The Agent Skills Directory

[PROMPT_INJECTION]: Indirect prompt injection vulnerability. The skill fetches untrusted data (GitHub issue title, body, and labels) and passes it directly to multiple subagents without boundary markers or sanitization.
Ingestion points: Issue data is fetched via gh issue view in Step 3.
Boundary markers: None. Data is interpolated directly into instructions for the validator, researcher, and detector subagents.
Capability inventory: Subagents are granted access to the codebase (current working directory) and can communicate results back to the orchestrator.
Sanitization: No evidence of sanitization or escaping of external content before it is passed to subagents.
[COMMAND_EXECUTION]: The skill uses shell commands (gh) with variables derived from user-supplied input ($ARGUMENTS). If the agent fails to strictly validate that the extracted issue ID is numeric, it could lead to command injection when executing gh issue view or gh issue comment.
[COMMAND_EXECUTION]: The skill performs multiple operations on a GitHub project board using the gh api graphql tool. While these operations use IDs fetched from earlier CLI calls, the overall workflow involves significant automated modification of repository metadata and project states based on AI-generated proposals.

triage