Merge pull request #69 from OpenManus/fix/gaia-output-format

revise gaia code to make it align with React output format + add test…

Author: mannaandpoem

openmanus_rl/agentgym/agentenv-gaia/agentenv_gaia/environment.py

@@ -17,7 +17,14 @@
 
 # List of default tools to include in every environment
 DEFAULT_TOOLS = ["web_search", "bash", "python_execute", "browser_use", "terminate"]
-
+# Define valid action types
+VALID_ACTIONS = {
+    "web_search": {"param": "query"},
+    "bash": {"param": "command"},
+    "python_execute": {"param": "code"},
+    "browser_use": {"param": "input"},
+    "terminate": {"param": "answer"}
+}
 
 
 # GaiaEnvServer class to manage environment instances
@@ -239,32 +246,45 @@ def step(self, env_id: str, action: str):
         with self.env_locks[env_id]:
             env = self.env_instances[env_id]
 
-            # Parse and process the action
-            action_type, action_input = self._parse_action(action)
+            try:
+                # Parse and process the action
+                action_type, action_input = self._parse_action(action)
 
-            # Update state
-            env["state"]["steps_taken"] += 1
+                # Update state
+                env["state"]["steps_taken"] += 1
 
-            # Process the action using available tools
-            observation, reward, done = self._process_action(env_id, action_type, action_input)
+                # Process the action using available tools
+                observation, reward, done = self._process_action(env_id, action_type, action_input)
 
-            # Store action and result in memory
-            env["state"]["memory"].append({
-                "action": action,
-                "result": observation,
-                "step": env["state"]["steps_taken"]
-            })
+                # Store action and result in memory
+                env["state"]["memory"].append({
+                    "action": action,
+                    "result": observation,
+                    "step": env["state"]["steps_taken"]
+                })
 
-            # Update state with results
-            env["state"]["reward"] += reward
-            env["state"]["done"] = done
+                # Update state with results
+                env["state"]["reward"] += reward
+                env["state"]["done"] = done
 
-            info = {
-                "steps_taken": env["state"]["steps_taken"],
-                "action_processed": action,
-            }
+                info = {
+                    "steps_taken": env["state"]["steps_taken"],
+                    "action_processed": action,
+                }
 
-            return observation, reward, done, info
+                return observation, reward, done, info
+
+            except ValueError as e:
+                # Handle validation errors
+                error_msg = str(e)
+                observation = f"Error: {error_msg}"
+                reward = -0.1  # Penalty for invalid action
+                done = False
+                info = {
+                    "steps_taken": env["state"]["steps_taken"],
+                    "error": error_msg
+                }
+                return observation, reward, done, info
 
     def _parse_action(self, action: str) -> Tuple[str, Dict[str, Any]]:
         """
@@ -275,53 +295,71 @@ def _parse_action(self, action: str) -> Tuple[str, Dict[str, Any]]:
 
         Returns:
             tuple: (action_type, action_parameters)
+
+        Raises:
+            ValueError: If action format is invalid
         """
         try:
+            # Split by newlines and find the action line
+            lines = action.split('\n')
+            action_line = None
+            for line in lines:
+                if line.strip().startswith('Action:'):
+                    action_line = line.strip()
+                    break
+            
+            if not action_line:
+                raise ValueError("No valid action found in the input")
+
             # Parse actions in format "Action: action_type with Action Input: action_input"
-            if "Action:" in action and "Action Input:" in action:
-                action_parts = action.split("Action Input:", 1)
+            if "Action:" in action_line and "Action Input:" in action_line:
+                action_parts = action_line.split("Action Input:", 1)
                 action_type = action_parts[0].replace("Action:", "").strip()
                 action_input = action_parts[1].strip()
-                return action_type, {"query": action_input} if action_type == "web_search" else {
-                    "command": action_input} if action_type == "bash" else {
-                    "code": action_input} if action_type == "python_execute" else {"input": action_input}
+                
+                # Validate action type
+                if action_type not in VALID_ACTIONS:
+                    raise ValueError(f"Invalid action type: {action_type}. Valid types are: {', '.join(VALID_ACTIONS.keys())}")
+                
+                # Return with appropriate parameter name
+                param_name = VALID_ACTIONS[action_type]["param"]
+                return action_type, {param_name: action_input}
 
             # Parse JSON format with tool_name and parameters
-            elif action.strip().startswith("{") and action.strip().endswith("}"):
+            elif action_line.strip().startswith("{") and action_line.strip().endswith("}"):
                 try:
-                    action_json = json.loads(action)
-                    if "tool_name" in action_json:
-                        tool_name = action_json.pop("tool_name")
-                        return tool_name, action_json
+                    action_json = json.loads(action_line)
+                    if "tool_name" not in action_json:
+                        raise ValueError("JSON format must include 'tool_name' field")
+                    
+                    tool_name = action_json.pop("tool_name")
+                    if tool_name not in VALID_ACTIONS:
+                        raise ValueError(f"Invalid tool name: {tool_name}. Valid tools are: {', '.join(VALID_ACTIONS.keys())}")
+                    
+                    return tool_name, action_json
                 except json.JSONDecodeError:
-                    pass
+                    raise ValueError("Invalid JSON format")
 
