@loadmill/droid-cua 2.8.0 → 2.9.0

package/build/src/prompts/execution.js

@@ -1,19 +1,19 @@
 import { appendCustomSections, buildBaseSystemPrompt, buildStrictModeRuntimeSection, isBrowserTarget } from "./base.js";
-export function buildAppContextSection(briefing) {
-    const text = typeof briefing === "string" ? briefing.trim() : "";
-    if (!text) {
+export function buildAppContextSection(text) {
+    const trimmed = typeof text === "string" ? text.trim() : "";
+    if (!trimmed) {
         return "";
     }
-    return `APP CONTEXT BRIEFING:
-The following is a condensed description of the app you are testing, relevant to the current task.
+    return `APP CONTEXT:
+The following describes the app you are testing.
 Use this to understand screen layouts, terminology, navigation, and expected behavior.
 
-${text}`;
+${trimmed}`;
 }
-export function buildExecutionModePrompt(deviceInfo, customInstructions = {}, appContextBriefing = "", runtimeOptions = {}) {
+export function buildExecutionModePrompt(deviceInfo, customInstructions = {}, appContextText = "", runtimeOptions = {}) {
     const executionCustomText = typeof customInstructions.executionModeInstructions === "string" ? customInstructions.executionModeInstructions.trim() : "";
     const basePrompt = buildBaseSystemPrompt(deviceInfo, customInstructions);
-    const appContextSection = buildAppContextSection(appContextBriefing);
+    const appContextSection = buildAppContextSection(appContextText);
     const strictModeSection = buildStrictModeRuntimeSection(runtimeOptions);
     const browserTarget = isBrowserTarget(deviceInfo);
     const prompt = `${basePrompt}
@@ -104,101 +104,3 @@ Remember:
 - You must include either "ASSERTION RESULT: PASS" or "ASSERTION RESULT: FAIL" in your response.
 - Be thorough but efficient. Only take the actions necessary to validate the assertion.`;
 }
-export function buildAppContextCompactionInput({ contextDocument, taskDescription, tokenBudget }) {
-    return [
-        {
-            role: "system",
-            content: [{
-                    type: "input_text",
-                    text: `You are compressing an app context document for an app testing agent.
-
-You will receive:
-1. A context document
-2. A test task
-
-Your job is to SELECT only the facts from the context document that are useful for the given task.
-The output will be injected into a system prompt with a strict token budget.
-
-CRITICAL:
-- Use only facts explicitly supported by the context document
-- Never invent, infer, normalize, substitute, or improve credentials, labels, screen names, button names, or numeric values
-- Preserve exact values verbatim when present in the source
-- Prefer facts that help the agent act correctly when they are not obvious from the task alone
-- Do not restate, paraphrase, summarize, or reorganize the test task
-- The output must not read like instructions or a test plan
-- Do not describe what the agent should do
-- Output only reference knowledge about the app
-- If a line could be copied from the task with minor wording changes, omit it
-- Prefer copying source facts verbatim or near-verbatim over rewriting them
-- Do not collapse multiple specific source facts into one generic summary if that removes useful distinctions
-
-Selection priority:
-1. Facts the agent would NOT know from the test script alone
-2. Facts that are hard to infer from screenshots
-3. Non-obvious navigation or interaction details
-4. Exact visible labels needed to act correctly
-5. Credentials and other exact values
-
-High-value facts:
-- exact UI labels
-- how state, mode, or account selection is performed
-- where logout is located
-- hidden or non-obvious navigation
-- which menu items are decorative or non-functional
-- screen titles and section labels used to confirm location
-- exact credentials and role labels
-
-Low-value facts:
-- restating the test steps
-- repeating literal values already present in the task
-- generic summaries like "approve the transaction"
-
-When the task involves authentication, switching state or mode, opening menus, or moving between major areas of the app, strongly prefer including:
-- how account, state, or mode selection is performed
-- exact visible labels for the relevant controls
-- where exit or sign-out actions are located
-- the screen or section labels that confirm the agent is in the right place
-
-Rules:
-- Output plain text only
-- No markdown, no bullet symbols, no numbering, no headers
-- Use terse, factual language: one fact per line, no filler words
-- Blank lines only to separate logical groups
-- Prefer exact visible UI labels over summaries
-- Do not describe step-by-step procedures
-- Do not restate the test workflow
-- State only facts about screens, elements, hidden interactions, entities, credentials, and navigation
-- If a useful fact is not explicitly stated in the context document, omit it
-- Include only information relevant to this task
-- Do not waste space repeating the task itself
-- If the task already states a value or action, include it only when the context adds non-obvious execution details
-- Return a short result or an empty string if little is relevant
-- Target: under ${tokenBudget} tokens
-
-Bad output patterns to avoid:
-- generic summaries that remove actionable details
-- lines that restate the task in generic prose
-- lines that describe obvious workflow steps instead of app knowledge
-- lines that replace exact source labels or mechanisms with broad summaries
-
-Good output characteristics:
-- preserves the exact label or mechanism from the source when it matters
-- keeps distinctions like dropdown vs tabs, drawer vs visible button, exact section titles, exact button text
-- includes hidden or non-obvious navigation details when relevant
-
-Return only the briefing text.`
-                }]
-        },
-        {
-            role: "user",
-            content: [{
-                    type: "input_text",
-                    text: `APP CONTEXT DOCUMENT:
-${contextDocument}
-
-TASK:
-${taskDescription}`
-                }]
-        }
-    ];
-}

package/build/src/utils/console-output.js

@@ -1,3 +1,8 @@
+import { classifyCaptureSource } from "../core/capture-sources.js";
+function isHiddenExecutionOutput(item) {
+    const captureSource = item?.payload?.captureSource;
+    return classifyCaptureSource(captureSource).mediaKind === "hidden-debug";
+}
 export function formatCliOutput(item) {
     if (item == null) {
         return null;
@@ -8,6 +13,9 @@ export function formatCliOutput(item) {
     if (typeof item !== "object") {
         return String(item);
     }
+    if (isHiddenExecutionOutput(item)) {
+        return null;
+    }
     const text = typeof item.text === "string" ? item.text : null;
     const type = typeof item.type === "string" ? item.type : "";
     const eventType = typeof item.eventType === "string" ? item.eventType : "";

package/build/src/utils/cua-debug-tracer.js

@@ -94,7 +94,6 @@ function redactRequestForFileLog(requestParams) {
                         ...item.output,
                         image_url: `data:image/png;base64,[${base64Data.length} chars]`
                     },
-                    current_url: item.current_url,
                     acknowledged_safety_checks: item.acknowledged_safety_checks
                 };
             }

package/build/src/utils/structured-debug-log-manager.js

@@ -20,7 +20,10 @@ function compactTimestamp(iso) {
 }
 function sanitizeValue(value, pathParts = [], eventName = "") {
     if (typeof value === "string") {
-        if (eventName === "cua.request.full" || eventName === "cua.response.full" || eventName === "app_context.briefing.full") {
+        if (eventName === "cua.request.full" ||
+            eventName === "cua.response.full" ||
+            eventName === "app_context.briefing.full" ||
+            eventName === "app_context.loaded") {
             return value;
         }
         if (isPromptContentPath(pathParts)) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@loadmill/droid-cua",
-  "version": "2.8.0",
+  "version": "2.9.0",
   "description": "AI-powered Android testing agent using OpenAI's computer-use model and ADB",
   "main": "build/index.js",
   "type": "module",