SquidWard

Monorepo scaffold for NemoClaw agent projects with shared OpenShell sandbox policy libraries. Prototype locally, deploy to a DGX Spark over SSH, inference entirely local to the Spark.

Built in 8h 14min using BeadHive

Quick start

just                      # list recipes
just each setup           # sync every project's venv
just each check           # lint + format + typecheck everywhere
just each test             # tests everywhere
just a hello-agent run     # run one agent's service locally
just dashboard-setup       # install the SquidWard dashboard dependencies
just dashboard-dev         # run the dashboard locally
just dashboard-demo        # run the real local ingestion pipeline and dashboard
just demo-pipeline         # generate → ingest → detect → recommend → dashboard

just demo-pipeline starts the durable SQLite ingestion API, posts 100 deterministic input events, runs them through deterministic rules and the promoted Isolation Forest, persists rolling risk assessments for every event, stores the resulting finding for OpenClaw investigation, and launches the dashboard at http://127.0.0.1:8300. Data remains available in data/demo-pipeline.db after the demo stops. Set DEMO_PIPELINE_RESET=0 to reuse the existing database instead of resetting it at startup.

Fixture-backed CVE, asset, model, and analyst-feedback pages are enabled by default for presentations. Set VITE_ENABLE_DEMO_PAGES=false before starting the dashboard or app stack to remove those pages from navigation.

To populate the database used by an already-running ingestion API, or remove only that synthetic run while preserving real captures:

just s ingestion run              # terminal 1; SQLite is durable
just demo-seed                     # terminal 2; posts 26 events through the API
just demo-clear                    # removes demo events/findings/actions only
just demo-live                     # continuously add traffic and active findings
just demo-seed http://HOST:8100    # target another running appliance

just demo-live emits mostly normal traffic with a suspicious correlated burst every third cycle. Each burst is scored by the real deterministic processing pipeline and creates a pending deny_destination recommendation. It runs until Ctrl-C; pass an ingestion URL as the recipe argument to target another host.

Demo cleanup identifies the synthetic run-synthetic-001 marker and also removes findings, recommendations, decisions, enforcement results, and rules derived exclusively from those events. It does not truncate the database.

On the GB10, the ingestion API reads GPU utilization from nvidia-smi and unified-memory usage from /proc/meminfo. Configure .env, then run:

just doctor local
just gb10-up
just gb10-app-up

From a workstation, just doctor <ssh-host> performs the same readiness checks over SSH. The dashboard binds to MGMT_BIND_ADDR; ingestion, FastMCP, OpenClaw, and LiteLLM remain on host-local or private container interfaces.

Add an agent

just new my-agent

Copies agents/hello-agent/ to agents/my-agent/, renames the package, and leaves you a self-contained project: its own pyproject.toml, agents.yaml (NemoClaw manifest), policy.yaml (OpenShell policy), and justfile. It gets no lockfile of its own — it resolves and locks into the single root uv.lock (see "The one rule" below).

The one rule

One owner per component directory — agents/<name>/, services/<name>/, or libs/<name>/. Stay inside yours and parallel worktrees merge cleanly.

The shared surfaces are CODEOWNERS-gated and need review: contracts/, the root pyproject.toml, uv.lock, and the root justfile.

uv.lock is a single root lock shared by every Python member, so it will conflict. It is machine-generated — never hand-edit it. Take one side and regenerate:

just relock          # or: just relock theirs

Deploy to the Spark

just check && just each test          # gate locally first
just deploy hello-agent spark.local   # rsync source + apply on the box
just deploy hello-agent spark.local --image   # ship a built image instead

Layout

Path	What
agents/<name>/	One self-contained agent project. Own manifest + policy, shares the root uv.lock.
services/dashboard/	SquidWard React dashboard. Own pnpm lockfile and build.
services/ingestion/	Durable SQLite ingestion API; GPU/memory telemetry, findings, recommendations.
services/processing/	Deterministic rules + Isolation Forest/autoencoder detection pipeline.
services/demo/	Demo traffic generation and live/simulate CLIs (just s demo ...).
services/collector/	Event collection service.
libs/agentkit/	Shared Python package — FastAPI service factory, agents.yaml/policy.yaml validation.
libs/policies/	Reusable OpenShell policy fragments.
libs/skills/	Shared OpenClaw/Claude skills, symlinked into claude-plugin/.
contracts/	Shared, review-gated schema/contract definitions.
infra/	GB10 deployment (Ansible/Docker) and OpenClaw desired state.
scripts/deploy.sh	rsync+apply, or image push.