chore: update ji local changes

frontend-web/src/components/workspace/recent-chat-list.tsx

@@ -213,7 +213,7 @@ export function RecentChatList() {
                         className="block w-full whitespace-nowrap group-hover/side-menu-item:overflow-hidden"
                         href={pathOfThread(thread)}
                       >
-                        {desensitize(titleOfThread(thread))}
+                        {desensitize(titleOfThread(thread, t.pages.newChat))}
                       </Link>
                     </SidebarMenuButton>
                     {env.NEXT_PUBLIC_STATIC_WEBSITE_ONLY !== "true" && (
@@ -236,7 +236,7 @@ export function RecentChatList() {
                             onSelect={() =>
                               handleRenameClick(
                                 thread.thread_id,
-                                titleOfThread(thread),
+                                titleOfThread(thread, t.pages.newChat),
                               )
                             }
                           >

frontend-web/src/core/auth/api.ts

@@ -64,24 +64,50 @@ export async function loginByUsername(
   return response.json() as Promise<LoginByUsernameResponse>;
 }
 
+// The upstream token issuer and the verifier (token_info_url) run on separate
+// servers whose clocks can drift ~3-4s apart. A token verified right after it
+// is issued can look "not yet valid / expired" to the verifier, which surfaces
+// as a 401 here. That window closes on its own once real time advances past the
+// skew, so we silently retry the exchange a couple of times before giving up —
+// the LoginPage keeps showing its loading state, so the user never notices.
+const TOKEN_LOGIN_MAX_ATTEMPTS = 3; // initial try + 2 retries → covers up to ~6s skew
+const TOKEN_LOGIN_RETRY_DELAY_MS = 3000;
+
+const wait = (ms: number) => new Promise<void>((resolve) => setTimeout(resolve, ms));
+
 export async function loginByToken(
   token: string,
 ): Promise<LoginByUsernameResponse> {
   // Token-exchange login: the gateway resolves the upstream token to a username
   // and signs the matching local account in.
-  const response = await apiFetch(
-    `${getBackendBaseURL()}/api/v1/auth/login/token?token=${encodeURIComponent(token)}`,
-    {
+  const url = `${getBackendBaseURL()}/api/v1/auth/login/token?token=${encodeURIComponent(token)}`;
+  let lastError = "Login failed.";
+
+  for (let attempt = 1; attempt <= TOKEN_LOGIN_MAX_ATTEMPTS; attempt += 1) {
+    const response = await apiFetch(url, {
       method: "POST",
       headers: {
         "Content-Type": "application/json",
       },
-    },
-  );
-  if (!response.ok) {
+    });
+    if (response.ok) {
+      return response.json() as Promise<LoginByUsernameResponse>;
+    }
+
     const error = await response.json().catch(() => null);
-    throw new Error(extractErrorMessage(error));
+    lastError = extractErrorMessage(error);
+
+    // Only 401 is plausibly the transient clock-skew case and worth waiting on.
+    // 403 (disabled), 429 (rate-limited), 500 (misconfigured) won't self-heal,
+    // so fail fast instead of stalling the user.
+    const isTransient = response.status === 401;
+    if (!isTransient || attempt === TOKEN_LOGIN_MAX_ATTEMPTS) {
+      throw new Error(lastError);
+    }
+    await wait(TOKEN_LOGIN_RETRY_DELAY_MS);
   }
-  return response.json() as Promise<LoginByUsernameResponse>;
+
+  // Unreachable: the loop either returns or throws above.
+  throw new Error(lastError);
 }
 

frontend-web/src/core/threads/utils.ts

@@ -62,6 +62,21 @@ export function textOfMessage(message: Message) {
   return null;
 }
 
