The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/impact.py dynamically loads the tree-sitting engine by searching through a list of candidate directories and prepending the discovered path to sys.path. This enables the skill to import the CodeCache class from a peer skill at runtime.
[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests and processes untrusted data from the local repository being analyzed.
Ingestion points: The find_text_refs and parse_features_file functions in scripts/impact.py perform recursive file reads on the target repository.
Boundary markers: The report generated for the LLM includes extracted source code and documentation snippets without distinct boundary markers or delimiters to separate data from instructions.
Capability inventory: The script performs extensive file system reads and outputs the results to the standard output stream for LLM consumption.
Sanitization: While the script uses errors='replace' during file reading and slices text snippets to 140 characters, it lacks sanitization or escaping mechanisms to prevent embedded instructions from influencing the agent's behavior.
[COMMAND_EXECUTION]: The SKILL.md setup instructions include shell commands to create a Python virtual environment and install the tree-sitter dependency.
[EXTERNAL_DOWNLOADS]: The skill uses the uv package manager to install the tree-sitter library from the official Python Package Index.

assessing-impact