Merge remote-tracking branch 'origin/main' into refactor/migrate-react-window-to-tanstack-virtual

# Conflicts:
#	web/pnpm-lock.yaml

.agents/skills/component-refactoring/SKILL.md

@@ -480,4 +480,4 @@ const useButtonState = () => {
 ### Related Skills
 
 - `frontend-testing` - For testing refactored components
-- `web/testing/testing.md` - Testing specification
+- `web/docs/test.md` - Testing specification

.agents/skills/frontend-testing/SKILL.md

@@ -7,7 +7,7 @@ description: Generate Vitest + React Testing Library tests for Dify frontend com
 
 This skill enables Claude to generate high-quality, comprehensive frontend tests for the Dify project following established conventions and best practices.
 
-> **⚠️ Authoritative Source**: This skill is derived from `web/testing/testing.md`. Use Vitest mock/timer APIs (`vi.*`).
+> **⚠️ Authoritative Source**: This skill is derived from `web/docs/test.md`. Use Vitest mock/timer APIs (`vi.*`).
 
 ## When to Apply This Skill
 
@@ -309,7 +309,7 @@ For more detailed information, refer to:
 
 ### Primary Specification (MUST follow)
 
-- **`web/testing/testing.md`** - The canonical testing specification. This skill is derived from this document.
+- **`web/docs/test.md`** - The canonical testing specification. This skill is derived from this document.
 
 ### Reference Examples in Codebase
 

.agents/skills/frontend-testing/references/workflow.md

@@ -4,7 +4,7 @@ This guide defines the workflow for generating tests, especially for complex com
 
 ## Scope Clarification
 
-This guide addresses **multi-file workflow** (how to process multiple test files). For coverage requirements within a single test file, see `web/testing/testing.md` § Coverage Goals.
+This guide addresses **multi-file workflow** (how to process multiple test files). For coverage requirements within a single test file, see `web/docs/test.md` § Coverage Goals.
 
 | Scope | Rule |
 |-------|------|

.github/workflows/api-tests.yml

@@ -72,6 +72,7 @@ jobs:
           OPENDAL_FS_ROOT: /tmp/dify-storage
         run: |
           uv run --project api pytest \
+            -n auto \
             --timeout "${PYTEST_TIMEOUT:-180}" \
             api/tests/integration_tests/workflow \
             api/tests/integration_tests/tools \

.github/workflows/build-push.yml

@@ -8,6 +8,7 @@ on:
       - "build/**"
       - "release/e-*"
       - "hotfix/**"
+      - "feat/hitl-backend"
     tags:
       - "*"
 

.github/workflows/style.yml

@@ -47,13 +47,9 @@ jobs:
         if: steps.changed-files.outputs.any_changed == 'true'
         run: uv run --directory api --dev lint-imports
 
-      - name: Run Basedpyright Checks
+      - name: Run Type Checks
         if: steps.changed-files.outputs.any_changed == 'true'
-        run: dev/basedpyright-check
-
-      - name: Run Mypy Type Checks
-        if: steps.changed-files.outputs.any_changed == 'true'
-        run: uv --directory api run mypy --exclude-gitignore --exclude 'tests/' --exclude 'migrations/' --check-untyped-defs --disable-error-code=import-untyped .
+        run: make type-check
 
       - name: Dotenv check
         if: steps.changed-files.outputs.any_changed == 'true'

AGENTS.md

@@ -7,7 +7,7 @@ Dify is an open-source platform for developing LLM applications with an intuitiv
 The codebase is split into:
 
 - **Backend API** (`/api`): Python Flask application organized with Domain-Driven Design
-- **Frontend Web** (`/web`): Next.js 15 application using TypeScript and React 19
+- **Frontend Web** (`/web`): Next.js application using TypeScript and React
 - **Docker deployment** (`/docker`): Containerized deployment configurations
 
 ## Backend Workflow
@@ -18,36 +18,7 @@ The codebase is split into:
 
 ## Frontend Workflow
 
-```bash
-cd web
-pnpm lint:fix
-pnpm type-check:tsgo
-pnpm test
-```
-
-### Frontend Linting
-
-ESLint is used for frontend code quality. Available commands:
-
-```bash
-# Lint all files (report only)
-pnpm lint
-
-# Lint and auto-fix issues
-pnpm lint:fix
-
-# Lint specific files or directories
-pnpm lint:fix app/components/base/button/
-pnpm lint:fix app/components/base/button/index.tsx
-
-# Lint quietly (errors only, no warnings)
-pnpm lint:quiet
-
-# Check code complexity
-pnpm lint:complexity
-```
-
-**Important**: Always run `pnpm lint:fix` before committing. The pre-commit hook runs `lint-staged` which only lints staged files.
+- Read `web/AGENTS.md` for details
 
 ## Testing & Quality Practices
 

CONTRIBUTING.md

@@ -77,7 +77,7 @@ How we prioritize:
 
 For setting up the frontend service, please refer to our comprehensive [guide](https://github.com/langgenius/dify/blob/main/web/README.md) in the `web/README.md` file. This document provides detailed instructions to help you set up the frontend environment properly.
 
-**Testing**: All React components must have comprehensive test coverage. See [web/testing/testing.md](https://github.com/langgenius/dify/blob/main/web/testing/testing.md) for the canonical frontend testing guidelines and follow every requirement described there.
+**Testing**: All React components must have comprehensive test coverage. See [web/docs/test.md](https://github.com/langgenius/dify/blob/main/web/docs/test.md) for the canonical frontend testing guidelines and follow every requirement described there.
 
 #### Backend
 

Makefile

@@ -68,9 +68,11 @@ lint:
 	@echo "✅ Linting complete"
 
 type-check:
-	@echo "📝 Running type check with basedpyright..."
-	@uv run --directory api --dev basedpyright
-	@echo "✅ Type check complete"
+	@echo "📝 Running type checks (basedpyright + mypy + ty)..."
+	@./dev/basedpyright-check $(PATH_TO_CHECK)
+	@uv --directory api run mypy --exclude-gitignore --exclude 'tests/' --exclude 'migrations/' --check-untyped-defs --disable-error-code=import-untyped .
+	@cd api && uv run ty check
+	@echo "✅ Type checks complete"
 
 test:
 	@echo "🧪 Running backend unit tests..."
@@ -78,7 +80,7 @@ test:
 		echo "Target: $(TARGET_TESTS)"; \
 		uv run --project api --dev pytest $(TARGET_TESTS); \
 	else \
-		uv run --project api --dev dev/pytest/pytest_unit_tests.sh; \
+		PYTEST_XDIST_ARGS="-n auto" uv run --project api --dev dev/pytest/pytest_unit_tests.sh; \
 	fi
 	@echo "✅ Tests complete"
 
@@ -130,7 +132,7 @@ help:
 	@echo "  make format         - Format code with ruff"
 	@echo "  make check          - Check code with ruff"
 	@echo "  make lint           - Format, fix, and lint code (ruff, imports, dotenv)"
-	@echo "  make type-check     - Run type checking with basedpyright"
+	@echo "  make type-check     - Run type checks (basedpyright, mypy, ty)"
 	@echo "  make test           - Run backend unit tests (or TARGET_TESTS=./api/tests/<target_tests>)"
 	@echo ""
 	@echo "Docker Build Targets:"

api/.env.example

@@ -617,6 +617,7 @@ PLUGIN_DAEMON_URL=http://127.0.0.1:5002
 PLUGIN_REMOTE_INSTALL_PORT=5003
 PLUGIN_REMOTE_INSTALL_HOST=localhost
 PLUGIN_MAX_PACKAGE_SIZE=15728640
+PLUGIN_MODEL_SCHEMA_CACHE_TTL=3600
 INNER_API_KEY_FOR_PLUGIN=QaHbTe77CtuXmsfyhR7+vRjI/+XbV1AaFy691iy+kGDv2Jvy0/eAh8Y1
 
 # Marketplace configuration
@@ -717,3 +718,27 @@ SANDBOX_EXPIRED_RECORDS_CLEAN_BATCH_SIZE=1000
 SANDBOX_EXPIRED_RECORDS_RETENTION_DAYS=30
 SANDBOX_EXPIRED_RECORDS_CLEAN_TASK_LOCK_TTL=90000
 
+
+# Redis URL used for PubSub between API and
+# celery worker
+# defaults to url constructed from `REDIS_*`
+# configurations
+PUBSUB_REDIS_URL=
+# Pub/sub channel type for streaming events.
+# valid options are:
+#
+# - pubsub: for normal Pub/Sub
+# - sharded: for sharded Pub/Sub
+#
+# It's highly recommended to use sharded Pub/Sub AND redis cluster
+# for large deployments.
+PUBSUB_REDIS_CHANNEL_TYPE=pubsub
+# Whether to use Redis cluster mode while running
+# PubSub.
+#  It's highly recommended to enable this for large deployments.
+PUBSUB_REDIS_USE_CLUSTERS=false
+
+# Whether to Enable human input timeout check task
+ENABLE_HUMAN_INPUT_TIMEOUT_TASK=true
+# Human input timeout check interval in minutes
+HUMAN_INPUT_TIMEOUT_TASK_INTERVAL=1

api/.importlinter

@@ -36,6 +36,8 @@ ignore_imports =
     core.workflow.nodes.loop.loop_node -> core.workflow.graph_engine
     core.workflow.nodes.loop.loop_node -> core.workflow.graph
     core.workflow.nodes.loop.loop_node -> core.workflow.graph_engine.command_channels
+    # TODO(QuantumGhost): fix the import violation later
+    core.workflow.entities.pause_reason -> core.workflow.nodes.human_input.entities
 
 [importlinter:contract:workflow-infrastructure-dependencies]
 name = Workflow Infrastructure Dependencies
@@ -58,6 +60,8 @@ ignore_imports =
     core.workflow.graph_engine.command_channels.redis_channel -> extensions.ext_redis
     core.workflow.graph_engine.manager -> extensions.ext_redis
     core.workflow.nodes.knowledge_retrieval.knowledge_retrieval_node -> extensions.ext_redis
+    # TODO(QuantumGhost): use DI to avoid depending on global DB.
+    core.workflow.nodes.human_input.human_input_node -> extensions.ext_database
 
 [importlinter:contract:workflow-external-imports]
 name = Workflow External Imports
@@ -145,6 +149,7 @@ ignore_imports =
     core.workflow.nodes.agent.agent_node -> core.agent.entities
     core.workflow.nodes.agent.agent_node -> core.agent.plugin_entities
     core.workflow.nodes.base.node -> core.app.entities.app_invoke_entities
+    core.workflow.nodes.human_input.human_input_node -> core.app.entities.app_invoke_entities
     core.workflow.nodes.knowledge_index.knowledge_index_node -> core.app.entities.app_invoke_entities
     core.workflow.nodes.knowledge_retrieval.knowledge_retrieval_node -> core.app.app_config.entities
     core.workflow.nodes.knowledge_retrieval.knowledge_retrieval_node -> core.app.entities.app_invoke_entities
@@ -227,6 +232,9 @@ ignore_imports =
     core.workflow.nodes.knowledge_index.entities -> core.rag.retrieval.retrieval_methods
     core.workflow.nodes.knowledge_index.knowledge_index_node -> core.rag.retrieval.retrieval_methods
     core.workflow.nodes.knowledge_index.knowledge_index_node -> models.dataset
+    core.workflow.nodes.knowledge_index.knowledge_index_node -> services.summary_index_service
+    core.workflow.nodes.knowledge_index.knowledge_index_node -> tasks.generate_summary_index_task
+    core.workflow.nodes.knowledge_index.knowledge_index_node -> core.rag.index_processor.processor.paragraph_index_processor
     core.workflow.nodes.knowledge_retrieval.knowledge_retrieval_node -> core.rag.retrieval.retrieval_methods
     core.workflow.nodes.llm.node -> models.dataset
     core.workflow.nodes.agent.agent_node -> core.tools.utils.message_transformer
@@ -245,6 +253,7 @@ ignore_imports =
     core.workflow.nodes.document_extractor.node -> core.variables.segments
     core.workflow.nodes.http_request.executor -> core.variables.segments
     core.workflow.nodes.http_request.node -> core.variables.segments
+    core.workflow.nodes.human_input.entities -> core.variables.consts
     core.workflow.nodes.iteration.iteration_node -> core.variables
     core.workflow.nodes.iteration.iteration_node -> core.variables.segments
     core.workflow.nodes.iteration.iteration_node -> core.variables.variables
@@ -291,6 +300,8 @@ ignore_imports =
     core.workflow.nodes.llm.llm_utils -> extensions.ext_database
     core.workflow.nodes.llm.node -> extensions.ext_database
     core.workflow.nodes.tool.tool_node -> extensions.ext_database
+    core.workflow.nodes.human_input.human_input_node -> extensions.ext_database
+    core.workflow.nodes.human_input.human_input_node -> core.repositories.human_input_repository
     core.workflow.workflow_entry -> extensions.otel.runtime
     core.workflow.nodes.agent.agent_node -> models
     core.workflow.nodes.base.node -> models.enums
@@ -300,6 +311,58 @@ ignore_imports =
     core.workflow.nodes.agent.agent_node -> services
     core.workflow.nodes.tool.tool_node -> services
 
+[importlinter:contract:model-runtime-no-internal-imports]
+name = Model Runtime Internal Imports
+type = forbidden
+source_modules =
+    core.model_runtime
+forbidden_modules =
+    configs
+    controllers
+    extensions
+    models
+    services
+    tasks
+    core.agent
+    core.app
+    core.base
+    core.callback_handler
+    core.datasource
+    core.db
+    core.entities
+    core.errors
+    core.extension
+    core.external_data_tool
+    core.file
+    core.helper
+    core.hosting_configuration
+    core.indexing_runner
+    core.llm_generator
+    core.logging
+    core.mcp
+    core.memory
+    core.model_manager
+    core.moderation
+    core.ops
+    core.plugin
+    core.prompt
+    core.provider_manager
+    core.rag
+    core.repositories
+    core.schemas
+    core.tools
+    core.trigger
+    core.variables
+    core.workflow
+ignore_imports =
+    core.model_runtime.model_providers.__base.ai_model -> configs
+    core.model_runtime.model_providers.__base.ai_model -> extensions.ext_redis
+    core.model_runtime.model_providers.__base.large_language_model -> configs
+    core.model_runtime.model_providers.__base.text_embedding_model -> core.entities.embedding_type
+    core.model_runtime.model_providers.model_provider_factory -> configs
+    core.model_runtime.model_providers.model_provider_factory -> extensions.ext_redis
+    core.model_runtime.model_providers.model_provider_factory -> models.provider_ids
+
 [importlinter:contract:rsc]
 name = RSC
 type = layers

api/app.py

@@ -1,4 +1,12 @@
+from __future__ import annotations
+
 import sys
+from typing import TYPE_CHECKING, cast
+
+if TYPE_CHECKING:
+    from celery import Celery
+
+    celery: Celery
 
 
 def is_db_command() -> bool:
@@ -23,7 +31,7 @@ else:
     from app_factory import create_app
 
     app = create_app()
-    celery = app.extensions["celery"]
+    celery = cast("Celery", app.extensions["celery"])
 
 if __name__ == "__main__":
     app.run(host="0.0.0.0", port=5001)

api/app_factory.py

@@ -149,7 +149,7 @@ def initialize_extensions(app: DifyApp):
             logger.info("Loaded %s (%s ms)", short_name, round((end_time - start_time) * 1000, 2))
 
 
-def create_migrations_app():
+def create_migrations_app() -> DifyApp:
     app = create_flask_app_with_configs()
     from extensions import ext_database, ext_migrate
 

api/configs/feature/__init__.py

@@ -1,3 +1,4 @@
+from datetime import timedelta
 from enum import StrEnum
 from typing import Literal
 
@@ -48,6 +49,16 @@ class SecurityConfig(BaseSettings):
         default=5,
     )
 
+    WEB_FORM_SUBMIT_RATE_LIMIT_MAX_ATTEMPTS: PositiveInt = Field(
+        description="Maximum number of web form submissions allowed per IP within the rate limit window",
+        default=30,
+    )
+
+    WEB_FORM_SUBMIT_RATE_LIMIT_WINDOW_SECONDS: PositiveInt = Field(
+        description="Time window in seconds for web form submission rate limiting",
+        default=60,
+    )
+
     LOGIN_DISABLED: bool = Field(
         description="Whether to disable login checks",
         default=False,
@@ -82,6 +93,12 @@ class AppExecutionConfig(BaseSettings):
         default=0,
     )
 
+    HUMAN_INPUT_GLOBAL_TIMEOUT_SECONDS: PositiveInt = Field(
+        description="Maximum seconds a workflow run can stay paused waiting for human input before global timeout.",
+        default=int(timedelta(days=7).total_seconds()),
+        ge=1,
+    )
+
 
 class CodeExecutionSandboxConfig(BaseSettings):
     """
@@ -243,6 +260,11 @@ class PluginConfig(BaseSettings):
         default=15728640 * 12,
     )
 
+    PLUGIN_MODEL_SCHEMA_CACHE_TTL: PositiveInt = Field(
+        description="TTL in seconds for caching plugin model schemas in Redis",
+        default=60 * 60,
+    )
+
 
 class MarketplaceConfig(BaseSettings):
     """
@@ -1129,6 +1151,14 @@ class CeleryScheduleTasksConfig(BaseSettings):
         description="Enable queue monitor task",
         default=False,
     )
+    ENABLE_HUMAN_INPUT_TIMEOUT_TASK: bool = Field(
+        description="Enable human input timeout check task",
+        default=True,
+    )
+    HUMAN_INPUT_TIMEOUT_TASK_INTERVAL: PositiveInt = Field(
+        description="Human input timeout check interval in minutes",
+        default=1,
+    )
     ENABLE_CHECK_UPGRADABLE_PLUGIN_TASK: bool = Field(
         description="Enable check upgradable plugin task",
         default=True,

api/configs/middleware/__init__.py

@@ -6,6 +6,7 @@ from pydantic import Field, NonNegativeFloat, NonNegativeInt, PositiveFloat, Pos
 from pydantic_settings import BaseSettings
 
 from .cache.redis_config import RedisConfig
+from .cache.redis_pubsub_config import RedisPubSubConfig
 from .storage.aliyun_oss_storage_config import AliyunOSSStorageConfig
 from .storage.amazon_s3_storage_config import S3StorageConfig
 from .storage.azure_blob_storage_config import AzureBlobStorageConfig
@@ -317,6 +318,7 @@ class MiddlewareConfig(
     CeleryConfig,  # Note: CeleryConfig already inherits from DatabaseConfig
     KeywordStoreConfig,
     RedisConfig,
+    RedisPubSubConfig,
     # configs of storage and storage providers
     StorageConfig,
     AliyunOSSStorageConfig,

api/configs/middleware/cache/redis_pubsub_config.py

@@ -0,0 +1,96 @@
+from typing import Literal, Protocol
+from urllib.parse import quote_plus, urlunparse
+
+from pydantic import Field
+from pydantic_settings import BaseSettings
+
+
+class RedisConfigDefaults(Protocol):
+    REDIS_HOST: str
+    REDIS_PORT: int
+    REDIS_USERNAME: str | None
+    REDIS_PASSWORD: str | None
+    REDIS_DB: int
+    REDIS_USE_SSL: bool
+    REDIS_USE_SENTINEL: bool | None
+    REDIS_USE_CLUSTERS: bool
+
+
+class RedisConfigDefaultsMixin:
+    def _redis_defaults(self: RedisConfigDefaults) -> RedisConfigDefaults:
+        return self
+
+
+class RedisPubSubConfig(BaseSettings, RedisConfigDefaultsMixin):
+    """
+    Configuration settings for Redis pub/sub streaming.
+    """
+
+    PUBSUB_REDIS_URL: str | None = Field(
+        alias="PUBSUB_REDIS_URL",
+        description=(
+            "Redis connection URL for pub/sub streaming events between API "
+            "and celery worker, defaults to url constructed from "
+            "`REDIS_*` configurations"
+        ),
+        default=None,
+    )
+
+    PUBSUB_REDIS_USE_CLUSTERS: bool = Field(
+        description=(
+            "Enable Redis Cluster mode for pub/sub streaming. It's highly "
+            "recommended to enable this for large deployments."
+        ),
+        default=False,
+    )
+
+    PUBSUB_REDIS_CHANNEL_TYPE: Literal["pubsub", "sharded"] = Field(
+        description=(
+            "Pub/sub channel type for streaming events. "
+            "Valid options are:\n"
+            "\n"
+            " - pubsub: for normal Pub/Sub\n"
+            " - sharded: for sharded Pub/Sub\n"
+            "\n"
+            "It's highly recommended to use sharded Pub/Sub AND redis cluster "
+            "for large deployments."
+        ),
+        default="pubsub",
+    )
+
+    def _build_default_pubsub_url(self) -> str:
+        defaults = self._redis_defaults()
+        if not defaults.REDIS_HOST or not defaults.REDIS_PORT:
+            raise ValueError("PUBSUB_REDIS_URL must be set when default Redis URL cannot be constructed")
+
+        scheme = "rediss" if defaults.REDIS_USE_SSL else "redis"
+        username = defaults.REDIS_USERNAME or None
+        password = defaults.REDIS_PASSWORD or None
+
+        userinfo = ""
+        if username:
+            userinfo = quote_plus(username)
+        if password:
+            password_part = quote_plus(password)
+            userinfo = f"{userinfo}:{password_part}" if userinfo else f":{password_part}"
+        if userinfo:
+            userinfo = f"{userinfo}@"
+
+        host = defaults.REDIS_HOST
+        port = defaults.REDIS_PORT
+        db = defaults.REDIS_DB
+
+        netloc = f"{userinfo}{host}:{port}"
+        return urlunparse((scheme, netloc, f"/{db}", "", "", ""))
+
+    @property
+    def normalized_pubsub_redis_url(self) -> str:
+        pubsub_redis_url = self.PUBSUB_REDIS_URL
+        if pubsub_redis_url:
+            cleaned = pubsub_redis_url.strip()
+            pubsub_redis_url = cleaned or None
+
+        if pubsub_redis_url:
+            return pubsub_redis_url
+
+        return self._build_default_pubsub_url()

api/contexts/__init__.py

@@ -6,7 +6,6 @@ from contexts.wrapper import RecyclableContextVar
 
 if TYPE_CHECKING:
     from core.datasource.__base.datasource_provider import DatasourcePluginProviderController
-    from core.model_runtime.entities.model_entities import AIModelEntity
     from core.plugin.entities.plugin_daemon import PluginModelProviderEntity
     from core.tools.plugin_tool.provider import PluginToolProviderController
     from core.trigger.provider import PluginTriggerProviderController
@@ -29,12 +28,6 @@ plugin_model_providers_lock: RecyclableContextVar[Lock] = RecyclableContextVar(
     ContextVar("plugin_model_providers_lock")
 )
 
-plugin_model_schema_lock: RecyclableContextVar[Lock] = RecyclableContextVar(ContextVar("plugin_model_schema_lock"))
-
-plugin_model_schemas: RecyclableContextVar[dict[str, "AIModelEntity"]] = RecyclableContextVar(
-    ContextVar("plugin_model_schemas")
-)
-
 datasource_plugin_providers: RecyclableContextVar[dict[str, "DatasourcePluginProviderController"]] = (
     RecyclableContextVar(ContextVar("datasource_plugin_providers"))
 )

api/controllers/console/__init__.py

@@ -37,6 +37,7 @@ from . import (
     apikey,
     extension,
     feature,
+    human_input_form,
     init_validate,
     ping,
     setup,
@@ -171,6 +172,7 @@ __all__ = [
     "forgot_password",
     "generator",
     "hit_testing",
+    "human_input_form",
     "init_validate",
     "installed_app",
     "load_balancing_config",

api/controllers/console/admin.py

@@ -243,15 +243,13 @@ class InsertExploreBannerApi(Resource):
     def post(self):
         payload = InsertExploreBannerPayload.model_validate(console_ns.payload)
 
-        content = {
-            "category": payload.category,
-            "title": payload.title,
-            "description": payload.description,
-            "img-src": payload.img_src,
-        }
-
         banner = ExporleBanner(
-            content=content,
+            content={
+                "category": payload.category,
+                "title": payload.title,
+                "description": payload.description,
+                "img-src": payload.img_src,
+            },
             link=payload.link,
             sort=payload.sort,
             language=payload.language,

api/controllers/console/app/conversation.py

@@ -89,6 +89,7 @@ status_count_model = console_ns.model(
         "success": fields.Integer,
         "failed": fields.Integer,
         "partial_success": fields.Integer,
+        "paused": fields.Integer,
     },
 )
 

api/controllers/console/app/generator.py

@@ -1,5 +1,4 @@
 from collections.abc import Sequence
-from typing import Any
 
 from flask_restx import Resource
 from pydantic import BaseModel, Field
@@ -12,10 +11,12 @@ from controllers.console.app.error import (
     ProviderQuotaExceededError,
 )
 from controllers.console.wraps import account_initialization_required, setup_required
+from core.app.app_config.entities import ModelConfig
 from core.errors.error import ModelCurrentlyNotSupportError, ProviderTokenNotInitError, QuotaExceededError
 from core.helper.code_executor.code_node_provider import CodeNodeProvider
 from core.helper.code_executor.javascript.javascript_code_provider import JavascriptCodeProvider
 from core.helper.code_executor.python3.python3_code_provider import Python3CodeProvider
+from core.llm_generator.entities import RuleCodeGeneratePayload, RuleGeneratePayload, RuleStructuredOutputPayload
 from core.llm_generator.llm_generator import LLMGenerator
 from core.model_runtime.errors.invoke import InvokeError
 from extensions.ext_database import db
@@ -26,28 +27,13 @@ from services.workflow_service import WorkflowService
 DEFAULT_REF_TEMPLATE_SWAGGER_2_0 = "#/definitions/{model}"
 
 
-class RuleGeneratePayload(BaseModel):
-    instruction: str = Field(..., description="Rule generation instruction")
-    model_config_data: dict[str, Any] = Field(..., alias="model_config", description="Model configuration")
-    no_variable: bool = Field(default=False, description="Whether to exclude variables")
-
-
-class RuleCodeGeneratePayload(RuleGeneratePayload):
-    code_language: str = Field(default="javascript", description="Programming language for code generation")
-
-
-class RuleStructuredOutputPayload(BaseModel):
-    instruction: str = Field(..., description="Structured output generation instruction")
-    model_config_data: dict[str, Any] = Field(..., alias="model_config", description="Model configuration")
-
-
 class InstructionGeneratePayload(BaseModel):
     flow_id: str = Field(..., description="Workflow/Flow ID")
     node_id: str = Field(default="", description="Node ID for workflow context")
     current: str = Field(default="", description="Current instruction text")
     language: str = Field(default="javascript", description="Programming language (javascript/python)")
     instruction: str = Field(..., description="Instruction for generation")
-    model_config_data: dict[str, Any] = Field(..., alias="model_config", description="Model configuration")
+    model_config_data: ModelConfig = Field(..., alias="model_config", description="Model configuration")
     ideal_output: str = Field(default="", description="Expected ideal output")
 
 
@@ -64,6 +50,7 @@ reg(RuleCodeGeneratePayload)
 reg(RuleStructuredOutputPayload)
 reg(InstructionGeneratePayload)
 reg(InstructionTemplatePayload)
+reg(ModelConfig)
 
 
 @console_ns.route("/rule-generate")
@@ -82,12 +69,7 @@ class RuleGenerateApi(Resource):
         _, current_tenant_id = current_account_with_tenant()
 
         try:
-            rules = LLMGenerator.generate_rule_config(
-                tenant_id=current_tenant_id,
-                instruction=args.instruction,
-                model_config=args.model_config_data,
-                no_variable=args.no_variable,
-            )
+            rules = LLMGenerator.generate_rule_config(tenant_id=current_tenant_id, args=args)
         except ProviderTokenNotInitError as ex:
             raise ProviderNotInitializeError(ex.description)
         except QuotaExceededError:
@@ -118,9 +100,7 @@ class RuleCodeGenerateApi(Resource):
         try:
             code_result = LLMGenerator.generate_code(
                 tenant_id=current_tenant_id,
-                instruction=args.instruction,
-                model_config=args.model_config_data,
-                code_language=args.code_language,
+                args=args,
             )
         except ProviderTokenNotInitError as ex:
             raise ProviderNotInitializeError(ex.description)
@@ -152,8 +132,7 @@ class RuleStructuredOutputGenerateApi(Resource):
         try:
             structured_output = LLMGenerator.generate_structured_output(
                 tenant_id=current_tenant_id,
-                instruction=args.instruction,
-                model_config=args.model_config_data,
+                args=args,
             )
         except ProviderTokenNotInitError as ex:
             raise ProviderNotInitializeError(ex.description)
@@ -204,23 +183,29 @@ class InstructionGenerateApi(Resource):
                     case "llm":
                         return LLMGenerator.generate_rule_config(
                             current_tenant_id,
-                            instruction=args.instruction,
-                            model_config=args.model_config_data,
-                            no_variable=True,
+                            args=RuleGeneratePayload(
+                                instruction=args.instruction,
+                                model_config=args.model_config_data,
+                                no_variable=True,
+                            ),
                         )
                     case "agent":
                         return LLMGenerator.generate_rule_config(
                             current_tenant_id,
-                            instruction=args.instruction,
-                            model_config=args.model_config_data,
-                            no_variable=True,
+                            args=RuleGeneratePayload(
+                                instruction=args.instruction,
+                                model_config=args.model_config_data,
+                                no_variable=True,
+                            ),
                         )
                     case "code":
                         return LLMGenerator.generate_code(
                             tenant_id=current_tenant_id,
-                            instruction=args.instruction,
-                            model_config=args.model_config_data,
-                            code_language=args.language,
+                            args=RuleCodeGeneratePayload(
+                                instruction=args.instruction,
+                                model_config=args.model_config_data,
+                                code_language=args.language,
+                            ),
                         )
                     case _:
                         return {"error": f"invalid node type: {node_type}"}

api/controllers/console/app/message.py

@@ -32,7 +32,7 @@ from libs.login import current_account_with_tenant, login_required
 from models.model import AppMode, Conversation, Message, MessageAnnotation, MessageFeedback
 from services.errors.conversation import ConversationNotExistsError
 from services.errors.message import MessageNotExistsError, SuggestedQuestionsAfterAnswerDisabledError
-from services.message_service import MessageService
+from services.message_service import MessageService, attach_message_extra_contents
 
 logger = logging.getLogger(__name__)
 DEFAULT_REF_TEMPLATE_SWAGGER_2_0 = "#/definitions/{model}"
@@ -198,6 +198,7 @@ message_detail_model = console_ns.model(
         "created_at": TimestampField,
         "agent_thoughts": fields.List(fields.Nested(agent_thought_model)),
         "message_files": fields.List(fields.Nested(message_file_model)),
+        "extra_contents": fields.List(fields.Raw),
         "metadata": fields.Raw(attribute="message_metadata_dict"),
         "status": fields.String,
         "error": fields.String,
@@ -290,6 +291,7 @@ class ChatMessageListApi(Resource):
             has_more = False
 
         history_messages = list(reversed(history_messages))
+        attach_message_extra_contents(history_messages)
 
         return InfiniteScrollPagination(data=history_messages, limit=args.limit, has_more=has_more)
 
@@ -474,4 +476,5 @@ class MessageApi(Resource):
         if not message:
             raise NotFound("Message Not Exists.")
 
