Add leader progress ledger, replan cap, and mailbox with actor attribution (Q2 WP2)

Why: Q2 acceptance requires leader runs to detect stuck delegation loops and
cap replans instead of spinning to the turn limit, and a mailbox primitive
whose messages can never resolve a pending permission approval (acceptance 3).

Evidence: squad conformance tests (leader_mailbox_conformance_test.go):
ledger stuck->replan, replan cap finalize, mailbox delivery after hold,
mailbox-cannot-approve-permission (structural + actor second defense),
post guards; full suite green with -race.

Impact: LeaderState gains Progress/StuckTurns/Replans/MaxReplans/ReplanNote
(persisted in leader.json); RunState gains Mailbox (persisted in
runstate.json); RespondDecision gains Actor - non-human actors are rejected
on choice gates; step briefs and leader task briefs include pending mail.

Refs: Q2 plan WP2 (2026-10-07-agent-stack-q2q3-plans.md); task card
.task/feature-agent-stack-q2.md

adapters/intent/gate.go

@@ -33,11 +33,13 @@ type Options struct {
 
 // NewFromOverlay 从 JSON overlay 构造闸门，并把 guard 域标成受保护域。
 // 命令行用它，这样命令本身不必 import agent-intent。
+// NFKC 开启：全角/兼容等价输入与词条同形折叠（与 backend intentgate 一致）。
 func NewFromOverlay(ctx context.Context, fsys fs.FS, roots ...string) (*Gate, error) {
 	return New(ctx, Options{
 		IntentOptions: intent.Options{
 			GuardDomains:     []string{"guard"},
 			ProtectedDomains: []string{"guard"},
+			NFKC:             true,
 		},
 		FS:    fsys,
 		Roots: roots,

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,
 			})
 		}
 	}

adapters/sqlite/go.mod

@@ -4,10 +4,12 @@ go 1.27
 
 require (
 	agent-native v0.0.0
+	agent-squad v0.0.0
 	modernc.org/sqlite v1.37.0
 )
 
 require (
+	error-code-system v0.0.0 // indirect
 	github.com/dustin/go-humanize v1.0.1 // indirect
 	github.com/google/uuid v1.6.0 // indirect
 	github.com/mattn/go-isatty v0.0.20 // indirect

adapters/sqlite/platform_test.go

@@ -11,7 +11,7 @@ import (
 	"testing"
 	"time"
 
-	"agent-native/runtime/llm"
+	agentsquad "agent-squad/squad"
 )
 
 func TestOrchestrationFlow(t *testing.T) {
@@ -116,12 +116,10 @@ func TestOrchestrationFlow(t *testing.T) {
 	if err != nil {
 		t.Fatal(err)
 	}
-	text, err := s.RunInProcess(context.Background(), woke, work, func(ctx context.Context, model string, messages []llm.Message, td []llm.ToolDecl, opts llm.StreamOptions) <-chan llm.Event {
-		ch := make(chan llm.Event, 1)
-		ch <- llm.Event{Kind: llm.KindMessageEnd, Message: llm.Message{Role: llm.RoleAssistant, Content: "done", StopReason: llm.StopEnd}}
-		close(ch)
-		return ch
-	})
+	text, err := s.RunInProcess(context.Background(), woke, work, stubExecutor{result: &agentsquad.ExecResult{
+		Success: true, Logs: "done",
+		Patches: map[string][]byte{"output.md": []byte("done")},
+	}})
 	if err != nil || text != "done" {
 		t.Fatalf("in-process %q %v", text, err)
 	}

adapters/sqlite/runstore.go

@@ -0,0 +1,230 @@
+package platform
+
+import (
+	"database/sql"
+	"encoding/json"
+	"fmt"
+	"os"
+	"time"
+
+	"agent-squad/squad"
+)
+
+// SQLiteStore 把 squad 的 Store 端口落到本库（Q1 WP5）：run/gate/checkpoint/leader
+// 与平台其余表同一个 SQLite 文件——运行状态与平台状态不再各记一份。
+// 表名带 squad_run_ 前缀，与平台既有 23 张表并存不冲突。
+type SQLiteStore struct {
+	db *sql.DB
+}
+
+// OpenRunStore 在本 Store 的底层连接上提供 squad.Store 端口（schema 随 migrate 建）。
+func (s *Store) RunStore() *SQLiteStore {
+	return &SQLiteStore{db: s.db}
+}
+
+const runSchema = `
+CREATE TABLE IF NOT EXISTS squad_run_state (
+  ws_dir TEXT PRIMARY KEY,
+  meta TEXT NOT NULL, workflow TEXT NOT NULL, squad TEXT NOT NULL, bb TEXT NOT NULL,
+  template TEXT, approved TEXT, repair_budget_remaining INTEGER NOT NULL DEFAULT 0,
+  repair_budget_granted INTEGER NOT NULL DEFAULT 0,
+  clarifications TEXT, steering TEXT, usage TEXT, updated_at TEXT NOT NULL
+);
+CREATE TABLE IF NOT EXISTS squad_run_gate (
+  ws_dir TEXT PRIMARY KEY, gate TEXT NOT NULL
+);
+CREATE TABLE IF NOT EXISTS squad_run_checkpoint (
+  ws_dir TEXT NOT NULL, superstep INTEGER NOT NULL, at TEXT NOT NULL,
+  graph_hash TEXT NOT NULL, steps TEXT NOT NULL,
+  PRIMARY KEY (ws_dir, superstep)
+);
+CREATE TABLE IF NOT EXISTS squad_run_leader (
+  ws_dir TEXT PRIMARY KEY, data TEXT NOT NULL
+);
+`
+
+func (s *SQLiteStore) migrate() error {
+	_, err := s.db.Exec(runSchema)
+	return err
+}
+
+// OpenSQLiteStore 独立打开一个 squad.Store（不经 platform.Store 的场景，如测试）。
+func OpenSQLiteStore(path string) (*SQLiteStore, error) {
+	db, err := sql.Open("sqlite", path+"?_pragma=busy_timeout(5000)&_pragma=journal_mode(WAL)")
+	if err != nil {
+		return nil, err
+	}
+	s := &SQLiteStore{db: db}
+	if err := s.migrate(); err != nil {
+		_ = db.Close()
+		return nil, err
+	}
+	return s, nil
+}
+
+func j(v any) string {
+	b, err := json.Marshal(v)
+	if err != nil {
+		return ""
+	}
+	return string(b)
+}
+
+func (s *SQLiteStore) SaveRun(st *squad.RunState) error {
+	if err := s.migrate(); err != nil {
+		return err
+	}
+	_, err := s.db.Exec(`INSERT INTO squad_run_state
+	  (ws_dir, meta, workflow, squad, bb, template, approved, repair_budget_remaining, repair_budget_granted, clarifications, steering, usage, updated_at)
+	  VALUES (?,?,?,?,?,?,?,?,?,?,?,?,?)
+	  ON CONFLICT(ws_dir) DO UPDATE SET
+	    meta=excluded.meta, workflow=excluded.workflow, squad=excluded.squad, bb=excluded.bb,
+	    template=excluded.template, approved=excluded.approved,
+	    repair_budget_remaining=excluded.repair_budget_remaining,
+	    repair_budget_granted=excluded.repair_budget_granted,
+	    clarifications=excluded.clarifications, steering=excluded.steering,
+	    usage=excluded.usage, updated_at=excluded.updated_at`,
+		st.Meta.WorkDir, j(st.Meta), j(st.Plan.Workflow), j(st.Plan.Squad), j(st.BB),
+		st.Template, j(st.Approved), st.RepairBudgetRemaining, st.RepairBudgetGranted,
+		j(st.Clarifications), j(st.Steering), j(st.Usage), now())
+	return err
+}
+
+func (s *SQLiteStore) LoadRun(wsDir string) (*squad.RunState, error) {
+	if err := s.migrate(); err != nil {
+		return nil, err
+	}
+	st := &squad.RunState{}
+	var meta, wf, sq, bb, template, approved, clar, steering, usage string
+	row := s.db.QueryRow(`SELECT meta, workflow, squad, bb, template, approved,
+	  repair_budget_remaining, repair_budget_granted, clarifications, steering, usage
+	  FROM squad_run_state WHERE ws_dir = ?`, wsDir)
+	var granted int
+	err := row.Scan(&meta, &wf, &sq, &bb, &template, &approved,
+		&st.RepairBudgetRemaining, &granted, &clar, &steering, &usage)
+	if err != nil {
+		if err == sql.ErrNoRows {
+			return nil, fmt.Errorf("load run %s: %w", wsDir, os.ErrNotExist)
+		}
+		return nil, err
+	}
+	st.RepairBudgetGranted = granted
+	for pair, dst := range map[string]any{
+		meta: &st.Meta, wf: &st.Plan.Workflow, sq: &st.Plan.Squad, bb: &st.BB,
+		approved: &st.Approved, clar: &st.Clarifications, steering: &st.Steering, usage: &st.Usage,
+	} {
+		if err := json.Unmarshal([]byte(pair), dst); err != nil {
+			return nil, fmt.Errorf("load run %s: parse: %w", wsDir, err)
+		}
+	}
+	st.Template = template
+	if st.Approved == nil {
+		st.Approved = map[string]bool{}
+	}
+	return st, nil
+}
+
+func (s *SQLiteStore) SaveGate(wsDir string, gate *squad.HitlGate) error {
+	if err := s.migrate(); err != nil {
+		return err
+	}
+	if gate == nil {
+		_, err := s.db.Exec(`DELETE FROM squad_run_gate WHERE ws_dir = ?`, wsDir)
+		return err
+	}
+	_, err := s.db.Exec(`INSERT INTO squad_run_gate (ws_dir, gate) VALUES (?,?)
+	  ON CONFLICT(ws_dir) DO UPDATE SET gate=excluded.gate`, wsDir, j(gate))
+	return err
+}
+
+func (s *SQLiteStore) LoadGate(wsDir string) (*squad.HitlGate, error) {
+	if err := s.migrate(); err != nil {
+		return nil, err
+	}
+	var raw string
+	err := s.db.QueryRow(`SELECT gate FROM squad_run_gate WHERE ws_dir = ?`, wsDir).Scan(&raw)
+	if err != nil {
+		if err == sql.ErrNoRows {
+			return nil, nil
+		}
+		return nil, err
+	}
+	gate := &squad.HitlGate{}
+	if err := json.Unmarshal([]byte(raw), gate); err != nil {
+		return nil, fmt.Errorf("parse gate: %w", err)
+	}
+	return gate, nil
+}
+
+func (s *SQLiteStore) SaveCheckpoint(wsDir string, cp squad.Checkpoint) error {
+	if err := s.migrate(); err != nil {
+		return err
+	}
+	_, err := s.db.Exec(`INSERT INTO squad_run_checkpoint (ws_dir, superstep, at, graph_hash, steps)
+	  VALUES (?,?,?,?,?)
+	  ON CONFLICT(ws_dir, superstep) DO UPDATE SET at=excluded.at, graph_hash=excluded.graph_hash, steps=excluded.steps`,
+		wsDir, cp.Superstep, cp.At.UTC().Format("2006-01-02T15:04:05Z07:00"), cp.GraphHash, j(cp.StepStates))
+	return err
+}
+
+func (s *SQLiteStore) LoadLatestCheckpoint(wsDir string) (*squad.Checkpoint, error) {
+	if err := s.migrate(); err != nil {
+		return nil, err
+	}
+	var cp squad.Checkpoint
+	var at, steps string
+	err := s.db.QueryRow(`SELECT superstep, at, graph_hash, steps FROM squad_run_checkpoint
+	  WHERE ws_dir = ? ORDER BY superstep DESC LIMIT 1`, wsDir).
+		Scan(&cp.Superstep, &at, &cp.GraphHash, &steps)
+	if err != nil {
+		if err == sql.ErrNoRows {
+			return nil, nil
+		}
+		return nil, err
+	}
+	t, _ := parseTime(at)
+	cp.At = t
+	if err := json.Unmarshal([]byte(steps), &cp.StepStates); err != nil {
+		return nil, fmt.Errorf("parse checkpoint steps: %w", err)
+	}
+	return &cp, nil
+}
+
+func (s *SQLiteStore) SaveLeader(wsDir string, data []byte) error {
+	if err := s.migrate(); err != nil {
+		return err
+	}
+	_, err := s.db.Exec(`INSERT INTO squad_run_leader (ws_dir, data) VALUES (?,?)
+	  ON CONFLICT(ws_dir) DO UPDATE SET data=excluded.data`, wsDir, string(data))
+	return err
+}
+
+func (s *SQLiteStore) LoadLeader(wsDir string) ([]byte, error) {
+	if err := s.migrate(); err != nil {
+		return nil, err
+	}
+	var raw string
+	err := s.db.QueryRow(`SELECT data FROM squad_run_leader WHERE ws_dir = ?`, wsDir).Scan(&raw)
+	if err != nil {
+		if err == sql.ErrNoRows {
+			return nil, nil
+		}
+		return nil, err
+	}
+	return []byte(raw), nil
+}
+
+// Close 只在独立打开（OpenSQLiteStore）时调用；经 platform.Store 打开时由平台统一关闭。
+func (s *SQLiteStore) Close() error { return s.db.Close() }
+
+var _ squad.Store = (*SQLiteStore)(nil)
+
+// parseTime 解析检查点时间（RFC3339Nano 与 RFC3339 两态）。
+func parseTime(s string) (time.Time, error) {
+	for _, layout := range []string{time.RFC3339Nano, time.RFC3339} {
+		if t, err := time.Parse(layout, s); err == nil {
+			return t, nil
+		}
+	}
+	return time.Time{}, fmt.Errorf("bad time %q", s)
+}

