The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The configuration documentation recommends using npx -y to download and execute various Model Context Protocol (MCP) servers from the npm registry.
Evidence includes suggested packages from well-known sources such as @anthropics/exa-mcp, @modelcontextprotocol/server-github, and @modelcontextprotocol/server-filesystem.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates a workflow where a primary agent can launch sub-agents by passing user-supplied prompts.
Ingestion points: The launch-subagent.sh script accepts a "Prompt" argument that is passed directly to the sub-agent session.
Boundary markers: No explicit boundary markers or "ignore" instructions are added to the sub-agent prompts by the script.
Capability inventory: The script utilizes agent-deck, tmux, and jq to manage sessions and send messages.
Sanitization: The script performs sanitization on the session title to prevent path traversal or shell issues, though the prompt content itself is passed through as-is.
[DYNAMIC_EXECUTION]: The config.toml reference describes how to define custom tools and MCP servers that execute commands at runtime.
Examples include running local scripts or binary commands via executable fields in the [mcps.*] and [tools.*] sections.
[PRIVILEGE_ESCALATION]: The documentation mentions a dangerous_mode = true setting for Claude Code.
This setting is intended to skip permission dialogs for file operations and tool use to enable seamless automation, which reduces human-in-the-loop security oversight.

agent-deck