@loadmill/droid-cua 2.10.1-rc.9 → 2.12.0

package/build/src/prompts/execution.js

@@ -40,7 +40,9 @@ CRITICAL RULES:
 Your process:
 1. Read the instruction
 2. Choose the right tool when the instruction is really a higher-level capability:
-   - Use \`assert\` for assertions and verification intent, including legacy \`assert:\` lines and natural language like verify, validate, ensure, or confirm
+   - Use \`assert\` for assertions and verification intent, including legacy \`assert:\` lines and natural language like verify, validate, ensure, or confirm. When calling \`assert\`, include a \`context\` value with the prior details needed to interpret the condition. If no prior context is needed, say that explicitly.
+   - Before calling \`assert\`, make sure the current screen contains the evidence needed to validate the condition. If needed, perform only minimal preparation first, such as taking a screenshot, waiting briefly, closing the keyboard, scrolling within the relevant view, or opening the relevant menu.
+   - Once you call \`assert\`, validation is read-only. The assertion tool will not click, type, wait, scroll, request screenshots, or change the device state.
    - Use \`navigate\` for direct browser navigation instead of trying to manipulate the address bar
    - Use \`loadmill\` only when the instruction explicitly references the word "loadmill"
 3. Execute the required actions
@@ -88,24 +90,39 @@ ${transitionLine}${transcriptSection}
 
 [IMPORTANT: Continue the current test from this new page. Use the prior context above, but rely on the latest screenshot for the current UI.]`;
 }
-export function buildAssertionSystemPrompt(baseSystemPrompt, assertionPrompt) {
+export function buildAssertionSystemPrompt(baseSystemPrompt, assertionPrompt, { assertionContext = "", transcript = "", } = {}) {
+    const normalizedAssertionContext = typeof assertionContext === "string" && assertionContext.trim()
+        ? assertionContext.trim()
+        : "No additional assertion context was provided.";
+    const normalizedTranscript = typeof transcript === "string" && transcript.trim()
+        ? transcript.trim()
+        : "No prior session transcript was provided.";
     return `${baseSystemPrompt}
 
 ASSERTION MODE:
-You are now validating an assertion. The user has provided an assertion statement that you must verify.
+You are validating one assertion as part of an automated test run.
+This is a read-only check.
 
 Your task:
-1. Take screenshots and perform LIMITED actions if needed to validate the assertion.
-2. Determine if the assertion is TRUE or FALSE based on the current state.
-3. You MUST respond with a clear verdict in this exact format:
+1. Use only the provided screenshot, assertion context, app context already included above, and session transcript.
+2. Do not ask for more screenshots, do not click, do not type, do not wait, and do not propose actions.
+3. Determine if the assertion is TRUE or FALSE based on the current state and provided context.
+4. If the evidence is insufficient, treat the assertion as false.
+5. You MUST respond with a clear verdict in this exact format:
    - If the assertion is true, include the text: "ASSERTION RESULT: PASS"
    - If the assertion is false or cannot be confidently validated, include: "ASSERTION RESULT: FAIL"
-4. After the verdict, provide a brief explanation (1-2 sentences) of why it passed or failed.
+6. After the verdict, provide a brief explanation (1-2 sentences) of why it passed or failed.
 
 The assertion to validate is: "${assertionPrompt}"
 
+ASSERTION CONTEXT:
+${normalizedAssertionContext}
+
+SESSION TRANSCRIPT SO FAR:
+${normalizedTranscript}
+
 Remember:
 - If you cannot confidently validate the assertion, treat it as FAIL.
 - 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.`;
+- This is read-only evaluation. Return the verdict and reason only.`;
 }

