changelog

what shipped in Hermes Agent and how to use it, in my words. updated when a loop closes publicly: feature ships, walkthrough lands, configs arrive here.

2026-09-07

• Hermes Wingtips collected: the series, #1 to #65, as-shipped X text, newest first. llms.txt is the same file for agents. cards stay on the original posts. make wingtips rebuilds both from docs/wingtips/*.md so the two urls stay current.

2026-07-09

• fourth recipe shipped: the sovereign GPU agent. the generalised version of my own 24/7 deployment on an RTX 5090: llama-server and the gateway as boot-persistent systemd services (with the context-sizing arithmetic and the --parallel 1 slot gotcha), SOUL/USER/MEMORY identity templates carrying the directive patterns that stop an agent freelancing, a starter skill pack (a morning brief whose news digest only writes from machine-dated candidates, a Pillow card renderer for a browserless box, a markdown memory vault with CPU semantic search), a trap-guarded drain/restore script with a one-service sudoers scope, and the zero-inbound-ports security posture. every script re-tested in its generalised form on the same box before landing here. tested on Ubuntu Server 26.04, an RTX 5090 32GB, Hermes Agent v0.15+.

2026-06-17

• first recipe shipped: 24/7 Hermes Agent on a Mac Mini M4. running Hermes 24/7 under launchd with a watchdog, scheduled jobs that only ping you when it matters (the [SILENT] pattern), a git-synced workspace, and a hard human-in-the-loop approval rule. includes the macOS launchd gotcha where hermes gateway start can silently downgrade supervision (exit 5) and the two-line launchctl fix. tested on Hermes Agent v0.16.0, macOS 15.

2026-06-10

• repo initialised. structure: recipes/ holds one folder per walkthrough with runnable configs and scripts. nothing lands here until it has run on real hardware.