fix: list multiple @mentions

fix: simplify @llm warning
feat: 展示@agent引用
2026-01-16 00:19:28 +08:00 · 2026-01-16 00:08:51 +08:00 · 2026-01-15 23:32:14 +08:00 · 2026-01-15 17:53:28 +08:00 · 2026-01-15 17:12:30 +08:00 · 2026-01-15 16:44:15 +08:00
413 changed files with 37389 additions and 7775 deletions
@@ -0,0 +1 @@
+../.claude/skills
@@ -5,5 +5,18 @@
    "typescript-lsp@claude-plugins-official": true,
    "pyright-lsp@claude-plugins-official": true,
    "ralph-loop@claude-plugins-official": true
+  },
+  "hooks": {
+    "PreToolUse": [
+      {
+        "matcher": "Bash",
+        "hooks": [
+          {
+            "type": "command",
+            "command": "npx -y block-no-verify@1.1.1"
+          }
+        ]
+      }
+    ]
  }
 }
@@ -0,0 +1,46 @@
+---
+name: orpc-contract-first
+description: Guide for implementing oRPC contract-first API patterns in Dify frontend. Triggers when creating new API contracts, adding service endpoints, integrating TanStack Query with typed contracts, or migrating legacy service calls to oRPC. Use for all API layer work in web/contract and web/service directories.
+---
+
+# oRPC Contract-First Development
+
+## Project Structure
+
+```
+web/contract/
+├── base.ts           # Base contract (inputStructure: 'detailed')
+├── router.ts         # Router composition & type exports
+├── marketplace.ts    # Marketplace contracts
+└── console/          # Console contracts by domain
+    ├── system.ts
+    └── billing.ts
+```
+
+## Workflow
+
+1. **Create contract** in `web/contract/console/{domain}.ts`
+   - Import `base` from `../base` and `type` from `@orpc/contract`
+   - Define route with `path`, `method`, `input`, `output`
+
+2. **Register in router** at `web/contract/router.ts`
+   - Import directly from domain file (no barrel files)
+   - Nest by API prefix: `billing: { invoices, bindPartnerStack }`
+
+3. **Create hooks** in `web/service/use-{domain}.ts`
+   - Use `consoleQuery.{group}.{contract}.queryKey()` for query keys
+   - Use `consoleClient.{group}.{contract}()` for API calls
+
+## Key Rules
+
+- **Input structure**: Always use `{ params, query?, body? }` format
+- **Path params**: Use `{paramName}` in path, match in `params` object
+- **Router nesting**: Group by API prefix (e.g., `/billing/*` → `billing: {}`)
+- **No barrel files**: Import directly from specific files
+- **Types**: Import from `@/types/`, use `type<T>()` helper
+
+## Type Export
+
+```typescript
+export type ConsoleInputs = InferContractRouterInputs<typeof consoleRouterContract>
+```
@@ -39,12 +39,6 @@ jobs:
      - name: Install dependencies
        run: uv sync --project api --dev

-      - name: Run pyrefly check
-        run: |
-          cd api
-          uv add --dev pyrefly
-          uv run pyrefly check || true
-
      - name: Run dify config tests
        run: uv run --project api dev/pytest/pytest_config_tests.py

@@ -0,0 +1,29 @@
+name: Deploy HITL
+
+on:
+  workflow_run:
+    workflows: ["Build and Push API & Web"]
+    branches:
+      - "feat/hitl-frontend"
+      - "feat/hitl-backend"
+    types:
+      - completed
+
+jobs:
+  deploy:
+    runs-on: ubuntu-latest
+    if: |
+      github.event.workflow_run.conclusion == 'success' &&
+      (
+        github.event.workflow_run.head_branch == 'feat/hitl-frontend' ||
+        github.event.workflow_run.head_branch == 'feat/hitl-backend'
+      )
+    steps:
+      - name: Deploy to server
+        uses: appleboy/ssh-action@v0.1.8
+        with:
+          host: ${{ secrets.HITL_SSH_HOST }}
+          username: ${{ secrets.SSH_USER }}
+          key: ${{ secrets.SSH_PRIVATE_KEY }}
+          script: |
+            ${{ vars.SSH_SCRIPT || secrets.SSH_SCRIPT }}
@@ -90,7 +90,7 @@ jobs:
        uses: actions/setup-node@v6
        if: steps.changed-files.outputs.any_changed == 'true'
        with:
-          node-version: 22
+          node-version: 24
          cache: pnpm
          cache-dependency-path: ./web/pnpm-lock.yaml

@@ -16,10 +16,6 @@ jobs:
    name: unit test for Node.js SDK
    runs-on: ubuntu-latest

-    strategy:
-      matrix:
-        node-version: [16, 18, 20, 22]
-
    defaults:
      run:
        working-directory: sdks/nodejs-client
@@ -29,10 +25,10 @@ jobs:
        with:
          persist-credentials: false

-      - name: Use Node.js ${{ matrix.node-version }}
+      - name: Use Node.js
        uses: actions/setup-node@v6
        with:
-          node-version: ${{ matrix.node-version }}
+          node-version: 24
          cache: ''
          cache-dependency-path: 'pnpm-lock.yaml'

@@ -57,7 +57,7 @@ jobs:
      - name: Set up Node.js
        uses: actions/setup-node@v6
        with:
-          node-version: 'lts/*'
+          node-version: 24
          cache: pnpm
          cache-dependency-path: ./web/pnpm-lock.yaml

@@ -31,7 +31,7 @@ jobs:
      - name: Setup Node.js
        uses: actions/setup-node@v6
        with:
-          node-version: 22
+          node-version: 24
          cache: pnpm
          cache-dependency-path: ./web/pnpm-lock.yaml

@@ -209,6 +209,7 @@ api/.vscode
 .history

 .idea/
+web/migration/

 # pnpm
 /.pnpm-store
@@ -1 +0,0 @@
-22.11.0
@@ -417,6 +417,8 @@ SMTP_USERNAME=123
 SMTP_PASSWORD=abc
 SMTP_USE_TLS=true
 SMTP_OPPORTUNISTIC_TLS=false
+# Optional: override the local hostname used for SMTP HELO/EHLO
+SMTP_LOCAL_HOSTNAME=
 # Sendgid configuration
 SENDGRID_API_KEY=
 # Sentry configuration
@@ -589,6 +591,7 @@ ENABLE_CLEAN_UNUSED_DATASETS_TASK=false
 ENABLE_CREATE_TIDB_SERVERLESS_TASK=false
 ENABLE_UPDATE_TIDB_SERVERLESS_STATUS_TASK=false
 ENABLE_CLEAN_MESSAGES=false
+ENABLE_WORKFLOW_RUN_CLEANUP_TASK=false
 ENABLE_MAIL_CLEAN_DOCUMENT_NOTIFY_TASK=false
 ENABLE_DATASETS_QUEUE_MONITOR=false
 ENABLE_CHECK_UPGRADABLE_PLUGIN_TASK=true
@@ -712,3 +715,4 @@ ANNOTATION_IMPORT_MAX_CONCURRENT=5
 SANDBOX_EXPIRED_RECORDS_CLEAN_GRACEFUL_PERIOD=21
 SANDBOX_EXPIRED_RECORDS_CLEAN_BATCH_SIZE=1000
 SANDBOX_EXPIRED_RECORDS_RETENTION_DAYS=30
+
@@ -1,7 +1,9 @@
 import base64
+import datetime
 import json
 import logging
 import secrets
+import time
 from typing import Any

 import click
@@ -34,7 +36,7 @@ from libs.rsa import generate_key_pair
 from models import Tenant
 from models.dataset import Dataset, DatasetCollectionBinding, DatasetMetadata, DatasetMetadataBinding, DocumentSegment
 from models.dataset import Document as DatasetDocument
-from models.model import Account, App, AppAnnotationSetting, AppMode, Conversation, MessageAnnotation, UploadFile
+from models.model import App, AppAnnotationSetting, AppMode, Conversation, MessageAnnotation, UploadFile
 from models.oauth import DatasourceOauthParamConfig, DatasourceProvider
 from models.provider import Provider, ProviderModel
 from models.provider_ids import DatasourceProviderID, ToolProviderID
@@ -45,6 +47,9 @@ from services.clear_free_plan_tenant_expired_logs import ClearFreePlanTenantExpi
 from services.plugin.data_migration import PluginDataMigration
 from services.plugin.plugin_migration import PluginMigration
 from services.plugin.plugin_service import PluginService
+from services.retention.conversation.messages_clean_policy import create_message_clean_policy
+from services.retention.conversation.messages_clean_service import MessagesCleanService
+from services.retention.workflow_run.clear_free_plan_expired_workflow_run_logs import WorkflowRunCleanup
 from tasks.remove_app_and_related_data_task import delete_draft_variables_batch

 logger = logging.getLogger(__name__)
@@ -62,8 +67,10 @@ def reset_password(email, new_password, password_confirm):
    if str(new_password).strip() != str(password_confirm).strip():
        click.echo(click.style("Passwords do not match.", fg="red"))
        return
+    normalized_email = email.strip().lower()
+
    with sessionmaker(db.engine, expire_on_commit=False).begin() as session:
-        account = session.query(Account).where(Account.email == email).one_or_none()
+        account = AccountService.get_account_by_email_with_case_fallback(email.strip(), session=session)

        if not account:
            click.echo(click.style(f"Account not found for email: {email}", fg="red"))
@@ -84,7 +91,7 @@ def reset_password(email, new_password, password_confirm):
        base64_password_hashed = base64.b64encode(password_hashed).decode()
        account.password = base64_password_hashed
        account.password_salt = base64_salt
-        AccountService.reset_login_error_rate_limit(email)
+        AccountService.reset_login_error_rate_limit(normalized_email)
        click.echo(click.style("Password reset successfully.", fg="green"))


@@ -100,20 +107,22 @@ def reset_email(email, new_email, email_confirm):
    if str(new_email).strip() != str(email_confirm).strip():
        click.echo(click.style("New emails do not match.", fg="red"))
        return
+    normalized_new_email = new_email.strip().lower()
+
    with sessionmaker(db.engine, expire_on_commit=False).begin() as session:
-        account = session.query(Account).where(Account.email == email).one_or_none()
+        account = AccountService.get_account_by_email_with_case_fallback(email.strip(), session=session)

        if not account:
            click.echo(click.style(f"Account not found for email: {email}", fg="red"))
            return

        try:
-            email_validate(new_email)
+            email_validate(normalized_new_email)
        except:
            click.echo(click.style(f"Invalid email: {new_email}", fg="red"))
            return

-        account.email = new_email
+        account.email = normalized_new_email
        click.echo(click.style("Email updated successfully.", fg="green"))


@@ -658,7 +667,7 @@ def create_tenant(email: str, language: str | None = None, name: str | None = No
        return

    # Create account
-    email = email.strip()
+    email = email.strip().lower()

    if "@" not in email:
        click.echo(click.style("Invalid email address.", fg="red"))
@@ -852,6 +861,61 @@ def clear_free_plan_tenant_expired_logs(days: int, batch: int, tenant_ids: list[
    click.echo(click.style("Clear free plan tenant expired logs completed.", fg="green"))


+@click.command("clean-workflow-runs", help="Clean expired workflow runs and related data for free tenants.")
+@click.option("--days", default=30, show_default=True, help="Delete workflow runs created before N days ago.")
+@click.option("--batch-size", default=200, show_default=True, help="Batch size for selecting workflow runs.")
+@click.option(
+    "--start-from",
+    type=click.DateTime(formats=["%Y-%m-%d", "%Y-%m-%dT%H:%M:%S"]),
+    default=None,
+    help="Optional lower bound (inclusive) for created_at; must be paired with --end-before.",
+)
+@click.option(
+    "--end-before",
+    type=click.DateTime(formats=["%Y-%m-%d", "%Y-%m-%dT%H:%M:%S"]),
+    default=None,
+    help="Optional upper bound (exclusive) for created_at; must be paired with --start-from.",
+)
+@click.option(
+    "--dry-run",
+    is_flag=True,
+    help="Preview cleanup results without deleting any workflow run data.",
+)
+def clean_workflow_runs(
+    days: int,
+    batch_size: int,
+    start_from: datetime.datetime | None,
+    end_before: datetime.datetime | None,
+    dry_run: bool,
+):
+    """
+    Clean workflow runs and related workflow data for free tenants.
+    """
+    if (start_from is None) ^ (end_before is None):
+        raise click.UsageError("--start-from and --end-before must be provided together.")
+
+    start_time = datetime.datetime.now(datetime.UTC)
+    click.echo(click.style(f"Starting workflow run cleanup at {start_time.isoformat()}.", fg="white"))
+
+    WorkflowRunCleanup(
+        days=days,
+        batch_size=batch_size,
+        start_from=start_from,
+        end_before=end_before,
+        dry_run=dry_run,
+    ).run()
+
+    end_time = datetime.datetime.now(datetime.UTC)
+    elapsed = end_time - start_time
+    click.echo(
+        click.style(
+            f"Workflow run cleanup completed. start={start_time.isoformat()} "
+            f"end={end_time.isoformat()} duration={elapsed}",
+            fg="green",
+        )
+    )
+
+
@click.option("-f", "--force", is_flag=True, help="Skip user confirmation and force the command to execute.")
@click.command("clear-orphaned-file-records", help="Clear orphaned file records.")
 def clear_orphaned_file_records(force: bool):
@@ -2111,3 +2175,79 @@ def migrate_oss(
            except Exception as e:
                db.session.rollback()
                click.echo(click.style(f"Failed to update DB storage_type: {str(e)}", fg="red"))
+
+
+@click.command("clean-expired-messages", help="Clean expired messages.")
+@click.option(
+    "--start-from",
+    type=click.DateTime(formats=["%Y-%m-%d", "%Y-%m-%dT%H:%M:%S"]),
+    required=True,
+    help="Lower bound (inclusive) for created_at.",
+)
+@click.option(
+    "--end-before",
+    type=click.DateTime(formats=["%Y-%m-%d", "%Y-%m-%dT%H:%M:%S"]),
+    required=True,
+    help="Upper bound (exclusive) for created_at.",
+)
+@click.option("--batch-size", default=1000, show_default=True, help="Batch size for selecting messages.")
+@click.option(
+    "--graceful-period",
+    default=21,
+    show_default=True,
+    help="Graceful period in days after subscription expiration, will be ignored when billing is disabled.",
+)
+@click.option("--dry-run", is_flag=True, default=False, help="Show messages logs would be cleaned without deleting")
+def clean_expired_messages(
+    batch_size: int,
+    graceful_period: int,
+    start_from: datetime.datetime,
+    end_before: datetime.datetime,
+    dry_run: bool,
+):
+    """
+    Clean expired messages and related data for tenants based on clean policy.
+    """
+    click.echo(click.style("clean_messages: start clean messages.", fg="green"))
+
+    start_at = time.perf_counter()
+
+    try:
+        # Create policy based on billing configuration
+        # NOTE: graceful_period will be ignored when billing is disabled.
+        policy = create_message_clean_policy(graceful_period_days=graceful_period)
+
+        # Create and run the cleanup service
+        service = MessagesCleanService.from_time_range(
+            policy=policy,
+            start_from=start_from,
+            end_before=end_before,
+            batch_size=batch_size,
+            dry_run=dry_run,
+        )
+        stats = service.run()
+
+        end_at = time.perf_counter()
+        click.echo(
+            click.style(
+                f"clean_messages: completed successfully\n"
+                f"  - Latency: {end_at - start_at:.2f}s\n"
+                f"  - Batches processed: {stats['batches']}\n"
+                f"  - Total messages scanned: {stats['total_messages']}\n"
+                f"  - Messages filtered: {stats['filtered_messages']}\n"
+                f"  - Messages deleted: {stats['total_deleted']}",
+                fg="green",
+            )
+        )
+    except Exception as e:
+        end_at = time.perf_counter()
+        logger.exception("clean_messages failed")
+        click.echo(
+            click.style(
+                f"clean_messages: failed after {end_at - start_at:.2f}s - {str(e)}",
+                fg="red",
+            )
+        )
+        raise
+
+    click.echo(click.style("messages cleanup completed.", fg="green"))
@@ -949,6 +949,12 @@ class MailConfig(BaseSettings):
        default=False,
    )

+    SMTP_LOCAL_HOSTNAME: str | None = Field(
+        description="Override the local hostname used in SMTP HELO/EHLO. "
+        "Useful behind NAT or when the default hostname causes rejections.",
+        default=None,
+    )
+
    EMAIL_SEND_IP_LIMIT_PER_MINUTE: PositiveInt = Field(
        description="Maximum number of emails allowed to be sent from the same IP address in a minute",
        default=50,
@@ -1101,6 +1107,10 @@ class CeleryScheduleTasksConfig(BaseSettings):
        description="Enable clean messages task",
        default=False,
    )
+    ENABLE_WORKFLOW_RUN_CLEANUP_TASK: bool = Field(
+        description="Enable scheduled workflow run cleanup task",
+        default=False,
+    )
    ENABLE_MAIL_CLEAN_DOCUMENT_NOTIFY_TASK: bool = Field(
        description="Enable mail clean document notify task",
        default=False,
@@ -4,7 +4,7 @@ from pydantic_settings import BaseSettings

 class VolcengineTOSStorageConfig(BaseSettings):
    """
-    Configuration settings for Volcengine Tinder Object Storage (TOS)
+    Configuration settings for Volcengine Torch Object Storage (TOS)
    """

    VOLCENGINE_TOS_BUCKET_NAME: str | None = Field(
@@ -592,9 +592,12 @@ def _get_conversation(app_model, conversation_id):
    if not conversation:
        raise NotFound("Conversation Not Exists.")

-    if not conversation.read_at:
-        conversation.read_at = naive_utc_now()
-        conversation.read_account_id = current_user.id
-        db.session.commit()
+    db.session.execute(
+        sa.update(Conversation)
+        .where(Conversation.id == conversation_id, Conversation.read_at.is_(None))
+        .values(read_at=naive_utc_now(), read_account_id=current_user.id)
+    )
+    db.session.commit()
+    db.session.refresh(conversation)

    return conversation
@@ -55,6 +55,35 @@ class InstructionTemplatePayload(BaseModel):
    type: str = Field(..., description="Instruction template type")


+class ContextGeneratePayload(BaseModel):
+    """Payload for generating extractor code node."""
+
+    workflow_id: str = Field(..., description="Workflow ID")
+    node_id: str = Field(..., description="Current tool/llm node ID")
+    parameter_name: str = Field(..., description="Parameter name to generate code for")
+    language: str = Field(default="python3", description="Code language (python3/javascript)")
+    prompt_messages: list[dict[str, Any]] = Field(
+        ..., description="Multi-turn conversation history, last message is the current instruction"
+    )
+    model_config_data: dict[str, Any] = Field(..., alias="model_config", description="Model configuration")
+
+
+class SuggestedQuestionsPayload(BaseModel):
+    """Payload for generating suggested questions."""
+
+    workflow_id: str = Field(..., description="Workflow ID")
+    node_id: str = Field(..., description="Current tool/llm node ID")
+    parameter_name: str = Field(..., description="Parameter name")
+    language: str = Field(
+        default="English", description="Language for generated questions (e.g. English, Chinese, Japanese)"
+    )
+    model_config_data: dict[str, Any] | None = Field(
+        default=None,
+        alias="model_config",
+        description="Model configuration (optional, uses system default if not provided)",
+    )
+
+
 def reg(cls: type[BaseModel]):
    console_ns.schema_model(cls.__name__, cls.model_json_schema(ref_template=DEFAULT_REF_TEMPLATE_SWAGGER_2_0))

@@ -64,6 +93,8 @@ reg(RuleCodeGeneratePayload)
 reg(RuleStructuredOutputPayload)
 reg(InstructionGeneratePayload)
 reg(InstructionTemplatePayload)
+reg(ContextGeneratePayload)
+reg(SuggestedQuestionsPayload)


@console_ns.route("/rule-generate")
@@ -278,3 +309,74 @@ class InstructionGenerationTemplateApi(Resource):
                return {"data": INSTRUCTION_GENERATE_TEMPLATE_CODE}
            case _:
                raise ValueError(f"Invalid type: {args.type}")
+
+
+@console_ns.route("/context-generate")
+class ContextGenerateApi(Resource):
+    @console_ns.doc("generate_with_context")
+    @console_ns.doc(description="Generate with multi-turn conversation context")
+    @console_ns.expect(console_ns.models[ContextGeneratePayload.__name__])
+    @console_ns.response(200, "Content generated successfully")
+    @console_ns.response(400, "Invalid request parameters or workflow not found")
+    @console_ns.response(402, "Provider quota exceeded")
+    @setup_required
+    @login_required
+    @account_initialization_required
+    def post(self):
+        from core.llm_generator.utils import deserialize_prompt_messages
+
+        args = ContextGeneratePayload.model_validate(console_ns.payload)
+        _, current_tenant_id = current_account_with_tenant()
+
+        prompt_messages = deserialize_prompt_messages(args.prompt_messages)
+
+        try:
+            return LLMGenerator.generate_with_context(
+                tenant_id=current_tenant_id,
+                workflow_id=args.workflow_id,
+                node_id=args.node_id,
+                parameter_name=args.parameter_name,
+                language=args.language,
+                prompt_messages=prompt_messages,
+                model_config=args.model_config_data,
+            )
+        except ProviderTokenNotInitError as ex:
+            raise ProviderNotInitializeError(ex.description)
+        except QuotaExceededError:
+            raise ProviderQuotaExceededError()
+        except ModelCurrentlyNotSupportError:
+            raise ProviderModelCurrentlyNotSupportError()
+        except InvokeError as e:
+            raise CompletionRequestError(e.description)
+
+
+@console_ns.route("/context-generate/suggested-questions")
+class SuggestedQuestionsApi(Resource):
+    @console_ns.doc("generate_suggested_questions")
+    @console_ns.doc(description="Generate suggested questions for context generation")
+    @console_ns.expect(console_ns.models[SuggestedQuestionsPayload.__name__])
+    @console_ns.response(200, "Questions generated successfully")
+    @setup_required
+    @login_required
+    @account_initialization_required
+    def post(self):
+        args = SuggestedQuestionsPayload.model_validate(console_ns.payload)
+        _, current_tenant_id = current_account_with_tenant()
+
+        try:
+            return LLMGenerator.generate_suggested_questions(
+                tenant_id=current_tenant_id,
+                workflow_id=args.workflow_id,
+                node_id=args.node_id,
+                parameter_name=args.parameter_name,
+                language=args.language,
+                model_config=args.model_config_data,
+            )
+        except ProviderTokenNotInitError as ex:
+            raise ProviderNotInitializeError(ex.description)
+        except QuotaExceededError:
+            raise ProviderQuotaExceededError()
+        except ModelCurrentlyNotSupportError:
+            raise ProviderModelCurrentlyNotSupportError()
+        except InvokeError as e:
+            raise CompletionRequestError(e.description)
@@ -63,10 +63,9 @@ class ActivateCheckApi(Resource):
        args = ActivateCheckQuery.model_validate(request.args.to_dict(flat=True))  # type: ignore

        workspaceId = args.workspace_id
-        reg_email = args.email
        token = args.token

-        invitation = RegisterService.get_invitation_if_token_valid(workspaceId, reg_email, token)
+        invitation = RegisterService.get_invitation_with_case_fallback(workspaceId, args.email, token)
        if invitation:
            data = invitation.get("data", {})
            tenant = invitation.get("tenant", None)
@@ -100,11 +99,12 @@ class ActivateApi(Resource):
    def post(self):
        args = ActivatePayload.model_validate(console_ns.payload)

-        invitation = RegisterService.get_invitation_if_token_valid(args.workspace_id, args.email, args.token)
+        normalized_request_email = args.email.lower() if args.email else None
+        invitation = RegisterService.get_invitation_with_case_fallback(args.workspace_id, args.email, args.token)
        if invitation is None:
            raise AlreadyActivateError()

-        RegisterService.revoke_token(args.workspace_id, args.email, args.token)
+        RegisterService.revoke_token(args.workspace_id, normalized_request_email, args.token)

        account = invitation["account"]
        account.name = args.name
@@ -1,7 +1,6 @@
 from flask import request
 from flask_restx import Resource
 from pydantic import BaseModel, Field, field_validator
-from sqlalchemy import select
 from sqlalchemy.orm import Session

 from configs import dify_config
@@ -62,6 +61,7 @@ class EmailRegisterSendEmailApi(Resource):
    @email_register_enabled
    def post(self):
        args = EmailRegisterSendPayload.model_validate(console_ns.payload)
+        normalized_email = args.email.lower()

        ip_address = extract_remote_ip(request)
        if AccountService.is_email_send_ip_limit(ip_address):
@@ -70,13 +70,12 @@ class EmailRegisterSendEmailApi(Resource):
        if args.language in languages:
            language = args.language

-        if dify_config.BILLING_ENABLED and BillingService.is_email_in_freeze(args.email):
+        if dify_config.BILLING_ENABLED and BillingService.is_email_in_freeze(normalized_email):
            raise AccountInFreezeError()

        with Session(db.engine) as session:
-            account = session.execute(select(Account).filter_by(email=args.email)).scalar_one_or_none()
-        token = None
-        token = AccountService.send_email_register_email(email=args.email, account=account, language=language)
+            account = AccountService.get_account_by_email_with_case_fallback(args.email, session=session)
+        token = AccountService.send_email_register_email(email=normalized_email, account=account, language=language)
        return {"result": "success", "data": token}


@@ -88,9 +87,9 @@ class EmailRegisterCheckApi(Resource):
    def post(self):
        args = EmailRegisterValidityPayload.model_validate(console_ns.payload)

-        user_email = args.email
+        user_email = args.email.lower()

-        is_email_register_error_rate_limit = AccountService.is_email_register_error_rate_limit(args.email)
+        is_email_register_error_rate_limit = AccountService.is_email_register_error_rate_limit(user_email)
        if is_email_register_error_rate_limit:
            raise EmailRegisterLimitError()

@@ -98,11 +97,14 @@ class EmailRegisterCheckApi(Resource):
        if token_data is None:
            raise InvalidTokenError()

-        if user_email != token_data.get("email"):
+        token_email = token_data.get("email")
+        normalized_token_email = token_email.lower() if isinstance(token_email, str) else token_email
+
+        if user_email != normalized_token_email:
            raise InvalidEmailError()

        if args.code != token_data.get("code"):
-            AccountService.add_email_register_error_rate_limit(args.email)
+            AccountService.add_email_register_error_rate_limit(user_email)
            raise EmailCodeError()

        # Verified, revoke the first token
@@ -113,8 +115,8 @@ class EmailRegisterCheckApi(Resource):
            user_email, code=args.code, additional_data={"phase": "register"}
        )

-        AccountService.reset_email_register_error_rate_limit(args.email)
-        return {"is_valid": True, "email": token_data.get("email"), "token": new_token}
+        AccountService.reset_email_register_error_rate_limit(user_email)
+        return {"is_valid": True, "email": normalized_token_email, "token": new_token}


@console_ns.route("/email-register")
@@ -141,22 +143,23 @@ class EmailRegisterResetApi(Resource):
        AccountService.revoke_email_register_token(args.token)

        email = register_data.get("email", "")
+        normalized_email = email.lower()

        with Session(db.engine) as session:
-            account = session.execute(select(Account).filter_by(email=email)).scalar_one_or_none()
+            account = AccountService.get_account_by_email_with_case_fallback(email, session=session)

            if account:
                raise EmailAlreadyInUseError()
            else:
-                account = self._create_new_account(email, args.password_confirm)
+                account = self._create_new_account(normalized_email, args.password_confirm)
                if not account:
                    raise AccountNotFoundError()
                token_pair = AccountService.login(account=account, ip_address=extract_remote_ip(request))
-                AccountService.reset_login_error_rate_limit(email)
+                AccountService.reset_login_error_rate_limit(normalized_email)

        return {"result": "success", "data": token_pair.model_dump()}

-    def _create_new_account(self, email, password) -> Account | None:
+    def _create_new_account(self, email: str, password: str) -> Account | None:
        # Create new account if allowed
        account = None
        try:
@@ -4,7 +4,6 @@ import secrets
 from flask import request
 from flask_restx import Resource, fields
 from pydantic import BaseModel, Field, field_validator
-from sqlalchemy import select
 from sqlalchemy.orm import Session

 from controllers.console import console_ns
@@ -21,7 +20,6 @@ from events.tenant_event import tenant_was_created
 from extensions.ext_database import db
 from libs.helper import EmailStr, extract_remote_ip
 from libs.password import hash_password, valid_password
-from models import Account
 from services.account_service import AccountService, TenantService
 from services.feature_service import FeatureService

@@ -76,6 +74,7 @@ class ForgotPasswordSendEmailApi(Resource):
    @email_password_login_enabled
    def post(self):
        args = ForgotPasswordSendPayload.model_validate(console_ns.payload)
+        normalized_email = args.email.lower()

        ip_address = extract_remote_ip(request)
        if AccountService.is_email_send_ip_limit(ip_address):
@@ -87,11 +86,11 @@ class ForgotPasswordSendEmailApi(Resource):
            language = "en-US"

        with Session(db.engine) as session:
-            account = session.execute(select(Account).filter_by(email=args.email)).scalar_one_or_none()
+            account = AccountService.get_account_by_email_with_case_fallback(args.email, session=session)

        token = AccountService.send_reset_password_email(
            account=account,
-            email=args.email,
+            email=normalized_email,
            language=language,
            is_allow_register=FeatureService.get_system_features().is_allow_register,
        )
@@ -122,9 +121,9 @@ class ForgotPasswordCheckApi(Resource):
    def post(self):
        args = ForgotPasswordCheckPayload.model_validate(console_ns.payload)

-        user_email = args.email
+        user_email = args.email.lower()

