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

from __future__ import annotations

import asyncio
import contextlib
from dataclasses import dataclass
from datetime import datetime, timezone
from typing import TYPE_CHECKING, Any, Awaitable, Callable

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.schemas.agui_input import extract_latest_user_payload
from core.agentscope.runtime.json_react_agent import JsonReActAgent
from core.agentscope.runtime.model_tracking import TrackingChatModel
from core.agentscope.runtime.stage_emitter import PipelineStageEmitter
from core.agentscope.tools.toolkit import build_toolkit
from core.agentscope.utils import (
    finalize_json_response,
    patch_agentscope_json_repair_compat,
)
from core.auth.credential_issuer import create_credential_issuer
from core.auth.tool_credential_context import (
    set_tool_credential,
    reset_tool_credential,
)
from core.config.settings import config
from core.db.session import AsyncSessionLocal
from models.llm import Llm
from models.llm_factory import LlmFactory
from models.system_agents import SystemAgents
from schemas.agent.forwarded_props import (
    ClientTimeContext,
    RuntimeMode,
    parse_forwarded_props_client_time,
    parse_forwarded_props_runtime_mode,
)
from schemas.agent.runtime_models import (
    RouterAgentOutput,
    WorkerAgentOutputLite,
)
from schemas.agent.skill_config import ProjectCliCommand, SkillName
from schemas.agent.system_agent import (
    AgentType,
    SystemAgentLLMConfig,
)
from schemas.domain.automation import RuntimeConfig
from schemas.domain.memory_content import UserMemoryContent, WorkProfileContent
from schemas.shared.user import UserContext
from services.llm_pricing.service import LlmPricingService
from sqlalchemy import select
from sqlalchemy.ext.asyncio import AsyncSession

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


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


class AgentScopeRunner:
    def __init__(self, *, llm_pricing_service: LlmPricingService | None = None) -> None:
        patch_agentscope_json_repair_compat()
        self._llm_pricing_service: LlmPricingService = (
            llm_pricing_service or LlmPricingService()
        )
        self._active_agent: JsonReActAgent | None = None
        self._active_agent_lock = asyncio.Lock()

    async def execute(
        self,
        *,
        user_context: UserContext,
        context_messages: list[Msg],
        pipeline: PipelineLike,
        run_input: RunAgentInput,
        runtime_config: RuntimeConfig,
        user_memory: UserMemoryContent | None = None,
        work_memory: WorkProfileContent | None = None,
        cancel_checker: Callable[[], Awaitable[bool]] | None = None,
    ) -> dict[str, Any]:
        runtime_client_time = self._resolve_runtime_client_time(run_input=run_input)
        runtime_mode = self._resolve_runtime_mode(run_input=run_input)
        stop_cancel_watch = asyncio.Event()
        cancel_watch_task: asyncio.Task[None] | None = None
        run_task = asyncio.current_task()

        if cancel_checker is not None and run_task is not None:
            cancel_watch_task = asyncio.create_task(
                self._watch_cancel_signal(
                    cancel_checker=cancel_checker,
                    stop_signal=stop_cancel_watch,
                    run_task=run_task,
                )
            )

        try:
            async with AsyncSessionLocal() as session:
                router_config = await self._load_stage_config(
                    session=session,
                    agent_type=AgentType.ROUTER,
                )
                worker_config = await self._load_stage_config(
                    session=session,
                    agent_type=AgentType.WORKER,
                )
                worker_toolkit = self._build_toolkit(
                    enabled_skills=runtime_config.enabled_skills,
                    allowed_commands=runtime_config.allowed_commands,
                )