adapters/sqlite/runstore_test.go

@@ -0,0 +1,188 @@
+package platform
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"testing"
+	"time"
+
+	"agent-squad/squad"
+)
+
+func newTestRunState(wsDir string) *squad.RunState {
+	return &squad.RunState{
+		Meta: squad.RunMetadata{
+			ID: "run_test", Status: "executing", WorkDir: wsDir,
+			CreatedAt: time.Now(), UpdatedAt: time.Now(),
+			Metadata: map[string]string{"mode": ""},
+		},
+		Plan: squad.SquadPlan{
+			Workflow: squad.WorkflowPlan{
+				WorkflowID: "wf1", Status: "in_progress",
+				Steps: []squad.StepSpec{
+					{ID: "s1", Role: "researcher", Status: "completed", Attempts: 1, AttemptID: "run_test/s1/1"},
+					{ID: "s2", Role: "writer", Status: "pending"},
+				},
+			},
+			Squad: squad.SquadConfig{SquadName: "demo"},
+		},
+		BB:                    squad.BlackboardState{Facts: []squad.FactItem{{ID: "f1", Claim: "x"}}},
+		Template:              "tpl",
+		Approved:              map[string]bool{"step:s1": true},
+		RepairBudgetRemaining: 1,
+		RepairBudgetGranted:   0,
+		Clarifications:        []string{"c1"},
+		Steering:              []string{"st1"},
+		Usage:                 map[string]*squad.Usage{"writer": {PromptTokens: 10, CompletionTokens: 5, Steps: 1}},
+	}
+}
+
+// TestSQLiteStoreRoundTrip：run/gate/checkpoint/leader 全链路（Q1 WP5 端口实现）。
+func TestSQLiteStoreRoundTrip(t *testing.T) {
+	st, err := OpenSQLiteStore(filepath.Join(t.TempDir(), "runs.db"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer st.Close()
+	ws := filepath.Join(t.TempDir(), "runs", "run_test")
+	orig := newTestRunState(ws)
+	if err := st.SaveRun(orig); err != nil {
+		t.Fatal(err)
+	}
+	got, err := st.LoadRun(ws)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if got.Meta.ID != "run_test" || got.Plan.Workflow.WorkflowID != "wf1" || got.Template != "tpl" {
+		t.Fatalf("loaded = %+v", got.Meta)
+	}
+	if !got.Approved["step:s1"] || got.RepairBudgetRemaining != 1 || len(got.Clarifications) != 1 || len(got.Steering) != 1 {
+		t.Fatalf("scalars = %+v", got)
+	}
+	if got.Usage["writer"].PromptTokens != 10 {
+		t.Fatalf("usage = %+v", got.Usage)
+	}
+	if _, err := st.LoadRun(filepath.Join(t.TempDir(), "nope")); err == nil {
+		t.Fatal("missing run should error")
+	}
+
+	// gate：写→读→删。
+	gate := &squad.HitlGate{PrevStatus: squad.RunExecuting}
+	if err := st.SaveGate(ws, gate); err != nil {
+		t.Fatal(err)
+	}
+	if g2, err := st.LoadGate(ws); err != nil || g2 == nil {
+		t.Fatalf("load gate: %v %+v", err, g2)
+	}
+	if err := st.SaveGate(ws, nil); err != nil {
+		t.Fatal(err)
+	}
+	if g2, err := st.LoadGate(ws); err != nil || g2 != nil {
+		t.Fatalf("gate after delete: %v %+v", err, g2)
+	}
+
+	// checkpoint：两层，latest 取最大 superstep。
+	if err := st.SaveCheckpoint(ws, squad.Checkpoint{Superstep: 1, At: time.Now(), GraphHash: "h1", StepStates: []squad.StepSnapshot{{ID: "s1", Status: "completed", AttemptID: "run_test/s1/1"}}}); err != nil {
+		t.Fatal(err)
+	}
+	if err := st.SaveCheckpoint(ws, squad.Checkpoint{Superstep: 2, At: time.Now(), GraphHash: "h2"}); err != nil {
+		t.Fatal(err)
+	}
+	latest, err := st.LoadLatestCheckpoint(ws)
+	if err != nil || latest == nil {
+		t.Fatalf("latest: %v", err)
+	}
+	if latest.Superstep != 2 || latest.GraphHash != "h2" {
+		t.Fatalf("latest = %+v", latest)
+	}
+	if l2, err := st.LoadLatestCheckpoint(filepath.Join(t.TempDir(), "nope")); err != nil || l2 != nil {
+		t.Fatalf("latest for unknown: %v %+v", err, l2)
+	}
+
+	// leader bytes。
+	if err := st.SaveLeader(ws, []byte(`{"turns":2}`)); err != nil {
+		t.Fatal(err)
+	}
+	data, err := st.LoadLeader(ws)
+	if err != nil || string(data) != `{"turns":2}` {
+		t.Fatalf("leader = %q %v", data, err)
+	}
+}
+
+// TestEngineRunsOnSQLiteStore：引擎注入 SQLiteStore 后运行状态进 SQLite、
+// 不再写 run 级 JSON 文件；第二引擎可从中恢复（验收第三条的实证形态）。
+func TestEngineRunsOnSQLiteStore(t *testing.T) {
+	dbPath := filepath.Join(t.TempDir(), "runs.db")
+	runStore, err := OpenSQLiteStore(dbPath)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer runStore.Close()
+	dir := t.TempDir()
+	reg := squad.NewRoleRegistry()
+	for _, spec := range []squad.RoleSpec{{Role: "researcher", Backend: "script"}, {Role: "writer", Backend: "script"}, {Role: "reviewer", Backend: "script"}} {
+		if err := reg.Register(spec); err != nil {
+			t.Fatal(err)
+		}
+	}
+	backends := squad.NewBackendResolver()
+	backends.RegisterInstance("script", stubExecutor{result: &squad.ExecResult{Success: false, Logs: "ok", Patches: map[string][]byte{"output.md": []byte(goodReportLite), "notes.md": []byte(goodReportLite), "fields.json": []byte("{}")}}})
+	blueprints := squad.NewBlueprintRegistry()
+	if err := blueprints.Register(squad.Blueprint{
+		Key: "demo", Title: "入库演示", Match: []string{"入库演示"},
+		Steps: []squad.StepSpec{
+			{ID: "s1", Role: "researcher", Title: "调研", Targets: []string{"notes.md", "fields.json"}},
+			{ID: "s2", Role: "writer", Title: "成文", Targets: []string{"output.md"}},
+			{ID: "s3", Role: "reviewer", Title: "终审", Targets: []string{"output.md"}},
+		},
+	}); err != nil {
+		t.Fatal(err)
+	}
+	eng, err := squad.NewEngine(squad.Options{
+		DataDir: dir, Registry: reg, Blueprints: blueprints,
+		Planner:  squad.NewRulePlanner(),
+		Backends: backends, Sink: squad.NopSink{},
+		Store: runStore,
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	res, err := eng.Run(context.Background(), squad.Goal{Text: "做一个三步入库演示", HitlMode: squad.HitlAuto}, squad.RunOptions{})
+	if err != nil || res.Status != squad.RunCompleted {
+		t.Fatalf("run: %v %+v", err, res)
+	}
+	// run 级 JSON 不在盘上（状态全进了 SQLite）。
+	for _, f := range []string{"workspace.json", "workflow.json", "squad.json", "blackboard.json", "runstate.json"} {
+		if _, err := os.Stat(filepath.Join(res.WorkDir, f)); !os.IsNotExist(err) {
+			t.Fatalf("%s should not be on disk with SQLite store (err=%v)", f, err)
+		}
+	}
+	// 状态可经端口读回。
+	st, err := runStore.LoadRun(res.WorkDir)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if st.Plan.Workflow.Status != "completed" {
+		t.Fatalf("workflow status = %s", st.Plan.Workflow.Status)
+	}
+	latest, err := runStore.LoadLatestCheckpoint(res.WorkDir)
+	if err != nil || latest == nil {
+		t.Fatalf("latest checkpoint: %v", err)
+	}
+	// 第二引擎（同库）恢复可用。
+	eng2, err := squad.NewEngine(squad.Options{
+		DataDir: dir, Registry: reg, Blueprints: blueprints,
+		Planner:  squad.NewRulePlanner(),
+		Backends: backends, Sink: squad.NopSink{},
+		Store: runStore,
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := eng2.Inspect(res.RunID); err != nil {
+		t.Fatalf("inspect from second engine: %v", err)
+	}
+}
+
+const goodReportLite = "# 测试交付物\n\n## 概述\n\n这是由脚本后端生成的合格交付物，长度超过五十个字符以满足最小质检要求。\n"

adapters/sqlite/runtime.go

@@ -9,8 +9,7 @@ import (
 	"path/filepath"
 	"strings"
 
-	"agent-native/runtime"
-	"agent-native/runtime/llm"
+	"agent-squad/squad"
 )
 
 // CLICommands are the external tools a daemon probes, in order.
@@ -196,36 +195,54 @@ func (s *Store) RunCommand(ctx context.Context, task Task, command, workdir stri
 	return s.FinishTask(task.ID, "completed", "")
 }
 
-// RunInProcess executes the task with AgentSession. stream is required; the
-// product supplies a provider, tests supply a script.
-func (s *Store) RunInProcess(ctx context.Context, task Task, workdir string, stream llm.StreamFunc) (string, error) {
+// RunInProcess executes the task through the squad.Executor port（Q1 WP5 修 X6：
+// 不再绕过执行端口直驱 agent-native runtime）。调用方经 adapters/native 等
+// 适配器构造 Executor；nil 即 fail-closed。
+func (s *Store) RunInProcess(ctx context.Context, task Task, workdir string, exec squad.Executor) (string, error) {
+	if exec == nil {
+		return "", fmt.Errorf("RunInProcess requires an executor (no bypass)")
+	}
 	if err := s.MarkRunning(task.ID); err != nil {
 		return "", err
 	}
 	issue, _ := s.GetIssue(task.IssueID)
-	agent := runtime.NewAgentSession(runtime.Config{
-		Model:    "agent-native",
-		Stream:   stream,
-		MaxTurns: 4,
-		OnEvent: func(ev runtime.Event) {
+	res, err := exec.Execute(ctx, squad.ExecRequest{
+		WorkDir: workdir,
+		Prompt:  issue.Title + "\n" + issue.Body,
+		Targets: []string{"output.md"},
+		OnEvent: func(ev squad.ProcessEvent) {
 			_ = s.AddTaskMessage(task.ID, ev.Type, ev.Text)
 		},
 	})
-	res, err := agent.Prompt(ctx, llm.Message{Role: llm.RoleUser, Content: issue.Title + "\n" + issue.Body})
 	if err != nil {
 		_ = s.FinishTask(task.ID, "failed", "config")
 		return "", err
 	}
-	if res.Err != nil {
-		_ = s.FinishTask(task.ID, "failed", "model")
-		return res.Text, res.Err
+	if res == nil || !res.Success {
+		reason := "model"
+		if res != nil && res.ErrorReason != "" {
+			reason = res.ErrorReason
+		}
+		_ = s.FinishTask(task.ID, "failed", reason)
+		if res != nil {
+			return res.Logs, fmt.Errorf("executor failed: %s", res.ErrorReason)
+		}
+		return "", fmt.Errorf("executor failed")
 	}
-	_ = s.AddUsage(task.ID, "agent-native", res.Usage.PromptTokens, res.Usage.CompletionTokens, res.Usage.CacheReadTokens, res.Usage.CacheWriteTokens)
+	if len(res.Patches) > 0 {
+		_, _ = squad.MaterializeTargets(workdir, []string{"output.md"}, res.Patches)
+	}
+	_ = s.AddUsage(task.ID, exec.Name(), res.PromptTokens, res.CompletionTokens, 0, 0)
 	if err := s.FinishTask(task.ID, "completed", ""); err != nil {
-		return res.Text, err
+		return res.Logs, err
 	}
-	_ = os.WriteFile(filepath.Join(workdir, "output.md"), []byte(res.Text), 0o644)
-	return res.Text, nil
+	if data, rerr := os.ReadFile(filepath.Join(workdir, "output.md")); rerr == nil {
+		return string(data), nil
+	}
+	if err := os.WriteFile(filepath.Join(workdir, "output.md"), []byte(res.Logs), 0o644); err != nil {
+		return res.Logs, err
+	}
+	return res.Logs, nil
 }
 
 // AgentCLI applies a task-token command: comment, status, sub-issue, or activity.

adapters/sqlite/stub_executor_test.go

@@ -0,0 +1,30 @@
+package platform
+
+import (
+	"context"
+
+	"agent-squad/squad"
+)
+
+// stubExecutor 是测试用的 squad.Executor（X6 后调用方经端口注入）。
+type stubExecutor struct {
+	result *squad.ExecResult
+	err    error
+}
+
+func (s stubExecutor) Name() string { return "stub" }
+
+func (s stubExecutor) Execute(ctx context.Context, req squad.ExecRequest) (*squad.ExecResult, error) {
+	if s.result == nil {
+		return nil, s.err
+	}
+	// 每次调用返回拷贝：调用方会改写 Success 字段，共享指针会跨步骤串味。
+	cp := *s.result
+	if s.result.Patches != nil {
+		cp.Patches = map[string][]byte{}
+		for k, v := range s.result.Patches {
+			cp.Patches[k] = v
+		}
+	}
+	return &cp, s.err
+}

cmd/agent-squad/cli/cli.go

@@ -86,6 +86,7 @@ type app struct {
 	files  *store.FileEventSink
 	st     *store.FileStore
 	log    *logsys.Logger
+	plat   *platform.Store
 }
 
 func dataDir() string {
@@ -159,12 +160,20 @@ func assemble(dir, plannerName, modelOverride, intentRoot string) (*app, error)
 			}
 		}
 	}
+	// 单库收编（Q1 WP5）：引擎运行状态与平台编排状态进同一个 SQLite。
+	// 库打不开即失败（与 serve 现状一致），不允许静默退回文件态。
+	plat, err := platform.Open(filepath.Join(dir, "orchestration.db"))
+	if err != nil {
+		return nil, fmt.Errorf("编排库打开失败: %w", err)
+	}
 	engine, err := squad.NewEngine(squad.Options{
 		DataDir: dir, Registry: reg, Blueprints: bps, Planner: planner,
 		Backends: backends, Sink: fileSink, IntentGate: gate,
 		IntentCatalog: catalog, IntentRunner: runner,
+		Store: plat.RunStore(),
 	})
 	if err != nil {
+		_ = plat.Close()
 		return nil, err
 	}
 	logCfg, err := config.Load("")
@@ -178,7 +187,7 @@ func assemble(dir, plannerName, modelOverride, intentRoot string) (*app, error)
 	if err != nil {
 		return nil, err
 	}
-	return &app{engine: engine, sink: fanout, files: fileSink, st: st, log: logger}, nil
+	return &app{engine: engine, sink: fanout, files: fileSink, st: st, log: logger, plat: plat}, nil
 }
 
 func (a *app) close() {
@@ -223,12 +232,9 @@ func cmdServe(args []string) int {
 	srv := server.New(a.engine, a.sink, a.st)
 	srv.SetToken(*token)
 	srv.SetLogger(a.log)
-	if plat, err := platform.Open(filepath.Join(*data, "orchestration.db")); err != nil {
-		fmt.Fprintln(os.Stderr, "编排库打开失败:", err)
-		return 1
-	} else {
-		srv.SetPlatform(plat)
-		defer plat.Close()
+	if a.plat != nil {
+		srv.SetPlatform(a.plat)
+		defer a.plat.Close()
 	}
 	stopSweep := srv.StartWakeupSweeper(15 * time.Second)
 	defer stopSweep()

cmd/agent-squad/cli/platform_cmd.go

@@ -9,8 +9,8 @@ import (
 	"path/filepath"
 	"time"
 
-	"agent-native/runtime/llm"
 	platform "agent-squad/adapters/sqlite"
+	"agent-squad/squad"
 )
 
 func cmdPlatform(args []string) int {
@@ -99,7 +99,7 @@ func cmdPlatform(args []string) int {
 				return 1
 			}
 		}
-		text, err := st.RunInProcess(context.Background(), task, fs.Arg(1), scriptedOK)
+		text, err := st.RunInProcess(context.Background(), task, fs.Arg(1), scriptedExecutor{})
 		return printPlatform(map[string]string{"text": text}, err)
 	default:
 		fmt.Fprintln(os.Stderr, "未知编排命令", args[0], args[1])
@@ -111,11 +111,16 @@ func taskByID(st *platform.Store, id string) (platform.Task, error) {
 	return st.TaskByToken(id)
 }
 
-func scriptedOK(ctx context.Context, model string, messages []llm.Message, td []llm.ToolDecl, opts llm.StreamOptions) <-chan llm.Event {
-	ch := make(chan llm.Event, 1)
-	ch <- llm.Event{Kind: llm.KindMessageEnd, Message: llm.Message{Role: llm.RoleAssistant, Content: "ok", StopReason: llm.StopEnd}}
-	close(ch)
-	return ch
+// scriptedExecutor 是 daemon run-native 冒烟路径的脚本执行器（X6 后经端口注入）。
+type scriptedExecutor struct{}
+
+func (scriptedExecutor) Name() string { return "scripted" }
+
+func (scriptedExecutor) Execute(ctx context.Context, req squad.ExecRequest) (*squad.ExecResult, error) {
+	return &squad.ExecResult{
+		Success: true, Logs: "ok",
+		Patches: map[string][]byte{"output.md": []byte("ok")},
+	}, nil
 }
 
 func printPlatform(v any, err error) int {

cmd/agent-squad/server/server.go

@@ -462,6 +462,8 @@ func (s *Server) StartWakeupSweeper(interval time.Duration) (stop func()) {
 				return
 			case <-t.C:
 				s.engine.SweepWakeups(context.Background())
+				// Q6：与事件唤醒并列，过期挂起按 ExpireAction 合成决策（此前是死代码）。
+				s.engine.SweepExpired(context.Background(), time.Now())
 			}
 		}
 	}()

