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

import json
from datetime import datetime, timedelta
from typing import cast

from sqlalchemy.orm import Session

from core_events import EventPublishContract, EventServiceClient, EventServiceClientError
from core_domain import (
    AgentSkillRefContract,
    AgentToolRefContract,
    ChatCompletionRequestContract,
    ChatMessageContract,
    MemoryCreateContract,
    MemoryScopeType,
    MemorySearchRequestContract,
    MemorySearchResultContract,
)
from core_shared import JSONValue

from app.bootstrap.settings import AgentServiceSettings
from app.db.models import AgentDefinition, AgentRun, AgentToolInvocation, AgentVersion
from app.domain.repositories import (
    AgentDefinitionRepository,
    AgentRunRepository,
    AgentToolInvocationRepository,
    AgentVersionRepository,
)
from app.infrastructure.model_gateway_client import ModelGatewayClient, ModelGatewayClientError
from app.infrastructure.memory_client import MemoryClient, MemoryClientError
from app.infrastructure.skill_client import SkillServiceClient, SkillServiceClientError
from app.infrastructure.tool_client import ToolServiceClient, ToolServiceClientError
from app.schemas.agent import (
    AgentCreateRequest,
    AgentRunCreateRequest,
    AgentRunExecuteRequest,
    AgentRunStatusUpdateRequest,
    AgentStatusUpdateRequest,
    AgentVersionCreateRequest,
)


class AgentApplicationService:
    def __init__(
        self,
        *,
        agent_repository: AgentDefinitionRepository,
        agent_version_repository: AgentVersionRepository,
        agent_run_repository: AgentRunRepository,
        agent_tool_invocation_repository: AgentToolInvocationRepository,
        model_gateway_client: ModelGatewayClient | None = None,
        memory_client: MemoryClient | None = None,
        tool_client: ToolServiceClient | None = None,
        skill_client: SkillServiceClient | None = None,
        event_client: EventServiceClient | None = None,
        react_max_steps: int = 5,
    ) -> None:
        self.agent_repository = agent_repository
        self.agent_version_repository = agent_version_repository
        self.agent_run_repository = agent_run_repository
        self.agent_tool_invocation_repository = agent_tool_invocation_repository
        self.model_gateway_client = model_gateway_client
        self.memory_client = memory_client
        self.tool_client = tool_client
        self.skill_client = skill_client
        self.event_client = event_client
        self.react_max_steps = react_max_steps

    def create_agent(self, payload: AgentCreateRequest) -> AgentDefinition:
        return self.agent_repository.create(
            tenant_id=payload.tenant_id,
            code=payload.code,
            name=payload.name,
            description=payload.description,
            agent_type=payload.agent_type,
            owner_user_id=payload.owner_user_id,
            metadata_json=payload.metadata_json,
        )

    def list_agents(self, *, tenant_id: str) -> list[AgentDefinition]:
        return self.agent_repository.list_by_tenant(tenant_id=tenant_id)

    def update_agent_status(
        self,
        *,
        agent_id: str,
        payload: AgentStatusUpdateRequest,
    ) -> AgentDefinition | None:
        return self.agent_repository.update_status(
            tenant_id=payload.tenant_id,
            agent_id=agent_id,
            status=payload.status,
        )

    def create_agent_version(self, payload: AgentVersionCreateRequest) -> AgentVersion:
        agent = self.agent_repository.get_by_id(
            tenant_id=payload.tenant_id,
            agent_id=payload.agent_id,
        )
        if agent is None:
            raise ValueError(f"agent not found: {payload.agent_id}")

        return self.agent_version_repository.create(
            tenant_id=payload.tenant_id,
            agent_id=payload.agent_id,
            status=payload.status,
            role=payload.role,
            goal=payload.goal,
            system_prompt=payload.system_prompt,
            model_config_json=payload.model_config_data.model_dump(mode="json"),
            memory_policy_json=payload.memory_policy.model_dump(mode="json"),
            tool_refs_json=[item.model_dump(mode="json") for item in payload.tool_refs],
            skill_refs_json=[item.model_dump(mode="json") for item in payload.skill_refs],
        )

    def list_agent_versions(self, *, tenant_id: str, agent_id: str) -> list[AgentVersion]:
        return self.agent_version_repository.list_by_agent(tenant_id=tenant_id, agent_id=agent_id)

    def create_agent_run(self, payload: AgentRunCreateRequest) -> AgentRun:
        agent_version = self._resolve_agent_version(
            tenant_id=payload.tenant_id,
            agent_id=payload.agent_id,
            agent_version_id=payload.agent_version_id,
        )
        if agent_version is None:
            raise ValueError("published agent version not found")

        agent_run = self.agent_run_repository.create(
            tenant_id=payload.tenant_id,
            agent_id=payload.agent_id,
            agent_version_id=agent_version.id,
            session_id=payload.session_id,
            input_text=payload.input_text,
            input_json=payload.input_json,
        )
        self._publish_event(
            event_type="agent.run.created",
            agent_run=agent_run,
            payload_json={"agent_run_id": agent_run.id, "status": agent_run.status},
        )
        return agent_run

