feat(api): Agent App type S1 — AppMode.AGENT + create flow + binding (#36829)

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-authored-by: autofix-ci[bot] <114827586+autofix-ci[bot]@users.noreply.github.com>
This commit is contained in:
zyssyz123
2026-06-02 03:50:10 +00:00
committed by GitHub
co-authored by Claude Opus 4.7 autofix-ci[bot]
parent e530e84772
commit e35d23c3cb
62 changed files with 3702 additions and 347 deletions
+3 -1
View File
@@ -237,7 +237,9 @@ class EasyUIBasedAppConfig(AppConfig):
"""
app_model_config_from: EasyUIBasedAppModelConfigFrom
app_model_config_id: str
# Optional: an Agent App has no legacy app_model_config row, so the id may be
# absent (persistence then stores NULL for the conversation's id).
app_model_config_id: str | None = None
app_model_config_dict: dict[str, Any]
model: ModelConfigEntity
prompt_template: PromptTemplateEntity
@@ -0,0 +1,106 @@
"""Build the EasyUI-style app config for an Agent App from its Agent Soul.
An Agent App has no legacy ``app_model_config``: its model / prompt live in the
bound Agent Soul snapshot. To ride the existing chat message + SSE pipeline we
synthesize an ``app_model_config``-shaped dict from the Soul (model + system
prompt) plus any app-level feature flags (opening statement, follow-up, …)
stored on ``app_model_config`` when present, then reuse the same sub-managers
the chat app type uses.
"""
from typing import Any, cast
from core.app.app_config.base_app_config_manager import BaseAppConfigManager
from core.app.app_config.common.sensitive_word_avoidance.manager import SensitiveWordAvoidanceConfigManager
from core.app.app_config.easy_ui_based_app.dataset.manager import DatasetConfigManager
from core.app.app_config.easy_ui_based_app.model_config.manager import ModelConfigManager
from core.app.app_config.easy_ui_based_app.prompt_template.manager import PromptTemplateConfigManager
from core.app.app_config.easy_ui_based_app.variables.manager import BasicVariablesConfigManager
from core.app.app_config.entities import (
EasyUIBasedAppConfig,
EasyUIBasedAppModelConfigFrom,
PromptTemplateEntity,
)
from models.agent_config_entities import AgentSoulConfig
from models.model import App, AppMode, AppModelConfig, AppModelConfigDict, Conversation
class AgentAppConfig(EasyUIBasedAppConfig):
"""Agent App config entity (EasyUI-shaped so it rides the chat pipeline).
``app_model_config_id`` is inherited as ``str | None``: an Agent App may have
no legacy ``app_model_config`` row, in which case persistence stores ``NULL``
for the conversation's ``app_model_config_id``.
"""
class AgentAppConfigManager(BaseAppConfigManager):
@classmethod
def get_app_config(
cls,
*,
app_model: App,
agent_soul: AgentSoulConfig,
app_model_config: AppModelConfig | None = None,
conversation: Conversation | None = None,
) -> AgentAppConfig:
"""Build the Agent App config from the Agent Soul (+ optional feature flags)."""
config_dict = cls._synthesize_config_dict(agent_soul, app_model_config)
# The synthesized dict is shaped like an app_model_config; the EasyUI
# sub-managers type their param as AppModelConfigDict (a TypedDict).
typed_config = cast(AppModelConfigDict, config_dict)
app_mode = AppMode.value_of(app_model.mode)
app_config = AgentAppConfig(
tenant_id=app_model.tenant_id,
app_id=app_model.id,
app_mode=app_mode,
# The config is derived from the Agent Soul snapshot, not a legacy
# app_model_config row; the id is informational only.
app_model_config_from=EasyUIBasedAppModelConfigFrom.APP_LATEST_CONFIG,
app_model_config_id=app_model_config.id if app_model_config else None,
app_model_config_dict=config_dict,
model=ModelConfigManager.convert(config=typed_config),
prompt_template=PromptTemplateConfigManager.convert(config=typed_config),
sensitive_word_avoidance=SensitiveWordAvoidanceConfigManager.convert(config=config_dict),
dataset=DatasetConfigManager.convert(config=typed_config),
additional_features=cls.convert_features(config_dict, app_mode),
)
app_config.variables, app_config.external_data_variables = BasicVariablesConfigManager.convert(
config=typed_config
)
return app_config
@staticmethod
def _synthesize_config_dict(
agent_soul: AgentSoulConfig,
app_model_config: AppModelConfig | None,
) -> dict[str, Any]:
"""Shape a Soul + feature flags into an ``app_model_config``-style dict.
Feature flags (opening statement / follow-up / tts / stt / citations /
moderation / annotation) come from ``app_model_config`` when present
(Q3: stored there), otherwise defaults; model + prompt always come from
the Agent Soul (the single source of truth for those).
