codex

Codex is a specialized skill for the Hermes Agent designed to bridge the agent with the OpenAI Codex CLI, enabling sophisticated, autonomous code manipulation directly within the terminal. It is primarily intended for software engineers, developers, and DevOps practitioners who need to automate complex programming workflows, manage technical debt, or handle repetitive batch operations across multiple projects. By leveraging the Codex CLI, the agent acts as an autonomous coding assistant capable of interacting with local file systems and version control histories.

• Perform one-shot coding tasks such as implementing new features, drafting refactoring suggestions, or writing unit tests.

• Conduct automated Pull Request (PR) reviews by cloning target branches into temporary workspaces and utilizing Codex to analyze diffs against the base branch.

• Execute batch issue fixing by spinning up parallel Git worktrees, allowing multiple instances of the agent to resolve distinct bugs simultaneously without interfering with each other.

• Support for multiple execution modes including full-auto (for auto-approved sandbox modifications) and yolo (for high-speed, direct modifications).

• Advanced process management capabilities allowing the agent to handle long-running operations in the background, poll for progress, and inject human-in-the-loop approvals via terminal input.

• Prerequisites: Ensure Node.js is installed to run the npm package @openai/codex and that a valid OpenAI API key is configured.

• Execution environment: The skill strictly requires a Git repository context to function; for scratch work, use the suggested pattern of creating a temporary directory and initializing a Git repo before running commands.

• PTY requirement: Always set pty=true in terminal calls to maintain interactivity, as Codex is a terminal-based application that requires a pseudo-terminal to avoid hanging.

• Resource management: Use background=true for long-horizon tasks and utilize the provided process management tools (poll, log, submit, kill) to monitor and control agent status effectively.

• Security: Exercise caution with the --yolo flag, as it bypasses sandbox protections and direct-writes to the filesystem; it is recommended for low-stakes or isolated testing environments only.