The Agent Skills Directory

[COMMAND_EXECUTION]: The skill configures Claude Code to automatically execute shell commands through the claude-flow CLI whenever tools such as Write, Edit, or Bash are used. These commands interpolate parameters like ${tool.params.command} and ${tool.params.file_path} directly into shell strings without sanitization, creating a vector for command injection.
[EXTERNAL_DOWNLOADS]: The skill's setup instructions require the global installation of the unverified package claude-flow@alpha via npm, which introduces a supply chain risk from a non-trusted source.
[DATA_EXFILTRATION]: The skill's 'Memory Coordination' and 'Neural Training' features are designed to store and synchronize development data, including 'edit context', 'task decisions', and 'neural patterns'. This data is shared across agents and MCP servers (claude-flow, ruv-swarm, flow-nexus), potentially exposing sensitive logic or proprietary code to external components.
[PROMPT_INJECTION]: The skill ingests untrusted data from edited project files and task descriptions to drive automated behaviors like agent spawning and pattern training. • Ingestion points: File content processed in post-edit and task descriptions in pre-task. • Boundary markers: No explicit delimiters or instructions to ignore embedded prompts are provided in the hook configurations. • Capability inventory: The skill can execute shell commands via npx, perform file system operations, and interact with MCP tools. • Sanitization: No validation or escaping of the interpolated tool parameters is described before they are passed to the shell or coordination system.

Hooks Automation