The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/run_pi.py invokes the external pi CLI tool using subprocess.run(). It passes arguments including user-provided prompts and the content of specified files. The command is constructed as a list, which prevents shell-level injection.
[EXTERNAL_DOWNLOADS]: The skill requires the global installation of the npm package @mariozechner/pi-coding-agent. This is a third-party dependency necessary for the skill to function.
[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes untrusted data from multiple sources to construct the final prompt for the downstream agent.
Ingestion points: Data is pulled from the prompt argument and file paths provided via --prompt-file and --session-file in scripts/run_pi.py.
Boundary markers: The build_prompt function in scripts/run_pi.py uses simple text headers (e.g., 'Execution constraint:', 'Prior conversation context to continue from:') to separate segments. These do not provide robust isolation against adversarial content within the ingested files.
Capability inventory: The underlying pi tool has significant capabilities, including the ability to read and edit files and execute shell commands (bash) in its default mode.
Sanitization: No sanitization or filtering is performed on the content of the prompt or files before they are passed to the external CLI tool.

pi-runner