Add Piston sandbox plugin

.github/CODEOWNERS

@@ -9,4 +9,5 @@
 # Plugins
 /plugins/data-designer-github/ @NVIDIA-NeMo/data_designer_reviewers @eric-tramel
 /plugins/data-designer-retrieval-sdg/ @NVIDIA-NeMo/data_designer_reviewers @shan-nvidia @oliverholworthy
+/plugins/data-designer-sandbox-piston/ @NVIDIA-NeMo/data_designer_reviewers @eric-tramel
 /plugins/data-designer-template/ @NVIDIA-NeMo/data_designer_reviewers

docs/plugins/data-designer-sandbox-piston/index.md

@@ -0,0 +1,140 @@
+# data-designer-sandbox-piston
+
+`data-designer-sandbox-piston` adds Piston-backed code execution to Data
+Designer. It provides a `code-sandbox` column type for batch workflows and a
+stdio MCP server that exposes the same Piston endpoint as a `run_code` tool for
+tool-calling LLM columns.
+
+The plugin is deployment-neutral: point it at any reachable Piston API URL. That
+URL can be a local Docker container on macOS or Linux, a service running beside a
+Data Designer worker, or a remote endpoint behind your own proxy.
+
+## Installation
+
+```bash
+uv add data-designer data-designer-sandbox-piston
+```
+
+## Column type
+
+Use the `code-sandbox` column type when a dataset already contains source code
+that should be executed by Piston.
+
+| Field | Required | Description |
+| --- | --- | --- |
+| `name` | Yes | Output column name. Each value is a dictionary with execution results. |
+| `target_column` | Yes | Existing column containing source code. |
+| `language` | Yes | Piston runtime language, such as `python` or `gcc`. |
+| `version` | No | Piston runtime version selector. Defaults to `*`. Required when `python_packages` is non-empty. |
+| `python_packages` | No | Optional Python package requirements for a prebuilt custom Python runtime. Only valid with `language="python"`; the deployment must provide the matching runtime before execution. |
+| `stdin` | No | Text passed to standard input. Defaults to an empty string. |
+| `args` | No | Command-line arguments passed to the program. Defaults to an empty list. |
+| `compile_timeout` | No | Compile wall-time limit in milliseconds. Defaults to `10000`. |
+| `run_timeout` | No | Run wall-time limit in milliseconds. Defaults to `3000`, matching stock Piston's default run limit. |
+| `compile_cpu_time` | No | Compile CPU-time limit in milliseconds. Defaults to `3000`. |
+| `run_cpu_time` | No | Run CPU-time limit in milliseconds. Defaults to `3000`. |
+| `sandbox_url` | Yes | HTTP or HTTPS Piston API base URL, such as `http://localhost:2000`. |
+
+## Output shape
+
+The output column contains a dictionary per row:
+
+```python
+{
+    "stdout": "42\n",
+    "stderr": "",
+    "output": "42\n",
+    "exit_code": 0,
+    "signal": None,
+    "message": None,
+    "status": None,
+    "cpu_time": 12.5,
+    "wall_time": 15.2,
+    "memory": 8192,
+}
+```
+
+Empty or missing source code returns `exit_code=-2`. Sandbox API failures return
+`exit_code=-1` with the final error in `stderr` and `message`.
+
+## Python Packages
+
+`python_packages` is declarative metadata. The plugin sends `language` and
+`version` to Piston; it does not install packages or build runtimes during
+generation. If you set a non-empty `python_packages` list, also set `version` to
+the exact custom Python runtime version that your deployment has already built
+and installed in Piston.
+
+## Example
+
+```python
+import pandas as pd
+from data_designer.config.config_builder import DataDesignerConfigBuilder
+from data_designer.config.seed_source_dataframe import DataFrameSeedSource
+
+builder = DataDesignerConfigBuilder()
+builder.with_seed_dataset(
+    DataFrameSeedSource(df=pd.DataFrame({"code": ["print(6 * 7)"]}))
+)
+builder.add_column(
+    name="sandbox_result",
+    column_type="code-sandbox",
+    target_column="code",
+    language="python",
+    version="*",
+    sandbox_url="http://localhost:2000",
+)
+```
+
+## MCP tool
+
+Use `SandboxMCPConfig` to create a `LocalStdioMCPProvider` for Data Designer
+tool-calling workflows:
+
+```python
+from data_designer_sandbox_piston import SandboxMCPConfig
+
+sandbox_mcp = SandboxMCPConfig(
+    name="sandbox",
+    sandbox_url="http://localhost:2000",
+    language="python",
+    result_fields=["stdout", "stderr", "exit_code"],
+)
+mcp_provider = sandbox_mcp.to_provider()
+tool_config = sandbox_mcp.to_tool_config()
+```
+
+The MCP process can also be launched directly:
+
+```bash
+SANDBOX_URL=http://localhost:2000 \
+SANDBOX_LANGUAGE=python \
+SANDBOX_VERSION='*' \
+SANDBOX_RUN_TIMEOUT=3000 \
+SANDBOX_RUN_CPU_TIME=3000 \
+SANDBOX_TOOL_DESCRIPTION='Execute Python code in a sandbox.' \
+SANDBOX_RESULT_FIELDS=stdout,stderr,exit_code \
+python -m data_designer_sandbox_piston.mcp_server
+```
+
+## Local and remote Piston
+
+For local development on macOS or Linux, run Piston in Docker and point
+`sandbox_url` at `http://localhost:2000`. The package includes a convenience
+script and Docker Compose example under `scripts/` and `docker/`.
+
+The local container stores Piston runtime packages under `/piston` in a Docker
+volume. A fresh volume has no runtimes installed; install the runtimes you need
+through Piston's package API, for example:
+
+```bash
+curl -X POST http://localhost:2000/api/v2/packages \
+  -H 'Content-Type: application/json' \
+  -d '{"language":"python","version":"3.12.0"}'
+```
+
+For remote deployment, build or run a Piston API image and expose port `2000`
+inside your deployment boundary. Piston must be run with the privileges and
+kernel support required by its own sandboxing model. See the
+[Piston project](https://github.com/engineer-man/piston) for current runtime
+installation and security guidance.

