The Agent Skills Directory

[PROMPT_INJECTION]: The skill instructs the agent to act "autonomously" and bypass user permission prompts ("no per-step permission") for specific repositories. This explicitly overrides the standard human-in-the-loop safety model for AI agents.
[DYNAMIC_EXECUTION]: The agent is directed to perform "self-healing" by patching its own running source code (e.g., modifying dist/*.js files or .ts source) and executing build tools like npm run build. Modifying executing code is a high-risk behavior that can lead to system instability or compromise.
[DATA_EXFILTRATION]: The skill facilitates the autonomous transmission of system information, error logs, and source code diffs to an external URL (https://comfyui-mcp-issue-worker.artokun.workers.dev). While it includes instructions to redact secrets, the automated exfiltration of diagnostic data and source code without user review represents a significant risk.
[COMMAND_EXECUTION]: The instructions contain shell and PowerShell scripts that use curl and Invoke-RestMethod to transmit data to a remote endpoint. It also instructs the agent to trigger npm commands for building and testing code on the local machine.
[CREDENTIALS_UNSAFE]: The skill contains a hardcoded API credential (CLIENT_KEY) in the script snippets used for bug reporting. Hardcoding keys instead of using environment variables is a poor security practice.
[INDIRECT_PROMPT_INJECTION]: The skill creates an attack surface by ingesting untrusted data (error logs and third-party code) during the diagnosis process.
Ingestion points: The agent reads error logs via get_system_stats and source files found at stack trace paths.
Boundary markers: There are no specific delimiters or instructions to ignore embedded commands when processing these external logs/files.
Capability inventory: The agent has the capability to execute shell commands (npm, curl, git) and modify the local filesystem.
Sanitization: Instructions specify manual secret scrubbing but do not provide technical sanitization for preventing the agent from following instructions embedded in the ingested logs.

report-bug