aiqtoolkit 1.2.0.dev0__py3-none-any.whl → 1.2.0rc2__py3-none-any.whl

aiq/tool/code_execution/code_sandbox.py

@@ -15,11 +15,16 @@
 import abc
 import json
 import logging
+import textwrap
+from typing import Any
 from urllib.parse import urljoin
 
 import requests
+import requests.adapters
 from pydantic import HttpUrl
 
+from aiq.utils.type_utils import override
+
 logger = logging.getLogger(__file__)
 
 
@@ -43,18 +48,18 @@ class Sandbox(abc.ABC):
         *,
         uri: HttpUrl,
     ):
-        self.url = self._get_execute_url(uri)
+        self.url: str = self._get_execute_url(uri)
         session = requests.Session()
         adapter = requests.adapters.HTTPAdapter(pool_maxsize=1500, pool_connections=1500, max_retries=3)
         session.mount('http://', adapter)
         session.mount('https://', adapter)
-        self.http_session = session
+        self.http_session: requests.Session = session
 
-    def _send_request(self, request, timeout):
+    def _send_request(self, request: dict[str, Any], timeout_seconds: float) -> dict[str, str]:
         output = self.http_session.post(
             url=self.url,
             data=json.dumps(request),
-            timeout=timeout,
+            timeout=timeout_seconds,
             headers={"Content-Type": "application/json"},
         )
         # retrying 502 errors
@@ -64,104 +69,180 @@ class Sandbox(abc.ABC):
         return self._parse_request_output(output)
 
     @abc.abstractmethod
-    def _parse_request_output(self, output):
+    def _parse_request_output(self, output: requests.Response) -> dict[str, str]:
         pass
 
     @abc.abstractmethod
-    def _get_execute_url(self, uri):
+    def _get_execute_url(self, uri: HttpUrl) -> str:
         pass
 
     @abc.abstractmethod
-    def _prepare_request(self, generated_code, timeout):
+    def _prepare_request(self, generated_code: str, timeout_seconds: float) -> dict[str, Any]:
         pass
 
     async def execute_code(
         self,
         generated_code: str,
-        timeout: float = 10.0,
+        timeout_seconds: float = 10.0,
         language: str = "python",
         max_output_characters: int = 1000,
-    ) -> tuple[dict, str]:
+    ) -> dict[str, str]:
 
-        generated_code = generated_code.lstrip().rstrip().lstrip("`").rstrip("`")
-        code_to_execute = """
-import traceback
-import json
-import os
-import warnings
-import contextlib
-import io
-warnings.filterwarnings('ignore')
-os.environ['OPENBLAS_NUM_THREADS'] = '16'
-"""
-
-        code_to_execute += f"""
-\ngenerated_code = {repr(generated_code)}\n
-stdout = io.StringIO()
-stderr = io.StringIO()
-
-with contextlib.redirect_stdout(stdout), contextlib.redirect_stderr(stderr):
-    try:
-        exec(generated_code)
-        status = "completed"
-    except Exception:
-        status = "error"
-        stderr.write(traceback.format_exc())
-stdout = stdout.getvalue()
-stderr = stderr.getvalue()
-if len(stdout) > {max_output_characters}:
-    stdout = stdout[:{max_output_characters}] + "<output cut>"
-if len(stderr) > {max_output_characters}:
-    stderr = stderr[:{max_output_characters}] + "<output cut>"
-if stdout:
-    stdout += "\\n"
-if stderr:
-    stderr += "\\n"
-output = {{"process_status": status, "stdout": stdout, "stderr": stderr}}
-print(json.dumps(output))
-"""
-        request = self._prepare_request(code_to_execute, timeout)
+        if language != "python":
+            raise ValueError(f"Language {language} not supported")
+
+        generated_code = generated_code.strip().strip("`")
+        code_to_execute = textwrap.dedent("""
+            import traceback
+            import json
+            import os
+            import warnings
+            import contextlib
+            import io
+            warnings.filterwarnings('ignore')
+            os.environ['OPENBLAS_NUM_THREADS'] = '16'
+        """).strip()
+
+        # Use json.dumps to properly escape the generated_code instead of repr()
+        escaped_code = json.dumps(generated_code)
+        code_to_execute += textwrap.dedent(f"""
+
+            generated_code = {escaped_code}
+
+            stdout = io.StringIO()
+            stderr = io.StringIO()
+
+            with contextlib.redirect_stdout(stdout), contextlib.redirect_stderr(stderr):
+                try:
+                    exec(generated_code)
+                    status = "completed"
+                except Exception:
+                    status = "error"
+                    stderr.write(traceback.format_exc())
+            stdout = stdout.getvalue()
+            stderr = stderr.getvalue()
+            if len(stdout) > {max_output_characters}:
+                stdout = stdout[:{max_output_characters}] + "<output cut>"
+            if len(stderr) > {max_output_characters}:
+                stderr = stderr[:{max_output_characters}] + "<output cut>"
+            if stdout:
+                stdout += "\\n"
+            if stderr:
+                stderr += "\\n"
+            output = {{"process_status": status, "stdout": stdout, "stderr": stderr}}
+            print(json.dumps(output))
+        """).strip()
+        request = self._prepare_request(code_to_execute, timeout_seconds)
         try:
-            output = self._send_request(request, timeout)
+            return self._send_request(request, timeout_seconds)
         except requests.exceptions.Timeout:
-            output = {"process_status": "timeout", "stdout": "", "stderr": "Timed out\n"}
-        return output
+            return {"process_status": "timeout", "stdout": "", "stderr": "Timed out\n"}
 
