Fix cross-action slot inheritance and fail-closed cancel errors

Why: review blockers B1/M2 - Suspend inherited the previous action's
accumulated slots into a new action, and applyCancel swallowed store
errors so an abandoned fill could become an unconfirmed real cancel.
Evidence: review-a1-run.json findings fixed; intentsession/server
tests pass with the reviewer's sequences (verify-a1fix.json).
Impact: different-action suspend starts fresh; non-stale cancel store
errors fail closed; all other flows unchanged.
Refs: feature/intent-multi-turn-a1

internal/intentsession/intentsession.go

@@ -185,12 +185,17 @@ func (s *Service) Pending(key Key) (*Session, error) {
 
 // Suspend 挂起一个可执行动作的预期并累计已抽到的槽位（续期）。已抽到的
 // 槽位不再计入仍缺清单（挂起即如实反映还差什么）。
+//
+// 累计只在同一个动作内成立：新预期指向别的动作（意图 id 或动作引用变了）
+// 时，上一轮已抽到的槽位一律丢弃——新任务从零开始，revision 照常推进，
+// 陈旧回答因此作废。否则上一轮抽到的运行号会跟着新动作一起派发：旧任务的
+// 槽位漏进新任务，正是本包纪律里明令禁止的那一类。
 func (s *Service) Suspend(key Key, expectation Expectation, slots map[string]string) (*Session, error) {
 	cur, err := s.load(key)
 	if err != nil {
 		return nil, err
 	}
-	if cur != nil {
+	if cur != nil && cur.Expectation != nil && sameExpectation(*cur.Expectation, expectation) {
 		slots = mergeSlots(cur.Slots, slots)
 	}
 	var missing []intentgate.SlotQuestion
@@ -408,6 +413,13 @@ func (s *Service) accepted(cur *Session, answer Answer) map[string]string {
 	return out
 }
 
+// sameExpectation 报告新旧预期是否同一个动作（意图 id 与动作引用都一致）：
+// 只有同一个动作再次挂起才继续累计已抽到的槽位。预期是「已取消」记录时
+// 动作种类为 cancel，与任何待补齐的动作都不相同——取消记录绝不被借走。
+func sameExpectation(a, b Expectation) bool {
+	return a.IntentID == b.IntentID && a.Action.Kind == b.Action.Kind && a.Action.ID == b.Action.ID
+}
+
 // missingOf 投影到存储形态（缺失槽位表）。
 func missingOf(list []intentgate.SlotQuestion) []store.IntentSlotQuestion {
 	out := make([]store.IntentSlotQuestion, 0, len(list))

internal/intentsession/intentsession_test.go

@@ -108,6 +108,88 @@ func TestSuspendThenAnswerCompletes(t *testing.T) {
 	}
 }
 
+// 换动作即换任务（跨动作槽位继承的回归序列）：A 抽到运行号后挂起 → 用户
+// 改口做另一个动作 B（B 缺别的槽位）→ 用户补齐 B → B 齐全时只带 B 自己
+// 的槽位。A 的运行号绝不跟着 B 一起派发。
+func TestSuspendDiscardsSlotsOfAnotherAction(t *testing.T) {
+	svc := newService(t)
+	key := cancelKey()
+	first, err := svc.Suspend(key, cancelExpectation(), map[string]string{"run_id": "R-7"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	report := Expectation{IntentID: "demo.report",
+		Action:   intentgate.ActionRef{Kind: intentgate.ActionCommand, ID: "demo.report"},
+		Missing:  []intentgate.SlotQuestion{{Key: "scope", Question: "哪个范围？"}},
+		Question: "哪个范围？"}
+	second, err := svc.Suspend(key, report, nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if second.Revision <= first.Revision {
+		t.Fatalf("换动作必须推进 revision（让在途回答作废）：%d vs %d", second.Revision, first.Revision)
+	}
+	if second.Expectation == nil || second.Expectation.IntentID != "demo.report" ||
+		second.Expectation.Action.ID != "demo.report" {
+		t.Fatalf("挂起的应是新动作：%+v", second)
+	}
+	if _, leaked := second.Slots["run_id"]; leaked {
+		t.Fatalf("上一个动作的槽位被继承：%+v", second.Slots)
+	}
+	if keys := questionKeys(second.Expectation.Missing); len(keys) != 1 || keys[0] != "scope" {
+		t.Fatalf("新动作的缺失清单：%v", keys)
+	}
+	// 旧动作的陈旧回答也作废（不得借新动作的壳补齐）。
+	if _, err := svc.ApplyAnswer(key, first.Revision,
+		Answer{Slots: map[string]string{"reason": "误触发"}}); !errors.Is(err, ErrStale) {
+		t.Fatalf("旧 revision 必须被拒：%v", err)
+	}
+	// 用户补齐 B：B 齐全，且槽位只有 B 自己的（没有 A 的 run_id）。
+	done, err := svc.ApplyAnswer(key, second.Revision, Answer{Slots: map[string]string{"scope": "华东"}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !done.Complete || done.Session != nil {
+		t.Fatalf("B 应被补齐并消费会话：%+v", done)
+	}
+	if len(done.Slots) != 1 || done.Slots["scope"] != "华东" {
+		t.Fatalf("派发槽位混进了别的动作的值：%v", done.Slots)
+	}
+}
+
+// 同一动作再次挂起仍累计已抽到的槽位（缺失清单变窄不算换动作）。
+func TestSuspendOfSameActionKeepsAccumulatedSlots(t *testing.T) {
+	svc := newService(t)
+	key := cancelKey()
+	first, err := svc.Suspend(key, cancelExpectation(), nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	// 同一意图与动作引用，只是路由面收敛了缺失清单（仍缺原因）。
+	narrowed := cancelExpectation()
+	narrowed.Missing = []intentgate.SlotQuestion{{Key: "reason", Question: "取消原因？"}}
+	again, err := svc.Suspend(key, narrowed, map[string]string{"run_id": "R-9"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if again.Revision <= first.Revision {
+		t.Fatalf("再次挂起必须推进 revision：%d vs %d", again.Revision, first.Revision)
+	}
+	if again.Slots["run_id"] != "R-9" {
+		t.Fatalf("同一动作的累计槽位被丢弃：%+v", again.Slots)
+	}
+	if keys := questionKeys(again.Expectation.Missing); len(keys) != 1 || keys[0] != "reason" {
+		t.Fatalf("同一动作的缺失清单：%v", keys)
+	}
+	done, err := svc.ApplyAnswer(key, again.Revision, Answer{Slots: map[string]string{"reason": "误触发"}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !done.Complete || done.Slots["run_id"] != "R-9" || done.Slots["reason"] != "误触发" {
+		t.Fatalf("同一动作补齐：%+v", done)
+	}
+}
+
 // 两轮部分填充：set_slot 落一个槽位，再由回答补另一个。
 func TestSetSlotThenAnswerCompletes(t *testing.T) {
 	svc := newService(t)

