The Agent Skills Directory

[DATA_EXFILTRATION]: The skill gathers system diagnostic information (OS version, Node.js version) and a list of installed skills (slugs and versions) to include in feedback reports sent to an external HTTPS endpoint. This risk is managed through a mandatory human-in-the-loop process: the agent is required to present the entire raw JSON payload to the user for review, and transmission only proceeds if the user provides explicit manual consent via a specific command-line flag.
[COMMAND_EXECUTION]: The script scripts/submit_feedback.py executes the command node --version using subprocess.run. This is a hardcoded, non-interactive diagnostic command used solely to provide environment context for bug reports.
[INDIRECT_PROMPT_INJECTION]: The skill processes metadata from other skills on the local filesystem (reading .claude/skills/*/origin.json and .dby/lock.json). While these ingestion points are protected by a strict attribute whitelist, they represent an external data source that enters the agent's context and is subsequently displayed to the user.
[SAFE]: The submission script includes a comprehensive secret scanner that uses regular expressions and Shannon entropy checks to inspect the entire report for potential credentials (e.g., AWS keys, GitHub tokens). The script will automatically abort the submission if a sensitive string is detected, ensuring that user secrets are not leaked during the feedback process.

dby-feedback