Python sandbox template

Use Python when an agent needs to turn a supplied dataset into an answer. Upload a CSV, let the agent write a small script, then return a compact statistical summary and the chart that supports it. Keep the model outside the sandbox; the interpreter needs CPU and memory, not a second model runtime.

• Pin Python and package versions in a reproducible environment; record the lockfile with the analysis.
• Place inputs under /work/input and outputs under /work/output. Give each task its own directory.
• Read a sample before loading an entire dataset. Select columns and explicit dtypes to control pandas memory use.

python /work/analysis.py

• A JSON result with counts, units and missing-value handling.
• PNG or SVG charts plus the generated source script.
• Bounded stdout, stderr and the process exit code.

Run a known fixture with missing values and assert the totals. Save the chart, verify its dimensions, and repeat with a failing script to ensure errors reach the agent.

Start with 2 vCPU and 4 GB of RAM, then measure peak memory and task duration on a representative fixture. These are workload planning values, not a benchmark or a provisioned configuration. Use the sandbox cost calculator to estimate running time and retained snapshots.

• A 4 GB starting profile does not mean a 4 GB CSV will fit; parsing and copies require additional memory.
• Reject arbitrary pickle files and executable input formats.
• Allow package downloads only during preparation, then remove network access for offline analysis.

Keep model inference separate from this execution profile. A hosted model or gpuOS can decide the next action while the CPU environment runs it. The quickstart describes the account workflow and the proposed runtime contract.