Files
smartmate/backend/newAgent/node/plan.go
LoveLosita 07d307fe07 Version: 0.9.4.dev.260407
后端:
1.粗排结果/预览语义修复(task_item suggested 保真 + existing/嵌入识别补全)
- 更新conv/schedule_state.go:LoadScheduleState 补齐 event.rel_id / schedules.embedded_task_id / task_item.embedded_time 三种“已落位”信号;嵌入任务强制 existing + 继承 host slots;补充 task_item duration/name/slot helper;Diff 相关英文注释改中文
- 更新conv/schedule_preview.go:预览层新增 shouldMarkSuggestedInPreview,pending 任务与 source=task_item 的建议态任务统一输出 suggested
2.newAgent 状态快照增强(ScheduleState/OriginalScheduleState 跨轮恢复)
- 更新model/state_store.go:AgentStateSnapshot 新增 ScheduleState / OriginalScheduleState
- 更新model/graph_run_state.go:AgentGraphRunInput/AgentGraphState 接入两份 schedule 状态;恢复旧快照时自动补 original clone
- 更新service/agentsvc/agent_newagent.go:loadOrCreateRuntimeState 返回并恢复 schedule/original;runNewAgentGraph 透传到 graph
- 更新node/agent_nodes.go:saveAgentState 一并保存 schedule/original 到 Redis 快照 3.Execute 链路纠偏(只写内存不落库 + 完整打点 + 恢复消息去重)
- 更新node/execute.go:AlwaysExecute/confirm resume 路径取消 PersistScheduleChanges,仅保留内存写;新增 execute LLM 完整上下文日志;新增工具调用前后 state 摘要日志;thinking 模式改为 enabled
- 更新node/chat.go:pending resume 不再重复写入同一轮 user message
- 更新service/agentsvc/agent_newagent.go:新增 deliver preview write/state 摘要日志,便于排查 suggested 丢失问题
4.AlwaysExecute 贯通 Plan→Graph→Execute
- 更新node/plan.go:PlanNodeInput 新增 AlwaysExecute;plan_done 后支持自动确认直接进入执行
- 更新graph/common_graph.go:branchAfterPlan 支持 PhaseExecuting/PhaseDone 分支
5.排课上下文补强(显式注入 task_class_ids,减少 Execute 误 ask_user)
- 更新prompt/execute.go:Plan/ReAct 两种 execute prompt 都显式写入任务类 ID,声明“上下文已完整,无需追问”
- 更新node/rough_build.go:粗排完成 pinned block 显式标注任务类 ID,避免 Execute 找不到 ID 来源
6.流式输出与预览调试工具修复
- 更新stream/emitter.go:保留换行,修复 pseudo stream 分片后文本黏连/双换行问题
- 更新infra/schedule_preview_viewer.html:升级预览工具,支持 candidate_plans / hybrid_entries

前端:无
仓库:
1.更新了infra内的html,适应了获取日程接口
2026-04-07 21:13:59 +08:00

281 lines
9.8 KiB
Go
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
package newagentnode
import (
"context"
"fmt"
"strings"
"time"
"github.com/google/uuid"
newagentllm "github.com/LoveLosita/smartflow/backend/newAgent/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"
"github.com/cloudwego/eino/schema"
)
const (
planStageName = "plan"
planStatusBlockID = "plan.status"
planSpeakBlockID = "plan.speak"
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 *newagentmodel.AgentRuntimeState
ConversationContext *newagentmodel.ConversationContext
UserInput string
Client *newagentllm.Client
ChunkEmitter *newagentstream.ChunkEmitter
ResumeNode string
AlwaysExecute bool // true 时计划生成后自动确认,不进入 confirm 节点
}
// 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. Phase 1快速评估不开 thinking让 LLM 同时产出复杂度评估和规划结果;
// 3. Phase 2深度规划若 LLM 自评需要深度思考且规划已完成,开 thinking 重跑;
// 4. 若模型先对用户说了话,则先把 speak 伪流式推给前端,并写回 history
// 5. 最后按 action 推进流程:
// 5.1 continue继续停留在 planning
// 5.2 ask_user打开 pending interaction后续交给 interrupt 收口;
// 5.3 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 := newagentprompt.BuildPlanMessages(flowState, conversationContext, input.UserInput)
// 3. Phase 1快速评估开 thinking让 LLM 同时产出复杂度评估和规划结果。
decision, rawResult, err := newagentllm.GenerateJSON[newagentmodel.PlanDecision](
ctx,
input.Client,
messages,
newagentllm.GenerateOptions{
Temperature: 0.2,
MaxTokens: 1600,
Thinking: newagentllm.ThinkingModeEnabled,
Metadata: map[string]any{
"stage": planStageName,
"phase": "assessment",
},
},
)
if err != nil {
if rawResult != nil && strings.TrimSpace(rawResult.Text) != "" {
return fmt.Errorf("规划评估解析失败,原始输出=%s错误=%w", strings.TrimSpace(rawResult.Text), err)
}
return fmt.Errorf("规划评估阶段模型调用失败: %w", err)
}
if err := decision.Validate(); err != nil {
return fmt.Errorf("规划评估决策不合法: %w", err)
