The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill mentions and facilitates the download of tools and dependencies from several unverified third-party GitHub accounts, including SpookyPirate, DeusData, headroomlabs-ai, obra, DietrichGebert, getagentseal, and morluto.
[REMOTE_CODE_EXECUTION]: Instructions in SKILL.md suggest using an external PowerShell installer (install.ps1) from the DeusData/codebase-memory-mcp repository. It also references Update-From-GitHub.ps1, a script intended for automated remote updates of the toolset.
[COMMAND_EXECUTION]: The skill enables and guides the agent to execute a variety of shell commands, including dotnet run, pip install, npx codeburn, and custom PowerShell scripts like Setup-HouseCarl.ps1, which can lead to system changes and installation of arbitrary software.
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: The discover_tools.ps1 script and references/reverse-engineering.md instructions involve reading and parsing local configuration files (INSTALLATION.json, kit.json) to determine executable paths.
Boundary markers: No specific boundary markers or instructions are provided to shield the agent from potentially adversarial content within these local configuration files.
Capability inventory: The skill allows full shell access via PowerShell and standard package managers (npm, pip, dotnet), creating a significant impact path if input data is poisoned.
Sanitization: The discovery process lacks sanitization or validation of the paths and parameters extracted from local JSON configuration files before they are used in command execution logic.

tool-discovery