The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to handle new or rare user requests by reasoning through the environment and generating SQL or tool orchestrations. This creates an attack surface for indirect prompt injection via user-supplied data or external database results. Any untrusted data processed here could potentially influence the agent's logic or subsequent tool use.
Ingestion points: User instructions and results returned from dynamic SQL queries.
Boundary markers: The skill defines operational boundaries (ontology scope) and resource budgets, but does not specify technical delimiters for separating instructions from data in the prompts.
Capability inventory: Execution of SQL (OLAP/OLTP), execution of the lp CLI tool, and writing documentation files to the .livepowers/explorations/ directory.
Sanitization: The instructions mandate the use of oltp-action-safety for write operations and suggest small-sample checks with LIMIT for read operations to prevent unintended data volume or side effects.
[COMMAND_EXECUTION]: The agent is instructed to interact with a specific CLI utility named lp (Livepowers) to manage tasks (lp task), record evidence (lp evidence), and manage intent aliases (lp intents). These commands are used to track costs, outcomes, and metadata of the exploration process.
[DYNAMIC_EXECUTION]: The skill involves the dynamic generation and execution of SQL queries based on reasoning and trial-and-error. It emphasizes a defensive approach, starting with small samples and verifying results against known values before full execution.

explore-with-evidence