}
// 4. Phase 2若 LLM 自评需要深度思考且本轮规划已完成,则开启 thinking 重跑。
// 条件NeedThinking=true + Action=plan_done → 说明 LLM 认为当前无 thinking 的计划质量不够。
// 其他 actioncontinue / ask_user不需要 thinking直接用 Phase 1 结果。
if decision.NeedThinking && decision.Action == newagentmodel.PlanActionDone {
if err := emitter.EmitStatus(
planStatusBlockID,
planStageName,
"deep_planning",
"正在深入思考,生成更完善的计划。",
false,
); err != nil {
return fmt.Errorf("深度规划状态推送失败: %w", err)
}
deepDecision, _, deepErr := newagentllm.GenerateJSON[newagentmodel.PlanDecision](
ctx,
input.Client,
messages,
newagentllm.GenerateOptions{
Temperature: 0.2,
MaxTokens: 3200,
Thinking: newagentllm.ThinkingModeEnabled,
Metadata: map[string]any{
"stage": planStageName,
"phase": "deep_planning",
},
},
)
if deepErr == nil && deepDecision != nil {
if validateErr := deepDecision.Validate(); validateErr == nil {
decision = deepDecision
}
}
// 深度规划失败时静默降级到 Phase 1 结果,不中断流程。
}
// 5. 若模型先对用户说了话,且不是 ask_userask_user 交给 interrupt 收口),则先以伪流式推送,再写回 history。
if strings.TrimSpace(decision.Speak) != "" && decision.Action != newagentmodel.PlanActionAskUser {
if err := emitter.EmitPseudoAssistantText(
ctx,
planSpeakBlockID,
planStageName,
decision.Speak,
newagentstream.DefaultPseudoStreamOptions(),
); err != nil {
return fmt.Errorf("规划文案推送失败: %w", err)
}
conversationContext.AppendHistory(schema.AssistantMessage(decision.Speak, nil))
}
// 6. 按规划动作推进流程状态。
switch decision.Action {
case newagentmodel.PlanActionContinue:
flowState.Phase = newagentmodel.PhasePlanning
return nil
case newagentmodel.PlanActionAskUser:
question := resolvePlanAskUserText(decision)
runtimeState.OpenAskUserInteraction(uuid.NewString(), question, strings.TrimSpace(input.ResumeNode))
return nil
case newagentmodel.PlanActionDone:
// 4.1 直接把结构化 PlanStep 固化到 CommonState避免 state 层丢失 done_when。
// 4.2 再把完整自然语言计划写入 pinned context保证后续 execute 优先看到。
// 4.3 若 LLM 识别到批量排课意图,把 NeedsRoughBuild 标记写入 CommonState
// Confirm 节点后的路由会据此决定是否跳入 RoughBuild 节点。
// 4.4 最后进入 waiting_confirm等待用户确认整体计划。
flowState.FinishPlan(decision.PlanSteps)
writePlanPinnedBlocks(conversationContext, decision.PlanSteps)
if decision.NeedsRoughBuild {
flowState.NeedsRoughBuild = true
// 以 LLM 决策中的 task_class_ids 为准(若非空则覆盖前端传入值)。
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。
// 这样可以与 Execute 节点的“写工具跳过确认”语义保持一致。
if input.AlwaysExecute {
flowState.ConfirmPlan()
_ = emitter.EmitStatus(
planStatusBlockID,
planStageName,
"plan_auto_confirmed",
"计划已自动确认,开始执行。",
false,
)
}
return nil
default:
// 1. LLM 输出了不支持的 action不应直接报错终止而应给它修正机会。
// 2. 使用通用修正函数追加错误反馈,让 Graph 继续循环。
// 3. LLM 下一轮会看到错误反馈并修正自己的输出。
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) (*newagentmodel.AgentRuntimeState, *newagentmodel.ConversationContext, *newagentstream.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 = newagentmodel.NewConversationContext("")
}
if input.ChunkEmitter == nil {
input.ChunkEmitter = newagentstream.NewChunkEmitter(newagentstream.NoopPayloadEmitter(), "", "", time.Now().Unix())
}
return input.RuntimeState, input.ConversationContext, input.ChunkEmitter, nil
}
func resolvePlanAskUserText(decision *newagentmodel.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 writePlanPinnedBlocks(ctx *newagentmodel.ConversationContext, steps []newagentmodel.PlanStep) {
if ctx == nil {
return
}
fullPlanText := buildPinnedPlanText(steps)
if strings.TrimSpace(fullPlanText) != "" {
ctx.UpsertPinnedBlock(newagentmodel.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(newagentmodel.ContextBlock{
Key: planCurrentStepKey,
Title: planCurrentStepTitle,
Content: firstStep,
})
}
func buildPinnedPlanText(steps []newagentmodel.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"))
}