feat(history): 会话消息持久化 + 主/次级会话（design/16）

- runtime/history: Store + HistoryStore 接口（Append/List 游标分页/Recent/DeleteSession/NewSubSession/ListSessions），主会话固定 web、次级 web:<uuid>
- db 迁移 0003: chat_sessions + chat_messages + idx_chat_msgs_session_created（事务幂等）
- gateway: GET/POST /sessions、GET /sessions/{id}/messages 热分页、DELETE /sessions/{id}（主不可删 403）；Deps.History 可空注入
- 编排层落库: session_persist.go（submitInternal 写用户消息、completeTurn 写助手消息，拒绝的消息不落库）
- WS: wsEvent 加 sessionId 定向 + history 回放帧（握手上行后回放主会话最新 50 条，thinking 剥离）
- wire.go 装配 History 到 gateway.Deps.History
- 端到端实测: POST /sessions 建次级 201 / GET /sessions 列表 / POST /chat 后 /messages 读到 user+assistant / DELETE 次级成功 / DELETE 主会话 403

cmd/cyrene/wire.go

@@ -23,6 +23,7 @@ import (
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/persona/loader"
 	personareg "git.yeij.top/AskaEth/Cyrene_v2/internal/persona/registry"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/platform"
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/history"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/agent"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/capacity"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/embedding"
@@ -92,6 +93,7 @@ type app struct {
 	backthink     *backthink.Backthink
 	vindex        vector.Index
 	owner         *user.OwnerProfile
+	history       *history.Store
 	gw            *gateway.Server
 	channelBridge *channelbridge.Bridge
 }
@@ -510,6 +512,11 @@ func (a *app) wireFrontend() error {
 		a.log.Warn("audit store unavailable, falling back to in-memory", "error", aerr)
 		auditStore = nil
 	}
+	hist, herr := history.NewStore(context.Background(), a.db, a.log)
+	if herr != nil {
+		return fmt.Errorf("history store: %w", herr)
+	}
+	a.history = hist
 	gw, err := gateway.New(a.cfg, gateway.Deps{
 		Log:         logger.WithSource(a.log, "gateway"),
 		Auth:        jwt,
@@ -519,6 +526,7 @@ func (a *app) wireFrontend() error {
 		Session:     a.sessions,
 		Personas:    a.personas,
 		Memories:    a.memStore,
+		History:     a.history,
 		IoT:         a.iot,
 		Scheduler:   a.scheduler,
 		Tasks:       a.tasks,

internal/db/migrations/0003_chat_history.down.sql

@@ -0,0 +1,4 @@
+-- 0003_chat_history: rolls back the chat history tables created by the up migration.
+DROP TABLE IF EXISTS chat_messages;
+DROP INDEX IF EXISTS idx_chat_msgs_session_created;
+DROP TABLE IF EXISTS chat_sessions;

internal/db/migrations/0003_chat_history.up.sql

@@ -0,0 +1,31 @@
+-- 0003_chat_history: persistent chat sessions + message stream (DevDocs/design/16 §2.1, §3.1).
+-- chat_sessions records the persistent sessions: one fixed main session plus arbitrarily
+-- many secondary (sub) sessions. chat_messages is the raw, durable message stream in
+-- millisecond-timestamp cursor order. The two are the backend truth for the Web frontend;
+-- they are NOT memory (design/16 §3.2 separates memory from the raw stream).
+--
+-- Its canonical schema is owned by internal/runtime/history, which reuses these tables. The
+-- migration is the authoritative path; internal/runtime/history.NewStore re-issues the same
+-- DDL idempotently so the store is self-contained (matching memory/session).
+
+CREATE TABLE IF NOT EXISTS chat_sessions (
+    id         TEXT PRIMARY KEY,            -- main = fixed id ("web"); sub = "web:<uuid>"
+    type       TEXT NOT NULL,               -- 'main' | 'sub'
+    title      TEXT NOT NULL DEFAULT '',
+    created_at INTEGER NOT NULL             -- ms timestamp
+);
+
+CREATE TABLE IF NOT EXISTS chat_messages (
+    id         TEXT PRIMARY KEY,            -- message unique id (uuid)
+    session_id TEXT NOT NULL,               -- owning session (main or sub)
+    role       TEXT NOT NULL,               -- user / assistant / system
+    content    TEXT NOT NULL,               -- raw message body (incl. <p>/<action>)
+    mode       TEXT,                        -- behavior mode (optional, bubble render)
+    mood       TEXT,                        -- mood (optional)
+    image      TEXT,                        -- attachment base64/URL (optional)
+    thinking   TEXT,                        -- chain of thought (optional, stored, not echoed by default)
+    created_at INTEGER NOT NULL             -- ms timestamp (pagination cursor)
+);
+
+CREATE INDEX IF NOT EXISTS idx_chat_msgs_session_created
+    ON chat_messages (session_id, created_at DESC);