-        is_forgot_password_error_rate_limit = AccountService.is_forgot_password_error_rate_limit(args.email)
+        is_forgot_password_error_rate_limit = AccountService.is_forgot_password_error_rate_limit(user_email)
        if is_forgot_password_error_rate_limit:
            raise EmailPasswordResetLimitError()

@@ -132,11 +131,16 @@ class ForgotPasswordCheckApi(Resource):
        if token_data is None:
            raise InvalidTokenError()

-        if user_email != token_data.get("email"):
+        token_email = token_data.get("email")
+        if not isinstance(token_email, str):
+            raise InvalidEmailError()
+        normalized_token_email = token_email.lower()
+
+        if user_email != normalized_token_email:
            raise InvalidEmailError()

        if args.code != token_data.get("code"):
-            AccountService.add_forgot_password_error_rate_limit(args.email)
+            AccountService.add_forgot_password_error_rate_limit(user_email)
            raise EmailCodeError()

        # Verified, revoke the first token
@@ -144,11 +148,11 @@ class ForgotPasswordCheckApi(Resource):

        # Refresh token data by generating a new token
        _, new_token = AccountService.generate_reset_password_token(
-            user_email, code=args.code, additional_data={"phase": "reset"}
+            token_email, code=args.code, additional_data={"phase": "reset"}
        )

-        AccountService.reset_forgot_password_error_rate_limit(args.email)
-        return {"is_valid": True, "email": token_data.get("email"), "token": new_token}
+        AccountService.reset_forgot_password_error_rate_limit(user_email)
+        return {"is_valid": True, "email": normalized_token_email, "token": new_token}


@console_ns.route("/forgot-password/resets")
@@ -187,9 +191,8 @@ class ForgotPasswordResetApi(Resource):
        password_hashed = hash_password(args.new_password, salt)

        email = reset_data.get("email", "")
-
        with Session(db.engine) as session:
-            account = session.execute(select(Account).filter_by(email=email)).scalar_one_or_none()
+            account = AccountService.get_account_by_email_with_case_fallback(email, session=session)

            if account:
                self._update_existing_account(account, password_hashed, salt, session)
@@ -90,32 +90,38 @@ class LoginApi(Resource):
    def post(self):
        """Authenticate user and login."""
        args = LoginPayload.model_validate(console_ns.payload)
+        request_email = args.email
+        normalized_email = request_email.lower()

-        if dify_config.BILLING_ENABLED and BillingService.is_email_in_freeze(args.email):
+        if dify_config.BILLING_ENABLED and BillingService.is_email_in_freeze(normalized_email):
            raise AccountInFreezeError()

-        is_login_error_rate_limit = AccountService.is_login_error_rate_limit(args.email)
+        is_login_error_rate_limit = AccountService.is_login_error_rate_limit(normalized_email)
        if is_login_error_rate_limit:
            raise EmailPasswordLoginLimitError()

+        invite_token = args.invite_token
        invitation_data: dict[str, Any] | None = None
-        if args.invite_token:
-            invitation_data = RegisterService.get_invitation_if_token_valid(None, args.email, args.invite_token)
+        if invite_token:
+            invitation_data = RegisterService.get_invitation_with_case_fallback(None, request_email, invite_token)
+            if invitation_data is None:
+                invite_token = None

        try:
            if invitation_data:
                data = invitation_data.get("data", {})
                invitee_email = data.get("email") if data else None
-                if invitee_email != args.email:
+                invitee_email_normalized = invitee_email.lower() if isinstance(invitee_email, str) else invitee_email
+                if invitee_email_normalized != normalized_email:
                    raise InvalidEmailError()
-                account = AccountService.authenticate(args.email, args.password, args.invite_token)
-            else:
-                account = AccountService.authenticate(args.email, args.password)
+            account = _authenticate_account_with_case_fallback(
+                request_email, normalized_email, args.password, invite_token
+            )
        except services.errors.account.AccountLoginError:
            raise AccountBannedError()
-        except services.errors.account.AccountPasswordError:
-            AccountService.add_login_error_rate_limit(args.email)
-            raise AuthenticationFailedError()
+        except services.errors.account.AccountPasswordError as exc:
+            AccountService.add_login_error_rate_limit(normalized_email)
+            raise AuthenticationFailedError() from exc
        # SELF_HOSTED only have one workspace
        tenants = TenantService.get_join_tenants(account)
        if len(tenants) == 0:
@@ -130,7 +136,7 @@ class LoginApi(Resource):
                }

        token_pair = AccountService.login(account=account, ip_address=extract_remote_ip(request))
-        AccountService.reset_login_error_rate_limit(args.email)
+        AccountService.reset_login_error_rate_limit(normalized_email)

        # Create response with cookies instead of returning tokens in body
        response = make_response({"result": "success"})
@@ -170,18 +176,19 @@ class ResetPasswordSendEmailApi(Resource):
    @console_ns.expect(console_ns.models[EmailPayload.__name__])
    def post(self):
        args = EmailPayload.model_validate(console_ns.payload)
+        normalized_email = args.email.lower()

        if args.language is not None and args.language == "zh-Hans":
            language = "zh-Hans"
        else:
            language = "en-US"
        try:
-            account = AccountService.get_user_through_email(args.email)
+            account = _get_account_with_case_fallback(args.email)
        except AccountRegisterError:
            raise AccountInFreezeError()

        token = AccountService.send_reset_password_email(
-            email=args.email,
+            email=normalized_email,
            account=account,
            language=language,
            is_allow_register=FeatureService.get_system_features().is_allow_register,
@@ -196,6 +203,7 @@ class EmailCodeLoginSendEmailApi(Resource):
    @console_ns.expect(console_ns.models[EmailPayload.__name__])
    def post(self):
        args = EmailPayload.model_validate(console_ns.payload)
+        normalized_email = args.email.lower()

        ip_address = extract_remote_ip(request)
        if AccountService.is_email_send_ip_limit(ip_address):
@@ -206,13 +214,13 @@ class EmailCodeLoginSendEmailApi(Resource):
        else:
            language = "en-US"
        try:
-            account = AccountService.get_user_through_email(args.email)
+            account = _get_account_with_case_fallback(args.email)
        except AccountRegisterError:
            raise AccountInFreezeError()

        if account is None:
            if FeatureService.get_system_features().is_allow_register:
-                token = AccountService.send_email_code_login_email(email=args.email, language=language)
+                token = AccountService.send_email_code_login_email(email=normalized_email, language=language)
            else:
                raise AccountNotFound()
        else:
@@ -229,14 +237,17 @@ class EmailCodeLoginApi(Resource):
    def post(self):
        args = EmailCodeLoginPayload.model_validate(console_ns.payload)

-        user_email = args.email
+        original_email = args.email
+        user_email = original_email.lower()
        language = args.language

        token_data = AccountService.get_email_code_login_data(args.token)
        if token_data is None:
            raise InvalidTokenError()

-        if token_data["email"] != args.email:
+        token_email = token_data.get("email")
+        normalized_token_email = token_email.lower() if isinstance(token_email, str) else token_email
+        if normalized_token_email != user_email:
            raise InvalidEmailError()

        if token_data["code"] != args.code:
@@ -244,7 +255,7 @@ class EmailCodeLoginApi(Resource):

        AccountService.revoke_email_code_login_token(args.token)
        try:
-            account = AccountService.get_user_through_email(user_email)
+            account = _get_account_with_case_fallback(original_email)
        except AccountRegisterError:
            raise AccountInFreezeError()
        if account:
@@ -275,7 +286,7 @@ class EmailCodeLoginApi(Resource):
            except WorkspacesLimitExceededError:
                raise WorkspacesLimitExceeded()
        token_pair = AccountService.login(account, ip_address=extract_remote_ip(request))
-        AccountService.reset_login_error_rate_limit(args.email)
+        AccountService.reset_login_error_rate_limit(user_email)

        # Create response with cookies instead of returning tokens in body
        response = make_response({"result": "success"})
@@ -309,3 +320,22 @@ class RefreshTokenApi(Resource):
            return response
        except Exception as e:
            return {"result": "fail", "message": str(e)}, 401
+
+
+def _get_account_with_case_fallback(email: str):
+    account = AccountService.get_user_through_email(email)
+    if account or email == email.lower():
+        return account
+
+    return AccountService.get_user_through_email(email.lower())
+
+
+def _authenticate_account_with_case_fallback(
+    original_email: str, normalized_email: str, password: str, invite_token: str | None
+):
+    try:
+        return AccountService.authenticate(original_email, password, invite_token)
+    except services.errors.account.AccountPasswordError:
+        if original_email == normalized_email:
+            raise
+        return AccountService.authenticate(normalized_email, password, invite_token)
@@ -3,7 +3,6 @@ import logging
 import httpx
 from flask import current_app, redirect, request
 from flask_restx import Resource
-from sqlalchemy import select
 from sqlalchemy.orm import Session
 from werkzeug.exceptions import Unauthorized

@@ -118,7 +117,10 @@ class OAuthCallback(Resource):
            invitation = RegisterService.get_invitation_by_token(token=invite_token)
            if invitation:
                invitation_email = invitation.get("email", None)
-                if invitation_email != user_info.email:
+                invitation_email_normalized = (
+                    invitation_email.lower() if isinstance(invitation_email, str) else invitation_email
+                )
+                if invitation_email_normalized != user_info.email.lower():
                    return redirect(f"{dify_config.CONSOLE_WEB_URL}/signin?message=Invalid invitation token.")

            return redirect(f"{dify_config.CONSOLE_WEB_URL}/signin/invite-settings?invite_token={invite_token}")
@@ -175,7 +177,7 @@ def _get_account_by_openid_or_email(provider: str, user_info: OAuthUserInfo) ->

    if not account:
        with Session(db.engine) as session:
-            account = session.execute(select(Account).filter_by(email=user_info.email)).scalar_one_or_none()
+            account = AccountService.get_account_by_email_with_case_fallback(user_info.email, session=session)

    return account

@@ -197,9 +199,10 @@ def _generate_account(provider: str, user_info: OAuthUserInfo) -> tuple[Account,
                tenant_was_created.send(new_tenant)

    if not account:
+        normalized_email = user_info.email.lower()
        oauth_new_user = True
        if not FeatureService.get_system_features().is_allow_register:
-            if dify_config.BILLING_ENABLED and BillingService.is_email_in_freeze(user_info.email):
+            if dify_config.BILLING_ENABLED and BillingService.is_email_in_freeze(normalized_email):
                raise AccountRegisterError(
                    description=(
                        "This email account has been deleted within the past "
@@ -210,7 +213,11 @@ def _generate_account(provider: str, user_info: OAuthUserInfo) -> tuple[Account,
                raise AccountRegisterError(description=("Invalid email or password"))
        account_name = user_info.name or "Dify"
        account = RegisterService.register(
-            email=user_info.email, name=account_name, password=None, open_id=user_info.id, provider=provider
+            email=normalized_email,
+            name=account_name,
+            password=None,
+            open_id=user_info.id,
+            provider=provider,
        )

        # Set interface language
@@ -7,7 +7,7 @@ from typing import Literal, cast
 import sqlalchemy as sa
 from flask import request
 from flask_restx import Resource, fields, marshal, marshal_with
-from pydantic import BaseModel
+from pydantic import BaseModel, Field
 from sqlalchemy import asc, desc, select
 from werkzeug.exceptions import Forbidden, NotFound

@@ -104,6 +104,15 @@ class DocumentRenamePayload(BaseModel):
    name: str


+class DocumentDatasetListParam(BaseModel):
+    page: int = Field(1, title="Page", description="Page number.")
+    limit: int = Field(20, title="Limit", description="Page size.")
+    search: str | None = Field(None, alias="keyword", title="Search", description="Search keyword.")
+    sort_by: str = Field("-created_at", alias="sort", title="SortBy", description="Sort by field.")
+    status: str | None = Field(None, title="Status", description="Document status.")
+    fetch_val: str = Field("false", alias="fetch")
+
+
 register_schema_models(
    console_ns,
    KnowledgeConfig,
@@ -225,14 +234,16 @@ class DatasetDocumentListApi(Resource):
    def get(self, dataset_id):
        current_user, current_tenant_id = current_account_with_tenant()
        dataset_id = str(dataset_id)
-        page = request.args.get("page", default=1, type=int)
-        limit = request.args.get("limit", default=20, type=int)
-        search = request.args.get("keyword", default=None, type=str)
-        sort = request.args.get("sort", default="-created_at", type=str)
-        status = request.args.get("status", default=None, type=str)
+        raw_args = request.args.to_dict()
+        param = DocumentDatasetListParam.model_validate(raw_args)
+        page = param.page
+        limit = param.limit
+        search = param.search
+        sort = param.sort_by
+        status = param.status
        # "yes", "true", "t", "y", "1" convert to True, while others convert to False.
        try:
-            fetch_val = request.args.get("fetch", default="false")
+            fetch_val = param.fetch_val
            if isinstance(fetch_val, bool):
                fetch = fetch_val
            else:
@@ -81,7 +81,7 @@ class ExternalKnowledgeApiPayload(BaseModel):
 class ExternalDatasetCreatePayload(BaseModel):
    external_knowledge_api_id: str
    external_knowledge_id: str
-    name: str = Field(..., min_length=1, max_length=40)
+    name: str = Field(..., min_length=1, max_length=100)
    description: str | None = Field(None, max_length=400)
    external_retrieval_model: dict[str, object] | None = None

@@ -84,10 +84,11 @@ class SetupApi(Resource):
            raise NotInitValidateError()

        args = SetupRequestPayload.model_validate(console_ns.payload)
+        normalized_email = args.email.lower()

        # setup
        RegisterService.setup(
-            email=args.email,
+            email=normalized_email,
            name=args.name,
            password=args.password,
            ip_address=extract_remote_ip(request),
@@ -41,7 +41,7 @@ from fields.member_fields import account_fields
 from libs.datetime_utils import naive_utc_now
 from libs.helper import EmailStr, TimestampField, extract_remote_ip, timezone
 from libs.login import current_account_with_tenant, login_required
-from models import Account, AccountIntegrate, InvitationCode
+from models import AccountIntegrate, InvitationCode
 from services.account_service import AccountService
 from services.billing_service import BillingService
 from services.errors.account import CurrentPasswordIncorrectError as ServiceCurrentPasswordIncorrectError
@@ -536,7 +536,8 @@ class ChangeEmailSendEmailApi(Resource):
        else:
            language = "en-US"
        account = None
-        user_email = args.email
+        user_email = None
+        email_for_sending = args.email.lower()
        if args.phase is not None and args.phase == "new_email":
            if args.token is None:
                raise InvalidTokenError()
@@ -546,16 +547,24 @@ class ChangeEmailSendEmailApi(Resource):
                raise InvalidTokenError()
            user_email = reset_data.get("email", "")

-            if user_email != current_user.email:
+            if user_email.lower() != current_user.email.lower():
                raise InvalidEmailError()
+
+            user_email = current_user.email
        else:
            with Session(db.engine) as session:
-                account = session.execute(select(Account).filter_by(email=args.email)).scalar_one_or_none()
+                account = AccountService.get_account_by_email_with_case_fallback(args.email, session=session)
            if account is None:
                raise AccountNotFound()
+            email_for_sending = account.email
+            user_email = account.email

        token = AccountService.send_change_email_email(
-            account=account, email=args.email, old_email=user_email, language=language, phase=args.phase
+            account=account,
+            email=email_for_sending,
+            old_email=user_email,
+            language=language,
+            phase=args.phase,
        )
        return {"result": "success", "data": token}

@@ -571,9 +580,9 @@ class ChangeEmailCheckApi(Resource):
        payload = console_ns.payload or {}
        args = ChangeEmailValidityPayload.model_validate(payload)

-        user_email = args.email
+        user_email = args.email.lower()

-        is_change_email_error_rate_limit = AccountService.is_change_email_error_rate_limit(args.email)
+        is_change_email_error_rate_limit = AccountService.is_change_email_error_rate_limit(user_email)
        if is_change_email_error_rate_limit:
            raise EmailChangeLimitError()

@@ -581,11 +590,13 @@ class ChangeEmailCheckApi(Resource):
        if token_data is None:
            raise InvalidTokenError()

-        if user_email != token_data.get("email"):
+        token_email = token_data.get("email")
+        normalized_token_email = token_email.lower() if isinstance(token_email, str) else token_email
+        if user_email != normalized_token_email:
            raise InvalidEmailError()

        if args.code != token_data.get("code"):
-            AccountService.add_change_email_error_rate_limit(args.email)
+            AccountService.add_change_email_error_rate_limit(user_email)
            raise EmailCodeError()

        # Verified, revoke the first token
@@ -596,8 +607,8 @@ class ChangeEmailCheckApi(Resource):
            user_email, code=args.code, old_email=token_data.get("old_email"), additional_data={}
        )

-        AccountService.reset_change_email_error_rate_limit(args.email)
-        return {"is_valid": True, "email": token_data.get("email"), "token": new_token}
+        AccountService.reset_change_email_error_rate_limit(user_email)
+        return {"is_valid": True, "email": normalized_token_email, "token": new_token}


@console_ns.route("/account/change-email/reset")
@@ -611,11 +622,12 @@ class ChangeEmailResetApi(Resource):
    def post(self):
        payload = console_ns.payload or {}
        args = ChangeEmailResetPayload.model_validate(payload)
+        normalized_new_email = args.new_email.lower()

-        if AccountService.is_account_in_freeze(args.new_email):
+        if AccountService.is_account_in_freeze(normalized_new_email):
            raise AccountInFreezeError()

-        if not AccountService.check_email_unique(args.new_email):
+        if not AccountService.check_email_unique(normalized_new_email):
            raise EmailAlreadyInUseError()

        reset_data = AccountService.get_change_email_data(args.token)
@@ -626,13 +638,13 @@ class ChangeEmailResetApi(Resource):

        old_email = reset_data.get("old_email", "")
        current_user, _ = current_account_with_tenant()
-        if current_user.email != old_email:
+        if current_user.email.lower() != old_email.lower():
            raise AccountNotFound()

-        updated_account = AccountService.update_account_email(current_user, email=args.new_email)
+        updated_account = AccountService.update_account_email(current_user, email=normalized_new_email)

        AccountService.send_change_email_completed_notify_email(
-            email=args.new_email,
+            email=normalized_new_email,
        )

        return updated_account
@@ -645,8 +657,9 @@ class CheckEmailUnique(Resource):
    def post(self):
        payload = console_ns.payload or {}
        args = CheckEmailUniquePayload.model_validate(payload)
-        if AccountService.is_account_in_freeze(args.email):
+        normalized_email = args.email.lower()
+        if AccountService.is_account_in_freeze(normalized_email):
            raise AccountInFreezeError()
-        if not AccountService.check_email_unique(args.email):
+        if not AccountService.check_email_unique(normalized_email):
            raise EmailAlreadyInUseError()
        return {"result": "success"}
@@ -116,26 +116,31 @@ class MemberInviteEmailApi(Resource):
            raise WorkspaceMembersLimitExceeded()

        for invitee_email in invitee_emails:
+            normalized_invitee_email = invitee_email.lower()
            try:
                if not inviter.current_tenant:
                    raise ValueError("No current tenant")
                token = RegisterService.invite_new_member(
-                    inviter.current_tenant, invitee_email, interface_language, role=invitee_role, inviter=inviter
+                    tenant=inviter.current_tenant,
+                    email=invitee_email,
+                    language=interface_language,
+                    role=invitee_role,
+                    inviter=inviter,
                )
-                encoded_invitee_email = parse.quote(invitee_email)
+                encoded_invitee_email = parse.quote(normalized_invitee_email)
                invitation_results.append(
                    {
                        "status": "success",
-                        "email": invitee_email,
+                        "email": normalized_invitee_email,
                        "url": f"{console_web_url}/activate?email={encoded_invitee_email}&token={token}",
                    }
                )
            except AccountAlreadyInTenantError:
                invitation_results.append(
-                    {"status": "success", "email": invitee_email, "url": f"{console_web_url}/signin"}
+                    {"status": "success", "email": normalized_invitee_email, "url": f"{console_web_url}/signin"}
                )
            except Exception as e:
-                invitation_results.append({"status": "failed", "email": invitee_email, "message": str(e)})
+                invitation_results.append({"status": "failed", "email": normalized_invitee_email, "message": str(e)})

        return {
            "result": "success",
@@ -4,7 +4,6 @@ import secrets
 from flask import request
 from flask_restx import Resource
 from pydantic import BaseModel, Field, field_validator
-from sqlalchemy import select
 from sqlalchemy.orm import Session

 from controllers.common.schema import register_schema_models
@@ -22,7 +21,7 @@ from controllers.web import web_ns
 from extensions.ext_database import db
 from libs.helper import EmailStr, extract_remote_ip
 from libs.password import hash_password, valid_password
-from models import Account
+from models.account import Account
 from services.account_service import AccountService


@@ -70,6 +69,9 @@ class ForgotPasswordSendEmailApi(Resource):
    def post(self):
        payload = ForgotPasswordSendPayload.model_validate(web_ns.payload or {})

+        request_email = payload.email
+        normalized_email = request_email.lower()
+
        ip_address = extract_remote_ip(request)
        if AccountService.is_email_send_ip_limit(ip_address):
            raise EmailSendIpLimitError()
@@ -80,12 +82,12 @@ class ForgotPasswordSendEmailApi(Resource):
            language = "en-US"

        with Session(db.engine) as session:
-            account = session.execute(select(Account).filter_by(email=payload.email)).scalar_one_or_none()
+            account = AccountService.get_account_by_email_with_case_fallback(request_email, session=session)
        token = None
        if account is None:
            raise AuthenticationFailedError()
        else:
-            token = AccountService.send_reset_password_email(account=account, email=payload.email, language=language)
+            token = AccountService.send_reset_password_email(account=account, email=normalized_email, language=language)

        return {"result": "success", "data": token}

@@ -104,9 +106,9 @@ class ForgotPasswordCheckApi(Resource):
    def post(self):
        payload = ForgotPasswordCheckPayload.model_validate(web_ns.payload or {})

-        user_email = payload.email
+        user_email = payload.email.lower()

-        is_forgot_password_error_rate_limit = AccountService.is_forgot_password_error_rate_limit(payload.email)
+        is_forgot_password_error_rate_limit = AccountService.is_forgot_password_error_rate_limit(user_email)
        if is_forgot_password_error_rate_limit:
            raise EmailPasswordResetLimitError()

@@ -114,11 +116,16 @@ class ForgotPasswordCheckApi(Resource):
        if token_data is None:
            raise InvalidTokenError()

-        if user_email != token_data.get("email"):
+        token_email = token_data.get("email")
+        if not isinstance(token_email, str):
+            raise InvalidEmailError()
+        normalized_token_email = token_email.lower()
+
+        if user_email != normalized_token_email:
            raise InvalidEmailError()

        if payload.code != token_data.get("code"):
-            AccountService.add_forgot_password_error_rate_limit(payload.email)
+            AccountService.add_forgot_password_error_rate_limit(user_email)
            raise EmailCodeError()

        # Verified, revoke the first token
@@ -126,11 +133,11 @@ class ForgotPasswordCheckApi(Resource):

        # Refresh token data by generating a new token
        _, new_token = AccountService.generate_reset_password_token(
-            user_email, code=payload.code, additional_data={"phase": "reset"}
+            token_email, code=payload.code, additional_data={"phase": "reset"}
        )

-        AccountService.reset_forgot_password_error_rate_limit(payload.email)
-        return {"is_valid": True, "email": token_data.get("email"), "token": new_token}
+        AccountService.reset_forgot_password_error_rate_limit(user_email)
+        return {"is_valid": True, "email": normalized_token_email, "token": new_token}


@web_ns.route("/forgot-password/resets")
@@ -174,7 +181,7 @@ class ForgotPasswordResetApi(Resource):
        email = reset_data.get("email", "")

        with Session(db.engine) as session:
-            account = session.execute(select(Account).filter_by(email=email)).scalar_one_or_none()
+            account = AccountService.get_account_by_email_with_case_fallback(email, session=session)

            if account:
                self._update_existing_account(account, password_hashed, salt, session)
@@ -197,25 +197,29 @@ class EmailCodeLoginApi(Resource):
        )
        args = parser.parse_args()

-        user_email = args["email"]
+        user_email = args["email"].lower()

        token_data = WebAppAuthService.get_email_code_login_data(args["token"])
        if token_data is None:
            raise InvalidTokenError()

-        if token_data["email"] != args["email"]:
+        token_email = token_data.get("email")
+        if not isinstance(token_email, str):
+            raise InvalidEmailError()
+        normalized_token_email = token_email.lower()
+        if normalized_token_email != user_email:
            raise InvalidEmailError()

        if token_data["code"] != args["code"]:
            raise EmailCodeError()

        WebAppAuthService.revoke_email_code_login_token(args["token"])
-        account = WebAppAuthService.get_user_through_email(user_email)
+        account = WebAppAuthService.get_user_through_email(token_email)
        if not account:
            raise AuthenticationFailedError()

        token = WebAppAuthService.login(account=account)
-        AccountService.reset_login_error_rate_limit(args["email"])
+        AccountService.reset_login_error_rate_limit(user_email)
        response = make_response({"result": "success", "data": {"access_token": token}})
        # set_access_token_to_cookie(request, response, token, samesite="None", httponly=False)
        return response
@@ -1,6 +1,7 @@
 import json
 import logging
 import uuid
+from decimal import Decimal
 from typing import Union, cast

 from sqlalchemy import select
@@ -41,6 +42,7 @@ from core.tools.tool_manager import ToolManager
 from core.tools.utils.dataset_retriever_tool import DatasetRetrieverTool
 from extensions.ext_database import db
 from factories import file_factory
+from models.enums import CreatorUserRole
 from models.model import Conversation, Message, MessageAgentThought, MessageFile

 logger = logging.getLogger(__name__)
@@ -289,6 +291,7 @@ class BaseAgentRunner(AppRunner):
        thought = MessageAgentThought(
            message_id=message_id,
            message_chain_id=None,
+            tool_process_data=None,
            thought="",
            tool=tool_name,
            tool_labels_str="{}",
@@ -296,20 +299,20 @@ class BaseAgentRunner(AppRunner):
            tool_input=tool_input,
            message=message,
            message_token=0,
-            message_unit_price=0,
-            message_price_unit=0,
+            message_unit_price=Decimal(0),
+            message_price_unit=Decimal("0.001"),
            message_files=json.dumps(messages_ids) if messages_ids else "",
            answer="",
            observation="",
            answer_token=0,
-            answer_unit_price=0,
-            answer_price_unit=0,
+            answer_unit_price=Decimal(0),
+            answer_price_unit=Decimal("0.001"),
            tokens=0,
-            total_price=0,
+            total_price=Decimal(0),
            position=self.agent_thought_count + 1,
            currency="USD",
            latency=0,
-            created_by_role="account",
+            created_by_role=CreatorUserRole.ACCOUNT,
            created_by=self.user_id,
        )

@@ -342,7 +345,8 @@ class BaseAgentRunner(AppRunner):
            raise ValueError("agent thought not found")

        if thought:
-            agent_thought.thought += thought
+            existing_thought = agent_thought.thought or ""
+            agent_thought.thought = f"{existing_thought}{thought}"

        if tool_name:
            agent_thought.tool = tool_name
@@ -440,21 +444,30 @@ class BaseAgentRunner(AppRunner):
            agent_thoughts: list[MessageAgentThought] = message.agent_thoughts
            if agent_thoughts:
                for agent_thought in agent_thoughts:
-                    tools = agent_thought.tool
-                    if tools:
-                        tools = tools.split(";")
+                    tool_names_raw = agent_thought.tool
+                    if tool_names_raw:
+                        tool_names = tool_names_raw.split(";")
                        tool_calls: list[AssistantPromptMessage.ToolCall] = []
                        tool_call_response: list[ToolPromptMessage] = []
-                        try:
-                            tool_inputs = json.loads(agent_thought.tool_input)
-                        except Exception:
-                            tool_inputs = {tool: {} for tool in tools}
-                        try:
-                            tool_responses = json.loads(agent_thought.observation)
-                        except Exception:
-                            tool_responses = dict.fromkeys(tools, agent_thought.observation)
+                        tool_input_payload = agent_thought.tool_input
+                        if tool_input_payload:
+                            try:
+                                tool_inputs = json.loads(tool_input_payload)
+                            except Exception:
+                                tool_inputs = {tool: {} for tool in tool_names}
+                        else:
+                            tool_inputs = {tool: {} for tool in tool_names}

-                        for tool in tools:
+                        observation_payload = agent_thought.observation
+                        if observation_payload:
+                            try:
+                                tool_responses = json.loads(observation_payload)
+                            except Exception:
+                                tool_responses = dict.fromkeys(tool_names, observation_payload)
+                        else:
+                            tool_responses = dict.fromkeys(tool_names, observation_payload)
+
+                        for tool in tool_names:
                            # generate a uuid for tool call
                            tool_call_id = str(uuid.uuid4())
                            tool_calls.append(
@@ -484,7 +497,7 @@ class BaseAgentRunner(AppRunner):
                                *tool_call_response,
                            ]
                        )
-                    if not tools:
+                    if not tool_names_raw:
                        result.append(AssistantPromptMessage(content=agent_thought.thought))
            else:
                if message.answer:
@@ -188,7 +188,7 @@ class FunctionCallAgentRunner(BaseAgentRunner):
                    ),
                )

-            assistant_message = AssistantPromptMessage(content="", tool_calls=[])
+            assistant_message = AssistantPromptMessage(content=response, tool_calls=[])
            if tool_calls:
                assistant_message.tool_calls = [
                    AssistantPromptMessage.ToolCall(
@@ -200,8 +200,6 @@ class FunctionCallAgentRunner(BaseAgentRunner):
                    )
                    for tool_call in tool_calls
                ]
-            else:
-                assistant_message.content = response

            self._current_thoughts.append(assistant_message)

@@ -1,4 +1,3 @@
-import json
 from collections.abc import Sequence
 from enum import StrEnum, auto
 from typing import Any, Literal
@@ -121,7 +120,7 @@ class VariableEntity(BaseModel):
    allowed_file_types: Sequence[FileType] | None = Field(default_factory=list)
    allowed_file_extensions: Sequence[str] | None = Field(default_factory=list)
    allowed_file_upload_methods: Sequence[FileTransferMethod] | None = Field(default_factory=list)
-    json_schema: str | None = Field(default=None)
+    json_schema: dict | None = Field(default=None)

    @field_validator("description", mode="before")
    @classmethod
@@ -135,17 +134,11 @@ class VariableEntity(BaseModel):

    @field_validator("json_schema")
    @classmethod
-    def validate_json_schema(cls, schema: str | None) -> str | None:
+    def validate_json_schema(cls, schema: dict | None) -> dict | None:
        if schema is None:
            return None
-
        try:
-            json_schema = json.loads(schema)
-        except json.JSONDecodeError:
-            raise ValueError(f"invalid json_schema value {schema}")
-
-        try:
-            Draft7Validator.check_schema(json_schema)
+            Draft7Validator.check_schema(schema)
        except SchemaError as e:
            raise ValueError(f"Invalid JSON schema: {e.message}")
        return schema
@@ -26,7 +26,6 @@ class AdvancedChatAppConfigManager(BaseAppConfigManager):
    @classmethod
    def get_app_config(cls, app_model: App, workflow: Workflow) -> AdvancedChatAppConfig:
        features_dict = workflow.features_dict
-
        app_mode = AppMode.value_of(app_model.mode)
        app_config = AdvancedChatAppConfig(
            tenant_id=app_model.tenant_id,
@@ -24,7 +24,7 @@ from core.app.layers.conversation_variable_persist_layer import ConversationVari
 from core.db.session_factory import session_factory
 from core.moderation.base import ModerationError
 from core.moderation.input_moderation import InputModeration
-from core.variables.variables import VariableUnion
+from core.variables.variables import Variable
 from core.workflow.enums import WorkflowType
 from core.workflow.graph_engine.command_channels.redis_channel import RedisChannel
 from core.workflow.graph_engine.layers.base import GraphEngineLayer
@@ -149,8 +149,8 @@ class AdvancedChatAppRunner(WorkflowBasedAppRunner):
                system_variables=system_inputs,
                user_inputs=inputs,
                environment_variables=self._workflow.environment_variables,
-                # Based on the definition of `VariableUnion`,
-                # `list[Variable]` can be safely used as `list[VariableUnion]` since they are compatible.
+                # Based on the definition of `Variable`,
+                # `VariableBase` instances can be safely used as `Variable` since they are compatible.
                conversation_variables=conversation_variables,
            )

@@ -318,7 +318,7 @@ class AdvancedChatAppRunner(WorkflowBasedAppRunner):
            trace_manager=app_generate_entity.trace_manager,
        )

-    def _initialize_conversation_variables(self) -> list[VariableUnion]:
+    def _initialize_conversation_variables(self) -> list[Variable]:
        """
        Initialize conversation variables for the current conversation.

