: Args.string({
-          description: "Optional filename to save to",
-          required: false,
-        }),
-      },
-      flags: {},
-      handler: async (args) => {
-        await agent.save(args.filename);
-      },
-    },
-
-    exit: {
-      description: "Exit the TestDriver agent",
-      args: {},
-      flags: {},
-      handler: async () => {
-        await agent.exit(false);
-      },
-    },
-
-    help: {
-      description: "Show help information",
-      args: {},
-      flags: {},
-      handler: async () => {
-        agent.showHelp();
-      },
-    },
-
-    version: {
-      description: "Show version information",
+    init: {
+      description: "Initialize a new TestDriver project with Vitest SDK examples",
       args: {},
       flags: {},
       handler: async () => {
-        const packageJson = require("../package.json");
-        console.log(`TestDriver.ai v${packageJson.version}`);
-      },
-    },
-
-    generate: {
-      description: "Generate test files based on current screen state",
-      args: {
-        prompt: Args.string({
-          description: "Multi-line text prompt describing what to generate",
-          required: false,
-        }),
-      },
-      flags: {
-        count: Flags.integer({
-          description: "Number of test files to generate",
-          default: 3,
-        }),
-        headless: Flags.boolean({
-          description: "Run in headless mode (no GUI)",
-          default: false,
-        }),
-        new: Flags.boolean({
-          description:
-            "Create a new sandbox instead of reconnecting to an existing one",
-          default: false,
-        }),
-        "sandbox-ami": Flags.string({
-          description: "Specify AMI ID for sandbox instance (e.g., ami-1234)",
-        }),
-        "sandbox-instance": Flags.string({
-          description: "Specify EC2 instance type for sandbox (e.g., i3.metal)",
-        }),
-        ip: Flags.string({
-          description:
-            "Connect directly to a sandbox at the specified IP address",
-        }),
-        os: Flags.string({
-          description: "Operating system for the sandbox (windows or linux)",
-          options: ["windows", "linux"],
-          default: "linux",
-        }),
-      },
-      handler: async (args, flags) => {
-        // Call generate with the count and prompt
-        await agent.generate(flags.count || 3, args.prompt);
+        // This handler is special - it doesn't need an agent instance
+        // It just scaffolds files, so it will be handled by the CLI command
+        throw new Error("Init mode should be handled by CLI interface");
       },
     },
 
-    init: {
-      description: "Initialize a new TestDriver project with Vitest SDK examples",
+    "setup": {
+      description:
+        "Set up TestDriver skills, agents, and MCP server for Claude Code",
       args: {},
       flags: {},
       handler: async () => {
-        // This handler is special - it doesn't need an agent instance
-        // It just scaffolds files, so it will be handled by the CLI command
-        throw new Error("Init mode should be handled by CLI interface");
+        throw new Error(
+          "setup should be handled by CLI interface",
+        );
       },
     },
   };

package/agent/lib/debugger-server.js

@@ -3,6 +3,7 @@ const http = require("http");
 const path = require("path");
 const fs = require("fs");
 const { eventsArray } = require("../events.js");
+const logger = require("./logger");
 
 let server = null;
 let wss = null;
@@ -131,7 +132,7 @@ function stopDebugger() {
   }
 
   clients.clear();
-  console.log("Debugger server stopped");
+  logger.log("Debugger server stopped");
 }
 
 module.exports = {

package/agent/lib/logger.js

@@ -0,0 +1,56 @@
+/**
+ * Logger utility for TestDriver
+ * 
+ * By default, outputs to stdout (console.log).
+ * When TD_STDIO=stderr is set, outputs to stderr (console.error).
+ * This is necessary for MCP servers which use stdout exclusively for JSON-RPC.
+ */
+
+const useStderr = process.env.TD_STDIO === 'stderr';
+
+/**
+ * Log a message - uses stdout by default, stderr if TD_STDIO=stderr
+ * @param {...any} args - Arguments to log
+ */
+function log(...args) {
+  if (useStderr) {
+    console.error(...args);
+  } else {
+    console.log(...args);
+  }
+}
+
+/**
+ * Log an error - always uses stderr
+ * @param {...any} args - Arguments to log
+ */
+function error(...args) {
+  console.error(...args);
+}
+
+/**
+ * Log a warning - uses stdout by default, stderr if TD_STDIO=stderr
+ * @param {...any} args - Arguments to log
+ */
+function warn(...args) {
+  if (useStderr) {
+    console.error(...args);
+  } else {
+    console.warn(...args);
+  }
+}
+
+/**
+ * Check if logger is configured to use stderr
+ * @returns {boolean}
+ */
+function isStderrMode() {
+  return useStderr;
+}
+
+module.exports = {
+  log,
+  error,
+  warn,
+  isStderrMode,
+};

package/agent/lib/sandbox.js

@@ -1,6 +1,7 @@
 const WebSocket = require("ws");
 const crypto = require("crypto");
 const { events } = require("../events");
+const logger = require("./logger");
 
 /**
  * Generate Sentry trace headers for distributed tracing
@@ -142,11 +143,9 @@ const createSandbox = (emitter, analytics, sessionInstance) => {
         // Log and store the Sentry trace ID for debugging
         if (reply.traceId) {
           this.traceId = reply.traceId;
-          console.log("");
-          console.log(`🔗 View Trace:`);
-          console.log(
-            `https://testdriver.sentry.io/explore/traces/trace/${reply.traceId}`,
-          );
+          logger.log('');
+          logger.log(`🔗 View Trace:`);
+          logger.log(`https://testdriver.sentry.io/explore/traces/trace/${reply.traceId}`);
         }
 
         emitter.emit(events.sandbox.authenticated, { traceId: reply.traceId });
@@ -239,8 +238,8 @@ const createSandbox = (emitter, analytics, sessionInstance) => {
         });
 
         this.socket.on("error", (err) => {
-          console.log("Socket Error");
-          err && console.log(err);
+          logger.log("Socket Error");
+          err && logger.log(err);
           clearInterval(this.heartbeat);
           emitter.emit(events.error.sandbox, err);
           this.apiSocketConnected = false;