 
 class LocalSandbox(Sandbox):
     """Locally hosted sandbox."""
 
-    def _get_execute_url(self, uri):
+    def __init__(self, *, uri: HttpUrl):
+        super().__init__(uri=uri)
+
+    @override
+    def _get_execute_url(self, uri: HttpUrl) -> str:
         return urljoin(str(uri), "execute")
 
-    def _parse_request_output(self, output):
+    @override
+    def _parse_request_output(self, output: requests.Response) -> dict[str, str]:
         try:
-            return output.json()
+            output_json = output.json()
+            assert isinstance(output_json, dict)
+            return output_json
         except json.JSONDecodeError as e:
-            logger.exception("Error  parsing output: %s. %s", output.text, e)
-            return {'process_status': 'error', 'stdout': '', 'stderr': 'Unknown error'}
+            logger.exception("Error parsing output: %s. %s", output.text, e)
+            return {'process_status': 'error', 'stdout': '', 'stderr': f'Unknown error: {e} \"{output.text}\"'}
 
-    def _prepare_request(self, generated_code, timeout, language='python', **kwargs):
-        return {
+    @override
+    def _prepare_request(self,
+                         generated_code: str,
+                         timeout_seconds: float,
+                         language: str = "python",
+                         **kwargs) -> dict[str, Any]:
+        request = {
             "generated_code": generated_code,
-            "timeout": timeout,
+            "timeout": timeout_seconds,
             "language": language,
         }
+        return request
+
+    @override
+    async def execute_code(
+        self,
+        generated_code: str,
+        timeout_seconds: float = 10.0,
+        language: str = "python",
+        max_output_characters: int = 1000,
+    ) -> dict[str, str]:
+        """Override execute_code to bypass the wrapper logic and send user code directly to our server."""
+
+        logger.debug("Raw input generated_code: %s", generated_code)
+
+        # The input appears to be a string representation of a dictionary
+        # We need to parse it and extract the actual code
+        try:
+            # Try to evaluate the string as a Python literal (dictionary)
+            import ast
+            parsed_dict = ast.literal_eval(generated_code)
+            if isinstance(parsed_dict, dict) and 'generated_code' in parsed_dict:
+                actual_code = parsed_dict['generated_code']
+                assert isinstance(actual_code, str)
+                logger.debug("Extracted code from dict: %s...", actual_code[:100])
+            else:
+                # If it's not a dict or doesn't have the expected key, use as-is
+                actual_code = generated_code
+                logger.debug("Using code as-is: %s...", actual_code[:100])
+        except (ValueError, SyntaxError):
+            # If parsing fails, use the input as-is
+            actual_code = generated_code
+            logger.debug("Failed to parse, using as-is: %s...", actual_code[:100])
+
+        # Clean the actual code more carefully to avoid removing backticks that are part of Python code
+        # remove all leading/trailing whitespace -- strip()
+        # remove all leading/trailing backticks -- strip("`")
+        # may potentially start with python, so just trim from the front.
+        POTENTIAL_PREFIXES = ["python"]
+        actual_code = actual_code.strip().strip("`")
+        for prefix in POTENTIAL_PREFIXES:
+            if actual_code.startswith(prefix):
+                actual_code = actual_code[len(prefix):]
+                break
+
+        # Send the user's code directly to our server without any wrapper logic
+        # Our server already handles stdout/stderr capture and error handling
+        request = self._prepare_request(actual_code, timeout_seconds, language)
+        try:
+            return self._send_request(request, timeout_seconds)
+        except requests.exceptions.Timeout:
+            return {"process_status": "timeout", "stdout": "", "stderr": "Timed out\n"}
 
 
 class PistonSandbox(Sandbox):
     """Piston sandbox (https://github.com/engineer-man/piston)"""
 
