Problem

This PR addresses four issues:

1. Cross-channel WeChat → Telegram: Telegram send failed: Timed out
2. Cross-channel Telegram → WeChat: Timeout context manager should be used inside a task
3. Session continuity on restart: Every gateway restart loses the active conversation
4. Proxy auto-detection: macOS system proxy doesn't detect SOCKS proxies

Root Causes & Fixes

1. _send_telegram() missing proxy support (send_message_tool.py)

The send_message tool's _send_telegram() created Bot(token=token) without proxy configuration, bypassing TELEGRAM_PROXY. The gateway adapter correctly uses the proxy, but the tool's direct send path didn't.

Fix: Added resolve_proxy_url("TELEGRAM_PROXY") + HTTPXRequest(proxy=...), matching the gateway adapter's pattern.

2. send_weixin_direct() event loop mismatch (weixin.py)

When send_weixin_direct() is called via _run_async() from a different async context (e.g. cross-channel Telegram→WeChat), _run_async spawns a new thread with asyncio.run() creating a fresh event loop. The code tries to reuse a live_adapter from _LIVE_ADAPTERS, but that adapter's aiohttp session is bound to the gateway's main event loop. Using it from the new loop causes aiohttp's timeout context manager to fail.

Fix: Added event loop compatibility check. If the live adapter's aiohttp session connector loop doesn't match the current running loop, skip adapter reuse and fall through to the existing fresh-session fallback path.

3. edit_message() missing finalize kwarg (telegram.py)

The gateway calls TelegramAdapter.edit_message(..., finalize=...) but the method signature didn't accept this keyword argument, causing TypeError.

Fix: Added finalize: bool = False parameter.

4. Session continuity on restart (gateway/run.py)

When gateway restart is called, _drain_active_agents() waits up to 60s for running agents to finish. The current agent can't self-terminate mid-response, so drain times out. The code then skips writing the .clean_shutdown marker, causing the next startup to call suspend_recently_active() — suspending all sessions and losing the conversation.

Fix: Always write the .clean_shutdown marker, even on drain timeout. The interrupt signal has been sent and agents handle it cleanly on their next iteration. The existing 3-restart stuck-loop detection (#7536) already suspends truly stuck sessions as a safety net. Losing every conversation on restart is far worse than resuming an interrupted one.

5. SOCKS proxy auto-detection (gateway/platforms/base.py)

_detect_macos_system_proxy() only checked HTTP/HTTPS proxy settings from scutil --proxy, ignoring SOCKS proxies. Many users (especially behind GFW) use SOCKS5 proxies (V2rayU, Clash, Shadowrocket).

Fix: Added SOCKS proxy detection (SOCKSEnable/SOCKSProxy/SOCKSPort) returning socks5:// URLs. Added explanatory comment for why http:// scheme is used for HTTPS proxies (HTTP CONNECT tunnelling). Env vars (ALL_PROXY, TELEGRAM_PROXY etc.) remain the recommended way to configure any proxy scheme.

Testing

• ✅ WeChat → Telegram: message delivered
• ✅ Telegram → WeChat: message delivered
• ✅ Gateway restart + message continues same session
• ✅ No new errors in gateway logs