"""
base: dict[str, Any] = dict(app_model_config.to_dict()) if app_model_config else {}
model = agent_soul.model
if model is not None:
base["model"] = {
"provider": model.model_provider,
"name": model.model,
"mode": "chat",
"completion_params": dict(model.model_settings or {}),
}
# The Agent Soul system prompt rides the EasyUI "simple" prompt slot; the
# agent backend is the real prompt authority, this only feeds the chat
# pipeline's bookkeeping (token counting, persistence).
base["prompt_type"] = PromptTemplateEntity.PromptType.SIMPLE.value
base["pre_prompt"] = agent_soul.prompt.system_prompt or ""
# Agent App takes the user message directly; no completion-style inputs form.
base.setdefault("user_input_form", [])
return base
__all__ = ["AgentAppConfig", "AgentAppConfigManager"]
@@ -0,0 +1,327 @@
"""Agent App generator: orchestrate one conversation turn for an Agent App.
Mirrors the agent_chat generator (conversation + message + queue + streamed
response over the EasyUI chat pipeline), but the backing config comes from the
bound Agent Soul and the answer is produced by ``AgentAppRunner`` calling the
dify-agent backend rather than an in-process LLM/ReAct loop.
"""
from __future__ import annotations
import contextvars
import logging
import threading
import uuid
from collections.abc import Generator, Mapping
from typing import Any
from flask import Flask, current_app
from sqlalchemy import select
from clients.agent_backend import AgentBackendRunEventAdapter
from clients.agent_backend.factory import create_agent_backend_run_client
from configs import dify_config
from constants import UUID_NIL
from core.app.app_config.easy_ui_based_app.model_config.converter import ModelConfigConverter
from core.app.apps.agent_app.app_config_manager import AgentAppConfigManager
from core.app.apps.agent_app.app_runner import AgentAppRunner
from core.app.apps.agent_app.generate_response_converter import AgentAppGenerateResponseConverter
from core.app.apps.agent_app.runtime_request_builder import AgentAppRuntimeRequestBuilder
from core.app.apps.agent_app.session_store import AgentAppRuntimeSessionStore
from core.app.apps.base_app_queue_manager import AppQueueManager, PublishFrom
from core.app.apps.exc import GenerateTaskStoppedError
from core.app.apps.message_based_app_generator import MessageBasedAppGenerator
from core.app.apps.message_based_app_queue_manager import MessageBasedAppQueueManager
from core.app.entities.app_invoke_entities import (
AgentAppGenerateEntity,
DifyRunContext,
InvokeFrom,
UserFrom,
)
from core.app.llm.model_access import build_dify_model_access
from core.ops.ops_trace_manager import TraceQueueManager
from extensions.ext_database import db
from models import Account, App, EndUser, Message
from models.agent import Agent, AgentConfigSnapshot, AgentScope, AgentSource, AgentStatus
from models.agent_config_entities import AgentSoulConfig
from services.conversation_service import ConversationService
logger = logging.getLogger(__name__)
class AgentAppGeneratorError(ValueError):
"""Raised when an Agent App turn cannot be set up."""
class AgentAppGenerator(MessageBasedAppGenerator):
def generate(
self,
*,
app_model: App,
user: Account | EndUser,
args: Mapping[str, Any],
invoke_from: InvokeFrom,
streaming: bool = True,
) -> Mapping[str, Any] | Generator[Mapping | str, None, None]:
if not streaming:
raise AgentAppGeneratorError("Agent App only supports streaming mode")
query = args.get("query")
if not isinstance(query, str) or not query.strip():
raise AgentAppGeneratorError("query is required")
query = query.replace("\x00", "")
inputs = args["inputs"]
# Resolve the bound roster Agent + its published Agent Soul snapshot.
agent, snapshot, agent_soul = self._resolve_agent(app_model)
conversation = None
conversation_id = args.get("conversation_id")
if conversation_id:
conversation = ConversationService.get_conversation(
app_model=app_model, conversation_id=conversation_id, user=user
)
