The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The library relies on standard Node.js packages including reachat, reablocks, zod, and optional peer dependencies like reaviz. These are installed via standard package managers and originate from the vendor or well-known open-source ecosystems.
[INDIRECT_PROMPT_INJECTION]: The Markdown component documentation indicates that rehypeRaw is enabled by default in the rendering pipeline. This allows the rendering of raw HTML strings, which serves as a potential attack surface for Indirect Prompt Injection (XSS) if the data provided by the agent or user is not properly sanitized before being passed to the component.
Ingestion points: MessageResponse and MessageQuestion components ingest and render external markdown strings.
Boundary markers: The library provides a systemPrompt generator within the ComponentCatalog to provide instructions to the LLM on how to format component specifications.
Capability inventory: The ComponentCatalog enables interactive components to trigger the sendMessage callback to communicate back to the agent. ChatInput facilitates file uploads.
Sanitization: The library includes remarkRedact for PII filtering (SSN, credit cards, bitcoin addresses), though the use of rehypeRaw explicitly permits HTML parsing.
[DYNAMIC_EXECUTION]: The ComponentCatalog feature implements a templated dynamic execution model where an LLM provides JSON specifications in fenced code blocks to render React components. These components are statically defined by the developer, and the library validates the incoming props against Zod schemas before rendering, mitigating arbitrary code execution risks.

reachat