agent-platform/services/runtime-service/app/application/services.py

from datetime import datetime, timedelta

from sqlalchemy.orm import Session

from core_dsl import parse_workflow_definition
from core_events import EventPublishContract, EventServiceClient, EventServiceClientError
from core_domain import (
    InitialNodeContract,
    NodeExecutionContextContract,
    NodeExecutionResultContract,
    NodeRunStatus,
    WorkflowVersionContract,
    WorkflowRunStatus,
)

from app.db.models import NodeRun, WorkflowRun
from app.domain.repositories import (
    ExecutionLogRepository,
    NodeArtifactRepository,
    NodeRunRepository,
    TraceSpanRepository,
    WorkflowRunRepository,
)
from app.infrastructure.executors import (
    NodeExecutionDispatcher,
    build_node_execution_dispatcher_with_clients,
)
from app.infrastructure.human_client import HumanServiceClient
from app.infrastructure.knowledge_client import KnowledgeServiceClient
from app.infrastructure.code_runner_client import CodeRunnerClient
from app.infrastructure.model_gateway_client import ModelGatewayClient
from app.infrastructure.planner import (
    derive_initial_node,
    derive_node_config,
    derive_successor_nodes,
)
from app.infrastructure.tool_client import ToolServiceClient
from app.infrastructure.workflow_client import WorkflowServiceClient
from app.bootstrap.settings import RuntimeServiceSettings
from app.schemas.run import (
    NodeRunExecuteRequest,
    HumanNodeResumeRequest,
    NodeRunStatusUpdateRequest,
    RunCreateRequest,
    RunExecuteRequest,
    WorkflowRunStatusUpdateRequest,
)
from core_shared import JSONValue, try_build_redis_client
from core_shared.task_queue import TaskQueuePublisher


class RuntimeApplicationService:
    def __init__(
        self,
        workflow_run_repository: WorkflowRunRepository,
        node_run_repository: NodeRunRepository,
        execution_log_repository: ExecutionLogRepository,
        node_artifact_repository: NodeArtifactRepository,
        trace_span_repository: TraceSpanRepository,
        execution_dispatcher: NodeExecutionDispatcher,
        workflow_client: WorkflowServiceClient | None = None,
        event_client: EventServiceClient | None = None,
        task_queue_publisher: TaskQueuePublisher | None = None,
    ) -> None:
        self.workflow_run_repository = workflow_run_repository
        self.node_run_repository = node_run_repository
        self.execution_log_repository = execution_log_repository
        self.node_artifact_repository = node_artifact_repository
        self.trace_span_repository = trace_span_repository
        self.execution_dispatcher = execution_dispatcher
        self.workflow_client = workflow_client
        self.event_client = event_client
        self.task_queue_publisher = task_queue_publisher

    def create_run(self, payload: RunCreateRequest) -> tuple[WorkflowRun, NodeRun | None]:
        initial_node = payload.initial_node or self._plan_initial_node(payload)
        workflow_run = self.workflow_run_repository.create(
            tenant_id=payload.tenant_id,
            app_id=payload.app_id,
            app_version_id=payload.app_version_id,
            workflow_id=payload.workflow_id,
            workflow_version_id=payload.workflow_version_id,
            session_id=payload.session_id,
            parent_run_id=payload.parent_run_id,
            root_run_id=payload.root_run_id,
            run_type=payload.run_type,
            trigger_type=payload.trigger_type,
            priority=payload.priority,
        )

        node_run = None
        if initial_node is not None:
            self.workflow_run_repository.update_node_count(
                run_id=workflow_run.id,
                current_node_count=1,
            )
            initial_config = self._resolve_node_config(
                tenant_id=payload.tenant_id,
                workflow_version_id=payload.workflow_version_id,
                node_id=initial_node.node_id,
            )
            scheduled_time, timeout_time = self._build_node_timing(initial_config)
            node_run = self.node_run_repository.create(
                tenant_id=payload.tenant_id,
                run_id=workflow_run.id,
                node_id=initial_node.node_id,
                node_type=initial_node.node_type,
                status=initial_node.status,
                scheduled_time=scheduled_time,
                timeout_time=timeout_time,
            )
            self._log_event(
                tenant_id=payload.tenant_id,
                run_id=workflow_run.id,
                node_run_id=node_run.id,
                event_type="node_queued",
                message=f"initial node queued: {initial_node.node_id}",
                detail_json={
                    "node_id": initial_node.node_id,
                    "node_type": initial_node.node_type,
                    "status": initial_node.status,
                },
            )
            self._publish_node_run_to_queue(node_run)

