Built-in tool for asking a human a question mid-execution.

Configuration

Configure via tool_config/1 in your agent:

def tool_config(Legion.Tools.HumanTool) do
  [handler: MyApp.ChatHandler, timeout: 30_000]
end

Options:

• :handler (required) — a pid or registered name that receives {:human_request, ref, from_pid, question, meta} and must send {:human_response, ref, answer} back to from_pid.
• :timeout — milliseconds to wait for a response. Defaults to :infinity.

ask/1 blocks inside the code evaluation, so the agent's sandbox_timeout (60 seconds by default) still applies and kills the eval regardless of this setting. Raise it, or set it to :infinity, for agents that wait on humans.

Usage (for LLM agent)

HumanTool.ask("What format do you prefer?")