testdriverai 6.0.16-canary.ea858d7.0 → 6.0.16

package/agent/events.js

@@ -1,6 +1,7 @@
 const { EventEmitter2 } = require("eventemitter2");
+const { censorSensitiveDataDeep } = require("./lib/censorship");
 
-// Factory function to create a new emitter instance
+// Factory function to create a new emitter instance with censoring middleware
 const createEmitter = () => {
   const emitter = new EventEmitter2({
     wildcard: true,
@@ -11,6 +12,15 @@ const createEmitter = () => {
     verboseMemoryLeak: false,
     ignoreErrors: false,
   });
+
+  // Override emit to censor sensitive data before emitting
+  const originalEmit = emitter.emit.bind(emitter);
+  emitter.emit = function (event, ...args) {
+    // Censor all arguments passed to emit
+    const censoredArgs = args.map(censorSensitiveDataDeep);
+    return originalEmit(event, ...censoredArgs);
+  };
+
   return emitter;
 };
 
@@ -33,7 +43,6 @@ const events = {
   vm: {
     show: "vm:show",
   },
-  narration: "narration",
   status: "status",
   log: {
     markdown: {
@@ -45,6 +54,7 @@ const events = {
     log: "log:log",
     warn: "log:warn",
     debug: "log:debug",
+    narration: "log:narration",
   },
   command: {
     start: "command:start",

package/agent/index.js

@@ -158,7 +158,7 @@ class TestDriverAgent extends EventEmitter2 {
   // single function to handle all program exits
   // allows us to save the current state, run lifecycle hooks, and track analytics
   async exit(failed = true, shouldSave = false, shouldRunPostrun = false) {
-    this.emitter.emit(events.log.log, theme.dim("exiting..."), true);
+    this.emitter.emit(events.log.narration, theme.dim("exiting..."), true);
 
     // Clean up redraw interval
     if (this.redraw && this.redraw.cleanup) {
@@ -290,7 +290,7 @@ class TestDriverAgent extends EventEmitter2 {
       image = null;
     }
 
-    this.emitter.emit(events.log.log, theme.dim("thinking..."), true);
+    this.emitter.emit(events.log.narration, theme.dim("thinking..."), true);
 
     const streamId = `error-${Date.now()}`;
     this.emitter.emit(events.log.markdown.start, streamId);
@@ -332,14 +332,14 @@ class TestDriverAgent extends EventEmitter2 {
 
     if (this.checkCount >= this.checkLimit) {
       this.emitter.emit(
-        events.log.log,
+        events.log.narration,
         theme.red("Exploratory loop detected. Exiting."),
       );
       await this.summarize("Check loop detected.");
       return await this.exit(true);
     }
 
-    this.emitter.emit(events.log.log, theme.dim("checking..."));
+    this.emitter.emit(events.log.narration, theme.dim("checking..."));
 
     // check asks the ai if the task is complete
     let thisScreenshot = await this.system.captureScreenBase64(1, false, true);
@@ -583,8 +583,6 @@ class TestDriverAgent extends EventEmitter2 {
           "check thinks more needs to be done",
         );
 
-        this.emitter.emit(events.log.log, theme.dim("not done yet!"));
-
         return await this.aiExecute(response, validateAndLoop);
       } else {
         this.emitter.emit(events.log.debug, "seems complete, returning");
@@ -719,7 +717,7 @@ class TestDriverAgent extends EventEmitter2 {
       }
     }
 
-    this.emitter.emit(events.log.log, theme.dim("thinking..."), true);
+    this.emitter.emit(events.log.narration, theme.dim("thinking..."), true);
 
     let response = `\`\`\`yaml
 commands:
@@ -747,7 +745,7 @@ commands:
 
     this.tasks.push(currentTask);
 
-    this.emitter.emit(events.log.log, theme.dim("thinking..."), true);
+    this.emitter.emit(events.log.narration, theme.dim("thinking..."), true);
 
     this.lastScreenshot = await this.system.captureScreenBase64();
 