internal/runtime/history/store.go

@@ -0,0 +1,390 @@
+// Package history implements the persistent chat-history session model described in
+// DevDocs/design/16-会话持久化与主次级会话.md.
+//
+// It is the storage layer, not the orchestration/context layer: it persists the raw
+// message stream (chat_messages) and the persistent session records (chat_sessions)
+// so the Web frontend can request + display history, and secondary (sub) sessions can
+// be created, listed and deleted. It intentionally does not touch memory (design/16
+// §3.2: memory is persona-scoped and lives in runtime/memory), does not build context
+// (runtime/loop) and does not schedule turns (gateway/session_orchestrator).
+//
+// Session model (design/16 §2.1): there is exactly one fixed main session (id "web") and
+// any number of sub sessions (id "web:<uuid>"). Main + sub share the same persona and
+// therefore the same memory; this package only models sessions + messages.
+package history
+
+import (
+	"context"
+	"crypto/rand"
+	"database/sql"
+	"encoding/hex"
+	"errors"
+	"fmt"
+	"io"
+	"log/slog"
+	"time"
+
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/logger"
+)
+
+// Session roles and types. These are the values persisted in the role / type columns.
+const (
+	// RoleUser, RoleAssistant, RoleSystem classify a message's author.
+	RoleUser      = "user"
+	RoleAssistant = "assistant"
+	RoleSystem    = "system"
+
+	// SessionMain and SessionSub classify a persistent session.
+	SessionMain = "main"
+	SessionSub  = "sub"
+)
+
+// MainSessionID is the fixed id of the one main session (design/16 §2.1). Sub session
+// ids are MainSessionID + ":" + uuid.
+const MainSessionID = "web"
+
+// defaultPageLimit is the page size used when the caller does not pass one.
+const defaultPageLimit = 50
+
+// maxPageLimit bounds a client-supplied limit so a single request cannot drain memory.
+const maxPageLimit = 200
+
+// Sentinels distinguishing history package failures. Callers wrap them with %w so
+// errors.Is works through the chain.
+var (
+	// ErrNilDB is returned when NewStore receives a nil *sql.DB.
+	ErrNilDB = errors.New("history: nil db")
+
+	// ErrImmutableMain is returned when a caller tries to delete the fixed main session.
+	ErrImmutableMain = errors.New("history: the main session cannot be deleted")
+)
+
+// Message is one persisted chat message. CreatedAt is a Unix millisecond timestamp and
+// doubles as the pagination cursor (design/16 §3.1): pages are ordered newest-first and
+// "before" is the created_at of the oldest message on the previous page.
+type Message struct {
+	ID        string
+	SessionID string
+	Role      string
+	Content   string
+	Mode      string
+	Mood      string
+	Image     string
+	Thinking  string
+	CreatedAt int64
+}
+
+// SessionMeta is the wire-facing metadata of one persistent session.
+type SessionMeta struct {
+	ID        string
+	Type      string // "main" | "sub"
+	Title     string
+	CreatedAt int64
+}
+
+// HistoryStore abstracts the chat-history store so callers (the gateway) and tests can
+// hold an interface rather than a concrete *Store.
+type HistoryStore interface {
+	// Append persists msg, assigning an id and timestamp when the caller omitted them.
+	Append(ctx context.Context, msg Message) error
+	// List returns up to limit messages for sessionID, newest first. before is a
+	// millisecond timestamp cursor: when nonzero, only messages older than before are
+	// returned (inclusive-free, matching the "before" bound on the frontend).
+	List(ctx context.Context, sessionID string, before int64, limit int) ([]Message, error)
+	// Recent returns the most recent n messages for sessionID in oldest-first order,
+	// for WS replay / context tail usage.
+	Recent(ctx context.Context, sessionID string, n int) ([]Message, error)
+	// DeleteSession removes a sub session and all its messages. The main session is
+	// immutable and returns ErrImmutableMain.
+	DeleteSession(ctx context.Context, sessionID string) error
+	// NewSubSession creates a new sub session and returns its id (MainSessionID:uuid).
+	NewSubSession(ctx context.Context, ownerID string) (string, error)
+	// ListSessions returns the main session first then every sub session, oldest first.
+	ListSessions(ctx context.Context) ([]SessionMeta, error)
+}
+
+// Store is the SQLite-backed chat-history store. It owns the chat_sessions and
+// chat_messages tables. Concurrency is delegated to *sql.DB (safe for concurrent use),
+// so Store holds no further lock. Every query is parameterized and reads list columns
+// explicitly (no SELECT *); reads take a context.
+type Store struct {
+	db  *sql.DB
+	log *slog.Logger
+}
+
+// rowScanner is satisfied by both *sql.Row and *sql.Rows, so a single scanMessage
+// helper serves List and Recent.
+type rowScanner interface {
+	Scan(dest ...any) error
+}
+
+// NewStore builds a Store on db, ensuring the chat_sessions/chat_messages schema exists
+// and that the fixed main session row is present. The DDL mirrors the 0003 migration;
+// it is idempotent so the store is self-contained and usable from any *sql.DB, matching
+// the memory and session packages. db must be live; nil returns ErrNilDB. log may be nil.
+func NewStore(ctx context.Context, db *sql.DB, log *slog.Logger) (*Store, error) {
+	if db == nil {
+		return nil, ErrNilDB
+	}
+	log = logger.WithSource(discardLogger(log), "history")
+	for _, ddl := range []string{
+		createSessionsTable,
+		createMessagesTable,
+		createMessagesIndex,
+	} {
+		if _, err := db.ExecContext(ctx, ddl); err != nil {
+			return nil, fmt.Errorf("history: apply schema: %w", err)
+		}
+	}
+	s := &Store{db: db, log: log}
+	if err := s.ensureMainSession(ctx); err != nil {
+		return nil, err
+	}
+	return s, nil
+}
+
+// Append persists msg, assigning a fresh id and a now-ms CreatedAt when the caller left
+// either empty/zero. It returns the error wrapped with the operation name.
+func (s *Store) Append(ctx context.Context, msg Message) error {
+	if msg.ID == "" {
+		id, err := newID()
+		if err != nil {
+			return fmt.Errorf("history: append id: %w", err)
+		}
+		msg.ID = id
+	}
+	if msg.CreatedAt == 0 {
+		msg.CreatedAt = time.Now().UnixMilli()
+	}
+	_, err := s.db.ExecContext(ctx, `INSERT INTO chat_messages
+		(id, session_id, role, content, mode, mood, image, thinking, created_at)
+		VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?)`,
+		msg.ID, msg.SessionID, msg.Role, msg.Content, msg.Mode, msg.Mood,
+		msg.Image, msg.Thinking, msg.CreatedAt)
+	if err != nil {
+		return fmt.Errorf("history: append %s to %s: %w", msg.Role, msg.SessionID, err)
+	}
+	return nil
+}
+
+// List returns up to limit messages for sessionID, newest first. When before is nonzero
+// it bounds the query to messages strictly older than before, so a client can page upward
+// by passing the oldest message's created_at from the previous page. The returned slice
+// is non-nil. limit is clamped to [1, maxPageLimit].
+func (s *Store) List(ctx context.Context, sessionID string, before int64, limit int) ([]Message, error) {
+	limit = clampLimit(limit)
+	query := `SELECT id, session_id, role, content, mode, mood, image, thinking, created_at
+		FROM chat_messages WHERE session_id = ?`
+	args := []any{sessionID}
+	if before > 0 {
+		query += ` AND created_at < ?`
+		args = append(args, before)
+	}
+	query += ` ORDER BY created_at DESC LIMIT ?`
+	args = append(args, limit)
+
+	rows, err := s.db.QueryContext(ctx, query, args...)
+	if err != nil {
+		return nil, fmt.Errorf("history: list %s: %w", sessionID, err)
+	}
+	defer func() { _ = rows.Close() }()
+
+	msgs := make([]Message, 0)
+	for rows.Next() {
+		m, err := scanMessage(rows)
+		if err != nil {
+			return nil, fmt.Errorf("history: scan list %s: %w", sessionID, err)
+		}
+		msgs = append(msgs, m)
+	}
+	if err := rows.Err(); err != nil {
+		return nil, fmt.Errorf("history: iterate list %s: %w", sessionID, err)
+	}
+	return msgs, nil
+}
+
+// Recent returns the most recent n messages for sessionID in oldest-first order (the
+// tail-most window), for WS replay / context tail. The returned slice is non-nil.
+func (s *Store) Recent(ctx context.Context, sessionID string, n int) ([]Message, error) {
+	if n < 1 {
+		n = defaultPageLimit
+	}
+	rows, err := s.db.QueryContext(ctx, `SELECT id, session_id, role, content, mode, mood, image, thinking, created_at
+		FROM chat_messages WHERE session_id = ? ORDER BY created_at DESC LIMIT ?`, sessionID, n)
+	if err != nil {
+		return nil, fmt.Errorf("history: recent %s: %w", sessionID, err)
+	}
+	defer func() { _ = rows.Close() }()
+
+	// Recent returns newest-last so the window is in chronological order; build it and
+	// reverse in place.
+	msgs := make([]Message, 0)
+	for rows.Next() {
+		m, err := scanMessage(rows)
+		if err != nil {
+			return nil, fmt.Errorf("history: scan recent %s: %w", sessionID, err)
+		}
+		msgs = append(msgs, m)
+	}
+	if err := rows.Err(); err != nil {
+		return nil, fmt.Errorf("history: iterate recent %s: %w", sessionID, err)
+	}
+	for i, j := 0, len(msgs)-1; i < j; i, j = i+1, j-1 {
+		msgs[i], msgs[j] = msgs[j], msgs[i]
+	}
+	return msgs, nil
+}
+
+// DeleteSession removes a sub session and all its messages in one transaction. The fixed
+// main session is immutable and returns ErrImmutableMain (design/16 §6).
+func (s *Store) DeleteSession(ctx context.Context, sessionID string) error {
+	if sessionID == MainSessionID {
+		return ErrImmutableMain
+	}
+	tx, err := s.db.BeginTx(ctx, nil)
+	if err != nil {
+		return fmt.Errorf("history: begin delete %s: %w", sessionID, err)
+	}
+	defer func() { _ = tx.Rollback() }()
+
+	if _, err := tx.ExecContext(ctx, `DELETE FROM chat_messages WHERE session_id = ?`, sessionID); err != nil {
+		return fmt.Errorf("history: delete messages %s: %w", sessionID, err)
+	}
+	if _, err := tx.ExecContext(ctx, `DELETE FROM chat_sessions WHERE id = ? AND type = ?`, sessionID, SessionSub); err != nil {
+		return fmt.Errorf("history: delete session %s: %w", sessionID, err)
+	}
+	if err := tx.Commit(); err != nil {
+		return fmt.Errorf("history: commit delete %s: %w", sessionID, err)
+	}
+	return nil
+}
+
+// NewSubSession creates a sub session and returns its id (MainSessionID:uuid). ownerID is
+// accepted for interface completeness (the Web owner is identified by JWT, not stored per
+// session); it is logged at debug and not persisted.
+func (s *Store) NewSubSession(ctx context.Context, ownerID string) (string, error) {
+	id, err := newID()
+	if err != nil {
+		return "", fmt.Errorf("history: new sub id: %w", err)
+	}
+	sessionID := MainSessionID + ":" + id
+	now := time.Now().UnixMilli()
+	if _, err := s.db.ExecContext(ctx, `INSERT INTO chat_sessions (id, type, title, created_at)
+		VALUES (?, ?, ?, ?)`, sessionID, SessionSub, "", now); err != nil {
+		return "", fmt.Errorf("history: insert sub session: %w", err)
+	}
+	if s.log != nil {
+		s.log.Debug("sub session created", "session_id", sessionID, "owner_id", ownerID)
+	}
+	return sessionID, nil
+}
+
+// ListSessions returns the main session first, then every sub session, oldest first. The
+// returned slice is non-nil.
+func (s *Store) ListSessions(ctx context.Context) ([]SessionMeta, error) {
+	rows, err := s.db.QueryContext(ctx, `SELECT id, type, title, created_at FROM chat_sessions
+		ORDER BY (type = ?) DESC, created_at ASC`, SessionMain)
+	if err != nil {
+		return nil, fmt.Errorf("history: list sessions: %w", err)
+	}
+	defer func() { _ = rows.Close() }()
+
+	sessions := make([]SessionMeta, 0)
+	for rows.Next() {
+		var (
+			meta      SessionMeta
+			typeStr   string
+			createdAt int64
+		)
+		if err := rows.Scan(&meta.ID, &typeStr, &meta.Title, &createdAt); err != nil {
+			return nil, fmt.Errorf("history: scan session: %w", err)
+		}
+		meta.Type = typeStr
+		meta.CreatedAt = createdAt
+		sessions = append(sessions, meta)
+	}
+	if err := rows.Err(); err != nil {
+		return nil, fmt.Errorf("history: iterate sessions: %w", err)
+	}
+	return sessions, nil
+}
+
+// ensureMainSession inserts the fixed main session row when it is absent, so ListSessions
+// always reports the main session even before any message is written.
+func (s *Store) ensureMainSession(ctx context.Context) error {
+	_, err := s.db.ExecContext(ctx, `INSERT OR IGNORE INTO chat_sessions (id, type, title, created_at)
+		VALUES (?, ?, ?, ?)`, MainSessionID, SessionMain, "", time.Now().UnixMilli())
+	if err != nil {
+		return fmt.Errorf("history: ensure main session: %w", err)
+	}
+	return nil
+}
+
+// scanMessage reads one chat_messages row. message columns are scanned positionally.
+func scanMessage(sc rowScanner) (Message, error) {
+	var m Message
+	var mode, mood, image, thinking string
+	if err := sc.Scan(&m.ID, &m.SessionID, &m.Role, &m.Content, &mode, &mood, &image, &thinking, &m.CreatedAt); err != nil {
+		return Message{}, err
+	}
+	m.Mode, m.Mood, m.Image, m.Thinking = mode, mood, image, thinking
+	return m, nil
+}
+
+// clampLimit bounds limit to [1, maxPageLimit]; a caller-supplied 0 or negative falls back
+// to defaultPageLimit.
+func clampLimit(limit int) int {
+	if limit < 1 {
+		return defaultPageLimit
+	}
+	if limit > maxPageLimit {
+		return maxPageLimit
+	}
+	return limit
+}
+
+// newID returns a 128-bit cryptographically random id, hex-encoded. It avoids a UUID
+// dependency so the package stays self-contained.
+func newID() (string, error) {
+	b := make([]byte, 16)
+	if _, err := rand.Read(b); err != nil {
+		return "", fmt.Errorf("history: generate id: %w", err)
+	}
+	return hex.EncodeToString(b), nil
+}
+
+// discardLogger returns log if non-nil, otherwise a logger that writes nowhere.
+func discardLogger(log *slog.Logger) *slog.Logger {
+	if log != nil {
+		return log
+	}
+	return slog.New(slog.NewTextHandler(io.Discard, nil))
+}
+
+// Schema DDL. The 0003 migration in internal/db is the authoritative mechanism for
+// domain tables, but it is intentionally also re-issued here idempotently so the store is
+// self-contained and usable from any *sql.DB (matching the memory and session packages).
+const createSessionsTable = `
+CREATE TABLE IF NOT EXISTS chat_sessions (
+    id         TEXT PRIMARY KEY,
+    type       TEXT NOT NULL,
+    title      TEXT NOT NULL DEFAULT '',
+    created_at INTEGER NOT NULL
+)`
+
+const createMessagesTable = `
+CREATE TABLE IF NOT EXISTS chat_messages (
+    id         TEXT PRIMARY KEY,
+    session_id TEXT NOT NULL,
+    role       TEXT NOT NULL,
+    content    TEXT NOT NULL,
+    mode       TEXT,
+    mood       TEXT,
+    image      TEXT,
+    thinking   TEXT,
+    created_at INTEGER NOT NULL
+)`
+
+const createMessagesIndex = `
+CREATE INDEX IF NOT EXISTS idx_chat_msgs_session_created
+    ON chat_messages (session_id, created_at DESC)`

