gutomec/nirvana-os-engine/nirvana-os-hermes

Nirvana-OS (Hermes bridge)

This is the Hermes-only bridge. It lives inside _shared/, which the installer links into every runtime's skills directory, so runtimes that discover SKILL.md recursively also see this file: its name is namespaced (nirvana-os-hermes) precisely so it can never collide with the first-class nirvana skill, which is the one every non-Hermes runtime must load.

The user runs Nirvana-OS, a Bun-native multi-agent orchestrator with three pillars: businesses (empresas — autonomous multi-agent organizations with org charts of employees), squads (portable agent teams with workflows), and mind-clones (persona DNA injected into employees). The nrv CLI is the single entry point. It reads the global registry at ~/businesses/ and ~/squads/.

Always answer in the user's language (default PT-BR). Run the commands below with your shell tool, then summarize the output for the user. Never invent business or squad names — only report what nrv actually prints.

Discovery — run the command, then report

• "Quais são minhas empresas?" / "What businesses do I have?" / "liste empresas"

→ nrv list-businesses (add --format=json only if you need to parse)

• "Quais são meus squads?" / "What squads do I have?" / "liste squads"

→ nrv list-squads (supports --format=table|json)

• "Quais mind-clones eu tenho?" / "list mind-clones" / "minhas personas/DNA"

→ nrv list-clones (aliases: list-mind-clones; --format=table|json) → nrv inspect-clone <slug> (details; add --dna for DNA layer counts) → nrv ask <clone-slug> "<pergunta>" (consult one mind-clone with its DNA injected)

• "O que o Nirvana pode fazer sobre X?" / capability search across all three pillars

→ nrv search "<topic>" (filter with --kind=business|squad|mind-clone; nrv find "<need>" for routing)

• Web cockpit / visão geral

→ nrv glance

• Anything else / full command surface

→ nrv --help (30+ subcommands: dispatch, ask, inspect, audit-view, export, …)

Orchestration — dispatch a production brief

When the user wants to actually produce something (a report, post, book, design, code, brand, analysis), the orchestration intelligence lives in the Nirvana harness. Do not produce the artifact yourself — hand the brief to nrv:

→ nrv dispatch --auto --exec "<the user's brief verbatim>" (full maestro: the router picks business/squad/mind-clone, runs the agent headless, runs the quality gate; without --exec it only scaffolds) → nrv ask "<question>" (quick consult)

The harness emits an audit chain under ~/.harness-logs/<date>/audit.jsonl for every dispatch, so the work is verifiable.

Honest limits in Hermes

Full in-process multi-agent dispatch is richest inside Claude Code (native subagent primitive). In Hermes, the maestro reasoning still works and nrv carries the deterministic pieces (routing, list, inspect, verify, quality-gate); employee dispatch degrades to sub-process, the same way it does in the Codex adapter. For read-only queries (list / inspect / search) there is no degradation — they work fully here.

Notes

• Requires the nrv CLI on PATH (command -v nrv). If missing, tell the user to install Nirvana-OS, do not improvise.
• This skill is a thin bridge: it routes to nrv, it does not hold the registry itself.