        self._log_event(
            tenant_id=payload.tenant_id,
            run_id=workflow_run.id,
            node_run_id=node_run.id if node_run is not None else None,
            event_type="run_created",
            message="workflow run created",
            detail_json={
                "workflow_id": payload.workflow_id,
                "workflow_version_id": payload.workflow_version_id,
                "session_id": payload.session_id,
            },
        )
        self._publish_event(
            event_type="workflow.run.created",
            workflow_run=workflow_run,
            payload_json={
                "run_id": workflow_run.id,
                "workflow_id": workflow_run.workflow_id,
                "workflow_version_id": workflow_run.workflow_version_id,
                "status": workflow_run.status,
            },
        )

        return workflow_run, node_run

    def list_runs(self, tenant_id: str, session_id: str | None = None) -> list[WorkflowRun]:
        return self.workflow_run_repository.list_by_scope(
            tenant_id=tenant_id,
            session_id=session_id,
        )

    def list_node_runs(self, tenant_id: str, run_id: str) -> list[NodeRun]:
        return self.node_run_repository.list_by_run(tenant_id=tenant_id, run_id=run_id)

    def list_execution_logs(
        self,
        tenant_id: str,
        run_id: str | None = None,
        node_run_id: str | None = None,
    ):
        return self.execution_log_repository.list_by_scope(
            tenant_id=tenant_id,
            run_id=run_id,
            node_run_id=node_run_id,
        )

    def list_node_artifacts(
        self,
        tenant_id: str,
        run_id: str | None = None,
        node_run_id: str | None = None,
        artifact_type: str | None = None,
    ):
        return self.node_artifact_repository.list_by_scope(
            tenant_id=tenant_id,
            run_id=run_id,
            node_run_id=node_run_id,
            artifact_type=artifact_type,
        )

    def list_trace_spans(
        self,
        tenant_id: str,
        run_id: str | None = None,
        node_run_id: str | None = None,
        span_type: str | None = None,
    ):
        return self.trace_span_repository.list_by_scope(
            tenant_id=tenant_id,
            run_id=run_id,
            node_run_id=node_run_id,
            span_type=span_type,
        )

    def update_run_status(
        self,
        run_id: str,
        payload: WorkflowRunStatusUpdateRequest,
    ) -> WorkflowRun | None:
        entity = self.workflow_run_repository.update_status(
            run_id=run_id,
            status=payload.status,
            error_code=payload.error_code,
            error_message=payload.error_message,
        )
        if entity is not None:
            self._publish_event(
                event_type=f"workflow.run.{entity.status}",
                workflow_run=entity,
                payload_json={
                    "run_id": entity.id,
                    "status": entity.status,
                    "error_code": entity.error_code,
                },
            )
        return entity

    def update_node_run_status(
        self,
        node_run_id: str,
        payload: NodeRunStatusUpdateRequest,
    ) -> NodeRun | None:
        node_run = self.node_run_repository.update_status(
            node_run_id=node_run_id,
            status=payload.status,
            worker_key=payload.worker_key,
            error_code=payload.error_code,
            error_message=payload.error_message,
            output_text=payload.output_text,
            output_json=payload.output_json,
        )
        if node_run is None:
            return None

        self._log_event(
            tenant_id=node_run.tenant_id,
            run_id=node_run.run_id,
            node_run_id=node_run.id,
            event_type="node_status_updated",
            message=f"node status updated to {payload.status}",
            detail_json={
                "node_id": node_run.node_id,
                "node_type": node_run.node_type,
                "status": payload.status,
                "error_code": payload.error_code,
            },
        )
        self._publish_event(
            event_type=f"workflow.node.{node_run.status}",
            workflow_run=None,
            node_run=node_run,
            payload_json={
                "run_id": node_run.run_id,
                "node_run_id": node_run.id,
                "node_id": node_run.node_id,
                "node_type": node_run.node_type,
                "status": node_run.status,
                "error_code": node_run.error_code,
            },
        )

        if payload.status == "completed":
            self._schedule_successor_nodes(node_run)
        if payload.status == "failed":
            workflow_run = self.workflow_run_repository.get_by_id(node_run.run_id)
            if workflow_run is not None:
                node_config = self._resolve_node_config(
                    tenant_id=node_run.tenant_id,
                    workflow_version_id=workflow_run.workflow_version_id,
                    node_id=node_run.node_id,
                )
                self._schedule_compensation_node(node_run=node_run, node_config=node_config)

        self._sync_workflow_run_status_from_nodes(
            tenant_id=node_run.tenant_id,
            run_id=node_run.run_id,
        )
        return node_run

    def execute_node_run(
        self,
        node_run_id: str,
        payload: NodeRunExecuteRequest,
    ) -> tuple[WorkflowRun, NodeRun, str] | None:
        node_run = self.node_run_repository.get_by_id(node_run_id)
        if node_run is None:
            return None

        workflow_run = self.workflow_run_repository.get_by_id(node_run.run_id)
        if workflow_run is None:
            return None

        if node_run.status in {"completed", "failed", "skipped"}:
            executor_name = self.execution_dispatcher.resolve_executor(
                node_run.node_type
            ).executor_name
            return workflow_run, node_run, executor_name

