#pragma once // Simulator: runs a Scenario between a SimulatedUser and an Agent, with // tools answered by a VirtualToolbox, then scores task success and safety // over the whole conversation with the Autorater. All loops are capped // (turns, tool rounds per turn, tool calls per session). #include #include #include "google/optimize/agent.h" #include "google/optimize/autorater.h" #include "google/optimize/caps.h" #include "google/optimize/scenario.h" #include "google/optimize/simulated_user.h" #include "google/optimize/virtual_tools.h" namespace molo::google::optimize { struct SimulationConfig { int max_turns = OptimizeCaps::kDefaultSimTurns; // 1..kMaxSimTurns void validate() const; }; struct SimulationResult { Conversation conversation; // "user_done" | "max_turns" | "tool_round_cap" | "tool_call_cap" std::string stop_reason; int tool_calls = 0; int unknown_tool_calls = 0; }; struct SimulationReport { Scenario scenario; SimulationResult run; Score task_success; Score safety; bool pass() const { return task_success.pass && safety.pass; } clients::body::Json to_json() const; }; class Simulator { public: // user_model plays the user; judge_model scores (may be the same object). Simulator(std::shared_ptr user_model, std::shared_ptr judge_model); SimulationResult run(const Scenario& scenario, Agent& agent, VirtualToolbox& tools, const SimulationConfig& config = {}) const; SimulationReport run_and_score(const Scenario& scenario, Agent& agent, VirtualToolbox& tools, const SimulationConfig& config = {}) const; private: SimulatedUser user_; Autorater judge_; }; } // namespace molo::google::optimize