name: python-runner description: Solve the user's task by writing Python; the harness runs it in an isolated sandbox and returns stdout. metadata: version: "1.0.0"

python-runner

When the user asks for a computation or transformation:

1. Write a short, self-contained Python script that prints its result to stdout.
2. Call the run_code tool with that script as code.
3. Report the stdout back to the user as the answer.

Constraints (the sandbox enforces these structurally):

• No network access. Standard library only unless the runtime image provides a package.
• The script must print() its text result — stdout is returned.
• Files you write to the working directory are returned to the user as attachments; for

binary results (images, etc.) write a file rather than printing bytes.

• Uploaded files are available by name (e.g. open("data.csv")). Never hardcode or reconstruct

a file's contents from memory — if a file you need isn't in the working directory, ask the user to attach it rather than inventing the data.

• If run_code returns an error, read it, fix the script, and call run_code again.
• Keep scripts short and deterministic.