@@ -789,7 +787,7 @@ commands:
   async generate(type, count, baseYaml, skipYaml = false) {
     this.emitter.emit(events.log.debug, "generate called, %s", type);
 
-    this.emitter.emit(events.log.log, theme.dim("thinking..."), true);
+    this.emitter.emit(events.log.narration, theme.dim("thinking..."), true);
 
     if (baseYaml && !skipYaml) {
       await this.runLifecycle("prerun");
@@ -866,7 +864,7 @@ commands:
 
   // this is the functinoality for "undo"
   async popFromHistory(fullStep) {
-    this.emitter.emit(events.log.log, theme.dim("undoing..."), true);
+    this.emitter.emit(events.log.narration, theme.dim("undoing..."), true);
 
     if (this.executionHistory.length) {
       if (fullStep) {
@@ -946,12 +944,16 @@ ${yml}
   async summarize(error = null) {
     this.analytics.track("summarize");
 
-    this.emitter.emit(events.log.log, theme.dim("reviewing test..."), true);
+    this.emitter.emit(
+      events.log.narration,
+      theme.dim("reviewing test..."),
+      true,
+    );
 
     // let text = prompts.summarize(tasks, error);
     let image = await this.system.captureScreenBase64();
 
-    this.emitter.emit(events.log.log, theme.dim("summarizing..."), true);
+    this.emitter.emit(events.log.narration, theme.dim("summarizing..."), true);
 
     const streamId = `summarize-${Date.now()}`;
     this.emitter.emit(events.log.markdown.start, streamId);
@@ -1124,7 +1126,7 @@ ${regression}
       timestamp: fileStartTime,
     });
 
-    this.emitter.emit(events.log.log, theme.cyan(`running ${file}...`));
+    this.emitter.emit(events.log.narration, theme.cyan(`running ${file}...`));
 
     let ymlObj = await this.loadYML(file);
 
@@ -1390,7 +1392,7 @@ ${regression}
         const mtime = new Date(stats.mtime);
         const now = new Date();
         const diffMinutes = (now - mtime) / (1000 * 60);
-        if (diffMinutes < 30) {
+        if (diffMinutes < 10) {
           const fileContent = fs.readFileSync(lastSandboxFile, "utf-8").trim();
 
           // Parse sandbox info (supports both old format and new format)
@@ -1463,8 +1465,8 @@ ${regression}
     // If instance already exists, do not build environment again
     if (this.instance) {
       this.emitter.emit(
-        events.log.log,
-        theme.dim("- sandbox instance already exists, skipping launch."),
+        events.log.narration,
+        theme.dim("sandbox instance already exists, skipping launch."),
       );
       return;
     }
@@ -1488,7 +1490,7 @@ ${regression}
       if (!this.config.CI) {
         this.emitter.emit(
           events.log.log,
-          theme.dim("-- `new` flag detected, will create a new sandbox"),
+          theme.dim("--`new` flag detected, will create a new sandbox"),
         );
       }
     }
@@ -1501,14 +1503,14 @@ ${regression}
     // Set sandbox ID for reconnection (only if not creating new and recent ID exists)
     if (!createNew && recentId) {
       this.emitter.emit(
-        events.log.log,
-        theme.dim(`- using recent sandbox: ${recentId}`),
+        events.log.narration,
+        theme.dim(`using recent sandbox: ${recentId}`),
       );
       this.sandboxId = recentId;
     } else if (!createNew) {
       this.emitter.emit(
-        events.log.log,
-        theme.dim(`- no recent sandbox found, creating a new one.`),
+        events.log.narration,
+        theme.dim(`no recent sandbox found, creating a new one.`),
       );
     }
 
