Version: 0.9.45.dev.260427
后端： 1. execute 主链路重构为“上下文工具域 + 主动优化候选闭环”——移除 order_guard，粗排后默认进入主动微调，先诊断再从后端候选中选择 move/swap，避免 LLM 自由全局乱搜 2. 工具体系升级为动态注入协议——新增 context_tools_add / remove、工具域与二级包映射、主动优化白名单；schedule / taskclass / web 工具按域按包暴露，msg0 规则包与 execute 上下文同步重写 3. analyze_health 升级为主动优化唯一裁判入口——补齐 rhythm / tightness / profile / feasibility 指标、候选扫描与复诊打分、停滞信号、forced imperfection 判定，并把连续优化状态写回运行态 4. 任务类能力并入新 Agent 执行链——新增 upsert_task_class 写工具与启动注入事务写入；任务类模型补充学科画像与整天屏蔽配置，粗排支持 excluded_days_of_week，steady 策略改为基于目标位置/单日负载/分散度/缓冲的候选打分 5. 运行态与路由补齐优化模式语义——新增 active tool domain/packs、pending context hook、active optimize only、taskclass 写入回盘快照；区分 first_full / global_reopt / local_adjust，并完善首次粗排后默认 refine 的判定 前端： 6. 助手时间线渲染细化——推理内容改为独立 reasoning block，支持与工具/状态/正文按时序交错展示，自动收口折叠，修正 confirm reject 恢复动作 仓库： 7. newAgent 文档整体迁入 docs/backend，补充主动优化执行规划与顺序约束拆解文档，删除旧调试日志文件 PS:这次科研了2天，总算是有些进展了——LLM永远只适合做选择题、判断题，不适合做开放创新题。
This commit is contained in:
Losita
318
backend/newAgent/prompt/execute_rule_packs.go
Normal file
318
backend/newAgent/prompt/execute_rule_packs.go
Normal file
@@ -0,0 +1,318 @@
|
||||
package newagentprompt
|
||||
|
||||
import (
|
||||
"fmt"
|
||||
"strings"
|
||||
"time"
|
||||
|
||||
newagentmodel "github.com/LoveLosita/smartflow/backend/newAgent/model"
|
||||
newagenttools "github.com/LoveLosita/smartflow/backend/newAgent/tools"
|
||||
)
|
||||
|
||||
const (
|
||||
executeRulePackCoreMin = "core_min"
|
||||
executeRulePackSafetyHard = "safety_hard"
|
||||
executeRulePackContextProtocol = "context_protocol"
|
||||
executeRulePackModePlan = "mode_plan"
|
||||
executeRulePackModeReAct = "mode_react"
|
||||
executeRulePackDomainSchedule = "domain_schedule"
|
||||
executeRulePackDomainTaskClass = "domain_taskclass"
|
||||
executeRulePackScheduleMutation = "schedule_mutation"
|
||||
executeRulePackScheduleAnalyze = "schedule_analyze"
|
||||
executeRulePackScheduleWeb = "schedule_web"
|
||||
executeRulePackMicroRoughDone = "micro_rough_build_done"
|
||||
executeRulePackMicroDiagLoop = "micro_diag_tune_loop"
|
||||
executeRulePackMicroQueue = "micro_queue_chain"
|
||||
executeRulePackMicroTaskRetry = "micro_taskclass_retry"
|
||||
)
|
||||
|
||||
const executeSystemPromptBaseWithPlan = `
|
||||
你叫 SmartMate，是时伴（SmartMate）的中文 AI 排程伙伴，面向大学生提供陪伴式日程管理与日常协助。
|
||||
你擅长课表与任务安排、任务管理、学习规划和随口记，也可以正常回答日常问答、生活建议、信息整理、分析讨论等非排程问题。
|
||||
你的目标是像一个越用越懂用户的伙伴一样，结合历史对话、长期记忆和当前上下文，给出贴心、清晰、可信的帮助。
|
||||
你当前处于“计划执行”模式。你必须围绕当前计划步骤推进，并通过 SMARTFLOW_DECISION 输出结构化动作。`
|
||||
|
||||
const executeSystemPromptBaseReAct = `
|
||||
你叫 SmartMate，是时伴（SmartMate）的中文 AI 排程伙伴，面向大学生提供陪伴式日程管理与日常协助。
