The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to use a specific set of command-line tools (lp task, lp outcome, lp registry) to manage task states, query status, and record costs. These operations are restricted to a specialized internal toolset and serve the primary purpose of the skill.
[INDIRECT_PROMPT_INJECTION]: The skill requires the agent to read and process diagnostic data from failed tasks, including exploration logs and task descriptions, to perform attribution and decide on the next steps. This introduces a surface where malicious content embedded in the failed task's context could influence the agent during the takeover process.
Ingestion points: Output from the lp task show command and diagnostic files located in .livepowers/explorations/.
Boundary markers: None explicitly defined in the instructions to separate untrusted task data from the agent's analytical logic.
Capability inventory: The agent is authorized to execute lp CLI commands for task state transitions (move, reset, close) and record financial/resource outcomes.
Sanitization: No specific sanitization or filtering of task diagnostics is mentioned before they are processed for the diagnostic package.

takeover-handling