    def list_agent_runs(
        self,
        *,
        tenant_id: str,
        agent_id: str | None = None,
        session_id: str | None = None,
    ) -> list[AgentRun]:
        return self.agent_run_repository.list_by_scope(
            tenant_id=tenant_id,
            agent_id=agent_id,
            session_id=session_id,
        )

    def list_agent_tool_invocations(
        self,
        *,
        tenant_id: str,
        agent_run_id: str,
    ) -> list[AgentToolInvocation]:
        return self.agent_tool_invocation_repository.list_by_run(
            tenant_id=tenant_id,
            agent_run_id=agent_run_id,
        )

    def update_agent_run_status(
        self,
        *,
        agent_run_id: str,
        payload: AgentRunStatusUpdateRequest,
    ) -> AgentRun | None:
        entity = self.agent_run_repository.get_by_id(
            tenant_id=payload.tenant_id,
            agent_run_id=agent_run_id,
        )
        if entity is None:
            return None
        return self.agent_run_repository.update_status(
            agent_run_id=agent_run_id,
            status=payload.status,
            worker_key=payload.worker_key,
            output_text=payload.output_text,
            output_json=payload.output_json,
            error_code=payload.error_code,
            error_message=payload.error_message,
        )

    def execute_agent_run(
        self,
        *,
        agent_run_id: str,
        payload: AgentRunExecuteRequest,
    ) -> AgentRun | None:
        agent_run = self.agent_run_repository.get_by_id(
            tenant_id=payload.tenant_id,
            agent_run_id=agent_run_id,
        )
        if agent_run is None:
            return None

        agent_version = self.agent_version_repository.get_by_id(
            tenant_id=payload.tenant_id,
            agent_version_id=agent_run.agent_version_id,
        )
        if agent_version is None:
            return self.agent_run_repository.update_status(
                agent_run_id=agent_run.id,
                status="failed",
                worker_key=payload.worker_key,
                error_code="agent_version_missing",
                error_message=f"agent version not found: {agent_run.agent_version_id}",
            )

        self.agent_run_repository.update_status(
            agent_run_id=agent_run.id,
            status="running",
            worker_key=payload.worker_key,
        )

        memory_results, memory_metadata = self._read_relevant_memories(
            agent_run=agent_run,
            agent_version=agent_version,
        )
        selected_tools = self._select_tool_refs(agent_run=agent_run, agent_version=agent_version)
        selected_skills = self._select_skill_refs(agent_run=agent_run, agent_version=agent_version)
        if payload.dry_run:
            messages = self._build_chat_messages(
                agent_run=agent_run,
                agent_version=agent_version,
                memory_results=memory_results,
                capability_context=self._format_capability_plan(
                    selected_tools=selected_tools,
                    selected_skills=selected_skills,
                ),
            )
            completed_run = self.agent_run_repository.update_status(
                agent_run_id=agent_run.id,
                status="completed",
                worker_key=payload.worker_key,
                output_text=self._build_dry_run_output(
                    agent_run=agent_run,
                    agent_version=agent_version,
                ),
                output_json={
                    "dry_run": True,
                    "agent_version_id": agent_version.id,
                    "message_count": len(messages),
                    "messages": [message.model_dump(mode="json") for message in messages],
                    "selected_tool_refs": [
                        tool_ref.model_dump(mode="json") for tool_ref in selected_tools
                    ],
                    "selected_skill_refs": [
                        skill_ref.model_dump(mode="json") for skill_ref in selected_skills
                    ],
                    **memory_metadata,
                },
            )
            if completed_run is not None:
                self._publish_event(
                    event_type="agent.run.completed",
                    agent_run=completed_run,
                    payload_json={
                        "agent_run_id": completed_run.id,
                        "dry_run": True,
                        "status": completed_run.status,
                    },
                )
            return completed_run

        if self._read_bool(agent_version.model_config_json, "react_enabled", default=False):
            return self._execute_react_agent_run(
                agent_run=agent_run,
                agent_version=agent_version,
                payload=payload,
                memory_results=memory_results,
                memory_metadata=memory_metadata,
                selected_tools=selected_tools,
                selected_skills=selected_skills,
            )

        tool_invocations = self._invoke_selected_tools(
            agent_run=agent_run,
            agent_version=agent_version,
            selected_tools=selected_tools,
        )
        skill_invocations = self._invoke_selected_skills(
            agent_run=agent_run,
            selected_skills=selected_skills,
            worker_key=payload.worker_key,
        )
        messages = self._build_chat_messages(
            agent_run=agent_run,
            agent_version=agent_version,
            memory_results=memory_results,
            capability_context=self._format_capability_results(
                tool_invocations=tool_invocations,
                skill_invocations=skill_invocations,
            ),
        )

