The Agent Skills Directory

[COMMAND_EXECUTION]: The skill manages the execution of external tools like git, claude, codex, and herdr using a supervisor script (scripts/headless-run.mjs). Commands are executed using child_process.spawn with argument arrays, which is a secure practice to prevent shell injection. The scripts/herdr-visible-run.mjs script includes a shellQuote function that correctly escapes arguments for commands run within Herdr panes.
[REMOTE_CODE_EXECUTION]: Code review tasks are performed by AI agents through their respective CLI tools. The Codex wrapper (scripts/headless-codex.mjs) explicitly configures the agent with sandbox_mode="read-only" to prevent accidental or malicious code execution. The Claude wrapper (scripts/headless-claude.mjs) manages writable modes specifically for simplification tasks, with a robust fingerprint-based restore mechanism that uses binary patches to revert the repository to its baseline state on failure.
[DATA_EXFILTRATION]: The skill handles repository diffs as part of its core functionality. It resolves git ranges to specific SHAs before transmission, ensuring only the intended code is reviewed. There is no evidence of unauthorized network activity or exfiltration of sensitive files like credentials or SSH keys.
[PROMPT_INJECTION]: While the skill processes external data (code diffs), it maintains security by defining clear review axes (Standards vs. Spec) and using structured command wrappers that prevent the LLM from deviating from the assigned task range or target SHAs.

review-gate