testchimp-runner-core 0.0.33 → 0.0.35

package/src/orchestrator/orchestrator-agent.ts

@@ -6,7 +6,8 @@
 import { expect } from '@playwright/test';
 import { LLMFacade } from '../llm-facade';
 import { ProgressReporter, TokenUsage, StepExecutionStatus } from '../progress-reporter';
-import { getEnhancedPageInfo } from '../utils/page-info-utils';
+import { getEnhancedPageInfo, PageInfo } from '../utils/page-info-utils';
+import { CoordinateConverter } from '../utils/coordinate-converter';
 import { ToolRegistry, ToolExecutionContext } from './tool-registry';
 import { DEFAULT_MODEL } from '../model-constants';
 import {
@@ -17,8 +18,16 @@ import {
   MemoryStep,
   OrchestratorStepResult,
   SelfReflection,
+  NoteToFutureSelf,
+  CoordinateAction,
+  ExplorationMode,
   DEFAULT_AGENT_CONFIG
 } from './types';
+import { OrchestratorPrompts } from './orchestrator-prompts';
+import { PageInfoRetry } from '../utils/page-info-retry';
+import { DecisionParser } from './decision-parser';
+import { PageSoMHandler } from './page-som-handler';
+import { SomCommand, CommandRunStatus, InteractionAction, isSomVerification, isSomCommand, SomVerification } from './som-types';
 
 /**
  * Orchestrator Agent - manages step execution with tool use and memory
@@ -30,6 +39,9 @@ export class OrchestratorAgent {
   private config: Required<AgentConfig>;
   private logger?: (message: string, level?: 'log' | 'error' | 'warn' | 'debug') => void;
   private debugMode: boolean = false;
+  private decisionParser: DecisionParser;
+  private somHandler?: PageSoMHandler;
+  private previousSomScreenshot?: string;  // Track previous iteration's screenshot
   
   constructor(
     llmFacade: LLMFacade,
@@ -45,6 +57,12 @@ export class OrchestratorAgent {
     this.progressReporter = progressReporter;
     this.logger = logger;
     this.debugMode = debugMode || false;
+    this.decisionParser = new DecisionParser(logger);
+    
+    // Initialize SoM handler if enabled
+    if (this.config.useSoM) {
+      this.somHandler = new PageSoMHandler(null as any, this.logger);
+    }
   }
   
   setDebugMode(enabled: boolean): void {
@@ -61,15 +79,18 @@ export class OrchestratorAgent {
     totalSteps: number,
     scenarioSteps: string[],
     memory: JourneyMemory,
-    jobId: string
+    jobId: string,
+    priorSteps?: string[],    // NEW: For repair mode (undefined for script gen)
+    nextSteps?: string[]      // NEW: For repair mode (undefined for script gen)
   ): Promise<OrchestratorStepResult> {
     this.logger?.(`\n[Orchestrator] ========== STEP ${stepNumber}/${totalSteps} ==========`);
     this.logger?.(`[Orchestrator] 🎯 Goal: ${stepDescription}`);
     
     let iteration = 0;
-    let previousReflection: SelfReflection | undefined = undefined;
+    let noteToSelf: NoteToFutureSelf | undefined = memory.latestNote;  // Start with note from previous step
     const commandsExecuted: string[] = [];
     let consecutiveFailures = 0;  // Track consecutive iterations with failed commands
+    let coordinateAttempts = 0;  // Track coordinate mode attempts (max 2)
     
     while (iteration < this.config.maxIterationsPerStep) {
       iteration++;
@@ -84,15 +105,23 @@ export class OrchestratorAgent {
         totalSteps,
         scenarioSteps,
         memory,
-        previousReflection,
-        consecutiveFailures
+        consecutiveFailures,
+        noteToSelf,  // Pass note from previous iteration
+        priorSteps,  // NEW: Pass repair context
+        nextSteps    // NEW: Pass repair context
       );
       
       // Call agent to make decision
-      const decision = await this.callAgent(context, jobId, stepNumber, iteration, consecutiveFailures);
+      const decision = await this.callAgent(
+        context,
+        jobId,
+        stepNumber,
+        iteration,
+        consecutiveFailures
+      );
       
       // Log agent's reasoning
-      this.logAgentDecision(decision, iteration);
+      this.decisionParser.log(decision, iteration);
       
       // Report progress
       await this.reportStepProgress(jobId, stepNumber, stepDescription, decision, iteration);
@@ -100,8 +129,39 @@ export class OrchestratorAgent {
       // Execute tools if requested (tools are READ-ONLY, they don't change state)
       let toolResults: Record<string, any> = {};
       
+      // ANTI-LOOP: Detect and BLOCK screenshot loops (PER STEP)
+      const screenshotsThisStep = memory.history.filter(s => 
+        s.stepNumber === stepNumber && 
+        (s.code.includes('take_screenshot') || s.action.includes('Screenshot'))
+      );
+      const recentScreenshots = memory.history.slice(-3).filter(s => 
+        s.code.includes('take_screenshot') || s.action.includes('Screenshot')
+      );
+      
+      if (screenshotsThisStep.length >= 3) {
+        this.logger?.(`[Orchestrator] 🚨 SCREENSHOT LOOP - ${screenshotsThisStep.length} screenshots THIS STEP! BLOCKING further screenshots`, 'error');
+      } else if (recentScreenshots.length >= 2 && iteration >= 3) {
+        this.logger?.(`[Orchestrator] ⚠️  WARNING: ${recentScreenshots.length} screenshots in last 3 iterations - agent may be looping`, 'warn');
+      }
+      
       if (decision.toolCalls && decision.toolCalls.length > 0) {
-        toolResults = await this.executeTools(decision.toolCalls, page, memory, stepNumber);
+        // ENFORCE: Block screenshot tool calls if too many taken IN THIS STEP
+        if (screenshotsThisStep.length >= 3) {
+          decision.toolCalls = decision.toolCalls.filter(tc => tc.name !== 'take_screenshot');
+          if (decision.toolCalls.length === 0) {
+            this.logger?.(`[Orchestrator] 🚫 REJECTED screenshot tool call - loop detected. Agent must ACT.`, 'warn');
+            toolResults = [{
+              toolName: 'take_screenshot',
+              success: false,
+              error: 'SYSTEM BLOCKED: Too many screenshots taken. You must use existing DOM snapshots and execute commands now. Analysis paralysis detected.',
+              data: null
+            }];
+          }
+        }
+        
+        if (decision.toolCalls.length > 0) {
+          toolResults = await this.executeTools(decision.toolCalls, page, memory, stepNumber, context.currentPageInfo.refMap);
+        }
         
         // If agent wants to wait for tool results before proceeding, call agent again
         if (decision.needsToolResults) {
@@ -125,7 +185,7 @@ export class OrchestratorAgent {
         this.logger?.(`[Orchestrator] 🚧 BLOCKER DETECTED: ${decision.blockerDetected.description}`);
         this.logger?.(`[Orchestrator] 🧹 Clearing blocker with ${decision.blockerDetected.clearingCommands.length} command(s)...`);
         
-        const blockerResult = await this.executeCommandsSequentially(
+        const blockerResult = await this.executeCommands(
           decision.blockerDetected.clearingCommands,
           page,
           memory,
@@ -148,9 +208,9 @@ export class OrchestratorAgent {
         }
       }
       
-      // Execute main commands (only if no blocker failure or continueOnFailure is true)
-      if (decision.commands && decision.commands.length > 0 && !iterationHadFailure) {
-        const executeResult = await this.executeCommandsSequentially(
+      // Execute main commands (only if no blocker failure)
+      if (!iterationHadFailure && decision.commands && decision.commands.length > 0) {
+        const executeResult = await this.executeCommands(
           decision.commands,
           page,
           memory,
@@ -171,8 +231,161 @@ export class OrchestratorAgent {
         }
       }
       