-export function titleOfThread(thread: AgentThread) {
-  return thread.values?.title ?? "Untitled";
+export function titleOfThread(thread: AgentThread, fallback = "Untitled") {
+  const title = thread.values?.title;
+  if (title) return title;
+
+  // No title yet (e.g. the very first turn is still streaming, or the backend
+  // title pass hasn't finished). Derive a sensible display title from the first
+  // user message so the user never sees a bare "Untitled" placeholder.
+  const messages = thread.values?.messages;
+  if (Array.isArray(messages)) {
+    const firstHuman = messages.find((m) => m.type === "human");
+    const text = firstHuman ? textOfMessage(firstHuman)?.trim() : null;
+    if (text) {
+      return text.length > 50 ? `${text.slice(0, 50).trimEnd()}…` : text;
+    }
+  }
+
+  return fallback;
 }

frontend-web/src/pages/ChatsPage.tsx

@@ -71,7 +71,7 @@ export default function ChatsPage() {
                   <Link key={thread.thread_id} href={pathOfThread(thread)}>
                     <div className="flex flex-col gap-2 border-b p-4">
                       <div>
-                        <div>{desensitize(titleOfThread(thread))}</div>
+                        <div>{desensitize(titleOfThread(thread, t.pages.newChat))}</div>
                       </div>
                       {thread.updated_at && (
                         <div className="text-muted-foreground text-sm">

offline-backend-20260512/backend/CLAUDE.md

@@ -166,7 +166,7 @@ Lead-agent middlewares are assembled in strict append order across `packages/har
 9. **SummarizationMiddleware** - Context reduction when approaching token limits (optional, if enabled)
 10. **TodoListMiddleware** - Task tracking with `write_todos` tool (optional, if plan_mode)
 11. **TokenUsageMiddleware** - Records token usage metrics when token tracking is enabled (optional)
-12. **TitleMiddleware** - Auto-generates thread title after first complete exchange and normalizes structured message content before prompting the title model
+12. **TitleMiddleware** - Two-phase auto-title (never shows "untitled"): `before_model` sets an instant question-derived *provisional* title the moment the first user message arrives (no LLM, zero latency), then `after_model` asynchronously generates the final LLM title and replaces it (tracked via the `title_provisional` state flag). Normalizes structured message content before prompting the title model
 13. **MemoryMiddleware** - Memory V2: `wrap_model_call` injects Hindsight recall transiently; `after_agent` auto-retains the turn to Hindsight
 14. **ViewImageMiddleware** - Injects base64 image data before LLM call (conditional on vision support)
 15. **DeferredToolFilterMiddleware** - Hides deferred tool schemas from the bound model until tool search is enabled (optional)
@@ -178,7 +178,7 @@ Lead-agent middlewares are assembled in strict append order across `packages/har
 10. **SummarizationMiddleware** - Context reduction when approaching token limits (optional, if enabled)
 11. **TodoListMiddleware** - Task tracking with `write_todos` tool (optional, if plan_mode)
 12. **TokenUsageMiddleware** - Records token usage metrics when token tracking is enabled (optional)
-13. **TitleMiddleware** - Auto-generates thread title after first complete exchange and normalizes structured message content before prompting the title model
+13. **TitleMiddleware** - Two-phase auto-title (never shows "untitled"): `before_model` sets an instant question-derived *provisional* title the moment the first user message arrives (no LLM, zero latency), then `after_model` asynchronously generates the final LLM title and replaces it (tracked via the `title_provisional` state flag). Normalizes structured message content before prompting the title model
 14. **MemoryMiddleware** - Queues conversations for async memory update (filters to user + final AI responses)
 15. **ViewImageMiddleware** - Injects base64 image data before LLM call (conditional on vision support)
 16. **DeferredToolFilterMiddleware** - Hides deferred tool schemas from the bound model until tool search is enabled (optional)

offline-backend-20260512/backend/packages/harness/deerflow/agents/middlewares/title_middleware.py

@@ -23,6 +23,7 @@ class TitleMiddlewareState(AgentState):
     """Compatible with the `ThreadState` schema."""
 
     title: NotRequired[str | None]
+    title_provisional: NotRequired[bool]
 
 
 class TitleMiddleware(AgentMiddleware[TitleMiddlewareState]):
@@ -61,16 +62,37 @@ class TitleMiddleware(AgentMiddleware[TitleMiddlewareState]):
 
         return ""
 
