The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local commands such as git and the codex CLI tool using Node.js subprocess methods. It also manages its own background worker processes using process.execPath. These operations are essential for the skill's documented functionality of repository analysis and task delegation.\n- [DATA_EXFILTRATION]: To provide context for its features, the skill reads repository metadata including git diff output, branch logs, and the content of untracked files (up to a 24KB limit). This data is passed to the OpenAI Codex CLI for analysis and processing. No unauthorized external network communication was detected.\n- [PROMPT_INJECTION]: The skill processes untrusted repository data which could contain malicious instructions designed to influence the model. It implements structured prompts with XML delimiters to mitigate this risk.\n
Ingestion points: Repository data collection in scripts/lib/git.mjs (diffs, status, untracked files).\n
Boundary markers: Uses XML tags such as <repository_context> and <task> in prompts/adversarial-review.md to delimit untrusted data.\n
Capability inventory: Supports file system writes via task --write and code execution through the codex CLI.\n
Sanitization: Implements file size limits for untracked files but relies on the underlying LLM's guardrails for content safety.\n- [EXTERNAL_DOWNLOADS]: The documentation directs users to install the @openai/codex CLI tool from npm. This is an expected dependency for an OpenAI-focused developer tool.

codex-companion