# Build the EasyUI-shaped config from the Agent Soul so the chat pipeline
# can persist usage; the answer itself comes from the agent backend.
app_model_config = app_model.app_model_config
app_config = AgentAppConfigManager.get_app_config(
app_model=app_model,
agent_soul=agent_soul,
app_model_config=app_model_config,
conversation=conversation,
)
model_conf = ModelConfigConverter.convert(app_config)
trace_manager = TraceQueueManager(app_model.id, user.id if isinstance(user, Account) else user.session_id)
application_generate_entity = AgentAppGenerateEntity(
task_id=str(uuid.uuid4()),
app_config=app_config,
model_conf=model_conf,
conversation_id=conversation.id if conversation else None,
inputs=self._prepare_user_inputs(
user_inputs=inputs, variables=app_config.variables, tenant_id=app_model.tenant_id
),
query=query,
files=[],
parent_message_id=(
args.get("parent_message_id")
if invoke_from not in {InvokeFrom.SERVICE_API, InvokeFrom.OPENAPI}
else UUID_NIL
),
user_id=user.id,
stream=streaming,
invoke_from=invoke_from,
extras={"auto_generate_conversation_name": args.get("auto_generate_name", True)},
call_depth=0,
trace_manager=trace_manager,
agent_id=agent.id,
agent_config_snapshot_id=snapshot.id,
)
conversation, message = self._init_generate_records(application_generate_entity, conversation)
queue_manager = MessageBasedAppQueueManager(
task_id=application_generate_entity.task_id,
user_id=application_generate_entity.user_id,
invoke_from=application_generate_entity.invoke_from,
conversation_id=conversation.id,
app_mode=conversation.mode,
message_id=message.id,
)
context = contextvars.copy_context()
worker_thread = threading.Thread(
target=self._generate_worker,
kwargs={
"flask_app": current_app._get_current_object(), # type: ignore
"context": context,
"application_generate_entity": application_generate_entity,
"queue_manager": queue_manager,
"conversation_id": conversation.id,
"message_id": message.id,
"user_from": UserFrom.ACCOUNT if isinstance(user, Account) else UserFrom.END_USER,
},
)
worker_thread.start()
response = self._handle_response(
application_generate_entity=application_generate_entity,
queue_manager=queue_manager,
conversation=conversation,
message=message,
user=user,
stream=streaming,
)
return AgentAppGenerateResponseConverter.convert(response=response, invoke_from=invoke_from)
def _generate_worker(
self,
*,
flask_app: Flask,
context: contextvars.Context,
application_generate_entity: AgentAppGenerateEntity,
queue_manager: AppQueueManager,
conversation_id: str,
message_id: str,
user_from: UserFrom,
) -> None:
from libs.flask_utils import preserve_flask_contexts
with preserve_flask_contexts(flask_app, context_vars=context):
try:
conversation = self._get_conversation(conversation_id)
message = self._get_message(message_id)
app_config = application_generate_entity.app_config
# Apply app-level input guards (content moderation + annotation
# reply) before reaching the Agent backend, mirroring the EasyUI
# chat / agent-chat runners. These can short-circuit the turn.
app_model = db.session.get(App, app_config.app_id)
if app_model is None:
raise AgentAppGeneratorError("App not found")
handled, query = self._run_input_guards(
application_generate_entity=application_generate_entity,
app_model=app_model,
message=message,
queue_manager=queue_manager,
)
if handled:
return
dify_context = DifyRunContext(
tenant_id=app_config.tenant_id,
app_id=app_config.app_id,
user_id=application_generate_entity.user_id,
user_from=user_from,
invoke_from=application_generate_entity.invoke_from,
)
credentials_provider, _ = build_dify_model_access(dify_context)
_, _, agent_soul = self._resolve_agent_by_id(
tenant_id=app_config.tenant_id,
agent_id=application_generate_entity.agent_id,
snapshot_id=application_generate_entity.agent_config_snapshot_id,
)
runner = AgentAppRunner(
request_builder=AgentAppRuntimeRequestBuilder(credentials_provider=credentials_provider),
agent_backend_client=create_agent_backend_run_client(
base_url=dify_config.AGENT_BACKEND_BASE_URL,
use_fake=dify_config.AGENT_BACKEND_USE_FAKE,
fake_scenario=dify_config.AGENT_BACKEND_FAKE_SCENARIO,
),
event_adapter=AgentBackendRunEventAdapter(),
session_store=AgentAppRuntimeSessionStore(),
)
runner.run(
dify_context=dify_context,
agent_id=application_generate_entity.agent_id,
agent_config_snapshot_id=application_generate_entity.agent_config_snapshot_id,
agent_soul=agent_soul,
conversation_id=conversation.id,
query=query,
message_id=message.id,
model_name=application_generate_entity.model_conf.model,
queue_manager=queue_manager,
)
except GenerateTaskStoppedError:
pass
except Exception as e:
logger.exception("Unknown Error in Agent App generate worker")
queue_manager.publish_error(e, PublishFrom.APPLICATION_MANAGER)
finally:
db.session.close()
def _run_input_guards(
self,
*,
application_generate_entity: AgentAppGenerateEntity,
app_model: App,
message: Message,
queue_manager: AppQueueManager,
) -> tuple[bool, str]:
"""Apply input moderation + annotation reply before the backend call.
Returns ``(handled, query)``: when ``handled`` is True a direct answer
has already been published (a blocked/preset moderation response or a
matched annotation) and the backend turn must be skipped. Otherwise
``query`` is the possibly moderation-overridden query to send onward.
