env subcommand, with all of their flags.
Prefix each command with cadenza (or cadenza-cli from a source install).
Inside the interactive shell you omit the prefix and
type the command directly.
Which commands need an extra? The core CLI (scaffold, show, stats, cache,
finetune-export, account) runs with no extra dependencies. Simulation needs
gym, LoRA needs lora, residual RL + distillation need rl, and real VLA
loading needs vla. See Configuration.
Commands that produce a governed verdict (lora eval, lora finetune --gate,
residual train/eval/bench, distill eval, vla grd/eval) talk to the
Cadenza API and require
sign-in.Projects
Run & analyze
On macOS the windowed viewer needs
mjpython. Pass --headless, or relaunch
with mjpython -m cadenzalabs. See Running missions.Actions (require sign-in)
Fine-tuning data & prompts
LoRA action head (lora extra)
Fine-tunes and governs the cadenza-lab LoRA action head for a project (on the
project’s own base/VLA if it ships a lora_encoder.py). See
Training & fine-tuning.
Residual RL (rl extra)
Closed-loop residual RL on the frozen base: a = clamp(a_base + α·gate·Δa).
Runs on a deterministic proxy task without cadenza-lab; --real targets the
gated sim seam. See Residual RL & distillation.
Distillation (rl extra)
Distill the residual teacher into a compact, base-free student that runs on
CPU/MPS (optional int8). See Residual RL & distillation.
VLA-only mode + GRD (lora extra; vla for real models)
A VLA-only project is a model + an action library in vla.json — no env.json,
no prompts. The GRD loop adapts only the model’s LoRA adapter in one governed
fine-tune + RL pass. See VLA mode & GRD.
Megan (require sign-in)
Use Cadenza as an API from your own project — your token is the API key. The Megan SDK (from cadenzalabs import MeganTK) is standard-library only; these
two CLI commands support it. See the Megan guide.