squad/attempt_test.go

@@ -0,0 +1,173 @@
+package squad
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"sync"
+	"testing"
+)
+
+// perStepRecordingBackend 按子工作区归因计数（steps 的 role→tasks/NN_role 布局）。
+type perStepRecordingBackend struct {
+	mu    sync.Mutex
+	calls map[string]int
+}
+
+func (b *perStepRecordingBackend) Name() string { return "recording" }
+
+func (b *perStepRecordingBackend) Execute(ctx context.Context, req ExecRequest) (*ExecResult, error) {
+	b.mu.Lock()
+	if b.calls == nil {
+		b.calls = map[string]int{}
+	}
+	b.calls[filepath.Base(req.WorkDir)]++
+	b.mu.Unlock()
+	out := map[string][]byte{}
+	for _, t := range req.Targets {
+		out[t] = []byte(goodReport)
+	}
+	if _, err := MaterializeTargets(req.WorkDir, req.Targets, out); err != nil {
+		return &ExecResult{Success: false, ErrorReason: err.Error()}, nil
+	}
+	return &ExecResult{Success: true, Logs: "ok"}, nil
+}
+
+func (b *perStepRecordingBackend) countOf(sub string) int {
+	b.mu.Lock()
+	defer b.mu.Unlock()
+	return b.calls[sub]
+}
+
+func newAttemptTestEngine(t *testing.T, backend *perStepRecordingBackend, dir string) *SquadEngine {
+	t.Helper()
+	reg := NewRoleRegistry()
+	for _, role := range []string{"researcher", "writer", "reviewer"} {
+		if err := reg.Register(RoleSpec{Role: role, Backend: "recording"}); err != nil {
+			t.Fatal(err)
+		}
+	}
+	backends := NewBackendResolver()
+	backends.RegisterInstance("recording", backend)
+	eng, err := NewEngine(Options{
+		DataDir: dir, Registry: reg, Planner: mockScriptPlanner(baseSteps()...),
+		Backends: backends, Sink: NopSink{},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	return eng
+}
+
+func flipToInterrupted(t *testing.T, workDir string, stepIdx int) {
+	t.Helper()
+	st, err := loadRunState(workDir)
+	if err != nil {
+		t.Fatal(err)
+	}
+	st.Meta.Status = string(RunExecuting)
+	st.Meta.CompletedAt = nil
+	wf := st.Workflow()
+	wf.Status = "running"
+	wf.Steps[stepIdx].Status = "running"
+	wf.Steps[stepIdx].Attempts = 1
+	st.setWorkflow(wf)
+	if err := st.save(); err != nil {
+		t.Fatal(err)
+	}
+}
+
+// TestAttemptIDFormat：attempt id 稳定落盘且形如 runID/stepID/N（Q1 WP3）。
+func TestAttemptIDFormat(t *testing.T) {
+	backend := &perStepRecordingBackend{}
+	dir := t.TempDir()
+	eng := newAttemptTestEngine(t, backend, dir)
+	res, err := eng.Run(context.Background(), Goal{Text: "格式"}, RunOptions{})
+	if err != nil || res.Status != RunCompleted {
+		t.Fatalf("run: %v %+v", err, res)
+	}
+	if res.Plan == nil {
+		t.Fatal("no plan")
+	}
+	for _, s := range res.Plan.Steps {
+		want := res.RunID + "/" + s.ID + "/1"
+		if s.AttemptID != want {
+			t.Fatalf("step %s attempt id = %q, want %q", s.ID, s.AttemptID, want)
+		}
+	}
+}
+
+// TestKilledAtSuperstepBoundarySideEffectsOnce：杀进程于 superstep 边界后恢复，
+// 已完成步骤不重新执行（执行次数=1）；running 且产物齐全的步骤按 attempt 判定
+// 完成、不 rerun——Q1 验收第一条（WP3）。
+func TestKilledAtSuperstepBoundarySideEffectsOnce(t *testing.T) {
+	backend := &perStepRecordingBackend{}
+	dir := t.TempDir()
+	eng := newAttemptTestEngine(t, backend, dir)
+	res, err := eng.Run(context.Background(), Goal{Text: "先跑完再中断", HitlMode: HitlAuto}, RunOptions{})
+	if err != nil || res.Status != RunCompleted {
+		t.Fatalf("setup run: %v %+v", err, res)
+	}
+	before := map[string]int{}
+	st, err := loadRunState(res.WorkDir)
+	if err != nil {
+		t.Fatal(err)
+	}
+	for _, s := range st.Plan.Workflow.Steps {
+		before[s.ID] = backend.countOf(filepath.Base(filepath.FromSlash(s.SubWorkspace)))
+	}
+
+	// 第二步停在 running（产物其实在盘上——首次运行真实写出的）。
+	flipToInterrupted(t, res.WorkDir, 1)
+
+	eng2 := newAttemptTestEngine(t, backend, dir)
+	res, err = eng2.Resume(context.Background(), res.RunID)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunCompleted {
+		t.Fatalf("resume: %+v", res)
+	}
+	for _, s := range res.Plan.Steps {
+		if s.Status != "completed" && s.Status != "skipped" {
+			t.Fatalf("step %s not finished: %s", s.ID, s.Status)
+		}
+		got := backend.countOf(filepath.Base(filepath.FromSlash(s.SubWorkspace)))
+		if got != before[s.ID] {
+			t.Fatalf("step %s executed %d times after kill (before=%d)——已完成/running 步骤被重复执行", s.ID, got, before[s.ID])
+		}
+	}
+}
+
+// TestInterruptedMissingProductsRetries：running 崩溃且产物缺失时按 auto 一次重试（WP3）。
+func TestInterruptedMissingProductsRetries(t *testing.T) {
+	backend := &perStepRecordingBackend{}
+	dir := t.TempDir()
+	eng := newAttemptTestEngine(t, backend, dir)
+	res, err := eng.Run(context.Background(), Goal{Text: "先跑完再中断", HitlMode: HitlAuto}, RunOptions{})
+	if err != nil || res.Status != RunCompleted {
+		t.Fatalf("setup run: %v %+v", err, res)
+	}
+	// 第二步停 running 并删掉其交付物（副作用未发生的形态）。
+	flipToInterrupted(t, res.WorkDir, 1)
+	st, err := loadRunState(res.WorkDir)
+	if err != nil {
+		t.Fatal(err)
+	}
+	sub := filepath.FromSlash(st.Plan.Workflow.Steps[1].SubWorkspace)
+	if err := os.Remove(filepath.Join(res.WorkDir, sub, "output.md")); err != nil {
+		t.Fatal(err)
+	}
+	eng2 := newAttemptTestEngine(t, backend, dir)
+	res, err = eng2.Resume(context.Background(), res.RunID)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunCompleted {
+		t.Fatalf("resume: %+v", res)
+	}
+	got := backend.countOf(filepath.Base(sub))
+	if got != 2 {
+		t.Fatalf("missing-product step executed %d times, want 2（auto 一次重试）", got)
+	}
+}

squad/checkpoint_test.go

@@ -0,0 +1,113 @@
+package squad
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+// TestCheckpointFilesAtBoundaries：sync 默认档——每个 superstep 边界有快照，
+// latest 指向最后一层且图指纹与重编译一致（Q1 WP4）。
+func TestCheckpointFilesAtBoundaries(t *testing.T) {
+	eng, _ := newTestEngine(t, mockScriptPlanner(baseSteps()...), map[string]map[string]string{
+		"01_researcher": {"notes.md": goodReport, "fields.json": "{}"},
+		"02_writer":     {"output.md": goodReport},
+		"03_reviewer":   {"output.md": goodReport},
+	}, NopSink{})
+	res, err := eng.Run(context.Background(), Goal{Text: "快照"}, RunOptions{})
+	if err != nil || res.Status != RunCompleted {
+		t.Fatalf("run: %v %+v", err, res)
+	}
+	latest, err := FileStore{}.LoadLatestCheckpoint(res.WorkDir)
+	if err != nil || latest == nil {
+		t.Fatalf("latest checkpoint: %v", err)
+	}
+	if latest.Superstep < 3 {
+		t.Fatalf("superstep = %d, want >= 3（三步各占一层）", latest.Superstep)
+	}
+	g, err := compileGraph(res.Plan.Steps)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if latest.GraphHash != g.hash() {
+		t.Fatalf("hash = %s, want %s", latest.GraphHash, g.hash())
+	}
+	if len(latest.StepStates) != 3 {
+		t.Fatalf("step states = %d", len(latest.StepStates))
+	}
+	if _, err := os.Stat(filepath.Join(res.WorkDir, "checkpoints", "001.json")); err != nil {
+		t.Fatalf("first boundary snapshot missing: %v", err)
+	}
+}
+
+// TestCheckpointAsyncAndOnExit：async 在退出前落完；on_exit 只有最终快照。
+func TestCheckpointAsyncAndOnExit(t *testing.T) {
+	for _, mode := range []string{"async", "on_exit"} {
+		eng, _ := newTestEngine(t, mockScriptPlanner(baseSteps()...), map[string]map[string]string{
+			"01_researcher": {"notes.md": goodReport, "fields.json": "{}"},
+			"02_writer":     {"output.md": goodReport},
+			"03_reviewer":   {"output.md": goodReport},
+		}, NopSink{})
+		res, err := eng.Run(context.Background(), Goal{Text: "快照"}, RunOptions{CheckpointMode: mode})
+		if err != nil || res.Status != RunCompleted {
+			t.Fatalf("%s run: %v %+v", mode, err, res)
+		}
+		latest, err := FileStore{}.LoadLatestCheckpoint(res.WorkDir)
+		if err != nil || latest == nil {
+			t.Fatalf("%s latest: %v", mode, err)
+		}
+		if latest.Superstep < 3 {
+			t.Fatalf("%s superstep = %d", mode, latest.Superstep)
+		}
+	}
+}
+
+// TestGraphContinuityRejectsStructuralChange：崩溃期间图形态被改则拒绝恢复（WP1 裁定）。
+func TestGraphContinuityRejectsStructuralChange(t *testing.T) {
+	eng, _ := newTestEngine(t, mockScriptPlanner(baseSteps()...), map[string]map[string]string{
+		"01_researcher": {"notes.md": goodReport, "fields.json": "{}"},
+		"02_writer":     {"output.md": goodReport},
+		"03_reviewer":   {"output.md": goodReport},
+	}, NopSink{})
+	dir := eng.opts.DataDir
+	res, err := eng.Run(context.Background(), Goal{Text: "先跑完再中断", HitlMode: HitlAuto}, RunOptions{})
+	if err != nil || res.Status != RunCompleted {
+		t.Fatalf("setup run: %v %+v", err, res)
+	}
+	// 翻回 running（仅状态变化，图结构不变——应先过连续性校验）。
+	st, err := loadRunState(res.WorkDir)
+	if err != nil {
+		t.Fatal(err)
+	}
+	st.Meta.Status = string(RunExecuting)
+	st.Meta.CompletedAt = nil
+	wf := st.Workflow()
+	wf.Status = "running"
+	wf.Steps[1].Status = "running"
+	wf.Steps[1].Attempts = 1
+	// 再改结构：把 s3 的 needs 改成另一个合法 id——编译能过但指纹必变。
+	wf.Steps[2].Needs = []string{"s1"}
+	st.setWorkflow(wf)
+	if err := st.save(); err != nil {
+		t.Fatal(err)
+	}
+	// 用一个新的引擎在同一目录上恢复（图编译本身合法：needs 引用存在的 id）。
+	reg := NewRoleRegistry()
+	for _, role := range []string{"researcher", "writer", "reviewer"} {
+		if err := reg.Register(RoleSpec{Role: role, Backend: "script"}); err != nil {
+			t.Fatal(err)
+		}
+	}
+	backends := NewBackendResolver()
+	backends.RegisterInstance("script", &scriptBackend{script: map[string]map[string]string{}})
+	eng2, err := NewEngine(Options{DataDir: dir, Registry: reg, Planner: eng.opts.Planner, Backends: backends, Sink: NopSink{}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	_, err = eng2.Resume(context.Background(), res.RunID)
+	if err == nil || !strings.Contains(err.Error(), "图形态不一致") && !strings.Contains(err.Error(), "graph continuity") {
+		t.Fatalf("expected continuity rejection, got %v", err)
+	}
+}

squad/engine.go

@@ -35,6 +35,9 @@ type Options struct {
 	HitlPolicy HitlPolicy
 	// RepairBudget 质检失败自动修复次数（默认 1）。
 	RepairBudget int
+	// Store 运行状态持久化端口（Q1 WP4/WP5）；nil 回落 FileStore。
+	// 注入 SQLiteStore 后运行状态进 SQLite（装配期一次，跨 run 一致）。
+	Store Store
 	// IntentGate 可选意图闸门：Run 入口一次确定性裁决（nil = 不启用，
 	// 行为与原实现一致）。见 intent.go；实现方为 squadintent 适配器。
 	IntentGate IntentGate
@@ -60,6 +63,31 @@ type engineDeps struct {
 	StepTimeout  time.Duration
 	HitlPolicy   HitlPolicy
 	IntentRunner IntentRunner
+	// Store 运行状态持久化端口（Q1 WP4）；nil 回落 FileStore（装配期注入替换实现）。
+	Store Store
+}
+
+// storeOr 返回注入的 Store，缺省 FileStore。
+func (d engineDeps) storeOr() Store {
+	if d.Store != nil {
+		return d.Store
+	}
+	return FileStore{}
+}
+
+// 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 {
+		return nil, err
+	}
+	st.store = e.deps.storeOr()
+	return st, nil
 }
 
 // RunOptions 单次运行的可选项。
@@ -75,6 +103,17 @@ type RunOptions struct {
 	LeaderRole string `json:"leader_role,omitempty"`
 	// MaxLeaderTurns leader 模式每运行的 leader 轮次上限（默认 8，防失控）。
 	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"`
 }
 
 // SquadEngine 小队引擎门面：Run / Respond / Resume / Inspect / ListRuns。
@@ -131,7 +170,7 @@ func NewEngine(opts Options) (*SquadEngine, error) {
 		Registry: opts.Registry, Backends: opts.Backends, Evaluator: opts.Evaluator,
 		Checker: opts.Checker, Sink: opts.Sink, Workspaces: ws,
 		StepTimeout: opts.StepTimeout, HitlPolicy: opts.HitlPolicy,
-		IntentRunner: opts.IntentRunner,
+		IntentRunner: opts.IntentRunner, Store: opts.Store,
 	}
 	return &SquadEngine{opts: opts, deps: deps, ws: ws, runLocks: map[string]*sync.Mutex{}}, nil
 }
@@ -206,6 +245,19 @@ func (e *SquadEngine) Run(ctx context.Context, goal Goal, runOpts RunOptions) (*
 			meta.Metadata["max_leader_turns"] = fmt.Sprintf("%d", runOpts.MaxLeaderTurns)
 		}
 	}
+	if runOpts.CheckpointMode != "" {
+		meta.Metadata["checkpoint_mode"] = runOpts.CheckpointMode
+	}
+	if runOpts.Topology != "" {
+		meta.Metadata["topology"] = runOpts.Topology
+	}
+	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)
+	}
 	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）。
@@ -239,12 +291,12 @@ func (e *SquadEngine) Run(ctx context.Context, goal Goal, runOpts RunOptions) (*
 		}
 	}
 
-	st := &runState{Meta: meta, Approved: map[string]bool{}, RepairBudgetRemaining: e.opts.RepairBudget}
+	st := &RunState{Meta: meta, Approved: map[string]bool{}, RepairBudgetRemaining: e.opts.RepairBudget, store: e.deps.storeOr()}
 	return e.planAndExecute(ctx, st, runOpts)
 }
 
 // planAndExecute 规划 → 挂起或执行；st 可为已有运行状态（澄清续跑复用）。
-func (e *SquadEngine) planAndExecute(ctx context.Context, st *runState, runOpts RunOptions) (*RunResult, error) {
+func (e *SquadEngine) planAndExecute(ctx context.Context, st *RunState, runOpts RunOptions) (*RunResult, error) {
 	drv := &flowDriver{opts: e.deps}
 	meta := &st.Meta
 
@@ -266,6 +318,10 @@ 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,
@@ -299,7 +355,7 @@ func (e *SquadEngine) planAndExecute(ctx context.Context, st *runState, runOpts
 			Context:   fmt.Sprintf("原因：%s\n问题：\n%s", planOut.Clarification.Reason, strings.Join(qs, "\n")),
 			CreatedAt: time.Now(), Status: "pending",
 		}
-		gate := &hitlGate{Request: *req, PrevStatus: RunPlanning}
+		gate := &HitlGate{Request: *req, PrevStatus: RunPlanning}
 		// 澄清挂起也需要工作区承载 pending.json。
 		if err := e.materializeWorkspace(meta, st, &SquadPlan{Workflow: WorkflowPlan{WorkflowID: NewPlanID(), Status: "pending", Steps: nil}}); err != nil {
 			return nil, err
@@ -339,7 +395,7 @@ func (e *SquadEngine) planAndExecute(ctx context.Context, st *runState, runOpts
 			CreatedAt: time.Now(), Status: "pending",
 			Reason: "plan",
 		}
-		gate := &hitlGate{Request: *req, Plan: planOut.Plan, PrevStatus: RunPlanning}
+		gate := &HitlGate{Request: *req, Plan: planOut.Plan, PrevStatus: RunPlanning}
 		if err := st.saveGate(gate); err != nil {
 			return nil, err
 		}
@@ -352,6 +408,20 @@ 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)
@@ -373,7 +443,7 @@ func (e *SquadEngine) planAndExecute(ctx context.Context, st *runState, runOpts
 }
 
 // initLeaderState 按运行选项初始化 leader 循环状态。
-func (e *SquadEngine) initLeaderState(st *runState) {
+func (e *SquadEngine) initLeaderState(st *RunState) {
 	leaderRole := st.Meta.Metadata["leader_role"]
 	if leaderRole == "" {
 		if _, ok := findAgent(st.Plan.Squad.Agents, "leader"); ok {
@@ -386,7 +456,13 @@ 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()
 }
 
@@ -402,12 +478,12 @@ func parseIntDefault(s string, def int) int {
 }
 
 // isLeaderRun 该运行是否处于 leader-delegation 模式。
-func (e *SquadEngine) isLeaderRun(st *runState) bool {
+func (e *SquadEngine) isLeaderRun(st *RunState) bool {
 	return st.Meta.Metadata["mode"] == "leader"
 }
 
 // resumeAfterHold 挂起解除后的统一续跑入口（按模式分发）。
-func (e *SquadEngine) resumeAfterHold(ctx context.Context, st *runState) (*RunResult, error) {
+func (e *SquadEngine) resumeAfterHold(ctx context.Context, st *RunState) (*RunResult, error) {
 	if e.isLeaderRun(st) {
 		out, gate, err := e.runLeaderLoop(ctx, st)
 		if err != nil {
@@ -428,16 +504,16 @@ func (e *SquadEngine) resumeAfterHold(ctx context.Context, st *runState) (*RunRe
 }
 
