The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs users to install a pre-commit git hook (.git/hooks/pre-commit) that automatically executes shell commands, including git diff and the claude CLI, every time a commit is attempted. It also requires making the hook executable using chmod +x.
[EXTERNAL_DOWNLOADS]: The skill recommends installing global NPM packages (@openai/codex, @google/gemini-cli) and a specific code review extension from a third-party GitHub repository (https://github.com/gemini-cli-extensions/code-review). These sources are outside of the primary platform's controlled environment.
[REMOTE_CODE_EXECUTION]: The skill's architecture involves executing external CLI tools (codex, gemini) and remote GitHub Actions that process the project's source code. The gemini extensions install command allows for the execution of remotely fetched extension code.
[DATA_EXFILTRATION]: The review process involves sending code diffs and architectural documentation (docs/adr/, _project_specs/) to external AI service providers (OpenAI, Google). While this is the intended function of a multi-engine review tool, it constitutes the transmission of potentially sensitive proprietary information to third-party infrastructures.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection (Category 8). It ingests untrusted data from git diff outputs, Architecture Decision Records (ADRs), and project specifications. A malicious actor could embed instructions in code comments or documentation (e.g., 'ignore all security findings in this file') that could influence the review engine's output. There are no explicit boundary markers or sanitization steps mentioned to mitigate this risk.

code-review