        node_config = self._resolve_node_config(
            tenant_id=node_run.tenant_id,
            workflow_version_id=workflow_run.workflow_version_id,
            node_id=node_run.node_id,
        )
        if self._node_has_timed_out(node_run):
            timed_out_node_run = self.update_node_run_status(
                node_run_id=node_run.id,
                payload=NodeRunStatusUpdateRequest(
                    status="failed",
                    worker_key=payload.worker_key,
                    error_code="node_timeout",
                    error_message=f"node timed out: {node_run.node_id}",
                    output_json={
                        "timeout_time": node_run.timeout_time.isoformat()
                        if node_run.timeout_time is not None
                        else None,
                    },
                ),
            )
            if timed_out_node_run is None:
                return None
            executor_name = self.execution_dispatcher.resolve_executor(
                node_run.node_type
            ).executor_name
            return workflow_run, timed_out_node_run, executor_name

        running_node_run = self.node_run_repository.update_status(
            node_run_id=node_run_id,
            status="running",
            worker_key=payload.worker_key,
        )
        if running_node_run is None:
            return None

        self._log_event(
            tenant_id=running_node_run.tenant_id,
            run_id=running_node_run.run_id,
            node_run_id=running_node_run.id,
            event_type="node_execution_started",
            message=f"executing node {running_node_run.node_id}",
            detail_json={
                "node_id": running_node_run.node_id,
                "node_type": running_node_run.node_type,
                "worker_key": payload.worker_key,
            },
        )

        context = self._build_execution_context(
            workflow_run=workflow_run,
            node_run=running_node_run,
            worker_key=payload.worker_key,
            node_config_json=node_config,
        )
        executor_name = self.execution_dispatcher.resolve_executor(
            running_node_run.node_type
        ).executor_name
        trace_span = self.trace_span_repository.start(
            tenant_id=running_node_run.tenant_id,
            run_id=running_node_run.run_id,
            node_run_id=running_node_run.id,
            parent_span_id=None,
            span_type="node_execution",
            name=f"{running_node_run.node_type}:{running_node_run.node_id}",
            attributes_json={
                "node_id": running_node_run.node_id,
                "node_type": running_node_run.node_type,
                "executor_name": executor_name,
                "worker_key": payload.worker_key,
            },
        )

        try:
            result, executor_name = self.execution_dispatcher.execute(
                context=context,
                request=payload,
            )
        except Exception as exc:
            result = NodeExecutionResultContract(
                status="failed",
                worker_key=payload.worker_key,
                error_code="executor_error",
                error_message=str(exc),
            )

        if result.status == "failed" and self._should_retry_node(
            node_run=running_node_run,
            node_config_json=context.node_config_json,
        ):
            retry_time, retry_timeout_time = self._build_retry_timing(context.node_config_json)
            retried_node_run = self.node_run_repository.requeue_for_retry(
                node_run_id=running_node_run.id,
                scheduled_time=retry_time,
                timeout_time=retry_timeout_time,
                error_code=result.error_code,
                error_message=result.error_message,
                output_text=result.output_text,
                output_json={
                    **(result.output_json or {}),
                    "retry_scheduled_time": retry_time.isoformat(),
                    "retry_reason": result.error_code or "node_failed",
                },
            )
            if retried_node_run is None:
                return None
            self._publish_node_run_to_queue(retried_node_run)
            self.trace_span_repository.finish(
                span_id=trace_span.id,
                status="error",
                error_code=result.error_code,
                error_message=result.error_message,
                attributes_json={
                    "node_status": "queued",
                    "executor_name": executor_name,
                    "retry_scheduled": True,
                    "attempt_no": retried_node_run.attempt_no,
                },
            )
            self._log_event(
                tenant_id=retried_node_run.tenant_id,
                run_id=retried_node_run.run_id,
                node_run_id=retried_node_run.id,
                event_type="node_retry_scheduled",
                message=f"node retry scheduled: {retried_node_run.node_id}",
                detail_json={
                    "node_id": retried_node_run.node_id,
                    "attempt_no": retried_node_run.attempt_no,
                    "scheduled_time": retry_time.isoformat(),
                    "error_code": result.error_code,
                },
            )
            self._sync_workflow_run_status_from_nodes(
                tenant_id=retried_node_run.tenant_id,
                run_id=retried_node_run.run_id,
            )
            workflow_run = self.workflow_run_repository.get_by_id(retried_node_run.run_id)
            if workflow_run is None:
                return None
            return workflow_run, retried_node_run, executor_name

        final_node_run = self.update_node_run_status(
            node_run_id=running_node_run.id,
            payload=NodeRunStatusUpdateRequest(
                status=result.status,
                worker_key=result.worker_key,
                error_code=result.error_code,
                error_message=result.error_message,
                output_text=result.output_text,
                output_json=result.output_json,
            ),
        )
        if final_node_run is None:
            return None

