code/scripts/aat_upstream_openai_runner.py

"""Upstream AssetOpsBench OpenAIAgentRunner smoke wrapper.

This is intentionally separate from ``scripts/aat_runner.py``. The production
Experiment 1 Cell A/B path uses the team-local OpenAI Agents SDK wrapper so the
direct and MCP arms share the same agent loop. This script is only for the
parity smoke required by #104: instantiate AssetOpsBench's upstream
``OpenAIAgentRunner`` Python API with this repo's Smart Grid MCP server paths,
then translate the result into the benchmark artifact schema.

The upstream ``openai-agent`` CLI does not expose a server-path override, so the
Python API is the narrowest way to run the upstream runner against Smart Grid.
The Python API still assumes AOB's own MCP server entry points; for this smoke
we patch its MCP server factory to launch this repo's Smart Grid servers through
the warmed Insomnia Python/bootstrap path used by the benchmark runner.
"""

from __future__ import annotations

import argparse
import asyncio
import importlib.util
import json
import logging
import os
import shlex
import sys
import time
import types
from pathlib import Path
from typing import Any

_REPO_ROOT = Path(__file__).resolve().parent.parent
if str(_REPO_ROOT) not in sys.path:
    sys.path.insert(0, str(_REPO_ROOT))

_LOG = logging.getLogger("aat_upstream_openai_runner")


SERVER_PATHS = {
    "iot": "mcp_servers/iot_server/server.py",
    "fmsr": "mcp_servers/fmsr_server/server.py",
    "tsfm": "mcp_servers/tsfm_server/server.py",
    "wo": "mcp_servers/wo_server/server.py",
}


def build_parser() -> argparse.ArgumentParser:
    parser = argparse.ArgumentParser(
        prog="aat_upstream_openai_runner",
        description="Run AOB's upstream OpenAIAgentRunner against Smart Grid MCP servers.",
    )
    parser.add_argument("--prompt", required=True)
    parser.add_argument("--output", required=True)
    parser.add_argument("--model-id", required=True)
    parser.add_argument("--aob-path", required=True)
    parser.add_argument("--max-turns", type=int, default=30)
    parser.add_argument("--verbose", action="store_true")
    return parser


def _setup_logging(verbose: bool) -> None:
    logging.basicConfig(
        level=logging.DEBUG if verbose else logging.INFO,
        format="%(asctime)s %(levelname)s %(name)s: %(message)s",
    )


def _bootstrap_aob(aob_path: Path) -> None:
    src = aob_path / "src"
    if not src.exists():
        raise FileNotFoundError(f"AssetOpsBench src/ not found under {aob_path}")
    if str(src) not in sys.path:
        sys.path.insert(0, str(src))
    _install_unused_claude_sdk_stub()


def _install_unused_claude_sdk_stub() -> None:
    """Avoid AOB's top-level Claude runner import for this OpenAI-only smoke.

    ``agent.openai_agent.runner`` is the only upstream surface used here, but
    importing it normally first executes ``agent.__init__``, which imports the
    Claude runner and requires ``claude_agent_sdk``. That SDK is irrelevant to
    the OpenAI parity smoke and is not available in the Insomnia AaT runtime.
    """
    if "claude_agent_sdk" in sys.modules:
        return
    if importlib.util.find_spec("claude_agent_sdk") is not None:
        _LOG.warning(
            "real claude_agent_sdk package is installed; not installing the "
            "OpenAI parity smoke import shim"
        )
        return
    stub = types.ModuleType("claude_agent_sdk")
    for name in (
        "AssistantMessage",
        "ClaudeAgentOptions",
        "HookMatcher",
        "ResultMessage",
        "TextBlock",
        "ToolUseBlock",
    ):
        setattr(stub, name, type(name, (), {}))

    async def _query(*_args: object, **_kwargs: object) -> None:
        raise RuntimeError(
            "claude_agent_sdk stub is unavailable in OpenAI parity smoke"
        )

    stub.query = _query
    sys.modules["claude_agent_sdk"] = stub


def _smartgrid_server_paths(repo_root: Path) -> dict[str, Path]:
    paths: dict[str, Path] = {}
    for name, rel in SERVER_PATHS.items():
        path = repo_root / rel
        if not path.exists():
            raise FileNotFoundError(f"Smart Grid MCP server missing: {path}")
        paths[name] = path
    return paths