 // 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 {
+func (e *SquadEngine) materializeWorkspace(meta *RunMetadata, st *RunState, plan *SquadPlan) error {
 	if st.Meta.WorkDir == "" {
 		st.Meta = *meta
-		wsDir, err := e.ws.PrepareWorkspace(&st.Meta, plan)
+		_, err := e.ws.PrepareWorkspace(&st.Meta, plan)
 		if err != nil {
 			return err
 		}
 		st.Template = plan.Template
-		bb, _ := loadBlackboard(wsDir)
-		st.BB = *bb
+		// 新运行的黑板在内存初始化；持久化由后续 st.save() 经 Store 端口完成。
+		st.BB = BlackboardState{Fields: map[string]string{}}
 	} else {
 		// 复用既有工作区：磁盘黑板为权威（澄清续跑/计划更新场景）。
 		if bb, err := loadBlackboard(st.Meta.WorkDir); err == nil {
@@ -489,7 +565,7 @@ func (e *SquadEngine) Respond(ctx context.Context, runID string, decision Respon
 
 func (e *SquadEngine) respondLocked(ctx context.Context, runID string, decision RespondDecision) (*RunResult, error) {
 	wsDir := filepath.Join(e.opts.DataDir, "runs", runID)
-	st, err := loadRunState(wsDir)
+	st, err := e.loadRun(wsDir)
 	if err != nil {
 		return nil, fmt.Errorf("load run %s: %w", runID, err)
 	}
@@ -509,6 +585,17 @@ 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)
 	}
@@ -543,7 +630,7 @@ func choiceHoldKind(kind HumanRequestKind) bool {
 	}
 }
 
-func (e *SquadEngine) noteAndKeepHold(st *runState, gate *hitlGate, decision RespondDecision) (*RunResult, error) {
+func (e *SquadEngine) noteAndKeepHold(st *RunState, gate *HitlGate, decision RespondDecision) (*RunResult, error) {
 	if err := st.addSteering(decision.Answer); err != nil {
 		return nil, err
 	}
@@ -555,7 +642,7 @@ func (e *SquadEngine) noteAndKeepHold(st *runState, gate *hitlGate, decision Res
 	return heldResult(st, gate), nil
 }
 
-func clearSameGate(st *runState, requestID string) error {
+func clearSameGate(st *RunState, requestID string) error {
 	current, err := loadGate(st.Meta.WorkDir)
 	if err != nil || current == nil || current.Request.ID != requestID {
 		return err
@@ -563,7 +650,7 @@ func clearSameGate(st *runState, requestID string) error {
 	return st.saveGate(nil)
 }
 
-func (e *SquadEngine) applyDecision(ctx context.Context, st *runState, gate *hitlGate, decision RespondDecision) (*RunResult, error) {
+func (e *SquadEngine) applyDecision(ctx context.Context, st *RunState, gate *HitlGate, decision RespondDecision) (*RunResult, error) {
 	req := &gate.Request
 	now := time.Now()
 	drv := &flowDriver{opts: e.deps}
@@ -694,8 +781,17 @@ func (e *SquadEngine) applyDecision(ctx context.Context, st *runState, gate *hit
 		case "approve":
 			return e.manualPass(st, decision)
 		case "retry":
-			// 追加一次修复预算并重跑交付步骤。
+			// 追加一次修复预算并重跑交付步骤。累计追加有上限（Q26），
+			// 上限内才加——无上限的人工 retry 会把运行变成永动机。
+			term := e.terminationOf(st)
+			if st.RepairBudgetGranted >= term.MaxRepairGranted {
+				drv.emit(st, Event{Type: EvPhase, Text: fmt.Sprintf("人工修复追加已达上限 %d，不再追加（Q26）", term.MaxRepairGranted)})
+				return &RunResult{RunID: st.Meta.ID, Status: RunFailed, Plan: planPtr(st),
+					Error: fmt.Sprintf("repair budget grants exhausted (%d/%d)", st.RepairBudgetGranted, term.MaxRepairGranted),
+					WorkDir: st.Meta.WorkDir}, nil
+			}
 			st.RepairBudgetRemaining++
+			st.RepairBudgetGranted++
 			st.Meta.Status = string(RunExecuting)
 			_ = st.save()
 			deliverStepID := lastCompletedStepID(st)
@@ -722,7 +818,7 @@ func (e *SquadEngine) applyDecision(ctx context.Context, st *runState, gate *hit
 }
 
 // respondClarification 澄清回应：答案并入目标后重新规划。
-func (e *SquadEngine) respondClarification(ctx context.Context, st *runState, gate *hitlGate, decision RespondDecision) (*RunResult, error) {
+func (e *SquadEngine) respondClarification(ctx context.Context, st *RunState, gate *HitlGate, decision RespondDecision) (*RunResult, error) {
 	choice := strings.ToLower(decision.Choice)
 	if choice == "abort" {
 		return e.cancelRun(st, "人工中止（澄清阶段）"), nil
@@ -757,7 +853,7 @@ func (e *SquadEngine) respondClarification(ctx context.Context, st *runState, ga
 }
 
 // bootstrapMinimalSquad 注册通用最小编制（organizer/specialist/writer/leader）。
-func (e *SquadEngine) bootstrapMinimalSquad(st *runState) {
+func (e *SquadEngine) bootstrapMinimalSquad(st *RunState) {
 	backend := st.Meta.Backend
 	if backend == "" {
 		backend = e.opts.DefaultBackend
@@ -777,7 +873,7 @@ func (e *SquadEngine) bootstrapMinimalSquad(st *runState) {
 }
 
 // manualPass 人工放行质检：标记 ManualPass 并收口。
-func (e *SquadEngine) manualPass(st *runState, decision RespondDecision) (*RunResult, error) {
+func (e *SquadEngine) manualPass(st *RunState, decision RespondDecision) (*RunResult, error) {
 	report := e.deps.Checker.Check(CheckRequest{
 		RunID: st.Meta.ID, Workspace: st.Meta.WorkDir, Deliverable: "output.md",
 		Plan: planPtr(st), Squad: &st.Plan.Squad, Template: st.Template,
@@ -806,7 +902,7 @@ func (e *SquadEngine) manualPass(st *runState, decision RespondDecision) (*RunRe
 		Summary: summarize(st, *report, status) + "（质检人工放行）"}, nil
 }
 
-func (e *SquadEngine) cancelRun(st *runState, reason string) *RunResult {
+func (e *SquadEngine) cancelRun(st *RunState, reason string) *RunResult {
 	now := time.Now()
 	st.Meta.Status = string(RunCancelled)
 	st.Meta.CompletedAt = &now
@@ -820,11 +916,11 @@ func (e *SquadEngine) cancelRun(st *runState, reason string) *RunResult {
 }
 
 // failStepResume 权限拒绝后的失败步骤处理（关键失败/非关键降级）。
-func (d *flowDriver) failStepResume(ctx context.Context, st *runState, idx int) (*RunResult, *hitlGate, error) {
+func (d *flowDriver) failStepResume(ctx context.Context, st *RunState, idx int) (*RunResult, *HitlGate, error) {
 	return d.failStep(ctx, st, idx)
 }
 
-func (e *SquadEngine) respondStepApproval(ctx context.Context, st *runState, req *HumanRequest, decision RespondDecision) (*RunResult, error) {
+func (e *SquadEngine) respondStepApproval(ctx context.Context, st *RunState, req *HumanRequest, decision RespondDecision) (*RunResult, error) {
 	switch strings.ToLower(decision.Choice) {
 	case "approve":
 		st.Approved["step:"+req.StepID] = true
@@ -842,7 +938,7 @@ func (e *SquadEngine) respondStepApproval(ctx context.Context, st *runState, req
 	return e.resumeAfterHold(ctx, st)
 }
 
-func (e *SquadEngine) respondInterrupted(ctx context.Context, st *runState, req *HumanRequest, decision RespondDecision) (*RunResult, error) {
+func (e *SquadEngine) respondInterrupted(ctx context.Context, st *RunState, req *HumanRequest, decision RespondDecision) (*RunResult, error) {
 	switch strings.ToLower(decision.Choice) {
 	case "retry":
 		if !markStep(st, req.StepID, "pending", "") {
@@ -862,7 +958,7 @@ func (e *SquadEngine) respondInterrupted(ctx context.Context, st *runState, req
 	return e.resumeAfterHold(ctx, st)
 }
 
-func markStep(st *runState, id, status, errText string) bool {
+func markStep(st *RunState, id, status, errText string) bool {
 	for i := range st.Plan.Workflow.Steps {
 		if st.Plan.Workflow.Steps[i].ID != id {
 			continue
@@ -876,7 +972,7 @@ func markStep(st *runState, id, status, errText string) bool {
 
 // recoverInterrupted 处理崩溃时停在 running 的步骤。
 // interactive 挂起确认，不自动再执行；auto 允许一次重试；disabled 直接失败。
-func (e *SquadEngine) recoverInterrupted(ctx context.Context, st *runState) (*RunResult, error) {
+func (e *SquadEngine) recoverInterrupted(ctx context.Context, st *RunState) (*RunResult, error) {
 	for i := range st.Plan.Workflow.Steps {
 		step := st.Plan.Workflow.Steps[i]
 		if step.Status != "running" {
@@ -903,6 +999,29 @@ func (e *SquadEngine) recoverInterrupted(ctx context.Context, st *runState) (*Ru
 				out, _, err := drv.failStep(ctx, st, i)
 				return out, err
 			}
+			// Q1 WP3 产物校验：步骤在 running 崩溃，但其交付物已全部落盘——
+			// 副作用实际发生过，直接判定完成而不是重新执行（副作用次数=1）。
+			subDir := filepath.Join(st.Meta.WorkDir, filepath.FromSlash(orDefault(step.SubWorkspace, fmt.Sprintf("tasks/%02d_%s", i+1, step.Role))))
+			targets := step.Targets
+			if len(targets) == 0 {
+				targets = []string{"output.md"}
+			}
+			if missingTargets(subDir, targets) == nil {
+				step.Status = "completed"
+				step.Error = ""
+				st.Plan.Workflow.Steps[i] = step
+				st.mu.Lock()
+				_ = mergeStepArtifacts(&st.BB, subDir)
+				err := st.save()
+				st.mu.Unlock()
+				if err != nil {
+					return nil, err
+				}
+				drv := &flowDriver{opts: e.deps}
+				drv.emit(st, Event{Type: EvStepCompleted, StepID: step.ID, Role: step.Role,
+					Text: fmt.Sprintf("中断恢复：产物齐全，按 attempt %s 判定完成", step.AttemptID)})
+				continue
+			}
 			st.Plan.Workflow.Steps[i].Status = "pending"
 			if err := st.save(); err != nil {
 				return nil, err
@@ -933,7 +1052,7 @@ func (e *SquadEngine) Resume(ctx context.Context, runID string) (*RunResult, err
 	defer release()
 
 	wsDir := filepath.Join(e.opts.DataDir, "runs", runID)
-	st, err := loadRunState(wsDir)
+	st, err := e.loadRun(wsDir)
 	if err != nil {
 		return nil, fmt.Errorf("load run %s: %w", runID, err)
 	}
@@ -944,6 +1063,11 @@ func (e *SquadEngine) Resume(ctx context.Context, runID string) (*RunResult, err
 	if RunStatus(st.Meta.Status).Terminal() {
 		return e.Inspect(runID)
 	}
+	// 图连续性校验（Q1 WP4/WP1 裁定）：崩溃期间的检查点图指纹与当前 Steps
+	// 重编译不一致 = 图形态被改过（动态调整或外力），拒绝恢复并提示重新规划。
+	if err := validateGraphContinuity(st); err != nil {
+		return nil, err
+	}
 	if out, err := e.recoverInterrupted(ctx, st); out != nil || err != nil {
 		return out, err
 	}
@@ -956,7 +1080,7 @@ func (e *SquadEngine) Inspect(runID string) (*RunResult, error) {
 		return nil, invalidID(runID)
 	}
 	wsDir := filepath.Join(e.opts.DataDir, "runs", runID)
-	st, err := loadRunState(wsDir)
+	st, err := e.loadRun(wsDir)
 	if err != nil {
 		return nil, fmt.Errorf("load run %s: %w", runID, err)
 	}
@@ -1003,7 +1127,7 @@ func (e *SquadEngine) ListRuns() ([]RunMetadata, error) {
 		if !ent.IsDir() {
 			continue
 		}
-		st, err := loadRunState(filepath.Join(runsDir, ent.Name()))
+		st, err := e.loadRun(filepath.Join(runsDir, ent.Name()))
 		if err != nil {
 			continue
 		}
@@ -1037,7 +1161,7 @@ func (e *SquadEngine) Steer(runID, text string) error {
 	}
 	defer release()
 
-	st, err := loadRunState(filepath.Join(e.opts.DataDir, "runs", runID))
+	st, err := e.loadRun(filepath.Join(e.opts.DataDir, "runs", runID))
 	if err != nil {
 		return fmt.Errorf("load run %s: %w", runID, err)
 	}
@@ -1061,7 +1185,7 @@ func (e *SquadEngine) SetBlackboardField(runID, key, value string) error {
 	}
 	defer release()
 
-	st, err := loadRunState(filepath.Join(e.opts.DataDir, "runs", runID))
+	st, err := e.loadRun(filepath.Join(e.opts.DataDir, "runs", runID))
 	if err != nil {
 		return fmt.Errorf("load run %s: %w", runID, err)
 	}
@@ -1143,7 +1267,7 @@ func (e *SquadEngine) fireWakeup(ctx context.Context, runID string) bool {
 	if err := writeJSON(filepath.Join(e.opts.DataDir, "runs", runID, "pending.json"), gate); err != nil {
 		return false
 	}
-	if gate.Request.WakeupFires >= MaxWakeupFires {
+	if gate.Request.WakeupFires >= e.terminationOf(nil).MaxWakeupFires {
 		_, err := e.respondLocked(ctx, runID, RespondDecision{RequestID: gate.Request.ID, Choice: "abort",
 			Answer: "自动唤醒次数达上限，自动中止"})
 		return err == nil
@@ -1248,7 +1372,7 @@ func appendAnswer(d RespondDecision) string {
 	return "（" + d.Answer + "）"
 }
 
-func lastCompletedStepID(st *runState) string {
+func lastCompletedStepID(st *RunState) string {
 	id := ""
 	for _, s := range st.Plan.Workflow.Steps {
 		if s.Status == "completed" {