        self.trace_span_repository.finish(
            span_id=trace_span.id,
            status="ok" if final_node_run.status == "completed" else "error",
            error_code=final_node_run.error_code,
            error_message=final_node_run.error_message,
            attributes_json={
                "node_status": final_node_run.status,
                "executor_name": executor_name,
                "has_output_text": final_node_run.output_text is not None,
                "has_output_json": final_node_run.output_json is not None,
            },
        )
        self._persist_node_execution_artifact(final_node_run)

        self._log_event(
            tenant_id=final_node_run.tenant_id,
            run_id=final_node_run.run_id,
            node_run_id=final_node_run.id,
            event_type="node_execution_finished",
            message=f"node execution finished with status {final_node_run.status}",
            detail_json={
                "node_id": final_node_run.node_id,
                "node_type": final_node_run.node_type,
                "executor_name": executor_name,
                "status": final_node_run.status,
            },
        )

        workflow_run = self.workflow_run_repository.get_by_id(final_node_run.run_id)
        if workflow_run is None:
            return None
        return workflow_run, final_node_run, executor_name

    def execute_next_node_run(
        self,
        tenant_id: str,
        run_id: str,
        payload: NodeRunExecuteRequest,
    ) -> tuple[WorkflowRun, NodeRun, str] | None:
        next_node_run = self.node_run_repository.get_next_queued_by_run(
            tenant_id=tenant_id,
            run_id=run_id,
        )
        if next_node_run is None:
            return None
        return self.execute_node_run(node_run_id=next_node_run.id, payload=payload)

    def execute_run(
        self,
        tenant_id: str,
        run_id: str,
        payload: RunExecuteRequest,
    ) -> tuple[WorkflowRun, list[NodeRun], list[str]] | None:
        workflow_run = self.workflow_run_repository.get_by_id(run_id)
        if workflow_run is None or workflow_run.tenant_id != tenant_id:
            return None

        executed_node_runs: list[NodeRun] = []
        executor_names: list[str] = []

        for _ in range(payload.max_steps):
            step_result = self.execute_next_node_run(
                tenant_id=tenant_id,
                run_id=run_id,
                payload=NodeRunExecuteRequest(worker_key=payload.worker_key),
            )
            if step_result is None:
                break

            workflow_run, node_run, executor_name = step_result
            executed_node_runs.append(node_run)
            executor_names.append(executor_name)

            if node_run.status != "completed":
                break

        final_run = self.workflow_run_repository.get_by_id(run_id)
        if final_run is None:
            return None
        return final_run, executed_node_runs, executor_names

    def resume_human_node_run(
        self,
        *,
        node_run_id: str,
        payload: HumanNodeResumeRequest,
    ) -> tuple[WorkflowRun, NodeRun, str] | None:
        node_run = self.node_run_repository.get_by_id(node_run_id)
        if node_run is None or node_run.tenant_id != payload.tenant_id:
            return None

        output_json = dict(node_run.output_json or {})
        existing_human_task_id = output_json.get("human_task_id")
        if existing_human_task_id is not None and existing_human_task_id != payload.human_task_id:
            return None

        if existing_human_task_id is None:
            output_json["human_task_id"] = payload.human_task_id
            self.node_run_repository.update_status(
                node_run_id=node_run.id,
                status="pending",
                worker_key=payload.worker_key,
                output_json=output_json,
            )

        return self.execute_node_run(
            node_run_id=node_run_id,
            payload=NodeRunExecuteRequest(worker_key=payload.worker_key),
        )

    def execute_next_claimed_node_run(
        self,
        *,
        worker_key: str,
        lease_seconds: int,
        redis_client: object | None = None,
    ) -> tuple[WorkflowRun, NodeRun, str, int] | None:
        released_lease_count = self.node_run_repository.release_expired_leases(
            now_time=datetime.utcnow(),
        )
        claimed_node_run = self.node_run_repository.claim_next_queued(
            worker_key=worker_key,
            lease_expire_time=datetime.utcnow() + timedelta(seconds=lease_seconds),
        )
        if claimed_node_run is None:
            return None

        if redis_client is not None:
            from core_shared.redis_primitives import DistributedLock, IdempotencyStore

            lock = DistributedLock(
                client=redis_client,
                name=f"node-run:{claimed_node_run.id}:lock",
                ttl_seconds=lease_seconds,
            )
            if not lock.acquire():
                return None
            idempotency_store = IdempotencyStore(
                client=redis_client,
                prefix="node-run-idempotency",
            )
            if not idempotency_store.begin(key=claimed_node_run.id):
                lock.release()
                return None
        else:
            lock = None
            idempotency_store = None

