后端: 1. AIHub 模型分级从 Worker/Strategist 两级重构为 Lite/Pro/Max 三级 - AIHub 结构体从 Worker + Strategist 改为 Lite + Pro + Max,分别对应轻量(标题生成)、标准(Chat 路由/闲聊/交付总结)、高能力(Plan 规划/Execute ReAct)三个能力层级 - config.example.yaml 新增 liteModel / proModel / maxModel 三个模型配置项,替代原 workerModel / strategistModel - 启动层 InitEino 改为创建三个独立模型实例,抽取公共 baseURL 和 apiKey 减少重复 - pickChatModel 统一返回 Pro 模型,旧 strategist 参数不再生效;pickTitleModel 从 Worker 切到 Lite - runNewAgentGraph 按 Plan/Execute→Max、Chat/Deliver→Pro 分级注入;Graph 出错回退也切到 Pro - Memory 模块初始化从 Worker 改为 Pro 2. Plan 节点从"两阶段评估"简化为"单轮深度规划",thinking 开关改为全配置化 - 移除 Phase 1(快速评估 1600 token)+ Phase 2(深度规划 3200 token)的两轮调用逻辑,改为单轮不限 token 深度规划 - PlanDecision 移除 need_thinking 字段,prompt 规则和 JSON contract 同步删除该字段 - 各节点(Plan / Execute / Deliver)thinking 开关从硬编码改为从 AgentGraphDeps 读取,由 config.yaml 的 agent.thinking 段按节点注入 - 新增 agent.thinking 配置段(plan / execute / deliver / memory 四个独立布尔开关),config.example.yaml 补齐默认值 - 新增 resolveThinkingMode 公共函数,plan / execute / deliver 和 memory 决策/抽取链路统一使用 3. Memory 模块 LLM 调用支持 thinking 开关 - Config 新增 LLMThinking 字段,config_loader 从 agent.thinking.memory 读取 - LLMDecisionOrchestrator.Compare 和 LLMWriteOrchestrator.ExtractFacts 的 thinking 模式从硬编码 Disabled 改为读取配置 前端: 1. 移除助手输入区模型选择器及全部偏好持久化逻辑 - 删除 ModelType 类型、selectedModel ref、MODEL_PREFERENCE_STORAGE_KEY 常量 - 删除 isModelType / loadModelPreferenceMap / persistModelPreferenceMap / savePreferredModel / resolvePreferredModel / applyPreferredModelForConversation 六个函数及 modelPreferenceMap ref - 删除 selectedModel watch 监听、发送消息时的 savePreferredModel 调用、切会话时的 applyPreferredModelForConversation 调用、会话迁移时的模型偏好迁移 - fetchChatStream 的 model 参数硬编码为 'worker' - 删除模板中"模型"下拉选择器(标准/策略)及对应的全局样式 .assistant-model-select-panel 2. 上下文窗口指示器简化为仅显示总占用 - ContextWindowMeter 移除 msg0~msg3 四段彩色分段逻辑(ContextSegment 接口、segments computed、v-for 渲染) - 进度条改为单一蓝色条,按 total/budget 比例填充;超预算时变红 - Tooltip 简化为仅显示"总计 X / 预算 Y(Z%)" 仓库:无
226 lines
6.8 KiB
Go
226 lines
6.8 KiB
Go
package newagentnode
|
||
|
||
import (
|
||
"context"
|
||
"fmt"
|
||
"strings"
|
||
"time"
|
||
|
||
"github.com/cloudwego/eino/schema"
|
||
|
||
infrallm "github.com/LoveLosita/smartflow/backend/infra/llm"
|
||
newagentmodel "github.com/LoveLosita/smartflow/backend/newAgent/model"
|
||
newagentprompt "github.com/LoveLosita/smartflow/backend/newAgent/prompt"
|
||
newagentstream "github.com/LoveLosita/smartflow/backend/newAgent/stream"
|
||
)
|
||
|
||
const (
|
||
deliverStageName = "deliver"
|
||
deliverStatusBlockID = "deliver.status"
|
||
deliverSpeakBlockID = "deliver.speak"
|
||
)
|
||
|
||
// DeliverNodeInput 描述交付节点单轮运行所需的最小依赖。
|
||
//
|
||
// 职责边界:
|
||
// 1. 只负责生成交付总结并推送给用户,不负责后续流程推进;
|
||
// 2. RuntimeState 提供计划步骤和执行状态;
|
||
// 3. ConversationContext 提供执行阶段的对话历史;
|
||
// 4. 交付完成后标记流程结束。
|
||
type DeliverNodeInput struct {
|
||
RuntimeState *newagentmodel.AgentRuntimeState
|
||
