The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes several command-line tools including git, uv, pnpm, gh, and cubic to perform repository management, run linting/tests, and interact with GitHub. These actions are standard for the skill's stated purpose of PR automation.
[INDIRECT_PROMPT_INJECTION]: The skill has an attack surface for indirect prompt injection as it is instructed to ingest and follow findings from external tool outputs (cubic review --json) and local skill files. While this could allow external content to influence agent behavior, the skill provides specific instructions for the agent to address findings and allows for manual disputes of incorrect comments.
Ingestion points: Output from cubic review and instructions in .agents/skills/polar-code-review/SKILL.md.
Boundary markers: The skill does not define specific delimiters for tool output but uses a structured review process.
Capability inventory: Shell command execution via git, uv, pnpm, gh, and cubic (SKILL.md).
Sanitization: Instructions include a manual check to ensure no PII is included and a mechanism to dispute factually wrong review comments.
[DATA_EXPOSURE]: The skill includes a safety requirement to verify that pull request bodies contain no PII, statistics, or organizational information, as pull requests are public. This follows security best practices for public code contributions.

open-pr