def _parse_parallel_tool_calls() -> bool | None:
    raw = os.environ.get("AAT_PARALLEL_TOOL_CALLS", "false").strip().lower()
    if raw in {"", "false", "0", "no", "off"}:
        return False
    if raw in {"true", "1", "yes", "on"}:
        return True
    if raw in {"auto", "default", "none"}:
        return None
    raise ValueError(
        "AAT_PARALLEL_TOOL_CALLS must be true, false, or auto; " f"got {raw!r}"
    )


def _is_watsonx_model(model_id: str) -> bool:
    return model_id.strip().lower().startswith("watsonx/")


def _configure_litellm_provider_compat(model_id: str) -> None:
    if not _is_watsonx_model(model_id):
        return
    try:
        import litellm  # type: ignore
    except ImportError:
        return
    litellm.drop_params = True


def _patch_aob_openai_runner(
    aob_openai_runner: Any,
    repo_root: Path,
    model_id: str = "",
) -> list[str]:
    """Patch AOB runner dependencies while leaving OpenAIAgentRunner.run intact."""
    from scripts.aat_system_prompt import AOB_SOURCE_SHA

    patches: list[str] = []

    build_mcp_servers = getattr(aob_openai_runner, "_build_mcp_servers", None)
    if not callable(build_mcp_servers):
        raise RuntimeError(
            "AssetOpsBench OpenAIAgentRunner patch precondition failed: "
            "agent.openai_agent.runner._build_mcp_servers is missing or not "
            f"callable at expected AOB source SHA {AOB_SOURCE_SHA}. "
            "Refusing to run parity smoke with the default upstream MCP launcher."
        )
    if not callable(getattr(aob_openai_runner, "Agent", None)):
        raise RuntimeError(
            "AssetOpsBench OpenAIAgentRunner patch precondition failed: "
            "agent.openai_agent.runner.Agent is missing or not callable at "
            f"expected AOB source SHA {AOB_SOURCE_SHA}. Refusing to run parity "
            "smoke without the local-vLLM parallel_tool_calls setting."
        )

    from agents import Agent as SDKAgent, ModelSettings
    from agents.mcp import MCPServerStdio
    from scripts.aat_tools_mcp import _client_timeout_seconds, _server_params

    def _build_smartgrid_mcp_servers(
        server_paths: dict[str, Path | str],
    ) -> list[MCPServerStdio]:
        client_timeout = _client_timeout_seconds()
        servers: list[MCPServerStdio] = []
        for name, spec in server_paths.items():
            path = Path(spec)
            if not path.is_absolute():
                path = repo_root / path
            params = _server_params(repo_root, path)
            command_line = [
                str(params["command"]),
                *[str(arg) for arg in params["args"]],
            ]
            print(
                "Upstream parity MCP server "
                f"{name}: timeout={client_timeout:g}s {shlex.join(command_line)}",
                file=sys.stderr,
            )
            servers.append(
                MCPServerStdio(
                    name=name,
                    params=params,
                    cache_tools_list=True,
                    client_session_timeout_seconds=client_timeout,
                )
            )
        return servers

    parallel_tool_calls = _parse_parallel_tool_calls()
    if _is_watsonx_model(model_id) and parallel_tool_calls is True:
        raise ValueError(
            "WatsonX does not support parallel_tool_calls; set "
            "AAT_PARALLEL_TOOL_CALLS=false or auto for hosted WatsonX runs."
        )
    _configure_litellm_provider_compat(model_id)
    effective_parallel_tool_calls = parallel_tool_calls
    if _is_watsonx_model(model_id) and effective_parallel_tool_calls is False:
        effective_parallel_tool_calls = None

    def _agent_with_model_settings(*args: Any, **kwargs: Any) -> Any:
        if effective_parallel_tool_calls is not None:
            kwargs.setdefault(
                "model_settings",
                ModelSettings(parallel_tool_calls=parallel_tool_calls),
            )
        return SDKAgent(*args, **kwargs)

    aob_openai_runner._build_mcp_servers = _build_smartgrid_mcp_servers
    patches.append("mcp_server_launch")

    aob_openai_runner.Agent = _agent_with_model_settings
    patches.append(f"parallel_tool_calls={parallel_tool_calls}")
    if _is_watsonx_model(model_id):
        patches.append("watsonx_drop_unsupported_params")