        if self.model_gateway_client is None:
            return self.agent_run_repository.update_status(
                agent_run_id=agent_run.id,
                status="failed",
                worker_key=payload.worker_key,
                error_code="model_gateway_missing",
                error_message="model gateway client is not configured",
                output_json={
                    "tool_invocations": tool_invocations,
                    "skill_invocations": skill_invocations,
                    **memory_metadata,
                },
            )

        try:
            response = self.model_gateway_client.create_chat_completion(
                ChatCompletionRequestContract(
                    model=self._read_optional_string(agent_version.model_config_json, "model"),
                    temperature=self._read_optional_float(
                        agent_version.model_config_json,
                        "temperature",
                    ),
                    max_tokens=self._read_optional_int(
                        agent_version.model_config_json,
                        "max_tokens",
                    ),
                    messages=messages,
                    metadata_json={
                        "tenant_id": agent_run.tenant_id,
                        "agent_id": agent_run.agent_id,
                        "agent_version_id": agent_version.id,
                        "agent_run_id": agent_run.id,
                    },
                )
            )
        except ModelGatewayClientError as exc:
            return self.agent_run_repository.update_status(
                agent_run_id=agent_run.id,
                status="failed",
                worker_key=payload.worker_key,
                error_code="model_gateway_error",
                error_message=str(exc),
            )

        memory_write_metadata = self._write_interaction_memory(
            agent_run=agent_run,
            agent_version=agent_version,
            output_text=response.content,
        )
        completed_run = self.agent_run_repository.update_status(
            agent_run_id=agent_run.id,
            status="completed",
            worker_key=payload.worker_key,
            output_text=response.content,
            output_json={
                "dry_run": False,
                "agent_version_id": agent_version.id,
                "model": response.model,
                "finish_reason": response.finish_reason,
                "usage_json": response.usage_json,
                "raw_response_json": response.raw_response_json,
                "tool_invocations": tool_invocations,
                "skill_invocations": skill_invocations,
                **memory_metadata,
                **memory_write_metadata,
            },
        )
        if completed_run is not None:
            self._publish_event(
                event_type="agent.run.completed",
                agent_run=completed_run,
                payload_json={
                    "agent_run_id": completed_run.id,
                    "dry_run": False,
                    "status": completed_run.status,
                },
            )
        return completed_run

    def _publish_event(
        self,
        *,
        event_type: str,
        agent_run: AgentRun,
        payload_json: dict[str, JSONValue],
    ) -> None:
        if self.event_client is None:
            return
        try:
            self.event_client.publish_event(
                EventPublishContract(
                    tenant_id=agent_run.tenant_id,
                    event_type=event_type,
                    source_service="agent-service",
                    aggregate_type="agent_run",
                    aggregate_id=agent_run.id,
                    correlation_id=agent_run.session_id,
                    payload_json={
                        **payload_json,
                        "agent_id": agent_run.agent_id,
                        "agent_version_id": agent_run.agent_version_id,
                    },
                )
            )
        except EventServiceClientError:
            return

    def _execute_react_agent_run(
        self,
        *,
        agent_run: AgentRun,
        agent_version: AgentVersion,
        payload: AgentRunExecuteRequest,
        memory_results: list[MemorySearchResultContract],
        memory_metadata: dict[str, JSONValue],
        selected_tools: list[AgentToolRefContract],
        selected_skills: list[AgentSkillRefContract],
    ) -> AgentRun | None:
        if self.model_gateway_client is None:
            return self.agent_run_repository.update_status(
                agent_run_id=agent_run.id,
                status="failed",
                worker_key=payload.worker_key,
                error_code="model_gateway_missing",
                error_message="model gateway client is not configured",
            )

        skill_invocations = self._invoke_selected_skills(
            agent_run=agent_run,
            selected_skills=selected_skills,
            worker_key=payload.worker_key,
        )
        messages = self._build_chat_messages(
            agent_run=agent_run,
            agent_version=agent_version,
            memory_results=memory_results,
            capability_context=self._format_react_instruction(
                selected_tools=selected_tools,
                skill_invocations=skill_invocations,
            ),
        )
        react_steps: list[dict[str, JSONValue]] = []
        tool_invocations: list[dict[str, JSONValue]] = []
        final_answer: str | None = None

        max_steps = self._read_int(
            agent_version.model_config_json,
            "react_max_steps",
            default=self.react_max_steps,
        )
        for step_index in range(max(max_steps, 1)):
            try:
                response = self.model_gateway_client.create_chat_completion(
                    self._build_chat_completion_request(
                        agent_run=agent_run,
                        agent_version=agent_version,
                        messages=messages,
                    )
                )
            except ModelGatewayClientError as exc:
                return self.agent_run_repository.update_status(
                    agent_run_id=agent_run.id,
                    status="failed",
                    worker_key=payload.worker_key,
                    error_code="model_gateway_error",
                    error_message=str(exc),
                    output_json={
                        "react_steps": react_steps,
                        "tool_invocations": tool_invocations,
                        "skill_invocations": skill_invocations,
                        **memory_metadata,
                    },
                )

