The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The scripts/ppq-image.ts script programmatically accesses the user's local application data directory (defaulting to $XDG_DATA_HOME/opencode) to extract API keys from opencode.db (an SQLite database) and auth.json. It specifically queries for credentials where the integration_id is 'ppq'.
[DATA_EXFILTRATION]: Extracted credentials and local image data are transmitted to the remote host api.ppq.ai. This behavior, while functional for the skill, moves sensitive system information and private data to an external service not explicitly verified as part of the vendor's infrastructure.
[COMMAND_EXECUTION]: The script utilizes execFileSync to invoke the identify command from the ImageMagick suite to verify image dimensions for processed files.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection via external image data and generative prompts.
Ingestion points: Untrusted image data enters via local files or remote URLs, and prompts are accepted from user input in scripts/ppq-image.ts.
Boundary markers: The skill instructions in SKILL.md require the agent to ask for tool selection and budget approval before performing operations, which provides a layer of human oversight.
Capability inventory: The skill can perform network operations (fetch), local command execution (identify), and file system access (readFile, writeFile).
Sanitization: The script includes logic to redact API keys from error messages and uses standard Node.js APIs for data handling.
[EXTERNAL_DOWNLOADS]: The script performs network requests to api.ppq.ai for model discovery and processing, and it is designed to fetch image data from remote URLs provided during execution.

jednadvacet-image-style