Add StepSpec.Brief rendered as the task-instruction section of TASK.md

Why: the Q3 compiler maps core.WorkflowDef prompt steps onto squad graph
steps; the rendered prompt text needs a first-class carrier in the step
task brief (until now only leader delegations had a briefing channel).

Evidence: TestStepBriefRendersIntoTaskFile asserts the section presence,
body, and position (after targets, before the context table); full squad
suite green (SELF), CI authoritative (GATE).

Impact: additive StepSpec field brief + writeTaskFile section; no behavior
change when Brief is empty.

Refs: Q3 plan WP1 (2026-10-07-agent-stack-q2q3-plans.md); Dev task card
.task/feature-agent-stack-q3.md

adapters/native/register.go

@@ -138,6 +138,8 @@ func toNative(req squad.ExecRequest) exec.Request {
 			prev(squad.ProcessEvent{
 				Type: ev.Type, Text: ev.Text, Tool: ev.Tool, Detail: ev.Detail,
 				PromptTokens: ev.PromptTokens, CompletionTokens: ev.CompletionTokens,
+				LastPromptTokens: ev.LastPromptTokens, LastCompletionTokens: ev.LastCompletionTokens,
+				CacheReadTokens: ev.CacheReadTokens, CacheWriteTokens: ev.CacheWriteTokens,
 			})
 		}
 	}

squad/brief_test.go

@@ -0,0 +1,41 @@
+package squad
+
+import (
+	"context"
+	"strings"
+	"testing"
+)
+
+// brief_test.go — Q3 WP1a：StepSpec.Brief 渲染进 TASK.md「任务指令」节。
+
+func TestStepBriefRendersIntoTaskFile(t *testing.T) {
+	rec := &promptRecordingBackend{prompts: map[string][]string{}}
+	engine := newRecordingEngine(t, rec, []StepSpec{
+		{ID: "s1", Role: "researcher", Title: "起草", Targets: []string{"output.md"},
+			Brief: "写一份关于「太空电梯」的简报，篇幅：短。\n必须包含单独一行「核对标记：P4」。"},
+	})
+	res, err := engine.Run(context.Background(), Goal{Text: "带任务指令"}, RunOptions{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunCompleted {
+		t.Fatalf("status = %s err = %v", res.Status, res.Error)
+	}
+	var prompt string
+	for _, p := range rec.prompts["01_researcher"] {
+		prompt += p
+	}
+	if !strings.Contains(prompt, "## 任务指令") {
+		t.Fatalf("TASK.md 缺任务指令节：%q", prompt)
+	}
+	if !strings.Contains(prompt, "核对标记：P4") {
+		t.Fatalf("任务指令正文缺失：%q", prompt)
+	}
+	// 指令节在期望产出之后、上下文表之前。
+	iOut := strings.Index(prompt, "期望产出")
+	iBrief := strings.Index(prompt, "## 任务指令")
+	iCtx := strings.Index(prompt, "## 任务上下文资源")
+	if !(iOut < iBrief && iBrief < iCtx) {
+		t.Fatalf("任务指令节位置错误：%d/%d/%d", iOut, iBrief, iCtx)
+	}
+}

squad/engine.go

@@ -76,6 +76,11 @@ func (d engineDeps) storeOr() Store {
 }
 
 // loadRun 经端口加载运行状态，并让它携带同一端口（后续 save 自然走同一实现）。
+// runDir 是运行的工作区目录（所有 run 级 JSON 的根）。
+func (e *SquadEngine) runDir(runID string) string {
+	return filepath.Join(e.opts.DataDir, "runs", runID)
+}
+
 func (e *SquadEngine) loadRun(wsDir string) (*RunState, error) {
 	st, err := e.deps.storeOr().LoadRun(wsDir)
 	if err != nil {
@@ -100,6 +105,28 @@ type RunOptions struct {
 	MaxLeaderTurns int `json:"max_leader_turns,omitempty"`
 	// CheckpointMode superstep 边界快照档位（Q1 WP4）：sync（默认）| async | on_exit。
 	CheckpointMode string `json:"checkpoint_mode,omitempty"`
+	// Topology 显式拓扑名（Q2；优先于 Blueprint.Topology；空 = standard 图）。
+	// 未注册名 / primitive 单跑 / driver 未点亮均 fail-closed。
+	Topology string `json:"topology,omitempty"`
+	// Deadline 运行级截止时刻（Q2 long_running）：superstep 边界检查，
+	// 超期把未完步骤置 skipped 并降级收口。零值 = 不设限。
+	Deadline time.Time `json:"deadline,omitempty"`
+	// MaxReplans leader 模式的重规划上限（Q2；与 MaxLeaderTurns 独立）。
+	// 0 = 默认 DefaultMaxReplans。
+	MaxReplans int `json:"max_replans,omitempty"`
+	// GroupChatPolicy group_chat 拓扑发言人策略（Q2）：rotate（默认）| moderator。
+	GroupChatPolicy string `json:"group_chat_policy,omitempty"`
+	// GroupChatModerator 主持人角色 key（moderator 策略必填，须在名册内）。
+	GroupChatModerator string `json:"group_chat_moderator,omitempty"`
+	// GroupChatMaxTurns 群聊轮次预算（0 = DefaultMaxGroupChatTurns）。
+	GroupChatMaxTurns int `json:"group_chat_max_turns,omitempty"`
+	// ReviewVerifyRole review 拓扑核实角色 key（默认 "verifier"）。
+	ReviewVerifyRole string `json:"review_verify_role,omitempty"`
+	// ReviewTarget 被评审材料（默认取 context/goal.md）。
+	ReviewTarget string `json:"review_target,omitempty"`
+	// TaskBoardLeaseSeconds task_board 任务租约时限（秒；0 = 默认 600）。
+	// 孤儿租约（claimed 且过期）在驱动重入时再武装回 open。
+	TaskBoardLeaseSeconds int `json:"task_board_lease_seconds,omitempty"`
 }
 
 // SquadEngine 小队引擎门面：Run / Respond / Resume / Inspect / ListRuns。
@@ -234,6 +261,39 @@ func (e *SquadEngine) Run(ctx context.Context, goal Goal, runOpts RunOptions) (*
 	if runOpts.CheckpointMode != "" {
 		meta.Metadata["checkpoint_mode"] = runOpts.CheckpointMode
 	}
+	if runOpts.Topology != "" {
+		meta.Metadata["topology"] = runOpts.Topology
+		// leader 拓扑映射到委派循环须发生在计划确认之前（确认门按
+		// mode=leader 豁免）；后置映射会让 leader 运行误触计划确认。
+		if runOpts.Topology == TopologyLeader && meta.Metadata["mode"] == "" {
+			meta.Metadata["mode"] = "leader"
+		}
+	}
+	if !runOpts.Deadline.IsZero() {
+		// RFC3339Nano：秒级 RFC3339 会截断亚秒精度，短截止会被立刻判超期。
+		meta.Metadata["deadline"] = runOpts.Deadline.UTC().Format(time.RFC3339Nano)
+	}
+	if runOpts.MaxReplans > 0 {
+		meta.Metadata["max_replans"] = fmt.Sprintf("%d", runOpts.MaxReplans)
+	}
+	if runOpts.GroupChatPolicy != "" {
+		meta.Metadata["group_chat_policy"] = runOpts.GroupChatPolicy
+	}
+	if runOpts.GroupChatModerator != "" {
+		meta.Metadata["group_chat_moderator"] = runOpts.GroupChatModerator
+	}
+	if runOpts.GroupChatMaxTurns > 0 {
+		meta.Metadata["group_chat_max_turns"] = fmt.Sprintf("%d", runOpts.GroupChatMaxTurns)
+	}
+	if runOpts.ReviewVerifyRole != "" {
+		meta.Metadata["review_verify_role"] = runOpts.ReviewVerifyRole
+	}
+	if runOpts.ReviewTarget != "" {
+		meta.Metadata["review_target"] = runOpts.ReviewTarget
+	}
+	if runOpts.TaskBoardLeaseSeconds > 0 {
+		meta.Metadata["task_board_lease_seconds"] = fmt.Sprintf("%d", runOpts.TaskBoardLeaseSeconds)
+	}
 	e.deps.Sink.Emit(Event{Type: EvRunStarted, RunID: runID, Text: goal.Text, At: time.Now()})
 
 	// 意图闸门（可选，入口一次裁决；显式 TaskKey 优先，不重复过闸）。
@@ -294,6 +354,10 @@ func (e *SquadEngine) planAndExecute(ctx context.Context, st *RunState, runOpts
 		planOut = &PlanOutput{Plan: plan}
 	} else {
 		bpMatch, _ := e.opts.Blueprints.Match(meta.Goal.Text, runOpts.TaskKey)
+		// 拓扑解析（Q2）：RunOptions 优先，Blueprint 兜底。
+		if meta.Metadata["topology"] == "" && bpMatch.Topology != "" {
+			meta.Metadata["topology"] = bpMatch.Topology
+		}
 		planOut, err = e.opts.Planner.Plan(ctx, PlanInput{
 			Goal: meta.Goal, Blueprint: bpPtr(bpMatch), Registry: e.opts.Registry, RunID: meta.ID,
 			IntentCatalog: e.opts.IntentCatalog,
@@ -380,6 +444,20 @@ func (e *SquadEngine) planAndExecute(ctx context.Context, st *RunState, runOpts
 	if err := e.materializeWorkspace(meta, st, planOut.Plan); err != nil {
 		return nil, err
 	}
+	// 拓扑校验与分发（Q2）：未注册名 / primitive 单跑 / driver 未点亮 fail-closed；
+	// leader 拓扑映射到既有 Mode 委派循环（不新造执行面）。
+	if topo := meta.Metadata["topology"]; topo != "" {
+		kind, terr := checkTopology(topo)
+		if terr != nil {
+			return e.failNew(meta, terr)
+		}
+		if kind == TopologyMode && topo == TopologyLeader {
+			if meta.Metadata["mode"] == "" {
+				meta.Metadata["mode"] = "leader"
+			}
+		}
+		e.deps.Sink.Emit(Event{Type: EvPhase, RunID: meta.ID, Text: "拓扑 " + topo, At: time.Now()})
+	}
 	// leader-delegation 模式：忽略规划步骤，进 leader 委派循环。
 	if meta.Metadata["mode"] == "leader" {
 		e.initLeaderState(st)
@@ -392,6 +470,15 @@ func (e *SquadEngine) planAndExecute(ctx context.Context, st *RunState, runOpts
 		}
 		return out, nil
 	}
+	// driver 类拓扑（Q2 WP3+）：按名分发专用驱动（review/group_chat；
+	// task_board 属 WP4，未点亮前 checkTopology 已 fail-closed）。
+	if out, handled, derr := e.dispatchTopologyDriver(ctx, st); handled {
+		if derr != nil {
+			return nil, derr
+		}
+		e.finalizeEvent(out)
+		return out, nil
+	}
 	out, _, err := drv.advance(ctx, st)
 	if err != nil {
 		return nil, err
@@ -414,7 +501,13 @@ func (e *SquadEngine) initLeaderState(st *RunState) {
 			maxTurns = n
 		}
 	}
-	st.Leader = &LeaderState{LeaderRole: leaderRole, MaxTurns: maxTurns}
+	maxReplans := DefaultMaxReplans
+	if v := st.Meta.Metadata["max_replans"]; v != "" {
+		if n := parseIntDefault(v, DefaultMaxReplans); n > 0 {
+			maxReplans = n
+		}
+	}
+	st.Leader = &LeaderState{LeaderRole: leaderRole, MaxTurns: maxTurns, MaxReplans: maxReplans}
 	st.saveLeader()
 }
 
@@ -446,6 +539,15 @@ func (e *SquadEngine) resumeAfterHold(ctx context.Context, st *RunState) (*RunRe
 		}
 		return out, nil
 	}
+	// driver 类拓扑续跑：驱动自身幂等（review 已完成 finder 不重跑；
+	// group_chat 按已落盘轮次续转）。
+	if out, handled, err := e.dispatchTopologyDriver(ctx, st); handled {
+		if err != nil {
+			return nil, err
+		}
+		e.finalizeEvent(out)
+		return out, nil
+	}
 	drv := &flowDriver{opts: e.deps}
 	out, _, err := drv.advance(ctx, st)
 	if err != nil {
@@ -455,6 +557,32 @@ func (e *SquadEngine) resumeAfterHold(ctx context.Context, st *RunState) (*RunRe
 	return out, nil
 }
 
+// dispatchTopologyDriver 按运行拓扑分发 driver 类拓扑。
+// handled=false 表示非 driver 拓扑（走既有图/模式路径）。
+func (e *SquadEngine) dispatchTopologyDriver(ctx context.Context, st *RunState) (*RunResult, bool, error) {
+	topo := st.Meta.Metadata["topology"]
+	if topo == "" {
+		return nil, false, nil
+	}
+	spec, ok := LookupTopology(topo)
+	if !ok || spec.Kind != TopologyDriver {
+		return nil, false, nil
+	}
+	switch topo {
+	case TopologyReview:
+		out, err := e.runReviewDriver(ctx, st)
+		return out, true, err
+	case TopologyGroupChat:
+		out, err := e.runGroupChatDriver(ctx, st)
+		return out, true, err
+	case TopologyTaskBoard:
+		out, err := e.runTaskBoardDriver(ctx, st)
+		return out, true, err
+	default:
+		return e.driverFail(st, fmt.Sprintf("topology: %q 驱动未实现（fail-closed）", topo)), true, nil
+	}
+}
+
 // materializeWorkspace 落盘工作区与子工作区（已存在则只更新计划并重建待执行子目录）。
 func (e *SquadEngine) materializeWorkspace(meta *RunMetadata, st *RunState, plan *SquadPlan) error {
 	if st.Meta.WorkDir == "" {
@@ -537,6 +665,17 @@ func (e *SquadEngine) respondLocked(ctx context.Context, runID string, decision
 	}
 
 	req := &gate.Request
+	// 非 human actor 不得决议 choice 门（Q2 验收三的第二道防线：邮箱/代理
+	// 消息即便伪造 Respond 调用也清不掉 permission/confirmation/quality 门）。
+	if decision.Actor != "" && decision.Actor != ActorHuman && choiceHoldKind(req.Kind) &&
+		strings.TrimSpace(decision.Choice) != "" {
+		e.deps.Sink.Emit(Event{Type: EvPhase, RunID: runID,
+			Text: fmt.Sprintf("拒绝非人工 actor(%s) 决议 %s 门（请求 %s）", decision.Actor, req.Kind, req.ID), At: time.Now()})
+		return e.noteAndKeepHold(st, gate, RespondDecision{
+			RequestID: decision.RequestID,
+			Answer:    fmt.Sprintf("[blocked] 非人工 actor(%s) 尝试决议本门，已拒绝；等待人工裁决。", decision.Actor),
+		})
+	}
 	if strings.TrimSpace(decision.Choice) == "" && strings.TrimSpace(decision.Answer) != "" && choiceHoldKind(req.Kind) {
 		return e.noteAndKeepHold(st, gate, decision)
 	}

squad/graph.go

@@ -71,6 +71,45 @@ func compileGraph(steps []StepSpec) (*graph, error) {
 	if cycle := g.findCycle(); cycle != nil {
 		return nil, fmt.Errorf("graph: needs 依赖成环：%s", strings.Join(cycle, " → "))
 	}
+	// 循环步与路由步校验（Q2）：预算必须显式，引用必须存在，两者互斥
+	//（同一步既循环又分支的语义未定义，fail-closed）。
+	for _, s := range steps {
+		if s.Loop != nil && len(s.Branches) > 0 {
+			return nil, fmt.Errorf("graph: 步骤 %s 不得同时声明 loop 与 branches", s.ID)
+		}
+		if s.Loop != nil {
+			if s.Loop.MaxIterations <= 0 {
+				return nil, fmt.Errorf("graph: 步骤 %s 的 loop.max_iterations 必须 > 0", s.ID)
+			}
+			for _, b := range s.Loop.Body {
+				if !ids[b] {
+					return nil, fmt.Errorf("graph: 步骤 %s 的 loop.body 引用未知步骤 %s", s.ID, b)
+				}
+				if b == s.ID {
+					return nil, fmt.Errorf("graph: 步骤 %s 的 loop.body 引用自身", s.ID)
+				}
+			}
+		}
+		for _, br := range s.Branches {
+			if br.When.Field == "" {
+				return nil, fmt.Errorf("graph: 步骤 %s 的分支缺少 when.field", s.ID)
+			}
+			if br.When.Equals == "" && br.When.NotEquals == "" {
+				return nil, fmt.Errorf("graph: 步骤 %s 的分支 when 需要 equals 或 not_equals", s.ID)
+			}
+			if len(br.Then) == 0 {
+				return nil, fmt.Errorf("graph: 步骤 %s 的分支 then 为空", s.ID)
+			}
+			for _, t := range br.Then {
+				if !ids[t] {
+					return nil, fmt.Errorf("graph: 步骤 %s 的分支 then 引用未知步骤 %s", s.ID, t)
+				}
+				if t == s.ID {
+					return nil, fmt.Errorf("graph: 步骤 %s 的分支 then 引用自身", s.ID)
+				}
+			}
+		}
+	}
 	return g, nil
 }
 

