Version: 0.9.75.dev.260505

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

backend/services/agent/tools/schedule/analyze_health_decision_v2.go

@@ -0,0 +1,124 @@
+package schedule
+
+import "strings"
+
+// buildAnalyzeHealthDecisionV2 生成 analyze_health 在主动优化场景下的最终裁决。
+//
+// 职责边界：
+// 1. 先尊重 base 层的判断：只有 base 明确允许继续优化时，才进入候选枚举。
+// 2. 候选只来自后端已经验证合法、并且复诊后确实变好的 move/swap 方案。
+// 3. 若没有真正改善的候选，则明确返回 close，避免把 LLM 推回开放式全窗搜索。
+func buildAnalyzeHealthDecisionV2(
+	state *ScheduleState,
+	snapshot analyzeHealthSnapshot,
+) analyzeHealthDecision {
+	base := buildAnalyzeHealthDecisionBase(state, snapshot)
+	decision := analyzeHealthDecision{
+		ShouldContinueOptimize: base.ShouldContinueOptimize,
+		PrimaryProblem:         base.PrimaryProblem,
+		ProblemScope:           base.ProblemScope,
+		IsForcedImperfection:   base.IsForcedImperfection,
+		RecommendedOperation:   base.RecommendedOperation,
+		ImprovementSignal: buildHealthImprovementSignal(
+			snapshot.Rhythm,
+			snapshot.Tightness,
+			base.ProblemScope,
+			base.RecommendedOperation,
+			snapshot.Profile,
+			snapshot.Feasibility,
+		),
+	}
+
+	if !shouldEnterHealthCandidateLoop(base) {
+		decision.Candidates = []analyzeHealthCandidate{
+			buildHealthCloseCandidate("保持当前安排并收口：当前不需要再进入主动优化候选。", snapshot, base),
+		}
+		decision.ShouldContinueOptimize = false
+		return decision
+	}
+
+	bestScan, ok := findBestHealthProblemScanResult(state, snapshot)
+	if !ok || bestScan.Problem.Kind != healthProblemHeavyAdjacent || bestScan.Problem.Pair == nil {
+		decision.Candidates = []analyzeHealthCandidate{
+			buildHealthCloseCandidate("保持当前安排并收口：当前没有值得继续处理的局部认知问题。", snapshot, base),
+		}
+		decision.ShouldContinueOptimize = false
+		decision.PrimaryProblem = "当前没有发现值得继续处理的局部认知问题"
+		decision.ProblemScope = nil
+		decision.RecommendedOperation = "close"
+		if snapshot.Tightness.TightnessLevel == "locked" || snapshot.Tightness.TightnessLevel == "tight" {
+			decision.IsForcedImperfection = true
+		}
+		decision.ImprovementSignal = buildHealthImprovementSignal(
+			snapshot.Rhythm,
+			snapshot.Tightness,
+			decision.ProblemScope,
+			decision.RecommendedOperation,
+			snapshot.Profile,
+			snapshot.Feasibility,
+		)
+		return decision
+	}
+
+	decision.PrimaryProblem = bestScan.Problem.Summary
+	decision.ProblemScope = bestScan.Problem.Scope
+	decision.Candidates = append(decision.Candidates, bestScan.Candidates...)
+	decision.Candidates = append(decision.Candidates,
+		buildHealthCloseCandidate("如果不想继续挪动，也可以保持当前安排并直接收口。", snapshot, base),
+	)
+	decision.ShouldContinueOptimize = true
+	decision.RecommendedOperation = strings.TrimSpace(bestScan.Candidates[0].Tool)
+	decision.ImprovementSignal = buildHealthImprovementSignal(
+		snapshot.Rhythm,
+		snapshot.Tightness,
+		decision.ProblemScope,
+		decision.RecommendedOperation,
+		snapshot.Profile,
+		snapshot.Feasibility,
+	)
+	return decision
+}
+
+// findBestHealthProblemScanResult 每轮重扫所有 heavy_adjacent 天，并选出当前收益最高的一天。
+//
+// 步骤化说明：
+// 1. 先收集所有仍需关注的 heavy_adjacent 天；这里只扫描问题天，不改候选类型。
+// 2. 再对每一天复用现有单天候选试算逻辑，保持“合法且复诊后确实变好”这一过滤语义不变。
+// 3. 最后只返回收益最高且达到最小阈值的一天；最终 decision.candidates 仍只来自这一天天然候选集。
+func findBestHealthProblemScanResult(
+	state *ScheduleState,
+	snapshot analyzeHealthSnapshot,
+) (analyzeHealthProblemScanResult, bool) {
+	problems := collectRepairableHeavyAdjacentProblems(state, snapshot)
+	if len(problems) == 0 {
+		return analyzeHealthProblemScanResult{}, false
+	}
+
+	results := make([]analyzeHealthProblemScanResult, 0, len(problems))
+	for _, problem := range problems {
+		scan, ok := buildHealthProblemScanResult(state, snapshot, problem)
+		if !ok {
+			continue
+		}
+		results = append(results, scan)
+	}
+	return selectBestHealthProblemScanResult(results)
+}
+
+// shouldEnterHealthCandidateLoop 判断本轮是否应进入“候选式主动优化”。
+//
+// 说明：
+// 1. 只有 base 已判定“值得继续优化”时才放行。
+// 2. 当前主动优化闭环只接受 move / swap 两类操作，其它动作不进入候选生成。
+// 3. 这样可以挡住 “ask_user / close / forced imperfection” 被后续枚举误覆盖的问题。
+func shouldEnterHealthCandidateLoop(base analyzeHealthDecisionBase) bool {
+	if !base.ShouldContinueOptimize {
+		return false
+	}
+	switch strings.TrimSpace(base.RecommendedOperation) {
+	case "move", "swap":
+		return true
+	default:
+		return false
+	}
+}