Stop killing live videos mid-save: never fake-complete or age-out jobs just because Comfy is briefly idle.

Co-authored-by: Cursor <cursoragent@cursor.com>

server/utils/studioQueue.ts

@@ -297,11 +297,9 @@ async function reapZombieLiveJobs() {
   if (queue.running > 0 || queue.pending > 0) return
   const { fetchHistory } = await import('~/server/utils/comfy')
   for (const job of listJobs()) {
+    // Never interrupt download/stitch/library save — Comfy is idle then by design.
     if (job.saving) continue
-    // Stuck extension loops claimed the GPU forever after a crash/sleep.
+    if (job.library?.chainContinuing) continue
-    if (job.library?.chainContinuing && jobAgeMs(job) >= ZOMBIE_EMPTY_COMFY_MS) {
-      job.library.chainContinuing = false
-    }
     if (job.status !== 'running' && job.status !== 'uploading' && job.status !== 'queued') continue
     if (jobIsLocallySubmitting(job)) continue
     const zombieMs = (job.kind === 'edit' || job.kind === 'music') ? 45_000 : ZOMBIE_EMPTY_COMFY_MS
@@ -321,29 +319,24 @@ async function reapZombieLiveJobs() {
       failZombieLiveJob(job, job.error || `ComfyUI ${status}`)
       continue
     }
-    // History still present + Comfy idle: waiter never settled (common for image jobs).
+    // History success + Comfy idle: the socket waiter usually finishes the save.
-    // Mark complete and advance the studio queue so waiting items can run.
+    // Never fake-complete video here — that skipped library save and left ghosts in the queue.
-    job.status = 'complete'
+    // Image/music: only settle if media was already persisted and the waiter died before settle.
-    job.error = undefined
+    if (job.kind === 'edit' && job.stillId) {
-    if (job.library) job.library.chainContinuing = false
+      job.status = 'complete'
-    await onLiveVideoSettled(job).catch(() => null)
+      job.error = undefined
+      await onLiveVideoSettled(job).catch(() => null)
+    } else if (job.kind === 'music' && (job.clipId || job.audio)) {
+      job.status = 'complete'
+      job.error = undefined
+      await onLiveVideoSettled(job).catch(() => null)
+    }
   }
 }
 
 function liveJobOwnsGpu(job: Job) {
   if (job.library?.stopAfterCurrent) return false
-  if (job.library?.chainContinuing) {
+  if (job.library?.chainContinuing) return true
-    // Never let a stuck chainContinuing flag block the queue after Comfy went idle.
-    if (job.status === 'complete' || job.status === 'error' || job.status === 'cancelled') {
-      job.library.chainContinuing = false
-      return false
-    }
-    if (jobAgeMs(job) >= ZOMBIE_EMPTY_COMFY_MS * 2) {
-      job.library.chainContinuing = false
-      return false
-    }
-    return true
-  }
   if (job.saving) return true
   if (job.status === 'uploading') return true
   if (job.status === 'running') {
@@ -355,19 +348,24 @@ function liveJobOwnsGpu(job: Job) {
   return false
 }
 
-/** Drop soft in-memory GPU claims when Comfy itself is idle so Start now can fire. */
+/**
-function clearSoftGpuClaimsWhenComfyIdle(opts?: { aggressive?: boolean }) {
+ * Force-start only: drop dead in-memory claims when Comfy is idle.
-  const ageCap = opts?.aggressive ? SUBMIT_WINDOW_MS : ZOMBIE_EMPTY_COMFY_MS
+ * Must never use job startedAt alone — real videos run far longer than ZOMBIE_EMPTY_COMFY_MS
+ * and sit with an empty Comfy queue while downloading/stitching.
+ */
+function clearDeadGpuClaimsForForceStart() {
   for (const live of listJobs()) {
-    if (live.library?.chainContinuing && jobAgeMs(live) >= ageCap) {
+    if (live.saving) continue
-      live.library.chainContinuing = false
-    }
-    if (live.saving && jobAgeMs(live) >= ZOMBIE_EMPTY_COMFY_MS) {
-      live.saving = false
-    }
-    if (live.status !== 'running' && live.status !== 'queued' && live.status !== 'uploading') continue
     if (jobIsLocallySubmitting(live)) continue
-    if (jobAgeMs(live) < ageCap) continue
+    if (live.status !== 'running' && live.status !== 'queued' && live.status !== 'uploading') {
+      if (live.library?.chainContinuing) live.library.chainContinuing = false
+      continue
+    }
+    // Active prompt still owned by a living waiter — leave it alone.
+    if (live.promptId && jobAgeMs(live) < QUEUED_GRACE_MS) continue
+    if (live.promptId) continue
+    if (live.library?.chainContinuing) live.library.chainContinuing = false
+    if (jobAgeMs(live) < SUBMIT_WINDOW_MS) continue
     failZombieLiveJob(live, 'Cleared stuck live job so the studio queue can start')
   }
 }
@@ -382,9 +380,7 @@ async function forceStartWaitingIfGpuIdle(owner: string, id: string) {
   const queue = await fetchLiveQueue().catch(() => null)
   const comfyRendering = Boolean(queue && queue.running > 0)
   if (comfyRendering) return false
-  // Soft ownership (chainContinuing / orphan running) used to block Start now forever
+  clearDeadGpuClaimsForForceStart()
-  // while Comfy's queue was already empty.
-  clearSoftGpuClaimsWhenComfyIdle({ aggressive: true })
   if (listJobs().some(liveJobOwnsGpu)) return false
   try {
     await startStudioJob(after)
@@ -987,16 +983,10 @@ async function dispatchStudioQueue() {
     })
   }
   if (await videoJobsBusy()) {
-    const queue = await fetchLiveQueue().catch(() => null)
+    if (owners.some(owner => readJobs(owner).some(job => job.status === 'waiting'))) {
-    if (queue && queue.running === 0 && queue.pending === 0) {
+      scheduleKickRetry()
-      clearSoftGpuClaimsWhenComfyIdle()
-    }
-    if (await videoJobsBusy()) {
-      if (owners.some(owner => readJobs(owner).some(job => job.status === 'waiting'))) {
-        scheduleKickRetry()
-      }
-      return
     }
+    return
   }
   // Re-read after videoJobsBusy — reap/settle may have rewritten disk rows.
   for (const owner of owners) {