The Agent Skills Directory

Dynamic Script Generation and Execution: The skill instructions direct the agent to write a traversal script (in Python or Bash) to the workspace and execute it. This script is used to perform a bottom-up traversal of the directory tree to automate repository mapping. While this is a standard automation pattern for security tools, it involves the runtime creation and execution of code within the agent's environment.
File System Modification: The skill is designed to write security-focused summaries (mantis-summary.md) directly into the source directories of the codebase. The skill includes a 'PINNED' mode that explicitly prevents writing when a read-only snapshot is detected, which serves as a security control. However, in other modes, it overwrites files in-place.
Indirect Prompt Injection Surface: The skill processes untrusted codebase content and historical security logs (historical_learnings.jsonl) to generate summaries via an LLM.
Ingestion points: Files are read from the codebase and the workspace/ directory.
Boundary markers: The instructions define a specific output schema (Core Components, API Endpoints, etc.) to structure the LLM's response.
Capability inventory: The skill can perform directory traversals, write scripts, and write markdown files to the repository.
Sanitization: No explicit sanitization of the source code content is performed before it is passed to the LLM for summarization.

mantis-summarize