@@ -1516,8 +1518,8 @@ ${regression}
     if (this.sandboxId && !this.config.CI && !createNew) {
       // Attempt to connect to known instance
       this.emitter.emit(
-        events.log.log,
-        theme.dim(`- connecting to sandbox ${this.sandboxId}...`),
+        events.log.narration,
+        theme.dim(`connecting to sandbox ${this.sandboxId}...`),
       );
 
       try {
@@ -1539,17 +1541,16 @@ ${regression}
       }
     }
 
-    this.emitter.emit(events.log.log, theme.dim(`- creating new sandbox...`));
     this.emitter.emit(
-      events.log.log,
-      theme.dim(`  (this can take between 10 - 240 seconds)`),
+      events.log.narration,
+      theme.dim(`creating new sandbox (can take up to 2 minutes)...`),
     );
     // We don't have resiliency/retries baked in, so let's at least give it 1 attempt
     // to see if that fixes the issue.
     let newSandbox = await this.createNewSandbox().catch(() => {
       this.emitter.emit(
-        events.log.log,
-        theme.dim(`  (double-checking sandbox availability)`),
+        events.log.narration,
+        theme.dim(`double-checking sandbox availability`),
       );
 
       return this.createNewSandbox();
@@ -1576,7 +1577,7 @@ ${regression}
       if (!debuggerStarted) {
         debuggerStarted = true; // Prevent multiple starts, especially when running test in parallel
         this.emitter.emit(
-          events.log.log,
+          events.log.narration,
           theme.green(`Starting debugger server...`),
         );
         debuggerProcess = await createDebuggerProcess(
@@ -1697,8 +1698,8 @@ ${regression}
 
   async connectToSandboxService() {
     this.emitter.emit(
-      events.log.log,
-      theme.gray(`- establishing connection...`),
+      events.log.narration,
+      theme.dim(`establishing connection...`),
     );
     let ableToBoot = await this.sandbox.boot(this.config.TD_API_ROOT);
 
@@ -1710,7 +1711,7 @@ Please check your network connection, TD_API_KEY, or the service status.`,
       );
     }
 
-    this.emitter.emit(events.log.log, theme.gray(`- authenticating...`));
+    this.emitter.emit(events.log.narration, theme.dim(`authenticating...`));
     let ableToAuth = await this.sandbox.auth(this.config.TD_API_KEY);
 
     if (!ableToAuth) {
@@ -1723,7 +1724,7 @@ Please check your network connection, TD_API_KEY, or the service status.`,
   }
 
   async connectToSandboxDirect(sandboxId, persist = false) {
-    this.emitter.emit(events.log.log, theme.gray(`- connecting...`));
+    this.emitter.emit(events.log.narration, theme.dim(`connecting...`));
     let instance = await this.sandbox.connect(sandboxId, persist);
     return instance;
   }

package/agent/lib/censorship.js

@@ -0,0 +1,70 @@
+// Shared censorship functionality for sensitive data
+// Uses @npmcli/redact for common patterns (URLs, auth headers, etc.)
+// Plus custom logic for environment variables and interpolation vars
+const { redactLog } = require("@npmcli/redact");
+
+let interpolationVars = JSON.parse(process.env.TD_INTERPOLATION_VARS || "{}");
+
+// Handle local `TD_*` variables
+for (const [key, value] of Object.entries(process.env)) {
+  if (key.startsWith("TD_") && key !== "TD_INTERPOLATION_VARS") {
+    interpolationVars[key] = value;
+  }
+}
+
+// Function to censor sensitive data in strings using both npmcli/redact and custom logic
+const censorSensitiveData = (message) => {
+  if (typeof message !== "string") {
+    return message;
+  }
+
+  // First, use npmcli/redact for common patterns:
+  // - URLs with credentials (https://user:pass@host)
+  // - Authorization headers
+  // - Common secret patterns in logs
+  let result = redactLog(message);
+
+  // Then apply our custom interpolation variable redaction
+  // This catches application-specific secrets from TD_* env vars
+  for (let value of Object.values(interpolationVars)) {
+    // Avoid replacing vars that are 0 or 1 characters
+    if (value && value.length >= 2) {
+      result = result.replaceAll(value, "****");
+    }
+  }
+
+  return result;
+};
+
+// Function to censor sensitive data in any value (recursive for objects/arrays)
+const censorSensitiveDataDeep = (value) => {
+  if (typeof value === "string") {
+    return censorSensitiveData(value);
+  } else if (Array.isArray(value)) {
+    return value.map(censorSensitiveDataDeep);
+  } else if (value && typeof value === "object") {
+    const result = {};
+    for (const [key, val] of Object.entries(value)) {
+      result[key] = censorSensitiveDataDeep(val);
+    }
+    return result;
+  }
+  return value;
+};
+
+// Function to update interpolation variables (for runtime updates)
+const updateInterpolationVars = (newVars) => {
+  interpolationVars = { ...interpolationVars, ...newVars };
+};
+
+// Function to get current interpolation variables (for debugging)
+const getInterpolationVars = () => {
+  return { ...interpolationVars };
+};
+
+module.exports = {
+  censorSensitiveData,
+  censorSensitiveDataDeep,
+  updateInterpolationVars,
+  getInterpolationVars,
+};

