Add swe-bench test - Not to run in regular PRs

requirements.txt

@@ -1,10 +1,66 @@
-openai==1.107.3
+aiohappyeyeballs==2.6.1
+aiohttp==3.12.15
+aiosignal==1.4.0
+annotated-types==0.7.0
 anthropic==0.75.0
-python-dotenv==1.1.1
+anyio==4.10.0
+attrs==25.3.0
+bashlex==0.18
+certifi==2025.8.3
+charset-normalizer==3.4.3
+click==8.3.0
+coverage==7.11.3
+datasets==4.5.0
+dill==0.4.0
+distro==1.9.0
 docker==7.1.0
+docstring_parser==0.17.0
 fastapi==0.116.1
-uvicorn==0.35.0
+filelock==3.20.3
+frozenlist==1.7.0
+fsspec==2025.10.0
+h11==0.16.0
+hf-xet==1.2.0
+httpcore==1.0.9
+httpx==0.28.1
+huggingface_hub==1.3.2
+idna==3.10
+iniconfig==2.1.0
+jiter==0.11.0
+markdown-it-py==4.0.0
+mdurl==0.1.2
+multidict==6.6.4
+multiprocess==0.70.18
+numpy==1.26.4
+openai==1.107.3
+packaging==25.0
+pandas==3.0.0
+pexpect==4.9.0
+pluggy==1.6.0
+propcache==0.3.2
+ptyprocess==0.7.0
+pyarrow==23.0.0
 pydantic==2.11.9
+pydantic_core==2.33.2
+Pygments==2.19.2
+pytest==8.4.2
+pytest-cov==7.0.0
+python-dateutil==2.9.0.post0
+python-dotenv==1.1.1
+python-multipart==0.0.20
+PyYAML==6.0.2
+requests==2.32.5
+rich==14.1.0
+shellingham==1.5.4
+six==1.17.0
+sniffio==1.3.1
+starlette==0.47.3
 swe-rex==1.4.0
-aiohttp==3.13.3
-pyyaml==6.0.2
+tqdm==4.67.1
+typer-slim==0.21.1
+typing-inspection==0.4.1
+typing_extensions==4.15.0
+urllib3==2.5.0
+uvicorn==0.35.0
+xxhash==3.6.0
+yarl==1.20.1

src/microbots/MicroBot.py

@@ -26,7 +26,8 @@
 You are a helpful agent well versed in software development and debugging.
 
 You will be provided with a coding or debugging task to complete inside a sandboxed shell environment.
-There is a shell session open for you. You will be provided with a task and you should achieve it using the shell commands.
+There is a shell session open for you.
+You will be provided with a task and you should achieve it using the shell commands.
 All your response must be in the following json format:
 {llm_output_format_str}
 The properties ( task_done, thoughts, command ) are mandatory on each response.
@@ -45,22 +46,6 @@
 1. If a command fails, analyze the error message and provide an alternative command in your next response. Same command will not pass again.
 2. Avoid using recursive commands like 'ls -R', 'rm -rf', 'tree', or 'find' without depth limits as they can produce excessive output or be destructive.
 3. You cannot run any interactive commands like vim, nano, etc.
-
-# TOOLS
-You have following special tools.
-
-    1. summarize_context: Use this tool if your input has too many irrelevant conversation turns.
-         You can use this tool to rewrite your own context in a concise manner focusing on important points only. For example, if you have a failed command output which you've solved in later steps deep down in the conversation, that is not required to be in the context. You can summarize the context to remove such irrelevant information. This tool will not update the system prompt. So, you can ignore details in the system prompt while summarizing.
-
-         Usage:
-            summarize_context <no_of_recent_turns_to_keep> "<your summary of the context>"
-
-            <no_of_recent_turns_to_keep> : Number of recent conversation turns to keep as is without summarizing. 1 means last user-assistant pair will be kept as is.
-            "<your summary of the context>" : Your summarized context in double quotes. The summary can be empty if you finished a sub-task and want to remove previous context.
-
-        Important Notes:
-            - The summarize tool call step will not be added to your history.
-            - Try to be very precise and concise in your summary.
 """
 
 
@@ -212,7 +197,7 @@ def run(
                     f" 💭  LLM thoughts: {LogTextColor.OKCYAN}{llm_response.thoughts}{LogTextColor.ENDC}",
                 )
             logger.info(
-                f" ➡️  LLM tool call : {LogTextColor.OKBLUE}{llm_response.command}{LogTextColor.ENDC}",
+                f" ➡️  LLM tool call : {LogTextColor.OKBLUE}{json.dumps(llm_response.command)}{LogTextColor.ENDC}",
             )
             # increment iteration count
             iteration_count += 1
@@ -233,19 +218,6 @@ def run(
                 return_value.error = f"Timeout of {timeout} seconds reached"
                 return return_value
 