                router_output = await self._execute_router_step(
                    pipeline=pipeline,
                    run_input=run_input,
                    user_context=user_context,
                    context_messages=context_messages,
                    stage_config=router_config,
                    runtime_client_time=runtime_client_time,
                    runtime_mode=runtime_mode,
                    user_memory=user_memory,
                )
                if cancel_checker is not None and await cancel_checker():
                    raise asyncio.CancelledError("run canceled by user")
                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,
                    runtime_client_time=runtime_client_time,
                    runtime_mode=runtime_mode,
                    work_memory=work_memory,
                )
                return {
                    "router": router_output.model_dump(mode="json", exclude_none=True),
                    "worker": worker_output.model_dump(mode="json", exclude_none=True),
                }
        finally:
            stop_cancel_watch.set()
            if cancel_watch_task is not None:
                cancel_watch_task.cancel()
                with contextlib.suppress(asyncio.CancelledError):
                    await cancel_watch_task

    async def _watch_cancel_signal(
        self,
        *,
        cancel_checker: Callable[[], Awaitable[bool]],
        stop_signal: asyncio.Event,
        run_task: asyncio.Task[object],
    ) -> None:
        while not stop_signal.is_set():
            should_cancel = False
            try:
                should_cancel = await cancel_checker()
            except Exception:
                should_cancel = False

            if should_cancel:
                async with self._active_agent_lock:
                    active_agent = self._active_agent
                if active_agent is not None:
                    with contextlib.suppress(Exception):
                        await active_agent.interrupt()
                if not run_task.done():
                    run_task.cancel("run canceled by user")
                return

            await asyncio.sleep(0.2)

    def _build_toolkit(
        self,
        *,
        enabled_skills: list[SkillName],
        allowed_commands: list[ProjectCliCommand],
    ) -> Any:
        enabled_skill_names = {str(skill.value) for skill in enabled_skills}
        allowed_command_names = {str(command.value) for command in allowed_commands}
        return build_toolkit(
            enabled_skill_names=enabled_skill_names if enabled_skill_names else None,
            allowed_commands=allowed_command_names if allowed_command_names else None,
        )

    async def _load_stage_config(
        self,
        *,
        session: AsyncSession,
        agent_type: AgentType,
    ) -> SystemAgentRuntimeConfig:
        stmt = (
            select(SystemAgents, Llm, LlmFactory)
            .join(Llm, SystemAgents.llm_id == Llm.id)
            .join(LlmFactory, Llm.factory_id == LlmFactory.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, factory = 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,
            api_base_url=factory.request_url,
            api_key=self._resolve_provider_api_key(factory_name=factory.name),
            llm_config=SystemAgentLLMConfig.model_validate(system_agent.config or {}),
            extra_context=None,
        )

    async def _execute_router_step(
        self,
        *,
        pipeline: PipelineLike,
        run_input: RunAgentInput,
        user_context: UserContext,
        context_messages: list[Msg],
        stage_config: SystemAgentRuntimeConfig,
        runtime_client_time: ClientTimeContext | None,
        runtime_mode: RuntimeMode,
        user_memory: UserMemoryContent | None,
    ) -> RouterAgentOutput:
        await self._emit_step_event(
            pipeline=pipeline,
            run_input=run_input,
            step_name=AgentType.ROUTER.value,
            event_type="STEP_STARTED",
            runtime_mode=runtime_mode,
        )
        router_result = await self._run_router_stage(
            user_context=user_context,
            context_messages=context_messages,
            stage_config=stage_config,
            runtime_client_time=runtime_client_time,
            user_memory=user_memory,
            run_input=run_input,
        )
        router_output = RouterAgentOutput.model_validate(router_result.payload)
        await self._emit_step_event(
            pipeline=pipeline,
            run_input=run_input,
            step_name=AgentType.ROUTER.value,
            event_type="STEP_FINISHED",
            runtime_mode=runtime_mode,
            extra_event={
                "_router_persist": {
                    "router_output": router_output.model_dump(
                        mode="json", exclude_none=True
                    ),
                    "response_metadata": router_result.response_metadata,
                }
            },
        )
        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,
        runtime_client_time: ClientTimeContext | None,
        runtime_mode: RuntimeMode,
        work_memory: WorkProfileContent | None,
    ) -> WorkerAgentOutputLite:
        await self._emit_step_event(
            pipeline=pipeline,
            run_input=run_input,
            step_name=AgentType.WORKER.value,
            event_type="STEP_STARTED",
            runtime_mode=runtime_mode,
        )
        worker_result = await self._run_worker_stage(
            user_context=user_context,
            input_messages=self._build_worker_input_messages(
                router_output=router_output
            ),
            toolkit=toolkit,
            run_input=run_input,
            stage_config=stage_config,
            worker_output_model=WorkerAgentOutputLite,
            pipeline=pipeline,
            runtime_client_time=runtime_client_time,
            runtime_mode=runtime_mode,
            work_memory=work_memory,
            requires_tool_evidence=router_output.requires_tool_evidence,
        )
        worker_output = WorkerAgentOutputLite.model_validate(worker_result.payload)
        await self._emit_step_event(
            pipeline=pipeline,
            run_input=run_input,
            step_name=AgentType.WORKER.value,
            event_type="STEP_FINISHED",
            runtime_mode=runtime_mode,
        )
        return worker_output

