The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the creation of agents that ingest untrusted user input while maintaining access to sensitive capabilities like file modification and command execution.
Ingestion points: Untrusted data enters the agent context via user-provided prompts in SKILL.md and references/working-examples.md, as well as data retrieved from external sources through tools or MCP servers.
Boundary markers: There are no specific instructions or boundary markers provided to delimit untrusted content or instruct the agent to ignore embedded commands.
Capability inventory: The skill enables significant capabilities across multiple files, including shell command execution (bash, execute), file editing (edit_file), and broad filesystem access via MCP servers as detailed in references/cli-agents-mcp.md.
Sanitization: No sanitization, escaping, or validation protocols for external content are documented in the instruction set or provided examples.
[EXTERNAL_DOWNLOADS]: The skill facilitates the installation of official SDK packages and tools from established services.
Downloads the official SDK packages including @github/copilot-sdk (NPM) and github-copilot-sdk (PyPI).
Provides examples for running local MCP servers using npx -y @modelcontextprotocol/server-filesystem.
[COMMAND_EXECUTION]: The documentation describes functionality for executing shell commands and managing local processes.
Documents the use of the bash and execute tool aliases for custom agents.
Provides instructions for running the Copilot CLI in headless mode via copilot --headless.

copilot-sdk