Install

Python 3.11, 3.12 or 3.13.

Not yet on PyPI — publishing is a pending decision, and nothing technical stands in its way. Until it is taken, install from source:

git clone https://github.com/TalkShopClub/morpheus-crl
cd morpheus-crl
pip install -e .

Extras

The core install is deliberately small; take only the layer you need.

Install	Adds	Pulls
pip install -e .	Rewards, configuration shifts, the trace record, the harness contract, the benchmark metrics, the command line, and a live MORPHEUS world	numpy, gymnasium, pydantic, pyyaml, httpx
pip install -e ".[train]"	The ported algorithm backends (ppo, ppo_ewc, ppo_lcm)	torch, transformers, peft
pip install -e ".[eval]" / ".[morpheus]"	Nothing — both layers need only the base install; the empty extras are kept so the commands work and say something true	—

Everything a base install claims is enforced by test: the metrics never need torch, reaching a live world needs only httpx, and naming a GPU backend in a configuration costs no import — the backend resolves only when a run actually starts, so documents validate on a laptop with no ML stack.

One platform note: macOS on Intel cannot install [train] — PyTorch's last x86_64 macOS wheel predates what transformers requires, a permanent ceiling rather than a resolution failure. Use an arm64 interpreter.

Verify it

python -c "import skyfall_crl; print(skyfall_crl.__version__)"
skyfall-crl --help
skyfall-crl demo

The third command is the five-minute tutorial: a complete continual-learning run, scored and explained. The package ships a py.typed marker, so editors and type checkers see the annotations without a stub package.

The command line

One skyfall-crl command, one subcommand per surface — demo, run, eval, aggregate, serve, conformance; the CLI reference has every flag. It also runs uninstalled from a checkout: python -m skyfall_crl.cli ….