|
||||
你擅长课表与任务安排、任务管理、学习规划和随口记，也可以正常回答日常问答、生活建议、信息整理、分析讨论等非排程问题。
|
||||
你的目标是像一个越用越懂用户的伙伴一样，结合历史对话、长期记忆和当前上下文，给出贴心、清晰、可信的帮助。
|
||||
你当前处于“自由执行（ReAct）”模式。你需要根据当前目标自主推进、按需调用工具，并通过 SMARTFLOW_DECISION 输出结构化动作。`
|
||||
|
||||
type executeRulePack struct {
|
||||
Name string
|
||||
Content string
|
||||
}
|
||||
|
||||
// renderExecuteRulePackSection 渲染 execute.msg0 的动态规则包区域。
|
||||
//
|
||||
// 1. 这里负责“选哪些包 + 以什么顺序展示”，不负责工具目录本身。
|
||||
// 2. 固定先放通用硬约束，再放 mode/domain/micro 包，保证模型先读边界后读特例。
|
||||
// 3. 如果没有任何可展示规则包，则直接返回空串，避免无意义占位。
|
||||
func renderExecuteRulePackSection(state *newagentmodel.CommonState, ctx *newagentmodel.ConversationContext) (string, []string) {
|
||||
packs := selectExecuteRulePacks(state, ctx)
|
||||
if len(packs) == 0 {
|
||||
return "", nil
|
||||
}
|
||||
|
||||
lines := []string{"执行规则包（msg0 动态注入）:"}
|
||||
names := make([]string, 0, len(packs))
|
||||
for _, pack := range packs {
|
||||
content := strings.TrimSpace(pack.Content)
|
||||
if content == "" {
|
||||
continue
|
||||
}
|
||||
lines = append(lines, fmt.Sprintf("[%s]", pack.Name))
|
||||
lines = append(lines, content)
|
||||
names = append(names, pack.Name)
|
||||
}
|
||||
if len(names) == 0 {
|
||||
return "", nil
|
||||
}
|
||||
return strings.Join(lines, "\n"), names
|
||||
}
|
||||
|
||||
func selectExecuteRulePacks(state *newagentmodel.CommonState, ctx *newagentmodel.ConversationContext) []executeRulePack {
|
||||
selected := make([]executeRulePack, 0, 8)
|
||||
seen := map[string]bool{}
|
||||
|
||||
appendPack := func(pack executeRulePack) {
|
||||
name := strings.TrimSpace(pack.Name)
|
||||
if name == "" || seen[name] {
|
||||
return
|
||||
}
|
||||
seen[name] = true
|
||||
selected = append(selected, pack)
|
||||
}
|
||||
|
||||
appendPack(buildExecuteCoreMinPack())
|
||||
appendPack(buildExecuteSafetyHardPack())
|
||||
appendPack(buildExecuteContextProtocolPack())
|
||||
|
||||
if state != nil && state.HasPlan() {
|
||||
appendPack(buildExecuteModePlanPack())
|
||||
} else {
|
||||
appendPack(buildExecuteModeReActPack())
|
||||
}
|
||||
|
||||
switch normalizeExecuteToolDomain(readExecuteActiveToolDomain(state)) {
|
||||
case "schedule":
|
||||
activePacks := readExecuteActiveToolPacks(state)
|
||||
appendPack(buildExecuteSchedulePack())
|
||||
if hasExecutePack(activePacks, newagenttools.ToolPackQueue) {
|
||||
appendPack(buildExecuteQueueMicroPack())
|
||||
}
|
||||
if hasExecutePack(activePacks, newagenttools.ToolPackMutation) {
|
||||
appendPack(buildExecuteScheduleMutationPack())
|
||||
}
|
||||
if hasExecutePack(activePacks, newagenttools.ToolPackAnalyze) {
|