squad/graph.go

@@ -0,0 +1,253 @@
+package squad
+
+import (
+	"fmt"
+	"hash/fnv"
+	"strings"
+)
+
+// graph 是一次执行图：由权威 Steps 现场编译（Steps 仍是唯一真相源，
+// 动态调整改 Steps 后下一轮自然生效）。只携带调度所需的派生信息：
+// 依赖、写集与可执行分层。
+type graph struct {
+	needs  map[string][]string // step id → 依赖 id（含相邻 Parallel 兼容映射产生的）
+	order  []string            // 声明序
+	writes map[string][]string // step id → 写集
+}
+
+// compileGraph 编译步骤图为调度图。规则：
+//   - 声明 Needs 的步骤：依赖即其 Needs（未知引用/环 fail-closed）。
+//   - 未声明 Needs 的步骤：保持现状语义——相邻 Parallel=true 的步骤同层
+//     （依赖 = 上一个「非本组」步骤），非 Parallel 步骤依赖前一个步骤。
+//   - 混合图允许：Needs 提供剪枝，未声明处保持线性。
+func compileGraph(steps []StepSpec) (*graph, error) {
+	g := &graph{
+		needs:  map[string][]string{},
+		order:  make([]string, 0, len(steps)),
+		writes: map[string][]string{},
+	}
+	ids := map[string]bool{}
+	for _, s := range steps {
+		if s.ID == "" {
+			return nil, fmt.Errorf("graph: 存在空 id 的步骤")
+		}
+		if ids[s.ID] {
+			return nil, fmt.Errorf("graph: 步骤 id 重复：%s", s.ID)
+		}
+		ids[s.ID] = true
+	}
+	prevNonGroup := -1 // 上一个「非本 Parallel 组」的步骤索引
+	for i, s := range steps {
+		g.order = append(g.order, s.ID)
+		g.writes[s.ID] = append([]string(nil), s.Writes...)
+		needsDecl := len(s.Needs) > 0
+		switch {
+		case needsDecl:
+			for _, n := range s.Needs {
+				if !ids[n] {
+					return nil, fmt.Errorf("graph: 步骤 %s 的 needs 引用未知步骤 %s", s.ID, n)
+				}
+				if n == s.ID {
+					return nil, fmt.Errorf("graph: 步骤 %s 的 needs 引用自身", s.ID)
+				}
+			}
+			g.needs[s.ID] = append([]string(nil), s.Needs...)
+			prevNonGroup = i
+		case s.Parallel:
+			// 相邻 Parallel 组：依赖上一个组外步骤（没有则无依赖）。
+			if prevNonGroup >= 0 {
+				g.needs[s.ID] = []string{steps[prevNonGroup].ID}
+			}
+			// Parallel 步骤不推进 prevNonGroup：相邻同伴同层。
+		default:
+			// 非 Parallel：依赖紧邻的前一步（与旧线性扫描的执行序一致——
+			// 整组 Parallel 完成后才轮到它，而不是只等组外前一步）。
+			if i > 0 {
+				g.needs[s.ID] = []string{steps[i-1].ID}
+			}
+			prevNonGroup = i
+		}
+	}
+	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
+}
+
+// findCycle 报告一条依赖环（没有则 nil）。
+func (g *graph) findCycle() []string {
+	const (
+		white = iota
+		gray
+		black
+	)
+	color := map[string]int{}
+	var stack []string
+	var visit func(id string) []string
+	visit = func(id string) []string {
+		if color[id] == gray {
+			return append(stack, id)
+		}
+		if color[id] == black {
+			return nil
+		}
+		color[id] = gray
+		stack = append(stack, id)
+		for _, n := range g.needs[id] {
+			if c := visit(n); c != nil {
+				return c
+			}
+		}
+		stack = stack[:len(stack)-1]
+		color[id] = black
+		return nil
+	}
+	for _, id := range g.order {
+		if c := visit(id); c != nil {
+			return c
+		}
+	}
+	return nil
+}
+
+// layer 是一个 superstep 的可执行集合：seq 为有写集冲突需按声明序串行的
+// 步骤索引，par 为可并行的步骤索引。先执行 par 组（runParallel 语义），
+// 再逐个执行 seq（含动态调整），与现状「并行组先于后续顺序步骤」一致。
+type layer struct {
+	par []int
+	seq []int
+}
+
+// stepPending 报告步骤是否仍待执行。蓝图登记的步骤不带 status（空串），
+// 旧线性扫描将其视为可执行——这里与之一致：空串等同 pending。
+func stepPending(status string) bool {
+	return status == "" || status == "pending" || status == "running"
+}
+
+// nextLayer 计算当前可执行层：状态为 pending/running 且全部依赖已「满足」。
+// 依赖满足 = completed/skipped/failed——failed 视为满足与现状「非关键失败
+// 降级继续」的线性扫描一致（关键失败时运行已终止，不会走到这里）。
+// 无可执行步骤时返回 (layer{}, false)：可能全部完成，也可能调度卡死
+// （由调用方用 pendingCount 区分）。
+func (g *graph) nextLayer(steps []StepSpec) (layer, bool) {
+	idxOf := map[string]int{}
+	for i, s := range steps {
+		idxOf[s.ID] = i
+	}
+	satisfied := func(id string) bool {
+		st := steps[idxOf[id]].Status
+		return st == "completed" || st == "skipped" || st == "failed"
+	}
+	var out layer
+	for i, s := range steps {
+		if !stepPending(s.Status) {
+			continue
+		}
+		ready := true
+		for _, n := range g.needs[s.ID] {
+			if !satisfied(n) {
+				ready = false
+				break
+			}
+		}
+		if !ready {
+			continue
+		}
+		if g.conflictsWithEarlier(out, i, steps) {
+			out.seq = append(out.seq, i)
+		} else {
+			out.par = append(out.par, i)
+		}
+	}
+	return out, len(out.par)+len(out.seq) > 0
+}
+
+// conflictsWithEarlier 报告步骤 i 的写集与本层中更早的可执行步骤相交。
+func (g *graph) conflictsWithEarlier(l layer, i int, steps []StepSpec) bool {
+	wi := g.writes[steps[i].ID]
+	if len(wi) == 0 {
+		return false
+	}
+	earlier := append(append([]int(nil), l.par...), l.seq...)
+	for _, j := range earlier {
+		for _, w := range g.writes[steps[j].ID] {
+			for _, v := range wi {
+				if w == v {
+					return true
+				}
+			}
+		}
+	}
+	return false
+}
+
+// pendingCount 统计仍未终结（pending/running/空串）的步骤数。
+func pendingCount(steps []StepSpec) int {
+	n := 0
+	for _, s := range steps {
+		if stepPending(s.Status) {
+			n++
+		}
+	}
+	return n
+}
+
+// hash 是图结构指纹：声明序、依赖与写集的稳定散列。
+// 恢复连续性校验用它比对崩溃前后的图形态（Q1 WP4；WP1 的图版本裁定落点）。
+func (g *graph) hash() string {
+	h := fnv.New64a()
+	for _, id := range g.order {
+		h.Write([]byte(id))
+		h.Write([]byte{0})
+		for _, n := range g.needs[id] {
+			h.Write([]byte(">"))
+			h.Write([]byte(n))
+		}
+		h.Write([]byte{1})
+		for _, w := range g.writes[id] {
+			h.Write([]byte("w"))
+			h.Write([]byte(w))
+		}
+		h.Write([]byte{2})
+	}
+	return fmt.Sprintf("%016x", h.Sum64())
+}

squad/graph_test.go

@@ -0,0 +1,139 @@
+package squad
+
+import (
+	"strings"
+	"testing"
+)
+
+func stepsOf(specs ...StepSpec) []StepSpec {
+	out := make([]StepSpec, len(specs))
+	copy(out, specs)
+	return out
+}
+
+func TestCompileGraphLinearAndAdjacentParallel(t *testing.T) {
+	steps := stepsOf(
+		StepSpec{ID: "a", Status: "pending"},
+		StepSpec{ID: "b", Status: "pending", Parallel: true},
+		StepSpec{ID: "c", Status: "pending", Parallel: true},
+		StepSpec{ID: "d", Status: "pending"},
+	)
+	g, err := compileGraph(steps)
+	if err != nil {
+		t.Fatal(err)
+	}
+	// 线性链：b/c 同层依赖 a，d 依赖 c。
+	if len(g.needs["a"]) != 0 {
+		t.Fatalf("a needs = %v", g.needs["a"])
+	}
+	if len(g.needs["b"]) != 1 || g.needs["b"][0] != "a" || g.needs["c"][0] != "a" {
+		t.Fatalf("parallel needs = %v %v", g.needs["b"], g.needs["c"])
+	}
+	if len(g.needs["d"]) != 1 || g.needs["d"][0] != "c" {
+		t.Fatalf("d needs = %v", g.needs["d"])
+	}
+	// 首轮可执行：只有 a。
+	layer, ok := g.nextLayer(steps)
+	if !ok || len(layer.par) != 1 || layer.par[0] != 0 || len(layer.seq) != 0 {
+		t.Fatalf("first layer = %+v", layer)
+	}
+	// a 完成后：b、c 同层并行。
+	steps[0].Status = "completed"
+	layer, ok = g.nextLayer(steps)
+	if !ok || len(layer.par) != 2 || layer.par[0] != 1 || layer.par[1] != 2 {
+		t.Fatalf("second layer = %+v", layer)
+	}
+}
+
+func TestCompileGraphNeedsPruneAndMixed(t *testing.T) {
+	steps := stepsOf(
+		StepSpec{ID: "a", Status: "pending"},
+		StepSpec{ID: "b", Status: "pending"},
+		StepSpec{ID: "c", Status: "pending", Needs: []string{"a"}},
+	)
+	g, err := compileGraph(steps)
+	if err != nil {
+		t.Fatal(err)
+	}
+	// a 完成后：b（线性依赖 a）与 c（needs a）同层。
+	steps[0].Status = "completed"
+	layer, ok := g.nextLayer(steps)
+	if !ok || len(layer.par) != 2 {
+		t.Fatalf("layer = %+v", layer)
+	}
+	// b 失败（非关键降级）后 c 仍按 needs 调度（与线性扫描的降级继续一致）。
+	steps[1].Status = "failed"
+	layer, ok = g.nextLayer(steps)
+	if !ok || len(layer.par)+len(layer.seq) == 0 {
+		t.Fatalf("after b failed: layer = %+v", layer)
+	}
+}
+
+func TestCompileGraphRejectsUnknownAndCycleAndDup(t *testing.T) {
+	if _, err := compileGraph(stepsOf(
+		StepSpec{ID: "a", Status: "pending", Needs: []string{"ghost"}},
+	)); err == nil || !strings.Contains(err.Error(), "未知步骤") {
+		t.Fatalf("unknown ref: %v", err)
+	}
+	if _, err := compileGraph(stepsOf(
+		StepSpec{ID: "a", Status: "pending", Needs: []string{"b"}},
+		StepSpec{ID: "b", Status: "pending", Needs: []string{"a"}},
+	)); err == nil || !strings.Contains(err.Error(), "成环") {
+		t.Fatalf("cycle: %v", err)
+	}
+	if _, err := compileGraph(stepsOf(
+		StepSpec{ID: "a", Status: "pending"},
+		StepSpec{ID: "a", Status: "pending"},
+	)); err == nil || !strings.Contains(err.Error(), "重复") {
+		t.Fatalf("dup: %v", err)
+	}
+	if _, err := compileGraph(stepsOf(
+		StepSpec{ID: "a", Status: "pending", Needs: []string{"a"}},
+	)); err == nil || !strings.Contains(err.Error(), "自身") {
+		t.Fatalf("self ref: %v", err)
+	}
+}
+
+func TestWriteSetConflictSerializes(t *testing.T) {
+	steps := stepsOf(
+		StepSpec{ID: "z", Status: "completed"},
+		StepSpec{ID: "a", Status: "pending", Parallel: true, Writes: []string{"report.md"}},
+		StepSpec{ID: "b", Status: "pending", Parallel: true, Writes: []string{"report.md"}},
+		StepSpec{ID: "c", Status: "pending", Parallel: true, Writes: []string{"notes.md"}},
+		StepSpec{ID: "d", Status: "pending", Parallel: true},
+	)
+	// 全部首轮可执行（无依赖）：b 与 a 冲突 → b 串行；c、d 无冲突 → 并行。
+	g, err := compileGraph(steps)
+	if err != nil {
+		t.Fatal(err)
+	}
+	layer, ok := g.nextLayer(steps)
+	if !ok {
+		t.Fatal("no layer")
+	}
+	if len(layer.seq) != 1 || layer.seq[0] != 2 {
+		t.Fatalf("seq = %v", layer.seq)
+	}
+	if len(layer.par) != 3 || layer.par[0] != 1 || layer.par[1] != 3 || layer.par[2] != 4 {
+		t.Fatalf("par = %v", layer.par)
+	}
+}
+
+func TestFailedDependencyStillSchedules(t *testing.T) {
+	// 与线性扫描的「非关键失败降级继续」一致：依赖 failed 仍视为满足。
+	steps := stepsOf(
+		StepSpec{ID: "b", Status: "failed"},
+		StepSpec{ID: "a", Status: "pending", Needs: []string{"b"}},
+	)
+	g, err := compileGraph(steps)
+	if err != nil {
+		t.Fatal(err)
+	}
+	layer, ok := g.nextLayer(steps)
+	if !ok || len(layer.par) != 1 || layer.par[0] != 1 {
+		t.Fatalf("failed dependency should still schedule: %+v", layer)
+	}
+	if pendingCount(steps) != 1 {
+		t.Fatalf("pending = %d", pendingCount(steps))
+	}
+}

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)
@@ -136,8 +147,8 @@ func expiredDecision(req HumanRequest, policy HitlPolicy, now time.Time) (Respon
 	return RespondDecision{RequestID: req.ID, Choice: choice, Answer: "expired"}, true
 }
 
-// hitlGate 引擎内部：一次挂起的完整上下文（落盘 pending.json）。
-type hitlGate struct {
+// HitlGate 引擎内部：一次挂起的完整上下文（落盘 pending.json）。
+type HitlGate struct {
 	Request    HumanRequest `json:"request"`
 	Plan       *SquadPlan   `json:"plan,omitempty"`     // confirmation 时携带候选计划
 	Steps      []StepSpec   `json:"steps,omitempty"`    // adjustment：建议追加的步骤

squad/leader.go

@@ -13,7 +13,7 @@ import (
 // leaderExecRequest 构造 leader 轮次的执行请求：与标准步骤同等装配
 // （过程事件桥、ToolProfile/Skills/工具白名单、执行标识与预算）——
 // 修平此前 leader 轮丢弃 OnEvent 与 ToolProfile 的缺口。
-func leaderExecRequest(drv *flowDriver, st *runState, leader RoleSpec, workDir, prompt, instructions string, targets []string, timeout time.Duration) ExecRequest {
+func leaderExecRequest(drv *flowDriver, st *RunState, leader RoleSpec, workDir, prompt, instructions string, targets []string, timeout time.Duration) ExecRequest {
 	stepID := fmt.Sprintf("leader_turn_%d", st.Leader.Turns)
 	var budget *ExecBudget
 	if timeout > 0 {
@@ -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）。
@@ -80,22 +102,24 @@ type delegation struct {
 
 const leaderSubWorkspace = "tasks/00_leader"
 
-func (st *runState) loadLeader() {
+func (st *RunState) loadLeader() {
 	if st.Leader != nil {
 		return
 	}
 	st.Leader = &LeaderState{MaxTurns: DefaultMaxLeaderTurns}
-	data, err := os.ReadFile(filepath.Join(st.Meta.WorkDir, "leader.json"))
-	if err == nil {
+	data, err := st.storeOr().LoadLeader(st.Meta.WorkDir)
+	if err == nil && data != nil {
 		_ = json.Unmarshal(data, st.Leader)
 	}
 }
 
-func (st *runState) saveLeader() {
+func (st *RunState) saveLeader() {
 	if st.Leader == nil {
 		return
 	}
-	_ = writeJSON(filepath.Join(st.Meta.WorkDir, "leader.json"), st.Leader)
+	if data, err := json.Marshal(st.Leader); err == nil {
+		_ = st.storeOr().SaveLeader(st.Meta.WorkDir, data)
+	}
 }
 
 // leaderPrompt 组装 leader 的系统指令（协调契约 + 通用守则）。
@@ -132,7 +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)
@@ -169,13 +193,28 @@ 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()
 }
 
 // runLeaderLoop leader-delegation 主循环：委派→执行→唤醒，直到
 // close/failed/上限/挂起（escalation/gap/权限等由 HITL 处理后重入本函数）。
-func (e *SquadEngine) runLeaderLoop(ctx context.Context, st *runState) (*RunResult, *hitlGate, error) {
+func (e *SquadEngine) runLeaderLoop(ctx context.Context, st *RunState) (*RunResult, *HitlGate, error) {
 	st.loadLeader()
 	drv := &flowDriver{opts: e.deps}
 	leader, ok := findAgent(st.Plan.Squad.Agents, st.Leader.LeaderRole)
@@ -204,7 +243,8 @@ func (e *SquadEngine) runLeaderLoop(ctx context.Context, st *runState) (*RunResu
 	}
 	_ = snapshotContextInto(st.Meta.WorkDir, subDir)
 
-	for st.Leader.Turns < st.Leader.MaxTurns {
+	term := e.terminationOf(st)
+	for st.Leader.Turns < term.MaxLeaderTurns {
 		// HITL：leader 角色权限升级（与标准模式同语义）。
 		if leader.BashPermission == "ask" && !st.Approved["role:"+leader.Role] {
 			if drv.policy(st).mode() != HitlInteractive {
@@ -220,7 +260,7 @@ func (e *SquadEngine) runLeaderLoop(ctx context.Context, st *runState) (*RunResu
 				},
 				CreatedAt: time.Now(), Status: "pending",
 			}
-			gate := &hitlGate{Request: *req, PrevStatus: RunExecuting}
+			gate := &HitlGate{Request: *req, PrevStatus: RunExecuting}
 			if err := st.saveGate(gate); err == nil {
 				st.Meta.Status = string(RunWaitingHuman)
 				_ = st.save()
@@ -230,9 +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
 		}
@@ -246,7 +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))
@@ -292,7 +343,7 @@ func (e *SquadEngine) runLeaderLoop(ctx context.Context, st *runState) (*RunResu
 				},
 				CreatedAt: time.Now(), Status: "pending",
 			}
-			gate := &hitlGate{Request: *req, PrevStatus: RunExecuting}
+			gate := &HitlGate{Request: *req, PrevStatus: RunExecuting}
 			if err := st.saveGate(gate); err == nil {
 				st.Meta.Status = string(RunWaitingHuman)
 				_ = st.save()
@@ -335,6 +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
@@ -352,6 +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）。
@@ -368,11 +464,21 @@ func (e *SquadEngine) runLeaderLoop(ctx context.Context, st *runState) (*RunResu
 			return e.leaderFail(st, fmt.Sprintf("leader 输出了未知 action %q", dec.Action)), nil, nil
 		}
 	}
-	return e.leaderFail(st, fmt.Sprintf("leader 轮次达上限 %d（防失控护栏），运行停止；可提高 MaxLeaderTurns 或人工介入", st.Leader.MaxTurns)), nil, nil
+	// Q23：达上限不再裸失败——按现有产出走 finalize 降级收口（质检/汇总路径与正常收口一致）。
+	drv2 := &flowDriver{opts: e.deps}
+	drv2.emit(st, Event{Type: EvPhase, Text: fmt.Sprintf("leader 轮次达上限 %d，按现有产出降级收口（Q23）", term.MaxLeaderTurns)})
+	out, gate2, err := drv2.finalize(ctx, st)
+	if err != nil {
+		return nil, nil, err
+	}
+	if gate2 != nil {
+		return heldResult(st, gate2), gate2, nil
+	}
+	return out, nil, nil
 }
 
