The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill utilizes a helper script (device-mcp.py) that downloads external code from an "operator-approved HTTPS repository" and uses uv sync to manage dependencies.\n- [REMOTE_CODE_EXECUTION]: The automated preparation process installs pinned source code and registers stdio entrypoints for the Model Context Protocol (MCP), creating a mechanism for executing remote-sourced code.\n- [COMMAND_EXECUTION]: The instructions direct the agent to run various shell commands for tool discovery and environmental setup, including python3, uv sync, command -v, and tool-specific health checks.\n- [INDIRECT_PROMPT_INJECTION]: The agent is instructed to read from historical state stores in ~/.moru/ (lessons) and ~/.moru/project-memory/. This ingestion of previously generated or recorded data provides a surface for indirect prompt injection.\n
Ingestion points: ~/.moru/ and ~/.moru/project-memory/.\n
Boundary markers: The guidance notes that metadata should be "content-free" and content is "bounded," but technical isolation markers are not defined.\n
Capability inventory: The agent can execute scripts, install packages via uv, and perform file system operations.\n
Sanitization: The policy specifies what should not be stored (prompts, secrets), but there is no mechanism for sanitizing data upon retrieval.\n- [PRIVILEGE_ESCALATION]: The documentation mentions that the tool may request "sandbox escalation" when performing device-level verification tasks.

local-tools