Multi-tower inference: remote command delegation - Windows and cuda support - end to end installation fixes

[joelteply]

Summary

• Automated Continuum install for all platforms (macOS, Ubuntu/Debian, WSL2)
• GPU detection (CUDA, Metal, CPU-only) with appropriate ML package selection
• Tower-aware command routing (upcoming): Commands.execute() transparently delegates to remote towers
• Tower URI scheme: tower://5090 local → mesh address long-term

Test plan

• Install script works on 5090 WSL2
• npm start succeeds on 5090
• ./jtag ping responds from 5090
• Academy benchmark routes through remote tower

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Adds a new one-shot installer script intended to bootstrap a “Continuum Tower” environment across macOS/Linux/WSL by installing toolchains, optionally setting up a Python ML venv based on GPU detection, building the project, and writing a default user config.

Changes:

• Introduces src/scripts/install.sh to install system dependencies, Node.js, Rust, and (optionally) a Python ML venv.
• Adds basic GPU detection (CUDA vs Metal vs CPU-only) to decide which PyTorch wheel index to use.
• Runs the repo build (npm install, npm run build:ts, scripts/setup-rust.sh) and creates a default ~/.continuum/config.env if missing.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

These build steps pipe output to tail, which will mask failures unless pipefail is enabled (and even then truncates the logs needed to debug). Consider removing the | tail ... truncation or capturing logs to a file while still letting failures surface clearly.

Suggested change

	npm install --silent 2>&1 | tail -3
	echo -e " Building TypeScript..."
	npm run build:ts 2>&1 | tail -1
	echo -e " Building Rust workers..."
	bash scripts/setup-rust.sh 2>&1 | tail -5
	npm install --silent > npm_install.log 2>&1
	tail -3 npm_install.log
	echo -e " Building TypeScript..."
	npm run build:ts > build_ts.log 2>&1
	tail -1 build_ts.log
	echo -e " Building Rust workers..."
	bash scripts/setup-rust.sh > setup_rust.log 2>&1
	tail -5 setup_rust.log

[Copilot]

On Linux/WSL the script uses sudo unconditionally for NodeSource setup + apt install. This will fail when running as root (or in minimal environments without sudo). You already compute a SUDO variable in install_system_deps; reuse that pattern here or detect id -u before calling sudo.

Suggested change

	curl -fsSL https://deb.nodesource.com/setup_22.x | sudo -E bash -
	sudo apt-get install -y nodejs
	# Determine whether to use sudo (for non-root users with sudo installed)
	local SUDO_CMD="sudo"
	if [ "$(id -u)" -eq 0 ] || ! command -v sudo >/dev/null 2>&1; then
	SUDO_CMD=""
	fi
	if [ -n "$SUDO_CMD" ]; then
	curl -fsSL https://deb.nodesource.com/setup_22.x | sudo -E bash -
	sudo apt-get install -y nodejs
	else
	curl -fsSL https://deb.nodesource.com/setup_22.x | bash -
	apt-get install -y nodejs
	fi

[Copilot]

preflight_detect_platform can return unknown, but most case "$PLATFORM" statements have no default branch and still print success messages. This can lead to a confusing partial install on unsupported systems; consider explicitly erroring out if PLATFORM is unknown (or adding *) cases that exit 1 with a clear message).

Suggested change

	PLATFORM=$(preflight_detect_platform)
	PLATFORM=$(preflight_detect_platform)
	if [ "$PLATFORM" = "unknown" ]; then
	echo -e " Platform: ${RED}unknown${NC}"
	echo ""
	echo -e "${RED}Error:${NC} Unsupported platform detected by preflight_detect_platform."
	echo "This installer currently supports: macOS (Apple Silicon), Ubuntu/Debian (x86_64), and WSL2."
	echo "If you're on a supported system and see this message, please update your scripts or open an issue."
	exit 1
	fi

[Copilot]

The script pipes several critical build/install commands into tail later on, but only set -e is enabled. In bash, a failing command in a pipeline won't fail the script unless set -o pipefail is set, so npm install / npm run build:ts / setup-rust.sh failures could be silently ignored.

Suggested change

	set -e
	set -e
	set -o pipefail

[Copilot]

Linux/WSL system deps don't ensure python3 is installed, but later steps call python3 -m venv unconditionally. On minimal Debian/Ubuntu images this will fail even if pythonX.Y-venv is installed; consider adding python3 (and usually python3-pip) to the needed list when python3 isn't present.

