The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides the mtd.py run tool, which is designed to execute arbitrary shell commands provided by the user or agent to run experiment scripts and record provenance.
[PERSISTENCE]: The skill can modify the agent's configuration file (.claude/settings.json) to install lifecycle hooks, enabling automatic execution of status checks during session start, file edits, and session termination.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input from project files and injects summaries into the agent context via hooks. Ingestion points: Reads files in problem/ (PDF/DOCX/MD), results/ (JSON), and paper/ (TeX/MD) directories. Boundary markers: None used in status report summaries. Capability inventory: Subprocess execution for experiments and tool invocation (Poppler, LaTeX, Git) and extensive file system access. Sanitization: Schema validation for JSON files is performed, but there is no specific sanitization of natural language text extracted from documents.

mathodology-workbench