Refactor Workflow and Chat Architecture (#68)
* refactor: overhaul workflow and chat architecture - Separate Chat and Workflow API endpoints and database models - Use JSONB to store workflow execution context in Postgres - Convert workflow engine to use pydantic-ai execution graphs inside a Ray task - Update frontend React components to support standalone workflow creation - Remove obsolete and broken workflow runner tests Co-authored-by: zhaoxi826 <198742034+zhaoxi826@users.noreply.github.com> * refactor: overhaul workflow and chat architecture - Separate Chat and Workflow API endpoints and database models - Use JSONB to store workflow execution context in Postgres - Convert workflow engine to use pydantic-ai execution graphs inside a Ray task - Update frontend React components to support standalone workflow creation - Remove obsolete and broken workflow runner tests Co-authored-by: zhaoxi826 <198742034+zhaoxi826@users.noreply.github.com> * refactor: overhaul workflow and chat architecture - Separate Chat and Workflow API endpoints and database models - Use JSONB to store workflow execution context in Postgres - Convert workflow engine to use pydantic-ai execution graphs inside a Ray task - Update frontend React components to support standalone workflow creation - Move workflow_engine inside workflow package to keep core root clean - Remove obsolete and broken workflow runner tests Co-authored-by: zhaoxi826 <198742034+zhaoxi826@users.noreply.github.com> --------- Co-authored-by: google-labs-jules[bot] <161369871+google-labs-jules[bot]@users.noreply.github.com> Co-authored-by: zhaoxi826 <198742034+zhaoxi826@users.noreply.github.com>
This commit is contained in:
GitHub
@@ -28,13 +28,11 @@ from pydantic_ai import Agent, RunContext
|
||||
from kilostar.core.global_state_machine.global_state_machine import GlobalStateMachine
|
||||
from kilostar.core.global_state_machine.model_provider.base_provider import Provider
|
||||
from kilostar.adapter.model_adapter.agent_factory import AgentFactory
|
||||
from kilostar.utils.ray_hook import ray_actor_hook
|
||||
|
||||
|
||||
@ray.remote
|
||||
class ConsciousnessNode:
|
||||
"""ConsciousnessNode 核心组件类。
|
||||
这是一个系统执行节点类，作为多智能体架构中的独立处理单元。它能够接收工作流上下文，根据内置的大模型策略进行意图理解和自主决策，从而驱动特定阶段的任务闭环。"""
|
||||
|
||||
def __init__(self) -> None:
|
||||
from kilostar.utils.logger import get_logger
|
||||
|
||||
@@ -48,19 +46,6 @@ class ConsciousnessNode:
|
||||
model_id: str,
|
||||
tools_list: list[str] = None,
|
||||
) -> None:
|
||||
"""
|
||||
create_agent方法，将agent对象装配到ConsciousnessNode的属性内
|
||||
该方法通过provider_title从global_state_machine中获取provider对象，然后从provider对象中取出供应商形象，装配为pydantic_ai的
|
||||
Agent实例，
|
||||
并挂载到self.agent属性
|
||||
Args:
|
||||
global_state_machine: 全局状态机
|
||||
provider_title: 供应商名
|
||||
model_id: 模型id
|
||||
|
||||
Returns:
|
||||
无返回
|
||||
"""
|
||||
system_prompt: str = (
|
||||
"你叫kilostar，是一个多智能体AI助手系统中的【意识节点 (Consciousness Node)】。\n"
|
||||
"你是系统的'高级规划师'和'架构师'，负责处理监控节点分配过来的复杂任务。\n"
|
||||
@@ -91,10 +76,6 @@ class ConsciousnessNode:
|
||||
|
||||
@self.agent.system_prompt
