The Agent Skills Directory

[COMMAND_EXECUTION]: The utility script scripts/export_config.sh uses the eval command to process key-value pairs extracted from the local configuration file ~/.bytehouse_config.json. While the values are escaped using the jq @sh filter, the use of eval on file contents presents a potential risk if the configuration file is modified by an untrusted process.
[INDIRECT_PROMPT_INJECTION]: The skill constructs SQL queries by interpolating potentially untrusted document content, titles, and metadata directly into string templates in scripts/search_client.py.
Ingestion points: Document content, URLs, and metadata are ingested via the insert_document and insert_batch_documents methods.
Boundary markers: The system does not utilize explicit delimiters or specialized prompt instructions to isolate untrusted data within the SQL strings.
Capability inventory: The client has the ability to execute DDL (CREATE), DML (INSERT), and DQL (SELECT) commands against the ByteHouse database instance.
Sanitization: The skill applies basic character escaping by replacing single quotes (.replace("'", "''")), but it does not use parameterized queries, which are the standard defense against SQL injection.

byted-bytehouse-multimodal-search