-    def _should_generate_title(self, state: TitleMiddlewareState) -> bool:
-        """Check if we should generate a title for this thread."""
+    def _should_set_provisional_title(self, state: TitleMiddlewareState) -> bool:
+        """Check if we should set an instant, question-derived placeholder title.
+
+        Runs in ``before_model`` so the thread shows a sensible title the moment
+        the user sends their first message — eliminating the "untitled" window
+        while the (possibly long) first answer streams. No LLM call, no latency.
+        """
         config = self._get_title_config()
         if not config.enabled:
             return False
 
-        # Check if thread already has a title in state
+        # Already has a title (provisional or final) — nothing to do.
         if state.get("title"):
             return False
 
+        # Only on the very first turn, as soon as the user message arrives.
+        messages = state.get("messages", [])
+        user_messages = [m for m in messages if m.type == "human"]
+        return len(user_messages) == 1
+
+    def _should_generate_title(self, state: TitleMiddlewareState) -> bool:
+        """Check if we should generate the final LLM title for this thread."""
+        config = self._get_title_config()
+        if not config.enabled:
+            return False
+
+        # Skip only if we already have a *finalized* title. A provisional
+        # placeholder set in before_model should still be upgraded by the LLM.
+        if state.get("title") and not state.get("title_provisional"):
+            return False
+
         # Check if this is the first turn (has at least one user message and one assistant response)
         messages = state.get("messages", [])
         if len(messages) < 2:
@@ -146,13 +168,27 @@ class TitleMiddleware(AgentMiddleware[TitleMiddlewareState]):
         config["tags"] = [*(config.get("tags") or []), "middleware:title"]
         return config
 
+    def _provisional_title_result(self, state: TitleMiddlewareState) -> dict | None:
+        """Set an instant placeholder title from the first user message (no LLM)."""
+        if not self._should_set_provisional_title(state):
+            return None
+
+        messages = state.get("messages", [])
+        user_msg_content = next((m.content for m in messages if m.type == "human"), "")
+        user_msg = self._normalize_content(user_msg_content)
+        if not user_msg.strip():
+            return None
+        title = self._fallback_title(user_msg)
+        logger.debug("TitleMiddleware: set provisional title %r before first model call", title)
+        return {"title": title, "title_provisional": True}
+
     def _generate_title_result(self, state: TitleMiddlewareState) -> dict | None:
         """Generate a local fallback title without blocking on an LLM call."""
         if not self._should_generate_title(state):
             return None
 
         _, user_msg = self._build_title_prompt(state)
-        return {"title": self._fallback_title(user_msg)}
+        return {"title": self._fallback_title(user_msg), "title_provisional": False}
 
     def _force_disable_thinking(self, model: object) -> None:
         """Force-disable thinking on the title model via `extra_body` overrides.
@@ -204,10 +240,18 @@ class TitleMiddleware(AgentMiddleware[TitleMiddlewareState]):
             response = await model.ainvoke(prompt, config=self._get_runnable_config())
             title = self._parse_title(response.content)
             if title:
-                return {"title": title}
+                return {"title": title, "title_provisional": False}
         except Exception:
             logger.debug("Failed to generate async title; falling back to local title", exc_info=True)
-        return {"title": self._fallback_title(user_msg)}
+        return {"title": self._fallback_title(user_msg), "title_provisional": False}
+
+    @override
+    def before_model(self, state: TitleMiddlewareState, runtime: Runtime) -> dict | None:
+        return self._provisional_title_result(state)
+
+    @override
+    async def abefore_model(self, state: TitleMiddlewareState, runtime: Runtime) -> dict | None:
+        return self._provisional_title_result(state)
 
     @override
     def after_model(self, state: TitleMiddlewareState, runtime: Runtime) -> dict | None:

offline-backend-20260512/backend/packages/harness/deerflow/agents/thread_state.py

@@ -69,6 +69,10 @@ class ThreadState(AgentState):
     sandbox: NotRequired[SandboxState | None]
     thread_data: NotRequired[ThreadDataState | None]
     title: NotRequired[str | None]
+    # True while `title` is a provisional, question-derived placeholder that the
+    # async LLM pass has not yet finalized. Lets TitleMiddleware show a title
+    # instantly (no "untitled" window) and upgrade it once the answer is ready.
+    title_provisional: NotRequired[bool]
     artifacts: Annotated[list[str], merge_artifacts]
     todos: NotRequired[list | None]
     uploaded_files: NotRequired[list[dict] | None]

offline-backend-20260512/backend/packages/harness/deerflow/runtime/scheduler/html_page.py

