The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/invoke_peer.py uses subprocess.run to execute the claude and codex CLI tools. This is the primary mechanism for cross-agent collaboration and includes security controls such as wrapping the command in sandbox-exec on macOS to enforce a read-only profile.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data including raw user requests and project artifacts, creating a potential surface for indirect injection.
Ingestion points: invoke_peer.py reads content from the handoff prompt and files specified in the artifact_manifest.
Boundary markers: The system uses specific XML-like tags (e.g., <raw_user_request>, <artifact_manifest>) to isolate untrusted data.
Capability inventory: The skill can execute peer CLI tools and perform atomic file operations (copy, replace) during installation via scripts/install_skill.py.
Sanitization: The implementation validates the existence of required markers and ensures artifact paths are absolute and within permitted directory scopes before processing.
[DYNAMIC_EXECUTION]: The unit tests scripts/test_install_skill.py and scripts/test_invoke_peer.py utilize importlib.util to dynamically load the skill's scripts for isolated testing. This is a standard development practice for Python testing frameworks.

dual-agent-collaboration