internal/rundiag/rundiag.go

@@ -1,488 +0,0 @@
-// Package rundiag 把单次 run 的诊断证据聚合成一份只读结论：卡在哪、谁负责、
-// 为什么在等、最后一个稳定检查点、已知副作用、下一步动作。
-//
-// 全包统一口径（任务卡 A2 的契约）：
-//   - 只读派生：Build 是纯函数，Gather 只跑 SELECT；本包不提供任何写路径。
-//   - 证据优先于推断：状态只由已落库的记录驱动。读不到就降级为 unknown + note，
-//     读不到隔离记录时更按 fail-closed 处理（不声称"可以安全重试"）。
-//   - 「没有记录」≠「没有发生」：side_effects 为空必须带 note，说明这是"未记录"
-//     而不是"未发生"。
-//   - 归属边界：HITL 决策卡、危险命令确认卡、命令审计信封都按 owner + 会话归属，
-//     其中决策卡与确认卡表里根本没有 run 列，审计信封虽有 runId 字段但宿主总线
-//     当前不写。因此"归到这个 run"是调用方按签发时间窗做的有界推断，报告以
-//     note 标出，绝不把会话里更早的记录算成本次 run 的副作用。
-package rundiag
-
-import (
-	"encoding/json"
-	"errors"
-	"fmt"
-	"time"
-
-	agentcommand "agent-command"
-	"agent-system-backend/internal/cmdbus"
-	"agent-system-backend/internal/store"
-	core "agent-system-core"
-)
-
-// responsibility.state 的取值域（任务卡固定）。
-const (
-	StateRunning             = "running"
-	StateWaitingHITL         = "waiting_hitl"
-	StateConfirmationPending = "confirmation_pending"
-	StateRecoveryHold        = "recovery_hold"
-	StateReconciling         = "reconciling"
-	StateFailed              = "failed"
-	StateCompleted           = "completed"
-	StateUnknown             = "unknown"
-)
-
-// responsibility.role：这一刻该谁动。
-const (
-	RoleRunner         = "runner"          // 执行器自己在推进
-	RoleHuman          = "human"           // 等人裁决（HITL / 危险命令确认）
-	RoleOperationOwner = "operation-owner" // 等副作用责任人对账与解除
-	RoleNone           = "none"            // 没有待办责任方
-)
-
-// next_step.action 的取值域。
-const (
-	ActionConfirmCard               = "confirm_card"
-	ActionReleaseRecoveryHold       = "release_recovery_hold"
-	ActionSafeRetry                 = "safe_retry"
-	ActionNeedsManualReconciliation = "needs_manual_reconciliation"
-	ActionNone                      = "none"
-)
-
-// side_effects[].kind：证据来源类别。
-const (
-	EffectOperation    = "operation"     // 命令审计信封（U7b 对账记录）
-	EffectRecoveryHold = "recovery_hold" // 恢复隔离行（U3b 的"副作用未知"证据）
-)
-
-// 读取上界：诊断是给人和值班看的，不是全量导出。
-const (
-	maxSideEffects = 50
-	maxAuditScan   = 200 // 审计信封按会话+时间窗有界扫描
-	maxConfirms    = 50
-)
-
-// Report 是诊断响应（JSON 契约，字段名用任务卡给定的 snake_case）。
-type Report struct {
-	Run            RunSummary     `json:"run"`
-	Responsibility Responsibility `json:"responsibility"`
-	LastCheckpoint *Checkpoint    `json:"last_checkpoint"` // 无记录时为 null（不编造）
-	SideEffects    []SideEffect   `json:"side_effects"`    // 恒为数组：空数组 = 未记录到
-	NextStep       NextStep       `json:"next_step"`
-	Notes          []string       `json:"notes,omitempty"` // 降级与边界说明
-}
-
-// RunSummary 是 run 自身的身份与时间戳。updated_at 的口径：已终态取 ended_at，
-// 未终态取最后一个已持久化事件的时间；两者都没有则为 null（并在 notes 说明）。
-type RunSummary struct {
-	ID           string  `json:"id"`
-	Status       string  `json:"status"`
-	Kind         string  `json:"kind,omitempty"`
-	Phase        string  `json:"phase,omitempty"`
-	Owner        string  `json:"owner"`
-	Project      string  `json:"project"`
-	Conversation string  `json:"conversation"`
-	CreatedAt    string  `json:"created_at"`
-	UpdatedAt    *string `json:"updated_at"`
-}
-
-// Responsibility 回答"卡在哪、谁负责、为什么等、从什么时候开始等"。
-type Responsibility struct {
-	State      string `json:"state"`
-	Role       string `json:"role"`
-	WaitReason string `json:"wait_reason"`
-	Since      string `json:"since,omitempty"` // 进入该状态的时间；存储没记就省略
-}
-
-// Checkpoint 指向最后一个稳定落盘点（事件流上的 seq 可直接作重放游标）。
-type Checkpoint struct {
-	ID   string `json:"id"`
-	At   string `json:"at"`
-	Type string `json:"type"`
-}
-
-// SideEffect 是一条已记录的外部副作用证据。operation 行为命令回执对账记录；
-// recovery_hold 行是 U3b 落下的"副作用结果未知"隔离证据（state = held/released）。
-type SideEffect struct {
-	Kind        string `json:"kind"`
-	OperationID string `json:"operation_id,omitempty"`
-	CommandID   string `json:"command_id,omitempty"`
-	State       string `json:"state"`
-	At          string `json:"at,omitempty"`
-	ReleasedAt  string `json:"released_at,omitempty"`
-}
-
-// NextStep 是唯一建议的下一步动作（诊断面只建议，不执行）。
-type NextStep struct {
-	Action string `json:"action"`
-	CardID string `json:"card_id,omitempty"` // confirm_card 可解析时给出卡 ID
-	Reason string `json:"reason"`
-}
-
-// Source 是诊断所需的只读端口（*store.Store 直接满足；测试可替身）。
-type Source interface {
-	GetRecoveryHold(runID string) (store.RecoveryHold, error)
-	ActiveHITLCardsByRun(runID string) ([]core.HITLCard, error)
-	PendingHumanDecisionsByConversation(owner, conversationID string, now time.Time) ([]store.HumanDecision, error)
-	ListOwnerPendingConfirms(owner string, limit int) ([]store.PendingConfirm, error)
-	LastRunCheckpoint(conversationID, runID string) (store.RunCheckpoint, bool, error)
-	ListAudits(owner, conversationID string, limit int) ([]agentcommand.AuditEnvelope, error)
-}
-
-// Evidence 是 Build 的全部输入（合成数据即可表驱动测试，无需真实存储）。
-type Evidence struct {
-	Run core.Run
-
-	Hold        *store.RecoveryHold // nil = 没有隔离行；HoldUnknown = 读失败
-	HoldUnknown bool
-	Cards       []core.HITLCard        // 该 run 未关闭的 HITL 卡
-	Decisions   []store.HumanDecision  // 会话内未过期的 pending 决策卡
-	Confirms    []store.PendingConfirm // owner 名下未确认的危险命令挂起行
-	Checkpoint  *store.RunCheckpoint   // 最后一个已持久化事件
-	Audits      []agentcommand.AuditEnvelope
-	Notes       []string
-}
-
-// Gather 只读地把一个 run 的证据读齐。任何一条可选查询失败都不让整个诊断失败：
-// 失败降级为 note；隔离记录读不到按 fail-closed 标记（HoldUnknown），因为"读不到
-// hold"绝不能被读成"没有 hold"。
-func Gather(src Source, run core.Run, now time.Time) Evidence {
-	ev := Evidence{Run: run}
-	if src == nil {
-		ev.HoldUnknown = true
-		return ev.withNote("diagnostics source unavailable; responsibility is fail-closed unknown")
-	}
-	owner := string(run.Owner)
-	conversation := string(run.ConversationID)
-
-	hold, err := src.GetRecoveryHold(string(run.ID))
-	switch {
-	case err == nil:
-		ev.Hold = &hold
-	case errors.Is(err, store.ErrNotFound):
-		// 没有隔离行是正常态，不记 note。
-	default:
-		ev.HoldUnknown = true
-		ev = ev.withNote("recovery hold unreadable: " + err.Error())
-	}
-
-	cards, err := src.ActiveHITLCardsByRun(string(run.ID))
-	if err != nil {
-		ev = ev.withNote("hitl cards unreadable: " + err.Error())
-	} else {
-		ev.Cards = cards
-	}
-
-	decisions, err := src.PendingHumanDecisionsByConversation(owner, conversation, now)
-	if err != nil {
-		ev = ev.withNote("human decisions unreadable: " + err.Error())
-	} else {
-		ev.Decisions = decisions
-	}
-
-	confirms, err := src.ListOwnerPendingConfirms(owner, maxConfirms)
-	if err != nil {
-		ev = ev.withNote("pending confirmations unreadable: " + err.Error())
-	} else {
-		ev.Confirms = confirms
-	}
-
-	cp, ok, err := src.LastRunCheckpoint(conversation, string(run.ID))
-	switch {
-	case err != nil:
-		ev = ev.withNote("last checkpoint unreadable: " + err.Error())
-	case ok:
-		ev.Checkpoint = &cp
-	}
-
-	audits, err := src.ListAudits(owner, conversation, maxAuditScan)
-	if err != nil {
-		ev = ev.withNote("command audit unreadable: " + err.Error())
-	} else {
-		ev.Audits = audits
-	}
-	return ev
-}
-
-// Build 把证据聚合成诊断结论（纯函数，不读写任何状态）。
-func Build(ev Evidence) Report {
-	rep := Report{Run: runSummary(ev), SideEffects: []SideEffect{}}
-	if ev.Checkpoint != nil {
-		rep.LastCheckpoint = &Checkpoint{ID: fmt.Sprintf("%d", ev.Checkpoint.Seq), At: ev.Checkpoint.TS, Type: ev.Checkpoint.Type}
-	}
-	rep.SideEffects = sideEffects(ev)
-	rep.Notes = append([]string(nil), ev.Notes...)
-
-	card := waitingCard(ev)
-	decision := waitingDecision(ev)
-	confirm := waitingConfirm(ev)
-
-	switch {
-	case ev.HoldUnknown:
-		// 隔离记录读不到：不猜没有 hold，直接 unknown，且不给"安全重试"。
-		rep.Notes = append(rep.Notes, "recovery hold state unknown; treat as quarantined until reconciled")
-		rep.Responsibility = Responsibility{
-			State: StateUnknown, Role: RoleNone,
-			WaitReason: "recovery hold state could not be read",
-		}
-		rep.NextStep = NextStep{Action: ActionNone, Reason: "cannot prove the run is free of an open recovery hold"}
-	case openHold(ev):
-		hold := ev.Hold
-		rep.Responsibility = Responsibility{
-			State: StateRecoveryHold, Role: RoleOperationOwner,
-			WaitReason: holdReason(hold), Since: hold.CreatedAt,
-		}
-		rep.NextStep = NextStep{Action: ActionReleaseRecoveryHold, Reason: holdReason(hold)}
-	case needsReconcile(ev):
-		reason := "recovery hold released but the run is still interrupted"
-		since := ""
-		if ev.Hold != nil {
-			since = ev.Hold.ReleasedAt
-			if ev.Hold.Reason != "" {
-				reason += "; hold reason: " + ev.Hold.Reason
-			}
-		}
-		rep.Responsibility = Responsibility{State: StateReconciling, Role: RoleOperationOwner, WaitReason: reason, Since: since}
-		rep.NextStep = NextStep{Action: ActionNeedsManualReconciliation, Reason: reason}
-	case ev.Run.Status == core.RunWaitingHITL || card != nil || decision != nil:
-		reason := "waiting for a human decision"
-		since := ""
-		if decision != nil {
-			reason = decision.Question
-			if reason == "" {
-				reason = "pending human decision on command " + decision.CommandID
-			}
-			since = decision.IssuedAt.UTC().Format(time.RFC3339)
-		} else if card != nil {
-			reason = card.Question
-			if reason == "" {
-				reason = "run is waiting on a HITL card"
-			}
-			since = card.CreatedAt
-		}
-		rep.Responsibility = Responsibility{State: StateWaitingHITL, Role: RoleHuman, WaitReason: reason, Since: since}
-		if card == nil && decision == nil {
-			// 只由 run 状态驱动：没有可解析的卡，也不知道等待从何时开始。
-			rep.Notes = append(rep.Notes, "run is marked waiting_hitl with no open card recorded; the wait entry time is not stored")
-		}
-		rep.NextStep = NextStep{Action: ActionConfirmCard, CardID: hitlCardID(card, decision), Reason: reason}
-	case confirm != nil:
-		reason := "dangerous command " + confirm.command + " awaits explicit confirmation"
-		rep.Responsibility = Responsibility{
-			State: StateConfirmationPending, Role: RoleHuman,
-			WaitReason: reason, Since: confirm.row.CreatedAt,
-		}
-		rep.NextStep = NextStep{Action: ActionConfirmCard, CardID: confirm.row.CardID, Reason: reason}
-	case ev.Run.Status == core.RunFailed:
-		reason := "run failed"
-		if ev.Run.Error != nil {
-			reason = fmt.Sprintf("run failed: %s: %s", ev.Run.Error.Code, ev.Run.Error.Message)
-		}
-		rep.Responsibility = Responsibility{State: StateFailed, Role: RoleRunner, WaitReason: reason, Since: endedAt(ev.Run)}
-		// 重试安全性看已记录的副作用证据：没有 open hold 且没有 unknown 回执时，
-		// 重试不会复制本次运行的后果；否则先对账再重试，不说"安全"。
-		if unknown := unknownOperations(rep.SideEffects); unknown > 0 {
-			reason := fmt.Sprintf("terminal failure, but %d recorded operation(s) have an unknown outcome; reconcile side effects before retrying", unknown)
-			rep.Notes = append(rep.Notes, reason)
-			rep.NextStep = NextStep{Action: ActionNeedsManualReconciliation, Reason: reason}
-		} else {
-			rep.NextStep = NextStep{Action: ActionSafeRetry, Reason: "terminal failure with no open recovery hold and no unknown receipt; retry is safe to attempt"}
-		}
-	case ev.Run.Status == core.RunCompleted:
-		rep.Responsibility = Responsibility{State: StateCompleted, Role: RoleNone, WaitReason: "run completed", Since: endedAt(ev.Run)}
-		rep.NextStep = NextStep{Action: ActionNone, Reason: "run already completed"}
-	case ev.Run.Status == core.RunRunning || ev.Run.Status == core.RunPending:
-		rep.Responsibility = Responsibility{
-			State: StateRunning, Role: RoleRunner,
-			WaitReason: "run is executing", Since: ev.Run.CreatedAt.UTC().Format(time.RFC3339),
-		}
-		rep.NextStep = NextStep{Action: ActionNone, Reason: "runner is executing; no operator action"}
-	default:
-		reason := fmt.Sprintf("run status %q has no diagnostic mapping", ev.Run.Status)
-		rep.Notes = append(rep.Notes, reason)
-		rep.Responsibility = Responsibility{State: StateUnknown, Role: RoleNone, WaitReason: reason, Since: endedAt(ev.Run)}
-		rep.NextStep = NextStep{Action: ActionNone, Reason: "status is unmapped; inspect the run record manually"}
-	}
-
-	if ev.Checkpoint == nil {
-		rep.Notes = append(rep.Notes, "no persisted checkpoint reference for this run (no durable stream event)")
-	}
-	if len(rep.SideEffects) == 0 {
-		rep.Notes = append(rep.Notes, "no recorded side effect evidence; absence of records is not proof of no effect")
-	}
-	return rep
-}
-
-func runSummary(ev Evidence) RunSummary {
-	summary := RunSummary{
-		ID: string(ev.Run.ID), Status: string(ev.Run.Status), Kind: string(ev.Run.Kind),
-		Phase: string(ev.Run.Phase), Owner: string(ev.Run.Owner), Project: string(ev.Run.ProjectID),
-		Conversation: string(ev.Run.ConversationID),
-	}
-	if !ev.Run.CreatedAt.IsZero() {
-		summary.CreatedAt = ev.Run.CreatedAt.UTC().Format(time.RFC3339)
-	}
-	// updated：终态取 ended_at，未终态取最后一次落盘（检查点）的时间。
-	if ended := endedAt(ev.Run); ended != "" {
-		summary.UpdatedAt = &ended
-	} else if ev.Checkpoint != nil && ev.Checkpoint.TS != "" {
-		ts := ev.Checkpoint.TS
-		summary.UpdatedAt = &ts
-	}
-	return summary
-}
-
-// sideEffects 汇总已记录的外部副作用证据：命令回执对账记录（U7b 审计信封）加上
-// 恢复隔离行（U3b 的"结果未知"证据）。
-func sideEffects(ev Evidence) []SideEffect {
-	out := []SideEffect{}
-	if ev.Hold != nil {
-		state := "held"
-		if ev.Hold.ReleasedAt != "" {
-			state = "released"
-		}
-		out = append(out, SideEffect{
-			Kind: EffectRecoveryHold, State: state, At: ev.Hold.CreatedAt, ReleasedAt: ev.Hold.ReleasedAt,
-		})
-	}
-	for _, env := range ev.Audits {
-		if !belongsToRun(ev, env) {
-			continue
-		}
-		out = append(out, SideEffect{
-			Kind: EffectOperation, OperationID: env.OperationID, CommandID: env.CommandID,
-			State: string(env.State), At: env.Time.UTC().Format(time.RFC3339),
-		})
-		if len(out) >= maxSideEffects {
-			break
-		}
-	}
-	return out
-}
-
-// belongsToRun 是审计信封的归属边界：信封自带 runId 时以它为准；否则只在"同一
-// owner + 会话且签发时间不早于 run 创建时间"的窗口内归属——会话里更早的记录属于
-// 更早的 run，不算进本次副作用。
-func belongsToRun(ev Evidence, env agentcommand.AuditEnvelope) bool {
-	if env.RunID != "" {
-		return env.RunID == string(ev.Run.ID)
-	}
-	if string(env.ConversationID) != "" && string(env.ConversationID) != string(ev.Run.ConversationID) {
-		return false
-	}
-	if !ev.Run.CreatedAt.IsZero() && !env.Time.IsZero() && env.Time.Before(ev.Run.CreatedAt) {
-		return false
-	}
-	return true
-}
-
-func openHold(ev Evidence) bool {
-	return ev.Hold != nil && ev.Hold.ReleasedAt == ""
-}
-
-// unknownOperations 统计结果未知的已记录 operation（四态语义里的 unknown：
-// 执行过但结果不确定）。它是"重试会不会复制副作用"的直接证据。
-func unknownOperations(effects []SideEffect) int {
-	n := 0
-	for _, e := range effects {
-		if e.Kind == EffectOperation && e.State == string(agentcommand.StateUnknown) {
-			n++
-		}
-	}
-	return n
-}
-
-// needsReconcile 是 U3b 隔离解除后的残留态：隔离已解除，但 run 仍停在中断态——
-// 没有人把它推进到终态，需要人工对账。
-func needsReconcile(ev Evidence) bool {
-	return ev.Hold != nil && ev.Hold.ReleasedAt != "" && ev.Run.Status == core.RunInterrupted
-}
-
-func holdReason(hold *store.RecoveryHold) string {
-	if hold == nil || hold.Reason == "" {
-		return "external side effect outcome unknown (recovery hold)"
-	}
-	return hold.Reason
-}
-
-func endedAt(run core.Run) string {
-	if run.EndedAt == nil {
-		return ""
-	}
-	return run.EndedAt.UTC().Format(time.RFC3339)
-}
-
-// waitingCard 取该 run 未关闭的 HITL 卡（最早一张；只看未关闭的）。
-func waitingCard(ev Evidence) *core.HITLCard {
-	for i := range ev.Cards {
-		if ev.Cards[i].Resolved || ev.Cards[i].Blocked {
-			continue
-		}
-		return &ev.Cards[i]
-	}
-	return nil
-}
-
-// waitingDecision 取本 run 生命周期内签发、仍未过期的人工决策卡。
-func waitingDecision(ev Evidence) *store.HumanDecision {
-	for i := range ev.Decisions {
-		rec := &ev.Decisions[i]
-		if rec.Status != "pending" {
-			continue
-		}
-		if !rec.IssuedAt.IsZero() && !ev.Run.CreatedAt.IsZero() && rec.IssuedAt.Before(ev.Run.CreatedAt) {
-			continue // 签发在本次 run 之前：属于更早的等待，不算本次
-		}
-		return rec
-	}
-	return nil
-}
-
-type scopedConfirm struct {
-	row     store.PendingConfirm
-	command string
-}
-
-// waitingConfirm 取本 run 生命周期内、同一会话挂起的危险命令确认行。载荷是总线
-// 落下的分发请求；解析不出会话的旧行按"无记录"处理，不回显原始载荷。
-func waitingConfirm(ev Evidence) *scopedConfirm {
-	for _, row := range ev.Confirms {
-		req := cmdbus.Request{}
-		if err := json.Unmarshal([]byte(row.Payload), &req); err != nil {
-			continue
-		}
-		if req.ConversationID != "" && req.ConversationID != string(ev.Run.ConversationID) {
-			continue
-		}
-		if req.Owner != "" && string(ev.Run.Owner) != "" && req.Owner != string(ev.Run.Owner) {
-			continue
-		}
-		created, err := time.Parse(time.RFC3339, row.CreatedAt)
-		if err == nil && !ev.Run.CreatedAt.IsZero() && created.Before(ev.Run.CreatedAt) {
-			continue // 挂起早于本次 run：属于更早的等待
-		}
-		return &scopedConfirm{row: row, command: req.ID}
-	}
-	return nil
-}
-
-// hitlCardID 给出可解析的卡标识：HITL 卡优先，其次决策卡 ID。
-func hitlCardID(card *core.HITLCard, decision *store.HumanDecision) string {
-	if card != nil {
-		return string(card.ID)
-	}
-	if decision != nil {
-		return decision.ID
-	}
-	return ""
-}
-
-func (ev Evidence) withNote(note string) Evidence {
-	ev.Notes = append(ev.Notes, note)
-	return ev
-}