ConversationContext *newagentmodel.ConversationContext
|
||
Client *infrallm.Client
|
||
ChunkEmitter *newagentstream.ChunkEmitter
|
||
ThinkingEnabled bool // 是否开启 thinking,由 config.yaml 的 agent.thinking.deliver 注入
|
||
}
|
||
|
||
// RunDeliverNode 执行一轮交付节点逻辑。
|
||
//
|
||
// 核心职责:
|
||
// 1. 调 LLM 基于原始计划 + 执行历史生成交付总结;
|
||
// 2. 伪流式推送总结给用户;
|
||
// 3. 写入对话历史,保证上下文连续;
|
||
// 4. 标记流程结束。
|
||
//
|
||
// 降级策略:
|
||
// 1. LLM 调用失败时,回退到机械格式化总结,不中断流程;
|
||
// 2. 机械总结包含计划步骤列表和完成进度。
|
||
func RunDeliverNode(ctx context.Context, input DeliverNodeInput) error {
|
||
runtimeState, conversationContext, emitter, err := prepareDeliverNodeInput(input)
|
||
if err != nil {
|
||
return err
|
||
}
|
||
flowState := runtimeState.EnsureCommonState()
|
||
|
||
// 1. 推送交付阶段状态,让前端知道正在生成总结。
|
||
if err := emitter.EmitStatus(
|
||
deliverStatusBlockID,
|
||
deliverStageName,
|
||
"summarizing",
|
||
"正在生成交付总结。",
|
||
false,
|
||
); err != nil {
|
||
return fmt.Errorf("交付阶段状态推送失败: %w", err)
|
||
}
|
||
|
||
// 2. 调 LLM 生成交付总结。
|
||
summary := generateDeliverSummary(ctx, input.Client, flowState, conversationContext, input.ThinkingEnabled)
|
||
|
||
// 3. 伪流式推送总结。
|
||
if strings.TrimSpace(summary) != "" {
|
||
if err := emitter.EmitPseudoAssistantText(
|
||
ctx,
|
||
deliverSpeakBlockID,
|
||
deliverStageName,
|
||
summary,
|
||
newagentstream.DefaultPseudoStreamOptions(),
|
||
); err != nil {
|
||
return fmt.Errorf("交付总结推送失败: %w", err)
|
||
}
|
||
conversationContext.AppendHistory(schema.AssistantMessage(summary, nil))
|
||
}
|
||
|
||
// 4. 推送最终完成状态。
|
||
_ = emitter.EmitStatus(
|
||
deliverStatusBlockID,
|
||
deliverStageName,
|
||
"done",
|
||
"本轮流程已结束。",
|
||
true,
|
||
)
|
||
|
||
return nil
|
||
}
|
||
|
||
// generateDeliverSummary 尝试调用 LLM 生成交付总结,失败时降级到机械格式化。
|
||
func generateDeliverSummary(
|
||
ctx context.Context,
|
||
client *infrallm.Client,
|
||
flowState *newagentmodel.CommonState,
|
||
conversationContext *newagentmodel.ConversationContext,
|
||
thinkingEnabled bool,
|
||
) string {
|
||
if flowState != nil {
|
||
switch {
|
||
case flowState.IsAborted():
|
||
return normalizeSpeak(buildAbortSummary(flowState))
|
||
case flowState.IsExhaustedTerminal():
|
||
return normalizeSpeak(buildExhaustedSummary(flowState))
|
||
}
|
||
}
|
||
|
||
if client == nil {
|
||
return buildMechanicalSummary(flowState)
|
||
}
|
||
|
||
messages := newagentprompt.BuildDeliverMessages(flowState, conversationContext)
|
||
result, err := client.GenerateText(
|
||
ctx,
|
||
messages,
|
||
infrallm.GenerateOptions{
|
||
Temperature: 0.5,
|
||
MaxTokens: 800,
|
||
Thinking: resolveThinkingMode(thinkingEnabled),
|
||
Metadata: map[string]any{
|
||
"stage": deliverStageName,
|
||
},
|
||
},
|
||
)
|
||
if err != nil || result == nil || strings.TrimSpace(result.Text) == "" {
|
||
return buildMechanicalSummary(flowState)
|
||
}
|
||
