d7184b776b
后端： 1.收口阶段 6 agent 结构迁移，将 newAgent 内核与 agentsvc 编排层迁入 services/agent - 切换 Agent 启动装配与 HTTP handler 直连 agent sv，移除旧 service agent bridge - 补齐 Agent 对 memory、task、task-class、schedule 的 RPC 适配与契约字段 - 扩展 schedule、task、task-class RPC/contract 支撑 Agent 查询、写入与 provider 切流 - 更新迁移文档、README 与相关注释，明确 agent 当前切流点和剩余 memory 迁移面
399 lines
13 KiB
Go
399 lines
13 KiB
Go
package agentnode
|
||
|
||
import (
|
||
"context"
|
||
"fmt"
|
||
"io"
|
||
"log"
|
||
"strings"
|
||
"time"
|
||
|
||
"github.com/google/uuid"
|
||
|
||
agentmodel "github.com/LoveLosita/smartflow/backend/services/agent/model"
|
||
agentprompt "github.com/LoveLosita/smartflow/backend/services/agent/prompt"
|
||
agentrouter "github.com/LoveLosita/smartflow/backend/services/agent/router"
|
||
agentstream "github.com/LoveLosita/smartflow/backend/services/agent/stream"
|
||
llmservice "github.com/LoveLosita/smartflow/backend/services/llm"
|
||
"github.com/cloudwego/eino/schema"
|
||
)
|
||
|
||
const (
|
||
planStageName = "plan"
|
||
planStatusBlockID = "plan.status"
|
||
planSpeakBlockID = "plan.speak"
|
||
planSummaryBlockID = "plan.summary"
|
||
planPinnedKey = "current_plan"
|
||
planCurrentStepKey = "current_step"
|
||
planCurrentStepTitle = "当前步骤"
|
||
planFullPlanTitle = "当前完整计划"
|
||
)
|
||
|
||
// PlanNodeInput 描述单轮规划节点执行所需的最小依赖。
|
||
type PlanNodeInput struct {
|
||
RuntimeState *agentmodel.AgentRuntimeState
|
||
ConversationContext *agentmodel.ConversationContext
|
||
UserInput string
|
||
Client *llmservice.Client
|
||
ChunkEmitter *agentstream.ChunkEmitter
|
||
ResumeNode string
|
||
AlwaysExecute bool // true 时计划生成后自动确认，不进入 confirm 节点
|
||
ThinkingEnabled bool // 是否开启 thinking，由 config.yaml 的 agent.thinking.plan 注入
|
||
CompactionStore agentmodel.CompactionStore // 上下文压缩持久化
|
||
PersistVisibleMessage agentmodel.PersistVisibleMessageFunc
|
||
}
|
||
|
||
// RunPlanNode 执行一轮规划节点逻辑。
|
||
//
|
||
// 步骤说明：
|
||
// 1. 先校验最小依赖，并推送一条"正在规划"的状态，避免用户空等；
|
||
// 2. 构造本轮规划输入，调用 LLM Stream 接口；
|
||
// 3. 从流中提取 <SMARTFLOW_DECISION> 标签内的 JSON 决策，同时流式推送 speak 正文；
|
||
// 4. 按 action 推进流程：
|
||
// 4.1 continue：继续停留在 planning；
|
||
// 4.2 ask_user：打开 pending interaction，后续交给 interrupt 收口；
|
||
// 4.3 plan_done：固化完整计划，刷新 pinned context，并进入 waiting_confirm。
|
||
func RunPlanNode(ctx context.Context, input PlanNodeInput) error {
|
||
runtimeState, conversationContext, emitter, err := preparePlanNodeInput(input)
|
||
if err != nil {
|
||
return err
|
||
}
|
||
flowState := runtimeState.EnsureCommonState()
|
||
|
||
// 1. 先发一条阶段状态，让前端知道当前已经进入规划环节。
|
||
if err := emitter.EmitStatus(
|
||
planStatusBlockID,
|
||
planStageName,
|
||
"planning",
|
||
"正在梳理目标并补全执行计划。",
|
||
false,
|
||
); err != nil {
|
||
return fmt.Errorf("规划阶段状态推送失败: %w", err)
|
||
}
|
||
|
||
// 2. 构造本轮规划输入。
|
||
messages := agentprompt.BuildPlanMessages(flowState, conversationContext, input.UserInput)
|
||
messages = compactUnifiedMessagesIfNeeded(ctx, messages, UnifiedCompactInput{
|
||
Client: input.Client,
|
||
CompactionStore: input.CompactionStore,
|
||
FlowState: flowState,
|
||
Emitter: emitter,
|
||
StageName: planStageName,
|
||
StatusBlockID: planStatusBlockID,
|
||
})
|
||
logNodeLLMContext(planStageName, "planning", flowState, messages)
|
||
|
||
// 3. 两阶段流式规划：从 LLM 流中先提取 <SMARTFLOW_DECISION> 决策标签，再流式推送 speak 正文。
|
||
