Docker Desktop's crash loop: 50 crashes, one setting
Docker Desktop on Windows refused to start after unclean shutdowns. The leftovers showed it had happened about 50 times, all traced to one unused feature.
The problem
After an unclean shutdown, Docker Desktop would fail to start with "file cannot be accessed by the system" on an internal socket file. Every self-hosted service on the machine (photos, media, file sync) went down with it.
The workaround
The broken file can't be deleted, but renaming its parent folder works every time. I scripted that into the Docker startup, so it runs automatically before Docker launches.
Finding the real cause
Each workaround left a renamed folder behind. When I counted them there were about 50 since August, which meant this was happening on nearly every unclean restart. The error always came from the same component: Docker's built-in AI "Inference manager", a feature I never used.
The fix
- Turned off Docker AI / Model Runner in Docker's settings.
- Removed a duplicate startup entry, so Docker only starts through the script that cleans up first.
After the change: a clean restart, zero new leftover folders, and every service came back on its own.
Lesson
Count your workarounds. One workaround is fine. Fifty means there's a root cause waiting to be found.