social-app/backend/src/core/agentscope/runtime/runner.py

from __future__ import annotations

from dataclasses import dataclass
from datetime import datetime, timezone
from typing import TYPE_CHECKING, Any
from uuid import UUID

from ag_ui.core.types import RunAgentInput
from agentscope.formatter import OpenAIChatFormatter
from agentscope.memory import InMemoryMemory
from agentscope.message import Msg
from agentscope.model import OpenAIChatModel
from core.agentscope.prompts.agent_prompt import build_worker_contract_prompt
from core.agentscope.prompts.system_prompt import build_system_prompt
from core.agentscope.runtime.json_react_agent import JsonReActAgent
from core.agentscope.runtime.model_tracking import TrackingChatModel
from core.agentscope.runtime.router_persistence import persist_router_message
from core.agentscope.runtime.stage_emitter import PipelineStageEmitter
from core.agentscope.tools.toolkit import build_stage_toolkit
from core.agentscope.utils import (
    finalize_json_response,
    patch_agentscope_json_repair_compat,
)
from core.db.session import AsyncSessionLocal
from core.logging import get_logger
from models.llm import Llm
from models.system_agents import SystemAgents
from schemas.agent.runtime_models import (
    RouterAgentOutput,
    WorkerAgentOutputLite,
    resolve_worker_output_model,
)
from schemas.agent.system_agent import AgentType, SystemAgentLLMConfig
from schemas.user import UserContext
from services.litellm.service import LiteLLMService
from sqlalchemy import select
from sqlalchemy.ext.asyncio import AsyncSession

if TYPE_CHECKING:
    from core.agentscope.runtime.orchestrator import PipelineLike

logger = get_logger("core.agentscope.runtime.runner")


@dataclass(frozen=True)
class SystemAgentRuntimeConfig:
    agent_type: AgentType
    model_code: str
    llm_config: SystemAgentLLMConfig


@dataclass(frozen=True)
class StageExecutionResult:
    message: Msg
    payload: dict[str, Any]
    response_metadata: dict[str, Any]


class AgentScopeRunner:
    def __init__(self, *, litellm_service: LiteLLMService | None = None) -> None:
        patch_agentscope_json_repair_compat()
        self._litellm_service = litellm_service or LiteLLMService()

    async def execute(
        self,
        *,
        user_context: UserContext,
        context_messages: list[Msg],
        pipeline: PipelineLike,
        run_input: RunAgentInput,
    ) -> dict[str, Any]:
        owner_id = UUID(user_context.id)

        async with AsyncSessionLocal() as session:
            worker_toolkit = self._build_worker_toolkit(
                session=session, owner_id=owner_id
            )
            router_config, worker_config = await self._load_stage_configs(
                session=session
            )

            router_output = await self._execute_router_step(
                session=session,
                pipeline=pipeline,
                run_input=run_input,
                user_context=user_context,
                context_messages=context_messages,
                stage_config=router_config,
            )
            worker_output = await self._execute_worker_step(
                pipeline=pipeline,
                run_input=run_input,
                user_context=user_context,
                router_output=router_output,
                toolkit=worker_toolkit,
                stage_config=worker_config,
            )

            return {
                "router": router_output.model_dump(mode="json", exclude_none=True),
                "worker": worker_output.model_dump(mode="json", exclude_none=True),
            }

    def _build_worker_toolkit(
        self,
        *,
        session: AsyncSession,
        owner_id: UUID,
    ) -> Any:
        return build_stage_toolkit(
            agent_type=AgentType.WORKER,
            session=session,
            owner_id=owner_id,
        )

    async def _load_stage_configs(
        self,
        *,
        session: AsyncSession,
    ) -> tuple[SystemAgentRuntimeConfig, SystemAgentRuntimeConfig]:
        router_config = await self._load_system_agent_config(
            session=session,
            agent_type=AgentType.ROUTER,
        )
        worker_config = await self._load_system_agent_config(
            session=session,
            agent_type=AgentType.WORKER,
        )
        return router_config, worker_config

