后端:
1.Execute 上下文修复(无限循环 / 重复确认根治)
- 更新node/execute.go:speak 写入历史(修复旧 TODO);confirm 动作 speak 不再丢失;
continue 无工具调用时写 reason 保证上下文推进;区分 tool_call 数组/JSON损坏两种
correction hint;goal_check hint 区分 plan/ReAct 模式
- 更新node/execute.go:新增 AlwaysExecute 字段,extra.always_execute=true 时写工具
跳过确认闸门直接执行并持久化
- 更新model/graph_run_state.go:AgentGraphRequest 新增 AlwaysExecute;新增
WriteSchedulePreviewFunc 类型和 WriteSchedulePreview Dep
- 更新service/agentsvc/agent.go:新增 readAgentExtraBool 辅助
2.粗排全链路修复
- 更新service/agentsvc/agent_newagent.go:makeRoughBuildFunc 改用 HybridScheduleEntry
而非 TaskClassItem.EmbeddedTime,普通时段放置不再被丢弃
- 更新conv/schedule_provider.go:LoadScheduleState 从 task class 日期范围推算多周
规划窗口,不再硬编码当前周 7 天;DayMapping 覆盖全部相关周,粗排跨周结果不再
被 WeekDayToDay 静默丢弃
- 更新node/rough_build.go:pinned block 区分有/无未覆盖 pending 任务两种情况,
有 pending 时明确操作顺序(find_free→place)和完成判定,防止 LLM 重复调
list_tasks;新增 countPendingTasks 辅助(只统计 Slots 为空的真正未覆盖任务)
- 更新model/common_state.go:新增 StartDirectExecute(),Chat 直接路由 execute 时
清空旧 PlanSteps,修复跨会话 HasPlan() 误判导致 ReAct 走 plan 模式的 bug
- 更新node/chat.go:handleRouteExecute 改用 StartDirectExecute()
3.排程预览缓存迁移至 Deliver 节点
- 更新node/agent_nodes.go:Deliver 节点完成后调用 WriteSchedulePreview,只有任务
真正完成才写预览缓存,中断路径不写中间态
- 更新service/agentsvc/agent_newagent.go:注入 makeWriteSchedulePreviewFunc;移除
graph 结束后的内联写入;makeRoughBuildFunc 注释修正
- 更新conv/schedule_preview.go:ScheduleStateToPreview 补设 GeneratedAt
- 更新model/agent.go:GetSchedulePlanPreviewResponse 新增 HybridEntries 字段
- 更新service/agentsvc/agent_schedule_preview.go:GET handler Redis/MySQL 两条路径
均透传 HybridEntries
4.Execute thinking 模式修复
- 更新newAgent/llm/ark_adapter.go:thinking 开启时强制 temperature=1,MaxTokens 自
动托底至 16000,调用方与适配层行为对齐
- 更新node/execute.go:调用参数同步改为 temperature=1.0 / MaxTokens=16000
undo:
1.流式推送换行未修复(undo)
2.上下文依然待审视
前端:无
仓库:无
212 lines
6.8 KiB
Go
212 lines
6.8 KiB
Go
package newagentprompt
|
||
|
||
import (
|
||
"fmt"
|
||
"strings"
|
||
|
||
newagentmodel "github.com/LoveLosita/smartflow/backend/newAgent/model"
|
||
"github.com/cloudwego/eino/schema"
|
||
)
|
||
|
||
// buildStageMessages 组装某个阶段通用的 messages。
|
||
//
|
||
// 消息排列策略(利用 LLM 近因效应):
|
||
// 1. system prompt(角色 + 阶段规则)— 始终最顶部,定义基本身份;
|
||
// 2. tool schemas(能力边界)— 稳定参考信息,放在 history 前即可;
|
||
// 3. history(对话历史、工具调用、修正反馈)— 按时间顺序排列;
|
||
// 4. pinned blocks(当前计划、当前步骤、粗排结果等最新约束)— 紧贴 user prompt,
|
||
// 利用近因效应让 LLM 优先关注本轮最相关的约束,而非被历史消息分散注意力;
|
||
// 5. user prompt(阶段性指令)— 始终在末尾,是本轮回答的核心触发。
|
||
func buildStageMessages(stageSystemPrompt string, ctx *newagentmodel.ConversationContext, runtimeUserPrompt string) []*schema.Message {
|
||
messages := make([]*schema.Message, 0, 4)
|
||
|
||
// 1. 合并 system prompt:基础角色约束 + 阶段规则,始终在最顶部。
|
||
mergedSystemPrompt := mergeSystemPrompts(ctx, stageSystemPrompt)
|
||
