The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes a roster configuration file (.agents/code-polish-roster.md) located within the repository's root directory. This file defines which skills, agents, or instruction files the agent should invoke. A malicious pull request or repository could modify this file to point the agent to follow adversarial instructions (via the file: directive) or execute unexpected tools, potentially leading to unauthorized behavior during the review process.
Ingestion points: SKILL.md (reading roster configuration from <repo root>/.agents/code-polish-roster.md and loading external instruction files).
Boundary markers: The instructions do not specify boundary markers or isolation for the instructions loaded via the roster.
Capability inventory: The skill has capabilities to read/write files (in --fix mode), execute other skills/agents, and run project-specific test suites/typecheckers.
Sanitization: The skill implements several structural safeguards to mitigate risks, including a requirement that only non-functional tokens (comments, docstrings, identifiers) are changed, automatic diff verification to revert unexpected changes, and a requirement that all tests and typechecks pass before changes are accepted.

code-polish