package/agent/lib/commander.js

@@ -80,20 +80,20 @@ commands:
       // this will actually interpret the command and execute it
       switch (object.command) {
         case "type":
-          emitter.emit(events.narration, `typing ${object.text}`);
+          emitter.emit(events.log.narration, `typing ${object.text}`);
           emitter.emit(events.log.log, generator.jsonToManual(object));
           response = await commands.type(object.text, object.delay);
           break;
         case "press-keys":
           emitter.emit(
-            events.narration,
+            events.log.narration,
             `pressing keys ${object.keys.join(",")}`,
           );
           emitter.emit(events.log.log, generator.jsonToManual(object));
           response = await commands["press-keys"](object.keys);
           break;
         case "scroll":
-          emitter.emit(events.narration, `scrolling ${object.direction}`);
+          emitter.emit(events.log.narration, `scrolling ${object.direction}`);
           emitter.emit(events.log.log, generator.jsonToManual(object));
           response = await commands.scroll(
             object.direction,
@@ -102,17 +102,20 @@ commands:
           );
           break;
         case "wait":
-          emitter.emit(events.narration, `waiting ${object.timeout} seconds`);
+          emitter.emit(
+            events.log.narration,
+            `waiting ${object.timeout} seconds`,
+          );
           emitter.emit(events.log.log, generator.jsonToManual(object));
           response = await commands.wait(object.timeout);
           break;
         case "click":
-          emitter.emit(events.narration, `${object.action}`);
+          emitter.emit(events.log.narration, `${object.action}`);
           emitter.emit(events.log.log, generator.jsonToManual(object));
           response = await commands["click"](object.x, object.y, object.action);
           break;
         case "hover":
-          emitter.emit(events.narration, `moving mouse`);
+          emitter.emit(events.log.narration, `moving mouse`);
           emitter.emit(events.log.log, generator.jsonToManual(object));
           response = await commands["hover"](object.x, object.y);
           break;
@@ -121,7 +124,10 @@ commands:
           response = await commands["drag"](object.x, object.y);
           break;
         case "hover-text":
-          emitter.emit(events.narration, `searching for ${object.description}`);
+          emitter.emit(
+            events.log.narration,
+            `searching for ${object.description}`,
+          );
           emitter.emit(events.log.log, generator.jsonToManual(object));
           response = await commands["hover-text"](
             object.text,
@@ -132,7 +138,7 @@ commands:
           break;
         case "hover-image":
           emitter.emit(
-            events.narration,
+            events.log.narration,
             `searching for image of ${object.description}`,
           );
           emitter.emit(events.log.log, generator.jsonToManual(object));
@@ -144,13 +150,16 @@ commands:
         case "match-image":
           emitter.emit(events.log.log, generator.jsonToManual(object));
           emitter.emit(
-            events.narration,
+            events.log.narration,
             `${object.action} image ${object.path}`,
           );
           response = await commands["match-image"](object.path, object.action);
           break;
         case "wait-for-image":
-          emitter.emit(events.narration, `waiting for ${object.description}`);
+          emitter.emit(
+            events.log.narration,
+            `waiting for ${object.description}`,
+          );
           emitter.emit(events.log.log, generator.jsonToManual(object));
           response = await commands["wait-for-image"](
             object.description,
@@ -158,7 +167,7 @@ commands:
           );
           break;
         case "wait-for-text":
-          emitter.emit(events.narration, `waiting for ${object.text}`);
+          emitter.emit(events.log.narration, `waiting for ${object.text}`);
           emitter.emit(events.log.log, generator.jsonToManual(object));
           copy.text = "*****";
           response = await commands["wait-for-text"](
@@ -168,7 +177,7 @@ commands:
           );
           break;
         case "scroll-until-text":
-          emitter.emit(events.narration, `scrolling until ${object.text}`);
+          emitter.emit(events.log.narration, `scrolling until ${object.text}`);
           emitter.emit(events.log.log, generator.jsonToManual(object));
           copy.text = "*****";
           response = await commands["scroll-until-text"](
@@ -181,7 +190,7 @@ commands:
           break;
         case "scroll-until-image": {
           const needle = object.description || object.path;
-          emitter.emit(events.narration, `scrolling until ${needle}`);
+          emitter.emit(events.log.narration, `scrolling until ${needle}`);
           emitter.emit(events.log.log, generator.jsonToManual(object));
           response = await commands["scroll-until-image"](
             object.description,
@@ -193,7 +202,7 @@ commands:
           break;
         }
         case "focus-application":
-          emitter.emit(events.narration, `focusing ${object.name}`);
+          emitter.emit(events.log.narration, `focusing ${object.name}`);
           emitter.emit(events.log.log, generator.jsonToManual(object));
           response = await commands["focus-application"](object.name);
           break;
@@ -205,12 +214,12 @@ commands:
           break;
         }
         case "assert":
-          emitter.emit(events.narration, `asserting ${object.expect}`);
+          emitter.emit(events.log.narration, `asserting ${object.expect}`);
           emitter.emit(events.log.log, generator.jsonToManual(object));
           response = await commands.assert(object.expect, object.async);
           break;
         case "exec":
-          emitter.emit(events.narration, `exec`);
+          emitter.emit(events.log.narration, `exec`);
           emitter.emit(
             events.log.log,
             generator.jsonToManual({

