fix: NoSuchKey, The specified key does not exist.

apps/dashboard/src/lib/features/course/components/lesson/video/hls-encoder.ts

@@ -339,9 +339,9 @@ export async function encodeAndUploadHls(opts: EncodeAndUploadOptions): Promise<
     emit({ stage: 'uploading', percent: 73 });
     await uploader.flush();
 
-    // HLS audio segments are enough for server-side Whisper prep — no second
-    // full pass through the source file for a separate audio.m4a upload.
-    const audioKey = audioTrack ? `${assetId}/audio/playlist.m3u8` : undefined;
+    // Prefer the dedicated HLS audio playlist when present; otherwise fall
+    // back to the highest uploaded video rendition (muxed audio is fine for Whisper).
+    const audioKey = audioTrack ? resolveTranscriptionAudioKey(assetId, rungs, uploader) : undefined;
 
     // --- Thumbnails ----------------------------------------------------
     emit({ stage: 'thumbnail', percent: 86 });
@@ -744,6 +744,7 @@ class HlsUploader {
   private uploadedCount = 0;
   private inflight: Promise<void>[] = [];
   private uploadFailures: unknown[] = [];
+  private readonly uploadedPaths = new Set<string>();
   private readonly uploadGate = new ConcurrencyGate(HLS_UPLOAD_CONCURRENCY);
 
   constructor(
@@ -758,6 +759,10 @@ class HlsUploader {
     }
   ) {}
 
+  hasUploaded(relativePath: string): boolean {
+    return this.uploadedPaths.has(relativePath);
+  }
+
   targetFor(path: string, mimeType: string): BufferTarget {
     return new BufferTarget({
       onFinalize: (buffer) => {
@@ -801,6 +806,7 @@ class HlsUploader {
       });
       this.options.presignedUrls.set(path, presigned);
 
+      this.uploadedPaths.add(path);
       this.totalBytes += u8.byteLength;
       this.uploadedCount += 1;
       this.options.onUploaded({ totalBytes: this.totalBytes, count: this.uploadedCount });
@@ -960,6 +966,28 @@ function pad(n: number, width: number): string {
   return n.toString().padStart(width, '0');
 }
 
+function resolveTranscriptionAudioKey(
+  assetId: string,
+  rungs: HlsRungConfig[],
+  uploader: HlsUploader
+): string | undefined {
+  if (uploader.hasUploaded('audio/playlist.m3u8')) {
+    return `${assetId}/audio/playlist.m3u8`;
+  }
+
+  for (let index = rungs.length - 1; index >= 0; index -= 1) {
+    const rung = rungs[index];
+    if (!rung) continue;
+
+    const playlistPath = `${rung.name}/playlist.m3u8`;
+    if (uploader.hasUploaded(playlistPath)) {
+      return `${assetId}/${playlistPath}`;
+    }
+  }
+
+  return undefined;
+}
+
 /** Rough byte budget for progress UI — actual HLS output varies with content complexity. */
 function estimateHlsUploadBytes(rungs: HlsRungConfig[], durationSeconds: number, hasAudio: boolean): number {
   const segmentCount = Math.max(1, Math.ceil(durationSeconds / HLS_TARGET_SEGMENT_SECONDS)) + 5;

apps/jobs/src/processors/media/extract-audio.ts

@@ -2,12 +2,13 @@ import { mkdir, readFile, rename, rm, stat, writeFile } from 'node:fs/promises';
 import path from 'node:path';
 import { tmpdir } from 'node:os';
 
-import { getJobStep, updateMediaJob, upsertJobStep } from '@cio/db/queries';
+import { getAssetById, getJobStep, updateMediaJob, upsertJobStep } from '@cio/db/queries';
 import { ZExtractAudioPayload, type TExtractAudioPayload } from '@cio/jobs/payloads/media';
 
 import {
   downloadObjectToTempFile,
   mediaBucket,
+  objectExists,
   safeUnlink,
   uploadFileToBucket,
   videosBucket
@@ -83,6 +84,60 @@ async function downloadHlsPlaylistToTempFile(
   return { inputPath, tempDir };
 }
 
