The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses Python's subprocess module in scripts/orchestrator_supervisor.py and scripts/resource_readiness.py to run shell commands. These commands are used to launch AI agent CLIs (Claude, Codex, Antigravity) and perform resource health checks. The logic includes executing user-provided commands passed via arguments to verify if resources are ready for task execution.
[REMOTE_CODE_EXECUTION]: The skill implements a local supervisor daemon (scripts/orchestrator_supervisor.py) that listens on a localhost port. This daemon receives commands via a TCP socket and executes them as new processes. While it uses a dynamically generated token for authentication (stored in server.json), this mechanism allows the agent to manage long-running tasks and inject follow-up prompts into active CLI sessions.
[PROMPT_INJECTION]: The skill's control plane relies on GitHub Issues or local Markdown files. The agent reads these external records to determine which tasks to run and which verification commands to execute. A malicious actor who can edit these records (e.g., by commenting on a GitHub issue) could potentially inject malicious commands or instructions that the orchestrator would then execute.
[DATA_EXFILTRATION]: The skill performs network operations using the GitHub CLI (gh) and PowerShell's Invoke-RestMethod to interact with the GitHub API. This involves reading issue data and posting handoff summaries, which transmits project-related information to GitHub's infrastructure.
[EXTERNAL_DOWNLOADS]: The documentation instructs users to install external system and language dependencies, such as tmux and pywinpty, to support interactive PTY backends for specific agent CLIs.

orchestrator-cli