Paginate chat history so long sessions open quickly.

Load the latest 80 messages first and fetch older pages on scroll-up via before_id, instead of pulling thousands of rows up front.

Co-authored-by: Cursor <cursoragent@cursor.com>

interfaces/admin/chat_api.py

@@ -66,7 +66,8 @@ def _wiki_root_from_config() -> Path | None:
         root = (Path(__file__).resolve().parents[2] / root).resolve()
     return root
 
-_CHAT_MSG_LIMIT = 5000
+# Default window for chat UI / admin: recent messages only (scroll-up loads older via before_id).
+_CHAT_MSG_LIMIT = 80
 _SESSION_TITLE_MAX_LEN = 120
 _AVATAR_UPLOAD_MAX_BYTES = 2 * 1024 * 1024
 _AVATAR_MIMES = frozenset({"image/png", "image/jpeg", "image/jpg", "image/webp", "image/gif"})
@@ -1056,6 +1057,7 @@ def include_chat_routes(router: APIRouter, *, resolve_auth: Callable[[SqliteStor
     def api_chat_messages(
         session_id: str,
         limit: int = Query(default=_CHAT_MSG_LIMIT, ge=1, le=20000),
+        before_id: int | None = Query(default=None, ge=1),
         authorization: str | None = Header(default=None),
     ) -> dict[str, Any]:
         store = get_assistant_store()
@@ -1066,11 +1068,24 @@ def include_chat_routes(router: APIRouter, *, resolve_auth: Callable[[SqliteStor
         if not sess:
             raise HTTPException(status_code=404, detail="session_not_found")
         meta = store.get_session_messages_meta(session_id)
-        msgs = store.get_messages(session_id=session_id, limit=int(limit))
+        lim = int(limit)
+        if before_id is not None:
+            msgs = store.get_messages_before_id(
+                session_id=session_id, before_id=int(before_id), limit=lim
+            )
+        else:
+            msgs = store.get_messages(session_id=session_id, limit=lim)
         msgs = _filter_internal_instruction_user_messages(msgs)
+        oldest_id = int(getattr(msgs[0], "id", 0) or 0) if msgs else 0
+        has_more = bool(oldest_id) and store.exists_message_before_id(
+            session_id=session_id, before_id=oldest_id
+        )
         return {
             "ok": True,
-            "message_count": len(msgs),
+            "message_count": int(meta.message_count or 0),
+            "returned_count": len(msgs),
+            "has_more": has_more,
+            "before_id": int(before_id) if before_id is not None else None,
             "messages": [_serialize_message(m) for m in msgs],
         }
 

interfaces/admin/static/chat.html

@@ -781,7 +781,7 @@
     <script defer src="/admin/assets/admin-theme.js"></script>
     <script
       defer
-      src="/admin/assets/chat.js?v=20260811-vendor"
+      src="/admin/assets/chat.js?v=20260811-history-page"
       onerror="(function(){var s=document.getElementById('chat-boot-splash');if(s){s.querySelector('.chat-boot-splash__title span:last-child').textContent='无法加载 chat.js';s.querySelector('.chat-boot-splash__muted').textContent='请确认网关已启动且 /admin/assets/chat.js 可访问。';}})()"
     ></script>
     <!-- Mermaid is larger; load after chat.js so login is not delayed by parse. -->

interfaces/admin/static/chat.js

@@ -1,7 +1,8 @@
 /* Standalone /chat page: same bearer + /admin/api/chat as admin SPA. */
 
 const PAGE_SIZE = 35;
-const CHAT_MESSAGES_FETCH_LIMIT = 5000;
+/** Initial / page size for session history (older messages load on scroll-up). */
+const CHAT_MESSAGES_FETCH_LIMIT = 80;
 