reader, err := input.Client.Stream(
|
||
ctx,
|
||
messages,
|
||
llmservice.GenerateOptions{
|
||
Temperature: 0.2,
|
||
// 显式设置上限，避免依赖框架默认值（默认 4096）导致长决策被截断。
|
||
// 注意：当前模型接口 max_tokens 上限为 131072，超过会 400。
|
||
MaxTokens: 131072,
|
||
Thinking: resolveThinkingMode(input.ThinkingEnabled),
|
||
Metadata: map[string]any{
|
||
"stage": planStageName,
|
||
"phase": "planning",
|
||
},
|
||
},
|
||
)
|
||
if err != nil {
|
||
return fmt.Errorf("规划阶段 Stream 调用失败: %w", err)
|
||
}
|
||
|
||
parser := agentrouter.NewStreamDecisionParser()
|
||
firstChunk := true
|
||
speakStreamed := false
|
||
reasoningDigestor, digestorErr := emitter.NewReasoningDigestor(ctx, planSpeakBlockID, planStageName)
|
||
if digestorErr != nil {
|
||
return fmt.Errorf("规划 thinking 摘要器初始化失败: %w", digestorErr)
|
||
}
|
||
defer func() {
|
||
if reasoningDigestor != nil {
|
||
_ = reasoningDigestor.Close(ctx)
|
||
}
|
||
}()
|
||
|
||
// 3.1 阶段一：解析决策标签。
|
||
for {
|
||
chunk, recvErr := reader.Recv()
|
||
if recvErr == io.EOF {
|
||
break
|
||
}
|
||
if recvErr != nil {
|
||
log.Printf("[WARN] plan stream recv error chat=%s err=%v", flowState.ConversationID, recvErr)
|
||
break
|
||
}
|
||
|
||
// thinking 内容只进入摘要器，不再把 raw reasoning_content 透传给前端。
|
||
if chunk != nil && strings.TrimSpace(chunk.ReasoningContent) != "" {
|
||
if reasoningDigestor != nil {
|
||
reasoningDigestor.Append(chunk.ReasoningContent)
|
||
}
|
||
}
|
||
|
||
content := ""
|
||
if chunk != nil {
|
||
content = chunk.Content
|
||
}
|
||
|
||
visible, ready, _ := parser.Feed(content)
|
||
if !ready {
|
||
continue
|
||
}
|
||
|
||
result := parser.Result()
|
||
if result.Fallback || result.ParseFailed {
|
||
return fmt.Errorf("规划解析失败，原始输出=%s", result.RawBuffer)
|
||
}
|
||
|
||
decision, parseErr := llmservice.ParseJSONObject[agentmodel.PlanDecision](result.DecisionJSON)
|
||
if parseErr != nil {
|
||
return fmt.Errorf("规划决策 JSON 解析失败: %w (raw=%s)", parseErr, result.RawBuffer)
|
||
}
|
||
if validateErr := decision.Validate(); validateErr != nil {
|
||
return fmt.Errorf("规划决策不合法: %w", validateErr)
|
||
}
|
||
|
||
// 3.2 阶段二：流式推送 speak（同一 reader 继续读取）。
|
||
var fullText strings.Builder
|
||
if visible != "" {
|
||
if reasoningDigestor != nil {
|
||
reasoningDigestor.MarkContentStarted()
|
||
}
|
||
if emitErr := emitter.EmitAssistantText(planSpeakBlockID, planStageName, visible, firstChunk); emitErr != nil {
|
||
return fmt.Errorf("规划文案推送失败: %w", emitErr)
|
||
}
|
||
speakStreamed = true
|
||
fullText.WriteString(visible)
|
||
firstChunk = false
|
||
}
|
||
for {
|
||
chunk2, recvErr2 := reader.Recv()
|
||
if recvErr2 == io.EOF {
|
||
break
|
||
}
|
||
if recvErr2 != nil {
|
||
log.Printf("[WARN] plan speak stream error chat=%s err=%v", flowState.ConversationID, recvErr2)
|
||
break
|
||
}
|
||
if chunk2 == nil {
|
||
continue
|
||
}
|
||
if strings.TrimSpace(chunk2.ReasoningContent) != "" {
|
||
if reasoningDigestor != nil {
|
||
reasoningDigestor.Append(chunk2.ReasoningContent)
|
||
}
|
||
}
|
||
if chunk2.Content != "" {
|
||
if reasoningDigestor != nil {
|
||
reasoningDigestor.MarkContentStarted()
|
||
}
|
||
if emitErr := emitter.EmitAssistantText(planSpeakBlockID, planStageName, chunk2.Content, firstChunk); emitErr != nil {
|
||
return fmt.Errorf("规划文案推送失败: %w", emitErr)
|
||
}
|
||
speakStreamed = true
|
||