        try:
            result = self.execute_node_run(
                node_run_id=claimed_node_run.id,
                payload=NodeRunExecuteRequest(worker_key=worker_key),
            )
            if idempotency_store is not None and result is not None:
                _, node_run, executor_name = result
                idempotency_store.complete(
                    key=claimed_node_run.id,
                    result={
                        "status": node_run.status,
                        "node_run_id": node_run.id,
                        "executor_name": executor_name,
                    },
                )
        finally:
            if lock is not None:
                lock.release()
        if result is None:
            return None

        workflow_run, node_run, executor_name = result
        return workflow_run, node_run, executor_name, released_lease_count

    def _persist_node_execution_artifact(self, node_run: NodeRun) -> None:
        if node_run.output_text is None and node_run.output_json is None:
            return

        size_bytes = len(node_run.output_text.encode("utf-8")) if node_run.output_text else None
        self.node_artifact_repository.create(
            tenant_id=node_run.tenant_id,
            run_id=node_run.run_id,
            node_run_id=node_run.id,
            node_id=node_run.node_id,
            artifact_type="execution_result",
            name=f"{node_run.node_id}-execution-result",
            mime_type="application/json" if node_run.output_json is not None else "text/plain",
            content_text=node_run.output_text,
            content_json=node_run.output_json,
            size_bytes=size_bytes,
        )

    def _plan_initial_node(self, payload: RunCreateRequest) -> InitialNodeContract | None:
        if self.workflow_client is None:
            return None
        workflow_version = self.workflow_client.get_workflow_version(
            tenant_id=payload.tenant_id,
            workflow_version_id=payload.workflow_version_id,
        )
        return derive_initial_node(workflow_version)

    def _resolve_node_config(
        self,
        *,
        tenant_id: str,
        workflow_version_id: str,
        node_id: str,
    ) -> dict[str, JSONValue]:
        if self.workflow_client is None:
            return {}
        workflow_version = self.workflow_client.get_workflow_version(
            tenant_id=tenant_id,
            workflow_version_id=workflow_version_id,
        )
        return derive_node_config(workflow_version, node_id)

    def _schedule_successor_nodes(self, node_run: NodeRun) -> None:
        if self.workflow_client is None:
            return

        workflow_run = self.workflow_run_repository.get_by_id(node_run.run_id)
        if workflow_run is None:
            return

        workflow_version = self.workflow_client.get_workflow_version(
            tenant_id=node_run.tenant_id,
            workflow_version_id=workflow_run.workflow_version_id,
        )
        run_state_json, node_output_json_by_node_id, node_output_text_by_node_id = (
            self._build_run_state_maps(
                tenant_id=node_run.tenant_id,
                run_id=node_run.run_id,
            )
        )
        successor_nodes = derive_successor_nodes(
            workflow_version,
            node_run.node_id,
            current_output_json=node_run.output_json,
            run_state_json=run_state_json,
            node_output_json_by_node_id=node_output_json_by_node_id,
            node_output_text_by_node_id=node_output_text_by_node_id,
        )
        if not successor_nodes:
            return

        existing_nodes = self.node_run_repository.list_by_run_and_node_ids(
            tenant_id=node_run.tenant_id,
            run_id=node_run.run_id,
            node_ids=[item.node_id for item in successor_nodes],
        )
        existing_node_counts: dict[str, int] = {}
        for item in existing_nodes:
            existing_node_counts[item.node_id] = existing_node_counts.get(item.node_id, 0) + 1

        for successor in successor_nodes:
            successor_config = derive_node_config(workflow_version, successor.node_id)
            if not self._is_join_ready(
                workflow_version=workflow_version,
                run_node_runs=self.node_run_repository.list_by_run(
                    tenant_id=node_run.tenant_id,
                    run_id=node_run.run_id,
                ),
                successor_node_id=successor.node_id,
                successor_node_type=successor.node_type,
                successor_config=successor_config,
            ):
                self._log_event(
                    tenant_id=node_run.tenant_id,
                    run_id=node_run.run_id,
                    node_run_id=None,
                    event_type="join_waiting",
                    message=f"join node waiting for predecessors: {successor.node_id}",
                    detail_json={
                        "node_id": successor.node_id,
                        "source_node_id": node_run.node_id,
                    },
                )
                continue
            if not self._can_schedule_repeated_node(
                successor_config,
                existing_count=existing_node_counts.get(successor.node_id, 0),
            ):
                continue
            scheduled_time, timeout_time = self._build_node_timing(successor_config)
            created = self.node_run_repository.create(
                tenant_id=node_run.tenant_id,
                run_id=node_run.run_id,
                parent_node_run_id=node_run.id,
                node_id=successor.node_id,
                node_type=successor.node_type,
                status=successor.status,
                scheduled_time=scheduled_time,
                timeout_time=timeout_time,
            )
            self._publish_node_run_to_queue(created)
            existing_node_counts[successor.node_id] = (
                existing_node_counts.get(successor.node_id, 0) + 1
            )
            self._log_event(
                tenant_id=node_run.tenant_id,
                run_id=node_run.run_id,
                node_run_id=None,
                event_type="node_queued",
                message=f"successor node queued: {successor.node_id}",
                detail_json={
                    "node_id": successor.node_id,
                    "node_type": successor.node_type,
                    "status": successor.status,
                    "source_node_id": node_run.node_id,
                },
            )