internal/rundiag/rundiag_test.go

@@ -1,440 +0,0 @@
-package rundiag
-
-import (
-	"encoding/json"
-	"errors"
-	"strings"
-	"testing"
-	"time"
-
-	agentcommand "agent-command"
-	"agent-system-backend/internal/cmdbus"
-	"agent-system-backend/internal/store"
-	core "agent-system-core"
-)
-
-var (
-	baseTime = time.Date(2026, 10, 20, 10, 0, 0, 0, time.UTC)
-	afterRun = baseTime.Add(5 * time.Minute)
-)
-
-// fakeSource 是只读端口的合成替身：每个字段都能单独触发"没有记录"或"读取失败"。
-type fakeSource struct {
-	hold      store.RecoveryHold
-	holdErr   error
-	cards     []core.HITLCard
-	decisions []store.HumanDecision
-	confirms  []store.PendingConfirm
-	cp        store.RunCheckpoint
-	cpOK      bool
-	audits    []agentcommand.AuditEnvelope
-}
-
-func (f fakeSource) GetRecoveryHold(string) (store.RecoveryHold, error) {
-	if f.holdErr != nil {
-		return store.RecoveryHold{}, f.holdErr
-	}
-	if f.hold.RunID == "" {
-		return store.RecoveryHold{}, store.ErrNotFound
-	}
-	return f.hold, nil
-}
-
-func (f fakeSource) ActiveHITLCardsByRun(string) ([]core.HITLCard, error) { return f.cards, nil }
-
-func (f fakeSource) PendingHumanDecisionsByConversation(string, string, time.Time) ([]store.HumanDecision, error) {
-	return f.decisions, nil
-}
-
-func (f fakeSource) ListOwnerPendingConfirms(string, int) ([]store.PendingConfirm, error) {
-	return f.confirms, nil
-}
-
-func (f fakeSource) LastRunCheckpoint(string, string) (store.RunCheckpoint, bool, error) {
-	return f.cp, f.cpOK, nil
-}
-
-func (f fakeSource) ListAudits(string, string, int) ([]agentcommand.AuditEnvelope, error) {
-	return f.audits, nil
-}
-
-func run(status core.RunStatus) core.Run {
-	return core.Run{
-		ID: "run_1", ConversationID: "cnv_1", ProjectID: "prj_1", Owner: "alice",
-		Kind: core.RunConversation, Status: status, Phase: core.PhaseResponding, CreatedAt: baseTime,
-	}
-}
-
-func ended(r core.Run, at time.Time) core.Run {
-	r.EndedAt = &at
-	return r
-}
-
-func checkpoint() *store.RunCheckpoint {
-	return &store.RunCheckpoint{Seq: 7, TS: baseTime.Add(2 * time.Minute).Format(time.RFC3339), Type: "agent.progress"}
-}
-
-// A2 契约：每个 responsibility.state 都有驱动它的记录，且 next_step 逐状态映射正确。
-func TestBuildResponsibilityStatesAndNextStep(t *testing.T) {
-	cases := []struct {
-		name     string
-		ev       Evidence
-		state    string
-		role     string
-		action   string
-		cardID   string
-		waitHint string
-		since    string
-	}{
-		{
-			name: "unreleased recovery hold wins over everything",
-			ev: Evidence{
-				Run: run(core.RunRunning),
-				Hold: &store.RecoveryHold{RunID: "run_1", Owner: "alice", Reason: "unknown external effect",
-					CreatedAt: afterRun.Format(time.RFC3339)},
-				Cards: []core.HITLCard{{ID: "card_hitl", Resolved: false, Question: "approve?"}},
-			},
-			state: StateRecoveryHold, role: RoleOperationOwner, action: ActionReleaseRecoveryHold,
-			waitHint: "unknown external effect", since: afterRun.Format(time.RFC3339),
-		},
-		{
-			name: "released hold on an interrupted run needs reconciliation",
-			ev: Evidence{
-				Run: run(core.RunInterrupted),
-				Hold: &store.RecoveryHold{RunID: "run_1", Owner: "alice", Reason: "unknown external effect",
-					CreatedAt: afterRun.Format(time.RFC3339), ReleasedAt: afterRun.Add(time.Minute).Format(time.RFC3339)},
-			},
-			state: StateReconciling, role: RoleOperationOwner, action: ActionNeedsManualReconciliation,
-			waitHint: "still interrupted", since: afterRun.Add(time.Minute).Format(time.RFC3339),
-		},
-		{
-			name: "waiting on a HITL card",
-			ev: Evidence{
-				Run:   run(core.RunWaitingHITL),
-				Cards: []core.HITLCard{{ID: "card_hitl", CreatedAt: afterRun.Format(time.RFC3339), Question: "approve the plan?"}},
-			},
-			state: StateWaitingHITL, role: RoleHuman, action: ActionConfirmCard, cardID: "card_hitl",
-			waitHint: "approve the plan?", since: afterRun.Format(time.RFC3339),
-		},
-		{
-			name: "waiting on a pending human decision without a run card",
-			ev: Evidence{
-				Run: run(core.RunRunning),
-				Decisions: []store.HumanDecision{{ID: "hd_1", CommandID: "rdm.confirm", Status: "pending",
-					Question: "approve delivery?", IssuedAt: afterRun, ExpiresAt: afterRun.Add(time.Hour)}},
-			},
-			state: StateWaitingHITL, role: RoleHuman, action: ActionConfirmCard, cardID: "hd_1",
-			waitHint: "approve delivery?", since: afterRun.Format(time.RFC3339),
-		},
-		{
-			name: "dangerous command awaiting confirmation",
-			ev: Evidence{
-				Run: run(core.RunRunning),
-				Confirms: []store.PendingConfirm{{CardID: "card_cmd", Owner: "alice", CreatedAt: afterRun.Format(time.RFC3339),
-					Payload: mustRequest(t, cmdbus.Request{ID: "rdm.docs.instantiate", Owner: "alice", ConversationID: "cnv_1"})}},
-			},
-			state: StateConfirmationPending, role: RoleHuman, action: ActionConfirmCard, cardID: "card_cmd",
-			waitHint: "rdm.docs.instantiate", since: afterRun.Format(time.RFC3339),
-		},
-		{
-			name: "terminal failure is a safe retry",
-			ev: Evidence{
-				Run: ended(func() core.Run {
-					r := run(core.RunFailed)
-					r.Error = &core.ErrorPart{Code: "run.tool_failed", Message: "tool crashed"}
-					return r
-				}(), afterRun),
-			},
-			state: StateFailed, role: RoleRunner, action: ActionSafeRetry,
-			waitHint: "run.tool_failed: tool crashed", since: afterRun.Format(time.RFC3339),
-		},
-		{
-			name: "terminal failure with an unknown receipt needs reconciliation, not a retry",
-			ev: Evidence{
-				Run: ended(run(core.RunFailed), afterRun),
-				Audits: []agentcommand.AuditEnvelope{{Time: afterRun, CommandID: "rdm.docs.instantiate",
-					OperationID: "op_unknown", State: agentcommand.StateUnknown, ConversationID: "cnv_1"}},
-			},
-			state: StateFailed, role: RoleRunner, action: ActionNeedsManualReconciliation,
-			waitHint: "run failed",
-		},
-		{
-			name:  "completed needs no action",
-			ev:    Evidence{Run: ended(run(core.RunCompleted), afterRun)},
-			state: StateCompleted, role: RoleNone, action: ActionNone,
-			waitHint: "completed", since: afterRun.Format(time.RFC3339),
-		},
-		{
-			name:  "executing run stays with the runner",
-			ev:    Evidence{Run: run(core.RunRunning)},
-			state: StateRunning, role: RoleRunner, action: ActionNone,
-			waitHint: "executing", since: baseTime.Format(time.RFC3339),
-		},
-		{
-			name:  "pending run reads as running",
-			ev:    Evidence{Run: run(core.RunPending)},
-			state: StateRunning, role: RoleRunner, action: ActionNone,
-			waitHint: "executing",
-		},
-		{
-			name:  "cancelled has no mapping and degrades to unknown",
-			ev:    Evidence{Run: ended(run(core.RunCancelled), afterRun)},
-			state: StateUnknown, role: RoleNone, action: ActionNone,
-			waitHint: "no diagnostic mapping",
-		},
-		{
-			name:  "interrupted without any hold record degrades to unknown",
-			ev:    Evidence{Run: run(core.RunInterrupted)},
-			state: StateUnknown, role: RoleNone, action: ActionNone,
-			waitHint: "no diagnostic mapping",
-		},
-		{
-			name:  "unreadable hold is fail-closed unknown, never a safe retry",
-			ev:    Evidence{Run: run(core.RunFailed), HoldUnknown: true},
-			state: StateUnknown, role: RoleNone, action: ActionNone,
-			waitHint: "could not be read",
-		},
-	}
-	for _, tc := range cases {
-		t.Run(tc.name, func(t *testing.T) {
-			rep := Build(tc.ev)
-			if rep.Responsibility.State != tc.state {
-				t.Fatalf("state=%q want %q (responsibility=%+v)", rep.Responsibility.State, tc.state, rep.Responsibility)
-			}
-			if rep.Responsibility.Role != tc.role {
-				t.Fatalf("role=%q want %q", rep.Responsibility.Role, tc.role)
-			}
-			if rep.NextStep.Action != tc.action {
-				t.Fatalf("next_step=%q want %q (reason=%q)", rep.NextStep.Action, tc.action, rep.NextStep.Reason)
-			}
-			if rep.NextStep.CardID != tc.cardID {
-				t.Fatalf("next_step.card_id=%q want %q", rep.NextStep.CardID, tc.cardID)
-			}
-			if tc.waitHint != "" && !strings.Contains(rep.Responsibility.WaitReason, tc.waitHint) {
-				t.Fatalf("wait_reason=%q does not contain %q", rep.Responsibility.WaitReason, tc.waitHint)
-			}
-			if tc.since != "" && rep.Responsibility.Since != tc.since {
-				t.Fatalf("since=%q want %q", rep.Responsibility.Since, tc.since)
-			}
-			if rep.NextStep.Reason == "" {
-				t.Fatal("next_step.reason must never be empty")
-			}
-		})
-	}
-}
-
-// 副作用聚合：命令回执对账记录并入；会话里更早的记录不算本次 run 的副作用。
-func TestBuildSideEffectsAggregateRecordedOperations(t *testing.T) {
-	ev := Evidence{
-		Run: run(core.RunFailed),
-		Audits: []agentcommand.AuditEnvelope{
-			{Time: baseTime.Add(-time.Hour), CommandID: "rdm.reqs.add", OperationID: "op_older",
-				State: agentcommand.StateCommitted, ConversationID: "cnv_1"},
-			{Time: afterRun, CommandID: "rdm.docs.instantiate", OperationID: "op_1",
-				State: agentcommand.StateUnknown, ConversationID: "cnv_1"},
-			{Time: afterRun, CommandID: "other.run.command", OperationID: "op_2",
-				State: agentcommand.StateCommitted, ConversationID: "cnv_other"},
-			{Time: afterRun, CommandID: "rdm.trace.confirm", OperationID: "op_3",
-				State: agentcommand.StateAccepted, ConversationID: "cnv_1", RunID: "run_1"},
-		},
-	}
-	rep := Build(ev)
-	if len(rep.SideEffects) != 2 {
-		t.Fatalf("side effects must hold only this run's receipts: %+v", rep.SideEffects)
-	}
-	got := map[string]SideEffect{}
-	for _, se := range rep.SideEffects {
-		got[se.OperationID] = se
-	}
-	recorded, ok := got["op_1"]
-	if !ok || recorded.Kind != EffectOperation || recorded.CommandID != "rdm.docs.instantiate" || recorded.State != string(agentcommand.StateUnknown) {
-		t.Fatalf("recorded operation missing from aggregation: %+v", rep.SideEffects)
-	}
-	if _, ok := got["op_older"]; ok {
-		t.Fatal("a receipt issued before the run started must not be attributed to it")
-	}
-	if _, ok := got["op_2"]; ok {
-		t.Fatal("a receipt from another conversation must not be attributed to this run")
-	}
-	if _, ok := got["op_3"]; !ok {
-		t.Fatal("a receipt carrying this run id must be attributed to it")
-	}
-}
-
-// hold 行本身也是"副作用结果未知"的持久证据（U3b），并标出是否已解除。
-func TestBuildSideEffectsIncludeRecoveryHoldEvidence(t *testing.T) {
-	ev := Evidence{Run: run(core.RunInterrupted), Hold: &store.RecoveryHold{RunID: "run_1", Owner: "alice",
-		Reason: "unknown", CreatedAt: afterRun.Format(time.RFC3339)}}
-	rep := Build(ev)
-	if len(rep.SideEffects) != 1 || rep.SideEffects[0].Kind != EffectRecoveryHold || rep.SideEffects[0].State != "held" {
-		t.Fatalf("open hold must show as held evidence: %+v", rep.SideEffects)
-	}
-	ev.Hold.ReleasedAt = afterRun.Add(time.Minute).Format(time.RFC3339)
-	rep = Build(ev)
-	if rep.SideEffects[0].State != "released" || rep.SideEffects[0].ReleasedAt == "" {
-		t.Fatalf("released hold must keep its release time: %+v", rep.SideEffects[0])
-	}
-}
-
-// 没有落盘检查点时给 null + note，绝不编一个；空副作用同样带 note。
-func TestBuildNeverFabricatesCheckpointOrAbsence(t *testing.T) {
-	rep := Build(Evidence{Run: run(core.RunRunning)})
-	if rep.LastCheckpoint != nil {
-		t.Fatalf("checkpoint must stay null when nothing is persisted: %+v", rep.LastCheckpoint)
-	}
-	if !containsNote(rep.Notes, "checkpoint") {
-		t.Fatalf("missing checkpoint note: %v", rep.Notes)
-	}
-	if !containsNote(rep.Notes, "absence of records is not proof") {
-		t.Fatalf("missing side effect absence note: %v", rep.Notes)
-	}
-	if rep.SideEffects == nil {
-		t.Fatal("side_effects must serialize as an array, never null")
-	}
-	withCheckpoint := Build(Evidence{Run: run(core.RunRunning), Checkpoint: checkpoint()})
-	if withCheckpoint.LastCheckpoint == nil || withCheckpoint.LastCheckpoint.ID != "7" || withCheckpoint.Run.UpdatedAt == nil {
-		t.Fatalf("persisted checkpoint must be surfaced: %+v", withCheckpoint)
-	}
-	if *withCheckpoint.Run.UpdatedAt != checkpoint().TS {
-		t.Fatalf("updated_at must fall back to the last persisted event: %v", *withCheckpoint.Run.UpdatedAt)
-	}
-}
-
-// Gather：只读端口的读齐与降级（读不到隔离记录按 fail-closed）。
-func TestGatherReadsEvidenceAndDegradesOnError(t *testing.T) {
-	src := fakeSource{
-		hold:     store.RecoveryHold{RunID: "run_1", Owner: "alice", Reason: "unknown", CreatedAt: afterRun.Format(time.RFC3339)},
-		cards:    []core.HITLCard{{ID: "card_hitl"}},
-		cp:       store.RunCheckpoint{Seq: 3, TS: afterRun.Format(time.RFC3339), Type: "agent.progress"},
-		cpOK:     true,
-		audits:   []agentcommand.AuditEnvelope{{Time: afterRun, CommandID: "rdm.confirm", State: agentcommand.StateAccepted}},
-		confirms: []store.PendingConfirm{{CardID: "card_cmd", Owner: "alice", CreatedAt: afterRun.Format(time.RFC3339), Payload: "{}"}},
-	}
-	ev := Gather(src, run(core.RunRunning), afterRun.Add(time.Minute))
-	if ev.Hold == nil || ev.HoldUnknown || len(ev.Cards) != 1 || ev.Checkpoint == nil || len(ev.Audits) != 1 || len(ev.Confirms) != 1 {
-		t.Fatalf("gather missed evidence: %+v", ev)
-	}
-	if rep := Build(ev); rep.Responsibility.State != StateRecoveryHold {
-		t.Fatalf("state=%q want %q", rep.Responsibility.State, StateRecoveryHold)
-	}
-
-	degraded := Gather(fakeSource{holdErr: errors.New("db gone")}, run(core.RunRunning), afterRun)
-	if !degraded.HoldUnknown || degraded.Hold != nil {
-		t.Fatalf("an unreadable hold must stay unknown, never absent: %+v", degraded)
-	}
-	rep := Build(degraded)
-	if rep.Responsibility.State != StateUnknown || rep.NextStep.Action == ActionSafeRetry {
-		t.Fatalf("unreadable hold must not yield a retry: %+v", rep)
-	}
-	if !containsNote(rep.Notes, "fail-closed") && !containsNote(rep.Notes, "quarantined") {
-		t.Fatalf("missing fail-closed note: %v", rep.Notes)
-	}
-
-	// 没有隔离行是正常态：不产生噪音 note。
-	normal := Build(Gather(fakeSource{}, run(core.RunRunning), afterRun))
-	if containsNote(normal.Notes, "recovery hold unreadable") {
-		t.Fatalf("a missing hold row is not a degraded read: %v", normal.Notes)
-	}
-}
-
-// 契约形状：snake_case 字段名固定，last_checkpoint 为 null 时显式出现。
-func TestReportJSONContract(t *testing.T) {
-	raw, err := json.Marshal(Build(Evidence{Run: run(core.RunRunning), Checkpoint: checkpoint()}))
-	if err != nil {
-		t.Fatal(err)
-	}
-	var out map[string]any
-	if err := json.Unmarshal(raw, &out); err != nil {
-		t.Fatal(err)
-	}
-	for _, key := range []string{"run", "responsibility", "last_checkpoint", "side_effects", "next_step"} {
-		if _, ok := out[key]; !ok {
-			t.Fatalf("missing contract key %q in %s", key, raw)
-		}
-	}
-	responsibility, _ := out["responsibility"].(map[string]any)
-	for _, key := range []string{"state", "role", "wait_reason", "since"} {
-		if _, ok := responsibility[key]; !ok {
-			t.Fatalf("missing responsibility key %q in %s", key, raw)
-		}
-	}
-	runSummary, _ := out["run"].(map[string]any)
-	for _, key := range []string{"id", "status", "owner", "project", "conversation", "created_at", "updated_at"} {
-		if _, ok := runSummary[key]; !ok {
-			t.Fatalf("missing run key %q in %s", key, raw)
-		}
-	}
-	nextStep, _ := out["next_step"].(map[string]any)
-	for _, key := range []string{"action", "reason"} {
-		if _, ok := nextStep[key]; !ok {
-			t.Fatalf("missing next_step key %q in %s", key, raw)
-		}
-	}
-
-	empty, err := json.Marshal(Build(Evidence{Run: run(core.RunRunning)}))
-	if err != nil {
-		t.Fatal(err)
-	}
-	var nulled map[string]any
-	if err := json.Unmarshal(empty, &nulled); err != nil {
-		t.Fatal(err)
-	}
-	if nulled["last_checkpoint"] != nil {
-		t.Fatalf("last_checkpoint must be null (not omitted) when nothing is persisted: %s", empty)
-	}
-	if _, ok := nulled["side_effects"].([]any); !ok {
-		t.Fatalf("side_effects must be an array: %s", empty)
-	}
-}
-
-// 归属边界：会话里更早签发的决策卡与确认行不算本次 run 在等的东西。
-func TestBuildIgnoresEvidenceIssuedBeforeTheRun(t *testing.T) {
-	ev := Evidence{
-		Run: run(core.RunRunning),
-		Decisions: []store.HumanDecision{{ID: "hd_old", Status: "pending", Question: "older?",
-			IssuedAt: baseTime.Add(-time.Hour), ExpiresAt: baseTime.Add(time.Hour)}},
-		Confirms: []store.PendingConfirm{{CardID: "card_old", Owner: "alice",
-			CreatedAt: baseTime.Add(-time.Hour).Format(time.RFC3339),
-			Payload:   mustRequest(t, cmdbus.Request{ID: "rdm.confirm", Owner: "alice", ConversationID: "cnv_1"})}},
-	}
-	rep := Build(ev)
-	if rep.Responsibility.State != StateRunning {
-		t.Fatalf("state=%q want %q: earlier conversation evidence must not hold this run", rep.Responsibility.State, StateRunning)
-	}
-}
-
-// 只有 run 状态写着 waiting_hitl、没有可解析的卡时：状态仍然成立，但必须说清楚
-// 等待起点没被存储。
-func TestBuildStatusDrivenWaitHasNoCardAndSaysSo(t *testing.T) {
-	rep := Build(Evidence{Run: run(core.RunWaitingHITL)})
-	if rep.Responsibility.State != StateWaitingHITL || rep.NextStep.Action != ActionConfirmCard {
-		t.Fatalf("status-driven wait: %+v", rep.Responsibility)
-	}
-	if rep.NextStep.CardID != "" {
-		t.Fatalf("an unresolvable card must not be invented: %q", rep.NextStep.CardID)
-	}
-	if rep.Responsibility.Since != "" {
-		t.Fatalf("unknown wait entry time must stay empty: %q", rep.Responsibility.Since)
-	}
-	if !containsNote(rep.Notes, "wait entry time is not stored") {
-		t.Fatalf("missing status-driven wait note: %v", rep.Notes)
-	}
-}
-
-func mustRequest(t *testing.T, req cmdbus.Request) string {
-	t.Helper()
-	raw, err := json.Marshal(req)
-	if err != nil {
-		t.Fatal(err)
-	}
-	return string(raw)
-}
-
-func containsNote(notes []string, want string) bool {
-	for _, n := range notes {
-		if strings.Contains(n, want) {
-			return true
-		}
-	}
-	return false
-}