 const I18N = {
   zh: {
@@ -73,7 +74,9 @@ const I18N = {
     "chat.marker.summary": "文件标记：指针 {p}，封套 {e}（封套内 {ep}）",
     "chat.marker.ttl": "TTL：turn {t} / session {s} / keep {k}",
     "chat.marker.reclaimed": "本轮回收 turn 标记 {n}",
-    "chat.historyTruncated": "仅显示最近 {shown} 条，共 {total} 条",
+    "chat.historyTruncated": "已显示最近 {shown}/{total} 条 · 上滑加载更早",
+    "chat.loadingOlder": "正在加载更早消息…",
+    "chat.historyComplete": "已到最早消息",
     "reasoning.summary": "模型推理片段",
     "reasoning.processNotes": "主通道过程说明",
     "tool.streamTitle": "工具与进度（本轮）",
@@ -263,7 +266,9 @@ const I18N = {
     "chat.marker.summary": "File markers: pointers {p}, envelope {e} (in-envelope {ep})",
     "chat.marker.ttl": "TTL: turn {t} / session {s} / keep {k}",
     "chat.marker.reclaimed": "Turn markers reclaimed {n}",
-    "chat.historyTruncated": "Showing last {shown} of {total} messages",
+    "chat.historyTruncated": "Showing latest {shown}/{total} · scroll up for older",
+    "chat.loadingOlder": "Loading older messages…",
+    "chat.historyComplete": "Reached the start of this chat",
     "reasoning.summary": "Model reasoning",
     "reasoning.processNotes": "Main-channel notes",
     "tool.streamTitle": "Tools & progress (this turn)",
@@ -3085,7 +3090,13 @@ async function appendMessageRow(messagesEl, m, options = {}) {
       else innerBubble.appendChild(bar);
     }
   }
+  if (options.prependBefore && options.prependBefore.parentNode === messagesEl) {
+    messagesEl.insertBefore(bubble, options.prependBefore);
+  } else if (options.prepend) {
+    messagesEl.insertBefore(bubble, messagesEl.firstChild);
+  } else {
     messagesEl.appendChild(bubble);
+  }
   hydrateMermaidIn(bubble);
 }
 
@@ -3414,6 +3425,9 @@ async function renderChatUi() {
   };
   messagesEl.addEventListener("scroll", () => {
     shouldFollowMessages = isNearBottom();
+    if (messagesEl.scrollTop <= 80) {
+      loadOlderMessagesForActive().catch(() => {});
+    }
   });
   messagesEl.addEventListener(
     "load",
@@ -4306,13 +4320,49 @@ async function renderChatUi() {
     },
   };
 