            action = self._parse_react_action(response.content)
            react_step: dict[str, JSONValue] = {
                "step_index": step_index,
                "model_content": response.content,
                "action": action,
            }
            react_steps.append(react_step)
            if action.get("action") == "finish":
                answer_value = action.get("answer")
                final_answer = answer_value if isinstance(answer_value, str) else response.content
                break

            if action.get("action") != "tool":
                final_answer = response.content
                break

            tool_code = action.get("tool_code")
            matching_tools = [
                item for item in selected_tools if item.tool_code == tool_code
            ]
            if not matching_tools:
                observation = f"tool not available: {tool_code}"
                react_step["observation"] = observation
                messages.append(ChatMessageContract(role="assistant", content=response.content))
                messages.append(ChatMessageContract(role="user", content=observation))
                continue

            tool_input = action.get("input_json")
            original_input_json = agent_run.input_json
            if isinstance(tool_input, dict):
                agent_run.input_json = {
                    str(item_key): item_value for item_key, item_value in tool_input.items()
                }
            current_invocations = self._invoke_selected_tools(
                agent_run=agent_run,
                agent_version=agent_version,
                selected_tools=matching_tools[:1],
            )
            agent_run.input_json = original_input_json
            tool_invocations.extend(current_invocations)
            observation = self._format_react_observation(current_invocations)
            react_step["observation"] = observation
            messages.append(ChatMessageContract(role="assistant", content=response.content))
            messages.append(ChatMessageContract(role="user", content=observation))

        if final_answer is None:
            final_answer = "ReAct loop reached max steps without a final answer."

        memory_write_metadata = self._write_interaction_memory(
            agent_run=agent_run,
            agent_version=agent_version,
            output_text=final_answer,
        )
        completed_run = self.agent_run_repository.update_status(
            agent_run_id=agent_run.id,
            status="completed",
            worker_key=payload.worker_key,
            output_text=final_answer,
            output_json={
                "dry_run": False,
                "agent_version_id": agent_version.id,
                "react_enabled": True,
                "react_steps": react_steps,
                "tool_invocations": tool_invocations,
                "skill_invocations": skill_invocations,
                **memory_metadata,
                **memory_write_metadata,
            },
        )
        if completed_run is not None:
            self._publish_event(
                event_type="agent.run.completed",
                agent_run=completed_run,
                payload_json={
                    "agent_run_id": completed_run.id,
                    "react_enabled": True,
                    "status": completed_run.status,
                },
            )
        return completed_run

    def execute_next_claimed_agent_run(
        self,
        *,
        worker_key: str,
        lease_seconds: int,
        dry_run: bool,
        redis_client: object | None = None,
    ) -> tuple[AgentRun, int] | None:
        released_lease_count = self.agent_run_repository.release_expired_leases(
            now_time=datetime.utcnow(),
        )
        claimed_agent_run = self.agent_run_repository.claim_next_queued(
            worker_key=worker_key,
            lease_expire_time=datetime.utcnow() + timedelta(seconds=lease_seconds),
        )
        if claimed_agent_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"agent-run:{claimed_agent_run.id}:lock",
                ttl_seconds=lease_seconds,
            )
            if not lock.acquire():
                return None
            idempotency_store = IdempotencyStore(
                client=redis_client,
                prefix="agent-run-idempotency",
            )
            if not idempotency_store.begin(key=claimed_agent_run.id):
                lock.release()
                return None
        else:
            lock = None
            idempotency_store = None

        try:
            result = self.execute_agent_run(
                agent_run_id=claimed_agent_run.id,
                payload=AgentRunExecuteRequest(
                    tenant_id=claimed_agent_run.tenant_id,
                    worker_key=worker_key,
                    dry_run=dry_run,
                ),
            )
            if idempotency_store is not None and result is not None:
                idempotency_store.complete(
                    key=claimed_agent_run.id,
                    result={"status": result.status, "agent_run_id": result.id},
                )
        finally:
            if lock is not None:
                lock.release()
        if result is None:
            return None
        return result, released_lease_count

    def _resolve_agent_version(
        self,
        *,
        tenant_id: str,
        agent_id: str,
        agent_version_id: str | None,
    ) -> AgentVersion | None:
        if agent_version_id is not None:
            return self.agent_version_repository.get_by_id(
                tenant_id=tenant_id,
                agent_version_id=agent_version_id,
            )
        return self.agent_version_repository.get_latest_published(
            tenant_id=tenant_id,
            agent_id=agent_id,
        )