+      // Handle coordinate-based actions (NEW - fallback when selectors fail) - ONLY if enabled
+      if (this.config.enableCoordinateMode && decision.coordinateAction && !iterationHadFailure) {
+        coordinateAttempts++;
+        
+        this.logger?.(`[Orchestrator] 🎯 Coordinate Action (attempt ${coordinateAttempts}/2): ${decision.coordinateAction.action} at (${decision.coordinateAction.xPercent}%, ${decision.coordinateAction.yPercent}%)`);
+        
+        try {
+          // BEFORE screenshot for visual verification
+          const beforeScreenshot = await page.screenshot({ encoding: 'base64', fullPage: false, type: 'jpeg', quality: 60 });
+          const beforeDataUrl = `data:image/jpeg;base64,${beforeScreenshot}`;
+          
+          // Generate Playwright commands from coordinate action
+          const coordCommands = await CoordinateConverter.generateCommands(decision.coordinateAction, page);
+          
+          this.logger?.(`[Orchestrator] Generated commands:`);
+          coordCommands.forEach(cmd => this.logger?.(`  ${cmd}`));
+          
+          // Execute coordinate commands
+          const coordResult = await this.executeCommands(
+            coordCommands,
+            page,
+            memory,
+            stepNumber,
+            iteration,
+            jobId
+          );
+          
+          commandsExecuted.push(...coordResult.executed);
+          
+          if (!coordResult.allSucceeded) {
+            this.logger?.(`[Orchestrator] ❌ Coordinate action failed (Playwright error)`);
+            consecutiveFailures++;
+            iterationHadFailure = true;
+            
+            // Give up after 2 coordinate attempts
+            if (coordinateAttempts >= 2) {
+              this.logger?.(`[Orchestrator] 🛑 Coordinate mode exhausted (2 attempts) - marking stuck`);
+              return {
+                success: false,
+                commands: commandsExecuted,
+                iterations: iteration,
+                terminationReason: 'agent_stuck',
+                memory,
+                error: 'Coordinate fallback failed after 2 attempts - unable to proceed'
+              };
+            }
+          } else {
+            this.logger?.(`[Orchestrator] ✅ Coordinate action succeeded (no Playwright error)`);
+            
+            // CRITICAL: Verify visually if goal was achieved (coordinates might have clicked wrong place)
+            // Wait for network idle (smarter than fixed timeout - waits as needed, max 10s)
+            try {
+              await page.waitForLoadState('networkidle', { timeout: 10000 });
+            } catch (e) {
+              // Network idle timeout - page may still be loading, but proceed with verification
+              this.logger?.(`[Orchestrator] ⚠️  Network idle timeout after 10s, proceeding with verification`, 'warn');
+            }
+            
+            const afterScreenshot = await page.screenshot({ encoding: 'base64', fullPage: false, type: 'jpeg', quality: 60 });
+            const afterDataUrl = `data:image/jpeg;base64,${afterScreenshot}`;
+            
+            this.logger?.(`[Orchestrator] 📸 Verifying coordinate action visually...`);
+            
+            const verificationRequest = {
+              model: 'gpt-5-mini',
+              systemPrompt: 'You are a visual verification expert for web automation. Compare before/after screenshots to determine if an action achieved its goal.',
+              userPrompt: `Goal: ${scenarioSteps[stepNumber - 1]}\n\nA coordinate-based action was just executed. Compare the BEFORE and AFTER screenshots.\n\nDid the action achieve the goal? Respond with JSON:\n{\n  "verified": boolean,\n  "reasoning": "What changed (or didn't change) between screenshots",\n  "visibleChanges": ["List of UI changes observed"],\n  "clickedWrongPlace": boolean\n}\n\nBe strict: Only return verified=true if you clearly see the expected change matching the goal.`,
+              images: [
+                { label: 'BEFORE', dataUrl: beforeDataUrl },
+                { label: 'AFTER', dataUrl: afterDataUrl }
+              ]
+            };
+            
+            const verificationResponse = await this.llmFacade.llmProvider.callLLM(verificationRequest);
+            const jsonMatch = verificationResponse.answer.match(/\{[\s\S]*\}/);
+            
+            if (jsonMatch) {
+              const verificationResult = JSON.parse(jsonMatch[0]);
+              const verified = verificationResult.verified === true;
+              const reasoning = verificationResult.reasoning || 'No reasoning provided';
+              const clickedWrongPlace = verificationResult.clickedWrongPlace === true;
+              
+              this.logger?.(`[Orchestrator] 📊 Visual verification: ${verified ? '✅ VERIFIED' : '❌ NOT VERIFIED'}`);
+              this.logger?.(`[Orchestrator] 💭 Reasoning: ${reasoning}`);
+              
+              if (verified) {
+                // Goal achieved! Can mark complete
+                consecutiveFailures = 0;
+                
+                // Store note for context
+                noteToSelf = {
+                  fromIteration: iteration,
+                  content: `Coordinate action verified successful: ${reasoning}`
+                };
+                memory.latestNote = noteToSelf;
+              } else {
+                // Goal NOT achieved despite no error
+                this.logger?.(`[Orchestrator] ⚠️  Coordinate click succeeded but goal NOT achieved`, 'warn');
+                consecutiveFailures++;
+                iterationHadFailure = true;
+                
+                // Store diagnostic note for next attempt
+                const diagnostic = clickedWrongPlace 
+                  ? `Clicked wrong place. ${reasoning}. Try different coordinates.`
+                  : `Action executed but goal not achieved. ${reasoning}. May need different approach.`;
+                
+                noteToSelf = {
+                  fromIteration: iteration,
+                  content: diagnostic
+                };
+                memory.latestNote = noteToSelf;
+                
+                // Give up after 2 coordinate attempts
+                if (coordinateAttempts >= 2) {
+                  this.logger?.(`[Orchestrator] 🛑 Coordinate mode exhausted (2 attempts, none achieved goal) - marking stuck`);
+                  return {
+                    success: false,
+                    commands: commandsExecuted,
+                    iterations: iteration,
+                    terminationReason: 'agent_stuck',
+                    memory,
+                    error: `Coordinate actions clicked but didn't achieve goal: ${reasoning}`
+                  };
+                }
+              }
+            } else {
+              this.logger?.(`[Orchestrator] ⚠️  Could not parse verification response - treating as unverified`, 'warn');
+              consecutiveFailures++;
+              iterationHadFailure = true;
+            }
+          }
+          
+        } catch (error: any) {
+          this.logger?.(`[Orchestrator] ❌ Coordinate action error: ${error.message}`, 'error');
+          consecutiveFailures++;
+          iterationHadFailure = true;
+          
+          // Give up after 2 coordinate attempts
+          if (coordinateAttempts >= 2) {
+            this.logger?.(`[Orchestrator] 🛑 Coordinate mode exhausted (2 attempts) - marking stuck`);
+            return {
+              success: false,
+              commands: commandsExecuted,
+              iterations: iteration,
+              terminationReason: 'agent_stuck',
+              memory,
+              error: 'Coordinate fallback failed after 2 attempts - unable to proceed'
+            };
+          }
+        }
+      }
+      
       // System-enforced stuck detection (agent might not detect it)
-      if (consecutiveFailures >= 3) {
+      // Allow 5 failures: 3 selector attempts + 2 coordinate attempts
+      if (consecutiveFailures >= 5) {
         this.logger?.(`[Orchestrator] 🛑 SYSTEM: ${consecutiveFailures} consecutive failures detected - forcing stuck`, 'warn');
         return {
           success: false,
@@ -205,23 +418,38 @@ export class OrchestratorAgent {
         }
       }
       
-      // Store self-reflection for next iteration
-      previousReflection = decision.selfReflection;
+      // Store note to future self (tactical memory across iterations AND steps)
+      if (decision.noteToFutureSelf) {
+        noteToSelf = {
+          fromIteration: iteration,
+          content: decision.noteToFutureSelf
+        };
+        memory.latestNote = noteToSelf;  // Persist in journey memory across steps
+        this.logger?.(`[Orchestrator] 📝 Note to self: ${decision.noteToFutureSelf}`);
+      }
       
       // Check termination
       if (decision.status !== 'continue') {
         this.logger?.(`[Orchestrator] 🎯 Status: ${decision.status}`);
         this.logger?.(`[Orchestrator] 💭 Reason: ${decision.statusReasoning}`);
         
-        return {
-          success: decision.status === 'complete',
-          commands: commandsExecuted,
-          iterations: iteration,
-          terminationReason: decision.status === 'complete' ? 'complete' :
-                           decision.status === 'stuck' ? 'agent_stuck' :
-                           'infeasible',
-          memory
-        };
+        // SAFETY CHECK: Don't allow "complete" if commands failed this iteration
+        if (decision.status === 'complete' && iterationHadFailure) {
+          this.logger?.(`[Orchestrator] ⚠️  OVERRIDE: Agent said "complete" but commands FAILED - forcing "continue"`, 'warn');
+          this.logger?.(`[Orchestrator] 💭 System: Commands must succeed before marking complete`);
+          // Don't return - continue to next iteration
+        } else {
+          // Valid termination
+          return {
+            success: decision.status === 'complete',
+            commands: commandsExecuted,
+            iterations: iteration,
+            terminationReason: decision.status === 'complete' ? 'complete' :
+                             decision.status === 'stuck' ? 'agent_stuck' :
+                             'infeasible',
+            memory
+          };
+        }
       }
     }
     
@@ -248,8 +476,10 @@ export class OrchestratorAgent {
     totalSteps: number,
     scenarioSteps: string[],
     memory: JourneyMemory,
-    previousReflection?: SelfReflection,
-    consecutiveFailures?: number
+    consecutiveFailures?: number,
+    noteFromPreviousIteration?: NoteToFutureSelf,
+    priorSteps?: string[],    // NEW: For repair mode
+    nextSteps?: string[]      // NEW: For repair mode
   ): Promise<AgentContext> {
     // Get fresh DOM
     const currentPageInfo = await getEnhancedPageInfo(page);
@@ -258,8 +488,41 @@ export class OrchestratorAgent {
     // Get recent steps
     const recentSteps = memory.history.slice(-this.config.recentStepsCount);
     
