The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes shell commands using the gh (GitHub CLI) for repository management, GraphQL queries, and REST API calls. It also executes a user-provided 'Gate command' (e.g., make check) to validate code changes before committing.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from multiple sources including inline comments, review bodies, and issue comments.
Ingestion points: Fetches comments via GraphQL and REST API in SKILL.md (Step 2) and references/graphql-queries.md.
Boundary markers: None explicitly defined in the prompts to separate code instructions from comment content.
Capability inventory: The skill has capabilities to read files, modify code (Step 4), commit and push changes (Step 6), and resolve threads via GraphQL.
Sanitization: The skill passes replies through an unslop utility before posting, but there is no evidence of sanitization for the instructions derived from comments before code modification.
[PRIVILEGE_ESCALATION]: The skill relies on an existing authenticated gh session. While it does not mint its own credentials, it uses the high-privilege access of the authenticated user to perform writes, pushes, and administrative actions (resolving reviews) on the repository.

review-loop