internal/runtime/history/store_test.go

@@ -0,0 +1,204 @@
+package history
+
+import (
+	"context"
+	"errors"
+	"io"
+	"log/slog"
+	"strings"
+	"testing"
+
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/db"
+)
+
+// testLogger returns a logger that writes nowhere, keeping test output clean.
+func testLogger() *slog.Logger {
+	return slog.New(slog.NewTextHandler(io.Discard, nil))
+}
+
+// newStore opens an isolated in-memory DB and returns a Store on it.
+func newStore(t *testing.T) *Store {
+	t.Helper()
+	ctx := context.Background()
+	conn, err := db.Open(ctx, ":memory:", testLogger())
+	if err != nil {
+		t.Fatalf("open memory db: %v", err)
+	}
+	t.Cleanup(func() { _ = conn.Close() })
+	st, err := NewStore(ctx, conn, testLogger())
+	if err != nil {
+		t.Fatalf("new store: %v", err)
+	}
+	return st
+}
+
+func TestNewStoreNilDB(t *testing.T) {
+	t.Parallel()
+	if _, err := NewStore(context.Background(), nil, testLogger()); !errors.Is(err, ErrNilDB) {
+		t.Fatalf("err = %v, want wrap of ErrNilDB", err)
+	}
+}
+
+func TestAppendAndListNewestFirst(t *testing.T) {
+	t.Parallel()
+	st := newStore(t)
+	ctx := context.Background()
+
+	for i, ts := range []int64{1000, 2000, 3000} {
+		m := Message{SessionID: MainSessionID, Role: RoleUser, Content: "hi", CreatedAt: ts}
+		if err := st.Append(ctx, m); err != nil {
+			t.Fatalf("append %d: %v", i, err)
+		}
+	}
+
+	msgs, err := st.List(ctx, MainSessionID, 0, 50)
+	if err != nil {
+		t.Fatalf("list: %v", err)
+	}
+	if len(msgs) != 3 {
+		t.Fatalf("len = %d, want 3", len(msgs))
+	}
+	// Newest first.
+	if msgs[0].CreatedAt != 3000 || msgs[1].CreatedAt != 2000 || msgs[2].CreatedAt != 1000 {
+		t.Fatalf("order = %+v, want [3000 2000 1000]", []int64{msgs[0].CreatedAt, msgs[1].CreatedAt, msgs[2].CreatedAt})
+	}
+	// Id/timestamp assigned when the caller omits them.
+	empty := Message{SessionID: MainSessionID, Role: RoleAssistant, Content: "ok"}
+	if err := st.Append(ctx, empty); err != nil {
+		t.Fatalf("append no ts: %v", err)
+	}
+	msgs, err = st.List(ctx, MainSessionID, 0, 50)
+	if err != nil {
+		t.Fatalf("list after append: %v", err)
+	}
+	if msgs[0].ID == "" || msgs[0].CreatedAt == 0 {
+		t.Fatalf("append did not fill id/created_at: %+v", msgs[0])
+	}
+}
+
+func TestListPaginationBefore(t *testing.T) {
+	t.Parallel()
+	st := newStore(t)
+	ctx := context.Background()
+
+	for _, ts := range []int64{1000, 2000, 3000, 4000, 5000, 6000} {
+		if err := st.Append(ctx, Message{SessionID: MainSessionID, Role: RoleUser, Content: "x", CreatedAt: ts}); err != nil {
+			t.Fatalf("append %d: %v", ts, err)
+		}
+	}
+
+	// First page: newest 2.
+	page, err := st.List(ctx, MainSessionID, 0, 2)
+	if err != nil {
+		t.Fatalf("first page: %v", err)
+	}
+	if len(page) != 2 || page[0].CreatedAt != 6000 || page[1].CreatedAt != 5000 {
+		t.Fatalf("first page = %+v, want [6000 5000]", page)
+	}
+	// Cursor on the page's oldest (5000) → older page [4000 3000].
+	next, err := st.List(ctx, MainSessionID, page[1].CreatedAt, 2)
+	if err != nil {
+		t.Fatalf("next page: %v", err)
+	}
+	if len(next) != 2 || next[0].CreatedAt != 4000 || next[1].CreatedAt != 3000 {
+		t.Fatalf("next page = %+v, want [4000 3000]", next)
+	}
+	// Past the start → empty.
+	tail, err := st.List(ctx, MainSessionID, 1000, 50)
+	if err != nil {
+		t.Fatalf("tail: %v", err)
+	}
+	if len(tail) != 0 {
+		t.Fatalf("tail = %+v, want empty", tail)
+	}
+}
+
+func TestRecentChronological(t *testing.T) {
+	t.Parallel()
+	st := newStore(t)
+	ctx := context.Background()
+
+	for _, ts := range []int64{1000, 2000, 3000, 4000, 5000} {
+		if err := st.Append(ctx, Message{SessionID: MainSessionID, Role: RoleUser, Content: "x", CreatedAt: ts}); err != nil {
+			t.Fatalf("append %d: %v", ts, err)
+		}
+	}
+	recent, err := st.Recent(ctx, MainSessionID, 3)
+	if err != nil {
+		t.Fatalf("recent: %v", err)
+	}
+	if len(recent) != 3 || recent[0].CreatedAt != 3000 || recent[1].CreatedAt != 4000 || recent[2].CreatedAt != 5000 {
+		t.Fatalf("recent = %+v, want [3000 4000 5000]", recent)
+	}
+}
+
+func TestNewSubSessionAndListSessions(t *testing.T) {
+	t.Parallel()
+	st := newStore(t)
+	ctx := context.Background()
+
+	// Main session exists by default.
+	sessions, err := st.ListSessions(ctx)
+	if err != nil {
+		t.Fatalf("list sessions initial: %v", err)
+	}
+	if len(sessions) != 1 || sessions[0].ID != MainSessionID || sessions[0].Type != SessionMain {
+		t.Fatalf("initial sessions = %+v, want [main web]", sessions)
+	}
+
+	id, err := st.NewSubSession(ctx, "owner")
+	if err != nil {
+		t.Fatalf("new sub: %v", err)
+	}
+	if !strings.HasPrefix(id, MainSessionID+":") {
+		t.Fatalf("sub id = %q, want prefix web:", id)
+	}
+
+	sessions, err = st.ListSessions(ctx)
+	if err != nil {
+		t.Fatalf("list sessions: %v", err)
+	}
+	if len(sessions) != 2 || sessions[0].ID != MainSessionID || sessions[0].Type != SessionMain {
+		t.Fatalf("sessions[0] = %+v, want main web first", sessions[0])
+	}
+	if sessions[1].ID != id || sessions[1].Type != SessionSub {
+		t.Fatalf("sessions[1] = %+v, want sub %s", sessions[1], id)
+	}
+}
+
+func TestDeleteSession(t *testing.T) {
+	t.Parallel()
+	st := newStore(t)
+	ctx := context.Background()
+
+	sub, err := st.NewSubSession(ctx, "owner")
+	if err != nil {
+		t.Fatalf("new sub: %v", err)
+	}
+	if err := st.Append(ctx, Message{SessionID: sub, Role: RoleUser, Content: "x", CreatedAt: 1000}); err != nil {
+		t.Fatalf("append: %v", err)
+	}
+
+	if err := st.DeleteSession(ctx, sub); err != nil {
+		t.Fatalf("delete sub: %v", err)
+	}
+	msgs, err := st.List(ctx, sub, 0, 50)
+	if err != nil {
+		t.Fatalf("list after delete: %v", err)
+	}
+	if len(msgs) != 0 {
+		t.Fatalf("messages after delete = %+v, want empty", msgs)
+	}
+	sessions, err := st.ListSessions(ctx)
+	if err != nil {
+		t.Fatalf("list sessions after delete: %v", err)
+	}
+	if len(sessions) != 1 {
+		t.Fatalf("sessions after delete = %+v, want [main]", sessions)
+	}
+
+	// Main session is immutable.
+	if err := st.DeleteSession(ctx, MainSessionID); !errors.Is(err, ErrImmutableMain) {
+		t.Fatalf("delete main err = %v, want ErrImmutableMain", err)
+	}
+}

