bug: CLI agent crashes on Chinese input containing spaces / CLI Agent 在中文含空格输入时崩溃 #2984

Summary

The CLI interactive agent (zeroclaw agent) crashes with Error reading input: stream did not contain valid UTF-8 when the user types Chinese (CJK) text that contains ASCII spaces. Input without spaces works fine.

概要

CLI 交互模式（zeroclaw agent）在用户输入包含空格的中文文本时崩溃，报错 Error reading input: stream did not contain valid UTF-8。不含空格的中文输入正常工作。

Reproduction / 复现步骤

# Enter the CLI agent (e.g. via kubectl exec or direct terminal)
zeroclaw agent

# Input WITHOUT spaces — works fine ✅
> 今天调度系统有多少失败的任务？

# Input WITH spaces — crashes ❌
> 今天调度系统有 多少失败的 任务？
Error reading input: stream did not contain valid UTF-8
# (agent exits immediately)

Root Cause / 根因分析

The interactive loop in src/agent/loop_.rs uses std::io::stdin().read_line(), which requires the entire input to be valid UTF-8. If any byte sequence is not valid UTF-8, it returns an error and the loop breaks (exits).

交互循环使用了 std::io::stdin().read_line()，该方法要求输入必须是合法 UTF-8。任何无效字节序列都会导致错误并退出循环。

When running through a PTY chain (e.g. kubectl exec -it, SSH, or remote terminals), the transport layer may split data frames at space (0x20) boundaries. If a multi-byte UTF-8 character (Chinese characters are 3 bytes each in UTF-8) is being transmitted and the frame boundary happens to interrupt the byte sequence, Rust's BufRead implementation detects incomplete UTF-8 and returns an InvalidData error immediately, rather than waiting for the remaining bytes.

当通过 PTY 链路运行时（如 kubectl exec -it、SSH 或远程终端），传输层可能在空格(0x20)边界处拆分数据帧。如果多字节 UTF-8 字符（中文字符在 UTF-8 中为 3 字节）正在传输时被帧边界切断，Rust 的 BufRead 实现会检测到不完整的 UTF-8 并立即返回 InvalidData 错误，而不是等待后续字节到达。

This explains why:

• No spaces → Chinese bytes are contiguous, no frame split mid-character → works
• With spaces → PTY may split at space boundary, possibly mid-character → crash

这解释了为什么：

• 无空格 → 中文字节连续，不会在字符中间拆帧 → 正常
• 有空格 → PTY 可能在空格边界拆帧，可能切断字符 → 崩溃

Affected Code / 受影响代码

// src/agent/loop_.rs (line ~3057)
let mut input = String::new();
match std::io::stdin().read_line(&mut input) {
    Ok(0) => break,
    Ok(_) => {}
    Err(e) => {
        eprintln!("\nError reading input: {e}\n");
        break; // ← exits the entire agent loop
    }
}

Proposed Fix / 修复方案

Replace stdin().read_line() with byte-level stdin().lock().read_until(b'\n') followed by String::from_utf8_lossy(). This reads raw bytes until newline without UTF-8 validation, then does lossy conversion (replacing any truly invalid bytes with U+FFFD instead of crashing).

将 stdin().read_line() 替换为字节级的 stdin().lock().read_until(b'\n')，然后使用 String::from_utf8_lossy() 转换。这会读取原始字节直到换行符，不进行 UTF-8 校验，然后做有损转换（将真正无效的字节替换为 U+FFFD 而非崩溃）。

Additionally, set ENV LANG=C.UTF-8 in the Dockerfile as defense-in-depth.

同时在 Dockerfile 中设置 ENV LANG=C.UTF-8 作为纵深防御。

Environment / 环境

• ZeroClaw v0.1.7+ (current master)
• Running in Docker container (Debian trixie-slim, user nobody)
• Accessed via kubectl exec -it (PTY over WebSocket)
• Also reproducible via SSH to containers
• Any CJK input with embedded ASCII spaces triggers the issue