 // leaderFail 统一的 leader 模式失败收口。
-func (e *SquadEngine) leaderFail(st *runState, reason string) *RunResult {
+func (e *SquadEngine) leaderFail(st *RunState, reason string) *RunResult {
 	now := time.Now()
 	st.Meta.Status = string(RunFailed)
 	st.Meta.CompletedAt = &now
@@ -386,7 +492,7 @@ func (e *SquadEngine) leaderFail(st *runState, reason string) *RunResult {
 }
 
 // snapshotDeliverablesToLeader 把成员交付文件复制进 leader 工作区 inbox/<步骤id>/。
-func snapshotDeliverablesToLeader(st *runState, step StepSpec) error {
+func snapshotDeliverablesToLeader(st *RunState, step StepSpec) error {
 	srcDir := filepath.Join(st.Meta.WorkDir, filepath.FromSlash(step.SubWorkspace))
 	dstDir := filepath.Join(st.Meta.WorkDir, filepath.FromSlash(leaderSubWorkspace), "inbox", step.ID)
 	for _, t := range step.Targets {

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/leader_test.go

@@ -200,8 +200,13 @@ func TestLeaderModeTurnCapGuard(t *testing.T) {
 	if err != nil {
 		t.Fatal(err)
 	}
-	if res.Status != RunFailed || !strings.Contains(res.Error, "上限") {
-		t.Fatalf("want turn-cap failure, got %s err=%s", res.Status, res.Error)
+	// Q23：达上限走 finalize 降级收口（completed/degraded/质检挂起），
+	// 不再是「运行停止」的裸失败。
+	if res.Status == RunFailed && strings.Contains(res.Error, "运行停止") {
+		t.Fatalf("Q23 closeout should not be a bare failure, got %s err=%s", res.Status, res.Error)
+	}
+	if res.Status != RunCompleted && res.Status != RunDegraded && res.Status != RunQualityReview {
+		t.Fatalf("Q23 closeout status = %s err=%s", res.Status, res.Error)
 	}
 	if b.turn < 3 {
 		t.Fatalf("leader should have run %d turns, ran %d", 3, b.turn)

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。
@@ -88,10 +91,16 @@ type StepSpec struct {
 	Mode         string   `json:"mode,omitempty"`          // 预留
 	Parallel     bool     `json:"parallel"`
 	Needs        []string `json:"needs,omitempty"`
+	// Writes 写集：本步骤会写入的工件相对路径（调度依据）。
+	// 同一 superstep 内写集相交的步骤强制串行，不相交才可并行（Q1 WP1）。
+	Writes []string `json:"writes,omitempty"`
 	Targets      []string `json:"targets,omitempty"` // 期望交付物相对路径，缺省 output.md
 	Status       string   `json:"status"`            // pending, running, completed, failed, skipped
 	// Attempts 该步骤进入 running 的次数。中断恢复用它决定 auto 策略是否还能重试。
-	Attempts   int    `json:"attempts,omitempty"`
+	Attempts int `json:"attempts,omitempty"`
+	// AttemptID 当前这次尝试的稳定标识：runID/stepID/N（N=Attempts）。
+	// 执行事件、产物与副作用记录都挂它；瞬态重试共享同一 AttemptID（Q1 WP3）。
+	AttemptID string `json:"attempt_id,omitempty"`
 	DurationMs int64  `json:"duration_ms,omitempty"`
 	Error      string `json:"error,omitempty"`
 	// RequiresApproval 为 true 时，该步骤启动前进入 HITL 计划确认。
@@ -109,6 +118,43 @@ 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"`
+}
+
+// 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/runstate.go

@@ -10,9 +10,9 @@ import (
 	"time"
 )
 
-// runState 引擎内部运行状态（磁盘即权威：workspace.json / workflow.json /
+// RunState 引擎内部运行状态（磁盘即权威：workspace.json / workflow.json /
 // squad.json / blackboard.json / pending.json 全部落盘，进程重启可恢复）。
-type runState struct {
+type RunState struct {
 	Meta     RunMetadata     `json:"meta"`
 	Plan     SquadPlan       `json:"plan"`
 	BB       BlackboardState `json:"bb"`
@@ -21,19 +21,35 @@ type runState struct {
 	Approved map[string]bool `json:"approved,omitempty"`
 	// RepairBudgetRemaining 质检失败后的自动重跑剩余次数。
 	RepairBudgetRemaining int `json:"repair_budget_remaining,omitempty"`
+	// RepairBudgetGranted 人工 retry 累计追加次数（Q26 上限判定用）。
+	RepairBudgetGranted int `json:"repair_budget_granted,omitempty"`
 	// Clarifications 已澄清问答（追加进 Goal 上下文）。
 	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"`
 	// 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）。
 	Leader *LeaderState `json:"-"`
+	// store 运行状态持久化端口（Q1 WP4）；nil 回落 FileStore（现状文件态）。
+	// 替换实现（如 SQLiteStore）在引擎装配期注入。
+	store Store `json:"-"`
 	// mu 保护并行步骤对共享状态（黑板/save/steering/usage）的变更。
 	mu sync.Mutex `json:"-"`
 }
 
-func (st *runState) policy() HitlPolicy {
+// storeOr 返回注入的 Store，缺省 FileStore。
+func (st *RunState) storeOr() Store {
+	if st.store != nil {
+		return st.store
+	}
+	return FileStore{}
+}
+
+func (st *RunState) policy() HitlPolicy {
 	return mergeHitlPolicy(HitlPolicy{}, st.Meta.HitlPolicyOf())
 }
 
@@ -49,8 +65,14 @@ func (m RunMetadata) HitlPolicyOf() *HitlPolicy {
 	return p
 }
 
-func (st *runState) save() error {
+func (st *RunState) save() error {
 	st.Meta.UpdatedAt = time.Now()
+	return st.storeOr().SaveRun(st)
+}
+
+// persist 是 FileStore 语义的低层落盘（全部写文件）。Store.SaveRun 调它；
+// RunState.save 走 Store 端口，替换实现（SQLiteStore）在装配期注入后自然接管。
+func (st *RunState) persist() error {
 	if err := writeJSON(filepath.Join(st.Meta.WorkDir, "workspace.json"), &st.Meta); err != nil {
 		return err
 	}
@@ -71,24 +93,28 @@ func (st *runState) save() error {
 	return writeJSON(filepath.Join(st.Meta.WorkDir, "runstate.json"), diskRunState{
 		Approved:              st.Approved,
 		RepairBudgetRemaining: st.RepairBudgetRemaining,
+		RepairBudgetGranted:   st.RepairBudgetGranted,
 		Clarifications:        st.Clarifications,
+		Mailbox:               st.Mailbox,
 	})
 }
 
 type diskRunState struct {
 	Approved              map[string]bool `json:"approved,omitempty"`
 	RepairBudgetRemaining int             `json:"repair_budget_remaining,omitempty"`
+	RepairBudgetGranted   int             `json:"repair_budget_granted,omitempty"`
 	Clarifications        []string        `json:"clarifications,omitempty"`
+	Mailbox               []MailMessage   `json:"mailbox,omitempty"`
 }
 
 // Workflow 返回计划的编排清单（简洁访问器）。
-func (st *runState) Workflow() WorkflowPlan { return st.Plan.Workflow }
+func (st *RunState) Workflow() WorkflowPlan { return st.Plan.Workflow }
 
 // setWorkflow 写回编排清单。
-func (st *runState) setWorkflow(wf WorkflowPlan) { st.Plan.Workflow = wf }
+func (st *RunState) setWorkflow(wf WorkflowPlan) { st.Plan.Workflow = wf }
 
 // saveGate 落盘 HITL 挂起上下文（pending.json）。
-func (st *runState) saveGate(gate *hitlGate) error {
+func (st *RunState) saveGate(gate *HitlGate) error {
 	if gate == nil {
 		err := os.Remove(filepath.Join(st.Meta.WorkDir, "pending.json"))
 		if os.IsNotExist(err) {
@@ -100,8 +126,8 @@ func (st *runState) saveGate(gate *hitlGate) error {
 }
 
 // loadRunState 从工作区恢复运行状态。
-func loadRunState(wsDir string) (*runState, error) {
-	st := &runState{}
+func loadRunState(wsDir string) (*RunState, error) {
+	st := &RunState{}
 	if err := readJSON(filepath.Join(wsDir, "workspace.json"), &st.Meta); err != nil {
 		return nil, fmt.Errorf("load workspace.json: %w", err)
 	}
@@ -131,7 +157,9 @@ func loadRunState(wsDir string) (*runState, error) {
 			st.Approved = disk.Approved
 		}
 		st.RepairBudgetRemaining = disk.RepairBudgetRemaining
+		st.RepairBudgetGranted = disk.RepairBudgetGranted
 		st.Clarifications = disk.Clarifications
+		st.Mailbox = disk.Mailbox
 	} else if !os.IsNotExist(err) {
 		return nil, err
 	}
@@ -143,7 +171,7 @@ func loadRunState(wsDir string) (*runState, error) {
 }
 
 // loadGate 恢复挂起上下文；无挂起返回 nil。
-func loadGate(wsDir string) (*hitlGate, error) {
+func loadGate(wsDir string) (*HitlGate, error) {
 	data, err := os.ReadFile(filepath.Join(wsDir, "pending.json"))
 	if err != nil {
 		if os.IsNotExist(err) {
@@ -151,7 +179,7 @@ func loadGate(wsDir string) (*hitlGate, error) {
 		}
 		return nil, err
 	}
-	gate := &hitlGate{}
+	gate := &HitlGate{}
 	if err := json.Unmarshal(data, gate); err != nil {
 		return nil, fmt.Errorf("parse pending.json: %w", err)
 	}
@@ -167,7 +195,7 @@ func readJSON(path string, v any) error {
 }
 
 // setStep 并行安全地回填步骤状态（与 save 的全量 marshal 互斥）。
-func (st *runState) setStep(idx int, s StepSpec) {
+func (st *RunState) setStep(idx int, s StepSpec) {
 	st.mu.Lock()
 	st.Plan.Workflow.Steps[idx] = s
 	st.mu.Unlock()
@@ -175,7 +203,7 @@ func (st *runState) setStep(idx int, s StepSpec) {
 
 // injectSteering 把挂起的运行中人工补充指令追加到 prompt（下一步骤边界注入），
 // 注入后清空并落盘。注入发生在步骤开始前，不会污染正在执行的步骤。
-func (st *runState) injectSteering(prompt string) string {
+func (st *RunState) injectSteering(prompt string) string {
 	st.mu.Lock()
 	defer st.mu.Unlock()
 	if len(st.Steering) == 0 {
@@ -194,7 +222,7 @@ func (st *runState) injectSteering(prompt string) string {
 }
 
 // addSteering 追加一条尚未注入的人工补充，并立刻落盘。
-func (st *runState) addSteering(text string) error {
+func (st *RunState) addSteering(text string) error {
 	text = strings.TrimSpace(text)
 	if text == "" {
 		return nil
@@ -206,7 +234,7 @@ func (st *runState) addSteering(text string) error {
 }
 
 // loadSteering 恢复未消费的补充指令。
-func (st *runState) loadSteering() {
+func (st *RunState) loadSteering() {
 	data, err := os.ReadFile(filepath.Join(st.Meta.WorkDir, "steering.json"))
 	if err != nil {
 		return
@@ -218,7 +246,7 @@ func (st *runState) loadSteering() {
 }
 
 // addUsage 累计角色用量。
-func (st *runState) addUsage(role string, promptTokens, completionTokens int) {
+func (st *RunState) addUsage(role string, promptTokens, completionTokens int) {
 	if promptTokens == 0 && completionTokens == 0 {
 		return
 	}
@@ -238,7 +266,7 @@ func (st *runState) addUsage(role string, promptTokens, completionTokens int) {
 }
 
 // usageSnapshot 输出用量快照（值拷贝）。
-func (st *runState) usageSnapshot() map[string]Usage {
+func (st *RunState) usageSnapshot() map[string]Usage {
 	if len(st.Usage) == 0 {
 		return nil
 	}

squad/runstore.go

@@ -0,0 +1,212 @@
+package squad
+
+import (
+	"encoding/json"
+	"fmt"
+	"os"
+	"path/filepath"
+	"time"
+)
+
+// Store 是运行状态持久化端口（Q1 WP4）：run/gate/checkpoint 的读写以
+// 运行工作区为键。现状文件态收编为 FileStore（行为与此前逐文件直写完全一致）；
+// SQLite 实现进 adapters/sqlite（WP5）。事件流仍归 store/eventsink（子包），
+// 不在本端口内重复。
+type Store interface {
+	SaveRun(st *RunState) error
+	LoadRun(wsDir string) (*RunState, error)
+	SaveGate(wsDir string, gate *HitlGate) error
+	LoadGate(wsDir string) (*HitlGate, error)
+	SaveCheckpoint(wsDir string, cp Checkpoint) error
+	LoadLatestCheckpoint(wsDir string) (*Checkpoint, error)
+	// SaveLeader/LoadLeader 持久化 leader 循环状态（raw JSON bytes；
+	// 文件态即 leader.json）。LoadLeader 不存在时返回 (nil, nil)。
+	SaveLeader(wsDir string, data []byte) error
+	LoadLeader(wsDir string) ([]byte, error)
+}
+
+// FileStore 把 runs/<id>/ 目录布局呈现为 Store（现状文件态）。
+type FileStore struct{}
+
+func (FileStore) SaveRun(st *RunState) error {
+	return st.persist()
+}
+
+func (FileStore) LoadRun(wsDir string) (*RunState, error) {
+	return loadRunState(wsDir)
+}
+
+func (FileStore) SaveGate(wsDir string, gate *HitlGate) error {
+	st := &RunState{}
+	st.Meta.WorkDir = wsDir
+	return st.saveGate(gate)
+}
+
+func (FileStore) LoadGate(wsDir string) (*HitlGate, error) {
+	return loadGate(wsDir)
+}
+
+func (FileStore) SaveLeader(wsDir string, data []byte) error {
+	return os.WriteFile(filepath.Join(wsDir, "leader.json"), data, 0o644)
+}
+
+func (FileStore) LoadLeader(wsDir string) ([]byte, error) {
+	data, err := os.ReadFile(filepath.Join(wsDir, "leader.json"))
+	if err != nil {
+		if os.IsNotExist(err) {
+			return nil, nil
+		}
+		return nil, err
+	}
+	return data, nil
+}
+
+func (FileStore) SaveCheckpoint(wsDir string, cp Checkpoint) error {
+	dir := filepath.Join(wsDir, "checkpoints")
+	if err := os.MkdirAll(dir, 0o755); err != nil {
+		return err
+	}
+	path := filepath.Join(dir, fmt.Sprintf("%03d.json", cp.Superstep))
+	if err := writeJSON(path, cp); err != nil {
+		return err
+	}
+	return writeJSON(filepath.Join(dir, "latest.json"), cp)
+}
+
+func (FileStore) LoadLatestCheckpoint(wsDir string) (*Checkpoint, error) {
+	data, err := os.ReadFile(filepath.Join(wsDir, "checkpoints", "latest.json"))
+	if err != nil {
+		if os.IsNotExist(err) {
+			return nil, nil
+		}
+		return nil, err
+	}
+	var cp Checkpoint
+	if err := json.Unmarshal(data, &cp); err != nil {
+		return nil, fmt.Errorf("parse latest checkpoint: %w", err)
+	}
+	return &cp, nil
+}
+
+// Checkpoint 是一个 superstep 边界的状态快照（Q1 WP4）。
+// GraphHash 是图结构指纹：恢复时与当前 Steps 重编译比对，不一致说明崩溃
+// 期间图形态被改过（动态调整或外力），按 WP1 裁定拒绝恢复并提示重新规划。
+type Checkpoint struct {
+	Superstep  int            `json:"superstep"`
+	At         time.Time      `json:"at"`
+	GraphHash  string         `json:"graph_hash"`
+	StepStates []StepSnapshot `json:"steps"`
+	wsDir      string         `json:"-"`
+}
+
+// write 立即同步落盘（经当前 Store 端口——SQLite 后端时进数据库）。
+func (cp Checkpoint) write(store Store) error {
+	return store.SaveCheckpoint(cp.wsDir, cp)
+}
+
+// StepSnapshot 是检查点里的单步快照。
+type StepSnapshot struct {
+	ID        string `json:"id"`
+	Status    string `json:"status"`
+	AttemptID string `json:"attempt_id,omitempty"`
+}
+
+// CheckpointMode 快照写盘档位（Q1 WP4 三档）。
+type CheckpointMode string
+
+const (
+	// CheckpointSync 每个 superstep 边界同步写（默认，与现状「逐步全量落盘」一致）。
+	CheckpointSync CheckpointMode = "sync"
+	// CheckpointAsync 边界快照进队列由单 goroutine 串行写（边界不阻塞；恢复粒度同 sync）。
+	CheckpointAsync CheckpointMode = "async"
+	// CheckpointOnExit 只在推进退出（终态/挂起/错误）时写最终快照。
+	CheckpointOnExit CheckpointMode = "on_exit"
+)
+
+// checkpointWriter 按档位执行边界快照；async 档单 goroutine 串行落盘。
+type checkpointWriter struct {
+	mode  CheckpointMode
+	store Store
+	ch    chan Checkpoint
+	done  chan error
+}
+
+func newCheckpointWriter(mode CheckpointMode, store Store) *checkpointWriter {
+	if mode == "" {
+		mode = CheckpointSync
+	}
+	if store == nil {
+		store = FileStore{}
+	}
+	w := &checkpointWriter{mode: mode, store: store}
+	if mode == CheckpointAsync {
+		w.ch = make(chan Checkpoint, 16)
+		w.done = make(chan error, 1)
+		go func() {
+			var first error
+			for cp := range w.ch {
+				if first != nil {
+					continue
+				}
+				if err := cp.write(w.store); err != nil {
+					first = err
+				}
+			}
+			w.done <- first
+		}()
+	}
+	return w
+}
+
+// boundary 由调用方（flowDriver）在 superstep 边界触发：
+// sync 立即写；async 入队（满即错，不静默丢）；on_exit 跳过。
+func (w *checkpointWriter) boundary(wsDir string, cp Checkpoint) error {
+	cp.wsDir = wsDir
+	switch w.mode {
+	case CheckpointAsync:
+		select {
+		case w.ch <- cp:
+		default:
+			return fmt.Errorf("checkpoint queue full at superstep %d", cp.Superstep)
+		}
+		return nil
+	case CheckpointOnExit:
+		return nil
+	default:
+		return cp.write(w.store)
+	}
+}
+
+// flush 在推进退出前调用：async 落完并报告首个错误；on_exit 写最终快照。
+func (w *checkpointWriter) flush(wsDir string, cp Checkpoint) error {
+	cp.wsDir = wsDir
+	switch w.mode {
+	case CheckpointAsync:
+		close(w.ch)
+		return <-w.done
+	case CheckpointOnExit:
+		return cp.write(w.store)
+	default:
+		return nil
+	}
+}
+
+// validateGraphContinuity 恢复连续性校验（Q1 WP4；WP1 的图版本裁定）。
+// 最新检查点存在且其图指纹与当前 Steps 重编译不一致时拒绝恢复——
+// 崩溃期间图形态被改过，继续推进会把已完成的结果接到错误的图上。
+// 检查点不存在（旧运行）或指纹一致则放行。
+func validateGraphContinuity(st *RunState) error {
+	latest, err := st.storeOr().LoadLatestCheckpoint(st.Meta.WorkDir)
+	if err != nil || latest == nil {
+		return err
+	}
+	g, gerr := compileGraph(st.Plan.Workflow.Steps)
+	if gerr != nil {
+		return fmt.Errorf("graph continuity: 当前 Steps 无法编译：%w（请人工介入或重新规划）", gerr)
+	}
+	if latest.GraphHash != g.hash() {
+		return fmt.Errorf("graph continuity: 崩溃前后图形态不一致（checkpoint superstep=%d hash=%s，当前=%s）——可能在动态调整或外部修改后崩溃，请重新规划或人工介入",
+			latest.Superstep, latest.GraphHash, g.hash())
+	}
+	return nil
+}

