PR #28895: fix(gateway): omit systemd 254+ Restart keys on older systemd

• Repository: NousResearch/hermes-agent
• Author: jasonnickel
• State: open | merged: False
• Link: https://github.com/NousResearch/hermes-agent/pull/28895

Description (problem / solution / changelog)

Summary

generate_systemd_unit() in hermes_cli/gateway.py hardcodes RestartMaxDelaySec=300 and RestartSteps=5. These keys were added in systemd 254. On systemd 252 (the version shipped with Debian 12 — the recommended host for hermes-gateway on production homelabs) systemd logs both as unknown and ignores them, polluting the journal with two Unknown key WARNINGs on every service start.

This PR detects the systemd version at unit-generation time via systemctl --version and emits the two keys only when running on systemd 254+. On older systemd the unit falls back to the existing Restart=always + RestartSec=5 behavior, which is equivalent for the common case where restarts succeed within the first attempt.

Why this matters

• Debian 12 is the recommended host distro for self-hosted hermes-gateway (used by NousResearch's own homelab guides + the existing Ansible playbooks).
• Two WARNINGs per service start mean every gateway restart pings whatever syslog/journalctl monitoring users have configured.
• Users on older distros currently have to maintain a local patch to silence these warnings (I've been carrying one for ~2 weeks on my homelab).

Test plan

• Generate unit on Debian 12 host (systemd 252): output contains Restart=always + RestartSec=5 only, no RestartMaxDelaySec / RestartSteps. No Unknown key warnings in journalctl -u hermes-gateway.
• Generate unit on Ubuntu 24.04 host (systemd 255): output identical to current behavior; both keys present.
• Service restarts cleanly on failure in both cases.
• No regression on macOS / non-systemd hosts (the helper falls through to return False on systemctl --version failure).

Backwards compatibility

The fallback path on systemd <254 is Restart=always + RestartSec=5 — same as systemd 254+ for any restart that succeeds on the first attempt. The only behavioral difference is exponential backoff between repeated failures, which only kicks in if the service is in a crash loop. For users in a crash loop, the journal warnings were never the priority issue anyway.

Closes #28880.

Changed files

• hermes_cli/gateway.py (modified, +33/-6)