Skip to main content
The CLI is organized around projects. A project is a directory with an env.json mission spec. You scaffold it, run it, analyze it, and turn it into training data.

​
The workflow

1

Scaffold

mkdir or env init create a project and its env.json.
2

Author

Edit env.json to define the scene, phases, success/fail predicates, and reward shaping. See the schema.
3

Run

env run executes the mission in simulation, logging every tick and scoring the result with the LLM judge.
4

Improve

env stats analyzes runs. env finetune, env train, and env lora turn them into better policies and prompts.

​
Command map

The full table with every flag is in the Command reference. The newer learning stages have their own guides: Residual RL & distillation, VLA mode & GRD, and the shared Governance gate that decides what’s safe to deploy. To call Cadenza from your own project, see Megan — your token as an API key.

​
Two ways to run

Inside the shell, omit the leading cadenza:
Press Tab in the shell to complete commands and env subcommands, and Ctrl+L to clear the screen / redraw the banner.