daily-brief

name	daily-brief
description	Compose the operator's morning brief: stack health + a dated news digest + a kickoff nudge. Triggered by the morning cron. OUTPUT: terse — bullet points, minimal prose, scannable. Under ~25 lines.
version	1.0.0
platforms	["linux"]
metadata	{"hermes":{"tags":["productivity","briefing","cron","monitoring"],"category":"productivity"}}

Daily Brief

When to Use

Triggered by the morning cron job (hermes cron add — see the recipe README). Produce ONE concise message with three sections, in the agent's voice. Keep the whole thing under ~35 lines.

Procedure

1. Stack health (local, via terminal tool — run these and summarize in 3-4 lines):
   • nvidia-smi --query-gpu=temperature.gpu,utilization.gpu,memory.used,memory.total --format=csv,noheader
   • systemctl is-active llama-server; systemctl is-active hermes-gateway
   • df -h / | tail -1
   • grep -iE "error|traceback|exception" ~/.hermes/logs/*.log 2>/dev/null | tail -5 (say "clean" if none) Flag anything red: a service not "active", disk >85%, GPU temp >80C.
2. News digest (local script — every candidate carries a machine-verified date): run ~/.hermes/scripts/news_digest.sh via the terminal tool. It prints dated candidates from configured sources (community threads ranked by real vote counts, fresh GitHub releases). Pick 2-3 genuinely interesting items FROM THAT LIST ONLY. For EACH item:
   • write ONE complete sentence, grounded only in what the candidate line actually says — no invented numbers or claims;
   • KEEP THE ITEM'S [date] in your line, e.g. "(Jul 7)" — never say "just released" or "new" for anything older than yesterday; the date decides, not you;
   • end the line with the item's real https://... source URL. You may web_extract a chosen item's URL to enrich the sentence, but never to add new items. If the script prints "no dated candidates", write "digest: quiet night" and continue — never block the brief, never fabricate sources.
3. Kickoff nudge: one short line — what's worth building today, or a question.

Why the digest is a script and not web search

An agent-driven web search returns undated snippets, and the model will narrate anything as fresh news — a month-old release reads exactly like yesterday's. The architecture rule this skill encodes: the model WRITES, it never researches. Machine-dated candidates come from a deterministic read-only script; the model's job is selection and one good sentence per item.

Pitfalls

• The high GPU utilization you see during this run is YOUR OWN inference generating this brief — never flag it as a runaway process.
• This runs unattended; if a command errors, note it briefly and move on rather than aborting.
• After timedatectl set-timezone, restart the gateway or this cron fires on the old timezone.