-            # Handle context summarization
-            if llm_response.command.startswith("summarize_context"):
-                parsed_args = self._parse_summarize_context_command(llm_response.command)
-                if parsed_args is None:
-                    # Invalid syntax - ask LLM to correct it
-                    error_msg = self._get_summarize_context_syntax_error(llm_response.command)
-                    llm_response = self.llm.ask(error_msg)
-                    continue
-                last_n_messages, summary = parsed_args
-                last_msg = self.llm.summarize_context(last_n_messages=last_n_messages, summary=summary)
-                llm_response = self.llm.ask(last_msg["content"])
-                continue
-
             # Validate command for dangerous operations
             is_safe, explanation = self._is_safe_command(llm_response.command)
             if not is_safe:
@@ -269,7 +241,7 @@ def run(
                     # HACK: anthropic-text-editor tool extra formats the output
                     try:
                         output_json = json.loads(llm_command_output.stdout)
-                        if isinstance(output_json, dict) and "content" in output_json:
+                        if "content" in output_json:
                             output_text = pformat(output_json["content"])
                     except json.JSONDecodeError:
                         pass
@@ -356,62 +328,6 @@ def _validate_folder_to_mount(self, folder_to_mount: Mount):
                 "Only MOUNT mount type is supported for folder_to_mount"
             )
 
-    def _parse_summarize_context_command(self, command: str) -> tuple[int, str] | None:
-        """
-        Parse the summarize_context command and extract arguments.
-
-        Expected format: summarize_context <n> "<summary>"
-        Where <n> is an integer and <summary> is a quoted string.
-
-        Returns:
-            tuple[int, str]: (last_n_messages, summary) if valid
-            None: if invalid syntax
-        """
-        import shlex
-        try:
-            # Use shlex to properly handle quoted strings
-            parts = shlex.split(command)
-            if len(parts) < 2 or len(parts) > 3:
-                return None
-
-            # First part should be 'summarize_context'
-            if parts[0] != "summarize_context":
-                return None
-
-            # Second part should be an integer
-            try:
-                last_n_messages = int(parts[1])
-            except ValueError:
-                return None
-
-            # Third part is the summary (optional, defaults to empty string)
-            summary = parts[2] if len(parts) > 2 else ""
-
-            return (last_n_messages, summary)
-
-        except ValueError:
-            # shlex.split can raise ValueError for malformed strings
-            return None
-
-    def _get_summarize_context_syntax_error(self, command: str) -> str:
-        """
-        Generate an error message for invalid summarize_context syntax.
-
-        Returns a detailed error message guiding the LLM to use correct syntax.
-        """
-        return f"""COMMAND_ERROR: Invalid summarize_context syntax.
-Your command: {command}
-
-Correct usage:
-    summarize_context <no_of_recent_turns_to_keep> "<your summary of the context>"
-
-Examples:
-    summarize_context 2 "Summary of previous work: explored files and found the bug"
-    summarize_context 0 ""
-    summarize_context 5 "Completed file exploration, ready to make changes"
-
-Please send the command again with correct syntax."""
-
     def _get_dangerous_command_explanation(self, command: str) -> Optional[str]:
         """Provides detailed explanation for why a command is dangerous and suggests alternatives.
 

src/microbots/llm/anthropic_api.py

@@ -24,7 +24,12 @@ def __init__(self, system_prompt, deployment_name=deployment_name, max_retries=3
             base_url=endpoint
         )
         self.deployment_name = deployment_name
-        super().__init__(system_prompt=system_prompt, max_retries=max_retries)
+        self.system_prompt = system_prompt
+        self.messages = []
+
+        # Set these values here. This logic will be handled in the parent class.
+        self.max_retries = max_retries
+        self.retries = 0
 
     def ask(self, message) -> LLMAskResponse:
         self.retries = 0  # reset retries for each ask. Handled in parent class.