"""
from core.app.apps.agent_app.app_runner import publish_text_answer
from core.app.entities.queue_entities import QueueAnnotationReplyEvent
from core.app.features.annotation_reply.annotation_reply import AnnotationReplyFeature
from core.moderation.base import ModerationError
from core.moderation.input_moderation import InputModeration
app_config = application_generate_entity.app_config
model_name = application_generate_entity.model_conf.model
query = application_generate_entity.query
# content moderation (sensitive_word_avoidance); a blocked input yields a
# preset answer, an "overridden" action returns a sanitized query.
try:
_, _, query = InputModeration().check(
app_id=app_config.app_id,
tenant_id=app_config.tenant_id,
app_config=app_config,
inputs=dict(application_generate_entity.inputs),
query=query or "",
message_id=message.id,
trace_manager=application_generate_entity.trace_manager,
)
except ModerationError as e:
publish_text_answer(queue_manager=queue_manager, model_name=model_name, answer=str(e))
return True, query
# annotation reply: a matching annotation answers the turn deterministically.
if query:
annotation_reply = AnnotationReplyFeature().query(
app_record=app_model,
message=message,
query=query,
user_id=application_generate_entity.user_id,
invoke_from=application_generate_entity.invoke_from,
)
if annotation_reply:
queue_manager.publish(
QueueAnnotationReplyEvent(message_annotation_id=annotation_reply.id),
PublishFrom.APPLICATION_MANAGER,
)
publish_text_answer(queue_manager=queue_manager, model_name=model_name, answer=annotation_reply.content)
return True, query
return False, query
def _resolve_agent(self, app_model: App) -> tuple[Agent, AgentConfigSnapshot, AgentSoulConfig]:
agent = db.session.scalar(
select(Agent).where(
Agent.app_id == app_model.id,
Agent.scope == AgentScope.ROSTER,
Agent.source == AgentSource.AGENT_APP,
Agent.status == AgentStatus.ACTIVE,
)
)
if agent is None:
raise AgentAppGeneratorError("Agent App has no bound Agent")
return self._resolve_agent_by_id(
tenant_id=app_model.tenant_id, agent_id=agent.id, snapshot_id=agent.active_config_snapshot_id
)
@staticmethod
def _resolve_agent_by_id(
*, tenant_id: str, agent_id: str, snapshot_id: str | None
) -> tuple[Agent, AgentConfigSnapshot, AgentSoulConfig]:
agent = db.session.scalar(select(Agent).where(Agent.id == agent_id, Agent.tenant_id == tenant_id))
if agent is None:
raise AgentAppGeneratorError("Agent not found")
if not snapshot_id:
raise AgentAppGeneratorError("Agent has no published version")
snapshot = db.session.scalar(select(AgentConfigSnapshot).where(AgentConfigSnapshot.id == snapshot_id))
if snapshot is None:
raise AgentAppGeneratorError("Agent published version not found")
agent_soul = AgentSoulConfig.model_validate(snapshot.config_snapshot_dict)
return agent, snapshot, agent_soul
__all__ = ["AgentAppGenerator", "AgentAppGeneratorError"]
+200
View File
@@ -0,0 +1,200 @@
"""Agent App runner: drive one conversation turn through the dify-agent backend.
Unlike the legacy ``AgentChatAppRunner`` (which runs an in-process ReAct loop),
this runner delegates to the Agent backend: build the run request from the
Agent Soul + conversation, create the run, consume its event stream, and
republish the assistant answer as chat queue events so the existing
EasyUI chat task pipeline persists the message and streams SSE. The conversation
``session_snapshot`` is saved on success for multi-turn continuity (S3).
"""
from __future__ import annotations
import json
import logging
from typing import Any
from pydantic import JsonValue
from clients.agent_backend import (
AgentBackendError,
AgentBackendInternalEventType,
AgentBackendRunClient,
AgentBackendRunEventAdapter,
AgentBackendRunSucceededInternalEvent,
AgentBackendStreamInternalEvent,
)
from core.app.apps.agent_app.runtime_request_builder import (
AgentAppRuntimeBuildContext,
AgentAppRuntimeRequestBuilder,
)
from core.app.apps.agent_app.session_store import AgentAppRuntimeSessionStore, AgentAppSessionScope
from core.app.apps.base_app_queue_manager import AppQueueManager, PublishFrom
from core.app.apps.exc import GenerateTaskStoppedError
from core.app.entities.app_invoke_entities import DifyRunContext
from core.app.entities.queue_entities import QueueLLMChunkEvent, QueueMessageEndEvent
from graphon.model_runtime.entities.llm_entities import LLMResult, LLMResultChunk, LLMResultChunkDelta, LLMUsage
from graphon.model_runtime.entities.message_entities import AssistantPromptMessage
from models.agent_config_entities import AgentSoulConfig
logger = logging.getLogger(__name__)
def publish_text_answer(*, queue_manager: AppQueueManager, model_name: str, answer: str) -> None:
"""Publish a complete assistant answer as one chunk + message-end.
