← all posts site ↗

a dead mount froze every tool

the agent kept telling me the same frustrating thing, tools are blocked again, the build is done, and two handoff steps are stuck behind a pre-tool-call callback that keeps timing out. i assumed it was hermes misbehaving and readied the usual restart. it was not hermes. it was one dead network mount.

the mount point points at a windows share that lives on a machine currently booted into the other os and offline for most of a day. the mount stayed mounted and dead. the nasty part is how the kernel handles it. a single stat call under that dead mount blocked for over twenty seconds, and the cleanup plugin stats every path token in a terminal command during the pre-tool-call hook. two tokens meant forty seconds, over the thirty second hook limit, so the hook failed closed and the tool got timed out. once the plugin is suppressed for a minute, every later tool call fails the same way, even the skill viewer which touches no files at all.

the fix had three parts. force-unmount the dead share so a stat returns instantly, make the empty mount point read-only so a stray write meant for the share can not land on local disk, and patch the cleanup plugin to skip any path that lives under a network mount, cifs, nfs, sshfs or 9p. the guard now runs in effectively zero time and the old path still takes five seconds. the share also gets unmounted before the machine reboots into the other os, so the dead state can not come back. the takeaway is that a tools hang is worth tracing to the filesystem before blaming the agent, because a roaming mkdir into a stale mount is exactly the silent failure that looks for all the world like a software bug.