@loadmill/droid-cua 2.18.0 → 2.21.0

package/build/src/integrations/loadmill/executor.js

@@ -57,13 +57,13 @@ ${futureScript}`
         }
     ];
 }
-async function selectRelevantParameterNames({ openAiClient, command, flowName, futureInstructions, parameters, }) {
+async function selectRelevantParameterNames({ openAiClient, helperModel, command, flowName, futureInstructions, parameters, }) {
     if (!openAiClient || !Array.isArray(parameters) || parameters.length === 0) {
         return [];
     }
     try {
         const response = await openAiClient.chat.completions.create({
-            model: LOADMILL_MODEL_ALIASES.DEFAULT_AGENT_MODEL,
+            model: helperModel || LOADMILL_MODEL_ALIASES.DEFAULT_AGENT_MODEL,
             messages: buildParameterSelectionMessages({
                 command,
                 flowName,
@@ -98,7 +98,7 @@ function findSelectedSuiteFlowRun(finalResult, selectedFlow) {
     }
     return flowRuns.find((flowRun) => flowRun?.testSuiteFlowId === selectedFlow?.id) || null;
 }
-async function buildRuntimeParameterContext({ loadmillClient, finalResult, selectedFlow, command, flowName, futureInstructions, openAiClient, }) {
+async function buildRuntimeParameterContext({ loadmillClient, finalResult, selectedFlow, command, flowName, futureInstructions, openAiClient, helperModel, }) {
     const selectedSuiteFlowRun = findSelectedSuiteFlowRun(finalResult, selectedFlow);
     if (!selectedSuiteFlowRun?.id || typeof loadmillClient.getTestSuiteFlowRun !== "function") {
         return null;
@@ -112,6 +112,7 @@ async function buildRuntimeParameterContext({ loadmillClient, finalResult, selec
             flowName,
             futureInstructions,
             parameters,
+            helperModel,
         });
         const candidates = selectLoadmillParameterCandidates(parameters, selectedParameterNames);
         const note = renderLoadmillParameterContextNote(candidates);
@@ -198,7 +199,7 @@ async function pollForCompletion(runId, loadmillClient, options = {}) {
  * @returns {Promise<{success: boolean, error?: string, result?: Object}>}
  */
 export async function executeLoadmillCommand(userInput, options = {}) {
-    const { onProgress = () => { }, loadmillClient = createLoadmillApiClient(), openAiClient = null, futureInstructions = [], pollTimeoutMs, pollIntervalMs, now, sleepFn, } = options;
+    const { onProgress = () => { }, loadmillClient = createLoadmillApiClient(), openAiClient = null, helperModel = null, futureInstructions = [], pollTimeoutMs, pollIntervalMs, now, sleepFn, } = options;
     // Check for API token
     if (!loadmillClient.getApiToken()) {
         return {
@@ -215,7 +216,10 @@ export async function executeLoadmillCommand(userInput, options = {}) {
     try {
         // Step 1: Interpret the command
         onProgress({ step: "interpreting", message: "Interpreting command..." });
-        const interpreted = await interpretLoadmillCommand(userInput, { client: openAiClient });
+        const interpreted = await interpretLoadmillCommand(userInput, {
+            client: openAiClient,
+            ...(helperModel ? { model: helperModel } : {}),
+        });
         logger.debug("Loadmill command interpreted", interpreted);
         // Step 2: Search for flows
         onProgress({ step: "searching", message: `Searching for flows matching "${interpreted.searchQuery}"...` });
@@ -229,7 +233,10 @@ export async function executeLoadmillCommand(userInput, options = {}) {
         }
         // Step 3: Select best match
         onProgress({ step: "selecting", message: `Found ${flows.length} flow(s). Selecting best match...` });
-        const { selectedFlow, confidence } = await selectBestFlow(flows, interpreted.searchQuery, { client: openAiClient });
+        const { selectedFlow, confidence } = await selectBestFlow(flows, interpreted.searchQuery, {
+            client: openAiClient,
+            ...(helperModel ? { model: helperModel } : {}),
+        });
         if (!selectedFlow) {
             return {
                 success: false,
@@ -272,6 +279,7 @@ export async function executeLoadmillCommand(userInput, options = {}) {
                 flowName,
                 futureInstructions,
                 openAiClient,
+                helperModel,
             })
             : null;
         return {