+  let historyHasMore = false;
+  let historyOldestId = 0;
+  let historyLoadedCount = 0;
+  let historyTotal = 0;
+  let historyLoadingOlder = false;
+  let historyCapEl = null;
+
+  const ensureHistoryCap = () => {
+    if (!historyHasMore && historyLoadedCount > 0 && historyTotal > 0 && historyLoadedCount >= historyTotal) {
+      if (historyCapEl && historyCapEl.parentNode) historyCapEl.remove();
+      historyCapEl = null;
+      return;
+    }
+    if (!historyHasMore && !historyLoadingOlder) {
+      if (historyCapEl && historyCapEl.parentNode) historyCapEl.remove();
+      historyCapEl = null;
+      return;
+    }
+    if (!historyCapEl) historyCapEl = el("div", { class: "muted chat-msg-cap" });
+    historyCapEl.textContent = historyLoadingOlder
+      ? t("chat.loadingOlder")
+      : t("chat.historyTruncated", {
+          shown: String(historyLoadedCount || 0),
+          total: String(historyTotal || historyLoadedCount || 0),
+        });
+    if (historyCapEl.parentNode !== messagesEl) {
+      messagesEl.insertBefore(historyCapEl, messagesEl.firstChild);
+    }
+  };
+
   const loadMessagesForActive = async (opts = {}) => {
     loadMessagesForActive._rid = (loadMessagesForActive._rid || 0) + 1;
     const rid = loadMessagesForActive._rid;
     shouldFollowMessages = true;
+    historyHasMore = false;
+    historyOldestId = 0;
+    historyLoadedCount = 0;
+    historyTotal = 0;
+    historyLoadingOlder = false;
     if (!activeId) {
       loadMessagesForActive._needsWsTextFallback = true;
       messagesEl.innerHTML = "";
+      historyCapEl = null;
       statusBar.textContent = sessions.length ? "" : t("chat.noSessions");
       if (!sessions.length) messagesEl.appendChild(el("div", { class: "muted", text: t("chat.empty") }));
       return 0;
@@ -4327,7 +4377,10 @@ async function renderChatUi() {
       const msgs = Array.isArray(resp.messages) ? resp.messages : [];
       const renderRows = _buildRenderRows(msgs);
       loadMessagesForActive._needsWsTextFallback = _needsWsTextFallbackFromRenderRows(renderRows);
-      const total = intOr(resp.message_count, msgs.length);
+      historyTotal = intOr(resp.message_count, msgs.length);
+      historyLoadedCount = msgs.length;
+      historyHasMore = !!resp.has_more;
+      historyOldestId = msgs.length ? intOr(msgs[0].id, 0) : 0;
       // End-of-turn hydrate: if the server returns nothing renderable yet (PG commit lag, transient API
       // glitch) but the caller still has a live stream worth keeping, do not wipe messagesEl — that
       // caused "stream flashes then entire dialog is empty" when reasoning/tool-output mode reloads.
@@ -4336,14 +4389,8 @@ async function renderChatUi() {
         return 0;
       }
       messagesEl.innerHTML = "";
-      if (total > msgs.length) {
-        messagesEl.appendChild(
-          el("div", {
-            class: "muted chat-msg-cap",
-            text: t("chat.historyTruncated", { shown: msgs.length, total }),
-          }),
-        );
-      }
+      historyCapEl = null;
+      ensureHistoryCap();
       if (!renderRows.length) {
         messagesEl.appendChild(el("div", { class: "muted", text: t("chat.empty") }));
       } else {
@@ -4363,6 +4410,49 @@ async function renderChatUi() {
     }
   };
 
+  const loadOlderMessagesForActive = async () => {
+    if (!activeId || !historyHasMore || historyLoadingOlder || !historyOldestId) return 0;
+    if (isChatStreaming()) return 0;
+    historyLoadingOlder = true;
+    ensureHistoryCap();
+    const prevHeight = messagesEl.scrollHeight;
+    const prevTop = messagesEl.scrollTop;
+    const rid = loadMessagesForActive._rid || 0;
+    try {
+      const resp = await apiGet(
+        `/admin/api/chat/sessions/${encodeURIComponent(activeId)}/messages?limit=${CHAT_MESSAGES_FETCH_LIMIT}&before_id=${encodeURIComponent(String(historyOldestId))}`,
+      );
+      if (rid !== loadMessagesForActive._rid) return 0;
+      const msgs = Array.isArray(resp.messages) ? resp.messages : [];
+      if (!msgs.length) {
+        historyHasMore = false;
+        ensureHistoryCap();
+        return 0;
+      }
+      const renderRows = _buildRenderRows(msgs);
+      historyTotal = intOr(resp.message_count, historyTotal);
+      historyLoadedCount += msgs.length;
+      historyHasMore = !!resp.has_more;
+      historyOldestId = intOr(msgs[0].id, historyOldestId);
+      ensureHistoryCap();
+      const anchor = historyCapEl && historyCapEl.parentNode === messagesEl ? historyCapEl.nextSibling : messagesEl.firstChild;
+      for (const m of renderRows) {
+        await appendMessageRow(messagesEl, m, { ...rowRenderOptions, prependBefore: anchor });
+      }
+      const delta = messagesEl.scrollHeight - prevHeight;
+      messagesEl.scrollTop = prevTop + delta;
+      return renderRows.length;
+    } catch (e) {
+      if (typeof showToast === "function") {
+        showToast(`${t("chat.error")}: ${String(e)}`, { kind: "error", ttlMs: 4500 });
+      }
+      return -1;
+    } finally {
+      historyLoadingOlder = false;
+      ensureHistoryCap();
+    }
+  };
+
   const reloadSessionsOnly = async () => {
     await refreshChannelPeerNameMap().catch(() => {});
     const resp = await apiGet(`/admin/api/chat/sessions?limit=${PAGE_SIZE}&offset=0`);