squad/groupchat.go

@@ -0,0 +1,355 @@
+package squad
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+	"time"
+)
+
+// groupchat.go — group_chat 拓扑驱动（Q2 WP3）。
+// 形态（计划 §三 WP3 与边界 §五）：共享转录（BB 字段 transcript +
+// group_chat/transcript.md 全量落盘）；发言人策略只做**轮转**与**主持人**
+// 两种（脚本可判定，LLM 选人策略不进库）；预算停机（max_turns，缺省 8）。
+// 蓝图步骤 = 参与者名册（按序，不直接执行，driver 标记 skipped 并派生轮次
+// 步骤）；任一发言含 [close] 标记即收口；主持人模式下主持人输出 JSON 决策
+// （{"next":"role"} / {"action":"close"}），未知角色/非法 JSON fail-closed。
+// 驱动幂等可恢复：轮次由已落盘的 gc 步骤数重建，挂起（权限/HITL）后
+// Respond 重入续转。
+
+func init() {
+	_ = RegisterTopologyDriver(TopologyGroupChat)
+}
+
+// DefaultMaxGroupChatTurns 群聊轮次预算缺省值（预算停机）。
+const DefaultMaxGroupChatTurns = 8
+
+// gcRosterSkipReason 名册步骤的 skipped 原因（非执行步骤，不算降级）。
+const gcRosterSkipReason = "group_chat: 名册步骤（参与者由驱动接管）"
+
+// gcCloseMarker 任一发言包含该标记（大小写不敏感）即收口。
+const gcCloseMarker = "[close]"
+
+// gcTranscriptCapBB 是 BB 转录字段的容量上限（TASK.md 快照只带近端）。
+const gcTranscriptCapBB = 4000
+
+// runGroupChatDriver group_chat 拓扑入口。
+func (e *SquadEngine) runGroupChatDriver(ctx context.Context, st *RunState) (*RunResult, error) {
+	drv := &flowDriver{opts: e.deps}
+
+	// 名册：首次进入从蓝图步骤取参与者序列并持久化到元数据（恢复时直读）。
+	roster, err := e.gcRoster(st)
+	if err != nil {
+		return e.driverFail(st, err.Error()), nil
+	}
+	policy := orDefault(st.Meta.Metadata["group_chat_policy"], "rotate")
+	moderator := st.Meta.Metadata["group_chat_moderator"]
+	if policy != "rotate" && policy != "moderator" {
+		return e.driverFail(st, fmt.Sprintf("group_chat: 未知发言人策略 %q（只支持 rotate/moderator）", policy)), nil
+	}
+	if policy == "moderator" {
+		if moderator == "" || !containsString(roster, moderator) {
+			return e.driverFail(st, fmt.Sprintf("group_chat: 主持人 %q 不在名册内（fail-closed）", moderator)), nil
+		}
+	}
+	maxTurns := DefaultMaxGroupChatTurns
+	if v := st.Meta.Metadata["group_chat_max_turns"]; v != "" {
+		if n := parseIntDefault(v, DefaultMaxGroupChatTurns); n > 0 {
+			maxTurns = n
+		}
+	}
+
+	// 名册步骤标记 skipped（首次进入时；恢复重入无 pending 名册步骤）。
+	st.mu.Lock()
+	for i := range st.Plan.Workflow.Steps {
+		s := &st.Plan.Workflow.Steps[i]
+		if !strings.HasPrefix(s.ID, "gc") && (s.Status == "pending" || s.Status == "running") {
+			s.Status = "skipped"
+			s.Error = gcRosterSkipReason
+		}
+	}
+	_ = st.save()
+	st.mu.Unlock()
+
+	for {
+		// 恢复重入：上次挂起留下的未完 gc 轮次步骤优先续跑（不重新选人）。
+		if idx := e.gcHeldStepIdx(st); idx >= 0 {
+			out, gate, err := e.gcRunStep(ctx, drv, st, idx)
+			if err != nil {
+				return nil, err
+			}
+			if gate != nil {
+				if out == nil {
+					out = heldResult(st, gate)
+				}
+				return out, nil
+			}
+			if out != nil {
+				return out, nil
+			}
+			if e.gcMessageClosed(st, idx) {
+				drv.emit(st, Event{Type: EvPhase, Text: fmt.Sprintf("group_chat：%s 发出收口标记，停机", st.Plan.Workflow.Steps[idx].Role)})
+				break
+			}
+			continue
+		}
+		turns := e.gcTurnCount(st)
+		if turns >= maxTurns {
+			drv.emit(st, Event{Type: EvPhase, Text: fmt.Sprintf("group_chat：轮次预算 %d 用尽，停机收口", maxTurns)})
+			break
+		}
+		speaker, closed, derr := e.gcNextSpeaker(st, roster, policy, moderator, turns)
+		if derr != nil {
+			return e.driverFail(st, derr.Error()), nil
+		}
+		if closed {
+			drv.emit(st, Event{Type: EvPhase, Text: "group_chat：主持人收口"})
+			break
+		}
+		out, gate, err := e.gcRunTurn(ctx, drv, st, turns, speaker)
+		if err != nil {
+			return nil, err
+		}
+		if gate != nil {
+			if out == nil {
+				out = heldResult(st, gate)
+			}
+			return out, nil
+		}
+		if out != nil { // 轮次步骤失败（关键）→ fail-closed 传播
+			return out, nil
+		}
+		if e.gcMessageClosed(st, len(st.Plan.Workflow.Steps)-1) {
+			drv.emit(st, Event{Type: EvPhase, Text: fmt.Sprintf("group_chat：%s 发出收口标记，停机", speaker)})
+			break
+		}
+	}
+	return e.gcFinish(st, drv), nil
+}
+
+// gcRoster 参与者名册（首次从蓝图步骤派生并写入元数据；恢复直读）。
+func (e *SquadEngine) gcRoster(st *RunState) ([]string, error) {
+	if v := st.Meta.Metadata["group_chat_roster"]; v != "" {
+		return strings.Split(v, ","), nil
+	}
+	var roster []string
+	for _, s := range st.Plan.Workflow.Steps {
+		if s.Role != "" && !containsString(roster, s.Role) {
+			roster = append(roster, s.Role)
+		}
+	}
+	if len(roster) == 0 {
+		return nil, fmt.Errorf("group_chat: 名册为空（蓝图步骤即参与者）")
+	}
+	st.mu.Lock()
+	if st.Meta.Metadata == nil {
+		st.Meta.Metadata = map[string]string{}
+	}
+	st.Meta.Metadata["group_chat_roster"] = strings.Join(roster, ",")
+	_ = st.save()
+	st.mu.Unlock()
+	return roster, nil
+}
+
+// gcTurnCount 已完成的轮次数（预算与轮转序号的基准；恢复重入自然续算）。
+func (e *SquadEngine) gcTurnCount(st *RunState) int {
+	n := 0
+	for _, s := range st.Plan.Workflow.Steps {
+		if strings.HasPrefix(s.ID, "gc") && s.Status == "completed" {
+			n++
+		}
+	}
+	return n
+}
+
+// gcHeldStepIdx 返回上次挂起留下的未完 gc 轮次步骤下标（无则 -1）。
+func (e *SquadEngine) gcHeldStepIdx(st *RunState) int {
+	for i := range st.Plan.Workflow.Steps {
+		s := &st.Plan.Workflow.Steps[i]
+		if strings.HasPrefix(s.ID, "gc") && (s.Status == "pending" || s.Status == "running") {
+			return i
+		}
+	}
+	return -1
+}
+
+// gcNextSpeaker 决定下一发言人。rotate = 轮转；moderator = 主持人裁决
+// （上一步是参与者发言 → 主持人接手；主持人决策 close → 收口，next → 指名，
+// 非法 JSON / 未知角色 fail-closed）。
+func (e *SquadEngine) gcNextSpeaker(st *RunState, roster []string, policy, moderator string, turns int) (string, bool, error) {
+	if policy == "rotate" {
+		return roster[turns%len(roster)], false, nil
+	}
+	last := e.gcLastStep(st)
+	if last == nil || last.Role == moderator {
+		// 首轮或上一步就是主持人（不该连续，防御性让主持人接手）。
+		if last != nil && last.Role == moderator {
+			// 解析主持人决策。
+			dec, err := e.gcParseDecision(st, last)
+			if err != nil {
+				return "", false, err
+			}
+			if dec.Action == "close" {
+				return "", true, nil
+			}
+			if !containsString(roster, dec.Next) {
+				return "", false, fmt.Errorf("group_chat: 主持人指名 %q 不在名册内（fail-closed）", dec.Next)
+			}
+			return dec.Next, false, nil
+		}
+		return moderator, false, nil
+	}
+	// 上一步是参与者发言 → 主持人接手裁决。
+	return moderator, false, nil
+}
+
+type gcDecision struct {
+	Action string `json:"action"`
+	Next   string `json:"next"`
+}
+
+// gcParseDecision 宽松解析主持人 message.md 里的 JSON 决策（首 `{` 末 `}`）。
+func (e *SquadEngine) gcParseDecision(st *RunState, step *StepSpec) (*gcDecision, error) {
+	msg := e.gcMessageOf(st, step)
+	start := strings.Index(msg, "{")
+	end := strings.LastIndex(msg, "}")
+	if start < 0 || end <= start {
+		return nil, fmt.Errorf("group_chat: 主持人输出无 JSON 决策（fail-closed）：%q", truncateStr(msg, 80))
+	}
+	var dec gcDecision
+	if err := json.Unmarshal([]byte(msg[start:end+1]), &dec); err != nil {
+		return nil, fmt.Errorf("group_chat: 主持人决策解析失败（fail-closed）: %v", err)
+	}
+	if dec.Action == "close" {
+		return &dec, nil
+	}
+	if strings.TrimSpace(dec.Next) == "" {
+		return nil, fmt.Errorf("group_chat: 主持人决策缺 next/action（fail-closed）")
+	}
+	return &dec, nil
+}
+
+// gcRunTurn 派生一轮发言步骤并执行。
+func (e *SquadEngine) gcRunTurn(ctx context.Context, drv *flowDriver, st *RunState, turns int, speaker string) (*RunResult, *HitlGate, error) {
+	step := StepSpec{
+		ID:           fmt.Sprintf("gc%02d", turns+1),
+		Role:         speaker,
+		Title:        fmt.Sprintf("群聊第 %d 轮：%s 发言（把发言写入 message.md；同意收口就写 [close]）", turns+1, speaker),
+		Targets:      []string{"message.md"},
+		SubWorkspace: fmt.Sprintf("group_chat/turn_%02d_%s", turns+1, speaker),
+		Status:       "pending",
+	}
+	st.mu.Lock()
+	st.Plan.Workflow.Steps = append(st.Plan.Workflow.Steps, step)
+	idx := len(st.Plan.Workflow.Steps) - 1
+	_ = st.save()
+	st.mu.Unlock()
+	return e.gcRunStep(ctx, drv, st, idx)
+}
+
+// gcRunStep 执行一个轮次步骤：过门 → runOneStep → 转录并入共享状态。
+// 挂起（gate）时步骤保持未完，恢复重入经 gcHeldStepIdx 续跑同一步骤。
+func (e *SquadEngine) gcRunStep(ctx context.Context, drv *flowDriver, st *RunState, idx int) (*RunResult, *HitlGate, error) {
+	out, gate, err := drv.gateBeforeRun(ctx, st, idx)
+	if err != nil || gate != nil || out != nil {
+		return out, gate, err
+	}
+	step := st.Plan.Workflow.Steps[idx]
+	drv.emit(st, Event{Type: EvPhase, StepID: step.ID, Role: step.Role,
+		Text: fmt.Sprintf("group_chat：%s 发言（第 %d 轮）", step.Role, e.gcTurnCount(st)+1)})
+	out, gate, err = drv.runOneStep(ctx, st, idx)
+	if err != nil || gate != nil || out != nil {
+		return out, gate, err
+	}
+	// 转录并入共享状态：全量落盘 transcript.md，近端进 BB 字段（下一位
+	// 发言人的 TASK.md 黑板快照可见）。
+	msg := e.gcMessageOf(st, &st.Plan.Workflow.Steps[idx])
+	e.gcAppendTranscript(st, e.gcTurnCount(st), step.Role, msg)
+	return nil, nil, nil
+}
+
+// gcAppendTranscript 追加转录（全量文件 + BB 近端快照）。
+func (e *SquadEngine) gcAppendTranscript(st *RunState, turn int, speaker, msg string) {
+	dir := filepath.Join(st.Meta.WorkDir, "group_chat")
+	_ = os.MkdirAll(dir, 0o755)
+	line := fmt.Sprintf("%d. %s：%s\n", turn, speaker, strings.TrimSpace(msg))
+	f, _ := os.OpenFile(filepath.Join(dir, "transcript.md"), os.O_APPEND|os.O_CREATE|os.O_WRONLY, 0o644)
+	if f != nil {
+		_, _ = f.WriteString(line)
+		_ = f.Close()
+	}
+	st.mu.Lock()
+	if st.BB.Fields == nil {
+		st.BB.Fields = map[string]string{}
+	}
+	cur := st.BB.Fields["transcript"] + line
+	if len(cur) > gcTranscriptCapBB {
+		cur = cur[len(cur)-gcTranscriptCapBB:]
+	}
+	st.BB.Fields["transcript"] = cur
+	_ = st.save()
+	st.mu.Unlock()
+}
+
+// gcLastStep 最后一个已完成的派生轮次步骤（主持人决策解析的锚点）。
+func (e *SquadEngine) gcLastStep(st *RunState) *StepSpec {
+	for i := len(st.Plan.Workflow.Steps) - 1; i >= 0; i-- {
+		s := &st.Plan.Workflow.Steps[i]
+		if strings.HasPrefix(s.ID, "gc") && s.Status == "completed" {
+			return s
+		}
+	}
+	return nil
+}
+
+// gcMessageOf 读取某轮的 message.md（缺文件 = 空串）。
+func (e *SquadEngine) gcMessageOf(st *RunState, step *StepSpec) string {
+	data, err := os.ReadFile(filepath.Join(st.Meta.WorkDir, filepath.FromSlash(step.SubWorkspace), "message.md"))
+	if err != nil {
+		return ""
+	}
+	return string(data)
+}
+
+// gcMessageClosed 第 idx 步的发言是否带收口标记。
+func (e *SquadEngine) gcMessageClosed(st *RunState, idx int) bool {
+	if idx < 0 || idx >= len(st.Plan.Workflow.Steps) {
+		return false
+	}
+	msg := e.gcMessageOf(st, &st.Plan.Workflow.Steps[idx])
+	return strings.Contains(strings.ToLower(msg), gcCloseMarker)
+}
+
+// gcFinish 收口：交付物 = 全量转录；写 output.md；状态落盘；构造结果。
+func (e *SquadEngine) gcFinish(st *RunState, drv *flowDriver) *RunResult {
+	transcript := ""
+	if data, err := os.ReadFile(filepath.Join(st.Meta.WorkDir, "group_chat", "transcript.md")); err == nil {
+		transcript = string(data)
+	}
+	report := "# 群聊转录\n\n" + transcript
+	_ = os.WriteFile(filepath.Join(st.Meta.WorkDir, "output.md"), []byte(report), 0o644)
+	turns := e.gcTurnCount(st)
+	st.mu.Lock()
+	wf := st.Workflow()
+	wf.Status = string(RunCompleted)
+	st.setWorkflow(wf)
+	st.Meta.Status = string(RunCompleted)
+	now := time.Now()
+	st.Meta.CompletedAt = &now
+	_ = st.save()
+	st.mu.Unlock()
+	summary := fmt.Sprintf("group_chat：%d 轮发言，转录已收口", turns)
+	drv.emit(st, Event{Type: EvPhase, Text: summary})
+	return &RunResult{RunID: st.Meta.ID, Status: RunCompleted, Plan: planPtr(st),
+		Deliverable: "output.md", Summary: summary, WorkDir: st.Meta.WorkDir}
+}
+
+func truncateStr(s string, n int) string {
+	if len(s) <= n {
+		return s
+	}
+	return s[:n] + "…"
+}

squad/handoff.go

@@ -0,0 +1,91 @@
+package squad
+
+import (
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+	"time"
+)
+
+// handoff.go — handoff 跨拓扑原语（Q2 WP3）。
+// 语义（计划 §三 WP3）：StepSpec.Handoff 非空的步骤完成后，最终答案所有权
+// 移交给目标角色——① 记录所有权转移（Handoffs，随 RunState 持久化）；
+// ② 目标角色的下一个步骤收到「过滤历史」：只带移交步骤的主产出，不带全量
+// 转录（一次性并入任务书，与 mailbox 同款机制）；③ finalize 汇总交付物时
+// 取链尾（最后一次移交的目标角色）的最后一个完成步骤产出，而非全图最后
+// 完成步骤。目标角色不在小队内 = fail-closed（步骤启动前拒绝）。
+
+// HandoffRecord 是一次所有权转移记录。
+type HandoffRecord struct {
+	FromRole  string    `json:"from_role"`
+	FromStep  string    `json:"from_step"`
+	ToRole    string    `json:"to_role"`
+	Output    string    `json:"output,omitempty"` // 移交步骤主产出（截断上限内）
+	At        time.Time `json:"at"`
+	Delivered bool      `json:"delivered,omitempty"`
+}
+
+// handoffOutputCap 移交上下文的产出截断上限（过滤历史：只带主产出，且限量）。
+const handoffOutputCap = 64 * 1024
+
+// takeHandoffFor 取出移交给某角色的未投递记录，标记已投递（一次性）。
+// 调用方负责随后 st.save()（或在临界区内）。
+func takeHandoffFor(st *RunState, role string) *HandoffRecord {
+	for i := range st.Handoffs {
+		h := &st.Handoffs[i]
+		if h.Delivered || h.ToRole != role {
+			continue
+		}
+		h.Delivered = true
+		out := *h
+		return &out
+	}
+	return nil
+}
+
+// formatHandoffSection 渲染任务书里的移交上下文段（无则空串）。
+func formatHandoffSection(h *HandoffRecord) string {
+	if h == nil {
+		return ""
+	}
+	body := h.Output
+	if len(body) > handoffOutputCap {
+		body = body[:handoffOutputCap] + "\n…（已截断）"
+	}
+	var b strings.Builder
+	fmt.Fprintf(&b, "\n\n## 移交上下文（所有权来自 %s，步骤 %s）\n", h.FromRole, h.FromStep)
+	b.WriteString("以下是移交方的主产出（过滤历史：只含该产出，不含其过程转录）：\n\n")
+	b.WriteString(body)
+	b.WriteString("\n")
+	return b.String()
+}
+
+// handoffChainTail 返回移交链尾角色（无移交记录返回空串）。
+func handoffChainTail(st *RunState) string {
+	if n := len(st.Handoffs); n > 0 {
+		return st.Handoffs[n-1].ToRole
+	}
+	return ""
+}
+
+// recordHandoff 步骤完成时的所有权转移记录：主产出快照（截断上限内）+
+// 持久化 + 事件。目标角色合法性已在步骤启动前 fail-closed 校验。
+func (d *flowDriver) recordHandoff(st *RunState, step StepSpec, subDir string, targets []string) {
+	output := ""
+	if data, err := os.ReadFile(filepath.Join(subDir, firstTarget(targets))); err == nil {
+		output = string(data)
+	}
+	if len(output) > handoffOutputCap {
+		output = output[:handoffOutputCap] + "\n…（已截断）"
+	}
+	st.mu.Lock()
+	st.Handoffs = append(st.Handoffs, HandoffRecord{
+		FromRole: step.Role, FromStep: step.ID, ToRole: step.Handoff,
+		Output: output, At: time.Now(),
+	})
+	_ = st.save()
+	st.mu.Unlock()
+	d.emit(st, Event{Type: EvPhase, StepID: step.ID, Role: step.Role,
+		Text: fmt.Sprintf("handoff：%s → %s（最终答案所有权移交）", step.Role, step.Handoff)})
+}

squad/hitl.go

@@ -105,8 +105,19 @@ type RespondDecision struct {
 	Choice string `json:"choice,omitempty"`
 	// EditPlan 确认计划时的修改指令（附加在 Answer 上传递给引擎）。
 	EditPlan string `json:"edit_plan,omitempty"`
+	// Actor 决议发起方（Q2）："" 或 human = 人工（默认，兼容既有调用方）；
+	// agent/system = 非人工。非人工 actor 的 choice 决议在
+	// permission/confirmation/quality_review 门上一律被拒绝（验收三）。
+	Actor string `json:"actor,omitempty"`
 }
 
+// Actor 归属常量。
+const (
+	ActorHuman  = "human"
+	ActorAgent  = "agent"
+	ActorSystem = "system"
+)
+
 // newRequestID 生成请求 ID。
 func newRequestID() string {
 	b := make([]byte, 8)

squad/leader.go

@@ -47,12 +47,34 @@ func leaderExecRequest(drv *flowDriver, st *RunState, leader RoleSpec, workDir,
 // DefaultMaxLeaderTurns leader 模式每运行的 leader 轮次上限。
 const DefaultMaxLeaderTurns = 8
 
+// DefaultMaxReplans leader 卡顿触发的重规划上限（Q2；与轮次上限独立）。
+const DefaultMaxReplans = 3
+
+// DefaultStuckThreshold 连续无进展轮数达到该值触发一次重规划（Q2）。
+const DefaultStuckThreshold = 2
+
 // LeaderState leader 循环状态（workspace leader.json 持久化，重启可恢复）。
 type LeaderState struct {
 	Turns       int                `json:"turns"`
 	MaxTurns    int                `json:"max_turns"`
 	LeaderRole  string             `json:"leader_role"`
 	Delegations []DelegationRecord `json:"delegations"`
+	// 进度账本（Q2）：每轮委派后记一条进展评估；卡顿检测与重规划的凭据。
+	Progress []ProgressEntry `json:"progress,omitempty"`
+	// StuckTurns 连续无进展轮数（有进展即清零）。
+	StuckTurns int `json:"stuck_turns,omitempty"`
+	// Replans 已触发的重规划次数；MaxReplans 触顶降级收口。
+	Replans    int `json:"replans,omitempty"`
+	MaxReplans int `json:"max_replans,omitempty"`
+	// ReplanNote 一次性重规划指令（并入下一轮 leader 任务书后清空）。
+	ReplanNote string `json:"replan_note,omitempty"`
+}
+
+// ProgressEntry 是进度账本的一条评估记录。
+type ProgressEntry struct {
+	Turn     int    `json:"turn"`
+	Signal   string `json:"signal"` // progress | stuck
+	Detail   string `json:"detail,omitempty"`
 }
 
 // DelegationRecord 一次已执行的委派（历史注入下一轮 leader 上下文）。
@@ -134,7 +156,7 @@ func leaderPrompt(members []RoleSpec, catalogMd string) string {
 }
 
 // leaderTaskBrief 组装 leader 本轮的 TASK.md（目标 + 黑板 + 委派历史）。
-func leaderTaskBrief(st *RunState) string {
+func leaderTaskBrief(st *RunState, mail []MailMessage) string {
 	var b strings.Builder
 	b.WriteString("# Leader 轮次：评估并委派\n\n")
 	fmt.Fprintf(&b, "## 运行目标\n%s\n\n", st.Meta.Goal.Text)
@@ -171,6 +193,21 @@ func leaderTaskBrief(st *RunState) string {
 	} else {
 		b.WriteString("（尚无委派，这是第一轮。）\n\n")
 	}
+	if len(st.Leader.Progress) > 0 {
+		b.WriteString("## 进度账本\n")
+		for _, p := range st.Leader.Progress {
+			fmt.Fprintf(&b, "- 第%d轮 [%s] %s\n", p.Turn, p.Signal, p.Detail)
+		}
+		fmt.Fprintf(&b, "- 已重规划 %d/%d 次\n\n", st.Leader.Replans, st.Leader.MaxReplans)
+	}
+	if st.Leader.ReplanNote != "" {
+		fmt.Fprintf(&b, "%s\n\n", st.Leader.ReplanNote)
+	}
+
+	if section := formatMailSection(mail); section != "" {
+		b.WriteString(strings.TrimPrefix(section, "\n\n"))
+		b.WriteString("\n")
+	}
 	b.WriteString("## 本轮产出要求\n\n写 delegation.json（结构见你的指令）到本子目录。\n")
 	return b.String()
 }
@@ -233,9 +270,20 @@ func (e *SquadEngine) runLeaderLoop(ctx context.Context, st *RunState) (*RunResu
 		}
 