    async def _run_router_stage(
        self,
        *,
        user_context: UserContext,
        context_messages: list[Msg],
        stage_config: SystemAgentRuntimeConfig,
        runtime_client_time: ClientTimeContext | None,
        user_memory: UserMemoryContent | None,
        run_input: RunAgentInput,
    ) -> StageExecutionResult:
        messages_for_router = self._build_router_messages(
            context_messages=context_messages,
            run_input=run_input,
        )
        tracking_model = self._build_model(stage_config=stage_config)
        response, payload = await finalize_json_response(
            model=tracking_model,
            formatter=OpenAIChatFormatter(),
            base_messages=[
                Msg(
                    "system",
                    build_system_prompt(
                        agent_type=AgentType.ROUTER,
                        llm_config=stage_config.llm_config,
                        user_context=user_context,
                        now_utc=datetime.now(timezone.utc),
                        runtime_client_time=runtime_client_time,
                        user_memory=user_memory,
                    ),
                    "system",
                ),
                *messages_for_router,
            ],
            output_model=RouterAgentOutput,
            retries=3,
        )
        response_msg = Msg(
            name="router",
            role="assistant",
            content=list(getattr(response, "content", [])),
            metadata=payload,
        )
        return StageExecutionResult(
            message=response_msg,
            payload=payload,
            response_metadata=self._llm_pricing_service.build_usage_metadata(
                model=stage_config.model_code,
                usage_summary=tracking_model.usage_summary(),
            ),
        )

    def _build_router_messages(
        self,
        *,
        context_messages: list[Msg],
        run_input: RunAgentInput,
    ) -> list[Msg]:
        if context_messages:
            last = context_messages[-1]
            if last.role == "user":
                return context_messages

        user_text, user_blocks = extract_latest_user_payload(run_input)
        if (
            user_blocks
            and isinstance(user_blocks[0], dict)
            and user_blocks[0].get("type") == "text"
        ):
            content: Any = user_text
        else:
            content = user_blocks

        user_msg = Msg(name="user", role="user", content=content)
        return [*context_messages, user_msg]

    async def _run_worker_stage(
        self,
        *,
        user_context: UserContext,
        input_messages: list[Msg],
        toolkit: Any,
        run_input: RunAgentInput,
        stage_config: SystemAgentRuntimeConfig,
        worker_output_model: type[WorkerAgentOutputLite],
        pipeline: PipelineLike,
        runtime_client_time: ClientTimeContext | None,
        runtime_mode: RuntimeMode,
        work_memory: WorkProfileContent | None,
        requires_tool_evidence: bool = False,
    ) -> StageExecutionResult:
        issuer = create_credential_issuer()
        credential = issuer.issue(
            owner_id=str(user_context.id),
            mode=runtime_mode.value,
        )
        credential_token = set_tool_credential(credential)

