Version: 0.7.0.dev.260319

 feat(agent): 新增智能排程 Agent 全链路 + ReAct 精排引擎

  🏗️ 智能排程 Graph 编排(阶段 1 基础链路)
  - 新增 scheduleplan 包:state / tool / prompt / nodes / runner / graph 六件套
  - 实现 plan → preview → materialize → apply → reflect → finalize 完整图编排
  - 通过函数注入解耦 agent 层与 service 层,避免循环依赖
  - 路由层新增 schedule_plan 动作,复用现有 SSE + 持久化链路

  🧠 ReAct 精排引擎(阶段 1.5 语义化微调)
  - 粗排后构建"混合日程"(既有课程 + 建议任务),统一为 HybridScheduleEntry
  - LLM 开启深度思考,通过 Swap / Move / TimeAvailable / GetAvailableSlots 四个 Tool 在内存中优化任务时间
  - reasoning_content 实时流式推送前端,用户可见 AI 思考过程
  - 精排结果仅预览不落库,向后兼容(未注入依赖时走原有 materialize 路径)

  📝 文档
  - 新增 ReAct 精排引擎决策记录

  ⚠️ 已知问题:深度思考模式耗时较长,超时策略待优化
This commit is contained in:
Losita
2026-03-19 23:16:35 +08:00
parent cd95aeeaaa
commit d3cec2a5b9
24 changed files with 2737 additions and 24 deletions

View File

@@ -0,0 +1,147 @@
package scheduleplan
import (
"context"
"github.com/cloudwego/eino-ext/components/model/ark"
"github.com/cloudwego/eino/compose"
"github.com/cloudwego/eino/schema"
)
// schedulePlanRunner 是"单次图运行"的请求级依赖容器。
//
// 设计目标:
// 1) 把节点运行所需依赖model/deps/emit/extra/history就近收口
// 2) 让 graph.go 只保留"节点连线"和"方法引用",提升可读性;
// 3) 避免在 graph.go 里重复出现内联闭包和参数透传。
type schedulePlanRunner struct {
chatModel *ark.ChatModel
deps SchedulePlanToolDeps
emitStage func(stage, detail string)
userMessage string
extra map[string]any
chatHistory []*schema.Message
// ── ReAct 精排所需 ──
outChan chan<- string // SSE 流式输出通道,用于推送 reasoning_content
modelName string // 模型名称,用于构造 OpenAI 兼容 chunk
}
// newSchedulePlanRunner 构造请求级 runner。
// 生命周期仅限一次 graph invoke不做跨请求复用。
func newSchedulePlanRunner(
chatModel *ark.ChatModel,
deps SchedulePlanToolDeps,
emitStage func(stage, detail string),
userMessage string,
extra map[string]any,
chatHistory []*schema.Message,
outChan chan<- string,
modelName string,
) *schedulePlanRunner {
return &schedulePlanRunner{
chatModel: chatModel,
deps: deps,
emitStage: emitStage,
userMessage: userMessage,
extra: extra,
chatHistory: chatHistory,
outChan: outChan,
modelName: modelName,
}
}
// ── 节点方法引用适配层 ──
func (r *schedulePlanRunner) planNode(ctx context.Context, st *SchedulePlanState) (*SchedulePlanState, error) {
return runPlanNode(ctx, st, r.chatModel, r.userMessage, r.extra, r.chatHistory, r.emitStage)
}
func (r *schedulePlanRunner) previewNode(ctx context.Context, st *SchedulePlanState) (*SchedulePlanState, error) {
return runPreviewNode(ctx, st, r.deps, r.emitStage)
}
func (r *schedulePlanRunner) materializeNode(ctx context.Context, st *SchedulePlanState) (*SchedulePlanState, error) {
return runMaterializeNode(ctx, st, r.chatModel, r.deps, r.emitStage)
}
func (r *schedulePlanRunner) applyNode(ctx context.Context, st *SchedulePlanState) (*SchedulePlanState, error) {
return runApplyNode(ctx, st, r.deps, r.emitStage)
}
func (r *schedulePlanRunner) reflectNode(ctx context.Context, st *SchedulePlanState) (*SchedulePlanState, error) {
return runReflectNode(ctx, st, r.chatModel, r.emitStage)
}
func (r *schedulePlanRunner) finalizeNode(ctx context.Context, st *SchedulePlanState) (*SchedulePlanState, error) {
return runFinalizeNode(ctx, st, r.chatModel, r.emitStage)
}
func (r *schedulePlanRunner) exitNode(_ context.Context, st *SchedulePlanState) (*SchedulePlanState, error) {
// exit 节点不做任何业务逻辑,仅把当前状态原样透传到 END。
return st, nil
}
// ── ReAct 精排节点适配层 ──
func (r *schedulePlanRunner) hybridBuildNode(ctx context.Context, st *SchedulePlanState) (*SchedulePlanState, error) {
return runHybridBuildNode(ctx, st, r.deps, r.emitStage)
}
func (r *schedulePlanRunner) reactRefineNode(ctx context.Context, st *SchedulePlanState) (*SchedulePlanState, error) {
return runReactRefineNode(ctx, st, r.chatModel, r.outChan, r.modelName, r.emitStage)
}
func (r *schedulePlanRunner) returnPreviewNode(ctx context.Context, st *SchedulePlanState) (*SchedulePlanState, error) {
return runReturnPreviewNode(ctx, st, r.emitStage)
}
// ── 分支决策适配层 ──
func (r *schedulePlanRunner) nextAfterPlan(_ context.Context, st *SchedulePlanState) (string, error) {
return selectNextAfterPlan(st), nil
}
func (r *schedulePlanRunner) nextAfterMaterialize(_ context.Context, st *SchedulePlanState) (string, error) {
// materialize 后:有 ApplyRequest 则去 apply否则去 exit。
if st == nil || st.ApplyRequest == nil || len(st.ApplyRequest.Items) == 0 {
return schedulePlanGraphNodeExit, nil
}
return schedulePlanGraphNodeApply, nil
}
func (r *schedulePlanRunner) nextAfterApply(_ context.Context, st *SchedulePlanState) (string, error) {
return selectNextAfterApply(st), nil
}
func (r *schedulePlanRunner) nextAfterReflect(_ context.Context, st *SchedulePlanState) (string, error) {
return selectNextAfterReflect(st), nil
}
// nextAfterPreview 根据 preview 结果决定下一步。
//
// 分支规则:
// 1) preview 失败(无候选方案)-> exit
// 2) HybridScheduleWithPlan 已注入 -> hybridBuild走 ReAct 精排路径)
// 3) 否则 -> materialize走原有落库路径向后兼容
func (r *schedulePlanRunner) nextAfterPreview(_ context.Context, st *SchedulePlanState) (string, error) {
if st == nil || len(st.CandidatePlans) == 0 {
return schedulePlanGraphNodeExit, nil
}
if r.deps.HybridScheduleWithPlan != nil {
return schedulePlanGraphNodeHybridBuild, nil
}
return schedulePlanGraphNodeMaterialize, nil
}
// nextAfterHybridBuild 根据 hybridBuild 结果决定下一步。
func (r *schedulePlanRunner) nextAfterHybridBuild(_ context.Context, st *SchedulePlanState) (string, error) {
if st == nil || len(st.HybridEntries) == 0 {
return schedulePlanGraphNodeExit, nil
}
return schedulePlanGraphNodeReactRefine, nil
}
// nextAfterFinalize 用于 finalize 分支——固定结束。
func (r *schedulePlanRunner) nextAfterFinalize(_ context.Context, _ *SchedulePlanState) (string, error) {
return compose.END, nil
}