package/agent/lib/commands.js

@@ -190,7 +190,7 @@ const createCommands = (
       }
     };
 
-    emitter.emit(events.log.log, `thinking...`);
+    emitter.emit(events.log.narration, `thinking...`);
 
     if (async) {
       await sdk
@@ -276,7 +276,7 @@ const createCommands = (
     }
 
     emitter.emit(
-      "log:debug",
+      events.log.narration,
       theme.dim(`${action} ${button} clicking at ${x}, ${y}...`),
       true,
     );
@@ -347,7 +347,7 @@ const createCommands = (
 
       description = description ? description.toString() : null;
 
-      emitter.emit(events.log.log, theme.dim("thinking..."), true);
+      emitter.emit(events.log.narration, theme.dim("thinking..."), true);
 
       let response = await sdk.req(
         "hover/text",
@@ -375,7 +375,7 @@ const createCommands = (
     // uses our api to find all images on screen
     "hover-image": async (description, action = "click") => {
       // take a screenshot
-      emitter.emit(events.log.log, theme.dim("thinking..."), true);
+      emitter.emit(events.log.narration, theme.dim("thinking..."), true);
 
       let response = await sdk.req(
         "hover/image",
@@ -446,7 +446,7 @@ const createCommands = (
     },
     "wait-for-image": async (description, timeout = 10000) => {
       emitter.emit(
-        events.log.log,
+        events.log.narration,
         theme.dim(
           `waiting for an image matching description "${description}"...`,
         ),
@@ -467,7 +467,7 @@ const createCommands = (
         durationPassed = new Date().getTime() - startTime;
         if (!passed) {
           emitter.emit(
-            events.log.log,
+            events.log.narration,
             theme.dim(
               `${niceSeconds(durationPassed)} seconds have passed without finding an image matching the description "${description}"`,
             ),
@@ -479,7 +479,7 @@ const createCommands = (
 