The EasyUI chat task pipeline consumes a QueueLLMChunkEvent stream followed
by a QueueMessageEndEvent; emitting the whole answer as a single chunk lets
both the backend-produced answer and short-circuited answers (moderation /
annotation reply) share the exact same persistence + SSE path.
"""
chunk = LLMResultChunk(
model=model_name,
prompt_messages=[],
delta=LLMResultChunkDelta(index=0, message=AssistantPromptMessage(content=answer)),
)
queue_manager.publish(QueueLLMChunkEvent(chunk=chunk), PublishFrom.APPLICATION_MANAGER)
queue_manager.publish(
QueueMessageEndEvent(
llm_result=LLMResult(
model=model_name,
prompt_messages=[],
message=AssistantPromptMessage(content=answer),
usage=LLMUsage.empty_usage(),
),
),
PublishFrom.APPLICATION_MANAGER,
)
class AgentAppRunner:
"""Runs one Agent App conversation turn against the Agent backend."""
def __init__(
self,
*,
request_builder: AgentAppRuntimeRequestBuilder,
agent_backend_client: AgentBackendRunClient,
event_adapter: AgentBackendRunEventAdapter,
session_store: AgentAppRuntimeSessionStore,
) -> None:
self._request_builder = request_builder
self._agent_backend_client = agent_backend_client
self._event_adapter = event_adapter
self._session_store = session_store
def run(
self,
*,
dify_context: DifyRunContext,
agent_id: str,
agent_config_snapshot_id: str,
agent_soul: AgentSoulConfig,
conversation_id: str,
query: str,
message_id: str,
model_name: str,
queue_manager: AppQueueManager,
) -> None:
scope = AgentAppSessionScope(
tenant_id=dify_context.tenant_id,
app_id=dify_context.app_id,
conversation_id=conversation_id,
agent_id=agent_id,
agent_config_snapshot_id=agent_config_snapshot_id,
)
session_snapshot = self._session_store.load_active_snapshot(scope)
runtime = self._request_builder.build(
AgentAppRuntimeBuildContext(
dify_context=dify_context,
agent_id=agent_id,
agent_config_snapshot_id=agent_config_snapshot_id,
agent_soul=agent_soul,
conversation_id=conversation_id,
user_query=query,
idempotency_key=message_id,
session_snapshot=session_snapshot,
)
)
create_response = self._agent_backend_client.create_run(runtime.request)
terminal = self._consume_stream(create_response.run_id, queue_manager=queue_manager)
if not isinstance(terminal, AgentBackendRunSucceededInternalEvent):
error = getattr(terminal, "error", None) or "Agent backend run did not complete successfully."
raise AgentBackendError(str(error))
answer = self._extract_answer(terminal.output)
self._publish_answer(queue_manager=queue_manager, model_name=model_name, answer=answer)
self._save_session(scope=scope, backend_run_id=terminal.run_id, snapshot=terminal.session_snapshot)
def _consume_stream(self, run_id: str, *, queue_manager: AppQueueManager):
terminal = None
for public_event in self._agent_backend_client.stream_events(run_id):
if queue_manager.is_stopped():
self._cancel_run(run_id)
raise GenerateTaskStoppedError()
for internal_event in self._event_adapter.adapt(public_event):
if queue_manager.is_stopped():
self._cancel_run(run_id)
raise GenerateTaskStoppedError()
if internal_event.type in (
AgentBackendInternalEventType.RUN_STARTED,
AgentBackendInternalEventType.STREAM_EVENT,
):
# Stream deltas are accumulated by the backend into the
# terminal output; token-level forwarding is an S3 refinement.
if isinstance(internal_event, AgentBackendStreamInternalEvent):
continue
continue
terminal = internal_event
break
if terminal is not None:
break
return terminal
def _cancel_run(self, run_id: str) -> None:
try:
self._agent_backend_client.cancel_run(run_id)
except Exception:
logger.warning("Failed to cancel stopped Agent App backend run: run_id=%s", run_id, exc_info=True)
def _publish_answer(self, *, queue_manager: AppQueueManager, model_name: str, answer: str) -> None:
# MVP: emit the full answer as a single chunk + message-end. The chat
# task pipeline streams the chunk over SSE and persists the message.
publish_text_answer(queue_manager=queue_manager, model_name=model_name, answer=answer)
def _save_session(self, *, scope: AgentAppSessionScope, backend_run_id: str, snapshot: Any) -> None:
try:
self._session_store.save_active_snapshot(scope=scope, backend_run_id=backend_run_id, snapshot=snapshot)
except Exception:
logger.warning(
"Failed to persist Agent App conversation session snapshot: "
"tenant_id=%s app_id=%s conversation_id=%s agent_id=%s",
scope.tenant_id,
scope.app_id,
scope.conversation_id,
scope.agent_id,
exc_info=True,
)
@staticmethod
def _extract_answer(output: JsonValue) -> str:
"""Normalize the backend's terminal output to assistant text.