-    def _get_execute_url(self, uri):
+    @override
+    def _get_execute_url(self, uri: HttpUrl) -> str:
         return urljoin(str(uri), "execute")
 
-    def _parse_request_output(self, output):
-        output = output.json()
-        if output['run']['signal'] == "SIGKILL":
-            return {'result': None, 'error_message': 'Unknown error: SIGKILL'}
-        return json.loads(output['run']['output'])
+    @override
+    def _parse_request_output(self, output: requests.Response) -> dict[str, str]:
+        output_json = output.json()
+        assert isinstance(output_json, dict)
+        assert 'run' in output_json
+        run_json = output_json['run']
+        assert isinstance(run_json, dict)
+        if run_json["code"] != 0:
+            return {'process_status': "error", 'stdout': run_json['stdout'], 'stderr': run_json['stderr']}
+        return {'process_status': "completed", 'stdout': run_json['stdout'], 'stderr': run_json['stderr']}
 
-    def _prepare_request(self, generated_code: str, timeout, **kwargs):
+    @override
+    def _prepare_request(self, generated_code: str, timeout_seconds: float, **kwargs) -> dict[str, Any]:
         return {
             "language": "py",
             "version": "3.10.0",
@@ -170,19 +251,17 @@ class PistonSandbox(Sandbox):
             }],
             "stdin": "",
             "args": [],
-            "run_timeout": timeout * 1000.0,  # milliseconds
+            "run_timeout": timeout_seconds * 1000.0,  # milliseconds
             "compile_memory_limit": -1,
             "run_memory_limit": -1,
         }
 
 
-sandboxes = {
-    'local': LocalSandbox,
-    'piston': PistonSandbox,
-}
-
-
 def get_sandbox(sandbox_type: str = "local", **kwargs):
     """A helper function to make it easier to set sandbox through cmd."""
+    sandboxes = {
+        'local': LocalSandbox,
+        'piston': PistonSandbox,
+    }
     sandbox_class = sandboxes[sandbox_type.lower()]
     return sandbox_class(**kwargs)

aiq/tool/code_execution/local_sandbox/.gitignore

@@ -0,0 +1 @@
+persistence_test.*

aiq/tool/code_execution/local_sandbox/local_sandbox_server.py

@@ -12,16 +12,59 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from __future__ import annotations
+
+import contextlib
 import logging
 import multiprocessing
+import os
 import resource
-import sys
+from enum import Enum
 from io import StringIO
 
 from flask import Flask
+from flask import Request
+from flask import Response
 from flask import request
+from pydantic import BaseModel
+from pydantic import Field
 
 app = Flask(__name__)
+logger = logging.getLogger(__name__)
+logger.setLevel(logging.WARNING)
+
+
+class CodeExecutionStatus(str, Enum):
+    """
+    Status of code execution.
+    """
+    COMPLETED = "completed"
+    ERROR = "error"
+    TIMEOUT = "timeout"
+
+
+class CodeExecutionResult(BaseModel):
+    """
+    Result of code execution.
+    """
+    process_status: CodeExecutionStatus = Field(default=CodeExecutionStatus.COMPLETED,
+                                                description="Status of the process")
+    stdout: str = Field(description="Standard output of the process")
+    stderr: str = Field(description="Standard error of the process")
+
+
+class CodeExecutionResponse(Response):
+    """
+    Response class that returns a JSON response with the given status code and result.
+    """
+
+    def __init__(self, status_code: int, result: CodeExecutionResult):
+        super().__init__(status=status_code, mimetype="application/json", response=result.model_dump_json())
+
+    @classmethod
+    def with_error(cls, status_code: int, error_message: str) -> 'CodeExecutionResponse':
+        return cls(status_code,
+                   CodeExecutionResult(process_status=CodeExecutionStatus.ERROR, stdout="", stderr=error_message))
 
 
 @app.after_request
@@ -34,50 +77,122 @@ def add_hsts_header(response):
     return response
 
 
-def execute_python(generated_code, timeout):
+def execute_python(generated_code: str, timeout: float) -> CodeExecutionResult:
+    """
+    Execute Python code in a subprocess.
+
+    Args:
+        generated_code: The code to execute
+        timeout: The timeout for the execution
+
+    Returns:
+        CodeExecutionResult object containing the execution result
+    """
+
     # running in a separate process to ensure any kind of crashes are properly handled
     queue = multiprocessing.Queue()
     process = multiprocessing.Process(target=execute_code_subprocess, args=(generated_code, queue))
+
     process.start()
+    # wait until the process finishes or the timeout expires
     process.join(timeout=timeout)
-
-    if process.is_alive():  # didn't finish successfully
+    if process.exitcode is None:
         process.kill()
-        return {"process_status": "timeout", "stdout": "", "stderr": "Timed out\n"}
+        return CodeExecutionResult(process_status=CodeExecutionStatus.TIMEOUT, stdout="", stderr="Timed out\n")
 
     return queue.get()
 