+    // SoM integration: Update markers and capture screenshot with visual IDs
+    let somScreenshot: string | undefined = undefined;
+    let somElementMap: string | undefined = undefined;
+    if (this.config.useSoM && this.somHandler) {
+      try {
+        if (!this.somHandler) {
+          this.somHandler = new PageSoMHandler(page, this.logger);
+        } else {
+          this.somHandler.setPage(page);
+        }
+        
+        // Wait briefly for page stability (handles first iteration + safety net for fast SPAs)
+        try {
+          await page.waitForLoadState('domcontentloaded', { timeout: 5000 });
+        } catch (error: any) {
+          // Page already loaded or timeout - continue
+        }
+        
+        // Update SoM markers
+        await this.somHandler.updateSom();
+        
+        // Get screenshot WITH markers (viewport only - agent can scroll or use take_screenshot for full page)
+        somScreenshot = await this.somHandler.getScreenshot(true, false, 60);
+        
+        // Get element map for disambiguation
+        somElementMap = this.somHandler.getSomElementMap();
+        
+        this.logger?.(`[Orchestrator] SoM screenshot captured for agent decision-making`, 'log');
+      } catch (error: any) {
+        this.logger?.(`[Orchestrator] Failed to capture SoM screenshot: ${error.message}`, 'error');
+      }
+    }
+    
     // Build context
-    return {
+    const context = {
       overallGoal: scenarioSteps.join('\n'),
       currentStepGoal,
       stepNumber,
@@ -271,8 +534,19 @@ export class OrchestratorAgent {
       recentSteps,
       experiences: memory.experiences,
       extractedData: memory.extractedData,
-      previousIterationGuidance: previousReflection
+      noteFromPreviousIteration,  // Pass tactical note from previous iteration
+      somScreenshot,  // SoM screenshot with visual markers (current)
+      somElementMap,  // SoM element details for disambiguation
+      priorSteps,     // NEW: Repair context (undefined for script gen)
+      nextSteps       // NEW: Repair context (undefined for script gen)
     };
+    
+    // Save current screenshot as previous for next iteration (for tool access)
+    if (somScreenshot) {
+      this.previousSomScreenshot = somScreenshot;
+    }
+    
+    return context;
   }
   
   /**
@@ -285,18 +559,50 @@ export class OrchestratorAgent {
     iteration: number,
     consecutiveFailures?: number
   ): Promise<AgentDecision> {
-    // Build prompt
-    const systemPrompt = this.buildSystemPrompt();
-    const userPrompt = this.buildUserPrompt(context, consecutiveFailures);
+    // Detect if coordinate mode should be activated
+    // Phase 1: Only 2 tiers (selectors → coordinates), so activate after 3 failures
+    // Phase 2: Will have 3 tiers (selectors → index → coordinates), threshold will be 5
+    const useCoordinateMode = consecutiveFailures !== undefined && consecutiveFailures >= 3;
+    
+    // Build appropriate system prompt based on mode
+    const toolDescriptions = this.toolRegistry.generateToolDescriptions();
+    let systemPrompt: string;
+    
+    if (this.config.useSoM) {
+      // SoM mode: Use visual element identification
+      systemPrompt = OrchestratorPrompts.buildSomSystemPrompt(this.config.somRestrictCoordinates);
+    } else if (useCoordinateMode) {
+      // Coordinate mode: Fallback when selectors fail
+      systemPrompt = OrchestratorPrompts.buildCoordinateSystemPrompt();
+    } else {
+      // Standard mode: DOM-based selectors
+      systemPrompt = OrchestratorPrompts.buildSystemPrompt(toolDescriptions, this.config.enableCoordinateMode);
+    }
+    
+    const userPrompt = OrchestratorPrompts.buildUserPrompt(context, consecutiveFailures, this.config.enableCoordinateMode);
+    
+    // Log prompt lengths for monitoring
+    const systemLength = systemPrompt.length;
+    const userLength = userPrompt.length;
+    const totalLength = systemLength + userLength;
+    const estimatedTokens = Math.ceil(totalLength / 4); // Rough estimate: 4 chars per token
+    
+    this.logger?.(`[Orchestrator] 📊 Prompt lengths: system=${systemLength} chars, user=${userLength} chars, total=${totalLength} chars (~${estimatedTokens} tokens)`, 'log');
     
     try {
       // Call LLM directly via provider
-      const llmRequest = {
+      const llmRequest: any = {
         model: DEFAULT_MODEL,
         systemPrompt,
         userPrompt
       };
       
+      // Include current SoM screenshot as image
+      if (context.somScreenshot) {
+        llmRequest.imageUrl = context.somScreenshot;
+        this.logger?.(`[Orchestrator] Including SoM screenshot in LLM request`, 'log');
+      }
+      
       const response = await this.llmFacade.llmProvider.callLLM(llmRequest);
       
       // Report token usage
@@ -318,7 +624,7 @@ export class OrchestratorAgent {
       }
       
       // Parse response
-      return this.parseAgentDecision(response.answer);
+      return this.decisionParser.parse(response.answer);
       
     } catch (error: any) {
       this.logger?.(`[Orchestrator] ✗ Agent call failed: ${error.message}`, 'error');
@@ -332,443 +638,6 @@ export class OrchestratorAgent {
     }
   }
   