Free-text Agent Apps return a plain string; if a structured output is
configured the value is a JSON object, which we serialize so the chat
message always has a string body.
"""
if isinstance(output, str):
return output
if isinstance(output, dict):
text = output.get("text")
if isinstance(text, str):
return text
return json.dumps(output, ensure_ascii=False)
return json.dumps(output, ensure_ascii=False)
__all__ = ["AgentAppRunner", "publish_text_answer"]
@@ -0,0 +1,15 @@
"""Response converter for the Agent App type.
The Agent App streams the same chatbot response shape as the chat / agent-chat
app types, so it reuses that converter wholesale; kept as a distinct subclass so
the app type owns its converter and can diverge later.
"""
from core.app.apps.agent_chat.generate_response_converter import AgentChatAppGenerateResponseConverter
class AgentAppGenerateResponseConverter(AgentChatAppGenerateResponseConverter):
pass
__all__ = ["AgentAppGenerateResponseConverter"]
@@ -0,0 +1,177 @@
"""Build dify-agent run requests for one Agent App conversation turn.
Mirrors the workflow ``WorkflowAgentRuntimeRequestBuilder`` but for the Agent
App surface: the user prompt is the chat message (no workflow-node job / no
previous-node context), and multi-turn continuity flows through the
conversation-keyed ``session_snapshot`` plus the history layer.
"""
from __future__ import annotations
from collections.abc import Mapping
from dataclasses import dataclass
from typing import Any, Protocol, cast
from agenton.compositor import CompositorSessionSnapshot
from dify_agent.layers.execution_context import DifyExecutionContextLayerConfig
from dify_agent.protocol import CreateRunRequest
from clients.agent_backend import (
AgentBackendAgentAppRunInput,
AgentBackendModelConfig,
AgentBackendRunRequestBuilder,
redact_for_agent_backend_log,
)
from core.app.entities.app_invoke_entities import DifyRunContext
from core.workflow.nodes.agent_v2.plugin_tools_builder import (
WorkflowAgentPluginToolsBuilder,
WorkflowAgentPluginToolsBuildError,
)
from models.agent_config_entities import AgentSoulConfig
from models.provider_ids import ModelProviderID
class AgentAppRuntimeRequestBuildError(ValueError):
"""Raised when Agent App state cannot be mapped to a valid run request."""
def __init__(self, error_code: str, message: str) -> None:
self.error_code = error_code
super().__init__(message)
class CredentialsProvider(Protocol):
def fetch(self, provider_name: str, model_name: str) -> dict[str, Any]: ...
@dataclass(frozen=True, slots=True)
class AgentAppRuntimeBuildContext:
dify_context: DifyRunContext
agent_id: str
agent_config_snapshot_id: str
agent_soul: AgentSoulConfig
conversation_id: str
user_query: str
idempotency_key: str
session_snapshot: CompositorSessionSnapshot | None = None
@dataclass(frozen=True, slots=True)
class AgentAppRuntimeRequest:
request: CreateRunRequest
redacted_request: dict[str, Any]
metadata: dict[str, Any]
class AgentAppRuntimeRequestBuilder:
"""Build dify-agent run requests from Agent App conversation state."""
