refactor FinishTool to builtin tools for agents (#39)

Author: xingyaoww

.openhands/microagents/repo.md

@@ -163,6 +163,7 @@ This repo has two python packages, with unit tests specifically written for each
 
 <CODE>
 - Avoid hacky trick like `sys.path.insert` when resolving package dependency
+- Use existing packages/libraries instead of implementing yourselves whenever possible.
 </CODE>
 
 <TESTING>

openhands/core/agent/codeact_agent/codeact_agent.py

@@ -8,52 +8,19 @@
     Message as LiteLLMMessage,
     ModelResponse,
 )
-from pydantic import Field, ValidationError
+from pydantic import ValidationError
 
 from openhands.core.context import EnvContext, PromptManager
 from openhands.core.conversation import ConversationCallbackType, ConversationState
 from openhands.core.llm import LLM, Message, TextContent, get_llm_metadata
 from openhands.core.logger import get_logger
-from openhands.core.tool import ActionBase, ObservationBase, Tool, ToolAnnotations
+from openhands.core.tool import BUILT_IN_TOOLS, ActionBase, FinishTool, ObservationBase, Tool
 
 from ..base import AgentBase
 
 
 logger = get_logger(__name__)
 
-"""Finish tool implementation."""
-
-
-class FinishAction(ActionBase):
-    message: str = Field(description="Final message to send to the user.")
-
-
-TOOL_DESCRIPTION = """Signals the completion of the current task or conversation.
-
-Use this tool when:
-- You have successfully completed the user's requested task
-- You cannot proceed further due to technical limitations or missing information
-
-The message should include:
-- A clear summary of actions taken and their results
-- Any next steps for the user
-- Explanation if you're unable to complete the task
-- Any follow-up questions if more information is needed
-"""
-
-
-FINISH_TOOL = Tool(
-    name="finish",
-    input_schema=FinishAction,
-    description=TOOL_DESCRIPTION,
-    annotations=ToolAnnotations(
-        title="finish",
-        readOnlyHint=True,
-        destructiveHint=False,
-        idempotentHint=True,
-        openWorldHint=False,
-    ),
-)
 
 
 class CodeActAgent(AgentBase):
@@ -65,8 +32,9 @@ def __init__(
         system_prompt_filename: str = "system_prompt.j2",
         cli_mode: bool = True,
     ) -> None:
-        assert FINISH_TOOL not in tools, "Finish tool is automatically included and should not be provided."
-        super().__init__(llm=llm, tools=tools + [FINISH_TOOL], env_context=env_context)
+        for tool in BUILT_IN_TOOLS:
+            assert tool not in tools, f"{tool} is automatically included and should not be provided."
+        super().__init__(llm=llm, tools=tools + BUILT_IN_TOOLS, env_context=env_context)
         self.prompt_manager = PromptManager(
             prompt_dir=os.path.join(os.path.dirname(__file__), "prompts"),
             system_prompt_filename=system_prompt_filename,
@@ -169,12 +137,6 @@ def _handle_tool_call(
             state.history.messages.append(Message(role="tool", name=tool.name, tool_call_id=tool_call.id, content=[TextContent(text=err)]))
             return state
 
-        # Early return for finish action (no need for tool execution)
-        if isinstance(action, FinishAction):
-            assert tool.name == FINISH_TOOL.name, "FinishAction must be used with the finish tool"
-            state.agent_finished = True
-            return state
-
         # Execute actions!
         if tool.executor is None:
             raise RuntimeError(f"Tool '{tool.name}' has no executor")
@@ -188,4 +150,8 @@ def _handle_tool_call(
         state.history.messages.append(tool_msg)
         if on_event:
             on_event(observation)
+
+        # Set conversation state
+        if tool.name == FinishTool.name:
+            state.agent_finished = True
         return state

openhands/core/context/manager.py

@@ -0,0 +1,8 @@
+
+class LLMContextManager:
+    """Context manager for messages we send to LLM.
+    
+    
+    """
+    def __init__(self) -> None:
+        pass

openhands/core/tests/tool/test_builtins.py

@@ -0,0 +1,16 @@
+from openhands.core.tool.builtins import BUILT_IN_TOOLS
+
+
+def test_all_tools_property():
+    for tool in BUILT_IN_TOOLS:
+        assert tool.description is not None
+        assert tool.input_schema is not None
+        assert tool.output_schema is not None
+        assert tool.executor is not None
+        assert tool.annotations is not None
+        # Annotations should have specific hints
+        # Builtin tools should have all these properties
+        assert tool.annotations.readOnlyHint
+        assert not tool.annotations.destructiveHint
+        assert tool.annotations.idempotentHint
+        assert not tool.annotations.openWorldHint

openhands/core/tool/__init__.py

@@ -1,5 +1,6 @@
 """OpenHands runtime package."""
 
+from .builtins import BUILT_IN_TOOLS, FinishTool
 from .tool import ActionBase, ObservationBase, Tool, ToolAnnotations, ToolExecutor
 
 
@@ -9,4 +10,6 @@
     "ToolExecutor",
     "ActionBase",
     "ObservationBase",
+    "FinishTool",
+    "BUILT_IN_TOOLS",
 ]

openhands/core/tool/builtins/__init__.py

@@ -0,0 +1,18 @@
+"""Implementing essential tools that doesn't interact with the environment.
+
+These are built in and are *required* for the agent to work.
+
+For tools that require interacting with the environment, add them to `openhands/tools`.
+"""
+
+from .finish import FinishAction, FinishObservation, FinishTool
+
+
+BUILT_IN_TOOLS = [FinishTool]
+
+__all__ = [
+    "BUILT_IN_TOOLS",
+    "FinishTool",
+    "FinishAction",
+    "FinishObservation",
+]

openhands/core/tool/builtins/finish.py

@@ -0,0 +1,47 @@
+from pydantic import Field
+
+from ..tool import ActionBase, ObservationBase, Tool, ToolAnnotations, ToolExecutor
+
+
+class FinishAction(ActionBase):
+    message: str = Field(description="Final message to send to the user.")
+
+class FinishObservation(ObservationBase):
+    message: str = Field(description="Final message sent to the user.")
+
+    @property
+    def agent_observation(self) -> str:
+        return self.message
+
+TOOL_DESCRIPTION = """Signals the completion of the current task or conversation.
+
+Use this tool when:
+- You have successfully completed the user's requested task
+- You cannot proceed further due to technical limitations or missing information
+
+The message should include:
+- A clear summary of actions taken and their results
+- Any next steps for the user
+- Explanation if you're unable to complete the task
+- Any follow-up questions if more information is needed
+"""
+
+class FinishExecutor(ToolExecutor):
+    def __call__(self, action: FinishAction) -> FinishObservation:
+        return FinishObservation(message=action.message)
+
+
+FinishTool = Tool(
+    name="finish",
+    input_schema=FinishAction,
+    output_schema=FinishObservation,
+    description=TOOL_DESCRIPTION,
+    executor=FinishExecutor(),
+    annotations=ToolAnnotations(
+        title="finish",
+        readOnlyHint=True,
+        destructiveHint=False,
+        idempotentHint=True,
+        openWorldHint=False,
+    ),
+)