The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill provides instructions in SKILL.md to install the agents package (npm install agents). This is a generic, unscoped package name that does not match the expected @cloudflare/agents scope used by other SDK components. The agents package on NPM is a legacy library unrelated to Cloudflare, posing a significant supply chain risk or potential installation of unintended code.
[PROMPT_INJECTION]: The skill documents an indirect prompt injection surface in references/email.md and references/mcp.md.
Ingestion points: The onEmail handler processes raw incoming email content, and the MCP integration ingests tool definitions and data from external servers.
Boundary markers: The provided implementation examples lack explicit delimiters (e.g., XML tags) or instructions to the LLM to ignore embedded commands in the untrusted data.
Capability inventory: The agent possesses capabilities to execute SQL queries (this.sql), manage schedules, and trigger durable workflows, which could be abused if the agent is influenced by malicious data.
Sanitization: There is no evidence of sanitization or validation of the external content before it is processed by the AI model.
[COMMAND_EXECUTION]: The skill documents the 'Code Mode' feature in references/codemode.md, which generates and executes JavaScript at runtime to orchestrate multiple tools. While this is presented as a library feature for optimization, it provides the agent with the capability to execute dynamic, LLM-generated logic within the Cloudflare Worker environment.
[SAFE]: The skill includes legitimate references to Cloudflare's official documentation and GitHub repositories (github.com/cloudflare/agents) for retrieval-augmented generation tasks.

agents-sdk