Background: sustained large network transfers even when Codex/agent is not actively used

**提交人**: 张老板
**客户端版本**: 0.1.71

---

Symptom
----------
The desktop client, when left running in the background (no active conversation, no Codex feature invoked), keeps performing large sustained uploads and downloads.

On one Windows machine, over an idle ~13h stretch the client downloaded ~770 MB and uploaded ~34 MB. Over roughly one day the proxy-side cumulative reached ~5.2 GB down / ~1.2 GB up, including single-minute bursts of ~38-287 MB.

Local-proxy attribution
----------
Traffic is attributed to the bundled runtime/process and points at OpenAI hosts (chatgpt.com, chat.openai.com, ws.chatgpt.com, persistent.oaistatic.com). It is being routed through the user's proxy, so it deducts from a metered/limited proxy allowance.

Expected
----------
When the agent is idle (not proactively working, not running user tasks), background network use should be minimal (keep-alive / heartbeat only). Sustained many-MB-to-GB-scale transfers from a client the user is not actively operating is surprising and can silently exhaust a user's proxy/data quota.

Notes
----------
- Observed in cloud mode on Windows.
- No tokens, accounts, or personal paths are included.
- Happy to provide a longer sanitized log snippet if useful for triage.
---
## 提交时的任务环境

仅代表提交时快照,不一定是故障环境。OS 来自提交客户端本机,不含 SSH 远端主机;Harness / 模型来自当前任务。与运行环境无关的反馈可忽略本段。
**OS**: win32 x64 (10.0.19045)
**Harness**: Claude Code
**Model ID**: ` ark-code-latest `
**界面语言**: zh-CN

Start by reproducing the reported idle behavior in the Windows cloud-mode desktop client and inspect the bundled runtime/process traffic to OpenAI hosts. Use the supplied duration and transfer measurements as the baseline; done means identifying the source of the sustained transfers and verifying that idle use is limited to expected keep-alive or heartbeat traffic.

Written by the indexing model from the issue text.