    return patches


def _tool_call_payload(tool_call: Any) -> dict[str, Any]:
    return {
        "name": getattr(tool_call, "name", "") or "",
        "arguments": getattr(tool_call, "input", {}) or {},
        "call_id": getattr(tool_call, "id", "") or "",
        "output": getattr(tool_call, "output", None),
    }


def _serialize_result(
    *,
    args: argparse.Namespace,
    prompt: str,
    result: Any,
    duration_seconds: float,
    server_paths: dict[str, Path],
    patches: list[str],
) -> dict[str, Any]:
    trajectory = getattr(result, "trajectory", None)
    turns = list(getattr(trajectory, "turns", []) or [])
    history = []
    tool_call_count = 0

    for index, turn in enumerate(turns, start=1):
        tool_calls = [
            _tool_call_payload(call)
            for call in list(getattr(turn, "tool_calls", []) or [])
        ]
        tool_call_count += len(tool_calls)
        history.append(
            {
                "turn": index,
                "role": "assistant",
                "content": getattr(turn, "text", "") or "",
                "tool_calls": tool_calls,
                "input_tokens": getattr(turn, "input_tokens", 0) or 0,
                "output_tokens": getattr(turn, "output_tokens", 0) or 0,
            }
        )

    answer = getattr(result, "answer", "") or ""
    max_turns_reached = bool(
        getattr(result, "max_turns_reached", False)
        or getattr(result, "max_turns_exhausted", False)
    )
    return {
        "question": prompt,
        "answer": answer,
        "success": (not max_turns_reached) and bool(answer),
        "failed_tools": [],
        "max_turns_exhausted": max_turns_reached,
        "turn_count": len(history),
        "tool_call_count": tool_call_count,
        "history": history,
        "runner_meta": {
            "runner": "AssetOpsBench OpenAIAgentRunner",
            "runner_source": "upstream_python_api",
            "aob_path": str(Path(args.aob_path).resolve()),
            "model_id": args.model_id,
            "max_turns": args.max_turns,
            "server_paths": {name: str(path) for name, path in server_paths.items()},
            "import_shims": ["claude_agent_sdk"],
            "aob_runner_patches": patches,
            "duration_seconds": duration_seconds,
        },
    }


def _write_output(path: Path, payload: dict[str, Any]) -> None:
    path.parent.mkdir(parents=True, exist_ok=True)
    path.write_text(json.dumps(payload, indent=2) + "\n", encoding="utf-8")


async def _main(args: argparse.Namespace) -> int:
    repo_root = Path(__file__).resolve().parent.parent
    aob_path = Path(args.aob_path).resolve()
    server_paths = _smartgrid_server_paths(repo_root)
    _bootstrap_aob(aob_path)

    from agent.openai_agent import runner as aob_openai_runner

    patches = _patch_aob_openai_runner(aob_openai_runner, repo_root, args.model_id)
    OpenAIAgentRunner = aob_openai_runner.OpenAIAgentRunner
    runner = OpenAIAgentRunner(
        server_paths=server_paths,
        model=args.model_id,
        max_turns=args.max_turns,
    )

    start = time.time()
    try:
        result = await runner.run(args.prompt)
    except Exception as exc:
        _LOG.exception("upstream OpenAIAgentRunner failed: %s", exc)
        _write_output(
            Path(args.output),
            {
                "question": args.prompt,
                "answer": "",
                "success": False,
                "error": f"{type(exc).__name__}: {exc}",
                "failed_tools": [],
                "max_turns_exhausted": False,
                "turn_count": 0,
                "tool_call_count": 0,
                "history": [],
                "runner_meta": {
                    "runner": "AssetOpsBench OpenAIAgentRunner",
                    "runner_source": "upstream_python_api",
                    "aob_path": str(aob_path),
                    "model_id": args.model_id,
                    "max_turns": args.max_turns,
                    "server_paths": {
                        name: str(path) for name, path in server_paths.items()
                    },
                    "import_shims": ["claude_agent_sdk"],
                    "aob_runner_patches": patches,
                    "duration_seconds": time.time() - start,
                },
            },
        )
        return 1

    payload = _serialize_result(
        args=args,
        prompt=args.prompt,
        result=result,
        duration_seconds=time.time() - start,
        server_paths=server_paths,
        patches=patches,
    )
    _write_output(Path(args.output), payload)
    return 0 if payload["success"] else 1


def main() -> None:
    args = build_parser().parse_args()
    _setup_logging(args.verbose)
    sys.exit(asyncio.run(_main(args)))


if __name__ == "__main__":
    main()