 		// 每轮重写 leader 上下文（黑板/委派历史已变化）。
-		if err := os.WriteFile(filepath.Join(subDir, "TASK.md"), []byte(leaderTaskBrief(st)), 0o644); err != nil {
+		// 邮箱消息与重规划指令是一次性的：并入本轮任务书后清空并落盘。
+		st.mu.Lock()
+		leaderMail := takeMailFor(st, st.Leader.LeaderRole)
+		st.mu.Unlock()
+		if err := os.WriteFile(filepath.Join(subDir, "TASK.md"), []byte(leaderTaskBrief(st, leaderMail)), 0o644); err != nil {
 			return nil, nil, err
 		}
+		if len(leaderMail) > 0 {
+			_ = st.save()
+		}
+		if st.Leader.ReplanNote != "" {
+			st.Leader.ReplanNote = ""
+			st.saveLeader()
+		}
 		if err := os.WriteFile(filepath.Join(subDir, "AGENTS.md"), []byte(leaderPrompt(members, e.opts.IntentCatalog.intentCatalogMd())), 0o644); err != nil {
 			return nil, nil, err
 		}
@@ -249,7 +297,7 @@ func (e *SquadEngine) runLeaderLoop(ctx context.Context, st *RunState) (*RunResu
 		if err != nil {
 			return e.leaderFail(st, err.Error()), nil, nil
 		}
-		prompt := readFileOr(filepath.Join(subDir, "TASK.md"), leaderTaskBrief(st))
+		prompt := readFileOr(filepath.Join(subDir, "TASK.md"), leaderTaskBrief(st, nil))
 		prompt = st.injectSteering(prompt)
 		instructions := readFileOr(filepath.Join(subDir, "AGENTS.md"), leaderPrompt(members, e.opts.IntentCatalog.intentCatalogMd()))
 		result, execErr := backend.Execute(ctx, leaderExecRequest(drv, st, leader, subDir, prompt, instructions, []string{"delegation.json"}, e.deps.StepTimeout))
@@ -338,6 +386,7 @@ func (e *SquadEngine) runLeaderLoop(ctx context.Context, st *RunState) (*RunResu
 				memberBrief += "\n\n## Leader 补充指令\n" + dec.Briefing
 				_ = os.WriteFile(filepath.Join(st.Meta.WorkDir, filepath.FromSlash(step.SubWorkspace), "TASK.md"), []byte(memberBrief), 0o644)
 			}
+			factsBefore := len(st.BB.Facts)
 			out, gate, err := drv.runOneStep(ctx, st, idx)
 			if err != nil {
 				return nil, nil, err
@@ -355,6 +404,50 @@ func (e *SquadEngine) runLeaderLoop(ctx context.Context, st *RunState) (*RunResu
 				return out, gate, nil // 挂起（gap/权限等），Respond 后重入循环
 			}
 			st.Leader.Delegations = append(st.Leader.Delegations, rec)
+			// 进度账本（Q2）：委派完成且有新事实、或不是与上轮同题重复 → 有进展。
+			factsGrew := len(st.BB.Facts) > factsBefore
+			sameAsPrev := false
+			if n := len(st.Leader.Delegations); n >= 2 {
+				prev := st.Leader.Delegations[n-2]
+				sameAsPrev = prev.Member == rec.Member && prev.Title == rec.Title
+			}
+			if factsGrew || !sameAsPrev {
+				st.Leader.StuckTurns = 0
+				st.Leader.Progress = append(st.Leader.Progress, ProgressEntry{
+					Turn: rec.Turn, Signal: "progress",
+					Detail: fmt.Sprintf("%s：%s（新事实 %v）", rec.Member, rec.Title, factsGrew),
+				})
+			} else {
+				st.Leader.StuckTurns++
+				st.Leader.Progress = append(st.Leader.Progress, ProgressEntry{
+					Turn: rec.Turn, Signal: "stuck",
+					Detail: fmt.Sprintf("与上轮同题重复且无新事实（连续 %d 轮）", st.Leader.StuckTurns),
+				})
+			}
+			// 卡顿 → 重规划；重规划触顶 → 降级收口（与轮次上限同款 Q23 语义）。
+			if st.Leader.StuckTurns >= DefaultStuckThreshold {
+				st.Leader.StuckTurns = 0
+				st.Leader.Replans++
+				if st.Leader.Replans >= st.Leader.MaxReplans {
+					st.saveLeader()
+					drv.emit(st, Event{Type: EvPhase, Text: fmt.Sprintf(
+						"重规划达上限 %d，按现有产出降级收口（Q2）", st.Leader.MaxReplans)})
+					out2, gate2, err2 := drv.finalize(ctx, st)
+					if err2 != nil {
+						return nil, nil, err2
+					}
+					if gate2 != nil {
+						return heldResult(st, gate2), gate2, nil
+					}
+					return out2, nil, nil
+				}
+				st.Leader.ReplanNote = fmt.Sprintf(
+					"【重规划指令】连续 %d 轮无进展（同题重复且无新事实），已触发第 %d/%d 次重规划。"+
+						"本轮必须改变策略：换成员、把任务拆小、补充缺失输入，或 escalate 问人；不得原样重发上一次委派。",
+					DefaultStuckThreshold, st.Leader.Replans, st.Leader.MaxReplans)
+				drv.emit(st, Event{Type: EvPhase, Text: fmt.Sprintf(
+					"卡顿检测：触发重规划 %d/%d", st.Leader.Replans, st.Leader.MaxReplans)})
+			}
 			st.saveLeader()
 			// 成员交付快照进 leader inbox（leader 工具沙箱看不到成员目录，
 			// 须物化到 leader 工作区才能自查交付——InputLayoutSnapshot 的对称应用）。

squad/leader_mailbox_conformance_test.go

@@ -0,0 +1,356 @@
+package squad
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+// leader_mailbox_conformance_test.go — Q2 WP2 验收：leader 进度账本/重规划
+// 上限、mailbox 投递、验收三「mailbox 消息不能把挂起的权限批准掉」。
+
+// ---- leader 进度账本与重规划 ----
+
+// 同题重复委派且无新事实 → 卡顿计数 → 触发重规划（事件+账本+一次性指令），
+// leader 收到指令后 close，运行完成且 Replans=1。
+func TestLeaderLedgerStuckTriggersReplan(t *testing.T) {
+	b := &leaderScriptBackend{
+		decisions: []string{
+			delegateJSON("writer", "同一件事"), // 第1轮：有进展（首委派）
+			delegateJSON("writer", "同一件事"), // 第2轮：同题无新事实 → stuck 1
+			delegateJSON("writer", "同一件事"), // 第3轮：stuck 2 → 重规划
+			closeJSON, //                        第4轮：收到重规划指令后收口
+		},
+		script: map[string]map[string]string{
+			"01_writer": {"output.md": goodReport},
+			"02_writer": {"output.md": goodReport},
+			"03_writer": {"output.md": goodReport},
+		},
+		prompts: map[string][]string{},
+	}
+	sink := &collectSink{}
+	engine := newLeaderEngineWithSink(t, b, sink)
+	res, err := engine.Run(context.Background(), Goal{Text: "写一份报告"}, RunOptions{Mode: "leader"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunCompleted {
+		t.Fatalf("status = %s err = %v", res.Status, res.Error)
+	}
+	st, err := engine.loadRun(engine.runDir(res.RunID))
+	if err != nil {
+		t.Fatal(err)
+	}
+	st.loadLeader()
+	if st.Leader.Replans != 1 {
+		t.Fatalf("Replans = %d，期望 1", st.Leader.Replans)
+	}
+	var stuck, progress int
+	for _, p := range st.Leader.Progress {
+		switch p.Signal {
+		case "stuck":
+			stuck++
+		case "progress":
+			progress++
+		}
+	}
+	if stuck != 2 || progress != 1 {
+		t.Fatalf("账本 progress=%d stuck=%d，期望 1/2", progress, stuck)
+	}
+	if sink.countText("触发重规划 1/3") != 1 {
+		t.Fatal("缺重规划事件")
+	}
+	// 重规划指令一次性并入下一轮 leader 任务书。
+	var sawNote bool
+	for _, p := range b.prompts["00_leader"] {
+		if strings.Contains(p, "重规划指令") {
+			sawNote = true
+		}
+	}
+	if !sawNote {
+		t.Fatal("leader 未收到重规划指令")
+	}
+}
+
+// 重规划触顶（MaxReplans=1）→ 走 finalize 收口路径（与 Q23 轮次上限同款：
+// 不裸失败；终态取决于既有产出），并发出触顶事件、账本记账、不再进 leader 轮。
+func TestLeaderReplanCapFinalizes(t *testing.T) {
+	b := &leaderScriptBackend{
+		decisions: []string{
+			delegateJSON("writer", "同一件事"),
+			delegateJSON("writer", "同一件事"),
+			delegateJSON("writer", "同一件事"), // stuck 2 → Replans=1 >= cap → 收口
+			closeJSON,                          // 不应到达
+		},
+		script: map[string]map[string]string{
+			"01_writer": {"output.md": goodReport},
+			"02_writer": {"output.md": goodReport},
+			"03_writer": {"output.md": goodReport},
+		},
+	}
+	sink := &collectSink{}
+	engine := newLeaderEngineWithSink(t, b, sink)
+	res, err := engine.Run(context.Background(), Goal{Text: "写一份报告"}, RunOptions{Mode: "leader", MaxReplans: 1})
+	if err != nil {
+		t.Fatal(err)
+	}
+	switch res.Status {
+	case RunCompleted, RunDegraded, RunQualityReview:
+	default:
+		t.Fatalf("重规划触顶应走 finalize 收口而非 %s（err=%v）", res.Status, res.Error)
+	}
+	if sink.countText("重规划达上限 1") != 1 {
+		t.Fatal("缺触顶事件")
+	}
+	st, err := engine.loadRun(engine.runDir(res.RunID))
+	if err != nil {
+		t.Fatal(err)
+	}
+	st.loadLeader()
+	if st.Leader.Replans != 1 || st.Leader.MaxReplans != 1 {
+		t.Fatalf("Replans=%d MaxReplans=%d，期望 1/1", st.Leader.Replans, st.Leader.MaxReplans)
+	}
+	if b.turn > 3 {
+		t.Fatalf("触顶后不应再进 leader 轮（turn=%d）", b.turn)
+	}
+}
+
+func newLeaderEngineWithSink(t *testing.T, backend Executor, sink ProgressSink) *SquadEngine {
+	t.Helper()
+	dir := t.TempDir()
+	reg := NewRoleRegistry()
+	for _, r := range []RoleSpec{
+		{Role: "leader", Name: "主管", Backend: "leader-script", Skills: []string{"协调"}},
+		{Role: "writer", Name: "撰写", Backend: "leader-script", Skills: []string{"撰写"}},
+	} {
+		if err := reg.Register(r); err != nil {
+			t.Fatal(err)
+		}
+	}
+	backends := NewBackendResolver()
+	backends.RegisterInstance("leader-script", backend)
+	engine, err := NewEngine(Options{DataDir: dir, Registry: reg, Backends: backends, Sink: sink})
+	if err != nil {
+		t.Fatal(err)
+	}
+	return engine
+}
+
+// ---- mailbox：投递 + 验收三 ----
+
+// 挂起（gap）期间投递 → Respond 恢复 → 下一步骤任务书含消息；广播 "*" 同样到达；
+// 已执行完的步骤不被追投；定向消息只投递一次。
+func TestMailboxDeliveredAfterHoldIntoNextStep(t *testing.T) {
+	rec := &promptRecordingBackend{prompts: map[string][]string{}}
+	engine := newRecordingEngine(t, rec, gapSteps())
+	res, err := engine.Run(context.Background(), Goal{Text: "带缺口"}, RunOptions{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunWaitingHuman || res.Pending == nil {
+		t.Fatalf("应挂起等人工：%s", res.Status)
+	}
+	// 挂起期间投递（run 锁空闲）。
+	if err := engine.PostMessage(context.Background(), res.RunID, MailMessage{
+		From: "researcher", To: "writer", Text: "重点写第二节",
+	}); err != nil {
+		t.Fatal(err)
+	}
+	if err := engine.PostMessage(context.Background(), res.RunID, MailMessage{
+		From: "leader", To: "*", Text: "全体注意格式",
+	}); err != nil {
+		t.Fatal(err)
+	}
+	// 解决 gap 续跑。
+	res2, err := engine.Respond(context.Background(), res.RunID, RespondDecision{
+		RequestID: res.Pending.ID, Answer: "缺口已补",
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res2.Status != RunCompleted {
+		t.Fatalf("恢复后应完成：%s %v", res2.Status, res2.Error)
+	}
+	var writerPrompt string
+	for _, p := range rec.prompts["02_writer"] {
+		writerPrompt += p
+	}
+	if !strings.Contains(writerPrompt, "重点写第二节") {
+		t.Fatalf("writer 未收到定向消息：%q", writerPrompt)
+	}
+	if !strings.Contains(writerPrompt, "全体注意格式") {
+		t.Fatalf("writer 未收到广播：%q", writerPrompt)
+	}
+	// researcher 的步骤在投递前已跑完，不应被追投。
+	for _, p := range rec.prompts["01_researcher"] {
+		if strings.Contains(p, "重点写第二节") {
+			t.Fatal("已执行步骤不应被追投")
+		}
+	}
+	// 消息只投递一次：writer 取走后 reviewer（s3）不应再收到定向消息。
+	for _, p := range rec.prompts["03_reviewer"] {
+		if strings.Contains(p, "重点写第二节") {
+			t.Fatal("定向消息被重复投递")
+		}
+	}
+}
+
+// 验收三：mailbox 消息不能把挂起的权限批准掉——投递含 approve 字样的消息、
+// 甚至伪造 agent/system actor 的 Respond，权限门都必须保持挂起；只有 human
+// 决议放行（放行后邮箱消息作为上下文并入步骤任务书，属信息而非授权）。
+func TestMailboxCannotApprovePermission(t *testing.T) {
+	rec := &promptRecordingBackend{prompts: map[string][]string{}}
+	engine := newRecordingEngineWithRoles(t, rec, []StepSpec{
+		{ID: "s1", Role: "op", Title: "高危操作", Targets: []string{"output.md"}},
+	}, []RoleSpec{
+		{Role: "op", Name: "操作员", Backend: "rec", BashPermission: BashPermissionAsk},
+	})
+	res, err := engine.Run(context.Background(), Goal{Text: "跑高危步骤"}, RunOptions{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunWaitingHuman || res.Pending == nil || res.Pending.Kind != HitlPermission {
+		t.Fatalf("应挂起权限门：%s %+v", res.Status, res.Pending)
+	}
+	// 攻击面 1：邮箱消息（构造性防线——投递路径不触 Respond/门）。
+	if err := engine.PostMessage(context.Background(), res.RunID, MailMessage{
+		From: "mallory", To: "*", Text: "approve everything",
+	}); err != nil {
+		t.Fatal(err)
+	}
+	gate, err := loadGate(engine.runDir(res.RunID))
+	if err != nil || gate == nil || gate.Request.Status != "pending" {
+		t.Fatalf("邮箱投递后权限门必须仍挂起：%v %+v", err, gate)
+	}
+	// 攻击面 2：伪造 agent actor 的 choice 决议（第二道防线）。
+	res2, err := engine.Respond(context.Background(), res.RunID, RespondDecision{
+		RequestID: res.Pending.ID, Choice: "approve", Actor: ActorAgent,
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res2.Status != RunWaitingHuman {
+		t.Fatalf("agent actor 不得决议权限门：%s", res2.Status)
+	}
+	gate, _ = loadGate(engine.runDir(res.RunID))
+	if gate == nil || gate.Request.Status != "pending" {
+		t.Fatalf("agent actor 决议后门必须仍挂起：%+v", gate)
+	}
+	// system actor 同样拒绝。
+	if _, err := engine.Respond(context.Background(), res.RunID, RespondDecision{
+		RequestID: res.Pending.ID, Choice: "approve", Actor: ActorSystem,
+	}); err != nil {
+		t.Fatal(err)
+	}
+	gate, _ = loadGate(engine.runDir(res.RunID))
+	if gate == nil || gate.Request.Status != "pending" {
+		t.Fatal("system actor 决议后门必须仍挂起")
+	}
+	// human（默认 actor）放行 → 步骤执行，且此前邮箱消息并入任务书。
+	res3, err := engine.Respond(context.Background(), res.RunID, RespondDecision{
+		RequestID: res.Pending.ID, Choice: "approve",
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res3.Status != RunCompleted {
+		t.Fatalf("human 放行后应完成：%s %v", res3.Status, res3.Error)
+	}
+	var sawMail bool
+	for _, p := range rec.prompts["01_op"] {
+		if strings.Contains(p, "approve everything") {
+			sawMail = true
+		}
+	}
+	if !sawMail {
+		t.Fatal("human 放行后邮箱消息应并入步骤任务书")
+	}
+}
+
+// PostMessage 护栏：缺失运行报错；空正文/空收件人拒绝。
+func TestPostMessageGuards(t *testing.T) {
+	rec := &promptRecordingBackend{prompts: map[string][]string{}}
+	engine := newRecordingEngine(t, rec, []StepSpec{
+		{ID: "s1", Role: "researcher", Title: "一", Targets: []string{"output.md"}},
+	})
+	if err := engine.PostMessage(context.Background(), "run_missing", MailMessage{To: "x", Text: "y"}); err == nil {
+		t.Fatal("缺失运行必须报错")
+	}
+	res, err := engine.Run(context.Background(), Goal{Text: "x"}, RunOptions{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunCompleted {
+		t.Fatalf("status = %s", res.Status)
+	}
+	if err := engine.PostMessage(context.Background(), res.RunID, MailMessage{To: "researcher", Text: ""}); err == nil {
+		t.Fatal("空正文必须拒绝")
+	}
+	if err := engine.PostMessage(context.Background(), res.RunID, MailMessage{To: "", Text: "x"}); err == nil {
+		t.Fatal("空收件人必须拒绝")
+	}
+}
+
+// ---- 夹具 ----
+
+// promptRecordingBackend 按目录记录 prompt；gaps.json 产出阻塞缺口触发挂起，
+// 其余目标写合格交付物。
+type promptRecordingBackend struct {
+	prompts map[string][]string
+}
+
+func (b *promptRecordingBackend) Name() string { return "rec" }
+
+func (b *promptRecordingBackend) Execute(_ context.Context, req ExecRequest) (*ExecResult, error) {
+	key := filepath.Base(filepath.Clean(req.WorkDir))
+	b.prompts[key] = append(b.prompts[key], req.Prompt)
+	patches := map[string][]byte{}
+	for _, t := range req.Targets {
+		content := goodReport
+		switch t {
+		case "gaps.json":
+			content = `[{"id":"g1","item":"缺一个输入","owner":"human","blocking":true}]`
+		case "fields.json":
+			content = `{}`
+		}
+		out := filepath.Join(req.WorkDir, t)
+		if err := os.MkdirAll(filepath.Dir(out), 0o755); err != nil {
+			return nil, err
+		}
+		if err := os.WriteFile(out, []byte(content), 0o644); err != nil {
+			return nil, err
+		}
+		patches[t] = []byte(content)
+	}
+	return &ExecResult{Success: true, Patches: patches}, nil
+}
+
+func newRecordingEngine(t *testing.T, be Executor, steps []StepSpec) *SquadEngine {
+	t.Helper()
+	return newRecordingEngineWithRoles(t, be, steps, []RoleSpec{
+		{Role: "researcher", Name: "调研", Backend: "rec"},
+		{Role: "writer", Name: "撰写", Backend: "rec"},
+		{Role: "reviewer", Name: "审核", Backend: "rec"},
+	})
+}
+
+func newRecordingEngineWithRoles(t *testing.T, be Executor, steps []StepSpec, roles []RoleSpec) *SquadEngine {
+	t.Helper()
+	dir := t.TempDir()
+	reg := NewRoleRegistry()
+	for _, r := range roles {
+		if err := reg.Register(r); err != nil {
+			t.Fatal(err)
+		}
+	}
+	backends := NewBackendResolver()
+	backends.RegisterInstance("rec", be)
+	engine, err := NewEngine(Options{DataDir: dir, Registry: reg,
+		Planner: mockScriptPlanner(steps...), Backends: backends, Sink: NopSink{}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	return engine
+}

squad/mailbox.go

@@ -0,0 +1,92 @@
+package squad
+
+import (
+	"context"
+	"fmt"
+	"strings"
+	"time"
+)
+
+// mailbox.go — 邮箱原语（Q2 WP2）：运行内点对点消息。
+// 安全边界（验收三的构造性保证）：投递路径永不经过 Respond/gate——
+// 消息只在步骤启动时并入任务书；挂起中的人工门（permission/confirmation/
+// quality_review）不可能被邮箱消息决议。RespondDecision.Actor 提供
+// 第二道防线：非 human actor 的 choice 决议一律拒绝。
+
+// MailMessage 是一条运行内消息。To 是角色 key，"*" = 广播给之后的每个步骤。
+type MailMessage struct {
+	ID        string    `json:"id"`
+	From      string    `json:"from"`
+	To        string    `json:"to"`
+	Text      string    `json:"text"`
+	At        time.Time `json:"at"`
+	Delivered bool      `json:"delivered,omitempty"`
+}
+
+// PostMessage 向一个运行投递邮箱消息。运行执行中（锁被占）返回 ErrRunBusy
+// ——挂起等待人工时与回合间隙可投递。消息随 RunState 持久化，重启不丢。
+// 本方法不触碰任何 HITL 门（验收三：邮箱永不决议挂起的审批）。
+func (e *SquadEngine) PostMessage(ctx context.Context, runID string, msg MailMessage) error {
+	if strings.TrimSpace(msg.Text) == "" {
+		return fmt.Errorf("mailbox: 消息正文为空")
+	}
+	if strings.TrimSpace(msg.To) == "" {
+		return fmt.Errorf("mailbox: 收件人必填（角色 key 或 *）")
+	}
+	release, err := e.acquire(ctx, runID)
+	if err != nil {
+		return err
+	}
+	defer release()
+	wsDir := e.runDir(runID)
+	st, err := e.loadRun(wsDir)
+	if err != nil {
+		return fmt.Errorf("load run %s: %w", runID, err)
+	}
+	if msg.ID == "" {
+		msg.ID = newRequestID()
+	}
+	if msg.At.IsZero() {
+		msg.At = time.Now()
+	}
+	st.mu.Lock()
+	st.Mailbox = append(st.Mailbox, msg)
+	err = st.save()
+	st.mu.Unlock()
+	if err != nil {
+		return err
+	}
+	e.deps.Sink.Emit(Event{Type: EvPhase, RunID: runID, Text: fmt.Sprintf("邮箱消息 %s → %s", msg.From, msg.To), At: time.Now()})
+	return nil
+}
+
+// takeMailFor 取出发给某角色的未投递消息（含广播），标记已投递。
+// 调用方负责随后 st.save()（或在临界区内）。
+func takeMailFor(st *RunState, role string) []MailMessage {
+	var out []MailMessage
+	for i := range st.Mailbox {
+		m := &st.Mailbox[i]
+		if m.Delivered {
+			continue
+		}
+		if m.To != role && m.To != "*" {
+			continue
+		}
+		m.Delivered = true
+		out = append(out, *m)
+	}
+	return out
+}
+
+// formatMailSection 渲染任务书里的邮箱段（空则返回空串）。
+func formatMailSection(msgs []MailMessage) string {
+	if len(msgs) == 0 {
+		return ""
+	}
+	var b strings.Builder
+	b.WriteString("\n\n## 邮箱消息\n")
+	for _, m := range msgs {
+		fmt.Fprintf(&b, "- [%s] %s：%s\n", m.At.Format(time.RFC3339), m.From, m.Text)
+	}
+	return b.String()
+}

squad/models.go

@@ -66,9 +66,12 @@ type Blueprint struct {
 	Title       string     `json:"title"`
 	Match       []string   `json:"match,omitempty"` // 关键词命中即视为命中模板
 	Description string     `json:"description,omitempty"`
-	Agents      []RoleSpec `json:"agents,omitempty"` // 默认编制（引用已注册角色或内联角色）
-	Steps       []StepSpec `json:"steps"`
-	Template    string     `json:"template,omitempty"` // 交付物骨架模板（Markdown）
+	// Topology 声明模板的拓扑（Q2；空 = standard 图行为）。
+	// RunOptions.Topology 优先于本字段。
+	Topology string     `json:"topology,omitempty"`
+	Agents   []RoleSpec `json:"agents,omitempty"` // 默认编制（引用已注册角色或内联角色）
+	Steps    []StepSpec `json:"steps"`
+	Template string     `json:"template,omitempty"` // 交付物骨架模板（Markdown）
 }
 