fullText.WriteString(chunk2.Content)
|
||
firstChunk = false
|
||
}
|
||
}
|
||
decision.Speak = fullText.String()
|
||
|
||
// 4. 若有 speak 且不是 ask_user（ask_user 交给 interrupt 收口），写入历史。
|
||
if strings.TrimSpace(decision.Speak) != "" && decision.Action != agentmodel.PlanActionAskUser {
|
||
msg := schema.AssistantMessage(decision.Speak, nil)
|
||
conversationContext.AppendHistory(msg)
|
||
persistVisibleAssistantMessage(ctx, input.PersistVisibleMessage, flowState, msg)
|
||
}
|
||
|
||
// 5. 按规划动作推进流程状态。
|
||
return handlePlanAction(ctx, input, runtimeState, conversationContext, emitter, flowState, decision, speakStreamed)
|
||
}
|
||
|
||
// 流结束但未找到决策标签。
|
||
return fmt.Errorf("规划阶段流结束但未提取到决策标签")
|
||
}
|
||
|
||
// handlePlanAction 根据 PlanDecision.Action 推进流程状态。
|
||
func handlePlanAction(
|
||
ctx context.Context,
|
||
input PlanNodeInput,
|
||
runtimeState *agentmodel.AgentRuntimeState,
|
||
conversationContext *agentmodel.ConversationContext,
|
||
emitter *agentstream.ChunkEmitter,
|
||
flowState *agentmodel.CommonState,
|
||
decision *agentmodel.PlanDecision,
|
||
askUserSpeakStreamed bool,
|
||
) error {
|
||
switch decision.Action {
|
||
case agentmodel.PlanActionContinue:
|
||
flowState.Phase = agentmodel.PhasePlanning
|
||
return nil
|
||
case agentmodel.PlanActionAskUser:
|
||
question := resolvePlanAskUserText(decision)
|
||
runtimeState.OpenAskUserInteraction(uuid.NewString(), question, strings.TrimSpace(input.ResumeNode))
|
||
// 1. plan 阶段若已流式推送过 ask_user 文本，interrupt 侧应避免重复正文输出；
|
||
// 2. plan 阶段 ask_user 不会提前写入 history，这里显式标记为 false。
|
||
runtimeState.SetPendingInteractionMetadata(agentmodel.PendingMetaAskUserSpeakStreamed, askUserSpeakStreamed)
|
||
runtimeState.SetPendingInteractionMetadata(agentmodel.PendingMetaAskUserHistoryAppended, false)
|
||
return nil
|
||
case agentmodel.PlanActionDone:
|
||
flowState.FinishPlan(decision.PlanSteps)
|
||
flowState.PendingContextHook = clonePlanContextHook(decision.ContextHook)
|
||
writePlanPinnedBlocks(conversationContext, decision.PlanSteps)
|
||
if decision.NeedsRoughBuild {
|
||
flowState.NeedsRoughBuild = true
|
||
if len(decision.TaskClassIDs) > 0 {
|
||
flowState.TaskClassIDs = decision.TaskClassIDs
|
||
}
|
||
}
|
||
// always_execute 开启时，计划层跳过确认闸门，直接进入执行阶段。
|
||
if input.AlwaysExecute {
|
||
summary := strings.TrimSpace(buildPlanSummary(decision.PlanSteps))
|
||
if summary != "" {
|
||
msg := schema.AssistantMessage(summary, nil)
|
||
if err := emitter.EmitPseudoAssistantText(
|
||
ctx,
|
||
planSummaryBlockID,
|
||
planStageName,
|
||
summary,
|
||
agentstream.DefaultPseudoStreamOptions(),
|
||
); err != nil {
|
||
return fmt.Errorf("自动执行前计划摘要推送失败: %w", err)
|
||
}
|
||
conversationContext.AppendHistory(msg)
|
||
persistVisibleAssistantMessage(ctx, input.PersistVisibleMessage, flowState, msg)
|
||
}
|
||
|
||
flowState.ConfirmPlan()
|
||
_ = emitter.EmitStatus(
|
||
planStatusBlockID,
|
||
planStageName,
|
||
"plan_auto_confirmed",
|
||
"计划已自动确认，开始执行。",
|
||
false,
|
||
)
|
||
}
|
||
return nil
|
||
default:
|
||
llmOutput := decision.Speak
|
||
if strings.TrimSpace(llmOutput) == "" {
|
||
llmOutput = decision.Reason
|
||
}
|
||
AppendLLMCorrectionWithHint(
|
||
conversationContext,
|
||
llmOutput,
|
||
fmt.Sprintf("你输出的 action \"%s\" 不是合法的执行动作。", decision.Action),
|
||