+        attach_message_extra_contents([message])
         return message

api/controllers/console/app/workflow.py

@@ -507,6 +507,179 @@ class WorkflowDraftRunLoopNodeApi(Resource):
             raise InternalServerError()
 
 
+class HumanInputFormPreviewPayload(BaseModel):
+    inputs: dict[str, Any] = Field(
+        default_factory=dict,
+        description="Values used to fill missing upstream variables referenced in form_content",
+    )
+
+
+class HumanInputFormSubmitPayload(BaseModel):
+    form_inputs: dict[str, Any] = Field(..., description="Values the user provides for the form's own fields")
+    inputs: dict[str, Any] = Field(
+        ...,
+        description="Values used to fill missing upstream variables referenced in form_content",
+    )
+    action: str = Field(..., description="Selected action ID")
+
+
+class HumanInputDeliveryTestPayload(BaseModel):
+    delivery_method_id: str = Field(..., description="Delivery method ID")
+    inputs: dict[str, Any] = Field(
+        default_factory=dict,
+        description="Values used to fill missing upstream variables referenced in form_content",
+    )
+
+
+reg(HumanInputFormPreviewPayload)
+reg(HumanInputFormSubmitPayload)
+reg(HumanInputDeliveryTestPayload)
+
+
+@console_ns.route("/apps/<uuid:app_id>/advanced-chat/workflows/draft/human-input/nodes/<string:node_id>/form/preview")
+class AdvancedChatDraftHumanInputFormPreviewApi(Resource):
+    @console_ns.doc("get_advanced_chat_draft_human_input_form")
+    @console_ns.doc(description="Get human input form preview for advanced chat workflow")
+    @console_ns.doc(params={"app_id": "Application ID", "node_id": "Node ID"})
+    @console_ns.expect(console_ns.models[HumanInputFormPreviewPayload.__name__])
+    @setup_required
+    @login_required
+    @account_initialization_required
+    @get_app_model(mode=[AppMode.ADVANCED_CHAT])
+    @edit_permission_required
+    def post(self, app_model: App, node_id: str):
+        """
+        Preview human input form content and placeholders
+        """
+        current_user, _ = current_account_with_tenant()
+        args = HumanInputFormPreviewPayload.model_validate(console_ns.payload or {})
+        inputs = args.inputs
+
+        workflow_service = WorkflowService()
+        preview = workflow_service.get_human_input_form_preview(
+            app_model=app_model,
+            account=current_user,
+            node_id=node_id,
+            inputs=inputs,
+        )
+        return jsonable_encoder(preview)
+
+
+@console_ns.route("/apps/<uuid:app_id>/advanced-chat/workflows/draft/human-input/nodes/<string:node_id>/form/run")
+class AdvancedChatDraftHumanInputFormRunApi(Resource):
+    @console_ns.doc("submit_advanced_chat_draft_human_input_form")
+    @console_ns.doc(description="Submit human input form preview for advanced chat workflow")
+    @console_ns.doc(params={"app_id": "Application ID", "node_id": "Node ID"})
+    @console_ns.expect(console_ns.models[HumanInputFormSubmitPayload.__name__])
+    @setup_required
+    @login_required
+    @account_initialization_required
+    @get_app_model(mode=[AppMode.ADVANCED_CHAT])
+    @edit_permission_required
+    def post(self, app_model: App, node_id: str):
+        """
+        Submit human input form preview
+        """
+        current_user, _ = current_account_with_tenant()
+        args = HumanInputFormSubmitPayload.model_validate(console_ns.payload or {})
+        workflow_service = WorkflowService()
+        result = workflow_service.submit_human_input_form_preview(
+            app_model=app_model,
+            account=current_user,
+            node_id=node_id,
+            form_inputs=args.form_inputs,
+            inputs=args.inputs,
+            action=args.action,
+        )
+        return jsonable_encoder(result)
+
+
+@console_ns.route("/apps/<uuid:app_id>/workflows/draft/human-input/nodes/<string:node_id>/form/preview")
+class WorkflowDraftHumanInputFormPreviewApi(Resource):
+    @console_ns.doc("get_workflow_draft_human_input_form")
+    @console_ns.doc(description="Get human input form preview for workflow")
+    @console_ns.doc(params={"app_id": "Application ID", "node_id": "Node ID"})
+    @console_ns.expect(console_ns.models[HumanInputFormPreviewPayload.__name__])
+    @setup_required
+    @login_required
+    @account_initialization_required
+    @get_app_model(mode=[AppMode.WORKFLOW])
+    @edit_permission_required
+    def post(self, app_model: App, node_id: str):
+        """
+        Preview human input form content and placeholders
+        """
+        current_user, _ = current_account_with_tenant()
+        args = HumanInputFormPreviewPayload.model_validate(console_ns.payload or {})
+        inputs = args.inputs
+
+        workflow_service = WorkflowService()
+        preview = workflow_service.get_human_input_form_preview(
+            app_model=app_model,
+            account=current_user,
+            node_id=node_id,
+            inputs=inputs,
+        )
+        return jsonable_encoder(preview)
+
+
+@console_ns.route("/apps/<uuid:app_id>/workflows/draft/human-input/nodes/<string:node_id>/form/run")
+class WorkflowDraftHumanInputFormRunApi(Resource):
+    @console_ns.doc("submit_workflow_draft_human_input_form")
+    @console_ns.doc(description="Submit human input form preview for workflow")
+    @console_ns.doc(params={"app_id": "Application ID", "node_id": "Node ID"})
+    @console_ns.expect(console_ns.models[HumanInputFormSubmitPayload.__name__])
+    @setup_required
+    @login_required
+    @account_initialization_required
+    @get_app_model(mode=[AppMode.WORKFLOW])
+    @edit_permission_required
+    def post(self, app_model: App, node_id: str):
+        """
+        Submit human input form preview
+        """
+        current_user, _ = current_account_with_tenant()
+        workflow_service = WorkflowService()
+        args = HumanInputFormSubmitPayload.model_validate(console_ns.payload or {})
+        result = workflow_service.submit_human_input_form_preview(
+            app_model=app_model,
+            account=current_user,
+            node_id=node_id,
+            form_inputs=args.form_inputs,
+            inputs=args.inputs,
+            action=args.action,
+        )
+        return jsonable_encoder(result)
+
+
+@console_ns.route("/apps/<uuid:app_id>/workflows/draft/human-input/nodes/<string:node_id>/delivery-test")
+class WorkflowDraftHumanInputDeliveryTestApi(Resource):
+    @console_ns.doc("test_workflow_draft_human_input_delivery")
+    @console_ns.doc(description="Test human input delivery for workflow")
+    @console_ns.doc(params={"app_id": "Application ID", "node_id": "Node ID"})
+    @console_ns.expect(console_ns.models[HumanInputDeliveryTestPayload.__name__])
+    @setup_required
+    @login_required
+    @account_initialization_required
+    @get_app_model(mode=[AppMode.WORKFLOW, AppMode.ADVANCED_CHAT])
+    @edit_permission_required
+    def post(self, app_model: App, node_id: str):
+        """
+        Test human input delivery
+        """
+        current_user, _ = current_account_with_tenant()
+        workflow_service = WorkflowService()
+        args = HumanInputDeliveryTestPayload.model_validate(console_ns.payload or {})
+        workflow_service.test_human_input_delivery(
+            app_model=app_model,
+            account=current_user,
+            node_id=node_id,
+            delivery_method_id=args.delivery_method_id,
+            inputs=args.inputs,
+        )
+        return jsonable_encoder({})
+
+
 @console_ns.route("/apps/<uuid:app_id>/workflows/draft/run")
 class DraftWorkflowRunApi(Resource):
     @console_ns.doc("run_draft_workflow")

api/controllers/console/app/workflow_run.py

@@ -5,10 +5,15 @@ from flask import request
 from flask_restx import Resource, fields, marshal_with
 from pydantic import BaseModel, Field, field_validator
 from sqlalchemy import select
+from sqlalchemy.orm import sessionmaker
 
+from configs import dify_config
 from controllers.console import console_ns
 from controllers.console.app.wraps import get_app_model
 from controllers.console.wraps import account_initialization_required, setup_required
+from controllers.web.error import NotFoundError
+from core.workflow.entities.pause_reason import HumanInputRequired
+from core.workflow.enums import WorkflowExecutionStatus
 from extensions.ext_database import db
 from fields.end_user_fields import simple_end_user_fields
 from fields.member_fields import simple_account_fields
@@ -27,9 +32,21 @@ from libs.custom_inputs import time_duration
 from libs.helper import uuid_value
 from libs.login import current_user, login_required
 from models import Account, App, AppMode, EndUser, WorkflowArchiveLog, WorkflowRunTriggeredFrom
+from models.workflow import WorkflowRun
+from repositories.factory import DifyAPIRepositoryFactory
 from services.retention.workflow_run.constants import ARCHIVE_BUNDLE_NAME
 from services.workflow_run_service import WorkflowRunService
 
+
+def _build_backstage_input_url(form_token: str | None) -> str | None:
+    if not form_token:
+        return None
+    base_url = dify_config.APP_WEB_URL
+    if not base_url:
+        return None
+    return f"{base_url.rstrip('/')}/form/{form_token}"
+
+
 # Workflow run status choices for filtering
 WORKFLOW_RUN_STATUS_CHOICES = ["running", "succeeded", "failed", "stopped", "partial-succeeded"]
 EXPORT_SIGNED_URL_EXPIRE_SECONDS = 3600
@@ -440,3 +457,63 @@ class WorkflowRunNodeExecutionListApi(Resource):
         )
 
         return {"data": node_executions}
+
+
+@console_ns.route("/workflow/<string:workflow_run_id>/pause-details")
+class ConsoleWorkflowPauseDetailsApi(Resource):
+    """Console API for getting workflow pause details."""
+
+    @account_initialization_required
+    @login_required
+    def get(self, workflow_run_id: str):
+        """
+        Get workflow pause details.
+
+        GET /console/api/workflow/<workflow_run_id>/pause-details
+
+        Returns information about why and where the workflow is paused.
+        """
+
+        # Query WorkflowRun to determine if workflow is suspended
+        session_maker = sessionmaker(bind=db.engine)
+        workflow_run_repo = DifyAPIRepositoryFactory.create_api_workflow_run_repository(session_maker=session_maker)
+        workflow_run = db.session.get(WorkflowRun, workflow_run_id)
+        if not workflow_run:
+            raise NotFoundError("Workflow run not found")
+
+        # Check if workflow is suspended
+        is_paused = workflow_run.status == WorkflowExecutionStatus.PAUSED
+        if not is_paused:
+            return {
+                "paused_at": None,
+                "paused_nodes": [],
+            }, 200
+
+        pause_entity = workflow_run_repo.get_workflow_pause(workflow_run_id)
+        pause_reasons = pause_entity.get_pause_reasons() if pause_entity else []
+
+        # Build response
+        paused_at = pause_entity.paused_at if pause_entity else None
+        paused_nodes = []
+        response = {
+            "paused_at": paused_at.isoformat() + "Z" if paused_at else None,
+            "paused_nodes": paused_nodes,
+        }
+
+        for reason in pause_reasons:
+            if isinstance(reason, HumanInputRequired):
+                paused_nodes.append(
+                    {
+                        "node_id": reason.node_id,
+                        "node_title": reason.node_title,
+                        "pause_type": {
+                            "type": "human_input",
+                            "form_id": reason.form_id,
+                            "backstage_input_url": _build_backstage_input_url(reason.form_token),
+                        },
+                    }
+                )
+            else:
+                raise AssertionError("unimplemented.")
+
+        return response, 200

api/controllers/console/datasets/datasets.py

@@ -148,6 +148,7 @@ class DatasetUpdatePayload(BaseModel):
     embedding_model: str | None = None
     embedding_model_provider: str | None = None
     retrieval_model: dict[str, Any] | None = None
+    summary_index_setting: dict[str, Any] | None = None
     partial_member_list: list[dict[str, str]] | None = None
     external_retrieval_model: dict[str, Any] | None = None
     external_knowledge_id: str | None = None
@@ -288,7 +289,14 @@ class DatasetListApi(Resource):
     @enterprise_license_required
     def get(self):
         current_user, current_tenant_id = current_account_with_tenant()
-        query = ConsoleDatasetListQuery.model_validate(request.args.to_dict())
+        # Convert query parameters to dict, handling list parameters correctly
+        query_params: dict[str, str | list[str]] = dict(request.args.to_dict())
+        # Handle ids and tag_ids as lists (Flask request.args.getlist returns list even for single value)
+        if "ids" in request.args:
+            query_params["ids"] = request.args.getlist("ids")
+        if "tag_ids" in request.args:
+            query_params["tag_ids"] = request.args.getlist("tag_ids")
+        query = ConsoleDatasetListQuery.model_validate(query_params)
         # provider = request.args.get("provider", default="vendor")
         if query.ids:
             datasets, total = DatasetService.get_datasets_by_ids(query.ids, current_tenant_id)

api/controllers/console/datasets/datasets_document.py

@@ -45,6 +45,7 @@ from models.dataset import DocumentPipelineExecutionLog
 from services.dataset_service import DatasetService, DocumentService
 from services.entities.knowledge_entities.knowledge_entities import KnowledgeConfig, ProcessRule, RetrievalModel
 from services.file_service import FileService
+from tasks.generate_summary_index_task import generate_summary_index_task
 
 from ..app.error import (
     ProviderModelCurrentlyNotSupportError,
@@ -103,6 +104,10 @@ class DocumentRenamePayload(BaseModel):
     name: str
 
 
+class GenerateSummaryPayload(BaseModel):
+    document_list: list[str]
+
+
 class DocumentBatchDownloadZipPayload(BaseModel):
     """Request payload for bulk downloading documents as a zip archive."""
 
@@ -125,6 +130,7 @@ register_schema_models(
     RetrievalModel,
     DocumentRetryPayload,
     DocumentRenamePayload,
+    GenerateSummaryPayload,
     DocumentBatchDownloadZipPayload,
 )
 
@@ -312,6 +318,13 @@ class DatasetDocumentListApi(Resource):
 
         paginated_documents = db.paginate(select=query, page=page, per_page=limit, max_per_page=100, error_out=False)
         documents = paginated_documents.items
+
+        DocumentService.enrich_documents_with_summary_index_status(
+            documents=documents,
+            dataset=dataset,
+            tenant_id=current_tenant_id,
+        )
+
         if fetch:
             for document in documents:
                 completed_segments = (
@@ -797,6 +810,7 @@ class DocumentApi(DocumentResource):
                 "display_status": document.display_status,
                 "doc_form": document.doc_form,
                 "doc_language": document.doc_language,
+                "need_summary": document.need_summary if document.need_summary is not None else False,
             }
         else:
             dataset_process_rules = DatasetService.get_process_rules(dataset_id)
@@ -832,6 +846,7 @@ class DocumentApi(DocumentResource):
                 "display_status": document.display_status,
                 "doc_form": document.doc_form,
                 "doc_language": document.doc_language,
+                "need_summary": document.need_summary if document.need_summary is not None else False,
             }
 
         return response, 200
@@ -1255,3 +1270,137 @@ class DocumentPipelineExecutionLogApi(DocumentResource):
             "input_data": log.input_data,
             "datasource_node_id": log.datasource_node_id,
         }, 200
+
+
+@console_ns.route("/datasets/<uuid:dataset_id>/documents/generate-summary")
+class DocumentGenerateSummaryApi(Resource):
+    @console_ns.doc("generate_summary_for_documents")
+    @console_ns.doc(description="Generate summary index for documents")
+    @console_ns.doc(params={"dataset_id": "Dataset ID"})
+    @console_ns.expect(console_ns.models[GenerateSummaryPayload.__name__])
+    @console_ns.response(200, "Summary generation started successfully")
+    @console_ns.response(400, "Invalid request or dataset configuration")
+    @console_ns.response(403, "Permission denied")
+    @console_ns.response(404, "Dataset not found")
+    @setup_required
+    @login_required
+    @account_initialization_required
+    @cloud_edition_billing_rate_limit_check("knowledge")
+    def post(self, dataset_id):
+        """
+        Generate summary index for specified documents.
+
+        This endpoint checks if the dataset configuration supports summary generation
+        (indexing_technique must be 'high_quality' and summary_index_setting.enable must be true),
+        then asynchronously generates summary indexes for the provided documents.
+        """
+        current_user, _ = current_account_with_tenant()
+        dataset_id = str(dataset_id)
+
+        # Get dataset
+        dataset = DatasetService.get_dataset(dataset_id)
+        if not dataset:
+            raise NotFound("Dataset not found.")
+
+        # Check permissions
+        if not current_user.is_dataset_editor:
+            raise Forbidden()
+
+        try:
+            DatasetService.check_dataset_permission(dataset, current_user)
+        except services.errors.account.NoPermissionError as e:
+            raise Forbidden(str(e))
+
+        # Validate request payload
+        payload = GenerateSummaryPayload.model_validate(console_ns.payload or {})
+        document_list = payload.document_list
+
+        if not document_list:
+            from werkzeug.exceptions import BadRequest
+
+            raise BadRequest("document_list cannot be empty.")
+
+        # Check if dataset configuration supports summary generation
+        if dataset.indexing_technique != "high_quality":
+            raise ValueError(
+                f"Summary generation is only available for 'high_quality' indexing technique. "
+                f"Current indexing technique: {dataset.indexing_technique}"
+            )
+
+        summary_index_setting = dataset.summary_index_setting
+        if not summary_index_setting or not summary_index_setting.get("enable"):
+            raise ValueError("Summary index is not enabled for this dataset. Please enable it in the dataset settings.")
+
+        # Verify all documents exist and belong to the dataset
+        documents = DocumentService.get_documents_by_ids(dataset_id, document_list)
+
+        if len(documents) != len(document_list):
+            found_ids = {doc.id for doc in documents}
+            missing_ids = set(document_list) - found_ids
+            raise NotFound(f"Some documents not found: {list(missing_ids)}")
+
+        # Dispatch async tasks for each document
+        for document in documents:
+            # Skip qa_model documents as they don't generate summaries
+            if document.doc_form == "qa_model":
+                logger.info("Skipping summary generation for qa_model document %s", document.id)
+                continue
+
+            # Dispatch async task
+            generate_summary_index_task.delay(dataset_id, document.id)
+            logger.info(
+                "Dispatched summary generation task for document %s in dataset %s",
+                document.id,
+                dataset_id,
+            )
+
+        return {"result": "success"}, 200
+
+
+@console_ns.route("/datasets/<uuid:dataset_id>/documents/<uuid:document_id>/summary-status")
+class DocumentSummaryStatusApi(DocumentResource):
+    @console_ns.doc("get_document_summary_status")
+    @console_ns.doc(description="Get summary index generation status for a document")
+    @console_ns.doc(params={"dataset_id": "Dataset ID", "document_id": "Document ID"})
+    @console_ns.response(200, "Summary status retrieved successfully")
+    @console_ns.response(404, "Document not found")
+    @setup_required
+    @login_required
+    @account_initialization_required
+    def get(self, dataset_id, document_id):
+        """
+        Get summary index generation status for a document.
+
+        Returns:
+        - total_segments: Total number of segments in the document
+        - summary_status: Dictionary with status counts
+          - completed: Number of summaries completed
+          - generating: Number of summaries being generated
+          - error: Number of summaries with errors
+          - not_started: Number of segments without summary records
+        - summaries: List of summary records with status and content preview
+        """
+        current_user, _ = current_account_with_tenant()
+        dataset_id = str(dataset_id)
+        document_id = str(document_id)
+
+        # Get dataset
+        dataset = DatasetService.get_dataset(dataset_id)
+        if not dataset:
+            raise NotFound("Dataset not found.")
+
+        # Check permissions
+        try:
+            DatasetService.check_dataset_permission(dataset, current_user)
+        except services.errors.account.NoPermissionError as e:
+            raise Forbidden(str(e))
+
+        # Get summary status detail from service
+        from services.summary_index_service import SummaryIndexService
+
+        result = SummaryIndexService.get_document_summary_status_detail(
+            document_id=document_id,
+            dataset_id=dataset_id,
+        )
+
+        return result, 200

api/controllers/console/datasets/datasets_segments.py

@@ -41,6 +41,17 @@ from services.errors.chunk import ChildChunkIndexingError as ChildChunkIndexingS
 from tasks.batch_create_segment_to_index_task import batch_create_segment_to_index_task
 
 
+def _get_segment_with_summary(segment, dataset_id):
+    """Helper function to marshal segment and add summary information."""
+    from services.summary_index_service import SummaryIndexService
+
+    segment_dict = dict(marshal(segment, segment_fields))
+    # Query summary for this segment (only enabled summaries)
+    summary = SummaryIndexService.get_segment_summary(segment_id=segment.id, dataset_id=dataset_id)
+    segment_dict["summary"] = summary.summary_content if summary else None
+    return segment_dict
+
+
 class SegmentListQuery(BaseModel):
     limit: int = Field(default=20, ge=1, le=100)
     status: list[str] = Field(default_factory=list)
@@ -63,6 +74,7 @@ class SegmentUpdatePayload(BaseModel):
     keywords: list[str] | None = None
     regenerate_child_chunks: bool = False
     attachment_ids: list[str] | None = None
+    summary: str | None = None  # Summary content for summary index
 
 
 class BatchImportPayload(BaseModel):
@@ -181,8 +193,25 @@ class DatasetDocumentSegmentListApi(Resource):
 
         segments = db.paginate(select=query, page=page, per_page=limit, max_per_page=100, error_out=False)
 
+        # Query summaries for all segments in this page (batch query for efficiency)
+        segment_ids = [segment.id for segment in segments.items]
+        summaries = {}
+        if segment_ids:
+            from services.summary_index_service import SummaryIndexService
+
+            summary_records = SummaryIndexService.get_segments_summaries(segment_ids=segment_ids, dataset_id=dataset_id)
+            # Only include enabled summaries (already filtered by service)
+            summaries = {chunk_id: summary.summary_content for chunk_id, summary in summary_records.items()}
+
+        # Add summary to each segment
+        segments_with_summary = []
+        for segment in segments.items:
+            segment_dict = dict(marshal(segment, segment_fields))
+            segment_dict["summary"] = summaries.get(segment.id)
+            segments_with_summary.append(segment_dict)
+
         response = {
-            "data": marshal(segments.items, segment_fields),
+            "data": segments_with_summary,
             "limit": limit,
             "total": segments.total,
             "total_pages": segments.pages,
@@ -328,7 +357,7 @@ class DatasetDocumentSegmentAddApi(Resource):
         payload_dict = payload.model_dump(exclude_none=True)
         SegmentService.segment_create_args_validate(payload_dict, document)
         segment = SegmentService.create_segment(payload_dict, document, dataset)
-        return {"data": marshal(segment, segment_fields), "doc_form": document.doc_form}, 200
+        return {"data": _get_segment_with_summary(segment, dataset_id), "doc_form": document.doc_form}, 200
 
 
 @console_ns.route("/datasets/<uuid:dataset_id>/documents/<uuid:document_id>/segments/<uuid:segment_id>")
@@ -390,10 +419,12 @@ class DatasetDocumentSegmentUpdateApi(Resource):
         payload = SegmentUpdatePayload.model_validate(console_ns.payload or {})
         payload_dict = payload.model_dump(exclude_none=True)
         SegmentService.segment_create_args_validate(payload_dict, document)
+
+        # Update segment (summary update with change detection is handled in SegmentService.update_segment)
         segment = SegmentService.update_segment(
             SegmentUpdateArgs.model_validate(payload.model_dump(exclude_none=True)), segment, document, dataset
         )
-        return {"data": marshal(segment, segment_fields), "doc_form": document.doc_form}, 200
+        return {"data": _get_segment_with_summary(segment, dataset_id), "doc_form": document.doc_form}, 200
 
     @setup_required
     @login_required

api/controllers/console/datasets/hit_testing.py

@@ -1,6 +1,13 @@
-from flask_restx import Resource
+from flask_restx import Resource, fields
 
 from controllers.common.schema import register_schema_model
+from fields.hit_testing_fields import (
+    child_chunk_fields,
+    document_fields,
+    files_fields,
+    hit_testing_record_fields,
+    segment_fields,
+)
 from libs.login import login_required
 
 from .. import console_ns
@@ -14,13 +21,45 @@ from ..wraps import (
 register_schema_model(console_ns, HitTestingPayload)
 
 
+def _get_or_create_model(model_name: str, field_def):
+    """Get or create a flask_restx model to avoid dict type issues in Swagger."""
+    existing = console_ns.models.get(model_name)
+    if existing is None:
+        existing = console_ns.model(model_name, field_def)
+    return existing
+
+
+# Register models for flask_restx to avoid dict type issues in Swagger
+document_model = _get_or_create_model("HitTestingDocument", document_fields)
+
+segment_fields_copy = segment_fields.copy()
+segment_fields_copy["document"] = fields.Nested(document_model)
+segment_model = _get_or_create_model("HitTestingSegment", segment_fields_copy)
+
+child_chunk_model = _get_or_create_model("HitTestingChildChunk", child_chunk_fields)
+files_model = _get_or_create_model("HitTestingFile", files_fields)
+
+hit_testing_record_fields_copy = hit_testing_record_fields.copy()
+hit_testing_record_fields_copy["segment"] = fields.Nested(segment_model)
+hit_testing_record_fields_copy["child_chunks"] = fields.List(fields.Nested(child_chunk_model))
+hit_testing_record_fields_copy["files"] = fields.List(fields.Nested(files_model))
+hit_testing_record_model = _get_or_create_model("HitTestingRecord", hit_testing_record_fields_copy)
+
+# Response model for hit testing API
+hit_testing_response_fields = {
+    "query": fields.String,
+    "records": fields.List(fields.Nested(hit_testing_record_model)),
+}
+hit_testing_response_model = _get_or_create_model("HitTestingResponse", hit_testing_response_fields)
+
+
 @console_ns.route("/datasets/<uuid:dataset_id>/hit-testing")
 class HitTestingApi(Resource, DatasetsHitTestingBase):
     @console_ns.doc("test_dataset_retrieval")
     @console_ns.doc(description="Test dataset knowledge retrieval")
     @console_ns.doc(params={"dataset_id": "Dataset ID"})
     @console_ns.expect(console_ns.models[HitTestingPayload.__name__])
-    @console_ns.response(200, "Hit testing completed successfully")
+    @console_ns.response(200, "Hit testing completed successfully", model=hit_testing_response_model)
     @console_ns.response(404, "Dataset not found")
     @console_ns.response(400, "Invalid parameters")
     @setup_required

api/controllers/console/human_input_form.py

@@ -0,0 +1,217 @@
+"""
+Console/Studio Human Input Form APIs.
+"""
+
+import json
+import logging
+from collections.abc import Generator
+
+from flask import Response, jsonify, request
+from flask_restx import Resource, reqparse
+from sqlalchemy import select
+from sqlalchemy.orm import Session, sessionmaker
+
+from controllers.console import console_ns
+from controllers.console.wraps import account_initialization_required, setup_required
+from controllers.web.error import InvalidArgumentError, NotFoundError
+from core.app.apps.advanced_chat.app_generator import AdvancedChatAppGenerator
+from core.app.apps.common.workflow_response_converter import WorkflowResponseConverter
+from core.app.apps.message_generator import MessageGenerator
+from core.app.apps.workflow.app_generator import WorkflowAppGenerator
+from extensions.ext_database import db
+from libs.login import current_account_with_tenant, login_required
+from models import App
+from models.enums import CreatorUserRole
+from models.human_input import RecipientType
+from models.model import AppMode
+from models.workflow import WorkflowRun
+from repositories.factory import DifyAPIRepositoryFactory
+from services.human_input_service import Form, HumanInputService
+from services.workflow_event_snapshot_service import build_workflow_event_stream
+
+logger = logging.getLogger(__name__)
+
+
+def _jsonify_form_definition(form: Form) -> Response:
+    payload = form.get_definition().model_dump()
+    payload["expiration_time"] = int(form.expiration_time.timestamp())
+    return Response(json.dumps(payload, ensure_ascii=False), mimetype="application/json")
+
+
+@console_ns.route("/form/human_input/<string:form_token>")
+class ConsoleHumanInputFormApi(Resource):
+    """Console API for getting human input form definition."""
+
+    @staticmethod
+    def _ensure_console_access(form: Form):
+        _, current_tenant_id = current_account_with_tenant()
+
+        if form.tenant_id != current_tenant_id:
+            raise NotFoundError("App not found")
+
+    @setup_required
+    @login_required
+    @account_initialization_required
+    def get(self, form_token: str):
+        """
+        Get human input form definition by form token.
+
+        GET /console/api/form/human_input/<form_token>
+        """
+        service = HumanInputService(db.engine)
+        form = service.get_form_definition_by_token_for_console(form_token)
+        if form is None:
+            raise NotFoundError(f"form not found, token={form_token}")
+
+        self._ensure_console_access(form)
+
+        return _jsonify_form_definition(form)
+
+    @account_initialization_required
+    @login_required
+    def post(self, form_token: str):
+        """
+        Submit human input form by form token.
+
+        POST /console/api/form/human_input/<form_token>
+
+        Request body:
+        {
+            "inputs": {
+                "content": "User input content"
+            },
+            "action": "Approve"
+        }
+        """
+        parser = reqparse.RequestParser()
+        parser.add_argument("inputs", type=dict, required=True, location="json")
+        parser.add_argument("action", type=str, required=True, location="json")
+        args = parser.parse_args()
+        current_user, _ = current_account_with_tenant()
+
+        service = HumanInputService(db.engine)
+        form = service.get_form_by_token(form_token)
+        if form is None:
+            raise NotFoundError(f"form not found, token={form_token}")
+
+        self._ensure_console_access(form)
+
+        recipient_type = form.recipient_type
+        if recipient_type not in {RecipientType.CONSOLE, RecipientType.BACKSTAGE}:
+            raise NotFoundError(f"form not found, token={form_token}")
+        # The type checker is not smart enought to validate the following invariant.
+        # So we need to assert it manually.
+        assert recipient_type is not None, "recipient_type cannot be None here."
+
+        service.submit_form_by_token(
+            recipient_type=recipient_type,
+            form_token=form_token,
+            selected_action_id=args["action"],
+            form_data=args["inputs"],
+            submission_user_id=current_user.id,
+        )
+
+        return jsonify({})
+
+
+@console_ns.route("/workflow/<string:workflow_run_id>/events")
+class ConsoleWorkflowEventsApi(Resource):
+    """Console API for getting workflow execution events after resume."""
+
+    @account_initialization_required
+    @login_required
+    def get(self, workflow_run_id: str):
+        """
+        Get workflow execution events stream after resume.
+
+        GET /console/api/workflow/<workflow_run_id>/events
+
+        Returns Server-Sent Events stream.
+        """
+
+        user, tenant_id = current_account_with_tenant()
+        session_maker = sessionmaker(db.engine)
+        repo = DifyAPIRepositoryFactory.create_api_workflow_run_repository(session_maker)
+        workflow_run = repo.get_workflow_run_by_id_and_tenant_id(
+            tenant_id=tenant_id,
+            run_id=workflow_run_id,
+        )
+        if workflow_run is None:
+            raise NotFoundError(f"WorkflowRun not found, id={workflow_run_id}")
+
+        if workflow_run.created_by_role != CreatorUserRole.ACCOUNT:
+            raise NotFoundError(f"WorkflowRun not created by account, id={workflow_run_id}")
+
+        if workflow_run.created_by != user.id:
+            raise NotFoundError(f"WorkflowRun not created by the current account, id={workflow_run_id}")
+
+        with Session(expire_on_commit=False, bind=db.engine) as session:
+            app = _retrieve_app_for_workflow_run(session, workflow_run)
+
+        if workflow_run.finished_at is not None:
+            # TODO(QuantumGhost): should we modify the handling for finished workflow run here?
+            response = WorkflowResponseConverter.workflow_run_result_to_finish_response(
+                task_id=workflow_run.id,
+                workflow_run=workflow_run,
+                creator_user=user,
+            )
+
+            payload = response.model_dump(mode="json")
+            payload["event"] = response.event.value
+
+            def _generate_finished_events() -> Generator[str, None, None]:
+                yield f"data: {json.dumps(payload)}\n\n"
+
+            event_generator = _generate_finished_events
+
+        else:
+            msg_generator = MessageGenerator()
+            if app.mode == AppMode.ADVANCED_CHAT:
+                generator = AdvancedChatAppGenerator()
+            elif app.mode == AppMode.WORKFLOW:
+                generator = WorkflowAppGenerator()
+            else:
+                raise InvalidArgumentError(f"cannot subscribe to workflow run, workflow_run_id={workflow_run.id}")
+
+            include_state_snapshot = request.args.get("include_state_snapshot", "false").lower() == "true"
+
+            def _generate_stream_events():
+                if include_state_snapshot:
+                    return generator.convert_to_event_stream(
+                        build_workflow_event_stream(
+                            app_mode=AppMode(app.mode),
+                            workflow_run=workflow_run,
+                            tenant_id=workflow_run.tenant_id,
+                            app_id=workflow_run.app_id,
+                            session_maker=session_maker,
+                        )
+                    )
+                return generator.convert_to_event_stream(
+                    msg_generator.retrieve_events(AppMode(app.mode), workflow_run.id),
+                )
+
+            event_generator = _generate_stream_events
+
+        return Response(
+            event_generator(),
+            mimetype="text/event-stream",
+            headers={
+                "Cache-Control": "no-cache",
+                "Connection": "keep-alive",
+            },
+        )
+
+
+def _retrieve_app_for_workflow_run(session: Session, workflow_run: WorkflowRun):
+    query = select(App).where(
+        App.id == workflow_run.app_id,
+        App.tenant_id == workflow_run.tenant_id,
+    )
+    app = session.scalars(query).first()
+    if app is None:
+        raise AssertionError(
+            f"App not found for WorkflowRun, workflow_run_id={workflow_run.id}, "
+            f"app_id={workflow_run.app_id}, tenant_id={workflow_run.tenant_id}"
+        )
+
+    return app

api/controllers/service_api/app/workflow.py

@@ -33,8 +33,9 @@ from core.workflow.graph_engine.manager import GraphEngineManager
 from extensions.ext_database import db
 from fields.workflow_app_log_fields import build_workflow_app_log_pagination_model
 from libs import helper
-from libs.helper import TimestampField
+from libs.helper import OptionalTimestampField, TimestampField
 from models.model import App, AppMode, EndUser
+from models.workflow import WorkflowRun
 from repositories.factory import DifyAPIRepositoryFactory
 from services.app_generate_service import AppGenerateService
 from services.errors.app import IsDraftWorkflowError, WorkflowIdFormatError, WorkflowNotFoundError
@@ -63,17 +64,32 @@ class WorkflowLogQuery(BaseModel):
 
 register_schema_models(service_api_ns, WorkflowRunPayload, WorkflowLogQuery)
 
+
+class WorkflowRunStatusField(fields.Raw):
+    def output(self, key, obj: WorkflowRun, **kwargs):
+        return obj.status.value
+
+
+class WorkflowRunOutputsField(fields.Raw):
+    def output(self, key, obj: WorkflowRun, **kwargs):
+        if obj.status == WorkflowExecutionStatus.PAUSED:
+            return {}
+
+        outputs = obj.outputs_dict
+        return outputs or {}
+
+
 workflow_run_fields = {
     "id": fields.String,
     "workflow_id": fields.String,
-    "status": fields.String,
+    "status": WorkflowRunStatusField,
     "inputs": fields.Raw,
-    "outputs": fields.Raw,
+    "outputs": WorkflowRunOutputsField,
     "error": fields.String,
     "total_steps": fields.Integer,
     "total_tokens": fields.Integer,
     "created_at": TimestampField,
-    "finished_at": TimestampField,
+    "finished_at": OptionalTimestampField,
     "elapsed_time": fields.Float,
 }
 

api/controllers/service_api/dataset/dataset.py

@@ -46,6 +46,7 @@ class DatasetCreatePayload(BaseModel):
     retrieval_model: RetrievalModel | None = None
     embedding_model: str | None = None
     embedding_model_provider: str | None = None
+    summary_index_setting: dict | None = None
 
 
 class DatasetUpdatePayload(BaseModel):
@@ -217,6 +218,7 @@ class DatasetListApi(DatasetApiResource):
                 embedding_model_provider=payload.embedding_model_provider,
                 embedding_model_name=payload.embedding_model,
                 retrieval_model=payload.retrieval_model,
+                summary_index_setting=payload.summary_index_setting,
             )
         except services.errors.dataset.DatasetNameDuplicateError:
             raise DatasetNameDuplicateError()

api/controllers/service_api/dataset/document.py

@@ -45,6 +45,7 @@ from services.entities.knowledge_entities.knowledge_entities import (
     Segmentation,
 )
 from services.file_service import FileService
+from services.summary_index_service import SummaryIndexService
 
 
 class DocumentTextCreatePayload(BaseModel):
@@ -508,6 +509,12 @@ class DocumentListApi(DatasetApiResource):
         )
         documents = paginated_documents.items
 
+        DocumentService.enrich_documents_with_summary_index_status(
+            documents=documents,
+            dataset=dataset,
+            tenant_id=tenant_id,
+        )
+
         response = {
             "data": marshal(documents, document_fields),
             "has_more": len(documents) == query_params.limit,
@@ -612,6 +619,16 @@ class DocumentApi(DatasetApiResource):
         if metadata not in self.METADATA_CHOICES:
             raise InvalidMetadataError(f"Invalid metadata value: {metadata}")
 