internal/server/intent_cancel_test.go

@@ -361,9 +361,8 @@ func TestIntentCancelRecordIsNotACancelTarget(t *testing.T) {
 	}
 	waitRunStatus(t, h, runID, string(core.RunCancelled))
 	gate.script("再取消一次", taskCancelIntent(nil))
-	// 第二轮是占槽位的澄清回合：等上一轮被取消的运行彻底收工再发车（忙 409 是
-	// 既有语义，见 postConversationTurn）。
-	code, out := postConversationTurn(t, srv, h, st, convID, "再取消一次")
+	code, out := doJSON(t, h, "POST", "/api/conversations/"+convID+"/messages",
+		map[string]any{"text": "再取消一次"})
 	if code != 200 {
 		t.Fatalf("第二轮取消：%d %v", code, out)
 	}

internal/server/intent_session.go

@@ -132,7 +132,11 @@ func (s *Server) fuseIntentSession(key intentsession.Key,
 	}
 	answer := answerFromRecognition(pending, recognized)
 	if answer.Cancel {
-		return recognized, s.applyCancel(svc, key, pending, answer), nil
+		outcome, cerr := s.applyCancel(svc, key, pending, answer)
+		if cerr != nil {
+			return recognized, nil, cerr
+		}
+		return recognized, outcome, nil
 	}
 	// 本轮对同一动作已是一次完整的新识别：它就是新的指令，按它派发；
 	// 旧累计槽位不混进新任务（会话就此作废）。
@@ -162,11 +166,23 @@ func (s *Server) fuseIntentSession(key intentsession.Key,
 }
 
 // applyCancel 消费库层 cancel：清空挂起预期并如实确认取消（不派发业务动作）。
+//
+// fail-closed：除 ErrStale（会话已被推进/取消，按「没有会话语义」如实处理）
+// 之外的任何存储错误一律向上传播。绝不把「清不掉挂起预期」当成「本轮没有
+// 会话语义」——那一行里的运行号仍会被 sessionRunTarget 解成取消靶子，一次
+// 未确认的「放弃补充」就会变成一次真的取消。
 func (s *Server) applyCancel(svc IntentSessions, key intentsession.Key,
-	pending *intentsession.Session, answer intentsession.Answer) *intentSessionOutcome {
+	pending *intentsession.Session, answer intentsession.Answer) (*intentSessionOutcome, error) {
 	out, err := svc.ApplyAnswer(key, pending.Revision, answer)
-	if errors.Is(err, intentsession.ErrStale) || err != nil || out == nil || !out.Cancelled {
-		return nil
+	if errors.Is(err, intentsession.ErrStale) {
+		// 陈旧取消作废：如实按「没有会话语义」处理。
+		return nil, nil
+	}
+	if err != nil {
+		return nil, err
+	}
+	if out == nil || !out.Cancelled {
+		return nil, nil
 	}
 	action, intent := "", pending.Expectation.IntentID
 	if pending.Expectation.Action.ID != "" {
@@ -176,7 +192,7 @@ func (s *Server) applyCancel(svc IntentSessions, key intentsession.Key,
 		cancelled: true,
 		note: &intentSessionNote{Kind: "cancelled", Action: action, Intent: intent, Pending: false,
 			Reply: "已取消本次槽位补充，未执行任何动作。"},
-	}
+	}, nil
 }
 
