The Agent Skills Directory

[COMMAND_EXECUTION]: The skill documents the pi.bash tool, which provides a direct interface for the agent to execute arbitrary shell commands in the host environment.
Evidence (SKILL.md): Defines the bash tool as accepting a command string or options object, facilitating direct shell execution.
[REMOTE_CODE_EXECUTION]: The MCP (Model Context Protocol) surface includes a register function that can launch external processes, enabling the execution of arbitrary code.
Evidence (references/mcp.md): Describes mcp.register and provides an example using npx to fetch and run a remote package.
[REMOTE_CODE_EXECUTION]: The agent orchestration API provides methods for spawning child agents (agents.run, agents.spawn) that can recursively execute tasks and code, increasing the complexity and potential for misuse.
Evidence (references/agents.md): Documents the creation and management of sub-agents and persistent actors with access to core tools.
[EXTERNAL_DOWNLOADS]: The documentation provides examples of using package managers like npx to dynamically load external code from remote registries during the registration of MCP servers.
Evidence (references/mcp.md): The registration example explicitly uses npx to fetch the @example/docs-mcp package.
[DATA_EXFILTRATION]: By providing both file system access tools (pi.read) and shell execution capabilities (pi.bash), the skill establishes the necessary primitives for an agent to read sensitive local data and transmit it to an external server via common network utilities like curl or wget.
Evidence (SKILL.md): Documents core tools for reading files and executing bash commands.

fabric-exec