    def _build_execution_context(
        self,
        *,
        workflow_run: WorkflowRun,
        node_run: NodeRun,
        worker_key: str | None,
        node_config_json: dict[str, JSONValue] | None = None,
    ) -> NodeExecutionContextContract:
        run_state_json, node_output_json_by_node_id, node_output_text_by_node_id = (
            self._build_run_state_maps(
                tenant_id=node_run.tenant_id,
                run_id=node_run.run_id,
            )
        )
        return NodeExecutionContextContract(
            tenant_id=node_run.tenant_id,
            run_id=node_run.run_id,
            node_run_id=node_run.id,
            node_id=node_run.node_id,
            node_type=node_run.node_type,
            node_config_json=node_config_json
            if node_config_json is not None
            else self._resolve_node_config(
                tenant_id=node_run.tenant_id,
                workflow_version_id=workflow_run.workflow_version_id,
                node_id=node_run.node_id,
            ),
            run_state_json=run_state_json,
            node_output_json_by_node_id=node_output_json_by_node_id,
            node_output_text_by_node_id=node_output_text_by_node_id,
            worker_key=worker_key,
        )

    def _build_run_state_maps(
        self,
        *,
        tenant_id: str,
        run_id: str,
    ) -> tuple[
        dict[str, JSONValue],
        dict[str, dict[str, JSONValue]],
        dict[str, str],
    ]:
        node_runs = self.node_run_repository.list_by_run(tenant_id=tenant_id, run_id=run_id)
        run_state_json: dict[str, JSONValue] = {}
        node_output_json_by_node_id: dict[str, dict[str, JSONValue]] = {}
        node_output_text_by_node_id: dict[str, str] = {}

        for item in node_runs:
            if item.output_json is not None:
                node_output_json_by_node_id[item.node_id] = dict(item.output_json)

                state_updates = item.output_json.get("state_updates")
                if isinstance(state_updates, dict):
                    for state_key, state_value in state_updates.items():
                        run_state_json[str(state_key)] = state_value

            if item.output_text is not None:
                node_output_text_by_node_id[item.node_id] = item.output_text

        return run_state_json, node_output_json_by_node_id, node_output_text_by_node_id

    def _build_node_timing(
        self,
        node_config_json: dict[str, JSONValue],
    ) -> tuple[datetime, datetime | None]:
        now = datetime.utcnow()
        delay_seconds = self._read_int_value(node_config_json, "delay_seconds", default=0)
        timeout_seconds = self._read_int_value(node_config_json, "timeout_seconds", default=0)
        scheduled_time = now + timedelta(seconds=max(delay_seconds, 0))
        timeout_time = (
            scheduled_time + timedelta(seconds=timeout_seconds)
            if timeout_seconds > 0
            else None
        )
        return scheduled_time, timeout_time

    def _node_has_timed_out(self, node_run: NodeRun) -> bool:
        return node_run.timeout_time is not None and node_run.timeout_time <= datetime.utcnow()

    def _should_retry_node(
        self,
        *,
        node_run: NodeRun,
        node_config_json: dict[str, JSONValue],
    ) -> bool:
        retry_policy = self._read_dict_value(node_config_json, "retry_policy")
        max_attempts = self._read_int_value(retry_policy, "max_attempts", default=1)
        return max_attempts > node_run.attempt_no

    def _read_retry_delay_seconds(self, node_config_json: dict[str, JSONValue]) -> int:
        retry_policy = self._read_dict_value(node_config_json, "retry_policy")
        return self._read_int_value(retry_policy, "retry_delay_seconds", default=0)

    def _build_retry_timing(
        self,
        node_config_json: dict[str, JSONValue],
    ) -> tuple[datetime, datetime | None]:
        retry_time = datetime.utcnow() + timedelta(
            seconds=self._read_retry_delay_seconds(node_config_json)
        )
        timeout_seconds = self._read_int_value(node_config_json, "timeout_seconds", default=0)
        timeout_time = (
            retry_time + timedelta(seconds=timeout_seconds)
            if timeout_seconds > 0
            else None
        )
        return retry_time, timeout_time

    def _is_join_ready(
        self,
        *,
        workflow_version: WorkflowVersionContract,
        run_node_runs: list[NodeRun],
        successor_node_id: str,
        successor_node_type: str,
        successor_config: dict[str, JSONValue],
    ) -> bool:
        join_policy = self._read_string_value(successor_config, "join_policy")
        if join_policy is None and successor_node_type != "join":
            return True
        workflow = self._parse_workflow(workflow_version)
        if workflow is None:
            return True
        predecessor_ids = [
            edge.source for edge in workflow.edges if edge.target == successor_node_id
        ]
        if not predecessor_ids:
            return True
        completed_node_ids = {
            item.node_id
            for item in run_node_runs
            if item.status in {"completed", "skipped"}
        }
        if join_policy in {None, "all_completed"}:
            return all(predecessor_id in completed_node_ids for predecessor_id in predecessor_ids)
        if join_policy == "any_completed":
            return any(predecessor_id in completed_node_ids for predecessor_id in predecessor_ids)
        return True

