The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to automatically execute shell commands found in the project's root directory. Specifically, it parses justfile recipes, Makefile targets, and package.json scripts to run tests, linters, and other pre-merge checks.
[INDIRECT_PROMPT_INJECTION]: The skill has a significant attack surface for indirect prompt injection because it relies on untrusted local files to determine its next actions. An attacker could craft a repository with malicious instructions inside project metadata files that the agent would then execute with local privileges.
Ingestion points: The skill reads justfile, Makefile, and package.json from the local environment using an 'Explore' subagent.
Boundary markers: No specific boundary markers or 'ignore' instructions are provided to the agent to prevent it from following commands embedded maliciously in these files.
Capability inventory: The agent has the capability to execute arbitrary subprocesses and interact with the GitHub CLI (gh).
Sanitization: There is no evidence of sanitization or validation of the extracted command strings before they are passed to the shell for execution.

ship