|
| 1 | +from abc import ABC, abstractmethod |
| 2 | +from dataclasses import dataclass |
| 3 | +import json |
| 4 | +import subprocess |
| 5 | +import tempfile |
| 6 | +from typing import Optional, Dict, Any, List |
| 7 | + |
| 8 | +from openai import OpenAI |
| 9 | + |
| 10 | + |
| 11 | +@dataclass |
| 12 | +class ExecutionResult: |
| 13 | + """Result of a command execution.""" |
| 14 | + |
| 15 | + output: str |
| 16 | + error: Optional[str] = None |
| 17 | + return_code: int = 0 |
| 18 | + |
| 19 | + |
| 20 | +class Executor(ABC): |
| 21 | + |
| 22 | + @abstractmethod |
| 23 | + def execute(self, command: str, timeout: int = 30) -> ExecutionResult: |
| 24 | + """Execute a command and return the result. |
| 25 | +
|
| 26 | + Args: |
| 27 | + command: The command to execute |
| 28 | + timeout: Timeout in seconds (default: 30) |
| 29 | +
|
| 30 | + Returns: |
| 31 | + ExecutionResult containing the command output and status |
| 32 | + """ |
| 33 | + pass |
| 34 | + |
| 35 | + |
| 36 | +class GuixExecutor(Executor): |
| 37 | + """Guix-based executor that runs commands in a sandboxed Guix shell environment.""" |
| 38 | + |
| 39 | + def __init__(self, working_dir: str, manifest_file: Optional[str] = None): |
| 40 | + """Initialize the Guix executor. |
| 41 | +
|
| 42 | + Args: |
| 43 | + working_dir: Working directory of the execution |
| 44 | + manifest_file: Path to a Guix manifest file specifying packages |
| 45 | + """ |
| 46 | + self.working_dir = working_dir |
| 47 | + self.manifest_file = manifest_file |
| 48 | + self.current_env = "" |
| 49 | + |
| 50 | + def execute( |
| 51 | + self, |
| 52 | + command: str, |
| 53 | + timeout: int = 30, |
| 54 | + ) -> ExecutionResult: |
| 55 | + """Execute a command in a sandboxed Guix shell.""" |
| 56 | + |
| 57 | + guix_cmd = ["guix", "shell"] |
| 58 | + |
| 59 | + if self.manifest_file: |
| 60 | + guix_cmd.extend(["-m", self.manifest_file]) |
| 61 | + |
| 62 | + with tempfile.NamedTemporaryFile(mode="w", suffix="_env.sh", delete=False) as env_file: |
| 63 | + env_file.write(self.current_env) |
| 64 | + |
| 65 | + with tempfile.NamedTemporaryFile(mode="w", suffix="_script.sh", delete=False) as script_file: |
| 66 | + script_file.write(f"source {env_file.name}\n") |
| 67 | + script_file.write("cd $PWD\n") |
| 68 | + script_file.write(f"{command}\n") |
| 69 | + script_file.write(f"export -p > {env_file.name}\n") |
| 70 | + |
| 71 | + # Add a very lightweight sandbox using https://github.com/containers/bubblewrap. |
| 72 | + # Originally we were using the guix shell --container sandbox for this, but there |
| 73 | + # are environments where that does not work (e.g. mounting the /proc filesystem |
| 74 | + # can fail in a GPU container). We might want to revisit this. |
| 75 | + guix_cmd.extend( |
| 76 | + # fmt: off |
| 77 | + [ |
| 78 | + "--", |
| 79 | + "bwrap", |
| 80 | + "--ro-bind", "/bin", "/bin", |
| 81 | + "--ro-bind", "/gnu", "/gnu", |
| 82 | + "--proc", "/proc", |
| 83 | + "--dev", "/dev", |
| 84 | + "--tmpfs", "/tmp", |
| 85 | + "--new-session", |
| 86 | + "--ro-bind", script_file.name, script_file.name, |
| 87 | + "--bind", env_file.name, env_file.name, |
| 88 | + "--ro-bind", "/etc/resolv.conf", "/etc/resolv.conf", |
| 89 | + "--bind", self.working_dir, "/home/skyrl", |
| 90 | + "--setenv", "HOME", "/home/skyrl/", |
| 91 | + "sh", |
| 92 | + script_file.name, |
| 93 | + ] |
| 94 | + # fmt: on |
| 95 | + ) |
| 96 | + |
| 97 | + try: |
| 98 | + result = subprocess.run( |