|
||||
async def dynamic_prompt(ctx: RunContext[ConsciousnessNodeDeps]):
|
||||
"""执行与 dynamic prompt 相关的核心业务流转操作。
|
||||
该方法封装了具体的算法策略或状态控制逻辑，确保操作能够在事务上下文中被原子且一致地执行。
|
||||
Args: ctx (RunContext[ConsciousnessNodeDeps]): 参与 dynamic prompt 逻辑运算或数据构建的上下文依赖对象。
|
||||
Returns: : 经由当前业务模型加工处理后所输出的具体数据实例或领域模型对象。"""
|
||||
prompt = system_prompt + "\n\n"
|
||||
prompt += (
|
||||
f"=== 当前任务上下文 ===\n"
|
||||
@@ -109,14 +90,66 @@ class ConsciousnessNode:
|
||||
|
||||
return prompt
|
||||
|
||||
async def start_workflow_design(self, trace_id: str, command: str):
|
||||
"""
|
||||
开始进行工作流设计的交互过程（与用户通过 SSE 进行确认或直接生成）
|
||||
目前简化为：直接根据 command 拆解并构建工作流，然后提交执行。
|
||||
"""
|
||||
self.logger.info(
|
||||
f"ConsciousnessNode: 开始为 trace_id {trace_id} 设计工作流。原始命令：{command}"
|
||||
)
|
||||
# 获取可用技能 (示例)
|
||||
postgres_database = ray_actor_hook("postgres_database").postgres_database
|
||||
skills_entities = await postgres_database.get_all_worker_individual.remote()
|
||||
available_skills = []
|
||||
if skills_entities:
|
||||
for skill in skills_entities:
|
||||
available_skills.append(
|
||||
{
|
||||
"agent_id": skill.agent_id,
|
||||
"name": skill.agent_name,
|
||||
"description": skill.description,
|
||||
}
|
||||
)
|
||||
|
||||
payload = ForWorkflowEngineInput(
|
||||
original_command=command, available_skills=available_skills
|
||||
)
|
||||
|
||||
# 通知 SSE 正在生成图结构
|
||||
global_workflow_manager = ray_actor_hook(
|
||||
"global_workflow_manager"
|
||||
).global_workflow_manager
|
||||
await global_workflow_manager.put_received.remote(
|
||||
trace_id, "正在为您构建并规划工作流任务节点，请稍候..."
|
||||
)
|
||||
|
||||
# 实际构建过程
|
||||
result = await self.working(payload)
|
||||
|
||||
if result and isinstance(result, ForWorkflowEngine):
|
||||
workflow = result.workflow
|
||||
workflow.trace_id = trace_id
|
||||
|
||||
await global_workflow_manager.put_received.remote(
|
||||
trace_id, "工作流构建完成，即将开始执行！"
|
||||
)
|
||||
|
||||
# 将生成的完整工作流提交执行
|
||||
workflow_engine = ray_actor_hook(
|
||||
"workflow_running_engine"
|
||||
).workflow_running_engine
|
||||
await workflow_engine.execute_workflow.remote(workflow)
|
||||
else:
|
||||
await global_workflow_manager.put_received.remote(
|
||||
trace_id, "很抱歉，工作流生成失败。"
|
||||
)
|
||||
await postgres_database.update_workflow_status.remote(trace_id, "failed")
|
||||
|
||||
async def working(
|
||||
self,
|
||||
payload: Union[ForWorkflowEngineInput, ForWorkflowInput, ForregulatoryInput],
|
||||
) -> Union[ForWorkflowEngine, ForWorkflow, ForregulatoryNode, None]:
|
||||
"""执行与 working 相关的核心业务流转操作。
|
||||
该方法封装了具体的算法策略或状态控制逻辑，确保操作能够在事务上下文中被原子且一致地执行。
|
||||
Args: payload (Union[ForWorkflowEngineInput, ForWorkflowInput, ForregulatoryInput]): 从客户端传递过来或由上游组件生成的核心业务数据体，通常需要进一步的清洗和结构化解析。
|
||||
Returns: (Union[ForWorkflowEngine, ForWorkflow, ForregulatoryNode, None]): 经由当前业务模型加工处理后所输出的具体数据实例或领域模型对象。"""
|
||||
try:
|
||||
result = await self._run(payload)
|
||||
if isinstance(result, (ForWorkflowEngine, ForWorkflow, ForregulatoryNode)):
|
||||
@@ -132,51 +165,20 @@ class ConsciousnessNode:
|
||||
|