squad/termination.go

@@ -0,0 +1,120 @@
+package squad
+
+import (
+	"fmt"
+	"time"
+)
+
+// Termination 运行级停止条件的单一注册点（Q1 WP6）：leader 轮次、修复预算、
+// 唤醒次数与步骤超时集中声明；复合判定用 And/Or 原语（AutoGen 终止条件语义）。
+// 默认值与现网一致；引擎选项在 terminationOf 覆盖对应字段。
+type Termination struct {
+	MaxLeaderTurns   int
+	RepairBudget     int
+	MaxRepairGranted int // 人工 retry 累计加预算上限（Q26）：默认 RepairBudget*2
+	MaxWakeupFires   int
+	StepTimeout      time.Duration
+}
+
+// DefaultTermination 与现网四处的硬编码取值一致。
+func DefaultTermination() Termination {
+	return Termination{
+		MaxLeaderTurns:   DefaultMaxLeaderTurns,
+		RepairBudget:     1,
+		MaxRepairGranted: 2,
+		MaxWakeupFires:   MaxWakeupFires,
+		StepTimeout:      10 * time.Minute,
+	}
+}
+
+// terminationOf 取本运行的停止条件：引擎选项与运行状态覆盖默认值。
+func (e *SquadEngine) terminationOf(st *RunState) Termination {
+	t := DefaultTermination()
+	if e.opts.RepairBudget > 0 {
+		t.RepairBudget = e.opts.RepairBudget
+		t.MaxRepairGranted = e.opts.RepairBudget * 2
+	}
+	if e.opts.StepTimeout > 0 {
+		t.StepTimeout = e.opts.StepTimeout
+	}
+	if st != nil && st.Leader != nil && st.Leader.MaxTurns > 0 {
+		t.MaxLeaderTurns = st.Leader.MaxTurns
+	}
+	return t
+}
+
+// Counters 运行期计数，由各判定处喂入（值语义快照）。
+type Counters struct {
+	LeaderTurns   int
+	RepairSpent   int
+	RepairGranted int
+	WakeupFires   int
+	Deadline      time.Time
+}
+
+// Cond 是一条停止条件：返回（原因，是否命中）。
+type Cond func(c Counters, now time.Time) (string, bool)
+
+// MaxTurnsC 轮次上限。
+func MaxTurnsC(n int) Cond {
+	return func(c Counters, _ time.Time) (string, bool) {
+		if c.LeaderTurns >= n {
+			return fmt.Sprintf("leader 轮次达上限 %d", n), true
+		}
+		return "", false
+	}
+}
+
+// RepairBudgetC 修复预算耗尽。
+func RepairBudgetC(n int) Cond {
+	return func(c Counters, _ time.Time) (string, bool) {
+		if c.RepairSpent > n {
+			return fmt.Sprintf("修复预算耗尽（%d/%d）", c.RepairSpent, n), true
+		}
+		return "", false
+	}
+}
+
+// WakeupsC 自动唤醒次数超限。
+func WakeupsC(n int) Cond {
+	return func(c Counters, _ time.Time) (string, bool) {
+		if c.WakeupFires > n {
+			return fmt.Sprintf("自动唤醒次数超限（%d/%d）", c.WakeupFires, n), true
+		}
+		return "", false
+	}
+}
+
+// DeadlineC 超过截止时刻。
+func DeadlineC() Cond {
+	return func(c Counters, now time.Time) (string, bool) {
+		if !c.Deadline.IsZero() && now.After(c.Deadline) {
+			return "超过截止时刻", true
+		}
+		return "", false
+	}
+}
+
+// And 全部条件命中才成立。
+func And(conds ...Cond) Cond {
+	return func(c Counters, now time.Time) (string, bool) {
+		for _, k := range conds {
+			if _, hit := k(c, now); !hit {
+				return "", false
+			}
+		}
+		return "all conditions met", true
+	}
+}
+
+// Or 任一条件命中即成立（默认的停止语义）。
+func Or(conds ...Cond) Cond {
+	return func(c Counters, now time.Time) (string, bool) {
+		for _, k := range conds {
+			if r, hit := k(c, now); hit {
+				return r, true
+			}
+		}
+		return "", false
+	}
+}

squad/termination_test.go

@@ -0,0 +1,82 @@
+package squad
+
+import (
+	"context"
+	"strings"
+	"testing"
+	"time"
+)
+
+// TestTerminationAndOr：And 全中才成立、Or 任一中即成立（WP6 原语）。
+func TestTerminationAndOr(t *testing.T) {
+	c := Counters{LeaderTurns: 8, RepairSpent: 2, WakeupFires: 0}
+	or := Or(MaxTurnsC(8), WakeupsC(1))
+	if _, hit := or(c, time.Now()); !hit {
+		t.Fatal("or should hit on turns")
+	}
+	and := And(MaxTurnsC(8), WakeupsC(1))
+	if _, hit := and(c, time.Now()); hit {
+		t.Fatal("and should not hit with zero wakeups")
+	}
+	and2 := And(MaxTurnsC(8), RepairBudgetC(1))
+	if _, hit := and2(c, time.Now()); !hit {
+		t.Fatal("and should hit when all met")
+	}
+	dl := DeadlineC()
+	past := Counters{Deadline: time.Now().Add(-time.Minute)}
+	if _, hit := dl(past, time.Now()); !hit {
+		t.Fatal("deadline should hit when past")
+	}
+}
+
+// TestRepairGrantCap：人工 retry 累计追加有上限（Q26）。
+func TestRepairGrantCap(t *testing.T) {
+	engine, _ := newTestEngine(t, mockScriptPlanner(baseSteps()...), map[string]map[string]string{
+		"01_researcher": {"notes.md": goodReport, "fields.json": "{}"},
+		"02_writer":     {"output.md": goodReport},
+		"03_reviewer":   {"output.md": goodReport},
+	}, NopSink{})
+	res, err := engine.Run(context.Background(), Goal{Text: "先跑完再验", HitlMode: HitlAuto}, RunOptions{})
+	if err != nil || res.Status != RunCompleted {
+		t.Fatalf("setup run: %v %+v", err, res)
+	}
+	// 预置：修复授予已达上限 + 一个真实的质检挂起门（与 finalize 的形态一致）。
+	st, err := loadRunState(res.WorkDir)
+	if err != nil {
+		t.Fatal(err)
+	}
+	st.RepairBudgetGranted = engine.terminationOf(st).MaxRepairGranted
+	st.Meta.CompletedAt = nil
+	st.Meta.Status = string(RunQualityReview)
+	req := &HumanRequest{
+		ID: newRequestID(), RunID: st.Meta.ID, Kind: HitlQualityReview, Severity: SeverityActionRequired,
+		Question: "交付物质检未通过且自动修复预算已用尽，请人工裁决：",
+		Options: []HumanOption{
+			{Value: "approve", Label: "人工放行（记录人工通过）"},
+			{Value: "retry", Label: "再修复一次（重跑交付步骤）"},
+			{Value: "abort", Label: "判定失败"},
+		},
+		CreatedAt: time.Now(), Status: "pending",
+	}
+	if err := st.saveGate(&HitlGate{Request: *req, PrevStatus: RunExecuting}); err != nil {
+		t.Fatal(err)
+	}
+	if err := st.save(); err != nil {
+		t.Fatal(err)
+	}
+	// 人工 retry 应被拒绝而不是再追加。
+	out, err := engine.Respond(context.Background(), res.RunID, RespondDecision{Choice: "retry"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if out == nil || !strings.Contains(out.Error, "grants exhausted") {
+		t.Fatalf("expected grant-cap failure, got %+v err=%v", out, err)
+	}
+	st2, err := loadRunState(res.WorkDir)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if st2.RepairBudgetGranted != engine.terminationOf(st2).MaxRepairGranted {
+		t.Fatalf("granted changed after cap: %d", st2.RepairBudgetGranted)
+	}
+}

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 +1,436 @@
+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"},
+		{"driver 未点亮", TopologyReview, "驱动未注册"},
+		{"driver 未点亮 group_chat", TopologyGroupChat, "驱动未注册"},
+		{"driver 未点亮 task_board", TopologyTaskBoard, "驱动未注册"},
+	}
+	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

@@ -13,96 +13,166 @@ import (
 	"error-code-system/errcode"
 )
 
-// flowDriver 工作流驱动器：推进 runState 直到终态或 HITL 挂起边界。
+// flowDriver 工作流驱动器：推进 RunState 直到终态或 HITL 挂起边界。
 type flowDriver struct {
 	opts engineDeps
 }
 
 // advance 执行到边界；返回 (结果, 挂起上下文)，二者互斥非 nil（infra 错误时都为 nil 且 err 非 nil）。
-func (d *flowDriver) advance(ctx context.Context, st *runState) (res *RunResult, gate *hitlGate, err error) {
+// 推进模型（Q1 WP1/WP2）：每轮从权威 Steps 现场编译调度图，按 superstep
+// 分层推进——层内无写集冲突的步骤并行（runParallel 复用），有冲突的按声明序
+// 串行（含动态调整）；动态调整改 Steps 后下一轮自然生效。
+func (d *flowDriver) advance(ctx context.Context, st *RunState) (res *RunResult, gate *HitlGate, err error) {
 	wf := st.Workflow()
 	wf.Status = "in_progress"
 	st.setWorkflow(wf)
 	d.emit(st, Event{Type: EvPhase, Text: "开始执行工作流"})
 
-	i := 0
-	for i < len(st.Plan.Workflow.Steps) {
-		step := st.Plan.Workflow.Steps[i]
-		switch step.Status {
-		case "completed", "skipped", "failed":
-			i++
-			continue
+	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
 		}
-
-		if out, g, err := d.gateBeforeRun(ctx, st, i); out != nil || g != nil || err != nil {
-			if out == nil && g != nil {
-				out = heldResult(st, g)
+	}
+	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
+	defer func() {
+		if lastGraph == nil {
+			if g, gerr := compileGraph(st.Plan.Workflow.Steps); gerr == nil {
+				lastGraph = g
 			}
-			return out, g, err
 		}
+		if lastGraph != nil {
+			_ = cpw.flush(st.Meta.WorkDir, snapshotOf(superNo, lastGraph, st.Plan.Workflow.Steps))
+		}
+	}()
+	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()
+			wf2.Status = "failed"
+			st.setWorkflow(wf2)
+			return &RunResult{RunID: st.Meta.ID, Status: RunFailed, Plan: planPtr(st),
+				Error: "graph: " + gerr.Error(), WorkDir: st.Meta.WorkDir}, nil, nil
+		}
+		layer, ok := g.nextLayer(st.Plan.Workflow.Steps)
+		if !ok {
+			if n := pendingCount(st.Plan.Workflow.Steps); n > 0 {
+				wf2 := st.Workflow()
+				wf2.Status = "failed"
+				st.setWorkflow(wf2)
+				return &RunResult{RunID: st.Meta.ID, Status: RunFailed, Plan: planPtr(st),
+					Error: fmt.Sprintf("graph: 依赖无法满足，仍有 %d 个步骤无法调度", n), WorkDir: st.Meta.WorkDir}, nil, nil
+			}
+			break
+		}
+		superNo++
+		d.emit(st, Event{Type: EvPhase, Text: fmt.Sprintf("superstep %d 推进（并行 %d 步，串行 %d 步）", superNo, len(layer.par), len(layer.seq))})
 
-		if step.Parallel {
-			// 收集相邻并行组。组内每一步都先过审批和权限，任一需要挂起则整组不执行。
-			var group []int
-			for i < len(st.Plan.Workflow.Steps) {
-				s := st.Plan.Workflow.Steps[i]
-				if s.Parallel && s.Status != "completed" && s.Status != "skipped" && s.Status != "failed" {
-					if out, g, err := d.gateBeforeRun(ctx, st, i); out != nil || g != nil || err != nil {
-						if out == nil && g != nil {
-							out = heldResult(st, g)
-						}
-						return out, g, err
+		// 并行段：组内每一步都先过审批和权限，任一需要挂起则整组不执行。
+		if len(layer.par) > 0 {
+			for _, idx := range layer.par {
+				if out, g2, err := d.gateBeforeRun(ctx, st, idx); out != nil || g2 != nil || err != nil {
+					if out == nil && g2 != nil {
+						out = heldResult(st, g2)
 					}
-					group = append(group, i)
-					i++
-					continue
+					return out, g2, err
 				}
-				break
 			}
-			if out, g, err := d.runParallel(ctx, st, group); out != nil || g != nil || err != nil {
-				if out == nil && g != nil {
-					out = heldResult(st, g)
+			if out, g2, err := d.runParallel(ctx, st, layer.par); out != nil || g2 != nil || err != nil {
+				if out == nil && g2 != nil {
+					out = heldResult(st, g2)
 				}
-				return out, g, err
+				return out, g2, err
 			}
 			if g2 := d.checkBlockingGaps(st); g2 != nil {
 				return heldResult(st, g2), g2, nil
 			}
-			continue
-		}
-
-		out, g, err := d.runOneStep(ctx, st, i)
-		if out != nil || g != nil || err != nil {
-			if out == nil && g != nil {
-				out = heldResult(st, g)
-			}
-			return out, g, err
-		}
-		i++
-
-		// 阻塞缺口挂起。
-		if g2 := d.checkBlockingGaps(st); g2 != nil {
-			return heldResult(st, g2), g2, nil
-		}
-
-		// 动态调整（仅顺序步骤后，避免并发竞态）。
-		if st.Plan.Workflow.Steps[i-1].AllowDynamicAdjustment && d.opts.Evaluator != nil {
-			out, g, err := d.applyDynamicAdjustment(ctx, st, i-1)
-			if out != nil || g != nil || err != nil {
-				if out == nil && g != nil {
-					out = heldResult(st, g)
+			// 单成员并行组实际是顺序执行：与旧线性扫描一致，其后可动态调整
+			// （多成员组不做，避免并发竞态）。
+			if len(layer.par) == 1 {
+				idx := layer.par[0]
+				if st.Plan.Workflow.Steps[idx].AllowDynamicAdjustment && d.opts.Evaluator != nil {
+					out, g2, err := d.applyDynamicAdjustment(ctx, st, idx)
+					if out != nil || g2 != nil || err != nil {
+						if out == nil && g2 != nil {
+							out = heldResult(st, g2)
+						}
+						return out, g2, err
+					}
 				}
-				return out, g, err
 			}
 		}
+
+		// 串行段（写集冲突步骤，按声明序逐个执行；顺序步骤后可动态调整）。
+		for _, idx := range layer.seq {
+			if out, g2, err := d.gateBeforeRun(ctx, st, idx); out != nil || g2 != nil || err != nil {
+				if out == nil && g2 != nil {
+					out = heldResult(st, g2)
+				}
+				return out, g2, err
+			}
+			out, g2, err := d.runOneStep(ctx, st, idx)
+			if out != nil || g2 != nil || err != nil {
+				if out == nil && g2 != nil {
+					out = heldResult(st, g2)
+				}
+				return out, g2, err
+			}
+			if g2 := d.checkBlockingGaps(st); g2 != nil {
+				return heldResult(st, g2), g2, nil
+			}
+			if st.Plan.Workflow.Steps[idx].AllowDynamicAdjustment && d.opts.Evaluator != nil {
+				out, g2, err := d.applyDynamicAdjustment(ctx, st, idx)
+				if out != nil || g2 != nil || err != nil {
+					if out == nil && g2 != nil {
+						out = heldResult(st, g2)
+					}
+					return out, g2, err
+				}
+			}
+		}
+
+		// superstep 边界快照（sync 立即 / async 入队；on_exit 由 defer 的 flush 写）。
+		lastGraph = g
+		if err := cpw.boundary(st.Meta.WorkDir, snapshotOf(superNo, g, st.Plan.Workflow.Steps)); err != nil {
+			return nil, nil, err
+		}
 	}
 