| 99 | + guix_cmd, |
| 100 | + shell=False, |
| 101 | + capture_output=True, |
| 102 | + text=True, |
| 103 | + timeout=timeout, |
| 104 | + cwd=self.working_dir, |
| 105 | + ) |
| 106 | + except Exception as e: |
| 107 | + return ExecutionResult( |
| 108 | + output="", |
| 109 | + error=f"Execution failed: {str(e)}", |
| 110 | + return_code=-1, |
| 111 | + ) |
| 112 | + |
| 113 | + with open(env_file.name, "r") as f: |
| 114 | + self.current_env = f.read() |
| 115 | + |
| 116 | + return ExecutionResult( |
| 117 | + output=result.stdout or "", |
| 118 | + error=result.stderr if result.stderr else None, |
| 119 | + return_code=result.returncode, |
| 120 | + ) |
| 121 | + |
| 122 | + |
| 123 | +@dataclass |
| 124 | +class ToolResult: |
| 125 | + """Result from executing a tool""" |
| 126 | + |
| 127 | + success: bool |
| 128 | + output: str |
| 129 | + error: Optional[str] = None |
| 130 | + |
| 131 | + |
| 132 | +class Tool(ABC): |
| 133 | + """Base class for all tools""" |
| 134 | + |
| 135 | + @abstractmethod |
| 136 | + def name(self) -> str: |
| 137 | + pass |
| 138 | + |
| 139 | + @abstractmethod |
| 140 | + def description(self) -> str: |
| 141 | + pass |
| 142 | + |
| 143 | + @abstractmethod |
| 144 | + def parameters(self) -> Dict[str, Any]: |
| 145 | + pass |
| 146 | + |
| 147 | + @abstractmethod |
| 148 | + def execute(self, **kwargs) -> ToolResult: |
| 149 | + pass |
| 150 | + |
| 151 | + |
| 152 | +class ShellCommandTool(Tool): |
| 153 | + """Tool for executing shell commands""" |
| 154 | + |
| 155 | + def __init__(self, executor: Executor): |
| 156 | + """Initialize the shell command tool with an executor. |
| 157 | +
|
| 158 | + Args: |
| 159 | + executor: The executor to use for running commands. |
| 160 | + """ |
| 161 | + self.executor = executor |
| 162 | + |
| 163 | + def name(self) -> str: |
| 164 | + return "execute_bash" |
| 165 | + |
| 166 | + def description(self) -> str: |
| 167 | + return "Execute a shell command and return the output" |
| 168 | + |
| 169 | + def parameters(self) -> Dict[str, Any]: |
| 170 | + return { |
| 171 | + "type": "object", |
| 172 | + "properties": {"command": {"type": "string", "description": "The shell command to execute"}}, |
| 173 | + "required": ["command"], |
| 174 | + } |
| 175 | + |
| 176 | + def execute(self, command: str, timeout: int = 30) -> ToolResult: |
| 177 | + """Execute a shell command using the configured executor.""" |
| 178 | + execution_result = self.executor.execute(command, timeout=timeout) |
| 179 | + |
| 180 | + return ToolResult( |
| 181 | + success=execution_result.return_code == 0, output=execution_result.output, error=execution_result.error |
| 182 | + ) |
| 183 | + |
| 184 | + |
| 185 | +class SimpleCoder: |
| 186 | + |
| 187 | + def __init__(self, api_key: str, model: str, executor: Executor): |
| 188 | + self.client = OpenAI(api_key=api_key) |
| 189 | + self.model = model |
| 190 | + self.executor = executor |
| 191 | + self.tools = { |
| 192 | + tool.name(): tool |
| 193 | + for tool in [ |
| 194 | + ShellCommandTool(executor=self.executor), |
| 195 | + ] |
| 196 | + } |
| 197 | + self.conversation_history = [] |
| 198 | + |
| 199 | + def _get_tool_definitions(self) -> List[Dict[str, Any]]: |
| 200 | + """Get OpenAI function definitions for all tools""" |
| 201 | + return [ |
| 202 | + { |
| 203 | + "type": "function", |
| 204 | + "function": {"name": tool.name(), "description": tool.description(), "parameters": tool.parameters()}, |