@@ -128,6 +128,45 @@ def extract_json_object(text: str | None) -> dict[str, Any] | None:
     return None
 
 
+def recover_html_document(text: str | None) -> str | None:
+    """Slice a complete HTML document out of a model reply.
+
+    Used as a fallback when the JSON wrapper around the ``html`` field can't be
+    parsed — the model commonly emits *literal* newlines and *unescaped* quotes
+    (``"zh-CN"``) inside the value, which is invalid JSON, so
+    :func:`extract_json_object` returns ``None``. Without this, the caller would
+    archive the whole ``{"title": ..., "html": "<!doctype html>…"}`` blob as the
+    page. We locate the ``<!doctype html>`` / ``<html …>`` start and the final
+    ``</html>`` and return just that span, dropping the surrounding JSON
+    scaffolding. Also accepts a reply that is already bare HTML. Returns
+    ``None`` when no HTML document can be found.
+    """
+    if not text or not isinstance(text, str):
+        return None
+    lowered = text.lower()
+    start = lowered.find("<!doctype html")
+    if start == -1:
+        start = lowered.find("<html")
+    if start == -1:
+        return None
+    end = lowered.rfind("</html>")
+    candidate = text[start:] if end == -1 else text[start : end + len("</html>")]
+    candidate = candidate.strip()
+    # If the html had been a properly-escaped JSON string value (the JSON failed
+    # to parse for some *other* reason), unescape the common sequences so they
+    # don't survive into the rendered page. Only triggers when escapes exist, so
+    # genuinely-literal HTML (the usual broken-JSON case) is left untouched.
+    if "\\n" in candidate or '\\"' in candidate or "\\/" in candidate or "\\t" in candidate:
+        candidate = (
+            candidate.replace('\\"', '"')
+            .replace("\\n", "\n")
+            .replace("\\r", "\r")
+            .replace("\\t", "\t")
+            .replace("\\/", "/")
+        )
+    return candidate or None
+
+
 def sanitize_html(html: str) -> tuple[str, list[str]]:
     """Remove script-execution vectors from ``html``.
 
@@ -242,11 +281,7 @@ _GENERATION_TEMPLATE = """你是一个内网页面生成助手。请根据输入
 
 用户页面要求：
 {page_prompt}
-
-限制条件：
-{negative_prompt}
-
-本次数据：
+{constraint_block}本次数据：
 {data}
 
 参考页面风格：
@@ -258,11 +293,12 @@ _GENERATION_TEMPLATE = """你是一个内网页面生成助手。请根据输入
 要求：
 1. 只输出 JSON，不要输出 Markdown，不要输出多余解释。
 2. JSON 字段必须包含 title、html、style_summary、layout_summary。
-3. html 必须是完整 HTML 文档，包含 <!doctype html>、html、head、body。
-4. 所有 CSS 写在 <style> 标签内，不要引用外部 CDN 或外链样式表。
-5. 不要包含 <script>、外链脚本、iframe、表单提交、on* 事件属性。
-6. 参考页面只用于风格和布局，页面内容必须来自“本次数据”，不要复用参考页旧数据。
-7. 页面使用中文，排版适合在浏览器中直接预览。
+3. html 字段的值必须是合法 JSON 字符串：其中所有换行必须写成 \\n、所有双引号必须转义为 \\"，不要在字符串里直接换行或使用未转义的引号。
+4. html 必须是完整 HTML 文档，包含 <!doctype html>、html、head、body。
+5. 所有 CSS 写在 <style> 标签内，不要引用外部 CDN 或外链样式表。
+6. 不要包含 <script>、外链脚本、iframe、表单提交、on* 事件属性。
+7. 参考页面只用于风格和布局，页面内容必须来自“本次数据”，不要复用参考页旧数据。
+8. 页面使用中文，排版适合在浏览器中直接预览。
 """
 