@@ -343,7 +343,7 @@ class AdvancedChatAppRunner(WorkflowBasedAppRunner):
            conversation_variables = [var.to_variable() for var in existing_variables]

            session.commit()
-            return cast(list[VariableUnion], conversation_variables)
+            return cast(list[Variable], conversation_variables)

    def _load_existing_conversation_variables(self, session: Session) -> list[ConversationVariable]:
        """
@@ -82,7 +82,7 @@ class AdvancedChatAppGenerateResponseConverter(AppGenerateResponseConverter):
                data = cls._error_to_stream_response(sub_stream_response.err)
                response_chunk.update(data)
            else:
-                response_chunk.update(sub_stream_response.model_dump(mode="json"))
+                response_chunk.update(sub_stream_response.model_dump(mode="json", exclude_none=True))
            yield response_chunk

    @classmethod
@@ -110,7 +110,7 @@ class AdvancedChatAppGenerateResponseConverter(AppGenerateResponseConverter):
            }

            if isinstance(sub_stream_response, MessageEndStreamResponse):
-                sub_stream_response_dict = sub_stream_response.model_dump(mode="json")
+                sub_stream_response_dict = sub_stream_response.model_dump(mode="json", exclude_none=True)
                metadata = sub_stream_response_dict.get("metadata", {})
                sub_stream_response_dict["metadata"] = cls._get_simple_metadata(metadata)
                response_chunk.update(sub_stream_response_dict)
@@ -120,6 +120,6 @@ class AdvancedChatAppGenerateResponseConverter(AppGenerateResponseConverter):
            elif isinstance(sub_stream_response, NodeStartStreamResponse | NodeFinishStreamResponse):
                response_chunk.update(sub_stream_response.to_ignore_detail_dict())
            else:
-                response_chunk.update(sub_stream_response.model_dump(mode="json"))
+                response_chunk.update(sub_stream_response.model_dump(mode="json", exclude_none=True))

            yield response_chunk
@@ -358,25 +358,6 @@ class AdvancedChatAppGenerateTaskPipeline(GraphRuntimeStateSupport):
        if node_finish_resp:
            yield node_finish_resp

-        # For ANSWER nodes, check if we need to send a message_replace event
-        # Only send if the final output differs from the accumulated task_state.answer
-        # This happens when variables were updated by variable_assigner during workflow execution
-        if event.node_type == NodeType.ANSWER and event.outputs:
-            final_answer = event.outputs.get("answer")
-            if final_answer is not None and final_answer != self._task_state.answer:
-                logger.info(
-                    "ANSWER node final output '%s' differs from accumulated answer '%s', sending message_replace event",
-                    final_answer,
-                    self._task_state.answer,
-                )
-                # Update the task state answer
-                self._task_state.answer = str(final_answer)
-                # Send message_replace event to update the UI
-                yield self._message_cycle_manager.message_replace_to_stream_response(
-                    answer=str(final_answer),
-                    reason="variable_update",
-                )
-
    def _handle_node_failed_events(
        self,
        event: Union[QueueNodeFailedEvent, QueueNodeExceptionEvent],
@@ -81,7 +81,7 @@ class AgentChatAppGenerateResponseConverter(AppGenerateResponseConverter):
                data = cls._error_to_stream_response(sub_stream_response.err)
                response_chunk.update(data)
            else:
-                response_chunk.update(sub_stream_response.model_dump(mode="json"))
+                response_chunk.update(sub_stream_response.model_dump(mode="json", exclude_none=True))
            yield response_chunk

    @classmethod
@@ -109,7 +109,7 @@ class AgentChatAppGenerateResponseConverter(AppGenerateResponseConverter):
            }

            if isinstance(sub_stream_response, MessageEndStreamResponse):
-                sub_stream_response_dict = sub_stream_response.model_dump(mode="json")
+                sub_stream_response_dict = sub_stream_response.model_dump(mode="json", exclude_none=True)
                metadata = sub_stream_response_dict.get("metadata", {})
                sub_stream_response_dict["metadata"] = cls._get_simple_metadata(metadata)
                response_chunk.update(sub_stream_response_dict)
@@ -117,6 +117,6 @@ class AgentChatAppGenerateResponseConverter(AppGenerateResponseConverter):
                data = cls._error_to_stream_response(sub_stream_response.err)
                response_chunk.update(data)
            else:
-                response_chunk.update(sub_stream_response.model_dump(mode="json"))
+                response_chunk.update(sub_stream_response.model_dump(mode="json", exclude_none=True))

            yield response_chunk
@@ -1,4 +1,3 @@
-import json
 from collections.abc import Generator, Mapping, Sequence
 from typing import TYPE_CHECKING, Any, Union, final

@@ -76,12 +75,24 @@ class BaseAppGenerator:
        user_inputs = {**user_inputs, **files_inputs, **file_list_inputs}

        # Check if all files are converted to File
-        if any(filter(lambda v: isinstance(v, dict), user_inputs.values())):
-            raise ValueError("Invalid input type")
-        if any(
-            filter(lambda v: isinstance(v, dict), filter(lambda item: isinstance(item, list), user_inputs.values()))
-        ):
-            raise ValueError("Invalid input type")
+        invalid_dict_keys = [
+            k
+            for k, v in user_inputs.items()
+            if isinstance(v, dict)
+            and entity_dictionary[k].type not in {VariableEntityType.FILE, VariableEntityType.JSON_OBJECT}
+        ]
+        if invalid_dict_keys:
+            raise ValueError(f"Invalid input type for {invalid_dict_keys}")
+
+        invalid_list_dict_keys = [
+            k
+            for k, v in user_inputs.items()
+            if isinstance(v, list)
+            and any(isinstance(item, dict) for item in v)
+            and entity_dictionary[k].type != VariableEntityType.FILE_LIST
+        ]
+        if invalid_list_dict_keys:
+            raise ValueError(f"Invalid input type for {invalid_list_dict_keys}")

        return user_inputs

@@ -178,12 +189,8 @@ class BaseAppGenerator:
                    elif value == 0:
                        value = False
            case VariableEntityType.JSON_OBJECT:
-                if not isinstance(value, str):
-                    raise ValueError(f"{variable_entity.variable} in input form must be a string")
-                try:
-                    json.loads(value)
-                except json.JSONDecodeError:
-                    raise ValueError(f"{variable_entity.variable} in input form must be a valid JSON object")
+                if value and not isinstance(value, dict):
+                    raise ValueError(f"{variable_entity.variable} in input form must be a dict")
            case _:
                raise AssertionError("this statement should be unreachable.")

@@ -81,7 +81,7 @@ class ChatAppGenerateResponseConverter(AppGenerateResponseConverter):
                data = cls._error_to_stream_response(sub_stream_response.err)
                response_chunk.update(data)
            else:
-                response_chunk.update(sub_stream_response.model_dump(mode="json"))
+                response_chunk.update(sub_stream_response.model_dump(mode="json", exclude_none=True))
            yield response_chunk

    @classmethod
@@ -109,7 +109,7 @@ class ChatAppGenerateResponseConverter(AppGenerateResponseConverter):
            }

            if isinstance(sub_stream_response, MessageEndStreamResponse):
-                sub_stream_response_dict = sub_stream_response.model_dump(mode="json")
+                sub_stream_response_dict = sub_stream_response.model_dump(mode="json", exclude_none=True)
                metadata = sub_stream_response_dict.get("metadata", {})
                sub_stream_response_dict["metadata"] = cls._get_simple_metadata(metadata)
                response_chunk.update(sub_stream_response_dict)
@@ -117,6 +117,6 @@ class ChatAppGenerateResponseConverter(AppGenerateResponseConverter):
                data = cls._error_to_stream_response(sub_stream_response.err)
                response_chunk.update(data)
            else:
-                response_chunk.update(sub_stream_response.model_dump(mode="json"))
+                response_chunk.update(sub_stream_response.model_dump(mode="json", exclude_none=True))

            yield response_chunk
@@ -70,6 +70,8 @@ class _NodeSnapshot:
    """Empty string means the node is not executing inside an iteration."""
    loop_id: str = ""
    """Empty string means the node is not executing inside a loop."""
+    mention_parent_id: str = ""
+    """Empty string means the node is not an extractor node."""


 class WorkflowResponseConverter:
@@ -131,6 +133,7 @@ class WorkflowResponseConverter:
            start_at=event.start_at,
            iteration_id=event.in_iteration_id or "",
            loop_id=event.in_loop_id or "",
+            mention_parent_id=event.in_mention_parent_id or "",
        )
        node_execution_id = NodeExecutionId(event.node_execution_id)
        self._node_snapshots[node_execution_id] = snapshot
@@ -287,6 +290,7 @@ class WorkflowResponseConverter:
                created_at=int(snapshot.start_at.timestamp()),
                iteration_id=event.in_iteration_id,
                loop_id=event.in_loop_id,
+                mention_parent_id=event.in_mention_parent_id,
                agent_strategy=event.agent_strategy,
            ),
        )
@@ -373,6 +377,7 @@ class WorkflowResponseConverter:
                files=self.fetch_files_from_node_outputs(event.outputs or {}),
                iteration_id=event.in_iteration_id,
                loop_id=event.in_loop_id,
+                mention_parent_id=event.in_mention_parent_id,
            ),
        )

@@ -422,6 +427,7 @@ class WorkflowResponseConverter:
                files=self.fetch_files_from_node_outputs(event.outputs or {}),
                iteration_id=event.in_iteration_id,
                loop_id=event.in_loop_id,
+                mention_parent_id=event.in_mention_parent_id,
                retry_index=event.retry_index,
            ),
        )
@@ -79,7 +79,7 @@ class CompletionAppGenerateResponseConverter(AppGenerateResponseConverter):
                data = cls._error_to_stream_response(sub_stream_response.err)
                response_chunk.update(data)
            else:
-                response_chunk.update(sub_stream_response.model_dump(mode="json"))
+                response_chunk.update(sub_stream_response.model_dump(mode="json", exclude_none=True))
            yield response_chunk

    @classmethod
@@ -106,7 +106,7 @@ class CompletionAppGenerateResponseConverter(AppGenerateResponseConverter):
            }

            if isinstance(sub_stream_response, MessageEndStreamResponse):
-                sub_stream_response_dict = sub_stream_response.model_dump(mode="json")
+                sub_stream_response_dict = sub_stream_response.model_dump(mode="json", exclude_none=True)
                metadata = sub_stream_response_dict.get("metadata", {})
                if not isinstance(metadata, dict):
                    metadata = {}
@@ -116,6 +116,6 @@ class CompletionAppGenerateResponseConverter(AppGenerateResponseConverter):
                data = cls._error_to_stream_response(sub_stream_response.err)
                response_chunk.update(data)
            else:
-                response_chunk.update(sub_stream_response.model_dump(mode="json"))
+                response_chunk.update(sub_stream_response.model_dump(mode="json", exclude_none=True))

            yield response_chunk
@@ -60,7 +60,7 @@ class WorkflowAppGenerateResponseConverter(AppGenerateResponseConverter):
                data = cls._error_to_stream_response(sub_stream_response.err)
                response_chunk.update(cast(dict, data))
            else:
-                response_chunk.update(sub_stream_response.model_dump())
+                response_chunk.update(sub_stream_response.model_dump(exclude_none=True))
            yield response_chunk

    @classmethod
@@ -91,5 +91,5 @@ class WorkflowAppGenerateResponseConverter(AppGenerateResponseConverter):
            elif isinstance(sub_stream_response, NodeStartStreamResponse | NodeFinishStreamResponse):
                response_chunk.update(cast(dict, sub_stream_response.to_ignore_detail_dict()))
            else:
-                response_chunk.update(sub_stream_response.model_dump())
+                response_chunk.update(sub_stream_response.model_dump(exclude_none=True))
            yield response_chunk
@@ -60,7 +60,7 @@ class WorkflowAppGenerateResponseConverter(AppGenerateResponseConverter):
                data = cls._error_to_stream_response(sub_stream_response.err)
                response_chunk.update(data)
            else:
-                response_chunk.update(sub_stream_response.model_dump(mode="json"))
+                response_chunk.update(sub_stream_response.model_dump(mode="json", exclude_none=True))
            yield response_chunk

    @classmethod
@@ -91,5 +91,5 @@ class WorkflowAppGenerateResponseConverter(AppGenerateResponseConverter):
            elif isinstance(sub_stream_response, NodeStartStreamResponse | NodeFinishStreamResponse):
                response_chunk.update(sub_stream_response.to_ignore_detail_dict())
            else:
-                response_chunk.update(sub_stream_response.model_dump(mode="json"))
+                response_chunk.update(sub_stream_response.model_dump(mode="json", exclude_none=True))
            yield response_chunk
@@ -385,6 +385,7 @@ class WorkflowBasedAppRunner:
                    start_at=event.start_at,
                    in_iteration_id=event.in_iteration_id,
                    in_loop_id=event.in_loop_id,
+                    in_mention_parent_id=event.in_mention_parent_id,
                    inputs=inputs,
                    process_data=process_data,
                    outputs=outputs,
@@ -405,6 +406,7 @@ class WorkflowBasedAppRunner:
                    start_at=event.start_at,
                    in_iteration_id=event.in_iteration_id,
                    in_loop_id=event.in_loop_id,
+                    in_mention_parent_id=event.in_mention_parent_id,
                    agent_strategy=event.agent_strategy,
                    provider_type=event.provider_type,
                    provider_id=event.provider_id,
@@ -428,6 +430,7 @@ class WorkflowBasedAppRunner:
                    execution_metadata=execution_metadata,
                    in_iteration_id=event.in_iteration_id,
                    in_loop_id=event.in_loop_id,
+                    in_mention_parent_id=event.in_mention_parent_id,
                )
            )
        elif isinstance(event, NodeRunFailedEvent):
@@ -444,6 +447,7 @@ class WorkflowBasedAppRunner:
                    execution_metadata=event.node_run_result.metadata,
                    in_iteration_id=event.in_iteration_id,
                    in_loop_id=event.in_loop_id,
+                    in_mention_parent_id=event.in_mention_parent_id,
                )
            )
        elif isinstance(event, NodeRunExceptionEvent):
@@ -460,6 +464,7 @@ class WorkflowBasedAppRunner:
                    execution_metadata=event.node_run_result.metadata,
                    in_iteration_id=event.in_iteration_id,
                    in_loop_id=event.in_loop_id,
+                    in_mention_parent_id=event.in_mention_parent_id,
                )
            )
        elif isinstance(event, NodeRunStreamChunkEvent):
@@ -469,6 +474,7 @@ class WorkflowBasedAppRunner:
                    from_variable_selector=list(event.selector),
                    in_iteration_id=event.in_iteration_id,
                    in_loop_id=event.in_loop_id,
+                    in_mention_parent_id=event.in_mention_parent_id,
                )
            )
        elif isinstance(event, NodeRunRetrieverResourceEvent):
@@ -477,6 +483,7 @@ class WorkflowBasedAppRunner:
                    retriever_resources=event.retriever_resources,
                    in_iteration_id=event.in_iteration_id,
                    in_loop_id=event.in_loop_id,
+                    in_mention_parent_id=event.in_mention_parent_id,
                )
            )
        elif isinstance(event, NodeRunAgentLogEvent):
@@ -190,6 +190,8 @@ class QueueTextChunkEvent(AppQueueEvent):
    """iteration id if node is in iteration"""
    in_loop_id: str | None = None
    """loop id if node is in loop"""
+    in_mention_parent_id: str | None = None
+    """parent node id if this is an extractor node event"""


 class QueueAgentMessageEvent(AppQueueEvent):
@@ -229,6 +231,8 @@ class QueueRetrieverResourcesEvent(AppQueueEvent):
    """iteration id if node is in iteration"""
    in_loop_id: str | None = None
    """loop id if node is in loop"""
+    in_mention_parent_id: str | None = None
+    """parent node id if this is an extractor node event"""


 class QueueAnnotationReplyEvent(AppQueueEvent):
@@ -306,6 +310,8 @@ class QueueNodeStartedEvent(AppQueueEvent):
    node_run_index: int = 1  # FIXME(-LAN-): may not used
    in_iteration_id: str | None = None
    in_loop_id: str | None = None
+    in_mention_parent_id: str | None = None
+    """parent node id if this is an extractor node event"""
    start_at: datetime
    agent_strategy: AgentNodeStrategyInit | None = None

@@ -328,6 +334,8 @@ class QueueNodeSucceededEvent(AppQueueEvent):
    """iteration id if node is in iteration"""
    in_loop_id: str | None = None
    """loop id if node is in loop"""
+    in_mention_parent_id: str | None = None
+    """parent node id if this is an extractor node event"""
    start_at: datetime

    inputs: Mapping[str, object] = Field(default_factory=dict)
@@ -383,6 +391,8 @@ class QueueNodeExceptionEvent(AppQueueEvent):
    """iteration id if node is in iteration"""
    in_loop_id: str | None = None
    """loop id if node is in loop"""
+    in_mention_parent_id: str | None = None
+    """parent node id if this is an extractor node event"""
    start_at: datetime

    inputs: Mapping[str, object] = Field(default_factory=dict)
@@ -407,6 +417,8 @@ class QueueNodeFailedEvent(AppQueueEvent):
    """iteration id if node is in iteration"""
    in_loop_id: str | None = None
    """loop id if node is in loop"""
+    in_mention_parent_id: str | None = None
+    """parent node id if this is an extractor node event"""
    start_at: datetime

    inputs: Mapping[str, object] = Field(default_factory=dict)
@@ -262,6 +262,7 @@ class NodeStartStreamResponse(StreamResponse):
        extras: dict[str, object] = Field(default_factory=dict)
        iteration_id: str | None = None
        loop_id: str | None = None
+        mention_parent_id: str | None = None
        agent_strategy: AgentNodeStrategyInit | None = None

    event: StreamEvent = StreamEvent.NODE_STARTED
@@ -285,6 +286,7 @@ class NodeStartStreamResponse(StreamResponse):
                "extras": {},
                "iteration_id": self.data.iteration_id,
                "loop_id": self.data.loop_id,
+                "mention_parent_id": self.data.mention_parent_id,
            },
        }

@@ -320,6 +322,7 @@ class NodeFinishStreamResponse(StreamResponse):
        files: Sequence[Mapping[str, Any]] | None = []
        iteration_id: str | None = None
        loop_id: str | None = None
+        mention_parent_id: str | None = None

    event: StreamEvent = StreamEvent.NODE_FINISHED
    workflow_run_id: str
@@ -349,6 +352,7 @@ class NodeFinishStreamResponse(StreamResponse):
                "files": [],
                "iteration_id": self.data.iteration_id,
                "loop_id": self.data.loop_id,
+                "mention_parent_id": self.data.mention_parent_id,
            },
        }

@@ -384,6 +388,7 @@ class NodeRetryStreamResponse(StreamResponse):
        files: Sequence[Mapping[str, Any]] | None = []
        iteration_id: str | None = None
        loop_id: str | None = None
+        mention_parent_id: str | None = None
        retry_index: int = 0

    event: StreamEvent = StreamEvent.NODE_RETRY
@@ -414,6 +419,7 @@ class NodeRetryStreamResponse(StreamResponse):
                "files": [],
                "iteration_id": self.data.iteration_id,
                "loop_id": self.data.loop_id,
+                "mention_parent_id": self.data.mention_parent_id,
                "retry_index": self.data.retry_index,
            },
        }
@@ -1,6 +1,6 @@
 import logging

-from core.variables import Variable
+from core.variables import VariableBase
 from core.workflow.constants import CONVERSATION_VARIABLE_NODE_ID
 from core.workflow.conversation_variable_updater import ConversationVariableUpdater
 from core.workflow.enums import NodeType
@@ -44,7 +44,7 @@ class ConversationVariablePersistenceLayer(GraphEngineLayer):
            if selector[0] != CONVERSATION_VARIABLE_NODE_ID:
                continue
            variable = self.graph_runtime_state.variable_pool.get(selector)
-            if not isinstance(variable, Variable):
+            if not isinstance(variable, VariableBase):
                logger.warning(
                    "Conversation variable not found in variable pool. selector=%s",
                    selector,
@@ -33,6 +33,10 @@ class MaxRetriesExceededError(ValueError):
    pass


+request_error = httpx.RequestError
+max_retries_exceeded_error = MaxRetriesExceededError
+
+
 def _create_proxy_mounts() -> dict[str, httpx.HTTPTransport]:
    return {
        "http://": httpx.HTTPTransport(
@@ -1,8 +1,8 @@
 import json
 import logging
 import re
-from collections.abc import Sequence
-from typing import Protocol, cast
+from collections.abc import Mapping, Sequence
+from typing import Any, Protocol, cast

 import json_repair

@@ -398,6 +398,488 @@ class LLMGenerator:
            logger.exception("Failed to invoke LLM model, model: %s", model_config.get("name"))
            return {"output": "", "error": f"An unexpected error occurred: {str(e)}"}