||||
@overload
|
||||
async def _run(self, payload: ForWorkflowEngineInput) -> ForWorkflowEngine:
|
||||
"""
|
||||
_run方法
|
||||
该分支应当在regulatory_node简单处理用户命令后，工作流创建前调用！
|
||||
Args:
|
||||
payload: 应当包含原始命令和可用技能等信息
|
||||
|
||||
Returns:
|
||||
ForWorkflowEngine对象，将被放到全局状态机后丢入WorkflowEngine的异步队列
|
||||
"""
|
||||
pass
|
||||
|
||||
@overload
|
||||
async def _run(self, payload: ForWorkflow) -> ForWorkflow:
|
||||
"""
|
||||
_run方法
|
||||
该分支应当在workflow运行时，由WorkflowEngine进行调用！
|
||||
Args:
|
||||
payload: 应当包含workflow中的WorkStep对象
|
||||
|
||||
Returns:
|
||||
ForWorkflow对象，作为ConsciousnessNode执行Workflow中的WorkStep的结果
|
||||
"""
|
||||
pass
|
||||
|
||||
@overload
|
||||
async def _run(self, payload: ForregulatoryInput) -> ForregulatoryNode:
|
||||
"""
|
||||
_run方法
|
||||
该分支应当在workflow运行完全结束后，由WorkflowEngine进行调用！
|
||||
Args:
|
||||
payload: 应当包含整个Workflow的情况
|
||||
|
||||
Returns:
|
||||
Forregulatory对象，作为ConsciousnessNode对于全工作流的技术性总结，返回给regulatoryNode
|
||||
"""
|
||||
pass
|
||||
|
||||
async def _run(
|
||||
self,
|
||||
payload: Union[ForregulatoryInput, ForWorkflowInput, ForWorkflowEngineInput],
|
||||
) -> Union[ForregulatoryNode, ForWorkflow, ForWorkflowEngine]:
|
||||
"""执行与 run 相关的核心业务流转操作。
|
||||
该方法封装了具体的算法策略或状态控制逻辑，确保操作能够在事务上下文中被原子且一致地执行。
|
||||
Args: payload (Union[ForregulatoryInput, ForWorkflowInput, ForWorkflowEngineInput]): 从客户端传递过来或由上游组件生成的核心业务数据体，通常需要进一步的清洗和结构化解析。
|
||||
Returns: (Union[ForregulatoryNode, ForWorkflow, ForWorkflowEngine]): 经由当前业务模型加工处理后所输出的具体数据实例或领域模型对象。"""
|
||||
try:
|
||||
self.agent.retries = 3
|
||||
if isinstance(payload, ForWorkflowEngineInput):
|
||||
|
||||
@@ -13,7 +13,7 @@
|
||||
# limitations under the License.
|
||||
|
||||
|
||||
from kilostar.core.workflow_running_engine.workflow import kilostarWorkflow, WorkStep
|
||||
from kilostar.core.work.workflow.workflow import KiloStarWorkflow, WorkflowStep
|
||||
from kilostar.utils.agent_model import ResponseModel, DepsModel, InputModel
|
||||
from pydantic import Field
|
||||
|
||||
@@ -28,7 +28,7 @@ class ConsciousnessNodeResponse(ResponseModel):
|
||||
class ForWorkflowEngine(ConsciousnessNodeResponse):
|
||||
"""生成workflow并放入WorkflowEngine"""
|
||||
|
||||
workflow: kilostarWorkflow = Field(
|
||||
workflow: KiloStarWorkflow = Field(
|
||||
..., description="生成好的符合规范的完整工作流对象。"
|
||||
)
|
||||
reasoning: str = Field(..., description="生成此工作流的原因和思路简述。")
|
||||
@@ -76,7 +76,7 @@ class ForWorkflowInput(ConsciousnessNodeInput):
|
||||
"""ForWorkflowInput 核心组件类。
|
||||
这是一个领域数据模型或功能封装类，承载了 ForWorkflowInput 相关的内聚属性定义与状态维护。它的存在隔离了局部的业务复杂性，并对外提供了类型安全的访问接口。"""
|
||||
|
||||
workflow_step: WorkStep
|
||||
workflow_step: WorkflowStep
|
||||
original_command: str
|
||||
|
||||
|
||||
@@ -84,5 +84,5 @@ class ForregulatoryInput(ConsciousnessNodeInput):
|
||||
"""ForregulatoryInput 核心组件类。
|
||||
这是一个领域数据模型或功能封装类，承载了 ForregulatoryInput 相关的内聚属性定义与状态维护。它的存在隔离了局部的业务复杂性，并对外提供了类型安全的访问接口。"""
|
||||
|
||||
workflow: kilostarWorkflow
|
||||
workflow: KiloStarWorkflow
|
||||
original_command: str
|
||||
|
||||
Reference in New Issue
Block a user