| name | nemoclaw-autoheal |
| description | Guide users through Hermes availability checks and the optional host-side auto-heal setup without crossing the sandbox-to-host boundary. |
nemoclaw-autoheal
Use this skill when a user says the Slack bot is not responding, asks whether
Hermes is healthy, reports inference 503/504 failures, asks about Outlook
connectivity, or wants to install the optional auto-heal services.
Boundary
You run inside the OpenShell sandbox. Do not run systemctl, recreate the
sandbox, inspect credentials, or try to install host services. Those actions
belong to the host operator. You may call nemoclaw_status to inspect the
local Hermes gateway, then give the operator the exact commands below.
Diagnose
- Call
nemoclaw_status first. If the gateway is stopped, say that the
sandbox gateway is unhealthy.
- For an inference error, explain that
503, 504, or a timeout can be an
upstream service or host-proxy failure.
- For a Slack failure, distinguish an unavailable gateway from an allowlist,
Socket Mode, or Slack API issue. Do not claim the bot received a message
unless the user supplied evidence.
- For Outlook, explain that a Graph connection closure may be sandbox egress
or a temporary proxy failure; it is not proof that the mailbox settings are
wrong.
Tell the host operator
Run these commands from the cloned example directory, not inside the sandbox:
bash scripts/autoheal/sanity-check.sh
bash scripts/autoheal/sanity-check.sh --repair
If auto-heal is not installed yet, first run:
bash scripts/autoheal/install.sh --check
bash scripts/autoheal/install.sh
For status and logs:
systemctl --user status nemoclaw-hermes-gateway-forward.service
systemctl --user list-timers 'nemoclaw-*'
journalctl --user -u nemoclaw-hermes-watchdog.service -n 80 --no-pager
If the user runs an inference endpoint through
http://host.openshell.internal:18080, remind them to set
NEMOCLAW_HOST_TLS_PROXY_UPSTREAM in the host .env before installing the
auto-heal services. The value is the HTTPS origin, for example
https://inference-api.nvidia.com.
Response style
Start with the current status or likely failure class, then give only the next
one or two host commands. Do not print or request Slack, Outlook, GitHub, or
inference credentials.
