The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input such as code diffs, pull request descriptions, and commit messages. While these inputs could potentially contain malicious instructions intended to influence the reviewer agents, the skill includes a mandatory 'Evidence Protocol' that instructs agents to treat the input as a claim rather than evidence and to verify findings against the full source code.
Ingestion points: Git diff outputs, GitHub CLI pull request data, and codebase files processed in SKILL.md and references/deep-review.md.
Boundary markers: The agent is explicitly instructed to read the entire file content and trace call paths rather than relying solely on the provided hunks or descriptions.
Capability inventory: The skill executes shell commands (git, gh, node), performs local file system operations in ~/.agents/sessions, and spawns subagents for parallel review.
Sanitization: The render-review.mjs script performs HTML escaping on data before generating the final report files.
[COMMAND_EXECUTION]: The skill utilizes several CLI tools to perform its tasks, including git for analyzing version history, gh for GitHub integration, and node for executing the bundled reporting and maintenance scripts. It also uses the open command to display the generated HTML reports to the user.
[EXTERNAL_DOWNLOADS]: The skill interacts with GitHub using the gh command-line tool to fetch pull request diffs and metadata. GitHub is a well-known and established service for software development.

code-review