Coding Env improvements for further development

[albertodepaola]

🎯 Summary

This PR adds in-repo development support for the coding_env environment and fixes the reward computation bug that prevented the transform pipeline from evaluating code and assigning rewards.

Testing

Runs the environment and prints rewards

$ python ./examples/local_coding_env.py
============================================================
CodingEnv.from_docker_image() Test
============================================================

Creating client from Docker image...
  CodingEnv.from_docker_image('coding-env:latest')

✓ Client created and container started!

Testing the environment:
------------------------------------------------------------

1. Reset:
   stdout: 
   stderr: 
   exit_code: 0
   State: episode_id=35c24391-037d-4d25-bd0b-85ee048de8fe, step_count=0

2. Execute Python code:
   1. Code: print('Hello, World!')...
      → stdout: Hello, World!
      → exit_code: 0
      → reward: 0.1
...

Running reward tests before the fix:

$ pytest ./tests/envs/test_python_codeact_rewards.py
...
=========== 29 failed, 2 passed in 0.46s ===========

After

$ pytest ./tests/envs/test_python_codeact_rewards.py
=========================================================================== test session starts ===========================================================================
platform darwin -- Python 3.12.9, pytest-8.4.2, pluggy-1.5.0
rootdir: /Users/betodepaola/projects/OpenEnv
configfile: pyproject.toml
plugins: asyncio-1.2.0, anyio-4.11.0
asyncio: mode=Mode.STRICT, debug=False, asyncio_default_fixture_loop_scope=None, asyncio_default_test_loop_scope=function
collected 31 items                                                                                                                                                        

tests/envs/test_python_codeact_rewards.py ...............................                                                                                           [100%]

===========  31 passed in 0.36s =========== 

---

✨ Features & Enhancements

• Added dual-mode import support for both standalone (PyPI) and in-repo development

  • Client, server, models, and transforms now support both import paths
  • Allows developers to work on coding_env within the OpenEnv monorepo without publishing to PyPI

• Updated Dockerfile for in-repo build mode

  • Supports BUILD_MODE argument for flexible Docker builds
  • Enables development workflow without PyPI package cycles

• Updated package dependencies

  • Bumped version in pyproject.toml, both for the env and main project, as the docker termination was updated.
  • Fixed relative imports for .models module

• Enhanced README with development instructions

  • Added setup guides for in-repo development workflow

• Improved container cleanup logic

  • Properly removes containers on exit to prevent resource leaks
  • Fixes issue where containers would accumulate after testing

---

🐛 Bug Fixes

Reward Computation Fix

• Fixed reward calculation in python_codeact_env.py
  • Added metadata={"last_code": action.code} to CodeObservation in step() method
  • Enables transform pipeline (CodeSafetyTransform and CodeQualityTransform) to evaluate code
  • Root cause: Transforms require code in metadata["last_code"] to calculate rewards, but it was never being set
  • Impact: Rewards now correctly computed instead of returning None
    • Safe + concise code (≤100 chars): reward = 0.1
    • Safe + verbose code (>100 chars): reward = 0.0
    • Dangerous patterns (import os, eval(), etc.): reward = -1.0
    • Syntax errors: reward = -0.2

Test Script Enhancement

• Added reward output to local_coding_env.py example
  • Displays result.reward for both successful executions and error scenarios
  • Provides visibility into the reward computation for debugging and verification
  • Helps users confirm the reward system is functioning correctly