feat: add environments/ and example_environments/

[cmunley1]

Summary

Introduces top-level environments/ and example_environments/ directories to give training and example environments a consistent home alongside their data and preparation scripts. Similar to benchmarks/ for evaluation environments.

Previously, training environment configs were scattered inside resources_servers/ or responses_api_agents/ making it hard to discover what environments exist and how to run them.

Structure

Environments are now located at:

benchmarks/
environments/
example_environments/

One example of each

• environments/workplace_assistant_simple_agent
• example_environments/example_single_tool_call

Each environment has a self-contained config.yaml, README.md, prepare.py, and example input data and rollouts. README copied directly from resources server.

Example usage

Start environment

config_paths="responses_api_models/openai_model/configs/openai_model.yaml,\
environments/workplace_assistant_simple_agent/config.yaml"
ng_run "+config_paths=[$config_paths]"

Collect rollouts

ng_collect_rollouts +agent_name=workplace_assistant_simple_agent \
    +input_jsonl_fpath=environments/workplace_assistant_simple_agent/data/example.jsonl \
    +output_jsonl_fpath=results/workplace_assistant_trajectory_collection.jsonl \
   +limit=1

Next steps

• Migration of existing environments
• CLI extension (ng_prepare_environment etc.)
• Environments hub site and public registry

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.