if mergedSystemPrompt != "" {
|
||
messages = append(messages, schema.SystemMessage(mergedSystemPrompt))
|
||
}
|
||
|
||
// 2. 工具摘要:稳定参考信息,放在 history 前即可。
|
||
if toolText := renderToolSchemas(ctx); toolText != "" {
|
||
messages = append(messages, schema.SystemMessage(toolText))
|
||
}
|
||
|
||
// 3. 对话历史:按时间顺序,包含工具调用结果和修正反馈。
|
||
if ctx != nil {
|
||
history := ctx.HistorySnapshot()
|
||
if len(history) > 0 {
|
||
// 兼容旧快照:裸 Tool 消息(无 ToolCallID)违反 OpenAI 兼容 API 格式约束,
|
||
// 会触发 API 拒绝请求导致连接断开。
|
||
// 这里将裸 Tool 消息降级为 User 消息,保证向后兼容。
|
||
for i, msg := range history {
|
||
if msg.Role == schema.Tool && msg.ToolCallID == "" {
|
||
history[i] = &schema.Message{
|
||
Role: schema.User,
|
||
Content: fmt.Sprintf("[工具执行结果]\n%s", msg.Content),
|
||
}
|
||
}
|
||
}
|
||
messages = append(messages, history...)
|
||
}
|
||
}
|
||
|
||
// 4. 置顶上下文块:当前计划、当前步骤、粗排结果等最新约束。
|
||
// 放在 history 之后、user prompt 之前,利用 LLM 近因效应提升对最新约束的注意力。
|
||
if pinnedText := renderPinnedBlocks(ctx); pinnedText != "" {
|
||
messages = append(messages, schema.SystemMessage(pinnedText))
|
||
}
|
||
|
||
// 5. 阶段性用户提示词:始终在末尾,是本轮回答的核心触发。
|
||
runtimeUserPrompt = strings.TrimSpace(runtimeUserPrompt)
|
||
if runtimeUserPrompt != "" {
|
||
messages = append(messages, schema.UserMessage(runtimeUserPrompt))
|
||
}
|
||
|
||
return messages
|
||
}
|
||
|
||
// renderStateSummary 把当前流程状态渲染成简洁文本。
|
||
func renderStateSummary(state *newagentmodel.CommonState) string {
|
||
if state == nil {
|
||
return "当前状态:state 缺失,请先做兜底处理。"
|
||
}
|
||
|
||
var sb strings.Builder
|
||
current, total := state.PlanProgress()
|
||
|
||
sb.WriteString(fmt.Sprintf("当前阶段:%s\n", state.Phase))
|
||
sb.WriteString(fmt.Sprintf("当前轮次:%d/%d\n", state.RoundUsed, state.MaxRounds))
|
||
|
||
if !state.HasPlan() {
|
||
sb.WriteString("当前完整 plan:暂无。\n")
|
||
} else {
|
||
sb.WriteString("当前完整 plan:\n")
|
||
for i, step := range state.PlanSteps {
|
||
sb.WriteString(fmt.Sprintf("%d. %s\n", i+1, strings.TrimSpace(step.Content)))
|
||
if strings.TrimSpace(step.DoneWhen) != "" {
|
||
sb.WriteString(fmt.Sprintf(" 完成判定:%s\n", strings.TrimSpace(step.DoneWhen)))
|
||
}
|
||
}
|
||
|
||
if step, ok := state.CurrentPlanStep(); ok {
|
||
sb.WriteString(fmt.Sprintf("当前步骤进度:%d/%d\n", current, total))
|
||
sb.WriteString("当前步骤内容:\n")
|
||
sb.WriteString(strings.TrimSpace(step.Content))
|
||
sb.WriteString("\n")
|
||
if strings.TrimSpace(step.DoneWhen) != "" {
|
||
sb.WriteString("当前步骤完成判定:\n")
|
||
sb.WriteString(strings.TrimSpace(step.DoneWhen))
|
||
sb.WriteString("\n")
|
||
}
|
||
} else {
|
||
sb.WriteString("当前步骤进度:暂时无有效当前步骤。\n")
|
||
}
|
||
}
|
||
|
||
// 渲染任务类约束元数据(如有),帮助 LLM 了解排程范围和策略,避免追问已有信息。
|
||
if len(state.TaskClasses) > 0 {
|
||
sb.WriteString("\n本次排课涉及的任务类约束:\n")
