The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the orchestrator to use several shell utilities including git (for change tracking), sed (for prompt templating), mkdir (for directory structure), and realpath (for path resolution). These are used to manage the development lifecycle and ensure subagent context is correctly isolated.
[REMOTE_CODE_EXECUTION]: Implementation and review tasks are dispatched to an external model via a local Python script (codex_bridge.py). While this involves external model execution, it is the primary intended function of the skill and is managed through controlled bridge invocations.
[DATA_EXFILTRATION]: Content from the workspace, project plans, and task specifications are sent to the Codex service. This data transmission is the core mechanism of the skill and is explicitly documented for the user.
[INDIRECT_PROMPT_INJECTION]: The skill introduces a surface for indirect injection where the orchestrator (Claude) must process and potentially act upon output from the Codex subagent. The skill mitigates this by requiring the orchestrator to independently verify Git diffs and run verification tests rather than relying on the subagent's summary reports.

subagent-driven-codex