The Agent Skills Directory

[SAFE]: The skill implementation follows security best practices for internal tool development. It utilizes a centralized WorkspaceService to handle file writes, which enforces path and extension restrictions (e.g., only allowing image extensions like .png). The use of environment variables for the image generation endpoint ensures that connectivity is managed through system configuration rather than hardcoded values. Furthermore, the skill implements a sensitive data redaction function (redactBase64) to prevent large image data strings from leaking into logs or terminal output during errors.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an ingestion surface as it processes natural language prompts and user messages to construct instructions for an external image generation engine. However, this is inherent to the skill's primary purpose. The implementation mitigates common risks by using a restricted workspace service for all persistent output, ensuring that even if the prompt logic is influenced, the resulting file operations remain within the designated project scope.
Ingestion points: Natural language input from toolCall.arguments.prompt and the user message ctx.message are used to assemble the final generation prompt.
Boundary markers: The skill does not use explicit delimiters to separate user input from system-generated character prompts, but uses structured logic to resolve subjects from a reference library.
Capability inventory: The skill can perform network requests to a configured image generation service and write image files to the workspace using writeFileBuffer.
Sanitization: File paths are validated against allowed extensions and the workspace root. Data URLs are redacted from error logs to prevent information leakage.

image-generation