    def _build_chat_messages(
        self,
        *,
        agent_run: AgentRun,
        agent_version: AgentVersion,
        memory_results: list[MemorySearchResultContract] | None = None,
        capability_context: str | None = None,
    ) -> list[ChatMessageContract]:
        messages = [
            ChatMessageContract(role="system", content=agent_version.system_prompt),
        ]
        if agent_version.goal:
            messages.append(
                ChatMessageContract(role="system", content=f"Goal: {agent_version.goal}")
            )
        if memory_results:
            messages.append(
                ChatMessageContract(
                    role="system",
                    content=self._format_memory_context(memory_results),
                )
            )
        if capability_context:
            messages.append(ChatMessageContract(role="system", content=capability_context))
        if agent_run.input_text:
            messages.append(ChatMessageContract(role="user", content=agent_run.input_text))
        if agent_run.input_json:
            messages.append(
                ChatMessageContract(
                    role="user",
                    content=f"Structured input: {agent_run.input_json}",
                )
            )
        return messages

    def _select_tool_refs(
        self,
        *,
        agent_run: AgentRun,
        agent_version: AgentVersion,
    ) -> list[AgentToolRefContract]:
        input_preview = self._build_input_preview(agent_run)
        selected: list[AgentToolRefContract] = []
        for item in agent_version.tool_refs_json:
            ref = AgentToolRefContract.model_validate(item)
            if (
                ref.required
                or self._read_bool(ref.config_json, "auto_invoke", default=False)
                or self._matches_selection_keywords(ref.config_json, input_preview)
            ):
                selected.append(ref)
        return selected

    def _select_skill_refs(
        self,
        *,
        agent_run: AgentRun,
        agent_version: AgentVersion,
    ) -> list[AgentSkillRefContract]:
        input_preview = self._build_input_preview(agent_run)
        selected: list[AgentSkillRefContract] = []
        for item in agent_version.skill_refs_json:
            ref = AgentSkillRefContract.model_validate(item)
            auto_invoke = self._read_bool(ref.config_json, "auto_invoke", default=True)
            if auto_invoke or self._matches_selection_keywords(ref.config_json, input_preview):
                selected.append(ref)
        return selected

    def _invoke_selected_tools(
        self,
        *,
        agent_run: AgentRun,
        agent_version: AgentVersion,
        selected_tools: list[AgentToolRefContract],
    ) -> list[dict[str, JSONValue]]:
        invocations: list[dict[str, JSONValue]] = []
        for ref in selected_tools:
            invocation = self.agent_tool_invocation_repository.create(
                tenant_id=agent_run.tenant_id,
                agent_run_id=agent_run.id,
                agent_id=agent_run.agent_id,
                agent_version_id=agent_version.id,
                tool_code=ref.tool_code,
                tool_binding_id=ref.tool_binding_id,
                status="selected",
                input_json=agent_run.input_json or {},
            )
            if ref.tool_binding_id is None:
                self.agent_tool_invocation_repository.update_status(
                    invocation_id=invocation.id,
                    status="skipped",
                    reason="tool_binding_id_missing",
                )
                invocations.append(
                    {
                        "status": "skipped",
                        "reason": "tool_binding_id_missing",
                        "tool_code": ref.tool_code,
                    }
                )
                continue
            if self.tool_client is None:
                self.agent_tool_invocation_repository.update_status(
                    invocation_id=invocation.id,
                    status="failed",
                    reason="tool_client_missing",
                    error_message="tool client is not configured",
                )
                invocations.append(
                    {
                        "status": "failed",
                        "reason": "tool_client_missing",
                        "tool_binding_id": ref.tool_binding_id,
                    }
                )
                continue
            try:
                self.agent_tool_invocation_repository.update_status(
                    invocation_id=invocation.id,
                    status="running",
                )
                detail = self.tool_client.get_tool_binding_detail(
                    tenant_id=agent_run.tenant_id,
                    binding_id=ref.tool_binding_id,
                )
                if not detail.binding.enabled:
                    self.agent_tool_invocation_repository.update_status(
                        invocation_id=invocation.id,
                        status="failed",
                        reason="tool_binding_disabled",
                        error_message="tool binding is disabled",
                    )
                    invocations.append(
                        {
                            "status": "failed",
                            "reason": "tool_binding_disabled",
                            "tool_binding_id": ref.tool_binding_id,
                        }
                    )
                    continue
                if detail.tool_definition.tool_type != "http":
                    self.agent_tool_invocation_repository.update_status(
                        invocation_id=invocation.id,
                        status="skipped",
                        reason="unsupported_tool_type",
                        error_message=detail.tool_definition.tool_type,
                    )
                    invocations.append(
                        {
                            "status": "skipped",
                            "reason": "unsupported_tool_type",
                            "tool_type": detail.tool_definition.tool_type,
                            "tool_binding_id": ref.tool_binding_id,
                        }
                    )
                    continue
                output_text, output_json = self.tool_client.invoke_http_tool(
                    detail=detail,
                    input_json=agent_run.input_json or {},
                    config_json=ref.config_json,
                )
            except ToolServiceClientError as exc:
                self.agent_tool_invocation_repository.update_status(
                    invocation_id=invocation.id,
                    status="failed",
                    reason="tool_service_error",
                    error_message=str(exc),
                )
                invocations.append(
                    {
                        "status": "failed",
                        "reason": str(exc),
                        "tool_binding_id": ref.tool_binding_id,
                    }
                )
                continue