+    @classmethod
+    def generate_with_context(
+        cls,
+        tenant_id: str,
+        workflow_id: str,
+        node_id: str,
+        parameter_name: str,
+        language: str,
+        prompt_messages: list[PromptMessage],
+        model_config: dict,
+    ) -> dict:
+        """
+        Generate extractor code node based on conversation context.
+
+        Args:
+            tenant_id: Tenant/workspace ID
+            workflow_id: Workflow ID
+            node_id: Current tool/llm node ID
+            parameter_name: Parameter name to generate code for
+            language: Code language (python3/javascript)
+            prompt_messages: Multi-turn conversation history (last message is instruction)
+            model_config: Model configuration (provider, name, completion_params)
+
+        Returns:
+            dict with CodeNodeData format:
+            - variables: Input variable selectors
+            - code_language: Code language
+            - code: Generated code
+            - outputs: Output definitions
+            - message: Explanation
+            - error: Error message if any
+        """
+        from sqlalchemy import select
+        from sqlalchemy.orm import Session
+
+        from services.workflow_service import WorkflowService
+
+        # Get workflow
+        with Session(db.engine) as session:
+            stmt = select(App).where(App.id == workflow_id)
+            app = session.scalar(stmt)
+            if not app:
+                return cls._error_response(f"App {workflow_id} not found")
+
+            workflow = WorkflowService().get_draft_workflow(app_model=app)
+            if not workflow:
+                return cls._error_response(f"Workflow for app {workflow_id} not found")
+
+        # Get upstream nodes via edge backtracking
+        upstream_nodes = cls._get_upstream_nodes(workflow.graph_dict, node_id)
+
+        # Get current node info
+        current_node = cls._get_node_by_id(workflow.graph_dict, node_id)
+        if not current_node:
+            return cls._error_response(f"Node {node_id} not found")
+
+        # Get parameter info
+        parameter_info = cls._get_parameter_info(
+            tenant_id=tenant_id,
+            node_data=current_node.get("data", {}),
+            parameter_name=parameter_name,
+        )
+
+        # Build system prompt
+        system_prompt = cls._build_extractor_system_prompt(
+            upstream_nodes=upstream_nodes,
+            current_node=current_node,
+            parameter_info=parameter_info,
+            language=language,
+        )
+
+        # Construct complete prompt_messages with system prompt
+        complete_messages: list[PromptMessage] = [
+            SystemPromptMessage(content=system_prompt),
+            *prompt_messages,
+        ]
+
+        from core.llm_generator.output_parser.structured_output import invoke_llm_with_structured_output
+
+        # Get model instance and schema
+        provider = model_config.get("provider", "")
+        model_name = model_config.get("name", "")
+        model_instance = ModelManager().get_model_instance(
+            tenant_id=tenant_id,
+            model_type=ModelType.LLM,
+            provider=provider,
+            model=model_name,
+        )
+
+        model_schema = model_instance.model_type_instance.get_model_schema(model_name, model_instance.credentials)
+        if not model_schema:
+            return cls._error_response(f"Model schema not found for {model_name}")
+
+        model_parameters = model_config.get("completion_params", {})
+        json_schema = cls._get_code_node_json_schema()
+
+        try:
+            response = invoke_llm_with_structured_output(
+                provider=provider,
+                model_schema=model_schema,
+                model_instance=model_instance,
+                prompt_messages=complete_messages,
+                json_schema=json_schema,
+                model_parameters=model_parameters,
+                stream=False,
+                tenant_id=tenant_id,
+            )
+
+            return cls._parse_code_node_output(
+                response.structured_output, language, parameter_info.get("type", "string")
+            )
+
+        except InvokeError as e:
+            return cls._error_response(str(e))
+        except Exception as e:
+            logger.exception("Failed to generate with context, model: %s", model_config.get("name"))
+            return cls._error_response(f"An unexpected error occurred: {str(e)}")
+
+    @classmethod
+    def _error_response(cls, error: str) -> dict:
+        """Return error response in CodeNodeData format."""
+        return {
+            "variables": [],
+            "code_language": "python3",
+            "code": "",
+            "outputs": {},
+            "message": "",
+            "error": error,
+        }
+
+    @classmethod
+    def generate_suggested_questions(
+        cls,
+        tenant_id: str,
+        workflow_id: str,
+        node_id: str,
+        parameter_name: str,
+        language: str,
+        model_config: dict | None = None,
+    ) -> dict:
+        """
+        Generate suggested questions for context generation.
+
+        Returns dict with questions array and error field.
+        """
+        from sqlalchemy import select
+        from sqlalchemy.orm import Session
+
+        from core.llm_generator.output_parser.structured_output import invoke_llm_with_structured_output
+        from services.workflow_service import WorkflowService
+
+        # Get workflow context (reuse existing logic)
+        with Session(db.engine) as session:
+            stmt = select(App).where(App.id == workflow_id)
+            app = session.scalar(stmt)
+            if not app:
+                return {"questions": [], "error": f"App {workflow_id} not found"}
+
+            workflow = WorkflowService().get_draft_workflow(app_model=app)
+            if not workflow:
+                return {"questions": [], "error": f"Workflow for app {workflow_id} not found"}
+
+        upstream_nodes = cls._get_upstream_nodes(workflow.graph_dict, node_id)
+        current_node = cls._get_node_by_id(workflow.graph_dict, node_id)
+        if not current_node:
+            return {"questions": [], "error": f"Node {node_id} not found"}
+
+        parameter_info = cls._get_parameter_info(
+            tenant_id=tenant_id,
+            node_data=current_node.get("data", {}),
+            parameter_name=parameter_name,
+        )
+
+        # Build prompt
+        system_prompt = cls._build_suggested_questions_prompt(
+            upstream_nodes=upstream_nodes,
+            current_node=current_node,
+            parameter_info=parameter_info,
+            language=language,
+        )
+
+        prompt_messages: list[PromptMessage] = [
+            SystemPromptMessage(content=system_prompt),
+        ]
+
+        # Get model instance - use default if model_config not provided
+        model_manager = ModelManager()
+        if model_config:
+            provider = model_config.get("provider", "")
+            model_name = model_config.get("name", "")
+            model_instance = model_manager.get_model_instance(
+                tenant_id=tenant_id,
+                model_type=ModelType.LLM,
+                provider=provider,
+                model=model_name,
+            )
+        else:
+            model_instance = model_manager.get_default_model_instance(
+                tenant_id=tenant_id,
+                model_type=ModelType.LLM,
+            )
+            model_name = model_instance.model
+
+        model_schema = model_instance.model_type_instance.get_model_schema(model_name, model_instance.credentials)
+        if not model_schema:
+            return {"questions": [], "error": f"Model schema not found for {model_name}"}
+
+        completion_params = model_config.get("completion_params", {}) if model_config else {}
+        model_parameters = {**completion_params, "max_tokens": 256}
+        json_schema = cls._get_suggested_questions_json_schema()
+
+        try:
+            response = invoke_llm_with_structured_output(
+                provider=model_instance.provider,
+                model_schema=model_schema,
+                model_instance=model_instance,
+                prompt_messages=prompt_messages,
+                json_schema=json_schema,
+                model_parameters=model_parameters,
+                stream=False,
+                tenant_id=tenant_id,
+            )
+
+            questions = response.structured_output.get("questions", []) if response.structured_output else []
+            return {"questions": questions, "error": ""}
+
+        except InvokeError as e:
+            return {"questions": [], "error": str(e)}
+        except Exception as e:
+            logger.exception("Failed to generate suggested questions, model: %s", model_name)
+            return {"questions": [], "error": f"An unexpected error occurred: {str(e)}"}
+
+    @classmethod
+    def _build_suggested_questions_prompt(
+        cls,
+        upstream_nodes: list[dict],
+        current_node: dict,
+        parameter_info: dict,
+        language: str = "English",
+    ) -> str:
+        """Build minimal prompt for suggested questions generation."""
+        # Simplify upstream nodes to reduce tokens
+        sources = [f"{n['title']}({','.join(n.get('outputs', {}).keys())})" for n in upstream_nodes[:5]]
+        param_type = parameter_info.get("type", "string")
+        param_desc = parameter_info.get("description", "")[:100]
+
+        return f"""Suggest 3 code generation questions for extracting data.
+Sources: {", ".join(sources)}
+Target: {parameter_info.get("name")}({param_type}) - {param_desc}
+Output 3 short, practical questions in {language}."""
+
+    @classmethod
+    def _get_suggested_questions_json_schema(cls) -> dict:
+        """Return JSON Schema for suggested questions."""
+        return {
+            "type": "object",
+            "properties": {
+                "questions": {
+                    "type": "array",
+                    "items": {"type": "string"},
+                    "minItems": 3,
+                    "maxItems": 3,
+                    "description": "3 suggested questions",
+                },
+            },
+            "required": ["questions"],
+        }
+
+    @classmethod
+    def _get_code_node_json_schema(cls) -> dict:
+        """Return JSON Schema for structured output."""
+        return {
+            "type": "object",
+            "properties": {
+                "variables": {
+                    "type": "array",
+                    "items": {
+                        "type": "object",
+                        "properties": {
+                            "variable": {"type": "string", "description": "Variable name in code"},
+                            "value_selector": {
+                                "type": "array",
+                                "items": {"type": "string"},
+                                "description": "Path like [node_id, output_name]",
+                            },
+                        },
+                        "required": ["variable", "value_selector"],
+                    },
+                },
+                "code": {"type": "string", "description": "Generated code with main function"},
+                "outputs": {
+                    "type": "object",
+                    "additionalProperties": {
+                        "type": "object",
+                        "properties": {"type": {"type": "string"}},
+                    },
+                    "description": "Output definitions, key is output name",
+                },
+                "explanation": {"type": "string", "description": "Brief explanation of the code"},
+            },
+            "required": ["variables", "code", "outputs", "explanation"],
+        }
+
+    @classmethod
+    def _get_upstream_nodes(cls, graph_dict: Mapping[str, Any], node_id: str) -> list[dict]:
+        """
+        Get all upstream nodes via edge backtracking.
+
+        Traverses the graph backwards from node_id to collect all reachable nodes.
+        """
+        from collections import defaultdict
+
+        nodes = {n["id"]: n for n in graph_dict.get("nodes", [])}
+        edges = graph_dict.get("edges", [])
+
+        # Build reverse adjacency list
+        reverse_adj: dict[str, list[str]] = defaultdict(list)
+        for edge in edges:
+            reverse_adj[edge["target"]].append(edge["source"])
+
+        # BFS to find all upstream nodes
+        visited: set[str] = set()
+        queue = [node_id]
+        upstream: list[dict] = []
+
+        while queue:
+            current = queue.pop(0)
+            for source in reverse_adj.get(current, []):
+                if source not in visited:
+                    visited.add(source)
+                    queue.append(source)
+                    if source in nodes:
+                        upstream.append(cls._extract_node_info(nodes[source]))
+
+        return upstream
+
+    @classmethod
+    def _get_node_by_id(cls, graph_dict: Mapping[str, Any], node_id: str) -> dict | None:
+        """Get node by ID from graph."""
+        for node in graph_dict.get("nodes", []):
+            if node["id"] == node_id:
+                return node
+        return None
+
+    @classmethod
+    def _extract_node_info(cls, node: dict) -> dict:
+        """Extract minimal node info with outputs based on node type."""
+        node_type = node["data"]["type"]
+        node_data = node.get("data", {})
+
+        # Build outputs based on node type (only type, no description to reduce tokens)
+        outputs: dict[str, str] = {}
+        match node_type:
+            case "start":
+                for var in node_data.get("variables", []):
+                    name = var.get("variable", var.get("name", ""))
+                    outputs[name] = var.get("type", "string")
+            case "llm":
+                outputs["text"] = "string"
+            case "code":
+                for name, output in node_data.get("outputs", {}).items():
+                    outputs[name] = output.get("type", "string")
+            case "http-request":
+                outputs = {"body": "string", "status_code": "number", "headers": "object"}
+            case "knowledge-retrieval":
+                outputs["result"] = "array[object]"
+            case "tool":
+                outputs = {"text": "string", "json": "object"}
+            case _:
+                outputs["output"] = "string"
+
+        info: dict = {
+            "id": node["id"],
+            "title": node_data.get("title", node["id"]),
+            "outputs": outputs,
+        }
+        # Only include description if not empty
+        desc = node_data.get("desc", "")
+        if desc:
+            info["desc"] = desc
+
+        return info
+
+    @classmethod
+    def _get_parameter_info(cls, tenant_id: str, node_data: dict, parameter_name: str) -> dict:
+        """Get parameter info from tool schema using ToolManager."""
+        default_info = {"name": parameter_name, "type": "string", "description": ""}
+
+        if node_data.get("type") != "tool":
+            return default_info
+
+        try:
+            from core.app.entities.app_invoke_entities import InvokeFrom
+            from core.tools.entities.tool_entities import ToolProviderType
+            from core.tools.tool_manager import ToolManager
+
+            provider_type_str = node_data.get("provider_type", "")
+            provider_type = ToolProviderType(provider_type_str) if provider_type_str else ToolProviderType.BUILT_IN
+
+            tool_runtime = ToolManager.get_tool_runtime(
+                provider_type=provider_type,
+                provider_id=node_data.get("provider_id", ""),
+                tool_name=node_data.get("tool_name", ""),
+                tenant_id=tenant_id,
+                invoke_from=InvokeFrom.DEBUGGER,
+            )
+
+            parameters = tool_runtime.get_merged_runtime_parameters()
+            for param in parameters:
+                if param.name == parameter_name:
+                    return {
+                        "name": param.name,
+                        "type": param.type.value if hasattr(param.type, "value") else str(param.type),
+                        "description": param.llm_description
+                        or (param.human_description.en_US if param.human_description else ""),
+                        "required": param.required,
+                    }
+        except Exception as e:
+            logger.debug("Failed to get parameter info from ToolManager: %s", e)
+
+        return default_info
+
+    @classmethod
+    def _build_extractor_system_prompt(
+        cls,
+        upstream_nodes: list[dict],
+        current_node: dict,
+        parameter_info: dict,
+        language: str,
+    ) -> str:
+        """Build system prompt for extractor code generation."""
+        upstream_json = json.dumps(upstream_nodes, indent=2, ensure_ascii=False)
+        param_type = parameter_info.get("type", "string")
+        return f"""You are a code generator for workflow automation.
+
+Generate {language} code to extract/transform upstream node outputs for the target parameter.
+
+## Upstream Nodes
+{upstream_json}
+
+## Target
+Node: {current_node["data"].get("title", current_node["id"])}
+Parameter: {parameter_info.get("name")} ({param_type}) - {parameter_info.get("description", "")}
+
+## Requirements
+- Write a main function that returns type: {param_type}
+- Use value_selector format: ["node_id", "output_name"]
+"""
+
+    @classmethod
+    def _parse_code_node_output(cls, content: Mapping[str, Any] | None, language: str, parameter_type: str) -> dict:
+        """
+        Parse structured output to CodeNodeData format.
+
+        Args:
+            content: Structured output dict from invoke_llm_with_structured_output
+            language: Code language
+            parameter_type: Expected parameter type
+
+        Returns dict with variables, code_language, code, outputs, message, error.
+        """
+        if content is None:
+            return cls._error_response("Empty or invalid response from LLM")
+
+        # Validate and normalize variables
+        variables = [
+            {"variable": v.get("variable", ""), "value_selector": v.get("value_selector", [])}
+            for v in content.get("variables", [])
+            if isinstance(v, dict)
+        ]
+
+        outputs = content.get("outputs", {"result": {"type": parameter_type}})
+
+        return {
+            "variables": variables,
+            "code_language": language,
+            "code": content.get("code", ""),
+            "outputs": outputs,
+            "message": content.get("explanation", ""),
+            "error": "",
+        }
+
    @staticmethod
    def instruction_modify_legacy(
        tenant_id: str, flow_id: str, current: str, instruction: str, model_config: dict, ideal_output: str | None
@@ -0,0 +1,188 @@
+"""
+File reference detection and conversion for structured output.
+
+This module provides utilities to:
+1. Detect file reference fields in JSON Schema (format: "dify-file-ref")
+2. Convert file ID strings to File objects after LLM returns
+"""
+
+import uuid
+from collections.abc import Mapping
+from typing import Any
+
+from core.file import File
+from core.variables.segments import ArrayFileSegment, FileSegment
+from factories.file_factory import build_from_mapping
+
+FILE_REF_FORMAT = "dify-file-ref"
+
+
+def is_file_ref_property(schema: dict) -> bool:
+    """Check if a schema property is a file reference."""
+    return schema.get("type") == "string" and schema.get("format") == FILE_REF_FORMAT
+
+
+def detect_file_ref_fields(schema: Mapping[str, Any], path: str = "") -> list[str]:
+    """
+    Recursively detect file reference fields in schema.
+
+    Args:
+        schema: JSON Schema to analyze
+        path: Current path in the schema (used for recursion)
+
+    Returns:
+        List of JSON paths containing file refs, e.g., ["image_id", "files[*]"]
+    """
+    file_ref_paths: list[str] = []
+    schema_type = schema.get("type")
+
+    if schema_type == "object":
+        for prop_name, prop_schema in schema.get("properties", {}).items():
+            current_path = f"{path}.{prop_name}" if path else prop_name
+
+            if is_file_ref_property(prop_schema):
+                file_ref_paths.append(current_path)
+            elif isinstance(prop_schema, dict):
+                file_ref_paths.extend(detect_file_ref_fields(prop_schema, current_path))
+
+    elif schema_type == "array":
+        items_schema = schema.get("items", {})
+        array_path = f"{path}[*]" if path else "[*]"
+
+        if is_file_ref_property(items_schema):
+            file_ref_paths.append(array_path)
+        elif isinstance(items_schema, dict):
+            file_ref_paths.extend(detect_file_ref_fields(items_schema, array_path))
+
+    return file_ref_paths
+
+
+def convert_file_refs_in_output(
+    output: Mapping[str, Any],
+    json_schema: Mapping[str, Any],
+    tenant_id: str,
+) -> dict[str, Any]:
+    """
+    Convert file ID strings to File objects based on schema.
+
+    Args:
+        output: The structured_output from LLM result
+        json_schema: The original JSON schema (to detect file ref fields)
+        tenant_id: Tenant ID for file lookup
+
+    Returns:
+        Output with file references converted to File objects
+    """
+    file_ref_paths = detect_file_ref_fields(json_schema)
+    if not file_ref_paths:
+        return dict(output)
+
+    result = _deep_copy_dict(output)
+
+    for path in file_ref_paths:
+        _convert_path_in_place(result, path.split("."), tenant_id)
+
+    return result
+
+
+def _deep_copy_dict(obj: Mapping[str, Any]) -> dict[str, Any]:
+    """Deep copy a mapping to a mutable dict."""
+    result: dict[str, Any] = {}
+    for key, value in obj.items():
+        if isinstance(value, Mapping):
+            result[key] = _deep_copy_dict(value)
+        elif isinstance(value, list):
+            result[key] = [_deep_copy_dict(item) if isinstance(item, Mapping) else item for item in value]
+        else:
+            result[key] = value
+    return result
+
+
+def _convert_path_in_place(obj: dict, path_parts: list[str], tenant_id: str) -> None:
+    """Convert file refs at the given path in place, wrapping in Segment types."""
+    if not path_parts:
+        return
+
+    current = path_parts[0]
+    remaining = path_parts[1:]
+
+    # Handle array notation like "files[*]"
+    if current.endswith("[*]"):
+        key = current[:-3] if current != "[*]" else None
+        target = obj.get(key) if key else obj
+
+        if isinstance(target, list):
+            if remaining:
+                # Nested array with remaining path - recurse into each item
+                for item in target:
+                    if isinstance(item, dict):
+                        _convert_path_in_place(item, remaining, tenant_id)
+            else:
+                # Array of file IDs - convert all and wrap in ArrayFileSegment
+                files: list[File] = []
+                for item in target:
+                    file = _convert_file_id(item, tenant_id)
+                    if file is not None:
+                        files.append(file)
+                # Replace the array with ArrayFileSegment
+                if key:
+                    obj[key] = ArrayFileSegment(value=files)
+        return
+
+    if not remaining:
+        # Leaf node - convert the value and wrap in FileSegment
+        if current in obj:
+            file = _convert_file_id(obj[current], tenant_id)
+            if file is not None:
+                obj[current] = FileSegment(value=file)
+            else:
+                obj[current] = None
+    else:
+        # Recurse into nested object
+        if current in obj and isinstance(obj[current], dict):
+            _convert_path_in_place(obj[current], remaining, tenant_id)
+
+
+def _convert_file_id(file_id: Any, tenant_id: str) -> File | None:
+    """
+    Convert a file ID string to a File object.
+
+    Tries multiple file sources in order:
+    1. ToolFile (files generated by tools/workflows)
+    2. UploadFile (files uploaded by users)
+    """
+    if not isinstance(file_id, str):
+        return None
+
+    # Validate UUID format
+    try:
+        uuid.UUID(file_id)
+    except ValueError:
+        return None
+
+    # Try ToolFile first (files generated by tools/workflows)
+    try:
+        return build_from_mapping(
+            mapping={
+                "transfer_method": "tool_file",
+                "tool_file_id": file_id,
+            },
+            tenant_id=tenant_id,
+        )
+    except ValueError:
+        pass
+
+    # Try UploadFile (files uploaded by users)
+    try:
+        return build_from_mapping(
+            mapping={
+                "transfer_method": "local_file",
+                "upload_file_id": file_id,
+            },
+            tenant_id=tenant_id,
+        )
+    except ValueError:
+        pass
+
+    # File not found in any source
+    return None
@@ -8,6 +8,7 @@ import json_repair
 from pydantic import TypeAdapter, ValidationError

 from core.llm_generator.output_parser.errors import OutputParserError
+from core.llm_generator.output_parser.file_ref import convert_file_refs_in_output
 from core.llm_generator.prompts import STRUCTURED_OUTPUT_PROMPT
 from core.model_manager import ModelInstance
 from core.model_runtime.callbacks.base_callback import Callback
@@ -57,6 +58,7 @@ def invoke_llm_with_structured_output(
    stream: Literal[True],
    user: str | None = None,
    callbacks: list[Callback] | None = None,
+    tenant_id: str | None = None,
 ) -> Generator[LLMResultChunkWithStructuredOutput, None, None]: ...
@overload
 def invoke_llm_with_structured_output(
@@ -72,6 +74,7 @@ def invoke_llm_with_structured_output(
    stream: Literal[False],
    user: str | None = None,
    callbacks: list[Callback] | None = None,
+    tenant_id: str | None = None,
 ) -> LLMResultWithStructuredOutput: ...
@overload
 def invoke_llm_with_structured_output(
@@ -87,6 +90,7 @@ def invoke_llm_with_structured_output(
    stream: bool = True,
    user: str | None = None,
    callbacks: list[Callback] | None = None,
+    tenant_id: str | None = None,
 ) -> LLMResultWithStructuredOutput | Generator[LLMResultChunkWithStructuredOutput, None, None]: ...
 def invoke_llm_with_structured_output(
    *,
@@ -101,20 +105,28 @@ def invoke_llm_with_structured_output(
    stream: bool = True,
    user: str | None = None,
    callbacks: list[Callback] | None = None,
+    tenant_id: str | None = None,
 ) -> LLMResultWithStructuredOutput | Generator[LLMResultChunkWithStructuredOutput, None, None]:
    """
-    Invoke large language model with structured output
-    1. This method invokes model_instance.invoke_llm with json_schema
-    2. Try to parse the result as structured output
+    Invoke large language model with structured output.

+    This method invokes model_instance.invoke_llm with json_schema and parses
+    the result as structured output.
+
+    :param provider: model provider name
+    :param model_schema: model schema entity
+    :param model_instance: model instance to invoke
    :param prompt_messages: prompt messages
-    :param json_schema: json schema
+    :param json_schema: json schema for structured output
    :param model_parameters: model parameters
    :param tools: tools for tool calling
    :param stop: stop words
    :param stream: is stream response
    :param user: unique user id
    :param callbacks: callbacks
+    :param tenant_id: tenant ID for file reference conversion. When provided and
+                      json_schema contains file reference fields (format: "dify-file-ref"),
+                      file IDs in the output will be automatically converted to File objects.
    :return: full response or stream response chunk generator result
    """

@@ -153,8 +165,18 @@ def invoke_llm_with_structured_output(
                f"Failed to parse structured output, LLM result is not a string: {llm_result.message.content}"
            )

+        structured_output = _parse_structured_output(llm_result.message.content)
+
+        # Convert file references if tenant_id is provided
+        if tenant_id is not None:
+            structured_output = convert_file_refs_in_output(
+                output=structured_output,
+                json_schema=json_schema,
+                tenant_id=tenant_id,
+            )
+
        return LLMResultWithStructuredOutput(
-            structured_output=_parse_structured_output(llm_result.message.content),
+            structured_output=structured_output,
            model=llm_result.model,
            message=llm_result.message,
            usage=llm_result.usage,
@@ -186,8 +208,18 @@ def invoke_llm_with_structured_output(
                    delta=event.delta,
                )