package/build/src/integrations/loadmill/interpreter.js

@@ -8,9 +8,9 @@ import { buildLoadmillCommandInterpretationMessages, buildLoadmillFlowSelectionM
  * @param {string} userInput - Natural language command
  * @returns {Promise<{searchQuery: string, parameters: Object}>}
  */
-export async function interpretLoadmillCommand(userInput, { client }) {
+export async function interpretLoadmillCommand(userInput, { client, model = LOADMILL_MODEL_ALIASES.FAST }) {
     const response = await client.chat.completions.create({
-        model: LOADMILL_MODEL_ALIASES.FAST,
+        model,
         messages: buildLoadmillCommandInterpretationMessages(userInput),
         response_format: { type: "json_object" }
     });
@@ -27,7 +27,7 @@ export async function interpretLoadmillCommand(userInput, { client }) {
  * @param {string} originalQuery - Original user query
  * @returns {Promise<{selectedFlow: Object|null, confidence: number}>}
  */
-export async function selectBestFlow(flows, originalQuery, { client }) {
+export async function selectBestFlow(flows, originalQuery, { client, model = LOADMILL_MODEL_ALIASES.FAST }) {
     // Ensure flows is an array
     if (!flows || !Array.isArray(flows) || flows.length === 0) {
         return { selectedFlow: null, confidence: 0 };
@@ -42,7 +42,7 @@ export async function selectBestFlow(flows, originalQuery, { client }) {
         return `${i + 1}. ID: ${f.id}, Name: "${name}"${suite ? `, Suite: "${suite}"` : ""}`;
     }).join("\n");
     const response = await client.chat.completions.create({
-        model: LOADMILL_MODEL_ALIASES.FAST,
+        model,
         messages: buildLoadmillFlowSelectionMessages(originalQuery, flowList),
         response_format: { type: "json_object" }
     });

package/build/src/modes/execution-mode.js

@@ -145,6 +145,19 @@ export class ExecutionMode {
                 });
                 return { success: false, error: 'Stopped by user' };
             }
+            if (runContext.waitIfPaused) {
+                const pauseResult = await runContext.waitIfPaused({
+                    boundary: 'before_instruction',
+                    refreshScreenOnResume: false,
+                    stepContext,
+                });
+                if (pauseResult?.status === 'stopped' || this.shouldStop) {
+                    this.emit(addOutput, 'info', 'Test execution stopped by user.', runContext, stepContext, {
+                        eventType: 'system_message'
+                    });
+                    return { success: false, error: 'Stopped by user' };
+                }
+            }
             const instruction = this.instructions[i];
             this.emit(addOutput, 'user', instruction, runContext, stepContext, {
                 eventType: 'instruction_started',
@@ -256,6 +269,9 @@ export class ExecutionMode {
                 return this.shouldStop;
             };
             const { responseId: newResponseId, summary: turnSummary } = await this.engine.runFullTurn(response, trackAction, instructionContext, stepContext, { currentScreenshot });
+            if (turnSummary.stopped || this.shouldStop) {
+                return { success: false, error: "Stopped by user" };
+            }
             if (turnSummary.assertionStatus === "passed") {
                 this.stats.assertionsPassed++;
             }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@loadmill/droid-cua",
-  "version": "2.18.0",
+  "version": "2.21.0",
   "description": "AI-powered Android testing agent using OpenAI's computer-use model and ADB",
   "main": "build/index.js",
   "type": "module",
@@ -22,6 +22,8 @@
     "bench:report:web": "node scripts/benchmark-report-web.js",
     "execution:profile": "node scripts/execution-profile.js",
     "test": "node --test unit-tests/*.test.js",
+    "test:aws-device-farm:android": "node integration-tests/aws-device-farm-android.test.js",
+    "test:aws-device-farm:ios": "node integration-tests/aws-device-farm-ios.test.js",
     "build": "tsc",
     "clean": "rm -rf build",
     "typecheck": "tsc --noEmit",
@@ -58,6 +60,7 @@
   "author": "Loadmill",
   "license": "UNLICENSED",
   "dependencies": {
+    "@aws-sdk/client-device-farm": "^3.1065.0",
     "dotenv": "^16.4.7",
     "ink": "^4.4.1",
     "ink-spinner": "^5.0.0",