    def _can_schedule_repeated_node(
        self,
        node_config_json: dict[str, JSONValue],
        *,
        existing_count: int,
    ) -> bool:
        if existing_count == 0:
            return True
        allow_loop = self._read_bool_value(node_config_json, "allow_loop", default=False)
        max_iterations = self._read_int_value(node_config_json, "max_iterations", default=1)
        return allow_loop and existing_count < max_iterations

    def _schedule_compensation_node(
        self,
        *,
        node_run: NodeRun,
        node_config: dict[str, JSONValue],
    ) -> None:
        compensation_node_id = self._read_string_value(node_config, "compensation_node_id")
        if compensation_node_id is None:
            compensation_config = self._read_dict_value(node_config, "compensation")
            compensation_node_id = self._read_string_value(compensation_config, "node_id")
        if compensation_node_id is None:
            return

        workflow_run = self.workflow_run_repository.get_by_id(node_run.run_id)
        if workflow_run is None:
            return
        compensation_config = self._resolve_node_config(
            tenant_id=node_run.tenant_id,
            workflow_version_id=workflow_run.workflow_version_id,
            node_id=compensation_node_id,
        )
        existing_nodes = self.node_run_repository.list_by_run_and_node_ids(
            tenant_id=node_run.tenant_id,
            run_id=node_run.run_id,
            node_ids=[compensation_node_id],
        )
        if existing_nodes and not self._can_schedule_repeated_node(
            compensation_config,
            existing_count=len(existing_nodes),
        ):
            return
        compensation_node_type = self._resolve_workflow_node_type(
            tenant_id=node_run.tenant_id,
            workflow_version_id=workflow_run.workflow_version_id,
            node_id=compensation_node_id,
        ) or "compensation"
        scheduled_time, timeout_time = self._build_node_timing(compensation_config)
        created = self.node_run_repository.create(
            tenant_id=node_run.tenant_id,
            run_id=node_run.run_id,
            parent_node_run_id=node_run.id,
            node_id=compensation_node_id,
            node_type=compensation_node_type,
            status="queued",
            scheduled_time=scheduled_time,
            timeout_time=timeout_time,
        )
        self._publish_node_run_to_queue(created)
        self._log_event(
            tenant_id=node_run.tenant_id,
            run_id=node_run.run_id,
            node_run_id=created.id,
            event_type="compensation_queued",
            message=f"compensation node queued: {compensation_node_id}",
            detail_json={
                "failed_node_id": node_run.node_id,
                "compensation_node_id": compensation_node_id,
            },
        )

    def _parse_workflow(self, workflow_version: WorkflowVersionContract):
        return parse_workflow_definition(workflow_version.dsl_json)

    def _resolve_workflow_node_type(
        self,
        *,
        tenant_id: str,
        workflow_version_id: str,
        node_id: str,
    ) -> str | None:
        if self.workflow_client is None:
            return None
        workflow_version = self.workflow_client.get_workflow_version(
            tenant_id=tenant_id,
            workflow_version_id=workflow_version_id,
        )
        workflow = self._parse_workflow(workflow_version)
        if workflow is None:
            return None
        for node in workflow.nodes:
            if node.id == node_id:
                return node.type
        return None

    def _read_string_value(self, payload: dict[str, JSONValue], key: str) -> str | None:
        value = payload.get(key)
        if isinstance(value, str) and value:
            return value
        return None

    def _read_bool_value(
        self,
        payload: dict[str, JSONValue],
        key: str,
        *,
        default: bool,
    ) -> bool:
        value = payload.get(key)
        if isinstance(value, bool):
            return value
        return default

    def _read_int_value(
        self,
        payload: dict[str, JSONValue],
        key: str,
        *,
        default: int,
    ) -> int:
        value = payload.get(key)
        if isinstance(value, int) and not isinstance(value, bool):
            return value
        return default

    def _read_dict_value(
        self,
        payload: dict[str, JSONValue],
        key: str,
    ) -> dict[str, JSONValue]:
        value = payload.get(key)
        if isinstance(value, dict):
            return {str(item_key): item_value for item_key, item_value in value.items()}
        return {}

    def _sync_workflow_run_status_from_nodes(self, *, tenant_id: str, run_id: str) -> None:
        node_runs = self.node_run_repository.list_by_run(tenant_id=tenant_id, run_id=run_id)
        if not node_runs:
            return