+            structured_output = _parse_structured_output(result_text)
+
+            # Convert file references if tenant_id is provided
+            if tenant_id is not None:
+                structured_output = convert_file_refs_in_output(
+                    output=structured_output,
+                    json_schema=json_schema,
+                    tenant_id=tenant_id,
+                )
+
            yield LLMResultChunkWithStructuredOutput(
-                structured_output=_parse_structured_output(result_text),
+                structured_output=structured_output,
                model=model_schema.model,
                prompt_messages=prompt_messages,
                system_fingerprint=system_fingerprint,
@@ -0,0 +1,45 @@
+"""Utility functions for LLM generator."""
+
+from core.model_runtime.entities.message_entities import (
+    AssistantPromptMessage,
+    PromptMessage,
+    PromptMessageRole,
+    SystemPromptMessage,
+    ToolPromptMessage,
+    UserPromptMessage,
+)
+
+
+def deserialize_prompt_messages(messages: list[dict]) -> list[PromptMessage]:
+    """
+    Deserialize list of dicts to list[PromptMessage].
+
+    Expected format:
+        [
+            {"role": "user", "content": "..."},
+            {"role": "assistant", "content": "..."},
+        ]
+    """
+    result: list[PromptMessage] = []
+    for msg in messages:
+        role = PromptMessageRole.value_of(msg["role"])
+        content = msg.get("content", "")
+
+        match role:
+            case PromptMessageRole.USER:
+                result.append(UserPromptMessage(content=content))
+            case PromptMessageRole.ASSISTANT:
+                result.append(AssistantPromptMessage(content=content))
+            case PromptMessageRole.SYSTEM:
+                result.append(SystemPromptMessage(content=content))
+            case PromptMessageRole.TOOL:
+                result.append(ToolPromptMessage(content=content, tool_call_id=msg.get("tool_call_id", "")))
+
+    return result
+
+
+def serialize_prompt_messages(messages: list[PromptMessage]) -> list[dict]:
+    """
+    Serialize list[PromptMessage] to list of dicts.
+    """
+    return [{"role": msg.role.value, "content": msg.content} for msg in messages]
@@ -0,0 +1,434 @@
+# Memory Module
+
+This module provides memory management for LLM conversations, enabling context retention across dialogue turns.
+
+## Overview
+
+The memory module contains two types of memory implementations:
+
+1. **TokenBufferMemory** - Conversation-level memory (existing)
+2. **NodeTokenBufferMemory** - Node-level memory (to be implemented, **Chatflow only**)
+
+> **Note**: `NodeTokenBufferMemory` is only available in **Chatflow** (advanced-chat mode).
+> This is because it requires both `conversation_id` and `node_id`, which are only present in Chatflow.
+> Standard Workflow mode does not have `conversation_id` and therefore cannot use node-level memory.
+
+```
+┌─────────────────────────────────────────────────────────────────────────────┐
+│                           Memory Architecture                               │
+├─────────────────────────────────────────────────────────────────────────────┤
+│                                                                             │
+│  ┌─────────────────────────────────────────────────────────────────────-┐   │
+│  │                      TokenBufferMemory                               │   │
+│  │  Scope: Conversation                                                 │   │
+│  │  Storage: Database (Message table)                                   │   │
+│  │  Key: conversation_id                                                │   │
+│  └─────────────────────────────────────────────────────────────────────-┘   │
+│                                                                             │
+│  ┌─────────────────────────────────────────────────────────────────────-┐   │
+│  │                    NodeTokenBufferMemory                             │   │
+│  │  Scope: Node within Conversation                                     │   │
+│  │  Storage: Object Storage (JSON file)                                 │   │
+│  │  Key: (app_id, conversation_id, node_id)                             │   │
+│  └─────────────────────────────────────────────────────────────────────-┘   │
+│                                                                             │
+└─────────────────────────────────────────────────────────────────────────────┘
+```
+
+---
+
+## TokenBufferMemory (Existing)
+
+### Purpose
+
+`TokenBufferMemory` retrieves conversation history from the `Message` table and converts it to `PromptMessage` objects for LLM context.
+
+### Key Features
+
+- **Conversation-scoped**: All messages within a conversation are candidates
+- **Thread-aware**: Uses `parent_message_id` to extract only the current thread (supports regeneration scenarios)
+- **Token-limited**: Truncates history to fit within `max_token_limit`
+- **File support**: Handles `MessageFile` attachments (images, documents, etc.)
+
+### Data Flow
+
+```
+Message Table                    TokenBufferMemory              LLM
+     │                                  │                        │
+     │  SELECT * FROM messages          │                        │
+     │  WHERE conversation_id = ?       │                        │
+     │  ORDER BY created_at DESC        │                        │
+     ├─────────────────────────────────▶│                        │
+     │                                  │                        │
+     │                    extract_thread_messages()              │
+     │                                  │                        │
+     │                    build_prompt_message_with_files()      │
+     │                                  │                        │
+     │                    truncate by max_token_limit            │
+     │                                  │                        │
+     │                                  │  Sequence[PromptMessage]
+     │                                  ├───────────────────────▶│
+     │                                  │                        │
+```
+
+### Thread Extraction
+
+When a user regenerates a response, a new thread is created:
+
+```
+Message A (user)
+    └── Message A' (assistant)
+            └── Message B (user)
+                    └── Message B' (assistant)
+            └── Message A'' (assistant, regenerated)  ← New thread
+                    └── Message C (user)
+                            └── Message C' (assistant)
+```
+
+`extract_thread_messages()` traces back from the latest message using `parent_message_id` to get only the current thread: `[A, A'', C, C']`
+
+### Usage
+
+```python
+from core.memory.token_buffer_memory import TokenBufferMemory
+
+memory = TokenBufferMemory(conversation=conversation, model_instance=model_instance)
+history = memory.get_history_prompt_messages(max_token_limit=2000, message_limit=100)
+```
+
+---
+
+## NodeTokenBufferMemory (To Be Implemented)
+
+### Purpose
+
+`NodeTokenBufferMemory` provides **node-scoped memory** within a conversation. Each LLM node in a workflow can maintain its own independent conversation history.
+
+### Use Cases
+
+1. **Multi-LLM Workflows**: Different LLM nodes need separate context
+2. **Iterative Processing**: An LLM node in a loop needs to accumulate context across iterations
+3. **Specialized Agents**: Each agent node maintains its own dialogue history
+
+### Design Decisions
+
+#### Storage: Object Storage for Messages (No New Database Table)
+
+| Aspect                    | Database             | Object Storage     |
+| ------------------------- | -------------------- | ------------------ |
+| Cost                      | High                 | Low                |
+| Query Flexibility         | High                 | Low                |
+| Schema Changes            | Migration required   | None               |
+| Consistency with existing | ConversationVariable | File uploads, logs |
+
+**Decision**: Store message data in object storage, but still use existing database tables for file metadata.
+
+**What is stored in Object Storage:**
+
+- Message content (text)
+- Message metadata (role, token_count, created_at)
+- File references (upload_file_id, tool_file_id, etc.)
+- Thread relationships (message_id, parent_message_id)
+
+**What still requires Database queries:**
+
+- File reconstruction: When reading node memory, file references are used to query
+  `UploadFile` / `ToolFile` tables via `file_factory.build_from_mapping()` to rebuild
+  complete `File` objects with storage_key, mime_type, etc.
+
+**Why this hybrid approach:**
+
+- No database migration required (no new tables)
+- Message data may be large, object storage is cost-effective
+- File metadata is already in database, no need to duplicate
+- Aligns with existing storage patterns (file uploads, logs)
+
+#### Storage Key Format
+
+```
+node_memory/{app_id}/{conversation_id}/{node_id}.json
+```
+
+#### Data Structure
+
+```json
+{
+  "version": 1,
+  "messages": [
+    {
+      "message_id": "msg-001",
+      "parent_message_id": null,
+      "role": "user",
+      "content": "Analyze this image",
+      "files": [
+        {
+          "type": "image",
+          "transfer_method": "local_file",
+          "upload_file_id": "file-uuid-123",
+          "belongs_to": "user"
+        }
+      ],
+      "token_count": 15,
+      "created_at": "2026-01-07T10:00:00Z"
+    },
+    {
+      "message_id": "msg-002",
+      "parent_message_id": "msg-001",
+      "role": "assistant",
+      "content": "This is a landscape image...",
+      "files": [],
+      "token_count": 50,
+      "created_at": "2026-01-07T10:00:01Z"
+    }
+  ]
+}
+```
+
+### Thread Support
+
+Node memory also supports thread extraction (for regeneration scenarios):
+
+```python
+def _extract_thread(
+    self,
+    messages: list[NodeMemoryMessage],
+    current_message_id: str
+) -> list[NodeMemoryMessage]:
+    """
+    Extract messages belonging to the thread of current_message_id.
+    Similar to extract_thread_messages() in TokenBufferMemory.
+    """
+    ...
+```
+
+### File Handling
+
+Files are stored as references (not full metadata):
+
+```python
+class NodeMemoryFile(BaseModel):
+    type: str                        # image, audio, video, document, custom
+    transfer_method: str             # local_file, remote_url, tool_file
+    upload_file_id: str | None       # for local_file
+    tool_file_id: str | None         # for tool_file
+    url: str | None                  # for remote_url
+    belongs_to: str                  # user / assistant
+```
+
+When reading, files are rebuilt using `file_factory.build_from_mapping()`.
+
+### API Design
+
+```python
+class NodeTokenBufferMemory:
+    def __init__(
+        self,
+        app_id: str,
+        conversation_id: str,
+        node_id: str,
+        model_instance: ModelInstance,
+    ):
+        """
+        Initialize node-level memory.
+
+        :param app_id: Application ID
+        :param conversation_id: Conversation ID
+        :param node_id: Node ID in the workflow
+        :param model_instance: Model instance for token counting
+        """
+        ...
+
+    def add_messages(
+        self,
+        message_id: str,
+        parent_message_id: str | None,
+        user_content: str,
+        user_files: Sequence[File],
+        assistant_content: str,
+        assistant_files: Sequence[File],
+    ) -> None:
+        """
+        Append a dialogue turn (user + assistant) to node memory.
+        Call this after LLM node execution completes.
+
+        :param message_id: Current message ID (from Message table)
+        :param parent_message_id: Parent message ID (for thread tracking)
+        :param user_content: User's text input
+        :param user_files: Files attached by user
+        :param assistant_content: Assistant's text response
+        :param assistant_files: Files generated by assistant
+        """
+        ...
+
+    def get_history_prompt_messages(
+        self,
+        current_message_id: str,
+        tenant_id: str,
+        max_token_limit: int = 2000,
+        file_upload_config: FileUploadConfig | None = None,
+    ) -> Sequence[PromptMessage]:
+        """
+        Retrieve history as PromptMessage sequence.
+
+        :param current_message_id: Current message ID (for thread extraction)
+        :param tenant_id: Tenant ID (for file reconstruction)
+        :param max_token_limit: Maximum tokens for history
+        :param file_upload_config: File upload configuration
+        :return: Sequence of PromptMessage for LLM context
+        """
+        ...
+
+    def flush(self) -> None:
+        """
+        Persist buffered changes to object storage.
+        Call this at the end of node execution.
+        """
+        ...
+
+    def clear(self) -> None:
+        """
+        Clear all messages in this node's memory.
+        """
+        ...
+```
+
+### Data Flow
+
+```
+Object Storage                  NodeTokenBufferMemory           LLM Node
+     │                                  │                           │
+     │                                  │◀── get_history_prompt_messages()
+     │  storage.load(key)               │                           │
+     │◀─────────────────────────────────┤                           │
+     │                                  │                           │
+     │  JSON data                       │                           │
+     ├─────────────────────────────────▶│                           │
+     │                                  │                           │
+     │                    _extract_thread()                         │
+     │                                  │                           │
+     │                    _rebuild_files() via file_factory         │
+     │                                  │                           │
+     │                    _build_prompt_messages()                  │
+     │                                  │                           │
+     │                    _truncate_by_tokens()                     │
+     │                                  │                           │
+     │                                  │  Sequence[PromptMessage]  │
+     │                                  ├──────────────────────────▶│
+     │                                  │                           │
+     │                                  │◀── LLM execution complete │
+     │                                  │                           │
+     │                                  │◀── add_messages()         │
+     │                                  │                           │
+     │  storage.save(key, data)         │                           │
+     │◀─────────────────────────────────┤                           │
+     │                                  │                           │
+```
+
+### Integration with LLM Node
+
+```python
+# In LLM Node execution
+
+# 1. Fetch memory based on mode
+if node_data.memory and node_data.memory.mode == MemoryMode.NODE:
+    # Node-level memory (Chatflow only)
+    memory = fetch_node_memory(
+        variable_pool=variable_pool,
+        app_id=app_id,
+        node_id=self.node_id,
+        node_data_memory=node_data.memory,
+        model_instance=model_instance,
+    )
+elif node_data.memory and node_data.memory.mode == MemoryMode.CONVERSATION:
+    # Conversation-level memory (existing behavior)
+    memory = fetch_memory(
+        variable_pool=variable_pool,
+        app_id=app_id,
+        node_data_memory=node_data.memory,
+        model_instance=model_instance,
+    )
+else:
+    memory = None
+
+# 2. Get history for context
+if memory:
+    if isinstance(memory, NodeTokenBufferMemory):
+        history = memory.get_history_prompt_messages(
+            current_message_id=current_message_id,
+            tenant_id=tenant_id,
+            max_token_limit=max_token_limit,
+        )
+    else:  # TokenBufferMemory
+        history = memory.get_history_prompt_messages(
+            max_token_limit=max_token_limit,
+        )
+    prompt_messages = [*history, *current_messages]
+else:
+    prompt_messages = current_messages
+
+# 3. Call LLM
+response = model_instance.invoke(prompt_messages)
+
+# 4. Append to node memory (only for NodeTokenBufferMemory)
+if isinstance(memory, NodeTokenBufferMemory):
+    memory.add_messages(
+        message_id=message_id,
+        parent_message_id=parent_message_id,
+        user_content=user_input,
+        user_files=user_files,
+        assistant_content=response.content,
+        assistant_files=response_files,
+    )
+    memory.flush()
+```
+
+### Configuration
+
+Add to `MemoryConfig` in `core/workflow/nodes/llm/entities.py`:
+
+```python
+class MemoryMode(StrEnum):
+    CONVERSATION = "conversation"  # Use TokenBufferMemory (default, existing behavior)
+    NODE = "node"                  # Use NodeTokenBufferMemory (new, Chatflow only)
+
+class MemoryConfig(BaseModel):
+    # Existing fields
+    role_prefix: RolePrefix | None = None
+    window: MemoryWindowConfig | None = None
+    query_prompt_template: str | None = None
+
+    # Memory mode (new)
+    mode: MemoryMode = MemoryMode.CONVERSATION
+```
+
+**Mode Behavior:**
+
+| Mode           | Memory Class          | Scope                    | Availability  |
+| -------------- | --------------------- | ------------------------ | ------------- |
+| `conversation` | TokenBufferMemory     | Entire conversation      | All app modes |
+| `node`         | NodeTokenBufferMemory | Per-node in conversation | Chatflow only |
+
+> When `mode=node` is used in a non-Chatflow context (no conversation_id), it should
+> fall back to no memory or raise a configuration error.
+
+---
+
+## Comparison
+
+| Feature        | TokenBufferMemory        | NodeTokenBufferMemory     |
+| -------------- | ------------------------ | ------------------------- |
+| Scope          | Conversation             | Node within Conversation  |
+| Storage        | Database (Message table) | Object Storage (JSON)     |
+| Thread Support | Yes                      | Yes                       |
+| File Support   | Yes (via MessageFile)    | Yes (via file references) |
+| Token Limit    | Yes                      | Yes                       |
+| Use Case       | Standard chat apps       | Complex workflows         |
+
+---
+
+## Future Considerations
+
+1. **Cleanup Task**: Add a Celery task to clean up old node memory files
+2. **Concurrency**: Consider Redis lock for concurrent node executions
+3. **Compression**: Compress large memory files to reduce storage costs
+4. **Extension**: Other nodes (Agent, Tool) may also benefit from node-level memory
@@ -0,0 +1,15 @@
+from core.memory.base import BaseMemory
+from core.memory.node_token_buffer_memory import (
+    NodeMemoryData,
+    NodeMemoryFile,
+    NodeTokenBufferMemory,
+)
+from core.memory.token_buffer_memory import TokenBufferMemory
+
+__all__ = [
+    "BaseMemory",
+    "NodeMemoryData",
+    "NodeMemoryFile",
+    "NodeTokenBufferMemory",
+    "TokenBufferMemory",
+]
@@ -0,0 +1,83 @@
+"""
+Base memory interfaces and types.
+
+This module defines the common protocol for memory implementations.
+"""
+
+from abc import ABC, abstractmethod
+from collections.abc import Sequence
+
+from core.model_runtime.entities import ImagePromptMessageContent, PromptMessage
+
+
+class BaseMemory(ABC):
+    """
+    Abstract base class for memory implementations.
+
+    Provides a common interface for both conversation-level and node-level memory.
+    """
+
+    @abstractmethod
+    def get_history_prompt_messages(
+        self,
+        *,
+        max_token_limit: int = 2000,
+        message_limit: int | None = None,
+    ) -> Sequence[PromptMessage]:
+        """
+        Get history prompt messages.
+
+        :param max_token_limit: Maximum tokens for history
+        :param message_limit: Maximum number of messages
+        :return: Sequence of PromptMessage for LLM context
+        """
+        pass
+
+    def get_history_prompt_text(
+        self,
+        human_prefix: str = "Human",
+        ai_prefix: str = "Assistant",
+        max_token_limit: int = 2000,
+        message_limit: int | None = None,
+    ) -> str:
+        """
+        Get history prompt as formatted text.
+
+        :param human_prefix: Prefix for human messages
+        :param ai_prefix: Prefix for assistant messages
+        :param max_token_limit: Maximum tokens for history
+        :param message_limit: Maximum number of messages
+        :return: Formatted history text
+        """
+        from core.model_runtime.entities import (
+            PromptMessageRole,
+            TextPromptMessageContent,
+        )
+
+        prompt_messages = self.get_history_prompt_messages(
+            max_token_limit=max_token_limit,
+            message_limit=message_limit,
+        )
+
+        string_messages = []
+        for m in prompt_messages:
+            if m.role == PromptMessageRole.USER:
+                role = human_prefix
+            elif m.role == PromptMessageRole.ASSISTANT:
+                role = ai_prefix
+            else:
+                continue
+
+            if isinstance(m.content, list):
+                inner_msg = ""
+                for content in m.content:
+                    if isinstance(content, TextPromptMessageContent):
+                        inner_msg += f"{content.data}\n"
+                    elif isinstance(content, ImagePromptMessageContent):
+                        inner_msg += "[image]\n"
+                string_messages.append(f"{role}: {inner_msg.strip()}")
+            else:
+                message = f"{role}: {m.content}"
+                string_messages.append(message)
+
+        return "\n".join(string_messages)
@@ -0,0 +1,353 @@
+"""
+Node-level Token Buffer Memory for Chatflow.
+
+This module provides node-scoped memory within a conversation.
+Each LLM node in a workflow can maintain its own independent conversation history.
+
+Note: This is only available in Chatflow (advanced-chat mode) because it requires
+both conversation_id and node_id.
+
+Design:
+- Storage is indexed by workflow_run_id (each execution stores one turn)
+- Thread tracking leverages Message table's parent_message_id structure
+- On read: query Message table for current thread, then filter Node Memory by workflow_run_ids
+"""
+
+import logging
+from collections.abc import Sequence
+
+from pydantic import BaseModel
+from sqlalchemy import select
+
+from core.file import File, FileTransferMethod
+from core.memory.base import BaseMemory
+from core.model_manager import ModelInstance
+from core.model_runtime.entities import (
+    AssistantPromptMessage,
+    ImagePromptMessageContent,
+    PromptMessage,
+    TextPromptMessageContent,
+    UserPromptMessage,
+)
+from core.prompt.utils.extract_thread_messages import extract_thread_messages
+from extensions.ext_database import db
+from extensions.ext_storage import storage
+from models.model import Message
+
+logger = logging.getLogger(__name__)
+
+
+class NodeMemoryFile(BaseModel):
+    """File reference stored in node memory."""
+
+    type: str  # image, audio, video, document, custom
+    transfer_method: str  # local_file, remote_url, tool_file
+    upload_file_id: str | None = None
+    tool_file_id: str | None = None
+    url: str | None = None
+
+
+class NodeMemoryTurn(BaseModel):
+    """A single dialogue turn (user + assistant) in node memory."""
+
+    user_content: str = ""
+    user_files: list[NodeMemoryFile] = []
+    assistant_content: str = ""
+    assistant_files: list[NodeMemoryFile] = []
+
+
+class NodeMemoryData(BaseModel):
+    """Root data structure for node memory storage."""
+
+    version: int = 1
+    # Key: workflow_run_id, Value: dialogue turn
+    turns: dict[str, NodeMemoryTurn] = {}
+
+
+class NodeTokenBufferMemory(BaseMemory):
+    """
+    Node-level Token Buffer Memory.
+
+    Provides node-scoped memory within a conversation. Each LLM node can maintain
+    its own independent conversation history, stored in object storage.
+
+    Key design: Thread tracking is delegated to Message table's parent_message_id.
+    Storage is indexed by workflow_run_id for easy filtering.
+
+    Storage key format: node_memory/{app_id}/{conversation_id}/{node_id}.json
+    """
+
+    def __init__(
+        self,
+        app_id: str,
+        conversation_id: str,
+        node_id: str,
+        tenant_id: str,
+        model_instance: ModelInstance,
+    ):
+        """
+        Initialize node-level memory.
+
+        :param app_id: Application ID
+        :param conversation_id: Conversation ID
+        :param node_id: Node ID in the workflow
+        :param tenant_id: Tenant ID for file reconstruction
+        :param model_instance: Model instance for token counting
+        """
+        self.app_id = app_id
+        self.conversation_id = conversation_id
+        self.node_id = node_id
+        self.tenant_id = tenant_id
+        self.model_instance = model_instance
+        self._storage_key = f"node_memory/{app_id}/{conversation_id}/{node_id}.json"
+        self._data: NodeMemoryData | None = None
+        self._dirty = False
+
+    def _load(self) -> NodeMemoryData:
+        """Load data from object storage."""
+        if self._data is not None:
+            return self._data
+
+        try:
+            raw = storage.load_once(self._storage_key)
+            self._data = NodeMemoryData.model_validate_json(raw)
+        except Exception:
+            # File not found or parse error, start fresh
+            self._data = NodeMemoryData()
+
+        return self._data
+
+    def _save(self) -> None:
+        """Save data to object storage."""
+        if self._data is not None:
+            storage.save(self._storage_key, self._data.model_dump_json().encode("utf-8"))
+            self._dirty = False
+
+    def _file_to_memory_file(self, file: File) -> NodeMemoryFile:
+        """Convert File object to NodeMemoryFile reference."""
+        return NodeMemoryFile(
+            type=file.type.value if hasattr(file.type, "value") else str(file.type),
+            transfer_method=(
+                file.transfer_method.value if hasattr(file.transfer_method, "value") else str(file.transfer_method)
+            ),
+            upload_file_id=file.related_id if file.transfer_method == FileTransferMethod.LOCAL_FILE else None,
+            tool_file_id=file.related_id if file.transfer_method == FileTransferMethod.TOOL_FILE else None,
+            url=file.remote_url if file.transfer_method == FileTransferMethod.REMOTE_URL else None,
+        )
+
+    def _memory_file_to_mapping(self, memory_file: NodeMemoryFile) -> dict:
+        """Convert NodeMemoryFile to mapping for file_factory."""
+        mapping: dict = {
+            "type": memory_file.type,
+            "transfer_method": memory_file.transfer_method,
+        }
+        if memory_file.upload_file_id:
+            mapping["upload_file_id"] = memory_file.upload_file_id
+        if memory_file.tool_file_id:
+            mapping["tool_file_id"] = memory_file.tool_file_id
+        if memory_file.url:
+            mapping["url"] = memory_file.url
+        return mapping
+
+    def _rebuild_files(self, memory_files: list[NodeMemoryFile]) -> list[File]:
+        """Rebuild File objects from NodeMemoryFile references."""
+        if not memory_files:
+            return []
+
+        from factories import file_factory
+
+        files = []
+        for mf in memory_files:
+            try:
+                mapping = self._memory_file_to_mapping(mf)
+                file = file_factory.build_from_mapping(mapping=mapping, tenant_id=self.tenant_id)
+                files.append(file)
+            except Exception as e:
+                logger.warning("Failed to rebuild file from memory: %s", e)
+                continue
+        return files
+
+    def _build_prompt_message(
+        self,
+        role: str,
+        content: str,
+        files: list[File],
+        detail: ImagePromptMessageContent.DETAIL = ImagePromptMessageContent.DETAIL.HIGH,
+    ) -> PromptMessage:
+        """Build PromptMessage from content and files."""
+        from core.file import file_manager
+
+        if not files:
+            if role == "user":
+                return UserPromptMessage(content=content)
+            else:
+                return AssistantPromptMessage(content=content)
+
+        # Build multimodal content
+        prompt_contents: list = []
+        for file in files:
+            try:
+                prompt_content = file_manager.to_prompt_message_content(file, image_detail_config=detail)
+                prompt_contents.append(prompt_content)
+            except Exception as e:
+                logger.warning("Failed to convert file to prompt content: %s", e)
+                continue
+
+        prompt_contents.append(TextPromptMessageContent(data=content))
+
+        if role == "user":
+            return UserPromptMessage(content=prompt_contents)
+        else:
+            return AssistantPromptMessage(content=prompt_contents)
+
+    def _get_thread_workflow_run_ids(self) -> list[str]:
+        """
+        Get workflow_run_ids for the current thread by querying Message table.
+
+        Returns workflow_run_ids in chronological order (oldest first).
+        """
+        # Query messages for this conversation
+        stmt = (
+            select(Message).where(Message.conversation_id == self.conversation_id).order_by(Message.created_at.desc())
+        )
+        messages = db.session.scalars(stmt.limit(500)).all()
+
+        if not messages:
+            return []
+
+        # Extract thread messages using existing logic
+        thread_messages = extract_thread_messages(messages)
+
+        # For newly created message, its answer is temporarily empty, skip it
+        if thread_messages and not thread_messages[0].answer and thread_messages[0].answer_tokens == 0:
+            thread_messages.pop(0)
+
+        # Reverse to get chronological order, extract workflow_run_ids
+        workflow_run_ids = []
+        for msg in reversed(thread_messages):
+            if msg.workflow_run_id:
+                workflow_run_ids.append(msg.workflow_run_id)
+
+        return workflow_run_ids
+
+    def add_messages(
+        self,
+        workflow_run_id: str,
+        user_content: str,
+        user_files: Sequence[File] | None = None,
+        assistant_content: str = "",
+        assistant_files: Sequence[File] | None = None,
+    ) -> None:
+        """
+        Add a dialogue turn to node memory.
+        Call this after LLM node execution completes.
+
+        :param workflow_run_id: Current workflow execution ID
+        :param user_content: User's text input
+        :param user_files: Files attached by user
+        :param assistant_content: Assistant's text response
+        :param assistant_files: Files generated by assistant
+        """
+        data = self._load()
+
+        # Convert files to memory file references
+        user_memory_files = [self._file_to_memory_file(f) for f in (user_files or [])]
+        assistant_memory_files = [self._file_to_memory_file(f) for f in (assistant_files or [])]
+
+        # Store the turn indexed by workflow_run_id
+        data.turns[workflow_run_id] = NodeMemoryTurn(
+            user_content=user_content,
+            user_files=user_memory_files,
+            assistant_content=assistant_content,
+            assistant_files=assistant_memory_files,
+        )
+
+        self._dirty = True
+
+    def get_history_prompt_messages(
+        self,
+        *,
+        max_token_limit: int = 2000,
+        message_limit: int | None = None,
+    ) -> Sequence[PromptMessage]:
+        """
+        Retrieve history as PromptMessage sequence.
+
+        Thread tracking is handled by querying Message table's parent_message_id structure.
+
+        :param max_token_limit: Maximum tokens for history
+        :param message_limit: unused, for interface compatibility
+        :return: Sequence of PromptMessage for LLM context
+        """
+        # message_limit is unused in NodeTokenBufferMemory (uses token limit instead)
+        _ = message_limit
+        detail = ImagePromptMessageContent.DETAIL.HIGH
+        data = self._load()
+
+        if not data.turns:
+            return []
+
+        # Get workflow_run_ids for current thread from Message table
+        thread_workflow_run_ids = self._get_thread_workflow_run_ids()
+
+        if not thread_workflow_run_ids:
+            return []
+
+        # Build prompt messages in thread order
+        prompt_messages: list[PromptMessage] = []
+        for wf_run_id in thread_workflow_run_ids:
+            turn = data.turns.get(wf_run_id)
+            if not turn:
+                # This workflow execution didn't have node memory stored
+                continue
+
+            # Build user message
+            user_files = self._rebuild_files(turn.user_files) if turn.user_files else []
+            user_msg = self._build_prompt_message(
+                role="user",
+                content=turn.user_content,
+                files=user_files,
+                detail=detail,
+            )
+            prompt_messages.append(user_msg)
+
+            # Build assistant message
+            assistant_files = self._rebuild_files(turn.assistant_files) if turn.assistant_files else []
+            assistant_msg = self._build_prompt_message(
+                role="assistant",
+                content=turn.assistant_content,
+                files=assistant_files,
+                detail=detail,
+            )
+            prompt_messages.append(assistant_msg)
+
+        if not prompt_messages:
+            return []
+
+        # Truncate by token limit
+        try:
+            current_tokens = self.model_instance.get_llm_num_tokens(prompt_messages)
+            while current_tokens > max_token_limit and len(prompt_messages) > 1:
+                prompt_messages.pop(0)
+                current_tokens = self.model_instance.get_llm_num_tokens(prompt_messages)
+        except Exception as e:
+            logger.warning("Failed to count tokens for truncation: %s", e)
+
+        return prompt_messages
+
+    def flush(self) -> None:
+        """
+        Persist buffered changes to object storage.
+        Call this at the end of node execution.
+        """
+        if self._dirty:
+            self._save()
+
+    def clear(self) -> None:
+        """Clear all messages in this node's memory."""
+        self._data = NodeMemoryData()
+        self._save()
+
+    def exists(self) -> bool:
+        """Check if node memory exists in storage."""
+        return storage.exists(self._storage_key)
@@ -5,12 +5,12 @@ from sqlalchemy.orm import sessionmaker

 from core.app.app_config.features.file_upload.manager import FileUploadConfigManager
 from core.file import file_manager
+from core.memory.base import BaseMemory
 from core.model_manager import ModelInstance
 from core.model_runtime.entities import (
    AssistantPromptMessage,
    ImagePromptMessageContent,
    PromptMessage,
-    PromptMessageRole,
    TextPromptMessageContent,
    UserPromptMessage,
 )
@@ -24,7 +24,7 @@ from repositories.api_workflow_run_repository import APIWorkflowRunRepository
 from repositories.factory import DifyAPIRepositoryFactory


-class TokenBufferMemory:
+class TokenBufferMemory(BaseMemory):
    def __init__(
        self,
        conversation: Conversation,
@@ -115,10 +115,14 @@ class TokenBufferMemory:
                return AssistantPromptMessage(content=prompt_message_contents)

    def get_history_prompt_messages(
-        self, max_token_limit: int = 2000, message_limit: int | None = None
+        self,
+        *,
+        max_token_limit: int = 2000,
+        message_limit: int | None = None,
    ) -> Sequence[PromptMessage]:
        """
        Get history prompt messages.
+
        :param max_token_limit: max token limit
        :param message_limit: message limit
        """
@@ -200,44 +204,3 @@ class TokenBufferMemory:
                curr_message_tokens = self.model_instance.get_llm_num_tokens(prompt_messages)

        return prompt_messages
-
-    def get_history_prompt_text(
-        self,
-        human_prefix: str = "Human",
-        ai_prefix: str = "Assistant",
-        max_token_limit: int = 2000,
-        message_limit: int | None = None,
-    ) -> str:
-        """
-        Get history prompt text.
-        :param human_prefix: human prefix
-        :param ai_prefix: ai prefix
-        :param max_token_limit: max token limit
-        :param message_limit: message limit
-        :return:
-        """
-        prompt_messages = self.get_history_prompt_messages(max_token_limit=max_token_limit, message_limit=message_limit)
-
-        string_messages = []
-        for m in prompt_messages:
-            if m.role == PromptMessageRole.USER:
-                role = human_prefix
-            elif m.role == PromptMessageRole.ASSISTANT:
-                role = ai_prefix
-            else:
-                continue
-
-            if isinstance(m.content, list):
-                inner_msg = ""
-                for content in m.content:
-                    if isinstance(content, TextPromptMessageContent):
-                        inner_msg += f"{content.data}\n"
-                    elif isinstance(content, ImagePromptMessageContent):
-                        inner_msg += "[image]\n"
-
-                string_messages.append(f"{role}: {inner_msg.strip()}")
-            else:
-                message = f"{role}: {m.content}"
-                string_messages.append(message)
-
-        return "\n".join(string_messages)
@@ -251,10 +251,7 @@ class AssistantPromptMessage(PromptMessage):

        :return: True if prompt message is empty, False otherwise
        """
-        if not super().is_empty() and not self.tool_calls:
-            return False
-
-        return True
+        return super().is_empty() and not self.tool_calls


 class SystemPromptMessage(PromptMessage):
@@ -1,6 +1,7 @@
 import logging
 from collections.abc import Sequence

+from opentelemetry.trace import SpanKind
 from sqlalchemy.orm import sessionmaker

 from core.ops.aliyun_trace.data_exporter.traceclient import (
@@ -54,7 +55,7 @@ from core.ops.entities.trace_entity import (
    ToolTraceInfo,
    WorkflowTraceInfo,
 )
-from core.repositories import SQLAlchemyWorkflowNodeExecutionRepository
+from core.repositories import DifyCoreRepositoryFactory
 from core.workflow.entities import WorkflowNodeExecution
 from core.workflow.enums import NodeType, WorkflowNodeExecutionMetadataKey
 from extensions.ext_database import db
@@ -151,6 +152,7 @@ class AliyunDataTrace(BaseTraceInstance):
            ),
            status=status,
            links=trace_metadata.links,
+            span_kind=SpanKind.SERVER,
        )
        self.trace_client.add_span(message_span)

@@ -273,7 +275,7 @@ class AliyunDataTrace(BaseTraceInstance):
        service_account = self.get_service_account_with_tenant(app_id)

        session_factory = sessionmaker(bind=db.engine)
-        workflow_node_execution_repository = SQLAlchemyWorkflowNodeExecutionRepository(
+        workflow_node_execution_repository = DifyCoreRepositoryFactory.create_workflow_node_execution_repository(
            session_factory=session_factory,
            user=service_account,
            app_id=app_id,
@@ -456,6 +458,7 @@ class AliyunDataTrace(BaseTraceInstance):
                ),
                status=status,
                links=trace_metadata.links,
+                span_kind=SpanKind.SERVER,
            )
            self.trace_client.add_span(message_span)

@@ -475,6 +478,7 @@ class AliyunDataTrace(BaseTraceInstance):
            ),
            status=status,
            links=trace_metadata.links,
+            span_kind=SpanKind.SERVER if message_span_id is None else SpanKind.INTERNAL,
        )
        self.trace_client.add_span(workflow_span)

@@ -166,7 +166,7 @@ class SpanBuilder:
            attributes=span_data.attributes,
            events=span_data.events,
            links=span_data.links,
-            kind=trace_api.SpanKind.INTERNAL,
+            kind=span_data.span_kind,
            status=span_data.status,
            start_time=span_data.start_time,
            end_time=span_data.end_time,
@@ -4,7 +4,7 @@ from typing import Any

 from opentelemetry import trace as trace_api
 from opentelemetry.sdk.trace import Event
-from opentelemetry.trace import Status, StatusCode
+from opentelemetry.trace import SpanKind, Status, StatusCode
 from pydantic import BaseModel, Field


@@ -34,3 +34,4 @@ class SpanData(BaseModel):
    status: Status = Field(default=Status(StatusCode.UNSET), description="The status of the span.")
    start_time: int | None = Field(..., description="The start time of the span in nanoseconds.")
    end_time: int | None = Field(..., description="The end time of the span in nanoseconds.")
+    span_kind: SpanKind = Field(default=SpanKind.INTERNAL, description="The OpenTelemetry SpanKind for this span.")
@@ -1,5 +1,6 @@
 from core.plugin.entities.endpoint import EndpointEntityWithInstance
 from core.plugin.impl.base import BasePluginClient
+from core.plugin.impl.exc import PluginDaemonInternalServerError


 class PluginEndpointClient(BasePluginClient):
@@ -70,18 +71,27 @@ class PluginEndpointClient(BasePluginClient):
    def delete_endpoint(self, tenant_id: str, user_id: str, endpoint_id: str):
        """
        Delete the given endpoint.
+
+        This operation is idempotent: if the endpoint is already deleted (record not found),
+        it will return True instead of raising an error.
        """
-        return self._request_with_plugin_daemon_response(
-            "POST",
-            f"plugin/{tenant_id}/endpoint/remove",
-            bool,
-            data={
-                "endpoint_id": endpoint_id,
-            },
-            headers={
-                "Content-Type": "application/json",
-            },
-        )
+        try:
+            return self._request_with_plugin_daemon_response(
+                "POST",
+                f"plugin/{tenant_id}/endpoint/remove",
+                bool,
+                data={
+                    "endpoint_id": endpoint_id,
+                },
+                headers={
+                    "Content-Type": "application/json",
+                },
+            )
+        except PluginDaemonInternalServerError as e:
+            # Make delete idempotent: if record is not found, consider it a success
+            if "record not found" in str(e.description).lower():
+                return True
+            raise

    def enable_endpoint(self, tenant_id: str, user_id: str, endpoint_id: str):
        """
@@ -5,7 +5,7 @@ from core.app.entities.app_invoke_entities import ModelConfigWithCredentialsEnti
 from core.file import file_manager
 from core.file.models import File
 from core.helper.code_executor.jinja2.jinja2_formatter import Jinja2Formatter
-from core.memory.token_buffer_memory import TokenBufferMemory
+from core.memory.base import BaseMemory
 from core.model_runtime.entities import (
    AssistantPromptMessage,
    PromptMessage,
@@ -43,7 +43,7 @@ class AdvancedPromptTransform(PromptTransform):
        files: Sequence[File],
        context: str | None,
        memory_config: MemoryConfig | None,
-        memory: TokenBufferMemory | None,
+        memory: BaseMemory | None,
        model_config: ModelConfigWithCredentialsEntity,
        image_detail_config: ImagePromptMessageContent.DETAIL | None = None,
    ) -> list[PromptMessage]:
@@ -84,7 +84,7 @@ class AdvancedPromptTransform(PromptTransform):
        files: Sequence[File],
        context: str | None,
        memory_config: MemoryConfig | None,
-        memory: TokenBufferMemory | None,
+        memory: BaseMemory | None,
        model_config: ModelConfigWithCredentialsEntity,
        image_detail_config: ImagePromptMessageContent.DETAIL | None = None,
    ) -> list[PromptMessage]:
@@ -145,7 +145,7 @@ class AdvancedPromptTransform(PromptTransform):
        files: Sequence[File],
        context: str | None,
        memory_config: MemoryConfig | None,
-        memory: TokenBufferMemory | None,
+        memory: BaseMemory | None,
        model_config: ModelConfigWithCredentialsEntity,
        image_detail_config: ImagePromptMessageContent.DETAIL | None = None,
    ) -> list[PromptMessage]:
@@ -270,7 +270,7 @@ class AdvancedPromptTransform(PromptTransform):

    def _set_histories_variable(
        self,
-        memory: TokenBufferMemory,
+        memory: BaseMemory,
        memory_config: MemoryConfig,
        raw_prompt: str,
        role_prefix: MemoryConfig.RolePrefix,
@@ -1,3 +1,4 @@
+from enum import StrEnum
 from typing import Literal

 from pydantic import BaseModel
@@ -5,6 +6,13 @@ from pydantic import BaseModel
 from core.model_runtime.entities.message_entities import PromptMessageRole


