The Agent Skills Directory

[COMMAND_EXECUTION]: The skill interacts with the local environment using the gh and git command-line tools. It dynamically constructs shell commands using user-provided arguments such as repository names (owner/repo) and pull request numbers.
Evidence: Commands such as gh pr list --repo <owner/repo>, gh pr view <number>, and git fetch origin pull/<number>/head are used to gather data and perform actions.
[INDIRECT_PROMPT_INJECTION]: The skill processes data from external sources (GitHub PRs) that can be controlled by third-party authors or commenters. This data could contain malicious instructions designed to manipulate the agent's review findings or influence its behavior.
Ingestion points: The skill reads PR titles, descriptions, and comments via gh pr view, gh api, and files like pr.json and comments.json (in report-only mode).
Boundary markers: The instructions do not specify the use of delimiters or 'ignore' instructions when interpolating PR content into the review prompts.
Capability inventory: The agent has the ability to execute shell commands, post comments back to GitHub via gh api --method POST, and dispatch tasks to other code review models.
Sanitization: There is no evidence of sanitization or filtering of the external PR content before it is analyzed or passed to the reviewer agent.
[DATA_EXPOSURE]: The skill reads local files within the project directory to determine business logic and model configuration.
Evidence: The skill reads docs/agents/issue-tracker.md to understand issue linking logic and docs/agents/delivery.md to identify the appropriate models for the PR code reviewer tier.
[DYNAMIC_EXECUTION]: The skill determines which AI model to dispatch for the code review pass at runtime based on the contents of local configuration files.
Evidence: In Stage 3, the model is selected based on the pr-code-reviewer Seats row or PR code reviewer Agents row in docs/agents/delivery.md.

review-prs