Add custom task system for BrowserGym environment

.gitignore

@@ -53,6 +53,7 @@ coverage.xml
 .env
 .venv
 venv/
+myenv/
 ENV/
 env.bak/
 venv.bak/

examples/browsergym_custom_example.py

@@ -0,0 +1,134 @@
+"""Example usage of custom BrowserGym tasks.
+
+This script demonstrates how to create and use custom tasks with the
+BrowserGym environment wrapper in OpenEnv.
+"""
+
+import sys
+import os
+import time
+
+# Add src to path for imports
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..', '..', '..'))
+
+from envs.browsergym_env.server.browsergym_environment import BrowserGymEnvironment
+from envs.browsergym_env.models import BrowserGymAction
+
+
+def multi_tab_copy_paste_example():
+    """Run the multi-tab copy-paste example."""
+
+    print("Multi-Tab Copy-Paste Task Example")
+    print("-" * 80)
+
+    # Create environment
+    env = BrowserGymEnvironment(
+        benchmark="custom",
+        task_name="copy-paste-multitab",
+        headless=False,
+        viewport_width=1280,
+        viewport_height=720,
+        timeout=10000.0,
+    )
+
+    # Reset environment
+    obs = env.reset()
+    print(f"Goal: {obs.goal}\n")
+
+    # Solve the multi-tab task- simulates user actions
+    steps = [
+        ("Select source text", "click('#source-text')"),
+        ("Select all text", "press('Control+A')"),
+        ("Copy text", "press('Control+C')"),
+        ("Navigate to target page", "click('#open-target-btn')"),
+        ("Click target input field", "click('#target-text')"),
+        ("Paste text", "press('Control+V')"),
+        ("Submit form", "click('#submit-btn')"),
+    ]
+
+    for i, (description, action_str) in enumerate(steps, 1):
+        print(f"Step {i}: {description}")
+        action = BrowserGymAction(action_str=action_str)
+        obs = env.step(action)
+
+        # Show which page we're on
+        current_page = "unknown"
+        if obs.metadata and 'custom_data' in obs.metadata:
+            current_page = obs.metadata['custom_data'].get('current_page', 'unknown')
+
+        print(f"  Reward: {obs.reward}, Done: {obs.done}, Page: {current_page}")
+
+        # Add delay to see the browser actions
+        time.sleep(1)
+
+        if obs.done:
+            print(f"\n✓ Task completed! Total reward: {env.state.cum_reward}")
+            break
+
+    env.close()
+    print("-" * 80)
+
+def single_tab_copy_paste_example():
+    """Run the single-tab copy-paste example."""
+
+    print("Custom BrowserGym Task Example: Copy-Paste")
+    print("-" * 80)
+
+    # Create environment
+    env = BrowserGymEnvironment(
+        benchmark="custom",
+        task_name="copy-paste",
+        headless=False,
+        viewport_width=1280,
+        viewport_height=720,
+        timeout=10000.0,
+    )
+
+    # Reset environment
+    obs = env.reset()
+    print(f"Goal: {obs.goal}\n")
+
+    # Solve the task
+    steps = [
+        ("Click source text field", "click('#source-text')"),
+        ("Select all text", "press('Control+A')"),
+        ("Copy text", "press('Control+C')"),
+        ("Click target field", "click('#target-text')"),
+        ("Paste text", "press('Control+V')"),
+        ("Click submit button", "click('#submit-btn')"),
+    ]
+
+    for i, (description, action_str) in enumerate(steps, 1):
+        print(f"Step {i}: {description}")
+        action = BrowserGymAction(action_str=action_str)
+        obs = env.step(action)
+        print(f"  Reward: {obs.reward}, Done: {obs.done}")
+
+        # Add delay to see the browser actions
+        time.sleep(1)
+
+        if obs.done:
+            print(f"\n✓ Task completed! Total reward: {env.state.cum_reward}")
+            break
+
+    env.close()
+    print("-" * 80)
+
+def main():
+    """Run the custom task example."""
+
+    # Run single-tab copy-paste example
+    print("Single-Tab Copy-Paste")
+    single_tab_copy_paste_example()
+
+    time.sleep(3)
+
+    # Run multi-tab copy-paste example
+    print("\nMulti-Tab Copy-Paste")
+    multi_tab_copy_paste_example()
+
+
+if __name__ == "__main__":
+    # Run main example
+    main()
+

rfcs/005-generic-task-integration.md

@@ -0,0 +1,82 @@
+# RFC: Generic Custom Task Integration
+
+**Status**: Draft
+**Created**: 06/12/2025
+**Authors**: @atchudhansg
+**RFC ID:** 004
+
+## Summary
+This RFC proposes a standardized mechanism for injecting custom task logic into OpenEnv environments at runtime. This allows users to define domain-specific tasks, rewards, and termination conditions without modifying the core environment code or rebuilding Docker images.
+
+## Motivation
+Currently, adding new tasks to an OpenEnv environment (like BrowserGym) typically requires:
+1.  Modifying the source code of the environment.
+2.  Rebuilding the Docker image.
+3.  Waiting for upstream PRs to be merged for official benchmarks.
+
+This slows down research and prototyping. Users often need to:
+-   Test agents on proprietary or local tasks.
+-   Rapidly iterate on reward functions.
+-   Create custom curricula of tasks.
+
+We need a way to "plug in" task definitions dynamically.
+
+## Proposal
+
+### 1. Task Interface
+We define a standard protocol that any custom task must implement. This should be generic enough for various environments (Browser, Terminal, etc.).
+
+```python
+class CustomTask:
+    def setup(self, config: dict) -> None:
+        """Initialize task resources."""
+        pass
+
+    def get_observation(self, env_state: Any) -> dict:
+        """Transform raw environment state into agent observation."""
+        pass
+
+    def calculate_reward(self, state: Any, action: Any, result: Any) -> float:
+        """Compute reward for the transition."""
+        pass
+
+    def check_done(self, state: Any) -> bool:
+        """Determine if the episode should terminate."""
+        pass
+```
+
+### 2. Injection Mechanism
+Environments should support loading these tasks from a specific directory or module path at runtime.
+
+-   **Volume Mount**: Users mount their task code to a standard path (e.g., `/opt/openenv/custom_tasks`).
+-   **Dynamic Loading**: The environment server scans this directory and registers valid task classes.
+-   **Configuration**: A standard environment variable (e.g., `OPENENV_CUSTOM_TASK_DIR`) tells the server where to look.
+
+### 3. Usage Example (BrowserGym)
+In the context of BrowserGym (as implemented in PR #X), this looks like:
+
+```python
+# User defines a task locally
+class MyTask(CustomTask):
+    ...
+
+# User runs the environment with the custom task
+env = BrowserGymEnv(
+    environment={
+        "BROWSERGYM_BENCHMARK": "custom",
+        "BROWSERGYM_TASK_NAME": "my-task"
+    },
+    volumes={
+        "/local/path/to/tasks": "/opt/openenv/custom_tasks"
+    }
+)
+```
+
+## Benefits
+-   **Decoupling**: Task logic is separate from environment infrastructure.
+-   **Speed**: No rebuilds required to change a reward function.
+-   **Flexibility**: Supports private/proprietary tasks that cannot be upstreamed.
+
+## Future Work
+-   Define schemas for task configuration.
+-   Support for remote task definitions (loading from URL/Git).