Merge branch 'main' into feat/agent-node-v2

This commit is contained in:
Novice
2026-01-07 17:34:23 +08:00
802 changed files with 41190 additions and 6172 deletions

View File

@ -9,7 +9,7 @@ Each instance uses a unique key for its command queue.
import json
from typing import TYPE_CHECKING, Any, final
from ..entities.commands import AbortCommand, CommandType, GraphEngineCommand, PauseCommand
from ..entities.commands import AbortCommand, CommandType, GraphEngineCommand, PauseCommand, UpdateVariablesCommand
if TYPE_CHECKING:
from extensions.ext_redis import RedisClientWrapper
@ -113,6 +113,8 @@ class RedisChannel:
return AbortCommand.model_validate(data)
if command_type == CommandType.PAUSE:
return PauseCommand.model_validate(data)
if command_type == CommandType.UPDATE_VARIABLES:
return UpdateVariablesCommand.model_validate(data)
# For other command types, use base class
return GraphEngineCommand.model_validate(data)

View File

@ -5,11 +5,12 @@ This package handles external commands sent to the engine
during execution.
"""
from .command_handlers import AbortCommandHandler, PauseCommandHandler
from .command_handlers import AbortCommandHandler, PauseCommandHandler, UpdateVariablesCommandHandler
from .command_processor import CommandProcessor
__all__ = [
"AbortCommandHandler",
"CommandProcessor",
"PauseCommandHandler",
"UpdateVariablesCommandHandler",
]

View File

@ -4,9 +4,10 @@ from typing import final
from typing_extensions import override
from core.workflow.entities.pause_reason import SchedulingPause
from core.workflow.runtime import VariablePool
from ..domain.graph_execution import GraphExecution
from ..entities.commands import AbortCommand, GraphEngineCommand, PauseCommand
from ..entities.commands import AbortCommand, GraphEngineCommand, PauseCommand, UpdateVariablesCommand
from .command_processor import CommandHandler
logger = logging.getLogger(__name__)
@ -31,3 +32,25 @@ class PauseCommandHandler(CommandHandler):
reason = command.reason
pause_reason = SchedulingPause(message=reason)
execution.pause(pause_reason)
@final
class UpdateVariablesCommandHandler(CommandHandler):
def __init__(self, variable_pool: VariablePool) -> None:
self._variable_pool = variable_pool
@override
def handle(self, command: GraphEngineCommand, execution: GraphExecution) -> None:
assert isinstance(command, UpdateVariablesCommand)
for update in command.updates:
try:
variable = update.value
self._variable_pool.add(variable.selector, variable)
logger.debug("Updated variable %s for workflow %s", variable.selector, execution.workflow_id)
except ValueError as exc:
logger.warning(
"Skipping invalid variable selector %s for workflow %s: %s",
getattr(update.value, "selector", None),
execution.workflow_id,
exc,
)

View File

@ -5,17 +5,21 @@ This module defines command types that can be sent to a running GraphEngine
instance to control its execution flow.
"""
from enum import StrEnum
from collections.abc import Sequence
from enum import StrEnum, auto
from typing import Any
from pydantic import BaseModel, Field
from core.variables.variables import VariableUnion
class CommandType(StrEnum):
"""Types of commands that can be sent to GraphEngine."""
ABORT = "abort"
PAUSE = "pause"
ABORT = auto()
PAUSE = auto()
UPDATE_VARIABLES = auto()
class GraphEngineCommand(BaseModel):
@ -37,3 +41,16 @@ class PauseCommand(GraphEngineCommand):
command_type: CommandType = Field(default=CommandType.PAUSE, description="Type of command")
reason: str = Field(default="unknown reason", description="reason for pause")
class VariableUpdate(BaseModel):
"""Represents a single variable update instruction."""
value: VariableUnion = Field(description="New variable value")
class UpdateVariablesCommand(GraphEngineCommand):
"""Command to update a group of variables in the variable pool."""
command_type: CommandType = Field(default=CommandType.UPDATE_VARIABLES, description="Type of command")
updates: Sequence[VariableUpdate] = Field(default_factory=list, description="Variable updates")

