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
105 lines
5.0 KiB
Go
105 lines
5.0 KiB
Go
package agentprompt
|
||
|
||
import (
|
||
"fmt"
|
||
"strings"
|
||
|
||
agentmodel "github.com/LoveLosita/smartflow/backend/services/agent/model"
|
||
agenttools "github.com/LoveLosita/smartflow/backend/services/agent/tools"
|
||
)
|
||
|
||
// renderExecuteNextStepHintV2 生成 execute.msg3 的轻量方向提示。
|
||
//
|
||
// 设计目标：
|
||
// 1. 主动优化模式下，只强调“先 analyze_health，再从 candidates 里选”，不再散发额外搜索暗示。
|
||
// 2. 普通链路仍保留必要的业务引导，避免误伤用户明确提出的普通调整请求。
|
||
// 3. 提示只给方向，不替模型代填最终写参数。
|
||
func renderExecuteNextStepHintV2(
|
||
state *agentmodel.CommonState,
|
||
latestAnalyze string,
|
||
latestMutation string,
|
||
roughBuildDone bool,
|
||
) string {
|
||
if state == nil {
|
||
return ""
|
||
}
|
||
|
||
activeDomain := strings.TrimSpace(state.ActiveToolDomain)
|
||
activePacks := agenttools.ResolveEffectiveToolPacks(state.ActiveToolDomain, state.ActiveToolPacks)
|
||
|
||
if state.ActiveOptimizeOnly {
|
||
switch {
|
||
case activeDomain == "" && roughBuildDone:
|
||
return "当前是粗排后主动优化专用模式；先激活 schedule，并只围绕 analyze_health -> move/swap 候选闭环推进。"
|
||
case !state.HealthCheckDone:
|
||
return "当前是粗排后主动优化专用模式；先调 analyze_health，等待后端给出 candidates，再做选择。"
|
||
case !state.HealthIsFeasible || strings.EqualFold(strings.TrimSpace(state.HealthRecommendedOperation), "ask_user"):
|
||
return "analyze_health 已判定当前更像时间窗或信息约束问题；不要继续挪动，先把冲突或缺失点明确告诉用户。"
|
||
case !state.HealthShouldContinueOptimize:
|
||
return "analyze_health 已判定当前无需继续主动优化；若用户没有新增要求，直接收口。"
|
||
default:
|
||
return "当前是粗排后主动优化专用模式；直接从 analyze_health 的 decision.candidates 里选一个合法 move/swap 执行，不要再自己搜索读工具。"
|
||
}
|
||
}
|
||
|
||
if activeDomain == "schedule" && state.HealthCheckDone {
|
||
switch {
|
||
case !state.HealthShouldContinueOptimize && state.HealthIsForcedImperfection:
|
||
return fmt.Sprintf(
|
||
"analyze_health 已判定当前更像约束代价：tightness=%s，主问题=%s。优先考虑收口。",
|
||
fallbackExecuteText(state.HealthTightnessLevel, "unknown"),
|
||
fallbackExecuteText(state.HealthPrimaryProblem, "无"),
|
||
)
|
||
case !state.HealthShouldContinueOptimize:
|
||
return fmt.Sprintf(
|
||
"analyze_health 已判定当前没有更值得继续处理的局部问题：%s。若用户未追加新要求，优先收口。",
|
||
fallbackExecuteText(state.HealthPrimaryProblem, "当前可直接收口"),
|
||
)
|
||
case state.HealthStagnationCount > 0:
|
||
return fmt.Sprintf(
|
||
"最近诊断已连续 %d 次无明显改善；若本轮仍不能让主问题变轻，优先收口。当前主问题：%s。",
|
||
state.HealthStagnationCount,
|
||
fallbackExecuteText(state.HealthPrimaryProblem, "无"),
|
||
)
|
||
case strings.EqualFold(strings.TrimSpace(state.HealthRecommendedOperation), "swap"):
|
||
return fmt.Sprintf(
|
||
"当前主问题：%s。优先在已有落位之间做局部 swap，别把问题扩散到更远的天数。",
|
||
fallbackExecuteText(state.HealthPrimaryProblem, "无"),
|
||
)
|
||
case strings.EqualFold(strings.TrimSpace(state.HealthRecommendedOperation), "move"):
|
||
return fmt.Sprintf(
|
||
"当前主问题：%s。若要 move，只在近范围合法落点里小修，不要做全窗口搜索。",
|
||
fallbackExecuteText(state.HealthPrimaryProblem, "无"),
|
||
)
|
||
}
|
||
}
|
||
|
||
if activeDomain == "" {
|
||
if roughBuildDone {
|
||
return `先激活 schedule 业务域；当前是粗排后的微调场景，通常至少需要 mutation+analyze。若要按统一条件逐个处理一批任务，再加 packs=["queue"]。`
|
||
}
|
||
return `先判断当前任务属于哪个业务域，再用 context_tools_add 激活对应工具。若用户只是在描述学习目标、总节数、难度、节次偏好、禁排时段、排除星期、内容拆分授权，默认先走 taskclass；只有用户明确要求“排进日程 / 给出具体时间安排 / 现在就排一版”时，才切 schedule。`
|
||
}
|
||
|
||
if activeDomain == "schedule" &&
|
||
strings.Contains(latestMutation, "batch_move") &&
|
||
(strings.Contains(latestMutation, "缺少") || strings.Contains(latestMutation, "无效")) {
|
||
return `当前 batch_move 路径受参数约束；若要处理一批符合同一条件的任务，优先加 packs=["queue"] 逐个处理。`
|
||
}
|
||
|
||
if activeDomain == "schedule" &&
|
||
latestAnalyze != "" &&
|
||
strings.Contains(latestAnalyze, "metrics") &&
|
||
!containsExecutePack(activePacks, agenttools.ToolPackQueue) {
|
||
return `若诊断已经完成，下一步应转入读事实或写操作，不要重复 analyze_health；涉及同类批量任务时优先考虑 packs=["queue"]。`
|
||
}
|
||
|
||
if activeDomain == "taskclass" &&
|
||
state.TaskClassUpsertLastTried &&
|
||
!state.TaskClassUpsertLastSuccess {
|
||
return `先判断 validation.issues 是“用户缺信息”还是“内部表示修正”；能从上下文补的先静默补齐，再用 confirm 重试 upsert_task_class，不要继续解释底层约束，更不要直接收口。`
|
||
}
|
||
|
||
return ""
|
||
}
|