Environment variables¶

What Hylo reads from the environment, and the third-party variables its setup relies on. Variables that belong to one machine (paths, renderers, caches) are best set per runtime in the execution profile, so they reach exactly the component that needs them.

Hylo¶

Variable	Read by	Meaning	Default
`HYLO_PROFILE`	`core.profile`	path of this machine's execution profile	`~/.config/hylo/profile.yaml`
`HYLO_STORE`	`core.store`	the run and artifact store directory; overrides `store:` in the profile	the profile's `store`, else `./runs`
`HYLO_NO_PLUGIN_AUTOLOAD`	`core.registry`	`1`: do not load plugins from entry points; only explicitly registered plugins and `module:attr` references resolve	unset (entry points load)
`HYLO_HARNESS`	`bench/envs/libero/setup.sh`	the harness checkout to install into the LIBERO venv	`../harness` (side by side)
`HYLO_GIT_BASE`	`harness/dev/bootstrap.sh`	clone missing repositories from `$HYLO_GIT_BASE/hylo-<name>.git`	unset: missing repositories are skipped
`HYLO_TEST_LLM`	the LLM reasoner tests	`1`: run the tests that call the local Ollama server	unset (skipped)

Tokens are never read from a fixed variable. Pass them on the command line (--token) or name a variable that holds one: hylo serve env|agent ... --token-env HYLO_TOKEN, or token_env: HYLO_TOKEN in a remote.ws agent's params. The templates use $HYLO_TOKEN by convention.

Rendering and simulators¶

Variable	Used by	Meaning
`MUJOCO_GL`	MuJoCo, LIBERO	renderer: `osmesa` (CPU, the user-space Mesa from `bench/scripts/install-osmesa-user.sh`) or `egl` (GPU). Unset: no off-screen rendering; image observations and the rendering tests are skipped
`PYOPENGL_PLATFORM`	LIBERO (robosuite)	must match `MUJOCO_GL` (`osmesa`)
`LD_LIBRARY_PATH`	the `libero-sim` runtime	where the user-space `libOSMesa.so` is (`~/.local/mesa/root/usr/lib/x86_64-linux-gnu`)
`LIBERO_CONFIG_PATH`	LIBERO	the non-interactive LIBERO config written by `setup.sh` (`bench/envs/libero/config`)
`OMP_NUM_THREADS`, `LP_NUM_THREADS`	numpy / OSMesa (llvmpipe)	threads per simulator process; kept low because benchmarks run many workers

Models¶

Variable	Used by	Meaning
`HF_HOME`	LeRobot, Hugging Face Hub	the model and dataset cache (`bench/envs/hf-home` in the `lerobot` runtime)
`HF_HUB_OFFLINE`	LeRobot	`1`: never download; use the cache. Set in the `lerobot` runtime and by the self-training learner (unless already set)
`TOKENIZERS_PARALLELISM`	tokenizers	`false`: no fork warnings in worker processes (set by the action-memory builder and the self-training learner)
`TORCHDYNAMO_DISABLE`	torch	`1`: no torch.compile; set by some LIBERO configs (`libero10_memory`, `libero10_noise`) and by self-training

Where to set them¶

In the profile (recommended)In a config (one-off)In the shell

uv run hylo profile set-runtime libero-sim --python bench/envs/libero/.venv/bin/python \
    --env MUJOCO_GL=osmesa --env PYOPENGL_PLATFORM=osmesa

env:
  type: libero
  remote: {runtime: libero-sim, env: {OMP_NUM_THREADS: "2", LP_NUM_THREADS: "4"}}

MUJOCO_GL=osmesa uv run pytest bench/adapters/hylo-env-mujoco

Variables in the profile or a config reach only the subprocess of that component; the parent process and other components are not affected. None of them enter the experiment digest.