[ENVIRONMENT] textarena wrapper env

[burtenshaw]

This environment wraps the library textarena with an OpenEnv server. This allows users to do RL on gams like wordle, and many more.

The PR contains:

• a server implementation
• DockerFile
• basic docs in readme
• a hello world example script
• an inference example using gpt-oss-120b

Open to feedback if it doesn't conform to any library norms.

The environment is deployed here
And I used it to train a model with GRPO in TRL here

Here's an example from a wordle inference:

🎯 Turn 5: model replied with -> [train]
   Parsed guess: [train]
   Feedback messages:
     [MESSAGE] [train]
     [MESSAGE] Player 0 submitted [train].
Feedback:
A R I S E
Y G Y X X

G R A I N
X G G G G

A R G U E
Y G X X X

B R A I N
X G G G G

T R A I N
X G G G G

You have 1 guesses left.

🎯 Turn 6: model replied with -> [drain]
   Parsed guess: [drain]
   Feedback messages:
     [MESSAGE] [drain]
     [MESSAGE] Player 0 won the game. Reason: Player 0 guessed the word correctly!
     [MESSAGE] The game ended in a draw. Reason: Turn limit reached.

✅ Game finished
   Reward: 0.0
   Done: True

[pankit-eng]

If you would like this env to be built as part of the repo build pipeline, then please update the .github/workflows/docker_build file to ensure it gets built the same way as other envs.

[burtenshaw]

Thanks for the review @pankit-eng . I updated the env to use the docker image and heard about consistency.

[rycerzes]

Hi @burtenshaw, I've been browsing through the PR and I had a quick question about the Wordle-specific reward signals you added: since the wrapper is designed to work with any TextArena game, I'm wondering if those auxiliary rewards (greens/yellows/etc.) are better kept in the environment for convenience, or if they should be handled in training code to keep things more modular. I think the wrapper should remain fully remain generic like other environments only providing the main reward value without any auxiliary/shaped reward signals computed in the environment.

[burtenshaw]

@rycerzes your feedback makes total sense. when I implemented this I wasn't sure about the consensus on sharing rewards for convenience vs many envs. in the last week, it seems like we're going for the latter. I'll open a new PR.