       if (passed) {
         emitter.emit(
-          events.log.log,
+          events.log.narration,
           theme.dim(
             `An image matching the description "${description}" found!`,
           ),
@@ -496,7 +496,7 @@ const createCommands = (
       await redraw.start();
 
       emitter.emit(
-        events.log.log,
+        events.log.narration,
         theme.dim(`waiting for text: "${text}"...`),
         true,
       );
@@ -525,7 +525,7 @@ const createCommands = (
         durationPassed = new Date().getTime() - startTime;
         if (!passed) {
           emitter.emit(
-            events.log.log,
+            events.log.narration,
             theme.dim(
               `${niceSeconds(durationPassed)} seconds have passed without finding "${text}"`,
             ),
@@ -536,7 +536,7 @@ const createCommands = (
       }
 
       if (passed) {
-        emitter.emit(events.log.log, theme.dim(`"${text}" found!`), true);
+        emitter.emit(events.log.narration, theme.dim(`"${text}" found!`), true);
         return;
       } else {
         throw new MatchError(
@@ -554,7 +554,7 @@ const createCommands = (
       await redraw.start();
 
       emitter.emit(
-        events.log.log,
+        events.log.narration,
         theme.dim(`scrolling for text: "${text}"...`),
         true,
       );
@@ -598,7 +598,7 @@ const createCommands = (
         passed = response.data;
         if (!passed) {
           emitter.emit(
-            events.log.log,
+            events.log.narration,
             theme.dim(
               `scrolling ${direction} ${incrementDistance}px. ${scrollDistance + incrementDistance}/${maxDistance}px scrolled...`,
             ),
@@ -610,7 +610,7 @@ const createCommands = (
       }
 
       if (passed) {
-        emitter.emit(events.log.log, theme.dim(`"${text}" found!`), true);
+        emitter.emit(events.log.narration, theme.dim(`"${text}" found!`), true);
         return;
       } else {
         throw new MatchError(
@@ -638,7 +638,7 @@ const createCommands = (
       }
 
       emitter.emit(
-        events.log.log,
+        events.log.narration,
         theme.dim(`scrolling for an image matching "${needle}"...`),
         true,
       );
@@ -665,7 +665,7 @@ const createCommands = (
 
         if (!passed) {
           emitter.emit(
-            events.log.log,
+            events.log.narration,
             theme.dim(`scrolling ${direction} ${incrementDistance} pixels...`),
             true,
           );
@@ -675,7 +675,11 @@ const createCommands = (
       }
 
       if (passed) {
-        emitter.emit(events.log.log, theme.dim(`"${needle}" found!`), true);
+        emitter.emit(
+          events.log.narration,
+          theme.dim(`"${needle}" found!`),
+          true,
+        );
         return;
       } else {
         throw new CommandError(
@@ -705,7 +709,7 @@ const createCommands = (
       return await assert(assertion, true, async);
     },
     exec: async (language, code, timeout, silent = false) => {
-      emitter.emit(events.log.log, theme.dim(`calling exec...`), true);
+      emitter.emit(events.log.narration, theme.dim(`calling exec...`), true);
 
       emitter.emit(events.log.log, code);
 
@@ -738,10 +742,10 @@ const createCommands = (
           return result.out?.stdout?.trim();
         }
       } else if (language == "js") {
-        emitter.emit(events.log.log, theme.dim(`running js...`), true);
+        emitter.emit(events.log.narration, theme.dim(`running js...`), true);
 
         emitter.emit(
-          events.log.log,
+          events.log.narration,
           theme.dim(`running value of \`${plat}\` in local JS vm...`),
           true,
         );

package/debugger/index.html

@@ -670,6 +670,8 @@
       // Handle window resize to recalculate scale (throttled)
       window.addEventListener("resize", throttle(resizeOverlay, 100));
 
+      setInterval(resizeOverlay, 1000);
+
       // Handle window blur/focus for screen locking
       window.addEventListener("blur", () => {
         showInteractionOverlay();