+        # Calculate summary_index_status if needed
+        summary_index_status = None
+        has_summary_index = dataset.summary_index_setting and dataset.summary_index_setting.get("enable") is True
+        if has_summary_index and document.need_summary is True:
+            summary_index_status = SummaryIndexService.get_document_summary_index_status(
+                document_id=document_id,
+                dataset_id=dataset_id,
+                tenant_id=tenant_id,
+            )
+
         if metadata == "only":
             response = {"id": document.id, "doc_type": document.doc_type, "doc_metadata": document.doc_metadata_details}
         elif metadata == "without":
@@ -646,6 +663,8 @@ class DocumentApi(DatasetApiResource):
                 "display_status": document.display_status,
                 "doc_form": document.doc_form,
                 "doc_language": document.doc_language,
+                "summary_index_status": summary_index_status,
+                "need_summary": document.need_summary if document.need_summary is not None else False,
             }
         else:
             dataset_process_rules = DatasetService.get_process_rules(dataset_id)
@@ -681,6 +700,8 @@ class DocumentApi(DatasetApiResource):
                 "display_status": document.display_status,
                 "doc_form": document.doc_form,
                 "doc_language": document.doc_language,
+                "summary_index_status": summary_index_status,
+                "need_summary": document.need_summary if document.need_summary is not None else False,
             }
 
         return response

api/controllers/web/__init__.py

@@ -23,6 +23,7 @@ from . import (
     feature,
     files,
     forgot_password,
+    human_input_form,
     login,
     message,
     passport,
@@ -30,6 +31,7 @@ from . import (
     saved_message,
     site,
     workflow,
+    workflow_events,
 )
 
 api.add_namespace(web_ns)
@@ -44,6 +46,7 @@ __all__ = [
     "feature",
     "files",
     "forgot_password",
+    "human_input_form",
     "login",
     "message",
     "passport",
@@ -52,4 +55,5 @@ __all__ = [
     "site",
     "web_ns",
     "workflow",
+    "workflow_events",
 ]

api/controllers/web/error.py

@@ -117,6 +117,12 @@ class InvokeRateLimitError(BaseHTTPException):
     code = 429
 
 
+class WebFormRateLimitExceededError(BaseHTTPException):
+    error_code = "web_form_rate_limit_exceeded"
+    description = "Too many form requests. Please try again later."
+    code = 429
+
+
 class NotFoundError(BaseHTTPException):
     error_code = "not_found"
     code = 404

api/controllers/web/human_input_form.py

@@ -0,0 +1,164 @@
+"""
+Web App Human Input Form APIs.
+"""
+
+import json
+import logging
+from datetime import datetime
+
+from flask import Response, request
+from flask_restx import Resource, reqparse
+from werkzeug.exceptions import Forbidden
+
+from configs import dify_config
+from controllers.web import web_ns
+from controllers.web.error import NotFoundError, WebFormRateLimitExceededError
+from controllers.web.site import serialize_app_site_payload
+from extensions.ext_database import db
+from libs.helper import RateLimiter, extract_remote_ip
+from models.account import TenantStatus
+from models.model import App, Site
+from services.human_input_service import Form, FormNotFoundError, HumanInputService
+
+logger = logging.getLogger(__name__)
+
+_FORM_SUBMIT_RATE_LIMITER = RateLimiter(
+    prefix="web_form_submit_rate_limit",
+    max_attempts=dify_config.WEB_FORM_SUBMIT_RATE_LIMIT_MAX_ATTEMPTS,
+    time_window=dify_config.WEB_FORM_SUBMIT_RATE_LIMIT_WINDOW_SECONDS,
+)
+_FORM_ACCESS_RATE_LIMITER = RateLimiter(
+    prefix="web_form_access_rate_limit",
+    max_attempts=dify_config.WEB_FORM_SUBMIT_RATE_LIMIT_MAX_ATTEMPTS,
+    time_window=dify_config.WEB_FORM_SUBMIT_RATE_LIMIT_WINDOW_SECONDS,
+)
+
+
+def _stringify_default_values(values: dict[str, object]) -> dict[str, str]:
+    result: dict[str, str] = {}
+    for key, value in values.items():
+        if value is None:
+            result[key] = ""
+        elif isinstance(value, (dict, list)):
+            result[key] = json.dumps(value, ensure_ascii=False)
+        else:
+            result[key] = str(value)
+    return result
+
+
+def _to_timestamp(value: datetime) -> int:
+    return int(value.timestamp())
+
+
+def _jsonify_form_definition(form: Form, site_payload: dict | None = None) -> Response:
+    """Return the form payload (optionally with site) as a JSON response."""
+    definition_payload = form.get_definition().model_dump()
+    payload = {
+        "form_content": definition_payload["rendered_content"],
+        "inputs": definition_payload["inputs"],
+        "resolved_default_values": _stringify_default_values(definition_payload["default_values"]),
+        "user_actions": definition_payload["user_actions"],
+        "expiration_time": _to_timestamp(form.expiration_time),
+    }
+    if site_payload is not None:
+        payload["site"] = site_payload
+    return Response(json.dumps(payload, ensure_ascii=False), mimetype="application/json")
+
+
+# TODO(QuantumGhost): disable authorization for web app
+# form api temporarily
+
+
+@web_ns.route("/form/human_input/<string:form_token>")
+# class HumanInputFormApi(WebApiResource):
+class HumanInputFormApi(Resource):
+    """API for getting and submitting human input forms via the web app."""
+
+    # def get(self, _app_model: App, _end_user: EndUser, form_token: str):
+    def get(self, form_token: str):
+        """
+        Get human input form definition by token.
+
+        GET /api/form/human_input/<form_token>
+        """
+        ip_address = extract_remote_ip(request)
+        if _FORM_ACCESS_RATE_LIMITER.is_rate_limited(ip_address):
+            raise WebFormRateLimitExceededError()
+        _FORM_ACCESS_RATE_LIMITER.increment_rate_limit(ip_address)
+
+        service = HumanInputService(db.engine)
+        # TODO(QuantumGhost): forbid submision for form tokens
+        # that are only for console.
+        form = service.get_form_by_token(form_token)
+
+        if form is None:
+            raise NotFoundError("Form not found")
+
+        service.ensure_form_active(form)
+        app_model, site = _get_app_site_from_form(form)
+
+        return _jsonify_form_definition(form, site_payload=serialize_app_site_payload(app_model, site, None))
+
+    # def post(self, _app_model: App, _end_user: EndUser, form_token: str):
+    def post(self, form_token: str):
+        """
+        Submit human input form by token.
+
+        POST /api/form/human_input/<form_token>
+
+        Request body:
+        {
+            "inputs": {
+                "content": "User input content"
+            },
+            "action": "Approve"
+        }
+        """
+        parser = reqparse.RequestParser()
+        parser.add_argument("inputs", type=dict, required=True, location="json")
+        parser.add_argument("action", type=str, required=True, location="json")
+        args = parser.parse_args()
+
+        ip_address = extract_remote_ip(request)
+        if _FORM_SUBMIT_RATE_LIMITER.is_rate_limited(ip_address):
+            raise WebFormRateLimitExceededError()
+        _FORM_SUBMIT_RATE_LIMITER.increment_rate_limit(ip_address)
+
+        service = HumanInputService(db.engine)
+        form = service.get_form_by_token(form_token)
+        if form is None:
+            raise NotFoundError("Form not found")
+
+        if (recipient_type := form.recipient_type) is None:
+            logger.warning("Recipient type is None for form, form_id=%", form.id)
+            raise AssertionError("Recipient type is None")
+
+        try:
+            service.submit_form_by_token(
+                recipient_type=recipient_type,
+                form_token=form_token,
+                selected_action_id=args["action"],
+                form_data=args["inputs"],
+                submission_end_user_id=None,
+                # submission_end_user_id=_end_user.id,
+            )
+        except FormNotFoundError:
+            raise NotFoundError("Form not found")
+
+        return {}, 200
+
+
+def _get_app_site_from_form(form: Form) -> tuple[App, Site]:
+    """Resolve App/Site for the form's app and validate tenant status."""
+    app_model = db.session.query(App).where(App.id == form.app_id).first()
+    if app_model is None or app_model.tenant_id != form.tenant_id:
+        raise NotFoundError("Form not found")
+
+    site = db.session.query(Site).where(Site.app_id == app_model.id).first()
+    if site is None:
+        raise Forbidden()
+
+    if app_model.tenant and app_model.tenant.status == TenantStatus.ARCHIVE:
+        raise Forbidden()
+
+    return app_model, site

api/controllers/web/site.py

@@ -1,4 +1,6 @@
-from flask_restx import fields, marshal_with
+from typing import cast
+
+from flask_restx import fields, marshal, marshal_with
 from werkzeug.exceptions import Forbidden
 
 from configs import dify_config
@@ -7,7 +9,7 @@ from controllers.web.wraps import WebApiResource
 from extensions.ext_database import db
 from libs.helper import AppIconUrlField
 from models.account import TenantStatus
-from models.model import Site
+from models.model import App, Site
 from services.feature_service import FeatureService
 
 
@@ -108,3 +110,14 @@ class AppSiteInfo:
                 "remove_webapp_brand": remove_webapp_brand,
                 "replace_webapp_logo": replace_webapp_logo,
             }
+
+
+def serialize_site(site: Site) -> dict:
+    """Serialize Site model using the same schema as AppSiteApi."""
+    return cast(dict, marshal(site, AppSiteApi.site_fields))
+
+
+def serialize_app_site_payload(app_model: App, site: Site, end_user_id: str | None) -> dict:
+    can_replace_logo = FeatureService.get_features(app_model.tenant_id).can_replace_logo
+    app_site_info = AppSiteInfo(app_model.tenant, app_model, site, end_user_id, can_replace_logo)
+    return cast(dict, marshal(app_site_info, AppSiteApi.app_fields))

api/controllers/web/workflow_events.py

@@ -0,0 +1,112 @@
+"""
+Web App Workflow Resume APIs.
+"""
+
+import json
+from collections.abc import Generator
+
+from flask import Response, request
+from sqlalchemy.orm import sessionmaker
+
+from controllers.web import api
+from controllers.web.error import InvalidArgumentError, NotFoundError
+from controllers.web.wraps import WebApiResource
+from core.app.apps.advanced_chat.app_generator import AdvancedChatAppGenerator
+from core.app.apps.base_app_generator import BaseAppGenerator
+from core.app.apps.common.workflow_response_converter import WorkflowResponseConverter
+from core.app.apps.message_generator import MessageGenerator
+from core.app.apps.workflow.app_generator import WorkflowAppGenerator
+from extensions.ext_database import db
+from models.enums import CreatorUserRole
+from models.model import App, AppMode, EndUser
+from repositories.factory import DifyAPIRepositoryFactory
+from services.workflow_event_snapshot_service import build_workflow_event_stream
+
+
+class WorkflowEventsApi(WebApiResource):
+    """API for getting workflow execution events after resume."""
+
+    def get(self, app_model: App, end_user: EndUser, task_id: str):
+        """
+        Get workflow execution events stream after resume.
+
+        GET /api/workflow/<task_id>/events
+
+        Returns Server-Sent Events stream.
+        """
+        workflow_run_id = task_id
+        session_maker = sessionmaker(db.engine)
+        repo = DifyAPIRepositoryFactory.create_api_workflow_run_repository(session_maker)
+        workflow_run = repo.get_workflow_run_by_id_and_tenant_id(
+            tenant_id=app_model.tenant_id,
+            run_id=workflow_run_id,
+        )
+
+        if workflow_run is None:
+            raise NotFoundError(f"WorkflowRun not found, id={workflow_run_id}")
+
+        if workflow_run.app_id != app_model.id:
+            raise NotFoundError(f"WorkflowRun not found, id={workflow_run_id}")
+
+        if workflow_run.created_by_role != CreatorUserRole.END_USER:
+            raise NotFoundError(f"WorkflowRun not created by end user, id={workflow_run_id}")
+
+        if workflow_run.created_by != end_user.id:
+            raise NotFoundError(f"WorkflowRun not created by the current end user, id={workflow_run_id}")
+
+        if workflow_run.finished_at is not None:
+            response = WorkflowResponseConverter.workflow_run_result_to_finish_response(
+                task_id=workflow_run.id,
+                workflow_run=workflow_run,
+                creator_user=end_user,
+            )
+
+            payload = response.model_dump(mode="json")
+            payload["event"] = response.event.value
+
+            def _generate_finished_events() -> Generator[str, None, None]:
+                yield f"data: {json.dumps(payload)}\n\n"
+
+            event_generator = _generate_finished_events
+        else:
+            app_mode = AppMode.value_of(app_model.mode)
+            msg_generator = MessageGenerator()
+            generator: BaseAppGenerator
+            if app_mode == AppMode.ADVANCED_CHAT:
+                generator = AdvancedChatAppGenerator()
+            elif app_mode == AppMode.WORKFLOW:
+                generator = WorkflowAppGenerator()
+            else:
+                raise InvalidArgumentError(f"cannot subscribe to workflow run, workflow_run_id={workflow_run.id}")
+
+            include_state_snapshot = request.args.get("include_state_snapshot", "false").lower() == "true"
+
+            def _generate_stream_events():
+                if include_state_snapshot:
+                    return generator.convert_to_event_stream(
+                        build_workflow_event_stream(
+                            app_mode=app_mode,
+                            workflow_run=workflow_run,
+                            tenant_id=app_model.tenant_id,
+                            app_id=app_model.id,
+                            session_maker=session_maker,
+                        )
+                    )
+                return generator.convert_to_event_stream(
+                    msg_generator.retrieve_events(app_mode, workflow_run.id),
+                )
+
+            event_generator = _generate_stream_events
+
+        return Response(
+            event_generator(),
+            mimetype="text/event-stream",
+            headers={
+                "Cache-Control": "no-cache",
+                "Connection": "keep-alive",
+            },
+        )
+
+
+# Register the APIs
+api.add_resource(WorkflowEventsApi, "/workflow/<string:task_id>/events")

api/core/app/apps/advanced_chat/app_generator.py

@@ -4,8 +4,8 @@ import contextvars
 import logging
 import threading
 import uuid
-from collections.abc import Generator, Mapping
-from typing import TYPE_CHECKING, Any, Literal, Union, overload
+from collections.abc import Generator, Mapping, Sequence
+from typing import TYPE_CHECKING, Any, Literal, TypeVar, Union, overload
 
 from flask import Flask, current_app
 from pydantic import ValidationError
@@ -29,21 +29,25 @@ 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 AdvancedChatAppGenerateEntity, InvokeFrom
 from core.app.entities.task_entities import ChatbotAppBlockingResponse, ChatbotAppStreamResponse
+from core.app.layers.pause_state_persist_layer import PauseStateLayerConfig, PauseStatePersistenceLayer
 from core.helper.trace_id_helper import extract_external_trace_id_from_args
 from core.model_runtime.errors.invoke import InvokeAuthorizationError
 from core.ops.ops_trace_manager import TraceQueueManager
 from core.prompt.utils.get_thread_messages_length import get_thread_messages_length
 from core.repositories import DifyCoreRepositoryFactory
+from core.workflow.graph_engine.layers.base import GraphEngineLayer
 from core.workflow.repositories.draft_variable_repository import (
     DraftVariableSaverFactory,
 )
 from core.workflow.repositories.workflow_execution_repository import WorkflowExecutionRepository
 from core.workflow.repositories.workflow_node_execution_repository import WorkflowNodeExecutionRepository
+from core.workflow.runtime import GraphRuntimeState
 from core.workflow.variable_loader import DUMMY_VARIABLE_LOADER, VariableLoader
 from extensions.ext_database import db
 from factories import file_factory
 from libs.flask_utils import preserve_flask_contexts
 from models import Account, App, Conversation, EndUser, Message, Workflow, WorkflowNodeExecutionTriggeredFrom
+from models.base import Base
 from models.enums import WorkflowRunTriggeredFrom
 from services.conversation_service import ConversationService
 from services.workflow_draft_variable_service import (
@@ -65,7 +69,9 @@ class AdvancedChatAppGenerator(MessageBasedAppGenerator):
         user: Union[Account, EndUser],
         args: Mapping[str, Any],
         invoke_from: InvokeFrom,
+        workflow_run_id: str,
         streaming: Literal[False],
+        pause_state_config: PauseStateLayerConfig | None = None,
     ) -> Mapping[str, Any]: ...
 
     @overload
@@ -74,9 +80,11 @@ class AdvancedChatAppGenerator(MessageBasedAppGenerator):
         app_model: App,
         workflow: Workflow,
         user: Union[Account, EndUser],
-        args: Mapping,
+        args: Mapping[str, Any],
         invoke_from: InvokeFrom,
+        workflow_run_id: str,
         streaming: Literal[True],
+        pause_state_config: PauseStateLayerConfig | None = None,
     ) -> Generator[Mapping | str, None, None]: ...
 
     @overload
@@ -85,9 +93,11 @@ class AdvancedChatAppGenerator(MessageBasedAppGenerator):
         app_model: App,
         workflow: Workflow,
         user: Union[Account, EndUser],
-        args: Mapping,
+        args: Mapping[str, Any],
         invoke_from: InvokeFrom,
+        workflow_run_id: str,
         streaming: bool,
+        pause_state_config: PauseStateLayerConfig | None = None,
     ) -> Mapping[str, Any] | Generator[str | Mapping, None, None]: ...
 
     def generate(
@@ -95,9 +105,11 @@ class AdvancedChatAppGenerator(MessageBasedAppGenerator):
         app_model: App,
         workflow: Workflow,
         user: Union[Account, EndUser],
-        args: Mapping,
+        args: Mapping[str, Any],
         invoke_from: InvokeFrom,
+        workflow_run_id: str,
         streaming: bool = True,
+        pause_state_config: PauseStateLayerConfig | None = None,
     ) -> Mapping[str, Any] | Generator[str | Mapping, None, None]:
         """
         Generate App response.
@@ -161,7 +173,6 @@ class AdvancedChatAppGenerator(MessageBasedAppGenerator):
             # always enable retriever resource in debugger mode
             app_config.additional_features.show_retrieve_source = True  # type: ignore
 
-        workflow_run_id = str(uuid.uuid4())
         # init application generate entity
         application_generate_entity = AdvancedChatAppGenerateEntity(
             task_id=str(uuid.uuid4()),
@@ -179,7 +190,7 @@ class AdvancedChatAppGenerator(MessageBasedAppGenerator):
             invoke_from=invoke_from,
             extras=extras,
             trace_manager=trace_manager,
-            workflow_run_id=workflow_run_id,
+            workflow_run_id=str(workflow_run_id),
         )
         contexts.plugin_tool_providers.set({})
         contexts.plugin_tool_providers_lock.set(threading.Lock())
@@ -216,6 +227,38 @@ class AdvancedChatAppGenerator(MessageBasedAppGenerator):
             workflow_node_execution_repository=workflow_node_execution_repository,
             conversation=conversation,
             stream=streaming,
+            pause_state_config=pause_state_config,
+        )
+
+    def resume(
+        self,
+        *,
+        app_model: App,
+        workflow: Workflow,
+        user: Union[Account, EndUser],
+        conversation: Conversation,
+        message: Message,
+        application_generate_entity: AdvancedChatAppGenerateEntity,
+        workflow_execution_repository: WorkflowExecutionRepository,
+        workflow_node_execution_repository: WorkflowNodeExecutionRepository,
+        graph_runtime_state: GraphRuntimeState,
+        pause_state_config: PauseStateLayerConfig | None = None,
+    ):
+        """
+        Resume a paused advanced chat execution.
+        """
+        return self._generate(
+            workflow=workflow,
+            user=user,
+            invoke_from=application_generate_entity.invoke_from,
+            application_generate_entity=application_generate_entity,
+            workflow_execution_repository=workflow_execution_repository,
+            workflow_node_execution_repository=workflow_node_execution_repository,
+            conversation=conversation,
+            message=message,
+            stream=application_generate_entity.stream,
+            pause_state_config=pause_state_config,
+            graph_runtime_state=graph_runtime_state,
         )
 
     def single_iteration_generate(
@@ -396,8 +439,12 @@ class AdvancedChatAppGenerator(MessageBasedAppGenerator):
         workflow_execution_repository: WorkflowExecutionRepository,
         workflow_node_execution_repository: WorkflowNodeExecutionRepository,
         conversation: Conversation | None = None,
+        message: Message | None = None,
         stream: bool = True,
         variable_loader: VariableLoader = DUMMY_VARIABLE_LOADER,
+        pause_state_config: PauseStateLayerConfig | None = None,
+        graph_runtime_state: GraphRuntimeState | None = None,
+        graph_engine_layers: Sequence[GraphEngineLayer] = (),
     ) -> Mapping[str, Any] | Generator[str | Mapping[str, Any], Any, None]:
         """
         Generate App response.
@@ -411,12 +458,12 @@ class AdvancedChatAppGenerator(MessageBasedAppGenerator):
         :param conversation: conversation
         :param stream: is stream
         """
-        is_first_conversation = False
-        if not conversation:
-            is_first_conversation = True
+        is_first_conversation = conversation is None
 
-        # init generate records
-        (conversation, message) = self._init_generate_records(application_generate_entity, conversation)
+        if conversation is not None and message is not None:
+            pass
+        else:
+            conversation, message = self._init_generate_records(application_generate_entity, conversation)
 
         if is_first_conversation:
             # update conversation features
@@ -439,6 +486,16 @@ class AdvancedChatAppGenerator(MessageBasedAppGenerator):
             message_id=message.id,
         )
 
+        graph_layers: list[GraphEngineLayer] = list(graph_engine_layers)
+        if pause_state_config is not None:
+            graph_layers.append(
+                PauseStatePersistenceLayer(
+                    session_factory=pause_state_config.session_factory,
+                    generate_entity=application_generate_entity,
+                    state_owner_user_id=pause_state_config.state_owner_user_id,
+                )
+            )
+
         # new thread with request context and contextvars
         context = contextvars.copy_context()
 
@@ -454,14 +511,25 @@ class AdvancedChatAppGenerator(MessageBasedAppGenerator):
                 "variable_loader": variable_loader,
                 "workflow_execution_repository": workflow_execution_repository,
                 "workflow_node_execution_repository": workflow_node_execution_repository,
+                "graph_engine_layers": tuple(graph_layers),
+                "graph_runtime_state": graph_runtime_state,
             },
         )
 
         worker_thread.start()
 
         # release database connection, because the following new thread operations may take a long time
-        db.session.refresh(workflow)
-        db.session.refresh(message)
+        with Session(bind=db.engine, expire_on_commit=False) as session:
+            workflow = _refresh_model(session, workflow)
+            message = _refresh_model(session, message)
+        #     workflow_ = session.get(Workflow, workflow.id)
+        #     assert workflow_ is not None
+        #     workflow = workflow_
+        #     message_ = session.get(Message, message.id)
+        #     assert message_ is not None
+        #     message = message_
+        # db.session.refresh(workflow)
+        # db.session.refresh(message)
         # db.session.refresh(user)
         db.session.close()
 
@@ -490,6 +558,8 @@ class AdvancedChatAppGenerator(MessageBasedAppGenerator):
         variable_loader: VariableLoader,
         workflow_execution_repository: WorkflowExecutionRepository,
         workflow_node_execution_repository: WorkflowNodeExecutionRepository,
+        graph_engine_layers: Sequence[GraphEngineLayer] = (),
+        graph_runtime_state: GraphRuntimeState | None = None,
     ):
         """
         Generate worker in a new thread.
@@ -547,6 +617,8 @@ class AdvancedChatAppGenerator(MessageBasedAppGenerator):
                 app=app,
                 workflow_execution_repository=workflow_execution_repository,
                 workflow_node_execution_repository=workflow_node_execution_repository,
+                graph_engine_layers=graph_engine_layers,
+                graph_runtime_state=graph_runtime_state,
             )
 
             try:
@@ -614,3 +686,13 @@ class AdvancedChatAppGenerator(MessageBasedAppGenerator):
             else:
                 logger.exception("Failed to process generate task pipeline, conversation_id: %s", conversation.id)
                 raise e
+
+
+_T = TypeVar("_T", bound=Base)
+
+
+def _refresh_model(session, model: _T) -> _T:
+    with Session(bind=db.engine, expire_on_commit=False) as session:
+        detach_model = session.get(type(model), model.id)
+        assert detach_model is not None
+        return detach_model

api/core/app/apps/advanced_chat/app_runner.py

@@ -66,6 +66,7 @@ class AdvancedChatAppRunner(WorkflowBasedAppRunner):
         workflow_execution_repository: WorkflowExecutionRepository,
         workflow_node_execution_repository: WorkflowNodeExecutionRepository,
         graph_engine_layers: Sequence[GraphEngineLayer] = (),
+        graph_runtime_state: GraphRuntimeState | None = None,
     ):
         super().__init__(
             queue_manager=queue_manager,
@@ -82,6 +83,7 @@ class AdvancedChatAppRunner(WorkflowBasedAppRunner):
         self._app = app
         self._workflow_execution_repository = workflow_execution_repository
         self._workflow_node_execution_repository = workflow_node_execution_repository
+        self._resume_graph_runtime_state = graph_runtime_state
 
     @trace_span(WorkflowAppRunnerHandler)
     def run(self):
@@ -110,7 +112,21 @@ class AdvancedChatAppRunner(WorkflowBasedAppRunner):
             invoke_from = InvokeFrom.DEBUGGER
         user_from = self._resolve_user_from(invoke_from)
 
-        if self.application_generate_entity.single_iteration_run or self.application_generate_entity.single_loop_run:
+        resume_state = self._resume_graph_runtime_state
+
+        if resume_state is not None:
+            graph_runtime_state = resume_state
+            variable_pool = graph_runtime_state.variable_pool
+            graph = self._init_graph(
+                graph_config=self._workflow.graph_dict,
+                graph_runtime_state=graph_runtime_state,
+                workflow_id=self._workflow.id,
+                tenant_id=self._workflow.tenant_id,
+                user_id=self.application_generate_entity.user_id,
+                invoke_from=invoke_from,
+                user_from=user_from,
+            )
+        elif self.application_generate_entity.single_iteration_run or self.application_generate_entity.single_loop_run:
             # Handle single iteration or single loop run
             graph, variable_pool, graph_runtime_state = self._prepare_single_node_execution(
                 workflow=self._workflow,

api/core/app/apps/advanced_chat/generate_task_pipeline.py

@@ -24,6 +24,8 @@ from core.app.entities.queue_entities import (
     QueueAgentLogEvent,
     QueueAnnotationReplyEvent,
     QueueErrorEvent,
+    QueueHumanInputFormFilledEvent,
+    QueueHumanInputFormTimeoutEvent,
     QueueIterationCompletedEvent,
     QueueIterationNextEvent,
     QueueIterationStartEvent,
@@ -42,6 +44,7 @@ from core.app.entities.queue_entities import (
     QueueTextChunkEvent,
     QueueWorkflowFailedEvent,
     QueueWorkflowPartialSuccessEvent,
+    QueueWorkflowPausedEvent,
     QueueWorkflowStartedEvent,
     QueueWorkflowSucceededEvent,
     WorkflowQueueMessage,
@@ -63,6 +66,8 @@ from core.base.tts import AppGeneratorTTSPublisher, AudioTrunk
 from core.model_runtime.entities.llm_entities import LLMUsage
 from core.model_runtime.utils.encoders import jsonable_encoder
 from core.ops.ops_trace_manager import TraceQueueManager
+from core.repositories.human_input_repository import HumanInputFormRepositoryImpl
+from core.workflow.entities.pause_reason import HumanInputRequired
 from core.workflow.enums import WorkflowExecutionStatus
 from core.workflow.nodes import NodeType
 from core.workflow.repositories.draft_variable_repository import DraftVariableSaverFactory
@@ -71,7 +76,8 @@ from core.workflow.system_variable import SystemVariable
 from extensions.ext_database import db
 from libs.datetime_utils import naive_utc_now
 from models import Account, Conversation, EndUser, Message, MessageFile
-from models.enums import CreatorUserRole
+from models.enums import CreatorUserRole, MessageStatus
+from models.execution_extra_content import HumanInputContent
 from models.workflow import Workflow
 
 logger = logging.getLogger(__name__)
@@ -128,6 +134,7 @@ class AdvancedChatAppGenerateTaskPipeline(GraphRuntimeStateSupport):
         )
 
         self._task_state = WorkflowTaskState()
+        self._seed_task_state_from_message(message)
         self._message_cycle_manager = MessageCycleManager(
             application_generate_entity=application_generate_entity, task_state=self._task_state
         )
@@ -135,6 +142,7 @@ class AdvancedChatAppGenerateTaskPipeline(GraphRuntimeStateSupport):
         self._application_generate_entity = application_generate_entity
         self._workflow_id = workflow.id
         self._workflow_features_dict = workflow.features_dict
+        self._workflow_tenant_id = workflow.tenant_id
         self._conversation_id = conversation.id
         self._conversation_mode = conversation.mode
         self._message_id = message.id
@@ -144,8 +152,13 @@ class AdvancedChatAppGenerateTaskPipeline(GraphRuntimeStateSupport):
         self._workflow_run_id: str = ""
         self._draft_var_saver_factory = draft_var_saver_factory
         self._graph_runtime_state: GraphRuntimeState | None = None
+        self._message_saved_on_pause = False
         self._seed_graph_runtime_state_from_queue_manager()
 
+    def _seed_task_state_from_message(self, message: Message) -> None:
+        if message.status == MessageStatus.PAUSED and message.answer:
+            self._task_state.answer = message.answer
+
     def process(self) -> Union[ChatbotAppBlockingResponse, Generator[ChatbotAppStreamResponse, None, None]]:
         """
         Process generate task pipeline.
@@ -308,6 +321,7 @@ class AdvancedChatAppGenerateTaskPipeline(GraphRuntimeStateSupport):
             task_id=self._application_generate_entity.task_id,
             workflow_run_id=run_id,
             workflow_id=self._workflow_id,
+            reason=event.reason,
         )
 
         yield workflow_start_resp
@@ -525,6 +539,35 @@ class AdvancedChatAppGenerateTaskPipeline(GraphRuntimeStateSupport):
         )
 
         yield workflow_finish_resp
+
+    def _handle_workflow_paused_event(
+        self,
+        event: QueueWorkflowPausedEvent,
+        **kwargs,
+    ) -> Generator[StreamResponse, None, None]:
+        """Handle workflow paused events."""
+        validated_state = self._ensure_graph_runtime_initialized()
+        responses = self._workflow_response_converter.workflow_pause_to_stream_response(
+            event=event,
+            task_id=self._application_generate_entity.task_id,
+            graph_runtime_state=validated_state,
+        )
+        for reason in event.reasons:
+            if isinstance(reason, HumanInputRequired):
+                self._persist_human_input_extra_content(form_id=reason.form_id, node_id=reason.node_id)
+        yield from responses
+        resolved_state: GraphRuntimeState | None = None
+        try:
+            resolved_state = self._ensure_graph_runtime_initialized()
+        except ValueError:
+            resolved_state = None
+
+        with self._database_session() as session:
+            self._save_message(session=session, graph_runtime_state=resolved_state)
+            message = self._get_message(session=session)
+            if message is not None:
+                message.status = MessageStatus.PAUSED
+            self._message_saved_on_pause = True
         self._base_task_pipeline.queue_manager.publish(QueueAdvancedChatMessageEndEvent(), PublishFrom.TASK_PIPELINE)
 
     def _handle_workflow_failed_event(
@@ -614,9 +657,10 @@ class AdvancedChatAppGenerateTaskPipeline(GraphRuntimeStateSupport):
                 reason=QueueMessageReplaceEvent.MessageReplaceReason.OUTPUT_MODERATION,
             )
 
-        # Save message
-        with self._database_session() as session:
-            self._save_message(session=session, graph_runtime_state=resolved_state)
+        # Save message unless it has already been persisted on pause.
+        if not self._message_saved_on_pause:
+            with self._database_session() as session:
+                self._save_message(session=session, graph_runtime_state=resolved_state)
 
         yield self._message_end_to_stream_response()
 
@@ -642,6 +686,65 @@ class AdvancedChatAppGenerateTaskPipeline(GraphRuntimeStateSupport):
         """Handle message replace events."""
         yield self._message_cycle_manager.message_replace_to_stream_response(answer=event.text, reason=event.reason)
 
+    def _handle_human_input_form_filled_event(
+        self, event: QueueHumanInputFormFilledEvent, **kwargs
+    ) -> Generator[StreamResponse, None, None]:
+        """Handle human input form filled events."""
+        self._persist_human_input_extra_content(node_id=event.node_id)
+        yield self._workflow_response_converter.human_input_form_filled_to_stream_response(
+            event=event, task_id=self._application_generate_entity.task_id
+        )
+
+    def _handle_human_input_form_timeout_event(
+        self, event: QueueHumanInputFormTimeoutEvent, **kwargs
+    ) -> Generator[StreamResponse, None, None]:
+        """Handle human input form timeout events."""
+        yield self._workflow_response_converter.human_input_form_timeout_to_stream_response(
+            event=event, task_id=self._application_generate_entity.task_id
+        )
+
+    def _persist_human_input_extra_content(self, *, node_id: str | None = None, form_id: str | None = None) -> None:
+        if not self._workflow_run_id or not self._message_id:
+            return
+
+        if form_id is None:
+            if node_id is None:
+                return
+            form_id = self._load_human_input_form_id(node_id=node_id)
+            if form_id is None:
+                logger.warning(
+                    "HumanInput form not found for workflow run %s node %s",
+                    self._workflow_run_id,
+                    node_id,
+                )
+                return
+
+        with self._database_session() as session:
+            exists_stmt = select(HumanInputContent).where(
+                HumanInputContent.workflow_run_id == self._workflow_run_id,
+                HumanInputContent.message_id == self._message_id,
+                HumanInputContent.form_id == form_id,
+            )
+            if session.scalar(exists_stmt) is not None:
+                return
+
+            content = HumanInputContent(
+                workflow_run_id=self._workflow_run_id,
+                message_id=self._message_id,
+                form_id=form_id,
+            )
+            session.add(content)
+
+    def _load_human_input_form_id(self, *, node_id: str) -> str | None:
+        form_repository = HumanInputFormRepositoryImpl(
+            session_factory=db.engine,
+            tenant_id=self._workflow_tenant_id,
+        )
+        form = form_repository.get_form(self._workflow_run_id, node_id)
+        if form is None:
+            return None
+        return form.id
+
     def _handle_agent_log_event(self, event: QueueAgentLogEvent, **kwargs) -> Generator[StreamResponse, None, None]:
         """Handle agent log events."""
         yield self._workflow_response_converter.handle_agent_log(
@@ -659,6 +762,7 @@ class AdvancedChatAppGenerateTaskPipeline(GraphRuntimeStateSupport):
             QueueWorkflowStartedEvent: self._handle_workflow_started_event,
             QueueWorkflowSucceededEvent: self._handle_workflow_succeeded_event,
             QueueWorkflowPartialSuccessEvent: self._handle_workflow_partial_success_event,
+            QueueWorkflowPausedEvent: self._handle_workflow_paused_event,
             QueueWorkflowFailedEvent: self._handle_workflow_failed_event,
             # Node events
             QueueNodeRetryEvent: self._handle_node_retry_event,
@@ -680,6 +784,8 @@ class AdvancedChatAppGenerateTaskPipeline(GraphRuntimeStateSupport):
             QueueMessageReplaceEvent: self._handle_message_replace_event,
             QueueAdvancedChatMessageEndEvent: self._handle_advanced_chat_message_end_event,
             QueueAgentLogEvent: self._handle_agent_log_event,
+            QueueHumanInputFormFilledEvent: self._handle_human_input_form_filled_event,
+            QueueHumanInputFormTimeoutEvent: self._handle_human_input_form_timeout_event,
         }
 
     def _dispatch_event(
@@ -747,6 +853,9 @@ class AdvancedChatAppGenerateTaskPipeline(GraphRuntimeStateSupport):
                 case QueueWorkflowFailedEvent():
                     yield from self._handle_workflow_failed_event(event, trace_manager=trace_manager)
                     break
+                case QueueWorkflowPausedEvent():
+                    yield from self._handle_workflow_paused_event(event)
+                    break
 
                 case QueueStopEvent():
                     yield from self._handle_stop_event(event, graph_runtime_state=None, trace_manager=trace_manager)
@@ -772,6 +881,11 @@ class AdvancedChatAppGenerateTaskPipeline(GraphRuntimeStateSupport):
 
     def _save_message(self, *, session: Session, graph_runtime_state: GraphRuntimeState | None = None):
         message = self._get_message(session=session)
+        if message is None:
+            return
+
+        if message.status == MessageStatus.PAUSED:
+            message.status = MessageStatus.NORMAL
 