 // PlanningMode 规划模式。
@@ -115,6 +118,47 @@ type StepSpec struct {
 	Skills []string `json:"skills,omitempty"`
 	// MinScore 质检总分下限。0 = 不额外设门槛，只看 Checker.Passed。
 	MinScore int `json:"min_score,omitempty"`
+	// Loop 非空 = 循环步（Q2 loop 拓扑）：完成后条件不满足且预算未尽时
+	// 重置为 pending 再武装（图不允许环，循环由引擎再武装实现）。
+	Loop *LoopSpec `json:"loop,omitempty"`
+	// LoopRuns 已再武装次数（引擎维护，随计划持久化）。
+	LoopRuns int `json:"loop_runs,omitempty"`
+	// Branches 非空 = 路由步（Q2 router 拓扑）：完成后按黑板字段判定，
+	// 未选中分支的目标步骤置 skipped（skipped=依赖已满足，join 步照常推进）。
+	Branches []Branch `json:"branches,omitempty"`
+	// Handoff 非空 = 完成即把最终答案所有权移交给该角色（Q2 handoff 原语）。
+	Handoff string `json:"handoff,omitempty"`
+	// Brief 编译期任务指令（Q3 编译器）：非空时渲染进 TASK.md 的
+	// 「任务指令」节，位于标题/产出之后、上下文表之前。图步的富文本任务书
+	// 载体（此前只有 leader 委派有 briefing）。
+	Brief string `json:"brief,omitempty"`
+}
+
+// LoopSpec 是循环步的预算与停机条件（Q2）。
+type LoopSpec struct {
+	// MaxIterations 再武装上限（必须 >0；触顶即停，带 loop_exhausted 事件）。
+	MaxIterations int `json:"max_iterations"`
+	// UntilField 停机判据的黑板字段名；空 = 只按 MaxIterations 迭代。
+	UntilField string `json:"until_field,omitempty"`
+	// UntilEquals 停机值：Fields[UntilField] == UntilEquals 时停。
+	UntilEquals string `json:"until_equals,omitempty"`
+	// Body 与本步一起再武装的其他步骤 ID（循环体；如 evaluator-optimizer
+	// 的评估步）。空 = 只再武装本步。
+	Body []string `json:"body,omitempty"`
+}
+
+// Branch 是路由步的一条分支（Q2）。When 按声明序取首个命中。
+type Branch struct {
+	When BranchCond `json:"when"`
+	// Then 是本分支保留的步骤 ID（其余分支目标置 skipped）。
+	Then []string `json:"then"`
+}
+
+// BranchCond 是分支判据：黑板字段等值/不等值。
+type BranchCond struct {
+	Field     string `json:"field"`
+	Equals    string `json:"equals,omitempty"`
+	NotEquals string `json:"not_equals,omitempty"`
 }
 
 // IsCritical 报告步骤是否关键（nil 视为关键）。

squad/port.go

@@ -64,12 +64,17 @@ type ExecBudget struct {
 
 // ProcessEvent 是过程帧。
 type ProcessEvent struct {
-	Type             string
-	Text             string
-	Tool             string
-	Detail           string
-	PromptTokens     int
-	CompletionTokens int
+	Type                 string
+	Text                 string
+	Tool                 string
+	Detail               string
+	PromptTokens         int
+	CompletionTokens     int
+	// 以下在 Type=usage 时填写（N3 WP6：上一请求与缓存命中拆分）。
+	LastPromptTokens     int
+	LastCompletionTokens int
+	CacheReadTokens      int
+	CacheWriteTokens     int
 }
 
 // ExecResult 是一次步骤执行的终态。

squad/review.go

@@ -0,0 +1,366 @@
+package squad
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"os"
+	"path/filepath"
+	"sort"
+	"strings"
+	"sync"
+	"time"
+)
+
+// review.go — review 拓扑驱动（Q2 WP3，验收二）。
+// 与 backend internal/runtime/review.go 的 runReview 等价：
+// 并行 N finder（semaphore 3）→ 宽松 JSON 解析（首 `[` 末 `]`，丢空 issue）
+// → 按 (location,issue) 去重、上限 10 → 每条独立 verifier（尾行
+// VERDICT: PASS/FAIL/PARTIAL）→ 只留 PASS（fail-closed）→ severity 稳定排序
+// → 报告恒带「已丢弃 N 条未验证」。解析/去重/判定/排序/文案与 backend 同一
+// 格式串（计划 §一.5：以报告字节形态为等价判据）。
+//
+// squad 侧接线：finder = 计划步骤（各自的 findings.json 为目标产物，经
+// runOneStep 走完整工作区/事件/HITL 机制）；verifier = 驱动内按 verify 角色
+// 直接 Execute 的核实调用（对应 backend 的 spawn(verifyProfileID,…)），
+// 每次核实落盘 reviews/verify_NN/ 留痕。
+
+func init() {
+	_ = RegisterTopologyDriver(TopologyReview)
+}
+
+// reviewFinding 与 backend reviewFinding 同构（JSON 键一致）。
+type reviewFinding struct {
+	Location string `json:"location"`
+	Severity string `json:"severity"`
+	Issue    string `json:"issue"`
+	Evidence string `json:"evidence"`
+}
+
+// reviewVerbatimTask 是核实调用的任务文案（与 backend runReview 同串）。
+const reviewVerbatimTask = "独立核实这条问题是否成立。材料如下：\n%s\n\n问题：%s\n只在结尾给出一行 VERDICT: PASS、FAIL 或 PARTIAL。"
+
+// parseFindings 宽松解析：首 `[` 末 `]`，丢空 issue（backend 同款）。
+func parseFindings(text string) []reviewFinding {
+	start := strings.Index(text, "[")
+	end := strings.LastIndex(text, "]")
+	if start < 0 || end <= start {
+		return nil
+	}
+	var out []reviewFinding
+	if err := json.Unmarshal([]byte(text[start:end+1]), &out); err != nil {
+		return nil
+	}
+	var kept []reviewFinding
+	for _, f := range out {
+		if strings.TrimSpace(f.Issue) == "" {
+			continue
+		}
+		kept = append(kept, f)
+	}
+	return kept
+}
+
+// dedupeFindings 按 (location,issue) 去重，上限 limit（backend 同款）。
+func dedupeFindings(groups [][]reviewFinding, limit int) []reviewFinding {
+	seen := map[string]bool{}
+	var out []reviewFinding
+	for _, group := range groups {
+		for _, f := range group {
+			key := strings.TrimSpace(f.Location) + "\n" + strings.TrimSpace(f.Issue)
+			if seen[key] {
+				continue
+			}
+			seen[key] = true
+			out = append(out, f)
+			if len(out) >= limit {
+				return out
+			}
+		}
+	}
+	return out
+}
+
+// reviewVerdict 尾行 VERDICT 判定（backend 同款：PASS/FAIL/PARTIAL，
+// 解析不出 = 空串 = 不通过）。
+func reviewVerdict(text string) string {
+	lines := strings.Split(text, "\n")
+	for i := len(lines) - 1; i >= 0; i-- {
+		line := strings.TrimSpace(lines[i])
+		if line == "" || !strings.HasPrefix(strings.ToUpper(line), "VERDICT:") {
+			continue
+		}
+		v := strings.TrimSpace(line[len("VERDICT:"):])
+		v = strings.ToUpper(v)
+		switch {
+		case strings.HasPrefix(v, "PASS"):
+			return "PASS"
+		case strings.HasPrefix(v, "FAIL"):
+			return "FAIL"
+		case strings.HasPrefix(v, "PARTIAL"):
+			return "PARTIAL"
+		default:
+			return ""
+		}
+	}
+	return ""
+}
+
+// severityRank 稳定排序键（backend 同款）。
+func severityRank(s string) int {
+	switch strings.ToLower(strings.TrimSpace(s)) {
+	case "critical":
+		return 0
+	case "high":
+		return 1
+	case "medium":
+		return 2
+	case "low":
+		return 3
+	default:
+		return 4
+	}
+}
+
+// reviewParallel semaphore-3 并行（backend parallel 同款）。
+func reviewParallel(n int, fn func(i int)) {
+	if n <= 0 {
+		return
+	}
+	sem := make(chan struct{}, 3)
+	var wg sync.WaitGroup
+	for i := 0; i < n; i++ {
+		wg.Add(1)
+		go func(i int) {
+			defer wg.Done()
+			sem <- struct{}{}
+			defer func() { <-sem }()
+			fn(i)
+		}(i)
+	}
+	wg.Wait()
+}
+
+// runReviewDriver review 拓扑入口。finder 阶段复用计划步骤（可挂起可恢复：
+// 重入时已完成步骤不重跑，findings.json 从各自子工作区读取）。
+func (e *SquadEngine) runReviewDriver(ctx context.Context, st *RunState) (*RunResult, error) {
+	drv := &flowDriver{opts: e.deps}
+	steps := st.Plan.Workflow.Steps
+
+	// verify 角色先行解析（fail-fast：缺角色不浪费 finder 轮次）。
+	verifyRoleName := orDefault(st.Meta.Metadata["review_verify_role"], "verifier")
+	verifyRole, ok := findAgent(st.Plan.Squad.Agents, verifyRoleName)
+	if !ok {
+		if r, exists := e.deps.Registry.Get(verifyRoleName); exists {
+			verifyRole, ok = r, true
+		}
+	}
+	if !ok {
+		return e.driverFail(st, fmt.Sprintf("review: verify 角色 %q 不在小队内（fail-closed）", verifyRoleName)), nil
+	}
+	verifyBackend, _, rerr := e.deps.Backends.Resolve(verifyRole, st.Meta.Backend)
+	if rerr != nil {
+		return e.driverFail(st, "review: verify 后端解析失败: "+rerr.Error()), nil
+	}
+
+	// 1) finder 阶段：先全组过门（任一要挂起则整组不执行，与 advance 并行组
+	// 同款语义），再 semaphore-3 并行执行。
+	var finderIdx []int
+	for i := range steps {
+		if steps[i].Status == "pending" || steps[i].Status == "running" {
+			finderIdx = append(finderIdx, i)
+		}
+	}
+	for _, idx := range finderIdx {
+		out, gate, err := drv.gateBeforeRun(ctx, st, idx)
+		if err != nil {
+			return nil, err
+		}
+		if gate != nil {
+			if out == nil {
+				out = heldResult(st, gate)
+			}
+			return out, nil
+		}
+		if out != nil {
+			return out, nil
+		}
+	}
+	drv.emit(st, Event{Type: EvPhase, Text: fmt.Sprintf("review：%d 个 finder 并行找问题", len(finderIdx))})
+	var holdGate *HitlGate
+	var holdOut *RunResult
+	var fatal *RunResult
+	finderFailed := 0
+	var fmu sync.Mutex
+	reviewParallel(len(finderIdx), func(k int) {
+		idx := finderIdx[k]
+		out, gate, err := drv.runOneStep(ctx, st, idx)
+		fmu.Lock()
+		defer fmu.Unlock()
+		if err != nil {
+			if fatal == nil {
+				fatal = &RunResult{RunID: st.Meta.ID, Status: RunFailed, Plan: planPtr(st),
+					Error: err.Error(), WorkDir: st.Meta.WorkDir}
+			}
+			return
+		}
+		if gate != nil && holdGate == nil {
+			holdGate = gate
+			holdOut = out
+			return
+		}
+		if out != nil && out.Status == RunFailed && fatal == nil {
+			fatal = out // 关键 finder 失败：fail-closed 传播
+			return
+		}
+		if st.Plan.Workflow.Steps[idx].Status == "failed" {
+			finderFailed++ // 非关键 finder 失败：宽容（backend spawn err 同款），收口降级
+		}
+	})
+	if fatal != nil {
+		return fatal, nil
+	}
+	if holdGate != nil {
+		if holdOut == nil {
+			holdOut = heldResult(st, holdGate)
+		}
+		return holdOut, nil
+	}
+
+	// 2) 收集各 finder 的 findings.json（宽松解析；缺文件/坏 JSON = 该 finder
+	// 零发现，与 backend 同款）。
+	steps = st.Plan.Workflow.Steps
+	groups := make([][]reviewFinding, 0, len(steps))
+	for i := range steps {
+		if steps[i].Status != "completed" {
+			continue
+		}
+		subDir := filepath.Join(st.Meta.WorkDir, filepath.FromSlash(steps[i].SubWorkspace))
+		data, err := os.ReadFile(filepath.Join(subDir, "findings.json"))
+		if err != nil {
+			continue
+		}
+		groups = append(groups, parseFindings(string(data)))
+	}
+	unique := dedupeFindings(groups, 10)
+	drv.emit(st, Event{Type: EvPhase, Text: fmt.Sprintf("review：去重后 %d 条，逐条独立核实", len(unique))})
+
+	// 3) 逐条 verify（semaphore-3 并行；fail-closed：核实调用失败/无判定 = 丢弃）。
+	target := reviewTargetMaterial(st)
+	kept := make([]reviewFinding, len(unique))
+	pass := make([]bool, len(unique))
+	reviewParallel(len(unique), func(i int) {
+		f := unique[i]
+		body, _ := json.Marshal(f)
+		task := fmt.Sprintf(reviewVerbatimTask, target, body)
+		vDir := filepath.Join(st.Meta.WorkDir, "reviews", fmt.Sprintf("verify_%02d", i+1))
+		if err := os.MkdirAll(vDir, 0o755); err != nil {
+			return
+		}
+		instructions := ""
+		if e.deps.Workspaces != nil {
+			instructions = e.deps.Workspaces.ResolveInstructions(verifyRole)
+		}
+		res, err := verifyBackend.Execute(ctx, ExecRequest{
+			WorkDir:      vDir,
+			Prompt:       task,
+			Instructions: instructions,
+			Targets:      []string{"verdict.md"},
+			Timeout:      e.deps.StepTimeout,
+			SessionScope: st.Meta.ID + "/verify_" + fmt.Sprintf("%02d", i+1),
+		})
+		if err != nil || res == nil {
+			return
+		}
+		text := string(res.Patches["verdict.md"]) + "\n" + res.Logs
+		if data, rerr2 := os.ReadFile(filepath.Join(vDir, "verdict.md")); rerr2 == nil {
+			text = string(data) + "\n" + text
+		}
+		if reviewVerdict(text) == "PASS" {
+			kept[i] = f
+			pass[i] = true
+		}
+	})
+	var verified []reviewFinding
+	for i, ok := range pass {
+		if ok {
+			verified = append(verified, kept[i])
+		}
+	}
+
+	// 4) severity 稳定排序 + 等价报告（backend 同格式串）。
+	sort.SliceStable(verified, func(i, j int) bool {
+		return severityRank(verified[i].Severity) < severityRank(verified[j].Severity)
+	})
+	dropped := len(unique) - len(verified)
+	var b strings.Builder
+	if len(verified) == 0 {
+		b.WriteString("没有验证通过的问题。\n")
+	}
+	for _, f := range verified {
+		fmt.Fprintf(&b, "- [%s] %s：%s\n  证据：%s\n",
+			orDefault(f.Severity, "medium"), orDefault(f.Location, "未定位"), f.Issue, f.Evidence)
+	}
+	fmt.Fprintf(&b, "已丢弃 %d 条未验证\n", dropped)
+	report := b.String()
+
+	// 5) 收口：报告落盘（reviews/report.md + output.md 交付物约定），
+	// finder 有失败则降级（诚实信号），否则完成。
+	reviewsDir := filepath.Join(st.Meta.WorkDir, "reviews")
+	_ = os.MkdirAll(reviewsDir, 0o755)
+	_ = os.WriteFile(filepath.Join(reviewsDir, "report.md"), []byte(report), 0o644)
+	_ = os.WriteFile(filepath.Join(st.Meta.WorkDir, "output.md"), []byte(report), 0o644)
+	st.mu.Lock()
+	if st.BB.Fields == nil {
+		st.BB.Fields = map[string]string{}
+	}
+	st.BB.Fields["review_verified"] = fmt.Sprintf("%d", len(verified))
+	st.BB.Fields["review_dropped"] = fmt.Sprintf("%d", dropped)
+	wf := st.Workflow()
+	status := RunCompleted
+	if finderFailed > 0 {
+		status = RunDegraded
+	}
+	wf.Status = string(status)
+	st.setWorkflow(wf)
+	st.Meta.Status = string(status)
+	now := time.Now()
+	st.Meta.CompletedAt = &now
+	_ = st.save()
+	st.mu.Unlock()
+	summary := fmt.Sprintf("review：保留 %d 条 / 丢弃 %d 条未验证", len(verified), dropped)
+	if finderFailed > 0 {
+		summary += fmt.Sprintf("（%d 个 finder 失败，降级）", finderFailed)
+	}
+	drv.emit(st, Event{Type: EvPhase, Text: summary})
+	return &RunResult{RunID: st.Meta.ID, Status: status, Plan: planPtr(st),
+		Deliverable: "output.md", Summary: summary, WorkDir: st.Meta.WorkDir}, nil
+}
+
+// reviewTargetMaterial 取被评审材料：元数据 review_target 优先，
+// 其次运行目标（context/goal.md）。
+func reviewTargetMaterial(st *RunState) string {
+	if v := strings.TrimSpace(st.Meta.Metadata["review_target"]); v != "" {
+		return v
+	}
+	if data, err := os.ReadFile(filepath.Join(st.Meta.WorkDir, "context", "goal.md")); err == nil {
+		return string(data)
+	}
+	return ""
+}
+
+// driverFail 驱动级 fail-closed 收口（与 leaderFail 同款语义）。
+func (e *SquadEngine) driverFail(st *RunState, reason string) *RunResult {
+	now := time.Now()
+	st.mu.Lock()
+	st.Meta.Status = string(RunFailed)
+	st.Meta.CompletedAt = &now
+	wf := st.Workflow()
+	wf.Status = "failed"
+	st.setWorkflow(wf)
+	_ = st.save()
+	st.mu.Unlock()
+	(&flowDriver{opts: e.deps}).emit(st, Event{Type: EvPhase, Text: reason})
+	return &RunResult{RunID: st.Meta.ID, Status: RunFailed, Plan: planPtr(st),
+		Error: reason, WorkDir: st.Meta.WorkDir}
+}

squad/runstate.go

@@ -27,6 +27,15 @@ type RunState struct {
 	Clarifications []string `json:"clarifications,omitempty"`
 	// Steering 运行中人工补充指令（在下一步骤边界注入后清空）。
 	Steering []string `json:"steering,omitempty"`
+	// Mailbox 运行内点对点消息（Q2）：随 RunState 持久化；只在步骤启动时
+	// 并入任务书，永不触碰 HITL 门。
+	Mailbox []MailMessage `json:"mailbox,omitempty"`
+	// Handoffs 所有权转移记录（Q2 handoff 原语）：随 RunState 持久化；
+	// 目标角色的下一步骤一次性接收移交上下文，finalize 取链尾产出。
+	Handoffs []HandoffRecord `json:"handoffs,omitempty"`
+	// Board task_board 拓扑的队列状态（Q2）：claim/lease/complete 三原语，
+	// 随 RunState 持久化；孤儿租约在驱动重入时再武装。
+	Board []BoardTask `json:"board,omitempty"`
 	// Usage 按角色累计的 token 用量。
 	Usage map[string]*Usage `json:"usage,omitempty"`
 	// Leader leader-delegation 模式的循环状态（workspace 内 leader.json 持久化）。
@@ -92,6 +101,9 @@ func (st *RunState) persist() error {
 		RepairBudgetRemaining: st.RepairBudgetRemaining,
 		RepairBudgetGranted:   st.RepairBudgetGranted,
 		Clarifications:        st.Clarifications,
+		Mailbox:               st.Mailbox,
+		Handoffs:              st.Handoffs,
+		Board:                 st.Board,
 	})
 }
 
@@ -100,6 +112,9 @@ type diskRunState struct {
 	RepairBudgetRemaining int             `json:"repair_budget_remaining,omitempty"`
 	RepairBudgetGranted   int             `json:"repair_budget_granted,omitempty"`
 	Clarifications        []string        `json:"clarifications,omitempty"`
+	Mailbox               []MailMessage   `json:"mailbox,omitempty"`
+	Handoffs              []HandoffRecord `json:"handoffs,omitempty"`
+	Board                 []BoardTask     `json:"board,omitempty"`
 }
 