 // answerFromRecognition 把一次识别投影成会话语义的答案：库层会话命令

internal/server/intent_session_test.go

@@ -3,7 +3,7 @@ package server
 import (
 	"context"
 	"encoding/json"
-	"fmt"
+	"errors"
 	"net/http"
 	"sort"
 	"strings"
@@ -115,105 +115,6 @@ func settle(t *testing.T, h http.Handler, out map[string]any) map[string]any {
 	return out
 }
 
-// 会话忙（409）是既有语义，不改：占槽位的回合（普通回合、澄清回合）在忙时如
-// 实 409，意图命中的命令则不占槽位、会话忙时也能用。测试侧要解决的是「等发得
-// 出去」，而不是放宽断言：上一轮的 run 行在结算提交那一刻就终态了，槽位要到收
-// 尾（补锚点、发终态事件、摘执行句柄）之后才释放，机器繁忙时这段窗口会被放大，
-// 于是「run 一终态就立刻发下一轮」会偶发撞上忙 409。
-
-// postConversationTurn 发一轮消息：先等会话空闲（见 waitConversationIdle），
-// 发车时若仍撞上忙（409）则在「忙拒没有留下痕迹」的前提下按 deadline 重发。
-// 忙不写消息与运行，但抢槽位发生在意图会话折叠之后——被忙拒掉的轮次仍可能已经
-// 消费或挂起了会话，再发一次走的就是另一条路径了；所以每次重发前先比对意图会
-// 话状态：变过了就如实失败，没变过才重发。
-func postConversationTurn(t *testing.T, srv *Server, h http.Handler, st *store.Store,
-	convID, text string) (int, map[string]any) {
-	t.Helper()
-	waitConversationIdle(t, srv, h, convID)
-	deadline := time.Now().Add(testRunTerminalTimeout)
-	for {
-		before := intentSessionTrace(t, st, convID)
-		code, out := doJSON(t, h, "POST", "/api/conversations/"+convID+"/messages", map[string]any{"text": text})
-		if code != http.StatusConflict {
-			return code, out
-		}
-		if after := intentSessionTrace(t, st, convID); after != before {
-			t.Fatalf("%s：忙（409）之前已改动意图会话，重发会走另一条路径：%s → %s", text, before, after)
-		}
-		if !time.Now().Before(deadline) {
-			t.Fatalf("%s：会话持续忙（409）超过 %v：%v", text, testRunTerminalTimeout, out)
-		}
-		time.Sleep(20 * time.Millisecond)
-	}
-}
-
-// waitConversationIdle 等会话真正让出槽位：既没有未终态运行，也没有仍在本地
-// 执行的 run——执行句柄在槽位释放之后才摘除（runmgr 的 execute 收尾），所以
-// 「没有活跃句柄」就是槽位已空的可观测证据，不用赌收尾窗口有多宽。
-func waitConversationIdle(t *testing.T, srv *Server, h http.Handler, convID string) {
-	t.Helper()
-	deadline := time.Now().Add(testRunTerminalTimeout)
-	for time.Now().Before(deadline) {
-		if code, out := doJSON(t, h, "GET", "/api/conversations/"+convID, nil); code == 200 &&
-			!conversationHasOpenRun(out) && !conversationHasActiveRun(srv, out) {
-			return
-		}
-		time.Sleep(5 * time.Millisecond)
-	}
-	t.Fatalf("会话 %s 在 %v 内未空闲", convID, testRunTerminalTimeout)
-}
-
-// conversationHasOpenRun 看会话里是否还有未终态运行（结算尚未提交）。
-func conversationHasOpenRun(detail map[string]any) bool {
-	for _, run := range conversationRuns(detail) {
-		switch run["status"] {
-		case string(core.RunCompleted), string(core.RunFailed),
-			string(core.RunCancelled), string(core.RunInterrupted):
-		default:
-			return true
-		}
-	}
-	return false
-}
-
-// conversationHasActiveRun 看会话里是否还有本地在执行的 run（槽位未让出）。
-func conversationHasActiveRun(srv *Server, detail map[string]any) bool {
-	for _, run := range conversationRuns(detail) {
-		if id, _ := run["id"].(string); id != "" && srv.mgr.RunActiveLocally(id) {
-			return true
-		}
-	}
-	return false
-}
-
-func conversationRuns(detail map[string]any) []map[string]any {
-	raw, _ := detail["runs"].([]any)
-	runs := make([]map[string]any, 0, len(raw))
-	for _, item := range raw {
-		if run, ok := item.(map[string]any); ok {
-			runs = append(runs, run)
-		}
-	}
-	return runs
-}
-
-// intentSessionTrace 描述会话上当前的意图会话状态（忙拒前后各取一次比对）。
-func intentSessionTrace(t *testing.T, st *store.Store, convID string) string {
-	t.Helper()
-	conv, err := st.GetConversation(store.OwnerDefault, convID)
-	if err != nil {
-		t.Fatal(err)
-	}
-	sess, err := st.GetIntentSession(store.OwnerDefault, string(conv.ProjectID), convID)
-	if err == store.ErrNotFound {
-		return "none"
-	}
-	if err != nil {
-		t.Fatal(err)
-	}
-	return fmt.Sprintf("rev=%d exp=%v slots=%v", sess.Revision, sess.Expectation, sess.Slots)
-}
-
 // 缺必填槽位 → 澄清并挂起；下一轮补齐 → 按同一条准入与派发路径执行。
 func TestIntentSessionAnswerCompletesPendingAction(t *testing.T) {
 	srv, st, gate, log := cancelServer(t)
@@ -223,7 +124,7 @@ func TestIntentSessionAnswerCompletesPendingAction(t *testing.T) {
 	gate.script("取消运行", cancelStart(nil, []intentgate.SlotQuestion{
 		{Key: "run_id", Question: runIDQ}, {Key: "reason", Question: "取消原因？"}}))
 
-	code, out := postConversationTurn(t, srv, h, st, convID, "取消运行")
+	code, out := doJSON(t, h, "POST", "/api/conversations/"+convID+"/messages", map[string]any{"text": "取消运行"})
 	if code != 200 {
 		t.Fatalf("第一轮：%d %v", code, out)
 	}
@@ -246,7 +147,7 @@ func TestIntentSessionAnswerCompletesPendingAction(t *testing.T) {
 	// 下一轮：仅补齐槽位（库层挂起预期命中分支的同形结果）。
 	answerText := "编号：R-7 原因：误触发"
 	gate.script(answerText, cancelAnswer(map[string]string{"run_id": "R-7", "reason": "误触发"}, nil))
-	code, out = postConversationTurn(t, srv, h, st, convID, answerText)
+	code, out = doJSON(t, h, "POST", "/api/conversations/"+convID+"/messages", map[string]any{"text": answerText})
 	if code != 200 || out["command"] != "demo.run-cancel" {
 		t.Fatalf("补齐后应派发同一动作：%d %v", code, out)
 	}
@@ -268,12 +169,12 @@ func TestIntentSessionAnswerCompletesPendingAction(t *testing.T) {
 
 // 两轮部分填充：先 set_slot 落一个槽位（仍缺 → 重新澄清），再由回答补齐。
 func TestIntentSessionSetSlotThenAnswerCompletes(t *testing.T) {
-	srv, st, gate, log := cancelServer(t)
+	srv, _, gate, log := cancelServer(t)
 	h := srv.Handler()
 	_, convID := setupConversation(t, h)
 	gate.script("取消运行", cancelStart(nil, []intentgate.SlotQuestion{
 		{Key: "run_id", Question: "取消哪个运行？"}, {Key: "reason", Question: "取消原因？"}}))
-	code, out := postConversationTurn(t, srv, h, st, convID, "取消运行")
+	code, out := doJSON(t, h, "POST", "/api/conversations/"+convID+"/messages", map[string]any{"text": "取消运行"})
 	if code != 200 {
 		t.Fatalf("第一轮：%d %v", code, out)
 	}
@@ -282,7 +183,7 @@ func TestIntentSessionSetSlotThenAnswerCompletes(t *testing.T) {
 	// 第二轮：库层 set_slot（只带一个槽位），仍缺原因 → 如实重新澄清。
 	gate.script("编号 R-7", intentgate.MatchResult{Branch: "expectation_hit", ID: "demo.run-cancel",
 		Commands: []intentgate.Command{{Kind: intentgate.CommandSetSlot, Name: "run_id", Value: "R-7"}}})
-	code, out = postConversationTurn(t, srv, h, st, convID, "编号 R-7")
+	code, out = doJSON(t, h, "POST", "/api/conversations/"+convID+"/messages", map[string]any{"text": "编号 R-7"})
 	if code != 200 {
 		t.Fatalf("set_slot 轮：%d %v", code, out)
 	}
@@ -304,7 +205,7 @@ func TestIntentSessionSetSlotThenAnswerCompletes(t *testing.T) {
 
 	// 第三轮：补上原因 → 派发，两个槽位都来自会话累计。
 	gate.script("原因：误触发", cancelAnswer(map[string]string{"reason": "误触发"}, nil))
-	code, out = postConversationTurn(t, srv, h, st, convID, "原因：误触发")
+	code, out = doJSON(t, h, "POST", "/api/conversations/"+convID+"/messages", map[string]any{"text": "原因：误触发"})
 	if code != 200 || out["command"] != "demo.run-cancel" {
 		t.Fatalf("补齐后应派发：%d %v", code, out)
 	}
@@ -321,7 +222,7 @@ func TestIntentSessionCancelClearsPendingAction(t *testing.T) {
 	_, convID := setupConversation(t, h)
 	gate.script("取消运行", cancelStart(nil, []intentgate.SlotQuestion{
 		{Key: "run_id", Question: "取消哪个运行？"}, {Key: "reason", Question: "取消原因？"}}))
-	code, out := postConversationTurn(t, srv, h, st, convID, "取消运行")
+	code, out := doJSON(t, h, "POST", "/api/conversations/"+convID+"/messages", map[string]any{"text": "取消运行"})
 	if code != 200 {
 		t.Fatalf("第一轮：%d %v", code, out)
 	}
@@ -329,7 +230,7 @@ func TestIntentSessionCancelClearsPendingAction(t *testing.T) {
 
 	gate.script("算了不取消了", intentgate.MatchResult{Branch: "matched", ID: "demo.run-cancel",
 		Commands: []intentgate.Command{{Kind: intentgate.CommandCancel, Target: "demo.run-cancel"}}})
-	code, out = postConversationTurn(t, srv, h, st, convID, "算了不取消了")
+	code, out = doJSON(t, h, "POST", "/api/conversations/"+convID+"/messages", map[string]any{"text": "算了不取消了"})
 	if code != 200 {
 		t.Fatalf("取消轮：%d %v", code, out)
 	}
@@ -354,7 +255,7 @@ func TestIntentSessionCancelClearsPendingAction(t *testing.T) {
 	// 取消之后的回答不得补齐旧动作（这里识别仍不完整 → 只澄清，不派发）。
 	gate.script("编号：R-7", cancelAnswer(map[string]string{"run_id": "R-7"},
 		[]intentgate.SlotQuestion{{Key: "reason", Question: "取消原因？"}}))
-	code, out = postConversationTurn(t, srv, h, st, convID, "编号：R-7")
+	code, out = doJSON(t, h, "POST", "/api/conversations/"+convID+"/messages", map[string]any{"text": "编号：R-7"})
 	if code != 200 {
 		t.Fatalf("取消后回答轮：%d %v", code, out)
 	}
@@ -376,7 +277,8 @@ func TestIntentSessionExpiryDoesNotCompleteOldAction(t *testing.T) {
 	// 第一轮：抽到 run_id，仍缺原因 → 挂起（累计槽位含 run_id）。
 	gate.script("取消运行 编号：R-7", cancelStart(map[string]string{"run_id": "R-7"},
 		[]intentgate.SlotQuestion{{Key: "reason", Question: "取消原因？"}}))
-	code, out := postConversationTurn(t, srv, h, st, convID, "取消运行 编号：R-7")
+	code, out := doJSON(t, h, "POST", "/api/conversations/"+convID+"/messages",
+		map[string]any{"text": "取消运行 编号：R-7"})
 	if code != 200 {
 		t.Fatalf("第一轮：%d %v", code, out)
 	}
@@ -397,7 +299,7 @@ func TestIntentSessionExpiryDoesNotCompleteOldAction(t *testing.T) {
 	// 第二轮只补了原因：过期的 run_id 不得被借来凑齐动作。
 	answerText := "原因：误触发"
 	gate.script(answerText, cancelStart(map[string]string{"reason": "误触发"}, nil))
-	code, out = postConversationTurn(t, srv, h, st, convID, answerText)
+	code, out = doJSON(t, h, "POST", "/api/conversations/"+convID+"/messages", map[string]any{"text": answerText})
 	if code != 200 {
 		t.Fatalf("过期后的回答轮：%d %v", code, out)
 	}
@@ -439,7 +341,7 @@ func TestIntentSessionExpiryDoesNotCompleteOldAction(t *testing.T) {
 		t.Fatal(err)
 	}
 	gate.script("清理缓存", otherIntentStart("demo.purge"))
-	code, out = postConversationTurn(t, srv, h, st, convID, "清理缓存")
+	code, out = doJSON(t, h, "POST", "/api/conversations/"+convID+"/messages", map[string]any{"text": "清理缓存"})
 	if code != 200 || out["command"] != "demo.purge" {
 		t.Fatalf("新任务应照常派发：%d %v", code, out)
 	}
@@ -448,6 +350,105 @@ func TestIntentSessionExpiryDoesNotCompleteOldAction(t *testing.T) {
 	}
 }
 