         # If there are assistant files, remove markdown image links from answer
         answer_text = self._task_state.answer

api/core/app/apps/base_app_generate_response_converter.py

@@ -79,6 +79,7 @@ class AppGenerateResponseConverter(ABC):
                         "document_name": resource["document_name"],
                         "score": resource["score"],
                         "content": resource["content"],
+                        "summary": resource.get("summary"),
                     }
                 )
             metadata["retriever_resources"] = updated_resources

api/core/app/apps/common/workflow_response_converter.py

@@ -5,9 +5,14 @@ from dataclasses import dataclass
 from datetime import datetime
 from typing import Any, NewType, Union
 
+from sqlalchemy import select
+from sqlalchemy.orm import Session
+
 from core.app.entities.app_invoke_entities import AdvancedChatAppGenerateEntity, InvokeFrom, WorkflowAppGenerateEntity
 from core.app.entities.queue_entities import (
     QueueAgentLogEvent,
+    QueueHumanInputFormFilledEvent,
+    QueueHumanInputFormTimeoutEvent,
     QueueIterationCompletedEvent,
     QueueIterationNextEvent,
     QueueIterationStartEvent,
@@ -19,9 +24,13 @@ from core.app.entities.queue_entities import (
     QueueNodeRetryEvent,
     QueueNodeStartedEvent,
     QueueNodeSucceededEvent,
+    QueueWorkflowPausedEvent,
 )
 from core.app.entities.task_entities import (
     AgentLogStreamResponse,
+    HumanInputFormFilledResponse,
+    HumanInputFormTimeoutResponse,
+    HumanInputRequiredResponse,
     IterationNodeCompletedStreamResponse,
     IterationNodeNextStreamResponse,
     IterationNodeStartStreamResponse,
@@ -31,7 +40,9 @@ from core.app.entities.task_entities import (
     NodeFinishStreamResponse,
     NodeRetryStreamResponse,
     NodeStartStreamResponse,
+    StreamResponse,
     WorkflowFinishStreamResponse,
+    WorkflowPauseStreamResponse,
     WorkflowStartStreamResponse,
 )
 from core.file import FILE_MODEL_IDENTITY, File
@@ -40,6 +51,8 @@ from core.tools.entities.tool_entities import ToolProviderType
 from core.tools.tool_manager import ToolManager
 from core.trigger.trigger_manager import TriggerManager
 from core.variables.segments import ArrayFileSegment, FileSegment, Segment
+from core.workflow.entities.pause_reason import HumanInputRequired
+from core.workflow.entities.workflow_start_reason import WorkflowStartReason
 from core.workflow.enums import (
     NodeType,
     SystemVariableKey,
@@ -51,8 +64,11 @@ from core.workflow.runtime import GraphRuntimeState
 from core.workflow.system_variable import SystemVariable
 from core.workflow.workflow_entry import WorkflowEntry
 from core.workflow.workflow_type_encoder import WorkflowRuntimeTypeConverter
+from extensions.ext_database import db
 from libs.datetime_utils import naive_utc_now
 from models import Account, EndUser
+from models.human_input import HumanInputForm
+from models.workflow import WorkflowRun
 from services.variable_truncator import BaseTruncator, DummyVariableTruncator, VariableTruncator
 
 NodeExecutionId = NewType("NodeExecutionId", str)
@@ -191,6 +207,7 @@ class WorkflowResponseConverter:
         task_id: str,
         workflow_run_id: str,
         workflow_id: str,
+        reason: WorkflowStartReason,
     ) -> WorkflowStartStreamResponse:
         run_id = self._ensure_workflow_run_id(workflow_run_id)
         started_at = naive_utc_now()
@@ -204,6 +221,7 @@ class WorkflowResponseConverter:
                 workflow_id=workflow_id,
                 inputs=self._workflow_inputs,
                 created_at=int(started_at.timestamp()),
+                reason=reason,
             ),
         )
 
@@ -264,6 +282,160 @@ class WorkflowResponseConverter:
             ),
         )
 
+    def workflow_pause_to_stream_response(
+        self,
+        *,
+        event: QueueWorkflowPausedEvent,
+        task_id: str,
+        graph_runtime_state: GraphRuntimeState,
+    ) -> list[StreamResponse]:
+        run_id = self._ensure_workflow_run_id()
+        started_at = self._workflow_started_at
+        if started_at is None:
+            raise ValueError(
+                "workflow_pause_to_stream_response called before workflow_start_to_stream_response",
+            )
+        paused_at = naive_utc_now()
+        elapsed_time = (paused_at - started_at).total_seconds()
+        encoded_outputs = self._encode_outputs(event.outputs) or {}
+        if self._application_generate_entity.invoke_from == InvokeFrom.SERVICE_API:
+            encoded_outputs = {}
+        pause_reasons = [reason.model_dump(mode="json") for reason in event.reasons]
+        human_input_form_ids = [reason.form_id for reason in event.reasons if isinstance(reason, HumanInputRequired)]
+        expiration_times_by_form_id: dict[str, datetime] = {}
+        if human_input_form_ids:
+            stmt = select(HumanInputForm.id, HumanInputForm.expiration_time).where(
+                HumanInputForm.id.in_(human_input_form_ids)
+            )
+            with Session(bind=db.engine) as session:
+                for form_id, expiration_time in session.execute(stmt):
+                    expiration_times_by_form_id[str(form_id)] = expiration_time
+
+        responses: list[StreamResponse] = []
+
+        for reason in event.reasons:
+            if isinstance(reason, HumanInputRequired):
+                expiration_time = expiration_times_by_form_id.get(reason.form_id)
+                if expiration_time is None:
+                    raise ValueError(f"HumanInputForm not found for pause reason, form_id={reason.form_id}")
+                responses.append(
+                    HumanInputRequiredResponse(
+                        task_id=task_id,
+                        workflow_run_id=run_id,
+                        data=HumanInputRequiredResponse.Data(
+                            form_id=reason.form_id,
+                            node_id=reason.node_id,
+                            node_title=reason.node_title,
+                            form_content=reason.form_content,
+                            inputs=reason.inputs,
+                            actions=reason.actions,
+                            display_in_ui=reason.display_in_ui,
+                            form_token=reason.form_token,
+                            resolved_default_values=reason.resolved_default_values,
+                            expiration_time=int(expiration_time.timestamp()),
+                        ),
+                    )
+                )
+
+        responses.append(
+            WorkflowPauseStreamResponse(
+                task_id=task_id,
+                workflow_run_id=run_id,
+                data=WorkflowPauseStreamResponse.Data(
+                    workflow_run_id=run_id,
+                    paused_nodes=list(event.paused_nodes),
+                    outputs=encoded_outputs,
+                    reasons=pause_reasons,
+                    status=WorkflowExecutionStatus.PAUSED.value,
+                    created_at=int(started_at.timestamp()),
+                    elapsed_time=elapsed_time,
+                    total_tokens=graph_runtime_state.total_tokens,
+                    total_steps=graph_runtime_state.node_run_steps,
+                ),
+            )
+        )
+
+        return responses
+
+    def human_input_form_filled_to_stream_response(
+        self, *, event: QueueHumanInputFormFilledEvent, task_id: str
+    ) -> HumanInputFormFilledResponse:
+        run_id = self._ensure_workflow_run_id()
+        return HumanInputFormFilledResponse(
+            task_id=task_id,
+            workflow_run_id=run_id,
+            data=HumanInputFormFilledResponse.Data(
+                node_id=event.node_id,
+                node_title=event.node_title,
+                rendered_content=event.rendered_content,
+                action_id=event.action_id,
+                action_text=event.action_text,
+            ),
+        )
+
+    def human_input_form_timeout_to_stream_response(
+        self, *, event: QueueHumanInputFormTimeoutEvent, task_id: str
+    ) -> HumanInputFormTimeoutResponse:
+        run_id = self._ensure_workflow_run_id()
+        return HumanInputFormTimeoutResponse(
+            task_id=task_id,
+            workflow_run_id=run_id,
+            data=HumanInputFormTimeoutResponse.Data(
+                node_id=event.node_id,
+                node_title=event.node_title,
+                expiration_time=int(event.expiration_time.timestamp()),
+            ),
+        )
+
+    @classmethod
+    def workflow_run_result_to_finish_response(
+        cls,
+        *,
+        task_id: str,
+        workflow_run: WorkflowRun,
+        creator_user: Account | EndUser,
+    ) -> WorkflowFinishStreamResponse:
+        run_id = workflow_run.id
+        elapsed_time = workflow_run.elapsed_time
+
+        encoded_outputs = workflow_run.outputs_dict
+        finished_at = workflow_run.finished_at
+        assert finished_at is not None
+
+        created_by: Mapping[str, object]
+        user = creator_user
+        if isinstance(user, Account):
+            created_by = {
+                "id": user.id,
+                "name": user.name,
+                "email": user.email,
+            }
+        else:
+            created_by = {
+                "id": user.id,
+                "user": user.session_id,
+            }
+
+        return WorkflowFinishStreamResponse(
+            task_id=task_id,
+            workflow_run_id=run_id,
+            data=WorkflowFinishStreamResponse.Data(
+                id=run_id,
+                workflow_id=workflow_run.workflow_id,
+                status=workflow_run.status.value,
+                outputs=encoded_outputs,
+                error=workflow_run.error,
+                elapsed_time=elapsed_time,
+                total_tokens=workflow_run.total_tokens,
+                total_steps=workflow_run.total_steps,
+                created_by=created_by,
+                created_at=int(workflow_run.created_at.timestamp()),
+                finished_at=int(finished_at.timestamp()),
+                files=cls.fetch_files_from_node_outputs(encoded_outputs),
+                exceptions_count=workflow_run.exceptions_count,
+            ),
+        )
+
     def workflow_node_start_to_stream_response(
         self,
         *,
@@ -592,7 +764,8 @@ class WorkflowResponseConverter:
             ),
         )
 
-    def fetch_files_from_node_outputs(self, outputs_dict: Mapping[str, Any] | None) -> Sequence[Mapping[str, Any]]:
+    @classmethod
+    def fetch_files_from_node_outputs(cls, outputs_dict: Mapping[str, Any] | None) -> Sequence[Mapping[str, Any]]:
         """
         Fetch files from node outputs
         :param outputs_dict: node outputs dict
@@ -601,7 +774,7 @@ class WorkflowResponseConverter:
         if not outputs_dict:
             return []
 
-        files = [self._fetch_files_from_variable_value(output_value) for output_value in outputs_dict.values()]
+        files = [cls._fetch_files_from_variable_value(output_value) for output_value in outputs_dict.values()]
         # Remove None
         files = [file for file in files if file]
         # Flatten list

api/core/app/apps/message_based_app_generator.py

@@ -1,6 +1,6 @@
 import json
 import logging
-from collections.abc import Generator
+from collections.abc import Callable, Generator, Mapping
 from typing import Union, cast
 
 from sqlalchemy import select
@@ -10,12 +10,14 @@ from core.app.app_config.entities import EasyUIBasedAppConfig, EasyUIBasedAppMod
 from core.app.apps.base_app_generator import BaseAppGenerator
 from core.app.apps.base_app_queue_manager import AppQueueManager
 from core.app.apps.exc import GenerateTaskStoppedError
+from core.app.apps.streaming_utils import stream_topic_events
 from core.app.entities.app_invoke_entities import (
     AdvancedChatAppGenerateEntity,
     AgentChatAppGenerateEntity,
     AppGenerateEntity,
     ChatAppGenerateEntity,
     CompletionAppGenerateEntity,
+    ConversationAppGenerateEntity,
     InvokeFrom,
 )
 from core.app.entities.task_entities import (
@@ -27,6 +29,8 @@ from core.app.entities.task_entities import (
 from core.app.task_pipeline.easy_ui_based_generate_task_pipeline import EasyUIBasedGenerateTaskPipeline
 from core.prompt.utils.prompt_template_parser import PromptTemplateParser
 from extensions.ext_database import db
+from extensions.ext_redis import get_pubsub_broadcast_channel
+from libs.broadcast_channel.channel import Topic
 from libs.datetime_utils import naive_utc_now
 from models import Account
 from models.enums import CreatorUserRole
@@ -156,6 +160,7 @@ class MessageBasedAppGenerator(BaseAppGenerator):
         query = application_generate_entity.query or "New conversation"
         conversation_name = (query[:20] + "…") if len(query) > 20 else query
 
+        created_new_conversation = conversation is None
         try:
             if not conversation:
                 conversation = Conversation(
@@ -232,6 +237,10 @@ class MessageBasedAppGenerator(BaseAppGenerator):
                 db.session.add_all(message_files)
 
             db.session.commit()
+
+            if isinstance(application_generate_entity, ConversationAppGenerateEntity):
+                application_generate_entity.conversation_id = conversation.id
+                application_generate_entity.is_new_conversation = created_new_conversation
             return conversation, message
         except Exception:
             db.session.rollback()
@@ -284,3 +293,29 @@ class MessageBasedAppGenerator(BaseAppGenerator):
             raise MessageNotExistsError("Message not exists")
 
         return message
+
+    @staticmethod
+    def _make_channel_key(app_mode: AppMode, workflow_run_id: str):
+        return f"channel:{app_mode}:{workflow_run_id}"
+
+    @classmethod
+    def get_response_topic(cls, app_mode: AppMode, workflow_run_id: str) -> Topic:
+        key = cls._make_channel_key(app_mode, workflow_run_id)
+        channel = get_pubsub_broadcast_channel()
+        topic = channel.topic(key)
+        return topic
+
+    @classmethod
+    def retrieve_events(
+        cls,
+        app_mode: AppMode,
+        workflow_run_id: str,
+        idle_timeout=300,
+        on_subscribe: Callable[[], None] | None = None,
+    ) -> Generator[Mapping | str, None, None]:
+        topic = cls.get_response_topic(app_mode, workflow_run_id)
+        return stream_topic_events(
+            topic=topic,
+            idle_timeout=idle_timeout,
+            on_subscribe=on_subscribe,
+        )

api/core/app/apps/message_generator.py

@@ -0,0 +1,36 @@
+from collections.abc import Callable, Generator, Mapping
+
+from core.app.apps.streaming_utils import stream_topic_events
+from extensions.ext_redis import get_pubsub_broadcast_channel
+from libs.broadcast_channel.channel import Topic
+from models.model import AppMode
+
+
+class MessageGenerator:
+    @staticmethod
+    def _make_channel_key(app_mode: AppMode, workflow_run_id: str):
+        return f"channel:{app_mode}:{str(workflow_run_id)}"
+
+    @classmethod
+    def get_response_topic(cls, app_mode: AppMode, workflow_run_id: str) -> Topic:
+        key = cls._make_channel_key(app_mode, workflow_run_id)
+        channel = get_pubsub_broadcast_channel()
+        topic = channel.topic(key)
+        return topic
+
+    @classmethod
+    def retrieve_events(
+        cls,
+        app_mode: AppMode,
+        workflow_run_id: str,
+        idle_timeout=300,
+        ping_interval: float = 10.0,
+        on_subscribe: Callable[[], None] | None = None,
+    ) -> Generator[Mapping | str, None, None]:
+        topic = cls.get_response_topic(app_mode, workflow_run_id)
+        return stream_topic_events(
+            topic=topic,
+            idle_timeout=idle_timeout,
+            ping_interval=ping_interval,
+            on_subscribe=on_subscribe,
+        )

api/core/app/apps/streaming_utils.py

@@ -0,0 +1,70 @@
+from __future__ import annotations
+
+import json
+import time
+from collections.abc import Callable, Generator, Iterable, Mapping
+from typing import Any
+
+from core.app.entities.task_entities import StreamEvent
+from libs.broadcast_channel.channel import Topic
+from libs.broadcast_channel.exc import SubscriptionClosedError
+
+
+def stream_topic_events(
+    *,
+    topic: Topic,
+    idle_timeout: float,
+    ping_interval: float | None = None,
+    on_subscribe: Callable[[], None] | None = None,
+    terminal_events: Iterable[str | StreamEvent] | None = None,
+) -> Generator[Mapping[str, Any] | str, None, None]:
+    # send a PING event immediately to prevent the connection staying in pending state for a long time.
+    #
+    # This simplify the debugging process as the DevTools in Chrome does not
+    # provide complete curl command for pending connections.
+    yield StreamEvent.PING.value
+
+    terminal_values = _normalize_terminal_events(terminal_events)
+    last_msg_time = time.time()
+    last_ping_time = last_msg_time
+    with topic.subscribe() as sub:
+        # on_subscribe fires only after the Redis subscription is active.
+        # This is used to gate task start and reduce pub/sub race for the first event.
+        if on_subscribe is not None:
+            on_subscribe()
+        while True:
+            try:
+                msg = sub.receive(timeout=0.1)
+            except SubscriptionClosedError:
+                return
+            if msg is None:
+                current_time = time.time()
+                if current_time - last_msg_time > idle_timeout:
+                    return
+                if ping_interval is not None and current_time - last_ping_time >= ping_interval:
+                    yield StreamEvent.PING.value
+                    last_ping_time = current_time
+                continue
+
+            last_msg_time = time.time()
+            last_ping_time = last_msg_time
+            event = json.loads(msg)
+            yield event
+            if not isinstance(event, dict):
+                continue
+
+            event_type = event.get("event")
+            if event_type in terminal_values:
+                return
+
+
+def _normalize_terminal_events(terminal_events: Iterable[str | StreamEvent] | None) -> set[str]:
+    if not terminal_events:
+        return {StreamEvent.WORKFLOW_FINISHED.value, StreamEvent.WORKFLOW_PAUSED.value}
+    values: set[str] = set()
+    for item in terminal_events:
+        if isinstance(item, StreamEvent):
+            values.add(item.value)
+        else:
+            values.add(str(item))
+    return values

api/core/app/apps/workflow/app_generator.py

@@ -25,6 +25,7 @@ from core.app.apps.workflow.generate_response_converter import WorkflowAppGenera
 from core.app.apps.workflow.generate_task_pipeline import WorkflowAppGenerateTaskPipeline
 from core.app.entities.app_invoke_entities import InvokeFrom, WorkflowAppGenerateEntity
 from core.app.entities.task_entities import WorkflowAppBlockingResponse, WorkflowAppStreamResponse
+from core.app.layers.pause_state_persist_layer import PauseStateLayerConfig, PauseStatePersistenceLayer
 from core.db.session_factory import session_factory
 from core.helper.trace_id_helper import extract_external_trace_id_from_args
 from core.model_runtime.errors.invoke import InvokeAuthorizationError
@@ -34,12 +35,15 @@ from core.workflow.graph_engine.layers.base import GraphEngineLayer
 from core.workflow.repositories.draft_variable_repository import DraftVariableSaverFactory
 from core.workflow.repositories.workflow_execution_repository import WorkflowExecutionRepository
 from core.workflow.repositories.workflow_node_execution_repository import WorkflowNodeExecutionRepository
+from core.workflow.runtime import GraphRuntimeState
 from core.workflow.variable_loader import DUMMY_VARIABLE_LOADER, VariableLoader
 from extensions.ext_database import db
 from factories import file_factory
 from libs.flask_utils import preserve_flask_contexts
-from models import Account, App, EndUser, Workflow, WorkflowNodeExecutionTriggeredFrom
+from models.account import Account
 from models.enums import WorkflowRunTriggeredFrom
+from models.model import App, EndUser
+from models.workflow import Workflow, WorkflowNodeExecutionTriggeredFrom
 from services.workflow_draft_variable_service import DraftVarLoader, WorkflowDraftVariableService
 
 if TYPE_CHECKING:
@@ -66,9 +70,11 @@ class WorkflowAppGenerator(BaseAppGenerator):
         invoke_from: InvokeFrom,
         streaming: Literal[True],
         call_depth: int,
+        workflow_run_id: str | uuid.UUID | None = None,
         triggered_from: WorkflowRunTriggeredFrom | None = None,
         root_node_id: str | None = None,
         graph_engine_layers: Sequence[GraphEngineLayer] = (),
+        pause_state_config: PauseStateLayerConfig | None = None,
     ) -> Generator[Mapping[str, Any] | str, None, None]: ...
 
     @overload
@@ -82,9 +88,11 @@ class WorkflowAppGenerator(BaseAppGenerator):
         invoke_from: InvokeFrom,
         streaming: Literal[False],
         call_depth: int,
+        workflow_run_id: str | uuid.UUID | None = None,
         triggered_from: WorkflowRunTriggeredFrom | None = None,
         root_node_id: str | None = None,
         graph_engine_layers: Sequence[GraphEngineLayer] = (),
+        pause_state_config: PauseStateLayerConfig | None = None,
     ) -> Mapping[str, Any]: ...
 
     @overload
@@ -98,9 +106,11 @@ class WorkflowAppGenerator(BaseAppGenerator):
         invoke_from: InvokeFrom,
         streaming: bool,
         call_depth: int,
+        workflow_run_id: str | uuid.UUID | None = None,
         triggered_from: WorkflowRunTriggeredFrom | None = None,
         root_node_id: str | None = None,
         graph_engine_layers: Sequence[GraphEngineLayer] = (),
+        pause_state_config: PauseStateLayerConfig | None = None,
     ) -> Union[Mapping[str, Any], Generator[Mapping[str, Any] | str, None, None]]: ...
 
     def generate(
@@ -113,9 +123,11 @@ class WorkflowAppGenerator(BaseAppGenerator):
         invoke_from: InvokeFrom,
         streaming: bool = True,
         call_depth: int = 0,
+        workflow_run_id: str | uuid.UUID | None = None,
         triggered_from: WorkflowRunTriggeredFrom | None = None,
         root_node_id: str | None = None,
         graph_engine_layers: Sequence[GraphEngineLayer] = (),
+        pause_state_config: PauseStateLayerConfig | None = None,
     ) -> Union[Mapping[str, Any], Generator[Mapping[str, Any] | str, None, None]]:
         files: Sequence[Mapping[str, Any]] = args.get("files") or []
 
@@ -150,7 +162,7 @@ class WorkflowAppGenerator(BaseAppGenerator):
         extras = {
             **extract_external_trace_id_from_args(args),
         }
-        workflow_run_id = str(uuid.uuid4())
+        workflow_run_id = str(workflow_run_id or uuid.uuid4())
         # FIXME (Yeuoly): we need to remove the SKIP_PREPARE_USER_INPUTS_KEY from the args
         # trigger shouldn't prepare user inputs
         if self._should_prepare_user_inputs(args):
@@ -216,13 +228,40 @@ class WorkflowAppGenerator(BaseAppGenerator):
             streaming=streaming,
             root_node_id=root_node_id,
             graph_engine_layers=graph_engine_layers,
+            pause_state_config=pause_state_config,
         )
 
-    def resume(self, *, workflow_run_id: str) -> None:
+    def resume(
+        self,
+        *,
+        app_model: App,
+        workflow: Workflow,
+        user: Union[Account, EndUser],
+        application_generate_entity: WorkflowAppGenerateEntity,
+        graph_runtime_state: GraphRuntimeState,
+        workflow_execution_repository: WorkflowExecutionRepository,
+        workflow_node_execution_repository: WorkflowNodeExecutionRepository,
+        graph_engine_layers: Sequence[GraphEngineLayer] = (),
+        pause_state_config: PauseStateLayerConfig | None = None,
+        variable_loader: VariableLoader = DUMMY_VARIABLE_LOADER,
+    ) -> Union[Mapping[str, Any], Generator[str | Mapping[str, Any], None, None]]:
         """
-        @TBD
+        Resume a paused workflow execution using the persisted runtime state.
         """
-        pass
+        return self._generate(
+            app_model=app_model,
+            workflow=workflow,
+            user=user,
+            application_generate_entity=application_generate_entity,
+            invoke_from=application_generate_entity.invoke_from,
+            workflow_execution_repository=workflow_execution_repository,
+            workflow_node_execution_repository=workflow_node_execution_repository,
+            streaming=application_generate_entity.stream,
+            variable_loader=variable_loader,
+            graph_engine_layers=graph_engine_layers,
+            graph_runtime_state=graph_runtime_state,
+            pause_state_config=pause_state_config,
+        )
 
     def _generate(
         self,
@@ -238,6 +277,8 @@ class WorkflowAppGenerator(BaseAppGenerator):
         variable_loader: VariableLoader = DUMMY_VARIABLE_LOADER,
         root_node_id: str | None = None,
         graph_engine_layers: Sequence[GraphEngineLayer] = (),
+        graph_runtime_state: GraphRuntimeState | None = None,
+        pause_state_config: PauseStateLayerConfig | None = None,
     ) -> Union[Mapping[str, Any], Generator[str | Mapping[str, Any], None, None]]:
         """
         Generate App response.
@@ -251,6 +292,8 @@ class WorkflowAppGenerator(BaseAppGenerator):
         :param workflow_node_execution_repository: repository for workflow node execution
         :param streaming: is stream
         """
+        graph_layers: list[GraphEngineLayer] = list(graph_engine_layers)
+
         # init queue manager
         queue_manager = WorkflowAppQueueManager(
             task_id=application_generate_entity.task_id,
@@ -259,6 +302,15 @@ class WorkflowAppGenerator(BaseAppGenerator):
             app_mode=app_model.mode,
         )
 
+        if pause_state_config is not None:
+            graph_layers.append(
+                PauseStatePersistenceLayer(
+                    session_factory=pause_state_config.session_factory,
+                    generate_entity=application_generate_entity,
+                    state_owner_user_id=pause_state_config.state_owner_user_id,
+                )
+            )
+
         # new thread with request context and contextvars
         context = contextvars.copy_context()
 
@@ -276,7 +328,8 @@ class WorkflowAppGenerator(BaseAppGenerator):
                 "root_node_id": root_node_id,
                 "workflow_execution_repository": workflow_execution_repository,
                 "workflow_node_execution_repository": workflow_node_execution_repository,
-                "graph_engine_layers": graph_engine_layers,
+                "graph_engine_layers": tuple(graph_layers),
+                "graph_runtime_state": graph_runtime_state,
             },
         )
 
@@ -378,6 +431,7 @@ class WorkflowAppGenerator(BaseAppGenerator):
             workflow_node_execution_repository=workflow_node_execution_repository,
             streaming=streaming,
             variable_loader=var_loader,
+            pause_state_config=None,
         )
 
     def single_loop_generate(
@@ -459,6 +513,7 @@ class WorkflowAppGenerator(BaseAppGenerator):
             workflow_node_execution_repository=workflow_node_execution_repository,
             streaming=streaming,
             variable_loader=var_loader,
+            pause_state_config=None,
         )
 
     def _generate_worker(
@@ -472,6 +527,7 @@ class WorkflowAppGenerator(BaseAppGenerator):
         workflow_node_execution_repository: WorkflowNodeExecutionRepository,
         root_node_id: str | None = None,
         graph_engine_layers: Sequence[GraphEngineLayer] = (),
+        graph_runtime_state: GraphRuntimeState | None = None,
     ) -> None:
         """
         Generate worker in a new thread.
@@ -517,6 +573,7 @@ class WorkflowAppGenerator(BaseAppGenerator):
                 workflow_node_execution_repository=workflow_node_execution_repository,
                 root_node_id=root_node_id,
                 graph_engine_layers=graph_engine_layers,
+                graph_runtime_state=graph_runtime_state,
             )
 
             try:

api/core/app/apps/workflow/app_runner.py

@@ -42,6 +42,7 @@ class WorkflowAppRunner(WorkflowBasedAppRunner):
         workflow_execution_repository: WorkflowExecutionRepository,
         workflow_node_execution_repository: WorkflowNodeExecutionRepository,
         graph_engine_layers: Sequence[GraphEngineLayer] = (),
+        graph_runtime_state: GraphRuntimeState | None = None,
     ):
         super().__init__(
             queue_manager=queue_manager,
@@ -55,6 +56,7 @@ class WorkflowAppRunner(WorkflowBasedAppRunner):
         self._root_node_id = root_node_id
         self._workflow_execution_repository = workflow_execution_repository
         self._workflow_node_execution_repository = workflow_node_execution_repository
+        self._resume_graph_runtime_state = graph_runtime_state
 
     @trace_span(WorkflowAppRunnerHandler)
     def run(self):
@@ -63,23 +65,28 @@ class WorkflowAppRunner(WorkflowBasedAppRunner):
         """
         app_config = self.application_generate_entity.app_config
         app_config = cast(WorkflowAppConfig, app_config)
-
-        system_inputs = SystemVariable(
-            files=self.application_generate_entity.files,
-            user_id=self._sys_user_id,
-            app_id=app_config.app_id,
-            timestamp=int(naive_utc_now().timestamp()),
-            workflow_id=app_config.workflow_id,
-            workflow_execution_id=self.application_generate_entity.workflow_execution_id,
-        )
-
         invoke_from = self.application_generate_entity.invoke_from
         # if only single iteration or single loop run is requested
         if self.application_generate_entity.single_iteration_run or self.application_generate_entity.single_loop_run:
             invoke_from = InvokeFrom.DEBUGGER
         user_from = self._resolve_user_from(invoke_from)
 
-        if self.application_generate_entity.single_iteration_run or self.application_generate_entity.single_loop_run:
+        resume_state = self._resume_graph_runtime_state
+
+        if resume_state is not None:
+            graph_runtime_state = resume_state
+            variable_pool = graph_runtime_state.variable_pool
+            graph = self._init_graph(
+                graph_config=self._workflow.graph_dict,
+                graph_runtime_state=graph_runtime_state,
+                workflow_id=self._workflow.id,
+                tenant_id=self._workflow.tenant_id,
+                user_id=self.application_generate_entity.user_id,
+                user_from=user_from,
+                invoke_from=invoke_from,
+                root_node_id=self._root_node_id,
+            )
+        elif self.application_generate_entity.single_iteration_run or self.application_generate_entity.single_loop_run:
             graph, variable_pool, graph_runtime_state = self._prepare_single_node_execution(
                 workflow=self._workflow,
                 single_iteration_run=self.application_generate_entity.single_iteration_run,
@@ -89,7 +96,14 @@ class WorkflowAppRunner(WorkflowBasedAppRunner):
             inputs = self.application_generate_entity.inputs
 
             # Create a variable pool.
-
+            system_inputs = SystemVariable(
+                files=self.application_generate_entity.files,
+                user_id=self._sys_user_id,
+                app_id=app_config.app_id,
+                timestamp=int(naive_utc_now().timestamp()),
+                workflow_id=app_config.workflow_id,
+                workflow_execution_id=self.application_generate_entity.workflow_execution_id,
+            )
             variable_pool = VariablePool(
                 system_variables=system_inputs,
                 user_inputs=inputs,
@@ -98,8 +112,6 @@ class WorkflowAppRunner(WorkflowBasedAppRunner):
             )
 
             graph_runtime_state = GraphRuntimeState(variable_pool=variable_pool, start_at=time.perf_counter())
-
-            # init graph
             graph = self._init_graph(
                 graph_config=self._workflow.graph_dict,
                 graph_runtime_state=graph_runtime_state,

api/core/app/apps/workflow/errors.py

@@ -0,0 +1,7 @@
+from libs.exception import BaseHTTPException
+
+
+class WorkflowPausedInBlockingModeError(BaseHTTPException):
+    error_code = "workflow_paused_in_blocking_mode"
+    description = "Workflow execution paused for human input; blocking response mode is not supported."
+    code = 400

api/core/app/apps/workflow/generate_task_pipeline.py

@@ -16,6 +16,8 @@ from core.app.entities.queue_entities import (
     MessageQueueMessage,
     QueueAgentLogEvent,
     QueueErrorEvent,
+    QueueHumanInputFormFilledEvent,
+    QueueHumanInputFormTimeoutEvent,
     QueueIterationCompletedEvent,
     QueueIterationNextEvent,
     QueueIterationStartEvent,
@@ -32,6 +34,7 @@ from core.app.entities.queue_entities import (
     QueueTextChunkEvent,
     QueueWorkflowFailedEvent,
     QueueWorkflowPartialSuccessEvent,
+    QueueWorkflowPausedEvent,
     QueueWorkflowStartedEvent,
     QueueWorkflowSucceededEvent,
     WorkflowQueueMessage,
@@ -46,11 +49,13 @@ from core.app.entities.task_entities import (
     WorkflowAppBlockingResponse,
     WorkflowAppStreamResponse,
     WorkflowFinishStreamResponse,
+    WorkflowPauseStreamResponse,
     WorkflowStartStreamResponse,
 )
 from core.app.task_pipeline.based_generate_task_pipeline import BasedGenerateTaskPipeline
 from core.base.tts import AppGeneratorTTSPublisher, AudioTrunk
 from core.ops.ops_trace_manager import TraceQueueManager
+from core.workflow.entities.workflow_start_reason import WorkflowStartReason
 from core.workflow.enums import WorkflowExecutionStatus
 from core.workflow.repositories.draft_variable_repository import DraftVariableSaverFactory
 from core.workflow.runtime import GraphRuntimeState
@@ -132,6 +137,25 @@ class WorkflowAppGenerateTaskPipeline(GraphRuntimeStateSupport):
         for stream_response in generator:
             if isinstance(stream_response, ErrorStreamResponse):
                 raise stream_response.err
+            elif isinstance(stream_response, WorkflowPauseStreamResponse):
+                response = WorkflowAppBlockingResponse(
+                    task_id=self._application_generate_entity.task_id,
+                    workflow_run_id=stream_response.data.workflow_run_id,
+                    data=WorkflowAppBlockingResponse.Data(
+                        id=stream_response.data.workflow_run_id,
+                        workflow_id=self._workflow.id,
+                        status=stream_response.data.status,
+                        outputs=stream_response.data.outputs or {},
+                        error=None,
+                        elapsed_time=stream_response.data.elapsed_time,
+                        total_tokens=stream_response.data.total_tokens,
+                        total_steps=stream_response.data.total_steps,
+                        created_at=stream_response.data.created_at,
+                        finished_at=None,
+                    ),
+                )
+
+                return response
             elif isinstance(stream_response, WorkflowFinishStreamResponse):
                 response = WorkflowAppBlockingResponse(
                     task_id=self._application_generate_entity.task_id,
@@ -146,7 +170,7 @@ class WorkflowAppGenerateTaskPipeline(GraphRuntimeStateSupport):
                         total_tokens=stream_response.data.total_tokens,
                         total_steps=stream_response.data.total_steps,
                         created_at=int(stream_response.data.created_at),
-                        finished_at=int(stream_response.data.finished_at),
+                        finished_at=int(stream_response.data.finished_at) if stream_response.data.finished_at else None,
                     ),
                 )
 