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

@@ -176,6 +176,17 @@ function extractResponseTexts(response) {
     })
         .filter(Boolean);
 }
+function summarizeContextManagement(contextManagement) {
+    const entries = Array.isArray(contextManagement) ? contextManagement : [];
+    const compaction = entries.find((entry) => entry?.type === "compaction") || null;
+    return {
+        enabled: Boolean(compaction),
+        compactThreshold: typeof compaction?.compact_threshold === "number"
+            ? compaction.compact_threshold
+            : null,
+        entries: entries.length > 0 ? entries : [],
+    };
+}
 export class CuaDebugTracer {
     constructor(logger) {
         this.logger = logger;
@@ -221,12 +232,14 @@ export class CuaDebugTracer {
         const missingCallIds = inputCallIds.filter((callId) => !lastResponseCallIds.includes(callId));
         const previousResponseIdMatchesLastResponseId = !previousResponseId || !lastResponseMeta?.id ? null : previousResponseId === lastResponseMeta.id;
         const allInputCallIdsFoundInLastResponse = inputCallIds.length === 0 ? true : missingCallIds.length === 0;
+        const contextManagement = summarizeContextManagement(requestParams.context_management);
         const requestConfig = {
             model: requestParams.model,
             tools: requestParams.tools,
             truncation: requestParams.truncation,
             reasoning: requestParams.reasoning,
-            store: requestParams.store
+            store: requestParams.store,
+            context_management: requestParams.context_management || null
         };
         const requestConfigHash = crypto.createHash("sha256").update(JSON.stringify(requestConfig)).digest("hex");
         emitDesktopDebug("cua.request", scope, ids, {
@@ -249,6 +262,7 @@ export class CuaDebugTracer {
                 acknowledgedSent: acknowledgedSafetyChecksSent,
                 acknowledgedSentCount: acknowledgedSafetyChecksSent.length
             },
+            contextManagement,
             chain: {
                 lastResponseId: lastResponseMeta?.id ?? null,
                 lastResponseOutputTypes: lastResponseMeta?.outputTypes ?? [],
@@ -269,6 +283,7 @@ export class CuaDebugTracer {
             lastResponseComputerCallIds: lastResponseCallIds,
             allInputCallIdsFoundInLastResponse,
             missingCallIds,
+            contextManagement,
             requestConfigHash
         });
         return {
@@ -278,6 +293,7 @@ export class CuaDebugTracer {
             localRequestId,
             requestLog,
             requestConfigHash,
+            contextManagement,
             previousPendingSafetyChecks,
             acknowledgedSafetyChecksSent,
             previousResponseIdMatchesLastResponseId,
@@ -322,7 +338,8 @@ export class CuaDebugTracer {
             output_length: totalItems,
             output_types: outputTypes,
             tool_calls: toolCalls.length > 0 ? toolCalls : "none",
-            pending_safety_checks: safetyChecks.length > 0 ? safetyChecks : "none"
+            pending_safety_checks: safetyChecks.length > 0 ? safetyChecks : "none",
+            usage: response.usage || null
         });
         if (accountedItems < totalItems) {
             this.logger.debug("UNACCOUNTED OUTPUT ITEMS - Full output array:", response.output);
@@ -334,6 +351,9 @@ export class CuaDebugTracer {
             outputTypes,
             output: summarizeResponseOutput(response),
             texts: extractResponseTexts(response),
+            usage: response.usage || null,
+            contextManagement: trace.contextManagement,
+            requestConfigHash: trace.requestConfigHash,
             safetyChecks: {
                 pending: safetyChecks,
                 pendingCount: safetyChecks.length
@@ -351,6 +371,7 @@ export class CuaDebugTracer {
             localRequestId: trace.localRequestId,
             request: trace.requestLog,
             requestConfigHash: trace.requestConfigHash,
+            contextManagement: trace.contextManagement,
             chain: {
                 lastResponseId: trace.lastResponseMeta?.id ?? null,
                 lastResponseOutputTypes: trace.lastResponseMeta?.outputTypes ?? [],
@@ -394,6 +415,7 @@ export class CuaDebugTracer {
                 missingCallIds: trace.missingCallIds
             },
             requestConfigHash: trace.requestConfigHash,
+            contextManagement: trace.contextManagement,
             details: responseError,
             safetyChecks: {
                 previousPending: trace.previousPendingSafetyChecks,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@loadmill/droid-cua",
-  "version": "2.10.1-rc.9",
+  "version": "2.12.0",
   "description": "AI-powered Android testing agent using OpenAI's computer-use model and ADB",
   "main": "build/index.js",
   "type": "module",