            self.agent_tool_invocation_repository.update_status(
                invocation_id=invocation.id,
                status="completed",
                output_text=output_text,
                output_json=output_json,
            )
            invocations.append(
                {
                    "status": "completed",
                    "tool_binding_id": ref.tool_binding_id,
                    "tool_code": detail.tool_definition.code,
                    "output_text": output_text,
                    "output_json": output_json,
                }
            )
        return invocations

    def _invoke_selected_skills(
        self,
        *,
        agent_run: AgentRun,
        selected_skills: list[AgentSkillRefContract],
        worker_key: str | None,
    ) -> list[dict[str, JSONValue]]:
        invocations: list[dict[str, JSONValue]] = []
        for ref in selected_skills:
            if self.skill_client is None:
                invocations.append(
                    {
                        "status": "failed",
                        "reason": "skill_client_missing",
                        "skill_id": ref.skill_id,
                        "skill_code": ref.skill_code,
                    }
                )
                continue

            skill_id = ref.skill_id or self._resolve_skill_id_by_code(
                tenant_id=agent_run.tenant_id,
                skill_code=ref.skill_code,
            )
            if skill_id is None:
                invocations.append(
                    {
                        "status": "failed",
                        "reason": "skill_id_missing",
                        "skill_code": ref.skill_code,
                    }
                )
                continue

            try:
                created_run = self.skill_client.create_skill_run(
                    tenant_id=agent_run.tenant_id,
                    skill_id=skill_id,
                    skill_version_id=self._read_optional_string(
                        ref.config_json,
                        "skill_version_id",
                    ),
                    installation_id=self._read_optional_string(
                        ref.config_json,
                        "installation_id",
                    ),
                    input_json=self._build_skill_input_json(agent_run=agent_run, ref=ref),
                )
                executed_run = self.skill_client.execute_skill_run(
                    tenant_id=agent_run.tenant_id,
                    skill_run_id=created_run.id,
                    worker_key=worker_key,
                )
            except SkillServiceClientError as exc:
                invocations.append(
                    {
                        "status": "failed",
                        "reason": str(exc),
                        "skill_id": skill_id,
                        "skill_code": ref.skill_code,
                    }
                )
                continue

            invocations.append(
                {
                    "status": executed_run.status,
                    "skill_id": skill_id,
                    "skill_code": ref.skill_code,
                    "skill_run_id": executed_run.id,
                    "output_text": executed_run.output_text,
                    "output_json": executed_run.output_json,
                    "error_code": executed_run.error_code,
                    "error_message": executed_run.error_message,
                }
            )
        return invocations

    def _format_capability_plan(
        self,
        *,
        selected_tools: list[AgentToolRefContract],
        selected_skills: list[AgentSkillRefContract],
    ) -> str:
        return (
            "Selected capability plan before model call:\n"
            f"Tools: {[item.model_dump(mode='json') for item in selected_tools]}\n"
            f"Skills: {[item.model_dump(mode='json') for item in selected_skills]}"
        )

    def _format_capability_results(
        self,
        *,
        tool_invocations: list[dict[str, JSONValue]],
        skill_invocations: list[dict[str, JSONValue]],
    ) -> str | None:
        if not tool_invocations and not skill_invocations:
            return None
        return (
            "Capability invocation results before model call:\n"
            f"Tools: {tool_invocations}\n"
            f"Skills: {skill_invocations}"
        )

    def _format_react_instruction(
        self,
        *,
        selected_tools: list[AgentToolRefContract],
        skill_invocations: list[dict[str, JSONValue]],
    ) -> str:
        return (
            "Use ReAct JSON only. Respond with one JSON object per turn.\n"
            "To call a tool: "
            '{"action":"tool","tool_code":"code","input_json":{...}}\n'
            "To finish: "
            '{"action":"finish","answer":"final answer"}\n'
            f"Available tools: {[item.model_dump(mode='json') for item in selected_tools]}\n"
            f"Pre-run skill results: {skill_invocations}"
        )

    def _format_react_observation(
        self,
        tool_invocations: list[dict[str, JSONValue]],
    ) -> str:
        return f"Observation: {tool_invocations}"

    def _parse_react_action(self, content: str) -> dict[str, JSONValue]:
        try:
            value = json.loads(content)
        except json.JSONDecodeError:
            start_index = content.find("{")
            end_index = content.rfind("}")
            if start_index < 0 or end_index <= start_index:
                return {"action": "finish", "answer": content}
            try:
                value = json.loads(content[start_index : end_index + 1])
            except json.JSONDecodeError:
                return {"action": "finish", "answer": content}
        if not isinstance(value, dict):
            return {"action": "finish", "answer": content}
        return {str(item_key): item_value for item_key, item_value in value.items()}