| 205 | + } |
| 206 | + for tool in self.tools.values() |
| 207 | + ] |
| 208 | + |
| 209 | + def _execute_tool(self, tool_name: str, arguments: Dict[str, Any]) -> ToolResult: |
| 210 | + """Execute a tool with given arguments""" |
| 211 | + if tool_name not in self.tools: |
| 212 | + return ToolResult(success=False, output="", error=f"Unknown tool: {tool_name}") |
| 213 | + |
| 214 | + tool = self.tools[tool_name] |
| 215 | + return tool.execute(**arguments) |
| 216 | + |
| 217 | + def run(self, task: str, max_iterations: int = 30): |
| 218 | + |
| 219 | + self.conversation_history = [ |
| 220 | + { |
| 221 | + "role": "system", |
| 222 | + "content": """You are a Software Engineering Agent. You can: |
| 223 | +1. Execute shell commands using execute_shell |
| 224 | +2. Read, write, or append to files using edit_file |
| 225 | +
|
| 226 | +Break down complex tasks into steps and use the appropriate tools to complete them. |
| 227 | +Always check the results of your actions and adapt your approach if needed.""", |
| 228 | + }, |
| 229 | + {"role": "user", "content": task}, |
| 230 | + ] |
| 231 | + |
| 232 | + for i in range(max_iterations): |
| 233 | + response = self.client.chat.completions.create( |
| 234 | + model=self.model, |
| 235 | + messages=self.conversation_history, |
| 236 | + tools=self._get_tool_definitions(), |
| 237 | + tool_choice="auto", |
| 238 | + ) |
| 239 | + assistant_message = response.choices[0].message |
| 240 | + self.conversation_history.append(assistant_message.model_dump()) |
| 241 | + |
| 242 | + # Check if the assistant wants to use tools |
| 243 | + if assistant_message.tool_calls: |
| 244 | + # Execute each tool call |
| 245 | + for tool_call in assistant_message.tool_calls: |
| 246 | + function_name = tool_call.function.name |
| 247 | + arguments = json.loads(tool_call.function.arguments) |
| 248 | + |
| 249 | + print(f"\n🔧 Executing {function_name} with args: {arguments}") |
| 250 | + |
| 251 | + # Execute the tool |
| 252 | + result = self._execute_tool(function_name, arguments) |
| 253 | + |
| 254 | + # Add tool result to conversation |
| 255 | + tool_message = { |
| 256 | + "role": "tool", |
| 257 | + "tool_call_id": tool_call.id, |
| 258 | + "content": json.dumps( |
| 259 | + {"success": result.success, "output": result.output, "error": result.error} |
| 260 | + ), |
| 261 | + } |
| 262 | + self.conversation_history.append(tool_message) |
| 263 | + |
| 264 | + print(f"✅ Result: {result.output}..." if result.success else f"❌ Error: {result.error}") |
| 265 | + else: |
| 266 | + print(f"\n🤖 Agent: {assistant_message.content}") |
| 267 | + return |
| 268 | + |
| 269 | + |
| 270 | +if __name__ == "__main__": |
| 271 | + import os |
| 272 | + import simplecoder |
| 273 | + |
| 274 | + manifest = os.path.abspath("manifest.scm") |
| 275 | + working_dir = os.path.abspath("test-repo") |
| 276 | + executor = simplecoder.GuixExecutor(working_dir, manifest) |
| 277 | + |
| 278 | + coder = simplecoder.SimpleCoder(os.environ["OPENAI_API_KEY"], "o4-mini", executor) |
| 279 | + task = """ |
| 280 | + I'm running missing_colon.py as follows: |
| 281 | +
|
| 282 | +division(23, 0) |
| 283 | +but I get the following error: |
| 284 | +
|
| 285 | + File "/Users/fuchur/Documents/24/git_sync/swe-agent-test-repo/tests/./missing_colon.py", line 4 |
| 286 | + def division(a: float, b: float) -> float |
| 287 | + ^ |
| 288 | +SyntaxError: invalid syntax |
| 289 | +""" |
| 290 | + coder.run(task) |
0 commit comments