+class MemoryMode(StrEnum):
+    """Memory mode for LLM nodes."""
+
+    CONVERSATION = "conversation"  # Use TokenBufferMemory (default, existing behavior)
+    NODE = "node"  # Use NodeTokenBufferMemory (Chatflow only)
+
+
 class ChatModelMessage(BaseModel):
    """
    Chat Message.
@@ -48,3 +56,4 @@ class MemoryConfig(BaseModel):
    role_prefix: RolePrefix | None = None
    window: WindowConfig
    query_prompt_template: str | None = None
+    mode: MemoryMode = MemoryMode.CONVERSATION
@@ -1,7 +1,7 @@
 from typing import Any

 from core.app.entities.app_invoke_entities import ModelConfigWithCredentialsEntity
-from core.memory.token_buffer_memory import TokenBufferMemory
+from core.memory.base import BaseMemory
 from core.model_manager import ModelInstance
 from core.model_runtime.entities.message_entities import PromptMessage
 from core.model_runtime.entities.model_entities import ModelPropertyKey
@@ -11,7 +11,7 @@ from core.prompt.entities.advanced_prompt_entities import MemoryConfig
 class PromptTransform:
    def _append_chat_histories(
        self,
-        memory: TokenBufferMemory,
+        memory: BaseMemory,
        memory_config: MemoryConfig,
        prompt_messages: list[PromptMessage],
        model_config: ModelConfigWithCredentialsEntity,
@@ -52,7 +52,7 @@ class PromptTransform:

    def _get_history_messages_from_memory(
        self,
-        memory: TokenBufferMemory,
+        memory: BaseMemory,
        memory_config: MemoryConfig,
        max_token_limit: int,
        human_prefix: str | None = None,
@@ -73,7 +73,7 @@ class PromptTransform:
        return memory.get_history_prompt_text(**kwargs)

    def _get_history_messages_list_from_memory(
-        self, memory: TokenBufferMemory, memory_config: MemoryConfig, max_token_limit: int
+        self, memory: BaseMemory, memory_config: MemoryConfig, max_token_limit: int
    ) -> list[PromptMessage]:
        """Get memory messages."""
        return list(
@@ -1047,6 +1047,8 @@ class ToolManager:
                        continue
                    tool_input = ToolNodeData.ToolInput.model_validate(tool_configurations.get(parameter.name, {}))
                    if tool_input.type == "variable":
+                        if not isinstance(tool_input.value, list):
+                            raise ToolParameterError(f"Invalid variable selector for {parameter.name}")
                        variable = variable_pool.get(tool_input.value)
                        if variable is None:
                            raise ToolParameterError(f"Variable {tool_input.value} does not exist")
@@ -1056,6 +1058,11 @@ class ToolManager:
                    elif tool_input.type == "mixed":
                        segment_group = variable_pool.convert_template(str(tool_input.value))
                        parameter_value = segment_group.text
+                    elif tool_input.type == "mention":
+                        # Mention type not supported in agent mode
+                        raise ToolParameterError(
+                            f"Mention type not supported in agent for parameter '{parameter.name}'"
+                        )
                    else:
                        raise ToolParameterError(f"Unknown tool input type '{tool_input.type}'")
                    runtime_parameters[parameter.name] = parameter_value
@@ -7,8 +7,8 @@ from typing import Any, cast

 from flask import has_request_context
 from sqlalchemy import select
-from sqlalchemy.orm import Session

+from core.db.session_factory import session_factory
 from core.file import FILE_MODEL_IDENTITY, File, FileTransferMethod
 from core.model_runtime.entities.llm_entities import LLMUsage, LLMUsageMetadata
 from core.tools.__base.tool import Tool
@@ -20,7 +20,6 @@ from core.tools.entities.tool_entities import (
    ToolProviderType,
 )
 from core.tools.errors import ToolInvokeError
-from extensions.ext_database import db
 from factories.file_factory import build_from_mapping
 from libs.login import current_user
 from models import Account, Tenant
@@ -230,30 +229,32 @@ class WorkflowTool(Tool):
        """
        Resolve user from database (worker/Celery context).
        """
+        with session_factory.create_session() as session:
+            tenant_stmt = select(Tenant).where(Tenant.id == self.runtime.tenant_id)
+            tenant = session.scalar(tenant_stmt)
+            if not tenant:
+                return None
+
+            user_stmt = select(Account).where(Account.id == user_id)
+            user = session.scalar(user_stmt)
+            if user:
+                user.current_tenant = tenant
+                session.expunge(user)
+                return user
+
+            end_user_stmt = select(EndUser).where(EndUser.id == user_id, EndUser.tenant_id == tenant.id)
+            end_user = session.scalar(end_user_stmt)
+            if end_user:
+                session.expunge(end_user)
+                return end_user

-        tenant_stmt = select(Tenant).where(Tenant.id == self.runtime.tenant_id)
-        tenant = db.session.scalar(tenant_stmt)
-        if not tenant:
            return None

-        user_stmt = select(Account).where(Account.id == user_id)
-        user = db.session.scalar(user_stmt)
-        if user:
-            user.current_tenant = tenant
-            return user
-
-        end_user_stmt = select(EndUser).where(EndUser.id == user_id, EndUser.tenant_id == tenant.id)
-        end_user = db.session.scalar(end_user_stmt)
-        if end_user:
-            return end_user
-
-        return None
-
    def _get_workflow(self, app_id: str, version: str) -> Workflow:
        """
        get the workflow by app id and version
        """
-        with Session(db.engine, expire_on_commit=False) as session, session.begin():
+        with session_factory.create_session() as session, session.begin():
            if not version:
                stmt = (
                    select(Workflow)
@@ -265,22 +266,24 @@ class WorkflowTool(Tool):
                stmt = select(Workflow).where(Workflow.app_id == app_id, Workflow.version == version)
                workflow = session.scalar(stmt)

-        if not workflow:
-            raise ValueError("workflow not found or not published")
+            if not workflow:
+                raise ValueError("workflow not found or not published")

-        return workflow
+            session.expunge(workflow)
+            return workflow

    def _get_app(self, app_id: str) -> App:
        """
        get the app by app id
        """
        stmt = select(App).where(App.id == app_id)
-        with Session(db.engine, expire_on_commit=False) as session, session.begin():
+        with session_factory.create_session() as session, session.begin():
            app = session.scalar(stmt)
-        if not app:
-            raise ValueError("app not found")
+            if not app:
+                raise ValueError("app not found")

-        return app
+            session.expunge(app)
+            return app

    def _transform_args(self, tool_parameters: dict) -> tuple[dict, list[dict]]:
        """
@@ -4,6 +4,7 @@ from .segments import (
    ArrayFileSegment,
    ArrayNumberSegment,
    ArrayObjectSegment,
+    ArrayPromptMessageSegment,
    ArraySegment,
    ArrayStringSegment,
    FileSegment,
@@ -20,6 +21,7 @@ from .variables import (
    ArrayFileVariable,
    ArrayNumberVariable,
    ArrayObjectVariable,
+    ArrayPromptMessageVariable,
    ArrayStringVariable,
    ArrayVariable,
    FileVariable,
@@ -30,6 +32,7 @@ from .variables import (
    SecretVariable,
    StringVariable,
    Variable,
+    VariableBase,
 )

 __all__ = [
@@ -41,6 +44,8 @@ __all__ = [
    "ArrayNumberVariable",
    "ArrayObjectSegment",
    "ArrayObjectVariable",
+    "ArrayPromptMessageSegment",
+    "ArrayPromptMessageVariable",
    "ArraySegment",
    "ArrayStringSegment",
    "ArrayStringVariable",
@@ -62,4 +67,5 @@ __all__ = [
    "StringSegment",
    "StringVariable",
    "Variable",
+    "VariableBase",
 ]
@@ -6,6 +6,7 @@ from typing import Annotated, Any, TypeAlias
 from pydantic import BaseModel, ConfigDict, Discriminator, Tag, field_validator

 from core.file import File
+from core.model_runtime.entities import PromptMessage

 from .types import SegmentType

@@ -208,6 +209,15 @@ class ArrayBooleanSegment(ArraySegment):
    value: Sequence[bool]


+class ArrayPromptMessageSegment(ArraySegment):
+    value_type: SegmentType = SegmentType.ARRAY_PROMPT_MESSAGE
+    value: Sequence[PromptMessage]
+
+    def to_object(self):
+        """Convert to JSON-serializable format for database storage and frontend."""
+        return [msg.model_dump() for msg in self.value]
+
+
 def get_segment_discriminator(v: Any) -> SegmentType | None:
    if isinstance(v, Segment):
        return v.value_type
@@ -232,7 +242,7 @@ def get_segment_discriminator(v: Any) -> SegmentType | None:
 # - All variants in `SegmentUnion` must inherit from the `Segment` class.
 # - The union must include all non-abstract subclasses of `Segment`, except:
 #   - `SegmentGroup`, which is not added to the variable pool.
-#   - `Variable` and its subclasses, which are handled by `VariableUnion`.
+#   - `VariableBase` and its subclasses, which are handled by `Variable`.
 SegmentUnion: TypeAlias = Annotated[
    (
        Annotated[NoneSegment, Tag(SegmentType.NONE)]
@@ -248,6 +258,7 @@ SegmentUnion: TypeAlias = Annotated[
        | Annotated[ArrayObjectSegment, Tag(SegmentType.ARRAY_OBJECT)]
        | Annotated[ArrayFileSegment, Tag(SegmentType.ARRAY_FILE)]
        | Annotated[ArrayBooleanSegment, Tag(SegmentType.ARRAY_BOOLEAN)]
+        | Annotated[ArrayPromptMessageSegment, Tag(SegmentType.ARRAY_PROMPT_MESSAGE)]
    ),
    Discriminator(get_segment_discriminator),
 ]
@@ -45,6 +45,7 @@ class SegmentType(StrEnum):
    ARRAY_OBJECT = "array[object]"
    ARRAY_FILE = "array[file]"
    ARRAY_BOOLEAN = "array[boolean]"
+    ARRAY_PROMPT_MESSAGE = "array[message]"

    NONE = "none"

@@ -3,8 +3,10 @@ from typing import Any

 import orjson

+from core.model_runtime.entities import PromptMessage
+
 from .segment_group import SegmentGroup
-from .segments import ArrayFileSegment, FileSegment, Segment
+from .segments import ArrayFileSegment, ArrayPromptMessageSegment, FileSegment, Segment


 def to_selector(node_id: str, name: str, paths: Iterable[str] = ()) -> Sequence[str]:
@@ -16,7 +18,7 @@ def to_selector(node_id: str, name: str, paths: Iterable[str] = ()) -> Sequence[

 def segment_orjson_default(o: Any):
    """Default function for orjson serialization of Segment types"""
-    if isinstance(o, ArrayFileSegment):
+    if isinstance(o, (ArrayFileSegment, ArrayPromptMessageSegment)):
        return [v.model_dump() for v in o.value]
    elif isinstance(o, FileSegment):
        return o.value.model_dump()
@@ -24,6 +26,8 @@ def segment_orjson_default(o: Any):
        return [segment_orjson_default(seg) for seg in o.value]
    elif isinstance(o, Segment):
        return o.value
+    elif isinstance(o, PromptMessage):
+        return o.model_dump()
    raise TypeError(f"Object of type {type(o).__name__} is not JSON serializable")


@@ -12,6 +12,7 @@ from .segments import (
    ArrayFileSegment,
    ArrayNumberSegment,
    ArrayObjectSegment,
+    ArrayPromptMessageSegment,
    ArraySegment,
    ArrayStringSegment,
    BooleanSegment,
@@ -27,7 +28,7 @@ from .segments import (
 from .types import SegmentType


-class Variable(Segment):
+class VariableBase(Segment):
    """
    A variable is a segment that has a name.

@@ -45,23 +46,23 @@ class Variable(Segment):
    selector: Sequence[str] = Field(default_factory=list)


-class StringVariable(StringSegment, Variable):
+class StringVariable(StringSegment, VariableBase):
    pass


-class FloatVariable(FloatSegment, Variable):
+class FloatVariable(FloatSegment, VariableBase):
    pass


-class IntegerVariable(IntegerSegment, Variable):
+class IntegerVariable(IntegerSegment, VariableBase):
    pass


-class ObjectVariable(ObjectSegment, Variable):
+class ObjectVariable(ObjectSegment, VariableBase):
    pass


-class ArrayVariable(ArraySegment, Variable):
+class ArrayVariable(ArraySegment, VariableBase):
    pass


@@ -89,16 +90,16 @@ class SecretVariable(StringVariable):
        return encrypter.obfuscated_token(self.value)


-class NoneVariable(NoneSegment, Variable):
+class NoneVariable(NoneSegment, VariableBase):
    value_type: SegmentType = SegmentType.NONE
    value: None = None


-class FileVariable(FileSegment, Variable):
+class FileVariable(FileSegment, VariableBase):
    pass


-class BooleanVariable(BooleanSegment, Variable):
+class BooleanVariable(BooleanSegment, VariableBase):
    pass


@@ -110,6 +111,10 @@ class ArrayBooleanVariable(ArrayBooleanSegment, ArrayVariable):
    pass


+class ArrayPromptMessageVariable(ArrayPromptMessageSegment, ArrayVariable):
+    pass
+
+
 class RAGPipelineVariable(BaseModel):
    belong_to_node_id: str = Field(description="belong to which node id, shared means public")
    type: str = Field(description="variable type, text-input, paragraph, select, number,  file, file-list")
@@ -139,13 +144,13 @@ class RAGPipelineVariableInput(BaseModel):
    value: Any


-# The `VariableUnion`` type is used to enable serialization and deserialization with Pydantic.
-# Use `Variable` for type hinting when serialization is not required.
+# The `Variable` type is used to enable serialization and deserialization with Pydantic.
+# Use `VariableBase` for type hinting when serialization is not required.
 #
 # Note:
-# - All variants in `VariableUnion` must inherit from the `Variable` class.
-# - The union must include all non-abstract subclasses of `Segment`, except:
-VariableUnion: TypeAlias = Annotated[
+# - All variants in `Variable` must inherit from the `VariableBase` class.
+# - The union must include all non-abstract subclasses of `VariableBase`.
+Variable: TypeAlias = Annotated[
    (
        Annotated[NoneVariable, Tag(SegmentType.NONE)]
        | Annotated[StringVariable, Tag(SegmentType.STRING)]
@@ -160,6 +165,7 @@ VariableUnion: TypeAlias = Annotated[
        | Annotated[ArrayObjectVariable, Tag(SegmentType.ARRAY_OBJECT)]
        | Annotated[ArrayFileVariable, Tag(SegmentType.ARRAY_FILE)]
        | Annotated[ArrayBooleanVariable, Tag(SegmentType.ARRAY_BOOLEAN)]
+        | Annotated[ArrayPromptMessageVariable, Tag(SegmentType.ARRAY_PROMPT_MESSAGE)]
        | Annotated[SecretVariable, Tag(SegmentType.SECRET)]
    ),
    Discriminator(get_segment_discriminator),
@@ -1,7 +1,7 @@
 import abc
 from typing import Protocol

-from core.variables import Variable
+from core.variables import VariableBase


 class ConversationVariableUpdater(Protocol):
@@ -20,12 +20,12 @@ class ConversationVariableUpdater(Protocol):
    """

    @abc.abstractmethod
-    def update(self, conversation_id: str, variable: "Variable"):
+    def update(self, conversation_id: str, variable: "VariableBase"):
        """
        Updates the value of the specified conversation variable in the underlying storage.

        :param conversation_id: The ID of the conversation to update. Typically references `ConversationVariable.id`.
-        :param variable: The `Variable` instance containing the updated value.
+        :param variable: The `VariableBase` instance containing the updated value.
        """
        pass

@@ -63,6 +63,7 @@ class NodeType(StrEnum):
    TRIGGER_SCHEDULE = "trigger-schedule"
    TRIGGER_PLUGIN = "trigger-plugin"
    HUMAN_INPUT = "human-input"
+    GROUP = "group"

    @property
    def is_trigger_node(self) -> bool:
@@ -211,6 +212,10 @@ class WorkflowExecutionStatus(StrEnum):
    def is_ended(self) -> bool:
        return self in _END_STATE

+    @classmethod
+    def ended_values(cls) -> list[str]:
+        return [status.value for status in _END_STATE]
+

 _END_STATE = frozenset(
    [
@@ -248,6 +253,7 @@ class WorkflowNodeExecutionMetadataKey(StrEnum):
    LOOP_VARIABLE_MAP = "loop_variable_map"  # single loop variable output
    DATASOURCE_INFO = "datasource_info"
    COMPLETED_REASON = "completed_reason"  # completed reason for loop node
+    MENTION_PARENT_ID = "mention_parent_id"  # parent node id for extractor nodes


 class WorkflowNodeExecutionStatus(StrEnum):
@@ -307,7 +307,14 @@ class Graph:
        if not node_configs:
            raise ValueError("Graph must have at least one node")

-        node_configs = [node_config for node_config in node_configs if node_config.get("type", "") != "custom-note"]
+        # Filter out UI-only node types:
+        # - custom-note: top-level type (node_config.type == "custom-note")
+        # - group: data-level type (node_config.data.type == "group")
+        node_configs = [
+            node_config
+            for node_config in node_configs
+            if node_config.get("type", "") != "custom-note" and node_config.get("data", {}).get("type", "") != "group"
+        ]

        # Parse node configurations
        node_configs_map = cls._parse_node_configs(node_configs)
@@ -11,7 +11,7 @@ from typing import Any

 from pydantic import BaseModel, Field

-from core.variables.variables import VariableUnion
+from core.variables.variables import Variable


 class CommandType(StrEnum):
@@ -46,7 +46,7 @@ class PauseCommand(GraphEngineCommand):
 class VariableUpdate(BaseModel):
    """Represents a single variable update instruction."""

-    value: VariableUnion = Field(description="New variable value")
+    value: Variable = Field(description="New variable value")


 class UpdateVariablesCommand(GraphEngineCommand):
@@ -93,8 +93,8 @@ class EventHandler:
        Args:
            event: The event to handle
        """
-        # Events in loops or iterations are always collected
-        if event.in_loop_id or event.in_iteration_id:
+        # Events in loops, iterations, or extractor groups are always collected
+        if event.in_loop_id or event.in_iteration_id or event.in_mention_parent_id:
            self._event_collector.collect(event)
            return
        return self._dispatch(event)
@@ -125,6 +125,11 @@ class EventHandler:
        Args:
            event: The node started event
        """
+        # Check if this is an extractor node (has parent_node_id)
+        if self._is_extractor_node(event.node_id):
+            self._handle_extractor_node_started(event)
+            return
+
        # Track execution in domain model
        node_execution = self._graph_execution.get_or_create_node_execution(event.node_id)
        is_initial_attempt = node_execution.retry_count == 0
@@ -164,6 +169,11 @@ class EventHandler:
        Args:
            event: The node succeeded event
        """
+        # Check if this is an extractor node (has parent_node_id)
+        if self._is_extractor_node(event.node_id):
+            self._handle_extractor_node_success(event)
+            return
+
        # Update domain model
        node_execution = self._graph_execution.get_or_create_node_execution(event.node_id)
        node_execution.mark_taken()
@@ -226,6 +236,11 @@ class EventHandler:
        Args:
            event: The node failed event
        """
+        # Check if this is an extractor node (has parent_node_id)
+        if self._is_extractor_node(event.node_id):
+            self._handle_extractor_node_failed(event)
+            return
+
        # Update domain model
        node_execution = self._graph_execution.get_or_create_node_execution(event.node_id)
        node_execution.mark_failed(event.error)
@@ -345,3 +360,57 @@ class EventHandler:
                    self._graph_runtime_state.set_output("answer", value)
            else:
                self._graph_runtime_state.set_output(key, value)
+
+    def _is_extractor_node(self, node_id: str) -> bool:
+        """
+        Check if node_id represents an extractor node (has parent_node_id).
+
+        Extractor nodes extract values from list[PromptMessage] for their parent node.
+        They have a parent_node_id field pointing to their parent node.
+        """
+        node = self._graph.nodes.get(node_id)
+        if node is None:
+            return False
+        return node.node_data.is_extractor_node
+
+    def _handle_extractor_node_started(self, event: NodeRunStartedEvent) -> None:
+        """
+        Handle extractor node started event.
+
+        Extractor nodes don't need full execution tracking, just collect the event.
+        """
+        # Track in response coordinator for stream ordering
+        self._response_coordinator.track_node_execution(event.node_id, event.id)
+
+        # Collect the event
+        self._event_collector.collect(event)
+
+    def _handle_extractor_node_success(self, event: NodeRunSucceededEvent) -> None:
+        """
+        Handle extractor node success event.
+
+        Extractor nodes need special handling:
+        - Store outputs in variable pool (for reference by other nodes)
+        - Accumulate token usage
+        - Collect the event for logging
+        - Do NOT process edges or enqueue next nodes (parent node handles that)
+        """
+        self._accumulate_node_usage(event.node_run_result.llm_usage)
+
+        # Store outputs in variable pool
+        self._store_node_outputs(event.node_id, event.node_run_result.outputs)
+
+        # Collect the event
+        self._event_collector.collect(event)
+
+    def _handle_extractor_node_failed(self, event: NodeRunFailedEvent) -> None:
+        """
+        Handle extractor node failed event.
+
+        Extractor node failures are collected for logging,
+        but the parent node is responsible for handling the error.
+        """
+        self._accumulate_node_usage(event.node_run_result.llm_usage)
+
+        # Collect the event for logging
+        self._event_collector.collect(event)
@@ -68,6 +68,7 @@ class _NodeRuntimeSnapshot:
    predecessor_node_id: str | None
    iteration_id: str | None
    loop_id: str | None
+    mention_parent_id: str | None
    created_at: datetime


@@ -230,6 +231,7 @@ class WorkflowPersistenceLayer(GraphEngineLayer):
        metadata = {
            WorkflowNodeExecutionMetadataKey.ITERATION_ID: event.in_iteration_id,
            WorkflowNodeExecutionMetadataKey.LOOP_ID: event.in_loop_id,
+            WorkflowNodeExecutionMetadataKey.MENTION_PARENT_ID: event.in_mention_parent_id,
        }

        domain_execution = WorkflowNodeExecution(
@@ -256,6 +258,7 @@ class WorkflowPersistenceLayer(GraphEngineLayer):
            predecessor_node_id=event.predecessor_node_id,
            iteration_id=event.in_iteration_id,
            loop_id=event.in_loop_id,
+            mention_parent_id=event.in_mention_parent_id,
            created_at=event.start_at,
        )
        self._node_snapshots[event.id] = snapshot
@@ -21,6 +21,12 @@ class GraphNodeEventBase(GraphEngineEvent):
    """iteration id if node is in iteration"""
    in_loop_id: str | None = None
    """loop id if node is in loop"""
+    in_mention_parent_id: str | None = None
+    """Parent node id if this is an extractor node event.
+    
+    When set, indicates this event belongs to an extractor node that
+    is extracting values for the specified parent node.
+    """

    # The version of the node, or "1" if not specified.
    node_version: str = "1"
@@ -12,11 +12,14 @@ from sqlalchemy.orm import Session
 from core.agent.entities import AgentToolEntity
 from core.agent.plugin_entities import AgentStrategyParameter
 from core.file import File, FileTransferMethod
+from core.memory.base import BaseMemory
+from core.memory.node_token_buffer_memory import NodeTokenBufferMemory
 from core.memory.token_buffer_memory import TokenBufferMemory
 from core.model_manager import ModelInstance, ModelManager
 from core.model_runtime.entities.llm_entities import LLMUsage, LLMUsageMetadata
 from core.model_runtime.entities.model_entities import AIModelEntity, ModelType
 from core.model_runtime.utils.encoders import jsonable_encoder
+from core.prompt.entities.advanced_prompt_entities import MemoryMode
 from core.provider_manager import ProviderManager
 from core.tools.entities.tool_entities import (
    ToolIdentity,
@@ -136,6 +139,9 @@ class AgentNode(Node[AgentNodeData]):
            )
            return

+        # Fetch memory for node memory saving
+        memory = self._fetch_memory_for_save()
+
        try:
            yield from self._transform_message(
                messages=message_stream,
@@ -149,6 +155,7 @@ class AgentNode(Node[AgentNodeData]):
                node_type=self.node_type,
                node_id=self._node_id,
                node_execution_id=self.id,
+                memory=memory,
            )
        except PluginDaemonClientSideError as e:
            transform_error = AgentMessageTransformError(
@@ -395,8 +402,20 @@ class AgentNode(Node[AgentNodeData]):
            icon = None
        return icon

-    def _fetch_memory(self, model_instance: ModelInstance) -> TokenBufferMemory | None:
-        # get conversation id
+    def _fetch_memory(self, model_instance: ModelInstance) -> BaseMemory | None:
+        """
+        Fetch memory based on configuration mode.
+
+        Returns TokenBufferMemory for conversation mode (default),
+        or NodeTokenBufferMemory for node mode (Chatflow only).
+        """
+        node_data = self.node_data
+        memory_config = node_data.memory
+
+        if not memory_config:
+            return None
+
+        # get conversation id (required for both modes in Chatflow)
        conversation_id_variable = self.graph_runtime_state.variable_pool.get(
            ["sys", SystemVariableKey.CONVERSATION_ID]
        )
@@ -404,16 +423,26 @@ class AgentNode(Node[AgentNodeData]):
            return None
        conversation_id = conversation_id_variable.value

-        with Session(db.engine, expire_on_commit=False) as session:
-            stmt = select(Conversation).where(Conversation.app_id == self.app_id, Conversation.id == conversation_id)
-            conversation = session.scalar(stmt)
-
-            if not conversation:
-                return None
-
-        memory = TokenBufferMemory(conversation=conversation, model_instance=model_instance)
-
-        return memory
+        # Return appropriate memory type based on mode
+        if memory_config.mode == MemoryMode.NODE:
+            # Node-level memory (Chatflow only)
+            return NodeTokenBufferMemory(
+                app_id=self.app_id,
+                conversation_id=conversation_id,
+                node_id=self._node_id,
+                tenant_id=self.tenant_id,
+                model_instance=model_instance,
+            )
+        else:
+            # Conversation-level memory (default)
+            with Session(db.engine, expire_on_commit=False) as session:
+                stmt = select(Conversation).where(
+                    Conversation.app_id == self.app_id, Conversation.id == conversation_id
+                )
+                conversation = session.scalar(stmt)
+                if not conversation:
+                    return None
+            return TokenBufferMemory(conversation=conversation, model_instance=model_instance)

    def _fetch_model(self, value: dict[str, Any]) -> tuple[ModelInstance, AIModelEntity | None]:
        provider_manager = ProviderManager()
@@ -457,6 +486,47 @@ class AgentNode(Node[AgentNodeData]):
        else:
            return [tool for tool in tools if tool.get("type") != ToolProviderType.MCP]

+    def _fetch_memory_for_save(self) -> BaseMemory | None:
+        """
+        Fetch memory instance for saving node memory.
+        This is a simplified version that doesn't require model_instance.
+        """
+        from core.model_manager import ModelManager
+        from core.model_runtime.entities.model_entities import ModelType
+
+        node_data = self.node_data
+        if not node_data.memory:
+            return None
+
+        # Get conversation_id
+        conversation_id_var = self.graph_runtime_state.variable_pool.get(["sys", SystemVariableKey.CONVERSATION_ID])
+        if not isinstance(conversation_id_var, StringSegment):
+            return None
+        conversation_id = conversation_id_var.value
+
+        # Return appropriate memory type based on mode
+        if node_data.memory.mode == MemoryMode.NODE:
+            # For node memory, we need a model_instance for token counting
+            # Use a simple default model for this purpose
+            try:
+                model_instance = ModelManager().get_default_model_instance(
+                    tenant_id=self.tenant_id,
+                    model_type=ModelType.LLM,
+                )
+            except Exception:
+                return None
+
+            return NodeTokenBufferMemory(
+                app_id=self.app_id,
+                conversation_id=conversation_id,
+                node_id=self._node_id,
+                tenant_id=self.tenant_id,
+                model_instance=model_instance,
+            )
+        else:
+            # Conversation-level memory doesn't need saving here
+            return None
+
    def _transform_message(
        self,
        messages: Generator[ToolInvokeMessage, None, None],
@@ -467,6 +537,7 @@ class AgentNode(Node[AgentNodeData]):
        node_type: NodeType,
        node_id: str,
        node_execution_id: str,
+        memory: BaseMemory | None = None,
    ) -> Generator[NodeEventBase, None, None]:
        """
        Convert ToolInvokeMessages into tuple[plain_text, files]
@@ -711,6 +782,21 @@ class AgentNode(Node[AgentNodeData]):
                is_final=True,
            )

+        # Save to node memory if in node memory mode
+        from core.workflow.nodes.llm import llm_utils
+
+        # Get user query from sys.query
+        user_query_var = self.graph_runtime_state.variable_pool.get(["sys", SystemVariableKey.QUERY])
+        user_query = user_query_var.text if user_query_var else ""
+
+        llm_utils.save_node_memory(
+            memory=memory,
+            variable_pool=self.graph_runtime_state.variable_pool,
+            user_query=user_query,
+            assistant_response=text,
+            assistant_files=files,
+        )
+
        yield StreamCompletedEvent(
            node_run_result=NodeRunResult(
                status=WorkflowNodeExecutionStatus.SUCCEEDED,
@@ -1,4 +1,10 @@
-from .entities import BaseIterationNodeData, BaseIterationState, BaseLoopNodeData, BaseLoopState, BaseNodeData
+from .entities import (
+    BaseIterationNodeData,
+    BaseIterationState,
+    BaseLoopNodeData,
+    BaseLoopState,
+    BaseNodeData,
+)
 from .usage_tracking_mixin import LLMUsageTrackingMixin

