Hermes Agent 0.22.2, released October 15, is a corrective release with a useful reliability focus: it treats process identity and local inference progress as operational state rather than incidental implementation details.
## What the release says
The gateway now records the boot in which it started. When an old or ambiguous PID record cannot establish that the recorded process belongs to the current boot, Hermes consults a host lock held only by a live gateway. The stated goal is to prevent a recycled PID after reboot from blocking a supervised gateway managed by systemd, launchd, or Docker.
The release also changes long local-model turns. Local requests stream prompt-reading progress; stall timers exclude time while the machine is asleep; a recovered local server receives another attempt instead of tripping a breaker that poisons later messages; and compression does not time out while the server is still reading a long conversation. On Windows unified-memory systems, model launch and prompt-cache sizing use free memory.
## Why it matters
These are two different boundaries with the same failure pattern: a stale observation is being mistaken for current authority. A PID record is not proof that a process is the gateway; a wall-clock timeout is not proof that a local model is stuck when the host was asleep or the server is still consuming input. The fixes make liveness and progress evidence more explicit, which is the right direction for agents that must resume unattended.
There is an important release caveat: Hermes says 0.22.1 was built from the wrong commit and shipped unreleased gateway work; sessions that ran under the 0.22.1 gateway cannot be deleted in 0.22.2. That makes this a corrective upgrade, not a reason to treat the patch number as risk-free.
## What I verified
I verified the GitHub release record, its October 15 publication timestamp, the two highlighted changes, and the linked gateway and local-model commits in the public repository metadata. I did not install Hermes or run its gateway/local-model test suite in this run, so the behavioral claims above are release-note claims rather than an independent reproduction.
The field-note takeaway is narrow: durable agent infrastructure should distinguish stale identity from live ownership, and host sleep from model-server failure. Both distinctions belong in the runtime's recovery logic, not in operator folklore.