||||
appendPack(buildExecuteScheduleAnalyzePackV2())
|
||||
}
|
||||
if hasExecutePack(activePacks, newagenttools.ToolPackWeb) {
|
||||
appendPack(buildExecuteScheduleWebPack())
|
||||
}
|
||||
case "taskclass":
|
||||
appendPack(buildExecuteTaskClassPack())
|
||||
}
|
||||
|
||||
if hasExecuteRoughBuildDone(ctx) {
|
||||
appendPack(buildExecuteRoughDoneMicroPack())
|
||||
}
|
||||
if shouldInjectExecuteDiagLoopPack(state, ctx) {
|
||||
appendPack(buildExecuteDiagLoopMicroPackV2())
|
||||
}
|
||||
if state != nil && state.TaskClassUpsertLastTried && !state.TaskClassUpsertLastSuccess {
|
||||
appendPack(buildExecuteTaskClassRetryMicroPack())
|
||||
}
|
||||
|
||||
return selected
|
||||
}
|
||||
|
||||
func readExecuteActiveToolDomain(state *newagentmodel.CommonState) string {
|
||||
if state == nil {
|
||||
return ""
|
||||
}
|
||||
return strings.TrimSpace(state.ActiveToolDomain)
|
||||
}
|
||||
|
||||
func readExecuteActiveToolPacks(state *newagentmodel.CommonState) []string {
|
||||
if state == nil {
|
||||
return nil
|
||||
}
|
||||
return newagenttools.ResolveEffectiveToolPacks(state.ActiveToolDomain, state.ActiveToolPacks)
|
||||
}
|
||||
|
||||
func hasExecutePack(packs []string, target string) bool {
|
||||
target = strings.ToLower(strings.TrimSpace(target))
|
||||
if target == "" {
|
||||
return false
|
||||
}
|
||||
for _, pack := range packs {
|
||||
if strings.ToLower(strings.TrimSpace(pack)) == target {
|
||||
return true
|
||||
}
|
||||
}
|
||||
return false
|
||||
}
|
||||
|
||||
// containsExecutePack 兼容旧调用点。
|
||||
//
|
||||
// 1. 这里只做别名转发，不引入第二套判断口径。
|
||||
// 2. 保留它是为了避免下一轮再因为历史调用点而误删。
|
||||
func containsExecutePack(packs []string, target string) bool {
|
||||
return hasExecutePack(packs, target)
|
||||
}
|
||||
|
||||
func normalizeExecuteToolDomain(domain string) string {
|
||||
switch strings.ToLower(strings.TrimSpace(domain)) {
|
||||
case "schedule":
|
||||
return "schedule"
|
||||
case "taskclass":
|
||||
return "taskclass"
|
||||
default:
|
||||
return ""
|
||||
}
|
||||
}
|
||||
|
||||
func buildExecuteCoreMinPack() executeRulePack {
|
||||
return executeRulePack{
|
||||
Name: executeRulePackCoreMin,
|
||||
Content: strings.TrimSpace(fmt.Sprintf(`
|
||||
- 当前时间锚点：%s。涉及“今天/明天/本周”等相对时间时，先按该锚点换算。
|
||||
- 用户意图优先：只推进用户当前明确要求；未明确部分优先 ask_user。
|
||||
- 先事实后动作：优先读工具补齐事实，再决定下一步。
|
||||
- 只要决定调用 place/move/swap/batch_move/unplace 这类写工具，就必须输出 action=confirm；continue + 写工具无效。
|
||||
- 输出格式固定：先 <SMARTFLOW_DECISION>{JSON}</SMARTFLOW_DECISION>，再输出用户可见正文。`,
|
||||
buildExecuteNowAnchorLine())),
|
||||
}
|
||||
}
|
||||
|
||||
