Retry library pulls after Comfy finishes and fall back to disk.

Long jobs were succeeding on Beast then failing the save step when /view briefly dropped; retries plus host-agent file serving keep finished MP4s from being orphaned.

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

server/utils/watch.ts

@@ -19,7 +19,7 @@ function classifyError(message: string) {
   if (lower.includes('out of memory') || (lower.includes('cuda') && lower.includes('alloc')) || lower.includes('vram')) {
     return 'ComfyUI VRAM allocation failed. Try Turbo or a smaller aspect ratio.'
   }
-  if (lower.includes('econnrefused') || lower.includes('unreachable') || lower.includes('fetch failed')) {
+  if (lower.includes('econnrefused') || lower.includes('unreachable') || lower.includes('fetch failed') || lower.includes('connection dropped')) {
     return 'ComfyUI host connection dropped. Confirm the desktop instance is running and reachable on the LAN.'
   }
   if (lower.includes('timeout')) {
@@ -28,6 +28,17 @@ function classifyError(message: string) {
   return message
 }
 
+function errorText(error: unknown) {
+  if (!error) return 'Unknown error'
+  if (typeof error === 'object') {
+    const rec = error as { statusMessage?: string; message?: string; data?: { cause?: string } }
+    const bits = [rec.statusMessage, rec.message, rec.data?.cause].filter(Boolean)
+    if (bits.length) return String(bits[0])
+  }
+  if (error instanceof Error) return error.message
+  return String(error)
+}
+
 function sleep(ms: number) {
   return new Promise(resolve => setTimeout(resolve, ms))
 }
@@ -175,8 +186,23 @@ export function watchComfyJob(job: Job, options: { persist?: boolean } = {}): Pr
       finishing = true
       job.saving = true
       try {
-        const history = await fetchHistory(job.promptId)
-        const video = extractVideo(history, job.promptId)
+        let history: Awaited<ReturnType<typeof fetchHistory>> | null = null
+        let video: ReturnType<typeof extractVideo> = null
+        for (let attempt = 1; attempt <= 6; attempt++) {
+          try {
+            history = await fetchHistory(job.promptId)
+            video = extractVideo(history, job.promptId)
+            if (video) break
+          } catch (error) {
+            if (attempt >= 6) throw error
+          }
+          emitLocal({
+            type: 'status',
+            message: `Waiting for ComfyUI history (${attempt}/6)...`,
+            progress: 96
+          })
+          await sleep(1000 * attempt)
+        }
         if (!video) {
           finishing = false
           job.saving = false
@@ -312,12 +338,20 @@ export function watchComfyJob(job: Job, options: { persist?: boolean } = {}): Pr
         } catch (saveError) {
           job.saving = false
           removeExtendTemp(job.library?.extendTmpDir)
-          const message = saveError instanceof Error ? saveError.message : String(saveError)
+          const message = classifyError(errorText(saveError))
           job.status = 'error'
           job.error = persist
             ? `Video generated but library save failed: ${message}`
             : `Segment finished but could not be prepared: ${message}`
           emitChainJob(job, { type: 'error', error: job.error, message: job.error })
+          // Keep pending so a later Recover / container restart can still pull the MP4.
+          if (job.library && job.promptId) {
+            writePendingJob(pendingFromJob(job, {
+              promptId: job.promptId,
+              currentClipId: job.clipId || job.library.parentClipId
+            }))
+          }
+          await syncQueueFromJob(job).catch(() => null)
         }
         resolve()
         return true
@@ -406,9 +440,19 @@ export function watchComfyJob(job: Job, options: { persist?: boolean } = {}): Pr
       if (payload.type === 'executing') {
         const node = data.node === null || data.node === undefined ? null : String(data.node)
         if (node === null) {
-          const saved = await succeed()
-          if (!saved) {
-            emitLocal({ type: 'status', message: 'Waiting for ComfyUI to save the MP4...', progress: 96 })
+          try {
+            const saved = await succeed()
+            if (!saved) {
+              emitLocal({ type: 'status', message: 'Waiting for ComfyUI to save the MP4...', progress: 96 })
+            }
+          } catch (error) {
+            finishing = false
+            job.saving = false
+            emitLocal({
+              type: 'status',
+              message: `Save hiccup (${errorText(error)}). Retrying...`,
+              progress: 96
+            })
           }
           return
         }