The Agent Skills Directory

[METADATA_POISONING]: The skill frontmatter and release highlights claim a futuristic version of the framework (v2.39.0) and a release date in September 2026, and it references non-existent model versions such as OpenAI GPT-6 and Claude 5.\n- [COMMAND_EXECUTION]: The documentation provides instructions for using the Model Context Protocol (MCP) to execute local shell commands and subprocesses via the MCPServerStdio and MCPToolset components.\n- [EXTERNAL_DOWNLOADS]: The skill describes how to fetch media content (images, audio, video, documents) from remote URLs, including the use of a force_download flag that enables server-side fetching.\n- [INDIRECT_PROMPT_INJECTION]: The framework exposes several surfaces where untrusted external data could be ingested into the agent context.\n
Ingestion points: Multimodal URLs defined in references/agents.md and results from search tools (Exa, Tavily, DuckDuckGo) or MCP tools in references/tools.md.\n
Boundary markers: The documentation in references/ui.md notes the existence of sanitize_messages within UI adapters.\n
Capability inventory: The framework supports subprocess creation (MCP), network operations (httpx), and vector search integrations.\n
Sanitization: UIAdapter.sanitize_messages is documented as a server-controlled policy to strip certain client-submitted flags.\n- [NO_CODE]: This skill is composed entirely of Markdown reference documentation and does not contain any executable code or configuration scripts.

pydantic-ai