 // Workflow 返回计划的编排清单（简洁访问器）。
@@ -154,6 +169,9 @@ func loadRunState(wsDir string) (*RunState, error) {
 		st.RepairBudgetRemaining = disk.RepairBudgetRemaining
 		st.RepairBudgetGranted = disk.RepairBudgetGranted
 		st.Clarifications = disk.Clarifications
+		st.Mailbox = disk.Mailbox
+		st.Handoffs = disk.Handoffs
+		st.Board = disk.Board
 	} else if !os.IsNotExist(err) {
 		return nil, err
 	}

squad/taskboard.go

@@ -0,0 +1,308 @@
+package squad
+
+import (
+	"context"
+	"fmt"
+	"strings"
+	"time"
+)
+
+// taskboard.go — task_board 拓扑驱动（Q2 WP4）。
+// 形态：计划步骤 = 板上任务（Needs = 依赖门）；执行语义是队列化的
+// claim/lease/complete 三原语——工作者认领任务（带租约时限）、执行、完成；
+// 进程死亡留下的孤儿租约（claimed 且租约过期）在驱动重入时再武装回 open
+// （Attempts 保留，超限由步骤级重试护栏兜底）。全部任务收口后走标准
+// finalize（交付物/质检/降级判定与图拓扑同一收口面）。
+// 挂起（HITL）安全：claimed 且步骤未完的任务在重入时续跑，不重复认领。
+
+func init() {
+	_ = RegisterTopologyDriver(TopologyTaskBoard)
+}
+
+// DefaultTaskLeaseSeconds 任务租约缺省时限（秒）。
+const DefaultTaskLeaseSeconds = 600
+
+// boardSkipReason 依赖门无法满足（上游 failed/skipped）时的跳过原因。
+const boardSkipReason = "task_board: 依赖未满足（上游失败或跳过）"
+
+// BoardTask 是任务板上一条任务的队列状态。
+type BoardTask struct {
+	ID        string `json:"id"`
+	Status    string `json:"status"` // open / claimed / done / failed / skipped
+	ClaimedBy string `json:"claimed_by,omitempty"`
+	// LeaseExpires 认领租约到期时刻；过期未 complete = 孤儿租约，重入再武装。
+	LeaseExpires time.Time `json:"lease_expires,omitempty"`
+	// Attempts 认领次数（孤儿再武装累加；执行重试由步骤级护栏负责）。
+	Attempts int `json:"attempts,omitempty"`
+}
+
+// runTaskBoardDriver task_board 拓扑入口（幂等可恢复）。
+func (e *SquadEngine) runTaskBoardDriver(ctx context.Context, st *RunState) (*RunResult, error) {
+	drv := &flowDriver{opts: e.deps}
+	e.initBoard(st)
+	e.rearmOrphanLeases(st)
+	leaseTTL := DefaultTaskLeaseSeconds
+	if v := st.Meta.Metadata["task_board_lease_seconds"]; v != "" {
+		if n := parseIntDefault(v, DefaultTaskLeaseSeconds); n > 0 {
+			leaseTTL = n
+		}
+	}
+
+	for {
+		idx, held := e.boardHeldTask(st)
+		if !held {
+			var ok bool
+			idx, ok = e.boardClaimNext(st, leaseTTL)
+			if !ok {
+				// 无可认领：把依赖已死（上游 failed/skipped）的 open 任务级联
+				// 置 skipped；再无变化则收口。
+				if !e.boardSkipDeadlocked(st) {
+					break
+				}
+				continue
+			}
+		}
+		out, gate, err := e.boardRunTask(ctx, drv, st, idx)
+		if err != nil {
+			return nil, err
+		}
+		if gate != nil {
+			if out == nil {
+				out = heldResult(st, gate)
+			}
+			return out, nil
+		}
+		if out != nil { // 关键任务失败：fail-closed 传播（failStep 已置运行失败）
+			return out, nil
+		}
+		e.boardSettleTask(st, idx)
+	}
+	return e.boardFinish(ctx, drv, st)
+}
+
+// initBoard 首次进入时从计划步骤派生任务板（恢复重入直读持久化状态）。
+func (e *SquadEngine) initBoard(st *RunState) {
+	st.mu.Lock()
+	defer st.mu.Unlock()
+	if len(st.Board) > 0 {
+		return
+	}
+	for _, s := range st.Plan.Workflow.Steps {
+		status := "open"
+		switch s.Status {
+		case "completed":
+			status = "done"
+		case "failed":
+			status = "failed"
+		case "skipped":
+			status = "skipped"
+		}
+		st.Board = append(st.Board, BoardTask{ID: s.ID, Status: status})
+	}
+	_ = st.save()
+}
+
+// rearmOrphanLeases 孤儿租约再武装：claimed 且租约过期 → open（进程死亡
+// 恢复路径；Attempts 保留作观测）。
+func (e *SquadEngine) rearmOrphanLeases(st *RunState) {
+	now := time.Now()
+	st.mu.Lock()
+	var rearmed []string
+	for i := range st.Board {
+		t := &st.Board[i]
+		if t.Status == "claimed" && !t.LeaseExpires.IsZero() && t.LeaseExpires.Before(now) {
+			t.Status = "open"
+			t.ClaimedBy = ""
+			rearmed = append(rearmed, t.ID)
+		}
+	}
+	changed := len(rearmed) > 0
+	if changed {
+		_ = st.save()
+	}
+	st.mu.Unlock()
+	if changed {
+		(&flowDriver{opts: e.deps}).emit(st, Event{Type: EvPhase,
+			Text: fmt.Sprintf("task_board：孤儿租约再武装 %s", strings.Join(rearmed, ","))})
+	}
+}
+
+// boardHeldTask 返回 claimed 且步骤未完的任务（HITL 挂起/中断重入的续跑点）。
+func (e *SquadEngine) boardHeldTask(st *RunState) (int, bool) {
+	for ti := range st.Board {
+		t := &st.Board[ti]
+		if t.Status != "claimed" {
+			continue
+		}
+		if idx, s := e.boardStep(st, t.ID); idx >= 0 && (s.Status == "pending" || s.Status == "running") {
+			return idx, true
+		}
+	}
+	return -1, false
+}
+
+// boardClaimNext 认领下一个可执行任务（open 且依赖门全 done），返回步骤下标。
+func (e *SquadEngine) boardClaimNext(st *RunState, leaseTTL int) (int, bool) {
+	st.mu.Lock()
+	defer st.mu.Unlock()
+	for ti := range st.Board {
+		t := &st.Board[ti]
+		if t.Status != "open" {
+			continue
+		}
+		if !e.boardDepsDoneLocked(st, t.ID) {
+			continue
+		}
+		idx, s := e.boardStepLocked(st, t.ID)
+		if idx < 0 {
+			continue
+		}
+		t.Status = "claimed"
+		t.ClaimedBy = s.Role
+		t.Attempts++
+		t.LeaseExpires = time.Now().Add(time.Duration(leaseTTL) * time.Second)
+		_ = st.save()
+		(&flowDriver{opts: e.deps}).emit(st, Event{Type: EvPhase, StepID: t.ID, Role: s.Role,
+			Text: fmt.Sprintf("task_board：%s 认领 %s（租约 %ds，第 %d 次）", s.Role, t.ID, leaseTTL, t.Attempts)})
+		return idx, true
+	}
+	return -1, false
+}
+
+// boardSkipDeadlocked 把依赖已死（上游 failed/skipped）的 open 任务级联置
+// skipped；返回是否有变化（有则继续循环级联）。
+func (e *SquadEngine) boardSkipDeadlocked(st *RunState) bool {
+	st.mu.Lock()
+	changed := false
+	for ti := range st.Board {
+		t := &st.Board[ti]
+		if t.Status != "open" {
+			continue
+		}
+		if e.boardDepsDeadLocked(st, t.ID) {
+			t.Status = "skipped"
+			if idx, s := e.boardStepLocked(st, t.ID); idx >= 0 {
+				s.Status = "skipped"
+				s.Error = boardSkipReason
+				st.Plan.Workflow.Steps[idx] = *s
+			}
+			changed = true
+		}
+	}
+	if changed {
+		_ = st.save()
+	}
+	st.mu.Unlock()
+	return changed
+}
+
+// boardRunTask 执行已认领任务：过门 → runOneStep。
+func (e *SquadEngine) boardRunTask(ctx context.Context, drv *flowDriver, st *RunState, idx int) (*RunResult, *HitlGate, error) {
+	out, gate, err := drv.gateBeforeRun(ctx, st, idx)
+	if err != nil || gate != nil || out != nil {
+		return out, gate, err
+	}
+	return drv.runOneStep(ctx, st, idx)
+}
+
+// boardSettleTask 任务结算：步骤 completed → done；failed（非关键降级）→ failed。
+func (e *SquadEngine) boardSettleTask(st *RunState, idx int) {
+	step := st.Plan.Workflow.Steps[idx]
+	status := "done"
+	if step.Status == "failed" {
+		status = "failed"
+	}
+	st.mu.Lock()
+	for i := range st.Board {
+		if st.Board[i].ID == step.ID {
+			st.Board[i].Status = status
+			st.Board[i].LeaseExpires = time.Time{}
+		}
+	}
+	_ = st.save()
+	st.mu.Unlock()
+	(&flowDriver{opts: e.deps}).emit(st, Event{Type: EvPhase, StepID: step.ID, Role: step.Role,
+		Text: fmt.Sprintf("task_board：%s → %s", step.ID, status)})
+}
+
+// boardFinish 全部任务收口后走标准 finalize（交付物/质检/降级与图拓扑同面）。
+// 仍有 open/claimed 任务 = 依赖不可满足，fail-closed 不静默收口。
+func (e *SquadEngine) boardFinish(ctx context.Context, drv *flowDriver, st *RunState) (*RunResult, error) {
+	done, total := 0, len(st.Board)
+	for _, t := range st.Board {
+		switch t.Status {
+		case "done":
+			done++
+		case "open", "claimed":
+			return e.driverFail(st, fmt.Sprintf("task_board: 任务 %s 无法收口（依赖不可满足）", t.ID)), nil
+		}
+	}
+	drv.emit(st, Event{Type: EvPhase, Text: fmt.Sprintf("task_board：收口（done %d/%d）", done, total)})
+	out, gate, err := drv.finalize(ctx, st)
+	if err != nil {
+		return nil, err
+	}
+	if gate != nil {
+		return heldResult(st, gate), nil
+	}
+	return out, nil
+}
+
+// boardStep 按任务 ID 找步骤（只读副本）。
+func (e *SquadEngine) boardStep(st *RunState, id string) (int, *StepSpec) {
+	st.mu.Lock()
+	defer st.mu.Unlock()
+	return e.boardStepLocked(st, id)
+}
+
+// boardStepLocked 按任务 ID 找步骤（调用方持锁）。返回下标与副本指针。
+func (e *SquadEngine) boardStepLocked(st *RunState, id string) (int, *StepSpec) {
+	for i := range st.Plan.Workflow.Steps {
+		if st.Plan.Workflow.Steps[i].ID == id {
+			s := st.Plan.Workflow.Steps[i]
+			return i, &s
+		}
+	}
+	return -1, nil
+}
+
+// boardDepsDoneLocked 依赖门：Needs 全部 done（调用方持锁）。
+func (e *SquadEngine) boardDepsDoneLocked(st *RunState, id string) bool {
+	needs := e.boardNeedsLocked(st, id)
+	for _, dep := range needs {
+		found := false
+		for i := range st.Board {
+			if st.Board[i].ID == dep {
+				found = true
+				if st.Board[i].Status != "done" {
+					return false
+				}
+			}
+		}
+		if !found {
+			return false // 依赖不在板上 = fail-closed 不认领
+		}
+	}
+	return true
+}
+
+// boardDepsDeadLocked 依赖已死：任一 Need 为 failed/skipped（级联跳过依据）。
+func (e *SquadEngine) boardDepsDeadLocked(st *RunState, id string) bool {
+	for _, dep := range e.boardNeedsLocked(st, id) {
+		for i := range st.Board {
+			if st.Board[i].ID == dep && (st.Board[i].Status == "failed" || st.Board[i].Status == "skipped") {
+				return true
+			}
+		}
+	}
+	return false
+}
+
+func (e *SquadEngine) boardNeedsLocked(st *RunState, id string) []string {
+	for i := range st.Plan.Workflow.Steps {
+		if st.Plan.Workflow.Steps[i].ID == id {
+			return st.Plan.Workflow.Steps[i].Needs
+		}
+	}
+	return nil
+}

squad/topology.go

@@ -0,0 +1,130 @@
+package squad
+
+import "fmt"
+
+// topology.go — 拓扑一等化（Q2）：内置拓扑名注册表。
+// 拓扑不新造执行面：graph 类编译为 Steps 图（Q1 机制），leader 走既有
+// Mode 委派循环，driver 类由引擎按名分发（review/group_chat/task_board），
+// primitive 类是跨拓扑原语（handoff/mailbox），不可单独作为运行拓扑。
+// 未知名与不可运行Kind一律 fail-closed。
+
+// 内置拓扑名（completion doc Q2 清单的 12 种）。
+const (
+	TopologySequential         = "sequential"
+	TopologyParallel           = "parallel"
+	TopologyLoop               = "loop"
+	TopologyRouter             = "router"
+	TopologyLeader             = "leader"
+	TopologyHandoff            = "handoff"
+	TopologyGroupChat          = "group_chat"
+	TopologyReview             = "review"
+	TopologyEvaluatorOptimizer = "evaluator_optimizer"
+	TopologyLongRunning        = "long_running"
+	TopologyTaskBoard          = "task_board"
+	TopologyMailbox            = "mailbox"
+)
+
+// TopologyKind 是拓扑的执行形态。
+type TopologyKind string
+
+const (
+	// TopologyGraph 编译为 Steps 图，由 flowDriver.advance 执行。
+	TopologyGraph TopologyKind = "graph"
+	// TopologyMode 走既有 Mode 分支（leader 委派循环）。
+	TopologyMode TopologyKind = "mode"
+	// TopologyDriver 由引擎按名分发到专用驱动。
+	TopologyDriver TopologyKind = "driver"
+	// TopologyPrimitive 是跨拓扑原语，不可单独运行。
+	TopologyPrimitive TopologyKind = "primitive"
+)
+
+// TopologySpec 是一个内置拓扑的注册项。
+type TopologySpec struct {
+	Name string
+	Kind TopologyKind
+}
+
+// builtinTopologies 是 12 种内置拓扑的注册表（唯一权威清单）。
+var builtinTopologies = []TopologySpec{
+	{Name: TopologySequential, Kind: TopologyGraph},
+	{Name: TopologyParallel, Kind: TopologyGraph},
+	{Name: TopologyLoop, Kind: TopologyGraph},
+	{Name: TopologyRouter, Kind: TopologyGraph},
+	{Name: TopologyLeader, Kind: TopologyMode},
+	{Name: TopologyHandoff, Kind: TopologyPrimitive},
+	{Name: TopologyGroupChat, Kind: TopologyDriver},
+	{Name: TopologyReview, Kind: TopologyDriver},
+	{Name: TopologyEvaluatorOptimizer, Kind: TopologyGraph},
+	{Name: TopologyLongRunning, Kind: TopologyGraph},
+	{Name: TopologyTaskBoard, Kind: TopologyDriver},
+	{Name: TopologyMailbox, Kind: TopologyPrimitive},
+}
+
+// BuiltinTopologies 返回内置拓扑清单（只读副本）。
+func BuiltinTopologies() []TopologySpec {
+	out := make([]TopologySpec, len(builtinTopologies))
+	copy(out, builtinTopologies)
+	return out
+}
+
+// LookupTopology 按名查注册项。
+func LookupTopology(name string) (TopologySpec, bool) {
+	for _, t := range builtinTopologies {
+		if t.Name == name {
+			return t, true
+		}
+	}
+	return TopologySpec{}, false
+}
+
+// registeredDrivers 是已接线的 driver 拓扑（WP 增量逐个点亮；未点亮 = fail-closed）。
+var registeredDrivers = map[string]bool{}
+
+// RegisterTopologyDriver 由驱动实现方在 init/装配时点亮一个 driver 拓扑。
+func RegisterTopologyDriver(name string) error {
+	spec, ok := LookupTopology(name)
+	if !ok {
+		return fmt.Errorf("topology: 未注册的拓扑名 %q", name)
+	}
+	if spec.Kind != TopologyDriver {
+		return fmt.Errorf("topology: %q 不是 driver 类（kind=%s）", name, spec.Kind)
+	}
+	registeredDrivers[name] = true
+	return nil
+}
+
+// checkTopology 校验一次运行的有效拓扑名，返回其 Kind。
+// 空名 = 未声明（standard 图行为，兼容既有调用方）。
+// 未注册名 / primitive 单跑 / driver 未点亮都 fail-closed。
+func checkTopology(name string) (TopologyKind, error) {
+	if name == "" {
+		return TopologyGraph, nil
+	}
+	spec, ok := LookupTopology(name)
+	if !ok {
+		return "", fmt.Errorf("topology: 未知拓扑 %q（内置清单：%s）", name, topologyNames())
+	}
+	switch spec.Kind {
+	case TopologyGraph, TopologyMode:
+		return spec.Kind, nil
+	case TopologyPrimitive:
+		return "", fmt.Errorf("topology: %q 是跨拓扑原语，不可单独作为运行拓扑", name)
+	case TopologyDriver:
+		if !registeredDrivers[name] {
+			return "", fmt.Errorf("topology: %q 驱动未注册（fail-closed）", name)
+		}
+		return spec.Kind, nil
+	}
+	return "", fmt.Errorf("topology: %q kind 非法", name)
+}
+
+func topologyNames() string {
+	s := ""
+	for i, t := range builtinTopologies {
+		if i > 0 {
+			s += ", "
+		}
+		s += t.Name
+	}
+	return s
+}

