The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The script scripts/aiicons.py fetches brand logos and icons from established content delivery networks including unpkg.com and simpleicons.org. These requests are used to retrieve SVG assets for diagram rendering.
[COMMAND_EXECUTION]: Extensive use of subprocess.run across many scripts to invoke system tools. Notable invocations include drawio for exporting diagrams, dot (Graphviz) for layout calculation, git for history analysis in scripts/timelapse.py and scripts/prdiff.py, and tred for graph reduction.
[DYNAMIC_EXECUTION]: Several scripts dynamically load and execute sibling Python files.
importlib.util is used in scripts/autolayout.py, scripts/c4.py, scripts/edgeports.py, scripts/tfimports.py, and scripts/tfstate.py to import functional logic from other files in the skill directory.
subprocess.run with sys.executable is used in scripts/compress.py, scripts/prdiff.py, scripts/raster2drawio.py, and scripts/timelapse.py to run bundled scripts as independent processes.
[INDIRECT_PROMPT_INJECTION]: The skill processes various untrusted external data sources which could serve as injection vectors if the agent is subsequently asked to reason about the generated diagrams.
Ingestion points: Reads user-provided source code (Python, JS, Go, Rust), infrastructure configurations (Terraform, Kubernetes, Docker Compose), SQL schemas, and runtime tool outputs (docker inspect, terraform show, kubectl get, git log).
Boundary markers: Boundary markers or explicit warnings to ignore instructions embedded in the processed data are absent.
Capability inventory: Almost all scripts have the capability to execute shell commands (subprocess.run) or write files, providing a significant capability surface if an injection were successful.
Sanitization: While the scripts perform XML and HTML entity escaping for label rendering, they do not specifically filter for natural language instructions embedded within the technical metadata being visualized.

drawio-skill