|
||
return normalizeSpeak(result.Text)
|
||
}
|
||
|
||
// buildAbortSummary 生成“流程已终止”的统一交付文案。
|
||
//
|
||
// 说明:
|
||
// 1. 第二轮开始,abort 的用户可见文案由终止方提前写入 CommonState;
|
||
// 2. deliver 不再重新猜测或改写业务异常,只做最终收口;
|
||
// 3. 若历史快照缺失 user_message,则回退到一份通用说明,避免前端收到空白结果。
|
||
func buildAbortSummary(state *newagentmodel.CommonState) string {
|
||
if state == nil || state.TerminalOutcome == nil {
|
||
return "本轮流程已终止。"
|
||
}
|
||
if msg := strings.TrimSpace(state.TerminalOutcome.UserMessage); msg != "" {
|
||
return msg
|
||
}
|
||
return "本轮流程已终止,请根据当前提示检查后再继续。"
|
||
}
|
||
|
||
// buildExhaustedSummary 生成“轮次耗尽”的统一收口文案。
|
||
func buildExhaustedSummary(state *newagentmodel.CommonState) string {
|
||
if state == nil {
|
||
return "本轮执行已达到安全轮次上限,当前先停止继续操作。"
|
||
}
|
||
|
||
prefix := "本轮执行已达到安全轮次上限,当前先停止继续操作。"
|
||
if state.TerminalOutcome != nil && strings.TrimSpace(state.TerminalOutcome.UserMessage) != "" {
|
||
prefix = strings.TrimSpace(state.TerminalOutcome.UserMessage)
|
||
}
|
||
if !state.HasPlan() {
|
||
return prefix
|
||
}
|
||
return prefix + "\n\n" + strings.TrimSpace(buildMechanicalSummary(state))
|
||
}
|
||
|
||
// buildMechanicalSummary 在 LLM 不可用时,机械拼接一份最小可用总结。
|
||
func buildMechanicalSummary(state *newagentmodel.CommonState) string {
|
||
if state == nil {
|
||
return "任务流程已结束。"
|
||
}
|
||
|
||
var sb strings.Builder
|
||
current, total := state.PlanProgress()
|
||
|
||
if !state.HasPlan() {
|
||
return "任务流程已结束。"
|
||
}
|
||
|
||
if state.IsExhaustedTerminal() {
|
||
sb.WriteString(fmt.Sprintf("任务因执行轮次耗尽提前结束,已完成 %d/%d 步。\n", current, total))
|
||
} else {
|
||
sb.WriteString("所有计划步骤已执行完毕。\n")
|
||
}
|
||
|
||
sb.WriteString("\n执行情况:\n")
|
||
for i, step := range state.PlanSteps {
|
||
marker := "[ ]"
|
||
if i < current {
|
||
marker = "[x]"
|
||
}
|
||
sb.WriteString(fmt.Sprintf("%s %s\n", marker, strings.TrimSpace(step.Content)))
|
||
}
|
||
|
||
if state.IsExhaustedTerminal() && current < total {
|
||
sb.WriteString("\n如需继续完成剩余步骤,可以告诉我继续。")
|
||
}
|
||
|
||
return sb.String()
|
||
}
|
||
|
||
// prepareDeliverNodeInput 校验并准备交付节点的运行态依赖。
|
||
func prepareDeliverNodeInput(input DeliverNodeInput) (
|
||
*newagentmodel.AgentRuntimeState,
|
||
*newagentmodel.ConversationContext,
|
||
*newagentstream.ChunkEmitter,
|
||
error,
|
||
) {
|
||
if input.RuntimeState == nil {
|
||
return nil, nil, nil, fmt.Errorf("deliver node: runtime state 不能为空")
|
||
}
|
||
|
||
input.RuntimeState.EnsureCommonState()
|
||
if input.ConversationContext == nil {
|
||
input.ConversationContext = newagentmodel.NewConversationContext("")
|
||
}
|
||
if input.ChunkEmitter == nil {
|
||
input.ChunkEmitter = newagentstream.NewChunkEmitter(
|
||
newagentstream.NoopPayloadEmitter(), "", "", time.Now().Unix(),
|
||
)
|
||
}
|
||
return input.RuntimeState, input.ConversationContext, input.ChunkEmitter, nil
|
||
}
|