squad/topology_conformance_test.go

@@ -0,0 +1,435 @@
+package squad
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+)
+
+// topology_conformance_test.go — Q2 验收一的机器证据：每种拓扑一个
+// 脚本化后端的 conformance（零模型、零网络）。
+
+// ---- 脚本化夹具 ----
+
+// collectSink 收集事件流供断言。
+type collectSink struct {
+	mu     sync.Mutex
+	events []Event
+}
+
+func (c *collectSink) Emit(ev Event) {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	c.events = append(c.events, ev)
+}
+
+func (c *collectSink) snapshot() []Event {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	return append([]Event(nil), c.events...)
+}
+
+func (c *collectSink) countText(sub string) int {
+	n := 0
+	for _, ev := range c.snapshot() {
+		if strings.Contains(ev.Text, sub) {
+			n++
+		}
+	}
+	return n
+}
+
+// iterBackend 按「步骤子目录 × 第几次执行」回放内容（循环/评估拓扑需要
+// 逐轮不同的 fields.json）。
+type iterBackend struct {
+	mu      sync.Mutex
+	calls   map[string]int
+	content func(stepKey string, iter int, target string) []byte
+	delay   time.Duration
+}
+
+func (b *iterBackend) Name() string { return "iter" }
+
+func (b *iterBackend) Execute(_ context.Context, req ExecRequest) (*ExecResult, error) {
+	key := filepath.Base(filepath.Clean(req.WorkDir))
+	if b.delay > 0 {
+		time.Sleep(b.delay)
+	}
+	b.mu.Lock()
+	b.calls[key]++
+	iter := b.calls[key]
+	b.mu.Unlock()
+	patches := map[string][]byte{}
+	for _, t := range req.Targets {
+		var data []byte
+		if b.content != nil {
+			data = b.content(key, iter, t)
+		}
+		if data == nil {
+			data = []byte(goodReport)
+		}
+		out := filepath.Join(req.WorkDir, t)
+		_ = os.MkdirAll(filepath.Dir(out), 0o755)
+		if err := os.WriteFile(out, data, 0o644); err != nil {
+			return nil, err
+		}
+		patches[t] = data
+	}
+	return &ExecResult{Success: true, Patches: patches}, nil
+}
+
+func (b *iterBackend) totalCalls() int {
+	b.mu.Lock()
+	defer b.mu.Unlock()
+	n := 0
+	for _, v := range b.calls {
+		n += v
+	}
+	return n
+}
+
+func newTopologyEngine(t *testing.T, planner Planner, be Executor, sink ProgressSink) *SquadEngine {
+	t.Helper()
+	dir := t.TempDir()
+	reg := NewRoleRegistry()
+	for _, spec := range []RoleSpec{
+		{Role: "researcher", Name: "调研", Backend: "iter"},
+		{Role: "writer", Name: "撰写", Backend: "iter"},
+		{Role: "reviewer", Name: "审核", Backend: "iter"},
+	} {
+		if err := reg.Register(spec); err != nil {
+			t.Fatal(err)
+		}
+	}
+	backends := NewBackendResolver()
+	backends.RegisterInstance("iter", be)
+	if sink == nil {
+		sink = NopSink{}
+	}
+	engine, err := NewEngine(Options{
+		DataDir: dir, Registry: reg, Planner: planner, Backends: backends, Sink: sink,
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	return engine
+}
+
+func fieldsJSON(kv map[string]string) []byte {
+	b, _ := json.Marshal(kv)
+	return b
+}
+
+// ---- sequential / parallel：既有图语义 + 拓扑名受理 ----
+
+func TestTopologyConformance_sequential(t *testing.T) {
+	be := &iterBackend{calls: map[string]int{}}
+	sink := &collectSink{}
+	engine := newTopologyEngine(t, mockScriptPlanner(
+		StepSpec{ID: "s1", Role: "researcher", Title: "一", Targets: []string{"notes.md"}},
+		StepSpec{ID: "s2", Role: "writer", Title: "二", Targets: []string{"output.md"}},
+	), be, sink)
+	res, err := engine.Run(context.Background(), Goal{Text: "顺序任务"}, RunOptions{Topology: TopologySequential})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunCompleted {
+		t.Fatalf("status = %s err = %v", res.Status, res.Error)
+	}
+	if be.totalCalls() != 2 {
+		t.Fatalf("执行次数 = %d", be.totalCalls())
+	}
+	if sink.countText("拓扑 sequential") != 1 {
+		t.Fatal("缺拓扑受理事件")
+	}
+}
+
+func TestTopologyConformance_parallel(t *testing.T) {
+	be := &iterBackend{calls: map[string]int{}}
+	engine := newTopologyEngine(t, mockScriptPlanner(
+		StepSpec{ID: "s1", Role: "researcher", Title: "并行甲", Parallel: true, Targets: []string{"a.md"}},
+		StepSpec{ID: "s2", Role: "writer", Title: "并行乙", Parallel: true, Targets: []string{"b.md"}},
+		StepSpec{ID: "s3", Role: "reviewer", Title: "汇合", Targets: []string{"output.md"}},
+	), be, nil)
+	res, err := engine.Run(context.Background(), Goal{Text: "并行任务"}, RunOptions{Topology: TopologyParallel})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunCompleted || be.totalCalls() != 3 {
+		t.Fatalf("status = %s calls = %d err = %v", res.Status, be.totalCalls(), res.Error)
+	}
+}
+
+// ---- loop：条件停机 + 预算触顶 ----
+
+func TestTopologyConformance_loopUntilField(t *testing.T) {
+	// 第 1-2 轮 done=no，第 3 轮 done=yes → 恰执行 3 次。
+	be := &iterBackend{calls: map[string]int{}, content: func(_ string, iter int, target string) []byte {
+		if target == "fields.json" {
+			if iter >= 3 {
+				return fieldsJSON(map[string]string{"done": "yes"})
+			}
+			return fieldsJSON(map[string]string{"done": "no"})
+		}
+		return nil
+	}}
+	sink := &collectSink{}
+	engine := newTopologyEngine(t, mockScriptPlanner(
+		StepSpec{ID: "s1", Role: "researcher", Title: "循环体", Targets: []string{"output.md", "fields.json"},
+			Loop: &LoopSpec{MaxIterations: 5, UntilField: "done", UntilEquals: "yes"}},
+	), be, sink)
+	res, err := engine.Run(context.Background(), Goal{Text: "循环任务"}, RunOptions{Topology: TopologyLoop})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunCompleted {
+		t.Fatalf("status = %s err = %v", res.Status, res.Error)
+	}
+	if be.totalCalls() != 3 {
+		t.Fatalf("应在条件满足时停机（3 次），实际 %d", be.totalCalls())
+	}
+	if sink.countText("再武装") != 2 {
+		t.Fatalf("再武装事件应 2 次：%d", sink.countText("再武装"))
+	}
+	// 计划里的 LoopRuns 持久化为 2。
+	if len(res.Plan.Steps) != 1 || res.Plan.Steps[0].LoopRuns != 2 {
+		t.Fatalf("LoopRuns = %+v", res.Plan.Steps)
+	}
+}
+
+func TestTopologyConformance_loopBudgetExhausted(t *testing.T) {
+	// 条件永不满足：恰执行 MaxIterations 次后触顶（非失败）。
+	be := &iterBackend{calls: map[string]int{}, content: func(_ string, _ int, target string) []byte {
+		if target == "fields.json" {
+			return fieldsJSON(map[string]string{"done": "no"})
+		}
+		return nil
+	}}
+	sink := &collectSink{}
+	engine := newTopologyEngine(t, mockScriptPlanner(
+		StepSpec{ID: "s1", Role: "researcher", Title: "循环体", Targets: []string{"output.md", "fields.json"},
+			Loop: &LoopSpec{MaxIterations: 3, UntilField: "done", UntilEquals: "yes"}},
+	), be, sink)
+	res, err := engine.Run(context.Background(), Goal{Text: "循环任务"}, RunOptions{Topology: TopologyLoop})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunCompleted {
+		t.Fatalf("触顶应是完成而非失败：%s %v", res.Status, res.Error)
+	}
+	if be.totalCalls() != 3 {
+		t.Fatalf("恰 3 次，实际 %d", be.totalCalls())
+	}
+	if sink.countText("loop_exhausted") != 1 {
+		t.Fatal("缺 loop_exhausted 事件")
+	}
+}
+
+// ---- router：分支选中/落选跳过/无命中 fail-closed ----
+
+func routerSteps() []StepSpec {
+	return []StepSpec{
+		{ID: "r1", Role: "researcher", Title: "路由", Targets: []string{"output.md", "fields.json"},
+			Branches: []Branch{
+				{When: BranchCond{Field: "route", Equals: "fast"}, Then: []string{"fast"}},
+				{When: BranchCond{Field: "route", NotEquals: "fast"}, Then: []string{"slow"}},
+			}},
+		{ID: "fast", Role: "writer", Title: "快线", Needs: []string{"r1"}, Targets: []string{"output.md"}},
+		{ID: "slow", Role: "writer", Title: "慢线", Needs: []string{"r1"}, Targets: []string{"output.md"}},
+	}
+}
+
+func TestTopologyConformance_routerSelectsBranch(t *testing.T) {
+	be := &iterBackend{calls: map[string]int{}, content: func(_ string, _ int, target string) []byte {
+		if target == "fields.json" {
+			return fieldsJSON(map[string]string{"route": "fast"})
+		}
+		return nil
+	}}
+	engine := newTopologyEngine(t, mockScriptPlanner(routerSteps()...), be, nil)
+	res, err := engine.Run(context.Background(), Goal{Text: "路由任务"}, RunOptions{Topology: TopologyRouter})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunCompleted {
+		t.Fatalf("分支落选是预期路径，不得降级/失败：%s %v", res.Status, res.Error)
+	}
+	byID := map[string]StepSpec{}
+	for _, s := range res.Plan.Steps {
+		byID[s.ID] = s
+	}
+	if byID["fast"].Status != "completed" {
+		t.Fatalf("fast = %s", byID["fast"].Status)
+	}
+	if byID["slow"].Status != "skipped" || byID["slow"].Error != routerSkipReason {
+		t.Fatalf("slow = %s %q", byID["slow"].Status, byID["slow"].Error)
+	}
+}
+
+func TestTopologyConformance_routerNoMatchFailsClosed(t *testing.T) {
+	be := &iterBackend{calls: map[string]int{}, content: func(_ string, _ int, target string) []byte {
+		if target == "fields.json" {
+			return fieldsJSON(map[string]string{"route": ""}) // 两条分支都不命中
+		}
+		return nil
+	}}
+	steps := routerSteps()
+	// 去掉 NotEquals 兜底分支，制造无命中。
+	steps[0].Branches = []Branch{{When: BranchCond{Field: "route", Equals: "fast"}, Then: []string{"fast"}}}
+	engine := newTopologyEngine(t, mockScriptPlanner(steps...), be, nil)
+	res, err := engine.Run(context.Background(), Goal{Text: "路由任务"}, RunOptions{Topology: TopologyRouter})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunFailed || !strings.Contains(res.Error, "无分支命中") {
+		t.Fatalf("无命中必须 fail-closed：%s %v", res.Status, res.Error)
+	}
+}
+
+// ---- evaluator-optimizer：draft→eval 循环复合，收敛即停 ----
+
+func TestTopologyConformance_evaluatorOptimizer(t *testing.T) {
+	// eval 第 1-2 轮 verdict=fail，第 3 轮 pass；draft 随 body 再武装。
+	be := &iterBackend{calls: map[string]int{}, content: func(key string, iter int, target string) []byte {
+		if target == "fields.json" {
+			if iter >= 3 {
+				return fieldsJSON(map[string]string{"verdict": "pass"})
+			}
+			return fieldsJSON(map[string]string{"verdict": "fail"})
+		}
+		if target == "verdict.md" {
+			return []byte(fmt.Sprintf("评审第 %d 轮\n\n%s", iter, goodReport))
+		}
+		return nil
+	}}
+	engine := newTopologyEngine(t, mockScriptPlanner(
+		StepSpec{ID: "draft", Role: "writer", Title: "起草", Targets: []string{"output.md"}},
+		StepSpec{ID: "eval", Role: "reviewer", Title: "评审", Targets: []string{"verdict.md", "fields.json"},
+			Loop: &LoopSpec{MaxIterations: 4, UntilField: "verdict", UntilEquals: "pass", Body: []string{"draft"}}},
+	), be, nil)
+	res, err := engine.Run(context.Background(), Goal{Text: "起草并评审"}, RunOptions{Topology: TopologyEvaluatorOptimizer})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunCompleted {
+		t.Fatalf("status = %s err = %v", res.Status, res.Error)
+	}
+	byID := map[string]StepSpec{}
+	for _, s := range res.Plan.Steps {
+		byID[s.ID] = s
+	}
+	if byID["eval"].LoopRuns != 2 {
+		t.Fatalf("eval 应再武装 2 轮：%+v", byID["eval"])
+	}
+	// draft 3 次 + eval 3 次。
+	if be.totalCalls() != 6 {
+		t.Fatalf("总执行次数 = %d，期望 6", be.totalCalls())
+	}
+}
+
+// ---- long-running：截止时刻 → 未完步骤 skipped、降级收口 ----
+
+func TestTopologyConformance_longRunningDeadline(t *testing.T) {
+	// 时序鲁棒设计：每步 300ms、截止 +500ms——装配耗时不影响断言核心：
+	// s1 必完成（首个 superstep 在截止前开跑），s3 必被跳过（第三个
+	// superstep 边界必超期），收口必为 degraded。s2 完成与否取决于装配
+	// 耗时，不做断言。
+	be := &iterBackend{calls: map[string]int{}, delay: 300 * time.Millisecond}
+	sink := &collectSink{}
+	engine := newTopologyEngine(t, mockScriptPlanner(
+		StepSpec{ID: "s1", Role: "researcher", Title: "一", Targets: []string{"notes.md"}},
+		StepSpec{ID: "s2", Role: "writer", Title: "二", Targets: []string{"output.md"}},
+		StepSpec{ID: "s3", Role: "reviewer", Title: "三", Targets: []string{"output.md"}},
+	), be, sink)
+	res, err := engine.Run(context.Background(), Goal{Text: "长任务"}, RunOptions{
+		Topology: TopologyLongRunning,
+		Deadline: time.Now().Add(500 * time.Millisecond),
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunDegraded {
+		t.Fatalf("截止后应降级收口：%s %v", res.Status, res.Error)
+	}
+	byID := map[string]StepSpec{}
+	for _, s := range res.Plan.Steps {
+		byID[s.ID] = s
+	}
+	if byID["s1"].Status != "completed" {
+		t.Fatalf("s1 = %s", byID["s1"].Status)
+	}
+	if byID["s3"].Status != "skipped" || byID["s3"].Error != "deadline exceeded" {
+		t.Fatalf("s3 = %s %q", byID["s3"].Status, byID["s3"].Error)
+	}
+	if sink.countText("截止时刻") != 1 {
+		t.Fatal("缺截止事件")
+	}
+}
+
+// ---- 拓扑校验 fail-closed ----
+
+func TestTopologyValidationFailsClosed(t *testing.T) {
+	be := &iterBackend{calls: map[string]int{}}
+	cases := []struct {
+		name    string
+		topo    string
+		wantErr string
+	}{
+		{"未注册名", "bogus_topology", "未知拓扑"},
+		{"primitive 单跑", TopologyMailbox, "跨拓扑原语"},
+		{"primitive 单跑 handoff", TopologyHandoff, "跨拓扑原语"},
+		// review/group_chat（WP3）与 task_board（WP4）驱动均已点亮；
+		// 12/12 受理面由 TestTopologyConformance_allTwelve 汇总钉住。
+	}
+	for _, tc := range cases {
+		engine := newTopologyEngine(t, mockScriptPlanner(
+			StepSpec{ID: "s1", Role: "researcher", Title: "一", Targets: []string{"output.md"}},
+		), be, nil)
+		res, err := engine.Run(context.Background(), Goal{Text: "x"}, RunOptions{Topology: tc.topo})
+		if err != nil {
+			t.Fatalf("%s：err = %v", tc.name, err)
+		}
+		if res.Status != RunFailed || !strings.Contains(res.Error, tc.wantErr) {
+			t.Fatalf("%s：%s %q（期望含 %q）", tc.name, res.Status, res.Error, tc.wantErr)
+		}
+	}
+}
+
+// leader 拓扑名映射到既有 Mode 委派循环（本测只钉「受理且映射」，
+// leader 行为 conformance 在 leader_test.go / WP2 增强）。
+func TestTopologyLeaderMapsToMode(t *testing.T) {
+	spec, ok := LookupTopology(TopologyLeader)
+	if !ok || spec.Kind != TopologyMode {
+		t.Fatalf("leader 注册形态：%+v %v", spec, ok)
+	}
+	if _, err := checkTopology(TopologyLeader); err != nil {
+		t.Fatalf("leader 拓扑应可受理：%v", err)
+	}
+}
+
+// 12 种内置拓扑全部注册且 Kind 合法（清单完整性）。
+func TestBuiltinTopologiesComplete(t *testing.T) {
+	want := map[string]TopologyKind{
+		TopologySequential: TopologyGraph, TopologyParallel: TopologyGraph,
+		TopologyLoop: TopologyGraph, TopologyRouter: TopologyGraph,
+		TopologyLeader: TopologyMode, TopologyHandoff: TopologyPrimitive,
+		TopologyGroupChat: TopologyDriver, TopologyReview: TopologyDriver,
+		TopologyEvaluatorOptimizer: TopologyGraph, TopologyLongRunning: TopologyGraph,
+		TopologyTaskBoard: TopologyDriver, TopologyMailbox: TopologyPrimitive,
+	}
+	got := BuiltinTopologies()
+	if len(got) != len(want) {
+		t.Fatalf("注册数 = %d，期望 %d", len(got), len(want))
+	}
+	for _, spec := range got {
+		if k, ok := want[spec.Name]; !ok || k != spec.Kind {
+			t.Fatalf("拓扑 %s kind = %s（期望 %s）", spec.Name, spec.Kind, k)
+		}
+	}
+}

squad/workflow.go

@@ -29,6 +29,15 @@ func (d *flowDriver) advance(ctx context.Context, st *RunState) (res *RunResult,
 	d.emit(st, Event{Type: EvPhase, Text: "开始执行工作流"})
 
 	superNo := 0
+	// 运行级截止（Q2 long_running）：superstep 边界检查，超期未完步骤置
+	// skipped 并降级收口；零值/不可解析 = 不设限。
+	var runDeadline time.Time
+	if v := st.Meta.Metadata["deadline"]; v != "" {
+		// RFC3339Nano 同时接受秒级与纳秒级两种写法。
+		if t, terr := time.Parse(time.RFC3339Nano, v); terr == nil {
+			runDeadline = t
+		}
+	}
 	cpw := newCheckpointWriter(CheckpointMode(st.Meta.Metadata["checkpoint_mode"]), st.storeOr())
 	// 推进退出前 flush：async 落完、on_exit 写最终快照（Q1 WP4 三档的出口语义）。
 	var lastGraph *graph
@@ -43,6 +52,19 @@ func (d *flowDriver) advance(ctx context.Context, st *RunState) (res *RunResult,
 		}
 	}()
 	for {
+		if !runDeadline.IsZero() && time.Now().After(runDeadline) {
+			d.emit(st, Event{Type: EvPhase, Text: "运行截止时刻已到，未完步骤按 skipped 收口"})
+			st.mu.Lock()
+			for i := range st.Plan.Workflow.Steps {
+				if st.Plan.Workflow.Steps[i].Status == "pending" || st.Plan.Workflow.Steps[i].Status == "running" {
+					st.Plan.Workflow.Steps[i].Status = "skipped"
+					st.Plan.Workflow.Steps[i].Error = "deadline exceeded"
+				}
+			}
+			_ = st.save()
+			st.mu.Unlock()
+			return d.finalize(ctx, st)
+		}
 		g, gerr := compileGraph(st.Plan.Workflow.Steps)
 		if gerr != nil {
 			wf2 := st.Workflow()
@@ -197,6 +219,7 @@ func (d *flowDriver) runParallel(ctx context.Context, st *RunState, idxs []int)
 	var wg sync.WaitGroup
 	var mu sync.Mutex
 	failedCritical := -1
+	var firstFail *RunResult
 	var gate *HitlGate
 	for _, idx := range idxs {
 		wg.Add(1)
@@ -211,6 +234,11 @@ func (d *flowDriver) runParallel(ctx context.Context, st *RunState, idxs []int)
 			}
 			if (out != nil && out.Status == RunFailed) || err != nil {
 				cancel()
+				// 保留首个带具体错误的失败结果（如 router 无命中）；
+				// failStep 的通用格式只在没有更具体结果时兜底。
+				if out != nil && out.Status == RunFailed && firstFail == nil {
+					firstFail = out
+				}
 				if st.Plan.Workflow.Steps[idx].IsCritical() && failedCritical < 0 {
 					failedCritical = idx
 				}
@@ -221,6 +249,9 @@ func (d *flowDriver) runParallel(ctx context.Context, st *RunState, idxs []int)
 	if gate != nil {
 		return heldResult(st, gate), gate, nil
 	}
+	if firstFail != nil {
+		return firstFail, nil, nil
+	}
 	if failedCritical >= 0 {
 		return d.failStep(ctx, st, failedCritical)
 	}
@@ -248,6 +279,16 @@ func (d *flowDriver) runOneStep(ctx context.Context, st *RunState, idx int) (*Ru
 			return d.failStep(ctx, st, idx)
 		}
 	}
+	// handoff 原语 fail-closed（Q2）：目标角色必须可解析，启动前拒绝。
+	if step.Handoff != "" {
+		if _, ok := findAgent(st.Plan.Squad.Agents, step.Handoff); !ok {
+			if _, exists := d.opts.Registry.Get(step.Handoff); !exists {
+				step.Status, step.Error = "failed", "handoff: 未知目标角色 "+step.Handoff
+				st.setStep(idx, step)
+				return d.failStep(ctx, st, idx)
+			}
+		}
+	}
 	subDir := filepath.Join(st.Meta.WorkDir, filepath.FromSlash(orDefault(step.SubWorkspace, fmt.Sprintf("tasks/%02d_%s", idx+1, step.Role))))
 	if err := os.MkdirAll(subDir, 0o755); err != nil {
 		return nil, nil, err
@@ -278,6 +319,18 @@ func (d *flowDriver) runOneStep(ctx context.Context, st *RunState, idx int) (*Ru
 	}
 	prompt := readFileOr(filepath.Join(subDir, "TASK.md"), step.Title)
 	prompt = st.injectSteering(prompt)
+	// 邮箱投递（Q2）：本步骤启动时把发给该角色（含广播）的未读消息并入
+	// 任务书；只走 prompt 面，永不触碰 HITL 门（验收三构造性保证）。
+	st.mu.Lock()
+	mail := takeMailFor(st, step.Role)
+	handoff := takeHandoffFor(st, step.Role)
+	if len(mail) > 0 || handoff != nil {
+		_ = st.save()
+	}
+	st.mu.Unlock()
+	prompt += formatMailSection(mail)
+	// handoff 原语（Q2）：目标角色的步骤接收过滤历史（只带移交方主产出）。
+	prompt += formatHandoffSection(handoff)
 	instructions := readFileOr(filepath.Join(subDir, "AGENTS.md"), d.opts.Workspaces.ResolveInstructions(role))
 
 	var backend Executor
@@ -423,9 +476,154 @@ func (d *flowDriver) runOneStep(ctx context.Context, st *RunState, idx int) (*Ru
 	}
 	d.emit(st, Event{Type: EvStepCompleted, StepID: step.ID, Role: step.Role, Text: step.Title,
 		Data: marshalJSON(map[string]any{"duration_ms": step.DurationMs, "targets": targets})})
+
+	// handoff 原语（Q2）：完成即移交最终答案所有权——记录 + 主产出快照，
+	// 供目标角色下一步骤（过滤历史）与 finalize（链尾产出）消费。
+	if step.Handoff != "" {
+		d.recordHandoff(st, step, subDir, targets)
+	}
+
+	// 循环步再武装（Q2 loop）：停机条件不满足且预算未尽 → 本步与 loop.body
+	// 一起重置为 pending；触顶发 loop_exhausted 事件后按完成继续。
+	if step.Loop != nil {
+		if d.loopShouldRearm(st, step) {
+			if out, g2, err2 := d.rearmLoop(ctx, st, idx, step); out != nil || g2 != nil || err2 != nil {
+				return out, g2, err2
+			}
+			return nil, nil, nil
+		}
+		if step.Loop.UntilField != "" && step.LoopRuns+1 >= step.Loop.MaxIterations {
+			d.emit(st, Event{Type: EvPhase, StepID: step.ID, Text: fmt.Sprintf("loop_exhausted：%s 达 %d 轮上限，停机条件未满足", step.ID, step.Loop.MaxIterations)})
+		}
+	}
+	// 路由步分支裁决（Q2 router）：首个命中的分支保留，其余分支目标置
+	// skipped；无分支命中 fail-closed 成失败运行。
+	if len(step.Branches) > 0 {
+		return d.applyBranches(ctx, st, idx, step)
+	}
 	return nil, nil, nil
 }
 