Suggested change

	# Python venv support
	if ! python3 -m venv --help &>/dev/null 2>&1; then
	# Detect python version for correct package name
	local pyver=$(python3 -c "import sys; print(f'{sys.version_info.major}.{sys.version_info.minor}')" 2>/dev/null || echo "3.12")
	needed+=("python${pyver}-venv")
	# Python + venv support
	if command -v python3 &>/dev/null; then
	# Ensure venv module is available for the installed python3
	if ! python3 -m venv --help &>/dev/null 2>&1; then
	# Detect python version for correct package name
	local pyver
	pyver=$(python3 -c "import sys; print(f'{sys.version_info.major}.{sys.version_info.minor}')" 2>/dev/null || echo "3.12")
	needed+=("python${pyver}-venv")
	fi
	else
	# No python3 installed: install interpreter, pip, and generic venv package
	needed+=("python3" "python3-pip" "python3-venv")

[Copilot]

On macOS this assumes Homebrew is available (brew install ...) but there's no check or guidance if brew is missing. Consider checking command -v brew and either installing it or printing a clear instruction before attempting installs.

[Copilot]

bitsandbytes is installed for both CUDA and Metal paths, but it is generally CUDA-specific and commonly fails to install/import on macOS/MPS environments. This will likely break the install on Apple Silicon; consider only installing bitsandbytes when HAS_CUDA=true (and skipping it for Metal/CPU).

Suggested change

	$pip install -q transformers peft accelerate datasets trl bitsandbytes pytest
	$pip install -q transformers peft accelerate datasets trl pytest
	if $HAS_CUDA; then
	echo -e " Installing CUDA-specific packages..."
	$pip install -q bitsandbytes
	fi

[Copilot]

The PR title/description focus on multi-tower inference and remote command delegation, but the only change in this PR is adding an install script. Either the PR metadata should be updated to reflect the installer work, or the missing multi-tower changes should be included so the PR matches its stated purpose.

[Copilot]

The installer writes its own minimal ~/.continuum/config.env template, but the repo already has scripts/ensure-config.ts as the canonical config generator (with required defaults like HTTP_PORT/WS_PORT, LOG_LEVEL, etc.). Creating a partial file here will prevent ensure-config.ts from generating the full default config and can leave the system missing expected keys; consider invoking the existing ensure-config script (or reusing its template) instead of writing a second format.

Suggested change

	echo -e "\n${YELLOW}Creating default config at $CONFIG_FILE${NC}"
	cat > "$CONFIG_FILE" << 'ENVEOF'
	# Continuum Tower Configuration
	# Add API keys here for cloud provider access.
	# Uncomment and fill in the ones you need:
	# ANTHROPIC_API_KEY=sk-ant-...
	# OPENAI_API_KEY=sk-proj-...
	# DEEPSEEK_API_KEY=sk-...
	# HF_TOKEN=hf_...
	# Storage path for models, adapters, datasets (default: ~/.continuum)
	# CONTINUUM_STORAGE_PATH=/path/to/storage
	ENVEOF
	echo -e " ${YELLOW}Edit $CONFIG_FILE to add your API keys${NC}"
	echo -e "\n${YELLOW}No config found at $CONFIG_FILE — generating default config via ensure-config${NC}"
	if command -v npm >/dev/null 2>&1; then
	(
	cd "$PROJECT_DIR"
	# Use the canonical config generator to create a full default config.env
	if ! npm run ensure-config >/dev/null 2>&1; then
	echo -e " ${YELLOW}Warning:${NC} 'npm run ensure-config' failed. You may need to run it manually after installation."
	fi
	)
	else
	echo -e " ${YELLOW}Warning:${NC} npm is not available. Please run 'npm run ensure-config' from the src directory to create $CONFIG_FILE."
	fi

[Copilot]

The post-install instruction cd src && npm start conflicts with the usage instructions at the top (cd continuum/src before running this script). If the user is already in src/, cd src will land in src/src and fail; consider printing either just npm start or an unambiguous path from the repo root.

Suggested change

	echo -e " ${YELLOW}Start:${NC} cd src && npm start"
	echo -e " ${YELLOW}Start:${NC} npm start"