| |
| """Run one RetroEnv episode against a server and print the trajectory. |
| |
| python rollout.py --server http://127.0.0.1:8000 --split eval --index 0 \\ |
| --provider anthropic --model claude-opus-5-5 |
| |
| Providers: anthropic (Messages API), openai (Responses API), or any |
| OpenAI-compatible chat endpoint via --endpoint (vLLM, HF router, OpenRouter). |
| For many tasks, use eval/run_eval.py in the project folder. |
| """ |
|
|
| from __future__ import annotations |
|
|
| import argparse |
| import json |
| import os |
|
|
| from retroenv_openenv import agent, agent_anthropic, agent_responses |
| from retroenv_openenv.client import RetroEnvClient |
|
|
|
|
| def main() -> int: |
| parser = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter) |
| parser.add_argument("--server", default=os.getenv("RETROENV_SERVER", "http://127.0.0.1:8000")) |
| parser.add_argument("--split", default="eval") |
| parser.add_argument("--index", type=int, default=0) |
| parser.add_argument("--provider", choices=("anthropic", "openai", "chat"), default="anthropic") |
| parser.add_argument("--model", default="claude-opus-5-5") |
| parser.add_argument("--endpoint", help="OpenAI-compatible base URL for --provider chat") |
| parser.add_argument("--api-key-env", default=None) |
| parser.add_argument("--max-turns", type=int, default=int(os.getenv("MAX_TURNS", "16"))) |
| args = parser.parse_args() |
|
|
| with RetroEnvClient(args.server) as env: |
| opening = env.reset(args.split, index=args.index) |
| print(f"task {opening['task_id']} ({args.split}[{args.index}]) target {opening['target_smiles']}") |
| print(f"tools: {', '.join(opening['tools'])}\n") |
| if args.provider == "anthropic": |
| import anthropic |
|
|
| llm = anthropic.Anthropic(api_key=os.getenv(args.api_key_env or "ANTHROPIC_API_KEY")) |
| result = agent_anthropic.run_episode( |
| llm, env, opening, agent_anthropic.ClaudeConfig(model=args.model, max_turns=args.max_turns) |
| ) |
| else: |
| from openai import OpenAI |
|
|
| if args.provider == "openai": |
| llm = OpenAI(api_key=os.getenv(args.api_key_env or "OPENAI_API_KEY")) |
| result = agent_responses.run_episode( |
| llm, env, opening, agent_responses.ResponsesConfig(model=args.model, max_turns=args.max_turns) |
| ) |
| else: |
| if not args.endpoint: |
| parser.error("--endpoint is required for --provider chat") |
| llm = OpenAI(base_url=args.endpoint, api_key=os.getenv(args.api_key_env or "OPENAI_API_KEY", "unused")) |
| result = agent.run_episode( |
| llm, env, opening, agent.AgentConfig(model=args.model, max_turns=args.max_turns) |
| ) |
|
|
| for message in result["transcript"]: |
| if message["role"] == "tool": |
| print(f" -> {message['content'][:200]}") |
| continue |
| content = message.get("content") |
| calls = message.get("tool_calls") or [ |
| block |
| for block in (content if isinstance(content, list) else []) |
| if block.get("type") in ("tool_use", "function_call") |
| ] |
| for call in calls: |
| name = call.get("name") or call.get("function", {}).get("name") |
| arguments = call.get("input") or call.get("arguments") or call.get("function", {}).get("arguments") |
| print(f" {name}({json.dumps(arguments)[:160] if not isinstance(arguments, str) else arguments[:160]})") |
| if isinstance(content, list): |
| for block in content: |
| if block.get("type") == "tool_result": |
| print(f" -> {str(block.get('content'))[:200]}") |
| print( |
| f"\nreward {result['reward']:.3f} valid {result['valid']} exact {result['exact_match']} " |
| f"turns {result['turns']} tool calls {result['tool_calls']}" |
| ) |
| if result["hard_failures"]: |
| print("failures: " + "; ".join(result["hard_failures"][:3])) |
| return 0 |
|
|
|
|
| if __name__ == "__main__": |
| raise SystemExit(main()) |
|
|