+type HlsAssetMetadata = {
+  hlsRenditions?: string[];
+};
+
+function readHlsRenditions(metadata: unknown): string[] {
+  if (typeof metadata !== 'object' || metadata === null || Array.isArray(metadata)) {
+    return [];
+  }
+
+  return (metadata as HlsAssetMetadata).hlsRenditions ?? [];
+}
+
+/**
+ * Pick the first videos-bucket object that actually exists. HLS assets may
+ * store a dedicated audio playlist, a legacy audio.m4a, or only muxed audio
+ * inside a video rendition playlist.
+ */
+async function resolveTranscriptionSourceKey(assetId: string, payloadStorageKey: string): Promise<string> {
+  const asset = await getAssetById(assetId);
+  if (!asset) {
+    throw new Error(`Asset ${assetId} not found`);
+  }
+
+  const candidates: string[] = [];
+  const push = (key: string | null | undefined) => {
+    if (key && !candidates.includes(key)) {
+      candidates.push(key);
+    }
+  };
+
+  push(asset.hlsAudioKey);
+  push(payloadStorageKey);
+  push(asset.storageKey);
+
+  if (asset.hlsManifestKey) {
+    const prefix = `${assetId}/`;
+    push(`${prefix}audio/playlist.m3u8`);
+    push(`${prefix}audio.m4a`);
+    for (const rung of [...readHlsRenditions(asset.metadata)].reverse()) {
+      push(`${prefix}${rung}/playlist.m3u8`);
+    }
+  }
+
+  const bucket = videosBucket();
+  for (const key of candidates) {
+    if (await objectExists(bucket, key)) {
+      log.info('extract-audio-source-resolved', { assetId, key });
+      return key;
+    }
+  }
+
+  throw new Error(`No transcription source object found for asset ${assetId} (tried ${candidates.join(', ')})`);
+}
+
 async function resolveAudioSourcePath(
   bucket: string,
   storageKey: string
@@ -125,7 +180,8 @@ export async function processExtractAudio(rawData: unknown): Promise<ExtractAudi
   let inputTempDir: string | undefined;
   let outputPath: string | undefined;
   try {
-    const source = await resolveAudioSourcePath(videosBucket(), storageKey);
+    const resolvedStorageKey = await resolveTranscriptionSourceKey(assetId, storageKey);
+    const source = await resolveAudioSourcePath(videosBucket(), resolvedStorageKey);
     inputPath = source.inputPath;
     inputTempDir = source.tempDir;
 

apps/jobs/src/utils/storage.ts

@@ -1,4 +1,4 @@
-import { GetObjectCommand, PutObjectCommand } from '@aws-sdk/client-s3';
+import { GetObjectCommand, HeadObjectCommand, NotFound, PutObjectCommand } from '@aws-sdk/client-s3';
 import { Readable } from 'node:stream';
 import { createWriteStream } from 'node:fs';
 import { mkdir, readFile, unlink } from 'node:fs/promises';
@@ -13,6 +13,24 @@ import { getS3Client, getStorageConfig } from '../config/storage';
  * Workers prefer disk over buffers so ffmpeg can seek without holding the
  * whole video in RAM.
  */
+export async function objectExists(bucket: string, key: string): Promise<boolean> {
+  try {
+    await getS3Client().send(new HeadObjectCommand({ Bucket: bucket, Key: key }));
+    return true;
+  } catch (error) {
+    if (error instanceof NotFound || (error as { name?: string }).name === 'NotFound') {
+      return false;
+    }
+
+    const status = (error as { $metadata?: { httpStatusCode?: number } }).$metadata?.httpStatusCode;
+    if (status === 404) {
+      return false;
+    }
+
+    throw error;
+  }
+}
+
 export async function downloadObjectToTempFile(bucket: string, key: string, fileNameHint?: string): Promise<string> {
   const dir = path.join(tmpdir(), 'cio-jobs');
   await mkdir(dir, { recursive: true });

packages/core/src/services/assets/assets.ts

@@ -1209,10 +1209,9 @@ export async function finalizeHls1080Service(orgId: string, assetId: string, dat
 }
 
 /**
- * HLS assets ship audio as a separate `audio.m4a` already extracted by
- * Mediabunny in the browser. The existing transcription pipeline accepts
- * any media key in the videos bucket — its extract-audio leg is a no-op
- * on input that's already audio.
+ * HLS assets ship audio as HLS segments (dedicated `audio/` playlist or muxed
+ * into a video rendition). The transcription pipeline demuxes those segments
+ * server-side before calling Whisper.
  */
 async function enqueueAudioTranscriptionForAsset(input: {
   organizationId: string;