internal/services/gateway/gateway_test.go

@@ -384,6 +384,33 @@ func TestWSUpgradeAndBroadcast(t *testing.T) {
 	}
 }
 
+func TestWsEventSessionIDField(t *testing.T) {
+	t.Parallel()
+	// When set, sessionId is marshalled so the client can filter by session.
+	set, err := json.Marshal(wsEvent{Type: "stream_end", SessionID: "web", Data: map[string]any{"stopReason": "stop"}})
+	if err != nil {
+		t.Fatalf("marshal set: %v", err)
+	}
+	if !strings.Contains(string(set), `"sessionId":"web"`) {
+		t.Fatalf("set frame = %s, want sessionId web", set)
+	}
+	// When empty, omitempty drops it so existing clients/tests see no change.
+	unset, err := json.Marshal(wsEvent{Type: "stream_end", Data: map[string]any{"stopReason": "stop"}})
+	if err != nil {
+		t.Fatalf("marshal unset: %v", err)
+	}
+	if strings.Contains(string(unset), "sessionId") {
+		t.Fatalf("unset frame = %s, want sessionId omitted", unset)
+	}
+	var ev wsEvent
+	if err := json.Unmarshal(unset, &ev); err != nil {
+		t.Fatalf("unmarshal: %v", err)
+	}
+	if ev.SessionID != "" {
+		t.Fatalf("decoded SessionID = %q, want empty", ev.SessionID)
+	}
+}
+
 // readWS reads one text frame and decodes it into a wsEvent.
 func readWS(t *testing.T, conn *websocket.Conn) wsEvent {
 	t.Helper()

internal/services/gateway/history.go

@@ -0,0 +1,165 @@
+package gateway
+
+import (
+	"errors"
+	"net/http"
+	"strconv"
+
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/auth"
+	apperr "git.yeij.top/AskaEth/Cyrene_v2/internal/errors"
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/history"
+)
+
+// historySessionSummary is the wire shape of one persistent session (main or sub) in the
+// GET /api/v1/sessions response.
+type historySessionSummary struct {
+	ID        string `json:"id"`
+	Type      string `json:"type"` // "main" | "sub"
+	Title     string `json:"title,omitempty"`
+	CreatedAt int64  `json:"createdAt"`
+}
+
+// historyMessage is the wire shape of one persisted chat message in the
+// GET /api/v1/sessions/{id}/messages response. Mode is carried so the frontend renders
+// the bubble without guessing (design/16 §5.1); thinking is stored and returned to the
+// owner/admin (the messages route is owner/admin-only, so it never reaches a viewer).
+type historyMessage struct {
+	ID        string `json:"id"`
+	SessionID string `json:"sessionId"`
+	Role      string `json:"role"`
+	Content   string `json:"content"`
+	Mode      string `json:"mode,omitempty"`
+	Mood      string `json:"mood,omitempty"`
+	Image     string `json:"image,omitempty"`
+	Thinking  string `json:"thinking,omitempty"`
+	CreatedAt int64  `json:"createdAt"`
+}
+
+// handleHistorySessions returns the main session first, then every sub session (oldest
+// first). It is the chat-history session list for the frontend (design/16 §2.2).
+func (s *Server) handleHistorySessions(w http.ResponseWriter, r *http.Request) {
+	meta, err := s.history.ListSessions(r.Context())
+	if err != nil {
+		s.log.Error("history list sessions", "err", err)
+		writeError(w, apperr.Wrap(apperr.ErrInternal, "sessions_failed", "failed to list sessions"))
+		return
+	}
+	out := make([]historySessionSummary, 0, len(meta))
+	for _, m := range meta {
+		out = append(out, historySessionSummary{
+			ID:        m.ID,
+			Type:      m.Type,
+			Title:     m.Title,
+			CreatedAt: m.CreatedAt,
+		})
+	}
+	writeJSON(w, http.StatusOK, out)
+}
+
+// handleCreateSubSession creates a new sub session and returns its sessionId. It is a
+// user-side write, so it is owner/admin-gated like /api/v1/chat.
+func (s *Server) handleCreateSubSession(w http.ResponseWriter, r *http.Request) {
+	if s.history == nil {
+		writeError(w, &apperr.AppError{Code: "not_implemented", Message: "会话历史服务未配置（未注入 HistoryStore）", HTTPCode: http.StatusNotImplemented})
+		return
+	}
+	ownerID := ""
+	if claims, ok := auth.ClaimsFromContext(r.Context()); ok {
+		ownerID = claims.Sub
+	}
+	id, err := s.history.NewSubSession(r.Context(), ownerID)
+	if err != nil {
+		s.log.Error("history create sub session", "err", err)
+		writeError(w, apperr.Wrap(apperr.ErrInternal, "session_failed", "failed to create session"))
+		return
+	}
+	writeJSON(w, http.StatusCreated, map[string]any{"sessionId": id})
+}
+
+// handleListMessages returns a page of messages for a session, newest first, with cursor
+// pagination. It is the hot-paging history endpoint (design/16 §2.2, §5.1): limit defaults
+// to 50 (clamped to 200), before is the oldest message's created_at from the previous page.
+func (s *Server) handleListMessages(w http.ResponseWriter, r *http.Request) {
+	if s.history == nil {
+		writeError(w, &apperr.AppError{Code: "not_implemented", Message: "会话历史服务未配置（未注入 HistoryStore）", HTTPCode: http.StatusNotImplemented})
+		return
+	}
+	id := r.PathValue("id")
+	if id == "" {
+		writeError(w, apperr.Wrap(apperr.ErrBadRequest, "bad_request", "session id is required"))
+		return
+	}
+	limitStr := r.URL.Query().Get("limit")
+	limit := historyDefaultLimit
+	if limitStr != "" {
+		v, err := strconv.Atoi(limitStr)
+		if err != nil || v < 1 {
+			writeError(w, apperr.Wrap(apperr.ErrBadRequest, "bad_request", "invalid limit"))
+			return
+		}
+		limit = v
+	}
+	var before int64
+	if b := r.URL.Query().Get("before"); b != "" {
+		v, err := strconv.ParseInt(b, 10, 64)
+		if err != nil {
+			writeError(w, apperr.Wrap(apperr.ErrBadRequest, "bad_request", "invalid before"))
+			return
+		}
+		before = v
+	}
+
+	msgs, err := s.history.List(r.Context(), id, before, limit)
+	if err != nil {
+		s.log.Error("history list messages", "err", err, "session", id)
+		writeError(w, apperr.Wrap(apperr.ErrInternal, "messages_failed", "failed to list messages"))
+		return
+	}
+	out := make([]historyMessage, 0, len(msgs))
+	for _, m := range msgs {
+		out = append(out, toHistoryMessage(m))
+	}
+	writeJSON(w, http.StatusOK, out)
+}
+
+// handleDeleteSession deletes a sub session and all its messages. The main session is
+// immutable (design/16 §6) and returns 403.
+func (s *Server) handleDeleteSession(w http.ResponseWriter, r *http.Request) {
+	if s.history == nil {
+		writeError(w, &apperr.AppError{Code: "not_implemented", Message: "会话历史服务未配置（未注入 HistoryStore）", HTTPCode: http.StatusNotImplemented})
+		return
+	}
+	id := r.PathValue("id")
+	if id == "" {
+		writeError(w, apperr.Wrap(apperr.ErrBadRequest, "bad_request", "session id is required"))
+		return
+	}
+	if err := s.history.DeleteSession(r.Context(), id); err != nil {
+		if errors.Is(err, history.ErrImmutableMain) {
+			writeError(w, apperr.Wrap(apperr.ErrForbidden, "forbidden", "the main session cannot be deleted"))
+			return
+		}
+		s.log.Error("history delete session", "err", err, "session", id)
+		writeError(w, apperr.Wrap(apperr.ErrInternal, "session_failed", "failed to delete session"))
+		return
+	}
+	writeJSON(w, http.StatusOK, map[string]any{"deleted": id})
+}
+
+// toHistoryMessage maps a stored history.Message to its wire shape.
+func toHistoryMessage(m history.Message) historyMessage {
+	return historyMessage{
+		ID:        m.ID,
+		SessionID: m.SessionID,
+		Role:      m.Role,
+		Content:   m.Content,
+		Mode:      m.Mode,
+		Mood:      m.Mood,
+		Image:     m.Image,
+		Thinking:  m.Thinking,
+		CreatedAt: m.CreatedAt,
+	}
+}
+
+// historyDefaultLimit is the page size used when the client omits the limit query param.
+const historyDefaultLimit = 50