    def _build_chat_completion_request(
        self,
        *,
        agent_run: AgentRun,
        agent_version: AgentVersion,
        messages: list[ChatMessageContract],
    ) -> ChatCompletionRequestContract:
        return ChatCompletionRequestContract(
            model=self._read_optional_string(agent_version.model_config_json, "model"),
            temperature=self._read_optional_float(
                agent_version.model_config_json,
                "temperature",
            ),
            max_tokens=self._read_optional_int(
                agent_version.model_config_json,
                "max_tokens",
            ),
            messages=messages,
            metadata_json={
                "tenant_id": agent_run.tenant_id,
                "agent_id": agent_run.agent_id,
                "agent_version_id": agent_version.id,
                "agent_run_id": agent_run.id,
            },
        )

    def _build_skill_input_json(
        self,
        *,
        agent_run: AgentRun,
        ref: AgentSkillRefContract,
    ) -> dict[str, JSONValue]:
        input_json: dict[str, JSONValue] = dict(agent_run.input_json or {})
        if agent_run.input_text:
            input_json.setdefault("input_text", agent_run.input_text)
        configured_input = ref.config_json.get("input_json")
        if isinstance(configured_input, dict):
            input_json.update(
                {str(item_key): item_value for item_key, item_value in configured_input.items()}
            )
        return input_json

    def _resolve_skill_id_by_code(self, *, tenant_id: str, skill_code: str | None) -> str | None:
        if skill_code is None or self.skill_client is None:
            return None
        try:
            skills = self.skill_client.list_skills(tenant_id=tenant_id)
        except SkillServiceClientError:
            return None
        for skill in skills:
            if skill.code == skill_code:
                return skill.id
        return None

    def _build_input_preview(self, agent_run: AgentRun) -> str:
        return f"{agent_run.input_text or ''} {agent_run.input_json or {}}".lower()

    def _matches_selection_keywords(
        self,
        config_json: dict[str, JSONValue],
        input_preview: str,
    ) -> bool:
        keywords = config_json.get("selection_keywords")
        if not isinstance(keywords, list):
            return False
        return any(
            isinstance(keyword, str) and keyword.lower() in input_preview
            for keyword in keywords
        )

    def _build_dry_run_output(self, *, agent_run: AgentRun, agent_version: AgentVersion) -> str:
        input_preview = agent_run.input_text or str(agent_run.input_json or {})
        return (
            f"[dry-run] Agent role={agent_version.role} "
            f"version={agent_version.version_no} received: {input_preview}"
        )

    def _read_optional_string(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_optional_float(self, payload: dict[str, JSONValue], key: str) -> float | None:
        value = payload.get(key)
        if isinstance(value, (int, float)) and not isinstance(value, bool):
            return float(value)
        return None

    def _read_optional_int(self, payload: dict[str, JSONValue], key: str) -> int | None:
        value = payload.get(key)
        if isinstance(value, int) and not isinstance(value, bool):
            return value
        return None

    def _read_relevant_memories(
        self,
        *,
        agent_run: AgentRun,
        agent_version: AgentVersion,
    ) -> tuple[list[MemorySearchResultContract], dict[str, JSONValue]]:
        if self.memory_client is None:
            return [], {"memory_read_enabled": False, "memory_read_reason": "client_missing"}
        if not self._read_bool(agent_version.memory_policy_json, "enabled", default=True):
            return [], {"memory_read_enabled": False, "memory_read_reason": "policy_disabled"}

        query = agent_run.input_text or str(agent_run.input_json or "")
        if not query:
            return [], {"memory_read_enabled": True, "memory_read_count": 0}

        scope = self._resolve_memory_scope(agent_run=agent_run, agent_version=agent_version)
        if scope is None:
            return [], {
                "memory_read_enabled": True,
                "memory_read_count": 0,
                "memory_read_reason": "scope_unavailable",
            }

        scope_type, scope_id = scope
        try:
            results = self.memory_client.search_memories(
                MemorySearchRequestContract(
                    tenant_id=agent_run.tenant_id,
                    query=query,
                    scope_type=scope_type,
                    scope_id=scope_id,
                    owner_agent_id=agent_run.agent_id,
                    session_id=agent_run.session_id,
                    limit=self._read_int(
                        agent_version.memory_policy_json,
                        "read_top_k",
                        default=8,
                    ),
                )
            )
        except MemoryClientError as exc:
            return [], {
                "memory_read_enabled": True,
                "memory_read_count": 0,
                "memory_read_error": str(exc),
            }

        return results, {
            "memory_read_enabled": True,
            "memory_read_count": len(results),
            "memory_scope_type": scope_type,
            "memory_scope_id": scope_id,
        }

    def _write_interaction_memory(
        self,
        *,
        agent_run: AgentRun,
        agent_version: AgentVersion,
        output_text: str,
    ) -> dict[str, JSONValue]:
        if self.memory_client is None:
            return {"memory_write_enabled": False, "memory_write_reason": "client_missing"}
        if not self._read_bool(agent_version.memory_policy_json, "write_enabled", default=True):
            return {"memory_write_enabled": False, "memory_write_reason": "policy_disabled"}

