Command-Line Reference
Commands
| Command |
What it does |
pyrualean api |
Print the API reference the agent reads (LIBERO-PRO's robot). --primitives no-vla leaves the VLA out. |
pyrualean prompt |
Print the prompt of a single shot (generate); --card renders it for a task card. A play episode saves the prompt its model read as prompt.txt in the run directory. |
pyrualean play |
Run one episode with an agent (Running Episodes). |
pyrualean card |
Boot only the simulator and save a task card. |
pyrualean generate |
Ask the model for a whole program from a task card (single shot). |
pyrualean run |
Run a program file for one episode. |
pyrualean summarize |
Tabulate run directories (Output Files). |
The task instance
play, card and run take the same options:
| Option |
Meaning (default) |
--robot |
libero, robotwin or robocasa (libero). play also takes robodojo_xpl_rgb, which starts RoboDojo's own client itself (it needs ROBODOJO_SOURCE_ROOT and one environment with both sides); RoboDojo's official evaluation runs through XPolicyLab (RoboDojo) |
--suite |
LIBERO suite, RoboTwin task config or RoboCasa split |
--task |
LIBERO task index, or the robot's task name |
--seed |
the scene seed (0; always pass it, since the robots' own defaults are not used) |
--out |
the run directory |
--cuda-device |
the GPU the episode's servers use |
--max-episode-steps |
the simulator step limit (10000) |
--rpent-root, --python |
the RPent checkout and the interpreter it is installed in |
play
| Option |
Meaning (default) |
--arm |
cells or program |
--runtime |
codex, claude or openai (codex; Agent Configuration) |
--model, --reasoning |
the model (gpt-6-astra) and its reasoning effort (xhigh) |
--base-url, --api-key-env |
an OpenAI-compatible endpoint, and the variable holding its key |
--max-turns |
the decision budget, in code cells |
--budget-s |
the wall-clock budget, in seconds (1200) |
--images |
on-motion, on-demand or none (on-motion) |
--feedback |
rich or pure (rich) |
--primitives |
full or no-vla (full) |
--max-programs |
the most programs on the program arm |
--guides, --guides-dir |
operating guides in the agent's workspace |
--vla-endpoint, --sam3-endpoint |
running model servers to attach to |
--codex-home, --codex-bin, --claude-cli |
the runtimes' homes and executables |
Environment variables
| Variable |
Meaning |
RPENT_ROOT |
the RPent checkout, when RPent is not installed as a package |
PYRUALEAN_RUNTIME |
default of --runtime |
PYRUALEAN_PRIMITIVES |
default of --primitives |
PYRUALEAN_CODEX_BIN, PYRUALEAN_CODEX_HOME |
defaults of --codex-bin, --codex-home |
PYRUALEAN_CLAUDE_CLI |
default of --claude-cli |
PYRUALEAN_OPENAI_EXTRA_BODY |
default of --openai-extra-body |
PYRUALEAN_CODEX_RETRIES |
retries of a request from Codex (play, generate) to a --base-url endpoint (10) |
OPENAI_BASE_URL, OPENAI_API_KEY |
the endpoint and key of --runtime openai when --base-url / --api-key-env are not given |
PYRUALEAN_OPENAI_RETRIES, PYRUALEAN_OPENAI_TIMEOUT_S |
retries (10) and request timeout in seconds (900) of --runtime openai |
CODEX_AUTO_COMPACT_TOKEN_LIMIT |
Codex's compaction threshold for the episode (Agent Configuration) |
pyrualean play passes --runtime and --primitives on only when they differ from codex and full, so a
PYRUALEAN_RUNTIME or PYRUALEAN_PRIMITIVES in the environment also wins over an explicit --runtime codex or
--primitives full: unset it to get those.
Each robot page lists the variables of its stack: checkpoints, assets and rendering.
Development
pip install -e '.[dev]'
pytest # offline tests against scripted backends
ruff check .