Summary

Hermes publishes llms.txt and llms-full.txt for LLM consumption, but neither is wired into the agent itself. The result: the agent hallucinates Hermes commands and then tries to recover by browsing the docs site — which fails in any sandbox backend with restricted egress.

Problem

Concrete first-run experience: a new user asks the agent to inspect the sandbox config. The agent confidently emits hermes config get sandbox, which does not exist (real commands are hermes config, hermes config edit, hermes config set, hermes config check — no get). When corrected, the agent reaches for the browser tool to consult hermes-agent.nousresearch.com. On Docker / Modal / Daytona / Vercel backends, DNS to the docs host is often blocked, so the recovery path also fails.

Net effect: confidently wrong answers about the very tool the user is asking about, and no graceful fallback. This is a strange failure mode for a self-improving agent — questions about itself should be the one thing it always gets right.

Idea

Ship a bundled skill hermes-docs in skills/ containing a snapshot of llms-full.txt, refreshed on each release via the existing bundled-skill sync (origin-hash logic already handles user edits cleanly).

skills/hermes-docs/
├── SKILL.md            # trigger description + when-to-use
└── references/
    └── docs.md         # llms-full.txt snapshot, ~1.8 MB

Progressive disclosure works as it does for every other bundled skill — Level 0 in the index, Level 1 SKILL.md when triggered, Level 2 docs.md pulled in only when needed. No new system, just a new skill.

Trigger surface: questions about Hermes CLI, config schema, sandbox backends, MCP setup, skills, gateway, terminal backends, or any "what does hermes <x> do" pattern.