The Agent Skills Directory

[COMMAND_EXECUTION]: The inventory scripts execute local Git commands to identify source files within the repository being analyzed.- Evidence: Calls to subprocess.run(['git', 'ls-files', ...]) in the Python-based inventory scripts (languages/python/inventory.py, languages/rust/inventory.py, languages/dart/inventory.py) and execSync('git ls-files ...') in the Node.js script (languages/typescript/inventory.mjs).- [DYNAMIC_EXECUTION]: The TypeScript inventory tool dynamically loads the typescript package from the target repository's local environment to perform AST analysis.- Evidence: languages/typescript/inventory.mjs uses createRequire to load the typescript library from the current working directory's package.json location.- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted source code from the repository to generate refactoring instructions for other AI models, creating a potential surface for instruction injection.- Ingestion points: All inventory scripts read raw source file content (.ts, .py, .rs, .dart) from the local filesystem.- Boundary markers: The inventory output is structured into Markdown tables and lists, but it lacks specific safety delimiters to isolate code-derived text from the agent's instructions.- Capability inventory: The executor model (Qwen3.6-27B) described in the program templates is expected to have repository-level editing capabilities including shell access.- Sanitization: The scripts do not sanitize or filter content extracted from source code (such as docstrings, comments, or variable names) before it is included in the generated Markdown refactor specifications.

decision-free-specs