 
 # need to memory-limit to avoid common errors of allocating too much
 # but this has to be done in a subprocess to not crush server itself
-def execute_code_subprocess(generated_code, queue):
-    limit = 1024 * 1024 * 1024 * 10  # 10gb - somehow with a smaller limit the server dies when numpy is used
-    resource.setrlimit(resource.RLIMIT_AS, (limit, limit))
-    resource.setrlimit(resource.RLIMIT_DATA, (limit, limit))
+def execute_code_subprocess(generated_code: str, queue):
+    """
+    Execute code in a subprocess.
+
+    Args:
+        generated_code: The code to execute
+        queue: The queue to put the result in
+    """
+
+    logger.debug("execute_code_subprocess started, PID: %s", os.getpid())
 
-    # this can be overriden inside generated code, so it's not a guaranteed protection
-    sys.stdout = StringIO()
     try:
-        exec(generated_code, {})  # pylint: disable=W0122
-        queue.put(sys.stdout.getvalue())
+        limit = 1024 * 1024 * 1024 * 10  # 10gb - somehow with a smaller limit the server dies when numpy is used
+        resource.setrlimit(resource.RLIMIT_AS, (limit, limit))
+        resource.setrlimit(resource.RLIMIT_DATA, (limit, limit))
     except Exception as e:
-        print(f"Error: {str(e)}")
-        queue.put({"process_status": "error", "stdout": "", "stderr": str(e) + "\n"})
+        logger.error("Failed to set resource limits, PID: %s, error: %s", os.getpid(), e)
+
+    stdout_capture = StringIO()
+    stderr_capture = StringIO()
+    try:
+        with contextlib.redirect_stdout(stdout_capture), contextlib.redirect_stderr(stderr_capture):
+            exec(generated_code, {})  # pylint: disable=W0122
+        logger.debug("execute_code_subprocess finished, PID: %s", os.getpid())
+        queue.put(CodeExecutionResult(stdout=stdout_capture.getvalue(), stderr=stderr_capture.getvalue()))
+    except Exception as e:
+        import traceback
+        with contextlib.redirect_stderr(stderr_capture):
+            traceback.print_exc()
+        logger.debug("execute_code_subprocess failed, PID: %s, error: %s", os.getpid(), e)
+        queue.put(
+            CodeExecutionResult(process_status=CodeExecutionStatus.ERROR,
+                                stdout=stdout_capture.getvalue(),
+                                stderr=stderr_capture.getvalue()))
+
+
+def do_execute(request: Request) -> CodeExecutionResponse:
+    """
+    Main function to handle execution requests.
+
+    Args:
+        request: Request object containing the execution request
+
+    Returns:
+        CodeExecutionResponse object containing the execution result
+    """
+    try:
+        # Check if request has JSON data
+        if not request.is_json:
+            return CodeExecutionResponse.with_error(400, "Request must be JSON")
+
+        # Get JSON data safely
+        json_data = request.get_json(silent=True)
+
+        if json_data is None:
+            return CodeExecutionResponse.with_error(400, "Invalid JSON data")
+
+        # Check for required fields
+        if 'generated_code' not in json_data:
+            return CodeExecutionResponse.with_error(400, "Missing required field: generated_code")
+
+        if 'timeout' not in json_data:
+            return CodeExecutionResponse.with_error(400, "Missing required field: timeout")
+
+        if 'language' not in json_data:
+            return CodeExecutionResponse.with_error(400, "Missing required field: language")
+
+        generated_code: str | None = json_data.get('generated_code', None)
+        assert generated_code is not None
+        timeout: float | None = json_data.get('timeout', None)
+        assert timeout is not None
+        language: str | None = json_data.get('language', None)
+        assert language is not None
+
+        if language != 'python':
+            return CodeExecutionResponse.with_error(400, "Only python execution is supported")
+
+        return CodeExecutionResponse(200, execute_python(generated_code, timeout))
+
+    except Exception as e:
+        return CodeExecutionResponse.with_error(500, f"Server error: {str(e)}")
 
