hermes-about

name	hermes-about
metadata	{"author":"rhoai3-coding-demo","version":"1.1.0","platform-family":"hermes","platform-baseline":"repo","ocp-baseline":"repo","skill-group":"Hermes Agent"}
description	Use when explaining Hermes Agent, choosing which Hermes mechanism fits a stage 080 need (skills vs memory, kanban vs delegate_task vs /goal, hooks vs plugins vs cron), checking platform/install support tiers, or verifying a claimed capability exists. Do NOT use for depth on any one mechanism — the matching hermes-* sibling skill owns it.

Hermes Agent Overview

Use this skill for orientation: what Hermes is, what it ships, which mechanism to reach for, and where it runs. Depth always lives in the sibling skills.

Source Grounding

Read references/source-capture.md for provenance and references/official-doc-extraction.md for the full validated extraction — the complete official feature inventory, the platform matrix, and all ten documented mechanism-selection comparisons quoted verbatim. Official Hermes Agent documentation (Nous Research) is the product authority.

What Hermes is (first-party)

"The self-improving AI agent built by Nous Research. A terminal-native autonomous coding and task agent with persistent memory, agent-created skills, and a messaging gateway that lives on 21+ messaging platforms… Runs on local, Docker, SSH, Daytona, Modal, or Singularity backends. Works with Nous Portal, OpenRouter, OpenAI, Anthropic, Google, or any OpenAI-compatible endpoint." (docs llms.txt — the only first-party positioning prose; the user-stories page is curated community quotes and must be cited as community voice, not Nous claims.)

Four front ends share one agent state: CLI, TUI (recommended interactive surface), Desktop App, Web Dashboard — "start a session in one and resume it in another".

Mechanism selection — consult before designing

The extraction carries all ten documented comparisons verbatim. The load-bearing one-liners:

Need	Reach for	Official rule
Short reasoning subtask, result back into parent context	delegate_task	"a function call"
Work crossing agent boundaries, surviving restarts, human input	Kanban	"a work queue where every handoff is a row any profile (or human) can see and edit"
Mechanical multi-step pipeline, no judgment	execute_code	"Use execute_code when you need mechanical data processing"
Keep iterating on one task in THIS chat	/goal	"single-session… never creates a kanban card"
One board card that iterates until acceptance	kanban --goal	"borrows the engine, not the board"
Key facts always in context	Memory	"~1,300 tokens total", always in prompt
"Did we discuss X last week?"	session_search	"Unlimited (all sessions)… no LLM calls"
Long procedures loaded only when relevant	Skills	skills/memory self-improvement split
Identity/tone that follows you everywhere	SOUL.md	"if it belongs to a project, it belongs in AGENTS.md"
Time-triggered automation	Cron	dependency-triggered work is Kanban's job
Intercept/observe lifecycle events	Hooks (4 systems)	see hermes-hooks for selection within
Add tools/commands/integrations	Plugins / MCP / config-driven	"Not everything is a Python plugin" — pick the surface

Platform and install facts that gate deployment

• Tier 1 (first priority, "we strive to never break"): macOS Apple Silicon, Windows 10/11, Linux/WSL2, Docker. Tier 2 (best-effort, "releases may break them"): Termux, Nix. Unsupported (PRs rejected): AUR, Intel Mac, pypi, brew.
• Docker installs don't support hermes update — upgrade by new image.
• OpenShift/Kubernetes is NOT an officially documented install surface. A third-party deployment pattern (Red Hat Developers article + community repo: UBI9 image, KServe/vLLM InferenceService, PVC-backed HERMES_HOME, restricted-SCC securityContext) is captured in references/openshift-deployment.md — WITH a verified-conflicts table; read that table before reusing any of it.
• Minimum model context: 64K tokens — smaller windows "will be rejected at startup" (hard constraint on any private/MaaS model wired in).
• Versioning is a PROCEDURE, not a number: hermes version vs the GitHub releases page; installs track origin/main. Never hardcode a version string in stage 080 docs (point-in-time anchor at capture: v0.20.0, 2026-08-03 — see hermes-cli).

Governance-relevant defaults (orientation level)

Blank Slate setup mode = provider/model + File Operations + Terminal only — the documented minimal-trust baseline. Plugins are opt-in (plugins.enabled; project-local plugins additionally gated by HERMES_ENABLE_PROJECT_PLUGINS). Nested delegation is opt-in (max_spawn_depth default 1). Kanban is deliberately single-host. Cron jobs can't create cron jobs, and the drift guard skips unattended runs rather than silently inheriting a model switch. Context files, SOUL.md, and memory writes are all injection-scanned before inclusion.

Workflow

1. Before recommending a mechanism, check the comparison tables in the extraction — never reason from first principles when a documented "use X when" exists.
2. Before claiming platform support, check the tier table and quote the tier language (Tier 2 carries no stability promise).
3. When describing capabilities, cite the Features Overview inventory in the extraction; features without an owning hermes-* skill are flagged there — say "not yet covered by a sibling skill", don't omit.
4. Route depth questions to the owning sibling immediately.
5. Cite the official section in the PR (stage 080 official-first rule).

Validation

Static checks (this is an orientation skill — no live seat needed):

• Every mechanism recommendation cites its comparison's official URL.
• No fixed Hermes version number asserted anywhere in stage 080 docs.
• Community-sourced positioning is attributed as community voice.

Pitfalls

• Citing the user-stories page as Nous positioning — it's a curated third-party quote wall; only llms.txt is first-party prose.
• Quoting the "21+ platforms" count as exact — the messaging gateway's own table lists more rows; reconcile before citing a number.
• Assuming Tier 2 (Nix!) stability for GitOps-style deployment plans — the docs' own words: "Breaks often… Best of luck~!"
• Wiring a <64K-context model — rejected at startup, not degraded.
• Treating Curator as a standalone subsystem — it's the skills-lifecycle maintenance daemon (owned by hermes-skills).
• Copying the third-party OpenShift manifests verbatim — they commit GATEWAY_ALLOW_ALL_USERS: "true" (officially: never in production), serve a 32K-context model (below the 64K startup floor), and wire the model via undocumented env vars instead of the config.yaml model: block. See references/openshift-deployment.md conflicts table.

Related Skills

hermes-configuration, hermes-managed-scope, hermes-kanban, hermes-skills, hermes-hooks, hermes-sessions, hermes-cli — the seven depth skills this overview routes into.

References

• references/source-capture.md
• references/official-doc-extraction.md
• references/openshift-deployment.md (SUPPLEMENTARY, third-party — OpenShift AI + vLLM deployment pattern with verified-conflicts table)