arch-machine

name	arch-machine
description	Operator surface for arch-machine: thin-first sentinel status, consent-gated thin install, threat-focused audit, expand/pull modules or profiles, archy control plane. Prefer archy + slash commands over gum tinfoil TUI. Use when user mentions arch-machine, archy, tinfoil, sentinel install, expand security/ml profile, thin bootstrap, or host audit.

arch-machine — agent-as-TUI (archy-first)

Mission: Drive arch-machine through Grok slash commands. Keep install thin by default; pull expandable tiers only with --yes. Prefer archy (Ratatui control plane) over tinfoil tui (gum legacy).

User surface = slash commands. bin/am-* are agent-internal — never tell the user to run am-init in a shell.

Cyclic: Grok plugin ↔ archy

Grok (this skill/plugin) ── /arch-* ──► arch-machine backends / archy
         ▲                                    │
         └──── archy G/p preload launch ──────┘

Direction	Operator action
Stay in Grok	/arch-status, /arch-audit, /arch-control, /arch-init, /arch-expand
Leave archy for Grok	In archy: job → NEXT [p] / brief Enter / G (preloaded prompt)

Detail: $PLUGIN/docs/CROSS-REF.md · host docs/archy.md section “Grok plugin”.

Architecture (host repo)

When the arch-machine checkout is present:

• archy = Eagle + Satellites + TEA (crates/archy, docs/archy.md)
• Audit = maintenance/security-audit.sh (malware · ports · supply · config)
• Grok Explain = interactive preload via archy (not bare empty Build)
• Skill in repo: .agents/skills/eagle-satellite-elomaxz

Design threads: Eagle+Satellites / offline jobs — see host crates/archy/README.md.

Resolve plugin + map

PLUGIN="${GROK_PLUGIN_ROOT:?GROK_PLUGIN_ROOT not set — open via installed plugin}"
MAP="$PLUGIN/core-map.json"

Immutable rails

1. Fail closed without consent: expand/init/install require /… --yes after clear operator agreement.
2. FSD / unsupervised may auto-run only: status, version, audit, map, path-probe, archy-print-root (core-map.json → modes.fsd.allowlist).
3. Never auto-install ml-dev or security-dev full profiles.
4. Prefer archy + security-audit.sh / CLI over tinfoil tui.
5. Confirm real command output; do not invent machine state.
6. Speak slash commands to the user (/arch-init --yes), not am-*.

Command map

Intent	User surface	Agent runs
Status / PATH	/arch-status	$PLUGIN/bin/am-status
Core map	/arch-status map	am-map
Control plane	/arch-control · --run	am-archy
Thin init	/arch-init · /arch-init --yes	am-init
Audit	/arch-audit [global|path|--dry-run]	am-audit → security-audit.sh
Expand	/arch-expand security then --yes	am-expand …

Supervised vs FSD

Mode	Behavior
Supervised (default)	Always status first; ask before mutate
FSD	Auto status/audit only; any expand/init without --yes → refuse

Core vs expandable

See $PLUGIN/docs/BOUNDARY.md and $PLUGIN/core-map.json. Thin core = sentinel + archy source/runtime; modules/profiles are expandable.

Remote pull + expand (real work)

Cache default: ~/.cache/arch-machine/src from p10ns11y/arch-machine (sentinel). Only with /arch-init --yes or expand with consent.

/arch-expand <module> --yes is not a no-op:

1. Ensure repo (clone/pull if needed; prefers ARCH_MACHINE_ROOT → ~/arch-machine over system install).
2. Run modules/<name>/install.sh --agent-expand when present.
3. Write .arch-expand-state/<name>.stamp.

Full ml-dev / security-dev still require explicit profile name + --yes (never auto).