-  /**
-   * Build system prompt with tool descriptions
-   */
-  private buildSystemPrompt(): string {
-    const toolDescriptions = this.toolRegistry.generateToolDescriptions();
-    
-    return `You are an intelligent test automation agent that executes web scenarios using Playwright.
-
-${toolDescriptions}
-
-YOUR RESPONSE FORMAT (JSON):
-{
-  "toolCalls": [{"name": "tool_name", "params": {...}}],  // Optional: Tools to use
-  "toolReasoning": "Why I need these tools",              // If using tools
-  "needsToolResults": true/false,                         // Wait for tool results before commands?
-  "commands": ["await page.click(...)", ...],             // Playwright commands to execute
-  "commandReasoning": "Why these commands",               // If executing commands
-  "selfReflection": {                                     // Guidance for your next iteration
-    "guidanceForNext": "What to focus on next",
-    "detectingLoop": false,                               // Set true if repeating same approach
-    "loopReasoning": "Why I think I'm looping"           // If detecting loop
-  },
-  "experiences": ["App-specific learning"],              // APP-SPECIFIC patterns (see rules below)
-  "status": "continue|complete|stuck|infeasible",        // Your decision
-  "statusReasoning": "Why this status",
-  "reasoning": "Overall reasoning for this iteration",
-  "blockerDetected": {                                   // OPTIONAL: If something blocks the current goal
-    "description": "What's blocking",                    // E.g., "Cookie consent modal"
-    "clearingCommands": ["await page.click(...)"]       // Commands to clear blocker FIRST
-  },
-  "stepReEvaluation": {                                  // OPTIONAL: If questioning step order (last resort)
-    "detected": true,                                    // Set true if you think step order is wrong
-    "issue": "prior_incomplete|already_done|wrong_order", // What's the issue?
-    "explanation": "Why step order seems wrong"          // Your reasoning
-  }
-}
-
-STATUS DECISION RULES (CRITICAL!):
-
-- "complete": ALL commands succeeded AND step goal achieved
-  ❌ NEVER mark complete if ANY command failed!
-  
-- "continue": Command failed OR need more actions
-  Use when: Need to retry with different selector, wait for something, or continue execution
-  
-- "stuck": Tried 3+ iterations, all failed
-  
-- "infeasible": Goal truly impossible (element doesn't exist anywhere)
-
-IF COMMAND FAILED → status must be "continue" (to retry) or "stuck" (if tried many times)
-IF COMMAND SUCCEEDED → status can be "complete" (if goal done) or "continue" (if more needed)
-
-STEP RE-EVALUATION (LAST RESORT - After 2+ Failures):
-
-When you've failed multiple times on current step, QUESTION YOUR ASSUMPTIONS:
-
-1. **LOOK BACK**: Did prior steps really complete?
-   - Check COMPLETED steps list vs current page state
-   - Example: Says "Login complete" but you see login form → PRIOR STEP FAILED
-   - If prior step incomplete: Execute commands to complete it, explain in reasoning
-
-2. **LOOK FORWARD**: Is current step already done or no longer relevant?
-   - Check if current step goal already achieved (element exists, state correct)
-   - Example: Step says "Click Submit" but form already submitted → STEP ALREADY DONE
-   - If already done: Mark "complete" and explain why
-
-3. **LOOK AHEAD**: Is next step more feasible?
-   - Sometimes UI flows differently than expected
-   - Example: Current step "Click Edit" fails, but next step "Fill Name" elements are visible
-   - If can do future step: Note in reasoning, but PREFER sticking to order
-
-STICK TO ORIGINAL PLAN unless clear evidence suggests otherwise.
-Only re-evaluate after repeated failures on same step.
-
-BLOCKER DETECTION (IMPORTANT!):
-
-Use "blockerDetected" when the CURRENT STEP GOAL is not immediately feasible because something unexpected is in the way:
-
-Examples of BLOCKERS:
-- Cookie consent modal appears before login
-- "What's New" tour overlay prevents clicking navigation
-- Email verification prompt blocks form submission
-- Session timeout modal requires re-login before continuing
-
-When you detect a blocker:
-1. Set "blockerDetected.description" to describe what's blocking
-2. Provide "clearingCommands" to dismiss/handle the blocker
-3. Set your main "commands" to execute AFTER blocker is cleared
-4. Agent will execute: clearingCommands FIRST, then your commands
-
-Example:
-{
-  "blockerDetected": {
-    "description": "Cookie consent modal blocking page interaction",
-    "clearingCommands": ["await page.getByRole('button', {name: 'Accept'}).click()"]
-  },
-  "commands": ["await page.getByRole('textbox', {name: 'Email'}).fill('user@example.com')"],
-  ...
-}
-
-DON'T use blockerDetected for:
-- Normal step execution (just use commands)
-- Wrong selectors (that's not a blocker, just fix the selector)
-- Missing elements (that's infeasible, not blocked)
-
-EXPERIENCES - APP-SPECIFIC PATTERNS ONLY:
-
-Guidelines:
-- Concise - combine related info into single learning
-- Focus on app quirks, not obvious practices
-- Max 1-2 experiences per iteration
-- Capture SEMANTIC SELECTOR patterns when discovered
-
-✅ GOOD - App-specific patterns:
-- "Login form fields accessible via getByLabel: 'Email' and 'Password'"
-- "Submit buttons consistently use role=button with text matching action"
-- "Form uses custom dropdowns (role=combobox) instead of native selects"
-- "Navigation items accessible via getByRole('link') with visible text"
-- "Site uses data-testid attributes consistently (pattern: action-element-name)"
-- "Input fields have clear placeholders - prefer getByPlaceholder over IDs"
-
-❌ BAD - Too specific or obvious:
-- Listing individual field selectors (combine them)
-- "Button exists with role=button" (obvious)
-- Generic automation advice
-- Noting auto-generated IDs like #«r3»-form-item (these are unreliable)
-
-COMBINE related learnings instead of listing each element separately.
-PRIORITIZE learning about semantic selector availability over CSS selectors.
-
-CRITICAL RULES FOR DECISION MAKING:
-
-1. **BE FOCUSED AND CONCISE** (CRITICAL!):
-   - Read the CURRENT STEP GOAL carefully - ONLY do what's explicitly asked for in THAT step
-   - DON'T add extra actions, verifications, or defensive checks not mentioned in the step goal
-   - If step says "Send a message" → ONLY send the message, DON'T verify it appeared
-   - If step says "Click Submit" → ONLY click, DON'T wait or verify submission success
-   - ONLY add expect() assertions if the step explicitly mentions "verify", "check", "ensure", or "confirm"
-   - Keep commands minimal - no unnecessary preparatory or validation actions
-
-2. **COMMAND EFFICIENCY**:
-   - DON'T click on input field before filling it - just use .fill() directly
-   - DON'T do preparatory actions unless necessary
-   - Example: page.fill works without clicking the input first
-
-2b. **VARIABLE SCOPE & AVAILABLE FUNCTIONS** (CRITICAL!):
-   - Commands in SAME BATCH share scope - variables persist within iteration
-   - To save data for later steps: Use extract_data tool, then access via extractedData in context
-   - Within same batch: Variables work across commands
-   - AVAILABLE IN CONTEXT: page (Playwright Page object) and expect (Playwright assertions)
-   - You can use expect() directly - it is already imported and available
-
-3. **WHEN COMMAND FAILS** (most important!):
-   - DON'T repeat the same command/selector
-   - DO analyze the error message
-   - DO look at DOM snapshot for SEMANTIC selectors (getByRole, getByLabel, getByText)
-   - AVOID falling back to auto-generated IDs with unicode characters (e.g., #«r3»-form-item)
-   - DO take screenshot if DOM doesn't help
-   - "Element is not a <select>" → It's a custom dropdown, use .click() then select from menu
-   - "Timeout waiting for locator" → Try semantic selectors from ARIA tree instead of CSS IDs
-
-4. **CHECK YOUR RECENT STEPS FOR PATTERNS**:
-   - If you see same selector failed 2+ times → YOU ARE LOOPING!
-   - Set detectingLoop: true and explain why
-   - Take screenshot to see what's actually on the page
-   - Use different SEMANTIC selector from DOM/ARIA tree (getByRole, getByLabel, getByText, data-testid)
-   - AVOID repeating with similar auto-generated ID patterns (e.g., #«r3»-form-item → #«r4»-form-item)
-
-5. **USE DOM SNAPSHOT** - ONLY USE WHAT'S ACTUALLY THERE (CRITICAL!):
-   - You get INTERACTIVE ELEMENTS list showing: position, tag, id, text, SUGGESTED SELECTORS
-   - ONLY use selectors from this list - DO NOT invent selectors with text/names not in the list!
-   - ❌ WRONG: List shows button "Submit" but you use getByRole('button', {name: 'SomeOtherText'})
-   - ✅ RIGHT: Use EXACT selector from the provided list
-   - ✅ RIGHT: If list shows "#submit-btn" → use that
-   - ✅ RIGHT: If list shows "getByRole('button', {name: 'X'})" → use exactly that
-   - If element you need is NOT in list → take_screenshot to find it
-   - ARIA tree shows roles/names that exist - inventing roles/names not in tree will fail!
-
-5b. **SELECTOR PREFERENCE** (CRITICAL - Playwright Best Practices):
-   ALWAYS prefer user-friendly, semantic selectors in this order:
-   
-   🏆 BEST (use these first):
-   1. page.getByRole('role', {name: 'text'})     → ✅ Accessible, semantic, resilient
-   2. page.getByLabel('label text')              → ✅ Great for form inputs
-   3. page.getByPlaceholder('placeholder')       → ✅ Good for inputs without labels
-   4. page.getByText('visible text')             → ✅ Clear and readable
-   5. page.getByTestId('test-id')                → ✅ Stable if available
-   
-   ⚠️ AVOID (last resort only):
-   - CSS selectors with auto-generated IDs: #r3-form-item, #«r3»-form-item
-   - CSS selectors with unicode: #«r4»-form-item
-   - Complex CSS paths: div.ant-form > div:nth-child(2)
-   
-   Examples:
-   ❌ BAD:  await page.fill('#«r3»-form-item', 'alice@example.com')
-   ✅ GOOD: await page.getByLabel('Email').fill('alice@example.com')
-   ✅ GOOD: await page.getByRole('textbox', {name: 'Email'}).fill('alice@example.com')
-   ✅ GOOD: await page.getByPlaceholder('Enter your email').fill('alice@example.com')
-   
-   ❌ BAD:  await page.click('#submit-btn-23847')
-   ✅ GOOD: await page.getByRole('button', {name: 'Submit'}).click()
-   ✅ GOOD: await page.getByText('Submit').click()
-   
-   WHY: Semantic selectors are more maintainable, self-documenting, and resilient to UI changes.