internal/services/gateway/history_handler_test.go

@@ -0,0 +1,228 @@
+package gateway
+
+import (
+	"context"
+	"encoding/json"
+	"io"
+	"log/slog"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/auth"
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/db"
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/history"
+)
+
+// testLogger returns a logger that writes nowhere, keeping test output clean.
+func testLogger() *slog.Logger {
+	return slog.New(slog.NewTextHandler(io.Discard, nil))
+}
+
+// newHistoryEnv builds a Server wired with a real in-memory history store, returning the
+// httptest server, the JWT service and the underlying history store.
+func newHistoryEnv(t *testing.T) (*Server, *httptest.Server, *auth.JWTService, *history.Store) {
+	t.Helper()
+	ctx := context.Background()
+	conn, err := db.Open(ctx, ":memory:", testLogger())
+	if err != nil {
+		t.Fatalf("open memory db: %v", err)
+	}
+	t.Cleanup(func() { _ = conn.Close() })
+	st, err := history.NewStore(ctx, conn, testLogger())
+	if err != nil {
+		t.Fatalf("new history store: %v", err)
+	}
+	srv, ts, jwt := buildEnv(t, cfgDefault(), Deps{History: st})
+	return srv, ts, jwt, st
+}
+
+// decodeDataSlice decodes the {data:[...]} envelope into the given slice target.
+func decodeDataSlice(t *testing.T, body []byte, out any) {
+	t.Helper()
+	if err := json.Unmarshal(decodeBody(t, body)["data"], out); err != nil {
+		t.Fatalf("decode data: %v (body=%s)", err, body)
+	}
+}
+
+func TestHistoryCreateAndListSessions(t *testing.T) {
+	t.Parallel()
+	_, ts, jwt, _ := newHistoryEnv(t)
+	tok := signAny(t, jwt, "acct", auth.RoleOwner)
+
+	// Fresh store always lists the main session.
+	resp, body := doReq(t, http.MethodGet, ts.URL+"/api/v1/sessions", nil, tok)
+	if resp.StatusCode != http.StatusOK {
+		t.Fatalf("list status = %d, want 200 (body=%s)", resp.StatusCode, body)
+	}
+	var initial []historySessionSummary
+	decodeDataSlice(t, body, &initial)
+	if len(initial) != 1 || initial[0].ID != history.MainSessionID || initial[0].Type != history.SessionMain {
+		t.Fatalf("initial sessions = %+v, want [main web]", initial)
+	}
+
+	// Create a sub session.
+	resp, body = doReq(t, http.MethodPost, ts.URL+"/api/v1/sessions", nil, tok)
+	if resp.StatusCode != http.StatusCreated {
+		t.Fatalf("create status = %d, want 201 (body=%s)", resp.StatusCode, body)
+	}
+	var created struct {
+		SessionID string `json:"sessionId"`
+	}
+	if err := json.Unmarshal(decodeBody(t, body)["data"], &created); err != nil {
+		t.Fatalf("decode create data: %v (body=%s)", err, body)
+	}
+	if !strings.HasPrefix(created.SessionID, history.MainSessionID+":") {
+		t.Fatalf("sessionId = %q, want prefix web:", created.SessionID)
+	}
+
+	// Now main first, then the sub.
+	resp, body = doReq(t, http.MethodGet, ts.URL+"/api/v1/sessions", nil, tok)
+	if resp.StatusCode != http.StatusOK {
+		t.Fatalf("list status = %d, want 200 (body=%s)", resp.StatusCode, body)
+	}
+	var sessions []historySessionSummary
+	decodeDataSlice(t, body, &sessions)
+	if len(sessions) != 2 || sessions[0].ID != history.MainSessionID || sessions[0].Type != history.SessionMain {
+		t.Fatalf("sessions[0] = %+v, want main web first", sessions[0])
+	}
+	if sessions[1].ID != created.SessionID || sessions[1].Type != history.SessionSub {
+		t.Fatalf("sessions[1] = %+v, want sub %s", sessions[1], created.SessionID)
+	}
+}
+
+func TestHistoryListMessagesPagination(t *testing.T) {
+	t.Parallel()
+	_, ts, jwt, st := newHistoryEnv(t)
+	ctx := context.Background()
+	// Seed the main session with messages at ascending ms timestamps.
+	for _, ts := range []int64{1000, 2000, 3000, 4000} {
+		if err := st.Append(ctx, history.Message{SessionID: history.MainSessionID, Role: history.RoleUser, Content: "x", CreatedAt: ts}); err != nil {
+			t.Fatalf("seed %d: %v", ts, err)
+		}
+	}
+	tok := signAny(t, jwt, "acct", auth.RoleOwner)
+
+	// Newest page of 2.
+	resp, body := doReq(t, http.MethodGet, ts.URL+"/api/v1/sessions/"+history.MainSessionID+"/messages?limit=2", nil, tok)
+	if resp.StatusCode != http.StatusOK {
+		t.Fatalf("first page status = %d, want 200 (body=%s)", resp.StatusCode, body)
+	}
+	var page []historyMessage
+	decodeDataSlice(t, body, &page)
+	if len(page) != 2 || page[0].CreatedAt != 4000 || page[1].CreatedAt != 3000 {
+		t.Fatalf("first page = %+v, want [4000 3000]", page)
+	}
+
+	// Cursor before the page's oldest (3000) → older page [2000 1000].
+	resp, body = doReq(t, http.MethodGet, ts.URL+"/api/v1/sessions/"+history.MainSessionID+"/messages?before=3000&limit=2", nil, tok)
+	if resp.StatusCode != http.StatusOK {
+		t.Fatalf("older page status = %d, want 200 (body=%s)", resp.StatusCode, body)
+	}
+	var older []historyMessage
+	decodeDataSlice(t, body, &older)
+	if len(older) != 2 || older[0].CreatedAt != 2000 || older[1].CreatedAt != 1000 {
+		t.Fatalf("older page = %+v, want [2000 1000]", older)
+	}
+
+	// Invalid limit/before → 400.
+	resp, body = doReq(t, http.MethodGet, ts.URL+"/api/v1/sessions/"+history.MainSessionID+"/messages?limit=abc", nil, tok)
+	if resp.StatusCode != http.StatusBadRequest {
+		t.Fatalf("invalid limit status = %d, want 400 (body=%s)", resp.StatusCode, body)
+	}
+}
+
+func TestHistoryDeleteSubSession(t *testing.T) {
+	t.Parallel()
+	_, ts, jwt, st := newHistoryEnv(t)
+	ctx := context.Background()
+	tok := signAny(t, jwt, "acct", auth.RoleOwner)
+
+	// Create a sub and seed a message.
+	resp, body := doReq(t, http.MethodPost, ts.URL+"/api/v1/sessions", nil, tok)
+	if resp.StatusCode != http.StatusCreated {
+		t.Fatalf("create status = %d, want 201 (body=%s)", resp.StatusCode, body)
+	}
+	var created struct {
+		SessionID string `json:"sessionId"`
+	}
+	if err := json.Unmarshal(decodeBody(t, body)["data"], &created); err != nil {
+		t.Fatalf("decode create data: %v", err)
+	}
+	if err := st.Append(ctx, history.Message{SessionID: created.SessionID, Role: history.RoleUser, Content: "x", CreatedAt: 1000}); err != nil {
+		t.Fatalf("seed sub: %v", err)
+	}
+
+	// Delete it.
+	resp, _ = doReq(t, http.MethodDelete, ts.URL+"/api/v1/sessions/"+created.SessionID, nil, tok)
+	if resp.StatusCode != http.StatusOK {
+		t.Fatalf("delete status = %d, want 200", resp.StatusCode)
+	}
+	// It is gone from the list.
+	resp, body = doReq(t, http.MethodGet, ts.URL+"/api/v1/sessions", nil, tok)
+	if resp.StatusCode != http.StatusOK {
+		t.Fatalf("list status = %d, want 200", resp.StatusCode)
+	}
+	var sessions []historySessionSummary
+	decodeDataSlice(t, body, &sessions)
+	if len(sessions) != 1 || sessions[0].ID != history.MainSessionID {
+		t.Fatalf("sessions after delete = %+v, want [main]", sessions)
+	}
+
+	// Main session is immutable → 403.
+	resp, body = doReq(t, http.MethodDelete, ts.URL+"/api/v1/sessions/"+history.MainSessionID, nil, tok)
+	if resp.StatusCode != http.StatusForbidden {
+		t.Fatalf("delete main status = %d, want 403 (body=%s)", resp.StatusCode, body)
+	}
+	wantErrorShape(t, body)
+}
+
+func TestHistoryNotWired(t *testing.T) {
+	t.Parallel()
+	_, ts, jwt := buildEnv(t, cfgDefault(), Deps{})
+	tok := signAny(t, jwt, "acct", auth.RoleOwner)
+
+	// Messages/create/delete answer a clear 501 when no history store is wired.
+	for _, tc := range []struct {
+		method, url string
+		want        int
+	}{
+		{http.MethodGet, "/api/v1/sessions/web/messages?limit=2", http.StatusNotImplemented},
+		{http.MethodPost, "/api/v1/sessions", http.StatusNotImplemented},
+	} {
+		resp, body := doReq(t, tc.method, ts.URL+tc.url, nil, tok)
+		if resp.StatusCode != tc.want {
+			t.Fatalf("%s %s status = %d, want %d (body=%s)", tc.method, tc.url, resp.StatusCode, tc.want, body)
+		}
+		if !strings.Contains(string(body), "not_implemented") {
+			t.Fatalf("%s %s body = %s, want not_implemented code", tc.method, tc.url, body)
+		}
+	}
+}
+
+func TestHistoryRBAC(t *testing.T) {
+	t.Parallel()
+	_, ts, jwt, _ := newHistoryEnv(t)
+	viewer := signAny(t, jwt, "acct", auth.RoleViewer)
+	owner := signAny(t, jwt, "acct", auth.RoleOwner)
+
+	// Viewer can list session metadata (readAuth) but not read/write history content.
+	resp, _ := doReq(t, http.MethodGet, ts.URL+"/api/v1/sessions", nil, viewer)
+	if resp.StatusCode != http.StatusOK {
+		t.Fatalf("viewer list sessions status = %d, want 200", resp.StatusCode)
+	}
+	resp, body := doReq(t, http.MethodGet, ts.URL+"/api/v1/sessions/web/messages", nil, viewer)
+	if resp.StatusCode != http.StatusForbidden {
+		t.Fatalf("viewer messages status = %d, want 403 (body=%s)", resp.StatusCode, body)
+	}
+	resp, body = doReq(t, http.MethodPost, ts.URL+"/api/v1/sessions", nil, viewer)
+	if resp.StatusCode != http.StatusForbidden {
+		t.Fatalf("viewer create status = %d, want 403 (body=%s)", resp.StatusCode, body)
+	}
+	// Owner can read messages.
+	resp, _ = doReq(t, http.MethodGet, ts.URL+"/api/v1/sessions/web/messages?limit=2", nil, owner)
+	if resp.StatusCode != http.StatusOK {
+		t.Fatalf("owner messages status = %d, want 200", resp.StatusCode)
+	}
+}

