diff --git a/admin/app/jobs/run_download_job.ts b/admin/app/jobs/run_download_job.ts
index a0ee4f9..12b3532 100644
--- a/admin/app/jobs/run_download_job.ts
+++ b/admin/app/jobs/run_download_job.ts
@@ -67,7 +67,7 @@ export class RunDownloadJob {
         if (val) {
           await cancelRedis.del(RunDownloadJob.cancelKey(job.id!))
           userCancelled = true
-          abortController.abort()
+          abortController.abort('user-cancel')
         }
       } catch {
         // Redis errors are non-fatal; in-process AbortController covers same-process cancels
@@ -184,7 +184,8 @@ export class RunDownloadJob {
     } catch (error: any) {
       // Only prevent retries for user-initiated cancellations. BullMQ lock mismatches
       // can also abort the stream, and those should be retried with backoff.
-      if (userCancelled) {
+      // Check both the flag (Redis poll) and abort reason (in-process cancel).
+      if (userCancelled || abortController.signal.reason === 'user-cancel') {
         throw new UnrecoverableError(`Download cancelled: ${error.message}`)
       }
       throw error
diff --git a/admin/app/services/download_service.ts b/admin/app/services/download_service.ts
index 40cdca3..ac9d02d 100644
--- a/admin/app/services/download_service.ts
+++ b/admin/app/services/download_service.ts
@@ -125,7 +125,7 @@ export class DownloadService {
     await RunDownloadJob.signalCancel(jobId)
 
     // Also try in-memory abort (works if worker is in same process)
-    RunDownloadJob.abortControllers.get(jobId)?.abort()
+    RunDownloadJob.abortControllers.get(jobId)?.abort('user-cancel')
     RunDownloadJob.abortControllers.delete(jobId)
 
     // Poll for terminal state (up to 4s at 250ms intervals) — cooperates with BullMQ's lifecycle