    async def _execute_router_step(
        self,
        *,
        session: AsyncSession,
        pipeline: PipelineLike,
        run_input: RunAgentInput,
        user_context: UserContext,
        context_messages: list[Msg],
        stage_config: SystemAgentRuntimeConfig,
    ) -> RouterAgentOutput:
        await self._emit_step_event(
            pipeline=pipeline,
            run_input=run_input,
            step_name="router",
            event_type="STEP_STARTED",
        )
        router_result = await self._run_router_stage(
            user_context=user_context,
            context_messages=context_messages,
            run_input=run_input,
            stage_config=stage_config,
        )
        router_output = RouterAgentOutput.model_validate(router_result.payload)
        await persist_router_message(
            session=session,
            thread_id=run_input.thread_id,
            run_id=run_input.run_id,
            model_code=stage_config.model_code,
            router_output=router_output,
            response_metadata=router_result.response_metadata,
        )
        await session.commit()
        await self._emit_step_event(
            pipeline=pipeline,
            run_input=run_input,
            step_name="router",
            event_type="STEP_FINISHED",
        )
        return router_output

    async def _execute_worker_step(
        self,
        *,
        pipeline: PipelineLike,
        run_input: RunAgentInput,
        user_context: UserContext,
        router_output: RouterAgentOutput,
        toolkit: Any,
        stage_config: SystemAgentRuntimeConfig,
    ) -> WorkerAgentOutputLite:
        worker_output_model = resolve_worker_output_model(router_output.ui.ui_mode)
        await self._emit_step_event(
            pipeline=pipeline,
            run_input=run_input,
            step_name="worker",
            event_type="STEP_STARTED",
        )
        worker_result = await self._run_worker_stage(
            user_context=user_context,
            router_output=router_output,
            toolkit=toolkit,
            run_input=run_input,
            stage_config=stage_config,
            worker_output_model=worker_output_model,
            pipeline=pipeline,
        )
        worker_output = worker_output_model.model_validate(worker_result.payload)
        await self._emit_step_event(
            pipeline=pipeline,
            run_input=run_input,
            step_name="worker",
            event_type="STEP_FINISHED",
        )
        return worker_output

    async def _load_system_agent_config(
        self,
        *,
        session: AsyncSession,
        agent_type: AgentType,
    ) -> SystemAgentRuntimeConfig:
        stmt = (
            select(SystemAgents, Llm)
            .join(Llm, SystemAgents.llm_id == Llm.id)
            .where(SystemAgents.agent_type == agent_type.value)
        )
        row = (await session.execute(stmt)).one_or_none()
        if row is None:
            raise RuntimeError(f"system agent config not found: {agent_type.value}")
        system_agent, llm = row
        status = str(system_agent.status).strip().lower()
        if status != "active":
            raise RuntimeError(f"system agent is not active: {agent_type.value}")
        return SystemAgentRuntimeConfig(
            agent_type=agent_type,
            model_code=llm.model_code,
            llm_config=SystemAgentLLMConfig.model_validate(system_agent.config or {}),
        )

    async def _run_router_stage(
        self,
        *,
        user_context: UserContext,
        context_messages: list[Msg],
        run_input: RunAgentInput,
        stage_config: SystemAgentRuntimeConfig,
    ) -> StageExecutionResult:
        tracking_model = self._build_model(stage_config=stage_config)
        system_prompt = build_system_prompt(
            agent_type=AgentType.ROUTER,
            user_context=user_context,
            now_utc=datetime.now(timezone.utc),
            tools=None,
        )
        response, payload = await finalize_json_response(
            model=tracking_model,
            formatter=OpenAIChatFormatter(),
            base_messages=[Msg("system", system_prompt, "system"), *context_messages],
            output_model=RouterAgentOutput,
            retries=0,
        )
        response_msg = Msg(
            name="router",
            role="assistant",
            content=list(getattr(response, "content", [])),
            metadata=payload,
        )