internal/services/gateway/rest.go

@@ -175,8 +175,15 @@ func (s *Server) handleChat(w http.ResponseWriter, r *http.Request) {
 	writeJSON(w, http.StatusOK, chatResponse{Reply: res.Text, Steps: res.Steps, StopReason: res.StopReason})
 }
 
-// handleListSessions returns the known sessions (or [] when no store is wired).
+// handleListSessions returns the persistent chat-history sessions (main + sub) when a
+// history store is wired, otherwise falls back to the channel-session view. The fallback
+// keeps the pre-design/16 contract: a server built without history still lists the known
+// channel sessions (or [] when no store is wired).
 func (s *Server) handleListSessions(w http.ResponseWriter, r *http.Request) {
+	if s.history != nil {
+		s.handleHistorySessions(w, r)
+		return
+	}
 	out := make([]SessionSummary, 0)
 	if s.sessions != nil {
 		list, err := s.sessions.List(r.Context())

internal/services/gateway/server.go

@@ -37,6 +37,7 @@ import (
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/logger"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/persona/loader"
 	personareg "git.yeij.top/AskaEth/Cyrene_v2/internal/persona/registry"
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/history"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/knowledge"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/loop"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/memory"
@@ -102,6 +103,19 @@ type MemoryStore interface {
 	List(ctx context.Context) ([]memory.MemoryItem, error)
 }
 
+// HistoryStore is the persistent chat-history store (design/16). *runtime/history.Store
+// is the production implementation. When nil the session/message endpoints degrade: list
+// sessions falls back to the channel-session view, and the create/messages/delete
+// endpoints answer with a clear 501.
+type HistoryStore interface {
+	Append(ctx context.Context, msg history.Message) error
+	List(ctx context.Context, sessionID string, before int64, limit int) ([]history.Message, error)
+	Recent(ctx context.Context, sessionID string, n int) ([]history.Message, error)
+	DeleteSession(ctx context.Context, sessionID string) error
+	NewSubSession(ctx context.Context, ownerID string) (string, error)
+	ListSessions(ctx context.Context) ([]history.SessionMeta, error)
+}
+
 // IoTStore lists IoT devices. *iot.Manager is the production implementation.
 type IoTStore interface {
 	List(ctx context.Context) ([]iot.Device, error)
@@ -179,6 +193,9 @@ type Deps struct {
 	Personas PersonaStore
 	// Memories lists memory entries. Optional.
 	Memories MemoryStore
+	// History is the persistent chat-history store (design/16). Optional; when nil the
+	// session/message endpoints degrade (list falls back to channel sessions, others 501).
+	History HistoryStore
 	// IoT lists IoT devices. Optional.
 	IoT IoTStore
 	// Scheduler audits sub-agent context for the subagent_ask WS route.
@@ -229,6 +246,7 @@ type Server struct {
 	sessions    SessionStore
 	personas    PersonaStore
 	memories    MemoryStore
+	history     HistoryStore
 	iot         IoTStore
 	auditor     SubagentAuditor
 	taskOutputs TaskOutputStore
@@ -270,6 +288,7 @@ func New(cfg *config.Config, deps Deps) (*Server, error) {
 		sessions:    deps.Session,
 		personas:    deps.Personas,
 		memories:    deps.Memories,
+		history:     deps.History,
 		iot:         deps.IoT,
 		auditor:     deps.Scheduler,
 		taskOutputs: deps.Tasks,
@@ -305,6 +324,15 @@ func (s *Server) Handler() http.Handler {
 	mux.Handle("GET /api/v1/sessions", readAuth(http.HandlerFunc(s.handleListSessions)))
 	mux.Handle("GET /api/v1/personas", readAuth(http.HandlerFunc(s.handleListPersonas)))
 
