The Agent Skills Directory

Local File Read Capability: The scripts/civic_graphql.py script utilizes a query_path parameter that allows the agent to read the contents of local files using Path(query_path).read_text(). Without path validation or sandboxing, this pattern could be used to access sensitive information on the host system if the agent is directed to read non-intended files.
Local File Write Capability: The skill allows writing GraphQL response data to a location specified by raw_output_path. The script uses path.write_text() and creates parent directories as needed. This functionality could potentially be used to overwrite existing files or create files in sensitive directories if the provided path is not restricted to a safe temporary location.
Indirect Prompt Injection Surface: The skill is designed to fetch and summarize data from an external GraphQL API (civicdb.org). As it processes external content and presents it to the agent, there is a potential for the agent to be influenced by data contained within the API responses. The skill includes basic truncation logic (_compact) to limit the depth and size of the processed data, which serves as a helpful boundary.
Command Execution via Subprocess: The SKILL.md file provides an example of executing the Python script via a shell command (echo ... | python scripts/civic_graphql.py). This establishes a pattern where the agent interacts with the underlying system via standard input/output, which is a common but powerful capability that should be monitored for unexpected command construction.

civic-skill