The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The framework facilitates workflows where untrusted data from user input or tool outputs (e.g., file content, API responses) is stored in session state and subsequently interpolated into agent instructions via template keys like {draft} or {validation_result}.
Ingestion points: User messages processed through runAsync in SKILL.md, file data retrieved via MCPToolset in references/tools.md, and remote API responses from OpenAPIToolset in references/tools.md.
Boundary markers: The documentation recommends implementing guardrails through beforeAgentCallback, using a SecurityPlugin for human-in-the-loop approvals, and utilizing the safety_v1 metric within the evaluation framework described in references/testing.md.
Capability inventory: The skill supports sandboxed code execution (CODE_EXECUTION), file system interactions through MCP servers, and network operations via OpenAPI tools.
Sanitization: The production checklist in references/deployment.md emphasizes input validation and human review checkpoints as best practices.
[COMMAND_EXECUTION]: The framework supports the Model Context Protocol (MCP) to execute shell commands for starting local servers.
Evidence: references/tools.md documents MCPToolset.fromServer examples that run commands such as npx -y @anthropic/mcp-server-filesystem. This utilizes tools from a well-known service for intended functionality.
[EXTERNAL_DOWNLOADS]: The skill instructs the agent to fetch configuration and packages from remote sources.
Evidence: OpenAPIToolset.fromUrl in references/tools.md is designed to download OpenAPI specifications from external URLs.
Evidence: The setup guide in SKILL.md fetches packages from the official npm registry, including @google/adk. These downloads target trusted organizations and official registries.
[DYNAMIC_EXECUTION]: The framework includes a built-in tool for runtime code execution.
Evidence: references/tools.md describes the CODE_EXECUTION tool for sandboxed code execution, which is a standard feature of the Agent Development Kit.

google-adk-typescript