The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes several local utility scripts written in Bash, Python, and Node.js (shoot_atlas.sh, build_standalone.py, validate_atlas_data.js, etc.) to automate the generation, validation, and rendering of architecture diagrams. It also invokes external tools like Chrome (in headless mode) and the Python built-in HTTP server to capture screenshots of the generated atlas. These operations are restricted to the local environment and the project directory.
[DYNAMIC_EXECUTION]: The validation scripts (validate_atlas_data.js and lint_atlas_layout.js) use the Node.js vm module to execute the ATLAS data object defined in the generated data.js file. This execution is performed within an isolated sandbox with an empty context, serving only to verify the schema and geometric constraints of the agent's output before delivery.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process arbitrary codebases to extract architectural patterns, which inherently involves ingesting untrusted data.
Ingestion points: The agent reads source code, build configuration files (e.g., package.json, Cargo.toml), and documentation during the parallel exploration phase.
Boundary markers: The instructions in SKILL.md and ontology-modeling.md enforce a strict reporting structure, requiring specific fields like entities, relations, and evidence, which helps maintain agent focus.
Capability inventory: The skill can write to the local file system, execute validation scripts, and run a local web server for rendering.
Sanitization: The skill implements a multi-step verification process, including mandatory grep checks by the agent to confirm evidence and automated schema validation of the final data structure.

nan-codebase-architecture-atlas