        logger.info(
            "router_reply_received",
            run_id=run_input.run_id,
            thread_id=run_input.thread_id,
            message_id=str(response_msg.id),
        )
        return StageExecutionResult(
            message=response_msg,
            payload=payload,
            response_metadata=self._litellm_service.build_usage_metadata(
                model=stage_config.model_code,
                usage_summary=tracking_model.usage_summary(),
            ),
        )

    async def _run_worker_stage(
        self,
        *,
        user_context: UserContext,
        router_output: RouterAgentOutput,
        toolkit: Any,
        run_input: RunAgentInput,
        stage_config: SystemAgentRuntimeConfig,
        worker_output_model: type[WorkerAgentOutputLite],
        pipeline: PipelineLike,
    ) -> StageExecutionResult:
        worker_input = self._build_worker_input_messages(router_output=router_output)
        tracking_model = self._build_model(stage_config=stage_config)
        emitter = PipelineStageEmitter(
            pipeline=pipeline,
            session_id=run_input.thread_id,
            run_id=run_input.run_id,
            stage="worker",
            emit_text_events=True,
            emit_tool_events=True,
        )
        agent = self._build_agent(
            agent_name="worker",
            system_prompt=build_system_prompt(
                agent_type=AgentType.WORKER,
                user_context=user_context,
                now_utc=datetime.now(timezone.utc),
                tools=None,
            ),
            toolkit=toolkit,
            model=tracking_model,
            emitter=emitter,
        )
        response_msg = await agent.reply_json(
            worker_input, output_model=worker_output_model
        )
        worker_payload = worker_output_model.model_validate(response_msg.metadata or {})
        response_metadata = self._litellm_service.build_usage_metadata(
            model=stage_config.model_code,
            usage_summary=tracking_model.usage_summary(),
        )
        await emitter.emit_final_text_end(
            worker_output=worker_payload.model_dump(mode="json", exclude_none=True),
            response_metadata=response_metadata,
        )
        return StageExecutionResult(
            message=response_msg,
            payload=worker_payload.model_dump(mode="json", exclude_none=True),
            response_metadata=response_metadata,
        )

    def _build_worker_input_messages(
        self,
        *,
        router_output: RouterAgentOutput,
    ) -> list[Msg]:
        return [
            Msg(
                name="router",
                role="user",
                content=build_worker_contract_prompt(router_output=router_output),
            )
        ]

    def _build_model(
        self, *, stage_config: SystemAgentRuntimeConfig
    ) -> TrackingChatModel:
        generate_kwargs: dict[str, Any] = {
            "temperature": stage_config.llm_config.temperature,
            "max_tokens": stage_config.llm_config.max_tokens,
            "timeout": stage_config.llm_config.timeout_seconds,
        }
        if stage_config.agent_type == AgentType.ROUTER:
            generate_kwargs["extra_body"] = {"enable_thinking": False}

        model = OpenAIChatModel(
            model_name=stage_config.model_code,
            api_key=self._litellm_service.proxy_api_key,
            stream=False,
            client_kwargs={"base_url": self._litellm_service.proxy_base_url},
            generate_kwargs=generate_kwargs,
        )
        return TrackingChatModel(model)

    def _build_agent(
        self,
        *,
        agent_name: str,
        system_prompt: str,
        toolkit: Any,
        model: TrackingChatModel,
        emitter: PipelineStageEmitter | None = None,
    ) -> JsonReActAgent:
        return JsonReActAgent(
            name=agent_name,
            sys_prompt=system_prompt,
            model=model,
            formatter=OpenAIChatFormatter(),
            toolkit=toolkit,
            memory=InMemoryMemory(),
            emitter=emitter,
        )

    async def _emit_step_event(
        self,
        *,
        pipeline: PipelineLike,
        run_input: RunAgentInput,
        step_name: str,
        event_type: str,
    ) -> None:
        await pipeline.emit(
            session_id=run_input.thread_id,
            event={
                "type": event_type,
                "threadId": run_input.thread_id,
                "runId": run_input.run_id,
                "stepName": step_name,
            },
        )


AgentScopeReActRunner = AgentScopeRunner