 
 # Main Flask endpoint to handle execution requests
 @app.route("/execute", methods=["POST"])
 def execute():
-    generated_code = request.json['generated_code']
-    timeout = request.json['timeout']
-    language = request.json.get('language', 'python')
-
-    if language == 'python':
-        return execute_python(generated_code, timeout)
-    return {"process_status": "error", "stdout": "", "stderr": "Only python execution is supported"}
+    return do_execute(request)
 
 
 if __name__ == '__main__':
-    log = logging.getLogger('werkzeug')
-    log.setLevel(logging.WARNING)
     app.run(port=6000)

aiq/tool/code_execution/local_sandbox/sandbox.requirements.txt

@@ -1,4 +1,6 @@
 numpy
 pandas
 scipy
-ipython
+ipython
+plotly
+pydantic

aiq/tool/code_execution/local_sandbox/start_local_sandbox.sh

@@ -14,12 +14,37 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+# Usage: ./start_local_sandbox.sh [SANDBOX_NAME] [OUTPUT_DATA_PATH]
 # NOTE: needs to run from the root of the repo!
 
+DOCKER_COMMAND=${DOCKER_COMMAND:-"docker"}
 SANDBOX_NAME=${1:-'local-sandbox'}
 NUM_THREADS=10
 
+# Get the output_data directory path for mounting
+# Priority: command line argument > environment variable > default path (current directory)
+OUTPUT_DATA_PATH=${2:-${OUTPUT_DATA_PATH:-$(pwd)}}
 
-docker build --tag=${SANDBOX_NAME} --build-arg="UWSGI_PROCESSES=$((${NUM_THREADS} * 10))" --build-arg="UWSGI_CHEAPER=${NUM_THREADS}" -f Dockerfile.sandbox .
+echo "Starting sandbox with container name: ${SANDBOX_NAME}"
+echo "Mounting output_data directory: ${OUTPUT_DATA_PATH}"
 
-docker run --network=host --rm --name=local-sandbox ${SANDBOX_NAME}
+# Verify the path exists before mounting, create if it doesn't
+if [ ! -d "${OUTPUT_DATA_PATH}" ]; then
+    echo "Output data directory does not exist, creating: ${OUTPUT_DATA_PATH}"
+    mkdir -p "${OUTPUT_DATA_PATH}"
+fi
+
+# Check if the Docker image already exists
+if ! ${DOCKER_COMMAND} images ${SANDBOX_NAME} | grep -q "${SANDBOX_NAME}"; then
+    echo "Docker image not found locally. Building ${SANDBOX_NAME}..."
+    ${DOCKER_COMMAND} build --tag=${SANDBOX_NAME} --build-arg="UWSGI_PROCESSES=$((${NUM_THREADS} * 10))" --build-arg="UWSGI_CHEAPER=${NUM_THREADS}" -f Dockerfile.sandbox .
+else
+    echo "Using existing Docker image: ${SANDBOX_NAME}"
+fi
+
+# Mount the output_data directory directly so files created in container appear in the local directory
+${DOCKER_COMMAND} run --rm --name=local-sandbox \
+  --network=host \
+  -v "${OUTPUT_DATA_PATH}:/workspace" \
+  -w /workspace \
+  ${SANDBOX_NAME}

aiq/tool/code_execution/register.py

@@ -46,7 +46,11 @@ async def code_execution_tool(config: CodeExecutionToolConfig, builder: Builder)
     class CodeExecutionInputSchema(BaseModel):
         generated_code: str = Field(description="String containing the code to be executed")
 
-    sandbox = get_sandbox(sandbox_type=config.sandbox_type, uri=config.uri)
+    # Create sandbox without working_directory
+    sandbox_kwargs = {"uri": config.uri}
+
+    sandbox = get_sandbox(sandbox_type=config.sandbox_type, **sandbox_kwargs)
+    logger.info(f"[DEBUG] Created sandbox of type: {config.sandbox_type}")
 
     async def _execute_code(generated_code: str) -> dict:
         logger.info("Executing code in the sandbox at %s", config.uri)
@@ -54,12 +58,12 @@ async def code_execution_tool(config: CodeExecutionToolConfig, builder: Builder)
             output = await sandbox.execute_code(
                 generated_code=generated_code,
                 language="python",
-                timeout=config.timeout,
+                timeout_seconds=config.timeout,
                 max_output_characters=config.max_output_characters,
             )
         except Exception as e:
             logger.exception("Error when executing code in the sandbox, %s", e)
-            return {"process_status": "error", "stdout": "", "stderr": e}
+            return {"process_status": "error", "stdout": "", "stderr": str(e)}
         return output
 
     yield FunctionInfo.from_fn(