@@ -259,13 +283,15 @@ class WorkflowAppGenerateTaskPipeline(GraphRuntimeStateSupport):
         run_id = self._extract_workflow_run_id(runtime_state)
         self._workflow_execution_id = run_id
 
-        with self._database_session() as session:
-            self._save_workflow_app_log(session=session, workflow_run_id=self._workflow_execution_id)
+        if event.reason == WorkflowStartReason.INITIAL:
+            with self._database_session() as session:
+                self._save_workflow_app_log(session=session, workflow_run_id=self._workflow_execution_id)
 
         start_resp = self._workflow_response_converter.workflow_start_to_stream_response(
             task_id=self._application_generate_entity.task_id,
             workflow_run_id=run_id,
             workflow_id=self._workflow.id,
+            reason=event.reason,
         )
         yield start_resp
 
@@ -440,6 +466,21 @@ class WorkflowAppGenerateTaskPipeline(GraphRuntimeStateSupport):
         )
         yield workflow_finish_resp
 
+    def _handle_workflow_paused_event(
+        self,
+        event: QueueWorkflowPausedEvent,
+        **kwargs,
+    ) -> Generator[StreamResponse, None, None]:
+        """Handle workflow paused events."""
+        self._ensure_workflow_initialized()
+        validated_state = self._ensure_graph_runtime_initialized()
+        responses = self._workflow_response_converter.workflow_pause_to_stream_response(
+            event=event,
+            task_id=self._application_generate_entity.task_id,
+            graph_runtime_state=validated_state,
+        )
+        yield from responses
+
     def _handle_workflow_failed_and_stop_events(
         self,
         event: Union[QueueWorkflowFailedEvent, QueueStopEvent],
@@ -495,6 +536,22 @@ class WorkflowAppGenerateTaskPipeline(GraphRuntimeStateSupport):
             task_id=self._application_generate_entity.task_id, event=event
         )
 
+    def _handle_human_input_form_filled_event(
+        self, event: QueueHumanInputFormFilledEvent, **kwargs
+    ) -> Generator[StreamResponse, None, None]:
+        """Handle human input form filled events."""
+        yield self._workflow_response_converter.human_input_form_filled_to_stream_response(
+            event=event, task_id=self._application_generate_entity.task_id
+        )
+
+    def _handle_human_input_form_timeout_event(
+        self, event: QueueHumanInputFormTimeoutEvent, **kwargs
+    ) -> Generator[StreamResponse, None, None]:
+        """Handle human input form timeout events."""
+        yield self._workflow_response_converter.human_input_form_timeout_to_stream_response(
+            event=event, task_id=self._application_generate_entity.task_id
+        )
+
     def _get_event_handlers(self) -> dict[type, Callable]:
         """Get mapping of event types to their handlers using fluent pattern."""
         return {
@@ -506,6 +563,7 @@ class WorkflowAppGenerateTaskPipeline(GraphRuntimeStateSupport):
             QueueWorkflowStartedEvent: self._handle_workflow_started_event,
             QueueWorkflowSucceededEvent: self._handle_workflow_succeeded_event,
             QueueWorkflowPartialSuccessEvent: self._handle_workflow_partial_success_event,
+            QueueWorkflowPausedEvent: self._handle_workflow_paused_event,
             # Node events
             QueueNodeRetryEvent: self._handle_node_retry_event,
             QueueNodeStartedEvent: self._handle_node_started_event,
@@ -520,6 +578,8 @@ class WorkflowAppGenerateTaskPipeline(GraphRuntimeStateSupport):
             QueueLoopCompletedEvent: self._handle_loop_completed_event,
             # Agent events
             QueueAgentLogEvent: self._handle_agent_log_event,
+            QueueHumanInputFormFilledEvent: self._handle_human_input_form_filled_event,
+            QueueHumanInputFormTimeoutEvent: self._handle_human_input_form_timeout_event,
         }
 
     def _dispatch_event(
@@ -602,6 +662,9 @@ class WorkflowAppGenerateTaskPipeline(GraphRuntimeStateSupport):
                 case QueueWorkflowFailedEvent():
                     yield from self._handle_workflow_failed_and_stop_events(event)
                     break
+                case QueueWorkflowPausedEvent():
+                    yield from self._handle_workflow_paused_event(event)
+                    break
 
                 case QueueStopEvent():
                     yield from self._handle_workflow_failed_and_stop_events(event)

api/core/app/apps/workflow_app_runner.py

@@ -1,3 +1,4 @@
+import logging
 import time
 from collections.abc import Mapping, Sequence
 from typing import Any, cast
@@ -7,6 +8,8 @@ from core.app.entities.app_invoke_entities import InvokeFrom
 from core.app.entities.queue_entities import (
     AppQueueEvent,
     QueueAgentLogEvent,
+    QueueHumanInputFormFilledEvent,
+    QueueHumanInputFormTimeoutEvent,
     QueueIterationCompletedEvent,
     QueueIterationNextEvent,
     QueueIterationStartEvent,
@@ -22,22 +25,27 @@ from core.app.entities.queue_entities import (
     QueueTextChunkEvent,
     QueueWorkflowFailedEvent,
     QueueWorkflowPartialSuccessEvent,
+    QueueWorkflowPausedEvent,
     QueueWorkflowStartedEvent,
     QueueWorkflowSucceededEvent,
 )
 from core.app.workflow.node_factory import DifyNodeFactory
 from core.workflow.entities import GraphInitParams
+from core.workflow.entities.pause_reason import HumanInputRequired
 from core.workflow.graph import Graph
 from core.workflow.graph_engine.layers.base import GraphEngineLayer
 from core.workflow.graph_events import (
     GraphEngineEvent,
     GraphRunFailedEvent,
     GraphRunPartialSucceededEvent,
+    GraphRunPausedEvent,
     GraphRunStartedEvent,
     GraphRunSucceededEvent,
     NodeRunAgentLogEvent,
     NodeRunExceptionEvent,
     NodeRunFailedEvent,
+    NodeRunHumanInputFormFilledEvent,
+    NodeRunHumanInputFormTimeoutEvent,
     NodeRunIterationFailedEvent,
     NodeRunIterationNextEvent,
     NodeRunIterationStartedEvent,
@@ -61,6 +69,9 @@ from core.workflow.variable_loader import DUMMY_VARIABLE_LOADER, VariableLoader,
 from core.workflow.workflow_entry import WorkflowEntry
 from models.enums import UserFrom
 from models.workflow import Workflow
+from tasks.mail_human_input_delivery_task import dispatch_human_input_email_task
+
+logger = logging.getLogger(__name__)
 
 
 class WorkflowBasedAppRunner:
@@ -327,7 +338,7 @@ class WorkflowBasedAppRunner:
         :param event: event
         """
         if isinstance(event, GraphRunStartedEvent):
-            self._publish_event(QueueWorkflowStartedEvent())
+            self._publish_event(QueueWorkflowStartedEvent(reason=event.reason))
         elif isinstance(event, GraphRunSucceededEvent):
             self._publish_event(QueueWorkflowSucceededEvent(outputs=event.outputs))
         elif isinstance(event, GraphRunPartialSucceededEvent):
@@ -338,6 +349,38 @@ class WorkflowBasedAppRunner:
             self._publish_event(QueueWorkflowFailedEvent(error=event.error, exceptions_count=event.exceptions_count))
         elif isinstance(event, GraphRunAbortedEvent):
             self._publish_event(QueueWorkflowFailedEvent(error=event.reason or "Unknown error", exceptions_count=0))
+        elif isinstance(event, GraphRunPausedEvent):
+            runtime_state = workflow_entry.graph_engine.graph_runtime_state
+            paused_nodes = runtime_state.get_paused_nodes()
+            self._enqueue_human_input_notifications(event.reasons)
+            self._publish_event(
+                QueueWorkflowPausedEvent(
+                    reasons=event.reasons,
+                    outputs=event.outputs,
+                    paused_nodes=paused_nodes,
+                )
+            )
+        elif isinstance(event, NodeRunHumanInputFormFilledEvent):
+            self._publish_event(
+                QueueHumanInputFormFilledEvent(
+                    node_execution_id=event.id,
+                    node_id=event.node_id,
+                    node_type=event.node_type,
+                    node_title=event.node_title,
+                    rendered_content=event.rendered_content,
+                    action_id=event.action_id,
+                    action_text=event.action_text,
+                )
+            )
+        elif isinstance(event, NodeRunHumanInputFormTimeoutEvent):
+            self._publish_event(
+                QueueHumanInputFormTimeoutEvent(
+                    node_id=event.node_id,
+                    node_type=event.node_type,
+                    node_title=event.node_title,
+                    expiration_time=event.expiration_time,
+                )
+            )
         elif isinstance(event, NodeRunRetryEvent):
             node_run_result = event.node_run_result
             inputs = node_run_result.inputs
@@ -544,5 +587,19 @@ class WorkflowBasedAppRunner:
                 )
             )
 
+    def _enqueue_human_input_notifications(self, reasons: Sequence[object]) -> None:
+        for reason in reasons:
+            if not isinstance(reason, HumanInputRequired):
+                continue
+            if not reason.form_id:
+                continue
+            try:
+                dispatch_human_input_email_task.apply_async(
+                    kwargs={"form_id": reason.form_id, "node_title": reason.node_title},
+                    queue="mail",
+                )
+            except Exception:  # pragma: no cover - defensive logging
+                logger.exception("Failed to enqueue human input email task for form %s", reason.form_id)
+
     def _publish_event(self, event: AppQueueEvent):
         self._queue_manager.publish(event, PublishFrom.APPLICATION_MANAGER)

api/core/app/entities/app_invoke_entities.py

@@ -132,7 +132,7 @@ class AppGenerateEntity(BaseModel):
     extras: dict[str, Any] = Field(default_factory=dict)
 
     # tracing instance
-    trace_manager: Optional["TraceQueueManager"] = None
+    trace_manager: Optional["TraceQueueManager"] = Field(default=None, exclude=True, repr=False)
 
 
 class EasyUIBasedAppGenerateEntity(AppGenerateEntity):
@@ -156,6 +156,7 @@ class ConversationAppGenerateEntity(AppGenerateEntity):
     """
 
     conversation_id: str | None = None
+    is_new_conversation: bool = False
     parent_message_id: str | None = Field(
         default=None,
         description=(

api/core/app/entities/queue_entities.py

@@ -8,6 +8,8 @@ from pydantic import BaseModel, ConfigDict, Field
 from core.model_runtime.entities.llm_entities import LLMResult, LLMResultChunk
 from core.rag.entities.citation_metadata import RetrievalSourceMetadata
 from core.workflow.entities import AgentNodeStrategyInit
+from core.workflow.entities.pause_reason import PauseReason
+from core.workflow.entities.workflow_start_reason import WorkflowStartReason
 from core.workflow.enums import WorkflowNodeExecutionMetadataKey
 from core.workflow.nodes import NodeType
 
@@ -46,6 +48,9 @@ class QueueEvent(StrEnum):
     PING = "ping"
     STOP = "stop"
     RETRY = "retry"
+    PAUSE = "pause"
+    HUMAN_INPUT_FORM_FILLED = "human_input_form_filled"
+    HUMAN_INPUT_FORM_TIMEOUT = "human_input_form_timeout"
 
 
 class AppQueueEvent(BaseModel):
@@ -261,6 +266,8 @@ class QueueWorkflowStartedEvent(AppQueueEvent):
     """QueueWorkflowStartedEvent entity."""
 
     event: QueueEvent = QueueEvent.WORKFLOW_STARTED
+    # Always present; mirrors GraphRunStartedEvent.reason for downstream consumers.
+    reason: WorkflowStartReason = WorkflowStartReason.INITIAL
 
 
 class QueueWorkflowSucceededEvent(AppQueueEvent):
@@ -484,6 +491,35 @@ class QueueStopEvent(AppQueueEvent):
         return reason_mapping.get(self.stopped_by, "Stopped by unknown reason.")
 
 
+class QueueHumanInputFormFilledEvent(AppQueueEvent):
+    """
+    QueueHumanInputFormFilledEvent entity
+    """
+
+    event: QueueEvent = QueueEvent.HUMAN_INPUT_FORM_FILLED
+
+    node_execution_id: str
+    node_id: str
+    node_type: NodeType
+    node_title: str
+    rendered_content: str
+    action_id: str
+    action_text: str
+
+
+class QueueHumanInputFormTimeoutEvent(AppQueueEvent):
+    """
+    QueueHumanInputFormTimeoutEvent entity
+    """
+
+    event: QueueEvent = QueueEvent.HUMAN_INPUT_FORM_TIMEOUT
+
+    node_id: str
+    node_type: NodeType
+    node_title: str
+    expiration_time: datetime
+
+
 class QueueMessage(BaseModel):
     """
     QueueMessage abstract entity
@@ -509,3 +545,14 @@ class WorkflowQueueMessage(QueueMessage):
     """
 
     pass
+
+
+class QueueWorkflowPausedEvent(AppQueueEvent):
+    """
+    QueueWorkflowPausedEvent entity
+    """
+
+    event: QueueEvent = QueueEvent.PAUSE
+    reasons: Sequence[PauseReason] = Field(default_factory=list)
+    outputs: Mapping[str, object] = Field(default_factory=dict)
+    paused_nodes: Sequence[str] = Field(default_factory=list)

api/core/app/entities/task_entities.py

@@ -7,7 +7,9 @@ from pydantic import BaseModel, ConfigDict, Field
 from core.model_runtime.entities.llm_entities import LLMResult, LLMUsage
 from core.rag.entities.citation_metadata import RetrievalSourceMetadata
 from core.workflow.entities import AgentNodeStrategyInit
+from core.workflow.entities.workflow_start_reason import WorkflowStartReason
 from core.workflow.enums import WorkflowNodeExecutionMetadataKey, WorkflowNodeExecutionStatus
+from core.workflow.nodes.human_input.entities import FormInput, UserAction
 
 
 class AnnotationReplyAccount(BaseModel):
@@ -69,6 +71,7 @@ class StreamEvent(StrEnum):
     AGENT_THOUGHT = "agent_thought"
     AGENT_MESSAGE = "agent_message"
     WORKFLOW_STARTED = "workflow_started"
+    WORKFLOW_PAUSED = "workflow_paused"
     WORKFLOW_FINISHED = "workflow_finished"
     NODE_STARTED = "node_started"
     NODE_FINISHED = "node_finished"
@@ -82,6 +85,9 @@ class StreamEvent(StrEnum):
     TEXT_CHUNK = "text_chunk"
     TEXT_REPLACE = "text_replace"
     AGENT_LOG = "agent_log"
+    HUMAN_INPUT_REQUIRED = "human_input_required"
+    HUMAN_INPUT_FORM_FILLED = "human_input_form_filled"
+    HUMAN_INPUT_FORM_TIMEOUT = "human_input_form_timeout"
 
 
 class StreamResponse(BaseModel):
@@ -205,6 +211,8 @@ class WorkflowStartStreamResponse(StreamResponse):
         workflow_id: str
         inputs: Mapping[str, Any]
         created_at: int
+        # Always present; mirrors QueueWorkflowStartedEvent.reason for SSE clients.
+        reason: WorkflowStartReason = WorkflowStartReason.INITIAL
 
     event: StreamEvent = StreamEvent.WORKFLOW_STARTED
     workflow_run_id: str
@@ -231,7 +239,7 @@ class WorkflowFinishStreamResponse(StreamResponse):
         total_steps: int
         created_by: Mapping[str, object] = Field(default_factory=dict)
         created_at: int
-        finished_at: int
+        finished_at: int | None
         exceptions_count: int | None = 0
         files: Sequence[Mapping[str, Any]] | None = []
 
@@ -240,6 +248,85 @@ class WorkflowFinishStreamResponse(StreamResponse):
     data: Data
 
 
+class WorkflowPauseStreamResponse(StreamResponse):
+    """
+    WorkflowPauseStreamResponse entity
+    """
+
+    class Data(BaseModel):
+        """
+        Data entity
+        """
+
+        workflow_run_id: str
+        paused_nodes: Sequence[str] = Field(default_factory=list)
+        outputs: Mapping[str, Any] = Field(default_factory=dict)
+        reasons: Sequence[Mapping[str, Any]] = Field(default_factory=list)
+        status: str
+        created_at: int
+        elapsed_time: float
+        total_tokens: int
+        total_steps: int
+
+    event: StreamEvent = StreamEvent.WORKFLOW_PAUSED
+    workflow_run_id: str
+    data: Data
+
+
+class HumanInputRequiredResponse(StreamResponse):
+    class Data(BaseModel):
+        """
+        Data entity
+        """
+
+        form_id: str
+        node_id: str
+        node_title: str
+        form_content: str
+        inputs: Sequence[FormInput] = Field(default_factory=list)
+        actions: Sequence[UserAction] = Field(default_factory=list)
+        display_in_ui: bool = False
+        form_token: str | None = None
+        resolved_default_values: Mapping[str, Any] = Field(default_factory=dict)
+        expiration_time: int = Field(..., description="Unix timestamp in seconds")
+
+    event: StreamEvent = StreamEvent.HUMAN_INPUT_REQUIRED
+    workflow_run_id: str
+    data: Data
+
+
+class HumanInputFormFilledResponse(StreamResponse):
+    class Data(BaseModel):
+        """
+        Data entity
+        """
+
+        node_id: str
+        node_title: str
+        rendered_content: str
+        action_id: str
+        action_text: str
+
+    event: StreamEvent = StreamEvent.HUMAN_INPUT_FORM_FILLED
+    workflow_run_id: str
+    data: Data
+
+
+class HumanInputFormTimeoutResponse(StreamResponse):
+    class Data(BaseModel):
+        """
+        Data entity
+        """
+
+        node_id: str
+        node_title: str
+        expiration_time: int
+
+    event: StreamEvent = StreamEvent.HUMAN_INPUT_FORM_TIMEOUT
+    workflow_run_id: str
+    data: Data
+
+
 class NodeStartStreamResponse(StreamResponse):
     """
     NodeStartStreamResponse entity
@@ -726,7 +813,7 @@ class WorkflowAppBlockingResponse(AppBlockingResponse):
         total_tokens: int
         total_steps: int
         created_at: int
-        finished_at: int
+        finished_at: int | None
 
     workflow_run_id: str
     data: Data

api/core/app/features/rate_limiting/rate_limit.py

@@ -1,3 +1,4 @@
+import contextlib
 import logging
 import time
 import uuid
@@ -103,6 +104,14 @@ class RateLimit:
             )
 
 
+@contextlib.contextmanager
+def rate_limit_context(rate_limit: RateLimit, request_id: str | None):
+    request_id = rate_limit.enter(request_id)
+    yield
+    if request_id is not None:
+        rate_limit.exit(request_id)
+
+
 class RateLimitGenerator:
     def __init__(self, rate_limit: RateLimit, generator: Generator[str, None, None], request_id: str):
         self.rate_limit = rate_limit

api/core/app/layers/pause_state_persist_layer.py

@@ -1,3 +1,4 @@
+from dataclasses import dataclass
 from typing import Annotated, Literal, Self, TypeAlias
 
 from pydantic import BaseModel, Field
@@ -52,6 +53,14 @@ class WorkflowResumptionContext(BaseModel):
         return self.generate_entity.entity
 
 
+@dataclass(frozen=True)
+class PauseStateLayerConfig:
+    """Configuration container for instantiating pause persistence layers."""
+
+    session_factory: Engine | sessionmaker[Session]
+    state_owner_user_id: str
+
+
 class PauseStatePersistenceLayer(GraphEngineLayer):
     def __init__(
         self,

api/core/app/task_pipeline/message_cycle_manager.py

@@ -82,10 +82,11 @@ class MessageCycleManager:
         if isinstance(self._application_generate_entity, CompletionAppGenerateEntity):
             return None
 
-        is_first_message = self._application_generate_entity.conversation_id is None
+        is_first_message = self._application_generate_entity.is_new_conversation
         extras = self._application_generate_entity.extras
         auto_generate_conversation_name = extras.get("auto_generate_conversation_name", True)
 
+        thread: Thread | None = None
         if auto_generate_conversation_name and is_first_message:
             # start generate thread
             # time.sleep not block other logic
@@ -101,9 +102,10 @@ class MessageCycleManager:
             thread.daemon = True
             thread.start()
 
-            return thread
+        if is_first_message:
+            self._application_generate_entity.is_new_conversation = False
 
-        return None
+        return thread
 
     def _generate_conversation_name_worker(self, flask_app: Flask, conversation_id: str, query: str):
         with flask_app.app_context():

api/core/entities/execution_extra_content.py

@@ -0,0 +1,54 @@
+from __future__ import annotations
+
+from collections.abc import Mapping, Sequence
+from typing import Any, TypeAlias
+
+from pydantic import BaseModel, ConfigDict, Field
+
+from core.workflow.nodes.human_input.entities import FormInput, UserAction
+from models.execution_extra_content import ExecutionContentType
+
+
+class HumanInputFormDefinition(BaseModel):
+    model_config = ConfigDict(frozen=True)
+
+    form_id: str
+    node_id: str
+    node_title: str
+    form_content: str
+    inputs: Sequence[FormInput] = Field(default_factory=list)
+    actions: Sequence[UserAction] = Field(default_factory=list)
+    display_in_ui: bool = False
+    form_token: str | None = None
+    resolved_default_values: Mapping[str, Any] = Field(default_factory=dict)
+    expiration_time: int
+
+
+class HumanInputFormSubmissionData(BaseModel):
+    model_config = ConfigDict(frozen=True)
+
+    node_id: str
+    node_title: str
+    rendered_content: str
+    action_id: str
+    action_text: str
+
+
+class HumanInputContent(BaseModel):
+    model_config = ConfigDict(frozen=True)
+
+    workflow_run_id: str
+    submitted: bool
+    form_definition: HumanInputFormDefinition | None = None
+    form_submission_data: HumanInputFormSubmissionData | None = None
+    type: ExecutionContentType = Field(default=ExecutionContentType.HUMAN_INPUT)
+
+
+ExecutionExtraContentDomainModel: TypeAlias = HumanInputContent
+
+__all__ = [
+    "ExecutionExtraContentDomainModel",
+    "HumanInputContent",
+    "HumanInputFormDefinition",
+    "HumanInputFormSubmissionData",
+]

api/core/entities/knowledge_entities.py

@@ -3,6 +3,7 @@ from pydantic import BaseModel, Field, field_validator
 
 class PreviewDetail(BaseModel):
     content: str
+    summary: str | None = None
     child_chunks: list[str] | None = None
 
 

api/core/entities/provider_configuration.py

@@ -28,8 +28,8 @@ from core.model_runtime.entities.provider_entities import (
 )
 from core.model_runtime.model_providers.__base.ai_model import AIModel
 from core.model_runtime.model_providers.model_provider_factory import ModelProviderFactory
-from extensions.ext_database import db
 from libs.datetime_utils import naive_utc_now
+from models.engine import db
 from models.provider import (
     LoadBalancingModelConfig,
     Provider,

api/core/file/file_manager.py

@@ -104,6 +104,8 @@ def download(f: File, /):
     ):
         return _download_file_content(f.storage_key)
     elif f.transfer_method == FileTransferMethod.REMOTE_URL:
+        if f.remote_url is None:
+            raise ValueError("Missing file remote_url")
         response = ssrf_proxy.get(f.remote_url, follow_redirects=True)
         response.raise_for_status()
         return response.content
@@ -134,6 +136,8 @@ def _download_file_content(path: str, /):
 def _get_encoded_string(f: File, /):
     match f.transfer_method:
         case FileTransferMethod.REMOTE_URL:
+            if f.remote_url is None:
+                raise ValueError("Missing file remote_url")
             response = ssrf_proxy.get(f.remote_url, follow_redirects=True)
             response.raise_for_status()
             data = response.content

api/core/helper/ssrf_proxy.py

@@ -4,8 +4,10 @@ Proxy requests to avoid SSRF
 
 import logging
 import time
+from typing import Any, TypeAlias
 
 import httpx
+from pydantic import TypeAdapter, ValidationError
 
 from configs import dify_config
 from core.helper.http_client_pooling import get_pooled_http_client
@@ -18,6 +20,9 @@ SSRF_DEFAULT_MAX_RETRIES = dify_config.SSRF_DEFAULT_MAX_RETRIES
 BACKOFF_FACTOR = 0.5
 STATUS_FORCELIST = [429, 500, 502, 503, 504]
 
+Headers: TypeAlias = dict[str, str]
+_HEADERS_ADAPTER = TypeAdapter(Headers)
+
 _SSL_VERIFIED_POOL_KEY = "ssrf:verified"
 _SSL_UNVERIFIED_POOL_KEY = "ssrf:unverified"
 _SSRF_CLIENT_LIMITS = httpx.Limits(
@@ -76,7 +81,7 @@ def _get_ssrf_client(ssl_verify_enabled: bool) -> httpx.Client:
     )
 
 
-def _get_user_provided_host_header(headers: dict | None) -> str | None:
+def _get_user_provided_host_header(headers: Headers | None) -> str | None:
     """
     Extract the user-provided Host header from the headers dict.
 
@@ -92,7 +97,7 @@ def _get_user_provided_host_header(headers: dict | None) -> str | None:
     return None
 
 
-def _inject_trace_headers(headers: dict | None) -> dict:
+def _inject_trace_headers(headers: Headers | None) -> Headers:
     """
     Inject W3C traceparent header for distributed tracing.
 
@@ -125,7 +130,7 @@ def _inject_trace_headers(headers: dict | None) -> dict:
     return headers
 
 
-def make_request(method, url, max_retries=SSRF_DEFAULT_MAX_RETRIES, **kwargs):
+def make_request(method: str, url: str, max_retries: int = SSRF_DEFAULT_MAX_RETRIES, **kwargs: Any) -> httpx.Response:
     # Convert requests-style allow_redirects to httpx-style follow_redirects
     if "allow_redirects" in kwargs:
         allow_redirects = kwargs.pop("allow_redirects")
@@ -142,10 +147,15 @@ def make_request(method, url, max_retries=SSRF_DEFAULT_MAX_RETRIES, **kwargs):
 
     # prioritize per-call option, which can be switched on and off inside the HTTP node on the web UI
     verify_option = kwargs.pop("ssl_verify", dify_config.HTTP_REQUEST_NODE_SSL_VERIFY)
+    if not isinstance(verify_option, bool):
+        raise ValueError("ssl_verify must be a boolean")
     client = _get_ssrf_client(verify_option)
 
     # Inject traceparent header for distributed tracing (when OTEL is not enabled)
-    headers = kwargs.get("headers") or {}
+    try:
+        headers: Headers = _HEADERS_ADAPTER.validate_python(kwargs.get("headers") or {})
+    except ValidationError as e:
+        raise ValueError("headers must be a mapping of string keys to string values") from e
     headers = _inject_trace_headers(headers)
     kwargs["headers"] = headers
 
@@ -198,25 +208,25 @@ def make_request(method, url, max_retries=SSRF_DEFAULT_MAX_RETRIES, **kwargs):
     raise MaxRetriesExceededError(f"Reached maximum retries ({max_retries}) for URL {url}")
 
 
-def get(url, max_retries=SSRF_DEFAULT_MAX_RETRIES, **kwargs):
+def get(url: str, max_retries: int = SSRF_DEFAULT_MAX_RETRIES, **kwargs: Any) -> httpx.Response:
     return make_request("GET", url, max_retries=max_retries, **kwargs)
 
 
-def post(url, max_retries=SSRF_DEFAULT_MAX_RETRIES, **kwargs):
+def post(url: str, max_retries: int = SSRF_DEFAULT_MAX_RETRIES, **kwargs: Any) -> httpx.Response:
     return make_request("POST", url, max_retries=max_retries, **kwargs)
 
 
-def put(url, max_retries=SSRF_DEFAULT_MAX_RETRIES, **kwargs):
+def put(url: str, max_retries: int = SSRF_DEFAULT_MAX_RETRIES, **kwargs: Any) -> httpx.Response:
     return make_request("PUT", url, max_retries=max_retries, **kwargs)
 
 
-def patch(url, max_retries=SSRF_DEFAULT_MAX_RETRIES, **kwargs):
+def patch(url: str, max_retries: int = SSRF_DEFAULT_MAX_RETRIES, **kwargs: Any) -> httpx.Response:
     return make_request("PATCH", url, max_retries=max_retries, **kwargs)
 
 
-def delete(url, max_retries=SSRF_DEFAULT_MAX_RETRIES, **kwargs):
+def delete(url: str, max_retries: int = SSRF_DEFAULT_MAX_RETRIES, **kwargs: Any) -> httpx.Response:
     return make_request("DELETE", url, max_retries=max_retries, **kwargs)
 
 
-def head(url, max_retries=SSRF_DEFAULT_MAX_RETRIES, **kwargs):
+def head(url: str, max_retries: int = SSRF_DEFAULT_MAX_RETRIES, **kwargs: Any) -> httpx.Response:
     return make_request("HEAD", url, max_retries=max_retries, **kwargs)

api/core/indexing_runner.py

@@ -311,14 +311,18 @@ class IndexingRunner:
         qa_preview_texts: list[QAPreviewDetail] = []
 
         total_segments = 0
+        # doc_form represents the segmentation method (general, parent-child, QA)
         index_type = doc_form
         index_processor = IndexProcessorFactory(index_type).init_index_processor()
+        # one extract_setting is one source document
         for extract_setting in extract_settings:
             # extract
             processing_rule = DatasetProcessRule(
                 mode=tmp_processing_rule["mode"], rules=json.dumps(tmp_processing_rule["rules"])
             )
+            # Extract document content
             text_docs = index_processor.extract(extract_setting, process_rule_mode=tmp_processing_rule["mode"])
+            # Cleaning and segmentation
             documents = index_processor.transform(
                 text_docs,
                 current_user=None,
@@ -361,6 +365,12 @@ class IndexingRunner:
 
         if doc_form and doc_form == "qa_model":
             return IndexingEstimate(total_segments=total_segments * 20, qa_preview=qa_preview_texts, preview=[])
+
+        # Generate summary preview
+        summary_index_setting = tmp_processing_rule.get("summary_index_setting")
+        if summary_index_setting and summary_index_setting.get("enable") and preview_texts:
+            preview_texts = index_processor.generate_summary_preview(tenant_id, preview_texts, summary_index_setting)
+
         return IndexingEstimate(total_segments=total_segments, preview=preview_texts)
 
     def _extract(

api/core/llm_generator/entities.py

@@ -0,0 +1,20 @@
+"""Shared payload models for LLM generator helpers and controllers."""
+
+from pydantic import BaseModel, Field
+
+from core.app.app_config.entities import ModelConfig
+
+
+class RuleGeneratePayload(BaseModel):
+    instruction: str = Field(..., description="Rule generation instruction")
+    model_config_data: ModelConfig = Field(..., alias="model_config", description="Model configuration")
+    no_variable: bool = Field(default=False, description="Whether to exclude variables")
+
+
+class RuleCodeGeneratePayload(RuleGeneratePayload):
+    code_language: str = Field(default="javascript", description="Programming language for code generation")
+
+
+class RuleStructuredOutputPayload(BaseModel):
+    instruction: str = Field(..., description="Structured output generation instruction")
+    model_config_data: ModelConfig = Field(..., alias="model_config", description="Model configuration")

api/core/llm_generator/llm_generator.py

@@ -6,6 +6,8 @@ from typing import Protocol, cast
 
 import json_repair
 
+from core.app.app_config.entities import ModelConfig
+from core.llm_generator.entities import RuleCodeGeneratePayload, RuleGeneratePayload, RuleStructuredOutputPayload
 from core.llm_generator.output_parser.rule_config_generator import RuleConfigGeneratorOutputParser
 from core.llm_generator.output_parser.suggested_questions_after_answer import SuggestedQuestionsAfterAnswerOutputParser
 from core.llm_generator.prompts import (
@@ -151,19 +153,19 @@ class LLMGenerator:
         return questions
 
