优化页面交互并增强 Provider 配置能力

- 表单下拉框不再用 label 包裹，改为 div + aria-label，避免点击后
  下拉面板瞬间关闭
- Provider 卡片新增「从上游获取模型」和「修改」入口，支持编辑
  base URL、API Key 与 api_mode，凭据写入对应的环境变量
- 服务端抽出 provider-model-discovery 统一不同 api_mode 的模型发现，
  旧的 custom_providers 与新的 providers 字典结构都能正确更新
- Agent Bridge 改为后台启动，HTTP 接口在 Python 运行时就绪前即可用；
  worker 连接被拒时按类型重试，不重放可能已产生副作用的请求
- 断线重连后恢复 profiles、models 与会话列表，健康检查改为自调度
  的 setTimeout，离线时 2 秒后重试
- 批量删除确认文案改用 Popconfirm content，避免变成侧栏触发器文本

docs/chat-chain-changes/2026-10-22-http-startup-recovery.md

@@ -0,0 +1,26 @@
+---
+date: 2026-10-22
+feature: HTTP startup and session reload after backend restart
+impact: Keep session APIs available during bridge startup and recover failed initial loads
+---
+
+PR/commit: pending.
+
+The HTTP listener, Web UI stores, routes and Socket.IO handlers are initialized
+before starting the managed Agent Bridge in the background. Bridge readiness
+continues to be checked by the existing chat handlers; the chat-run protocol is
+unchanged. The bridge manager is registered with shutdown before starting it.
+
+A failed session-list request no longer marks the list loaded or discards a
+route-selected session. Offline health polling retries after two seconds;
+reconnection reloads profiles, models and sessions. An already loaded session
+list retains its state on request failure. Chat, global-agent and history views
+handle reconnection. Batch-delete confirmation text uses Popconfirm content
+rather than becoming part of the sidebar trigger.
+
+Validation: app-store and session-order Vitest coverage; Playwright startup
+recovery (base route and direct session route), deletion confirmation and history
+deep links; build and harness checks. A real isolated built-server smoke check
+used a fake Python executable sleeping for 60 seconds and confirmed /health
+responded with status=ok before the bridge became ready. State, uploads and
+bridge endpoint were isolated under a temporary directory.

docs/chat-chain-changes/2026-10-22-worker-connect-readiness.md

@@ -0,0 +1,21 @@
+---
+date: 2026-10-22
+feature: Agent Bridge worker connection readiness
+impact: Recover refused worker connections before returning a chat error
+---
+
+# Worker connection readiness recovery
+
+- Date: 2026-10-22
+- PR/commit: pending
+- Feature: Agent Bridge client, cold and idle worker connection recovery.
+- Behavior: Retry broker responses explicitly typed as worker connection refusal
+  or missing socket, within the existing connect retry budget. Keep the original
+  payload (including session and conversation history). The broker's own ping
+  does not establish worker readiness. Previously only Node-to-broker connection
+  failures were retried; Python-to-worker connect failures escaped to chat.
+- Safety: Do not replay untyped agent errors, timeouts, closed responses, or
+  resets, because the worker might already have accepted the request. Vendored
+  Python and the chat-run event contract remain unchanged.
+- Validation: focused worker readiness and existing chat/bridge tests; build and
+  harness checks.