The Agent Skills Directory

[DATA_EXFILTRATION]: The skill reads session authentication keys and metadata from ~/.claude/sessions and Codex SQLite databases. This access is utilized to authenticate local inter-session communication and verify message delivery history.
[COMMAND_EXECUTION]: It executes local system binaries including uv, codex, and claude to manage message delivery, environment preparation, and worker execution.
[PERSISTENCE]: The skill registers a tool hook via hooks/hooks.json. This PreToolUse hook ensures that a security guard script (native_guard.py) validates cross-session messaging tool calls before they are processed by the agent.
[INDIRECT_PROMPT_INJECTION]: The skill ingests text messages from other agent sessions. It implements security boundaries by wrapping this external data in advisory XML-style tags and providing explicit instructions to treat peer coordination text as untrusted.
[DYNAMIC_EXECUTION]: It uses Python's importlib.util to dynamically load sibling modules for coordination and messaging logic at runtime.
[EXTERNAL_DOWNLOADS]: The skill utilizes external Node.js dependencies (agentpair, @agentpair/protocol) retrieved from standard registries to facilitate its network-based agent pairing and messaging features.

peer-message