     @classmethod
-    def generate_rule_config(cls, tenant_id: str, instruction: str, model_config: dict, no_variable: bool):
+    def generate_rule_config(cls, tenant_id: str, args: RuleGeneratePayload):
         output_parser = RuleConfigGeneratorOutputParser()
 
         error = ""
         error_step = ""
         rule_config = {"prompt": "", "variables": [], "opening_statement": "", "error": ""}
-        model_parameters = model_config.get("completion_params", {})
-        if no_variable:
+        model_parameters = args.model_config_data.completion_params
+        if args.no_variable:
             prompt_template = PromptTemplateParser(WORKFLOW_RULE_CONFIG_PROMPT_GENERATE_TEMPLATE)
 
             prompt_generate = prompt_template.format(
                 inputs={
-                    "TASK_DESCRIPTION": instruction,
+                    "TASK_DESCRIPTION": args.instruction,
                 },
                 remove_template_variables=False,
             )
@@ -175,8 +177,8 @@ class LLMGenerator:
             model_instance = model_manager.get_model_instance(
                 tenant_id=tenant_id,
                 model_type=ModelType.LLM,
-                provider=model_config.get("provider", ""),
-                model=model_config.get("name", ""),
+                provider=args.model_config_data.provider,
+                model=args.model_config_data.name,
             )
 
             try:
@@ -190,7 +192,7 @@ class LLMGenerator:
                 error = str(e)
                 error_step = "generate rule config"
             except Exception as e:
-                logger.exception("Failed to generate rule config, model: %s", model_config.get("name"))
+                logger.exception("Failed to generate rule config, model: %s", args.model_config_data.name)
                 rule_config["error"] = str(e)
 
             rule_config["error"] = f"Failed to {error_step}. Error: {error}" if error else ""
@@ -209,7 +211,7 @@ class LLMGenerator:
         # format the prompt_generate_prompt
         prompt_generate_prompt = prompt_template.format(
             inputs={
-                "TASK_DESCRIPTION": instruction,
+                "TASK_DESCRIPTION": args.instruction,
             },
             remove_template_variables=False,
         )
@@ -220,8 +222,8 @@ class LLMGenerator:
         model_instance = model_manager.get_model_instance(
             tenant_id=tenant_id,
             model_type=ModelType.LLM,
-            provider=model_config.get("provider", ""),
-            model=model_config.get("name", ""),
+            provider=args.model_config_data.provider,
+            model=args.model_config_data.name,
         )
 
         try:
@@ -250,7 +252,7 @@ class LLMGenerator:
             # the second step to generate the task_parameter and task_statement
             statement_generate_prompt = statement_template.format(
                 inputs={
-                    "TASK_DESCRIPTION": instruction,
+                    "TASK_DESCRIPTION": args.instruction,
                     "INPUT_TEXT": prompt_content.message.get_text_content(),
                 },
                 remove_template_variables=False,
@@ -276,7 +278,7 @@ class LLMGenerator:
                 error_step = "generate conversation opener"
 
         except Exception as e:
-            logger.exception("Failed to generate rule config, model: %s", model_config.get("name"))
+            logger.exception("Failed to generate rule config, model: %s", args.model_config_data.name)
             rule_config["error"] = str(e)
 
         rule_config["error"] = f"Failed to {error_step}. Error: {error}" if error else ""
@@ -284,16 +286,20 @@ class LLMGenerator:
         return rule_config
 
     @classmethod
-    def generate_code(cls, tenant_id: str, instruction: str, model_config: dict, code_language: str = "javascript"):
-        if code_language == "python":
+    def generate_code(
+        cls,
+        tenant_id: str,
+        args: RuleCodeGeneratePayload,
+    ):
+        if args.code_language == "python":
             prompt_template = PromptTemplateParser(PYTHON_CODE_GENERATOR_PROMPT_TEMPLATE)
         else:
             prompt_template = PromptTemplateParser(JAVASCRIPT_CODE_GENERATOR_PROMPT_TEMPLATE)
 
         prompt = prompt_template.format(
             inputs={
-                "INSTRUCTION": instruction,
-                "CODE_LANGUAGE": code_language,
+                "INSTRUCTION": args.instruction,
+                "CODE_LANGUAGE": args.code_language,
             },
             remove_template_variables=False,
         )
@@ -302,28 +308,28 @@ class LLMGenerator:
         model_instance = model_manager.get_model_instance(
             tenant_id=tenant_id,
             model_type=ModelType.LLM,
-            provider=model_config.get("provider", ""),
-            model=model_config.get("name", ""),
+            provider=args.model_config_data.provider,
+            model=args.model_config_data.name,
         )
 
         prompt_messages = [UserPromptMessage(content=prompt)]
-        model_parameters = model_config.get("completion_params", {})
+        model_parameters = args.model_config_data.completion_params
         try:
             response: LLMResult = model_instance.invoke_llm(
                 prompt_messages=list(prompt_messages), model_parameters=model_parameters, stream=False
             )
 
             generated_code = response.message.get_text_content()
-            return {"code": generated_code, "language": code_language, "error": ""}
+            return {"code": generated_code, "language": args.code_language, "error": ""}
 
         except InvokeError as e:
             error = str(e)
-            return {"code": "", "language": code_language, "error": f"Failed to generate code. Error: {error}"}
+            return {"code": "", "language": args.code_language, "error": f"Failed to generate code. Error: {error}"}
         except Exception as e:
             logger.exception(
-                "Failed to invoke LLM model, model: %s, language: %s", model_config.get("name"), code_language
+                "Failed to invoke LLM model, model: %s, language: %s", args.model_config_data.name, args.code_language
             )
-            return {"code": "", "language": code_language, "error": f"An unexpected error occurred: {str(e)}"}
+            return {"code": "", "language": args.code_language, "error": f"An unexpected error occurred: {str(e)}"}
 
     @classmethod
     def generate_qa_document(cls, tenant_id: str, query, document_language: str):
@@ -353,20 +359,20 @@ class LLMGenerator:
         return answer.strip()
 
     @classmethod
-    def generate_structured_output(cls, tenant_id: str, instruction: str, model_config: dict):
+    def generate_structured_output(cls, tenant_id: str, args: RuleStructuredOutputPayload):
         model_manager = ModelManager()
         model_instance = model_manager.get_model_instance(
             tenant_id=tenant_id,
             model_type=ModelType.LLM,
-            provider=model_config.get("provider", ""),
-            model=model_config.get("name", ""),
+            provider=args.model_config_data.provider,
+            model=args.model_config_data.name,
         )
 
         prompt_messages = [
             SystemPromptMessage(content=SYSTEM_STRUCTURED_OUTPUT_GENERATE),
-            UserPromptMessage(content=instruction),
+            UserPromptMessage(content=args.instruction),
         ]
-        model_parameters = model_config.get("model_parameters", {})
+        model_parameters = args.model_config_data.completion_params
 
         try:
             response: LLMResult = model_instance.invoke_llm(
@@ -390,12 +396,17 @@ class LLMGenerator:
             error = str(e)
             return {"output": "", "error": f"Failed to generate JSON Schema. Error: {error}"}
         except Exception as e:
-            logger.exception("Failed to invoke LLM model, model: %s", model_config.get("name"))
+            logger.exception("Failed to invoke LLM model, model: %s", args.model_config_data.name)
             return {"output": "", "error": f"An unexpected error occurred: {str(e)}"}
 
     @staticmethod
     def instruction_modify_legacy(
-        tenant_id: str, flow_id: str, current: str, instruction: str, model_config: dict, ideal_output: str | None
+        tenant_id: str,
+        flow_id: str,
+        current: str,
+        instruction: str,
+        model_config: ModelConfig,
+        ideal_output: str | None,
     ):
         last_run: Message | None = (
             db.session.query(Message).where(Message.app_id == flow_id).order_by(Message.created_at.desc()).first()
@@ -434,7 +445,7 @@ class LLMGenerator:
         node_id: str,
         current: str,
         instruction: str,
-        model_config: dict,
+        model_config: ModelConfig,
         ideal_output: str | None,
         workflow_service: WorkflowServiceInterface,
     ):
@@ -505,7 +516,7 @@ class LLMGenerator:
     @staticmethod
     def __instruction_modify_common(
         tenant_id: str,
-        model_config: dict,
+        model_config: ModelConfig,
         last_run: dict | None,
         current: str | None,
         error_message: str | None,
@@ -526,8 +537,8 @@ class LLMGenerator:
         model_instance = ModelManager().get_model_instance(
             tenant_id=tenant_id,
             model_type=ModelType.LLM,
-            provider=model_config.get("provider", ""),
-            model=model_config.get("name", ""),
+            provider=model_config.provider,
+            model=model_config.name,
         )
         match node_type:
             case "llm" | "agent":
@@ -570,7 +581,5 @@ class LLMGenerator:
             error = str(e)
             return {"error": f"Failed to generate code. Error: {error}"}
         except Exception as e:
-            logger.exception(
-                "Failed to invoke LLM model, model: %s", json.dumps(model_config.get("name")), exc_info=True
-            )
+            logger.exception("Failed to invoke LLM model, model: %s", json.dumps(model_config.name), exc_info=True)
             return {"error": f"An unexpected error occurred: {str(e)}"}

api/core/llm_generator/prompts.py

@@ -434,3 +434,20 @@ INSTRUCTION_GENERATE_TEMPLATE_PROMPT = """The output of this prompt is not as ex
 You should edit the prompt according to the IDEAL OUTPUT."""
 
 INSTRUCTION_GENERATE_TEMPLATE_CODE = """Please fix the errors in the {{#error_message#}}."""
+
+DEFAULT_GENERATOR_SUMMARY_PROMPT = (
+    """Summarize the following content. Extract only the key information and main points. """
+    """Remove redundant details.
+
+Requirements:
+1. Write a concise summary in plain text
+2. Use the same language as the input content
+3. Focus on important facts, concepts, and details
+4. If images are included, describe their key information
+5. Do not use words like "好的", "ok", "I understand", "This text discusses", "The content mentions"
+6. Write directly without extra words
+
+Output only the summary text. Start summarizing now:
+
+"""
+)

api/core/mcp/session/base_session.py

@@ -347,7 +347,7 @@ class BaseSession(
                         message.message.root.model_dump(by_alias=True, mode="json", exclude_none=True)
                     )
 
-                    responder = RequestResponder(
+                    responder = RequestResponder[ReceiveRequestT, SendResultT](
                         request_id=message.message.root.id,
                         request_meta=validated_request.root.params.meta if validated_request.root.params else None,
                         request=validated_request,

api/core/model_runtime/model_providers/__base/ai_model.py

@@ -1,10 +1,11 @@
 import decimal
 import hashlib
-from threading import Lock
+import logging
 
-from pydantic import BaseModel, ConfigDict, Field
+from pydantic import BaseModel, ConfigDict, Field, ValidationError
+from redis import RedisError
 
-import contexts
+from configs import dify_config
 from core.model_runtime.entities.common_entities import I18nObject
 from core.model_runtime.entities.defaults import PARAMETER_RULE_TEMPLATE
 from core.model_runtime.entities.model_entities import (
@@ -24,6 +25,9 @@ from core.model_runtime.errors.invoke import (
     InvokeServerUnavailableError,
 )
 from core.plugin.entities.plugin_daemon import PluginModelProviderEntity
+from extensions.ext_redis import redis_client
+
+logger = logging.getLogger(__name__)
 
 
 class AIModel(BaseModel):
@@ -144,34 +148,60 @@ class AIModel(BaseModel):
 
         plugin_model_manager = PluginModelClient()
         cache_key = f"{self.tenant_id}:{self.plugin_id}:{self.provider_name}:{self.model_type.value}:{model}"
-        # sort credentials
         sorted_credentials = sorted(credentials.items()) if credentials else []
         cache_key += ":".join([hashlib.md5(f"{k}:{v}".encode()).hexdigest() for k, v in sorted_credentials])
 
+        cached_schema_json = None
         try:
-            contexts.plugin_model_schemas.get()
-        except LookupError:
-            contexts.plugin_model_schemas.set({})
-            contexts.plugin_model_schema_lock.set(Lock())
-
-        with contexts.plugin_model_schema_lock.get():
-            if cache_key in contexts.plugin_model_schemas.get():
-                return contexts.plugin_model_schemas.get()[cache_key]
-
-            schema = plugin_model_manager.get_model_schema(
-                tenant_id=self.tenant_id,
-                user_id="unknown",
-                plugin_id=self.plugin_id,
-                provider=self.provider_name,
-                model_type=self.model_type.value,
-                model=model,
-                credentials=credentials or {},
+            cached_schema_json = redis_client.get(cache_key)
+        except (RedisError, RuntimeError) as exc:
+            logger.warning(
+                "Failed to read plugin model schema cache for model %s: %s",
+                model,
+                str(exc),
+                exc_info=True,
             )
+        if cached_schema_json:
+            try:
+                return AIModelEntity.model_validate_json(cached_schema_json)
+            except ValidationError:
+                logger.warning(
+                    "Failed to validate cached plugin model schema for model %s",
+                    model,
+                    exc_info=True,
+                )
+                try:
+                    redis_client.delete(cache_key)
+                except (RedisError, RuntimeError) as exc:
+                    logger.warning(
+                        "Failed to delete invalid plugin model schema cache for model %s: %s",
+                        model,
+                        str(exc),
+                        exc_info=True,
+                    )
 
-            if schema:
-                contexts.plugin_model_schemas.get()[cache_key] = schema
+        schema = plugin_model_manager.get_model_schema(
+            tenant_id=self.tenant_id,
+            user_id="unknown",
+            plugin_id=self.plugin_id,
+            provider=self.provider_name,
+            model_type=self.model_type.value,
+            model=model,
+            credentials=credentials or {},
+        )
 
-            return schema
+        if schema:
+            try:
+                redis_client.setex(cache_key, dify_config.PLUGIN_MODEL_SCHEMA_CACHE_TTL, schema.model_dump_json())
+            except (RedisError, RuntimeError) as exc:
+                logger.warning(
+                    "Failed to write plugin model schema cache for model %s: %s",
+                    model,
+                    str(exc),
+                    exc_info=True,
+                )
+
+        return schema
 
     def get_customizable_model_schema_from_credentials(self, model: str, credentials: dict) -> AIModelEntity | None:
         """

api/core/model_runtime/model_providers/__base/large_language_model.py

@@ -92,6 +92,10 @@ def _build_llm_result_from_first_chunk(
     Build a single `LLMResult` from the first returned chunk.
 
     This is used for `stream=False` because the plugin side may still implement the response via a chunked stream.
+
+    Note:
+        This function always drains the `chunks` iterator after reading the first chunk to ensure any underlying
+        streaming resources are released (e.g., HTTP connections owned by the plugin runtime).
     """
     content = ""
     content_list: list[PromptMessageContentUnionTypes] = []
@@ -99,18 +103,25 @@ def _build_llm_result_from_first_chunk(
     system_fingerprint: str | None = None
     tools_calls: list[AssistantPromptMessage.ToolCall] = []
 
-    first_chunk = next(chunks, None)
-    if first_chunk is not None:
-        if isinstance(first_chunk.delta.message.content, str):
-            content += first_chunk.delta.message.content
-        elif isinstance(first_chunk.delta.message.content, list):
-            content_list.extend(first_chunk.delta.message.content)
+    try:
+        first_chunk = next(chunks, None)
+        if first_chunk is not None:
+            if isinstance(first_chunk.delta.message.content, str):
+                content += first_chunk.delta.message.content
+            elif isinstance(first_chunk.delta.message.content, list):
+                content_list.extend(first_chunk.delta.message.content)
 
-        if first_chunk.delta.message.tool_calls:
-            _increase_tool_call(first_chunk.delta.message.tool_calls, tools_calls)
+            if first_chunk.delta.message.tool_calls:
+                _increase_tool_call(first_chunk.delta.message.tool_calls, tools_calls)
 
-        usage = first_chunk.delta.usage or LLMUsage.empty_usage()
-        system_fingerprint = first_chunk.system_fingerprint
+            usage = first_chunk.delta.usage or LLMUsage.empty_usage()
+            system_fingerprint = first_chunk.system_fingerprint
+    finally:
+        try:
+            for _ in chunks:
+                pass
+        except Exception:
+            logger.debug("Failed to drain non-stream plugin chunk iterator.", exc_info=True)
 
     return LLMResult(
         model=model,
@@ -283,7 +294,7 @@ class LargeLanguageModel(AIModel):
             # TODO
             raise self._transform_invoke_error(e)
 
-        if stream and isinstance(result, Generator):
+        if stream and not isinstance(result, LLMResult):
             return self._invoke_result_generator(
                 model=model,
                 result=result,

api/core/model_runtime/model_providers/model_provider_factory.py

@@ -5,7 +5,11 @@ import logging
 from collections.abc import Sequence
 from threading import Lock
 
+from pydantic import ValidationError
+from redis import RedisError
+
 import contexts
+from configs import dify_config
 from core.model_runtime.entities.model_entities import AIModelEntity, ModelType
 from core.model_runtime.entities.provider_entities import ProviderConfig, ProviderEntity, SimpleProviderEntity
 from core.model_runtime.model_providers.__base.ai_model import AIModel
@@ -18,6 +22,7 @@ from core.model_runtime.model_providers.__base.tts_model import TTSModel
 from core.model_runtime.schema_validators.model_credential_schema_validator import ModelCredentialSchemaValidator
 from core.model_runtime.schema_validators.provider_credential_schema_validator import ProviderCredentialSchemaValidator
 from core.plugin.entities.plugin_daemon import PluginModelProviderEntity
+from extensions.ext_redis import redis_client
 from models.provider_ids import ModelProviderID
 
 logger = logging.getLogger(__name__)
@@ -175,34 +180,60 @@ class ModelProviderFactory:
         """
         plugin_id, provider_name = self.get_plugin_id_and_provider_name_from_provider(provider)
         cache_key = f"{self.tenant_id}:{plugin_id}:{provider_name}:{model_type.value}:{model}"
-        # sort credentials
         sorted_credentials = sorted(credentials.items()) if credentials else []
         cache_key += ":".join([hashlib.md5(f"{k}:{v}".encode()).hexdigest() for k, v in sorted_credentials])
 
+        cached_schema_json = None
         try:
-            contexts.plugin_model_schemas.get()
-        except LookupError:
-            contexts.plugin_model_schemas.set({})
-            contexts.plugin_model_schema_lock.set(Lock())
-
-        with contexts.plugin_model_schema_lock.get():
-            if cache_key in contexts.plugin_model_schemas.get():
-                return contexts.plugin_model_schemas.get()[cache_key]
-
-            schema = self.plugin_model_manager.get_model_schema(
-                tenant_id=self.tenant_id,
-                user_id="unknown",
-                plugin_id=plugin_id,
-                provider=provider_name,
-                model_type=model_type.value,
-                model=model,
-                credentials=credentials or {},
+            cached_schema_json = redis_client.get(cache_key)
+        except (RedisError, RuntimeError) as exc:
+            logger.warning(
+                "Failed to read plugin model schema cache for model %s: %s",
+                model,
+                str(exc),
+                exc_info=True,
             )
+        if cached_schema_json:
+            try:
+                return AIModelEntity.model_validate_json(cached_schema_json)
+            except ValidationError:
+                logger.warning(
+                    "Failed to validate cached plugin model schema for model %s",
+                    model,
+                    exc_info=True,
+                )
+                try:
+                    redis_client.delete(cache_key)
+                except (RedisError, RuntimeError) as exc:
+                    logger.warning(
+                        "Failed to delete invalid plugin model schema cache for model %s: %s",
+                        model,
+                        str(exc),
+                        exc_info=True,
+                    )
 
-            if schema:
-                contexts.plugin_model_schemas.get()[cache_key] = schema
+        schema = self.plugin_model_manager.get_model_schema(
+            tenant_id=self.tenant_id,
+            user_id="unknown",
+            plugin_id=plugin_id,
+            provider=provider_name,
+            model_type=model_type.value,
+            model=model,
+            credentials=credentials or {},
+        )
 
-            return schema
+        if schema:
+            try:
+                redis_client.setex(cache_key, dify_config.PLUGIN_MODEL_SCHEMA_CACHE_TTL, schema.model_dump_json())
+            except (RedisError, RuntimeError) as exc:
+                logger.warning(
+                    "Failed to write plugin model schema cache for model %s: %s",
+                    model,
+                    str(exc),
+                    exc_info=True,
+                )
+
+        return schema
 
     def get_models(
         self,
@@ -283,6 +314,8 @@ class ModelProviderFactory:
         elif model_type == ModelType.TTS:
             return TTSModel.model_validate(init_params)
 
+        raise ValueError(f"Unsupported model type: {model_type}")
+
     def get_provider_icon(self, provider: str, icon_type: str, lang: str) -> tuple[bytes, str]:
         """
         Get provider icon

api/core/ops/ops_trace_manager.py

@@ -15,10 +15,7 @@ from sqlalchemy import select
 from sqlalchemy.orm import Session, sessionmaker
 
 from core.helper.encrypter import batch_decrypt_token, encrypt_token, obfuscated_token
-from core.ops.entities.config_entity import (
-    OPS_FILE_PATH,
-    TracingProviderEnum,
-)
+from core.ops.entities.config_entity import OPS_FILE_PATH, TracingProviderEnum
 from core.ops.entities.trace_entity import (
     DatasetRetrievalTraceInfo,
     GenerateNameTraceInfo,
@@ -31,8 +28,8 @@ from core.ops.entities.trace_entity import (
     WorkflowTraceInfo,
 )
 from core.ops.utils import get_message_data
-from extensions.ext_database import db
 from extensions.ext_storage import storage
+from models.engine import db
 from models.model import App, AppModelConfig, Conversation, Message, MessageFile, TraceAppConfig
 from models.workflow import WorkflowAppLog
 from tasks.ops_trace_task import process_trace_tasks
@@ -469,6 +466,8 @@ class TraceTask:
 
     @classmethod
     def _get_workflow_run_repo(cls):
+        from repositories.factory import DifyAPIRepositoryFactory
+
         if cls._workflow_run_repo is None:
             with cls._repo_lock:
                 if cls._workflow_run_repo is None:

api/core/ops/utils.py

@@ -5,7 +5,7 @@ from urllib.parse import urlparse
 
 from sqlalchemy import select
 
-from extensions.ext_database import db
+from models.engine import db
 from models.model import Message
 
 

api/core/plugin/backwards_invocation/app.py

@@ -1,3 +1,4 @@
+import uuid
 from collections.abc import Generator, Mapping
 from typing import Union
 
@@ -11,6 +12,7 @@ from core.app.apps.chat.app_generator import ChatAppGenerator
 from core.app.apps.completion.app_generator import CompletionAppGenerator
 from core.app.apps.workflow.app_generator import WorkflowAppGenerator
 from core.app.entities.app_invoke_entities import InvokeFrom
+from core.app.layers.pause_state_persist_layer import PauseStateLayerConfig
 from core.plugin.backwards_invocation.base import BaseBackwardsInvocation
 from extensions.ext_database import db
 from models import Account
@@ -101,6 +103,11 @@ class PluginAppBackwardsInvocation(BaseBackwardsInvocation):
             if not workflow:
                 raise ValueError("unexpected app type")
 
+            pause_config = PauseStateLayerConfig(
+                session_factory=db.engine,
+                state_owner_user_id=workflow.created_by,
+            )
+
             return AdvancedChatAppGenerator().generate(
                 app_model=app,
                 workflow=workflow,
@@ -112,7 +119,9 @@ class PluginAppBackwardsInvocation(BaseBackwardsInvocation):
                     "conversation_id": conversation_id,
                 },
                 invoke_from=InvokeFrom.SERVICE_API,
+                workflow_run_id=str(uuid.uuid4()),
                 streaming=stream,
+                pause_state_config=pause_config,
             )
         elif app.mode == AppMode.AGENT_CHAT:
             return AgentChatAppGenerator().generate(
@@ -159,6 +168,11 @@ class PluginAppBackwardsInvocation(BaseBackwardsInvocation):
         if not workflow:
             raise ValueError("unexpected app type")
 
+        pause_config = PauseStateLayerConfig(
+            session_factory=db.engine,
+            state_owner_user_id=workflow.created_by,
+        )
+
         return WorkflowAppGenerator().generate(
             app_model=app,
             workflow=workflow,
@@ -167,6 +181,7 @@ class PluginAppBackwardsInvocation(BaseBackwardsInvocation):
             invoke_from=InvokeFrom.SERVICE_API,
             streaming=stream,
             call_depth=1,
+            pause_state_config=pause_config,
         )
 
     @classmethod

api/core/rag/datasource/retrieval_service.py

@@ -24,7 +24,13 @@ from core.rag.rerank.rerank_type import RerankMode
 from core.rag.retrieval.retrieval_methods import RetrievalMethod
 from core.tools.signature import sign_upload_file
 from extensions.ext_database import db
-from models.dataset import ChildChunk, Dataset, DocumentSegment, SegmentAttachmentBinding
+from models.dataset import (
+    ChildChunk,
+    Dataset,
+    DocumentSegment,
+    DocumentSegmentSummary,
+    SegmentAttachmentBinding,
+)
 from models.dataset import Document as DatasetDocument
 from models.model import UploadFile
 from services.external_knowledge_service import ExternalDatasetService
@@ -389,15 +395,15 @@ class RetrievalService:
                 .all()
             }
 
-            records = []
-            include_segment_ids = set()
-            segment_child_map = {}
-
             valid_dataset_documents = {}
             image_doc_ids: list[Any] = []
             child_index_node_ids = []
             index_node_ids = []
             doc_to_document_map = {}
+            summary_segment_ids = set()  # Track segments retrieved via summary
+            summary_score_map: dict[str, float] = {}  # Map original_chunk_id to summary score
+
+            # First pass: collect all document IDs and identify summary documents
             for document in documents:
                 document_id = document.metadata.get("document_id")
                 if document_id not in dataset_documents:
@@ -408,16 +414,39 @@ class RetrievalService:
                     continue
                 valid_dataset_documents[document_id] = dataset_document
 
+                doc_id = document.metadata.get("doc_id") or ""
+                doc_to_document_map[doc_id] = document
+
+                # Check if this is a summary document
+                is_summary = document.metadata.get("is_summary", False)
+                if is_summary:
+                    # For summary documents, find the original chunk via original_chunk_id
+                    original_chunk_id = document.metadata.get("original_chunk_id")
+                    if original_chunk_id:
+                        summary_segment_ids.add(original_chunk_id)
+                        # Save summary's score for later use
+                        summary_score = document.metadata.get("score")
+                        if summary_score is not None:
+                            try:
+                                summary_score_float = float(summary_score)
+                                # If the same segment has multiple summary hits, take the highest score
+                                if original_chunk_id not in summary_score_map:
+                                    summary_score_map[original_chunk_id] = summary_score_float
+                                else:
+                                    summary_score_map[original_chunk_id] = max(
+                                        summary_score_map[original_chunk_id], summary_score_float
+                                    )
+                            except (ValueError, TypeError):
+                                # Skip invalid score values
+                                pass
+                    continue  # Skip adding to other lists for summary documents
+
                 if dataset_document.doc_form == IndexStructureType.PARENT_CHILD_INDEX:
-                    doc_id = document.metadata.get("doc_id") or ""
-                    doc_to_document_map[doc_id] = document
                     if document.metadata.get("doc_type") == DocType.IMAGE:
                         image_doc_ids.append(doc_id)
                     else:
                         child_index_node_ids.append(doc_id)
                 else:
-                    doc_id = document.metadata.get("doc_id") or ""
-                    doc_to_document_map[doc_id] = document
                     if document.metadata.get("doc_type") == DocType.IMAGE:
                         image_doc_ids.append(doc_id)
                     else:
@@ -433,6 +462,7 @@ class RetrievalService:
             attachment_map: dict[str, list[dict[str, Any]]] = {}
             child_chunk_map: dict[str, list[ChildChunk]] = {}
             doc_segment_map: dict[str, list[str]] = {}
+            segment_summary_map: dict[str, str] = {}  # Map segment_id to summary content
 
             with session_factory.create_session() as session:
                 attachments = cls.get_segment_attachment_infos(image_doc_ids, session)
@@ -447,6 +477,7 @@ class RetrievalService:
                         doc_segment_map[attachment["segment_id"]].append(attachment["attachment_id"])
                     else:
                         doc_segment_map[attachment["segment_id"]] = [attachment["attachment_id"]]
+
                 child_chunk_stmt = select(ChildChunk).where(ChildChunk.index_node_id.in_(child_index_node_ids))
                 child_index_nodes = session.execute(child_chunk_stmt).scalars().all()
 
@@ -470,6 +501,7 @@ class RetrievalService:
                     index_node_segments = session.execute(document_segment_stmt).scalars().all()  # type: ignore
                     for index_node_segment in index_node_segments:
                         doc_segment_map[index_node_segment.id] = [index_node_segment.index_node_id]
+
                 if segment_ids:
                     document_segment_stmt = select(DocumentSegment).where(
                         DocumentSegment.enabled == True,
@@ -481,6 +513,40 @@ class RetrievalService:
                 if index_node_segments:
                     segments.extend(index_node_segments)
 
+                # Handle summary documents: query segments by original_chunk_id
+                if summary_segment_ids:
+                    summary_segment_ids_list = list(summary_segment_ids)
+                    summary_segment_stmt = select(DocumentSegment).where(
+                        DocumentSegment.enabled == True,
+                        DocumentSegment.status == "completed",
+                        DocumentSegment.id.in_(summary_segment_ids_list),
+                    )
+                    summary_segments = session.execute(summary_segment_stmt).scalars().all()  # type: ignore
+                    segments.extend(summary_segments)
+                    # Add summary segment IDs to segment_ids for summary query
+                    for seg in summary_segments:
+                        if seg.id not in segment_ids:
+                            segment_ids.append(seg.id)
+
+                # Batch query summaries for segments retrieved via summary (only enabled summaries)
+                if summary_segment_ids:
+                    summaries = (
+                        session.query(DocumentSegmentSummary)
+                        .filter(
+                            DocumentSegmentSummary.chunk_id.in_(list(summary_segment_ids)),
+                            DocumentSegmentSummary.status == "completed",
+                            DocumentSegmentSummary.enabled == True,  # Only retrieve enabled summaries
+                        )
+                        .all()
+                    )
+                    for summary in summaries:
+                        if summary.summary_content:
+                            segment_summary_map[summary.chunk_id] = summary.summary_content
+
+            include_segment_ids = set()
+            segment_child_map: dict[str, dict[str, Any]] = {}
+            records: list[dict[str, Any]] = []
+
             for segment in segments:
                 child_chunks: list[ChildChunk] = child_chunk_map.get(segment.id, [])
                 attachment_infos: list[dict[str, Any]] = attachment_map.get(segment.id, [])
@@ -489,30 +555,44 @@ class RetrievalService:
                 if ds_dataset_document and ds_dataset_document.doc_form == IndexStructureType.PARENT_CHILD_INDEX:
                     if segment.id not in include_segment_ids:
                         include_segment_ids.add(segment.id)
+                        # Check if this segment was retrieved via summary
+                        # Use summary score as base score if available, otherwise 0.0
+                        max_score = summary_score_map.get(segment.id, 0.0)
+
                         if child_chunks or attachment_infos:
                             child_chunk_details = []
-                            max_score = 0.0
                             for child_chunk in child_chunks:
-                                document = doc_to_document_map[child_chunk.index_node_id]
+                                child_document: Document | None = doc_to_document_map.get(child_chunk.index_node_id)
+                                if child_document:
+                                    child_score = child_document.metadata.get("score", 0.0)
+                                else:
+                                    child_score = 0.0
                                 child_chunk_detail = {
                                     "id": child_chunk.id,
                                     "content": child_chunk.content,
                                     "position": child_chunk.position,
-                                    "score": document.metadata.get("score", 0.0) if document else 0.0,
+                                    "score": child_score,
                                 }
                                 child_chunk_details.append(child_chunk_detail)
-                                max_score = max(max_score, document.metadata.get("score", 0.0) if document else 0.0)
+                                max_score = max(max_score, child_score)
                             for attachment_info in attachment_infos:
-                                file_document = doc_to_document_map[attachment_info["id"]]
-                                max_score = max(
-                                    max_score, file_document.metadata.get("score", 0.0) if file_document else 0.0
-                                )
+                                file_document = doc_to_document_map.get(attachment_info["id"])
+                                if file_document:
+                                    max_score = max(max_score, file_document.metadata.get("score", 0.0))
 
                             map_detail = {
                                 "max_score": max_score,
                                 "child_chunks": child_chunk_details,
                             }
                             segment_child_map[segment.id] = map_detail
+                        else:
+                            # No child chunks or attachments, use summary score if available
+                            summary_score = summary_score_map.get(segment.id)
+                            if summary_score is not None:
+                                segment_child_map[segment.id] = {
+                                    "max_score": summary_score,
+                                    "child_chunks": [],
+                                }
                         record: dict[str, Any] = {
                             "segment": segment,
                         }
@@ -520,14 +600,23 @@ class RetrievalService:
                 else:
                     if segment.id not in include_segment_ids:
                         include_segment_ids.add(segment.id)
-                        max_score = 0.0
-                        segment_document = doc_to_document_map.get(segment.index_node_id)
-                        if segment_document:
-                            max_score = max(max_score, segment_document.metadata.get("score", 0.0))
+
+                        # Check if this segment was retrieved via summary
+                        # Use summary score if available (summary retrieval takes priority)
+                        max_score = summary_score_map.get(segment.id, 0.0)
+
+                        # If not retrieved via summary, use original segment's score
+                        if segment.id not in summary_score_map:
+                            segment_document = doc_to_document_map.get(segment.index_node_id)
+                            if segment_document:
+                                max_score = max(max_score, segment_document.metadata.get("score", 0.0))
+
+                        # Also consider attachment scores
                         for attachment_info in attachment_infos:
                             file_doc = doc_to_document_map.get(attachment_info["id"])
                             if file_doc:
                                 max_score = max(max_score, file_doc.metadata.get("score", 0.0))
+
                         record = {
                             "segment": segment,
                             "score": max_score,
@@ -576,9 +665,16 @@ class RetrievalService:
                     else None
                 )
 
+                # Extract summary if this segment was retrieved via summary
+                summary_content = segment_summary_map.get(segment.id)
+
                 # Create RetrievalSegments object
                 retrieval_segment = RetrievalSegments(
-                    segment=segment, child_chunks=child_chunks_list, score=score, files=files
+                    segment=segment,
+                    child_chunks=child_chunks_list,
+                    score=score,
+                    files=files,
+                    summary=summary_content,
                 )
                 result.append(retrieval_segment)
 

api/core/rag/embedding/retrieval.py

@@ -20,3 +20,4 @@ class RetrievalSegments(BaseModel):
     child_chunks: list[RetrievalChildChunk] | None = None
     score: float | None = None
     files: list[dict[str, str | int]] | None = None
+    summary: str | None = None  # Summary content if retrieved via summary index

api/core/rag/entities/citation_metadata.py

@@ -22,3 +22,4 @@ class RetrievalSourceMetadata(BaseModel):
     doc_metadata: dict[str, Any] | None = None
     title: str | None = None
     files: list[dict[str, Any]] | None = None
+    summary: str | None = None

api/core/rag/extractor/word_extractor.py

@@ -1,4 +1,7 @@
-"""Abstract interface for document loader implementations."""
+"""Word (.docx) document extractor used for RAG ingestion.
+
+Supports local file paths and remote URLs (downloaded via `core.helper.ssrf_proxy`).
+"""
 
 import logging
 import mimetypes
@@ -8,7 +11,6 @@ import tempfile
 import uuid
 from urllib.parse import urlparse
 
-import httpx
 from docx import Document as DocxDocument
 from docx.oxml.ns import qn
 from docx.text.run import Run
@@ -44,7 +46,7 @@ class WordExtractor(BaseExtractor):
 
         # If the file is a web path, download it to a temporary file, and use that
         if not os.path.isfile(self.file_path) and self._is_valid_url(self.file_path):
-            response = httpx.get(self.file_path, timeout=None)
+            response = ssrf_proxy.get(self.file_path)
 
             if response.status_code != 200:
                 response.close()
@@ -55,6 +57,7 @@ class WordExtractor(BaseExtractor):
             self.temp_file = tempfile.NamedTemporaryFile()  # noqa SIM115
             try:
                 self.temp_file.write(response.content)
+                self.temp_file.flush()
             finally:
                 response.close()
             self.file_path = self.temp_file.name

api/core/rag/index_processor/index_processor_base.py

@@ -13,6 +13,7 @@ from urllib.parse import unquote, urlparse
 import httpx
 
 from configs import dify_config
+from core.entities.knowledge_entities import PreviewDetail
 from core.helper import ssrf_proxy
 from core.rag.extractor.entity.extract_setting import ExtractSetting
 from core.rag.index_processor.constant.doc_type import DocType
@@ -45,6 +46,17 @@ class BaseIndexProcessor(ABC):
     def transform(self, documents: list[Document], current_user: Account | None = None, **kwargs) -> list[Document]:
         raise NotImplementedError
 
+    @abstractmethod
+    def generate_summary_preview(
+        self, tenant_id: str, preview_texts: list[PreviewDetail], summary_index_setting: dict
+    ) -> list[PreviewDetail]:
+        """
+        For each segment in preview_texts, generate a summary using LLM and attach it to the segment.
+        The summary can be stored in a new attribute, e.g., summary.
+        This method should be implemented by subclasses.
+        """
+        raise NotImplementedError
+
     @abstractmethod
     def load(
         self,

api/core/rag/index_processor/processor/paragraph_index_processor.py

@@ -1,9 +1,27 @@
 """Paragraph index processor."""
 
