fix: stop feeding running ASR output back as Whisper initial prompt (repetition loops)

apps/desktop/src/pipeline/providers/transcription/amical-cloud-provider.ts

@@ -1030,7 +1030,9 @@ export class AmicalCloudProvider implements TranscriptionProvider {
               vadProbs,
               languages: snapshot.currentLanguages,
               vocabulary: snapshot.currentVocabulary,
-              previousTranscription: snapshot.currentAggregatedTranscription,
+              // Not sent: conditioning on the running ASR output can seed a
+              // self-reinforcing repetition loop on the server side too.
+              previousTranscription: undefined,
               formatting: {
                 enabled: enableFormatting,
               },

apps/desktop/src/pipeline/providers/transcription/whisper-provider.ts

@@ -286,9 +286,15 @@ export class WhisperProvider implements TranscriptionProvider {
     aggregatedTranscription?: string,
     accessibilityContext?: TranscribeContext["accessibilityContext"],
   ): string {
+    // NOTE: previousTranscription (the running ASR output) is intentionally not
+    // fed back as the initial prompt. Conditioning Whisper on its own prior
+    // output lets a single hallucinated phrase self-reinforce into a repetition
+    // loop across windows. Vocabulary and cursor context are static, so they
+    // stay. See whisper repetition-hallucination behavior.
+    void aggregatedTranscription;
     const prompt = buildWhisperPrompt({
       vocabulary,
-      previousTranscription: aggregatedTranscription,
+      previousTranscription: undefined,
       beforeText:
         accessibilityContext?.context?.textSelection?.preSelectionText,
     });