Transcription workflow fails on Docker image with Node 24

[bradykirk]

Description

Self-hosted Docker deployment fails to run transcription workflows. The workflow queue crashes immediately with an ArrayBuffer error before reaching Deepgram or any actual transcription logic.

Error from logs:
[local world] Queue operation failed: TypeError: fetch failed
at async (.next/server/chunks/2242.js:30:10424) {
[cause]: TypeError: Cannot perform ArrayBuffer.prototype.slice on a detached ArrayBuffer
at ArrayBuffer.slice ()
}

Expected behavior: Transcription should process and generate captions for videos.

Root cause analysis: The workflow package (4.0.1-beta.42) appears incompatible with Node 24. The current Dockerfile uses node:24-alpine, and the ArrayBuffer detachment error is consistent with V8
changes in newer Node versions affecting buffer handling in the workflow package's local execution mode.

Reproduction

1. Deploy ghcr.io/capsoftware/cap-web:latest via Docker (using Coolify)
2. Set DEEPGRAM_API_KEY environment variable
3. Upload a video with audio
4. View the video share page
5. Observe repeated [local world] Queue operation failed errors in logs

Additional Context

• Cap version: ghcr.io/capsoftware/cap-web:latest
• Operating system, version: Docker on Coolify (Linux host)
• Device (optional): Self-hosted server

Container Node version: v24.12.0

Potential fix??? Downgrade to Node 20 or 22 LTS in the Dockerfile, or update the workflow package to a version compatible with Node 24.

[linear]

CAP-620

[tembo]

This appears to be a duplicate of #1539, which reports the same "detached ArrayBuffer" error with the identical stack trace in Docker self-hosted deployments.

The Node 24 upgrade was introduced in PR #1313 (merged October 2025), which changed the Dockerfile from node:22-alpine to node:24-alpine. The root package.json specifies node >=20, confirming Node 20 LTS remains the supported baseline.

Relevant files for a fix:

• apps/web/Dockerfile:3 - where the base image is specified
• apps/web/package.json:142 - contains the workflow package at version 4.0.1-beta.42

The fix would be reverting line 3 in apps/web/Dockerfile from:

FROM node:24-alpine AS base

to:

FROM node:22-alpine AS base

or using node:20-alpine to match the project's minimum engine requirement.

[akcware]

I can confirm this issue persists even after downgrading to Node 20.

Environment:

• Custom Docker image built with node:20-alpine (Node v20.20.0)
• Digital Ocean Spaces as S3 backend
• Ubuntu VPS with 4GB RAM

What I tried:

1. Built custom image with Node 20 instead of Node 24
2. Verified Node version inside container: v20.20.0
3. Confirmed S3 files are publicly accessible (HTTP 200)
4. Confirmed Deepgram API is reachable from container

Result:
Same ArrayBuffer.prototype.slice on a detached ArrayBuffer error persists.

[local world] Queue operation failed: TypeError: fetch failed
[cause]: TypeError: Cannot perform ArrayBuffer.prototype.slice on a detached ArrayBuffer

This suggests the issue is not solely related to Node 24's V8 changes, but likely a deeper incompatibility in the workflow package's local execution mode with the Docker/self-hosted environment.

Video upload and playback work fine - only transcription is broken.

[VirtualRoy]

I can confirm this issue exists.

[Deducer]

I'm experiencing the exact same problem on my self-hosted Cap instance:

Error:

[local world] Queue operation failed: TypeError: fetch failed
[cause]: TypeError: Cannot perform ArrayBuffer.prototype.slice on a detached ArrayBuffer

Setup:

Deployment: Docker via Coolify (same as OP)
Images: ghcr.io/capsoftware/cap-web:latest + ghcr.io/capsoftware/cap-media-server:latest
Storage: Cloudflare R2 (S3-compatible)
Database: MySQL 8.0
Queue: Redis 7.2 with maxmemory-policy: noeviction
API: Deepgram configured and valid

Observations:

✅ Videos upload successfully
✅ Media server processes videos with FFmpeg successfully
❌ Transcription workflow fails immediately with this error
Error repeats continuously in logs

[jadasi]

Same problem with a fresh deploy to Railway: transcription and summary features are broken.