src/microbots/llm/llm.py

@@ -21,17 +21,6 @@ class LLMAskResponse:
     command: str = ""
 
 class LLMInterface(ABC):
-    def __init__(self, system_prompt: str, max_retries: int = 3):
-        self.system_prompt = system_prompt
-        self.max_retries = max_retries
-        self.retries = 0
-        self.messages = [
-            {
-                "role": "system",
-                "content": self.system_prompt,
-            }
-        ]
-
     @abstractmethod
     def ask(self, message: str) -> LLMAskResponse:
         pass
@@ -55,7 +44,7 @@ def _validate_llm_response(self, response: str) -> tuple[bool, LLMAskResponse]:
             return False, None
 
         if all(key in response_dict for key in LLMAskResponse.__annotations__.keys()):
-            logger.debug("The llm response is %s ", response_dict)
+            logger.info("The llm response is %s ", response_dict)
 
             if response_dict.get("task_done") not in [True, False]:
                 self.retries += 1
@@ -95,45 +84,3 @@ def _validate_llm_response(self, response: str) -> tuple[bool, LLMAskResponse]:
             logger.warning("LLM response is missing required fields. Retrying... (%d/%d)", self.retries, self.max_retries)
             self.messages.append({"role": "user", "content": "LLM_RES_ERROR: LLM response is missing required fields. Please respond in the correct JSON format.\n" + llm_output_format_str})
             return False, None
-
-    def summarize_context(self, last_n_messages: int = 10, summary: str="") -> dict:
-        """
-        It is a helper function for the LLM to summarize its own context.
-        Leave the last N messages and add the summary between system prompt and the last N messages.
-
-        summary can be empty. If empty, empty summary will be added.
-        """
-        logger.debug("Messages : %s", self.messages)
-        # Keep the system prompt
-        msg0 = self.messages[0]["content"]
-        # Pop the last message which asked for summarization
-        self.messages.pop()
-        # Get the last N conversations (user + assistant)
-        # If there are not enough messages, take all except system prompt
-        if (len(self.messages) > (last_n_messages*2 - 1)):
-            logger.debug("Summarizing last %d messages", last_n_messages)
-            recent_messages = self.messages[-(last_n_messages*2 - 1):]
-        else:
-            logger.debug("Not enough messages to summarize, taking all except system prompt")
-            recent_messages = self.messages[1:]
-        logger.debug("Recent messages that will not be summarized: %s", recent_messages)
-
-        # Update system prompt if it already has a summary
-        # summary will be between __summary__ and __end_summary__
-        if "__summary__" in msg0:
-            system_prompt = msg0.split("__summary__")[0]
-            old_summary = msg0.split("__end_summary__")[0].split("__summary__")[1]
-            logger.debug("Old summary found: %s", old_summary)
-            combined_summary = old_summary + "\n" + summary
-        else:
-            system_prompt = msg0
-            combined_summary = summary
-
-        new_system_prompt = f"{system_prompt}\n__summary__\n{combined_summary}\n__end_summary__"
-
-        # Append without previous user message
-        self.messages = [{"role": "system", "content": new_system_prompt}] + recent_messages[:-1]
-        logger.debug("Context summarized. New system prompt: %s", new_system_prompt)
-
-        logger.debug("Last message before summarization: %s", recent_messages[-1])
-        return recent_messages[-1]  # return the last user message that given before summarization

src/microbots/llm/ollama_local.py

@@ -46,11 +46,15 @@ class OllamaLocal(LLMInterface):
     def __init__(self, system_prompt, model_name=None, model_port=None, max_retries=3):
         self.model_name = model_name or os.environ.get("LOCAL_MODEL_NAME")
         self.model_port = model_port or os.environ.get("LOCAL_MODEL_PORT")
+        self.system_prompt = system_prompt
+        self.messages = [{"role": "system", "content": system_prompt}]
 
         if not self.model_name or not self.model_port:
             raise ValueError("LOCAL_MODEL_NAME and LOCAL_MODEL_PORT environment variables must be set or passed as arguments to OllamaLocal.")
 
-        super().__init__(system_prompt=system_prompt, max_retries=max_retries)
+        # Set these values here. This logic will be handled in the parent class.
+        self.max_retries = max_retries
+        self.retries = 0
 
     def ask(self, message) -> LLMAskResponse:
         self.retries = 0  # reset retries for each ask. Handled in parent class.

src/microbots/llm/openai_api.py

@@ -16,9 +16,14 @@
 class OpenAIApi(LLMInterface):
 
     def __init__(self, system_prompt, deployment_name=deployment_name, max_retries=3):
-        self.deployment_name = deployment_name
         self.ai_client = OpenAI(base_url=f"{endpoint}", api_key=api_key)
-        super().__init__(system_prompt=system_prompt, max_retries=max_retries)
+        self.deployment_name = deployment_name
+        self.system_prompt = system_prompt
+        self.messages = [{"role": "system", "content": system_prompt}]
+
+        # Set these values here. This logic will be handled in the parent class.
+        self.max_retries = max_retries
+        self.retries = 0
 
     def ask(self, message) -> LLMAskResponse:
         self.retries = 0 # reset retries for each ask. Handled in parent class.