def __init__(
self,
*,
credentials_provider: CredentialsProvider,
request_builder: AgentBackendRunRequestBuilder | None = None,
plugin_tools_builder: WorkflowAgentPluginToolsBuilder | None = None,
) -> None:
self._credentials_provider = credentials_provider
self._request_builder = request_builder or AgentBackendRunRequestBuilder()
self._plugin_tools_builder = plugin_tools_builder or WorkflowAgentPluginToolsBuilder()
def build(self, context: AgentAppRuntimeBuildContext) -> AgentAppRuntimeRequest:
agent_soul = context.agent_soul
if agent_soul.model is None:
raise AgentAppRuntimeRequestBuildError(
"agent_model_not_configured",
"Agent App requires the Agent Soul model to be configured.",
)
metadata = self._build_metadata(context)
credentials = self._credentials_provider.fetch(agent_soul.model.model_provider, agent_soul.model.model)
try:
tools_layer = self._plugin_tools_builder.build(
tenant_id=context.dify_context.tenant_id,
app_id=context.dify_context.app_id,
user_id=context.dify_context.user_id,
tools=agent_soul.tools,
invoke_from=context.dify_context.invoke_from,
)
except WorkflowAgentPluginToolsBuildError as error:
raise AgentAppRuntimeRequestBuildError(error.error_code, str(error)) from error
if tools_layer is not None:
metadata["agent_tools"] = {
"dify_tool_count": len(tools_layer.tools),
"dify_tool_names": [tool.name or tool.tool_name for tool in tools_layer.tools],
}
request = self._request_builder.build_for_agent_app(
AgentBackendAgentAppRunInput(
model=AgentBackendModelConfig(
plugin_id=self._plugin_daemon_plugin_id(
plugin_id=agent_soul.model.plugin_id,
model_provider=agent_soul.model.model_provider,
),
model_provider=self._plugin_daemon_provider_name(agent_soul.model.model_provider),
model=agent_soul.model.model,
credentials=self._normalize_credentials(credentials),
model_settings=agent_soul.model.model_settings,
),
execution_context=DifyExecutionContextLayerConfig(
tenant_id=context.dify_context.tenant_id,
user_id=context.dify_context.user_id,
app_id=context.dify_context.app_id,
conversation_id=context.conversation_id,
agent_id=context.agent_id,
agent_config_version_id=context.agent_config_snapshot_id,
invoke_from="agent_app",
),
agent_soul_prompt=agent_soul.prompt.system_prompt or None,
user_prompt=context.user_query,
tools=tools_layer,
session_snapshot=context.session_snapshot,
idempotency_key=context.idempotency_key,
metadata=metadata,
)
)
redacted = cast(dict[str, Any], redact_for_agent_backend_log(request))
return AgentAppRuntimeRequest(request=request, redacted_request=redacted, metadata=metadata)
@staticmethod
def _build_metadata(context: AgentAppRuntimeBuildContext) -> dict[str, Any]:
return {
"tenant_id": context.dify_context.tenant_id,
"app_id": context.dify_context.app_id,
"conversation_id": context.conversation_id,
"agent_id": context.agent_id,
"agent_config_snapshot_id": context.agent_config_snapshot_id,
}
@staticmethod
def _plugin_daemon_plugin_id(*, plugin_id: str, model_provider: str) -> str:
"""Return the transport plugin id expected by plugin-daemon headers."""
if plugin_id.count("/") == 1:
return plugin_id
if plugin_id:
return ModelProviderID(plugin_id).plugin_id
return ModelProviderID(model_provider).plugin_id
@staticmethod
def _plugin_daemon_provider_name(model_provider: str) -> str:
"""Return the provider name expected by plugin-daemon dispatch payloads."""
return ModelProviderID(model_provider).provider_name
@staticmethod
def _normalize_credentials(credentials: Mapping[str, Any]) -> dict[str, str | int | float | bool | None]:
normalized: dict[str, str | int | float | bool | None] = {}
for key, value in credentials.items():
if isinstance(value, str | int | float | bool) or value is None:
normalized[key] = value
else:
normalized[key] = str(value)
return normalized
__all__ = [
"AgentAppRuntimeBuildContext",
"AgentAppRuntimeRequest",
"AgentAppRuntimeRequestBuildError",
"AgentAppRuntimeRequestBuilder",
]
@@ -0,0 +1,106 @@
"""Conversation-keyed Agent backend session store for the Agent App type.
Shares the unified ``agent_runtime_sessions`` table with the workflow Agent
Node store, but owns rows with ``owner_type = conversation``: one Agent App
conversation maps to one Agent session, so multi-turn chat re-enters the same
``session_snapshot``. Cross-conversation memory (PRD Global / Per app) is a
phase-2 concern and not modeled here.
"""
from __future__ import annotations
from dataclasses import dataclass
from agenton.compositor import CompositorSessionSnapshot
from sqlalchemy import select
from core.db.session_factory import session_factory
from libs.datetime_utils import naive_utc_now
from models.agent import (
AgentRuntimeSession,
AgentRuntimeSessionOwnerType,
AgentRuntimeSessionStatus,
)
@dataclass(frozen=True, slots=True)
class AgentAppSessionScope:
"""Identity of one Agent App conversation session."""
tenant_id: str
app_id: str
conversation_id: str
agent_id: str
agent_config_snapshot_id: str
class AgentAppRuntimeSessionStore:
"""Persists Agent backend session snapshots for Agent App conversations."""