 __all__ = [
@@ -175,6 +175,16 @@ class BaseNodeData(ABC, BaseModel):
    default_value: list[DefaultValue] | None = None
    retry_config: RetryConfig = RetryConfig()

+    # Parent node ID when this node is used as an extractor.
+    # If set, this node is an "attached" extractor node that extracts values
+    # from list[PromptMessage] for the parent node's parameters.
+    parent_node_id: str | None = None
+
+    @property
+    def is_extractor_node(self) -> bool:
+        """Check if this node is an extractor node (has parent_node_id)."""
+        return self.parent_node_id is not None
+
    @property
    def default_value_dict(self) -> dict[str, Any]:
        if self.default_value:
@@ -270,10 +270,87 @@ class Node(Generic[NodeDataT]):
        """Check if execution should be stopped."""
        return self.graph_runtime_state.stop_event.is_set()

+    def _find_extractor_node_configs(self) -> list[dict[str, Any]]:
+        """
+        Find all extractor node configurations that have parent_node_id == self._node_id.
+
+        Returns:
+            List of node configuration dicts for extractor nodes
+        """
+        nodes = self.graph_config.get("nodes", [])
+        extractor_configs = []
+        for node_config in nodes:
+            node_data = node_config.get("data", {})
+            if node_data.get("parent_node_id") == self._node_id:
+                extractor_configs.append(node_config)
+        return extractor_configs
+
+    def _execute_extractor_nodes(self) -> Generator[GraphNodeEventBase, None, None]:
+        """
+        Execute all extractor nodes associated with this node.
+
+        Extractor nodes are nodes with parent_node_id == self._node_id.
+        They are executed before the main node to extract values from list[PromptMessage].
+        """
+        from core.workflow.nodes.node_mapping import LATEST_VERSION, NODE_TYPE_CLASSES_MAPPING
+
+        extractor_configs = self._find_extractor_node_configs()
+        logger.debug("[Extractor] Found %d extractor nodes for parent '%s'", len(extractor_configs), self._node_id)
+        if not extractor_configs:
+            return
+
+        for config in extractor_configs:
+            node_id = config.get("id")
+            node_data = config.get("data", {})
+            node_type_str = node_data.get("type")
+
+            if not node_id or not node_type_str:
+                continue
+
+            # Get node class
+            try:
+                node_type = NodeType(node_type_str)
+            except ValueError:
+                continue
+
+            node_mapping = NODE_TYPE_CLASSES_MAPPING.get(node_type)
+            if not node_mapping:
+                continue
+
+            node_version = str(node_data.get("version", "1"))
+            node_cls = node_mapping.get(node_version) or node_mapping.get(LATEST_VERSION)
+            if not node_cls:
+                continue
+
+            # Instantiate and execute the extractor node
+            extractor_node = node_cls(
+                id=node_id,
+                config=config,
+                graph_init_params=self._graph_init_params,
+                graph_runtime_state=self.graph_runtime_state,
+            )
+
+            # Execute and process extractor node events
+            for event in extractor_node.run():
+                # Tag event with parent node id for stream ordering and history tracking
+                if isinstance(event, GraphNodeEventBase):
+                    event.in_mention_parent_id = self._node_id
+
+                if isinstance(event, NodeRunSucceededEvent):
+                    # Store extractor node outputs in variable pool
+                    outputs: Mapping[str, Any] = event.node_run_result.outputs
+                    for variable_name, variable_value in outputs.items():
+                        self.graph_runtime_state.variable_pool.add((node_id, variable_name), variable_value)
+                if not isinstance(event, NodeRunStreamChunkEvent):
+                    yield event
+
    def run(self) -> Generator[GraphNodeEventBase, None, None]:
        execution_id = self.ensure_execution_id()
        self._start_at = naive_utc_now()

+        # Step 1: Execute associated extractor nodes before main node execution
+        yield from self._execute_extractor_nodes()
+
        # Create and push start event with required fields
        start_event = NodeRunStartedEvent(
            id=execution_id,
@@ -17,6 +17,7 @@ from core.helper import ssrf_proxy
 from core.variables.segments import ArrayFileSegment, FileSegment
 from core.workflow.runtime import VariablePool

+from ..protocols import FileManagerProtocol, HttpClientProtocol
 from .entities import (
    HttpRequestNodeAuthorization,
    HttpRequestNodeData,
@@ -78,6 +79,8 @@ class Executor:
        timeout: HttpRequestNodeTimeout,
        variable_pool: VariablePool,
        max_retries: int = dify_config.SSRF_DEFAULT_MAX_RETRIES,
+        http_client: HttpClientProtocol = ssrf_proxy,
+        file_manager: FileManagerProtocol = file_manager,
    ):
        # If authorization API key is present, convert the API key using the variable pool
        if node_data.authorization.type == "api-key":
@@ -104,6 +107,8 @@ class Executor:
        self.data = None
        self.json = None
        self.max_retries = max_retries
+        self._http_client = http_client
+        self._file_manager = file_manager

        # init template
        self.variable_pool = variable_pool
@@ -200,7 +205,7 @@ class Executor:
                    if file_variable is None:
                        raise FileFetchError(f"cannot fetch file with selector {file_selector}")
                    file = file_variable.value
-                    self.content = file_manager.download(file)
+                    self.content = self._file_manager.download(file)
                case "x-www-form-urlencoded":
                    form_data = {
                        self.variable_pool.convert_template(item.key).text: self.variable_pool.convert_template(
@@ -239,7 +244,7 @@ class Executor:
                            ):
                                file_tuple = (
                                    file.filename,
-                                    file_manager.download(file),
+                                    self._file_manager.download(file),
                                    file.mime_type or "application/octet-stream",
                                )
                                if key not in files:
@@ -332,19 +337,18 @@ class Executor:
        do http request depending on api bundle
        """
        _METHOD_MAP = {
-            "get": ssrf_proxy.get,
-            "head": ssrf_proxy.head,
-            "post": ssrf_proxy.post,
-            "put": ssrf_proxy.put,
-            "delete": ssrf_proxy.delete,
-            "patch": ssrf_proxy.patch,
+            "get": self._http_client.get,
+            "head": self._http_client.head,
+            "post": self._http_client.post,
+            "put": self._http_client.put,
+            "delete": self._http_client.delete,
+            "patch": self._http_client.patch,
        }
        method_lc = self.method.lower()
        if method_lc not in _METHOD_MAP:
            raise InvalidHttpMethodError(f"Invalid http method {self.method}")

        request_args = {
-            "url": self.url,
            "data": self.data,
            "files": self.files,
            "json": self.json,
@@ -357,8 +361,12 @@ class Executor:
        }
        # request_args = {k: v for k, v in request_args.items() if v is not None}
        try:
-            response: httpx.Response = _METHOD_MAP[method_lc](**request_args, max_retries=self.max_retries)
-        except (ssrf_proxy.MaxRetriesExceededError, httpx.RequestError) as e:
+            response: httpx.Response = _METHOD_MAP[method_lc](
+                url=self.url,
+                **request_args,
+                max_retries=self.max_retries,
+            )
+        except (self._http_client.max_retries_exceeded_error, self._http_client.request_error) as e:
            raise HttpRequestNodeError(str(e)) from e
        # FIXME: fix type ignore, this maybe httpx type issue
        return response
@@ -1,10 +1,11 @@
 import logging
 import mimetypes
-from collections.abc import Mapping, Sequence
-from typing import Any
+from collections.abc import Callable, Mapping, Sequence
+from typing import TYPE_CHECKING, Any

 from configs import dify_config
-from core.file import File, FileTransferMethod
+from core.file import File, FileTransferMethod, file_manager
+from core.helper import ssrf_proxy
 from core.tools.tool_file_manager import ToolFileManager
 from core.variables.segments import ArrayFileSegment
 from core.workflow.enums import NodeType, WorkflowNodeExecutionStatus
@@ -13,6 +14,7 @@ from core.workflow.nodes.base import variable_template_parser
 from core.workflow.nodes.base.entities import VariableSelector
 from core.workflow.nodes.base.node import Node
 from core.workflow.nodes.http_request.executor import Executor
+from core.workflow.nodes.protocols import FileManagerProtocol, HttpClientProtocol
 from factories import file_factory

 from .entities import (
@@ -30,10 +32,35 @@ HTTP_REQUEST_DEFAULT_TIMEOUT = HttpRequestNodeTimeout(

 logger = logging.getLogger(__name__)

+if TYPE_CHECKING:
+    from core.workflow.entities import GraphInitParams
+    from core.workflow.runtime import GraphRuntimeState
+

 class HttpRequestNode(Node[HttpRequestNodeData]):
    node_type = NodeType.HTTP_REQUEST

+    def __init__(
+        self,
+        id: str,
+        config: Mapping[str, Any],
+        graph_init_params: "GraphInitParams",
+        graph_runtime_state: "GraphRuntimeState",
+        *,
+        http_client: HttpClientProtocol = ssrf_proxy,
+        tool_file_manager_factory: Callable[[], ToolFileManager] = ToolFileManager,
+        file_manager: FileManagerProtocol = file_manager,
+    ) -> None:
+        super().__init__(
+            id=id,
+            config=config,
+            graph_init_params=graph_init_params,
+            graph_runtime_state=graph_runtime_state,
+        )
+        self._http_client = http_client
+        self._tool_file_manager_factory = tool_file_manager_factory
+        self._file_manager = file_manager
+
    @classmethod
    def get_default_config(cls, filters: Mapping[str, object] | None = None) -> Mapping[str, object]:
        return {
@@ -71,6 +98,8 @@ class HttpRequestNode(Node[HttpRequestNodeData]):
                timeout=self._get_request_timeout(self.node_data),
                variable_pool=self.graph_runtime_state.variable_pool,
                max_retries=0,
+                http_client=self._http_client,
+                file_manager=self._file_manager,
            )
            process_data["request"] = http_executor.to_log()

@@ -199,7 +228,7 @@ class HttpRequestNode(Node[HttpRequestNodeData]):
        mime_type = (
            content_disposition_type or content_type or mimetypes.guess_type(filename)[0] or "application/octet-stream"
        )
-        tool_file_manager = ToolFileManager()
+        tool_file_manager = self._tool_file_manager_factory()

        tool_file = tool_file_manager.create_file_by_raw(
            user_id=self.user_id,
@@ -11,7 +11,7 @@ from typing_extensions import TypeIs
 from core.model_runtime.entities.llm_entities import LLMUsage
 from core.variables import IntegerVariable, NoneSegment
 from core.variables.segments import ArrayAnySegment, ArraySegment
-from core.variables.variables import VariableUnion
+from core.variables.variables import Variable
 from core.workflow.constants import CONVERSATION_VARIABLE_NODE_ID
 from core.workflow.enums import (
    NodeExecutionType,
@@ -240,7 +240,7 @@ class IterationNode(LLMUsageTrackingMixin, Node[IterationNodeData]):
                        datetime,
                        list[GraphNodeEventBase],
                        object | None,
-                        dict[str, VariableUnion],
+                        dict[str, Variable],
                        LLMUsage,
                    ]
                ],
@@ -308,7 +308,7 @@ class IterationNode(LLMUsageTrackingMixin, Node[IterationNodeData]):
        item: object,
        flask_app: Flask,
        context_vars: contextvars.Context,
-    ) -> tuple[datetime, list[GraphNodeEventBase], object | None, dict[str, VariableUnion], LLMUsage]:
+    ) -> tuple[datetime, list[GraphNodeEventBase], object | None, dict[str, Variable], LLMUsage]:
        """Execute a single iteration in parallel mode and return results."""
        with preserve_flask_contexts(flask_app=flask_app, context_vars=context_vars):
            iter_start_at = datetime.now(UTC).replace(tzinfo=None)
@@ -515,11 +515,11 @@ class IterationNode(LLMUsageTrackingMixin, Node[IterationNodeData]):

        return variable_mapping

-    def _extract_conversation_variable_snapshot(self, *, variable_pool: VariablePool) -> dict[str, VariableUnion]:
+    def _extract_conversation_variable_snapshot(self, *, variable_pool: VariablePool) -> dict[str, Variable]:
        conversation_variables = variable_pool.variable_dictionary.get(CONVERSATION_VARIABLE_NODE_ID, {})
        return {name: variable.model_copy(deep=True) for name, variable in conversation_variables.items()}

-    def _sync_conversation_variables_from_snapshot(self, snapshot: dict[str, VariableUnion]) -> None:
+    def _sync_conversation_variables_from_snapshot(self, snapshot: dict[str, Variable]) -> None:
        parent_pool = self.graph_runtime_state.variable_pool
        parent_conversations = parent_pool.variable_dictionary.get(CONVERSATION_VARIABLE_NODE_ID, {})

@@ -1,7 +1,7 @@
 from collections.abc import Mapping, Sequence
-from typing import Any, Literal
+from typing import Annotated, Any, Literal, TypeAlias

-from pydantic import BaseModel, Field, field_validator
+from pydantic import BaseModel, ConfigDict, Field, field_validator

 from core.model_runtime.entities import ImagePromptMessageContent, LLMMode
 from core.prompt.entities.advanced_prompt_entities import ChatModelMessage, CompletionModelPromptTemplate, MemoryConfig
@@ -58,9 +58,28 @@ class LLMNodeCompletionModelPromptTemplate(CompletionModelPromptTemplate):
    jinja2_text: str | None = None


+class PromptMessageContext(BaseModel):
+    """Context variable reference in prompt template.
+
+    YAML/JSON format: { "$context": ["node_id", "variable_name"] }
+    This will be expanded to list[PromptMessage] at runtime.
+    """
+
+    model_config = ConfigDict(populate_by_name=True)
+
+    value_selector: Sequence[str] = Field(alias="$context")
+
+
+# Union type for prompt template items (static message or context variable reference)
+PromptTemplateItem: TypeAlias = Annotated[
+    LLMNodeChatModelMessage | PromptMessageContext,
+    Field(discriminator=None),
+]
+
+
 class LLMNodeData(BaseNodeData):
    model: ModelConfig
-    prompt_template: Sequence[LLMNodeChatModelMessage] | LLMNodeCompletionModelPromptTemplate
+    prompt_template: Sequence[PromptTemplateItem] | LLMNodeCompletionModelPromptTemplate
    prompt_config: PromptConfig = Field(default_factory=PromptConfig)
    memory: MemoryConfig | None = None
    context: ContextConfig
@@ -8,12 +8,13 @@ from configs import dify_config
 from core.app.entities.app_invoke_entities import ModelConfigWithCredentialsEntity
 from core.entities.provider_entities import ProviderQuotaType, QuotaUnit
 from core.file.models import File
-from core.memory.token_buffer_memory import TokenBufferMemory
+from core.memory import NodeTokenBufferMemory, TokenBufferMemory
+from core.memory.base import BaseMemory
 from core.model_manager import ModelInstance, ModelManager
 from core.model_runtime.entities.llm_entities import LLMUsage
 from core.model_runtime.entities.model_entities import ModelType
 from core.model_runtime.model_providers.__base.large_language_model import LargeLanguageModel
-from core.prompt.entities.advanced_prompt_entities import MemoryConfig
+from core.prompt.entities.advanced_prompt_entities import MemoryConfig, MemoryMode
 from core.variables.segments import ArrayAnySegment, ArrayFileSegment, FileSegment, NoneSegment, StringSegment
 from core.workflow.enums import SystemVariableKey
 from core.workflow.nodes.llm.entities import ModelConfig
@@ -86,25 +87,100 @@ def fetch_files(variable_pool: VariablePool, selector: Sequence[str]) -> Sequenc


 def fetch_memory(
-    variable_pool: VariablePool, app_id: str, node_data_memory: MemoryConfig | None, model_instance: ModelInstance
-) -> TokenBufferMemory | None:
+    variable_pool: VariablePool,
+    app_id: str,
+    tenant_id: str,
+    node_data_memory: MemoryConfig | None,
+    model_instance: ModelInstance,
+    node_id: str = "",
+) -> BaseMemory | None:
+    """
+    Fetch memory based on configuration mode.
+
+    Returns TokenBufferMemory for conversation mode (default),
+    or NodeTokenBufferMemory for node mode (Chatflow only).
+
+    :param variable_pool: Variable pool containing system variables
+    :param app_id: Application ID
+    :param tenant_id: Tenant ID
+    :param node_data_memory: Memory configuration
+    :param model_instance: Model instance for token counting
+    :param node_id: Node ID in the workflow (required for node mode)
+    :return: Memory instance or None if not applicable
+    """
    if not node_data_memory:
        return None

-    # get conversation id
+    # Get conversation_id from variable pool (required for both modes in Chatflow)
    conversation_id_variable = variable_pool.get(["sys", SystemVariableKey.CONVERSATION_ID])
    if not isinstance(conversation_id_variable, StringSegment):
        return None
    conversation_id = conversation_id_variable.value

-    with Session(db.engine, expire_on_commit=False) as session:
-        stmt = select(Conversation).where(Conversation.app_id == app_id, Conversation.id == conversation_id)
-        conversation = session.scalar(stmt)
-        if not conversation:
+    # Return appropriate memory type based on mode
+    if node_data_memory.mode == MemoryMode.NODE:
+        # Node-level memory (Chatflow only)
+        if not node_id:
            return None
+        return NodeTokenBufferMemory(
+            app_id=app_id,
+            conversation_id=conversation_id,
+            node_id=node_id,
+            tenant_id=tenant_id,
+            model_instance=model_instance,
+        )
+    else:
+        # Conversation-level memory (default)
+        with Session(db.engine, expire_on_commit=False) as session:
+            stmt = select(Conversation).where(Conversation.app_id == app_id, Conversation.id == conversation_id)
+            conversation = session.scalar(stmt)
+            if not conversation:
+                return None
+        return TokenBufferMemory(conversation=conversation, model_instance=model_instance)

-    memory = TokenBufferMemory(conversation=conversation, model_instance=model_instance)
-    return memory
+
+def save_node_memory(
+    memory: BaseMemory | None,
+    variable_pool: VariablePool,
+    user_query: str,
+    assistant_response: str,
+    user_files: Sequence["File"] | None = None,
+    assistant_files: Sequence["File"] | None = None,
+) -> None:
+    """
+    Save dialogue turn to node memory if applicable.
+
+    This function handles the storage logic for NodeTokenBufferMemory.
+    For TokenBufferMemory (conversation-level), no action is taken as it uses
+    the Message table which is managed elsewhere.
+
+    :param memory: Memory instance (NodeTokenBufferMemory or TokenBufferMemory)
+    :param variable_pool: Variable pool containing system variables
+    :param user_query: User's input text
+    :param assistant_response: Assistant's response text
+    :param user_files: Files attached by user (optional)
+    :param assistant_files: Files generated by assistant (optional)
+    """
+    if not isinstance(memory, NodeTokenBufferMemory):
+        return
+
+    # Get workflow_run_id as the key for this execution
+    workflow_run_id_var = variable_pool.get(["sys", SystemVariableKey.WORKFLOW_EXECUTION_ID])
+    if not isinstance(workflow_run_id_var, StringSegment):
+        return
+
+    workflow_run_id = workflow_run_id_var.value
+    if not workflow_run_id:
+        return
+
+    memory.add_messages(
+        workflow_run_id=workflow_run_id,
+        user_content=user_query,
+        user_files=list(user_files) if user_files else None,
+        assistant_content=assistant_response,
+        assistant_files=list(assistant_files) if assistant_files else None,
+    )
+    memory.flush()


 def deduct_llm_quota(tenant_id: str, model_instance: ModelInstance, usage: LLMUsage):
@@ -7,7 +7,7 @@ import logging
 import re
 import time
 from collections.abc import Generator, Mapping, Sequence
-from typing import TYPE_CHECKING, Any, Literal
+from typing import TYPE_CHECKING, Any, Literal, cast

 from sqlalchemy import select

@@ -16,10 +16,11 @@ from core.file import File, FileTransferMethod, FileType, file_manager
 from core.helper.code_executor import CodeExecutor, CodeLanguage
 from core.llm_generator.output_parser.errors import OutputParserError
 from core.llm_generator.output_parser.structured_output import invoke_llm_with_structured_output
-from core.memory.token_buffer_memory import TokenBufferMemory
+from core.memory.base import BaseMemory
 from core.model_manager import ModelInstance, ModelManager
 from core.model_runtime.entities import (
    ImagePromptMessageContent,
+    MultiModalPromptMessageContent,
    PromptMessage,
    PromptMessageContentType,
    TextPromptMessageContent,
@@ -51,6 +52,7 @@ from core.rag.entities.citation_metadata import RetrievalSourceMetadata
 from core.tools.signature import sign_upload_file
 from core.variables import (
    ArrayFileSegment,
+    ArrayPromptMessageSegment,
    ArraySegment,
    FileSegment,
    NoneSegment,
@@ -87,6 +89,7 @@ from .entities import (
    LLMNodeCompletionModelPromptTemplate,
    LLMNodeData,
    ModelConfig,
+    PromptMessageContext,
 )
 from .exc import (
    InvalidContextStructureError,
@@ -159,8 +162,9 @@ class LLMNode(Node[LLMNodeData]):
        variable_pool = self.graph_runtime_state.variable_pool

        try:
-            # init messages template
-            self.node_data.prompt_template = self._transform_chat_messages(self.node_data.prompt_template)
+            # Parse prompt template to separate static messages and context references
+            prompt_template = self.node_data.prompt_template
+            static_messages, context_refs, template_order = self._parse_prompt_template()

            # fetch variables and fetch values from variable pool
            inputs = self._fetch_inputs(node_data=self.node_data)
@@ -208,8 +212,10 @@ class LLMNode(Node[LLMNodeData]):
            memory = llm_utils.fetch_memory(
                variable_pool=variable_pool,
                app_id=self.app_id,
+                tenant_id=self.tenant_id,
                node_data_memory=self.node_data.memory,
                model_instance=model_instance,
+                node_id=self._node_id,
            )

            query: str | None = None
@@ -220,21 +226,40 @@ class LLMNode(Node[LLMNodeData]):
                ):
                    query = query_variable.text

-            prompt_messages, stop = LLMNode.fetch_prompt_messages(
-                sys_query=query,
-                sys_files=files,
-                context=context,
-                memory=memory,
-                model_config=model_config,
-                prompt_template=self.node_data.prompt_template,
-                memory_config=self.node_data.memory,
-                vision_enabled=self.node_data.vision.enabled,
-                vision_detail=self.node_data.vision.configs.detail,
-                variable_pool=variable_pool,
-                jinja2_variables=self.node_data.prompt_config.jinja2_variables,
-                tenant_id=self.tenant_id,
-                context_files=context_files,
-            )
+            # Get prompt messages
+            prompt_messages: Sequence[PromptMessage]
+            stop: Sequence[str] | None
+            if isinstance(prompt_template, list) and context_refs:
+                prompt_messages, stop = self._build_prompt_messages_with_context(
+                    context_refs=context_refs,
+                    template_order=template_order,
+                    static_messages=static_messages,
+                    query=query,
+                    files=files,
+                    context=context,
+                    memory=memory,
+                    model_config=model_config,
+                    context_files=context_files,
+                )
+            else:
+                prompt_messages, stop = LLMNode.fetch_prompt_messages(
+                    sys_query=query,
+                    sys_files=files,
+                    context=context,
+                    memory=memory,
+                    model_config=model_config,
+                    prompt_template=cast(
+                        Sequence[LLMNodeChatModelMessage] | LLMNodeCompletionModelPromptTemplate,
+                        self.node_data.prompt_template,
+                    ),
+                    memory_config=self.node_data.memory,
+                    vision_enabled=self.node_data.vision.enabled,
+                    vision_detail=self.node_data.vision.configs.detail,
+                    variable_pool=variable_pool,
+                    jinja2_variables=self.node_data.prompt_config.jinja2_variables,
+                    tenant_id=self.tenant_id,
+                    context_files=context_files,
+                )

            # handle invoke result
            generator = LLMNode.invoke_llm(
@@ -250,6 +275,7 @@ class LLMNode(Node[LLMNodeData]):
                node_id=self._node_id,
                node_type=self.node_type,
                reasoning_format=self.node_data.reasoning_format,
+                tenant_id=self.tenant_id,
            )

            structured_output: LLMStructuredOutput | None = None
@@ -301,12 +327,25 @@ class LLMNode(Node[LLMNodeData]):
                "reasoning_content": reasoning_content,
                "usage": jsonable_encoder(usage),
                "finish_reason": finish_reason,
+                "context": self._build_context(prompt_messages, clean_text),
            }
            if structured_output:
                outputs["structured_output"] = structured_output.structured_output
            if self._file_outputs:
                outputs["files"] = ArrayFileSegment(value=self._file_outputs)

+            # Write to Node Memory if in node memory mode
+            # Resolve the query template to get actual user content
+            actual_query = variable_pool.convert_template(query or "").text
+            llm_utils.save_node_memory(
+                memory=memory,
+                variable_pool=variable_pool,
+                user_query=actual_query,
+                assistant_response=clean_text,
+                user_files=files,
+                assistant_files=self._file_outputs,
+            )
+
            # Send final chunk event to indicate streaming is complete
            yield StreamChunkEvent(
                selector=[self._node_id, "text"],
@@ -367,6 +406,7 @@ class LLMNode(Node[LLMNodeData]):
        node_id: str,
        node_type: NodeType,
        reasoning_format: Literal["separated", "tagged"] = "tagged",
+        tenant_id: str | None = None,
    ) -> Generator[NodeEventBase | LLMStructuredOutput, None, None]:
        model_schema = model_instance.model_type_instance.get_model_schema(
            node_data_model.name, model_instance.credentials
@@ -390,6 +430,7 @@ class LLMNode(Node[LLMNodeData]):
                stop=list(stop or []),
                stream=True,
                user=user_id,
+                tenant_id=tenant_id,
            )
        else:
            request_start_time = time.perf_counter()
@@ -566,6 +607,48 @@ class LLMNode(Node[LLMNodeData]):
        # Separated mode: always return clean text and reasoning_content
        return clean_text, reasoning_content or ""

+    @staticmethod
+    def _build_context(
+        prompt_messages: Sequence[PromptMessage],
+        assistant_response: str,
+    ) -> list[PromptMessage]:
+        """
+        Build context from prompt messages and assistant response.
+        Excludes system messages and includes the current LLM response.
+        Returns list[PromptMessage] for use with ArrayPromptMessageSegment.
+
+        Note: Multi-modal content base64 data is truncated to avoid storing large data in context.
+        """
+        context_messages: list[PromptMessage] = [
+            LLMNode._truncate_multimodal_content(m) for m in prompt_messages if m.role != PromptMessageRole.SYSTEM
+        ]
+        context_messages.append(AssistantPromptMessage(content=assistant_response))
+        return context_messages
+
+    @staticmethod
+    def _truncate_multimodal_content(message: PromptMessage) -> PromptMessage:
+        """
+        Truncate multi-modal content base64 data in a message to avoid storing large data.
+        Preserves the PromptMessage structure for ArrayPromptMessageSegment compatibility.
+        """
+        content = message.content
+        if content is None or isinstance(content, str):
+            return message
+
+        # Process list content, truncating multi-modal base64 data
+        new_content: list[PromptMessageContentUnionTypes] = []
+        for item in content:
+            if isinstance(item, MultiModalPromptMessageContent):
+                # Truncate base64_data similar to prompt_messages_to_prompt_for_saving
+                truncated_base64 = ""
+                if item.base64_data:
+                    truncated_base64 = item.base64_data[:10] + "...[TRUNCATED]..." + item.base64_data[-10:]
+                new_content.append(item.model_copy(update={"base64_data": truncated_base64}))
+            else:
+                new_content.append(item)
+
+        return message.model_copy(update={"content": new_content})
+
    def _transform_chat_messages(
        self, messages: Sequence[LLMNodeChatModelMessage] | LLMNodeCompletionModelPromptTemplate, /
    ) -> Sequence[LLMNodeChatModelMessage] | LLMNodeCompletionModelPromptTemplate:
@@ -581,6 +664,106 @@ class LLMNode(Node[LLMNodeData]):

        return messages

+    def _parse_prompt_template(
+        self,
+    ) -> tuple[list[LLMNodeChatModelMessage], list[PromptMessageContext], list[tuple[int, str]]]:
+        """
+        Parse prompt_template to separate static messages and context references.
+
+        Returns:
+            Tuple of (static_messages, context_refs, template_order)
+            - static_messages: list of LLMNodeChatModelMessage
+            - context_refs: list of PromptMessageContext
+            - template_order: list of (index, type) tuples preserving original order
+        """
+        prompt_template = self.node_data.prompt_template
+        static_messages: list[LLMNodeChatModelMessage] = []
+        context_refs: list[PromptMessageContext] = []
+        template_order: list[tuple[int, str]] = []
+
+        if isinstance(prompt_template, list):
+            for idx, item in enumerate(prompt_template):
+                if isinstance(item, PromptMessageContext):
+                    context_refs.append(item)
+                    template_order.append((idx, "context"))
+                else:
+                    static_messages.append(item)
+                    template_order.append((idx, "static"))
+            # Transform static messages for jinja2
+            if static_messages:
+                self.node_data.prompt_template = self._transform_chat_messages(static_messages)
+
+        return static_messages, context_refs, template_order
+
+    def _build_prompt_messages_with_context(
+        self,
+        *,
+        context_refs: list[PromptMessageContext],
+        template_order: list[tuple[int, str]],
+        static_messages: list[LLMNodeChatModelMessage],
+        query: str | None,
+        files: Sequence[File],
+        context: str | None,
+        memory: BaseMemory | None,
+        model_config: ModelConfigWithCredentialsEntity,
+        context_files: list[File],
+    ) -> tuple[list[PromptMessage], Sequence[str] | None]:
+        """
+        Build prompt messages by combining static messages and context references in DSL order.
+
+        Returns:
+            Tuple of (prompt_messages, stop_sequences)
+        """
+        variable_pool = self.graph_runtime_state.variable_pool
+
+        # Build a map from context index to its messages
+        context_messages_map: dict[int, list[PromptMessage]] = {}
+        context_idx = 0
+        for idx, type_ in template_order:
+            if type_ == "context":
+                ctx_ref = context_refs[context_idx]
+                ctx_var = variable_pool.get(ctx_ref.value_selector)
+                if ctx_var is None:
+                    raise VariableNotFoundError(f"Variable {'.'.join(ctx_ref.value_selector)} not found")
+                if not isinstance(ctx_var, ArrayPromptMessageSegment):
+                    raise InvalidVariableTypeError(f"Variable {'.'.join(ctx_ref.value_selector)} is not array[message]")
+                context_messages_map[idx] = list(ctx_var.value)
+                context_idx += 1
+
+        # Process static messages
+        static_prompt_messages: Sequence[PromptMessage] = []
+        stop: Sequence[str] | None = None
+        if static_messages:
+            static_prompt_messages, stop = LLMNode.fetch_prompt_messages(
+                sys_query=query,
+                sys_files=files,
+                context=context,
+                memory=memory,
+                model_config=model_config,
+                prompt_template=cast(Sequence[LLMNodeChatModelMessage], self.node_data.prompt_template),
+                memory_config=self.node_data.memory,
+                vision_enabled=self.node_data.vision.enabled,
+                vision_detail=self.node_data.vision.configs.detail,
+                variable_pool=variable_pool,
+                jinja2_variables=self.node_data.prompt_config.jinja2_variables,
+                tenant_id=self.tenant_id,
+                context_files=context_files,
+            )
+
+        # Combine messages according to original DSL order
+        combined_messages: list[PromptMessage] = []
+        static_msg_iter = iter(static_prompt_messages)
+        for idx, type_ in template_order:
+            if type_ == "context":
+                combined_messages.extend(context_messages_map[idx])
+            else:
+                if msg := next(static_msg_iter, None):
+                    combined_messages.append(msg)
+        # Append any remaining static messages (e.g., memory messages)
+        combined_messages.extend(static_msg_iter)
+
+        return combined_messages, stop
+
    def _fetch_jinja_inputs(self, node_data: LLMNodeData) -> dict[str, str]:
        variables: dict[str, Any] = {}

