The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill bundles multiple pre-compiled ELF shared libraries in the parsers/ directory (e.g., libtree_sitter_html.so, libtree_sitter_python.so). These are dynamically loaded at runtime using ctypes.CDLL within scripts/engine.py to provide language-specific parsing functionality.
[COMMAND_EXECUTION]: The test suite, specifically tests/test_cli_cache.py, utilizes subprocess.run to execute the CLI tool during integration testing. While primarily for development, this patterns allows for shell command execution.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and parse third-party codebases, which constitutes a significant attack surface. Malicious instructions could be embedded in code comments or docstrings within a repository being explored. When the agent retrieves these symbols or source snippets, it may inadvertently follow instructions contained within that untrusted data.
Ingestion points: Codebase files read into memory via CodeCache.scan and CodeCache.references in scripts/engine.py.
Boundary markers: Extracted symbol signatures and docstrings are returned to the agent without specific delimiters or isolation instructions.
Capability inventory: The agent can read files, search symbols, and find references across the filesystem root.
Sanitization: No filtering or sanitization is performed on extracted docstrings or comments before they are presented to the agent.

tree-sitting