 _REPAIR_TEMPLATE = """你是一个 HTML 修复助手。下面的 HTML 存在问题，请修复后重新输出完整页面。
@@ -312,14 +348,27 @@ def build_generation_prompt(
     reference_layout: str | None,
     skill_instructions: str | None = None,
     data_limit: int = 16000,
-    skill_limit: int = 16000,
+    skill_limit: int = 30000,
 ) -> str:
     skill_block = ""
     if skill_instructions and skill_instructions.strip():
         skill_block = "\n可用页面生成技能规范（请严格遵循）：\n" + _clip(skill_instructions.strip(), skill_limit) + "\n"
+    # The constraint block is only emitted when the user actually supplied one,
+    # and is framed as a top-priority hard rule that overrides the page
+    # requirements / reference page / data on conflict — otherwise the model
+    # tends to treat "限制条件" as a soft hint and ignore it.
+    constraint_block = ""
+    neg = (negative_prompt or "").strip()
+    if neg:
+        constraint_block = (
+            "\n限制条件（最高优先级的硬性约束，必须逐条严格遵守；"
+            "若与上面的页面要求、参考页面或本次数据发生冲突，一律以这里的限制条件为准）：\n"
+            + neg
+            + "\n\n"
+        )
     return _GENERATION_TEMPLATE.format(
         page_prompt=(page_prompt or "（未提供，使用通用看板风格）").strip(),
-        negative_prompt=(negative_prompt or "无").strip(),
+        constraint_block=constraint_block,
         data=_clip(data, data_limit),
         reference_style=(reference_style or "无").strip(),
         reference_layout=(reference_layout or "无").strip(),

offline-backend-20260512/backend/packages/harness/deerflow/runtime/scheduler/service.py

@@ -966,6 +966,7 @@ class ScheduledTaskService:
         """
         cfg = self._html_config(task)
         page_prompt = str(cfg.get("page_prompt") or "").strip()
+        negative_prompt = str(cfg.get("negative_prompt") or "").strip()
         gather_instruction = (
             "请运用你自身的检索/搜索等技能，围绕下述页面需求获取最新、真实的资料，"
             "并整理成结构化的中文要点（可用列表/小标题），供后续生成 HTML 页面使用，本步骤不要生成 HTML：\n"
@@ -976,6 +977,11 @@ class ScheduledTaskService:
         base_prompt = str(task.get("prompt") or "").strip()
         if base_prompt:
             parts.append(f"补充说明：{base_prompt}")
+        # Apply the user's constraints to data gathering too, not just rendering
+        # — a "限制条件" like "只要最近 24 小时的新闻" or "排除某主题" must shape
+        # which sources/data the agent collects, otherwise it silently leaks in.
+        if negative_prompt:
+            parts.append(f"限制条件（最高优先级，必须严格遵守）：{negative_prompt}")
         gather_task = {**task, "prompt": "\n".join(parts)}
         # Never apply the markdown gate to the gather pass.
         gather_ctx = dict(task.get("execution_context") or {})
@@ -1100,11 +1106,14 @@ class ScheduledTaskService:
         title = str(parsed.get("title") or task.get("name") or "生成页面").strip()
         html = parsed.get("html")
         if not isinstance(html, str) or "<" not in html:
-            # GLM5 may have replied with bare HTML instead of JSON — accept it.
-            if isinstance(raw, str) and ("<html" in raw.lower() or "<!doctype" in raw.lower()):
-                html = raw
-            else:
-                raise ValueError("GLM5 未返回有效的 HTML(JSON 解析失败)")
+            # The JSON wrapper couldn't be parsed (the model emitted literal
+            # newlines / unescaped quotes inside the "html" value), or it replied
+            # with bare HTML. Either way, recover just the HTML document so we
+            # never archive the {"title":...,"html":"..."} JSON blob as the page.
+            recovered = html_page_utils.recover_html_document(raw)
+            if recovered is None:
+                raise ValueError("模型未返回有效的 HTML(JSON 解析失败)")
+            html = recovered
         style_summary = str(parsed.get("style_summary") or "").strip() or None
         layout_summary = str(parsed.get("layout_summary") or "").strip() or None