CLAUDE.md

Agent Instructions

You're working inside the WAT framework (Workflows, Agents, Tools).

Project Context

This repository is a setup guide for running NemoClaw (NVIDIA's sandboxed OpenClaw agent framework) on a DGX Spark (GB10, 128GB unified memory, CUDA CC 12.1, ARM64).

• DGX Spark IP: 192.168.42.2
• SSH alias: dgx-spark (user: adadrag, key: ~/.ssh/dgx_spark)
• Sibling repo: qwen3.5-dgx-spark (Atlas benchmarks for Qwen3.5-35B-A3B)

The WAT Architecture

Layer 1: Workflows (The Instructions)

• The README phases are the SOPs: cloud quick-start, Ollama local, Atlas advanced
• Each phase is a linear walkthrough with exact commands

Layer 2: Agents (The Decision-Maker)

• You orchestrate the setup, run benchmarks, and update documentation
• Read the relevant phase, execute commands via SSH, handle failures

Layer 3: Tools (The Execution)

• scripts/setup-nemoclaw.sh — automates NemoClaw + Ollama install
• scripts/setup-atlas.sh — automates Atlas inference engine setup
• benchmarks/benchmark-nemotron.py — Nemotron 3 Super 120B benchmark suite

Critical Rules

1. Never run multiple inference engines simultaneously on the DGX Spark. Stop one before starting another (OOM risk — causes system freeze requiring power cycle).
2. Always stop Ollama before starting Atlas: systemctl stop ollama
3. Always stop Atlas before starting Ollama: docker stop atlas
4. HuggingFace cache is owned by root — fix with: docker run --rm -v $HOME/.cache/huggingface:/hf alpine chown -R $(id -u):$(id -g) /hf
5. NemoClaw local inference requires: NEMOCLAW_EXPERIMENTAL=1

How to Operate

1. Look for existing tools first — check scripts/ and benchmarks/
2. When things fail: read the error, fix the script, retest, update the README
3. Keep the README current with any new findings