-   Auto-generated IDs (especially with unicode like «r3») break when component instances change.
-
-5c. **ASSERTIONS AND VERIFICATION** (CRITICAL - Use expect):
-   Use expect() for verification ONLY when the step explicitly asks for verification.
-   
-   🏆 WHEN TO USE expect():
-   - Step says "Verify message appears" → await expect(locator).toBeVisible()
-   - Step says "Check field is empty" → await expect(locator).toBeEmpty()
-   - Step says "Confirm URL changed" → await expect(page).toHaveURL('...')
-   - Step explicitly mentions: "verify", "check", "ensure", "confirm", "validate"
-   
-   ❌ WHEN NOT TO USE expect():
-   - Step says "Send a message" → DON'T add expect() for message visibility
-   - Step says "Click Submit" → DON'T add expect() for button state or result
-   - Step says "Fill Email field" → DON'T add expect() to verify it's filled
-   - Step says "Navigate to Dashboard" → DON'T add expect() for URL (unless step asks)
-   
-   🏆 expect() syntax (when needed):
-   - await expect(locator).toBeVisible()
-   - await expect(locator).toHaveText('expected text')
-   - await expect(locator).toHaveValue('expected value')
-   - await expect(locator).toBeEmpty()
-   - await expect(page).toHaveURL('expected-url')
-   
-   ❌ BAD - Manual checks:
-   const value = await page.getByPlaceholder('Message...').inputValue();
-   if (value !== '') throw new Error(\`Expected empty but found: '\${value}'\`);
-   
-   ✅ GOOD - Use expect (but ONLY if step asks to verify):
-   await expect(page.getByPlaceholder('Message...')).toBeEmpty();
-   
-   WHY: expect() has built-in retries and better error messages.
-   BUT: Only use it when verification is explicitly requested in the step goal.
-
-6. **TOOLS ARE READ-ONLY (Information Gathering)**:
-   AVAILABLE TOOLS (for information only):
-   - take_screenshot: Get visual context when DOM isn't enough or stuck 2+ times
-     * DEFAULT: Captures FULL PAGE (elements below fold included)
-     * Set isFullPage=false ONLY if you need just viewport
-     * Use for: Planning, finding hidden elements, verifying layout
-   - recall_history: Remember earlier steps in journey
-   - inspect_page: Force fresh DOM snapshot (rarely needed, already in context)
-   - check_page_ready: Verify page loaded (use sparingly!)
-   - extract_data: Save data from page to memory for later steps
-   
-   ALL STATE CHANGES USE PLAYWRIGHT COMMANDS:
-   - Navigation: await page.goto('url')
-   - Go back: await page.goBack()
-   - Refresh: await page.reload()
-   - Fill inputs: await page.fill() or page.getByRole().fill()
-   - Click buttons: await page.click() or page.getByRole().click()
-   - Wait: await page.waitForURL(), page.waitForLoadState()
-   
-   Tools gather info → Commands change state
-
-7. **ERROR-SPECIFIC RESPONSES**:
-   - "Timeout waiting for locator" → Selector doesn't exist, try different one from DOM
-   - "Element is not a <select>" → It's a custom dropdown (button with role=combobox), use .click()
-   - "Multiple elements" → Add more specific selector from DOM (use #id)
-
-8. **WHEN TO RUN COMMANDS vs TOOLS**:
-   - Confident about selectors from DOM → Run commands directly
-   - Unsure or failed 2+ times → Take screenshot first
-   - First iteration of a step → Usually can run commands from DOM`;
-  }
-  
-  /**
-   * Build user prompt with context
-   */
-  private buildUserPrompt(context: AgentContext, consecutiveFailures?: number): string {
-    const parts: string[] = [];
-    
-    parts.push('=== YOUR CURRENT CONTEXT ===\n');
-    
-    // System warnings for accumulated failures
-    if (consecutiveFailures && consecutiveFailures >= 2) {
-      parts.push(`⚠️⚠️⚠️ SYSTEM WARNING ⚠️⚠️⚠️`);
-      parts.push(`You've failed ${consecutiveFailures} iterations in a row!`);
-      parts.push(`Next failure will force STUCK. You MUST:`);
-      parts.push(`1. Take screenshot if you haven't already`);
-      parts.push(`2. Try completely different selector strategy`);
-      parts.push(`3. QUESTION ASSUMPTIONS: Am I at the right step?`);
-      parts.push(`   - LOOK BACK: Did prior steps really complete? Check page state vs COMPLETED list`);
-      parts.push(`   - LOOK FORWARD: Is current step already done? Check if goal already achieved`);
-      parts.push(`4. OR mark status="stuck" if truly can't proceed`);
-      parts.push(`⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️\n`);
-    }
-    
-    // Goals - make current step very prominent
-    parts.push(`🎯 CURRENT STEP GOAL (${context.stepNumber}/${context.totalSteps}):`);
-    parts.push(`${context.currentStepGoal}`);
-    parts.push(`^^ DO ONLY WHAT THIS STEP ASKS - NO EXTRA ACTIONS OR VERIFICATIONS ^^`);
-    parts.push(`^^ If step doesn't say "verify/check/confirm" → DON'T add expect() assertions ^^`);
-    parts.push(`^^ Mark "complete" ONLY if commands succeeded ^^`);
-    parts.push(``);
-    parts.push(`OVERALL SCENARIO: ${context.overallGoal}\n`);
-    
-    if (context.completedSteps.length > 0) {
-      parts.push(`COMPLETED: ${context.completedSteps.join(', ')}`);
-    }
-    if (context.remainingSteps.length > 0) {
-      parts.push(`REMAINING: ${context.remainingSteps.join(', ')}\n`);
-    }
-    
-    // Current page state
-    parts.push(`\nCURRENT PAGE:`);
-    parts.push(`URL: ${context.currentURL}`);
-    parts.push(`Title: ${context.currentPageInfo.title}`);
-    parts.push(`\nINTERACTIVE ELEMENTS (with positions and selectors):`);
-    parts.push(context.currentPageInfo.formattedElements);
-    parts.push(`\nARIA TREE (hierarchical structure):`);
-    parts.push(JSON.stringify(context.currentPageInfo.ariaSnapshot, null, 2).substring(0, 5000));
-    if (JSON.stringify(context.currentPageInfo.ariaSnapshot).length > 5000) {
-      parts.push('... (truncated)');
-    }
-    parts.push('')
-    
-    // Recent steps
-    if (context.recentSteps.length > 0) {
-      parts.push(`\nRECENT STEPS (last ${context.recentSteps.length}):`);
-      for (const step of context.recentSteps) {
-        const status = step.result === 'success' ? '✓' : '✗';
-        parts.push(`  ${status} ${step.stepNumber}.${step.iteration || ''} ${step.action}`);
-        parts.push(`     Code: ${step.code}`);
-        if (step.result === 'failure' && step.error) {
-          parts.push(`     ❌ ERROR: ${step.error}`);
-          parts.push(`     ^^ THIS SELECTOR FAILED - TRY DIFFERENT APPROACH ^^`);
-        } else {
-          parts.push(`     Result: ${step.observation}`);
-        }
-      }
-      parts.push('');
-      
-      // Detect repeated failures
-      const recentFailures = context.recentSteps.filter(s => s.result === 'failure');
-      if (recentFailures.length >= 2) {
-        const sameSelector = recentFailures.slice(-2).every((s, i, arr) => 
-          i === 0 || s.code === arr[i-1].code
-        );
-        if (sameSelector) {
-          parts.push(`⚠️ WARNING: You've tried the same selector multiple times and it failed!`);
-          parts.push(`   Last failed selector: ${recentFailures[recentFailures.length - 1].code}`);
-          parts.push(`   YOU MUST try a completely different selector this time!\n`);
-        }
-      }
-    }
-    
-    // Experiences
-    if (context.experiences.length > 0) {
-      parts.push(`\nLEARNINGS (${context.experiences.length}):`);
-      for (const exp of context.experiences) {
-        parts.push(`  - ${exp}`);
-      }
-      parts.push('');
-    }
-    
-    // Extracted data
-    if (Object.keys(context.extractedData).length > 0) {
-      parts.push(`\nEXTRACTED DATA:`);
-      for (const [key, value] of Object.entries(context.extractedData)) {
-        parts.push(`  ${key}: ${value}`);
-      }
-      parts.push('');
-    }
-    
-    // Previous reflection
-    if (context.previousIterationGuidance) {
-      parts.push(`\nYOUR PREVIOUS GUIDANCE TO YOURSELF:`);
-      parts.push(`Focus: ${context.previousIterationGuidance.guidanceForNext}`);
-      if (context.previousIterationGuidance.detectingLoop) {
-        parts.push(`⚠ YOU DETECTED A LOOP: ${context.previousIterationGuidance.loopReasoning}`);
-      }
-      parts.push('');
-    }
-    
-    // Tool results (if any)
-    if (context.toolResults) {
-      parts.push(`\n========== TOOL RESULTS ==========`);
-      for (const [toolName, result] of Object.entries(context.toolResults)) {
-        parts.push(`\n${toolName}: ${result.success ? '✓ SUCCESS' : '✗ FAILED'}`);
-        
-        if (result.learning) {
-          parts.push(`\n📊 ANALYSIS FROM ${toolName.toUpperCase()}:`);
-          parts.push(result.learning);
-          parts.push(`^^ USE THIS INFORMATION TO DECIDE YOUR NEXT COMMANDS ^^`);
-        }
-        
-        if (result.error) {
-          parts.push(`  Error: ${result.error}`);
-        }
-      }
-      parts.push(`\n===================================\n`);
-    }
-    
-    parts.push('=== WHAT DO YOU WANT TO DO? ===');
-    parts.push('Respond with JSON following the format specified in the system prompt.');
-    
-    return parts.join('\n');
-  }
-  
-  /**
-   * Parse agent decision from LLM response
-   */
-  private parseAgentDecision(response: string): AgentDecision {
-    try {
-      // Extract JSON from response
-      const jsonMatch = response.match(/\{[\s\S]*\}/);
-      if (!jsonMatch) {
-        throw new Error('No JSON found in response');
-      }
-      
-      const parsed = JSON.parse(jsonMatch[0]);
-      
-      // Validate required fields
-      if (!parsed.status || !parsed.reasoning) {
-        throw new Error('Missing required fields: status and reasoning');
-      }
-      
-      return parsed as AgentDecision;
-      
-    } catch (error: any) {
-      this.logger?.(`[Orchestrator] ✗ Failed to parse agent decision: ${error.message}`, 'error');
-      this.logger?.(`[Orchestrator] Raw response: ${response.substring(0, 200)}...`, 'log');
-      
-      // Return fallback
-      return {
-        status: 'stuck',
-        statusReasoning: 'Failed to parse agent response',
-        reasoning: `Parse error: ${error.message}`
-      };
-    }
-  }
-  
   /**
    * Execute tools
    */
@@ -776,16 +645,20 @@ CRITICAL RULES FOR DECISION MAKING:
     toolCalls: any[],
     page: any,
     memory: JourneyMemory,
-    stepNumber: number
+    stepNumber: number,
+    refMap?: Map<string, any>
   ): Promise<Record<string, any>> {
     this.logger?.(`[Orchestrator] 🔧 Executing ${toolCalls.length} tool(s)`);
     
