The Agent Skills Directory

[COMMAND_EXECUTION]: The skill's core functionality involves orchestrating external developer tools and Git operations via subprocess. It invokes command-line interfaces for various LLM backends (Codex, Grok, Claude) and manages isolated filesystems using git worktree and rsync. These operations are strictly scoped to the tool's intended purpose of project management and backend invocation.\n- [INDIRECT_PROMPT_INJECTION]: As an ensemble orchestrator, the skill ingests user-provided "decision packets" which are then processed by multiple sub-agents. To mitigate potential injection risks, the skill implements a robust security architecture: \n
Ingestion points: Data enters via the main scripts/dispatch.py script from stdin or a specified prompt file, which is then passed to backend adapters.\n
Boundary markers: Input packets are delimited by explicit header markers within the dispatcher logic to separate system instructions from user-provided content.\n
Capability inventory: The skill possesses significant capabilities, including the ability to write to and execute code within isolated worktrees, and to spawn sub-processes for external LLM command-line tools.\n
Sanitization: The skill employs multiple layers of protection, including a whitelist-based environment variable sanitizer (sanitized_env in scripts/backends/native_headless.py) that strips credential-shaped variables, and regex-based redaction (redact_error in scripts/dispatch.py and redact in scripts/backends/claude_headless.py) to prevent API keys or tokens from appearing in logs or agent output.

moa