The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of untrusted natural language prompts that can influence agent behavior and tool usage.
Ingestion points: Untrusted data enters the agent context via the session.sendAndWait method described in SKILL.md.
Boundary markers: The documentation and examples do not demonstrate the use of delimiters or specific instructions to disregard embedded commands in user prompts.
Capability inventory: The SDK enables high-privilege capabilities including file system access (via MCP), arbitrary tool execution (via defineTool), and network operations (via BYOK providers).
Sanitization: No explicit input sanitization or validation logic is provided in the core usage patterns.
[COMMAND_EXECUTION]: The skill provides mechanisms to execute local commands and manage subprocesses.
The documentation illustrates the configuration of local Stdio servers using the npx command to spawn external processes (e.g., @modelcontextprotocol/server-filesystem).
It supports session hooks like onPreToolUse that can programmatically allow or deny shell access, though the default behavior depends on user implementation.
[EXTERNAL_DOWNLOADS]: The skill references and demonstrates the retrieval of external dependencies and protocol implementations.
The installation section points to official packages for Node.js, Python, Go, and .NET from standard registries.
Examples show the dynamic execution of the Model Context Protocol (MCP) filesystem server from the official registry.

copilot-sdk