The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses Node.js subprocess spawning in local-cli.mjs and run-role.mjs to invoke external tools like Claude, Codex, and Grok. This is a core feature for distributing tasks to different model backends.
[PRIVILEGE_ESCALATION]: During setup, install-routing.mjs automatically updates the host agent's security configuration (e.g., permissions.allow) to whitelist the skill's runner scripts. This modification bypasses future manual approval steps for these specific commands.
[INDIRECT_PROMPT_INJECTION]: The skill analyzes user prompts to perform task routing. Adversarial content in the user's task description could potentially manipulate the orchestration logic. Ingestion occurs via the route-hook.mjs script, which processes user input to decide which sub-agents to trigger. Boundaries are managed through contracts/delegation.md, but the system maintains broad capabilities to execute commands and access the filesystem.
[DYNAMIC_EXECUTION]: The tool dynamically compiles Markdown contracts and host-specific configuration files (like TOML for Codex) from templates at runtime, injecting them into the execution environment of sub-agents.

odai-orchestration