update: swe

metagpt/roles/di/role_zero.py

@@ -4,7 +4,7 @@ import inspect
 import json
 import re
 import traceback
-from typing import Callable, Dict, List, Literal, Tuple, Union
+from typing import Callable, Dict, List, Literal, Tuple
 
 from pydantic import model_validator
 
@@ -167,7 +167,7 @@ class RoleZero(Role):
         if self.use_fixed_sop:
             return await super()._act()
 
-        commands, ok = await self._get_commands()
+        commands, ok = await self._parse_commands()
         if not ok:
             error_msg = commands
             return error_msg
@@ -202,7 +202,7 @@ class RoleZero(Role):
             actions_taken += 1
         return rsp  # return output from the last action
 
-    async def _get_commands(self) -> Tuple[Union[UserMessage, List[Dict]], bool]:
+    async def _parse_commands(self) -> Tuple[List[Dict], bool]:
         """Retrieves commands from the Large Language Model (LLM).
 
         This function attempts to retrieve a list of commands from the LLM by
@@ -211,7 +211,6 @@ class RoleZero(Role):
 
         Returns:
             A tuple containing:
-                - A `UserMessage` object or dict representing the commands.
                 - A boolean flag indicating success (True) or failure (False).
         """
         try:

metagpt/roles/di/swe.py

@@ -22,6 +22,7 @@ class SWE(RoleZero):
     _system_msg: str = SWE_AGENT_SYSTEM_TEMPLATE
     system_msg: list[str] = [_system_msg.format(WINDOW=_bash_window_size)]
     _instruction: str = NEXT_STEP_TEMPLATE
+    # tools: list[str] = ["Bash", "Browser"]
     tools: list[str] = ["Bash"]
     terminal: Bash = Field(default_factory=Bash, exclude=True)
     output_diff: str = ""
@@ -35,11 +36,23 @@ class SWE(RoleZero):
         return res
 
     def _set_system_msg(self):
+        """
+        Sets the system message for the SWE agent.
+
+        Sets the `_bash_window_size` from the environment variable `WINDOW` if it exists.
+        Formats the `_system_msg` template with the current `_bash_window_size`.
+        """
         if os.getenv("WINDOW"):
             self._bash_window_size = int(os.getenv("WINDOW"))
         self.system_msg = [self._system_msg.format(WINDOW=self._bash_window_size)]
 
     def _format_instruction(self):
+        """
+        Formats the instruction message for the SWE agent.
+
+        Runs the "state" command in the terminal, parses its output as JSON,
+        and uses it to format the `_instruction` template.
+        """
         state_output = self.terminal.run("state")
         bash_state = json.loads(state_output)
 
@@ -50,7 +63,15 @@ class SWE(RoleZero):
         return self.instruction
 
     async def _handle_action(self):
-        commands, ok = await self._get_commands()
+        """
+        Handles actions based on parsed commands.
+
+        Parses commands, checks for a "submit" action, and generates a patch using `git diff`.
+        Stores the cleaned patch in `output_diff`. Logs any exceptions.
+
+        This function is specifically added for SWE bench evaluation.
+        """
+        commands, ok = await self._parse_commands()
         if not ok:
             return
         for cmd in commands: