fix: terminal hang on exit and MCP connection leak

[thecannabisapp]

问题描述 / Issue Description

在使用较长会话后，执行 /exit 或输入 exit 时，终端会无响应并挂起，需强制关闭终端窗口。同时，退出后系统防火墙会报告来自已终止进程的孤立 MCP 连接。

After extended sessions, executing /exit or typing exit causes the terminal to hang and become unresponsive, requiring the terminal window to be force-closed. Additionally, the system firewall reports orphaned MCP connections from terminated processes after exit.

根本原因 / Root Cause

1. TTY 阻塞导致退出挂起 / TTY Block Causes Exit Hang

_cursor_position_unix() 在 tty.setcbreak(fd) 之后调用 os.read(fd, 32)。虽然上游代码已引入 select.select() 超时，但在以下情况 os.read() 仍可能阻塞：

• asyncio 任务取消信号无法中断 os.read() 系统调用
• prompt_toolkit 的 stdin reader 与 _cursor_position_unix() 竞争读取 CPR 响应时，数据可能在 select 返回后到 os.read 执行前被消耗

_cursor_position_unix() calls os.read(fd, 32) after tty.setcbreak(fd). Although upstream code already introduced a select.select() timeout, os.read() can still block because:

• asyncio task cancellation cannot interrupt the os.read() system call
• prompt_toolkit's stdin reader races with _cursor_position_unix() for the CPR response; data may be consumed between select returning and os.read executing

2. TTY 恢复不完整 / Incomplete TTY Restoration

ensure_tty_sane() 恢复 c_lflag 中的 ISIG | IEXTEN | ICANON | ECHO，但未重置 c_cc 中的 VMIN/VTIME。若进程在 cbreak 模式下崩溃或挂起，终端会保持在 VMIN=1, VTIME=0 状态，导致 Ctrl+C 发送字节 0x03 而非 SIGINT。

ensure_tty_sane() restores ISIG | IEXTEN | ICANON | ECHO in c_lflag, but does not reset VMIN/VTIME in c_cc. If the process crashes or hangs in cbreak mode, the terminal remains in VMIN=1, VTIME=0 state, causing Ctrl+C to send byte 0x03 instead of SIGINT.

3. MCP 客户端泄漏 / MCP Client Leak

KimiToolset.cleanup()（负责关闭所有 fastmcp.Client 连接）在 KimiCLI.shutdown_background_tasks() 或 shell 拆卸路径中从未被调用。这导致 stdio/WebSocket 传输在 CLI 进程退出后仍保持连接。

KimiToolset.cleanup(), which closes all fastmcp.Client connections, is never invoked in KimiCLI.shutdown_background_tasks() or the shell teardown path. This leaves stdio/WebSocket transports connected after the CLI process exits.

修复方案 / Proposed Fix

1. 在 _cursor_position_unix() 中使用 os.set_blocking(fd, False) 使读取非阻塞，并在 finally 中恢复

2. 在 ensure_tty_sane() 中重置 VMIN=1, VTIME=0

3. 在 shutdown_background_tasks() 中调用 KimiToolset.cleanup()

4. Use os.set_blocking(fd, False) in _cursor_position_unix() to make reads non-blocking, restoring it in finally

5. Reset VMIN=1, VTIME=0 in ensure_tty_sane()

6. Call KimiToolset.cleanup() in shutdown_background_tasks()

环境 / Environment

• macOS 15.7.3 (Ghostty 1.3.1)
• Python 3.13.7
• kimi-cli 1.37.0
• prompt-toolkit 3.0.52

[thecannabisapp]

请维护者将此 issue 的类型设置为 Bug，谢谢！

Could a maintainer please set the issue type to Bug? Thank you!