+// reportSpec 是第二个动作（生成报告）：必填 scope，并声明了可选 run_id。
+// 声明过的键即准入会映射的键，所以上一轮继承来的运行号会在派发参数里
+// 现形——这正是回归要断言的那条泄漏路径。
+var reportSpec = agentcommand.CommandSpec{ID: "demo.report", Title: "生成报告",
+	ParamsSchema: json.RawMessage(`{"type":"object","required":["text","scope"],
+		"properties":{"scope":{"type":"string"},"run_id":{"type":"string"}}}`)}
+
+// reportStart 是第二个动作的一次识别（缺 scope → 路由层澄清 + 挂起）。
+func reportStart(slots map[string]string, missing []intentgate.SlotQuestion) intentgate.MatchResult {
+	return intentgate.MatchResult{Branch: "matched", ID: "demo.report", Slots: slots, Missing: missing,
+		Commands: []intentgate.Command{{Kind: intentgate.CommandStart,
+			Action: intentgate.ActionRef{Kind: intentgate.ActionAuto, ID: "demo.report"}, Slots: slots}}}
+}
+
+// reportAnswer 是对第二个动作挂起预期的回答（库层 expectation_hit 同形）。
+func reportAnswer(slots map[string]string) intentgate.MatchResult {
+	return intentgate.MatchResult{Branch: "expectation_hit", ID: "demo.report", Slots: slots}
+}
+
+// 换动作即换任务（跨动作槽位继承的回归序列）：A 抽到运行号后仍缺原因而
+// 挂起 → 用户改口做另一个动作 B（B 缺 scope）→ 用户补齐 B 的 scope →
+// B 派发时只带 B 自己的槽位，A 的运行号绝不跟着过去。
+func TestIntentSessionNewActionDoesNotInheritOldSlots(t *testing.T) {
+	srv, st, gate, log := cancelServer(t)
+	if err := srv.Bus.RegisterBuiltin(reportSpec, "", func(_ context.Context, req cmdbus.Request) (any, error) {
+		log.record(req)
+		return map[string]any{"ok": true}, nil
+	}); err != nil {
+		t.Fatal(err)
+	}
+	h := srv.Handler()
+	projectID, convID := setupConversation(t, h)
+
+	// T1：动作 A（demo.run-cancel）抽到 run_id，仍缺原因 → 挂起。
+	gate.script("取消运行 编号：R-7", cancelStart(map[string]string{"run_id": "R-7"},
+		[]intentgate.SlotQuestion{{Key: "reason", Question: "取消原因？"}}))
+	code, out := doJSON(t, h, "POST", "/api/conversations/"+convID+"/messages",
+		map[string]any{"text": "取消运行 编号：R-7"})
+	if code != 200 {
+		t.Fatalf("A 轮：%d %v", code, out)
+	}
+	settle(t, h, out)
+	pending := pendingIntent(t, st, projectID, convID)
+	if pending == nil || pending.Expectation.Action.ID != "demo.run-cancel" || pending.Slots["run_id"] != "R-7" {
+		t.Fatalf("A 未挂起/未累计槽位：%+v", pending)
+	}
+
+	// T2：用户改口做动作 B（demo.report，缺 scope）→ 新预期，不继承 A 的槽位。
+	scopeQ := "哪个范围？"
+	gate.script("生成报告", reportStart(nil, []intentgate.SlotQuestion{{Key: "scope", Question: scopeQ}}))
+	code, out = doJSON(t, h, "POST", "/api/conversations/"+convID+"/messages",
+		map[string]any{"text": "生成报告"})
+	if code != 200 {
+		t.Fatalf("B 轮：%d %v", code, out)
+	}
+	settle(t, h, out)
+	if _, executed := out["command"]; executed {
+		t.Fatalf("槽位未就绪不得派发：%v", out)
+	}
+	if c := clarifyOf(t, out); c["command"] != "demo.report" || c["question"] != scopeQ {
+		t.Fatalf("B 的澄清诉求：%v", c)
+	}
+	if len(log.snapshot()) != 0 {
+		t.Fatalf("澄清轮不得派发：%v", log.snapshot())
+	}
+	pending = pendingIntent(t, st, projectID, convID)
+	if pending == nil || pending.Expectation == nil || pending.Expectation.Action.ID != "demo.report" {
+		t.Fatalf("B 未挂起预期：%+v", pending)
+	}
+	if _, leaked := pending.Slots["run_id"]; leaked {
+		t.Fatalf("A 的槽位被 B 继承：%+v", pending.Slots)
+	}
+
+	// T3：用户补齐 B 的 scope → B 派发，参数里只有 B 的槽位（没有 A 的 run_id）。
+	answerText := "范围：华东"
+	gate.script(answerText, reportAnswer(map[string]string{"scope": "华东"}))
+	code, out = doJSON(t, h, "POST", "/api/conversations/"+convID+"/messages",
+		map[string]any{"text": answerText})
+	if code != 200 || out["command"] != "demo.report" {
+		t.Fatalf("补齐后应派发 B：%d %v", code, out)
+	}
+	if _, asked := out["clarify"]; asked {
+		t.Fatalf("补齐后不应再澄清：%v", out)
+	}
+	calls := log.snapshot()
+	if len(calls) != 1 {
+		t.Fatalf("派发次数：%d", len(calls))
+	}
+	if calls[0]["scope"] != "华东" || calls[0]["text"] != answerText {
+		t.Fatalf("B 的派发参数：%v", calls[0])
+	}
+	if _, leaked := calls[0]["run_id"]; leaked {
+		t.Fatalf("A 的运行号漏进了 B 的派发参数：%v", calls[0])
+	}
+	if pendingIntent(t, st, projectID, convID) != nil {
+		t.Fatal("补齐后会话应被消费")
+	}
+}
+
 // otherIntentStart 是另一个动作的完整识别（换任务）。
 func otherIntentStart(id string) intentgate.MatchResult {
 	return intentgate.MatchResult{Branch: "matched", ID: id,
@@ -463,7 +464,7 @@ func TestIntentSessionIsolatesConversations(t *testing.T) {
 	projectB, convB := setupConversation(t, h)
 	gate.script("取消运行", cancelStart(nil, []intentgate.SlotQuestion{
 		{Key: "run_id", Question: "取消哪个运行？"}, {Key: "reason", Question: "取消原因？"}}))
-	code, out := postConversationTurn(t, srv, h, st, convA, "取消运行")
+	code, out := doJSON(t, h, "POST", "/api/conversations/"+convA+"/messages", map[string]any{"text": "取消运行"})
 	if code != 200 {
 		t.Fatalf("A 会话第一轮：%d %v", code, out)
 	}
@@ -478,7 +479,7 @@ func TestIntentSessionIsolatesConversations(t *testing.T) {
 	// B 会话收到同样的回答：A 的槽位不得被借走（没有 A 的累计值）。
 	answer := "编号：R-7 原因：误触发"
 	gate.script(answer, cancelAnswer(map[string]string{"run_id": "R-7", "reason": "误触发"}, nil))
-	code, out = postConversationTurn(t, srv, h, st, convB, answer)
+	code, out = doJSON(t, h, "POST", "/api/conversations/"+convB+"/messages", map[string]any{"text": answer})
 	if code != 200 {
 		t.Fatalf("B 会话回答轮：%d %v", code, out)
 	}
@@ -495,19 +496,19 @@ func TestIntentSessionIsolatesConversations(t *testing.T) {
 func TestIntentSessionStillRequiresHighRiskConfirmation(t *testing.T) {
 	risky := cancelSpec
 	risky.Danger = agentcommand.DangerHigh
-	srv, st, gate, log := cancelServerWith(t, risky)
+	srv, _, gate, log := cancelServerWith(t, risky)
 	h := srv.Handler()
 	_, convID := setupConversation(t, h)
 	gate.script("取消运行", cancelStart(nil, []intentgate.SlotQuestion{
 		{Key: "run_id", Question: "取消哪个运行？"}, {Key: "reason", Question: "取消原因？"}}))
-	code, out := postConversationTurn(t, srv, h, st, convID, "取消运行")
+	code, out := doJSON(t, h, "POST", "/api/conversations/"+convID+"/messages", map[string]any{"text": "取消运行"})
 	if code != 200 {
 		t.Fatalf("第一轮：%d %v", code, out)
 	}
 	settle(t, h, out)
 	answer := "编号：R-7 原因：误触发"
 	gate.script(answer, cancelAnswer(map[string]string{"run_id": "R-7", "reason": "误触发"}, nil))
-	code, out = postConversationTurn(t, srv, h, st, convID, answer)
+	code, out = doJSON(t, h, "POST", "/api/conversations/"+convID+"/messages", map[string]any{"text": answer})
 	if code != 200 || out["command"] != "demo.run-cancel" {
 		t.Fatalf("补齐后应派发到总线：%d %v", code, out)
 	}
@@ -526,7 +527,7 @@ func TestIntentSessionInvalidAnswerKeepsClarifying(t *testing.T) {
 	_, convID := setupConversation(t, h)
 	gate.script("取消运行", cancelStart(nil, []intentgate.SlotQuestion{
 		{Key: "run_id", Question: "取消哪个运行？"}}))
-	code, out := postConversationTurn(t, srv, h, st, convID, "取消运行")
+	code, out := doJSON(t, h, "POST", "/api/conversations/"+convID+"/messages", map[string]any{"text": "取消运行"})
 	if code != 200 {
 		t.Fatalf("第一轮：%d %v", code, out)
 	}
@@ -535,7 +536,7 @@ func TestIntentSessionInvalidAnswerKeepsClarifying(t *testing.T) {
 	gate.script("编号：R-9", intentgate.MatchResult{Branch: "expectation_hit", ID: "demo.run-cancel",
 		Slots:   map[string]string{"run_id": "R-9"},
 		Invalid: []intentgate.SlotDefect{{Key: "run_id", Reason: "unknown run", Question: "取消哪个运行？", Invalid: true}}})
-	code, out = postConversationTurn(t, srv, h, st, convID, "编号：R-9")
+	code, out = doJSON(t, h, "POST", "/api/conversations/"+convID+"/messages", map[string]any{"text": "编号：R-9"})
 	if code != 200 {
 		t.Fatalf("非法补充轮：%d %v", code, out)
 	}
@@ -595,7 +596,7 @@ func TestIntentSessionLegacyResponsesUnchanged(t *testing.T) {
 	}
 	run := func(t *testing.T, sessions IntentSessions) []string {
 		t.Helper()
-		srv, st, gate, log := cancelServer(t)
+		srv, _, gate, log := cancelServer(t)
 		if sessions != nil {
 			srv.SetIntentSessionService(sessions)
 		}
@@ -604,7 +605,7 @@ func TestIntentSessionLegacyResponsesUnchanged(t *testing.T) {
 		_, convID := setupConversation(t, h)
 		var summaries []string
 		for _, text := range []string{"取消运行 编号：R-3 原因：误触发", "今天天气如何", "算了", "取消运行"} {
-			code, out := postConversationTurn(t, srv, h, st, convID, text)
+			code, out := doJSON(t, h, "POST", "/api/conversations/"+convID+"/messages", map[string]any{"text": text})
 			if code != 200 {
 				t.Fatalf("%s：%d %v", text, code, out)
 			}
@@ -636,6 +637,62 @@ func TestIntentSessionLegacyResponsesUnchanged(t *testing.T) {
 	}
 }
 
