feat: add gym-anything (CUA-World) environment adapter

[dzorlu]

Summary

• Adds GymAnythingEnvClient wrapping gym-anything's 250+ Docker-based desktop software environments as an OpenEnv-compatible environment
• Provides reset_async()/step_async() with screenshot observations, coordinate scaling, and programmatic verifier rewards
• Compatible with existing Fleet computer_use pipeline (same tool/observation format)

Test plan

• Smoke test with one environment (e.g., gimp_all_fast_env)

🤖 Generated with Claude Code

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

Bugbot Autofix prepared a fix for the issue found in the latest run.

• ✅ Fixed: Step count tracked but max_steps never enforced
  • Added a max-step check in step_async and included it in the done condition so episodes now terminate when _step_count reaches max_steps.

Or push these changes by commenting:

@cursor push a858d8aac1

Preview (a858d8aac1)

diff --git a/src/envs/gym_anything_env/client.py b/src/envs/gym_anything_env/client.py
--- a/src/envs/gym_anything_env/client.py
+++ b/src/envs/gym_anything_env/client.py
@@ -122,6 +122,7 @@
     async def step_async(self, action: Dict[str, Any]) -> Tuple[Dict, float, bool, Dict]:
         """Execute action, return (obs, reward, done, info)."""
         self._step_count += 1
+        max_steps_reached = self._step_count >= self.max_steps
         is_done = action.get("done", False)
 
         tool_name = action.get("tool", "")
@@ -148,11 +149,13 @@
                 {"type": "image_url", "image_url": {"url": f"data:image/png;base64,{screenshot}"}}
             ]
 
+        done = done or is_done or max_steps_reached
+
         # Normalize reward from 0-100 to 0-1
         if done and "verifier" in info:
             reward = info["verifier"].get("score", 0) / 100.0
 
-        return observation, reward, done or is_done, info
+        return observation, reward, done, info
 
     def close(self):
         if self.ga_env:

_{This Bugbot Autofix run was free. To enable autofix for future PRs, go to the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 4bcd4ac. Configure here.}

[cursor]

Step count tracked but max_steps never enforced

Medium Severity

max_steps is accepted as a constructor parameter and _step_count is incremented every step_async call, but the step count is never compared against max_steps to terminate the episode. The analogous fleet_env task_env.py correctly checks self._step_count >= self.max_steps and sets done accordingly. Without this check, the environment never signals completion due to exceeding the step limit, potentially causing infinite agent loops.

Additional Locations (1)

• src/envs/gym_anything_env/client.py#L75-L87

 

^{Reviewed by Cursor Bugbot for commit 4bcd4ac. Configure here.}