fix(core): Prevent loop detection false positives on lists with long shared prefixes (#18975)

packages/core/src/services/loopDetectionService.test.ts

@@ -210,6 +210,25 @@ describe('LoopDetectionService', () => {
       expect(loggers.logLoopDetected).toHaveBeenCalledTimes(1);
     });
 
+    it('should not detect a loop for a list with a long shared prefix', () => {
+      service.reset('');
+      let isLoop = false;
+      const longPrefix =
+        'projects/my-google-cloud-project-12345/locations/us-central1/services/';
+
+      let listContent = '';
+      for (let i = 0; i < 15; i++) {
+        listContent += `- ${longPrefix}${i}\n`;
+      }
+
+      // Simulate receiving the list in a single large chunk or a few chunks
+      // This is the specific case where the issue occurs, as list boundaries might not reset tracking properly
+      isLoop = service.addAndCheck(createContentEvent(listContent));
+
+      expect(isLoop).toBe(false);
+      expect(loggers.logLoopDetected).not.toHaveBeenCalled();
+    });
+
     it('should not detect a loop if repetitions are very far apart', () => {
       service.reset('');
       const repeatedContent = createRepetitiveContent(1, CONTENT_CHUNK_SIZE);

packages/core/src/services/loopDetectionService.ts

@@ -379,7 +379,30 @@ export class LoopDetectionService {
     const averageDistance = totalDistance / (CONTENT_LOOP_THRESHOLD - 1);
     const maxAllowedDistance = CONTENT_CHUNK_SIZE * 5;
 
-    return averageDistance <= maxAllowedDistance;
+    if (averageDistance > maxAllowedDistance) {
+      return false;
+    }
+
+    // Verify that the sequence is actually repeating, not just sharing a common prefix.
+    // For a true loop, the text between occurrences of the chunk (the period) should be highly repetitive.
+    const periods = new Set<string>();
+    for (let i = 0; i < recentIndices.length - 1; i++) {
+      periods.add(
+        this.streamContentHistory.substring(
+          recentIndices[i],
+          recentIndices[i + 1],
+        ),
+      );
+    }
+
+    // If the periods are mostly unique, it's a list of distinct items with a shared prefix.
+    // A true loop will have a small number of unique periods (usually 1, sometimes 2 or 3).
+    // We use Math.floor(CONTENT_LOOP_THRESHOLD / 2) as a safe threshold.
+    if (periods.size > Math.floor(CONTENT_LOOP_THRESHOLD / 2)) {
+      return false;
+    }
+
+    return true;
   }
 
   /**