+	// Session/model reads and writes are the owner's private history; owner/admin only
+	// (chat uses the same roles). Session metadata in GET /sessions stays readAuth for
+	// any authenticated role, but the message content (and thinking) is owner/admin-only
+	// per design/16 §6. Creating and deleting a sub session mutate persistent state.
+	historyAuth := s.auth.Authorize(auth.RoleOwner, auth.RoleAdmin)
+	mux.Handle("GET /api/v1/sessions/{id}/messages", historyAuth(http.HandlerFunc(s.handleListMessages)))
+	mux.Handle("POST /api/v1/sessions", historyAuth(http.HandlerFunc(s.handleCreateSubSession)))
+	mux.Handle("DELETE /api/v1/sessions/{id}", historyAuth(http.HandlerFunc(s.handleDeleteSession)))
+
 	// Realtime.
 	mux.HandleFunc("/ws/chat", s.handleWS)
 
@@ -425,9 +453,10 @@ func newOrchestrator(s *Server) *Orchestrator {
 			return s.hub.runTurn(ctx, sessionID, messages, env)
 		},
 		Emit: func(sessionID, typ string, data map[string]any) {
-			s.hub.broadcast(wsEvent{Type: typ, Data: data})
+			s.hub.broadcast(wsEvent{Type: typ, SessionID: sessionID, Data: data})
 		},
-		Log: s.log,
+		Log:     s.log,
+		History: s.history,
 	})
 }
 

internal/services/gateway/session_orchestrator.go

@@ -104,6 +104,10 @@ type OrchestratorOptions struct {
 	Log        *slog.Logger
 	MaxPending int
 	MaxTail    int
+	// History is the optional persistent chat-history store (design/16 §3.3).
+	// When nil every Append is skipped, so the orchestrator runs without storage
+	// (tests / a gateway built without a history store).
+	History HistoryStore
 }
 
 // sessionState is the per-session message state (design/15 §2). Every mutation
@@ -135,6 +139,7 @@ type Orchestrator struct {
 	run        TurnRunner
 	emit       StateSink
 	log        *slog.Logger
+	history    HistoryStore
 	maxPending int
 	maxTail    int
 }
@@ -160,6 +165,7 @@ func NewOrchestrator(opts OrchestratorOptions) *Orchestrator {
 		run:        opts.Run,
 		emit:       opts.Emit,
 		log:        opts.Log,
+		history:    opts.History,
 		maxPending: maxPending,
 		maxTail:    maxTail,
 	}
@@ -212,6 +218,7 @@ func (o *Orchestrator) submitInternal(ctx context.Context, sessionID string, mes
 		depth := len(st.pending)
 		mode := st.mode
 		o.mu.Unlock()
+		o.persistUserMessages(ctx, sessionID, messages)
 		o.emitState(sessionID, EventPendingAck, map[string]any{"sessionId": sessionID, "mode": mode, "queueDepth": depth})
 		return SubmitResult{Accepted: true, Pending: true, QueueDepth: depth}, nil
 	}
@@ -225,6 +232,7 @@ func (o *Orchestrator) submitInternal(ctx context.Context, sessionID string, mes
 	o.mu.Unlock()
 
 	o.emitState(sessionID, EventSessionBusy, map[string]any{"sessionId": sessionID, "mode": ModeThinking})
+	o.persistUserMessages(ctx, sessionID, messages)
 
 	if sync {
 		res, err := o.run(runCtx, sessionID, messages, env)
@@ -257,6 +265,7 @@ func (o *Orchestrator) goTurn(sessionID string, gen uint64, messages []model.Mes
 // (design/15 §2.2: incremental, not full replay). A stale generation (an
 // interrupt or a newer top-level turn superseded this one) is a no-op.
 func (o *Orchestrator) completeTurn(sessionID string, gen uint64, res *loop.Result, runErr error) error {
+	o.persistAssistantReply(sessionID, res)
 	o.mu.Lock()
 	st := o.states[sessionID]
 	if st == nil || st.gen != gen {

internal/services/gateway/session_orchestrator_test.go

@@ -8,6 +8,7 @@ import (
 	"testing"
 	"time"
 
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/history"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/loop"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/model"
 )
@@ -149,6 +150,89 @@ func readOnlyEnv() loop.SessionEnv {
 	return loop.SessionEnv{MemWritable: false}
 }
 
+// fakeHistory records every Append so tests assert the orchestrator persists the
+// user stream and the assistant reply into the chat-history store.
+type fakeHistory struct {
+	mu      sync.Mutex
+	appends []history.Message
+}
+
+func (f *fakeHistory) Append(_ context.Context, msg history.Message) error {
+	f.mu.Lock()
+	f.appends = append(f.appends, msg)
+	f.mu.Unlock()
+	return nil
+}
+
+func (f *fakeHistory) List(context.Context, string, int64, int) ([]history.Message, error) {
+	return nil, nil
+}
+
+func (f *fakeHistory) Recent(context.Context, string, int) ([]history.Message, error) {
+	return nil, nil
+}
+
+func (f *fakeHistory) DeleteSession(context.Context, string) error {
+	return nil
+}
+
+func (f *fakeHistory) NewSubSession(context.Context, string) (string, error) {
+	return "", nil
+}
+
+func (f *fakeHistory) ListSessions(context.Context) ([]history.SessionMeta, error) {
+	return nil, nil
+}
+
+func (f *fakeHistory) records() []history.Message {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	return append([]history.Message(nil), f.appends...)
+}
+
+func TestSubmitPersistsUserAndAssistant(t *testing.T) {
+	t.Parallel()
+	run := newGatedRunner(1)
+	hist := &fakeHistory{}
+	o := NewOrchestrator(OrchestratorOptions{
+		Run:     run.Run,
+		Log:     slog.Default(),
+		History: hist,
+	})
+	run.tickets <- struct{}{}
+	sr, res := o.SubmitSync(context.Background(), "web", []model.Message{userMsg("你好")}, loop.OwnerEnv())
+	if !sr.Accepted || !sr.Immediate {
+		t.Fatalf("submit = %+v, want Accepted+Immediate", sr)
+	}
+	if res == nil {
+		t.Fatal("sync result is nil")
+	}
+	recs := hist.records()
+	if len(recs) != 2 {
+		t.Fatalf("appends = %d, want 2 (user + assistant)", len(recs))
+	}
+	if recs[0].SessionID != "web" || recs[0].Role != history.RoleUser || recs[0].Content != "你好" {
+		t.Fatalf("user record = %+v", recs[0])
+	}
+	if recs[1].SessionID != "web" || recs[1].Role != history.RoleAssistant || recs[1].Content == "" {
+		t.Fatalf("assistant record = %+v", recs[1])
+	}
+}
+
+func TestSubmitDoesNotPersistRejectedTurn(t *testing.T) {
+	t.Parallel()
+	hist := &fakeHistory{}
+	o := NewOrchestrator(OrchestratorOptions{Run: newGatedRunner(0).Run, Log: slog.Default(), History: hist, MaxPending: 1})
+	// Empty turn is rejected before any persistence.
+	sr := o.Submit(context.Background(), "web", nil, loop.OwnerEnv())
+	if sr.Accepted || !sr.Empty {
+		t.Fatalf("empty submit = %+v, want Empty reject", sr)
+	}
+	if n := len(hist.records()); n != 0 {
+		t.Fatalf("appends = %d, want 0 for a rejected turn", n)
+	}
+}
+
 func TestSubmitIdleStartsImmediate(t *testing.T) {
 	t.Parallel()
 	run := newGatedRunner(1)