@@ -778,7 +961,7 @@ class LLMNode(Node[LLMNodeData]):
        sys_query: str | None = None,
        sys_files: Sequence[File],
        context: str | None = None,
-        memory: TokenBufferMemory | None = None,
+        memory: BaseMemory | None = None,
        model_config: ModelConfigWithCredentialsEntity,
        prompt_template: Sequence[LLMNodeChatModelMessage] | LLMNodeCompletionModelPromptTemplate,
        memory_config: MemoryConfig | None = None,
@@ -1337,7 +1520,7 @@ def _calculate_rest_token(

 def _handle_memory_chat_mode(
    *,
-    memory: TokenBufferMemory | None,
+    memory: BaseMemory | None,
    memory_config: MemoryConfig | None,
    model_config: ModelConfigWithCredentialsEntity,
 ) -> Sequence[PromptMessage]:
@@ -1354,7 +1537,7 @@ def _handle_memory_chat_mode(

 def _handle_memory_completion_mode(
    *,
-    memory: TokenBufferMemory | None,
+    memory: BaseMemory | None,
    memory_config: MemoryConfig | None,
    model_config: ModelConfigWithCredentialsEntity,
 ) -> str:
@@ -1,16 +1,21 @@
-from collections.abc import Sequence
+from collections.abc import Callable, Sequence
 from typing import TYPE_CHECKING, final

 from typing_extensions import override

 from configs import dify_config
+from core.file import file_manager
+from core.helper import ssrf_proxy
 from core.helper.code_executor.code_executor import CodeExecutor
 from core.helper.code_executor.code_node_provider import CodeNodeProvider
+from core.tools.tool_file_manager import ToolFileManager
 from core.workflow.enums import NodeType
 from core.workflow.graph import NodeFactory
 from core.workflow.nodes.base.node import Node
 from core.workflow.nodes.code.code_node import CodeNode
 from core.workflow.nodes.code.limits import CodeNodeLimits
+from core.workflow.nodes.http_request.node import HttpRequestNode
+from core.workflow.nodes.protocols import FileManagerProtocol, HttpClientProtocol
 from core.workflow.nodes.template_transform.template_renderer import (
    CodeExecutorJinja2TemplateRenderer,
    Jinja2TemplateRenderer,
@@ -43,6 +48,9 @@ class DifyNodeFactory(NodeFactory):
        code_providers: Sequence[type[CodeNodeProvider]] | None = None,
        code_limits: CodeNodeLimits | None = None,
        template_renderer: Jinja2TemplateRenderer | None = None,
+        http_request_http_client: HttpClientProtocol = ssrf_proxy,
+        http_request_tool_file_manager_factory: Callable[[], ToolFileManager] = ToolFileManager,
+        http_request_file_manager: FileManagerProtocol = file_manager,
    ) -> None:
        self.graph_init_params = graph_init_params
        self.graph_runtime_state = graph_runtime_state
@@ -61,6 +69,9 @@ class DifyNodeFactory(NodeFactory):
            max_object_array_length=dify_config.CODE_MAX_OBJECT_ARRAY_LENGTH,
        )
        self._template_renderer = template_renderer or CodeExecutorJinja2TemplateRenderer()
+        self._http_request_http_client = http_request_http_client
+        self._http_request_tool_file_manager_factory = http_request_tool_file_manager_factory
+        self._http_request_file_manager = http_request_file_manager

    @override
    def create_node(self, node_config: dict[str, object]) -> Node:
@@ -113,6 +124,7 @@ class DifyNodeFactory(NodeFactory):
                code_providers=self._code_providers,
                code_limits=self._code_limits,
            )
+
        if node_type == NodeType.TEMPLATE_TRANSFORM:
            return TemplateTransformNode(
                id=node_id,
@@ -122,6 +134,17 @@ class DifyNodeFactory(NodeFactory):
                template_renderer=self._template_renderer,
            )

+        if node_type == NodeType.HTTP_REQUEST:
+            return HttpRequestNode(
+                id=node_id,
+                config=node_config,
+                graph_init_params=self.graph_init_params,
+                graph_runtime_state=self.graph_runtime_state,
+                http_client=self._http_request_http_client,
+                tool_file_manager_factory=self._http_request_tool_file_manager_factory,
+                file_manager=self._http_request_file_manager,
+            )
+
        return node_class(
            id=node_id,
            config=node_config,
@@ -7,7 +7,7 @@ from typing import Any, cast

 from core.app.entities.app_invoke_entities import ModelConfigWithCredentialsEntity
 from core.file import File
-from core.memory.token_buffer_memory import TokenBufferMemory
+from core.memory.base import BaseMemory
 from core.model_manager import ModelInstance
 from core.model_runtime.entities import ImagePromptMessageContent
 from core.model_runtime.entities.llm_entities import LLMUsage
@@ -145,8 +145,10 @@ class ParameterExtractorNode(Node[ParameterExtractorNodeData]):
        memory = llm_utils.fetch_memory(
            variable_pool=variable_pool,
            app_id=self.app_id,
+            tenant_id=self.tenant_id,
            node_data_memory=node_data.memory,
            model_instance=model_instance,
+            node_id=self._node_id,
        )

        if (
@@ -244,6 +246,14 @@ class ParameterExtractorNode(Node[ParameterExtractorNodeData]):
        # transform result into standard format
        result = self._transform_result(data=node_data, result=result or {})

+        # Save to node memory if in node memory mode
+        llm_utils.save_node_memory(
+            memory=memory,
+            variable_pool=variable_pool,
+            user_query=query,
+            assistant_response=json.dumps(result, ensure_ascii=False),
+        )
+
        return NodeRunResult(
            status=WorkflowNodeExecutionStatus.SUCCEEDED,
            inputs=inputs,
@@ -299,7 +309,7 @@ class ParameterExtractorNode(Node[ParameterExtractorNodeData]):
        query: str,
        variable_pool: VariablePool,
        model_config: ModelConfigWithCredentialsEntity,
-        memory: TokenBufferMemory | None,
+        memory: BaseMemory | None,
        files: Sequence[File],
        vision_detail: ImagePromptMessageContent.DETAIL | None = None,
    ) -> tuple[list[PromptMessage], list[PromptMessageTool]]:
@@ -381,7 +391,7 @@ class ParameterExtractorNode(Node[ParameterExtractorNodeData]):
        query: str,
        variable_pool: VariablePool,
        model_config: ModelConfigWithCredentialsEntity,
-        memory: TokenBufferMemory | None,
+        memory: BaseMemory | None,
        files: Sequence[File],
        vision_detail: ImagePromptMessageContent.DETAIL | None = None,
    ) -> list[PromptMessage]:
@@ -419,7 +429,7 @@ class ParameterExtractorNode(Node[ParameterExtractorNodeData]):
        query: str,
        variable_pool: VariablePool,
        model_config: ModelConfigWithCredentialsEntity,
-        memory: TokenBufferMemory | None,
+        memory: BaseMemory | None,
        files: Sequence[File],
        vision_detail: ImagePromptMessageContent.DETAIL | None = None,
    ) -> list[PromptMessage]:
@@ -453,7 +463,7 @@ class ParameterExtractorNode(Node[ParameterExtractorNodeData]):
        query: str,
        variable_pool: VariablePool,
        model_config: ModelConfigWithCredentialsEntity,
-        memory: TokenBufferMemory | None,
+        memory: BaseMemory | None,
        files: Sequence[File],
        vision_detail: ImagePromptMessageContent.DETAIL | None = None,
    ) -> list[PromptMessage]:
@@ -681,7 +691,7 @@ class ParameterExtractorNode(Node[ParameterExtractorNodeData]):
        node_data: ParameterExtractorNodeData,
        query: str,
        variable_pool: VariablePool,
-        memory: TokenBufferMemory | None,
+        memory: BaseMemory | None,
        max_token_limit: int = 2000,
    ) -> list[ChatModelMessage]:
        model_mode = ModelMode(node_data.model.mode)
@@ -708,7 +718,7 @@ class ParameterExtractorNode(Node[ParameterExtractorNodeData]):
        node_data: ParameterExtractorNodeData,
        query: str,
        variable_pool: VariablePool,
-        memory: TokenBufferMemory | None,
+        memory: BaseMemory | None,
        max_token_limit: int = 2000,
    ):
        model_mode = ModelMode(node_data.model.mode)
@@ -0,0 +1,29 @@
+from typing import Protocol
+
+import httpx
+
+from core.file import File
+
+
+class HttpClientProtocol(Protocol):
+    @property
+    def max_retries_exceeded_error(self) -> type[Exception]: ...
+
+    @property
+    def request_error(self) -> type[Exception]: ...
+
+    def get(self, url: str, max_retries: int = ..., **kwargs: object) -> httpx.Response: ...
+
+    def head(self, url: str, max_retries: int = ..., **kwargs: object) -> httpx.Response: ...
+
+    def post(self, url: str, max_retries: int = ..., **kwargs: object) -> httpx.Response: ...
+
+    def put(self, url: str, max_retries: int = ..., **kwargs: object) -> httpx.Response: ...
+
+    def delete(self, url: str, max_retries: int = ..., **kwargs: object) -> httpx.Response: ...
+
+    def patch(self, url: str, max_retries: int = ..., **kwargs: object) -> httpx.Response: ...
+
+
+class FileManagerProtocol(Protocol):
+    def download(self, f: File, /) -> bytes: ...
@@ -4,7 +4,7 @@ from collections.abc import Mapping, Sequence
 from typing import TYPE_CHECKING, Any

 from core.app.entities.app_invoke_entities import ModelConfigWithCredentialsEntity
-from core.memory.token_buffer_memory import TokenBufferMemory
+from core.memory.base import BaseMemory
 from core.model_manager import ModelInstance
 from core.model_runtime.entities import LLMUsage, ModelPropertyKey, PromptMessageRole
 from core.model_runtime.utils.encoders import jsonable_encoder
@@ -96,8 +96,10 @@ class QuestionClassifierNode(Node[QuestionClassifierNodeData]):
        memory = llm_utils.fetch_memory(
            variable_pool=variable_pool,
            app_id=self.app_id,
+            tenant_id=self.tenant_id,
            node_data_memory=node_data.memory,
            model_instance=model_instance,
+            node_id=self._node_id,
        )
        # fetch instruction
        node_data.instruction = node_data.instruction or ""
@@ -203,6 +205,14 @@ class QuestionClassifierNode(Node[QuestionClassifierNodeData]):
                "usage": jsonable_encoder(usage),
            }

+            # Save to node memory if in node memory mode
+            llm_utils.save_node_memory(
+                memory=memory,
+                variable_pool=variable_pool,
+                user_query=query or "",
+                assistant_response=f"class_name: {category_name}, class_id: {category_id}",
+            )
+
            return NodeRunResult(
                status=WorkflowNodeExecutionStatus.SUCCEEDED,
                inputs=variables,
@@ -312,7 +322,7 @@ class QuestionClassifierNode(Node[QuestionClassifierNodeData]):
        self,
        node_data: QuestionClassifierNodeData,
        query: str,
-        memory: TokenBufferMemory | None,
+        memory: BaseMemory | None,
        max_token_limit: int = 2000,
    ):
        model_mode = ModelMode(node_data.model.mode)
@@ -1,4 +1,3 @@
-import json
 from typing import Any

 from jsonschema import Draft7Validator, ValidationError
@@ -43,25 +42,22 @@ class StartNode(Node[StartNodeData]):
            if value is None and variable.required:
                raise ValueError(f"{key} is required in input form")

+            # If no value provided, skip further processing for this key
+            if not value:
+                continue
+
+            if not isinstance(value, dict):
+                raise ValueError(f"JSON object for '{key}' must be an object")
+
+            # Overwrite with normalized dict to ensure downstream consistency
+            node_inputs[key] = value
+
+            # If schema exists, then validate against it
            schema = variable.json_schema
            if not schema:
                continue

-            if not value:
-                continue
-
            try:
-                json_schema = json.loads(schema)
-            except json.JSONDecodeError as e:
-                raise ValueError(f"{schema} must be a valid JSON object")
-
-            try:
-                json_value = json.loads(value)
-            except json.JSONDecodeError as e:
-                raise ValueError(f"{value} must be a valid JSON object")
-
-            try:
-                Draft7Validator(json_schema).validate(json_value)
+                Draft7Validator(schema).validate(value)
            except ValidationError as e:
                raise ValueError(f"JSON object for '{key}' does not match schema: {e.message}")
-            node_inputs[key] = json_value
@@ -1,11 +1,63 @@
-from typing import Any, Literal, Union
+import re
+from collections.abc import Sequence
+from typing import Any, Literal, Self, Union

-from pydantic import BaseModel, field_validator
+from pydantic import BaseModel, field_validator, model_validator
 from pydantic_core.core_schema import ValidationInfo

 from core.tools.entities.tool_entities import ToolProviderType
 from core.workflow.nodes.base.entities import BaseNodeData

+# Pattern to match mention value format: {{@node.context@}}instruction
+# The placeholder {{@node.context@}} must appear at the beginning
+# Format: {{@agent_node_id.context@}} where agent_node_id is dynamic, context is fixed
+MENTION_VALUE_PATTERN = re.compile(r"^\{\{@([a-zA-Z0-9_]+)\.context@\}\}(.*)$", re.DOTALL)
+
+
+def parse_mention_value(value: str) -> tuple[str, str]:
+    """Parse mention value into (node_id, instruction).
+
+    Args:
+        value: The mention value string like "{{@llm.context@}}extract keywords"
+
+    Returns:
+        Tuple of (node_id, instruction)
+
+    Raises:
+        ValueError: If value format is invalid
+    """
+    match = MENTION_VALUE_PATTERN.match(value)
+    if not match:
+        raise ValueError(
+            "For mention type, value must start with {{@node.context@}} placeholder, "
+            "e.g., '{{@llm.context@}}extract keywords'"
+        )
+    return match.group(1), match.group(2)
+
+
+class MentionConfig(BaseModel):
+    """Configuration for extracting value from context variable.
+
+    Used when a tool parameter needs to be extracted from list[PromptMessage]
+    context using an extractor LLM node.
+
+    Note: instruction is embedded in the value field as "{{@node.context@}}instruction"
+    """
+
+    # ID of the extractor LLM node
+    extractor_node_id: str
+
+    # Output variable selector from extractor node
+    # e.g., ["text"], ["structured_output", "query"]
+    output_selector: Sequence[str]
+
+    # Strategy when output is None
+    null_strategy: Literal["raise_error", "use_default"] = "raise_error"
+
+    # Default value when null_strategy is "use_default"
+    # Type should match the parameter's expected type
+    default_value: Any = None
+

 class ToolEntity(BaseModel):
    provider_id: str
@@ -35,7 +87,9 @@ class ToolNodeData(BaseNodeData, ToolEntity):
    class ToolInput(BaseModel):
        # TODO: check this type
        value: Union[Any, list[str]]
-        type: Literal["mixed", "variable", "constant"]
+        type: Literal["mixed", "variable", "constant", "mention"]
+        # Required config for mention type, extracting value from context variable
+        mention_config: MentionConfig | None = None

        @field_validator("type", mode="before")
        @classmethod
@@ -48,6 +102,9 @@ class ToolNodeData(BaseNodeData, ToolEntity):

            if typ == "mixed" and not isinstance(value, str):
                raise ValueError("value must be a string")
+            elif typ == "mention":
+                # Skip here, will be validated in model_validator
+                pass
            elif typ == "variable":
                if not isinstance(value, list):
                    raise ValueError("value must be a list")
@@ -58,6 +115,26 @@ class ToolNodeData(BaseNodeData, ToolEntity):
                raise ValueError("value must be a string, int, float, bool or dict")
            return typ

+        @model_validator(mode="after")
+        def check_mention_type(self) -> Self:
+            """Validate mention type with mention_config."""
+            if self.type != "mention":
+                return self
+
+            value = self.value
+            if value is None:
+                return self
+
+            if not isinstance(value, str):
+                raise ValueError("value must be a string for mention type")
+            # For mention type, value must match format: {{@node.context@}}instruction
+            # This will raise ValueError if format is invalid
+            parse_mention_value(value)
+            # mention_config is required for mention type
+            if self.mention_config is None:
+                raise ValueError("mention_config is required for mention type")
+            return self
+
    tool_parameters: dict[str, ToolInput]
    # The version of the tool parameter.
    # If this value is None, it indicates this is a previous version
@@ -1,7 +1,10 @@
+import logging
 from collections.abc import Generator, Mapping, Sequence
 from typing import TYPE_CHECKING, Any

 from sqlalchemy import select
+
+logger = logging.getLogger(__name__)
 from sqlalchemy.orm import Session

 from core.callback_handler.workflow_tool_callback_handler import DifyWorkflowCallbackHandler
@@ -184,6 +187,7 @@ class ToolNode(Node[ToolNodeData]):
            tool_parameters (Sequence[ToolParameter]): The list of tool parameters.
            variable_pool (VariablePool): The variable pool containing the variables.
            node_data (ToolNodeData): The data associated with the tool node.
+            for_log (bool): Whether to generate parameters for logging.

        Returns:
            Mapping[str, Any]: A dictionary containing the generated parameters.
@@ -199,14 +203,37 @@ class ToolNode(Node[ToolNodeData]):
                continue
            tool_input = node_data.tool_parameters[parameter_name]
            if tool_input.type == "variable":
-                variable = variable_pool.get(tool_input.value)
+                if not isinstance(tool_input.value, list):
+                    raise ToolParameterError(f"Invalid variable selector for parameter '{parameter_name}'")
+                selector = tool_input.value
+                variable = variable_pool.get(selector)
                if variable is None:
                    if parameter.required:
-                        raise ToolParameterError(f"Variable {tool_input.value} does not exist")
+                        raise ToolParameterError(f"Variable {selector} does not exist")
                    continue
                parameter_value = variable.value
+            elif tool_input.type == "mention":
+                # Mention type: get value from extractor node's output
+                if tool_input.mention_config is None:
+                    raise ToolParameterError(
+                        f"mention_config is required for mention type parameter '{parameter_name}'"
+                    )
+                mention_config = tool_input.mention_config.model_dump()
+                try:
+                    parameter_value, found = variable_pool.resolve_mention(
+                        mention_config, parameter_name=parameter_name
+                    )
+                    if not found and parameter.required:
+                        raise ToolParameterError(
+                            f"Extractor output not found for required parameter '{parameter_name}'"
+                        )
+                    if not found:
+                        continue
+                except ValueError as e:
+                    raise ToolParameterError(str(e)) from e
            elif tool_input.type in {"mixed", "constant"}:
-                segment_group = variable_pool.convert_template(str(tool_input.value))
+                template = str(tool_input.value)
+                segment_group = variable_pool.convert_template(template)
                parameter_value = segment_group.log if for_log else segment_group.text
            else:
                raise ToolParameterError(f"Unknown tool input type '{tool_input.type}'")
@@ -488,8 +515,12 @@ class ToolNode(Node[ToolNodeData]):
                for selector in selectors:
                    result[selector.variable] = selector.value_selector
            elif input.type == "variable":
-                selector_key = ".".join(input.value)
-                result[f"#{selector_key}#"] = input.value
+                if isinstance(input.value, list):
+                    selector_key = ".".join(input.value)
+                    result[f"#{selector_key}#"] = input.value
+            elif input.type == "mention":
+                # Mention type: value is handled by extractor node, no direct variable reference
+                pass
            elif input.type == "constant":
                pass

@@ -1,7 +1,7 @@
 from collections.abc import Mapping, Sequence
 from typing import TYPE_CHECKING, Any

-from core.variables import SegmentType, Variable
+from core.variables import SegmentType, VariableBase
 from core.workflow.constants import CONVERSATION_VARIABLE_NODE_ID
 from core.workflow.entities import GraphInitParams
 from core.workflow.enums import NodeType, WorkflowNodeExecutionStatus
@@ -33,6 +33,15 @@ class VariableAssignerNode(Node[VariableAssignerData]):
            graph_runtime_state=graph_runtime_state,
        )

+    def blocks_variable_output(self, variable_selectors: set[tuple[str, ...]]) -> bool:
+        """
+        Check if this Variable Assigner node blocks the output of specific variables.
+
+        Returns True if this node updates any of the requested conversation variables.
+        """
+        assigned_selector = tuple(self.node_data.assigned_variable_selector)
+        return assigned_selector in variable_selectors
+
    @classmethod
    def version(cls) -> str:
        return "1"
@@ -64,7 +73,7 @@ class VariableAssignerNode(Node[VariableAssignerData]):
        assigned_variable_selector = self.node_data.assigned_variable_selector
        # Should be String, Number, Object, ArrayString, ArrayNumber, ArrayObject
        original_variable = self.graph_runtime_state.variable_pool.get(assigned_variable_selector)
-        if not isinstance(original_variable, Variable):
+        if not isinstance(original_variable, VariableBase):
            raise VariableOperatorNodeError("assigned variable not found")

        match self.node_data.write_mode:
@@ -2,7 +2,7 @@ import json
 from collections.abc import Mapping, MutableMapping, Sequence
 from typing import TYPE_CHECKING, Any

-from core.variables import SegmentType, Variable
+from core.variables import SegmentType, VariableBase
 from core.variables.consts import SELECTORS_LENGTH
 from core.workflow.constants import CONVERSATION_VARIABLE_NODE_ID
 from core.workflow.enums import NodeType, WorkflowNodeExecutionStatus
@@ -118,7 +118,7 @@ class VariableAssignerNode(Node[VariableAssignerNodeData]):
                # ==================== Validation Part

                # Check if variable exists
-                if not isinstance(variable, Variable):
+                if not isinstance(variable, VariableBase):
                    raise VariableNotFoundError(variable_selector=item.variable_selector)

                # Check if operation is supported
@@ -192,7 +192,7 @@ class VariableAssignerNode(Node[VariableAssignerNodeData]):

        for selector in updated_variable_selectors:
            variable = self.graph_runtime_state.variable_pool.get(selector)
-            if not isinstance(variable, Variable):
+            if not isinstance(variable, VariableBase):
                raise VariableNotFoundError(variable_selector=selector)
            process_data[variable.name] = variable.value

@@ -213,7 +213,7 @@ class VariableAssignerNode(Node[VariableAssignerNodeData]):
    def _handle_item(
        self,
        *,
-        variable: Variable,
+        variable: VariableBase,
        operation: Operation,
        value: Any,
    ):
@@ -9,10 +9,10 @@ from typing import Annotated, Any, Union, cast
 from pydantic import BaseModel, Field

 from core.file import File, FileAttribute, file_manager
-from core.variables import Segment, SegmentGroup, Variable
+from core.variables import Segment, SegmentGroup, VariableBase
 from core.variables.consts import SELECTORS_LENGTH
 from core.variables.segments import FileSegment, ObjectSegment
-from core.variables.variables import RAGPipelineVariableInput, VariableUnion
+from core.variables.variables import RAGPipelineVariableInput, Variable
 from core.workflow.constants import (
    CONVERSATION_VARIABLE_NODE_ID,
    ENVIRONMENT_VARIABLE_NODE_ID,
@@ -32,7 +32,7 @@ class VariablePool(BaseModel):
    # The first element of the selector is the node id, it's the first-level key in the dictionary.
    # Other elements of the selector are the keys in the second-level dictionary. To get the key, we hash the
    # elements of the selector except the first one.
-    variable_dictionary: defaultdict[str, Annotated[dict[str, VariableUnion], Field(default_factory=dict)]] = Field(
+    variable_dictionary: defaultdict[str, Annotated[dict[str, Variable], Field(default_factory=dict)]] = Field(
        description="Variables mapping",
        default=defaultdict(dict),
    )
@@ -46,13 +46,13 @@ class VariablePool(BaseModel):
        description="System variables",
        default_factory=SystemVariable.empty,
    )
-    environment_variables: Sequence[VariableUnion] = Field(
+    environment_variables: Sequence[Variable] = Field(
        description="Environment variables.",
-        default_factory=list[VariableUnion],
+        default_factory=list[Variable],
    )
-    conversation_variables: Sequence[VariableUnion] = Field(
+    conversation_variables: Sequence[Variable] = Field(
        description="Conversation variables.",
-        default_factory=list[VariableUnion],
+        default_factory=list[Variable],
    )
    rag_pipeline_variables: list[RAGPipelineVariableInput] = Field(
        description="RAG pipeline variables.",
@@ -105,7 +105,7 @@ class VariablePool(BaseModel):
                f"got {len(selector)} elements"
            )

-        if isinstance(value, Variable):
+        if isinstance(value, VariableBase):
            variable = value
        elif isinstance(value, Segment):
            variable = variable_factory.segment_to_variable(segment=value, selector=selector)
@@ -114,9 +114,9 @@ class VariablePool(BaseModel):
            variable = variable_factory.segment_to_variable(segment=segment, selector=selector)

        node_id, name = self._selector_to_keys(selector)
-        # Based on the definition of `VariableUnion`,
-        # `list[Variable]` can be safely used as `list[VariableUnion]` since they are compatible.
-        self.variable_dictionary[node_id][name] = cast(VariableUnion, variable)
+        # Based on the definition of `Variable`,
+        # `VariableBase` instances can be safely used as `Variable` since they are compatible.
+        self.variable_dictionary[node_id][name] = cast(Variable, variable)

    @classmethod
    def _selector_to_keys(cls, selector: Sequence[str]) -> tuple[str, str]:
@@ -268,6 +268,58 @@ class VariablePool(BaseModel):
                continue
            self.add(selector, value)

+    def resolve_mention(
+        self,
+        mention_config: Mapping[str, Any],
+        /,
+        *,
+        parameter_name: str = "",
+    ) -> tuple[Any, bool]:
+        """
+        Resolve a mention parameter value from an extractor node's output.
+
+        Mention parameters reference values extracted by an extractor LLM node
+        from list[PromptMessage] context.
+
+        Args:
+            mention_config: A dict containing:
+                - extractor_node_id: ID of the extractor LLM node
+                - output_selector: Selector path for the output variable (e.g., ["text"])
+                - null_strategy: "raise_error" or "use_default"
+                - default_value: Value to use when null_strategy is "use_default"
+            parameter_name: Name of the parameter being resolved (for error messages)
+
+        Returns:
+            Tuple of (resolved_value, found):
+                - resolved_value: The extracted value, or default_value if not found
+                - found: True if value was found, False if using default
+
+        Raises:
+            ValueError: If extractor_node_id is missing, or if null_strategy is
+                       "raise_error" and the value is not found
+        """
+        extractor_node_id = mention_config.get("extractor_node_id")
+        if not extractor_node_id:
+            raise ValueError(f"Missing extractor_node_id for mention parameter '{parameter_name}'")
+
+        output_selector = list(mention_config.get("output_selector", []))
+        null_strategy = mention_config.get("null_strategy", "raise_error")
+        default_value = mention_config.get("default_value")
+
+        # Build full selector: [extractor_node_id, ...output_selector]
+        full_selector = [extractor_node_id] + output_selector
+        variable = self.get(full_selector)
+
+        if variable is None:
+            if null_strategy == "use_default":
+                return default_value, False
+            raise ValueError(
+                f"Extractor node '{extractor_node_id}' output '{'.'.join(output_selector)}' "
+                f"not found for parameter '{parameter_name}'"
+            )
+
+        return variable.value, True
+
    @classmethod
    def empty(cls) -> VariablePool:
        """Create an empty variable pool."""
--- a/Show More
+++ b/Show More