Skip to content

quickstart.md

Latest commit

 

History

History
87 lines (68 loc) · 3.17 KB

quickstart.md

File metadata and controls

87 lines (68 loc) · 3.17 KB
Raw

Quickstart: iN2N Production-Mode Enforcement & Durable Runtime

How an operator turns a risk into an honestly-enforced, durable production risk — and how to verify each guarantee. Assumes a working 056 risk (Border + members) and the installed defenseclaw + openshell CLIs.

1. Make the runtime durable (any mode)

# Generate + enable the mesh daemon unit and one unit per always-on member
python3 scripts/in2n-services.py generate
python3 scripts/in2n-services.py enable
systemctl --user status 'netclaw-*.service'   # mesh + per-member, all active, Restart=always

Verify resilience (SC-005):

systemctl --user kill netclaw-mesh.service        # simulate a crash
sleep 10 && systemctl --user is-active netclaw-mesh.service   # -> active (auto-restarted)
# log out / reboot -> all netclaw-*.service come back on their own

2. Turn on production enforcement

# in ~/.openclaw/mesh.systemd.env (and per-member env)
N2N_RISK_MODE=production
# optional: block on ANY missing control, including audit
# N2N_STRICT_ALL=1
systemctl --user restart netclaw-mesh.service

3. Check the posture (US1)

# via the n2n-mcp status tool (or the HUD posture panel / operator heartbeat)
openclaw agent -m "what is my risk posture?"
# -> "production — enforced" with sandbox, model-guard, audit all active

Force a degraded state and confirm honesty (SC-001):

# stop DefenseClaw's guard path, then:
openclaw agent -m "risk posture?"
# -> "production — DEGRADED (model-guard missing)"  (never "enforced")

4. Verify each control

Sandbox (US2, SC-002) — a member runs under OpenShell in production:

openclaw agent -m "how many snapshots in IP Fabric?"   # routes to the ipfabric member
# the member process is a child of `openshell`; with openshell unavailable the
# delegation is REFUSED (fail-closed), posture -> degraded(sandbox)

Model-guard + scan (US3, SC-003) — with DefenseClaw down, a delegation fails closed:

# stop defenseclaw guard, then delegate -> result.enforcement == "refused:model-guard"
# (never silently runs through a direct provider)

GAIT trail (US4, SC-004) — every federation event is a git commit + a SQLite row:

git -C ~/.openclaw/n2n/gait log --oneline | head
# one commit per delegation/enrollment/removal/quarantine; SHAs match remote_invocation_record.gait_ref

5. Degraded policy (SC-008)

Missing control Default behavior result.enforcement
sandbox OR model-guard (containment) delegation refused refused:<control>
GAIT (audit only) delegation runs, loudly flagged audit-degraded
any, with N2N_STRICT_ALL=1 delegation refused refused:<control>

6. Fault isolation (US6, SC-006)

# daemon down  -> heartbeat: "federation-layer/daemon fault"     (not a member flap)
# member down  -> heartbeat: "<member> down (will cold-start)"
# backend down -> a task reports "backend unreachable", NOT a federation fault

7. Regression guard (SC-007)

cd /home/johncapobianco/netclaw && pytest tests/n2n/   # 44 eN2N regression + 45 iN2N + new 057 tests all green