The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes a local Python script to collect repository information via shell commands. The script scripts/metadata.py uses subprocess.run() to call git remote, git rev-parse, and git branch to populate plan metadata. These commands are hardcoded and do not process external user input, which mitigates injection risks.
[PROMPT_INJECTION]: The skill processes external files that could contain malicious instructions, creating a surface for indirect prompt injection. 1. Ingestion points: Reads user-referenced files and research documents from the repository. 2. Boundary markers: Instructions mandate reading files fully and trusting research documents as the source of truth. 3. Capability inventory: The agent has access to file-reading tools, file-writing capabilities, and local command execution via metadata.py. 4. Sanitization: No logic is present to filter or sanitize instructions embedded within the analyzed documents.

rpi-plan