     const results: Record<string, any> = {};
-    const toolContext: ToolExecutionContext = {
+    const toolContext: ToolExecutionContext & { refMap?: Map<string, any>; previousSomScreenshot?: string; somHandler?: any } = {
       page,
       memory,
       stepNumber,
-      logger: this.logger
+      logger: this.logger,
+      refMap,  // Pass refMap for interact_with_ref tool
+      previousSomScreenshot: this.previousSomScreenshot,  // For view_previous_screenshot tool
+      somHandler: this.somHandler  // For refresh_som_markers tool
     };
     
     for (const toolCall of toolCalls.slice(0, this.config.maxToolCallsPerIteration)) {
@@ -805,71 +678,259 @@ CRITICAL RULES FOR DECISION MAKING:
   }
   
   /**
-   * Execute commands sequentially with SHARED context (variables persist across commands)
+   * Parse SomCommand from command object
    */
-  private async executeCommandsSequentially(
-    commands: string[],
+  private parseSomCommand(cmd: any): SomCommand | null {
+    if (typeof cmd === 'object' && cmd.action) {
+      // Valid if: has elementRef, OR has coord, OR is navigation action
+      const isNavigationAction = ['navigate', 'goBack', 'goForward', 'reload'].includes(cmd.action);
+      const hasValidTarget = cmd.elementRef || cmd.coord || isNavigationAction;
+      
+      if (hasValidTarget) {
+        return {
+          elementRef: cmd.elementRef,
+          coord: cmd.coord,
+          action: cmd.action,
+          value: cmd.value,
+          fromCoord: cmd.fromCoord,
+          toCoord: cmd.toCoord,
+          force: cmd.force,
+          scrollAmount: cmd.scrollAmount,
+          scrollDirection: cmd.scrollDirection,
+          button: cmd.button,
+          clickCount: cmd.clickCount,
+          modifiers: cmd.modifiers,
+          delay: cmd.delay,
+          timeout: cmd.timeout
+        };
+      }
+    }
+    return null;
+  }
+  
+  /**
+   * Execute commands (mix of ref and playwright commands)
+   */
+  private async executeCommands(
+    commands: string[] | any[],
     page: any,
     memory: JourneyMemory,
     stepNumber: number,
     iteration: number,
     jobId: string
   ): Promise<{ executed: string[]; allSucceeded: boolean }> {
-    this.logger?.(`[Orchestrator] 📝 Executing ${commands.length} command(s) in shared context`);
+    this.logger?.(`[Orchestrator] 📝 Executing ${commands.length} command(s)`);
     
     const executed: string[] = [];
-    const limitedCommands = commands.slice(0, this.config.maxCommandsPerIteration);
     
-    // Build execution with shared context (all commands share scope - variables persist)
-    const commandsWithTracking = limitedCommands.map((cmd, i) => {
-      return `
-// Command ${i + 1}/${limitedCommands.length}
+    if (commands.length === 0) {
+      return { executed: [], allSucceeded: true };
+    }
+    
+    // SoM mode: Execute commands through PageSoMHandler
+    if (this.config.useSoM && this.somHandler) {
+      this.logger?.(`[Orchestrator] Using SoM mode for command execution`, 'log');
+      
+      for (let i = 0; i < commands.length; i++) {
+        const cmd = commands[i];
+        
+        // Check if verification or action command
+        if (isSomVerification(cmd)) {
+          // Handle verification command
+          try {
+            const result = await this.somHandler.executeVerification(cmd);
+            
+            // Always add command to executed array (even if verification failed)
+            // Scripts should contain the expect even if it fails during generation
+            if (result.playwrightCommand) {
+              executed.push(result.playwrightCommand);
+            }
+            
+            if (result.success) {
+              this.logger?.(`[Orchestrator]   ✓ [${i + 1}/${commands.length}] Verification passed`, 'log');
+              
+              memory.history.push({
+                stepNumber,
+                iteration,
+                action: `Verification ${i + 1}/${commands.length}: ${cmd.verificationType}`,
+                code: result.playwrightCommand,
+                result: 'success',
+                observation: `Verified: ${cmd.description || cmd.expected}`,
+                url: page.url(),
+                timestamp: Date.now()
+              });
+            } else {
+              this.logger?.(`[Orchestrator]   ✗ [${i + 1}/${commands.length}] Verification failed (non-fatal): ${result.error}`, 'warn');
+              
+              memory.history.push({
+                stepNumber,
+                iteration,
+                action: `Verification ${i + 1}/${commands.length} - FAILED`,
+                code: result.playwrightCommand || JSON.stringify(cmd),
+                result: 'failure',
+                observation: `Failed: ${result.error}`,
+                error: result.error,
+                url: page.url(),
+                timestamp: Date.now()
+              });
+              
+              // Continue anyway - verification failures are non-blocking for script generation
+            }
+            
+            // Small delay between commands
+            if (i < commands.length - 1) {
+              await page.waitForTimeout(300);
+            }
+            
+          } catch (error: any) {
+            this.logger?.(`[Orchestrator]   ✗ [${i + 1}/${commands.length}] Verification exception: ${error.message}`, 'error');
+          }
+          
+        } else if (isSomCommand(cmd)) {
+          // Handle action command (existing logic)
+          const somCommand = cmd as SomCommand;
+          
+          try {
+            const result = await this.somHandler.runCommand(
+              somCommand,
+              this.config.somUseSomIdBasedCommands || false
+            );
+            
+            if (result.status === CommandRunStatus.SUCCESS && result.successAttempt) {
+              this.logger?.(`[Orchestrator]   ✓ [${i + 1}/${commands.length}] SoM action succeeded`, 'log');
+              executed.push(result.successAttempt.command!);
+              
+              memory.history.push({
+                stepNumber,
+                iteration,
+                action: `SoM Action ${i + 1}/${commands.length}: ${somCommand.action}`,
+                code: result.successAttempt.command!,
+                result: 'success',
+                observation: 'Executed successfully',
+                url: page.url(),
+                timestamp: Date.now()
+              });
+              
+              // Small delay for form validation/animations
+              if (i < commands.length - 1) {
+                await page.waitForTimeout(300);
+              }
+            } else {
+              this.logger?.(`[Orchestrator]   ✗ [${i + 1}/${commands.length}] SoM action failed: ${result.error}`, 'error');
+              
+              memory.history.push({
+                stepNumber,
+                iteration,
+                action: `SoM Action ${i + 1}/${commands.length}: ${somCommand.action} - FAILED`,
+                code: JSON.stringify(somCommand),
+                result: 'failure',
+                observation: `Failed: ${result.error}`,
+                error: result.error,
+                url: page.url(),
+                timestamp: Date.now()
+              });
+              
+              // Refresh SoM after batch (DOM may have changed)
+              if (this.somHandler && page) {
+                this.somHandler.setPage(page);
+                await this.somHandler.updateSom();
+              }
+              
+              return { executed, allSucceeded: false };
+            }
+          } catch (error: any) {
+            this.logger?.(`[Orchestrator]   ✗ [${i + 1}/${commands.length}] SoM action exception: ${error.message}`, 'error');
+            
+            memory.history.push({
+              stepNumber,
+              iteration,
+              action: `SoM Action ${i + 1}/${commands.length} - EXCEPTION`,
+              code: JSON.stringify(somCommand),
+              result: 'failure',
+              observation: `Exception: ${error.message}`,
+              error: error.message,
+              url: page.url(),
+              timestamp: Date.now()
+            });
+            
+            // Refresh SoM after batch (DOM may have changed)
+            if (this.somHandler && page) {
+              this.somHandler.setPage(page);
+              await this.somHandler.updateSom();
+            }
+            
+            return { executed, allSucceeded: false };
+          }
+        } else {
+          this.logger?.(`[Orchestrator]   ⚠ [${i + 1}/${commands.length}] Not a valid SoM command/verification, skipping`, 'warn');
+        }
+      }
+      
+      // Always wait for page to stabilize after command batch
+      // This handles both explicit navigation AND clicks that trigger navigation/SPA routes
+      try {
+        this.logger?.(`[Orchestrator] Waiting for page to stabilize...`, 'log');
+        // Use networkidle with short timeout to catch navigation without blocking on SPAs with continuous requests
+        await page.waitForLoadState('networkidle', { timeout: 3000 });
+        this.logger?.(`[Orchestrator] Page stabilized (networkidle)`, 'log');
+      } catch (error: any) {
+        // If networkidle times out, fall back to domcontentloaded
+        try {
+          await page.waitForLoadState('domcontentloaded', { timeout: 2000 });
+          this.logger?.(`[Orchestrator] Page loaded (domcontentloaded)`, 'log');
+        } catch (error2: any) {
+          this.logger?.(`[Orchestrator] Page load wait timeout (continuing anyway)`, 'warn');
+        }
+      }
+      
+      // Refresh SoM after batch (DOM may have changed and page is now stable)
+      if (this.somHandler && page) {
+        this.somHandler.setPage(page);
+        await this.somHandler.updateSom();
+      }
+      
+      return { executed, allSucceeded: true };
+    }
+    
+    // Standard mode: Execute all commands in sequence with small delay between them
+    // Delay helps with form validation, button enabling, and animations
+    const wrappedCode = (commands as string[]).map((cmd, i) => `
+// Command ${i + 1}/${commands.length}
 try {
   ${cmd}
   __results.push({ index: ${i}, success: true });
+  ${i < commands.length - 1 ? 'await page.waitForTimeout(300);' : ''} // Small delay for form validation/animations
 } catch (error) {
   __results.push({ index: ${i}, success: false, error: error.message });
-  throw error; // Stop on first failure
-}`;
-    }).join('\n');
+  throw error;
+}`).join('\n');
     
