The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes various shell commands as part of its core loop.
It uses git diff, git commit, and git revert for local version control operations.
It interacts with GitHub using gh pr diff and gh pr view to fetch pull request data.
It allows for a user-specified or auto-detected verify_cmd to be executed after each fix round to gate commits.
[INDIRECT_PROMPT_INJECTION]: As a code review tool, the skill has a significant attack surface for indirect prompt injection.
Ingestion points: Untrusted data enters the agent context through git diff or gh pr diff calls in SKILL.md.
Boundary markers: Prompts in references/reproduction-prompt.md and references/reviewer-prompt.md use delimiters and instructions to guide subagents.
Capability inventory: The skill can perform git commit, git revert, and execute arbitrary commands via the verify_cmd or subagent tools.
Sanitization: There is no explicit sanitization of the code being reviewed; however, the iterative review and verification loop acts as a functional safeguard.
[EXTERNAL_DOWNLOADS]: The skill mentions and facilitates the installation of dependencies or complementary sub-skills.
It points users to the mrshu-skills marketplace or the mrshu/agent-skills repository on GitHub to install missing reviewer backends (e.g., codex-exec). These are vendor-owned resources.
[DYNAMIC_EXECUTION]: The skill manages the execution of subagents and validation scripts at runtime.
It uses a shell wrapper scripts/run-reviewer.sh to execute reviewer commands (like codex exec) with timeouts and protocol enforcement.
It uses subprocess.run in scripts/test-clarity-output.py to invoke the clarity-pass validator script during testing.

review-anvil