textarena_env: parse TEXTARENA_KW_* values as JSON

docs/source/environments/textarena.md

@@ -180,7 +180,7 @@ The server supports configuration via environment variables:
 - `TEXTARENA_NUM_PLAYERS` - Number of players (default: 1)
 - `TEXTARENA_MAX_TURNS` - Maximum turns per episode
 - `TEXTARENA_DOWNLOAD_NLTK` - Download NLTK data (default: "1")
-- `TEXTARENA_KW_*` - Pass additional kwargs to TextArena (e.g., `TEXTARENA_KW_difficulty=hard`)
+- `TEXTARENA_KW_*` - Pass additional kwargs to TextArena. Values are read as JSON, so numbers and booleans keep their type (e.g., `TEXTARENA_KW_CLUES=65` for Sudoku, `TEXTARENA_KW_difficulty=hard`)
 
 ## Development & Testing
 

envs/textarena_env/README.md

@@ -192,7 +192,7 @@ The server supports configuration via environment variables:
 - `TEXTARENA_NUM_PLAYERS` - Number of players (default: 1)
 - `TEXTARENA_MAX_TURNS` - Maximum turns per episode
 - `TEXTARENA_DOWNLOAD_NLTK` - Download NLTK data (default: "1")
-- `TEXTARENA_KW_*` - Pass additional kwargs to TextArena (e.g., `TEXTARENA_KW_difficulty=hard`)
+- `TEXTARENA_KW_*` - Pass additional kwargs to TextArena. Values are read as JSON, so numbers and booleans keep their type (e.g., `TEXTARENA_KW_CLUES=65` for Sudoku, `TEXTARENA_KW_difficulty=hard`)
 
 ## Development & Testing
 

envs/textarena_env/server/app.py

@@ -9,8 +9,10 @@
 from __future__ import annotations
 
 import inspect
+import json
 import logging
 import os
+from typing import Any
 
 from openenv.core.env_server.http_server import create_app
 
@@ -31,14 +33,21 @@
     from .gradio_ui import build_textarena_gradio_app
 
 
-def _parse_env_kwargs(prefix: str = "TEXTARENA_KW_") -> dict[str, str]:
-    """Collect arbitrary environment kwargs from the process environment."""
+def _parse_env_kwargs(prefix: str = "TEXTARENA_KW_") -> dict[str, Any]:
+    """Collect arbitrary environment kwargs from the process environment.
 
-    env_kwargs: dict[str, str] = {}
+    Values are read as JSON, so `TEXTARENA_KW_CLUES=65` passes the integer 65. Values that are not JSON, like
+    `TEXTARENA_KW_DIFFICULTY=hard`, are passed as strings.
+    """
+
+    env_kwargs: dict[str, Any] = {}
     for key, value in os.environ.items():
         if key.startswith(prefix):
             env_key = key[len(prefix) :].lower()
-            env_kwargs[env_key] = value
+            try:
+                env_kwargs[env_key] = json.loads(value)
+            except json.JSONDecodeError:
+                env_kwargs[env_key] = value
     return env_kwargs
 
 

tests/envs/test_textarena_environment.py

@@ -536,3 +536,25 @@ def test_bandit_seeded_trajectory_is_reproducible_across_sessions():
 
     assert sum(len(v) for v in _bandit_history(first).values()) == len(actions)
     assert _bandit_history(first) == _bandit_history(second)
+
+
+def test_env_kwargs_are_parsed_as_json(monkeypatch):
+    pytest.importorskip("textarena", reason="textarena not installed")
+    from textarena_env.server.app import _parse_env_kwargs
+
+    monkeypatch.setenv("TEXTARENA_KW_CLUES", "65")
+    monkeypatch.setenv("TEXTARENA_KW_DIFFICULTY", "hard")
+    monkeypatch.setenv("TEXTARENA_KW_HARDCORE", "true")
+
+    assert _parse_env_kwargs() == {"clues": 65, "difficulty": "hard", "hardcore": True}
+
+
+def test_sudoku_clues_from_env_kwargs(monkeypatch):
+    pytest.importorskip("textarena", reason="textarena not installed")
+    from textarena_env.server.app import _parse_env_kwargs
+
+    monkeypatch.setenv("TEXTARENA_KW_CLUES", "65")
+    env = TextArenaEnvironment(env_id="Sudoku-v0", env_kwargs=_parse_env_kwargs())
+    board = env.reset().messages[0].content.split("Game Board:")[-1]
+
+    assert board.count(".") <= 81 - 65