The Agent Skills Directory

[PROMPT_INJECTION]: The skill facilitates building agents that ingest untrusted user data, creating a surface for indirect prompt injection.\n
Ingestion points: Untrusted data enters the agent context through agent.sendMessage() calls (seen in SKILL.md, references/agent-patterns.md, and references/react-integration.md).\n
Boundary markers: The instructions mention using system_prompt to define agent behavior, which can serve as a delimiter.\n
Capability inventory: The defined agents can perform web searches, execute code (via internal tools), and call external webhooks.\n
Sanitization: Examples do not explicitly demonstrate sanitization or escaping of the user-provided content before interpolation.\n- [REMOTE_CODE_EXECUTION]: Documentation examples for implementing a 'calculator' tool use the JavaScript eval() function to process mathematical expressions. This pattern allows for the execution of arbitrary logic assembled at runtime if the expression argument contains unsanitized input. Evidence of this pattern is present in references/agent-patterns.md and references/tool-builder.md.\n- [COMMAND_EXECUTION]: The skill provides shell commands for installing the SDK and adding other skills to the environment. Evidence includes npm install @inferencesh/sdk and npx skills add belt-sh/cli in SKILL.md.

javascript-sdk