+// brokenCancelStore 是「放弃补充时写库失败」的替身：其余方法照常委派给
+// 真实会话服务，只有 cancel 的折叠返回非 ErrStale 的存储错误（真实写库
+// 故障的注入点）。
+type brokenCancelStore struct{ IntentSessions }
+
+func (b brokenCancelStore) ApplyAnswer(key intentsession.Key, revision int,
+	answer intentsession.Answer) (*intentsession.Outcome, error) {
+	if answer.Cancel {
+		return nil, errors.New("clear intent session: injected store failure")
+	}
+	return b.IntentSessions.ApplyAnswer(key, revision, answer)
+}
+
+// 放弃补充时写库失败 → fail-closed：本轮报错，绝不降级成「没有会话语义」
+// 去规划/派发一次取消（挂起行里的运行号仍在，会被解成取消靶子）。
+func TestIntentSessionCancelFailsClosedOnStoreError(t *testing.T) {
+	srv, st, gate, log := cancelServer(t)
+	h := srv.Handler()
+	projectID, convID := setupConversation(t, h)
+	runID := liveRun(t, h, convID)
+	seedBoundRun(t, st, projectID, convID, runID)
+	srv.SetIntentSessionService(brokenCancelStore{IntentSessions: intentsession.New(st)})
+
+	// 放弃本次槽位补充（库层 cancel，指向正在挂起的 A）——写库故障就在此刻。
+	gate.script("算了不取消了", intentgate.MatchResult{Branch: "matched", ID: "demo.run-cancel",
+		Commands: []intentgate.Command{{Kind: intentgate.CommandCancel, Target: "demo.run-cancel"}}})
+	code, out := doJSON(t, h, "POST", "/api/conversations/"+convID+"/messages",
+		map[string]any{"text": "算了不取消了"})
+	if code == 200 {
+		t.Fatalf("写库失败的一轮不得照常应答：%d %v", code, out)
+	}
+	if _, asked := out["cancel"]; asked {
+		t.Fatalf("写库失败时不得回执一次取消：%v", out)
+	}
+	if _, ran := out["run"]; ran {
+		t.Fatalf("写库失败时不得起新回合：%v", out)
+	}
+	if len(log.snapshot()) != 0 {
+		t.Fatalf("写库失败时不得派发：%v", log.snapshot())
+	}
+	// 挂起行确实没被清掉（故障没被伪装成成功），这一轮因此更不能取消它。
+	if pendingIntent(t, st, projectID, convID) == nil {
+		t.Fatal("注入的写库失败不应真的清空挂起预期")
+	}
+	// 绝不能真的取消那个运行：给可能的误派发一点时间，再如实断言。
+	deadline := time.Now().Add(500 * time.Millisecond)
+	for time.Now().Before(deadline) {
+		if code, d := doJSON(t, h, "GET", "/api/runs/"+runID, nil); code == 200 {
+			if r, _ := d["run"].(map[string]any); r["status"] == string(core.RunCancelled) {
+				t.Fatalf("写库失败的一轮仍然取消了运行 %s：%v", runID, r)
+			}
+		}
+		time.Sleep(20 * time.Millisecond)
+	}
+}
+
 // inertSessions 是「关闭会话语义」的替身：任何调用都如实报告没有会话。
 type inertSessions struct{}
 

internal/server/run_diagnostics.go

@@ -1,32 +0,0 @@
-package server
-
-import (
-	"net/http"
-	"time"
-
-	"agent-system-backend/internal/rundiag"
-)
-
-// getRunDiagnostics 是单次 run 的只读诊断入口（feature unit A2）：
-//
-//	GET /api/runs/{id}/diagnostics
-//
-// 它回答值班/排障最常问的六个问题：这个 run 卡在哪（responsibility.state）、
-// 谁该动（responsibility.role）、为什么在等（responsibility.wait_reason）、最后一个
-// 稳定落盘点（last_checkpoint）、已知的外部副作用（side_effects）、下一步动作
-// （next_step）。
-//
-// 授权口径与既有 run 读入口逐字同款：owner 只取自认证上下文，绝不接受请求参数；
-// run 由 store.GetRun(owner, id) 做租户过滤，跨租户与不存在同为 404（不泄露存在性）。
-// handler 路径上只有 SELECT——不给写操作、不改状态、不触发任何副作用；聚合逻辑在
-// internal/rundiag，证据缺失一律降级为 unknown + notes，不编造检查点，也不把"没有
-// 记录"读成"没有发生"。
-func (s *Server) getRunDiagnostics(w http.ResponseWriter, r *http.Request) {
-	run, err := s.st.GetRun(s.owner(r), r.PathValue("id"))
-	if err != nil {
-		writeStoreErr(w, err)
-		return
-	}
-	evidence := rundiag.Gather(s.st, run, time.Now().UTC())
-	writeJSON(w, http.StatusOK, rundiag.Build(evidence))
-}

internal/server/run_diagnostics_test.go

@@ -1,226 +0,0 @@
-package server
-
-import (
-	"encoding/json"
-	"net/http"
-	"net/http/httptest"
-	"strings"
-	"testing"
-	"time"
-
-	agentcommand "agent-command"
-	"agent-system-backend/internal/store"
-	core "agent-system-core"
-)
-
-// diagnosticsFixture 建一个 owner 名下的项目/会话/run（合成数据，无执行器）。
-func diagnosticsFixture(t *testing.T, st *store.Store, owner string, status core.RunStatus) (core.Run, core.Conversation) {
-	t.Helper()
-	project, err := st.CreateProject(owner, "诊断项目")
-	if err != nil {
-		t.Fatal(err)
-	}
-	conversation, err := st.CreateConversation(owner, project.ID, "诊断会话", "")
-	if err != nil {
-		t.Fatal(err)
-	}
-	run := core.Run{
-		ConversationID: conversation.ID, ProjectID: project.ID, Owner: owner,
-		Kind: core.RunConversation, Status: status, Phase: core.PhaseActing, CreatedAt: time.Now().UTC(),
-	}
-	if err := st.CreateRun(&run); err != nil {
-		t.Fatal(err)
-	}
-	return run, conversation
-}
-
-func getDiagnostics(t *testing.T, h http.Handler, key, path string) *httptest.ResponseRecorder {
-	t.Helper()
-	req := httptest.NewRequest(http.MethodGet, path, nil)
-	if key != "" {
-		req.Header.Set("Authorization", "Bearer "+key)
-	}
-	rec := httptest.NewRecorder()
-	h.ServeHTTP(rec, req)
-	return rec
-}
-
-func decodeReport(t *testing.T, rec *httptest.ResponseRecorder) map[string]any {
-	t.Helper()
-	if rec.Code != http.StatusOK {
-		t.Fatalf("diagnostics: %d %s", rec.Code, rec.Body.String())
-	}
-	var out map[string]any
-	if err := json.Unmarshal(rec.Body.Bytes(), &out); err != nil {
-		t.Fatalf("decode diagnostics: %v (%s)", err, rec.Body.String())
-	}
-	return out
-}
-
-// 授权口径与既有 run 读入口同款：未认证 401、跨 owner 404（不泄露存在性）、
-// 不存在的 run 同样 404。
-func TestRunDiagnosticsOwnerScopingMatchesRunReads(t *testing.T) {
-	srv, st := newAuthTestServer(t)
-	h := srv.Handler()
-	aliceKey, err := createLegacyAPIKeyFixture(t, st, "alice", "e2e")
-	if err != nil {
-		t.Fatal(err)
-	}
-	bobKey, err := createLegacyAPIKeyFixture(t, st, "bob", "e2e")
-	if err != nil {
-		t.Fatal(err)
-	}
-	run, _ := diagnosticsFixture(t, st, "alice", core.RunRunning)
-	path := "/api/runs/" + string(run.ID) + "/diagnostics"
-
-	if rec := getDiagnostics(t, h, "", path); rec.Code != http.StatusUnauthorized {
-		t.Fatalf("unauthenticated diagnostics must be rejected: %d", rec.Code)
-	}
-	if rec := getDiagnostics(t, h, bobKey, path); rec.Code != http.StatusNotFound {
-		t.Fatalf("cross-owner diagnostics must not leak the payload: %d %s", rec.Code, rec.Body.String())
-	}
-	if rec := getDiagnostics(t, h, aliceKey, "/api/runs/run_missing/diagnostics"); rec.Code != http.StatusNotFound {
-		t.Fatalf("missing run must be 404: %d", rec.Code)
-	}
-	rec := getDiagnostics(t, h, aliceKey, path)
-	report := decodeReport(t, rec)
-	runSection, _ := report["run"].(map[string]any)
-	if runSection["id"] != string(run.ID) || runSection["owner"] != "alice" || runSection["project"] != string(run.ProjectID) {
-		t.Fatalf("run identity wrong: %v", runSection)
-	}
-	responsibility, _ := report["responsibility"].(map[string]any)
-	if responsibility["state"] != "running" {
-		t.Fatalf("a running run with no blockers reads as running: %v", responsibility)
-	}
-}
-
-// 隔离中的 run：责任在副作用责任人，下一步是解除隔离；检查点与已记录的命令
-// 回执一并给出。
-func TestRunDiagnosticsReportsHoldCheckpointAndSideEffects(t *testing.T) {
-	srv, st := newAuthTestServer(t)
-	h := srv.Handler()
-	key, err := createLegacyAPIKeyFixture(t, st, "alice", "e2e")
-	if err != nil {
-		t.Fatal(err)
-	}
-	run, conversation := diagnosticsFixture(t, st, "alice", core.RunRunning)
-	if _, err := st.AppendEvents(string(conversation.ID), []core.StreamEvent{
-		{RunID: run.ID, Type: core.EvThinkingDelta, Payload: json.RawMessage(`{"index":0,"text":"核对差异"}`)},
-		{RunID: run.ID, Type: core.EvAgentProgress, Payload: json.RawMessage(`{"note":"核对中"}`)},
-	}); err != nil {
-		t.Fatal(err)
-	}
-	if err := st.AppendAudit("alice", agentcommand.AuditEnvelope{
-		Time: time.Now().UTC(), Actor: "alice", CommandID: "rdm.docs.instantiate",
-		OperationID: "op_diag_1", ConversationID: string(conversation.ID), State: agentcommand.StateUnknown,
-	}); err != nil {
-		t.Fatal(err)
-	}
-	if err := st.SetRecoveryHold(string(run.ID), "alice", string(conversation.ID), "外部副作用结果未知"); err != nil {
-		t.Fatal(err)
-	}
-
-	report := decodeReport(t, getDiagnostics(t, h, key, "/api/runs/"+string(run.ID)+"/diagnostics"))
-	responsibility, _ := report["responsibility"].(map[string]any)
-	if responsibility["state"] != "recovery_hold" || responsibility["role"] != "operation-owner" {
-		t.Fatalf("held run responsibility: %v", responsibility)
-	}
-	nextStep, _ := report["next_step"].(map[string]any)
-	if nextStep["action"] != "release_recovery_hold" {
-		t.Fatalf("held run next step: %v", nextStep)
-	}
-	checkpoint, _ := report["last_checkpoint"].(map[string]any)
-	if checkpoint == nil || checkpoint["type"] != string(core.EvAgentProgress) {
-		t.Fatalf("last persisted checkpoint must be surfaced: %v", report["last_checkpoint"])
-	}
-	effects, _ := report["side_effects"].([]any)
-	var foundOperation bool
-	for _, raw := range effects {
-		effect, _ := raw.(map[string]any)
-		if effect["kind"] == "operation" && effect["operation_id"] == "op_diag_1" {
-			foundOperation = true
-			if effect["command_id"] != "rdm.docs.instantiate" || effect["state"] != string(agentcommand.StateUnknown) {
-				t.Fatalf("operation projection: %v", effect)
-			}
-		}
-	}
-	if !foundOperation {
-		t.Fatalf("recorded operation missing from side effects: %v", effects)
-	}
-}
-
-// 只读契约：GET 不改任何状态，写方法不注册。
-func TestRunDiagnosticsIsReadOnly(t *testing.T) {
-	srv, st := newAuthTestServer(t)
-	h := srv.Handler()
-	key, err := createLegacyAPIKeyFixture(t, st, "alice", "e2e")
-	if err != nil {
-		t.Fatal(err)
-	}
-	run, conversation := diagnosticsFixture(t, st, "alice", core.RunRunning)
-	if err := st.SetRecoveryHold(string(run.ID), "alice", string(conversation.ID), "unknown"); err != nil {
-		t.Fatal(err)
-	}
-	before, err := st.GetRun("alice", string(run.ID))
-	if err != nil {
-		t.Fatal(err)
-	}
-	path := "/api/runs/" + string(run.ID) + "/diagnostics"
-	if rec := getDiagnostics(t, h, key, path); rec.Code != http.StatusOK {
-		t.Fatalf("diagnostics: %d %s", rec.Code, rec.Body.String())
-	}
-	after, err := st.GetRun("alice", string(run.ID))
-	if err != nil {
-		t.Fatal(err)
-	}
-	if before.Status != after.Status || before.Phase != after.Phase || !before.CreatedAt.Equal(after.CreatedAt) {
-		t.Fatalf("read-only endpoint mutated the run: %+v -> %+v", before, after)
-	}
-	hold, err := st.GetRecoveryHold(string(run.ID))
-	if err != nil || hold.ReleasedAt != "" {
-		t.Fatalf("read-only endpoint changed the recovery hold: %+v err=%v", hold, err)
-	}
-	for _, method := range []string{http.MethodPost, http.MethodPut, http.MethodDelete, http.MethodPatch} {
-		req := httptest.NewRequest(method, path, nil)
-		req.Header.Set("Authorization", "Bearer "+key)
-		rec := httptest.NewRecorder()
-		h.ServeHTTP(rec, req)
-		if rec.Code == http.StatusOK || rec.Code == http.StatusAccepted {
-			t.Fatalf("%s must not be served on a read-only diagnostics path: %d", method, rec.Code)
-		}
-	}
-}
-
-// 读不到的证据降级为 unknown + note，不让整个诊断失败，也不给出安全重试。
-func TestRunDiagnosticsUnknownStateDegradesWithNote(t *testing.T) {
-	srv, st := newAuthTestServer(t)
-	h := srv.Handler()
-	key, err := createLegacyAPIKeyFixture(t, st, "alice", "e2e")
-	if err != nil {
-		t.Fatal(err)
-	}
-	run, _ := diagnosticsFixture(t, st, "alice", core.RunCancelled)
-
-	report := decodeReport(t, getDiagnostics(t, h, key, "/api/runs/"+string(run.ID)+"/diagnostics"))
-	responsibility, _ := report["responsibility"].(map[string]any)
-	if responsibility["state"] != "unknown" {
-		t.Fatalf("unmapped status must degrade to unknown: %v", responsibility)
-	}
-	if report["last_checkpoint"] != nil {
-		t.Fatalf("no persisted event means no checkpoint reference: %v", report["last_checkpoint"])
-	}
-	notes, _ := report["notes"].([]any)
-	var sawStatusNote, sawEffectNote bool
-	for _, raw := range notes {
-		note, _ := raw.(string)
-		if strings.Contains(note, "no diagnostic mapping") {
-			sawStatusNote = true
-		}
-		if strings.Contains(note, "absence of records is not proof") {
-			sawEffectNote = true
-		}
-	}
-	if !sawStatusNote || !sawEffectNote {
-		t.Fatalf("missing degradation notes: %v", notes)
-	}
-}

internal/server/server.go

@@ -392,8 +392,6 @@ func (s *Server) Handler() http.Handler {
 	mux.HandleFunc("DELETE /api/prompt-commands/{id}", s.deleteUserPromptCommand)
 	mux.HandleFunc("POST /api/runs/{id}/cancel", s.cancelRun)
 	mux.HandleFunc("GET /api/runs/{id}", s.getRun)
-	// 只读运行诊断（A2）：卡在哪、谁负责、已知副作用、下一步；无写操作。
-	mux.HandleFunc("GET /api/runs/{id}/diagnostics", s.getRunDiagnostics)
 