|
||
for _, tc := range state.TaskClasses {
|
||
line := fmt.Sprintf("- [ID=%d] %s:策略=%s,总时段预算=%d", tc.ID, tc.Name, tc.Strategy, tc.TotalSlots)
|
||
if tc.StartDate != "" || tc.EndDate != "" {
|
||
line += fmt.Sprintf(",日期范围=%s ~ %s", tc.StartDate, tc.EndDate)
|
||
}
|
||
if tc.AllowFillerCourse {
|
||
line += ",允许嵌入水课"
|
||
}
|
||
if len(tc.ExcludedSlots) > 0 {
|
||
line += fmt.Sprintf(",排除时段=%v", tc.ExcludedSlots)
|
||
}
|
||
sb.WriteString(line + "\n")
|
||
}
|
||
}
|
||
|
||
return sb.String()
|
||
}
|
||
|
||
// renderPinnedBlocks 把 ConversationContext 中的置顶块渲染成独立的 system 文本。
|
||
func renderPinnedBlocks(ctx *newagentmodel.ConversationContext) string {
|
||
if ctx == nil {
|
||
return ""
|
||
}
|
||
|
||
blocks := ctx.PinnedBlocksSnapshot()
|
||
if len(blocks) == 0 {
|
||
return ""
|
||
}
|
||
|
||
var sb strings.Builder
|
||
sb.WriteString("以下是后端置顶注入的上下文,请优先遵守:\n")
|
||
for _, block := range blocks {
|
||
title := strings.TrimSpace(block.Title)
|
||
if title == "" {
|
||
title = strings.TrimSpace(block.Key)
|
||
}
|
||
if title != "" {
|
||
sb.WriteString("【")
|
||
sb.WriteString(title)
|
||
sb.WriteString("】\n")
|
||
}
|
||
sb.WriteString(strings.TrimSpace(block.Content))
|
||
sb.WriteString("\n")
|
||
}
|
||
return strings.TrimSpace(sb.String())
|
||
}
|
||
|
||
// renderToolSchemas 把工具摘要渲染成独立文本块。
|
||
func renderToolSchemas(ctx *newagentmodel.ConversationContext) string {
|
||
if ctx == nil {
|
||
return ""
|
||
}
|
||
|
||
schemas := ctx.ToolSchemasSnapshot()
|
||
if len(schemas) == 0 {
|
||
return ""
|
||
}
|
||
|
||
var sb strings.Builder
|
||
sb.WriteString("以下是当前可用工具摘要,仅供你在规划时参考能力边界:\n")
|
||
for _, item := range schemas {
|
||
name := strings.TrimSpace(item.Name)
|
||
desc := strings.TrimSpace(item.Desc)
|
||
schemaText := strings.TrimSpace(item.SchemaText)
|
||
|
||
if name != "" {
|
||
sb.WriteString("- 工具名:")
|
||
sb.WriteString(name)
|
||
sb.WriteString("\n")
|
||
}
|
||
if desc != "" {
|
||
sb.WriteString(" 说明:")
|
||
sb.WriteString(desc)
|
||
sb.WriteString("\n")
|
||
}
|
||
if schemaText != "" {
|
||
sb.WriteString(" 参数摘要:")
|
||
sb.WriteString(schemaText)
|
||
sb.WriteString("\n")
|
||
}
|
||
}
|
||
|
||
return strings.TrimSpace(sb.String())
|
||
}
|
||
|
||
func mergeSystemPrompts(ctx *newagentmodel.ConversationContext, stageSystemPrompt string) string {
|
||
base := ""
|
||
if ctx != nil {
|
||
base = strings.TrimSpace(ctx.SystemPrompt)
|
||
}
|
||
stageSystemPrompt = strings.TrimSpace(stageSystemPrompt)
|
||
|
||
switch {
|
||
case base == "" && stageSystemPrompt == "":
|
||
return ""
|
||
case base == "":
|
||
return stageSystemPrompt
|
||
case stageSystemPrompt == "":
|
||
return base
|
||
default:
|
||
return base + "\n\n" + stageSystemPrompt
|
||
}
|
||
}
|