def load_active_snapshot(self, scope: AgentAppSessionScope) -> CompositorSessionSnapshot | None:
with session_factory.create_session() as session:
row = session.scalar(self._active_stmt(scope))
if row is None:
return None
return CompositorSessionSnapshot.model_validate_json(row.session_snapshot)
def save_active_snapshot(
self,
*,
scope: AgentAppSessionScope,
backend_run_id: str,
snapshot: CompositorSessionSnapshot | None,
) -> None:
if snapshot is None:
return
snapshot_json = snapshot.model_dump_json()
with session_factory.create_session() as session:
row = session.scalar(self._scope_stmt(scope))
if row is None:
row = AgentRuntimeSession(
tenant_id=scope.tenant_id,
app_id=scope.app_id,
owner_type=AgentRuntimeSessionOwnerType.CONVERSATION,
agent_id=scope.agent_id,
agent_config_snapshot_id=scope.agent_config_snapshot_id,
conversation_id=scope.conversation_id,
backend_run_id=backend_run_id,
session_snapshot=snapshot_json,
composition_layer_specs="[]",
status=AgentRuntimeSessionStatus.ACTIVE,
)
session.add(row)
else:
row.backend_run_id = backend_run_id
row.session_snapshot = snapshot_json
row.status = AgentRuntimeSessionStatus.ACTIVE
row.cleaned_at = None
session.commit()
def mark_cleaned(self, *, scope: AgentAppSessionScope, backend_run_id: str | None = None) -> None:
with session_factory.create_session() as session:
row = session.scalar(self._active_stmt(scope))
if row is None:
return
if backend_run_id is not None:
row.backend_run_id = backend_run_id
row.status = AgentRuntimeSessionStatus.CLEANED
row.cleaned_at = naive_utc_now()
session.commit()
@staticmethod
def _scope_stmt(scope: AgentAppSessionScope):
return select(AgentRuntimeSession).where(
AgentRuntimeSession.owner_type == AgentRuntimeSessionOwnerType.CONVERSATION,
AgentRuntimeSession.tenant_id == scope.tenant_id,
AgentRuntimeSession.conversation_id == scope.conversation_id,
AgentRuntimeSession.agent_id == scope.agent_id,
AgentRuntimeSession.agent_config_snapshot_id == scope.agent_config_snapshot_id,
)
@classmethod
def _active_stmt(cls, scope: AgentAppSessionScope):
return cls._scope_stmt(scope).where(AgentRuntimeSession.status == AgentRuntimeSessionStatus.ACTIVE)
__all__ = ["AgentAppRuntimeSessionStore", "AgentAppSessionScope"]
@@ -134,6 +134,10 @@ class AppQueueManager(ABC):
self._check_for_sqlalchemy_models(event.model_dump())
self._publish(event, pub_from)
def is_stopped(self) -> bool:
"""Return whether the current task has been manually stopped."""
return self._is_stopped()
@abstractmethod
def _publish(self, event: AppQueueEvent, pub_from: PublishFrom) -> None:
"""
@@ -200,6 +200,21 @@ class AgentChatAppGenerateEntity(ConversationAppGenerateEntity, EasyUIBasedAppGe
pass
class AgentAppGenerateEntity(ChatAppGenerateEntity):
"""
Agent App (new Agent app type) Generate Entity.
Subclasses ``ChatAppGenerateEntity`` so it rides the exact same EasyUI chat
pipeline (generator, task pipeline, message cycle) without widening every
accepted-entity union. The answer is produced by the dify-agent backend
rather than an in-process LLM call; ``model_conf`` is synthesized from the
bound Agent Soul model so the chat task pipeline can persist usage.
"""
agent_id: str
agent_config_snapshot_id: str
class AdvancedChatAppGenerateEntity(ConversationAppGenerateEntity):
"""
Advanced Chat Application Generate Entity.
@@ -10,6 +10,7 @@ from clients.agent_backend.request_builder import CleanupLayerSpec
from core.db.session_factory import session_factory
from libs.datetime_utils import naive_utc_now
from models.agent import (
AgentRuntimeSessionOwnerType,
WorkflowAgentRuntimeSession,
WorkflowAgentRuntimeSessionStatus,
)
@@ -83,13 +84,16 @@ class WorkflowAgentRuntimeSessionStore:
scope=WorkflowAgentSessionScope(
tenant_id=row.tenant_id,
app_id=row.app_id,
workflow_id=row.workflow_id,
# These columns are nullable on the unified runtime-session
# table (workflow_run ⊕ conversation owner), but are always
# populated for a workflow-owned row; coerce for the typed scope.
workflow_id=row.workflow_id or "",
workflow_run_id=row.workflow_run_id,
node_id=row.node_id,
node_id=row.node_id or "",
node_execution_id=row.node_execution_id or "",
binding_id=row.binding_id,
binding_id=row.binding_id or "",
agent_id=row.agent_id,
agent_config_snapshot_id=row.agent_config_snapshot_id,
agent_config_snapshot_id=row.agent_config_snapshot_id or "",
),
session_snapshot=CompositorSessionSnapshot.model_validate_json(row.session_snapshot),
backend_run_id=row.backend_run_id,
@@ -125,6 +129,7 @@ class WorkflowAgentRuntimeSessionStore:
row = WorkflowAgentRuntimeSession(
tenant_id=scope.tenant_id,
app_id=scope.app_id,
owner_type=AgentRuntimeSessionOwnerType.WORKFLOW_RUN,
workflow_id=scope.workflow_id,
workflow_run_id=scope.workflow_run_id,
node_id=scope.node_id,