feat: 任务详情 LLM 进度与失败原因标识（列表失败原因摘要）

cloud/console-react/src/sections/TaskDetailPanel.tsx

@@ -23,6 +23,29 @@ import { CoordPanel, AgentsPanel, ApprovalPanel, PeerPanel, ThreadPanel } from "
 //   只是没被挂进页签）。
 // Data, actions and sub-sections are unchanged; only grouping
 // and visibility changed.
+// LLM failure summary: the raw error_text is usually a codex JSON
+// blob (e.g. a 503 from the relay). The overview must say, at a
+// glance, whether the model ever ran and why it stopped — previously
+// the only copy lived in the "操作与结果" <pre>, so a failed task
+// looked like "failed" with no reason. This is a display mapping
+// only; it never claims a connection state the server does not
+// report (there is no "LLM connected" field — running progress
+// below is elapsed time + last activity, not a live socket state).
+function errorSummary(err: string): string {
+  if (/503|Service Unavailable/.test(err)) return "上游模型服务暂时不可用（503）：模型一轮都没跑出来，不是任务逻辑失败；等上游恢复或换模型重试";
+  if (/429|rate limit/i.test(err)) return "上游限流（429）：模型拒绝了本次调用，可稍后重试";
+  if (/timeout|timed out/i.test(err)) return "上游调用超时：模型响应未在限时内返回";
+  if (/401|403|Invalid API key|unauthorized/i.test(err)) return "上游鉴权失败：网关 key 或权限有问题";
+  if (/ECONNREFUSED|ENOTFOUND|network|fetch failed/i.test(err)) return "连不上上游模型服务：网络或网关地址有问题";
+  return "任务失败（原因见原始错误文本）";
+}
+
+function fmtAge(sec: number | null): string {
+  if (sec === null || sec < 0) return "未知";
+  if (sec < 60) return Math.floor(sec) + " 秒前";
+  if (sec < 3600) return Math.floor(sec / 60) + " 分钟前";
+  return (sec / 3600).toFixed(1) + " 小时前";
+}
 type DetailTab = "overview" | "collab" | "review" | "threads" | "browser";
 const DETAIL_TABS: Array<[DetailTab, string]> = [
   ["overview", "概览"], ["collab", "协作"], ["review", "评审与审批"], ["threads", "线程"], ["browser", "浏览器"],
@@ -82,6 +105,22 @@ export default function TaskDetailPanel() {
           </div>
         </div>
       </Card>
+      {(t.status === "running" || t.status === "queued") ? (() => {
+        const now = Date.now() / 1000;
+        const last = t.last_heartbeat || t.updated_at || null;
+        const idle = last ? now - last : null;
+        return (<Card title="LLM 进度">
+          <div>状态 {t.status === "running" ? "进行中" : "排队中"} · 已用 {dur}{idle !== null ? " · 上次活动 " + fmtAge(idle) : ""}</div>
+          {idle !== null && idle > 180
+            ? <div className="muted">超过 3 分钟无活动：可能在等上游模型响应（上游 503 时 worker 会重试后失败，失败后这里会显示失败原因；无活动不等于已断连，服务端没有“LLM 已连接”字段，不作此断言）</div>
+            : <div className="muted">首轮无输出时属正常：模型第一轮回复到达前 trace 里没有消息；回复到达后会在会话区出现，成本里的 LLM 请求数随之增加</div>}
+        </Card>);
+      })() : null}
+      {t.error_text ? (<Card title="失败原因">
+        <div>{errorSummary(t.error_text)}</div>
+        <pre style={{ whiteSpace: "pre-wrap", overflowWrap: "anywhere" }}>{t.error_text.slice(0, 1000)}</pre>
+        <div className="muted">完整错误文本仍在下方“操作与结果”里；模型 {t.model || "默认"} · executor {t.executor || ""}</div>
+      </Card>) : null}
       <Card title="基本信息">
         <dl className="detail-kv">
           <div><dt>任务</dt><dd>{t.id}</dd></div>

cloud/console-react/src/sections/TaskListSection.tsx

@@ -72,6 +72,9 @@ export default function TaskListSection({ onNew }: { onNew: () => void }) {
                     {" · "}<span title={durTitle}>耗时 {dur}</span>
                     {" · "}<span title={tokTitle}>{tok}</span>
                   </div>
+                  {(t.status === "failed" || t.status === "interrupted" || t.status === "merge-failed") && t.error_text ? (
+                    <div className="issue-meta" title={t.error_text}>{/503|Service Unavailable/.test(t.error_text) ? "上游 503：模型服务暂时不可用" : t.error_text.replace(/\s+/g, " ").slice(0, 60)}</div>
+                  ) : null}
                 </span>
               </li>
               );