The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes multiple commands via the GitHub CLI (gh) and Elixir build system (mix). These include gh pr view, gh api graphql, gh api (POST), mix compile, and mix test. These commands are used for their intended purpose of PR management and code verification.
[DATA_EXFILTRATION]: The skill reads pull request metadata and review comments from GitHub. It also accesses the local file system to read and modify Elixir code files to address reviewer feedback. It posts responses back to the GitHub API, which is a trusted and expected behavior for a PR review tool.
[PROMPT_INJECTION]: The skill is subject to indirect prompt injection as it processes untrusted data from GitHub pull request comments. An attacker with review permissions could attempt to embed malicious instructions within a comment to influence the agent's code fixes or responses. This risk is significantly mitigated by the skill's 'Iron Laws' and the requirement for explicit user confirmation before any code is modified or any response is posted to GitHub.
Ingestion points: Fetches comment bodies from GitHub using paginated GraphQL queries in SKILL.md.
Boundary markers: Includes 'Iron Laws' (e.g., 'NEVER auto-post responses', 'NEVER claim a fix without a shown diff') and a mandatory confirmation step (Step 3.4) to verify actions.
Capability inventory: Capabilities include file writing (code fixes), shell command execution (mix), and network writes via the GitHub API.
Sanitization: The skill relies on user review and confirmation rather than automated sanitization of comment text.

pr-review