-    const wrappedCode = `
-const __results = [];
-${commandsWithTracking}
-return __results;
-`;
+    const fullCode = `const __results = []; ${wrappedCode} return __results;`;
     
     try {
-      // Execute in shared context - variables declared here persist for entire scenario
-      // Pass both page and expect to make Playwright assertions available
-      const func = new Function('page', 'expect', 'return (async () => { ' + wrappedCode + ' })()');
-      const results = await func(page, expect);
+      const func = new Function('page', 'expect', 'return (async () => { ' + fullCode + ' })()');
+      const results = await func(page, (global as any).expect);
       
-      // Record results for each command
-      for (let i = 0; i < limitedCommands.length; i++) {
-        const cmd = limitedCommands[i];
+      for (let i = 0; i < commands.length; i++) {
+        const cmd = commands[i];
         const result = results[i];
         
         if (result && result.success) {
-          this.logger?.(`[Orchestrator]   ✓ [${i + 1}/${limitedCommands.length}] Success`);
-          
-          // Record in history
+          this.logger?.(`[Orchestrator]   ✓ [${i + 1}/${commands.length}] Success`);
           memory.history.push({
             stepNumber,
             iteration,
-            action: `Command ${i + 1}/${limitedCommands.length}`,
+            action: `Command ${i + 1}/${commands.length}`,
             code: cmd,
             result: 'success',
             observation: 'Executed successfully',
             url: page.url(),
             timestamp: Date.now()
           });
-          
           executed.push(cmd);
         }
       }
       
-      // Cap history
       if (memory.history.length > this.config.maxHistorySize) {
         memory.history = memory.history.slice(-this.config.maxHistorySize);
       }
@@ -877,132 +938,25 @@ return __results;
       return { executed, allSucceeded: true };
       
     } catch (error: any) {
-      // One of the commands failed - find which one
       const errorMessage = error.message || String(error);
+      this.logger?.(`[Orchestrator] ❌ Command execution failed: ${errorMessage}`, 'error');
       
-      // Capture page state for debug logging
-      let pageStateDebug = '';
-      if (this.debugMode) {
-        try {
-          const pageInfo = await getEnhancedPageInfo(page);
-          pageStateDebug = `
-=== DEBUG: PAGE STATE AT FAILURE ===
-URL: ${page.url()}
-Title: ${pageInfo.title}
-
-INTERACTIVE ELEMENTS:
-${pageInfo.formattedElements}
-
-ARIA SNAPSHOT:
-${JSON.stringify(pageInfo.ariaSnapshot, null, 2)}
-
-====================================`;
-        } catch (debugError: any) {
-          pageStateDebug = `Failed to capture page state: ${debugError.message}`;
-        }
-      }
-      
-      // Record all that succeeded, then the failure
-      for (let i = 0; i < limitedCommands.length; i++) {
-        const cmd = limitedCommands[i];
-        
-        // This is a failed command (error happened here or earlier)
-        if (executed.length <= i) {
-          this.logger?.(`[Orchestrator]   ✗ [${i + 1}/${limitedCommands.length}] Failed: ${errorMessage}`, 'error');
-          
-          // Log detailed debug info
-          if (this.debugMode && pageStateDebug) {
-            this.logger?.(pageStateDebug, 'debug');
-          }
-          
-          memory.history.push({
-            stepNumber,
-            iteration,
-            action: `Command ${i + 1}/${limitedCommands.length} - FAILED`,
-            code: cmd,
-            result: 'failure',
-            observation: `Failed with error: ${errorMessage}. This selector likely doesn't exist or is incorrect.`,
-            error: errorMessage,
-            url: page.url(),
-            timestamp: Date.now()
-          });
-          
-          if (i < limitedCommands.length - 1) {
-            this.logger?.(`[Orchestrator]   ⚠ Skipping remaining ${limitedCommands.length - i - 1} command(s)`, 'warn');
-          }
-          
-          break;
-        }
-      }
+      memory.history.push({
+        stepNumber,
+        iteration,
+        action: `Command - FAILED`,
+        code: commands[executed.length] || '',
+        result: 'failure',
+        observation: `Failed: ${errorMessage}`,
+        error: errorMessage,
+        url: page.url(),
+        timestamp: Date.now()
+      });
       
