The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill handles untrusted user input which is subsequently interpolated into prompts sent to external LLM providers, creating a potential surface for indirect injection.
Ingestion points: The user_question and query variables in functions such as call_with_prompt_cache, smart_llm_call, and semantic_cache_lookup within SKILL.md accept external data.
Boundary markers: There are no implementation-level delimiters or instructions provided to the agent to treat the cached content or user queries as untrusted data.
Capability inventory: The skill utilizes network-enabled libraries (openai, anthropic, qdrant-client) to communicate with remote vector databases and model APIs.
Sanitization: The provided code snippets lack input validation or sanitization mechanisms to filter malicious instructions embedded in the processed queries.

llm-caching