The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted data from the repository being reviewed, including source code, diffs, and project-specific instruction files like REVIEW.md and CLAUDE.md. The documentation in reference/codex-integration.md explicitly states that project-specific rules take precedence, which creates an inherent surface for indirect prompt injection. However, the skill mitigates this risk by employing a tri-engine consensus model and a mandatory grounding step where findings are verified against the actual source code before being reported.
[COMMAND_EXECUTION]: The skill utilizes the Bash tool to execute local command-line utilities for code analysis, including git, codex, and the Claude Code CLI. These executions are scoped to the intended functionality of the skill and are accompanied by instructions to use subscription-based authentication rather than passing sensitive API keys as command arguments.
[DYNAMIC_EXECUTION]: To ensure compatibility across different development environments, the skill dynamically probes for the presence of review binaries in various potential installation paths, such as ~/.bun/bin or ~/.local/bin. This involves executing these binaries with version-check flags, which is a common and documented pattern for tools in this ecosystem to confirm availability.

judge