        self.workflow_run_repository.update_node_count(
            run_id=run_id,
            current_node_count=len(node_runs),
        )

        next_status, error_code, error_message = self._derive_run_status(node_runs)
        self.workflow_run_repository.update_status(
            run_id=run_id,
            status=next_status,
            error_code=error_code,
            error_message=error_message,
        )
        self._log_event(
            tenant_id=tenant_id,
            run_id=run_id,
            node_run_id=None,
            event_type="run_status_synced",
            message=f"workflow run status synced to {next_status}",
            detail_json={
                "status": next_status,
                "error_code": error_code,
            },
        )

    def _derive_run_status(
        self,
        node_runs: list[NodeRun],
    ) -> tuple[WorkflowRunStatus, str | None, str | None]:
        statuses = {node_run.status for node_run in node_runs}

        active_statuses: set[NodeRunStatus] = {"pending", "queued", "running"}
        if statuses.intersection(active_statuses):
            return "running", None, None

        if "failed" in statuses:
            failed_node = next((item for item in node_runs if item.status == "failed"), None)
            error_code = failed_node.error_code if failed_node is not None else None
            error_message = failed_node.error_message if failed_node is not None else None
            return "failed", error_code, error_message

        terminal_statuses: set[NodeRunStatus] = {"completed", "skipped"}
        if statuses and statuses.issubset(terminal_statuses):
            return "completed", None, None

        return "running", None, None

    def _log_event(
        self,
        *,
        tenant_id: str,
        run_id: str,
        node_run_id: str | None,
        event_type: str,
        message: str,
        detail_json: dict[str, JSONValue] | None,
        level: str = "info",
    ) -> None:
        self.execution_log_repository.create(
            tenant_id=tenant_id,
            run_id=run_id,
            node_run_id=node_run_id,
            event_type=event_type,
            level=level,
            message=message,
            detail_json=detail_json,
        )

    def _publish_event(
        self,
        *,
        event_type: str,
        payload_json: dict[str, JSONValue],
        workflow_run: WorkflowRun | None = None,
        node_run: NodeRun | None = None,
    ) -> None:
        if self.event_client is None:
            return
        tenant_id = workflow_run.tenant_id if workflow_run is not None else None
        if tenant_id is None and node_run is not None:
            tenant_id = node_run.tenant_id
        if tenant_id is None:
            return
        aggregate_id = workflow_run.id if workflow_run is not None else None
        if aggregate_id is None and node_run is not None:
            aggregate_id = node_run.id
        aggregate_type = "workflow_run" if workflow_run is not None else "node_run"
        correlation_id = workflow_run.session_id if workflow_run is not None else None
        if correlation_id is None and node_run is not None:
            correlation_id = node_run.run_id
        try:
            self.event_client.publish_event(
                EventPublishContract(
                    tenant_id=tenant_id,
                    event_type=event_type,
                    source_service="runtime-service",
                    aggregate_type=aggregate_type,
                    aggregate_id=aggregate_id,
                    correlation_id=correlation_id,
                    payload_json=payload_json,
                )
            )
        except EventServiceClientError:
            return

    def _publish_node_run_to_queue(self, node_run: NodeRun) -> None:
        if node_run.status != "queued" or self.task_queue_publisher is None:
            return
        self.task_queue_publisher.publish_runtime_node_run(
            tenant_id=node_run.tenant_id,
            node_run_id=node_run.id,
        )


def build_runtime_application_service(
    *,
    db: Session,
    settings: RuntimeServiceSettings,
) -> RuntimeApplicationService:
    redis_client = try_build_redis_client(settings.redis_url)
    return RuntimeApplicationService(
        workflow_run_repository=WorkflowRunRepository(db),
        node_run_repository=NodeRunRepository(db),
        execution_log_repository=ExecutionLogRepository(db),
        node_artifact_repository=NodeArtifactRepository(db),
        trace_span_repository=TraceSpanRepository(db),
        execution_dispatcher=build_node_execution_dispatcher_with_clients(
            code_runner_client=CodeRunnerClient(base_url=settings.code_runner_service_url),
            model_gateway_client=ModelGatewayClient(base_url=settings.model_gateway_service_url),
            tool_client=ToolServiceClient(base_url=settings.tool_service_url),
            human_client=HumanServiceClient(
                base_url=settings.human_service_url,
                timeout_seconds=settings.human_service_timeout_seconds,
            ),
            knowledge_client=KnowledgeServiceClient(
                base_url=settings.knowledge_service_url,
                timeout_seconds=settings.knowledge_service_timeout_seconds,
            ),
        ),
        workflow_client=WorkflowServiceClient(base_url=settings.workflow_service_url),
        event_client=EventServiceClient(
            base_url=settings.event_service_url,
            timeout_seconds=settings.event_service_timeout_seconds,
        ),
        task_queue_publisher=(
            TaskQueuePublisher(client=redis_client) if redis_client is not None else None
        ),
    )