The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill's setup instructions guide the user to install the @openai/codex global NPM package and several platform-specific plugins (openai/codex-plugin-cc, codex@openai-codex). These resources are hosted by a well-known service provider and are documented as part of the intended functionality.
[COMMAND_EXECUTION]: The orchestrator model is instructed to execute the /codex:rescue command, which interfaces with the Codex CLI. This allows the agent to perform engineering tasks and solve problems through an external command-line interface.
[INDIRECT_PROMPT_INJECTION]: The skill acts as an orchestrator that ingests and processes potentially untrusted data to delegate tasks, creating a surface for indirect injection.
Ingestion points: Untrusted data enters the agent's context through user-provided 'Goal' descriptions and codebase 'Context' files during the orchestration process.
Boundary markers: The delegation briefs defined in the instructions do not use explicit delimiters or instructions to prevent the sub-agents from executing commands that might be embedded within the processed data.
Capability inventory: The skill maintains the capability to modify project files (via the fast-worker sub-agent) and execute shell-level commands (via the Codex CLI integration).
Sanitization: There is no evidence of validation or sanitization of input data before it is interpolated into the prompts used to brief the sub-agents.

multi-model-orchestrator