       return { executed, allSucceeded: false };
     }
   }
-  
-  /**
-   * Execute a single command
-   */
-  private async executeCommand(cmd: string, page: any): Promise<void> {
-    // Wrap in async function and execute
-    const wrapped = `(async () => { ${cmd} })()`;
-    
-    try {
-      await eval(wrapped);
-    } catch (error: any) {
-      // If eval fails, try direct execution with page context
-      // Pass both page and expect to make Playwright assertions available
-      const func = new Function('page', 'expect', `return (async () => { ${cmd} })()`);
-      await func(page, expect);
-    }
-  }
-  
-  /**
-   * Log agent decision
-   */
-  private logAgentDecision(decision: AgentDecision, iteration: number): void {
-    this.logger?.(`[Orchestrator] 💭 REASONING: ${decision.reasoning}`);
-    
-    if (decision.selfReflection) {
-      this.logger?.(`[Orchestrator] 🧠 SELF-REFLECTION:`);
-      this.logger?.(`[Orchestrator]    Next: ${decision.selfReflection.guidanceForNext}`);
-      if (decision.selfReflection.detectingLoop) {
-        this.logger?.(`[Orchestrator]    🔄 LOOP DETECTED: ${decision.selfReflection.loopReasoning}`, 'warn');
-      }
-    }
-    
-    if (decision.toolCalls && decision.toolCalls.length > 0) {
-      this.logger?.(`[Orchestrator] 🔧 TOOLS: ${decision.toolCalls.map(t => t.name).join(', ')}`);
-      if (decision.toolReasoning) {
-        this.logger?.(`[Orchestrator] 📋 Why: ${decision.toolReasoning}`);
-      }
-    }
-    
-    if (decision.blockerDetected) {
-      this.logger?.(`[Orchestrator] 🚧 BLOCKER: ${decision.blockerDetected.description}`, 'warn');
-      this.logger?.(`[Orchestrator] 🧹 Clearing with ${decision.blockerDetected.clearingCommands.length} command(s)`);
-    }
-    
-    if (decision.stepReEvaluation?.detected) {
-      this.logger?.(`[Orchestrator] 🔍 STEP RE-EVALUATION: ${decision.stepReEvaluation.issue}`, 'warn');
-      this.logger?.(`[Orchestrator] 📝 Explanation: ${decision.stepReEvaluation.explanation}`);
-    }
-    
-    if (decision.commands && decision.commands.length > 0) {
-      this.logger?.(`[Orchestrator] 📝 COMMANDS (${decision.commands.length}):`);
-      decision.commands.slice(0, 3).forEach((cmd, i) => {
-        this.logger?.(`[Orchestrator]    ${i + 1}. ${cmd.substring(0, 80)}...`);
-      });
-      if (decision.commands.length > 3) {
-        this.logger?.(`[Orchestrator]    ... and ${decision.commands.length - 3} more`);
-      }
-      if (decision.commandReasoning) {
-        this.logger?.(`[Orchestrator] 💡 Why: ${decision.commandReasoning}`);
-      }
-    }
-    
-    // Experiences will be logged when added to memory, no need to log here
-  }
-  
+
   /**
    * Report step progress
    */
@@ -1032,6 +986,297 @@ ${JSON.stringify(pageInfo.ariaSnapshot, null, 2)}
       agentStatus: decision.status
     });
   }
+  
+  /**
+   * Execute exploration mode - agent autonomously explores to achieve journey goal
+   * Fires onStepProgress callbacks for each autonomous action (transparent to caller)
+   */
+  async executeExploration(
+    page: any,
+    explorationConfig: ExplorationMode,
+    jobId: string
+  ): Promise<OrchestratorStepResult> {
+    this.logger?.(`\n[Orchestrator] ========== EXPLORATION MODE ==========`);
+    this.logger?.(`[Orchestrator] 🎯 Journey Goal: ${explorationConfig.explorationPrompt}`);
+    if (explorationConfig.testDataPrompt) {
+      this.logger?.(`[Orchestrator] 📋 Test Data: ${explorationConfig.testDataPrompt}`);
+    }
+    
+    const memory: JourneyMemory = {
+      history: [],
+      experiences: [],
+      extractedData: {}
+    };
+    
+    const maxSteps = explorationConfig.maxExplorationSteps || 50;
+    let stepNumber = 0;
+    const commandsExecuted: string[] = [];
+    
+    while (stepNumber < maxSteps) {
+      stepNumber++;
+      
+      this.logger?.(`\n[Orchestrator] === Exploration Step ${stepNumber}/${maxSteps} ===`);
+      
+      // Build exploratory context
+      const context = await this.buildExploratoryContext(
+        page,
+        explorationConfig.explorationPrompt,
+        explorationConfig.testDataPrompt,
+        memory,
+        stepNumber,
+        maxSteps
+      );
+      
+      // Call agent with exploratory prompt
+      const decision = await this.callExploratoryAgent(
+        context,
+        jobId,
+        stepNumber
+      );
+      
+      this.decisionParser.log(decision, stepNumber);
+      
+      // Report step start (fires JourneyRunner's beforeStepStart callback)
+      if (this.progressReporter?.onStepProgress) {
+        const stepInfo = {
+          jobId,
+          stepNumber,
+          stepId: `exploration-${stepNumber}-${Date.now()}`,
+          description: decision.reasoning,
+          code: '',  // Will be filled after commands execute
+          status: StepExecutionStatus.IN_PROGRESS,
+          wasRepaired: false
+        };
+        await this.progressReporter.onStepProgress(stepInfo);
+      }
+      
+      // Execute tools if requested
+      if (decision.toolCalls && decision.toolCalls.length > 0) {
+        const toolResults = await this.executeTools(decision.toolCalls, page, memory, stepNumber);
+        
+        // If needs tool results, call agent again
+        if (decision.needsToolResults) {
+          const updatedContext = { ...context, toolResults };
+          const continuedDecision = await this.callExploratoryAgent(updatedContext, jobId, stepNumber);
+          
+          decision.commands = continuedDecision.commands || decision.commands;
+          decision.commandReasoning = continuedDecision.commandReasoning || decision.commandReasoning;
+          decision.status = continuedDecision.status;
+        }
+      }
+      
+      // Handle blocker clearing
+      if (decision.blockerDetected && decision.blockerDetected.clearingCommands) {
+        this.logger?.(`[Orchestrator] 🚧 Clearing blocker: ${decision.blockerDetected.description}`);
+        const blockerResult = await this.executeCommands(
+          decision.blockerDetected.clearingCommands,
+          page,
+          memory,
+          stepNumber,
+          1,
+          jobId
+        );
+        commandsExecuted.push(...blockerResult.executed);
+      }
+      
+      // Execute exploration commands
+      let commandsSucceeded = true;
+      if (decision.commands && decision.commands.length > 0) {
+        const executeResult = await this.executeCommands(
+          decision.commands,
+          page,
+          memory,
+          stepNumber,
+          1,
+          jobId
+        );
+        commandsExecuted.push(...executeResult.executed);
+        commandsSucceeded = executeResult.allSucceeded;
+      }
+      
+      // Report step completion (fires JourneyRunner's onStepComplete callback)
+      if (this.progressReporter?.onStepProgress) {
+        const stepInfo = {
+          jobId,
+          stepNumber,
+          stepId: `exploration-${stepNumber}-${Date.now()}`,
+          description: decision.reasoning,
+          code: decision.commands?.join('\n') || '',
+          status: commandsSucceeded ? StepExecutionStatus.SUCCESS : StepExecutionStatus.FAILURE,
+          error: commandsSucceeded ? undefined : 'Command execution failed',
+          wasRepaired: false
+        };
+        await this.progressReporter.onStepProgress(stepInfo);
+      }
+      
+      // Add experiences (both app patterns AND exploration progress)
+      if (decision.experiences) {
+        memory.experiences.push(...decision.experiences);
+        if (memory.experiences.length > this.config.maxExperiences) {
+          memory.experiences = memory.experiences.slice(-this.config.maxExperiences);
+        }
+      }
+      
+      // Store note for next iteration
+      if (decision.noteToFutureSelf) {
+        memory.latestNote = {
+          fromIteration: stepNumber,
+          content: decision.noteToFutureSelf
+        };
+      }
+      
+      // Check termination
+      if (decision.status === 'complete') {
+        this.logger?.(`[Orchestrator] ✅ Journey exploration complete: ${decision.statusReasoning}`);
+        return {
+          success: true,
+          commands: commandsExecuted,
+          iterations: stepNumber,
+          terminationReason: 'complete',
+          memory
+        };
+      } else if (decision.status === 'stuck') {
+        this.logger?.(`[Orchestrator] ❌ Exploration stuck: ${decision.statusReasoning}`);
+        return {
+          success: false,
+          commands: commandsExecuted,
+          iterations: stepNumber,
+          terminationReason: 'agent_stuck',
+          memory,
+          error: decision.statusReasoning
+        };
+      }
+    }
+    
+    // Hit max steps - not necessarily a failure
+    this.logger?.(`[Orchestrator] ⚠ Maximum exploration steps reached (budget limit)`);
+    return {
+      success: true,  // Not a failure - just budget limit
+      commands: commandsExecuted,
+      iterations: stepNumber,
+      terminationReason: 'system_limit',
+      memory
+    };
+  }
+  
+  private async buildExploratoryContext(
+    page: any,
+    explorationPrompt: string,
+    testDataPrompt: string | undefined,
+    memory: JourneyMemory,
+    stepNumber: number,
+    maxSteps: number
+  ): Promise<AgentContext> {
+    // Wait for page to be ready and elements to appear (especially important after navigation)
+    const currentPageInfo = await PageInfoRetry.getWithRetry(page);
+    const currentURL = page.url();
+    const recentSteps = memory.history.slice(-this.config.recentStepsCount);
+    
+    // SoM integration for exploratory mode
+    let somScreenshot: string | undefined = undefined;
+    let somElementMap: string | undefined = undefined;
+    if (this.config.useSoM && this.somHandler) {
+      try {
+        this.somHandler.setPage(page);
+        
+        // Wait briefly for page stability (handles first iteration + safety net for fast SPAs)
+        try {
+          await page.waitForLoadState('domcontentloaded', { timeout: 2000 });
+        } catch (error: any) {
+          // Page already loaded or timeout - continue
+        }
+        
+        // Update SoM markers
+        await this.somHandler.updateSom();
+        somScreenshot = await this.somHandler.getScreenshot(true, false, 60);  // Viewport only - agent can scroll or request full page
+        
+        // Get element map for disambiguation
+        somElementMap = this.somHandler.getSomElementMap();
+        
+        this.logger?.(`[Orchestrator] SoM screenshot captured for exploratory agent`, 'log');
+      } catch (error: any) {
+        this.logger?.(`[Orchestrator] Failed to capture SoM screenshot: ${error.message}`, 'error');
+      }
+    }
+    
+    const context = {
+      overallGoal: explorationPrompt,
+      currentStepGoal: explorationPrompt,  // Same as overall for single journey
+      stepNumber,
+      totalSteps: maxSteps,
+      completedSteps: [],
+      remainingSteps: [],
+      currentPageInfo,
+      currentURL,
+      recentSteps,
+      experiences: memory.experiences,
+      extractedData: memory.extractedData,
+      noteFromPreviousIteration: memory.latestNote,
+      testDataPrompt,  // CRITICAL: Store testDataPrompt in context
+      somScreenshot,  // SoM screenshot for exploratory mode (current)
+      somElementMap   // SoM element details for disambiguation
+    };
+    
+    // Save current screenshot as previous for next iteration (for tool access)
+    if (somScreenshot) {
+      this.previousSomScreenshot = somScreenshot;
+    }
+    
+    return context;
+  }
+  
+  private async callExploratoryAgent(
+    context: AgentContext,
+    jobId: string,
+    stepNumber: number
+  ): Promise<AgentDecision> {
+    const toolDescriptions = this.toolRegistry.generateToolDescriptions();
+    
+    // Use SoM system prompt if in SoM mode, otherwise use standard exploratory prompt
+    const systemPrompt = this.config.useSoM
+      ? OrchestratorPrompts.buildSomSystemPrompt(this.config.somRestrictCoordinates)
+      : OrchestratorPrompts.buildExploratorySystemPrompt(toolDescriptions);
+      
+    const userPrompt = OrchestratorPrompts.buildExploratoryUserPrompt(
+      context,
+      context.overallGoal,
+      context.testDataPrompt,  // Pass testDataPrompt from context
+      stepNumber,
+      context.totalSteps
+    );
+    
+    const llmRequest: any = {
+      model: DEFAULT_MODEL,
+      systemPrompt,
+      userPrompt
+    };
+    
+    // Include current SoM screenshot as image
+    if (context.somScreenshot) {
+      llmRequest.imageUrl = context.somScreenshot;
+      this.logger?.(`[Orchestrator] Including SoM screenshot in exploratory LLM request`, 'log');
+    }
+    
+    const response = await this.llmFacade.llmProvider.callLLM(llmRequest);
+    
+    // Report token usage
+    if (response.usage && this.progressReporter?.onTokensUsed) {
+      await this.progressReporter.onTokensUsed({
+        jobId,
+        stepNumber,
+        iteration: 1,
+        inputTokens: response.usage.inputTokens,
+        outputTokens: response.usage.outputTokens,
+        includesImage: false,
+        model: DEFAULT_MODEL,
+        timestamp: Date.now()
+      });
+    }
+    
+    // Parse response (same JSON format as regular mode)
+    const decision = this.decisionParser.parse(response.answer);
+    return decision;
+  }
 }