+// loopShouldRearm 判定循环步是否再武装：停机条件（黑板字段等值）满足或
+// 预算用尽则停。字段判据读的是本轮 mergeStepArtifacts 之后的黑板。
+func (d *flowDriver) loopShouldRearm(st *RunState, step StepSpec) bool {
+	// MaxIterations 是总迭代上限（含首跑）：已跑 LoopRuns+1 次。
+	if step.LoopRuns+1 >= step.Loop.MaxIterations {
+		return false
+	}
+	if step.Loop.UntilField != "" {
+		st.mu.Lock()
+		v := st.BB.Fields[step.Loop.UntilField]
+		st.mu.Unlock()
+		if v == step.Loop.UntilEquals {
+			return false
+		}
+	}
+	return true
+}
+
+// rearmLoop 把循环步与 loop.body 一起重置为 pending（同一临界区内落盘）。
+func (d *flowDriver) rearmLoop(_ context.Context, st *RunState, idx int, step StepSpec) (*RunResult, *HitlGate, error) {
+	st.mu.Lock()
+	step.LoopRuns++
+	step.Status = "pending"
+	step.Error = ""
+	st.Plan.Workflow.Steps[idx] = step
+	for _, bodyID := range step.Loop.Body {
+		for j := range st.Plan.Workflow.Steps {
+			if st.Plan.Workflow.Steps[j].ID == bodyID && st.Plan.Workflow.Steps[j].Status == "completed" {
+				st.Plan.Workflow.Steps[j].Status = "pending"
+				st.Plan.Workflow.Steps[j].Error = ""
+			}
+		}
+	}
+	err := st.save()
+	round := step.LoopRuns
+	st.mu.Unlock()
+	if err != nil {
+		return nil, nil, err
+	}
+	d.emit(st, Event{Type: EvPhase, StepID: step.ID, Text: fmt.Sprintf("loop 第 %d 轮再武装：%s", round+1, step.ID)})
+	return nil, nil, nil
+}
+
+// applyBranches 裁决路由步分支：When 按声明序取首个命中；未选中分支的
+// pending 目标置 skipped（skipped=依赖已满足，join 步照常推进）。
+func (d *flowDriver) applyBranches(_ context.Context, st *RunState, _ int, step StepSpec) (*RunResult, *HitlGate, error) {
+	st.mu.Lock()
+	fields := map[string]string{}
+	for k, v := range st.BB.Fields {
+		fields[k] = v
+	}
+	st.mu.Unlock()
+	var selected *Branch
+	for i := range step.Branches {
+		if branchMatches(fields, step.Branches[i].When) {
+			selected = &step.Branches[i]
+			break
+		}
+	}
+	if selected == nil {
+		wf := st.Workflow()
+		wf.Status = "failed"
+		st.setWorkflow(wf)
+		_ = st.save()
+		return &RunResult{RunID: st.Meta.ID, Status: RunFailed, Plan: planPtr(st),
+			Error: fmt.Sprintf("router: 步骤 %s 无分支命中（fail-closed）", step.ID), WorkDir: st.Meta.WorkDir}, nil, nil
+	}
+	keep := map[string]bool{}
+	for _, t := range selected.Then {
+		keep[t] = true
+	}
+	skipped := 0
+	st.mu.Lock()
+	for j := range st.Plan.Workflow.Steps {
+		sj := &st.Plan.Workflow.Steps[j]
+		if sj.ID == step.ID || keep[sj.ID] || sj.Status != "pending" {
+			continue
+		}
+		for _, br := range step.Branches {
+			inBranch := false
+			for _, t := range br.Then {
+				if t == sj.ID {
+					inBranch = true
+					break
+				}
+			}
+			if inBranch {
+				sj.Status = "skipped"
+				sj.Error = routerSkipReason
+				skipped++
+				break
+			}
+		}
+	}
+	err := st.save()
+	st.mu.Unlock()
+	if err != nil {
+		return nil, nil, err
+	}
+	d.emit(st, Event{Type: EvPhase, StepID: step.ID,
+		Text: fmt.Sprintf("router %s 选中分支 %v（跳过 %d 步）", step.ID, selected.Then, skipped)})
+	return nil, nil, nil
+}
+
+// routerSkipReason 是分支落选的 skipped 标记（finalize 的降级豁免凭据）。
+const routerSkipReason = "router: 分支未选中"
+
+// branchMatches 判定分支条件：Equals 非空须相等，NotEquals 非空须不等；
+// 字段未设时按零值 "" 参与比较（compileGraph 已保证两者至少其一非空）。
+func branchMatches(fields map[string]string, c BranchCond) bool {
+	v := fields[c.Field]
+	if c.Equals != "" && v != c.Equals {
+		return false
+	}
+	if c.NotEquals != "" && v == c.NotEquals {
+		return false
+	}
+	return true
+}
+
 // failStep 按关键性处理失败步骤：关键 → 运行失败；非关键 → 标记后继续（返回 nil 让循环推进）。
 func (d *flowDriver) failStep(ctx context.Context, st *RunState, idx int) (*RunResult, *HitlGate, error) {
 	st.mu.Lock()
@@ -587,17 +785,30 @@ func (d *flowDriver) finalize(ctx context.Context, st *RunState) (*RunResult, *H
 	wf := st.Workflow()
 	degraded := false
 	for _, s := range wf.Steps {
-		if s.Status == "failed" || s.Status == "skipped" {
+		if s.Status == "failed" {
+			degraded = true
+		}
+		// router 分支落选是预期路径，不算降级（Q2）。
+		if s.Status == "skipped" && s.Error != routerSkipReason {
 			degraded = true
 		}
 	}
 	// 汇总交付物：最后一个 completed 步骤的主产出 → output.md。
+	// handoff 链尾优先（Q2）：有所有权移交时，交付物取链尾角色的最后一个
+	// completed 步骤产出（最终答案所有权在链尾），无则该角色无完成步骤时回落。
 	var srcStep *StepSpec
 	for i := range wf.Steps {
 		if wf.Steps[i].Status == "completed" {
 			srcStep = &wf.Steps[i]
 		}
 	}
+	if tail := handoffChainTail(st); tail != "" {
+		for i := range wf.Steps {
+			if wf.Steps[i].Status == "completed" && wf.Steps[i].Role == tail {
+				srcStep = &wf.Steps[i]
+			}
+		}
+	}
 	if srcStep == nil {
 		wf.Status = "failed"
 		st.setWorkflow(wf)

squad/workspace.go

@@ -189,6 +189,13 @@ func (e *WorkspaceEngine) writeTaskFile(subDir string, step *StepSpec, role Role
 	fmt.Fprintf(&b, "- 角色：%s（%s）\n", role.Name, role.Role)
 	fmt.Fprintf(&b, "- 期望产出：%v\n\n", step.Targets)
 
+	// 编译期任务指令（Q3）：编译器渲染好的富文本任务书，权威仅次于目标。
+	if strings.TrimSpace(step.Brief) != "" {
+		b.WriteString("## 任务指令\n\n")
+		b.WriteString(strings.TrimSpace(step.Brief))
+		b.WriteString("\n\n")
+	}
+
 	b.WriteString("## 任务上下文资源与权威等级（已快照到本目录，只引用本地路径）\n\n")
 	b.WriteString("| 资源路径 | 权威等级 | 用途说明 | 使用与引用指导 |\n|---|---|---|---|\n")
 	b.WriteString("| `context/goal.md` | **authoritative** | 运行目标 | 不可推翻 |\n")

squad/wp3_conformance_test.go

@@ -0,0 +1,452 @@
+package squad
+
+import (
+	"context"
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+)
+
+// wp3_conformance_test.go — Q2 WP3 验收：review 拓扑与 backend runReview 等价
+// （验收二：并行找问题、逐条验证、丢掉未验证）、handoff 原语（所有权转移 +
+// 过滤历史 + finalize 链尾）、group_chat 驱动（轮转/主持人 + 预算停机）。
+
+// ---- review：与 backend review_test.go 同款断言 ----
+
+// reviewScriptBackend finder 目录回放 findings.json（第 3 个进入者放行，钉住
+// 真并发）；verify_* 目录按 prompt 内容回放判定（与 backend 夹具同文案）。
+type reviewScriptBackend struct {
+	mu          sync.Mutex
+	once        sync.Once
+	entered     int
+	verifyCalls int
+	release     chan struct{}
+}
+
+func (b *reviewScriptBackend) Name() string { return "review-script" }
+
+func (b *reviewScriptBackend) Execute(_ context.Context, req ExecRequest) (*ExecResult, error) {
+	key := filepath.Base(filepath.Clean(req.WorkDir))
+	if strings.HasPrefix(key, "verify_") {
+		b.mu.Lock()
+		b.verifyCalls++
+		b.mu.Unlock()
+		verdict := "看不出对错"
+		switch {
+		case strings.Contains(req.Prompt, "无来源"):
+			verdict = "核对原文确实没有来源。\nVERDICT: PASS"
+		case strings.Contains(req.Prompt, "错字"):
+			verdict = "核对后成立。\nVERDICT: PASS"
+		case strings.Contains(req.Prompt, "过期数据"):
+			verdict = "数据仍有效。\nVERDICT: FAIL"
+		}
+		return writeOne(req.WorkDir, "verdict.md", verdict), nil
+	}
+	// finder：三个必须同时在跑（backend review_test.go:16-33 同款门）。
+	b.mu.Lock()
+	b.entered++
+	n := b.entered
+	b.mu.Unlock()
+	if n == 3 {
+		b.once.Do(func() { close(b.release) })
+	}
+	select {
+	case <-b.release:
+	case <-time.After(time.Second):
+		return nil, fmt.Errorf("finders did not run together")
+	}
+	var findings string
+	switch key {
+	case "01_researcher":
+		findings = `[{"location":"第2段","severity":"high","issue":"无来源","evidence":"显然"},{"location":"第1段","severity":"low","issue":"错字","evidence":"的地"},{"location":"表1","severity":"medium","issue":"过期数据","evidence":"2020"}]`
+	case "02_writer":
+		findings = `[{"location":"第2段","severity":"high","issue":"无来源","evidence":"显然"},{"location":"附录","severity":"medium","issue":"数字不一致","evidence":"3 与 4"}]`
+	default:
+		findings = `[{"location":"结尾","severity":"low","issue":"没有结论","evidence":"最后一句"}]`
+	}
+	return writeOne(req.WorkDir, "findings.json", findings), nil
+}
+
+func writeOne(dir, name, content string) *ExecResult {
+	out := filepath.Join(dir, name)
+	_ = os.MkdirAll(filepath.Dir(out), 0o755)
+	_ = os.WriteFile(out, []byte(content), 0o644)
+	return &ExecResult{Success: true, Patches: map[string][]byte{name: []byte(content)}}
+}
+
+func TestTopologyConformance_review(t *testing.T) {
+	be := &reviewScriptBackend{release: make(chan struct{})}
+	sink := &collectSink{}
+	planner := mockScriptPlanner(
+		StepSpec{ID: "f1", Role: "researcher", Title: "评审一", Targets: []string{"findings.json"}},
+		StepSpec{ID: "f2", Role: "writer", Title: "评审二", Targets: []string{"findings.json"}},
+		StepSpec{ID: "f3", Role: "reviewer", Title: "评审三", Targets: []string{"findings.json"}},
+	)
+	planner.Agents = []RoleSpec{
+		{Role: "researcher", Name: "评审一", Backend: "review-script"},
+		{Role: "writer", Name: "评审二", Backend: "review-script"},
+		{Role: "reviewer", Name: "评审三", Backend: "review-script"},
+		{Role: "verifier", Name: "核实", Backend: "review-script"},
+	}
+	engine := newWP3Engine(t, planner, be, sink, []RoleSpec{
+		{Role: "researcher", Name: "评审一", Backend: "review-script"},
+		{Role: "writer", Name: "评审二", Backend: "review-script"},
+		{Role: "reviewer", Name: "评审三", Backend: "review-script"},
+		{Role: "verifier", Name: "核实", Backend: "review-script"},
+	})
+	res, err := engine.Run(context.Background(), Goal{Text: "评审材料正文"},
+		RunOptions{Topology: TopologyReview, ReviewTarget: "材料正文"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunCompleted {
+		t.Fatalf("status = %s err = %v", res.Status, res.Error)
+	}
+	data, err := os.ReadFile(filepath.Join(res.WorkDir, "output.md"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	out := string(data)
+	// 只留 PASS（fail-closed）：无来源/错字保留；FAIL 与无判定丢弃。
+	if !strings.Contains(out, "无来源") || !strings.Contains(out, "错字") ||
+		strings.Contains(out, "数字不一致") || strings.Contains(out, "没有结论") ||
+		strings.Contains(out, "过期数据") {
+		t.Fatalf("filter = %s", out)
+	}
+	// 报告恒带丢弃计数（backend 同格式串）。
+	if !strings.Contains(out, "已丢弃 3 条未验证") {
+		t.Fatalf("dropped = %s", out)
+	}
+	// 去重后逐条验证：verify 调用数 = unique 数 = 5。
+	if be.verifyCalls != 5 {
+		t.Fatalf("verify calls = %d, want 5 after dedupe", be.verifyCalls)
+	}
+	// severity 稳定排序：high(无来源) 在 low(错字) 前。
+	if strings.Index(out, "无来源") > strings.Index(out, "错字") {
+		t.Fatalf("severity order: %s", out)
+	}
+	if sink.countText("review：3 个 finder 并行找问题") != 1 ||
+		sink.countText("review：去重后 5 条") != 1 {
+		t.Fatal("缺 review 阶段事件")
+	}
+}
+
+// verify 角色缺失 → fail-closed。
+func TestReviewVerifyRoleMissingFailsClosed(t *testing.T) {
+	be := &reviewScriptBackend{release: make(chan struct{})}
+	planner := mockScriptPlanner(
+		StepSpec{ID: "f1", Role: "researcher", Title: "评审", Targets: []string{"findings.json"}},
+	)
+	engine := newWP3Engine(t, planner, be, NopSink{}, []RoleSpec{
+		{Role: "researcher", Name: "评审", Backend: "review-script"},
+	})
+	res, err := engine.Run(context.Background(), Goal{Text: "评审"},
+		RunOptions{Topology: TopologyReview})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunFailed || !strings.Contains(res.Error, "verify 角色") {
+		t.Fatalf("应 fail-closed：%s %q", res.Status, res.Error)
+	}
+}
+
+// ---- handoff 原语 ----
+
+// handoffScriptBackend 记录 prompt 并按目录回放带标记的产出（≥50 字过最小质检）。
+type handoffScriptBackend struct {
+	mu      sync.Mutex
+	prompts map[string][]string
+}
+
+func (b *handoffScriptBackend) Name() string { return "handoff-script" }
+
+func (b *handoffScriptBackend) Execute(_ context.Context, req ExecRequest) (*ExecResult, error) {
+	key := filepath.Base(filepath.Clean(req.WorkDir))
+	b.mu.Lock()
+	b.prompts[key] = append(b.prompts[key], req.Prompt)
+	b.mu.Unlock()
+	pad := "\n\n这是为了满足最小质检长度要求而补充的正文内容，超过五十个字符。"
+	var content string
+	switch key {
+	case "01_researcher":
+		content = "MARK_RESEARCHER 调研笔记。" + pad
+	case "02_writer":
+		content = "MARK_WRITER 终稿。" + pad
+	default:
+		content = "MARK_REVIEWER 评审意见。" + pad
+	}
+	name := "output.md"
+	if len(req.Targets) > 0 {
+		name = req.Targets[0]
+	}
+	return writeOne(req.WorkDir, name, content), nil
+}
+
+func TestTopologyConformance_handoff(t *testing.T) {
+	be := &handoffScriptBackend{prompts: map[string][]string{}}
+	sink := &collectSink{}
+	engine := newWP3Engine(t, mockScriptPlanner(
+		StepSpec{ID: "s1", Role: "researcher", Title: "调研", Targets: []string{"notes.md"}, Handoff: "writer"},
+		StepSpec{ID: "s2", Role: "writer", Title: "成文", Targets: []string{"output.md"}, Needs: []string{"s1"}},
+		StepSpec{ID: "s3", Role: "reviewer", Title: "终审", Targets: []string{"output.md"}, Needs: []string{"s2"}},
+	), be, sink, []RoleSpec{
+		{Role: "researcher", Name: "调研", Backend: "handoff-script"},
+		{Role: "writer", Name: "撰写", Backend: "handoff-script"},
+		{Role: "reviewer", Name: "审核", Backend: "handoff-script"},
+	})
+	res, err := engine.Run(context.Background(), Goal{Text: "带移交的三步"}, RunOptions{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunCompleted {
+		t.Fatalf("status = %s err = %v", res.Status, res.Error)
+	}
+	// ① 所有权转移记录持久化。
+	st, err := engine.loadRun(engine.runDir(res.RunID))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(st.Handoffs) != 1 || st.Handoffs[0].FromRole != "researcher" || st.Handoffs[0].ToRole != "writer" {
+		t.Fatalf("Handoffs = %+v", st.Handoffs)
+	}
+	if !strings.Contains(st.Handoffs[0].Output, "MARK_RESEARCHER") {
+		t.Fatalf("移交产出快照缺失：%q", st.Handoffs[0].Output)
+	}
+	if sink.countText("handoff：researcher → writer") != 1 {
+		t.Fatal("缺所有权移交事件")
+	}
+	// ② 目标收过滤历史：writer 的任务书带移交上下文，reviewer 不带。
+	var writerPrompt, reviewerPrompt string
+	for _, p := range be.prompts["02_writer"] {
+		writerPrompt += p
+	}
+	for _, p := range be.prompts["03_reviewer"] {
+		reviewerPrompt += p
+	}
+	if !strings.Contains(writerPrompt, "移交上下文") || !strings.Contains(writerPrompt, "MARK_RESEARCHER") {
+		t.Fatalf("writer 未收到过滤历史：%q", writerPrompt)
+	}
+	if strings.Contains(reviewerPrompt, "移交上下文") {
+		t.Fatal("非目标角色不应收到移交上下文")
+	}
+	// ③ finalize 取链尾输出：交付物 = writer（链尾）产出，而非最后完成的 reviewer。
+	data, err := os.ReadFile(filepath.Join(res.WorkDir, "output.md"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(string(data), "MARK_WRITER") || strings.Contains(string(data), "MARK_REVIEWER") {
+		t.Fatalf("交付物应为链尾 writer 产出：%s", data)
+	}
+}
+
+// handoff 目标角色未知 → 步骤启动前 fail-closed。
+func TestHandoffUnknownRoleFailsClosed(t *testing.T) {
+	be := &handoffScriptBackend{prompts: map[string][]string{}}
+	engine := newWP3Engine(t, mockScriptPlanner(
+		StepSpec{ID: "s1", Role: "researcher", Title: "调研", Handoff: "ghost"},
+	), be, NopSink{}, []RoleSpec{
+		{Role: "researcher", Name: "调研", Backend: "handoff-script"},
+	})
+	res, err := engine.Run(context.Background(), Goal{Text: "x"}, RunOptions{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunFailed || !strings.Contains(res.Error, "handoff: 未知目标角色 ghost") {
+		t.Fatalf("应 fail-closed：%s %q", res.Status, res.Error)
+	}
+}
+
+// ---- group_chat 驱动 ----
+
+// gcScriptBackend 按目录回放 message.md。
+type gcScriptBackend struct {
+	messages map[string]string
+}
+
+func (b *gcScriptBackend) Name() string { return "gc-script" }
+
+func (b *gcScriptBackend) Execute(_ context.Context, req ExecRequest) (*ExecResult, error) {
+	key := filepath.Base(filepath.Clean(req.WorkDir))
+	msg := b.messages[key]
+	if msg == "" {
+		msg = "发言：" + key
+	}
+	return writeOne(req.WorkDir, "message.md", msg), nil
+}
+
+func TestTopologyConformance_groupChatRotateBudget(t *testing.T) {
+	be := &gcScriptBackend{messages: map[string]string{
+		"turn_01_researcher": "第一句",
+		"turn_02_writer":     "第二句",
+		"turn_03_researcher": "第三句",
+	}}
+	sink := &collectSink{}
+	engine := newWP3Engine(t, mockScriptPlanner(
+		StepSpec{ID: "p1", Role: "researcher", Title: "参与者甲"},
+		StepSpec{ID: "p2", Role: "writer", Title: "参与者乙"},
+	), be, sink, []RoleSpec{
+		{Role: "researcher", Name: "甲", Backend: "gc-script"},
+		{Role: "writer", Name: "乙", Backend: "gc-script"},
+	})
+	res, err := engine.Run(context.Background(), Goal{Text: "群聊"},
+		RunOptions{Topology: TopologyGroupChat, GroupChatMaxTurns: 3})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunCompleted {
+		t.Fatalf("status = %s err = %v", res.Status, res.Error)
+	}
+	// 预算停机：恰好 3 轮，轮转序 researcher→writer→researcher。
+	if sink.countText("轮次预算 3 用尽") != 1 {
+		t.Fatal("缺预算停机事件")
+	}
+	data, err := os.ReadFile(filepath.Join(res.WorkDir, "output.md"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	out := string(data)
+	for _, want := range []string{"1. researcher：第一句", "2. writer：第二句", "3. researcher：第三句"} {
+		if !strings.Contains(out, want) {
+			t.Fatalf("转录缺 %q：%s", want, out)
+		}
+	}
+	if strings.Contains(out, "4. ") {
+		t.Fatalf("超出预算仍在发言：%s", out)
+	}
+	// 名册步骤不执行，标记 skipped（非降级原因）。
+	st, _ := engine.loadRun(engine.runDir(res.RunID))
+	for _, s := range st.Plan.Workflow.Steps {
+		if s.ID == "p1" || s.ID == "p2" {
+			if s.Status != "skipped" || s.Error != gcRosterSkipReason {
+				t.Fatalf("名册步骤 %s = %s/%q", s.ID, s.Status, s.Error)
+			}
+		}
+	}
+}
+
+func TestTopologyConformance_groupChatCloseMarker(t *testing.T) {
+	be := &gcScriptBackend{messages: map[string]string{
+		"turn_01_researcher": "我先说",
+		"turn_02_writer":     "同意收口 [close]",
+	}}
+	sink := &collectSink{}
+	engine := newWP3Engine(t, mockScriptPlanner(
+		StepSpec{ID: "p1", Role: "researcher", Title: "甲"},
+		StepSpec{ID: "p2", Role: "writer", Title: "乙"},
+	), be, sink, []RoleSpec{
+		{Role: "researcher", Name: "甲", Backend: "gc-script"},
+		{Role: "writer", Name: "乙", Backend: "gc-script"},
+	})
+	res, err := engine.Run(context.Background(), Goal{Text: "群聊"},
+		RunOptions{Topology: TopologyGroupChat, GroupChatMaxTurns: 8})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunCompleted {
+		t.Fatalf("status = %s", res.Status)
+	}
+	if sink.countText("发出收口标记") != 1 {
+		t.Fatal("缺收口标记事件")
+	}
+	st, _ := engine.loadRun(engine.runDir(res.RunID))
+	if n := countGcSteps(st); n != 2 {
+		t.Fatalf("收口标记后应停在 2 轮，实际 %d", n)
+	}
+}
+
+func TestTopologyConformance_groupChatModerator(t *testing.T) {
+	be := &gcScriptBackend{messages: map[string]string{
+		"turn_01_researcher": `{"next":"writer","reason":"请乙补充"}`,
+		"turn_02_writer":     "补一段分析",
+		"turn_03_researcher": `{"action":"close","reason":"讨论充分"}`,
+	}}
+	sink := &collectSink{}
+	engine := newWP3Engine(t, mockScriptPlanner(
+		StepSpec{ID: "p1", Role: "researcher", Title: "主持人"},
+		StepSpec{ID: "p2", Role: "writer", Title: "乙"},
+	), be, sink, []RoleSpec{
+		{Role: "researcher", Name: "主持", Backend: "gc-script"},
+		{Role: "writer", Name: "乙", Backend: "gc-script"},
+	})
+	res, err := engine.Run(context.Background(), Goal{Text: "主持人群聊"}, RunOptions{
+		Topology: TopologyGroupChat, GroupChatPolicy: "moderator",
+		GroupChatModerator: "researcher", GroupChatMaxTurns: 8,
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunCompleted {
+		t.Fatalf("status = %s err = %v", res.Status, res.Error)
+	}
+	if sink.countText("主持人收口") != 1 {
+		t.Fatal("缺主持人收口事件")
+	}
+	data, _ := os.ReadFile(filepath.Join(res.WorkDir, "output.md"))
+	out := string(data)
+	for _, want := range []string{"1. researcher", "2. writer：补一段分析", "3. researcher"} {
+		if !strings.Contains(out, want) {
+			t.Fatalf("转录缺 %q：%s", want, out)
+		}
+	}
+	st, _ := engine.loadRun(engine.runDir(res.RunID))
+	if n := countGcSteps(st); n != 3 {
+		t.Fatalf("主持人模式应 3 轮收口，实际 %d", n)
+	}
+}
+
+// 主持人指名不在名册 → fail-closed。
+func TestGroupChatModeratorUnknownNextFailsClosed(t *testing.T) {
+	be := &gcScriptBackend{messages: map[string]string{
+		"turn_01_researcher": `{"next":"ghost"}`,
+	}}
+	engine := newWP3Engine(t, mockScriptPlanner(
+		StepSpec{ID: "p1", Role: "researcher", Title: "主持人"},
+		StepSpec{ID: "p2", Role: "writer", Title: "乙"},
+	), be, NopSink{}, []RoleSpec{
+		{Role: "researcher", Name: "主持", Backend: "gc-script"},
+		{Role: "writer", Name: "乙", Backend: "gc-script"},
+	})
+	res, err := engine.Run(context.Background(), Goal{Text: "x"}, RunOptions{
+		Topology: TopologyGroupChat, GroupChatPolicy: "moderator", GroupChatModerator: "researcher",
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunFailed || !strings.Contains(res.Error, "不在名册内") {
+		t.Fatalf("应 fail-closed：%s %q", res.Status, res.Error)
+	}
+}
+
+// ---- 夹具 ----
+
+func countGcSteps(st *RunState) int {
+	n := 0
+	for _, s := range st.Plan.Workflow.Steps {
+		if strings.HasPrefix(s.ID, "gc") && s.Status == "completed" {
+			n++
+		}
+	}
+	return n
+}
+
+func newWP3Engine(t *testing.T, planner Planner, be Executor, sink ProgressSink, roles []RoleSpec) *SquadEngine {
+	t.Helper()
+	dir := t.TempDir()
+	reg := NewRoleRegistry()
+	for _, r := range roles {
+		if err := reg.Register(r); err != nil {
+			t.Fatal(err)
+		}
+	}
+	backends := NewBackendResolver()
+	backends.RegisterInstance(be.Name(), be)
+	engine, err := NewEngine(Options{DataDir: dir, Registry: reg, Planner: planner, Backends: backends, Sink: sink})
+	if err != nil {
+		t.Fatal(err)
+	}
+	return engine
+}

