The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from pull request diffs and project source code across multiple specialized auditor agents, creating a potential vector for malicious instructions to influence the review outcome.
Ingestion points: PR diffs and codebase files are read and processed in references/pr-review.md and references/team-review.md.
Boundary markers: Markdown headers and code blocks are used for delimitation, but no explicit instructions are provided to agents to ignore potentially malicious embedded content within the analyzed files.
Capability inventory: The skill can execute shell commands for linting, interact with GitHub via gh api, and perform automated file modifications and commits via a fix loop.
Sanitization: There is no evidence of content sanitization or escaping before project data is interpolated into auditor prompts.
[COMMAND_EXECUTION]: The skill orchestrates the execution of several command-line development tools on project files, which could be exploited if filenames are maliciously crafted.
Evidence: references/pr-review.md instructs the agent to run ruff, vulture, npx knip, npx tsc, and git check-ignore on identified changed files.
Evidence: references/code-review-delivery.md describes the use of gh api to post comments and suggestions to GitHub.
[DYNAMIC_EXECUTION]: The review workflow involves tools and sub-processes that may execute code or load dependencies from the project under review.
Evidence: The use of npx knip and npx tsc in references/pr-review.md involves tools that interact with the project's dependency tree and source code.
Evidence: The fix loop described in references/code-review-delivery.md involves running project-defined tests to verify fixes before committing them.

review-method