"合法的 action 包括：continue（继续当前步骤）、ask_user（追问用户）、next_plan（推进到下一步）、done（任务完成）。",
|
||
)
|
||
return nil
|
||
}
|
||
}
|
||
|
||
func preparePlanNodeInput(input PlanNodeInput) (*agentmodel.AgentRuntimeState, *agentmodel.ConversationContext, *agentstream.ChunkEmitter, error) {
|
||
if input.RuntimeState == nil {
|
||
return nil, nil, nil, fmt.Errorf("plan node: runtime state 不能为空")
|
||
}
|
||
if input.Client == nil {
|
||
return nil, nil, nil, fmt.Errorf("plan node: plan client 未注入")
|
||
}
|
||
|
||
input.RuntimeState.EnsureCommonState()
|
||
if input.ConversationContext == nil {
|
||
input.ConversationContext = agentmodel.NewConversationContext("")
|
||
}
|
||
if input.ChunkEmitter == nil {
|
||
input.ChunkEmitter = agentstream.NewChunkEmitter(agentstream.NoopPayloadEmitter(), "", "", time.Now().Unix())
|
||
}
|
||
return input.RuntimeState, input.ConversationContext, input.ChunkEmitter, nil
|
||
}
|
||
|
||
func resolvePlanAskUserText(decision *agentmodel.PlanDecision) string {
|
||
if decision == nil {
|
||
return "我还缺一点关键信息，想先向你确认一下。"
|
||
}
|
||
if strings.TrimSpace(decision.Speak) != "" {
|
||
return strings.TrimSpace(decision.Speak)
|
||
}
|
||
if strings.TrimSpace(decision.Reason) != "" {
|
||
return strings.TrimSpace(decision.Reason)
|
||
}
|
||
return "我还缺一点关键信息，想先向你确认一下。"
|
||
}
|
||
|
||
func clonePlanContextHook(hook *agentmodel.ContextHook) *agentmodel.ContextHook {
|
||
if hook == nil {
|
||
return nil
|
||
}
|
||
cloned := *hook
|
||
if len(hook.Packs) > 0 {
|
||
cloned.Packs = append([]string(nil), hook.Packs...)
|
||
}
|
||
cloned.Normalize()
|
||
if cloned.Domain == "" {
|
||
return nil
|
||
}
|
||
return &cloned
|
||
}
|
||
|
||
func writePlanPinnedBlocks(ctx *agentmodel.ConversationContext, steps []agentmodel.PlanStep) {
|
||
if ctx == nil {
|
||
return
|
||
}
|
||
|
||
fullPlanText := buildPinnedPlanText(steps)
|
||
if strings.TrimSpace(fullPlanText) != "" {
|
||
ctx.UpsertPinnedBlock(agentmodel.ContextBlock{
|
||
Key: planPinnedKey,
|
||
Title: planFullPlanTitle,
|
||
Content: fullPlanText,
|
||
})
|
||
}
|
||
|
||
if len(steps) == 0 {
|
||
return
|
||
}
|
||
|
||
firstStep := strings.TrimSpace(steps[0].Content)
|
||
if strings.TrimSpace(steps[0].DoneWhen) != "" {
|
||
firstStep = fmt.Sprintf("%s\n完成判定：%s", firstStep, strings.TrimSpace(steps[0].DoneWhen))
|
||
}
|
||
ctx.UpsertPinnedBlock(agentmodel.ContextBlock{
|
||
Key: planCurrentStepKey,
|
||
Title: planCurrentStepTitle,
|
||
Content: firstStep,
|
||
})
|
||
}
|
||
|
||
func buildPinnedPlanText(steps []agentmodel.PlanStep) string {
|
||
if len(steps) == 0 {
|
||
return ""
|
||
}
|
||
|
||
lines := make([]string, 0, len(steps))
|
||
for i, step := range steps {
|
||
content := strings.TrimSpace(step.Content)
|
||
if content == "" {
|
||
continue
|
||
}
|
||
|
||
line := fmt.Sprintf("%d. %s", i+1, content)
|
||
if strings.TrimSpace(step.DoneWhen) != "" {
|
||
line += fmt.Sprintf("\n完成判定：%s", strings.TrimSpace(step.DoneWhen))
|
||
}
|
||
lines = append(lines, line)
|
||
}
|
||
return strings.TrimSpace(strings.Join(lines, "\n\n"))
|
||
}
|
||
|
||
// resolveThinkingMode 根据配置布尔值返回对应的 ThinkingMode。
|
||
// 供 plan / execute / deliver 节点统一使用。
|
||
func resolveThinkingMode(enabled bool) llmservice.ThinkingMode {
|
||
if enabled {
|
||
return llmservice.ThinkingModeEnabled
|
||
}
|
||
return llmservice.ThinkingModeDisabled
|
||
}
|