The Agent Skills Directory

COMMAND_EXECUTION (LOW): The skill provides examples such as searchFiles and gitStatus that utilize bun's shell ($) to execute system commands like find and git. While Bun provides automatic escaping for variables, shell execution is a high-privilege capability. This is considered LOW severity as it is the primary purpose of an SDK development skill.
EXTERNAL_DOWNLOADS (LOW): The fetchUrl tool example demonstrates the use of the fetch API to make network requests to arbitrary URLs. This can be used to download external content into the agent's context.
DATA_EXFILTRATION (LOW): Examples like contextInfo and projectInfo provide access to session metadata and local project structures. If combined with the network capabilities of fetchUrl, this could potentially be used for data exfiltration.
PROMPT_INJECTION (LOW): The skill is vulnerable to Indirect Prompt Injection (Category 8c) through the fetchUrl tool.
Ingestion points: External data is ingested via the fetch call in examples/basic-tool.ts.
Boundary markers: The example lacks explicit boundary markers or instructions to ignore embedded prompts in the fetched content.
Capability inventory: The skill has access to shell execution (bun shell) and file system listing.
Sanitization: No sanitization of the fetched URL body is performed before returning it to the agent.

OpenCode SDK Development