svc/persistence/assistant_store_protocol.py

@@ -81,6 +81,8 @@ class AssistantStoreProtocol(Protocol):
     def get_llm_profile_secret(self, profile_id: 'str') -> 'Optional[str]': ...
     def get_messages(self, session_id: 'str', limit: 'int' = 200) -> 'list[ChatMessage]': ...
     def get_messages_after_id(self, *, session_id: 'str', after_id: 'int', limit: 'int' = 200) -> 'list[ChatMessage]': ...
+    def get_messages_before_id(self, *, session_id: 'str', before_id: 'int', limit: 'int' = 200) -> 'list[ChatMessage]': ...
+    def exists_message_before_id(self, *, session_id: 'str', before_id: 'int') -> 'bool': ...
     def get_or_create_channel_session(self, *, tenant_id: 'str', channel: 'str', external_chat_id: 'str', external_user_id: 'str', session_title: 'str') -> 'str': ...
     def get_or_create_channel_session_v2(self, *, tenant_id: 'str', channel: 'str', account_id: 'str', external_chat_id: 'str', external_user_id: 'str', session_title: 'str') -> 'str': ...
     def get_secret(self, key: 'str') -> 'Optional[str]': ...

svc/persistence/sa_repos/chat_messages.py

@@ -237,6 +237,102 @@ class ChatMessagesSaRepository:
                 rows.insert(0, dict(arow))
         return [_row_to_chat_message(r) for r in rows]
 
+    def get_messages_before_id(
+        self, *, session_id: str, before_id: int, limit: int
+    ) -> list[ChatMessage]:
+        """Return up to ``limit`` messages with id < ``before_id``, ASC within the window.
+
+        Same orphan-tool prepend behavior as :meth:`get_messages_recent_asc` so a page
+        that starts mid-tool-turn still includes the parent assistant row.
+        """
+        if limit <= 0:
+            return []
+        sid = str(session_id or "").strip()
+        if not sid:
+            return []
+        bid = int(before_id or 0)
+        if bid <= 0:
+            return []
+        lim = max(1, min(int(limit), 2000))
+        ids_sq = (
+            select(chat_message.c.id)
+            .where(chat_message.c.session_id == sid, chat_message.c.id < bid)
+            .order_by(chat_message.c.id.desc())
+            .limit(lim)
+            .scalar_subquery()
+        )
+        stmt = (
+            select(
+                chat_message.c.id,
+                chat_message.c.session_id,
+                chat_message.c.role,
+                chat_message.c.content,
+                chat_message.c.tool_calls,
+                chat_message.c.attachments,
+                chat_message.c.turn_uuid,
+                chat_message.c.event_type,
+                chat_message.c.event_payload,
+                chat_message.c.timestamp,
+            )
+            .where(chat_message.c.session_id == sid, chat_message.c.id.in_(ids_sq))
+            .order_by(chat_message.c.id.asc())
+        )
+        prepended: set[int] = set()
+        with self._engine.connect() as conn:
+            rows: list[dict[str, Any]] = [dict(r) for r in conn.execute(stmt).mappings().all()]
+            while rows:
+                first = rows[0]
+                if str(first.get("role") or "") != "tool":
+                    break
+                aid = _tool_row_assistant_message_id(first.get("tool_calls"))
+                if aid is None:
+                    break
+                first_id = int(first["id"])
+                if aid >= first_id:
+                    break
+                if any(int(r["id"]) == int(aid) for r in rows):
+                    break
+                if int(aid) in prepended:
+                    break
+                arow = conn.execute(
+                    select(
+                        chat_message.c.id,
+                        chat_message.c.session_id,
+                        chat_message.c.role,
+                        chat_message.c.content,
+                        chat_message.c.tool_calls,
+                        chat_message.c.attachments,
+                        chat_message.c.turn_uuid,
+                        chat_message.c.event_type,
+                        chat_message.c.event_payload,
+                        chat_message.c.timestamp,
+                    )
+                    .where(chat_message.c.session_id == sid, chat_message.c.id == int(aid))
+                    .limit(1)
+                ).mappings().first()
+                if not arow:
+                    break
+                prepended.add(int(aid))
+                rows.insert(0, dict(arow))
+        return [_row_to_chat_message(r) for r in rows]
+
+    def exists_message_before_id(self, *, session_id: str, before_id: int) -> bool:
+        sid = str(session_id or "").strip()
+        bid = int(before_id or 0)
+        if not sid or bid <= 0:
+            return False
+        with self._engine.connect() as conn:
+            return bool(
+                conn.execute(
+                    select(
+                        exists().where(
+                            chat_message.c.session_id == sid,
+                            chat_message.c.id < bid,
+                        )
+                    )
+                ).scalar()
+            )
+
     def get_messages_after_id(
         self, *, session_id: str, after_id: int, limit: int
     ) -> list[ChatMessage]:

svc/persistence/sqlite_store.py

@@ -2282,6 +2282,26 @@ class SqliteStore(ScheduledJobStoreMixin):
             session_id=sid, after_id=aid, limit=lim
         )
 
+    def get_messages_before_id(self, *, session_id: str, before_id: int, limit: int = 200) -> list[ChatMessage]:
+        """Return up to ``limit`` messages with id < ``before_id``, ASC within the window."""
+        sid = str(session_id or "").strip()
+        if not sid:
+            return []
+        bid = int(before_id or 0)
+        if bid <= 0:
+            return []
+        lim = max(1, min(int(limit), 2000))
+        return self._chat_messages_repo().get_messages_before_id(
+            session_id=sid, before_id=bid, limit=lim
+        )
+
+    def exists_message_before_id(self, *, session_id: str, before_id: int) -> bool:
+        sid = str(session_id or "").strip()
+        bid = int(before_id or 0)
+        if not sid or bid <= 0:
+            return False
+        return self._chat_messages_repo().exists_message_before_id(session_id=sid, before_id=bid)
+
     def add_tool_log(
         self,
         session_id: str,

tests/test_sa_chat_messages.py

@@ -89,3 +89,19 @@ def test_sa_chat_message_tool_window_prepend(fresh_sqlite_store: SqliteStore) ->
     assert len(win) == 2
     assert win[0].id == asst.id
     assert win[1].role == "tool"
+
+
+def test_sa_chat_message_before_id_window(fresh_sqlite_store: SqliteStore) -> None:
+    s = fresh_sqlite_store
+    sess = s.create_session("P")
+    ids = []
+    for i in range(6):
+        ids.append(s.add_message(sess.id, "user", f"m{i}").id)
+    recent = s.get_messages(sess.id, limit=2)
+    assert [m.content for m in recent] == ["m4", "m5"]
+    older = s.get_messages_before_id(session_id=sess.id, before_id=recent[0].id, limit=2)
+    assert [m.content for m in older] == ["m2", "m3"]
+    assert s.exists_message_before_id(session_id=sess.id, before_id=older[0].id) is True
+    head = s.get_messages_before_id(session_id=sess.id, before_id=ids[1], limit=10)
+    assert [m.content for m in head] == ["m0"]
+    assert s.exists_message_before_id(session_id=sess.id, before_id=head[0].id) is False