View File

@ -5,9 +5,12 @@ This engine uses a modular architecture with separated packages following
Domain-Driven Design principles for improved maintainability and testability.
"""
from __future__ import annotations
import contextvars
import logging
import queue
import threading
from collections.abc import Generator
from typing import TYPE_CHECKING, cast, final
@ -30,8 +33,13 @@ from core.workflow.runtime import GraphRuntimeState, ReadOnlyGraphRuntimeStateWr
if TYPE_CHECKING: # pragma: no cover - used only for static analysis
from core.workflow.runtime.graph_runtime_state import GraphProtocol
from .command_processing import AbortCommandHandler, CommandProcessor, PauseCommandHandler
from .entities.commands import AbortCommand, PauseCommand
from .command_processing import (
AbortCommandHandler,
CommandProcessor,
PauseCommandHandler,
UpdateVariablesCommandHandler,
)
from .entities.commands import AbortCommand, PauseCommand, UpdateVariablesCommand
from .error_handler import ErrorHandler
from .event_management import EventHandler, EventManager
from .graph_state_manager import GraphStateManager
@ -70,10 +78,13 @@ class GraphEngine:
scale_down_idle_time: float | None = None,
) -> None:
"""Initialize the graph engine with all subsystems and dependencies."""
# stop event
self._stop_event = threading.Event()
# Bind runtime state to current workflow context
self._graph = graph
self._graph_runtime_state = graph_runtime_state
self._graph_runtime_state.stop_event = self._stop_event
self._graph_runtime_state.configure(graph=cast("GraphProtocol", graph))
self._command_channel = command_channel
@ -140,6 +151,9 @@ class GraphEngine:
pause_handler = PauseCommandHandler()
self._command_processor.register_handler(PauseCommand, pause_handler)
update_variables_handler = UpdateVariablesCommandHandler(self._graph_runtime_state.variable_pool)
self._command_processor.register_handler(UpdateVariablesCommand, update_variables_handler)
# === Extensibility ===
# Layers allow plugins to extend engine functionality
self._layers: list[GraphEngineLayer] = []
@ -169,6 +183,7 @@ class GraphEngine:
max_workers=self._max_workers,
scale_up_threshold=self._scale_up_threshold,
scale_down_idle_time=self._scale_down_idle_time,
stop_event=self._stop_event,
)
# === Orchestration ===
@ -199,6 +214,7 @@ class GraphEngine:
event_handler=self._event_handler_registry,
execution_coordinator=self._execution_coordinator,
event_emitter=self._event_manager,
stop_event=self._stop_event,
)
# === Validation ===
@ -212,9 +228,16 @@ class GraphEngine:
if id(node.graph_runtime_state) != expected_state_id:
raise ValueError(f"GraphRuntimeState consistency violation: Node '{node.id}' has a different instance")
def layer(self, layer: GraphEngineLayer) -> "GraphEngine":
def _bind_layer_context(
self,
layer: GraphEngineLayer,
) -> None:
layer.initialize(ReadOnlyGraphRuntimeStateWrapper(self._graph_runtime_state), self._command_channel)
def layer(self, layer: GraphEngineLayer) -> GraphEngine:
"""Add a layer for extending functionality."""
self._layers.append(layer)
self._bind_layer_context(layer)
return self
def run(self) -> Generator[GraphEngineEvent, None, None]:
@ -301,14 +324,7 @@ class GraphEngine:
def _initialize_layers(self) -> None:
"""Initialize layers with context."""
self._event_manager.set_layers(self._layers)
# Create a read-only wrapper for the runtime state
read_only_state = ReadOnlyGraphRuntimeStateWrapper(self._graph_runtime_state)
for layer in self._layers:
try:
layer.initialize(read_only_state, self._command_channel)
except Exception as e:
logger.warning("Failed to initialize layer %s: %s", layer.__class__.__name__, e)
try:
layer.on_graph_start()
except Exception as e:
@ -316,6 +332,7 @@ class GraphEngine:
def _start_execution(self, *, resume: bool = False) -> None:
"""Start execution subsystems."""
self._stop_event.clear()
paused_nodes: list[str] = []
if resume:
paused_nodes = self._graph_runtime_state.consume_paused_nodes()
@ -343,13 +360,12 @@ class GraphEngine:
def _stop_execution(self) -> None:
"""Stop execution subsystems."""
self._stop_event.set()
self._dispatcher.stop()
self._worker_pool.stop()
# Don't mark complete here as the dispatcher already does it
# Notify layers
logger = logging.getLogger(__name__)
for layer in self._layers:
try:
layer.on_graph_end(self._graph_execution.error)

View File

@ -60,6 +60,7 @@ class SkipPropagator:
if edge_states["has_taken"]:
# Enqueue node
self._state_manager.enqueue_node(downstream_node_id)
self._state_manager.start_execution(downstream_node_id)
return
# All edges are skipped, propagate skip to this node

View File

@ -8,7 +8,7 @@ Pluggable middleware for engine extensions.
Abstract base class for layers.
- `initialize()` - Receive runtime context
- `initialize()` - Receive runtime context (runtime state is bound here and always available to hooks)
- `on_graph_start()` - Execution start hook
- `on_event()` - Process all events
- `on_graph_end()` - Execution end hook
@ -34,6 +34,9 @@ engine.layer(debug_layer)
engine.run()
```
`engine.layer()` binds the read-only runtime state before execution, so
`graph_runtime_state` is always available inside layer hooks.
## Custom Layers
```python