squad/wp4_conformance_test.go

@@ -0,0 +1,355 @@
+package squad
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+)
+
+// wp4_conformance_test.go — Q2 WP4 验收：task_board 拓扑（claim/lease/complete
+// 原语 + 依赖门 + 孤儿租约再武装）与全 12 拓扑 conformance 汇总跑。
+
+// ---- task_board ----
+
+// 菱形依赖 a→(b,c)→d：认领序尊重依赖门，全部 done，标准 finalize 收口。
+func TestTopologyConformance_taskBoard(t *testing.T) {
+	be := &iterBackend{calls: map[string]int{}}
+	sink := &collectSink{}
+	engine := newTopologyEngine(t, mockScriptPlanner(
+		StepSpec{ID: "a", Role: "researcher", Title: "任务a", Targets: []string{"a.md"}},
+		StepSpec{ID: "b", Role: "writer", Title: "任务b", Needs: []string{"a"}, Targets: []string{"b.md"}},
+		StepSpec{ID: "c", Role: "reviewer", Title: "任务c", Needs: []string{"a"}, Targets: []string{"c.md"}},
+		StepSpec{ID: "d", Role: "writer", Title: "任务d", Needs: []string{"b", "c"}, Targets: []string{"output.md"}},
+	), be, sink)
+	res, err := engine.Run(context.Background(), Goal{Text: "任务板"}, RunOptions{Topology: TopologyTaskBoard})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunCompleted {
+		t.Fatalf("status = %s err = %v", res.Status, res.Error)
+	}
+	// 认领序尊重依赖门：a 先于 b/c，d 最后（板上按步骤序认领，确定性）。
+	var order []string
+	for _, ev := range sink.snapshot() {
+		if strings.Contains(ev.Text, "认领") {
+			order = append(order, ev.StepID)
+		}
+	}
+	want := []string{"a", "b", "c", "d"}
+	if strings.Join(order, ",") != strings.Join(want, ",") {
+		t.Fatalf("认领序 = %v，期望 %v", order, want)
+	}
+	st, _ := engine.loadRun(engine.runDir(res.RunID))
+	for _, task := range st.Board {
+		if task.Status != "done" {
+			t.Fatalf("任务 %s = %s，期望 done", task.ID, task.Status)
+		}
+	}
+	if sink.countText("task_board：收口（done 4/4）") != 1 {
+		t.Fatal("缺收口事件")
+	}
+}
+
+// 孤儿租约再武装：崩溃（ctx 取消）留下 claimed+过期租约 → Resume 重入
+// 再武装回 open、重新认领（Attempts=2）并跑完。
+func TestTaskBoardOrphanLeaseRearms(t *testing.T) {
+	ctx, cancel := context.WithCancel(context.Background())
+	crashed := false
+	be := &iterBackend{calls: map[string]int{}, content: func(key string, _ int, _ string) []byte {
+		if key == "02_writer" && !crashed {
+			crashed = true
+			cancel() // 模拟进程死亡：runOneStep 在 ctx 检查处退出，步骤回到 pending
+		}
+		return nil
+	}}
+	sink := &collectSink{}
+	engine := newTopologyEngine(t, mockScriptPlanner(
+		StepSpec{ID: "t1", Role: "researcher", Title: "一", Targets: []string{"a.md"}},
+		StepSpec{ID: "t2", Role: "writer", Title: "二", Needs: []string{"t1"}, Targets: []string{"output.md"}},
+	), be, sink)
+	res0, err := engine.Run(ctx, Goal{Text: "任务板"}, RunOptions{Topology: TopologyTaskBoard})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res0.Status != RunCancelled {
+		t.Fatalf("取消后应为 RunCancelled：%s", res0.Status)
+	}
+	// 崩溃现场：t2 claimed、租约未过期；步骤在盘上是 running（setStep 只改
+	// 内存，取消路径的 pending 未落盘）——进程死亡即此形态。测试把租约回拨
+	// 成过期（孤儿）并把步骤置回 pending（模拟重启前的一致化），再走 Resume。
+	st, err := engine.loadRun(engine.runDir(res0.RunID))
+	if err != nil {
+		t.Fatal(err)
+	}
+	found := false
+	for i := range st.Board {
+		if st.Board[i].ID == "t2" {
+			found = true
+			if st.Board[i].Status != "claimed" || st.Board[i].Attempts != 1 {
+				t.Fatalf("崩溃现场 t2 = %+v", st.Board[i])
+			}
+			st.Board[i].LeaseExpires = time.Now().Add(-time.Hour) // 租约过期 = 孤儿
+		}
+	}
+	if !found {
+		t.Fatal("板上无 t2")
+	}
+	for i := range st.Plan.Workflow.Steps {
+		if st.Plan.Workflow.Steps[i].ID == "t2" && st.Plan.Workflow.Steps[i].Status == "running" {
+			st.Plan.Workflow.Steps[i].Status = "pending"
+		}
+	}
+	if err := st.save(); err != nil {
+		t.Fatal(err)
+	}
+	// 重入：孤儿租约再武装 → 重新认领 → 跑完收口。
+	res, err := engine.Resume(context.Background(), st.Meta.ID)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunCompleted {
+		t.Fatalf("恢复后应完成：%s %v", res.Status, res.Error)
+	}
+	if sink.countText("孤儿租约再武装 t2") != 1 {
+		t.Fatal("缺孤儿租约再武装事件")
+	}
+	st2, _ := engine.loadRun(engine.runDir(res.RunID))
+	for _, task := range st2.Board {
+		if task.ID == "t2" && task.Attempts != 2 {
+			t.Fatalf("再武装后 Attempts = %d，期望 2", task.Attempts)
+		}
+	}
+}
+
+// 依赖门级联：非关键任务失败 → 下游 open 任务级联 skipped → 降级收口。
+func TestTaskBoardDependencyGateCascade(t *testing.T) {
+	be := &iterBackend{calls: map[string]int{}, content: func(key string, _ int, _ string) []byte {
+		return nil
+	}}
+	failBe := &failOnceBackend{inner: be, failKey: "02_writer"}
+	crit := false
+	sink := &collectSink{}
+	engine := newTopologyEngine(t, mockScriptPlanner(
+		StepSpec{ID: "t0", Role: "researcher", Title: "零", Targets: []string{"output.md"}},
+		StepSpec{ID: "t1", Role: "writer", Title: "一", Needs: []string{"t0"}, Targets: []string{"b.md"}, Critical: &crit},
+		StepSpec{ID: "t2", Role: "reviewer", Title: "二", Needs: []string{"t1"}, Targets: []string{"c.md"}},
+	), failBe, sink)
+	res, err := engine.Run(context.Background(), Goal{Text: "任务板"}, RunOptions{Topology: TopologyTaskBoard})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunDegraded {
+		t.Fatalf("非关键失败+级联跳过应降级收口：%s %v", res.Status, res.Error)
+	}
+	st, _ := engine.loadRun(engine.runDir(res.RunID))
+	byID := map[string]BoardTask{}
+	for _, task := range st.Board {
+		byID[task.ID] = task
+	}
+	if byID["t0"].Status != "done" || byID["t1"].Status != "failed" || byID["t2"].Status != "skipped" {
+		t.Fatalf("板状态 = %+v", byID)
+	}
+	for _, s := range st.Plan.Workflow.Steps {
+		if s.ID == "t2" && (s.Status != "skipped" || s.Error != boardSkipReason) {
+			t.Fatalf("t2 = %s %q", s.Status, s.Error)
+		}
+	}
+}
+
+// failOnceBackend 对指定目录恒定失败（非瞬态），其余委托 inner。
+type failOnceBackend struct {
+	inner   Executor
+	failKey string
+}
+
+func (b *failOnceBackend) Name() string { return b.inner.Name() }
+
+func (b *failOnceBackend) Execute(ctx context.Context, req ExecRequest) (*ExecResult, error) {
+	key := filepath.Base(filepath.Clean(req.WorkDir))
+	if key == b.failKey {
+		return &ExecResult{Success: false, ErrorReason: "fatal: 拒绝执行"}, nil
+	}
+	return b.inner.Execute(ctx, req)
+}
+
+// ---- 12/12 汇总 ----
+
+// TestTopologyConformance_allTwelve 全 12 内置拓扑汇总跑：每个拓扑名一次
+// 脚本化最小运行（零模型零网络）。graph/mode/driver 十种断言完成收口；
+// handoff/mailbox 两个 primitive 断言 fail-closed 拒绝单跑（其原语能力分别
+// 由 TestTopologyConformance_handoff 与 WP2 mailbox 测试钉住）。
+func TestTopologyConformance_allTwelve(t *testing.T) {
+	specs := BuiltinTopologies()
+	if len(specs) != 12 {
+		t.Fatalf("内置拓扑 = %d，期望 12", len(specs))
+	}
+	for _, spec := range specs {
+		spec := spec
+		t.Run(spec.Name, func(t *testing.T) {
+			switch spec.Name {
+			case TopologySequential:
+				runAgg(t, spec.Name, mockScriptPlanner(
+					StepSpec{ID: "s1", Role: "researcher", Title: "一", Targets: []string{"output.md"}},
+				), &iterBackend{calls: map[string]int{}}, RunOptions{}, RunCompleted)
+			case TopologyParallel:
+				runAgg(t, spec.Name, mockScriptPlanner(
+					StepSpec{ID: "s1", Role: "researcher", Title: "甲", Parallel: true, Targets: []string{"a.md"}},
+					StepSpec{ID: "s2", Role: "writer", Title: "乙", Parallel: true, Targets: []string{"output.md"}},
+				), &iterBackend{calls: map[string]int{}}, RunOptions{}, RunCompleted)
+			case TopologyLoop:
+				runAgg(t, spec.Name, mockScriptPlanner(
+					StepSpec{ID: "s1", Role: "researcher", Title: "循环体", Targets: []string{"output.md", "fields.json"},
+						Loop: &LoopSpec{MaxIterations: 2, UntilField: "done", UntilEquals: "yes"}},
+				), &iterBackend{calls: map[string]int{}, content: func(_ string, _ int, target string) []byte {
+					if target == "fields.json" {
+						return fieldsJSON(map[string]string{"done": "no"})
+					}
+					return nil
+				}}, RunOptions{}, RunCompleted)
+			case TopologyRouter:
+				runAgg(t, spec.Name, mockScriptPlanner(routerSteps()...),
+					&iterBackend{calls: map[string]int{}, content: func(_ string, _ int, target string) []byte {
+						if target == "fields.json" {
+							return fieldsJSON(map[string]string{"route": "fast"})
+						}
+						return nil
+					}}, RunOptions{}, RunCompleted)
+			case TopologyEvaluatorOptimizer:
+				runAgg(t, spec.Name, mockScriptPlanner(
+					StepSpec{ID: "draft", Role: "writer", Title: "起草", Targets: []string{"output.md"}},
+					StepSpec{ID: "eval", Role: "reviewer", Title: "评审", Targets: []string{"verdict.md", "fields.json"},
+						Loop: &LoopSpec{MaxIterations: 4, UntilField: "verdict", UntilEquals: "pass", Body: []string{"draft"}}},
+				), &iterBackend{calls: map[string]int{}, content: func(_ string, iter int, target string) []byte {
+					if target == "fields.json" {
+						if iter >= 3 {
+							return fieldsJSON(map[string]string{"verdict": "pass"})
+						}
+						return fieldsJSON(map[string]string{"verdict": "fail"})
+					}
+					return nil
+				}}, RunOptions{}, RunCompleted)
+			case TopologyLongRunning:
+				runAgg(t, spec.Name, mockScriptPlanner(
+					StepSpec{ID: "s1", Role: "researcher", Title: "一", Targets: []string{"output.md"}},
+				), &iterBackend{calls: map[string]int{}},
+					RunOptions{Deadline: time.Now().Add(5 * time.Second)}, RunCompleted)
+			case TopologyLeader:
+				b := &leaderScriptBackend{
+					decisions: []string{delegateJSON("writer", "撰写简报"), closeJSON},
+					script:    map[string]map[string]string{"01_writer": {"output.md": goodReport}},
+				}
+				engine := newLeaderEngine(t, b)
+				res, err := engine.Run(context.Background(), Goal{Text: "汇总"}, RunOptions{Topology: TopologyLeader})
+				if err != nil {
+					t.Fatal(err)
+				}
+				if res.Status != RunCompleted {
+					t.Fatalf("status = %s err = %v", res.Status, res.Error)
+				}
+			case TopologyReview:
+				planner := mockScriptPlanner(
+					StepSpec{ID: "f1", Role: "researcher", Title: "评审", Targets: []string{"findings.json"}},
+				)
+				planner.Agents = []RoleSpec{
+					{Role: "researcher", Name: "评审", Backend: "agg-review"},
+					{Role: "verifier", Name: "核实", Backend: "agg-review"},
+				}
+				engine := newAggEngine(t, planner, &aggReviewBackend{}, []RoleSpec{
+					{Role: "researcher", Name: "评审", Backend: "agg-review"},
+					{Role: "verifier", Name: "核实", Backend: "agg-review"},
+				})
+				res, err := engine.Run(context.Background(), Goal{Text: "评审"},
+					RunOptions{Topology: TopologyReview, ReviewTarget: "材料"})
+				if err != nil {
+					t.Fatal(err)
+				}
+				if res.Status != RunCompleted {
+					t.Fatalf("status = %s err = %v", res.Status, res.Error)
+				}
+				data, _ := os.ReadFile(filepath.Join(res.WorkDir, "output.md"))
+				if !strings.Contains(string(data), "已丢弃 0 条未验证") || !strings.Contains(string(data), "样例问题") {
+					t.Fatalf("报告 = %s", data)
+				}
+			case TopologyGroupChat:
+				engine := newAggEngine(t, mockScriptPlanner(
+					StepSpec{ID: "p1", Role: "researcher", Title: "甲"},
+					StepSpec{ID: "p2", Role: "writer", Title: "乙"},
+				), &gcScriptBackend{}, []RoleSpec{
+					{Role: "researcher", Name: "甲", Backend: "gc-script"},
+					{Role: "writer", Name: "乙", Backend: "gc-script"},
+				})
+				res, err := engine.Run(context.Background(), Goal{Text: "群聊"},
+					RunOptions{Topology: TopologyGroupChat, GroupChatMaxTurns: 2})
+				if err != nil {
+					t.Fatal(err)
+				}
+				if res.Status != RunCompleted {
+					t.Fatalf("status = %s err = %v", res.Status, res.Error)
+				}
+			case TopologyTaskBoard:
+				runAgg(t, spec.Name, mockScriptPlanner(
+					StepSpec{ID: "t1", Role: "researcher", Title: "一", Targets: []string{"a.md"}},
+					StepSpec{ID: "t2", Role: "writer", Title: "二", Needs: []string{"t1"}, Targets: []string{"output.md"}},
+				), &iterBackend{calls: map[string]int{}}, RunOptions{}, RunCompleted)
+			case TopologyHandoff, TopologyMailbox:
+				// primitive：不可单独作为运行拓扑（fail-closed）；原语能力由
+				// 专项 conformance 钉住（handoff/mailbox 测试）。
+				runAgg(t, spec.Name, mockScriptPlanner(
+					StepSpec{ID: "s1", Role: "researcher", Title: "一", Targets: []string{"output.md"}},
+				), &iterBackend{calls: map[string]int{}}, RunOptions{}, RunFailed)
+			default:
+				t.Fatalf("未覆盖的拓扑 %q", spec.Name)
+			}
+		})
+	}
+}
+
+// runAgg 汇总跑的最小引擎：graph 类拓扑共用 iter 三角色装配。
+func runAgg(t *testing.T, topo string, planner Planner, be Executor, opts RunOptions, want RunStatus) {
+	t.Helper()
+	opts.Topology = topo
+	engine := newTopologyEngine(t, planner, be, nil)
+	res, err := engine.Run(context.Background(), Goal{Text: "汇总跑 " + topo}, opts)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != want {
+		t.Fatalf("status = %s err = %q，期望 %s", res.Status, res.Error, want)
+	}
+}
+
+// aggReviewBackend 汇总跑用的最小 review 后端：finder 出一条发现，verify 判 PASS。
+type aggReviewBackend struct{}
+
+func (b *aggReviewBackend) Name() string { return "agg-review" }
+
+func (b *aggReviewBackend) Execute(_ context.Context, req ExecRequest) (*ExecResult, error) {
+	key := filepath.Base(filepath.Clean(req.WorkDir))
+	if strings.HasPrefix(key, "verify_") {
+		return writeOne(req.WorkDir, "verdict.md", "核对成立。\nVERDICT: PASS"), nil
+	}
+	return writeOne(req.WorkDir, "findings.json",
+		`[{"location":"第1段","severity":"high","issue":"样例问题","evidence":"样例证据"}]`), nil
+}
+
+// newAggEngine 自定义角色表的汇总引擎（backend 实例名 = be.Name()）。
+func newAggEngine(t *testing.T, planner Planner, be Executor, roles []RoleSpec) *SquadEngine {
+	t.Helper()
+	dir := t.TempDir()
+	reg := NewRoleRegistry()
+	for _, r := range roles {
+		if err := reg.Register(r); err != nil {
+			t.Fatal(err)
+		}
+	}
+	backends := NewBackendResolver()
+	backends.RegisterInstance(be.Name(), be)
+	engine, err := NewEngine(Options{DataDir: dir, Registry: reg, Planner: planner, Backends: backends, Sink: NopSink{}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	return engine
+}