The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill documents an experimental 'Code Mode' feature that generates executable JavaScript at runtime based on LLM output to orchestrate multi-tool workflows.
Evidence: references/codemode.md describes the generation of an executeTask() function that orchestrates tools like getWeather and sendEmail, which is then executed via the CodeModeProxy binding.
[INDIRECT_PROMPT_INJECTION]: The skill describes processing external, untrusted data that is subsequently passed to an LLM without explicit boundary markers, creating a vulnerability surface.
Ingestion points: Data enters the system via onEmail (raw email content) in references/email.md and this.mcp.getAITools() (external MCP server tool definitions) in references/codemode.md.
Boundary markers: Code examples show direct interpolation of message history and tool results into model prompts without specific delimiters or instructions to ignore embedded commands.
Capability inventory: The agent context includes capabilities for database operations via this.sql, state modification with this.setState, network requests via fetch, and runtime script execution through Code Mode.
Sanitization: The skill demonstrates the use of a globalOutbound fetch handler in references/codemode.md to implement a hostname blocklist for security filtering.
[COMMAND_EXECUTION]: The documentation provides standard shell commands for environment setup and package verification.
Evidence: SKILL.md and references/mcp.md include commands such as npm install agents, npm ls agents, and npm install @modelcontextprotocol/sdk zod.
[EXTERNAL_DOWNLOADS]: The skill fetches configuration and documentation from trusted external sources and uses common package registries.
Evidence: The skill references documentation and source code hosted at github.com/cloudflare/agents and official Cloudflare developer documentation at developers.cloudflare.com.

agents-sdk