View File

@ -13,6 +13,14 @@ from core.workflow.nodes.base.node import Node
from core.workflow.runtime import ReadOnlyGraphRuntimeState
class GraphEngineLayerNotInitializedError(Exception):
"""Raised when a layer's runtime state is accessed before initialization."""
def __init__(self, layer_name: str | None = None) -> None:
name = layer_name or "GraphEngineLayer"
super().__init__(f"{name} runtime state is not initialized. Bind the layer to a GraphEngine before access.")
class GraphEngineLayer(ABC):
"""
Abstract base class for GraphEngine layers.
@ -28,22 +36,27 @@ class GraphEngineLayer(ABC):
def __init__(self) -> None:
"""Initialize the layer. Subclasses can override with custom parameters."""
self.graph_runtime_state: ReadOnlyGraphRuntimeState | None = None
self._graph_runtime_state: ReadOnlyGraphRuntimeState | None = None
self.command_channel: CommandChannel | None = None
@property
def graph_runtime_state(self) -> ReadOnlyGraphRuntimeState:
if self._graph_runtime_state is None:
raise GraphEngineLayerNotInitializedError(type(self).__name__)
return self._graph_runtime_state
def initialize(self, graph_runtime_state: ReadOnlyGraphRuntimeState, command_channel: CommandChannel) -> None:
"""
Initialize the layer with engine dependencies.
Called by GraphEngine before execution starts to inject the read-only runtime state
and command channel. This allows layers to observe engine context and send
commands, but prevents direct state modification.
Called by GraphEngine to inject the read-only runtime state and command channel.
This is invoked when the layer is registered with a `GraphEngine` instance.
Implementations should be idempotent.
Args:
graph_runtime_state: Read-only view of the runtime state
command_channel: Channel for sending commands to the engine
"""
self.graph_runtime_state = graph_runtime_state
self._graph_runtime_state = graph_runtime_state
self.command_channel = command_channel
@abstractmethod

View File

@ -109,10 +109,8 @@ class DebugLoggingLayer(GraphEngineLayer):
self.logger.info("=" * 80)
self.logger.info("🚀 GRAPH EXECUTION STARTED")
self.logger.info("=" * 80)
if self.graph_runtime_state:
# Log initial state
self.logger.info("Initial State:")
# Log initial state
self.logger.info("Initial State:")
@override
def on_event(self, event: GraphEngineEvent) -> None:
@ -243,8 +241,7 @@ class DebugLoggingLayer(GraphEngineLayer):
self.logger.info(" Node retries: %s", self.retry_count)
# Log final state if available
if self.graph_runtime_state and self.include_outputs:
if self.graph_runtime_state.outputs:
self.logger.info("Final outputs: %s", self._format_dict(self.graph_runtime_state.outputs))
if self.include_outputs and self.graph_runtime_state.outputs:
self.logger.info("Final outputs: %s", self._format_dict(self.graph_runtime_state.outputs))
self.logger.info("=" * 80)

View File

@ -337,8 +337,6 @@ class WorkflowPersistenceLayer(GraphEngineLayer):
if update_finished:
execution.finished_at = naive_utc_now()
runtime_state = self.graph_runtime_state
if runtime_state is None:
return
execution.total_tokens = runtime_state.total_tokens
execution.total_steps = runtime_state.node_run_steps
execution.outputs = execution.outputs or runtime_state.outputs
@ -404,6 +402,4 @@ class WorkflowPersistenceLayer(GraphEngineLayer):
def _system_variables(self) -> Mapping[str, Any]:
runtime_state = self.graph_runtime_state
if runtime_state is None:
return {}
return runtime_state.variable_pool.get_by_prefix(SYSTEM_VARIABLE_NODE_ID)

View File

@ -3,14 +3,20 @@ GraphEngine Manager for sending control commands via Redis channel.
This module provides a simplified interface for controlling workflow executions
using the new Redis command channel, without requiring user permission checks.
Supports stop, pause, and resume operations.
"""
import logging
from collections.abc import Sequence
from typing import final
from core.workflow.graph_engine.command_channels.redis_channel import RedisChannel
from core.workflow.graph_engine.entities.commands import AbortCommand, GraphEngineCommand, PauseCommand
from core.workflow.graph_engine.entities.commands import (
AbortCommand,
GraphEngineCommand,
PauseCommand,
UpdateVariablesCommand,
VariableUpdate,
)
from extensions.ext_redis import redis_client
logger = logging.getLogger(__name__)
@ -23,7 +29,6 @@ class GraphEngineManager:
This class provides a simple interface for controlling workflow executions
by sending commands through Redis channels, without user validation.
Supports stop and pause operations.
"""
@staticmethod
@ -45,6 +50,16 @@ class GraphEngineManager:
pause_command = PauseCommand(reason=reason or "User requested pause")
GraphEngineManager._send_command(task_id, pause_command)
@staticmethod
def send_update_variables_command(task_id: str, updates: Sequence[VariableUpdate]) -> None:
"""Send a command to update variables in a running workflow."""
if not updates:
return
update_command = UpdateVariablesCommand(updates=updates)
GraphEngineManager._send_command(task_id, update_command)
@staticmethod
def _send_command(task_id: str, command: GraphEngineCommand) -> None:
"""Send a command to the workflow-specific Redis channel."""

