The Agent Skills Directory

[Command Execution]: The skill's scripts utilize subprocess.run to interact with the gh CLI and git. This is the intended design for its functionality, allowing it to fetch pull request metadata, read repository files, and post review comments. The commands are built using parameters like pull request numbers and repository names obtained from the GitHub environment.
[External Data Interaction]: The skill fetches data from the GitHub API using the gh tool to analyze pull request diffs, code content, and existing comments. This interaction is central to its purpose as a code review assistant.
[Indirect Prompt Injection Surface]: By ingesting external data such as pull request code and comments, the skill has an attack surface for indirect prompt injection. This risk is mitigated by a mandatory human approval step (Step 6), where the user reviews all proposed comments before they are posted. Additionally, sub-agents are restricted by a 'depth limit' that prevents them from executing code or auditing third-party dependencies.
[Defensive Reporting Practices]: The build_report.py script implements a csv_safe function to protect against spreadsheet formula injection (CSV Injection). It prepends a tab character to any data starting with characters that could trigger formula execution (=, +, -, or @) when the report is opened in tools like Excel.
[Automated Verification]: The verify_findings.py script provides a layer of validation by checking that proposed review comments accurately reflect the code at the specified lines. It also filters out findings that are deemed too complex or 'expensive' for a human to quickly verify, which helps reduce the impact of potential AI hallucinations.

github-pr-review