+import logging
+import re
 import uuid
 from collections.abc import Mapping
-from typing import Any
+from typing import Any, cast
 
+logger = logging.getLogger(__name__)
+
+from core.entities.knowledge_entities import PreviewDetail
+from core.file import File, FileTransferMethod, FileType, file_manager
+from core.llm_generator.prompts import DEFAULT_GENERATOR_SUMMARY_PROMPT
+from core.model_manager import ModelInstance
+from core.model_runtime.entities.llm_entities import LLMResult, LLMUsage
+from core.model_runtime.entities.message_entities import (
+    ImagePromptMessageContent,
+    PromptMessage,
+    PromptMessageContentUnionTypes,
+    TextPromptMessageContent,
+    UserPromptMessage,
+)
+from core.model_runtime.entities.model_entities import ModelFeature, ModelType
+from core.provider_manager import ProviderManager
 from core.rag.cleaner.clean_processor import CleanProcessor
 from core.rag.datasource.keyword.keyword_factory import Keyword
 from core.rag.datasource.retrieval_service import RetrievalService
@@ -17,12 +35,17 @@ from core.rag.index_processor.index_processor_base import BaseIndexProcessor
 from core.rag.models.document import AttachmentDocument, Document, MultimodalGeneralStructureChunk
 from core.rag.retrieval.retrieval_methods import RetrievalMethod
 from core.tools.utils.text_processing_utils import remove_leading_symbols
+from core.workflow.nodes.llm import llm_utils
+from extensions.ext_database import db
+from factories.file_factory import build_from_mapping
 from libs import helper
+from models import UploadFile
 from models.account import Account
-from models.dataset import Dataset, DatasetProcessRule
+from models.dataset import Dataset, DatasetProcessRule, DocumentSegment, SegmentAttachmentBinding
 from models.dataset import Document as DatasetDocument
 from services.account_service import AccountService
 from services.entities.knowledge_entities.knowledge_entities import Rule
+from services.summary_index_service import SummaryIndexService
 
 
 class ParagraphIndexProcessor(BaseIndexProcessor):
@@ -108,6 +131,29 @@ class ParagraphIndexProcessor(BaseIndexProcessor):
                 keyword.add_texts(documents)
 
     def clean(self, dataset: Dataset, node_ids: list[str] | None, with_keywords: bool = True, **kwargs):
+        # Note: Summary indexes are now disabled (not deleted) when segments are disabled.
+        # This method is called for actual deletion scenarios (e.g., when segment is deleted).
+        # For disable operations, disable_summaries_for_segments is called directly in the task.
+        # Only delete summaries if explicitly requested (e.g., when segment is actually deleted)
+        delete_summaries = kwargs.get("delete_summaries", False)
+        if delete_summaries:
+            if node_ids:
+                # Find segments by index_node_id
+                segments = (
+                    db.session.query(DocumentSegment)
+                    .filter(
+                        DocumentSegment.dataset_id == dataset.id,
+                        DocumentSegment.index_node_id.in_(node_ids),
+                    )
+                    .all()
+                )
+                segment_ids = [segment.id for segment in segments]
+                if segment_ids:
+                    SummaryIndexService.delete_summaries_for_segments(dataset, segment_ids)
+            else:
+                # Delete all summaries for the dataset
+                SummaryIndexService.delete_summaries_for_segments(dataset, None)
+
         if dataset.indexing_technique == "high_quality":
             vector = Vector(dataset)
             if node_ids:
@@ -227,3 +273,322 @@ class ParagraphIndexProcessor(BaseIndexProcessor):
             }
         else:
             raise ValueError("Chunks is not a list")
+
+    def generate_summary_preview(
+        self, tenant_id: str, preview_texts: list[PreviewDetail], summary_index_setting: dict
+    ) -> list[PreviewDetail]:
+        """
+        For each segment, concurrently call generate_summary to generate a summary
+        and write it to the summary attribute of PreviewDetail.
+        In preview mode (indexing-estimate), if any summary generation fails, the method will raise an exception.
+        """
+        import concurrent.futures
+
+        from flask import current_app
+
+        # Capture Flask app context for worker threads
+        flask_app = None
+        try:
+            flask_app = current_app._get_current_object()  # type: ignore
+        except RuntimeError:
+            logger.warning("No Flask application context available, summary generation may fail")
+
+        def process(preview: PreviewDetail) -> None:
+            """Generate summary for a single preview item."""
+            if flask_app:
+                # Ensure Flask app context in worker thread
+                with flask_app.app_context():
+                    summary, _ = self.generate_summary(tenant_id, preview.content, summary_index_setting)
+                    preview.summary = summary
+            else:
+                # Fallback: try without app context (may fail)
+                summary, _ = self.generate_summary(tenant_id, preview.content, summary_index_setting)
+                preview.summary = summary
+
+        # Generate summaries concurrently using ThreadPoolExecutor
+        # Set a reasonable timeout to prevent hanging (60 seconds per chunk, max 5 minutes total)
+        timeout_seconds = min(300, 60 * len(preview_texts))
+        errors: list[Exception] = []
+
+        with concurrent.futures.ThreadPoolExecutor(max_workers=min(10, len(preview_texts))) as executor:
+            futures = [executor.submit(process, preview) for preview in preview_texts]
+            # Wait for all tasks to complete with timeout
+            done, not_done = concurrent.futures.wait(futures, timeout=timeout_seconds)
+
+            # Cancel tasks that didn't complete in time
+            if not_done:
+                timeout_error_msg = (
+                    f"Summary generation timeout: {len(not_done)} chunks did not complete within {timeout_seconds}s"
+                )
+                logger.warning("%s. Cancelling remaining tasks...", timeout_error_msg)
+                # In preview mode, timeout is also an error
+                errors.append(TimeoutError(timeout_error_msg))
+                for future in not_done:
+                    future.cancel()
+                # Wait a bit for cancellation to take effect
+                concurrent.futures.wait(not_done, timeout=5)
+
+            # Collect exceptions from completed futures
+            for future in done:
+                try:
+                    future.result()  # This will raise any exception that occurred
+                except Exception as e:
+                    logger.exception("Error in summary generation future")
+                    errors.append(e)
+
+        # In preview mode (indexing-estimate), if there are any errors, fail the request
+        if errors:
+            error_messages = [str(e) for e in errors]
+            error_summary = (
+                f"Failed to generate summaries for {len(errors)} chunk(s). "
+                f"Errors: {'; '.join(error_messages[:3])}"  # Show first 3 errors
+            )
+            if len(errors) > 3:
+                error_summary += f" (and {len(errors) - 3} more)"
+            logger.error("Summary generation failed in preview mode: %s", error_summary)
+            raise ValueError(error_summary)
+
+        return preview_texts
+
+    @staticmethod
+    def generate_summary(
+        tenant_id: str,
+        text: str,
+        summary_index_setting: dict | None = None,
+        segment_id: str | None = None,
+    ) -> tuple[str, LLMUsage]:
+        """
+        Generate summary for the given text using ModelInstance.invoke_llm and the default or custom summary prompt,
+        and supports vision models by including images from the segment attachments or text content.
+
+        Args:
+            tenant_id: Tenant ID
+            text: Text content to summarize
+            summary_index_setting: Summary index configuration
+            segment_id: Optional segment ID to fetch attachments from SegmentAttachmentBinding table
+
+        Returns:
+            Tuple of (summary_content, llm_usage) where llm_usage is LLMUsage object
+        """
+        if not summary_index_setting or not summary_index_setting.get("enable"):
+            raise ValueError("summary_index_setting is required and must be enabled to generate summary.")
+
+        model_name = summary_index_setting.get("model_name")
+        model_provider_name = summary_index_setting.get("model_provider_name")
+        summary_prompt = summary_index_setting.get("summary_prompt")
+
+        if not model_name or not model_provider_name:
+            raise ValueError("model_name and model_provider_name are required in summary_index_setting")
+
+        # Import default summary prompt
+        if not summary_prompt:
+            summary_prompt = DEFAULT_GENERATOR_SUMMARY_PROMPT
+
+        provider_manager = ProviderManager()
+        provider_model_bundle = provider_manager.get_provider_model_bundle(
+            tenant_id, model_provider_name, ModelType.LLM
+        )
+        model_instance = ModelInstance(provider_model_bundle, model_name)
+
+        # Get model schema to check if vision is supported
+        model_schema = model_instance.model_type_instance.get_model_schema(model_name, model_instance.credentials)
+        supports_vision = model_schema and model_schema.features and ModelFeature.VISION in model_schema.features
+
+        # Extract images if model supports vision
+        image_files = []
+        if supports_vision:
+            # First, try to get images from SegmentAttachmentBinding (preferred method)
+            if segment_id:
+                image_files = ParagraphIndexProcessor._extract_images_from_segment_attachments(tenant_id, segment_id)
+
+            # If no images from attachments, fall back to extracting from text
+            if not image_files:
+                image_files = ParagraphIndexProcessor._extract_images_from_text(tenant_id, text)
+
+        # Build prompt messages
+        prompt_messages = []
+
+        if image_files:
+            # If we have images, create a UserPromptMessage with both text and images
+            prompt_message_contents: list[PromptMessageContentUnionTypes] = []
+
+            # Add images first
+            for file in image_files:
+                try:
+                    file_content = file_manager.to_prompt_message_content(
+                        file, image_detail_config=ImagePromptMessageContent.DETAIL.LOW
+                    )
+                    prompt_message_contents.append(file_content)
+                except Exception as e:
+                    logger.warning("Failed to convert image file to prompt message content: %s", str(e))
+                    continue
+
+            # Add text content
+            if prompt_message_contents:  # Only add text if we successfully added images
+                prompt_message_contents.append(TextPromptMessageContent(data=f"{summary_prompt}\n{text}"))
+                prompt_messages.append(UserPromptMessage(content=prompt_message_contents))
+            else:
+                # If image conversion failed, fall back to text-only
+                prompt = f"{summary_prompt}\n{text}"
+                prompt_messages.append(UserPromptMessage(content=prompt))
+        else:
+            # No images, use simple text prompt
+            prompt = f"{summary_prompt}\n{text}"
+            prompt_messages.append(UserPromptMessage(content=prompt))
+
+        result = model_instance.invoke_llm(
+            prompt_messages=cast(list[PromptMessage], prompt_messages), model_parameters={}, stream=False
+        )
+
+        # Type assertion: when stream=False, invoke_llm returns LLMResult, not Generator
+        if not isinstance(result, LLMResult):
+            raise ValueError("Expected LLMResult when stream=False")
+
+        summary_content = getattr(result.message, "content", "")
+        usage = result.usage
+
+        # Deduct quota for summary generation (same as workflow nodes)
+        try:
+            llm_utils.deduct_llm_quota(tenant_id=tenant_id, model_instance=model_instance, usage=usage)
+        except Exception as e:
+            # Log but don't fail summary generation if quota deduction fails
+            logger.warning("Failed to deduct quota for summary generation: %s", str(e))
+
+        return summary_content, usage
+
+    @staticmethod
+    def _extract_images_from_text(tenant_id: str, text: str) -> list[File]:
+        """
+        Extract images from markdown text and convert them to File objects.
+
+        Args:
+            tenant_id: Tenant ID
+            text: Text content that may contain markdown image links
+
+        Returns:
+            List of File objects representing images found in the text
+        """
+        # Extract markdown images using regex pattern
+        pattern = r"!\[.*?\]\((.*?)\)"
+        images = re.findall(pattern, text)
+
+        if not images:
+            return []
+
+        upload_file_id_list = []
+
+        for image in images:
+            # For data before v0.10.0
+            pattern = r"/files/([a-f0-9\-]+)/image-preview(?:\?.*?)?"
+            match = re.search(pattern, image)
+            if match:
+                upload_file_id = match.group(1)
+                upload_file_id_list.append(upload_file_id)
+                continue
+
+            # For data after v0.10.0
+            pattern = r"/files/([a-f0-9\-]+)/file-preview(?:\?.*?)?"
+            match = re.search(pattern, image)
+            if match:
+                upload_file_id = match.group(1)
+                upload_file_id_list.append(upload_file_id)
+                continue
+
+            # For tools directory - direct file formats (e.g., .png, .jpg, etc.)
+            pattern = r"/files/tools/([a-f0-9\-]+)\.([a-zA-Z0-9]+)(?:\?[^\s\)\"\']*)?"
+            match = re.search(pattern, image)
+            if match:
+                # Tool files are handled differently, skip for now
+                continue
+
+        if not upload_file_id_list:
+            return []
+
+        # Get unique IDs for database query
+        unique_upload_file_ids = list(set(upload_file_id_list))
+        upload_files = (
+            db.session.query(UploadFile)
+            .where(UploadFile.id.in_(unique_upload_file_ids), UploadFile.tenant_id == tenant_id)
+            .all()
+        )
+
+        # Create File objects from UploadFile records
+        file_objects = []
+        for upload_file in upload_files:
+            # Only process image files
+            if not upload_file.mime_type or "image" not in upload_file.mime_type:
+                continue
+
+            mapping = {
+                "upload_file_id": upload_file.id,
+                "transfer_method": FileTransferMethod.LOCAL_FILE.value,
+                "type": FileType.IMAGE.value,
+            }
+
+            try:
+                file_obj = build_from_mapping(
+                    mapping=mapping,
+                    tenant_id=tenant_id,
+                )
+                file_objects.append(file_obj)
+            except Exception as e:
+                logger.warning("Failed to create File object from UploadFile %s: %s", upload_file.id, str(e))
+                continue
+
+        return file_objects
+
+    @staticmethod
+    def _extract_images_from_segment_attachments(tenant_id: str, segment_id: str) -> list[File]:
+        """
+        Extract images from SegmentAttachmentBinding table (preferred method).
+        This matches how DatasetRetrieval gets segment attachments.
+
+        Args:
+            tenant_id: Tenant ID
+            segment_id: Segment ID to fetch attachments for
+
+        Returns:
+            List of File objects representing images found in segment attachments
+        """
+        from sqlalchemy import select
+
+        # Query attachments from SegmentAttachmentBinding table
+        attachments_with_bindings = db.session.execute(
+            select(SegmentAttachmentBinding, UploadFile)
+            .join(UploadFile, UploadFile.id == SegmentAttachmentBinding.attachment_id)
+            .where(
+                SegmentAttachmentBinding.segment_id == segment_id,
+                SegmentAttachmentBinding.tenant_id == tenant_id,
+            )
+        ).all()
+
+        if not attachments_with_bindings:
+            return []
+
+        file_objects = []
+        for _, upload_file in attachments_with_bindings:
+            # Only process image files
+            if not upload_file.mime_type or "image" not in upload_file.mime_type:
+                continue
+
+            try:
+                # Create File object directly (similar to DatasetRetrieval)
+                file_obj = File(
+                    id=upload_file.id,
+                    filename=upload_file.name,
+                    extension="." + upload_file.extension,
+                    mime_type=upload_file.mime_type,
+                    tenant_id=tenant_id,
+                    type=FileType.IMAGE,
+                    transfer_method=FileTransferMethod.LOCAL_FILE,
+                    remote_url=upload_file.source_url,
+                    related_id=upload_file.id,
+                    size=upload_file.size,
+                    storage_key=upload_file.key,
+                )
+                file_objects.append(file_obj)
+            except Exception as e:
+                logger.warning("Failed to create File object from UploadFile %s: %s", upload_file.id, str(e))
+                continue
+
+        return file_objects

api/core/rag/index_processor/processor/parent_child_index_processor.py

@@ -1,11 +1,14 @@
 """Paragraph index processor."""
 
 import json
+import logging
 import uuid
 from collections.abc import Mapping
 from typing import Any
 
 from configs import dify_config
+from core.db.session_factory import session_factory
+from core.entities.knowledge_entities import PreviewDetail
 from core.model_manager import ModelInstance
 from core.rag.cleaner.clean_processor import CleanProcessor
 from core.rag.datasource.retrieval_service import RetrievalService
@@ -25,6 +28,9 @@ from models.dataset import ChildChunk, Dataset, DatasetProcessRule, DocumentSegm
 from models.dataset import Document as DatasetDocument
 from services.account_service import AccountService
 from services.entities.knowledge_entities.knowledge_entities import ParentMode, Rule
+from services.summary_index_service import SummaryIndexService
+
+logger = logging.getLogger(__name__)
 
 
 class ParentChildIndexProcessor(BaseIndexProcessor):
@@ -135,6 +141,30 @@ class ParentChildIndexProcessor(BaseIndexProcessor):
 
     def clean(self, dataset: Dataset, node_ids: list[str] | None, with_keywords: bool = True, **kwargs):
         # node_ids is segment's node_ids
+        # Note: Summary indexes are now disabled (not deleted) when segments are disabled.
+        # This method is called for actual deletion scenarios (e.g., when segment is deleted).
+        # For disable operations, disable_summaries_for_segments is called directly in the task.
+        # Only delete summaries if explicitly requested (e.g., when segment is actually deleted)
+        delete_summaries = kwargs.get("delete_summaries", False)
+        if delete_summaries:
+            if node_ids:
+                # Find segments by index_node_id
+                with session_factory.create_session() as session:
+                    segments = (
+                        session.query(DocumentSegment)
+                        .filter(
+                            DocumentSegment.dataset_id == dataset.id,
+                            DocumentSegment.index_node_id.in_(node_ids),
+                        )
+                        .all()
+                    )
+                    segment_ids = [segment.id for segment in segments]
+                    if segment_ids:
+                        SummaryIndexService.delete_summaries_for_segments(dataset, segment_ids)
+            else:
+                # Delete all summaries for the dataset
+                SummaryIndexService.delete_summaries_for_segments(dataset, None)
+
         if dataset.indexing_technique == "high_quality":
             delete_child_chunks = kwargs.get("delete_child_chunks") or False
             precomputed_child_node_ids = kwargs.get("precomputed_child_node_ids")
@@ -326,3 +356,91 @@ class ParentChildIndexProcessor(BaseIndexProcessor):
             "preview": preview,
             "total_segments": len(parent_childs.parent_child_chunks),
         }
+
+    def generate_summary_preview(
+        self, tenant_id: str, preview_texts: list[PreviewDetail], summary_index_setting: dict
+    ) -> list[PreviewDetail]:
+        """
+        For each parent chunk in preview_texts, concurrently call generate_summary to generate a summary
+        and write it to the summary attribute of PreviewDetail.
+        In preview mode (indexing-estimate), if any summary generation fails, the method will raise an exception.
+
+        Note: For parent-child structure, we only generate summaries for parent chunks.
+        """
+        import concurrent.futures
+
+        from flask import current_app
+
+        # Capture Flask app context for worker threads
+        flask_app = None
+        try:
+            flask_app = current_app._get_current_object()  # type: ignore
+        except RuntimeError:
+            logger.warning("No Flask application context available, summary generation may fail")
+
+        def process(preview: PreviewDetail) -> None:
+            """Generate summary for a single preview item (parent chunk)."""
+            from core.rag.index_processor.processor.paragraph_index_processor import ParagraphIndexProcessor
+
+            if flask_app:
+                # Ensure Flask app context in worker thread
+                with flask_app.app_context():
+                    summary, _ = ParagraphIndexProcessor.generate_summary(
+                        tenant_id=tenant_id,
+                        text=preview.content,
+                        summary_index_setting=summary_index_setting,
+                    )
+                    preview.summary = summary
+            else:
+                # Fallback: try without app context (may fail)
+                summary, _ = ParagraphIndexProcessor.generate_summary(
+                    tenant_id=tenant_id,
+                    text=preview.content,
+                    summary_index_setting=summary_index_setting,
+                )
+                preview.summary = summary
+
+        # Generate summaries concurrently using ThreadPoolExecutor
+        # Set a reasonable timeout to prevent hanging (60 seconds per chunk, max 5 minutes total)
+        timeout_seconds = min(300, 60 * len(preview_texts))
+        errors: list[Exception] = []
+
+        with concurrent.futures.ThreadPoolExecutor(max_workers=min(10, len(preview_texts))) as executor:
+            futures = [executor.submit(process, preview) for preview in preview_texts]
+            # Wait for all tasks to complete with timeout
+            done, not_done = concurrent.futures.wait(futures, timeout=timeout_seconds)
+
+            # Cancel tasks that didn't complete in time
+            if not_done:
+                timeout_error_msg = (
+                    f"Summary generation timeout: {len(not_done)} chunks did not complete within {timeout_seconds}s"
+                )
+                logger.warning("%s. Cancelling remaining tasks...", timeout_error_msg)
+                # In preview mode, timeout is also an error
+                errors.append(TimeoutError(timeout_error_msg))
+                for future in not_done:
+                    future.cancel()
+                # Wait a bit for cancellation to take effect
+                concurrent.futures.wait(not_done, timeout=5)
+
+            # Collect exceptions from completed futures
+            for future in done:
+                try:
+                    future.result()  # This will raise any exception that occurred
+                except Exception as e:
+                    logger.exception("Error in summary generation future")
+                    errors.append(e)
+
+        # In preview mode (indexing-estimate), if there are any errors, fail the request
+        if errors:
+            error_messages = [str(e) for e in errors]
+            error_summary = (
+                f"Failed to generate summaries for {len(errors)} chunk(s). "
+                f"Errors: {'; '.join(error_messages[:3])}"  # Show first 3 errors
+            )
+            if len(errors) > 3:
+                error_summary += f" (and {len(errors) - 3} more)"
+            logger.error("Summary generation failed in preview mode: %s", error_summary)
+            raise ValueError(error_summary)
+
+        return preview_texts

api/core/rag/index_processor/processor/qa_index_processor.py

@@ -11,6 +11,8 @@ import pandas as pd
 from flask import Flask, current_app
 from werkzeug.datastructures import FileStorage
 
+from core.db.session_factory import session_factory
+from core.entities.knowledge_entities import PreviewDetail
 from core.llm_generator.llm_generator import LLMGenerator
 from core.rag.cleaner.clean_processor import CleanProcessor
 from core.rag.datasource.retrieval_service import RetrievalService
@@ -25,9 +27,10 @@ from core.rag.retrieval.retrieval_methods import RetrievalMethod
 from core.tools.utils.text_processing_utils import remove_leading_symbols
 from libs import helper
 from models.account import Account
-from models.dataset import Dataset
+from models.dataset import Dataset, DocumentSegment
 from models.dataset import Document as DatasetDocument
 from services.entities.knowledge_entities.knowledge_entities import Rule
+from services.summary_index_service import SummaryIndexService
 
 logger = logging.getLogger(__name__)
 
@@ -144,6 +147,31 @@ class QAIndexProcessor(BaseIndexProcessor):
                 vector.create_multimodal(multimodal_documents)
 
     def clean(self, dataset: Dataset, node_ids: list[str] | None, with_keywords: bool = True, **kwargs):
+        # Note: Summary indexes are now disabled (not deleted) when segments are disabled.
+        # This method is called for actual deletion scenarios (e.g., when segment is deleted).
+        # For disable operations, disable_summaries_for_segments is called directly in the task.
+        # Note: qa_model doesn't generate summaries, but we clean them for completeness
+        # Only delete summaries if explicitly requested (e.g., when segment is actually deleted)
+        delete_summaries = kwargs.get("delete_summaries", False)
+        if delete_summaries:
+            if node_ids:
+                # Find segments by index_node_id
+                with session_factory.create_session() as session:
+                    segments = (
+                        session.query(DocumentSegment)
+                        .filter(
+                            DocumentSegment.dataset_id == dataset.id,
+                            DocumentSegment.index_node_id.in_(node_ids),
+                        )
+                        .all()
+                    )
+                    segment_ids = [segment.id for segment in segments]
+                    if segment_ids:
+                        SummaryIndexService.delete_summaries_for_segments(dataset, segment_ids)
+            else:
+                # Delete all summaries for the dataset
+                SummaryIndexService.delete_summaries_for_segments(dataset, None)
+
         vector = Vector(dataset)
         if node_ids:
             vector.delete_by_ids(node_ids)
@@ -212,6 +240,17 @@ class QAIndexProcessor(BaseIndexProcessor):
             "total_segments": len(qa_chunks.qa_chunks),
         }
 
+    def generate_summary_preview(
+        self, tenant_id: str, preview_texts: list[PreviewDetail], summary_index_setting: dict
+    ) -> list[PreviewDetail]:
+        """
+        QA model doesn't generate summaries, so this method returns preview_texts unchanged.
+
+        Note: QA model uses question-answer pairs, which don't require summary generation.
+        """
+        # QA model doesn't generate summaries, return as-is
+        return preview_texts
+
     def _format_qa_document(self, flask_app: Flask, tenant_id: str, document_node, all_qa_documents, document_language):
         format_documents = []
         if document_node.page_content is None or not document_node.page_content.strip():

api/core/rag/retrieval/dataset_retrieval.py

@@ -236,20 +236,24 @@ class DatasetRetrieval:
             if records:
                 for record in records:
                     segment = record.segment
+                    # Build content: if summary exists, add it before the segment content
                     if segment.answer:
-                        document_context_list.append(
-                            DocumentContext(
-                                content=f"question:{segment.get_sign_content()} answer:{segment.answer}",
-                                score=record.score,
-                            )
-                        )
+                        segment_content = f"question:{segment.get_sign_content()} answer:{segment.answer}"
                     else:
-                        document_context_list.append(
-                            DocumentContext(
-                                content=segment.get_sign_content(),
-                                score=record.score,
-                            )
+                        segment_content = segment.get_sign_content()
+
+                    # If summary exists, prepend it to the content
+                    if record.summary:
+                        final_content = f"{record.summary}\n{segment_content}"
+                    else:
+                        final_content = segment_content
+
+                    document_context_list.append(
+                        DocumentContext(
+                            content=final_content,
+                            score=record.score,
                         )
+                    )
                     if vision_enabled:
                         attachments_with_bindings = db.session.execute(
                             select(SegmentAttachmentBinding, UploadFile)
@@ -316,6 +320,9 @@ class DatasetRetrieval:
                                 source.content = f"question:{segment.content} \nanswer:{segment.answer}"
                             else:
                                 source.content = segment.content
+                            # Add summary if this segment was retrieved via summary
+                            if hasattr(record, "summary") and record.summary:
+                                source.summary = record.summary
                             retrieval_resource_list.append(source)
         if hit_callback and retrieval_resource_list:
             retrieval_resource_list = sorted(retrieval_resource_list, key=lambda x: x.score or 0.0, reverse=True)

api/core/repositories/__init__.py

@@ -1,19 +1,18 @@
-"""
-Repository implementations for data access.
+"""Repository implementations for data access."""
 
-This package contains concrete implementations of the repository interfaces
-defined in the core.workflow.repository package.
-"""
+from __future__ import annotations
 
-from core.repositories.celery_workflow_execution_repository import CeleryWorkflowExecutionRepository
-from core.repositories.celery_workflow_node_execution_repository import CeleryWorkflowNodeExecutionRepository
-from core.repositories.factory import DifyCoreRepositoryFactory, RepositoryImportError
-from core.repositories.sqlalchemy_workflow_node_execution_repository import SQLAlchemyWorkflowNodeExecutionRepository
+from .celery_workflow_execution_repository import CeleryWorkflowExecutionRepository
+from .celery_workflow_node_execution_repository import CeleryWorkflowNodeExecutionRepository
+from .factory import DifyCoreRepositoryFactory, RepositoryImportError
+from .sqlalchemy_workflow_execution_repository import SQLAlchemyWorkflowExecutionRepository
+from .sqlalchemy_workflow_node_execution_repository import SQLAlchemyWorkflowNodeExecutionRepository
 