internal/services/gateway/session_persist.go

@@ -0,0 +1,62 @@
+package gateway
+
+import (
+	"context"
+
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/history"
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/loop"
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/model"
+)
+
+// This file owns the orchestrator's chat-history persistence hooks (design/16
+// §3.3): the user stream is written at submit time and the assistant reply when a
+// turn settles. Both are silent no-ops when no history store is wired, so the
+// orchestrator runs unchanged without a store (tests / an unwired gateway).
+
+// persistUserMessages writes each RoleUser message of a submitted turn to the
+// history store so the raw user stream is durable before the turn resolves. It is
+// a no-op when no history store is wired. Rejected turns (empty, pending-full)
+// never reach it, so nothing vacuous is persisted.
+func (o *Orchestrator) persistUserMessages(ctx context.Context, sessionID string, messages []model.Message) {
+	if o.history == nil {
+		return
+	}
+	for _, m := range messages {
+		if m.Role != model.RoleUser {
+			continue
+		}
+		text, image := messageTextImage(m)
+		if text == "" && image == "" {
+			continue
+		}
+		if err := o.history.Append(ctx, history.Message{SessionID: sessionID, Role: history.RoleUser, Content: text, Image: image}); err != nil {
+			o.log.Error("orchestrator persist user", "err", err, "session", sessionID)
+		}
+	}
+}
+
+// persistAssistantReply writes the final assistant text of a settled turn to the
+// history store. It is a no-op when no history store is wired or the turn
+// produced no text (interrupted/errored turns leave res nil). The write runs
+// under a fresh background context so a canceled turn context never aborts it.
+func (o *Orchestrator) persistAssistantReply(sessionID string, res *loop.Result) {
+	if o.history == nil || res == nil || res.Text == "" {
+		return
+	}
+	if err := o.history.Append(context.Background(), history.Message{SessionID: sessionID, Role: history.RoleAssistant, Content: res.Text}); err != nil {
+		o.log.Error("orchestrator persist assistant", "err", err, "session", sessionID)
+	}
+}
+
+// messageTextImage returns the concatenated text and the first image URL of a
+// message's content parts, for persistence into the chat_messages columns.
+func messageTextImage(m model.Message) (text, image string) {
+	for _, c := range m.Content {
+		if c.Type == model.ContentTypeText {
+			text += c.Text
+		} else if c.Type == model.ContentTypeImage && image == "" {
+			image = c.ImageURL
+		}
+	}
+	return text, image
+}

internal/services/gateway/ws_history.go

@@ -0,0 +1,34 @@
+package gateway
+
+import (
+	"context"
+)
+
+// wsReplayLimit is the number of most-recent messages replayed to a connecting
+// client. It is intentionally small so a reconnect is cheap; the frontend pulls
+// the full active-session history via REST (design/16 §4.1) and the replay is
+// only there to give a brand-new connection immediate context.
+const wsReplayLimit = 50
+
+// replayHistory sends the most recent messages for sessionID to a single,
+// freshly-registered connection as a ws "history" event. It is best-effort: a
+// nil store or a read error skips the replay silently. Thinking is NOT included
+// (design/16 §6) so the frame never leaks chain-of-thought to a connection whose
+// role the gateway did not restrict on upgrade.
+func (h *hub) replayHistory(c *wsConn, sessionID string) {
+	if h.history == nil {
+		return
+	}
+	msgs, err := h.history.Recent(context.Background(), sessionID, wsReplayLimit)
+	if err != nil {
+		h.log.Warn("history replay", "err", err, "session", sessionID)
+		return
+	}
+	out := make([]historyMessage, 0, len(msgs))
+	for _, m := range msgs {
+		wire := toHistoryMessage(m)
+		wire.Thinking = ""
+		out = append(out, wire)
+	}
+	c.enqueueJSON(wsEvent{Type: "history", SessionID: sessionID, Data: map[string]any{"messages": out}})
+}

internal/services/gateway/ws_hub.go

@@ -17,6 +17,7 @@ import (
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/auth"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/bus"
 	apperr "git.yeij.top/AskaEth/Cyrene_v2/internal/errors"
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/history"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/loop"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/model"
 )
@@ -41,10 +42,11 @@ const (
 // used by the streaming path so a client can correlate a burst of streamed
 // events to one assistant turn and keep them in order.
 type wsEvent struct {
-	Type   string `json:"type"`
-	TurnID string `json:"turnId,omitempty"`
-	Seq    int    `json:"seq,omitempty"`
-	Data   any    `json:"data,omitempty"`
+	Type      string `json:"type"`
+	TurnID    string `json:"turnId,omitempty"`
+	Seq       int    `json:"seq,omitempty"`
+	SessionID string `json:"sessionId,omitempty"`
+	Data      any    `json:"data,omitempty"`
 }
 
 // wsInbound is the client→server frame. Data carries the per-type payload.
@@ -68,15 +70,16 @@ type wsConn struct {
 
 // hub owns the connection set and relays bus events to every client.
 type hub struct {
-	mu          sync.RWMutex
-	conns       map[*wsConn]struct{}
-	bus         bus.Bus
-	runner      Chatter
-	auditor     SubagentAuditor
-	taskOutputs TaskOutputStore
+	mu           sync.RWMutex
+	conns        map[*wsConn]struct{}
+	bus          bus.Bus
+	runner       Chatter
+	auditor      SubagentAuditor
+	taskOutputs  TaskOutputStore
 	voice        VoiceService
 	streaming    bool
 	orchestrator *Orchestrator
+	history      HistoryStore
 	log          *slog.Logger
 	upgrader     websocket.Upgrader
 	origins      map[string]bool
@@ -96,6 +99,7 @@ func newHub(s *Server) *hub {
 		taskOutputs: s.taskOutputs,
 		voice:       s.voice,
 		streaming:   s.streaming,
+		history:     s.history,
 		log:         s.log,
 		origins:     origins,
 		allowAll:    allowAll,
@@ -222,7 +226,7 @@ func (h *hub) streamChat(ctx context.Context, payload any) {
 		_ = json.Unmarshal(b, &req)
 	}
 	if h.runner == nil || (strings.TrimSpace(req.Text) == "" && strings.TrimSpace(req.Image) == "") {
-		h.broadcast(wsEvent{Type: "stream_end", TurnID: newTurnID(), Seq: int(h.streamSeq.Add(1)), Data: map[string]any{"stopReason": "idle"}})
+		h.broadcast(wsEvent{Type: "stream_end", TurnID: newTurnID(), Seq: int(h.streamSeq.Add(1)), SessionID: req.SessionID, Data: map[string]any{"stopReason": "idle"}})
 		return
 	}
 	if req.SessionID == "" {
@@ -248,19 +252,19 @@ func (h *hub) streamChat(ctx context.Context, payload any) {
 func (h *hub) runTurn(ctx context.Context, sessionID string, msgs []model.Message, env loop.SessionEnv) (*loop.Result, error) {
 	turnID := newTurnID()
 	if h.runner == nil || len(msgs) == 0 {
-		h.broadcast(wsEvent{Type: "stream_end", TurnID: turnID, Seq: int(h.streamSeq.Add(1)), Data: map[string]any{"stopReason": "idle"}})
+		h.broadcast(wsEvent{Type: "stream_end", TurnID: turnID, Seq: int(h.streamSeq.Add(1)), SessionID: sessionID, Data: map[string]any{"stopReason": "idle"}})
 		return &loop.Result{StopReason: "idle"}, nil
 	}
 	res, err := h.runner.Run(loop.WithSessionEnv(ctx, env), sessionID, msgs)
 	if err != nil {
 		h.log.Error("chat run", "err", err, "session", sessionID)
-		h.broadcast(wsEvent{Type: "stream_end", TurnID: turnID, Seq: int(h.streamSeq.Add(1)), Data: map[string]any{"stopReason": "error", "error": err.Error()}})
+		h.broadcast(wsEvent{Type: "stream_end", TurnID: turnID, Seq: int(h.streamSeq.Add(1)), SessionID: sessionID, Data: map[string]any{"stopReason": "error", "error": err.Error()}})
 		return nil, err
 	}
 	if !h.streaming {
-		h.broadcast(wsEvent{Type: "stream_chunk", TurnID: turnID, Seq: int(h.streamSeq.Add(1)), Data: map[string]any{"delta": res.Text}})
+		h.broadcast(wsEvent{Type: "stream_chunk", TurnID: turnID, Seq: int(h.streamSeq.Add(1)), SessionID: sessionID, Data: map[string]any{"delta": res.Text}})
 	}
-	h.broadcast(wsEvent{Type: "stream_end", TurnID: turnID, Seq: int(h.streamSeq.Add(1)), Data: map[string]any{"stopReason": res.StopReason, "steps": res.Steps}})
+	h.broadcast(wsEvent{Type: "stream_end", TurnID: turnID, Seq: int(h.streamSeq.Add(1)), SessionID: sessionID, Data: map[string]any{"stopReason": res.StopReason, "steps": res.Steps}})
 	return res, nil
 }
 
@@ -330,6 +334,7 @@ func (s *Server) handleWS(w http.ResponseWriter, r *http.Request) {
 		claims: claims,
 	}
 	s.hub.register(c)
+	s.hub.replayHistory(c, history.MainSessionID)
 	go c.writePump()
 	go c.readPump()
 }