Skip to main content
An Environment is its own Python package under environments/, a uv workspace member with its own pyproject.toml. demo is the minimal shape. math is the full worked example.

​
Package shape

Register the loader as an entry point:
Implement the protocol: splits(), grade(), fingerprint(), and the optional install() and verify_install().

​
Rules that bite

  • Never depend on monte. Declaring monte in your pyproject.toml cycles the uv workspace graph. And the resources server runs in its own venv with no monte installed, so a top-level import monte crashes it. Import monte lazily inside functions, never at module top.
  • The prompt template is yours, and it is frozen. The Environment owns its prompt file, and the prompt bytes are part of the fingerprint. An edit to the prompt is a Measurement change, not a tweak.
  • One name, once. A duplicate Environment name across installed packages refuses at discovery. A broken load_environment() raises a Refusal that names the distribution.

​
Frozen data

Real task data lives on the durable data root, never in the repo. It is seeded by a manifest-verified copy and never derived on a box. Your install() checks the files against manifest.json and registers the Gym benchmark configs. Install must be idempotent.

​
The two twins

Two different things get called a twin. Keep them apart:
  • The Gym row twin: the *.gym.jsonl files, re-derived from the frozen source through the current prompt renderer. These are the rows Gym consumes.
  • The resources-server plant: your FastAPI grading app plus its configs, copied to a name-derived Gym path. Training and eval both grade through this one server.

​
Test what you wrote

  1. uv run pytest: the test_env_* files cover grader and install-hook behavior.
  2. gym env test --resources-server <name> exercises the resources server. gym env start plus a POST to /verify checks the live plant. These are Gym commands, not monte ones.
  3. monte env list confirms discovery. monte env install <name> on the box verifies the data.
  4. A smoke eval against a scratch Measurement closes the loop.

Environment

The concept: what an Environment is and guarantees.

Quickstart

See the loop run against the demo Environment.