The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses shell commands and Node.js to perform project setup. It queries git metadata using git rev-parse, creates directories with mkdir -p, and executes internal tooling such as build-codex-artifacts.js to generate documentation.
[INDIRECT_PROMPT_INJECTION]: The skill generates the AGENTS.md file by ingesting project-specific information from local files.
Ingestion points: Reads the project name and test scripts from package.json and CLAUDE.md to customize the agent's core behavioral requirements.
Boundary markers: The generated file uses standard Markdown headers for structure but does not implement explicit "ignore instructions" delimiters for the interpolated project data.
Capability inventory: The skill has authority to modify the filesystem (Write/Edit) and execute shell commands (Bash) within the repository.
Sanitization: The skill verifies that the generated output contains no unresolved placeholders and stays within a 24 KiB size limit to prevent prompt bloat or unexpected behavior.
[DYNAMIC_EXECUTION]: The skill installs and executes git hooks (commit-msg and pre-push). The bash stanzas for these hooks are implemented with specific hardening measures, including the use of privileged mode (bash -p) and the removal of sensitive environment variables (such as BASH_ENV and SHELLOPTS) to prevent environment-based hijacking by a committer.

codex-setup