The Agent Skills Directory

[COMMAND_EXECUTION]: The calculatorTool example in SKILL.md implements the JavaScript eval() function to process input. If an agent uses this code as a template, it would allow for arbitrary code execution within the agent's runtime environment via malicious strings passed to the tool.
[PROMPT_INJECTION]: The skill documents the nextTurnParams feature, which allows tools to programmatically modify the agent's system instructions for subsequent turns. This creates a significant surface for indirect prompt injection.
Ingestion points: External data enters the context via model inputs and results from tools such as web_search (described in SKILL.md).
Boundary markers: There are no explicit markers or delimiters shown in the examples to isolate external tool outputs from the instructions.
Capability inventory: The skill enables network communication with the OpenRouter API and demonstrates local command execution capabilities via the eval example.
Sanitization: No sanitization or input validation is present in the nextTurnParams or eval implementation examples.
[EXTERNAL_DOWNLOADS]: The skill documents the installation of the @openrouter/sdk package from the NPM registry. This package is a recognized resource for the skill vendor.

openrouter-typescript-sdk