docs/plugins/data-designer-sandbox-piston/usage.md

@@ -0,0 +1,93 @@
+# Usage
+
+This plugin is a client for a running Piston API. It does not install Piston
+runtimes for you, which keeps the Data Designer package portable across local,
+cluster, and remote service deployments.
+
+## Start a local sandbox
+
+Use Docker on macOS or Linux:
+
+```bash
+bash plugins/data-designer-sandbox-piston/scripts/run-local-piston.sh
+```
+
+The script starts a container on `http://localhost:2000`. You can also use the
+Docker Compose example:
+
+```bash
+docker compose -f plugins/data-designer-sandbox-piston/docker/docker-compose.yml up --build
+```
+
+Piston runtime packages are stored under `/piston` inside a Docker volume. A new
+Piston package volume starts empty, so install the runtimes your workflow needs:
+
+```bash
+curl -X POST http://localhost:2000/api/v2/packages \
+  -H 'Content-Type: application/json' \
+  -d '{"language":"python","version":"3.12.0"}'
+```
+
+## Execute a code column
+
+```python
+from data_designer.config.config_builder import DataDesignerConfigBuilder
+
+builder = DataDesignerConfigBuilder()
+builder.add_column(
+    name="result",
+    column_type="code-sandbox",
+    target_column="python_code",
+    language="python",
+    sandbox_url="http://localhost:2000",
+)
+```
+
+Each output value is a JSON-serializable dictionary containing stdout, stderr,
+exit code, status, timing, and memory fields.
+
+If a workflow needs Python packages, build or provide a custom Piston Python
+runtime first, then set `version` to that exact runtime version and list the
+expected packages in `python_packages`. The package list is metadata for humans
+and tool descriptions; execution still depends on the configured Piston runtime.
+
+## Add an MCP `run_code` tool
+
+```python
+from data_designer_sandbox_piston import SandboxMCPConfig
+
+sandbox_mcp = SandboxMCPConfig(
+    name="sandbox",
+    sandbox_url="http://localhost:2000",
+    language="python",
+    result_fields=["stdout", "stderr", "exit_code"],
+)
+
+mcp_provider = sandbox_mcp.to_provider()
+tool_config = sandbox_mcp.to_tool_config()
+```
+
+Pass `mcp_provider` and `tool_config` into the Data Designer configuration path
+that configures MCP providers and tool aliases for your LLM columns.
+
+If the sandbox URL is assigned by your launcher at runtime, omit `sandbox_url`
+from `SandboxMCPConfig` and set `SANDBOX_URL` in the environment inherited by the
+MCP subprocess.
+
+## Remote endpoint
+
+For a remote deployment, use the same configuration with a different URL:
+
+```python
+builder.add_column(
+    name="result",
+    column_type="code-sandbox",
+    target_column="python_code",
+    language="python",
+    sandbox_url="https://piston.example.internal",
+)
+```
+
+Keep the Piston endpoint private to trusted Data Designer workers or put it
+behind your own authentication proxy. The plugin forwards code to the configured
+endpoint and relies on that service for isolation and runtime availability.