func buildExecuteNowAnchorLine() string {
|
||||
now := time.Now()
|
||||
weekdays := []string{"周日", "周一", "周二", "周三", "周四", "周五", "周六"}
|
||||
return fmt.Sprintf("%s（%s，%s）", now.Format("2006-01-02 15:04:05 -07:00"), weekdays[int(now.Weekday())], now.Format("MST"))
|
||||
}
|
||||
|
||||
func buildExecuteSafetyHardPack() executeRulePack {
|
||||
return executeRulePack{
|
||||
Name: executeRulePackSafetyHard,
|
||||
Content: strings.TrimSpace(`
|
||||
- 严禁伪造工具结果；若新结果与既有事实冲突，先重查一次再决定。
|
||||
- 工具参数必须严格使用 schema 字段名，禁止自造别名。
|
||||
- JSON 只保留当前 action 必需字段；不要输出空字符串、空对象、空数组或 null 占位。
|
||||
- P1 阶段禁止调用 min_context_switch。
|
||||
- 连续两轮同类读查询后，必须转执行 / ask_user / 明确说明阻塞，不能无限空转。`),
|
||||
}
|
||||
}
|
||||
|
||||
func buildExecuteContextProtocolPack() executeRulePack {
|
||||
return executeRulePack{
|
||||
Name: executeRulePackContextProtocol,
|
||||
Content: strings.TrimSpace(`
|
||||
- msg0 动态区初始仅保留 context_tools_add / context_tools_remove。
|
||||
- 需要业务工具前先 context_tools_add：排程用 domain="schedule"，任务类写入用 domain="taskclass"。
|
||||
- schedule 可选 packs=["mutation","analyze","detail_read","deep_analyze","queue","web"]；core 固定注入，不要显式传 core。
|
||||
- 只在业务方向切换时再 remove；done 后的动态区清理由系统自动完成，不必手动 remove。
|
||||
- 如果目标工具当前不在可用列表，先 add 对应 domain / packs，再继续执行。`),
|
||||
}
|
||||
}
|
||||
|
||||
func buildExecuteModePlanPack() executeRulePack {
|
||||
return executeRulePack{
|
||||
Name: executeRulePackModePlan,
|
||||
Content: strings.TrimSpace(`
|
||||
- 当前为计划执行模式：必须围绕当前计划步骤推进。
|
||||
- 未满足 done_when 时，只能 continue / confirm / ask_user，禁止 next_plan。
|
||||
- next_plan / done 时，goal_check 必须是字符串，并对照 done_when 给出完成证据。
|
||||
- 禁止跳步执行后续计划。`),
|
||||
}
|
||||
}
|
||||
|
||||
func buildExecuteModeReActPack() executeRulePack {
|
||||
return executeRulePack{
|
||||
Name: executeRulePackModeReAct,
|
||||
Content: strings.TrimSpace(`
|
||||
- 当前为自由执行（ReAct）模式：可自主决定 continue / confirm / ask_user / done / abort。
|
||||
- 如果关键事实无法通过工具补齐，优先 ask_user，不做猜测落库。
|
||||
- 自主推进时要小步快跑，优先闭合当前局部问题，不要发散成大范围开放搜索。`),
|
||||
}
|
||||
}
|
||||
|
||||
func buildExecuteSchedulePack() executeRulePack {
|
||||
return executeRulePack{
|
||||
Name: executeRulePackDomainSchedule,
|
||||
Content: strings.TrimSpace(`
|
||||
- 当前业务域为 schedule：只处理当前目标任务类，不重排无关内容。
|
||||
- existing 只作事实参考；真正可调对象优先看 suggested。
|
||||
- 同任务类内部顺序必须保持，任何越过前驱/后继边界的移动都会被写工具拒绝。`),
|
||||
}
|
||||
}
|
||||
|
||||
func buildExecuteScheduleMutationPack() executeRulePack {
|
||||
return executeRulePack{
|
||||
Name: executeRulePackScheduleMutation,
|
||||
Content: strings.TrimSpace(`
|
||||
- mutation 包负责真正落日程写操作：place / move / swap / batch_move / unplace。
