The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill documents the installation of the Pi CLI using a remote shell script from the vendor's official domain (https://pi.dev/install.sh | sh). This is a standard installation pattern for developer tools and originates from the tool's primary service domain.
[COMMAND_EXECUTION]: Pi utilizes a !command syntax within configuration files (models.json, auth.json) and the interactive editor to execute shell commands for secret resolution (e.g., fetching keys from a password manager) and environment setup. This is a core architectural feature of the tool and is clearly documented for the user.
[EXTERNAL_DOWNLOADS]: The skill details the usage of a package manager (pi install) that can fetch extensions and prompt templates from npm, GitHub, and remote URLs. It explicitly warns users to review third-party code before installation.
[INDIRECT_PROMPT_INJECTION]: The pi-web-access extension introduces an attack surface for indirect prompt injection by fetching and processing untrusted content from the web, GitHub repositories, and YouTube transcripts. The documentation acknowledges these capabilities and describes the use of isolated containers (Docker, OpenShell) for handling untrusted work.
[SAFE_PRACTICE]: The installation instructions recommend the --ignore-scripts flag when installing via npm to prevent the execution of potentially malicious lifecycle scripts during dependency resolution.

pi-agent