The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary objective is to ingest large volumes of untrusted data from multiple repositories to create a context hub. This creates a significant ingestion surface for indirect instructions.
Ingestion points: scripts/scan_repo.py reads repository files including READMEs and manifests; scripts/build_knowledge_graph.py parses markdown tables and Mermaid diagrams from hub documents.
Boundary markers: The skill implements strict templates (e.g., assets/catalog-compiled-truth-template.md) that use horizontal rules and formal citation formats ([Source: ...]) to separate synthesized agent content from source evidence logs.
Capability inventory: The skill performs extensive file system read/write operations and utilizes subprocess for git commands and internal tool orchestration across multiple Python scripts.
Sanitization: Ingested data is normalized into structured JSON; reporting tools in scripts/export_graph_report.py use html.escape and regex-based sanitization for Mermaid diagram generation.
[COMMAND_EXECUTION]: Several scripts, including scripts/build_artifact_set.py, scripts/scan_portfolio.py, and scripts/incremental_update.py, utilize the subprocess module to execute tasks.
These calls are restricted to executing internal Python scripts within the skill or standard git commands for analysis and change detection. No arbitrary shell execution on user-provided strings was detected.
[DYNAMIC_EXECUTION]: The regression test suite in scripts/test_knowledge_graph_regressions.py uses importlib.util to dynamically load the skill's local Python modules.
This implementation is used exclusively for testing purposes and is restricted to loading known local script files.

dev-context-multi-repo