 __all__ = [
     "CeleryWorkflowExecutionRepository",
     "CeleryWorkflowNodeExecutionRepository",
     "DifyCoreRepositoryFactory",
     "RepositoryImportError",
+    "SQLAlchemyWorkflowExecutionRepository",
     "SQLAlchemyWorkflowNodeExecutionRepository",
 ]

api/core/repositories/human_input_repository.py

@@ -0,0 +1,553 @@
+import dataclasses
+import json
+from collections.abc import Mapping, Sequence
+from datetime import datetime
+from typing import Any
+
+from sqlalchemy import Engine, select
+from sqlalchemy.orm import Session, selectinload, sessionmaker
+
+from core.workflow.nodes.human_input.entities import (
+    DeliveryChannelConfig,
+    EmailDeliveryMethod,
+    EmailRecipients,
+    ExternalRecipient,
+    FormDefinition,
+    HumanInputNodeData,
+    MemberRecipient,
+    WebAppDeliveryMethod,
+)
+from core.workflow.nodes.human_input.enums import (
+    DeliveryMethodType,
+    HumanInputFormKind,
+    HumanInputFormStatus,
+)
+from core.workflow.repositories.human_input_form_repository import (
+    FormCreateParams,
+    FormNotFoundError,
+    HumanInputFormEntity,
+    HumanInputFormRecipientEntity,
+)
+from libs.datetime_utils import naive_utc_now
+from libs.uuid_utils import uuidv7
+from models.account import Account, TenantAccountJoin
+from models.human_input import (
+    BackstageRecipientPayload,
+    ConsoleDeliveryPayload,
+    ConsoleRecipientPayload,
+    EmailExternalRecipientPayload,
+    EmailMemberRecipientPayload,
+    HumanInputDelivery,
+    HumanInputForm,
+    HumanInputFormRecipient,
+    RecipientType,
+    StandaloneWebAppRecipientPayload,
+)
+
+
+@dataclasses.dataclass(frozen=True)
+class _DeliveryAndRecipients:
+    delivery: HumanInputDelivery
+    recipients: Sequence[HumanInputFormRecipient]
+
+
+@dataclasses.dataclass(frozen=True)
+class _WorkspaceMemberInfo:
+    user_id: str
+    email: str
+
+
+class _HumanInputFormRecipientEntityImpl(HumanInputFormRecipientEntity):
+    def __init__(self, recipient_model: HumanInputFormRecipient):
+        self._recipient_model = recipient_model
+
+    @property
+    def id(self) -> str:
+        return self._recipient_model.id
+
+    @property
+    def token(self) -> str:
+        if self._recipient_model.access_token is None:
+            raise AssertionError(f"access_token should not be None for recipient {self._recipient_model.id}")
+        return self._recipient_model.access_token
+
+
+class _HumanInputFormEntityImpl(HumanInputFormEntity):
+    def __init__(self, form_model: HumanInputForm, recipient_models: Sequence[HumanInputFormRecipient]):
+        self._form_model = form_model
+        self._recipients = [_HumanInputFormRecipientEntityImpl(recipient) for recipient in recipient_models]
+        self._web_app_recipient = next(
+            (
+                recipient
+                for recipient in recipient_models
+                if recipient.recipient_type == RecipientType.STANDALONE_WEB_APP
+            ),
+            None,
+        )
+        self._console_recipient = next(
+            (recipient for recipient in recipient_models if recipient.recipient_type == RecipientType.CONSOLE),
+            None,
+        )
+        self._submitted_data: Mapping[str, Any] | None = (
+            json.loads(form_model.submitted_data) if form_model.submitted_data is not None else None
+        )
+
+    @property
+    def id(self) -> str:
+        return self._form_model.id
+
+    @property
+    def web_app_token(self):
+        if self._console_recipient is not None:
+            return self._console_recipient.access_token
+        if self._web_app_recipient is None:
+            return None
+        return self._web_app_recipient.access_token
+
+    @property
+    def recipients(self) -> list[HumanInputFormRecipientEntity]:
+        return list(self._recipients)
+
+    @property
+    def rendered_content(self) -> str:
+        return self._form_model.rendered_content
+
+    @property
+    def selected_action_id(self) -> str | None:
+        return self._form_model.selected_action_id
+
+    @property
+    def submitted_data(self) -> Mapping[str, Any] | None:
+        return self._submitted_data
+
+    @property
+    def submitted(self) -> bool:
+        return self._form_model.submitted_at is not None
+
+    @property
+    def status(self) -> HumanInputFormStatus:
+        return self._form_model.status
+
+    @property
+    def expiration_time(self) -> datetime:
+        return self._form_model.expiration_time
+
+
+@dataclasses.dataclass(frozen=True)
+class HumanInputFormRecord:
+    form_id: str
+    workflow_run_id: str | None
+    node_id: str
+    tenant_id: str
+    app_id: str
+    form_kind: HumanInputFormKind
+    definition: FormDefinition
+    rendered_content: str
+    created_at: datetime
+    expiration_time: datetime
+    status: HumanInputFormStatus
+    selected_action_id: str | None
+    submitted_data: Mapping[str, Any] | None
+    submitted_at: datetime | None
+    submission_user_id: str | None
+    submission_end_user_id: str | None
+    completed_by_recipient_id: str | None
+    recipient_id: str | None
+    recipient_type: RecipientType | None
+    access_token: str | None
+
+    @property
+    def submitted(self) -> bool:
+        return self.submitted_at is not None
+
+    @classmethod
+    def from_models(
+        cls, form_model: HumanInputForm, recipient_model: HumanInputFormRecipient | None
+    ) -> "HumanInputFormRecord":
+        definition_payload = json.loads(form_model.form_definition)
+        if "expiration_time" not in definition_payload:
+            definition_payload["expiration_time"] = form_model.expiration_time
+        return cls(
+            form_id=form_model.id,
+            workflow_run_id=form_model.workflow_run_id,
+            node_id=form_model.node_id,
+            tenant_id=form_model.tenant_id,
+            app_id=form_model.app_id,
+            form_kind=form_model.form_kind,
+            definition=FormDefinition.model_validate(definition_payload),
+            rendered_content=form_model.rendered_content,
+            created_at=form_model.created_at,
+            expiration_time=form_model.expiration_time,
+            status=form_model.status,
+            selected_action_id=form_model.selected_action_id,
+            submitted_data=json.loads(form_model.submitted_data) if form_model.submitted_data else None,
+            submitted_at=form_model.submitted_at,
+            submission_user_id=form_model.submission_user_id,
+            submission_end_user_id=form_model.submission_end_user_id,
+            completed_by_recipient_id=form_model.completed_by_recipient_id,
+            recipient_id=recipient_model.id if recipient_model else None,
+            recipient_type=recipient_model.recipient_type if recipient_model else None,
+            access_token=recipient_model.access_token if recipient_model else None,
+        )
+
+
+class _InvalidTimeoutStatusError(ValueError):
+    pass
+
+
+class HumanInputFormRepositoryImpl:
+    def __init__(
+        self,
+        session_factory: sessionmaker | Engine,
+        tenant_id: str,
+    ):
+        if isinstance(session_factory, Engine):
+            session_factory = sessionmaker(bind=session_factory)
+        self._session_factory = session_factory
+        self._tenant_id = tenant_id
+
+    def _delivery_method_to_model(
+        self,
+        session: Session,
+        form_id: str,
+        delivery_method: DeliveryChannelConfig,
+    ) -> _DeliveryAndRecipients:
+        delivery_id = str(uuidv7())
+        delivery_model = HumanInputDelivery(
+            id=delivery_id,
+            form_id=form_id,
+            delivery_method_type=delivery_method.type,
+            delivery_config_id=delivery_method.id,
+            channel_payload=delivery_method.model_dump_json(),
+        )
+        recipients: list[HumanInputFormRecipient] = []
+        if isinstance(delivery_method, WebAppDeliveryMethod):
+            recipient_model = HumanInputFormRecipient(
+                form_id=form_id,
+                delivery_id=delivery_id,
+                recipient_type=RecipientType.STANDALONE_WEB_APP,
+                recipient_payload=StandaloneWebAppRecipientPayload().model_dump_json(),
+            )
+            recipients.append(recipient_model)
+        elif isinstance(delivery_method, EmailDeliveryMethod):
+            email_recipients_config = delivery_method.config.recipients
+            recipients.extend(
+                self._build_email_recipients(
+                    session=session,
+                    form_id=form_id,
+                    delivery_id=delivery_id,
+                    recipients_config=email_recipients_config,
+                )
+            )
+
+        return _DeliveryAndRecipients(delivery=delivery_model, recipients=recipients)
+
+    def _build_email_recipients(
+        self,
+        session: Session,
+        form_id: str,
+        delivery_id: str,
+        recipients_config: EmailRecipients,
+    ) -> list[HumanInputFormRecipient]:
+        member_user_ids = [
+            recipient.user_id for recipient in recipients_config.items if isinstance(recipient, MemberRecipient)
+        ]
+        external_emails = [
+            recipient.email for recipient in recipients_config.items if isinstance(recipient, ExternalRecipient)
+        ]
+        if recipients_config.whole_workspace:
+            members = self._query_all_workspace_members(session=session)
+        else:
+            members = self._query_workspace_members_by_ids(session=session, restrict_to_user_ids=member_user_ids)
+
+        return self._create_email_recipients_from_resolved(
+            form_id=form_id,
+            delivery_id=delivery_id,
+            members=members,
+            external_emails=external_emails,
+        )
+
+    @staticmethod
+    def _create_email_recipients_from_resolved(
+        *,
+        form_id: str,
+        delivery_id: str,
+        members: Sequence[_WorkspaceMemberInfo],
+        external_emails: Sequence[str],
+    ) -> list[HumanInputFormRecipient]:
+        recipient_models: list[HumanInputFormRecipient] = []
+        seen_emails: set[str] = set()
+
+        for member in members:
+            if not member.email:
+                continue
+            if member.email in seen_emails:
+                continue
+            seen_emails.add(member.email)
+            payload = EmailMemberRecipientPayload(user_id=member.user_id, email=member.email)
+            recipient_models.append(
+                HumanInputFormRecipient.new(
+                    form_id=form_id,
+                    delivery_id=delivery_id,
+                    payload=payload,
+                )
+            )
+
+        for email in external_emails:
+            if not email:
+                continue
+            if email in seen_emails:
+                continue
+            seen_emails.add(email)
+            recipient_models.append(
+                HumanInputFormRecipient.new(
+                    form_id=form_id,
+                    delivery_id=delivery_id,
+                    payload=EmailExternalRecipientPayload(email=email),
+                )
+            )
+
+        return recipient_models
+
+    def _query_all_workspace_members(
+        self,
+        session: Session,
+    ) -> list[_WorkspaceMemberInfo]:
+        stmt = (
+            select(Account.id, Account.email)
+            .join(TenantAccountJoin, TenantAccountJoin.account_id == Account.id)
+            .where(TenantAccountJoin.tenant_id == self._tenant_id)
+        )
+        rows = session.execute(stmt).all()
+        return [_WorkspaceMemberInfo(user_id=account_id, email=email) for account_id, email in rows]
+
+    def _query_workspace_members_by_ids(
+        self,
+        session: Session,
+        restrict_to_user_ids: Sequence[str],
+    ) -> list[_WorkspaceMemberInfo]:
+        unique_ids = {user_id for user_id in restrict_to_user_ids if user_id}
+        if not unique_ids:
+            return []
+
+        stmt = (
+            select(Account.id, Account.email)
+            .join(TenantAccountJoin, TenantAccountJoin.account_id == Account.id)
+            .where(TenantAccountJoin.tenant_id == self._tenant_id)
+        )
+        stmt = stmt.where(Account.id.in_(unique_ids))
+
+        rows = session.execute(stmt).all()
+        return [_WorkspaceMemberInfo(user_id=account_id, email=email) for account_id, email in rows]
+
+    def create_form(self, params: FormCreateParams) -> HumanInputFormEntity:
+        form_config: HumanInputNodeData = params.form_config
+
+        with self._session_factory(expire_on_commit=False) as session, session.begin():
+            # Generate unique form ID
+            form_id = str(uuidv7())
+            start_time = naive_utc_now()
+            node_expiration = form_config.expiration_time(start_time)
+            form_definition = FormDefinition(
+                form_content=form_config.form_content,
+                inputs=form_config.inputs,
+                user_actions=form_config.user_actions,
+                rendered_content=params.rendered_content,
+                expiration_time=node_expiration,
+                default_values=dict(params.resolved_default_values),
+                display_in_ui=params.display_in_ui,
+                node_title=form_config.title,
+            )
+            form_model = HumanInputForm(
+                id=form_id,
+                tenant_id=self._tenant_id,
+                app_id=params.app_id,
+                workflow_run_id=params.workflow_execution_id,
+                form_kind=params.form_kind,
+                node_id=params.node_id,
+                form_definition=form_definition.model_dump_json(),
+                rendered_content=params.rendered_content,
+                expiration_time=node_expiration,
+                created_at=start_time,
+            )
+            session.add(form_model)
+            recipient_models: list[HumanInputFormRecipient] = []
+            for delivery in params.delivery_methods:
+                delivery_and_recipients = self._delivery_method_to_model(
+                    session=session,
+                    form_id=form_id,
+                    delivery_method=delivery,
+                )
+                session.add(delivery_and_recipients.delivery)
+                session.add_all(delivery_and_recipients.recipients)
+                recipient_models.extend(delivery_and_recipients.recipients)
+            if params.console_recipient_required and not any(
+                recipient.recipient_type == RecipientType.CONSOLE for recipient in recipient_models
+            ):
+                console_delivery_id = str(uuidv7())
+                console_delivery = HumanInputDelivery(
+                    id=console_delivery_id,
+                    form_id=form_id,
+                    delivery_method_type=DeliveryMethodType.WEBAPP,
+                    delivery_config_id=None,
+                    channel_payload=ConsoleDeliveryPayload().model_dump_json(),
+                )
+                console_recipient = HumanInputFormRecipient(
+                    form_id=form_id,
+                    delivery_id=console_delivery_id,
+                    recipient_type=RecipientType.CONSOLE,
+                    recipient_payload=ConsoleRecipientPayload(
+                        account_id=params.console_creator_account_id,
+                    ).model_dump_json(),
+                )
+                session.add(console_delivery)
+                session.add(console_recipient)
+                recipient_models.append(console_recipient)
+            if params.backstage_recipient_required and not any(
+                recipient.recipient_type == RecipientType.BACKSTAGE for recipient in recipient_models
+            ):
+                backstage_delivery_id = str(uuidv7())
+                backstage_delivery = HumanInputDelivery(
+                    id=backstage_delivery_id,
+                    form_id=form_id,
+                    delivery_method_type=DeliveryMethodType.WEBAPP,
+                    delivery_config_id=None,
+                    channel_payload=ConsoleDeliveryPayload().model_dump_json(),
+                )
+                backstage_recipient = HumanInputFormRecipient(
+                    form_id=form_id,
+                    delivery_id=backstage_delivery_id,
+                    recipient_type=RecipientType.BACKSTAGE,
+                    recipient_payload=BackstageRecipientPayload(
+                        account_id=params.console_creator_account_id,
+                    ).model_dump_json(),
+                )
+                session.add(backstage_delivery)
+                session.add(backstage_recipient)
+                recipient_models.append(backstage_recipient)
+            session.flush()
+
+        return _HumanInputFormEntityImpl(form_model=form_model, recipient_models=recipient_models)
+
+    def get_form(self, workflow_execution_id: str, node_id: str) -> HumanInputFormEntity | None:
+        form_query = select(HumanInputForm).where(
+            HumanInputForm.workflow_run_id == workflow_execution_id,
+            HumanInputForm.node_id == node_id,
+            HumanInputForm.tenant_id == self._tenant_id,
+        )
+        with self._session_factory(expire_on_commit=False) as session:
+            form_model: HumanInputForm | None = session.scalars(form_query).first()
+            if form_model is None:
+                return None
+
+            recipient_query = select(HumanInputFormRecipient).where(HumanInputFormRecipient.form_id == form_model.id)
+            recipient_models = session.scalars(recipient_query).all()
+        return _HumanInputFormEntityImpl(form_model=form_model, recipient_models=recipient_models)
+
+
+class HumanInputFormSubmissionRepository:
+    """Repository for fetching and submitting human input forms."""
+
+    def __init__(self, session_factory: sessionmaker | Engine):
+        if isinstance(session_factory, Engine):
+            session_factory = sessionmaker(bind=session_factory)
+        self._session_factory = session_factory
+
+    def get_by_token(self, form_token: str) -> HumanInputFormRecord | None:
+        query = (
+            select(HumanInputFormRecipient)
+            .options(selectinload(HumanInputFormRecipient.form))
+            .where(HumanInputFormRecipient.access_token == form_token)
+        )
+        with self._session_factory(expire_on_commit=False) as session:
+            recipient_model = session.scalars(query).first()
+            if recipient_model is None or recipient_model.form is None:
+                return None
+            return HumanInputFormRecord.from_models(recipient_model.form, recipient_model)
+
+    def get_by_form_id_and_recipient_type(
+        self,
+        form_id: str,
+        recipient_type: RecipientType,
+    ) -> HumanInputFormRecord | None:
+        query = (
+            select(HumanInputFormRecipient)
+            .options(selectinload(HumanInputFormRecipient.form))
+            .where(
+                HumanInputFormRecipient.form_id == form_id,
+                HumanInputFormRecipient.recipient_type == recipient_type,
+            )
+        )
+        with self._session_factory(expire_on_commit=False) as session:
+            recipient_model = session.scalars(query).first()
+            if recipient_model is None or recipient_model.form is None:
+                return None
+            return HumanInputFormRecord.from_models(recipient_model.form, recipient_model)
+
+    def mark_submitted(
+        self,
+        *,
+        form_id: str,
+        recipient_id: str | None,
+        selected_action_id: str,
+        form_data: Mapping[str, Any],
+        submission_user_id: str | None,
+        submission_end_user_id: str | None,
+    ) -> HumanInputFormRecord:
+        with self._session_factory(expire_on_commit=False) as session, session.begin():
+            form_model = session.get(HumanInputForm, form_id)
+            if form_model is None:
+                raise FormNotFoundError(f"form not found, id={form_id}")
+
+            recipient_model = session.get(HumanInputFormRecipient, recipient_id) if recipient_id else None
+
+            form_model.selected_action_id = selected_action_id
+            form_model.submitted_data = json.dumps(form_data)
+            form_model.submitted_at = naive_utc_now()
+            form_model.status = HumanInputFormStatus.SUBMITTED
+            form_model.submission_user_id = submission_user_id
+            form_model.submission_end_user_id = submission_end_user_id
+            form_model.completed_by_recipient_id = recipient_id
+
+            session.add(form_model)
+            session.flush()
+            session.refresh(form_model)
+            if recipient_model is not None:
+                session.refresh(recipient_model)
+
+            return HumanInputFormRecord.from_models(form_model, recipient_model)
+
+    def mark_timeout(
+        self,
+        *,
+        form_id: str,
+        timeout_status: HumanInputFormStatus,
+        reason: str | None = None,
+    ) -> HumanInputFormRecord:
+        with self._session_factory(expire_on_commit=False) as session, session.begin():
+            form_model = session.get(HumanInputForm, form_id)
+            if form_model is None:
+                raise FormNotFoundError(f"form not found, id={form_id}")
+
+            if timeout_status not in {HumanInputFormStatus.TIMEOUT, HumanInputFormStatus.EXPIRED}:
+                raise _InvalidTimeoutStatusError(f"invalid timeout status: {timeout_status}")
+
+            # already handled or submitted
+            if form_model.status in {HumanInputFormStatus.TIMEOUT, HumanInputFormStatus.EXPIRED}:
+                return HumanInputFormRecord.from_models(form_model, None)
+
+            if form_model.submitted_at is not None or form_model.status == HumanInputFormStatus.SUBMITTED:
+                raise FormNotFoundError(f"form already submitted, id={form_id}")
+
+            form_model.status = timeout_status
+            form_model.selected_action_id = None
+            form_model.submitted_data = None
+            form_model.submission_user_id = None
+            form_model.submission_end_user_id = None
+            form_model.completed_by_recipient_id = None
+            # Reason is recorded in status/error downstream; not stored on form.
+            session.add(form_model)
+            session.flush()
+            session.refresh(form_model)
+
+            return HumanInputFormRecord.from_models(form_model, None)

api/core/repositories/sqlalchemy_workflow_node_execution_repository.py

@@ -488,6 +488,7 @@ class SQLAlchemyWorkflowNodeExecutionRepository(WorkflowNodeExecutionRepository)
                 WorkflowNodeExecutionModel.workflow_run_id == workflow_run_id,
                 WorkflowNodeExecutionModel.tenant_id == self._tenant_id,
                 WorkflowNodeExecutionModel.triggered_from == triggered_from,
+                WorkflowNodeExecutionModel.status != WorkflowNodeExecutionStatus.PAUSED,
             )
 
             if self._app_id:

api/core/tools/errors.py

@@ -1,4 +1,5 @@
 from core.tools.entities.tool_entities import ToolInvokeMeta
+from libs.exception import BaseHTTPException
 
 
 class ToolProviderNotFoundError(ValueError):
@@ -37,6 +38,12 @@ class ToolCredentialPolicyViolationError(ValueError):
     pass
 
 
+class WorkflowToolHumanInputNotSupportedError(BaseHTTPException):
+    error_code = "workflow_tool_human_input_not_supported"
+    description = "Workflow with Human Input nodes cannot be published as a workflow tool."
+    code = 400
+
+
 class ToolEngineInvokeError(Exception):
     meta: ToolInvokeMeta
 

api/core/tools/utils/dataset_retriever/dataset_retriever_tool.py

@@ -169,20 +169,24 @@ class DatasetRetrieverTool(DatasetRetrieverBaseTool):
                 if records:
                     for record in records:
                         segment = record.segment
+                        # Build content: if summary exists, add it before the segment content
                         if segment.answer:
-                            document_context_list.append(
-                                DocumentContext(
-                                    content=f"question:{segment.get_sign_content()} answer:{segment.answer}",
-                                    score=record.score,
-                                )
-                            )
+                            segment_content = f"question:{segment.get_sign_content()} answer:{segment.answer}"
                         else:
-                            document_context_list.append(
-                                DocumentContext(
-                                    content=segment.get_sign_content(),
-                                    score=record.score,
-                                )
+                            segment_content = segment.get_sign_content()
+
+                        # If summary exists, prepend it to the content
+                        if record.summary:
+                            final_content = f"{record.summary}\n{segment_content}"
+                        else:
+                            final_content = segment_content
+
+                        document_context_list.append(
+                            DocumentContext(
+                                content=final_content,
+                                score=record.score,
                             )
+                        )
 
                     if self.return_resource:
                         for record in records:
@@ -216,6 +220,9 @@ class DatasetRetrieverTool(DatasetRetrieverBaseTool):
                                     source.content = f"question:{segment.content} \nanswer:{segment.answer}"
                                 else:
                                     source.content = segment.content
+                                # Add summary if this segment was retrieved via summary
+                                if hasattr(record, "summary") and record.summary:
+                                    source.summary = record.summary
                                 retrieval_resource_list.append(source)
 
             if self.return_resource and retrieval_resource_list:

api/core/tools/utils/workflow_configuration_sync.py

@@ -3,6 +3,8 @@ from typing import Any
 
 from core.app.app_config.entities import VariableEntity
 from core.tools.entities.tool_entities import WorkflowToolParameterConfiguration
+from core.tools.errors import WorkflowToolHumanInputNotSupportedError
+from core.workflow.enums import NodeType
 from core.workflow.nodes.base.entities import OutputVariableEntity
 
 
@@ -50,6 +52,13 @@ class WorkflowToolConfigurationUtils:
 
         return [outputs_by_variable[variable] for variable in variable_order]
 
+    @classmethod
+    def ensure_no_human_input_nodes(cls, graph: Mapping[str, Any]) -> None:
+        nodes = graph.get("nodes", [])
+        for node in nodes:
+            if node.get("data", {}).get("type") == NodeType.HUMAN_INPUT:
+                raise WorkflowToolHumanInputNotSupportedError()
+
     @classmethod
     def check_is_synced(
         cls, variables: list[VariableEntity], tool_configurations: list[WorkflowToolParameterConfiguration]

api/core/tools/workflow_as_tool/tool.py

@@ -98,6 +98,10 @@ class WorkflowTool(Tool):
             invoke_from=self.runtime.invoke_from,
             streaming=False,
             call_depth=self.workflow_call_depth + 1,
+            # NOTE(QuantumGhost): We explicitly set `pause_state_config` to `None`
+            # because workflow pausing mechanisms (such as HumanInput) are not
+            # supported within WorkflowTool execution context.
+            pause_state_config=None,
         )
         assert isinstance(result, dict)
         data = result.get("data", {})

api/core/trigger/debug/event_bus.py

@@ -23,8 +23,8 @@ class TriggerDebugEventBus:
     """
 
     # LUA_SELECT: Atomic poll or register for event
-    # KEYS[1] = trigger_debug_inbox:{tenant_id}:{address_id}
-    # KEYS[2] = trigger_debug_waiting_pool:{tenant_id}:...
+    # KEYS[1] = trigger_debug_inbox:{<tenant_id>}:<address_id>
+    # KEYS[2] = trigger_debug_waiting_pool:{<tenant_id>}:...
     # ARGV[1] = address_id
     LUA_SELECT = (
         "local v=redis.call('GET',KEYS[1]);"
@@ -35,7 +35,7 @@ class TriggerDebugEventBus:
     )
 
     # LUA_DISPATCH: Dispatch event to all waiting addresses
-    # KEYS[1] = trigger_debug_waiting_pool:{tenant_id}:...
+    # KEYS[1] = trigger_debug_waiting_pool:{<tenant_id>}:...
     # ARGV[1] = tenant_id
     # ARGV[2] = event_json
     LUA_DISPATCH = (
@@ -43,7 +43,7 @@ class TriggerDebugEventBus:
         "if #a==0 then return 0 end;"
         "redis.call('DEL',KEYS[1]);"
         "for i=1,#a do "
-        f"redis.call('SET','trigger_debug_inbox:'..ARGV[1]..':'..a[i],ARGV[2],'EX',{TRIGGER_DEBUG_EVENT_TTL});"
+        f"redis.call('SET','trigger_debug_inbox:{{'..ARGV[1]..'}}'..':'..a[i],ARGV[2],'EX',{TRIGGER_DEBUG_EVENT_TTL});"
         "end;"
         "return #a"
     )
@@ -108,7 +108,7 @@ class TriggerDebugEventBus:
             Event object if available, None otherwise
         """
         address_id: str = hashlib.sha256(f"{user_id}|{app_id}|{node_id}".encode()).hexdigest()
-        address: str = f"trigger_debug_inbox:{tenant_id}:{address_id}"
+        address: str = f"trigger_debug_inbox:{{{tenant_id}}}:{address_id}"
 
         try:
             event_data = redis_client.eval(

api/core/trigger/debug/events.py

@@ -42,7 +42,7 @@ def build_webhook_pool_key(tenant_id: str, app_id: str, node_id: str) -> str:
         app_id: App ID
         node_id: Node ID
     """
-    return f"{TriggerDebugPoolKey.WEBHOOK}:{tenant_id}:{app_id}:{node_id}"
+    return f"{TriggerDebugPoolKey.WEBHOOK}:{{{tenant_id}}}:{app_id}:{node_id}"
 
 
 class PluginTriggerDebugEvent(BaseDebugEvent):
@@ -64,4 +64,4 @@ def build_plugin_pool_key(tenant_id: str, provider_id: str, subscription_id: str
         provider_id: Provider ID
         subscription_id: Subscription ID
     """
-    return f"{TriggerDebugPoolKey.PLUGIN}:{tenant_id}:{str(provider_id)}:{subscription_id}:{name}"
+    return f"{TriggerDebugPoolKey.PLUGIN}:{{{tenant_id}}}:{str(provider_id)}:{subscription_id}:{name}"

api/core/workflow/entities/__init__.py

@@ -2,10 +2,12 @@ from .agent import AgentNodeStrategyInit
 from .graph_init_params import GraphInitParams
 from .workflow_execution import WorkflowExecution
 from .workflow_node_execution import WorkflowNodeExecution
+from .workflow_start_reason import WorkflowStartReason
 
 __all__ = [
     "AgentNodeStrategyInit",
     "GraphInitParams",
     "WorkflowExecution",
     "WorkflowNodeExecution",
+    "WorkflowStartReason",
 ]

api/core/workflow/entities/graph_init_params.py

@@ -5,6 +5,16 @@ from pydantic import BaseModel, Field
 
 
 class GraphInitParams(BaseModel):
+    """GraphInitParams encapsulates the configurations and contextual information
+    that remain constant throughout a single execution of the graph engine.
+
+    A single execution is defined as follows: as long as the execution has not reached
+    its conclusion, it is considered one execution. For instance, if a workflow is suspended
+    and later resumed, it is still regarded as a single execution, not two.
+
+    For the state diagram of workflow execution, refer to `WorkflowExecutionStatus`.
+    """
+
     # init params
     tenant_id: str = Field(..., description="tenant / workspace id")
     app_id: str = Field(..., description="app id")

api/core/workflow/entities/pause_reason.py

@@ -1,8 +1,11 @@
+from collections.abc import Mapping
 from enum import StrEnum, auto
-from typing import Annotated, Literal, TypeAlias
+from typing import Annotated, Any, Literal, TypeAlias
 
 from pydantic import BaseModel, Field
 
+from core.workflow.nodes.human_input.entities import FormInput, UserAction
+
 
 class PauseReasonType(StrEnum):
     HUMAN_INPUT_REQUIRED = auto()
@@ -11,10 +14,31 @@ class PauseReasonType(StrEnum):
 
 class HumanInputRequired(BaseModel):
     TYPE: Literal[PauseReasonType.HUMAN_INPUT_REQUIRED] = PauseReasonType.HUMAN_INPUT_REQUIRED
-
     form_id: str
-    # The identifier of the human input node causing the pause.
+    form_content: str
+    inputs: list[FormInput] = Field(default_factory=list)
+    actions: list[UserAction] = Field(default_factory=list)
+    display_in_ui: bool = False
     node_id: str
+    node_title: str
+
+    # The `resolved_default_values` stores the resolved values of variable defaults. It's a mapping from
+    # `output_variable_name` to their resolved values.
+    #
+    # For example, The form contains a input with output variable name `name` and placeholder type `VARIABLE`, its
+    # selector is ["start", "name"]. While the HumanInputNode is executed, the correspond value of variable
+    # `start.name` in variable pool is `John`. Thus, the resolved value of the output variable `name` is `John`. The
+    # `resolved_default_values` is `{"name": "John"}`.
+    #
+    # Only form inputs with default value type `VARIABLE` will be resolved and stored in `resolved_default_values`.
+    resolved_default_values: Mapping[str, Any] = Field(default_factory=dict)
+
+    # The `form_token` is the token used to submit the form via UI surfaces. It corresponds to
+    # `HumanInputFormRecipient.access_token`.
+    #
+    # This field is `None` if webapp delivery is not set and not
+    # in orchestrating mode.
+    form_token: str | None = None
 
 
 class SchedulingPause(BaseModel):

api/core/workflow/entities/workflow_start_reason.py

@@ -0,0 +1,8 @@
+from enum import StrEnum
+
+
+class WorkflowStartReason(StrEnum):
+    """Reason for workflow start events across graph/queue/SSE layers."""
+
+    INITIAL = "initial"  # First start of a workflow run.
+    RESUMPTION = "resumption"  # Start triggered after resuming a paused run.

api/core/workflow/graph_engine/_engine_utils.py

@@ -0,0 +1,15 @@
+import time
+
+
+def get_timestamp() -> float:
+    """Retrieve a timestamp as a float point numer representing the number of seconds
+    since the Unix epoch.
+
+    This function is primarily used to measure the execution time of the workflow engine.
+    Since workflow execution may be paused and resumed on a different machine,
+    `time.perf_counter` cannot be used as it is inconsistent across machines.
+
+    To address this, the function uses the wall clock as the time source.
+    However, it assumes that the clocks of all servers are properly synchronized.
+    """
+    return round(time.time())

api/core/workflow/graph_engine/config.py

@@ -2,12 +2,14 @@
 GraphEngine configuration models.
 """
 
-from pydantic import BaseModel
+from pydantic import BaseModel, ConfigDict
 
 
 class GraphEngineConfig(BaseModel):
     """Configuration for GraphEngine worker pool scaling."""
 
+    model_config = ConfigDict(frozen=True)
+
     min_workers: int = 1
     max_workers: int = 5
     scale_up_threshold: int = 3

api/core/workflow/graph_engine/event_management/event_handlers.py

@@ -192,9 +192,13 @@ class EventHandler:
             self._event_collector.collect(edge_event)
 
         # Enqueue ready nodes
-        for node_id in ready_nodes:
-            self._state_manager.enqueue_node(node_id)
-            self._state_manager.start_execution(node_id)
+        if self._graph_execution.is_paused:
+            for node_id in ready_nodes:
+                self._graph_runtime_state.register_deferred_node(node_id)
+        else:
+            for node_id in ready_nodes:
+                self._state_manager.enqueue_node(node_id)
+                self._state_manager.start_execution(node_id)
 
         # Update execution tracking
         self._state_manager.finish_execution(event.node_id)

api/core/workflow/graph_engine/graph_engine.py

@@ -14,6 +14,7 @@ from collections.abc import Generator
 from typing import TYPE_CHECKING, cast, final
 
 from core.workflow.context import capture_current_context
+from core.workflow.entities.workflow_start_reason import WorkflowStartReason
 from core.workflow.enums import NodeExecutionType
 from core.workflow.graph import Graph
 from core.workflow.graph_events import (
@@ -56,6 +57,9 @@ if TYPE_CHECKING:
 logger = logging.getLogger(__name__)
 
 
+_DEFAULT_CONFIG = GraphEngineConfig()
+
+
 @final
 class GraphEngine:
     """
@@ -71,7 +75,7 @@ class GraphEngine:
         graph: Graph,
         graph_runtime_state: GraphRuntimeState,
         command_channel: CommandChannel,
-        config: GraphEngineConfig,
+        config: GraphEngineConfig = _DEFAULT_CONFIG,
     ) -> None:
         """Initialize the graph engine with all subsystems and dependencies."""
         # stop event
@@ -235,7 +239,9 @@ class GraphEngine:
                 self._graph_execution.paused = False
                 self._graph_execution.pause_reasons = []
 
-            start_event = GraphRunStartedEvent()
+            start_event = GraphRunStartedEvent(
+                reason=WorkflowStartReason.RESUMPTION if is_resume else WorkflowStartReason.INITIAL,
+            )
             self._event_manager.notify_layers(start_event)
             yield start_event
 
@@ -304,15 +310,17 @@ class GraphEngine:
         for layer in self._layers:
             try:
                 layer.on_graph_start()
-            except Exception as e:
-                logger.warning("Layer %s failed on_graph_start: %s", layer.__class__.__name__, e)
+            except Exception:
+                logger.exception("Layer %s failed on_graph_start", layer.__class__.__name__)
 
     def _start_execution(self, *, resume: bool = False) -> None:
         """Start execution subsystems."""
         self._stop_event.clear()
         paused_nodes: list[str] = []
+        deferred_nodes: list[str] = []
         if resume:
             paused_nodes = self._graph_runtime_state.consume_paused_nodes()
+            deferred_nodes = self._graph_runtime_state.consume_deferred_nodes()
 
         # Start worker pool (it calculates initial workers internally)
         self._worker_pool.start()
@@ -328,7 +336,11 @@ class GraphEngine:
             self._state_manager.enqueue_node(root_node.id)
             self._state_manager.start_execution(root_node.id)
         else:
-            for node_id in paused_nodes:
+            seen_nodes: set[str] = set()
+            for node_id in paused_nodes + deferred_nodes:
+                if node_id in seen_nodes:
+                    continue
+                seen_nodes.add(node_id)
                 self._state_manager.enqueue_node(node_id)
                 self._state_manager.start_execution(node_id)
 
@@ -346,8 +358,8 @@ class GraphEngine:
         for layer in self._layers:
             try:
                 layer.on_graph_end(self._graph_execution.error)
-            except Exception as e:
-                logger.warning("Layer %s failed on_graph_end: %s", layer.__class__.__name__, e)
+            except Exception:
+                logger.exception("Layer %s failed on_graph_end", layer.__class__.__name__)
 
     # Public property accessors for attributes that need external access
     @property

api/core/workflow/graph_engine/graph_state_manager.py

@@ -224,6 +224,8 @@ class GraphStateManager:
         Returns:
             Number of executing nodes
         """
+        # This count is a best-effort snapshot and can change concurrently.
+        # Only use it for pause-drain checks where scheduling is already frozen.
         with self._lock:
             return len(self._executing_nodes)
 

api/core/workflow/graph_engine/orchestration/dispatcher.py

@@ -83,12 +83,12 @@ class Dispatcher:
         """Main dispatcher loop."""
         try:
             self._process_commands()
+            paused = False
             while not self._stop_event.is_set():
-                if (
-                    self._execution_coordinator.aborted
-                    or self._execution_coordinator.paused
-                    or self._execution_coordinator.execution_complete
-                ):
+                if self._execution_coordinator.aborted or self._execution_coordinator.execution_complete:
+                    break
+                if self._execution_coordinator.paused:
+                    paused = True
                     break
 
                 self._execution_coordinator.check_scaling()
@@ -101,13 +101,10 @@ class Dispatcher:
                     time.sleep(0.1)
 
             self._process_commands()
-            while True:
-                try:
-                    event = self._event_queue.get(block=False)
-                    self._event_handler.dispatch(event)
-                    self._event_queue.task_done()
-                except queue.Empty:
-                    break
+            if paused:
+                self._drain_events_until_idle()
+            else:
+                self._drain_event_queue()
 
         except Exception as e:
             logger.exception("Dispatcher error")
@@ -122,3 +119,24 @@ class Dispatcher:
     def _process_commands(self, event: GraphNodeEventBase | None = None):
         if event is None or isinstance(event, self._COMMAND_TRIGGER_EVENTS):
             self._execution_coordinator.process_commands()
+
+    def _drain_event_queue(self) -> None:
+        while True:
+            try:
+                event = self._event_queue.get(block=False)
+                self._event_handler.dispatch(event)
+                self._event_queue.task_done()
+            except queue.Empty:
+                break
+
+    def _drain_events_until_idle(self) -> None:
+        while not self._stop_event.is_set():
+            try:
+                event = self._event_queue.get(timeout=0.1)
+                self._event_handler.dispatch(event)
+                self._event_queue.task_done()
+                self._process_commands(event)
+            except queue.Empty:
+                if not self._execution_coordinator.has_executing_nodes():
+                    break
+        self._drain_event_queue()