View File

@ -44,6 +44,7 @@ class Dispatcher:
event_queue: queue.Queue[GraphNodeEventBase],
event_handler: "EventHandler",
execution_coordinator: ExecutionCoordinator,
stop_event: threading.Event,
event_emitter: EventManager | None = None,
) -> None:
"""
@ -61,7 +62,7 @@ class Dispatcher:
self._event_emitter = event_emitter
self._thread: threading.Thread | None = None
self._stop_event = threading.Event()
self._stop_event = stop_event
self._start_time: float | None = None
def start(self) -> None:
@ -69,16 +70,14 @@ class Dispatcher:
if self._thread and self._thread.is_alive():
return
self._stop_event.clear()
self._start_time = time.time()
self._thread = threading.Thread(target=self._dispatcher_loop, name="GraphDispatcher", daemon=True)
self._thread.start()
def stop(self) -> None:
"""Stop the dispatcher thread."""
self._stop_event.set()
if self._thread and self._thread.is_alive():
self._thread.join(timeout=10.0)
self._thread.join(timeout=2.0)
def _dispatcher_loop(self) -> None:
"""Main dispatcher loop."""

View File

@ -2,6 +2,8 @@
Factory for creating ReadyQueue instances from serialized state.
"""
from __future__ import annotations
from typing import TYPE_CHECKING
from .in_memory import InMemoryReadyQueue
@ -11,7 +13,7 @@ if TYPE_CHECKING:
from .protocol import ReadyQueue
def create_ready_queue_from_state(state: ReadyQueueState) -> "ReadyQueue":
def create_ready_queue_from_state(state: ReadyQueueState) -> ReadyQueue:
"""
Create a ReadyQueue instance from a serialized state.

View File

@ -5,6 +5,8 @@ This module contains the private ResponseSession class used internally
by ResponseStreamCoordinator to manage streaming sessions.
"""
from __future__ import annotations
from dataclasses import dataclass
from core.workflow.nodes.answer.answer_node import AnswerNode
@ -27,7 +29,7 @@ class ResponseSession:
index: int = 0 # Current position in the template segments
@classmethod
def from_node(cls, node: Node) -> "ResponseSession":
def from_node(cls, node: Node) -> ResponseSession:
"""
Create a ResponseSession from an AnswerNode or EndNode.

View File

@ -42,6 +42,7 @@ class Worker(threading.Thread):
event_queue: queue.Queue[GraphNodeEventBase],
graph: Graph,
layers: Sequence[GraphEngineLayer],
stop_event: threading.Event,
worker_id: int = 0,
flask_app: Flask | None = None,
context_vars: contextvars.Context | None = None,
@ -65,13 +66,16 @@ class Worker(threading.Thread):
self._worker_id = worker_id
self._flask_app = flask_app
self._context_vars = context_vars
self._stop_event = threading.Event()
self._last_task_time = time.time()
self._stop_event = stop_event
self._layers = layers if layers is not None else []
def stop(self) -> None:
"""Signal the worker to stop processing."""
self._stop_event.set()
"""Worker is controlled via shared stop_event from GraphEngine.
This method is a no-op retained for backward compatibility.
"""
pass
@property
def is_idle(self) -> bool:

View File

@ -41,6 +41,7 @@ class WorkerPool:
event_queue: queue.Queue[GraphNodeEventBase],
graph: Graph,
layers: list[GraphEngineLayer],
stop_event: threading.Event,
flask_app: "Flask | None" = None,
context_vars: "Context | None" = None,
min_workers: int | None = None,
@ -81,6 +82,7 @@ class WorkerPool:
self._worker_counter = 0
self._lock = threading.RLock()
self._running = False
self._stop_event = stop_event
# No longer tracking worker states with callbacks to avoid lock contention
@ -135,7 +137,7 @@ class WorkerPool:
# Wait for workers to finish
for worker in self._workers:
if worker.is_alive():
worker.join(timeout=10.0)
worker.join(timeout=2.0)
self._workers.clear()
@ -152,6 +154,7 @@ class WorkerPool:
worker_id=worker_id,
flask_app=self._flask_app,
context_vars=self._context_vars,
stop_event=self._stop_event,
)
worker.start()