The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it processes untrusted project data to guide an external implementation agent.
Ingestion points: In Step 2, the skill reads project-specific configuration files (CLAUDE.md, AGENTS.md) and user-provided task descriptions to synthesize a 'brief' for the Codex CLI.
Boundary markers: The generated brief utilizes XML tags based on codex:gpt-5-4-prompting guidelines and includes a dedicated 'Hard constraints' section to restrict the external agent's operations.
Capability inventory: The skill executes the Codex engine via node, performs git operations for branch management, and runs project-defined commands such as npm install, pip install, and various test suites in Steps 0, 3, and 5.
Sanitization: The skill's workflow includes a mandatory manual review step (Step 6) where the orchestrating agent is instructed to personally verify all code changes and avoid delegating the review back to the Codex engine.
[COMMAND_EXECUTION]: The skill performs multiple command-line operations to setup the environment and execute the Codex implementation logic.
Evidence: Step 0 involves locating and executing a script (codex-companion.mjs) from the user's plugin cache ($HOME/.claude/plugins/cache/openai-codex/) using node.
Evidence: The skill modifies the global configuration file ~/.codex/config.toml to enable network_access for the workspace-write sandbox environment.
Evidence: The skill executes arbitrary project setup, linting, and testing commands within dynamically created git worktrees in Steps 3 and 5.

codex-dev-native