 	// ---- HITL ----
 	mux.HandleFunc("POST /api/hitl/{id}/resolve", s.resolveHITL)

internal/store/run_diagnostics.go

@@ -1,110 +0,0 @@
-package store
-
-import (
-	"database/sql"
-	"errors"
-	"time"
-)
-
-// 运行诊断的只读投影（feature unit A2）。本文件只放 SELECT：不加表、不加索引、
-// 不动迁移，也不引入新的写语义。诊断面回答的是"已经落库的证据是什么"，因此
-// 每个查询都遵守同一条口径：查不到就是"没有记录"，调用方不得把它读成"没发生"
-// （与 /api/commands/operations 的对账查询同源，见 cmdbus.OperationQuery）。
-
-// RunCheckpoint 是 run 事件流上最后一个已持久化事件的稳定引用。
-//
-// 诚实边界：runs 行本身没有"检查点"列——事件流（events）才是 run 唯一的持久
-// 进度痕迹（续跑、SSE 重放都以它为准）。所以这里给的是"最后一次稳定落盘"的
-// 定位点（seq 可作游标），不是推断出来的中间状态；run 没有任何事件时不存在这条
-// 记录，调用方必须显式处理"没有"而不是编一个。
-type RunCheckpoint struct {
-	Seq  int64  // 会话事件序（可作重放游标）
-	TS   string // RFC3339（写入方原样落库）
-	Type string // 事件类型
-}
-
-// LastRunCheckpoint 返回该 run 在会话事件流里的最后一条已持久化事件。
-// 没有事件时 ok=false，err 为 nil。
-func (s *Store) LastRunCheckpoint(conversationID, runID string) (RunCheckpoint, bool, error) {
-	if conversationID == "" || runID == "" {
-		return RunCheckpoint{}, false, nil
-	}
-	var cp RunCheckpoint
-	err := s.db.QueryRow(s.q(`SELECT seq, ts, type FROM events
-		WHERE conversation_id=? AND run_id=? ORDER BY seq DESC LIMIT 1`), conversationID, runID).
-		Scan(&cp.Seq, &cp.TS, &cp.Type)
-	if errors.Is(err, sql.ErrNoRows) {
-		return RunCheckpoint{}, false, nil
-	}
-	if err != nil {
-		return RunCheckpoint{}, false, err
-	}
-	return cp, true, nil
-}
-
-// PendingHumanDecisionsByConversation 列出会话内未过期、尚未裁决的人工决策卡
-// （owner + 会话双谓词，跨租户天然不可见；过期卡按 PendingHumanDecisions 的同一
-// 口径过滤）。诊断面据此判断 run 是否在等人工裁决。
-//
-// 口径说明：决策卡按命令与会话归属，表里没有 run 列——把它归到某个 run 是调用方
-// 的时间窗推断，不是存储事实（见 rundiag 的归属边界注释）。
-func (s *Store) PendingHumanDecisionsByConversation(owner, conversationID string, now time.Time) ([]HumanDecision, error) {
-	if owner == "" || conversationID == "" {
-		return []HumanDecision{}, nil
-	}
-	rows, err := s.db.Query(s.q(`SELECT id,plugin_id,audience,command_id,human_actor,owner,project_id,proposal_id,
-		proposal_revision,payload_hash,question,status,decision,proof_id,conversation_id,issued_at,expires_at,responded_at
-		FROM human_decisions WHERE owner=? AND conversation_id=? AND status='pending' ORDER BY issued_at`),
-		owner, conversationID)
-	if err != nil {
-		return nil, err
-	}
-	defer rows.Close()
-	out := []HumanDecision{}
-	for rows.Next() {
-		rec, err := scanHumanDecision(rows)
-		if err != nil {
-			return nil, err
-		}
-		if now.Before(rec.ExpiresAt) {
-			out = append(out, rec)
-		}
-	}
-	return out, rows.Err()
-}
-
-// PendingConfirm 是危险命令确认挂起态（pending_confirms）的只读行。
-// Payload 是 cmdbus 落下的分发请求 JSON（确认后据此二次分发）；诊断面只解析出
-// 归属与命令标识，不回显原始载荷。
-type PendingConfirm struct {
-	CardID    string
-	Owner     string
-	Payload   string
-	CreatedAt string
-}
-
-// ListOwnerPendingConfirms 列出该租户名下尚未确认的危险命令挂起行（时间正序）。
-// owner 为空视为无作用域：不返回任何行（纵深防御，认证层恒注入非空 owner）。
-func (s *Store) ListOwnerPendingConfirms(owner string, limit int) ([]PendingConfirm, error) {
-	if owner == "" {
-		return []PendingConfirm{}, nil
-	}
-	if limit <= 0 || limit > 200 {
-		limit = 50
-	}
-	rows, err := s.db.Query(s.q(`SELECT card_id, owner, payload, created_at FROM pending_confirms
-		WHERE owner=? ORDER BY created_at LIMIT ?`), owner, limit)
-	if err != nil {
-		return nil, err
-	}
-	defer rows.Close()
-	out := []PendingConfirm{}
-	for rows.Next() {
-		var p PendingConfirm
-		if err := rows.Scan(&p.CardID, &p.Owner, &p.Payload, &p.CreatedAt); err != nil {
-			return nil, err
-		}
-		out = append(out, p)
-	}
-	return out, rows.Err()
-}

internal/store/run_diagnostics_test.go

@@ -1,120 +0,0 @@
-package store
-
-import (
-	"encoding/json"
-	"testing"
-	"time"
-
-	core "agent-system-core"
-)
-
-// 运行诊断只读投影：检查点定位、HITL 决策卡的会话+租户作用域、确认挂起行的租户
-// 作用域。口径与既有读入口一致：查不到就是"没有记录"，不推断"没发生"。
-func TestRunDiagnosticsReadHelpers(t *testing.T) {
-	st := openTestStore(t)
-	base := time.Now().UTC().Truncate(time.Second)
-
-	if err := st.EnsureUser("alice", ""); err != nil {
-		t.Fatal(err)
-	}
-	if err := st.EnsureUser("bob", ""); err != nil {
-		t.Fatal(err)
-	}
-	project, err := st.CreateProject("alice", "诊断项目")
-	if err != nil {
-		t.Fatal(err)
-	}
-	conversation, err := st.CreateConversation("alice", project.ID, "诊断会话", "")
-	if err != nil {
-		t.Fatal(err)
-	}
-	run := core.Run{ConversationID: conversation.ID, ProjectID: project.ID, Owner: "alice",
-		Kind: core.RunConversation, Status: core.RunRunning, CreatedAt: base}
-	if err := st.CreateRun(&run); err != nil {
-		t.Fatal(err)
-	}
-	other := core.Run{ConversationID: conversation.ID, ProjectID: project.ID, Owner: "alice",
-		Kind: core.RunSubagent, Status: core.RunRunning, CreatedAt: base}
-	if err := st.CreateRun(&other); err != nil {
-		t.Fatal(err)
-	}
-
-	// 检查点：只认本 run 的最后一条事件，且不跨 run 串味。
-	if cp, ok, err := st.LastRunCheckpoint(string(conversation.ID), string(run.ID)); err != nil || ok {
-		t.Fatalf("a run without events has no checkpoint: %+v ok=%v err=%v", cp, ok, err)
-	}
-	if _, err := st.AppendEvents(string(conversation.ID), []core.StreamEvent{
-		{RunID: run.ID, Type: core.EvThinkingDelta, TS: base.Format(time.RFC3339), Payload: json.RawMessage(`{}`)},
-		{RunID: other.ID, Type: core.EvToolOutput, TS: base.Format(time.RFC3339), Payload: json.RawMessage(`{}`)},
-		{RunID: run.ID, Type: core.EvAgentProgress, TS: base.Format(time.RFC3339), Payload: json.RawMessage(`{}`)},
-	}); err != nil {
-		t.Fatal(err)
-	}
-	cp, ok, err := st.LastRunCheckpoint(string(conversation.ID), string(run.ID))
-	if err != nil || !ok {
-		t.Fatalf("checkpoint read: ok=%v err=%v", ok, err)
-	}
-	if cp.Type != string(core.EvAgentProgress) || cp.TS == "" || cp.Seq <= 0 {
-		t.Fatalf("checkpoint must point at the newest persisted event of this run: %+v", cp)
-	}
-	if child, ok, err := st.LastRunCheckpoint(string(conversation.ID), string(other.ID)); err != nil || !ok || child.Type != string(core.EvToolOutput) {
-		t.Fatalf("child run checkpoint: %+v ok=%v err=%v", child, ok, err)
-	}
-	if cp, ok, err := st.LastRunCheckpoint("cnv_other", string(run.ID)); err != nil || ok {
-		t.Fatalf("another conversation must not expose this run's events: %+v ok=%v err=%v", cp, ok, err)
-	}
-
-	// HITL 决策卡：pending + 未过期才可见，且按 owner+会话双谓词隔离。
-	now := base.Add(time.Minute)
-	issue := func(id, owner, conversationID string, ttl time.Duration) {
-		t.Helper()
-		if _, err := st.IssueHumanDecision(HumanDecision{PluginID: "p", Audience: "a", CommandID: "rdm.confirm",
-			HumanActor: owner, Owner: owner, ProjectID: string(project.ID), ProposalID: id, ProposalRevision: 1,
-			PayloadHash: "h" + id, Question: "approve " + id + "?", ConversationID: conversationID},
-			now.Add(-time.Minute)); err != nil {
-			t.Fatal(err)
-		}
-		// 过期口径：把 expires_at 改到过去，模拟签发即过期的卡（走方言绑定）。
-		if ttl < 0 {
-			if _, err := st.DB().Exec(st.q(`UPDATE human_decisions SET expires_at=? WHERE proposal_id=?`),
-				now.Add(-time.Hour).Format(time.RFC3339Nano), id); err != nil {
-				t.Fatal(err)
-			}
-		}
-	}
-	issue("alive", "alice", string(conversation.ID), time.Hour)
-	issue("expired", "alice", string(conversation.ID), -time.Hour)
-	issue("otherowner", "bob", "cnv_bob", time.Hour)
-
-	pending, err := st.PendingHumanDecisionsByConversation("alice", string(conversation.ID), now)
-	if err != nil {
-		t.Fatal(err)
-	}
-	if len(pending) != 1 || pending[0].ProposalID != "alive" || pending[0].Status != "pending" {
-		t.Fatalf("only live pending cards in scope: %+v", pending)
-	}
-	if cross, err := st.PendingHumanDecisionsByConversation("bob", string(conversation.ID), now); err != nil || len(cross) != 0 {
-		t.Fatalf("cross-owner cards must stay invisible: %+v err=%v", cross, err)
-	}
-	if none, err := st.PendingHumanDecisionsByConversation("", string(conversation.ID), now); err != nil || len(none) != 0 {
-		t.Fatalf("an empty owner has no scope: %+v err=%v", none, err)
-	}
-
-	// 确认挂起行：按租户隔离，空 owner 无作用域。
-	if err := st.SavePendingConfirm("card_alice", "alice", `{"ID":"rdm.confirm"}`); err != nil {
-		t.Fatal(err)
-	}
-	if err := st.SavePendingConfirm("card_bob", "bob", `{"ID":"rdm.confirm"}`); err != nil {
-		t.Fatal(err)
-	}
-	confirms, err := st.ListOwnerPendingConfirms("alice", 0)
-	if err != nil {
-		t.Fatal(err)
-	}
-	if len(confirms) != 1 || confirms[0].CardID != "card_alice" || confirms[0].Owner != "alice" || confirms[0].Payload == "" || confirms[0].CreatedAt == "" {
-		t.Fatalf("owner-scoped pending confirms: %+v", confirms)
-	}
-	if none, err := st.ListOwnerPendingConfirms("", 0); err != nil || len(none) != 0 {
-		t.Fatalf("an empty owner has no scope: %+v err=%v", none, err)
-	}
-}