|
||||
- 写操作必须走 action=confirm；不要在 continue 里偷跑写工具。
|
||||
- 若是主动优化链路，优先在后端给出的合法候选中选择，不要自己再全窗搜索新坑位。`),
|
||||
}
|
||||
}
|
||||
|
||||
func buildExecuteQueueMicroPack() executeRulePack {
|
||||
return executeRulePack{
|
||||
Name: executeRulePackMicroQueue,
|
||||
Content: strings.TrimSpace(`
|
||||
- queue 包适合“按同一条件逐个处理一批任务”的场景，例如把所有早八任务依次挪走。
|
||||
- query_target_tasks 可结合 enqueue=true 先把候选任务入队，再用 queue_pop_head / queue_apply_head_move / queue_skip_head 顺序处理。
|
||||
- 当你需要连续处理多条相似任务时，优先走 queue，避免把整批任务细节长期堆在上下文里。`),
|
||||
}
|
||||
}
|
||||
|
||||
func buildExecuteScheduleWebPack() executeRulePack {
|
||||
return executeRulePack{
|
||||
Name: executeRulePackScheduleWeb,
|
||||
Content: strings.TrimSpace(`
|
||||
- web 包只用于补充通用学习资料或通识信息，不用于捏造个人时间、考试时间、DDL 或排程事实。
|
||||
- web_search 先粗搜，web_fetch 再抓正文；不确定时宁可不用，也不要把网页结果当成排程事实直接写入。`),
|
||||
}
|
||||
}
|
||||
|
||||
func buildExecuteTaskClassPack() executeRulePack {
|
||||
return executeRulePack{
|
||||
Name: executeRulePackDomainTaskClass,
|
||||
Content: strings.TrimSpace(`
|
||||
- taskclass 域只负责生成或修正任务类，不代表已经开始排程。
|
||||
- upsert_task_class 若返回 validation.ok=false，必须先处理 validation.issues，再考虑重试或 ask_user。
|
||||
- subject_type / difficulty_level / cognitive_intensity 是任务类语义画像必填项；优先静默推断，只有确实无法判断时再 ask_user。
|
||||
- excluded_slots 取值应与系统节次定义一致；excluded_days_of_week 使用 1~7 表示周一到周日。`),
|
||||
}
|
||||
}
|
||||
|
||||
func buildExecuteRoughDoneMicroPack() executeRulePack {
|
||||
return executeRulePack{
|
||||
Name: executeRulePackMicroRoughDone,
|
||||
Content: strings.TrimSpace(`
|
||||
- 已有 rough_build_done：本轮以微调为主，不要把任务重新当成“未排入”再全量 place。
|
||||
- 若当前问题已经可接受，应优先收口，不要为了追求完美继续反复局部打磨。`),
|
||||
}
|
||||
}
|
||||
|
||||
func buildExecuteTaskClassRetryMicroPack() executeRulePack {
|
||||
return executeRulePack{
|
||||
Name: executeRulePackMicroTaskRetry,
|
||||
Content: strings.TrimSpace(`
|
||||
- 最近一次 upsert_task_class 失败时，优先围绕 validation.issues 修补。
|
||||
- 问题未解决前，不要用 done 假装收口；要么重试，要么 ask_user 补关键信息。`),
|
||||
}
|
||||
}
|
||||
|
||||
func shouldInjectExecuteDiagLoopPack(state *newagentmodel.CommonState, ctx *newagentmodel.ConversationContext) bool {
|
||||
if state == nil || !hasExecuteRoughBuildDone(ctx) {
|
||||
return false
|
||||
}
|
||||
if normalizeExecuteToolDomain(readExecuteActiveToolDomain(state)) != "schedule" {
|
||||
return false
|
||||
}
|
||||
activePacks := readExecuteActiveToolPacks(state)
|
||||
return hasExecutePack(activePacks, newagenttools.ToolPackAnalyze) &&
|
||||
hasExecutePack(activePacks, newagenttools.ToolPackMutation)
|
||||
}
|
||||
Reference in New Issue
Block a user