The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion and processing of untrusted data from users and external tools, creating a potential surface for indirect prompt injection.
Ingestion points: The skill processes messages, commands, and state objects provided via POST requests in route handlers.
Boundary markers: Data is partitioned into structured chunks (e.g., text-delta, tool-call, data), providing structural boundaries, though it lacks explicit instructions to ignore embedded commands in the text parts.
Capability inventory: The skill allows writing to stream responses, interacting with Redis/ioredis for state persistence, and executing tool calls (e.g., get_weather).
Sanitization: Documentation encourages the use of zod for parameter validation in client-side tools.
[EXTERNAL_DOWNLOADS]: The skill references and downloads dependencies from official and trusted sources.
Fetches the assistant-stream package from standard registries (NPM and PyPI).
References official documentation and assets hosted on assistant-ui.com and GitHub repositories under the assistant-ui and vercel-labs organizations.
[COMMAND_EXECUTION]: The resumable stream feature utilizes Redis EVAL commands to run Lua scripts for atomic state finalization. This is a standard implementation pattern for distributed synchronization in storage adapters.

streaming