Free Env

[xingdi-eric-yuan]

This pull request introduces a new "FreeEnv" environment and "FreeAgent" agent, enabling open-ended, container-based exploration sessions in debug_gym. It adds support for running arbitrary container images with customized setup and toolchains, including human-in-the-loop operation. The changes also include configuration and tooling updates to support flexible session management and error handling.

New Environment and Agent:

• Added FreeEnv, a lightweight environment for running arbitrary container images, with support for custom instructions, workspace directories, setup commands, and optional git initialization. (debug_gym/gym/envs/free_env.py)
• Introduced FreeAgent, a minimal reasoning agent designed for FreeEnv sessions, including improved error handling for environment startup issues. (debug_gym/agents/free_agent.py)

Integration and Selection:

• Registered FreeEnv and FreeAgent in their respective module initializers, and updated environment selection logic to support the new "free" type. (debug_gym/gym/envs/__init__.py, debug_gym/agents/__init__.py) [1] [2] [3]

Tooling and Session Management:

• Updated the SubmitTool to allow disabling automatic evaluation upon submission, facilitating open-ended sessions. (debug_gym/gym/tools/submit.py)
• Added a configuration file and interactive demo script for FreeEnv sessions, supporting both automated and human-driven workflows. (scripts/config_free_env.yaml, scripts/free_env_human.py) [1] [2]

BaseAgent Improvements:

• Improved BaseAgent to handle memory_size configuration and more robust rewrite step termination logic. (debug_gym/agents/base_agent.py) [1] [2] [3]

[xingdi-eric-yuan]

Missing tests.
update: added tests.

[MarcCote]

What's that for?