docs/plugins/index.md

@@ -27,6 +27,17 @@ Browse available Data Designer plugins by what they add to your data generation
       <span class="plugin-doc-card__chips"><span class="plugin-doc-chip">document-chunker</span><span class="plugin-doc-chip">embedding-dedup</span></span>
     </span>
   </a>
+  <a class="plugin-doc-card" href="data-designer-sandbox-piston/" aria-label="Open data-designer-sandbox-piston documentation">
+    <span class="plugin-doc-card__header">
+      <span class="plugin-doc-card__title">data-designer-sandbox-piston</span>
+      <span class="plugin-doc-card__version">v0.1.0</span>
+    </span>
+    <span class="plugin-doc-card__description">Piston-backed code execution columns and MCP tools for Data Designer</span>
+    <span class="plugin-doc-card__section">
+      <span class="plugin-doc-card__label">Entry points</span>
+      <span class="plugin-doc-card__chips"><span class="plugin-doc-chip">code-sandbox</span></span>
+    </span>
+  </a>
   <a class="plugin-doc-card" href="data-designer-template/" aria-label="Open data-designer-template documentation">
     <span class="plugin-doc-card__header">
       <span class="plugin-doc-card__title">data-designer-template</span>

plugins/data-designer-sandbox-piston/CODEOWNERS

@@ -0,0 +1,3 @@
+# Owner(s) of this plugin — used to generate the root CODEOWNERS file.
+# GitHub accepts @username, @org/team, or email format.
+* @NVIDIA-NeMo/data_designer_reviewers @eric-tramel

plugins/data-designer-sandbox-piston/README.md

@@ -0,0 +1,45 @@
+# data-designer-sandbox-piston
+
+Piston-backed code execution for Data Designer.
+
+This package provides:
+
+- A `code-sandbox` column type that executes code from an existing column via a
+  local or remote Piston API.
+- A stdio MCP server that exposes the same sandbox as a `run_code` tool.
+- Local Docker helper examples for macOS and Linux development.
+
+## Installation
+
+```bash
+uv add data-designer data-designer-sandbox-piston
+```
+
+## Column usage
+
+```python
+builder.add_column(
+    name="sandbox_result",
+    column_type="code-sandbox",
+    target_column="python_code",
+    language="python",
+    sandbox_url="http://localhost:2000",
+)
+```
+
+## MCP usage
+
+```python
+from data_designer_sandbox_piston import SandboxMCPConfig
+
+sandbox_mcp = SandboxMCPConfig(
+    name="sandbox",
+    sandbox_url="http://localhost:2000",
+    language="python",
+)
+mcp_provider = sandbox_mcp.to_provider()
+tool_config = sandbox_mcp.to_tool_config()
+```
+
+Plugin documentation for the repository site lives in this package's `docs/`
+directory.

plugins/data-designer-sandbox-piston/docker/Dockerfile

@@ -0,0 +1,7 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+ARG PISTON_IMAGE=ghcr.io/engineer-man/piston:latest
+FROM ${PISTON_IMAGE}
+
+EXPOSE 2000

plugins/data-designer-sandbox-piston/docker/docker-compose.yml

@@ -0,0 +1,25 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+services:
+  piston:
+    build:
+      context: .
+      dockerfile: Dockerfile
+    container_name: data-designer-piston
+    privileged: true
+    ports:
+      - "${PISTON_PORT:-2000}:2000"
+    volumes:
+      - piston-data:/piston
+    restart: unless-stopped
+    healthcheck:
+      test:
+        - CMD-SHELL
+        - "python3 -c 'import urllib.request; urllib.request.urlopen(\"http://127.0.0.1:2000/api/v2/runtimes\", timeout=1)'"
+      interval: 10s
+      timeout: 2s
+      retries: 12
+
+volumes:
+  piston-data: