The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill involves processing untrusted external data such as pull request comments, descriptions, and automated feedback. This ingestion of external content creates a surface for indirect prompt injection. The skill provides clear mitigation instructions, directing the agent to treat such feedback as adversarial input rather than authority and requiring verification against the original project scope.\n- Ingestion points: Data is ingested via GitHub CLI commands including gh pr view and gh issue view.\n- Boundary markers: The skill does not use specific structural delimiters for external content but establishes a logical trust boundary via instruction.\n- Capability inventory: The skill has capabilities to read and edit files, and to execute shell commands for repository testing (e.g., just fast-check, uv run).\n- Sanitization: No technical sanitization of external text is implemented; the skill relies on the agent's reasoning and the explicit 'Scope Discipline' section to filter out out-of-scope or malicious instructions.

pull-request