 	return d.finalize(ctx, st)
 }
 
+// snapshotOf 组装一个 superstep 边界快照：序号、图结构指纹与全量步骤状态。
+func snapshotOf(superNo int, g *graph, steps []StepSpec) Checkpoint {
+	cp := Checkpoint{Superstep: superNo, At: time.Now(), GraphHash: g.hash()}
+	for _, s := range steps {
+		cp.StepStates = append(cp.StepStates, StepSnapshot{ID: s.ID, Status: s.Status, AttemptID: s.AttemptID})
+	}
+	return cp
+}
+
 // gateBeforeRun 在步骤真正执行前检查权限升级和显式审批。
 // 返回非空结果或挂起时，调用方不得执行该步骤，也不得执行它所在的并行组。
-func (d *flowDriver) gateBeforeRun(ctx context.Context, st *runState, i int) (*RunResult, *hitlGate, error) {
+func (d *flowDriver) gateBeforeRun(ctx context.Context, st *RunState, i int) (*RunResult, *HitlGate, error) {
 	step := st.Plan.Workflow.Steps[i]
 	role, _ := findAgent(st.Plan.Squad.Agents, step.Role)
 	if role.BashPermission == BashPermissionAsk && !st.Approved["role:"+role.Role] {
@@ -143,13 +213,14 @@ func (d *flowDriver) gateBeforeRun(ctx context.Context, st *runState, i int) (*R
 }
 
 // runParallel 并发执行一个并行组；关键步骤失败即终止整个运行。
-func (d *flowDriver) runParallel(ctx context.Context, st *runState, idxs []int) (*RunResult, *hitlGate, error) {
+func (d *flowDriver) runParallel(ctx context.Context, st *RunState, idxs []int) (*RunResult, *HitlGate, error) {
 	ctx, cancel := context.WithCancel(ctx)
 	defer cancel()
 	var wg sync.WaitGroup
 	var mu sync.Mutex
 	failedCritical := -1
-	var gate *hitlGate
+	var firstFail *RunResult
+	var gate *HitlGate
 	for _, idx := range idxs {
 		wg.Add(1)
 		go func(idx int) {
@@ -163,6 +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
 				}
@@ -173,6 +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)
 	}
@@ -180,7 +259,7 @@ func (d *flowDriver) runParallel(ctx context.Context, st *runState, idxs []int)
 }
 
 // runOneStep 执行单个步骤（含瞬态重试与交付物校验）。
-func (d *flowDriver) runOneStep(ctx context.Context, st *runState, idx int) (*RunResult, *hitlGate, error) {
+func (d *flowDriver) runOneStep(ctx context.Context, st *RunState, idx int) (*RunResult, *HitlGate, error) {
 	if err := ctx.Err(); err != nil {
 		return &RunResult{RunID: st.Meta.ID, Status: RunCancelled, Error: err.Error()}, nil, err
 	}
@@ -211,6 +290,7 @@ func (d *flowDriver) runOneStep(ctx context.Context, st *runState, idx int) (*Ru
 	d.emit(st, Event{Type: EvStepStarted, StepID: step.ID, Role: step.Role, Text: step.Title})
 	step.Status = "running"
 	step.Attempts++
+	step.AttemptID = fmt.Sprintf("%s/%s/%d", st.Meta.ID, step.ID, step.Attempts)
 	step.Error = ""
 	// 与收口落盘同一把锁：并行步骤一边改 Steps、一边全量 marshal 会数据竞争。
 	st.mu.Lock()
@@ -229,6 +309,15 @@ 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)
+	if len(mail) > 0 {
+		_ = st.save()
+	}
+	st.mu.Unlock()
+	prompt += formatMailSection(mail)
 	instructions := readFileOr(filepath.Join(subDir, "AGENTS.md"), d.opts.Workspaces.ResolveInstructions(role))
 
 	var backend Executor
@@ -283,7 +372,7 @@ func (d *flowDriver) runOneStep(ctx context.Context, st *runState, idx int) (*Ru
 			ToolAllowlist: role.Tools,
 			Skills:        firstNonEmptySlice(step.Skills, role.Skills),
 			SessionScope:  st.Meta.ID + "/" + step.ID,
-			ExecutionID:   st.Meta.ID + "-" + step.ID,
+			ExecutionID:   step.AttemptID,
 			Budget:        stepBudget,
 		}
 		if step.IntentID != "" {
@@ -374,11 +463,150 @@ 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})})
+
+	// 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) {
+func (d *flowDriver) failStep(ctx context.Context, st *RunState, idx int) (*RunResult, *HitlGate, error) {
 	st.mu.Lock()
 	step := st.Plan.Workflow.Steps[idx]
 	critical := step.IsCritical()
@@ -402,7 +630,7 @@ func (d *flowDriver) failStep(ctx context.Context, st *runState, idx int) (*RunR
 }
 
 // checkBlockingGaps 检查未解决阻塞缺口；interactive 挂起，auto 记录兜底假设并继续。
-func (d *flowDriver) checkBlockingGaps(st *runState) *hitlGate {
+func (d *flowDriver) checkBlockingGaps(st *RunState) *HitlGate {
 	gaps := st.BB.UnresolvedBlockingGaps()
 	if len(gaps) == 0 {
 		return nil
@@ -444,7 +672,7 @@ func (d *flowDriver) checkBlockingGaps(st *runState) *hitlGate {
 		Context:   strings.Join(lines, "\n"),
 		CreatedAt: time.Now(), Status: "pending",
 	}
-	gate := &hitlGate{Request: *req, PrevStatus: RunExecuting}
+	gate := &HitlGate{Request: *req, PrevStatus: RunExecuting}
 	if err := st.saveGate(gate); err == nil {
 		st.Meta.Status = string(RunWaitingHuman)
 		_ = st.save()
@@ -454,7 +682,7 @@ func (d *flowDriver) checkBlockingGaps(st *runState) *hitlGate {
 }
 
 // applyDynamicAdjustment 评估步骤产出并应用/审批动态调整。
-func (d *flowDriver) applyDynamicAdjustment(ctx context.Context, st *runState, idx int) (*RunResult, *hitlGate, error) {
+func (d *flowDriver) applyDynamicAdjustment(ctx context.Context, st *RunState, idx int) (*RunResult, *HitlGate, error) {
 	step := st.Plan.Workflow.Steps[idx]
 	subDir := filepath.Join(st.Meta.WorkDir, filepath.FromSlash(step.SubWorkspace))
 	output := readFileOr(filepath.Join(subDir, firstTarget(step.Targets)), "")
@@ -494,7 +722,7 @@ func (d *flowDriver) applyDynamicAdjustment(ctx context.Context, st *runState, i
 				evalRes.Reason, strings.Join(addLines, "\n"), strings.Join(skipLines, "\n")),
 			CreatedAt: time.Now(), Status: "pending",
 		}
-		gate := &hitlGate{Request: *req, Steps: evalRes.DynamicStepsToAdd, SkipIDs: evalRes.StepsToSkip, PrevStatus: RunExecuting}
+		gate := &HitlGate{Request: *req, Steps: evalRes.DynamicStepsToAdd, SkipIDs: evalRes.StepsToSkip, PrevStatus: RunExecuting}
 		if err := st.saveGate(gate); err == nil {
 			st.Meta.Status = string(RunWaitingHuman)
 			_ = st.save()
@@ -507,7 +735,7 @@ func (d *flowDriver) applyDynamicAdjustment(ctx context.Context, st *runState, i
 }
 
 // applyAdjustment 应用动态调整：新步骤插到最后一步之前，跳过步骤标记 skipped。
-func applyAdjustment(st *runState, add []StepSpec, skipIDs []string) {
+func applyAdjustment(st *RunState, add []StepSpec, skipIDs []string) {
 	steps := st.Plan.Workflow.Steps
 	for _, id := range skipIDs {
 		for i := range steps {
@@ -534,11 +762,15 @@ func applyAdjustment(st *runState, add []StepSpec, skipIDs []string) {
 }
 
 // finalize 收口：汇总交付物 → 质检 → 自动修复 → （必要时）人工裁决 → 清单。
-func (d *flowDriver) finalize(ctx context.Context, st *runState) (*RunResult, *hitlGate, error) {
+func (d *flowDriver) finalize(ctx context.Context, st *RunState) (*RunResult, *HitlGate, error) {
 	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
 		}
 	}
@@ -616,7 +848,7 @@ func (d *flowDriver) finalize(ctx context.Context, st *runState) (*RunResult, *h
 				Context:   string(marshalJSON(report)),
 				CreatedAt: time.Now(), Status: "pending",
 			}
-			gate := &hitlGate{Request: *req, PrevStatus: RunExecuting}
+			gate := &HitlGate{Request: *req, PrevStatus: RunExecuting}
 			if err := st.saveGate(gate); err == nil {
 				st.Meta.Status = string(RunQualityReview)
 				_ = st.save()
@@ -650,7 +882,7 @@ func (d *flowDriver) finalize(ctx context.Context, st *runState) (*RunResult, *h
 }
 
 // rerunForRepair 重置交付物源步骤并重新执行到收口。
-func (d *flowDriver) rerunForRepair(ctx context.Context, st *runState, stepID string) (*RunResult, *hitlGate, error) {
+func (d *flowDriver) rerunForRepair(ctx context.Context, st *RunState, stepID string) (*RunResult, *HitlGate, error) {
 	for i := range st.Plan.Workflow.Steps {
 		if st.Plan.Workflow.Steps[i].ID == stepID {
 			st.Plan.Workflow.Steps[i].Status = "pending"
@@ -668,7 +900,7 @@ func (d *flowDriver) rerunForRepair(ctx context.Context, st *runState, stepID st
 }
 
 // heldResult 由挂起上下文构造对外结果（挂起路径统一返回非 nil 结果）。
-func heldResult(st *runState, gate *hitlGate) *RunResult {
+func heldResult(st *RunState, gate *HitlGate) *RunResult {
 	status := RunWaitingHuman
 	switch gate.Request.Kind {
 	case HitlClarification:
@@ -698,7 +930,7 @@ func waitConditionFrom(gaps []GapItem) *WaitCondition {
 }
 
 // hold 生成挂起请求并返回结果。
-func (d *flowDriver) hold(st *runState, kind HumanRequestKind, stepID, role, question string, options []HumanOption, context, reason string, extra *hitlGate) (*RunResult, *hitlGate, error) {
+func (d *flowDriver) hold(st *RunState, kind HumanRequestKind, stepID, role, question string, options []HumanOption, context, reason string, extra *HitlGate) (*RunResult, *HitlGate, error) {
 	req := &HumanRequest{
 		ID: newRequestID(), RunID: st.Meta.ID, StepID: stepID, Kind: kind,
 		Severity: SeverityActionRequired,
@@ -714,7 +946,7 @@ func (d *flowDriver) hold(st *runState, kind HumanRequestKind, stepID, role, que
 	}
 	gate := extra
 	if gate == nil {
-		gate = &hitlGate{}
+		gate = &HitlGate{}
 	}
 	gate.Request = *req
 	gate.PrevStatus = RunExecuting
@@ -731,7 +963,7 @@ func (d *flowDriver) hold(st *runState, kind HumanRequestKind, stepID, role, que
 // policy 合并引擎全局策略与 Goal 覆盖（窄者胜）。
 // policyRole 在引擎全局/Goal 链之上叠加角色级 HitlPolicy
 // （窄者胜；Goal 仍最高）。
-func (d *flowDriver) policyRole(st *runState, role RoleSpec) HitlPolicy {
+func (d *flowDriver) policyRole(st *RunState, role RoleSpec) HitlPolicy {
 	p := d.policy(st)
 	if role.HitlPolicy != nil {
 		p = mergeHitlPolicy(p, role.HitlPolicy)
@@ -739,11 +971,11 @@ func (d *flowDriver) policyRole(st *runState, role RoleSpec) HitlPolicy {
 	return p
 }
 
-func (d *flowDriver) policy(st *runState) HitlPolicy {
+func (d *flowDriver) policy(st *RunState) HitlPolicy {
 	return mergeHitlPolicy(d.opts.HitlPolicy, st.Meta.HitlPolicyOf())
 }
 
-func (d *flowDriver) emit(st *runState, ev Event) {
+func (d *flowDriver) emit(st *RunState, ev Event) {
 	if d.opts.Sink == nil {
 		return
 	}
@@ -754,11 +986,11 @@ func (d *flowDriver) emit(st *runState, ev Event) {
 	d.opts.Sink.Emit(ev)
 }
 
-func (d *flowDriver) emitHuman(st *runState, req *HumanRequest) {
+func (d *flowDriver) emitHuman(st *RunState, req *HumanRequest) {
 	d.emit(st, Event{Type: EvHumanInputRequired, Text: req.Question, Data: marshalJSON(req)})
 }
 
-func summarize(st *runState, report QualityReport, status RunStatus) string {
+func summarize(st *RunState, report QualityReport, status RunStatus) string {
 	var b strings.Builder
 	fmt.Fprintf(&b, "运行 %s：%s。步骤 %d 个（完成 %d/跳过 %d/失败 %d）。质检得分 %.0f。",
 		st.Meta.ID, status, len(st.Plan.Workflow.Steps),
@@ -769,7 +1001,7 @@ func summarize(st *runState, report QualityReport, status RunStatus) string {
 	return b.String()
 }
 
-func countStatus(st *runState, status string) int {
+func countStatus(st *RunState, status string) int {
 	n := 0
 	for _, s := range st.Plan.Workflow.Steps {
 		if s.Status == status {
@@ -779,7 +1011,7 @@ func countStatus(st *runState, status string) int {
 	return n
 }
 
-func buildManifest(st *runState, deliverable string, report QualityReport) Manifest {
+func buildManifest(st *RunState, deliverable string, report QualityReport) Manifest {
 	files := map[string]string{}
 	_ = filepath.Walk(st.Meta.WorkDir, func(path string, info os.FileInfo, err error) error {
 		if err != nil || info.IsDir() {
@@ -838,7 +1070,7 @@ func appendLog(dir, line string) {
 	_, _ = f.WriteString(line)
 }
 
-func planPtr(st *runState) *WorkflowPlan {
+func planPtr(st *RunState) *WorkflowPlan {
 	wf := st.Workflow()
 	return &wf
 }

squad/workspace.go

@@ -34,18 +34,9 @@ func (e *WorkspaceEngine) PrepareWorkspace(meta *RunMetadata, plan *SquadPlan) (
 	}
 	meta.WorkDir = wsDir
 	meta.UpdatedAt = time.Now()
-	if err := writeJSON(filepath.Join(wsDir, "workspace.json"), meta); err != nil {
-		return "", err
-	}
-	if err := writeJSON(filepath.Join(wsDir, "squad.json"), &plan.Squad); err != nil {
-		return "", err
-	}
-	if err := writeJSON(filepath.Join(wsDir, "workflow.json"), &plan.Workflow); err != nil {
-		return "", err
-	}
-	if err := saveBlackboard(wsDir, &BlackboardState{Fields: map[string]string{}}); err != nil {
-		return "", err
-	}
+	// 注意：run 级状态 JSON 不在此处写——由 st.save() 经 Store 端口统一落盘
+	// （Q1 WP5：SQLite 后端时这些状态进数据库而非文件）。这里只建目录与
+	// 上下文/模板等运行无关状态的工作区资产。
 	ctxIdx, err := e.writeContextAssets(wsDir, &meta.Goal, plan)
 	if err != nil {
 		return "", err