The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is entirely centered around executing shell commands via the bash tool to run interactive terminal applications like Codex, Claude Code, and Pi. It provides instructions for complex command sequences, background process management, and direct stdin manipulation.
[PRIVILEGE_ESCALATION]: The skill documentation explicitly defines and encourages the use of an elevated parameter for the bash tool, which is intended to run processes on the host machine instead of within the standard sandbox environment. Furthermore, it promotes the use of the --yolo flag for the Codex CLI, which is described as disabling both the sandbox and all user approvals, significantly increasing the risk of unintended or malicious system modifications.
[EXTERNAL_DOWNLOADS]: The skill references the installation of a specific external Node.js package (@mariozechner/pi-coding-agent) for use as a coding agent.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to perform tasks such as 'Reviewing PRs' and 'Parallel Issue Fixing' by running AI agents on untrusted external codebases. This creates a significant surface for indirect prompt injection, where malicious instructions embedded in a pull request or repository could be executed by the sub-agent with the permissions granted by the skill.
Ingestion points: External code repositories and pull requests via git clone, gh pr checkout, and git worktree (found in SKILL.md).
Boundary markers: None identified in the instructions provided to the sub-agents.
Capability inventory: Full shell execution capabilities via the bash tool, including background processing and potential host access via the elevated flag.
Sanitization: No sanitization or filtering of external content is mentioned before it is processed by the coding agents.

coding-agent