        try:
            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=stage_config.agent_type.value,
                runtime_mode=runtime_mode.value,
                emit_text_events=True,
                emit_tool_events=True,
            )
            agent = self._build_agent(
                agent_name=stage_config.agent_type.value,
                system_prompt=build_system_prompt(
                    agent_type=stage_config.agent_type,
                    llm_config=stage_config.llm_config,
                    user_context=user_context,
                    now_utc=datetime.now(timezone.utc),
                    runtime_client_time=runtime_client_time,
                    extra_context=stage_config.extra_context,
                    work_memory=work_memory,
                ),
                toolkit=toolkit,
                model=tracking_model,
                emitter=emitter,
                force_tool_on_first_reasoning=requires_tool_evidence,
            )
            async with self._active_agent_lock:
                self._active_agent = agent
            try:
                response_msg = await agent.reply_json(
                    input_messages, output_model=worker_output_model
                )
            finally:
                async with self._active_agent_lock:
                    if self._active_agent is agent:
                        self._active_agent = None
            worker_payload = worker_output_model.model_validate(response_msg.metadata or {})
            response_metadata = self._llm_pricing_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,
            )
        finally:
            reset_tool_credential(credential_token)

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

    def _build_model(
        self, *, stage_config: SystemAgentRuntimeConfig
    ) -> TrackingChatModel:
        generate_kwargs: dict[str, Any] = {
            "timeout": stage_config.llm_config.timeout_seconds,
            "extra_body": {"enable_thinking": False},
        }
        if stage_config.llm_config.temperature is not None:
            generate_kwargs["temperature"] = stage_config.llm_config.temperature
        if stage_config.llm_config.max_tokens is not None:
            generate_kwargs["max_tokens"] = stage_config.llm_config.max_tokens

        model = OpenAIChatModel(
            model_name=stage_config.model_code,
            api_key=stage_config.api_key,
            stream=False,
            client_kwargs={"base_url": stage_config.api_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,
        force_tool_on_first_reasoning: bool = False,
    ) -> JsonReActAgent:
        return JsonReActAgent(
            name=agent_name,
            sys_prompt=system_prompt,
            model=model,
            formatter=OpenAIChatFormatter(),
            toolkit=toolkit,
            memory=InMemoryMemory(),
            emitter=emitter,
            force_tool_on_first_reasoning=force_tool_on_first_reasoning,
        )

    async def _emit_step_event(
        self,
        *,
        pipeline: PipelineLike,
        run_input: RunAgentInput,
        step_name: str,
        event_type: str,
        runtime_mode: RuntimeMode,
        extra_event: dict[str, Any] | None = None,
    ) -> None:
        payload: dict[str, Any] = {
            "type": event_type,
            "threadId": run_input.thread_id,
            "runId": run_input.run_id,
            "runtime_mode": runtime_mode.value,
            "stepName": step_name,
        }
        if extra_event:
            payload.update(extra_event)
        await pipeline.emit(
            session_id=run_input.thread_id,
            event=payload,
        )

    def _resolve_runtime_client_time(
        self, *, run_input: RunAgentInput
    ) -> ClientTimeContext | None:
        return parse_forwarded_props_client_time(
            getattr(run_input, "forwarded_props", None)
        )

    @staticmethod
    def _resolve_runtime_mode(*, run_input: RunAgentInput) -> RuntimeMode:
        try:
            return parse_forwarded_props_runtime_mode(
                getattr(run_input, "forwarded_props", None)
            )
        except ValueError:
            return RuntimeMode.CHAT

    @staticmethod
    def _resolve_provider_api_key(*, factory_name: str) -> str:
        normalized_factory_name = factory_name.strip().upper()
        if normalized_factory_name == "VOLCENGINE":
            normalized_factory_name = "ARK"

        provider_keys = {
            str(key).strip().upper(): str(value).strip()
            for key, value in config.llm.provider_keys.items()
            if str(value).strip()
        }
        api_key = provider_keys.get(normalized_factory_name, "")
        if not api_key:
            raise RuntimeError(f"provider api key missing for factory: {factory_name}")
        return api_key


@dataclass(frozen=True)
class SystemAgentRuntimeConfig:
    agent_type: AgentType
    model_code: str
    api_base_url: str
    api_key: str
    llm_config: SystemAgentLLMConfig
    extra_context: str | None = None


AgentScopeReActRunner = AgentScopeRunner