        scope = self._resolve_memory_scope(agent_run=agent_run, agent_version=agent_version)
        if scope is None:
            return {"memory_write_enabled": True, "memory_write_reason": "scope_unavailable"}

        scope_type, scope_id = scope
        try:
            memory = self.memory_client.create_memory(
                MemoryCreateContract(
                    tenant_id=agent_run.tenant_id,
                    scope_type=scope_type,
                    scope_id=scope_id,
                    memory_type="conversation",
                    content_text=self._format_interaction_memory(
                        agent_run=agent_run,
                        output_text=output_text,
                    ),
                    content_json={
                        "agent_run_id": agent_run.id,
                        "agent_version_id": agent_version.id,
                        "input_text": agent_run.input_text,
                        "output_text": output_text,
                    },
                    metadata_json={
                        "source": "agent-service",
                        "role": agent_version.role,
                        "version_no": agent_version.version_no,
                    },
                    owner_agent_id=agent_run.agent_id,
                    session_id=agent_run.session_id,
                    source_ref=f"agent_run:{agent_run.id}",
                    importance_score=self._read_nested_int(
                        agent_version.memory_policy_json,
                        "config_json",
                        "write_importance_score",
                        default=50,
                    ),
                )
            )
        except MemoryClientError as exc:
            return {
                "memory_write_enabled": True,
                "memory_write_error": str(exc),
            }

        return {
            "memory_write_enabled": True,
            "memory_written_id": memory.id,
            "memory_scope_type": scope_type,
            "memory_scope_id": scope_id,
        }

    def _resolve_memory_scope(
        self,
        *,
        agent_run: AgentRun,
        agent_version: AgentVersion,
    ) -> tuple[MemoryScopeType, str] | None:
        scope_value = self._read_optional_string(
            agent_version.memory_policy_json,
            "memory_scope",
        ) or "session"
        if scope_value == "tenant":
            return "tenant", agent_run.tenant_id
        if scope_value == "agent":
            return "agent", agent_run.agent_id
        if scope_value == "session" and agent_run.session_id:
            return "session", agent_run.session_id
        if scope_value == "user":
            user_id = self._read_input_json_string(agent_run=agent_run, key="user_id")
            if user_id is not None:
                return "user", user_id
        if scope_value == "team":
            team_id = self._read_input_json_string(agent_run=agent_run, key="team_id")
            if team_id is not None:
                return "team", team_id
        return None

    def _format_memory_context(self, memory_results: list[MemorySearchResultContract]) -> str:
        lines = ["Relevant memories:"]
        for index, result in enumerate(memory_results, start=1):
            lines.append(f"{index}. {result.item.content_text}")
        return "\n".join(lines)

    def _format_interaction_memory(self, *, agent_run: AgentRun, output_text: str) -> str:
        input_text = agent_run.input_text or str(agent_run.input_json or {})
        return f"User input: {input_text}\nAgent output: {output_text}"

    def _read_bool(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(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_nested_int(
        self,
        payload: dict[str, JSONValue],
        parent_key: str,
        child_key: str,
        *,
        default: int,
    ) -> int:
        parent_value = payload.get(parent_key)
        if not isinstance(parent_value, dict):
            return default
        return self._read_int(
            cast(dict[str, JSONValue], parent_value),
            child_key,
            default=default,
        )

    def _read_input_json_string(self, *, agent_run: AgentRun, key: str) -> str | None:
        if agent_run.input_json is None:
            return None
        value = agent_run.input_json.get(key)
        if isinstance(value, str) and value:
            return value
        return None


def build_agent_application_service(
    *,
    db: Session,
    settings: AgentServiceSettings,
) -> AgentApplicationService:
    return AgentApplicationService(
        agent_repository=AgentDefinitionRepository(db),
        agent_version_repository=AgentVersionRepository(db),
        agent_run_repository=AgentRunRepository(db),
        agent_tool_invocation_repository=AgentToolInvocationRepository(db),
        model_gateway_client=ModelGatewayClient(
            base_url=settings.model_gateway_service_url,
            timeout_seconds=settings.model_gateway_timeout_seconds,
        ),
        memory_client=MemoryClient(
            base_url=settings.memory_service_url,
            timeout_seconds=settings.memory_service_timeout_seconds,
        ),
        tool_client=ToolServiceClient(
            base_url=settings.tool_service_url,
            timeout_seconds=settings.tool_service_timeout_seconds,
        ),
        skill_client=SkillServiceClient(
            base_url=settings.skill_service_url,
            timeout_seconds=settings.skill_service_timeout_seconds,
        ),
        event_client=EventServiceClient(
            base_url=settings.event_service_url,
            timeout_seconds=settings.event_service_timeout_seconds,
        ),
        react_max_steps=settings.react_max_steps,
    )