             # Parse direct tool calls (e.g., "web_search: query")
-            elif ":" in action:
-                tool_name, input_text = action.split(":", 1)
+            elif ":" in action_line:
+                tool_name, input_text = action_line.split(":", 1)
                 tool_name = tool_name.strip()
                 input_text = input_text.strip()
 
-                # Map to appropriate parameter name based on tool
-                if tool_name.lower() in ["web_search", "search_web"]:
-                    return tool_name, {"query": input_text}
-                elif tool_name.lower() == "bash":
-                    return tool_name, {"command": input_text}
-                elif tool_name.lower() == "python_execute":
-                    return tool_name, {"code": input_text}
-                elif tool_name.lower() == "browser_use":
-                    return tool_name, {"url": input_text}
-                elif tool_name.lower() == "terminate":
-                    return "terminate", {"answer": input_text}
-                else:
-                    return tool_name, {"input": input_text}
+                if tool_name not in VALID_ACTIONS:
+                    raise ValueError(f"Invalid tool name: {tool_name}. Valid tools are: {', '.join(VALID_ACTIONS.keys())}")
 
-            # Handle plain text as a default case
-            return "unknown", {"input": action}
+                param_name = VALID_ACTIONS[tool_name]["param"]
+                return tool_name, {param_name: input_text}
+
+            else:
+                raise ValueError("Invalid action format. Must be one of:\n" +
+                              "1. 'Action: tool_name with Action Input: input'\n" +
+                              "2. JSON format with tool_name and parameters\n" +
+                              "3. 'tool_name: input'")
 
         except Exception as e:
-            print(f"Error parsing action: {e}")
-            return "unknown", {"input": action}
+            raise ValueError(f"Error parsing action: {str(e)}")
 
     def _process_action(self, env_id: str, action_type: str, action_input: Dict[str, Any]) -> Tuple[str, float, bool]:
         """
@@ -495,6 +533,7 @@ def _format_new_task(self, env_id: str):
         observation += "\nUse tools in the format: Action: tool_name with Action Input: your_input\n"
         observation += "\nGive me one action."
 
+
         return observation
 
     def _format_observation(self, env_id: str):

openmanus_rl/agentgym/agentenv-gaia/test_gaia_react.py

@@ -0,0 +1,87 @@
+import requests
+import json
+
+def test_gaia_react_format():
+    # 服务器地址
+    base_url = "http://127.0.0.1:8081"
+    
+    # 1. 创建新环境
+    create_response = requests.post(
+        f"{base_url}/create",
+        json={
+            "data_dir": "data/",
+            "level": "level1",
+            "dataset": "validation"
+        }
+    )
+    print(f"Create response: {create_response.text}")
+    env_idx = create_response.text.strip('"')  # 移除引号
+    print(f"Created environment with ID: {env_idx}")
+    
+    if not env_idx:
+        print("Failed to create environment")
+        return
+    
+    # 2. 获取初始观察
+    obs_response = requests.get(
+        f"{base_url}/observation",
+        params={"env_idx": env_idx}
+    )
+    print(f"Observation response: {obs_response.text}")
+    try:
+        initial_observation = obs_response.json().get("observation")
+    except:
+        initial_observation = obs_response.text
+    print(f"\nInitial observation:\n{initial_observation}")
+    
+    # 3. 测试ReAct格式的响应
+    # 使用完整的ReAct格式，包含Thought和Action
+    react_response = """Thought: I need to search for information about the capital of France.
+Action: web_search Action Input: What is the capital of France?"""
+    
+    # 4. 执行动作
+    step_response = requests.post(
+        f"{base_url}/step",
+        json={
+            "env_idx": env_idx,
+            "action": react_response
+        }
+    )
+    print(f"Step response: {step_response.text}")
+    try:
+        result = step_response.json()
+    except:
+        result = {"response": step_response.text}
+    print(f"\nStep result:\n{json.dumps(result, indent=2)}")
+    
+    # 5. 验证输出格式
+    print("\nValidating ReAct format:")
+    print("1. Response contains 'Thought:' section:", "Thought:" in react_response)
+    print("2. Response contains 'Action:' section:", "Action:" in react_response)
+    print("3. Action format is correct:", "Action Input:" in react_response)
+    
+    # 提取并验证action部分
+    action_line = None
+    for line in react_response.split('\n'):
+        if line.strip().startswith('Action:'):
+            action_line = line.strip()
+            break
+    
+    if action_line:
+        print("4. Action line format is valid:", action_line.startswith('Action:'))
+        print("5. Tool name is valid:", action_line.split("Action:")[1].split("Action Input:")[0].strip() in ["web_search", "bash", "python_execute", "browser_use", "terminate"])
+    
+    # 6. 获取可用动作
+    actions_response = requests.get(
+        f"{base_url}/available_actions",
+        params={"env_idx": env_idx}
+    )
+    print(f"Available actions response: {actions_response.text}")
+    try:
+        available_actions = actions_response.json()
+    except:
+        available_actions = {"response": actions_response.text}
+    print(f"\nAvailable actions:\n{json.dumps(available_actions, indent=2)}")
+
+if __name__ == "__main__":
+    test_gaia_react_format()